540 (29.51 % of base) - System.Runtime.Intrinsics.Vector512:Dot[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):ubyte
; Assembly listing for method System.Runtime.Intrinsics.Vector512:Dot[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):ubyte (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
-; partially interruptible
+; fully interruptible
; No PGO data
-; 0 inlinees with PGO data; 38 single block inlinees; 47 inlinees without PGO data
+; 0 inlinees with PGO data; 42 single block inlinees; 24 inlinees without PGO data
; Final local variable assignments
;
;* V00 arg0 [V00 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[ubyte]>
;* V01 arg1 [V01 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[ubyte]>
-;# V02 OutArgs [V02 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V03 tmp1 [V03,T24] ( 2, 4 ) int -> rax "impAppendStmt"
-; V04 tmp2 [V04,T47] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V05 tmp3 [V05,T48] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V06 tmp4 [V06,T25] ( 2, 4 ) int -> rax "impAppendStmt"
-; V07 tmp5 [V07,T49] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V08 tmp6 [V08,T50] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V09 tmp7 [V09,T26] ( 2, 4 ) int -> rax "impAppendStmt"
-; V10 tmp8 [V10,T08] ( 8, 8 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
-;* V11 tmp9 [V11,T39] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V12 tmp10 [V12 ] ( 9, 18 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V13 tmp11 [V13 ] ( 9, 18 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V14 tmp12 [V14 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V15 tmp13 [V15,T00] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V16 tmp14 [V16 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V17 tmp15 [V17 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V18 tmp16 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V19 tmp17 [V19 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V20 tmp18 [V20 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V21 tmp19 [V21 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V22 tmp20 [V22,T16] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V23 tmp21 [V23 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V24 tmp22 [V24 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V25 tmp23 [V25 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V26 tmp24 [V26 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V27 tmp25 [V27 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+; V02 OutArgs [V02 ] ( 1, 1 ) struct (16) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V03 tmp1 [V03,T20] ( 2, 4 ) int -> rbx "impAppendStmt"
+; V04 tmp2 [V04,T55] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V05 tmp3 [V05,T56] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V06 tmp4 [V06,T21] ( 2, 4 ) int -> rbx "impAppendStmt"
+; V07 tmp5 [V07,T57] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V08 tmp6 [V08,T58] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V09 tmp7 [V09,T22] ( 2, 4 ) int -> rsi "impAppendStmt"
+;* V10 tmp8 [V10 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V11 tmp9 [V11 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V12 tmp10 [V12 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V13 tmp11 [V13 ] ( 2, 5 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V14 tmp12 [V14,T12] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
+; V15 tmp13 [V15 ] ( 2, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V16 tmp14 [V16 ] ( 2, 10 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V17 tmp15 [V17 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V18 tmp16 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+;* V19 tmp17 [V19 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V20 tmp18 [V20 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V21 tmp19 [V21 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V22 tmp20 [V22,T04] ( 16, 16 ) ubyte -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V23 tmp21 [V23,T47] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V24 tmp22 [V24 ] ( 9, 18 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V25 tmp23 [V25 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V26 tmp24 [V26 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V27 tmp25 [V27 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V28 tmp26 [V28 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V29 tmp27 [V29,T09] ( 8, 8 ) ubyte -> rcx ld-addr-op "Inline ldloca(s) first use temp"
-;* V30 tmp28 [V30,T40] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V31 tmp29 [V31 ] ( 9, 18 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V32 tmp30 [V32 ] ( 9, 18 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V33 tmp31 [V33 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V34 tmp32 [V34,T01] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V35 tmp33 [V35 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V36 tmp34 [V36 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V37 tmp35 [V37 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V38 tmp36 [V38 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V39 tmp37 [V39 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V40 tmp38 [V40 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V41 tmp39 [V41,T17] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V42 tmp40 [V42 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V43 tmp41 [V43 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V44 tmp42 [V44 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V45 tmp43 [V45 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V46 tmp44 [V46 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V29 tmp27 [V29 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V30 tmp28 [V30 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V31 tmp29 [V31 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V32 tmp30 [V32 ] ( 2, 5 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V33 tmp31 [V33,T13] ( 5, 17 ) int -> rdi "Inline stloc first use temp"
+; V34 tmp32 [V34 ] ( 2, 10 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V35 tmp33 [V35 ] ( 2, 10 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V36 tmp34 [V36 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V37 tmp35 [V37 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+;* V38 tmp36 [V38 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V39 tmp37 [V39 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V40 tmp38 [V40 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V41 tmp39 [V41,T05] ( 16, 16 ) ubyte -> rdi ld-addr-op "Inline ldloca(s) first use temp"
+;* V42 tmp40 [V42,T48] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V43 tmp41 [V43 ] ( 9, 18 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V44 tmp42 [V44 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V45 tmp43 [V45 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V46 tmp44 [V46 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V47 tmp45 [V47 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V48 tmp46 [V48 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V49 tmp47 [V49 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V50 tmp48 [V50 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V51 tmp49 [V51,T51] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V52 tmp50 [V52,T52] ( 3, 6 ) simd16 -> [rbp-0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V53 tmp51 [V53,T27] ( 2, 4 ) int -> rcx "impAppendStmt"
-; V54 tmp52 [V54,T10] ( 8, 8 ) ubyte -> rcx ld-addr-op "Inline ldloca(s) first use temp"
-;* V55 tmp53 [V55,T41] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V56 tmp54 [V56 ] ( 9, 18 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V57 tmp55 [V57 ] ( 9, 18 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V58 tmp56 [V58 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V59 tmp57 [V59,T02] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V60 tmp58 [V60 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V61 tmp59 [V61 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V62 tmp60 [V62 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V63 tmp61 [V63 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V64 tmp62 [V64 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V65 tmp63 [V65 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V66 tmp64 [V66,T18] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V67 tmp65 [V67 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V68 tmp66 [V68 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V69 tmp67 [V69 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V70 tmp68 [V70 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V71 tmp69 [V71 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+; V51 tmp49 [V51,T59] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V52 tmp50 [V52,T60] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V53 tmp51 [V53,T23] ( 2, 4 ) int -> r15 "impAppendStmt"
+;* V54 tmp52 [V54 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V55 tmp53 [V55 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V56 tmp54 [V56 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V57 tmp55 [V57 ] ( 2, 5 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V58 tmp56 [V58,T14] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
+; V59 tmp57 [V59 ] ( 2, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V60 tmp58 [V60 ] ( 2, 10 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V61 tmp59 [V61 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V62 tmp60 [V62 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+;* V63 tmp61 [V63 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V64 tmp62 [V64 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V65 tmp63 [V65 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V66 tmp64 [V66,T06] ( 16, 16 ) ubyte -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V67 tmp65 [V67,T49] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V68 tmp66 [V68 ] ( 9, 18 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V69 tmp67 [V69 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V70 tmp68 [V70 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V71 tmp69 [V71 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V72 tmp70 [V72 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V73 tmp71 [V73,T11] ( 8, 8 ) ubyte -> rdx ld-addr-op "Inline ldloca(s) first use temp"
-;* V74 tmp72 [V74,T42] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V75 tmp73 [V75 ] ( 9, 18 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V76 tmp74 [V76 ] ( 9, 18 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V77 tmp75 [V77 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V78 tmp76 [V78,T03] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V79 tmp77 [V79 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V80 tmp78 [V80 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V81 tmp79 [V81 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V82 tmp80 [V82 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V83 tmp81 [V83 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V84 tmp82 [V84 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V85 tmp83 [V85,T19] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V86 tmp84 [V86 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V87 tmp85 [V87 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V88 tmp86 [V88 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V89 tmp87 [V89 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V90 tmp88 [V90 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V91 tmp89 [V91 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V92 tmp90 [V92 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V93 tmp91 [V93 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V94 tmp92 [V94 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V95 tmp93 [V95 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V96 tmp94 [V96 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V97 tmp95 [V97,T53] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V98 tmp96 [V98,T54] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V99 tmp97 [V99,T28] ( 2, 4 ) int -> rcx "impAppendStmt"
-; V100 tmp98 [V100,T55] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V101 tmp99 [V101,T56] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V102 tmp100 [V102,T29] ( 2, 4 ) int -> rcx "impAppendStmt"
-; V103 tmp101 [V103,T12] ( 8, 8 ) ubyte -> rcx ld-addr-op "Inline ldloca(s) first use temp"
-;* V104 tmp102 [V104,T43] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V105 tmp103 [V105 ] ( 9, 18 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V106 tmp104 [V106 ] ( 9, 18 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V73 tmp71 [V73 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V74 tmp72 [V74 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V75 tmp73 [V75 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V76 tmp74 [V76 ] ( 2, 5 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V77 tmp75 [V77,T15] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
+; V78 tmp76 [V78 ] ( 2, 10 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V79 tmp77 [V79 ] ( 2, 10 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V80 tmp78 [V80 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V81 tmp79 [V81 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+;* V82 tmp80 [V82 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V83 tmp81 [V83 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V84 tmp82 [V84 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V85 tmp83 [V85,T07] ( 16, 16 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
+;* V86 tmp84 [V86,T50] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V87 tmp85 [V87 ] ( 9, 18 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V88 tmp86 [V88 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+; V89 tmp87 [V89,T61] ( 3, 6 ) simd32 -> [rbp-0x110] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V90 tmp88 [V90,T62] ( 3, 6 ) simd32 -> [rbp-0x130] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V91 tmp89 [V91,T24] ( 2, 4 ) int -> r15 "impAppendStmt"
+; V92 tmp90 [V92,T63] ( 3, 6 ) simd16 -> [rbp-0x140] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V93 tmp91 [V93,T64] ( 3, 6 ) simd16 -> [rbp-0x150] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V94 tmp92 [V94,T25] ( 2, 4 ) int -> r15 "impAppendStmt"
+;* V95 tmp93 [V95 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V96 tmp94 [V96 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V97 tmp95 [V97 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V98 tmp96 [V98 ] ( 2, 5 ) struct ( 8) [rbp-0x158] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V99 tmp97 [V99,T01] ( 7, 25 ) int -> r15 "Inline stloc first use temp"
+; V100 tmp98 [V100 ] ( 2, 10 ) struct ( 8) [rbp-0x160] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V101 tmp99 [V101 ] ( 2, 10 ) struct ( 8) [rbp-0x168] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V102 tmp100 [V102 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V103 tmp101 [V103 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V104 tmp102 [V104,T08] ( 16, 16 ) ubyte -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V105 tmp103 [V105,T51] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V106 tmp104 [V106 ] ( 9, 18 ) struct ( 8) [rbp-0x170] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V107 tmp105 [V107 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V108 tmp106 [V108,T04] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V109 tmp107 [V109 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V110 tmp108 [V110 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V111 tmp109 [V111 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V112 tmp110 [V112 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V113 tmp111 [V113 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V114 tmp112 [V114 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V115 tmp113 [V115,T20] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V116 tmp114 [V116 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V117 tmp115 [V117 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V118 tmp116 [V118 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V119 tmp117 [V119 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V120 tmp118 [V120 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V121 tmp119 [V121 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V122 tmp120 [V122,T13] ( 8, 8 ) ubyte -> rdx ld-addr-op "Inline ldloca(s) first use temp"
-;* V123 tmp121 [V123,T44] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V124 tmp122 [V124 ] ( 9, 18 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V125 tmp123 [V125 ] ( 9, 18 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V126 tmp124 [V126 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V127 tmp125 [V127,T05] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V128 tmp126 [V128 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V129 tmp127 [V129 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V130 tmp128 [V130 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V131 tmp129 [V131 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V132 tmp130 [V132 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V133 tmp131 [V133 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V134 tmp132 [V134,T21] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V135 tmp133 [V135 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V136 tmp134 [V136 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V137 tmp135 [V137 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V138 tmp136 [V138 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V139 tmp137 [V139 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V140 tmp138 [V140 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V141 tmp139 [V141 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V142 tmp140 [V142 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V143 tmp141 [V143 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V144 tmp142 [V144,T57] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V145 tmp143 [V145,T58] ( 3, 6 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V146 tmp144 [V146,T30] ( 2, 4 ) int -> rdx "impAppendStmt"
-; V147 tmp145 [V147,T14] ( 8, 8 ) ubyte -> rdx ld-addr-op "Inline ldloca(s) first use temp"
-;* V148 tmp146 [V148,T45] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V149 tmp147 [V149 ] ( 9, 18 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V150 tmp148 [V150 ] ( 9, 18 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V151 tmp149 [V151 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V152 tmp150 [V152,T06] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V153 tmp151 [V153 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V154 tmp152 [V154 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V155 tmp153 [V155 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V156 tmp154 [V156 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V157 tmp155 [V157 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V158 tmp156 [V158 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V159 tmp157 [V159,T22] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V160 tmp158 [V160 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V161 tmp159 [V161 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V162 tmp160 [V162 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V163 tmp161 [V163 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V164 tmp162 [V164 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V165 tmp163 [V165 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V166 tmp164 [V166,T15] ( 8, 8 ) ubyte -> rdi ld-addr-op "Inline ldloca(s) first use temp"
-;* V167 tmp165 [V167,T46] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V168 tmp166 [V168 ] ( 9, 18 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V169 tmp167 [V169 ] ( 9, 18 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V170 tmp168 [V170 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V171 tmp169 [V171,T07] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V172 tmp170 [V172 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V173 tmp171 [V173 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V174 tmp172 [V174 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V175 tmp173 [V175 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V176 tmp174 [V176 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V177 tmp175 [V177 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V178 tmp176 [V178,T23] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V179 tmp177 [V179 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V180 tmp178 [V180 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V181 tmp179 [V181 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V182 tmp180 [V182 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V183 tmp181 [V183 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V184 tmp182 [V184 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V185 tmp183 [V185 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V186 tmp184 [V186 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V187 tmp185 [V187 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V188 tmp186 [V188 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V189 tmp187 [V189 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V190 tmp188 [V190 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V191 tmp189 [V191 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V192 tmp190 [V192,T59] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V00._lower (fldOffset=0x0)" P-INDEP
-; V193 tmp191 [V193,T60] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V00._upper (fldOffset=0x20)" P-INDEP
-; V194 tmp192 [V194,T61] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
-; V195 tmp193 [V195,T62] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
-; V196 tmp194 [V196 ] ( 9, 17 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
-; V197 tmp195 [V197 ] ( 9, 17 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
-; V198 tmp196 [V198 ] ( 9, 17 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V199 tmp197 [V199 ] ( 9, 17 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V32._00 (fldOffset=0x0)" P-DEP
-; V200 tmp198 [V200 ] ( 9, 17 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V56._00 (fldOffset=0x0)" P-DEP
-; V201 tmp199 [V201 ] ( 9, 17 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V57._00 (fldOffset=0x0)" P-DEP
-; V202 tmp200 [V202 ] ( 9, 17 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V75._00 (fldOffset=0x0)" P-DEP
-; V203 tmp201 [V203 ] ( 9, 17 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V76._00 (fldOffset=0x0)" P-DEP
-; V204 tmp202 [V204 ] ( 9, 17 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V105._00 (fldOffset=0x0)" P-DEP
-; V205 tmp203 [V205 ] ( 9, 17 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V106._00 (fldOffset=0x0)" P-DEP
-; V206 tmp204 [V206 ] ( 9, 17 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V124._00 (fldOffset=0x0)" P-DEP
-; V207 tmp205 [V207 ] ( 9, 17 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V125._00 (fldOffset=0x0)" P-DEP
-; V208 tmp206 [V208 ] ( 9, 17 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V149._00 (fldOffset=0x0)" P-DEP
-; V209 tmp207 [V209 ] ( 9, 17 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V150._00 (fldOffset=0x0)" P-DEP
-; V210 tmp208 [V210 ] ( 9, 17 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V168._00 (fldOffset=0x0)" P-DEP
-; V211 tmp209 [V211 ] ( 9, 17 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V169._00 (fldOffset=0x0)" P-DEP
-; V212 cse0 [V212,T31] ( 2, 2 ) int -> rcx "CSE #01: conservative"
-; V213 cse1 [V213,T32] ( 2, 2 ) int -> rdx "CSE #02: conservative"
-; V214 cse2 [V214,T33] ( 2, 2 ) int -> rdx "CSE #03: conservative"
-; V215 cse3 [V215,T34] ( 2, 2 ) int -> rdi "CSE #04: conservative"
-; V216 cse4 [V216,T35] ( 2, 2 ) int -> rdx "CSE #05: conservative"
-; V217 cse5 [V217,T36] ( 2, 2 ) int -> rdi "CSE #06: conservative"
-; V218 cse6 [V218,T37] ( 2, 2 ) int -> rdi "CSE #07: conservative"
-; V219 cse7 [V219,T38] ( 2, 2 ) int -> rsi "CSE #08: conservative"
+;* V108 tmp106 [V108 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V109 tmp107 [V109 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V110 tmp108 [V110 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V111 tmp109 [V111 ] ( 2, 5 ) struct ( 8) [rbp-0x178] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V112 tmp110 [V112,T02] ( 7, 25 ) int -> r14 "Inline stloc first use temp"
+; V113 tmp111 [V113 ] ( 2, 10 ) struct ( 8) [rbp-0x180] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V114 tmp112 [V114 ] ( 2, 10 ) struct ( 8) [rbp-0x188] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V115 tmp113 [V115 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V116 tmp114 [V116 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V117 tmp115 [V117,T09] ( 16, 16 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
+;* V118 tmp116 [V118,T52] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V119 tmp117 [V119 ] ( 9, 18 ) struct ( 8) [rbp-0x190] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V120 tmp118 [V120 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+; V121 tmp119 [V121,T65] ( 3, 6 ) simd16 -> [rbp-0x1A0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V122 tmp120 [V122,T66] ( 3, 6 ) simd16 -> [rbp-0x1B0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V123 tmp121 [V123,T26] ( 2, 4 ) int -> r14 "impAppendStmt"
+;* V124 tmp122 [V124 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V125 tmp123 [V125 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V126 tmp124 [V126 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V127 tmp125 [V127 ] ( 2, 5 ) struct ( 8) [rbp-0x1B8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V128 tmp126 [V128,T03] ( 7, 25 ) int -> r14 "Inline stloc first use temp"
+; V129 tmp127 [V129 ] ( 2, 10 ) struct ( 8) [rbp-0x1C0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V130 tmp128 [V130 ] ( 2, 10 ) struct ( 8) [rbp-0x1C8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V131 tmp129 [V131 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V132 tmp130 [V132 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V133 tmp131 [V133,T10] ( 16, 16 ) ubyte -> r14 ld-addr-op "Inline ldloca(s) first use temp"
+;* V134 tmp132 [V134,T53] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V135 tmp133 [V135 ] ( 9, 18 ) struct ( 8) [rbp-0x1D0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V136 tmp134 [V136 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V137 tmp135 [V137 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V138 tmp136 [V138 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V139 tmp137 [V139 ] ( 2, 4 ) struct ( 8) [rbp-0x1D8] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V140 tmp138 [V140,T11] ( 16, 16 ) ubyte -> r13 ld-addr-op "Inline ldloca(s) first use temp"
+;* V141 tmp139 [V141,T54] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V142 tmp140 [V142 ] ( 9, 18 ) struct ( 8) [rbp-0x1E0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V143 tmp141 [V143 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V144 tmp142 [V144 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V145 tmp143 [V145 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V146 tmp144 [V146 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V147 tmp145 [V147,T67] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V00._lower (fldOffset=0x0)" P-INDEP
+; V148 tmp146 [V148,T68] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V00._upper (fldOffset=0x20)" P-INDEP
+; V149 tmp147 [V149,T69] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
+; V150 tmp148 [V150,T70] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
+; V151 tmp149 [V151,T30] ( 2, 2 ) long -> rsi "field V10._00 (fldOffset=0x0)" P-INDEP
+; V152 tmp150 [V152,T31] ( 2, 2 ) long -> rdi "field V11._00 (fldOffset=0x0)" P-INDEP
+;* V153 tmp151 [V153 ] ( 0, 0 ) long -> zero-ref "field V12._00 (fldOffset=0x0)" P-INDEP
+; V154 tmp152 [V154 ] ( 2, 5 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V155 tmp153 [V155 ] ( 2, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V156 tmp154 [V156 ] ( 2, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V157 tmp155 [V157 ] ( 9, 17 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V24._00 (fldOffset=0x0)" P-DEP
+; V158 tmp156 [V158,T32] ( 2, 2 ) long -> rdi "field V29._00 (fldOffset=0x0)" P-INDEP
+; V159 tmp157 [V159,T33] ( 2, 2 ) long -> rax "field V30._00 (fldOffset=0x0)" P-INDEP
+;* V160 tmp158 [V160 ] ( 0, 0 ) long -> zero-ref "field V31._00 (fldOffset=0x0)" P-INDEP
+; V161 tmp159 [V161 ] ( 2, 5 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V32._00 (fldOffset=0x0)" P-DEP
+; V162 tmp160 [V162 ] ( 2, 9 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V34._00 (fldOffset=0x0)" P-DEP
+; V163 tmp161 [V163 ] ( 2, 9 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V164 tmp162 [V164 ] ( 9, 17 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V43._00 (fldOffset=0x0)" P-DEP
+; V165 tmp163 [V165,T34] ( 2, 2 ) long -> rsi "field V54._00 (fldOffset=0x0)" P-INDEP
+; V166 tmp164 [V166,T35] ( 2, 2 ) long -> rdi "field V55._00 (fldOffset=0x0)" P-INDEP
+;* V167 tmp165 [V167 ] ( 0, 0 ) long -> zero-ref "field V56._00 (fldOffset=0x0)" P-INDEP
+; V168 tmp166 [V168 ] ( 2, 5 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V57._00 (fldOffset=0x0)" P-DEP
+; V169 tmp167 [V169 ] ( 2, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V59._00 (fldOffset=0x0)" P-DEP
+; V170 tmp168 [V170 ] ( 2, 9 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V60._00 (fldOffset=0x0)" P-DEP
+; V171 tmp169 [V171 ] ( 9, 17 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V68._00 (fldOffset=0x0)" P-DEP
+; V172 tmp170 [V172,T36] ( 2, 2 ) long -> rsi "field V73._00 (fldOffset=0x0)" P-INDEP
+; V173 tmp171 [V173,T37] ( 2, 2 ) long -> rdi "field V74._00 (fldOffset=0x0)" P-INDEP
+;* V174 tmp172 [V174 ] ( 0, 0 ) long -> zero-ref "field V75._00 (fldOffset=0x0)" P-INDEP
+; V175 tmp173 [V175 ] ( 2, 5 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V76._00 (fldOffset=0x0)" P-DEP
+; V176 tmp174 [V176 ] ( 2, 9 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V78._00 (fldOffset=0x0)" P-DEP
+; V177 tmp175 [V177 ] ( 2, 9 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V79._00 (fldOffset=0x0)" P-DEP
+; V178 tmp176 [V178 ] ( 9, 17 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V87._00 (fldOffset=0x0)" P-DEP
+; V179 tmp177 [V179,T38] ( 2, 2 ) long -> rdi "field V95._00 (fldOffset=0x0)" P-INDEP
+; V180 tmp178 [V180,T39] ( 2, 2 ) long -> rsi "field V96._00 (fldOffset=0x0)" P-INDEP
+;* V181 tmp179 [V181 ] ( 0, 0 ) long -> zero-ref "field V97._00 (fldOffset=0x0)" P-INDEP
+; V182 tmp180 [V182 ] ( 2, 5 ) long -> [rbp-0x158] do-not-enreg[X] addr-exposed "field V98._00 (fldOffset=0x0)" P-DEP
+; V183 tmp181 [V183 ] ( 2, 9 ) long -> [rbp-0x160] do-not-enreg[X] addr-exposed "field V100._00 (fldOffset=0x0)" P-DEP
+; V184 tmp182 [V184 ] ( 2, 9 ) long -> [rbp-0x168] do-not-enreg[X] addr-exposed "field V101._00 (fldOffset=0x0)" P-DEP
+; V185 tmp183 [V185 ] ( 9, 17 ) long -> [rbp-0x170] do-not-enreg[X] addr-exposed "field V106._00 (fldOffset=0x0)" P-DEP
+; V186 tmp184 [V186,T40] ( 2, 2 ) long -> rdi "field V108._00 (fldOffset=0x0)" P-INDEP
+; V187 tmp185 [V187,T41] ( 2, 2 ) long -> rsi "field V109._00 (fldOffset=0x0)" P-INDEP
+;* V188 tmp186 [V188 ] ( 0, 0 ) long -> zero-ref "field V110._00 (fldOffset=0x0)" P-INDEP
+; V189 tmp187 [V189 ] ( 2, 5 ) long -> [rbp-0x178] do-not-enreg[X] addr-exposed "field V111._00 (fldOffset=0x0)" P-DEP
+; V190 tmp188 [V190 ] ( 2, 9 ) long -> [rbp-0x180] do-not-enreg[X] addr-exposed "field V113._00 (fldOffset=0x0)" P-DEP
+; V191 tmp189 [V191 ] ( 2, 9 ) long -> [rbp-0x188] do-not-enreg[X] addr-exposed "field V114._00 (fldOffset=0x0)" P-DEP
+; V192 tmp190 [V192 ] ( 9, 17 ) long -> [rbp-0x190] do-not-enreg[X] addr-exposed "field V119._00 (fldOffset=0x0)" P-DEP
+; V193 tmp191 [V193,T42] ( 2, 2 ) long -> rdi "field V124._00 (fldOffset=0x0)" P-INDEP
+; V194 tmp192 [V194,T43] ( 2, 2 ) long -> rsi "field V125._00 (fldOffset=0x0)" P-INDEP
+;* V195 tmp193 [V195 ] ( 0, 0 ) long -> zero-ref "field V126._00 (fldOffset=0x0)" P-INDEP
+; V196 tmp194 [V196 ] ( 2, 5 ) long -> [rbp-0x1B8] do-not-enreg[X] addr-exposed "field V127._00 (fldOffset=0x0)" P-DEP
+; V197 tmp195 [V197 ] ( 2, 9 ) long -> [rbp-0x1C0] do-not-enreg[X] addr-exposed "field V129._00 (fldOffset=0x0)" P-DEP
+; V198 tmp196 [V198 ] ( 2, 9 ) long -> [rbp-0x1C8] do-not-enreg[X] addr-exposed "field V130._00 (fldOffset=0x0)" P-DEP
+; V199 tmp197 [V199 ] ( 9, 17 ) long -> [rbp-0x1D0] do-not-enreg[X] addr-exposed "field V135._00 (fldOffset=0x0)" P-DEP
+;* V200 tmp198 [V200 ] ( 0, 0 ) long -> zero-ref "field V137._00 (fldOffset=0x0)" P-INDEP
+;* V201 tmp199 [V201 ] ( 0, 0 ) long -> zero-ref "field V138._00 (fldOffset=0x0)" P-INDEP
+; V202 tmp200 [V202,T29] ( 2, 3 ) long -> [rbp-0x1D8] do-not-enreg[H] hidden-struct-arg "field V139._00 (fldOffset=0x0)" P-DEP
+; V203 tmp201 [V203 ] ( 9, 17 ) long -> [rbp-0x1E0] do-not-enreg[X] addr-exposed "field V142._00 (fldOffset=0x0)" P-DEP
+; V204 tmp202 [V204,T27] ( 2, 4 ) int -> rsi "argument with side effect"
+; V205 tmp203 [V205,T00] ( 16, 32 ) int -> rax "argument with side effect"
+; V206 tmp204 [V206,T28] ( 2, 4 ) int -> rsi "argument with side effect"
+; V207 cse0 [V207,T44] ( 2, 2 ) int -> rsi "CSE #02: conservative"
+; V208 cse1 [V208,T45] ( 2, 2 ) int -> rdi "CSE #04: conservative"
+; V209 cse2 [V209,T46] ( 2, 2 ) int -> rsi "CSE #06: conservative"
+; V210 cse3 [V210,T16] ( 4, 16 ) long -> rax "CSE #01: moderate"
+; V211 cse4 [V211,T17] ( 4, 16 ) long -> rcx "CSE #03: moderate"
+; V212 cse5 [V212,T18] ( 4, 16 ) long -> rax "CSE #05: moderate"
+; V213 cse6 [V213,T19] ( 4, 16 ) long -> rax "CSE #07: moderate"
;
-; Lcl frame size = 256
+; Lcl frame size = 464
G_M29434_IG01:
push rbp
- sub rsp, 256
- lea rbp, [rsp+0x100]
- ;; size=16 bbWeight=1 PerfScore 1.75
+ push r15
+ push r14
+ push r13
+ push rbx
+ sub rsp, 464
+ lea rbp, [rsp+0x1F0]
+ ;; size=23 bbWeight=1 PerfScore 5.75
G_M29434_IG02:
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x50]
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x30], xmm2
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x28], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x30], rax
- movzx rax, byte ptr [rbp-0x28]
- movzx rcx, byte ptr [rbp-0x30]
- imul eax, ecx
- movzx rax, al
- movzx rcx, byte ptr [rbp-0x27]
- movzx rdx, byte ptr [rbp-0x2F]
- imul ecx, edx
- movzx rcx, cl
- add eax, ecx
- movzx rcx, al
- movzx rax, cl
- movzx rcx, byte ptr [rbp-0x26]
- movzx rdx, byte ptr [rbp-0x2E]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x25]
- movzx rdi, byte ptr [rbp-0x2D]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add eax, ecx
- movzx rax, al
- movzx rdx, byte ptr [rbp-0x24]
- movzx rcx, byte ptr [rbp-0x2C]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x23]
- movzx rdi, byte ptr [rbp-0x2B]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add ecx, eax
- movzx rax, cl
- movzx rdx, byte ptr [rbp-0x22]
- movzx rcx, byte ptr [rbp-0x2A]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x21]
- movzx rdi, byte ptr [rbp-0x29]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add ecx, eax
- movzx rax, cl
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x38], rcx
- mov rcx, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x40], rcx
- movzx rcx, byte ptr [rbp-0x38]
- movzx rdx, byte ptr [rbp-0x40]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x37]
- movzx rdi, byte ptr [rbp-0x3F]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- movzx rdx, cl
- movzx rcx, dl
- movzx rdx, byte ptr [rbp-0x36]
- movzx rdi, byte ptr [rbp-0x3E]
- ;; size=250 bbWeight=1 PerfScore 66.75
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ mov rsi, qword ptr [rbp-0x30]
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rsi
+ mov qword ptr [rbp-0x58], rdi
+ xor esi, esi
+ align [0 bytes for IG03]
+ ;; size=46 bbWeight=1 PerfScore 14.75
G_M29434_IG03:
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x35]
- movzx rsi, byte ptr [rbp-0x3D]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add ecx, edx
- movzx rcx, cl
- movzx rdi, byte ptr [rbp-0x34]
- movzx rdx, byte ptr [rbp-0x3C]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x33]
- movzx rsi, byte ptr [rbp-0x3B]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- movzx rdi, byte ptr [rbp-0x32]
- movzx rdx, byte ptr [rbp-0x3A]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x31]
- movzx rsi, byte ptr [rbp-0x39]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- add eax, ecx
- movzx rax, al
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x60], xmm0
- mov rcx, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x68], rcx
- mov rcx, qword ptr [rbp-0x60]
- mov qword ptr [rbp-0x70], rcx
- movzx rcx, byte ptr [rbp-0x68]
- movzx rdx, byte ptr [rbp-0x70]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x67]
- movzx rdi, byte ptr [rbp-0x6F]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- movzx rdx, cl
- movzx rcx, dl
- movzx rdx, byte ptr [rbp-0x66]
- movzx rdi, byte ptr [rbp-0x6E]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x65]
- movzx rsi, byte ptr [rbp-0x6D]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add ecx, edx
- movzx rcx, cl
- movzx rdi, byte ptr [rbp-0x64]
- movzx rdx, byte ptr [rbp-0x6C]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x63]
- movzx rsi, byte ptr [rbp-0x6B]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- ;; size=266 bbWeight=1 PerfScore 64.00
+ lea rdi, [rbp-0x50]
+ movsxd rax, esi
+ movzx rdi, byte ptr [rdi+rax]
+ lea rcx, [rbp-0x58]
+ movzx rcx, byte ptr [rcx+rax]
+ imul edi, ecx
+ lea rcx, [rbp-0x48]
+ mov byte ptr [rcx+rax], dil
+ inc esi
+ cmp esi, 8
+ jl SHORT G_M29434_IG03
+ ;; size=38 bbWeight=4 PerfScore 41.00
G_M29434_IG04:
- movzx rdi, byte ptr [rbp-0x62]
- movzx rdx, byte ptr [rbp-0x6A]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x61]
- movzx rsi, byte ptr [rbp-0x69]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- mov rdx, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x78], rdx
- mov rdx, qword ptr [rbp-0x58]
- mov qword ptr [rbp-0x80], rdx
- movzx rdx, byte ptr [rbp-0x78]
- movzx rdi, byte ptr [rbp-0x80]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x77]
- movzx rsi, byte ptr [rbp-0x7F]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- movzx rdi, dl
- movzx rdx, dil
- movzx rdi, byte ptr [rbp-0x76]
- movzx rsi, byte ptr [rbp-0x7E]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0x75]
- movzx r8, byte ptr [rbp-0x7D]
- imul esi, r8d
+ mov rsi, qword ptr [rbp-0x48]
+ mov qword ptr [rbp-0x60], rsi
+ movzx rsi, byte ptr [rbp-0x60]
+ movzx rdi, byte ptr [rbp-0x5F]
+ add esi, edi
movzx rsi, sil
+ movzx rdi, byte ptr [rbp-0x5E]
add edi, esi
- add edx, edi
- movzx rdx, dl
- movzx rsi, byte ptr [rbp-0x74]
- movzx rdi, byte ptr [rbp-0x7C]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0x73]
- movzx r8, byte ptr [rbp-0x7B]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0x5D]
add edi, esi
- add edi, edx
- movzx rdx, dil
- movzx rsi, byte ptr [rbp-0x72]
- movzx rdi, byte ptr [rbp-0x7A]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0x71]
- movzx r8, byte ptr [rbp-0x79]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0x5C]
add edi, esi
- add edi, edx
- movzx rdx, dil
- add ecx, edx
- add eax, ecx
- movzx rax, al
- vmovups ymm0, ymmword ptr [rbp+0x30]
- vmovups ymm1, ymmword ptr [rbp+0x70]
- vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x90], xmm2
- vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0xA0], xmm2
- mov rcx, qword ptr [rbp-0x90]
- mov qword ptr [rbp-0xA8], rcx
- mov rcx, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xB0], rcx
- ;; size=294 bbWeight=1 PerfScore 65.50
-G_M29434_IG05:
- movzx rcx, byte ptr [rbp-0xA8]
- movzx rdx, byte ptr [rbp-0xB0]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0xA7]
- movzx rdi, byte ptr [rbp-0xAF]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- movzx rdx, cl
- movzx rcx, dl
- movzx rdx, byte ptr [rbp-0xA6]
- movzx rdi, byte ptr [rbp-0xAE]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0xA5]
- movzx rsi, byte ptr [rbp-0xAD]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add ecx, edx
- movzx rcx, cl
- movzx rdi, byte ptr [rbp-0xA4]
- movzx rdx, byte ptr [rbp-0xAC]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0xA3]
- movzx rsi, byte ptr [rbp-0xAB]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- movzx rdi, byte ptr [rbp-0xA2]
- movzx rdx, byte ptr [rbp-0xAA]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0xA1]
- movzx rsi, byte ptr [rbp-0xA9]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- mov rdx, qword ptr [rbp-0x88]
- mov qword ptr [rbp-0xB8], rdx
- mov rdx, qword ptr [rbp-0x98]
- mov qword ptr [rbp-0xC0], rdx
- movzx rdx, byte ptr [rbp-0xB8]
- movzx rdi, byte ptr [rbp-0xC0]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0xB7]
- movzx rsi, byte ptr [rbp-0xBF]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- movzx rdi, dl
- movzx rdx, dil
- movzx rdi, byte ptr [rbp-0xB6]
- movzx rsi, byte ptr [rbp-0xBE]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xB5]
- movzx r8, byte ptr [rbp-0xBD]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0x5B]
add edi, esi
- add edx, edi
- movzx rdx, dl
- movzx rsi, byte ptr [rbp-0xB4]
- movzx rdi, byte ptr [rbp-0xBC]
- imul edi, esi
- ;; size=357 bbWeight=1 PerfScore 63.50
-G_M29434_IG06:
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xB3]
- movzx r8, byte ptr [rbp-0xBB]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0x5A]
add edi, esi
- add edi, edx
- movzx rdx, dil
- movzx rsi, byte ptr [rbp-0xB2]
- movzx rdi, byte ptr [rbp-0xBA]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xB1]
- movzx r8, byte ptr [rbp-0xB9]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0x59]
add edi, esi
- add edi, edx
- movzx rdx, dil
- add ecx, edx
- movzx rcx, cl
+ movzx rsi, dil
+ mov rdi, qword ptr [rbp-0x28]
+ mov rax, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x70], rdi
+ mov qword ptr [rbp-0x78], rax
+ xor edi, edi
+ align [0 bytes for IG05]
+ ;; size=108 bbWeight=1 PerfScore 17.75
+G_M29434_IG05:
+ lea rax, [rbp-0x70]
+ movsxd rcx, edi
+ movzx rax, byte ptr [rax+rcx]
+ lea rdx, [rbp-0x78]
+ movzx rdx, byte ptr [rdx+rcx]
+ imul eax, edx
+ lea rdx, [rbp-0x68]
+ mov byte ptr [rdx+rcx], al
+ inc edi
+ cmp edi, 8
+ jl SHORT G_M29434_IG05
+ ;; size=36 bbWeight=4 PerfScore 41.00
+G_M29434_IG06:
+ mov rdi, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0x80], rdi
+ movzx rdi, byte ptr [rbp-0x80]
+ movzx rax, byte ptr [rbp-0x7F]
+ add edi, eax
+ movzx rdi, dil
+ movzx rax, byte ptr [rbp-0x7E]
+ add eax, edi
+ movzx rdi, al
+ movzx rax, byte ptr [rbp-0x7D]
+ add eax, edi
+ movzx rdi, al
+ movzx rax, byte ptr [rbp-0x7C]
+ add eax, edi
+ movzx rdi, al
+ movzx rax, byte ptr [rbp-0x7B]
+ add eax, edi
+ movzx rdi, al
+ movzx rax, byte ptr [rbp-0x7A]
+ add eax, edi
+ movzx rdi, al
+ movzx rax, byte ptr [rbp-0x79]
+ add eax, edi
+ movzx rdi, al
+ add esi, edi
+ movzx rbx, sil
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0xD0], xmm0
+ vmovaps xmmword ptr [rbp-0x90], xmm0
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0xE0], xmm0
- mov rdx, qword ptr [rbp-0xD0]
- mov qword ptr [rbp-0xE8], rdx
- mov rdx, qword ptr [rbp-0xE0]
- mov qword ptr [rbp-0xF0], rdx
- movzx rdx, byte ptr [rbp-0xE8]
- movzx rdi, byte ptr [rbp-0xF0]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0xE7]
- movzx rsi, byte ptr [rbp-0xEF]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- movzx rdi, dl
- movzx rdx, dil
- movzx rdi, byte ptr [rbp-0xE6]
- movzx rsi, byte ptr [rbp-0xEE]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xE5]
- movzx r8, byte ptr [rbp-0xED]
- imul esi, r8d
+ vmovaps xmmword ptr [rbp-0xA0], xmm0
+ mov rsi, qword ptr [rbp-0x90]
+ mov rdi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xB0], rsi
+ mov qword ptr [rbp-0xB8], rdi
+ xor esi, esi
+ align [0 bytes for IG07]
+ ;; size=147 bbWeight=1 PerfScore 24.25
+G_M29434_IG07:
+ lea rdi, [rbp-0xB0]
+ movsxd rax, esi
+ movzx rdi, byte ptr [rdi+rax]
+ lea rcx, [rbp-0xB8]
+ movzx rcx, byte ptr [rcx+rax]
+ imul edi, ecx
+ lea rcx, [rbp-0xA8]
+ mov byte ptr [rcx+rax], dil
+ inc esi
+ cmp esi, 8
+ jl SHORT G_M29434_IG07
+ ;; size=47 bbWeight=4 PerfScore 41.00
+G_M29434_IG08:
+ mov rsi, qword ptr [rbp-0xA8]
+ mov qword ptr [rbp-0xC0], rsi
+ movzx rsi, byte ptr [rbp-0xC0]
+ movzx rdi, byte ptr [rbp-0xBF]
+ add esi, edi
movzx rsi, sil
+ movzx rdi, byte ptr [rbp-0xBE]
add edi, esi
- add edx, edi
- movzx rdx, dl
- movzx rsi, byte ptr [rbp-0xE4]
- movzx rdi, byte ptr [rbp-0xEC]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xE3]
- movzx r8, byte ptr [rbp-0xEB]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0xBD]
add edi, esi
- add edi, edx
- movzx rdx, dil
- movzx rsi, byte ptr [rbp-0xE2]
- movzx rdi, byte ptr [rbp-0xEA]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xE1]
- movzx r8, byte ptr [rbp-0xE9]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0xBC]
add edi, esi
- add edi, edx
- movzx rdx, dil
- mov rdi, qword ptr [rbp-0xC8]
- ;; size=377 bbWeight=1 PerfScore 63.00
-G_M29434_IG07:
- mov qword ptr [rbp-0xF8], rdi
- mov rdi, qword ptr [rbp-0xD8]
- mov qword ptr [rbp-0x100], rdi
- movzx rdi, byte ptr [rbp-0xF8]
- movzx rsi, byte ptr [rbp-0x100]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0xF7]
- movzx r8, byte ptr [rbp-0xFF]
- imul esi, r8d
- movzx rsi, sil
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0xBB]
add edi, esi
movzx rsi, dil
- movzx rdi, sil
- movzx rsi, byte ptr [rbp-0xF6]
- movzx r8, byte ptr [rbp-0xFE]
- imul esi, r8d
- movzx rsi, sil
- movzx r8, byte ptr [rbp-0xF5]
- movzx r9, byte ptr [rbp-0xFD]
- imul r8d, r9d
- movzx r8, r8b
- add esi, r8d
+ movzx rdi, byte ptr [rbp-0xBA]
add edi, esi
- movzx rdi, dil
- movzx r8, byte ptr [rbp-0xF4]
- movzx rsi, byte ptr [rbp-0xFC]
- imul esi, r8d
- movzx rsi, sil
- movzx r8, byte ptr [rbp-0xF3]
- movzx r9, byte ptr [rbp-0xFB]
- imul r8d, r9d
- movzx r8, r8b
- add esi, r8d
- add esi, edi
- movzx rdi, sil
- movzx r8, byte ptr [rbp-0xF2]
- movzx rsi, byte ptr [rbp-0xFA]
- imul esi, r8d
- movzx rsi, sil
- movzx r8, byte ptr [rbp-0xF1]
- movzx r9, byte ptr [rbp-0xF9]
- imul r8d, r9d
- movzx r8, r8b
- add esi, r8d
- add esi, edi
- movzx rdi, sil
- add edx, edi
- add ecx, edx
- add eax, ecx
+ movzx rsi, dil
+ movzx rdi, byte ptr [rbp-0xB9]
+ add edi, esi
+ movzx rsi, dil
+ mov r15d, esi
+ mov rsi, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xD0], rsi
+ mov qword ptr [rbp-0xD8], rdi
+ xor esi, esi
+ align [0 bytes for IG09]
+ ;; size=153 bbWeight=1 PerfScore 18.00
+G_M29434_IG09:
+ lea rdi, [rbp-0xD0]
+ movsxd rax, esi
+ movzx rdi, byte ptr [rdi+rax]
+ lea rcx, [rbp-0xD8]
+ movzx rcx, byte ptr [rcx+rax]
+ imul edi, ecx
+ lea rcx, [rbp-0xC8]
+ mov byte ptr [rcx+rax], dil
+ inc esi
+ cmp esi, 8
+ jl SHORT G_M29434_IG09
+ ;; size=47 bbWeight=4 PerfScore 41.00
+G_M29434_IG10:
+ mov rsi, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0xE0], rsi
+ movzx rsi, byte ptr [rbp-0xE0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDF]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDE]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDD]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDC]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDB]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDA]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xD9]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rdi, r15b
+ mov esi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov esi, eax
+ movzx rdi, bl
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ vmovups ymm0, ymmword ptr [rbp+0x30]
+ vmovups ymmword ptr [rbp-0x110], ymm0
+ vmovups ymm1, ymmword ptr [rbp+0x70]
+ vmovups ymmword ptr [rbp-0x130], ymm1
+ vmovaps ymm2, ymm0
+ vmovaps xmmword ptr [rbp-0x140], xmm2
+ vmovaps ymm2, ymm1
+ vmovaps xmmword ptr [rbp-0x150], xmm2
+ mov rdi, qword ptr [rbp-0x140]
+ mov rsi, qword ptr [rbp-0x150]
+ mov qword ptr [rbp-0x160], rdi
+ mov qword ptr [rbp-0x168], rsi
+ xor r15d, r15d
+ ;; size=309 bbWeight=1 PerfScore 62.50
+G_M29434_IG11:
+ lea rdi, [rbp-0x160]
+ movsxd rsi, r15d
+ movzx rdi, byte ptr [rdi+rsi]
+ lea rsi, [rbp-0x168]
+ movsxd rax, r15d
+ movzx rsi, byte ptr [rsi+rax]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0x158]
+ movsxd rdi, r15d
+ mov byte ptr [rsi+rdi], al
+ inc r15d
+ cmp r15d, 8
+ jl SHORT G_M29434_IG11
+ ;; size=64 bbWeight=4 PerfScore 48.00
+G_M29434_IG12:
+ mov rsi, qword ptr [rbp-0x158]
+ mov qword ptr [rbp-0x170], rsi
+ movzx rsi, byte ptr [rbp-0x170]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x16F]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x16E]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x16D]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x16C]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x16B]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x16A]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0x169]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ mov rdi, qword ptr [rbp-0x138]
+ mov rsi, qword ptr [rbp-0x148]
+ mov qword ptr [rbp-0x180], rdi
+ mov qword ptr [rbp-0x188], rsi
+ xor r14d, r14d
+ ;; size=252 bbWeight=1 PerfScore 44.25
+G_M29434_IG13:
+ lea rdi, [rbp-0x180]
+ movsxd rsi, r14d
+ movzx rdi, byte ptr [rdi+rsi]
+ lea rsi, [rbp-0x188]
+ movsxd rax, r14d
+ movzx rsi, byte ptr [rsi+rax]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0x178]
+ movsxd rdi, r14d
+ mov byte ptr [rsi+rdi], al
+ inc r14d
+ cmp r14d, 8
+ jl SHORT G_M29434_IG13
+ ;; size=64 bbWeight=4 PerfScore 48.00
+G_M29434_IG14:
+ mov rsi, qword ptr [rbp-0x178]
+ mov qword ptr [rbp-0x190], rsi
+ movzx rsi, byte ptr [rbp-0x190]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x18F]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x18E]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x18D]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x18C]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x18B]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x18A]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x189]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rdi, r15b
+ mov esi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ vmovups ymm0, ymmword ptr [rbp-0x110]
+ vextractf128 xmm0, ymm0, 1
+ vmovaps xmmword ptr [rbp-0x1A0], xmm0
+ vmovups ymm1, ymmword ptr [rbp-0x130]
+ vextractf128 xmm0, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x1B0], xmm0
+ mov rdi, qword ptr [rbp-0x1A0]
+ mov rsi, qword ptr [rbp-0x1B0]
+ mov qword ptr [rbp-0x1C0], rdi
+ mov qword ptr [rbp-0x1C8], rsi
+ xor r14d, r14d
+ ;; size=286 bbWeight=1 PerfScore 60.25
+G_M29434_IG15:
+ lea rdi, [rbp-0x1C0]
+ movsxd rsi, r14d
+ movzx rdi, byte ptr [rdi+rsi]
+ lea rsi, [rbp-0x1C8]
+ movsxd rax, r14d
+ movzx rsi, byte ptr [rsi+rax]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0x1B8]
+ movsxd rdi, r14d
+ mov byte ptr [rsi+rdi], al
+ inc r14d
+ cmp r14d, 8
+ jl SHORT G_M29434_IG15
+ ;; size=64 bbWeight=4 PerfScore 48.00
+G_M29434_IG16:
+ mov rsi, qword ptr [rbp-0x1B8]
+ mov qword ptr [rbp-0x1D0], rsi
+ movzx rsi, byte ptr [rbp-0x1D0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1CF]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1CE]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1CD]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1CC]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1CB]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1CA]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ movzx rsi, byte ptr [rbp-0x1C9]
+ mov edi, r14d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r14d, eax
+ mov rdi, qword ptr [rbp-0x198]
+ mov qword ptr [rsp], rdi
+ mov rdi, qword ptr [rbp-0x1A8]
+ mov qword ptr [rsp+0x08], rdi
+ lea rdi, [rbp-0x1D8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64`1[ubyte]:op_Multiply(System.Runtime.Intrinsics.Vector64`1[ubyte],System.Runtime.Intrinsics.Vector64`1[ubyte]):System.Runtime.Intrinsics.Vector64`1[ubyte]
+ call [rax]System.Runtime.Intrinsics.Vector64`1[ubyte]:op_Multiply(System.Runtime.Intrinsics.Vector64`1[ubyte],System.Runtime.Intrinsics.Vector64`1[ubyte]):System.Runtime.Intrinsics.Vector64`1[ubyte]
+ mov rdi, qword ptr [rbp-0x1D8]
+ mov qword ptr [rbp-0x1E0], rdi
+ lea rdi, [rbp-0x1E0]
+ xor esi, esi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 1
+ ;; size=329 bbWeight=1 PerfScore 58.50
+G_M29434_IG17:
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 2
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 3
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 4
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 5
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 6
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ lea rdi, [rbp-0x1E0]
+ mov esi, 7
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Vector64:GetElementUnsafe[ubyte](byref,int):ubyte
+ mov esi, eax
+ mov edi, r13d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r13d, eax
+ movzx rdi, r14b
+ mov esi, r13d
+ ;; size=303 bbWeight=1 PerfScore 55.75
+G_M29434_IG18:
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov esi, eax
+ movzx rdi, r15b
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ add eax, ebx
movzx rax, al
- ;; size=258 bbWeight=1 PerfScore 41.00
-G_M29434_IG08:
+ ;; size=35 bbWeight=1 PerfScore 7.50
+G_M29434_IG19:
vzeroupper
- add rsp, 256
+ add rsp, 464
+ pop rbx
+ pop r13
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=19 bbWeight=1 PerfScore 4.75
-; Total bytes of code 1830, prolog size 16, PerfScore 368.25, instruction count 422, allocated bytes for code 1830 (MethodHash=b8b18d05) for method System.Runtime.Intrinsics.Vector512:Dot[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):ubyte (FullOpts)
+; Total bytes of code 2370, prolog size 23, PerfScore 682.00, instruction count 496, allocated bytes for code 2370 (MethodHash=b8b18d05) for method System.Runtime.Intrinsics.Vector512:Dot[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):ubyte (FullOpts)451 (58.12 % of base) - System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector,T>.Dot(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):ubyte
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Dot(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):ubyte (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
-; partially interruptible
+; fully interruptible
; No PGO data
-; 0 inlinees with PGO data; 19 single block inlinees; 23 inlinees without PGO data
+; 0 inlinees with PGO data; 23 single block inlinees; 9 inlinees without PGO data
; Final local variable assignments
;
-; V00 arg0 [V00,T27] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V01 arg1 [V01,T28] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V00 arg0 [V00,T32] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V01 arg1 [V01,T33] ( 2, 2 ) simd32 -> [rbp+0x30] single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
;# V02 OutArgs [V02 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V03 tmp1 [V03,T12] ( 2, 4 ) int -> rax "impAppendStmt"
-; V04 tmp2 [V04,T23] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V05 tmp3 [V05,T24] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V06 tmp4 [V06,T13] ( 2, 4 ) int -> rax "impAppendStmt"
-; V07 tmp5 [V07,T04] ( 8, 8 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
-;* V08 tmp6 [V08,T19] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V09 tmp7 [V09 ] ( 9, 18 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V10 tmp8 [V10 ] ( 9, 18 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V11 tmp9 [V11 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V12 tmp10 [V12,T00] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V13 tmp11 [V13 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V14 tmp12 [V14 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V15 tmp13 [V15 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V16 tmp14 [V16 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V17 tmp15 [V17 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V18 tmp16 [V18 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V19 tmp17 [V19,T08] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V20 tmp18 [V20 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V21 tmp19 [V21 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V22 tmp20 [V22 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V23 tmp21 [V23 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V24 tmp22 [V24 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V25 tmp23 [V25 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V26 tmp24 [V26,T05] ( 8, 8 ) ubyte -> rcx ld-addr-op "Inline ldloca(s) first use temp"
-;* V27 tmp25 [V27,T20] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V28 tmp26 [V28 ] ( 9, 18 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V29 tmp27 [V29 ] ( 9, 18 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V03 tmp1 [V03,T12] ( 2, 4 ) int -> rbx "impAppendStmt"
+; V04 tmp2 [V04,T28] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V05 tmp3 [V05,T29] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V06 tmp4 [V06,T13] ( 2, 4 ) int -> rbx "impAppendStmt"
+;* V07 tmp5 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V08 tmp6 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V09 tmp7 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V10 tmp8 [V10 ] ( 2, 5 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V11 tmp9 [V11,T04] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
+; V12 tmp10 [V12 ] ( 2, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V13 tmp11 [V13 ] ( 2, 10 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V14 tmp12 [V14 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V15 tmp13 [V15 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+;* V16 tmp14 [V16 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V17 tmp15 [V17 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V18 tmp16 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V19 tmp17 [V19,T00] ( 16, 16 ) ubyte -> rbx ld-addr-op "Inline ldloca(s) first use temp"
+;* V20 tmp18 [V20,T24] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V21 tmp19 [V21 ] ( 9, 18 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V22 tmp20 [V22 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V23 tmp21 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V24 tmp22 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V25 tmp23 [V25 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V26 tmp24 [V26 ] ( 2, 5 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V27 tmp25 [V27,T05] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V28 tmp26 [V28 ] ( 2, 10 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V29 tmp27 [V29 ] ( 2, 10 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V30 tmp28 [V30 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V31 tmp29 [V31,T01] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V32 tmp30 [V32 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V33 tmp31 [V33 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V34 tmp32 [V34 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V35 tmp33 [V35 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V36 tmp34 [V36 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V37 tmp35 [V37 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V38 tmp36 [V38,T09] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V39 tmp37 [V39 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V40 tmp38 [V40 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V41 tmp39 [V41 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V42 tmp40 [V42 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V43 tmp41 [V43 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V44 tmp42 [V44 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V45 tmp43 [V45 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V46 tmp44 [V46 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V47 tmp45 [V47 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V48 tmp46 [V48,T25] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V49 tmp47 [V49,T26] ( 3, 6 ) simd16 -> [rbp-0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V50 tmp48 [V50,T14] ( 2, 4 ) int -> rcx "impAppendStmt"
-; V51 tmp49 [V51,T06] ( 8, 8 ) ubyte -> rcx ld-addr-op "Inline ldloca(s) first use temp"
-;* V52 tmp50 [V52,T21] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V53 tmp51 [V53 ] ( 9, 18 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V54 tmp52 [V54 ] ( 9, 18 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V55 tmp53 [V55 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V56 tmp54 [V56,T02] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V57 tmp55 [V57 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V58 tmp56 [V58 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V59 tmp57 [V59 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V60 tmp58 [V60 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V61 tmp59 [V61 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V62 tmp60 [V62 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V63 tmp61 [V63,T10] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V64 tmp62 [V64 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V65 tmp63 [V65 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V66 tmp64 [V66 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V67 tmp65 [V67 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V68 tmp66 [V68 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V69 tmp67 [V69 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V70 tmp68 [V70,T07] ( 8, 8 ) ubyte -> rdx ld-addr-op "Inline ldloca(s) first use temp"
-;* V71 tmp69 [V71,T22] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V72 tmp70 [V72 ] ( 9, 18 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V73 tmp71 [V73 ] ( 9, 18 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V74 tmp72 [V74 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V75 tmp73 [V75,T03] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V76 tmp74 [V76 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V77 tmp75 [V77 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V78 tmp76 [V78 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V79 tmp77 [V79 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V80 tmp78 [V80 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V81 tmp79 [V81 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V82 tmp80 [V82,T11] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V83 tmp81 [V83 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V84 tmp82 [V84 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V85 tmp83 [V85 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V86 tmp84 [V86 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V87 tmp85 [V87 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V88 tmp86 [V88 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V89 tmp87 [V89 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V90 tmp88 [V90 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V91 tmp89 [V91 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V92 tmp90 [V92 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V93 tmp91 [V93 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V94 tmp92 [V94 ] ( 9, 17 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V95 tmp93 [V95 ] ( 9, 17 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
-; V96 tmp94 [V96 ] ( 9, 17 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V28._00 (fldOffset=0x0)" P-DEP
-; V97 tmp95 [V97 ] ( 9, 17 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
-; V98 tmp96 [V98 ] ( 9, 17 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V53._00 (fldOffset=0x0)" P-DEP
-; V99 tmp97 [V99 ] ( 9, 17 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V54._00 (fldOffset=0x0)" P-DEP
-; V100 tmp98 [V100 ] ( 9, 17 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V72._00 (fldOffset=0x0)" P-DEP
-; V101 tmp99 [V101 ] ( 9, 17 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V73._00 (fldOffset=0x0)" P-DEP
-; V102 cse0 [V102,T15] ( 2, 2 ) int -> rcx "CSE #01: moderate"
-; V103 cse1 [V103,T16] ( 2, 2 ) int -> rdx "CSE #02: moderate"
-; V104 cse2 [V104,T17] ( 2, 2 ) int -> rdx "CSE #03: moderate"
-; V105 cse3 [V105,T18] ( 2, 2 ) int -> rdi "CSE #04: moderate"
+;* V31 tmp29 [V31 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V32 tmp30 [V32,T01] ( 16, 16 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
+;* V33 tmp31 [V33,T25] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V34 tmp32 [V34 ] ( 9, 18 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V35 tmp33 [V35 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+; V36 tmp34 [V36,T30] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V37 tmp35 [V37,T31] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V38 tmp36 [V38,T14] ( 2, 4 ) int -> r15 "impAppendStmt"
+;* V39 tmp37 [V39 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V40 tmp38 [V40 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V41 tmp39 [V41 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V42 tmp40 [V42 ] ( 2, 5 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V43 tmp41 [V43,T06] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V44 tmp42 [V44 ] ( 2, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V45 tmp43 [V45 ] ( 2, 10 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V46 tmp44 [V46 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V47 tmp45 [V47 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V48 tmp46 [V48,T02] ( 16, 16 ) ubyte -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V49 tmp47 [V49,T26] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V50 tmp48 [V50 ] ( 9, 18 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V51 tmp49 [V51 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V52 tmp50 [V52 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V53 tmp51 [V53 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V54 tmp52 [V54 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V55 tmp53 [V55 ] ( 2, 5 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V56 tmp54 [V56,T07] ( 5, 17 ) int -> r14 "Inline stloc first use temp"
+; V57 tmp55 [V57 ] ( 2, 10 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V58 tmp56 [V58 ] ( 2, 10 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V59 tmp57 [V59 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V60 tmp58 [V60 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V61 tmp59 [V61,T03] ( 16, 16 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
+;* V62 tmp60 [V62,T27] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V63 tmp61 [V63 ] ( 9, 18 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V64 tmp62 [V64 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+; V65 tmp63 [V65,T16] ( 2, 2 ) long -> rsi "field V07._00 (fldOffset=0x0)" P-INDEP
+; V66 tmp64 [V66,T17] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+;* V67 tmp65 [V67 ] ( 0, 0 ) long -> zero-ref "field V09._00 (fldOffset=0x0)" P-INDEP
+; V68 tmp66 [V68 ] ( 2, 5 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V69 tmp67 [V69 ] ( 2, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V70 tmp68 [V70 ] ( 2, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V71 tmp69 [V71 ] ( 9, 17 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
+; V72 tmp70 [V72,T18] ( 2, 2 ) long -> rdi "field V23._00 (fldOffset=0x0)" P-INDEP
+; V73 tmp71 [V73,T19] ( 2, 2 ) long -> rsi "field V24._00 (fldOffset=0x0)" P-INDEP
+;* V74 tmp72 [V74 ] ( 0, 0 ) long -> zero-ref "field V25._00 (fldOffset=0x0)" P-INDEP
+; V75 tmp73 [V75 ] ( 2, 5 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
+; V76 tmp74 [V76 ] ( 2, 9 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V28._00 (fldOffset=0x0)" P-DEP
+; V77 tmp75 [V77 ] ( 2, 9 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V78 tmp76 [V78 ] ( 9, 17 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V34._00 (fldOffset=0x0)" P-DEP
+; V79 tmp77 [V79,T20] ( 2, 2 ) long -> rdi "field V39._00 (fldOffset=0x0)" P-INDEP
+; V80 tmp78 [V80,T21] ( 2, 2 ) long -> rsi "field V40._00 (fldOffset=0x0)" P-INDEP
+;* V81 tmp79 [V81 ] ( 0, 0 ) long -> zero-ref "field V41._00 (fldOffset=0x0)" P-INDEP
+; V82 tmp80 [V82 ] ( 2, 5 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V42._00 (fldOffset=0x0)" P-DEP
+; V83 tmp81 [V83 ] ( 2, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V44._00 (fldOffset=0x0)" P-DEP
+; V84 tmp82 [V84 ] ( 2, 9 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
+; V85 tmp83 [V85 ] ( 9, 17 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V50._00 (fldOffset=0x0)" P-DEP
+; V86 tmp84 [V86,T22] ( 2, 2 ) long -> rdi "field V52._00 (fldOffset=0x0)" P-INDEP
+; V87 tmp85 [V87,T23] ( 2, 2 ) long -> rsi "field V53._00 (fldOffset=0x0)" P-INDEP
+;* V88 tmp86 [V88 ] ( 0, 0 ) long -> zero-ref "field V54._00 (fldOffset=0x0)" P-INDEP
+; V89 tmp87 [V89 ] ( 2, 5 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V55._00 (fldOffset=0x0)" P-DEP
+; V90 tmp88 [V90 ] ( 2, 9 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V57._00 (fldOffset=0x0)" P-DEP
+; V91 tmp89 [V91 ] ( 2, 9 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V58._00 (fldOffset=0x0)" P-DEP
+; V92 tmp90 [V92 ] ( 9, 17 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V63._00 (fldOffset=0x0)" P-DEP
+; V93 tmp91 [V93,T15] ( 2, 4 ) int -> rsi "argument with side effect"
+; V94 cse0 [V94,T08] ( 4, 16 ) long -> rax "CSE #01: aggressive"
+; V95 cse1 [V95,T09] ( 4, 16 ) long -> r14 "CSE #02: aggressive"
+; V96 cse2 [V96,T10] ( 4, 16 ) long -> r14 "CSE #03: aggressive"
+; V97 cse3 [V97,T11] ( 4, 16 ) long -> r13 "CSE #04: aggressive"
;
-; Lcl frame size = 128
+; Lcl frame size = 192
G_M42821_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
+ push r15
+ push r14
+ push r13
+ push rbx
+ sub rsp, 192
+ lea rbp, [rsp+0xE0]
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x30]
- ;; size=26 bbWeight=1 PerfScore 9.75
+ ;; size=33 bbWeight=1 PerfScore 13.75
G_M42821_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ vmovups ymmword ptr [rbp+0x30], ymm1
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x28], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x30], rax
- movzx rax, byte ptr [rbp-0x28]
- movzx rcx, byte ptr [rbp-0x30]
- imul eax, ecx
- movzx rax, al
- movzx rcx, byte ptr [rbp-0x27]
- movzx rdx, byte ptr [rbp-0x2F]
- imul ecx, edx
- movzx rcx, cl
- add eax, ecx
- movzx rcx, al
- movzx rax, cl
- movzx rcx, byte ptr [rbp-0x26]
- movzx rdx, byte ptr [rbp-0x2E]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x25]
- movzx rdi, byte ptr [rbp-0x2D]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add eax, ecx
- movzx rax, al
- movzx rdx, byte ptr [rbp-0x24]
- movzx rcx, byte ptr [rbp-0x2C]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x23]
- movzx rdi, byte ptr [rbp-0x2B]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add ecx, eax
- movzx rax, cl
- movzx rdx, byte ptr [rbp-0x22]
- movzx rcx, byte ptr [rbp-0x2A]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x21]
- movzx rdi, byte ptr [rbp-0x29]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add ecx, eax
- movzx rax, cl
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x38], rcx
- mov rcx, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x40], rcx
- movzx rcx, byte ptr [rbp-0x38]
- movzx rdx, byte ptr [rbp-0x40]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x37]
- movzx rdi, byte ptr [rbp-0x3F]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- movzx rdx, cl
- movzx rcx, dl
- movzx rdx, byte ptr [rbp-0x36]
- movzx rdi, byte ptr [rbp-0x3E]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x35]
- ;; size=251 bbWeight=1 PerfScore 62.00
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ mov rsi, qword ptr [rbp-0x30]
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rsi
+ mov qword ptr [rbp-0x58], rdi
+ xor esi, esi
+ align [0 bytes for IG03]
+ ;; size=46 bbWeight=1 PerfScore 8.75
G_M42821_IG03:
- movzx rsi, byte ptr [rbp-0x3D]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add ecx, edx
- movzx rcx, cl
- movzx rdi, byte ptr [rbp-0x34]
- movzx rdx, byte ptr [rbp-0x3C]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x33]
- movzx rsi, byte ptr [rbp-0x3B]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- movzx rdi, byte ptr [rbp-0x32]
- movzx rdx, byte ptr [rbp-0x3A]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x31]
- movzx rsi, byte ptr [rbp-0x39]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- add eax, ecx
- movzx rax, al
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x60], xmm0
- mov rcx, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x68], rcx
- mov rcx, qword ptr [rbp-0x60]
- mov qword ptr [rbp-0x70], rcx
- movzx rcx, byte ptr [rbp-0x68]
- movzx rdx, byte ptr [rbp-0x70]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x67]
- movzx rdi, byte ptr [rbp-0x6F]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- movzx rdx, cl
- movzx rcx, dl
- movzx rdx, byte ptr [rbp-0x66]
- movzx rdi, byte ptr [rbp-0x6E]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x65]
- movzx rsi, byte ptr [rbp-0x6D]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add ecx, edx
- movzx rcx, cl
- movzx rdi, byte ptr [rbp-0x64]
- movzx rdx, byte ptr [rbp-0x6C]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x63]
- movzx rsi, byte ptr [rbp-0x6B]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- movzx rdi, byte ptr [rbp-0x62]
- movzx rdx, byte ptr [rbp-0x6A]
- imul edx, edi
- ;; size=267 bbWeight=1 PerfScore 64.75
+ lea rdi, [rbp-0x50]
+ movsxd rax, esi
+ movzx rdi, byte ptr [rdi+rax]
+ lea rcx, [rbp-0x58]
+ movzx rcx, byte ptr [rcx+rax]
+ imul edi, ecx
+ lea rcx, [rbp-0x48]
+ mov byte ptr [rcx+rax], dil
+ inc esi
+ cmp esi, 8
+ jl SHORT G_M42821_IG03
+ ;; size=38 bbWeight=4 PerfScore 41.00
G_M42821_IG04:
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x61]
- movzx rsi, byte ptr [rbp-0x69]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- mov rdx, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x78], rdx
- mov rdx, qword ptr [rbp-0x58]
- mov qword ptr [rbp-0x80], rdx
- movzx rdx, byte ptr [rbp-0x78]
- movzx rdi, byte ptr [rbp-0x80]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x77]
+ mov rsi, qword ptr [rbp-0x48]
+ mov qword ptr [rbp-0x60], rsi
+ movzx rsi, byte ptr [rbp-0x60]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x5F]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x5E]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x5D]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x5C]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x5B]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x5A]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x59]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ mov rdi, qword ptr [rbp-0x28]
+ mov rsi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x70], rdi
+ mov qword ptr [rbp-0x78], rsi
+ xor r15d, r15d
+ ;; size=195 bbWeight=1 PerfScore 44.25
+G_M42821_IG05:
+ lea rdi, [rbp-0x70]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0x78]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0x68]
+ mov byte ptr [rsi+r14], al
+ inc r15d
+ cmp r15d, 8
+ jl SHORT G_M42821_IG05
+ ;; size=50 bbWeight=4 PerfScore 46.00
+G_M42821_IG06:
+ mov rsi, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0x80], rsi
+ movzx rsi, byte ptr [rbp-0x80]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
movzx rsi, byte ptr [rbp-0x7F]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- movzx rdi, dl
- movzx rdx, dil
- movzx rdi, byte ptr [rbp-0x76]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
movzx rsi, byte ptr [rbp-0x7E]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0x75]
- movzx r8, byte ptr [rbp-0x7D]
- imul esi, r8d
- movzx rsi, sil
- add edi, esi
- add edx, edi
- movzx rdx, dl
- movzx rsi, byte ptr [rbp-0x74]
- movzx rdi, byte ptr [rbp-0x7C]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0x73]
- movzx r8, byte ptr [rbp-0x7B]
- imul esi, r8d
- movzx rsi, sil
- add edi, esi
- add edi, edx
- movzx rdx, dil
- movzx rsi, byte ptr [rbp-0x72]
- movzx rdi, byte ptr [rbp-0x7A]
- imul edi, esi
- movzx rdi, dil
- movzx rsi, byte ptr [rbp-0x71]
- movzx r8, byte ptr [rbp-0x79]
- imul esi, r8d
- movzx rsi, sil
- add edi, esi
- add edi, edx
- movzx rdx, dil
- add ecx, edx
- add eax, ecx
- movzx rax, al
- ;; size=220 bbWeight=1 PerfScore 47.00
-G_M42821_IG05:
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x7D]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x7C]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x7B]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x7A]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x79]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rdi, bl
+ mov esi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ vmovups ymm0, ymmword ptr [rbp+0x10]
+ vextractf128 xmm0, ymm0, 1
+ vmovaps xmmword ptr [rbp-0x90], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x30]
+ vextractf128 xmm0, ymm1, 1
+ vmovaps xmmword ptr [rbp-0xA0], xmm0
+ mov rdi, qword ptr [rbp-0x90]
+ mov rsi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xB0], rdi
+ mov qword ptr [rbp-0xB8], rsi
+ xor r15d, r15d
+ ;; size=249 bbWeight=1 PerfScore 60.25
+G_M42821_IG07:
+ lea rdi, [rbp-0xB0]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0xB8]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0xA8]
+ mov byte ptr [rsi+r14], al
+ inc r15d
+ cmp r15d, 8
+ jl SHORT G_M42821_IG07
+ ;; size=59 bbWeight=4 PerfScore 46.00
+G_M42821_IG08:
+ mov rsi, qword ptr [rbp-0xA8]
+ mov qword ptr [rbp-0xC0], rsi
+ movzx rsi, byte ptr [rbp-0xC0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xBF]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xBE]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xBD]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xBC]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xBB]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xBA]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ movzx rsi, byte ptr [rbp-0xB9]
+ mov edi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov r15d, eax
+ mov rdi, qword ptr [rbp-0x88]
+ mov rsi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xD0], rdi
+ mov qword ptr [rbp-0xD8], rsi
+ xor r14d, r14d
+ ;; size=252 bbWeight=1 PerfScore 44.25
+G_M42821_IG09:
+ lea rdi, [rbp-0xD0]
+ movsxd r13, r14d
+ movzx rdi, byte ptr [rdi+r13]
+ lea rsi, [rbp-0xD8]
+ movzx rsi, byte ptr [rsi+r13]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0xC8]
+ mov byte ptr [rsi+r13], al
+ inc r14d
+ cmp r14d, 8
+ jl SHORT G_M42821_IG09
+ ;; size=59 bbWeight=4 PerfScore 46.00
+G_M42821_IG10:
+ mov rsi, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0xE0], rsi
+ movzx rsi, byte ptr [rbp-0xE0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDF]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDE]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDD]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDC]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDB]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xDA]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0xD9]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rdi, r15b
+ mov esi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov esi, eax
+ movzx rdi, bl
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ nop
+ ;; size=227 bbWeight=1 PerfScore 45.75
+G_M42821_IG11:
vzeroupper
- add rsp, 128
+ add rsp, 192
+ pop rbx
+ pop r13
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=19 bbWeight=1 PerfScore 4.75
-; Total bytes of code 776, prolog size 16, PerfScore 186.25, instruction count 214, allocated bytes for code 776 (MethodHash=d63158ba) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Dot(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):ubyte (FullOpts)
+; Total bytes of code 1227, prolog size 33, PerfScore 400.75, instruction count 263, allocated bytes for code 1227 (MethodHash=d63158ba) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Dot(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):ubyte (FullOpts)411 (78.59 % of base) - System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector,T>.ShiftLeft(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftLeft(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 16 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 16 single block inlinees; 10 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V02 arg1 [V02,T02] ( 3, 3 ) int -> rsi single-def
+; V00 RetBuf [V00,T02] ( 4, 4 ) byref -> r15 single-def
+; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V02 arg1 [V02,T00] ( 27, 27 ) int -> rbx single-def
;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> mm1 "impAppendStmt"
+; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V07 tmp4 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V09 tmp6 [V09 ] ( 9, 9 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V09 tmp6 [V09 ] ( 9, 9 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V10 tmp7 [V10,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V11 tmp8 [V11 ] ( 9, 18 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V11 tmp8 [V11 ] ( 9, 18 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V12 tmp9 [V12 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
;* V13 tmp10 [V13 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V14 tmp11 [V14 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V15 tmp12 [V15 ] ( 9, 9 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V15 tmp12 [V15 ] ( 9, 9 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V16 tmp13 [V16,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V17 tmp14 [V17 ] ( 9, 18 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V17 tmp14 [V17 ] ( 9, 18 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V18 tmp15 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V19 tmp16 [V19 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V20 tmp17 [V20 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V21 tmp18 [V21,T14] ( 3, 3 ) simd16 -> [rbp-0x40] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V22 tmp19 [V22,T12] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V24 tmp21 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V25 tmp22 [V25 ] ( 9, 9 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V26 tmp23 [V26,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V27 tmp24 [V27 ] ( 9, 18 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V28 tmp25 [V28 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V29 tmp26 [V29 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V30 tmp27 [V30 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V31 tmp28 [V31 ] ( 9, 9 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V32 tmp29 [V32,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 9, 18 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V34 tmp31 [V34 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V35 tmp32 [V35 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V36 tmp33 [V36 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V37 tmp34 [V37,T15] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V38 tmp35 [V38 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V39 tmp36 [V39,T03] ( 2, 2 ) long -> rax "field V07._00 (fldOffset=0x0)" P-INDEP
-; V40 tmp37 [V40,T04] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V41 tmp38 [V41 ] ( 9, 9 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V42 tmp39 [V42 ] ( 9, 17 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V43 tmp40 [V43 ] ( 9, 9 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V44 tmp41 [V44 ] ( 9, 17 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V45 tmp42 [V45,T05] ( 2, 2 ) long -> rax "field V23._00 (fldOffset=0x0)" P-INDEP
-; V46 tmp43 [V46,T06] ( 2, 2 ) long -> rcx "field V24._00 (fldOffset=0x0)" P-INDEP
-; V47 tmp44 [V47 ] ( 9, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp45 [V48 ] ( 9, 17 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
-; V49 tmp46 [V49 ] ( 9, 9 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V50 tmp47 [V50 ] ( 9, 17 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V51 cse0 [V51,T00] ( 33, 33 ) int -> rsi "CSE #01: aggressive"
+; V19 tmp16 [V19,T14] ( 3, 3 ) simd16 -> [rbp-0x70] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V20 tmp17 [V20,T12] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+;* V21 tmp18 [V21 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V22 tmp19 [V22 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V23 tmp20 [V23 ] ( 9, 9 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V24 tmp21 [V24,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 9, 18 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V27 tmp24 [V27 ] ( 9, 9 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V28 tmp25 [V28,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 9, 18 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V30 tmp27 [V30 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V31 tmp28 [V31,T15] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+;* V32 tmp29 [V32 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V33 tmp30 [V33,T03] ( 2, 2 ) long -> r14 "field V07._00 (fldOffset=0x0)" P-INDEP
+; V34 tmp31 [V34,T04] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+; V35 tmp32 [V35 ] ( 9, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V36 tmp33 [V36 ] ( 9, 17 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
+; V37 tmp34 [V37 ] ( 9, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V38 tmp35 [V38 ] ( 9, 17 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V39 tmp36 [V39,T05] ( 2, 2 ) long -> r14 "field V21._00 (fldOffset=0x0)" P-INDEP
+; V40 tmp37 [V40,T06] ( 2, 2 ) long -> rax "field V22._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp38 [V41 ] ( 9, 9 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V42 tmp39 [V42 ] ( 9, 17 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 9, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V44 tmp41 [V44 ] ( 9, 17 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V45 cse0 [V45,T01] ( 9, 9 ) int -> rsi "CSE #01: aggressive"
;
-; Lcl frame size = 128
+; Lcl frame size = 152
G_M41805_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
+ push r15
+ push r14
+ push rbx
+ sub rsp, 152
+ lea rbp, [rsp+0xB0]
+ mov r15, rdi
+ mov ebx, esi
vmovups ymm0, ymmword ptr [rbp+0x10]
- ;; size=21 bbWeight=1 PerfScore 5.75
+ ;; size=31 bbWeight=1 PerfScore 9.25
G_M41805_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm1, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm1
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x20], rax
- movzx rax, byte ptr [rbp-0x20]
+ vmovaps xmmword ptr [rbp-0x40], xmm1
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rdi
+ movzx rdi, byte ptr [rbp-0x50]
+ mov esi, ebx
and esi, 7
- shlx eax, eax, esi
- mov byte ptr [rbp-0x18], al
- movzx rax, byte ptr [rbp-0x1F]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x17], al
- movzx rax, byte ptr [rbp-0x1E]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x16], al
- movzx rax, byte ptr [rbp-0x1D]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x15], al
- movzx rax, byte ptr [rbp-0x1C]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x14], al
- movzx rax, byte ptr [rbp-0x1B]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x13], al
- movzx rax, byte ptr [rbp-0x1A]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x12], al
- movzx rax, byte ptr [rbp-0x19]
- shlx eax, eax, esi
- mov byte ptr [rbp-0x11], al
- mov rax, qword ptr [rbp-0x18]
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x30], rcx
- movzx rcx, byte ptr [rbp-0x30]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x28], cl
- movzx rcx, byte ptr [rbp-0x2F]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x27], cl
- movzx rcx, byte ptr [rbp-0x2E]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x26], cl
- movzx rcx, byte ptr [rbp-0x2D]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x25], cl
- movzx rcx, byte ptr [rbp-0x2C]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x24], cl
- movzx rcx, byte ptr [rbp-0x2B]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x23], cl
- movzx rcx, byte ptr [rbp-0x2A]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x22], cl
- movzx rcx, byte ptr [rbp-0x29]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x21], cl
- mov rcx, qword ptr [rbp-0x28]
- mov qword ptr [rbp-0x40], rax
- mov qword ptr [rbp-0x38], rcx
- ;; size=236 bbWeight=1 PerfScore 49.50
-G_M41805_IG03:
- vmovaps xmm1, xmmword ptr [rbp-0x40]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- mov rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x60], rax
- movzx rax, byte ptr [rbp-0x60]
- shlx eax, eax, esi
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x48], dil
+ movzx rdi, byte ptr [rbp-0x4F]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x47], dil
+ movzx rdi, byte ptr [rbp-0x4E]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x46], dil
+ movzx rdi, byte ptr [rbp-0x4D]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x45], dil
+ movzx rdi, byte ptr [rbp-0x4C]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x44], dil
+ movzx rdi, byte ptr [rbp-0x4B]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x43], dil
+ movzx rdi, byte ptr [rbp-0x4A]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x42], dil
+ movzx rdi, byte ptr [rbp-0x49]
+ shlx edi, edi, esi
+ mov byte ptr [rbp-0x41], dil
+ mov r14, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x60], rdi
+ movzx rdi, byte ptr [rbp-0x60]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x58], al
- movzx rax, byte ptr [rbp-0x5F]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5F]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x57], al
- movzx rax, byte ptr [rbp-0x5E]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5E]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x56], al
- movzx rax, byte ptr [rbp-0x5D]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5D]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ ;; size=236 bbWeight=1 PerfScore 48.75
+G_M41805_IG03:
mov byte ptr [rbp-0x55], al
- movzx rax, byte ptr [rbp-0x5C]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x54], al
- movzx rax, byte ptr [rbp-0x5B]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5B]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x53], al
- movzx rax, byte ptr [rbp-0x5A]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5A]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x52], al
- movzx rax, byte ptr [rbp-0x59]
- shlx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x59]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
mov byte ptr [rbp-0x51], al
- mov rax, qword ptr [rbp-0x58]
- mov rcx, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x70], rcx
- movzx rcx, byte ptr [rbp-0x70]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x68], cl
- movzx rcx, byte ptr [rbp-0x6F]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x67], cl
- movzx rcx, byte ptr [rbp-0x6E]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x66], cl
- movzx rcx, byte ptr [rbp-0x6D]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x65], cl
- movzx rcx, byte ptr [rbp-0x6C]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x64], cl
- movzx rcx, byte ptr [rbp-0x6B]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x63], cl
- movzx rcx, byte ptr [rbp-0x6A]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x62], cl
- movzx rcx, byte ptr [rbp-0x69]
- shlx ecx, ecx, esi
- mov byte ptr [rbp-0x61], cl
- mov rcx, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0x80], rax
- mov qword ptr [rbp-0x78], rcx
- ;; size=240 bbWeight=1 PerfScore 54.00
+ mov rdi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], r14
+ mov qword ptr [rbp-0x68], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x70]
+ vmovaps xmmword ptr [rbp-0x30], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x80], xmm1
+ mov rdi, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0x90], rdi
+ movzx rdi, byte ptr [rbp-0x90]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x88], al
+ movzx rdi, byte ptr [rbp-0x8F]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x87], al
+ movzx rdi, byte ptr [rbp-0x8E]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x86], al
+ movzx rdi, byte ptr [rbp-0x8D]
+ mov esi, ebx
+ ;; size=234 bbWeight=1 PerfScore 56.75
G_M41805_IG04:
- vinserti128 ymm0, ymm1, xmmword ptr [rbp-0x80], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=14 bbWeight=1 PerfScore 6.25
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x85], al
+ movzx rdi, byte ptr [rbp-0x8C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x84], al
+ movzx rdi, byte ptr [rbp-0x8B]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x83], al
+ movzx rdi, byte ptr [rbp-0x8A]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x82], al
+ movzx rdi, byte ptr [rbp-0x89]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x81], al
+ mov r14, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xA0], rdi
+ movzx rdi, byte ptr [rbp-0xA0]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x98], al
+ movzx rdi, byte ptr [rbp-0x9F]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x97], al
+ movzx rdi, byte ptr [rbp-0x9E]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x96], al
+ movzx rdi, byte ptr [rbp-0x9D]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ ;; size=252 bbWeight=1 PerfScore 47.25
G_M41805_IG05:
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x95], al
+ movzx rdi, byte ptr [rbp-0x9C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x94], al
+ movzx rdi, byte ptr [rbp-0x9B]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x93], al
+ movzx rdi, byte ptr [rbp-0x9A]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x92], al
+ movzx rdi, byte ptr [rbp-0x99]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftLeft(ubyte,int):ubyte
+ mov byte ptr [rbp-0x91], al
+ mov rax, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], r14
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm0, xmmword ptr [rbp-0x30]
+ vinserti128 ymm0, ymm0, xmmword ptr [rbp-0xB0], 1
+ vmovups ymmword ptr [r15], ymm0
+ mov rax, r15
+ ;; size=164 bbWeight=1 PerfScore 38.25
+G_M41805_IG06:
vzeroupper
- add rsp, 128
+ add rsp, 152
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 523, prolog size 16, PerfScore 118.25, instruction count 129, allocated bytes for code 523 (MethodHash=5ba45cb2) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftLeft(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)
+; Total bytes of code 934, prolog size 21, PerfScore 204.50, instruction count 190, allocated bytes for code 934 (MethodHash=5ba45cb2) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftLeft(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)411 (78.59 % of base) - System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 16 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 16 single block inlinees; 10 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V02 arg1 [V02,T02] ( 3, 3 ) int -> rsi single-def
+; V00 RetBuf [V00,T02] ( 4, 4 ) byref -> r15 single-def
+; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V02 arg1 [V02,T00] ( 27, 27 ) int -> rbx single-def
;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> mm1 "impAppendStmt"
+; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V07 tmp4 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V09 tmp6 [V09 ] ( 9, 9 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V09 tmp6 [V09 ] ( 9, 9 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V10 tmp7 [V10,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V11 tmp8 [V11 ] ( 9, 18 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V11 tmp8 [V11 ] ( 9, 18 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V12 tmp9 [V12 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
;* V13 tmp10 [V13 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V14 tmp11 [V14 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V15 tmp12 [V15 ] ( 9, 9 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V15 tmp12 [V15 ] ( 9, 9 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V16 tmp13 [V16,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V17 tmp14 [V17 ] ( 9, 18 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V17 tmp14 [V17 ] ( 9, 18 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V18 tmp15 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V19 tmp16 [V19 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V20 tmp17 [V20 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V21 tmp18 [V21,T14] ( 3, 3 ) simd16 -> [rbp-0x40] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V22 tmp19 [V22,T12] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V24 tmp21 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V25 tmp22 [V25 ] ( 9, 9 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V26 tmp23 [V26,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V27 tmp24 [V27 ] ( 9, 18 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V28 tmp25 [V28 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V29 tmp26 [V29 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V30 tmp27 [V30 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V31 tmp28 [V31 ] ( 9, 9 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V32 tmp29 [V32,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 9, 18 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V34 tmp31 [V34 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V35 tmp32 [V35 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V36 tmp33 [V36 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V37 tmp34 [V37,T15] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V38 tmp35 [V38 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V39 tmp36 [V39,T03] ( 2, 2 ) long -> rax "field V07._00 (fldOffset=0x0)" P-INDEP
-; V40 tmp37 [V40,T04] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V41 tmp38 [V41 ] ( 9, 9 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V42 tmp39 [V42 ] ( 9, 17 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V43 tmp40 [V43 ] ( 9, 9 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V44 tmp41 [V44 ] ( 9, 17 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V45 tmp42 [V45,T05] ( 2, 2 ) long -> rax "field V23._00 (fldOffset=0x0)" P-INDEP
-; V46 tmp43 [V46,T06] ( 2, 2 ) long -> rcx "field V24._00 (fldOffset=0x0)" P-INDEP
-; V47 tmp44 [V47 ] ( 9, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp45 [V48 ] ( 9, 17 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
-; V49 tmp46 [V49 ] ( 9, 9 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V50 tmp47 [V50 ] ( 9, 17 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V51 cse0 [V51,T00] ( 33, 33 ) int -> rsi "CSE #01: aggressive"
+; V19 tmp16 [V19,T14] ( 3, 3 ) simd16 -> [rbp-0x70] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V20 tmp17 [V20,T12] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+;* V21 tmp18 [V21 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V22 tmp19 [V22 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V23 tmp20 [V23 ] ( 9, 9 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V24 tmp21 [V24,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 9, 18 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V27 tmp24 [V27 ] ( 9, 9 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V28 tmp25 [V28,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 9, 18 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V30 tmp27 [V30 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V31 tmp28 [V31,T15] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+;* V32 tmp29 [V32 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V33 tmp30 [V33,T03] ( 2, 2 ) long -> r14 "field V07._00 (fldOffset=0x0)" P-INDEP
+; V34 tmp31 [V34,T04] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+; V35 tmp32 [V35 ] ( 9, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V36 tmp33 [V36 ] ( 9, 17 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
+; V37 tmp34 [V37 ] ( 9, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V38 tmp35 [V38 ] ( 9, 17 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V39 tmp36 [V39,T05] ( 2, 2 ) long -> r14 "field V21._00 (fldOffset=0x0)" P-INDEP
+; V40 tmp37 [V40,T06] ( 2, 2 ) long -> rax "field V22._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp38 [V41 ] ( 9, 9 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V42 tmp39 [V42 ] ( 9, 17 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 9, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V44 tmp41 [V44 ] ( 9, 17 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V45 cse0 [V45,T01] ( 9, 9 ) int -> rsi "CSE #01: aggressive"
;
-; Lcl frame size = 128
+; Lcl frame size = 152
G_M45350_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
+ push r15
+ push r14
+ push rbx
+ sub rsp, 152
+ lea rbp, [rsp+0xB0]
+ mov r15, rdi
+ mov ebx, esi
vmovups ymm0, ymmword ptr [rbp+0x10]
- ;; size=21 bbWeight=1 PerfScore 5.75
+ ;; size=31 bbWeight=1 PerfScore 9.25
G_M45350_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm1, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm1
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x20], rax
- movzx rax, byte ptr [rbp-0x20]
+ vmovaps xmmword ptr [rbp-0x40], xmm1
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rdi
+ movzx rdi, byte ptr [rbp-0x50]
+ mov esi, ebx
and esi, 7
- sarx eax, eax, esi
- mov byte ptr [rbp-0x18], al
- movzx rax, byte ptr [rbp-0x1F]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x17], al
- movzx rax, byte ptr [rbp-0x1E]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x16], al
- movzx rax, byte ptr [rbp-0x1D]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x15], al
- movzx rax, byte ptr [rbp-0x1C]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x14], al
- movzx rax, byte ptr [rbp-0x1B]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x13], al
- movzx rax, byte ptr [rbp-0x1A]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x12], al
- movzx rax, byte ptr [rbp-0x19]
- sarx eax, eax, esi
- mov byte ptr [rbp-0x11], al
- mov rax, qword ptr [rbp-0x18]
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x30], rcx
- movzx rcx, byte ptr [rbp-0x30]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x28], cl
- movzx rcx, byte ptr [rbp-0x2F]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x27], cl
- movzx rcx, byte ptr [rbp-0x2E]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x26], cl
- movzx rcx, byte ptr [rbp-0x2D]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x25], cl
- movzx rcx, byte ptr [rbp-0x2C]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x24], cl
- movzx rcx, byte ptr [rbp-0x2B]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x23], cl
- movzx rcx, byte ptr [rbp-0x2A]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x22], cl
- movzx rcx, byte ptr [rbp-0x29]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x21], cl
- mov rcx, qword ptr [rbp-0x28]
- mov qword ptr [rbp-0x40], rax
- mov qword ptr [rbp-0x38], rcx
- ;; size=236 bbWeight=1 PerfScore 49.50
-G_M45350_IG03:
- vmovaps xmm1, xmmword ptr [rbp-0x40]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- mov rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x60], rax
- movzx rax, byte ptr [rbp-0x60]
- sarx eax, eax, esi
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x48], dil
+ movzx rdi, byte ptr [rbp-0x4F]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x47], dil
+ movzx rdi, byte ptr [rbp-0x4E]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x46], dil
+ movzx rdi, byte ptr [rbp-0x4D]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x45], dil
+ movzx rdi, byte ptr [rbp-0x4C]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x44], dil
+ movzx rdi, byte ptr [rbp-0x4B]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x43], dil
+ movzx rdi, byte ptr [rbp-0x4A]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x42], dil
+ movzx rdi, byte ptr [rbp-0x49]
+ sarx edi, edi, esi
+ mov byte ptr [rbp-0x41], dil
+ mov r14, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x60], rdi
+ movzx rdi, byte ptr [rbp-0x60]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x58], al
- movzx rax, byte ptr [rbp-0x5F]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5F]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x57], al
- movzx rax, byte ptr [rbp-0x5E]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5E]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x56], al
- movzx rax, byte ptr [rbp-0x5D]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5D]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ ;; size=236 bbWeight=1 PerfScore 48.75
+G_M45350_IG03:
mov byte ptr [rbp-0x55], al
- movzx rax, byte ptr [rbp-0x5C]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x54], al
- movzx rax, byte ptr [rbp-0x5B]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5B]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x53], al
- movzx rax, byte ptr [rbp-0x5A]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x5A]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x52], al
- movzx rax, byte ptr [rbp-0x59]
- sarx eax, eax, esi
+ movzx rdi, byte ptr [rbp-0x59]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
mov byte ptr [rbp-0x51], al
- mov rax, qword ptr [rbp-0x58]
- mov rcx, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x70], rcx
- movzx rcx, byte ptr [rbp-0x70]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x68], cl
- movzx rcx, byte ptr [rbp-0x6F]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x67], cl
- movzx rcx, byte ptr [rbp-0x6E]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x66], cl
- movzx rcx, byte ptr [rbp-0x6D]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x65], cl
- movzx rcx, byte ptr [rbp-0x6C]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x64], cl
- movzx rcx, byte ptr [rbp-0x6B]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x63], cl
- movzx rcx, byte ptr [rbp-0x6A]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x62], cl
- movzx rcx, byte ptr [rbp-0x69]
- sarx ecx, ecx, esi
- mov byte ptr [rbp-0x61], cl
- mov rcx, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0x80], rax
- mov qword ptr [rbp-0x78], rcx
- ;; size=240 bbWeight=1 PerfScore 54.00
+ mov rdi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], r14
+ mov qword ptr [rbp-0x68], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x70]
+ vmovaps xmmword ptr [rbp-0x30], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x80], xmm1
+ mov rdi, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0x90], rdi
+ movzx rdi, byte ptr [rbp-0x90]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x88], al
+ movzx rdi, byte ptr [rbp-0x8F]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x87], al
+ movzx rdi, byte ptr [rbp-0x8E]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x86], al
+ movzx rdi, byte ptr [rbp-0x8D]
+ mov esi, ebx
+ ;; size=234 bbWeight=1 PerfScore 56.75
G_M45350_IG04:
- vinserti128 ymm0, ymm1, xmmword ptr [rbp-0x80], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=14 bbWeight=1 PerfScore 6.25
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x85], al
+ movzx rdi, byte ptr [rbp-0x8C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x84], al
+ movzx rdi, byte ptr [rbp-0x8B]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x83], al
+ movzx rdi, byte ptr [rbp-0x8A]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x82], al
+ movzx rdi, byte ptr [rbp-0x89]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x81], al
+ mov r14, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xA0], rdi
+ movzx rdi, byte ptr [rbp-0xA0]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x98], al
+ movzx rdi, byte ptr [rbp-0x9F]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x97], al
+ movzx rdi, byte ptr [rbp-0x9E]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x96], al
+ movzx rdi, byte ptr [rbp-0x9D]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ ;; size=252 bbWeight=1 PerfScore 47.25
G_M45350_IG05:
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x95], al
+ movzx rdi, byte ptr [rbp-0x9C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x94], al
+ movzx rdi, byte ptr [rbp-0x9B]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x93], al
+ movzx rdi, byte ptr [rbp-0x9A]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x92], al
+ movzx rdi, byte ptr [rbp-0x99]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:ShiftRightArithmetic(ubyte,int):ubyte
+ mov byte ptr [rbp-0x91], al
+ mov rax, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], r14
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm0, xmmword ptr [rbp-0x30]
+ vinserti128 ymm0, ymm0, xmmword ptr [rbp-0xB0], 1
+ vmovups ymmword ptr [r15], ymm0
+ mov rax, r15
+ ;; size=164 bbWeight=1 PerfScore 38.25
+G_M45350_IG06:
vzeroupper
- add rsp, 128
+ add rsp, 152
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 523, prolog size 16, PerfScore 118.25, instruction count 129, allocated bytes for code 523 (MethodHash=9dbe4ed9) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)
+; Total bytes of code 934, prolog size 21, PerfScore 204.50, instruction count 190, allocated bytes for code 934 (MethodHash=9dbe4ed9) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[ubyte],int):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)406 (82.35 % of base) - System.Runtime.Intrinsics.Vector512:Dot[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):long
; Assembly listing for method System.Runtime.Intrinsics.Vector512:Dot[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):long (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
-; rsp based frame
+; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 22 single block inlinees; 23 inlinees without PGO data
+; 0 inlinees with PGO data; 43 single block inlinees; 24 inlinees without PGO data
; Final local variable assignments
;
;* V00 arg0 [V00 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[long]>
;* V01 arg1 [V01 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[long]>
-;# V02 OutArgs [V02 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-;* V03 tmp1 [V03 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V04 tmp2 [V04 ] ( 0, 0 ) simd32 -> zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-;* V05 tmp3 [V05 ] ( 0, 0 ) simd32 -> zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V06 tmp4 [V06,T00] ( 2, 4 ) long -> rax "impAppendStmt"
-; V07 tmp5 [V07,T08] ( 3, 6 ) simd16 -> [rsp+0xF0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V08 tmp6 [V08,T09] ( 3, 6 ) simd16 -> [rsp+0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V09 tmp7 [V09,T01] ( 2, 4 ) long -> rax "impAppendStmt"
-;* V10 tmp8 [V10 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V11 tmp9 [V11 ] ( 2, 4 ) struct ( 8) [rsp+0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V12 tmp10 [V12 ] ( 2, 4 ) struct ( 8) [rsp+0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V13 tmp11 [V13 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V14 tmp12 [V14 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V15 tmp13 [V15 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V16 tmp14 [V16 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V17 tmp15 [V17 ] ( 2, 4 ) struct ( 8) [rsp+0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V18 tmp16 [V18 ] ( 2, 4 ) struct ( 8) [rsp+0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V19 tmp17 [V19 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V20 tmp18 [V20 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V21 tmp19 [V21 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V22 tmp20 [V22 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V23 tmp21 [V23 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V24 tmp22 [V24,T10] ( 3, 6 ) simd16 -> [rsp+0xB0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V25 tmp23 [V25,T11] ( 3, 6 ) simd16 -> [rsp+0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V26 tmp24 [V26,T02] ( 2, 4 ) long -> rcx "impAppendStmt"
-;* V27 tmp25 [V27 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V28 tmp26 [V28 ] ( 2, 4 ) struct ( 8) [rsp+0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V29 tmp27 [V29 ] ( 2, 4 ) struct ( 8) [rsp+0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V30 tmp28 [V30 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V31 tmp29 [V31 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V32 tmp30 [V32 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V33 tmp31 [V33 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V34 tmp32 [V34 ] ( 2, 4 ) struct ( 8) [rsp+0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V35 tmp33 [V35 ] ( 2, 4 ) struct ( 8) [rsp+0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V36 tmp34 [V36 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V37 tmp35 [V37 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V38 tmp36 [V38 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V39 tmp37 [V39 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V40 tmp38 [V40 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V41 tmp39 [V41,T06] ( 2, 2 ) long -> rax "Inline return value spill temp"
-;* V42 tmp40 [V42 ] ( 0, 0 ) simd32 -> zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-;* V43 tmp41 [V43 ] ( 0, 0 ) simd32 -> zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V44 tmp42 [V44,T03] ( 2, 4 ) long -> rcx "impAppendStmt"
-; V45 tmp43 [V45,T12] ( 3, 6 ) simd16 -> [rsp+0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V46 tmp44 [V46,T13] ( 3, 6 ) simd16 -> [rsp+0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V47 tmp45 [V47,T04] ( 2, 4 ) long -> rcx "impAppendStmt"
-;* V48 tmp46 [V48 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V49 tmp47 [V49 ] ( 2, 4 ) struct ( 8) [rsp+0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V50 tmp48 [V50 ] ( 2, 4 ) struct ( 8) [rsp+0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V51 tmp49 [V51 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V52 tmp50 [V52 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V53 tmp51 [V53 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V54 tmp52 [V54 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V55 tmp53 [V55 ] ( 2, 4 ) struct ( 8) [rsp+0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V56 tmp54 [V56 ] ( 2, 4 ) struct ( 8) [rsp+0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V57 tmp55 [V57 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+; V02 OutArgs [V02 ] ( 1, 1 ) struct (16) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V03 tmp1 [V03,T00] ( 2, 4 ) long -> rbx "impAppendStmt"
+; V04 tmp2 [V04,T50] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V05 tmp3 [V05,T51] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V06 tmp4 [V06,T01] ( 2, 4 ) long -> rbx "impAppendStmt"
+; V07 tmp5 [V07,T52] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V08 tmp6 [V08,T53] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V09 tmp7 [V09,T02] ( 2, 4 ) long -> rbx "impAppendStmt"
+;* V10 tmp8 [V10 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V11 tmp9 [V11 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V12 tmp10 [V12 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V13 tmp11 [V13 ] ( 2, 2 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V14 tmp12 [V14,T35] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V15 tmp13 [V15 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V16 tmp14 [V16 ] ( 2, 4 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V17 tmp15 [V17 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V18 tmp16 [V18 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V19 tmp17 [V19 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V20 tmp18 [V20 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V21 tmp19 [V21,T09] ( 2, 2 ) long -> rbx ld-addr-op "Inline ldloca(s) first use temp"
+;* V22 tmp20 [V22,T36] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V23 tmp21 [V23 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V24 tmp22 [V24 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V25 tmp23 [V25 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V26 tmp24 [V26 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+;* V27 tmp25 [V27 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V28 tmp26 [V28 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V29 tmp27 [V29 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V30 tmp28 [V30 ] ( 2, 2 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V31 tmp29 [V31,T37] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V32 tmp30 [V32 ] ( 2, 4 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V33 tmp31 [V33 ] ( 2, 4 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V34 tmp32 [V34 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V35 tmp33 [V35 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V36 tmp34 [V36 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V37 tmp35 [V37 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V38 tmp36 [V38,T10] ( 2, 2 ) long -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V39 tmp37 [V39,T38] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V40 tmp38 [V40 ] ( 2, 4 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V41 tmp39 [V41 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V42 tmp40 [V42 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V43 tmp41 [V43 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+;* V44 tmp42 [V44 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V45 tmp43 [V45 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V46 tmp44 [V46,T54] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V47 tmp45 [V47,T55] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V48 tmp46 [V48,T03] ( 2, 4 ) long -> r15 "impAppendStmt"
+;* V49 tmp47 [V49 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V50 tmp48 [V50 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V51 tmp49 [V51 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V52 tmp50 [V52 ] ( 2, 2 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V53 tmp51 [V53,T39] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V54 tmp52 [V54 ] ( 2, 4 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V55 tmp53 [V55 ] ( 2, 4 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V56 tmp54 [V56 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V57 tmp55 [V57 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
;* V58 tmp56 [V58 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
;* V59 tmp57 [V59 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V60 tmp58 [V60 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V61 tmp59 [V61 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V62 tmp60 [V62,T14] ( 3, 6 ) simd16 -> [rsp+0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V63 tmp61 [V63,T15] ( 3, 6 ) simd16 -> [rsp+0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V64 tmp62 [V64,T05] ( 2, 4 ) long -> rdx "impAppendStmt"
-;* V65 tmp63 [V65 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V66 tmp64 [V66 ] ( 2, 4 ) struct ( 8) [rsp+0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V67 tmp65 [V67 ] ( 2, 4 ) struct ( 8) [rsp+0x10] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V68 tmp66 [V68 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V69 tmp67 [V69 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V70 tmp68 [V70 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V71 tmp69 [V71 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V72 tmp70 [V72 ] ( 2, 4 ) struct ( 8) [rsp+0x08] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V73 tmp71 [V73 ] ( 2, 4 ) struct ( 8) [rsp+0x00] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V74 tmp72 [V74 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+; V60 tmp58 [V60,T11] ( 2, 2 ) long -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V61 tmp59 [V61,T40] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V62 tmp60 [V62 ] ( 2, 4 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V63 tmp61 [V63 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V64 tmp62 [V64 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V65 tmp63 [V65 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+;* V66 tmp64 [V66 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V67 tmp65 [V67 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V68 tmp66 [V68 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V69 tmp67 [V69 ] ( 2, 2 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V70 tmp68 [V70,T41] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V71 tmp69 [V71 ] ( 2, 4 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V72 tmp70 [V72 ] ( 2, 4 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V73 tmp71 [V73 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V74 tmp72 [V74 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
;* V75 tmp73 [V75 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
;* V76 tmp74 [V76 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V77 tmp75 [V77 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V78 tmp76 [V78 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V79 tmp77 [V79,T07] ( 2, 2 ) long -> rcx "Inline return value spill temp"
-;* V80 tmp78 [V80 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-; V81 tmp79 [V81,T16] ( 2, 2 ) simd32 -> mm0 single-def "field V00._lower (fldOffset=0x0)" P-INDEP
-; V82 tmp80 [V82,T17] ( 2, 2 ) simd32 -> mm1 single-def "field V00._upper (fldOffset=0x20)" P-INDEP
-; V83 tmp81 [V83,T18] ( 2, 2 ) simd32 -> mm2 single-def "field V01._lower (fldOffset=0x0)" P-INDEP
-; V84 tmp82 [V84,T19] ( 2, 2 ) simd32 -> mm3 single-def "field V01._upper (fldOffset=0x20)" P-INDEP
-; V85 tmp83 [V85 ] ( 2, 3 ) long -> [rsp+0xD8] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V86 tmp84 [V86 ] ( 2, 3 ) long -> [rsp+0xD0] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
-; V87 tmp85 [V87 ] ( 2, 3 ) long -> [rsp+0xC8] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V88 tmp86 [V88 ] ( 2, 3 ) long -> [rsp+0xC0] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
-; V89 tmp87 [V89 ] ( 2, 3 ) long -> [rsp+0x98] do-not-enreg[X] addr-exposed "field V28._00 (fldOffset=0x0)" P-DEP
-; V90 tmp88 [V90 ] ( 2, 3 ) long -> [rsp+0x90] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
-; V91 tmp89 [V91 ] ( 2, 3 ) long -> [rsp+0x88] do-not-enreg[X] addr-exposed "field V34._00 (fldOffset=0x0)" P-DEP
-; V92 tmp90 [V92 ] ( 2, 3 ) long -> [rsp+0x80] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
-; V93 tmp91 [V93 ] ( 2, 3 ) long -> [rsp+0x58] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
-; V94 tmp92 [V94 ] ( 2, 3 ) long -> [rsp+0x50] do-not-enreg[X] addr-exposed "field V50._00 (fldOffset=0x0)" P-DEP
-; V95 tmp93 [V95 ] ( 2, 3 ) long -> [rsp+0x48] do-not-enreg[X] addr-exposed "field V55._00 (fldOffset=0x0)" P-DEP
-; V96 tmp94 [V96 ] ( 2, 3 ) long -> [rsp+0x40] do-not-enreg[X] addr-exposed "field V56._00 (fldOffset=0x0)" P-DEP
-; V97 tmp95 [V97 ] ( 2, 3 ) long -> [rsp+0x18] do-not-enreg[X] addr-exposed "field V66._00 (fldOffset=0x0)" P-DEP
-; V98 tmp96 [V98 ] ( 2, 3 ) long -> [rsp+0x10] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
-; V99 tmp97 [V99 ] ( 2, 3 ) long -> [rsp+0x08] do-not-enreg[X] addr-exposed "field V72._00 (fldOffset=0x0)" P-DEP
-; V100 tmp98 [V100 ] ( 2, 3 ) long -> [rsp+0x00] do-not-enreg[X] addr-exposed "field V73._00 (fldOffset=0x0)" P-DEP
+; V77 tmp75 [V77,T12] ( 2, 2 ) long -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V78 tmp76 [V78,T42] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V79 tmp77 [V79 ] ( 2, 4 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V80 tmp78 [V80 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+; V81 tmp79 [V81,T56] ( 3, 6 ) simd32 -> [rbp-0x110] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V82 tmp80 [V82,T57] ( 3, 6 ) simd32 -> [rbp-0x130] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V83 tmp81 [V83,T04] ( 2, 4 ) long -> r15 "impAppendStmt"
+; V84 tmp82 [V84,T58] ( 3, 6 ) simd16 -> [rbp-0x140] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V85 tmp83 [V85,T59] ( 3, 6 ) simd16 -> [rbp-0x150] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V86 tmp84 [V86,T05] ( 2, 4 ) long -> r15 "impAppendStmt"
+;* V87 tmp85 [V87 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V88 tmp86 [V88 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V89 tmp87 [V89 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V90 tmp88 [V90 ] ( 2, 2 ) struct ( 8) [rbp-0x158] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V91 tmp89 [V91,T43] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V92 tmp90 [V92 ] ( 2, 4 ) struct ( 8) [rbp-0x160] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V93 tmp91 [V93 ] ( 2, 4 ) struct ( 8) [rbp-0x168] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V94 tmp92 [V94 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V95 tmp93 [V95 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V96 tmp94 [V96,T13] ( 2, 2 ) long -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V97 tmp95 [V97,T44] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V98 tmp96 [V98 ] ( 2, 4 ) struct ( 8) [rbp-0x170] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V99 tmp97 [V99 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V100 tmp98 [V100 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V101 tmp99 [V101 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V102 tmp100 [V102 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V103 tmp101 [V103 ] ( 2, 2 ) struct ( 8) [rbp-0x178] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V104 tmp102 [V104,T45] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V105 tmp103 [V105 ] ( 2, 4 ) struct ( 8) [rbp-0x180] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V106 tmp104 [V106 ] ( 2, 4 ) struct ( 8) [rbp-0x188] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V107 tmp105 [V107 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V108 tmp106 [V108 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V109 tmp107 [V109,T14] ( 2, 2 ) long -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V110 tmp108 [V110,T46] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V111 tmp109 [V111 ] ( 2, 4 ) struct ( 8) [rbp-0x190] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V112 tmp110 [V112 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+; V113 tmp111 [V113,T60] ( 3, 6 ) simd16 -> [rbp-0x1A0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V114 tmp112 [V114,T61] ( 3, 6 ) simd16 -> [rbp-0x1B0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V115 tmp113 [V115,T06] ( 2, 4 ) long -> r14 "impAppendStmt"
+;* V116 tmp114 [V116 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V117 tmp115 [V117 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V118 tmp116 [V118 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V119 tmp117 [V119 ] ( 2, 2 ) struct ( 8) [rbp-0x1B8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V120 tmp118 [V120,T47] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V121 tmp119 [V121 ] ( 2, 4 ) struct ( 8) [rbp-0x1C0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V122 tmp120 [V122 ] ( 2, 4 ) struct ( 8) [rbp-0x1C8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V123 tmp121 [V123 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V124 tmp122 [V124 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V125 tmp123 [V125,T15] ( 2, 2 ) long -> r14 ld-addr-op "Inline ldloca(s) first use temp"
+;* V126 tmp124 [V126,T48] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V127 tmp125 [V127 ] ( 2, 4 ) struct ( 8) [rbp-0x1D0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V128 tmp126 [V128 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V129 tmp127 [V129 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V130 tmp128 [V130 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V131 tmp129 [V131 ] ( 2, 4 ) struct ( 8) [rbp-0x1D8] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V132 tmp130 [V132,T16] ( 2, 2 ) long -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V133 tmp131 [V133,T49] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V134 tmp132 [V134 ] ( 2, 4 ) struct ( 8) [rbp-0x1E0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V135 tmp133 [V135 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V136 tmp134 [V136 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V137 tmp135 [V137 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V138 tmp136 [V138,T62] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V00._lower (fldOffset=0x0)" P-INDEP
+; V139 tmp137 [V139,T63] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V00._upper (fldOffset=0x20)" P-INDEP
+; V140 tmp138 [V140,T64] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
+; V141 tmp139 [V141,T65] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
+; V142 tmp140 [V142,T18] ( 2, 2 ) long -> rsi "field V10._00 (fldOffset=0x0)" P-INDEP
+; V143 tmp141 [V143,T19] ( 2, 2 ) long -> rdi "field V11._00 (fldOffset=0x0)" P-INDEP
+;* V144 tmp142 [V144 ] ( 0, 0 ) long -> zero-ref "field V12._00 (fldOffset=0x0)" P-INDEP
+; V145 tmp143 [V145 ] ( 2, 2 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V146 tmp144 [V146 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V147 tmp145 [V147 ] ( 2, 3 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V148 tmp146 [V148 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V149 tmp147 [V149,T20] ( 2, 2 ) long -> rsi "field V27._00 (fldOffset=0x0)" P-INDEP
+; V150 tmp148 [V150,T21] ( 2, 2 ) long -> rdi "field V28._00 (fldOffset=0x0)" P-INDEP
+;* V151 tmp149 [V151 ] ( 0, 0 ) long -> zero-ref "field V29._00 (fldOffset=0x0)" P-INDEP
+; V152 tmp150 [V152 ] ( 2, 2 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V30._00 (fldOffset=0x0)" P-DEP
+; V153 tmp151 [V153 ] ( 2, 3 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V32._00 (fldOffset=0x0)" P-DEP
+; V154 tmp152 [V154 ] ( 2, 3 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V155 tmp153 [V155 ] ( 2, 3 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V40._00 (fldOffset=0x0)" P-DEP
+; V156 tmp154 [V156,T22] ( 2, 2 ) long -> rsi "field V49._00 (fldOffset=0x0)" P-INDEP
+; V157 tmp155 [V157,T23] ( 2, 2 ) long -> rdi "field V50._00 (fldOffset=0x0)" P-INDEP
+;* V158 tmp156 [V158 ] ( 0, 0 ) long -> zero-ref "field V51._00 (fldOffset=0x0)" P-INDEP
+; V159 tmp157 [V159 ] ( 2, 2 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V52._00 (fldOffset=0x0)" P-DEP
+; V160 tmp158 [V160 ] ( 2, 3 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V54._00 (fldOffset=0x0)" P-DEP
+; V161 tmp159 [V161 ] ( 2, 3 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V55._00 (fldOffset=0x0)" P-DEP
+; V162 tmp160 [V162 ] ( 2, 3 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V62._00 (fldOffset=0x0)" P-DEP
+; V163 tmp161 [V163,T24] ( 2, 2 ) long -> rsi "field V66._00 (fldOffset=0x0)" P-INDEP
+; V164 tmp162 [V164,T25] ( 2, 2 ) long -> rdi "field V67._00 (fldOffset=0x0)" P-INDEP
+;* V165 tmp163 [V165 ] ( 0, 0 ) long -> zero-ref "field V68._00 (fldOffset=0x0)" P-INDEP
+; V166 tmp164 [V166 ] ( 2, 2 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V69._00 (fldOffset=0x0)" P-DEP
+; V167 tmp165 [V167 ] ( 2, 3 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V71._00 (fldOffset=0x0)" P-DEP
+; V168 tmp166 [V168 ] ( 2, 3 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V72._00 (fldOffset=0x0)" P-DEP
+; V169 tmp167 [V169 ] ( 2, 3 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V79._00 (fldOffset=0x0)" P-DEP
+; V170 tmp168 [V170,T26] ( 2, 2 ) long -> rdi "field V87._00 (fldOffset=0x0)" P-INDEP
+; V171 tmp169 [V171,T27] ( 2, 2 ) long -> rsi "field V88._00 (fldOffset=0x0)" P-INDEP
+;* V172 tmp170 [V172 ] ( 0, 0 ) long -> zero-ref "field V89._00 (fldOffset=0x0)" P-INDEP
+; V173 tmp171 [V173 ] ( 2, 2 ) long -> [rbp-0x158] do-not-enreg[X] addr-exposed "field V90._00 (fldOffset=0x0)" P-DEP
+; V174 tmp172 [V174 ] ( 2, 3 ) long -> [rbp-0x160] do-not-enreg[X] addr-exposed "field V92._00 (fldOffset=0x0)" P-DEP
+; V175 tmp173 [V175 ] ( 2, 3 ) long -> [rbp-0x168] do-not-enreg[X] addr-exposed "field V93._00 (fldOffset=0x0)" P-DEP
+; V176 tmp174 [V176 ] ( 2, 3 ) long -> [rbp-0x170] do-not-enreg[X] addr-exposed "field V98._00 (fldOffset=0x0)" P-DEP
+; V177 tmp175 [V177,T28] ( 2, 2 ) long -> rdi "field V100._00 (fldOffset=0x0)" P-INDEP
+; V178 tmp176 [V178,T29] ( 2, 2 ) long -> rsi "field V101._00 (fldOffset=0x0)" P-INDEP
+;* V179 tmp177 [V179 ] ( 0, 0 ) long -> zero-ref "field V102._00 (fldOffset=0x0)" P-INDEP
+; V180 tmp178 [V180 ] ( 2, 2 ) long -> [rbp-0x178] do-not-enreg[X] addr-exposed "field V103._00 (fldOffset=0x0)" P-DEP
+; V181 tmp179 [V181 ] ( 2, 3 ) long -> [rbp-0x180] do-not-enreg[X] addr-exposed "field V105._00 (fldOffset=0x0)" P-DEP
+; V182 tmp180 [V182 ] ( 2, 3 ) long -> [rbp-0x188] do-not-enreg[X] addr-exposed "field V106._00 (fldOffset=0x0)" P-DEP
+; V183 tmp181 [V183 ] ( 2, 3 ) long -> [rbp-0x190] do-not-enreg[X] addr-exposed "field V111._00 (fldOffset=0x0)" P-DEP
+; V184 tmp182 [V184,T30] ( 2, 2 ) long -> rdi "field V116._00 (fldOffset=0x0)" P-INDEP
+; V185 tmp183 [V185,T31] ( 2, 2 ) long -> rsi "field V117._00 (fldOffset=0x0)" P-INDEP
+;* V186 tmp184 [V186 ] ( 0, 0 ) long -> zero-ref "field V118._00 (fldOffset=0x0)" P-INDEP
+; V187 tmp185 [V187 ] ( 2, 2 ) long -> [rbp-0x1B8] do-not-enreg[X] addr-exposed "field V119._00 (fldOffset=0x0)" P-DEP
+; V188 tmp186 [V188 ] ( 2, 3 ) long -> [rbp-0x1C0] do-not-enreg[X] addr-exposed "field V121._00 (fldOffset=0x0)" P-DEP
+; V189 tmp187 [V189 ] ( 2, 3 ) long -> [rbp-0x1C8] do-not-enreg[X] addr-exposed "field V122._00 (fldOffset=0x0)" P-DEP
+; V190 tmp188 [V190 ] ( 2, 3 ) long -> [rbp-0x1D0] do-not-enreg[X] addr-exposed "field V127._00 (fldOffset=0x0)" P-DEP
+;* V191 tmp189 [V191 ] ( 0, 0 ) long -> zero-ref "field V129._00 (fldOffset=0x0)" P-INDEP
+;* V192 tmp190 [V192 ] ( 0, 0 ) long -> zero-ref "field V130._00 (fldOffset=0x0)" P-INDEP
+; V193 tmp191 [V193,T17] ( 2, 3 ) long -> [rbp-0x1D8] do-not-enreg[H] hidden-struct-arg "field V131._00 (fldOffset=0x0)" P-DEP
+; V194 tmp192 [V194 ] ( 2, 3 ) long -> [rbp-0x1E0] do-not-enreg[X] addr-exposed "field V134._00 (fldOffset=0x0)" P-DEP
+; V195 tmp193 [V195,T07] ( 2, 4 ) long -> rsi "argument with side effect"
+; V196 tmp194 [V196,T08] ( 2, 4 ) long -> rsi "argument with side effect"
+; V197 cse0 [V197,T32] ( 2, 2 ) long -> rbx "CSE #01: conservative"
+; V198 cse1 [V198,T33] ( 2, 2 ) long -> r15 "CSE #02: conservative"
+; V199 cse2 [V199,T34] ( 2, 2 ) long -> r15 "CSE #03: conservative"
;
-; Lcl frame size = 264
+; Lcl frame size = 472
G_M64954_IG01:
- sub rsp, 264
- vmovups ymm0, ymmword ptr [rsp+0x110]
- vmovups ymm1, ymmword ptr [rsp+0x130]
- vmovups ymm2, ymmword ptr [rsp+0x150]
- vmovups ymm3, ymmword ptr [rsp+0x170]
- ;; size=43 bbWeight=1 PerfScore 16.25
+ push rbp
+ push r15
+ push r14
+ push rbx
+ sub rsp, 472
+ lea rbp, [rsp+0x1F0]
+ ;; size=21 bbWeight=1 PerfScore 4.75
G_M64954_IG02:
- vmovaps ymm4, ymm0
- vmovaps xmmword ptr [rsp+0xF0], xmm4
- vmovaps ymm4, ymm2
- vmovaps xmmword ptr [rsp+0xE0], xmm4
- mov rax, qword ptr [rsp+0xF0]
- mov qword ptr [rsp+0xD8], rax
- mov rax, qword ptr [rsp+0xE0]
- mov qword ptr [rsp+0xD0], rax
- mov rax, qword ptr [rsp+0xD8]
- imul rax, qword ptr [rsp+0xD0]
- mov rcx, qword ptr [rsp+0xF8]
- mov qword ptr [rsp+0xC8], rcx
- mov rcx, qword ptr [rsp+0xE8]
- mov qword ptr [rsp+0xC0], rcx
- mov rcx, qword ptr [rsp+0xC8]
- imul rcx, qword ptr [rsp+0xC0]
- add rax, rcx
+ vmovups ymm0, ymmword ptr [rbp+0x10]
+ vmovups ymm1, ymmword ptr [rbp+0x50]
+ vmovaps ymm2, ymm0
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ vmovaps ymm2, ymm1
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ mov rsi, qword ptr [rbp-0x30]
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rsi
+ mov qword ptr [rbp-0x58], rdi
+ mov rsi, qword ptr [rbp-0x50]
+ imul rsi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x48], rsi
+ mov rsi, qword ptr [rbp-0x48]
+ mov qword ptr [rbp-0x60], rsi
+ mov rbx, qword ptr [rbp-0x60]
+ mov rsi, qword ptr [rbp-0x28]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x70], rsi
+ mov qword ptr [rbp-0x78], rdi
+ mov rsi, qword ptr [rbp-0x70]
+ imul rsi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0x68], rsi
+ mov rsi, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0x80], rsi
+ mov r15, qword ptr [rbp-0x80]
+ add rbx, r15
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rsp+0xB0], xmm0
- vextractf128 xmm0, ymm2, 1
- vmovaps xmmword ptr [rsp+0xA0], xmm0
- mov rcx, qword ptr [rsp+0xB0]
- mov qword ptr [rsp+0x98], rcx
- mov rcx, qword ptr [rsp+0xA0]
- mov qword ptr [rsp+0x90], rcx
- mov rcx, qword ptr [rsp+0x98]
- imul rcx, qword ptr [rsp+0x90]
- mov rdx, qword ptr [rsp+0xB8]
- mov qword ptr [rsp+0x88], rdx
- mov rdx, qword ptr [rsp+0xA8]
- mov qword ptr [rsp+0x80], rdx
- add rax, rcx
- mov rcx, qword ptr [rsp+0x88]
- imul rcx, qword ptr [rsp+0x80]
- add rax, rcx
- vmovaps ymm0, ymm1
- vmovaps xmmword ptr [rsp+0x70], xmm0
- vmovaps ymm0, ymm3
- vmovaps xmmword ptr [rsp+0x60], xmm0
- mov rcx, qword ptr [rsp+0x70]
- mov qword ptr [rsp+0x58], rcx
- mov rcx, qword ptr [rsp+0x60]
- mov qword ptr [rsp+0x50], rcx
- mov rcx, qword ptr [rsp+0x58]
- imul rcx, qword ptr [rsp+0x50]
- mov rdx, qword ptr [rsp+0x78]
- mov qword ptr [rsp+0x48], rdx
- mov rdx, qword ptr [rsp+0x68]
- mov qword ptr [rsp+0x40], rdx
- mov rdx, qword ptr [rsp+0x48]
- imul rdx, qword ptr [rsp+0x40]
- add rcx, rdx
+ vmovaps xmmword ptr [rbp-0x90], xmm0
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rsp+0x30], xmm0
- vextractf128 xmm0, ymm3, 1
- vmovaps xmmword ptr [rsp+0x20], xmm0
- mov rdx, qword ptr [rsp+0x30]
- mov qword ptr [rsp+0x18], rdx
- mov rdx, qword ptr [rsp+0x20]
- mov qword ptr [rsp+0x10], rdx
- mov rdx, qword ptr [rsp+0x18]
- imul rdx, qword ptr [rsp+0x10]
- mov rdi, qword ptr [rsp+0x38]
- ;; size=406 bbWeight=1 PerfScore 82.00
+ vmovaps xmmword ptr [rbp-0xA0], xmm0
+ mov rsi, qword ptr [rbp-0x90]
+ mov rdi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xB0], rsi
+ mov qword ptr [rbp-0xB8], rdi
+ mov rsi, qword ptr [rbp-0xB0]
+ imul rsi, qword ptr [rbp-0xB8]
+ mov qword ptr [rbp-0xA8], rsi
+ mov rsi, qword ptr [rbp-0xA8]
+ mov qword ptr [rbp-0xC0], rsi
+ mov r15, qword ptr [rbp-0xC0]
+ mov rsi, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xD0], rsi
+ mov qword ptr [rbp-0xD8], rdi
+ mov rsi, qword ptr [rbp-0xD0]
+ imul rsi, qword ptr [rbp-0xD8]
+ mov qword ptr [rbp-0xC8], rsi
+ mov rsi, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0xE0], rsi
+ mov rsi, qword ptr [rbp-0xE0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, r15
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ ;; size=331 bbWeight=1 PerfScore 76.75
G_M64954_IG03:
- mov qword ptr [rsp+0x08], rdi
- mov rdi, qword ptr [rsp+0x28]
- mov qword ptr [rsp], rdi
- add rcx, rdx
- mov rdx, qword ptr [rsp+0x08]
- imul rdx, qword ptr [rsp]
- add rcx, rdx
- add rax, rcx
- ;; size=33 bbWeight=1 PerfScore 8.75
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rbx, rax
+ vmovups ymm0, ymmword ptr [rbp+0x30]
+ vmovups ymmword ptr [rbp-0x110], ymm0
+ vmovups ymm1, ymmword ptr [rbp+0x70]
+ vmovups ymmword ptr [rbp-0x130], ymm1
+ vmovaps ymm2, ymm0
+ vmovaps xmmword ptr [rbp-0x140], xmm2
+ vmovaps ymm2, ymm1
+ vmovaps xmmword ptr [rbp-0x150], xmm2
+ mov rdi, qword ptr [rbp-0x140]
+ mov rsi, qword ptr [rbp-0x150]
+ mov qword ptr [rbp-0x160], rdi
+ mov qword ptr [rbp-0x168], rsi
+ mov rdi, qword ptr [rbp-0x160]
+ mov rsi, qword ptr [rbp-0x168]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x158], rax
+ mov rsi, qword ptr [rbp-0x158]
+ mov qword ptr [rbp-0x170], rsi
+ mov rsi, qword ptr [rbp-0x170]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov r15, rax
+ mov rdi, qword ptr [rbp-0x138]
+ mov rsi, qword ptr [rbp-0x148]
+ mov qword ptr [rbp-0x180], rdi
+ mov qword ptr [rbp-0x188], rsi
+ mov rdi, qword ptr [rbp-0x180]
+ mov rsi, qword ptr [rbp-0x188]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x178], rax
+ mov rsi, qword ptr [rbp-0x178]
+ mov qword ptr [rbp-0x190], rsi
+ mov rsi, qword ptr [rbp-0x190]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, r15
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov r15, rax
+ vmovups ymm0, ymmword ptr [rbp-0x110]
+ vextractf128 xmm0, ymm0, 1
+ vmovaps xmmword ptr [rbp-0x1A0], xmm0
+ vmovups ymm1, ymmword ptr [rbp-0x130]
+ vextractf128 xmm0, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x1B0], xmm0
+ mov rdi, qword ptr [rbp-0x1A0]
+ mov rsi, qword ptr [rbp-0x1B0]
+ mov qword ptr [rbp-0x1C0], rdi
+ mov qword ptr [rbp-0x1C8], rsi
+ mov rdi, qword ptr [rbp-0x1C0]
+ mov rsi, qword ptr [rbp-0x1C8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ ;; size=369 bbWeight=1 PerfScore 76.75
G_M64954_IG04:
+ mov qword ptr [rbp-0x1B8], rax
+ mov rsi, qword ptr [rbp-0x1B8]
+ mov qword ptr [rbp-0x1D0], rsi
+ mov rsi, qword ptr [rbp-0x1D0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov r14, rax
+ mov rdi, qword ptr [rbp-0x198]
+ mov qword ptr [rsp], rdi
+ mov rdi, qword ptr [rbp-0x1A8]
+ mov qword ptr [rsp+0x08], rdi
+ lea rdi, [rbp-0x1D8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64`1[long]:op_Multiply(System.Runtime.Intrinsics.Vector64`1[long],System.Runtime.Intrinsics.Vector64`1[long]):System.Runtime.Intrinsics.Vector64`1[long]
+ call [rax]System.Runtime.Intrinsics.Vector64`1[long]:op_Multiply(System.Runtime.Intrinsics.Vector64`1[long],System.Runtime.Intrinsics.Vector64`1[long]):System.Runtime.Intrinsics.Vector64`1[long]
+ mov rsi, qword ptr [rbp-0x1D8]
+ mov qword ptr [rbp-0x1E0], rsi
+ mov rsi, qword ptr [rbp-0x1E0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, r14
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, r15
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ add rax, rbx
+ ;; size=161 bbWeight=1 PerfScore 29.75
+G_M64954_IG05:
vzeroupper
- add rsp, 264
+ add rsp, 472
+ pop rbx
+ pop r14
+ pop r15
+ pop rbp
ret
- ;; size=11 bbWeight=1 PerfScore 2.25
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 493, prolog size 7, PerfScore 109.25, instruction count 79, allocated bytes for code 493 (MethodHash=fe850245) for method System.Runtime.Intrinsics.Vector512:Dot[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):long (FullOpts)
+; Total bytes of code 899, prolog size 21, PerfScore 192.25, instruction count 164, allocated bytes for code 899 (MethodHash=fe850245) for method System.Runtime.Intrinsics.Vector512:Dot[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):long (FullOpts)294 (88.55 % of base) - System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],short):System.Runtime.Intrinsics.Vector256`1[short]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],short):System.Runtime.Intrinsics.Vector256`1[short] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 16 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 16 single block inlinees; 9 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[short]>
+; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> rbx single-def
+; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[short]>
; V02 arg1 [V02,T02] ( 3, 3 ) short -> rsi single-def
;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> mm1 "impAppendStmt"
+; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
;* V07 tmp4 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V09 tmp6 [V09 ] ( 5, 5 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V09 tmp6 [V09 ] ( 5, 5 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V10 tmp7 [V10,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V11 tmp8 [V11 ] ( 5, 10 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V11 tmp8 [V11 ] ( 5, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V12 tmp9 [V12 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-;* V13 tmp10 [V13 ] ( 0, 0 ) short -> zero-ref "Inline return value spill temp"
-;* V14 tmp11 [V14 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V15 tmp12 [V15 ] ( 5, 5 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V16 tmp13 [V16,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V17 tmp14 [V17 ] ( 5, 10 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V18 tmp15 [V18 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-;* V19 tmp16 [V19 ] ( 0, 0 ) short -> zero-ref "Inline return value spill temp"
-;* V20 tmp17 [V20 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V21 tmp18 [V21,T14] ( 3, 3 ) simd16 -> [rbp-0x40] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V22 tmp19 [V22,T12] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V24 tmp21 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V25 tmp22 [V25 ] ( 5, 5 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V26 tmp23 [V26,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V27 tmp24 [V27 ] ( 5, 10 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V13 tmp10 [V13 ] ( 5, 5 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V14 tmp11 [V14,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V15 tmp12 [V15 ] ( 5, 10 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V16 tmp13 [V16 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
+; V17 tmp14 [V17,T14] ( 3, 3 ) simd16 -> [rbp-0x70] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V18 tmp15 [V18,T12] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+;* V19 tmp16 [V19 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V21 tmp18 [V21 ] ( 5, 5 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V22 tmp19 [V22,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V23 tmp20 [V23 ] ( 5, 10 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V24 tmp21 [V24 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 5, 5 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V26 tmp23 [V26,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V27 tmp24 [V27 ] ( 5, 10 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V28 tmp25 [V28 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-;* V29 tmp26 [V29 ] ( 0, 0 ) short -> zero-ref "Inline return value spill temp"
-;* V30 tmp27 [V30 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V31 tmp28 [V31 ] ( 5, 5 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V32 tmp29 [V32,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 5, 10 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V34 tmp31 [V34 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-;* V35 tmp32 [V35 ] ( 0, 0 ) short -> zero-ref "Inline return value spill temp"
-;* V36 tmp33 [V36 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V37 tmp34 [V37,T15] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-;* V38 tmp35 [V38 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V39 tmp36 [V39,T03] ( 2, 2 ) long -> rsi "field V07._00 (fldOffset=0x0)" P-INDEP
-; V40 tmp37 [V40,T04] ( 2, 2 ) long -> rax "field V08._00 (fldOffset=0x0)" P-INDEP
-; V41 tmp38 [V41 ] ( 5, 5 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V42 tmp39 [V42 ] ( 5, 9 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V43 tmp40 [V43 ] ( 5, 5 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V44 tmp41 [V44 ] ( 5, 9 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V45 tmp42 [V45,T05] ( 2, 2 ) long -> rsi "field V23._00 (fldOffset=0x0)" P-INDEP
-; V46 tmp43 [V46,T06] ( 2, 2 ) long -> rax "field V24._00 (fldOffset=0x0)" P-INDEP
-; V47 tmp44 [V47 ] ( 5, 5 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp45 [V48 ] ( 5, 9 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
-; V49 tmp46 [V49 ] ( 5, 5 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V50 tmp47 [V50 ] ( 5, 9 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V51 cse0 [V51,T00] ( 17, 17 ) int -> rcx "CSE #01: aggressive"
+; V29 tmp26 [V29,T15] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+;* V30 tmp27 [V30 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[short]>
+; V31 tmp28 [V31,T03] ( 2, 2 ) long -> r14 "field V07._00 (fldOffset=0x0)" P-INDEP
+; V32 tmp29 [V32,T04] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+; V33 tmp30 [V33 ] ( 5, 5 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V34 tmp31 [V34 ] ( 5, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
+; V35 tmp32 [V35 ] ( 5, 5 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V36 tmp33 [V36 ] ( 5, 9 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V37 tmp34 [V37,T05] ( 2, 2 ) long -> r14 "field V19._00 (fldOffset=0x0)" P-INDEP
+; V38 tmp35 [V38,T06] ( 2, 2 ) long -> rax "field V20._00 (fldOffset=0x0)" P-INDEP
+; V39 tmp36 [V39 ] ( 5, 5 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
+; V40 tmp37 [V40 ] ( 5, 9 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V41 tmp38 [V41 ] ( 5, 5 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V42 tmp39 [V42 ] ( 5, 9 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V43 cse0 [V43,T00] ( 17, 17 ) int -> r15 "CSE #01: aggressive"
;
-; Lcl frame size = 128
+; Lcl frame size = 152
G_M25609_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
+ push r15
+ push r14
+ push rbx
+ sub rsp, 152
+ lea rbp, [rsp+0xB0]
+ mov rbx, rdi
vmovups ymm0, ymmword ptr [rbp+0x10]
- ;; size=21 bbWeight=1 PerfScore 5.75
+ ;; size=29 bbWeight=1 PerfScore 9.00
G_M25609_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm1, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm1
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x20], rax
- movsx rax, word ptr [rbp-0x20]
- movsx rcx, si
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x18], ax
- movsx rax, word ptr [rbp-0x1E]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x16], ax
- movsx rax, word ptr [rbp-0x1C]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x14], ax
- movsx rax, word ptr [rbp-0x1A]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x12], ax
- mov rsi, qword ptr [rbp-0x18]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x30], rax
- movsx rax, word ptr [rbp-0x30]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x28], ax
- movsx rax, word ptr [rbp-0x2E]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x26], ax
- movsx rax, word ptr [rbp-0x2C]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x24], ax
- movsx rax, word ptr [rbp-0x2A]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x22], ax
- mov rax, qword ptr [rbp-0x28]
- mov qword ptr [rbp-0x40], rsi
- mov qword ptr [rbp-0x38], rax
- vmovaps xmm1, xmmword ptr [rbp-0x40]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- mov rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x60], rax
- movsx rax, word ptr [rbp-0x60]
- cdq
- idiv edx:eax, ecx
+ vmovaps xmmword ptr [rbp-0x40], xmm1
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rdi
+ movsx rdi, word ptr [rbp-0x50]
+ movsx r15, si
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x48], ax
+ movsx rdi, word ptr [rbp-0x4E]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x46], ax
+ movsx rdi, word ptr [rbp-0x4C]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x44], ax
+ movsx rdi, word ptr [rbp-0x4A]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x42], ax
+ mov r14, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x60], rdi
+ movsx rdi, word ptr [rbp-0x60]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
mov word ptr [rbp-0x58], ax
- movsx rax, word ptr [rbp-0x5E]
- cdq
- idiv edx:eax, ecx
+ movsx rdi, word ptr [rbp-0x5E]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
mov word ptr [rbp-0x56], ax
- movsx rax, word ptr [rbp-0x5C]
- cdq
- idiv edx:eax, ecx
+ movsx rdi, word ptr [rbp-0x5C]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
mov word ptr [rbp-0x54], ax
- movsx rax, word ptr [rbp-0x5A]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x52], ax
- ;; size=213 bbWeight=1 PerfScore 371.50
+ movsx rdi, word ptr [rbp-0x5A]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ ;; size=224 bbWeight=1 PerfScore 63.50
G_M25609_IG03:
- mov rsi, qword ptr [rbp-0x58]
- mov rax, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x70], rax
- movsx rax, word ptr [rbp-0x70]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x68], ax
- movsx rax, word ptr [rbp-0x6E]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x66], ax
- movsx rax, word ptr [rbp-0x6C]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x64], ax
- movsx rax, word ptr [rbp-0x6A]
- cdq
- idiv edx:eax, ecx
- mov word ptr [rbp-0x62], ax
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0x80], rsi
- mov qword ptr [rbp-0x78], rax
- vinserti128 ymm0, ymm1, xmmword ptr [rbp-0x80], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=86 bbWeight=1 PerfScore 130.25
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x52], ax
+ mov rdi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], r14
+ mov qword ptr [rbp-0x68], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x70]
+ vmovaps xmmword ptr [rbp-0x30], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x80], xmm1
+ mov rdi, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0x90], rdi
+ movsx rdi, word ptr [rbp-0x90]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x88], ax
+ movsx rdi, word ptr [rbp-0x8E]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x86], ax
+ movsx rdi, word ptr [rbp-0x8C]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x84], ax
+ movsx rdi, word ptr [rbp-0x8A]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x82], ax
+ mov r14, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xA0], rdi
+ movsx rdi, word ptr [rbp-0xA0]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x98], ax
+ movsx rdi, word ptr [rbp-0x9E]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x96], ax
+ movsx rdi, word ptr [rbp-0x9C]
+ mov esi, r15d
+ ;; size=264 bbWeight=1 PerfScore 71.25
G_M25609_IG04:
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x94], ax
+ movsx rdi, word ptr [rbp-0x9A]
+ mov esi, r15d
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x92], ax
+ mov rax, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], r14
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm0, xmmword ptr [rbp-0x30]
+ vinserti128 ymm0, ymm0, xmmword ptr [rbp-0xB0], 1
+ vmovups ymmword ptr [rbx], ymm0
+ mov rax, rbx
+ ;; size=92 bbWeight=1 PerfScore 24.00
+G_M25609_IG05:
vzeroupper
- add rsp, 128
+ add rsp, 152
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 332, prolog size 16, PerfScore 510.25, instruction count 97, allocated bytes for code 332 (MethodHash=57b79bf6) for method System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],short):System.Runtime.Intrinsics.Vector256`1[short] (FullOpts)
+; Total bytes of code 626, prolog size 21, PerfScore 172.00, instruction count 124, allocated bytes for code 626 (MethodHash=57b79bf6) for method System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],short):System.Runtime.Intrinsics.Vector256`1[short] (FullOpts)267 (126.54 % of base) - System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector,T>.Dot(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Dot(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
-; rsp based frame
+; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 11 single block inlinees; 11 inlinees without PGO data
+; 0 inlinees with PGO data; 23 single block inlinees; 9 inlinees without PGO data
; Final local variable assignments
;
-; V00 arg0 [V00,T08] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
-; V01 arg1 [V01,T09] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V00 arg0 [V00,T28] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V01 arg1 [V01,T29] ( 2, 2 ) simd32 -> [rbp+0x30] single-def <System.Runtime.Intrinsics.Vector256`1[long]>
;# V02 OutArgs [V02 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V03 tmp1 [V03,T00] ( 2, 4 ) long -> rax "impAppendStmt"
-; V04 tmp2 [V04,T04] ( 3, 6 ) simd16 -> [rsp+0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V05 tmp3 [V05,T05] ( 3, 6 ) simd16 -> [rsp+0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V06 tmp4 [V06,T01] ( 2, 4 ) long -> rax "impAppendStmt"
-;* V07 tmp5 [V07 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V08 tmp6 [V08 ] ( 2, 4 ) struct ( 8) [rsp+0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V09 tmp7 [V09 ] ( 2, 4 ) struct ( 8) [rsp+0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V10 tmp8 [V10 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V11 tmp9 [V11 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V12 tmp10 [V12 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V13 tmp11 [V13 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V14 tmp12 [V14 ] ( 2, 4 ) struct ( 8) [rsp+0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V15 tmp13 [V15 ] ( 2, 4 ) struct ( 8) [rsp+0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V16 tmp14 [V16 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V17 tmp15 [V17 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V18 tmp16 [V18 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V19 tmp17 [V19 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V20 tmp18 [V20 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V21 tmp19 [V21,T06] ( 3, 6 ) simd16 -> [rsp+0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V22 tmp20 [V22,T07] ( 3, 6 ) simd16 -> [rsp+0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V23 tmp21 [V23,T02] ( 2, 4 ) long -> rcx "impAppendStmt"
-;* V24 tmp22 [V24 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V25 tmp23 [V25 ] ( 2, 4 ) struct ( 8) [rsp+0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V26 tmp24 [V26 ] ( 2, 4 ) struct ( 8) [rsp+0x10] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V27 tmp25 [V27 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V28 tmp26 [V28 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V29 tmp27 [V29 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V30 tmp28 [V30 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V31 tmp29 [V31 ] ( 2, 4 ) struct ( 8) [rsp+0x08] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V32 tmp30 [V32 ] ( 2, 4 ) struct ( 8) [rsp+0x00] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V33 tmp31 [V33 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V34 tmp32 [V34 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V35 tmp33 [V35 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V36 tmp34 [V36 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V37 tmp35 [V37 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V38 tmp36 [V38,T03] ( 2, 2 ) long -> rax "Inline return value spill temp"
-; V39 tmp37 [V39 ] ( 2, 3 ) long -> [rsp+0x58] do-not-enreg[X] addr-exposed "field V08._00 (fldOffset=0x0)" P-DEP
-; V40 tmp38 [V40 ] ( 2, 3 ) long -> [rsp+0x50] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V41 tmp39 [V41 ] ( 2, 3 ) long -> [rsp+0x48] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
-; V42 tmp40 [V42 ] ( 2, 3 ) long -> [rsp+0x40] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V43 tmp41 [V43 ] ( 2, 3 ) long -> [rsp+0x18] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V44 tmp42 [V44 ] ( 2, 3 ) long -> [rsp+0x10] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
-; V45 tmp43 [V45 ] ( 2, 3 ) long -> [rsp+0x08] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V46 tmp44 [V46 ] ( 2, 3 ) long -> [rsp+0x00] do-not-enreg[X] addr-exposed "field V32._00 (fldOffset=0x0)" P-DEP
+; V03 tmp1 [V03,T00] ( 2, 4 ) long -> rbx "impAppendStmt"
+; V04 tmp2 [V04,T24] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V05 tmp3 [V05,T25] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V06 tmp4 [V06,T01] ( 2, 4 ) long -> rbx "impAppendStmt"
+;* V07 tmp5 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V08 tmp6 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V09 tmp7 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V10 tmp8 [V10 ] ( 2, 2 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V11 tmp9 [V11,T16] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V12 tmp10 [V12 ] ( 2, 4 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V13 tmp11 [V13 ] ( 2, 4 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V14 tmp12 [V14 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V15 tmp13 [V15 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V16 tmp14 [V16 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V17 tmp15 [V17 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V18 tmp16 [V18,T04] ( 2, 2 ) long -> rbx ld-addr-op "Inline ldloca(s) first use temp"
+;* V19 tmp17 [V19,T17] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V20 tmp18 [V20 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V21 tmp19 [V21 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V22 tmp20 [V22 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V23 tmp21 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V24 tmp22 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V25 tmp23 [V25 ] ( 2, 2 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V26 tmp24 [V26,T18] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V27 tmp25 [V27 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V28 tmp26 [V28 ] ( 2, 4 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V29 tmp27 [V29 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V30 tmp28 [V30 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V31 tmp29 [V31,T05] ( 2, 2 ) long -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V32 tmp30 [V32,T19] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V33 tmp31 [V33 ] ( 2, 4 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V34 tmp32 [V34 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+; V35 tmp33 [V35,T26] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V36 tmp34 [V36,T27] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V37 tmp35 [V37,T02] ( 2, 4 ) long -> r15 "impAppendStmt"
+;* V38 tmp36 [V38 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V39 tmp37 [V39 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V40 tmp38 [V40 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V41 tmp39 [V41 ] ( 2, 2 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V42 tmp40 [V42,T20] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V43 tmp41 [V43 ] ( 2, 4 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V44 tmp42 [V44 ] ( 2, 4 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V45 tmp43 [V45 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V46 tmp44 [V46 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V47 tmp45 [V47,T06] ( 2, 2 ) long -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V48 tmp46 [V48,T21] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V49 tmp47 [V49 ] ( 2, 4 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V50 tmp48 [V50 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V51 tmp49 [V51 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V52 tmp50 [V52 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V53 tmp51 [V53 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V54 tmp52 [V54 ] ( 2, 2 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V55 tmp53 [V55,T22] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V56 tmp54 [V56 ] ( 2, 4 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V57 tmp55 [V57 ] ( 2, 4 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V58 tmp56 [V58 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V59 tmp57 [V59 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V60 tmp58 [V60,T07] ( 2, 2 ) long -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V61 tmp59 [V61,T23] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V62 tmp60 [V62 ] ( 2, 4 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V63 tmp61 [V63 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+; V64 tmp62 [V64,T08] ( 2, 2 ) long -> rsi "field V07._00 (fldOffset=0x0)" P-INDEP
+; V65 tmp63 [V65,T09] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+;* V66 tmp64 [V66 ] ( 0, 0 ) long -> zero-ref "field V09._00 (fldOffset=0x0)" P-INDEP
+; V67 tmp65 [V67 ] ( 2, 2 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V68 tmp66 [V68 ] ( 2, 3 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V69 tmp67 [V69 ] ( 2, 3 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V70 tmp68 [V70 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V20._00 (fldOffset=0x0)" P-DEP
+; V71 tmp69 [V71,T10] ( 2, 2 ) long -> rdi "field V22._00 (fldOffset=0x0)" P-INDEP
+; V72 tmp70 [V72,T11] ( 2, 2 ) long -> rsi "field V23._00 (fldOffset=0x0)" P-INDEP
+;* V73 tmp71 [V73 ] ( 0, 0 ) long -> zero-ref "field V24._00 (fldOffset=0x0)" P-INDEP
+; V74 tmp72 [V74 ] ( 2, 2 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V75 tmp73 [V75 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V76 tmp74 [V76 ] ( 2, 3 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V28._00 (fldOffset=0x0)" P-DEP
+; V77 tmp75 [V77 ] ( 2, 3 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V78 tmp76 [V78,T12] ( 2, 2 ) long -> rdi "field V38._00 (fldOffset=0x0)" P-INDEP
+; V79 tmp77 [V79,T13] ( 2, 2 ) long -> rsi "field V39._00 (fldOffset=0x0)" P-INDEP
+;* V80 tmp78 [V80 ] ( 0, 0 ) long -> zero-ref "field V40._00 (fldOffset=0x0)" P-INDEP
+; V81 tmp79 [V81 ] ( 2, 2 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V41._00 (fldOffset=0x0)" P-DEP
+; V82 tmp80 [V82 ] ( 2, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V43._00 (fldOffset=0x0)" P-DEP
+; V83 tmp81 [V83 ] ( 2, 3 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V44._00 (fldOffset=0x0)" P-DEP
+; V84 tmp82 [V84 ] ( 2, 3 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
+; V85 tmp83 [V85,T14] ( 2, 2 ) long -> rdi "field V51._00 (fldOffset=0x0)" P-INDEP
+; V86 tmp84 [V86,T15] ( 2, 2 ) long -> rsi "field V52._00 (fldOffset=0x0)" P-INDEP
+;* V87 tmp85 [V87 ] ( 0, 0 ) long -> zero-ref "field V53._00 (fldOffset=0x0)" P-INDEP
+; V88 tmp86 [V88 ] ( 2, 2 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V54._00 (fldOffset=0x0)" P-DEP
+; V89 tmp87 [V89 ] ( 2, 3 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V56._00 (fldOffset=0x0)" P-DEP
+; V90 tmp88 [V90 ] ( 2, 3 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V57._00 (fldOffset=0x0)" P-DEP
+; V91 tmp89 [V91 ] ( 2, 3 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V62._00 (fldOffset=0x0)" P-DEP
+; V92 tmp90 [V92,T03] ( 2, 4 ) long -> rsi "argument with side effect"
;
-; Lcl frame size = 136
+; Lcl frame size = 192
G_M64869_IG01:
- sub rsp, 136
- vmovups ymm0, ymmword ptr [rsp+0x90]
- vmovups ymm1, ymmword ptr [rsp+0xB0]
- ;; size=25 bbWeight=1 PerfScore 8.25
+ push rbp
+ push r15
+ push rbx
+ sub rsp, 192
+ lea rbp, [rsp+0xD0]
+ vmovups ymm0, ymmword ptr [rbp+0x10]
+ vmovups ymm1, ymmword ptr [rbp+0x30]
+ ;; size=29 bbWeight=1 PerfScore 11.75
G_M64869_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rsp+0x70], xmm2
+ vmovaps xmmword ptr [rbp-0x20], xmm2
+ vmovups ymmword ptr [rbp+0x30], ymm1
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rsp+0x60], xmm2
- mov rax, qword ptr [rsp+0x70]
- mov qword ptr [rsp+0x58], rax
- mov rax, qword ptr [rsp+0x60]
- mov qword ptr [rsp+0x50], rax
- mov rax, qword ptr [rsp+0x58]
- imul rax, qword ptr [rsp+0x50]
- mov rcx, qword ptr [rsp+0x78]
- mov qword ptr [rsp+0x48], rcx
- mov rcx, qword ptr [rsp+0x68]
- mov qword ptr [rsp+0x40], rcx
- mov rcx, qword ptr [rsp+0x48]
- imul rcx, qword ptr [rsp+0x40]
- add rax, rcx
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ mov rsi, qword ptr [rbp-0x20]
+ mov rdi, qword ptr [rbp-0x30]
+ mov qword ptr [rbp-0x40], rsi
+ mov qword ptr [rbp-0x48], rdi
+ mov rsi, qword ptr [rbp-0x40]
+ imul rsi, qword ptr [rbp-0x48]
+ mov qword ptr [rbp-0x38], rsi
+ mov rsi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x50], rsi
+ mov rsi, qword ptr [rbp-0x50]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rbx, rax
+ mov rdi, qword ptr [rbp-0x18]
+ mov rsi, qword ptr [rbp-0x28]
+ mov qword ptr [rbp-0x60], rdi
+ mov qword ptr [rbp-0x68], rsi
+ mov rdi, qword ptr [rbp-0x60]
+ mov rsi, qword ptr [rbp-0x68]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x58], rax
+ mov rsi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], rsi
+ mov rsi, qword ptr [rbp-0x70]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rbx, rax
+ vmovups ymm0, ymmword ptr [rbp+0x10]
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rsp+0x30], xmm0
+ vmovaps xmmword ptr [rbp-0x80], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x30]
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rsp+0x20], xmm0
- mov rcx, qword ptr [rsp+0x30]
- mov qword ptr [rsp+0x18], rcx
- mov rcx, qword ptr [rsp+0x20]
- mov qword ptr [rsp+0x10], rcx
- mov rcx, qword ptr [rsp+0x18]
- imul rcx, qword ptr [rsp+0x10]
- mov rdx, qword ptr [rsp+0x38]
- mov qword ptr [rsp+0x08], rdx
- mov rdx, qword ptr [rsp+0x28]
- mov qword ptr [rsp], rdx
- add rax, rcx
- mov rcx, qword ptr [rsp+0x08]
- imul rcx, qword ptr [rsp]
- add rax, rcx
- ;; size=175 bbWeight=1 PerfScore 45.25
+ vmovaps xmmword ptr [rbp-0x90], xmm0
+ mov rdi, qword ptr [rbp-0x80]
+ mov rsi, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xA0], rdi
+ mov qword ptr [rbp-0xA8], rsi
+ mov rdi, qword ptr [rbp-0xA0]
+ mov rsi, qword ptr [rbp-0xA8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x98], rax
+ mov rsi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], rsi
+ mov rsi, qword ptr [rbp-0xB0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ ;; size=299 bbWeight=1 PerfScore 69.75
G_M64869_IG03:
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov r15, rax
+ mov rdi, qword ptr [rbp-0x78]
+ mov rsi, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0xC0], rdi
+ mov qword ptr [rbp-0xC8], rsi
+ mov rdi, qword ptr [rbp-0xC0]
+ mov rsi, qword ptr [rbp-0xC8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0xB8], rax
+ mov rsi, qword ptr [rbp-0xB8]
+ mov qword ptr [rbp-0xD0], rsi
+ mov rsi, qword ptr [rbp-0xD0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, r15
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ nop
+ ;; size=135 bbWeight=1 PerfScore 27.75
+G_M64869_IG04:
vzeroupper
- add rsp, 136
+ add rsp, 192
+ pop rbx
+ pop r15
+ pop rbp
ret
- ;; size=11 bbWeight=1 PerfScore 2.25
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 211, prolog size 7, PerfScore 55.75, instruction count 41, allocated bytes for code 211 (MethodHash=91e4029a) for method System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Dot(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long (FullOpts)
+; Total bytes of code 478, prolog size 19, PerfScore 113.00, instruction count 99, allocated bytes for code 478 (MethodHash=91e4029a) for method System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Dot(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long (FullOpts)266 (35.90 % of base) - System.Runtime.Intrinsics.Vector512:Divide[int](System.Runtime.Intrinsics.Vector512`1[int],System.Runtime.Intrinsics.Vector512`1[int]):System.Runtime.Intrinsics.Vector512`1[int]
; Assembly listing for method System.Runtime.Intrinsics.Vector512:Divide[int](System.Runtime.Intrinsics.Vector512`1[int],System.Runtime.Intrinsics.Vector512`1[int]):System.Runtime.Intrinsics.Vector512`1[int] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 43 single block inlinees; 27 inlinees without PGO data
+; 0 inlinees with PGO data; 38 single block inlinees; 21 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T00] ( 5, 5 ) byref -> rdi single-def
+; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rbx single-def
;* V01 arg0 [V01 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[int]>
;* V02 arg1 [V02 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[int]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T29] ( 2, 4 ) simd32 -> mm0 "impAppendStmt"
-; V05 tmp2 [V05,T30] ( 2, 4 ) simd32 -> mm1 "spilled call-like call argument"
-; V06 tmp3 [V06,T17] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V07 tmp4 [V07,T18] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V08 tmp5 [V08,T31] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T28] ( 2, 4 ) simd32 -> [rbp-0x30] spill-single-def "impAppendStmt"
+; V05 tmp2 [V05,T29] ( 2, 4 ) simd32 -> [rbp-0x50] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument"
+; V06 tmp3 [V06,T16] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
+; V07 tmp4 [V07,T17] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
+; V08 tmp5 [V08,T30] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
;* V09 tmp6 [V09 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V10 tmp7 [V10,T19] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V11 tmp8 [V11,T20] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V10 tmp7 [V10,T18] ( 3, 6 ) simd16 -> [rbp-0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V11 tmp8 [V11,T19] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
;* V12 tmp9 [V12 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V13 tmp10 [V13 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V14 tmp11 [V14 ] ( 3, 3 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V14 tmp11 [V14 ] ( 3, 3 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V15 tmp12 [V15,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V16 tmp13 [V16 ] ( 3, 6 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V17 tmp14 [V17 ] ( 3, 6 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V16 tmp13 [V16 ] ( 3, 6 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V17 tmp14 [V17 ] ( 3, 6 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V18 tmp15 [V18 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V19 tmp16 [V19 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
;* V20 tmp17 [V20 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
;* V21 tmp18 [V21 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V22 tmp19 [V22 ] ( 3, 3 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V22 tmp19 [V22 ] ( 3, 3 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V23 tmp20 [V23,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V24 tmp21 [V24 ] ( 3, 6 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V25 tmp22 [V25 ] ( 3, 6 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V24 tmp21 [V24 ] ( 3, 6 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V25 tmp22 [V25 ] ( 3, 6 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V26 tmp23 [V26 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V27 tmp24 [V27 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
;* V28 tmp25 [V28 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
;* V29 tmp26 [V29 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V30 tmp27 [V30,T33] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V31 tmp28 [V31,T21] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V32 tmp29 [V32,T22] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V30 tmp27 [V30,T32] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V31 tmp28 [V31,T20] ( 3, 6 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V32 tmp29 [V32,T21] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
;* V33 tmp30 [V33 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V34 tmp31 [V34 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V35 tmp32 [V35 ] ( 3, 3 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V35 tmp32 [V35 ] ( 3, 3 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V36 tmp33 [V36,T11] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V37 tmp34 [V37 ] ( 3, 6 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V38 tmp35 [V38 ] ( 3, 6 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V37 tmp34 [V37 ] ( 3, 6 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V38 tmp35 [V38 ] ( 3, 6 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V39 tmp36 [V39 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V40 tmp37 [V40 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
;* V41 tmp38 [V41 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
;* V42 tmp39 [V42 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V43 tmp40 [V43 ] ( 3, 3 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V43 tmp40 [V43 ] ( 3, 3 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V44 tmp41 [V44,T12] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V45 tmp42 [V45 ] ( 3, 6 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V46 tmp43 [V46 ] ( 3, 6 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V45 tmp42 [V45 ] ( 3, 6 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V46 tmp43 [V46 ] ( 3, 6 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V47 tmp44 [V47 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V48 tmp45 [V48 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
;* V49 tmp46 [V49 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
;* V50 tmp47 [V50 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V51 tmp48 [V51,T34] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V51 tmp48 [V51,T33] ( 3, 3 ) simd16 -> [rbp-0x110] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
;* V52 tmp49 [V52 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V53 tmp50 [V53,T23] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V54 tmp51 [V54,T24] ( 3, 6 ) simd32 -> mm2 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V55 tmp52 [V55,T32] ( 2, 4 ) simd16 -> mm3 "impAppendStmt"
+; V53 tmp50 [V53,T22] ( 3, 6 ) simd32 -> [rbp-0x130] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
+; V54 tmp51 [V54,T23] ( 3, 6 ) simd32 -> [rbp-0x150] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[int]>
+; V55 tmp52 [V55,T31] ( 2, 4 ) simd16 -> [rbp-0x160] spill-single-def "impAppendStmt"
;* V56 tmp53 [V56 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V57 tmp54 [V57,T25] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V58 tmp55 [V58,T26] ( 3, 6 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V57 tmp54 [V57,T24] ( 3, 6 ) simd16 -> [rbp-0x170] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V58 tmp55 [V58,T25] ( 3, 6 ) simd16 -> [rbp-0x180] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
;* V59 tmp56 [V59 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V60 tmp57 [V60 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V61 tmp58 [V61 ] ( 3, 3 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V61 tmp58 [V61 ] ( 3, 3 ) struct ( 8) [rbp-0x188] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V62 tmp59 [V62,T13] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V63 tmp60 [V63 ] ( 3, 6 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V64 tmp61 [V64 ] ( 3, 6 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V63 tmp60 [V63 ] ( 3, 6 ) struct ( 8) [rbp-0x190] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V64 tmp61 [V64 ] ( 3, 6 ) struct ( 8) [rbp-0x198] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V65 tmp62 [V65 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V66 tmp63 [V66 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V67 tmp64 [V67 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V68 tmp65 [V68 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V69 tmp66 [V69 ] ( 3, 3 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V70 tmp67 [V70,T14] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V71 tmp68 [V71 ] ( 3, 6 ) struct ( 8) [rbp-0x108] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V72 tmp69 [V72 ] ( 3, 6 ) struct ( 8) [rbp-0x110] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V73 tmp70 [V73 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V74 tmp71 [V74 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V75 tmp72 [V75 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V76 tmp73 [V76 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V77 tmp74 [V77,T35] ( 3, 3 ) simd16 -> [rbp-0x120] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V78 tmp75 [V78,T27] ( 3, 6 ) simd16 -> [rbp-0x130] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V79 tmp76 [V79,T28] ( 3, 6 ) simd16 -> [rbp-0x140] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-;* V80 tmp77 [V80 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V81 tmp78 [V81 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V82 tmp79 [V82 ] ( 3, 3 ) struct ( 8) [rbp-0x148] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V83 tmp80 [V83,T15] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V84 tmp81 [V84 ] ( 3, 6 ) struct ( 8) [rbp-0x150] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V85 tmp82 [V85 ] ( 3, 6 ) struct ( 8) [rbp-0x158] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V86 tmp83 [V86 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V87 tmp84 [V87 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V88 tmp85 [V88 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V89 tmp86 [V89 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V90 tmp87 [V90 ] ( 3, 3 ) struct ( 8) [rbp-0x160] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V91 tmp88 [V91,T16] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V92 tmp89 [V92 ] ( 3, 6 ) struct ( 8) [rbp-0x168] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V93 tmp90 [V93 ] ( 3, 6 ) struct ( 8) [rbp-0x170] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V94 tmp91 [V94 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V95 tmp92 [V95 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V96 tmp93 [V96 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V97 tmp94 [V97 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V98 tmp95 [V98,T36] ( 3, 3 ) simd16 -> [rbp-0x180] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-;* V99 tmp96 [V99 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[int]>
-;* V100 tmp97 [V100 ] ( 0, 0 ) struct (64) zero-ref ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[int]>
-; V101 tmp98 [V101,T39] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
-; V102 tmp99 [V102,T40] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
-; V103 tmp100 [V103,T41] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
-; V104 tmp101 [V104,T42] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
-; V105 tmp102 [V105,T01] ( 2, 2 ) long -> rcx "field V12._00 (fldOffset=0x0)" P-INDEP
-; V106 tmp103 [V106,T02] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
-; V107 tmp104 [V107 ] ( 3, 3 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
-; V108 tmp105 [V108 ] ( 3, 5 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
-; V109 tmp106 [V109 ] ( 3, 5 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V110 tmp107 [V110 ] ( 3, 3 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V22._00 (fldOffset=0x0)" P-DEP
-; V111 tmp108 [V111 ] ( 3, 5 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V24._00 (fldOffset=0x0)" P-DEP
-; V112 tmp109 [V112 ] ( 3, 5 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V113 tmp110 [V113,T03] ( 2, 2 ) long -> rcx "field V33._00 (fldOffset=0x0)" P-INDEP
-; V114 tmp111 [V114,T04] ( 2, 2 ) long -> rax "field V34._00 (fldOffset=0x0)" P-INDEP
-; V115 tmp112 [V115 ] ( 3, 3 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
-; V116 tmp113 [V116 ] ( 3, 5 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
-; V117 tmp114 [V117 ] ( 3, 5 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V38._00 (fldOffset=0x0)" P-DEP
-; V118 tmp115 [V118 ] ( 3, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V43._00 (fldOffset=0x0)" P-DEP
-; V119 tmp116 [V119 ] ( 3, 5 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
-; V120 tmp117 [V120 ] ( 3, 5 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
-; V121 tmp118 [V121,T05] ( 2, 2 ) long -> rcx "field V59._00 (fldOffset=0x0)" P-INDEP
-; V122 tmp119 [V122,T06] ( 2, 2 ) long -> rax "field V60._00 (fldOffset=0x0)" P-INDEP
-; V123 tmp120 [V123 ] ( 3, 3 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V61._00 (fldOffset=0x0)" P-DEP
-; V124 tmp121 [V124 ] ( 3, 5 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V63._00 (fldOffset=0x0)" P-DEP
-; V125 tmp122 [V125 ] ( 3, 5 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V64._00 (fldOffset=0x0)" P-DEP
-; V126 tmp123 [V126 ] ( 3, 3 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V69._00 (fldOffset=0x0)" P-DEP
-; V127 tmp124 [V127 ] ( 3, 5 ) long -> [rbp-0x108] do-not-enreg[X] addr-exposed "field V71._00 (fldOffset=0x0)" P-DEP
-; V128 tmp125 [V128 ] ( 3, 5 ) long -> [rbp-0x110] do-not-enreg[X] addr-exposed "field V72._00 (fldOffset=0x0)" P-DEP
-; V129 tmp126 [V129,T07] ( 2, 2 ) long -> rcx "field V80._00 (fldOffset=0x0)" P-INDEP
-; V130 tmp127 [V130,T08] ( 2, 2 ) long -> rax "field V81._00 (fldOffset=0x0)" P-INDEP
-; V131 tmp128 [V131 ] ( 3, 3 ) long -> [rbp-0x148] do-not-enreg[X] addr-exposed "field V82._00 (fldOffset=0x0)" P-DEP
-; V132 tmp129 [V132 ] ( 3, 5 ) long -> [rbp-0x150] do-not-enreg[X] addr-exposed "field V84._00 (fldOffset=0x0)" P-DEP
-; V133 tmp130 [V133 ] ( 3, 5 ) long -> [rbp-0x158] do-not-enreg[X] addr-exposed "field V85._00 (fldOffset=0x0)" P-DEP
-; V134 tmp131 [V134 ] ( 3, 3 ) long -> [rbp-0x160] do-not-enreg[X] addr-exposed "field V90._00 (fldOffset=0x0)" P-DEP
-; V135 tmp132 [V135 ] ( 3, 5 ) long -> [rbp-0x168] do-not-enreg[X] addr-exposed "field V92._00 (fldOffset=0x0)" P-DEP
-; V136 tmp133 [V136 ] ( 3, 5 ) long -> [rbp-0x170] do-not-enreg[X] addr-exposed "field V93._00 (fldOffset=0x0)" P-DEP
-; V137 tmp134 [V137,T37] ( 2, 2 ) simd32 -> mm0 "field V100._lower (fldOffset=0x0)" P-INDEP
-; V138 tmp135 [V138,T38] ( 2, 2 ) simd32 -> mm1 "field V100._upper (fldOffset=0x20)" P-INDEP
+; V67 tmp64 [V67 ] ( 3, 3 ) struct ( 8) [rbp-0x1A0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V68 tmp65 [V68,T14] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V69 tmp66 [V69 ] ( 3, 6 ) struct ( 8) [rbp-0x1A8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V70 tmp67 [V70 ] ( 3, 6 ) struct ( 8) [rbp-0x1B0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V71 tmp68 [V71 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V72 tmp69 [V72 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V73 tmp70 [V73,T34] ( 3, 3 ) simd16 -> [rbp-0x1C0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V74 tmp71 [V74,T26] ( 3, 6 ) simd16 -> [rbp-0x1D0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V75 tmp72 [V75,T27] ( 3, 6 ) simd16 -> [rbp-0x1E0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+;* V76 tmp73 [V76 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V77 tmp74 [V77 ] ( 2, 4 ) struct ( 8) [rbp-0x1E8] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V78 tmp75 [V78 ] ( 3, 3 ) struct ( 8) [rbp-0x1F0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V79 tmp76 [V79,T15] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V80 tmp77 [V80 ] ( 3, 6 ) struct ( 8) [rbp-0x1F8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V81 tmp78 [V81 ] ( 3, 6 ) struct ( 8) [rbp-0x200] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V82 tmp79 [V82 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V83 tmp80 [V83 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V84 tmp81 [V84 ] ( 3, 3 ) simd16 -> [rbp-0x210] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V85 tmp82 [V85 ] ( 3, 3 ) struct (64) [rbp-0x250] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[int]>
+; V86 tmp83 [V86,T35] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
+; V87 tmp84 [V87,T36] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
+; V88 tmp85 [V88,T37] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
+; V89 tmp86 [V89,T38] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
+; V90 tmp87 [V90,T02] ( 2, 2 ) long -> rdi "field V12._00 (fldOffset=0x0)" P-INDEP
+; V91 tmp88 [V91,T03] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
+; V92 tmp89 [V92 ] ( 3, 3 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
+; V93 tmp90 [V93 ] ( 3, 5 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V94 tmp91 [V94 ] ( 3, 5 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V95 tmp92 [V95 ] ( 3, 3 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V22._00 (fldOffset=0x0)" P-DEP
+; V96 tmp93 [V96 ] ( 3, 5 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V24._00 (fldOffset=0x0)" P-DEP
+; V97 tmp94 [V97 ] ( 3, 5 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V98 tmp95 [V98,T04] ( 2, 2 ) long -> rdi "field V33._00 (fldOffset=0x0)" P-INDEP
+; V99 tmp96 [V99,T05] ( 2, 2 ) long -> rsi "field V34._00 (fldOffset=0x0)" P-INDEP
+; V100 tmp97 [V100 ] ( 3, 3 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V101 tmp98 [V101 ] ( 3, 5 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
+; V102 tmp99 [V102 ] ( 3, 5 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V38._00 (fldOffset=0x0)" P-DEP
+; V103 tmp100 [V103 ] ( 3, 3 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V43._00 (fldOffset=0x0)" P-DEP
+; V104 tmp101 [V104 ] ( 3, 5 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
+; V105 tmp102 [V105 ] ( 3, 5 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
+; V106 tmp103 [V106,T06] ( 2, 2 ) long -> r15 "field V59._00 (fldOffset=0x0)" P-INDEP
+; V107 tmp104 [V107,T07] ( 2, 2 ) long -> rdi "field V60._00 (fldOffset=0x0)" P-INDEP
+; V108 tmp105 [V108 ] ( 3, 3 ) long -> [rbp-0x188] do-not-enreg[X] addr-exposed "field V61._00 (fldOffset=0x0)" P-DEP
+; V109 tmp106 [V109 ] ( 3, 5 ) long -> [rbp-0x190] do-not-enreg[X] addr-exposed "field V63._00 (fldOffset=0x0)" P-DEP
+; V110 tmp107 [V110 ] ( 3, 5 ) long -> [rbp-0x198] do-not-enreg[X] addr-exposed "field V64._00 (fldOffset=0x0)" P-DEP
+; V111 tmp108 [V111 ] ( 3, 3 ) long -> [rbp-0x1A0] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
+; V112 tmp109 [V112 ] ( 3, 5 ) long -> [rbp-0x1A8] do-not-enreg[X] addr-exposed "field V69._00 (fldOffset=0x0)" P-DEP
+; V113 tmp110 [V113 ] ( 3, 5 ) long -> [rbp-0x1B0] do-not-enreg[X] addr-exposed "field V70._00 (fldOffset=0x0)" P-DEP
+; V114 tmp111 [V114,T08] ( 2, 2 ) long -> r15 "field V76._00 (fldOffset=0x0)" P-INDEP
+; V115 tmp112 [V115,T01] ( 2, 3 ) long -> [rbp-0x1E8] do-not-enreg[H] hidden-struct-arg "field V77._00 (fldOffset=0x0)" P-DEP
+; V116 tmp113 [V116 ] ( 3, 3 ) long -> [rbp-0x1F0] do-not-enreg[X] addr-exposed "field V78._00 (fldOffset=0x0)" P-DEP
+; V117 tmp114 [V117 ] ( 3, 5 ) long -> [rbp-0x1F8] do-not-enreg[X] addr-exposed "field V80._00 (fldOffset=0x0)" P-DEP
+; V118 tmp115 [V118 ] ( 3, 5 ) long -> [rbp-0x200] do-not-enreg[X] addr-exposed "field V81._00 (fldOffset=0x0)" P-DEP
+; V119 tmp116 [V119 ] ( 3, 3 ) simd32 -> [rbp-0x250] do-not-enreg[XS] addr-exposed "field V85._lower (fldOffset=0x0)" P-DEP
+; V120 tmp117 [V120 ] ( 2, 2 ) simd32 -> [rbp-0x230] do-not-enreg[XS] addr-exposed "field V85._upper (fldOffset=0x20)" P-DEP
;
-; Lcl frame size = 384
+; Lcl frame size = 608
G_M33333_IG01:
push rbp
- sub rsp, 384
- lea rbp, [rsp+0x180]
- ;; size=16 bbWeight=1 PerfScore 1.75
+ push r15
+ push rbx
+ sub rsp, 608
+ lea rbp, [rsp+0x270]
+ mov rbx, rdi
+ ;; size=22 bbWeight=1 PerfScore 4.00
G_M33333_IG02:
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x50]
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x60], xmm2
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- mov eax, dword ptr [rbp-0x30]
+ vmovaps xmmword ptr [rbp-0x70], xmm2
+ mov rax, qword ptr [rbp-0x60]
+ mov qword ptr [rbp-0x80], rax
+ mov rax, qword ptr [rbp-0x70]
+ mov qword ptr [rbp-0x88], rax
+ mov eax, dword ptr [rbp-0x80]
cdq
- idiv edx:eax, dword ptr [rbp-0x38]
- mov dword ptr [rbp-0x28], eax
- mov eax, dword ptr [rbp-0x2C]
+ idiv edx:eax, dword ptr [rbp-0x88]
+ mov dword ptr [rbp-0x78], eax
+ mov eax, dword ptr [rbp-0x7C]
cdq
- idiv edx:eax, dword ptr [rbp-0x34]
- mov dword ptr [rbp-0x24], eax
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
- mov eax, dword ptr [rbp-0x48]
+ idiv edx:eax, dword ptr [rbp-0x84]
+ mov dword ptr [rbp-0x74], eax
+ mov rdi, qword ptr [rbp-0x78]
+ mov rax, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x98], rax
+ mov rax, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0xA0], rax
+ mov eax, dword ptr [rbp-0x98]
cdq
- idiv edx:eax, dword ptr [rbp-0x50]
- mov dword ptr [rbp-0x40], eax
- mov eax, dword ptr [rbp-0x44]
+ idiv edx:eax, dword ptr [rbp-0xA0]
+ mov dword ptr [rbp-0x90], eax
+ mov eax, dword ptr [rbp-0x94]
cdq
- idiv edx:eax, dword ptr [rbp-0x4C]
- mov dword ptr [rbp-0x3C], eax
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
+ idiv edx:eax, dword ptr [rbp-0x9C]
+ mov dword ptr [rbp-0x8C], eax
+ mov rax, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xB0], rdi
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm2, xmmword ptr [rbp-0xB0]
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
+ vmovaps xmmword ptr [rbp-0xC0], xmm0
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- mov eax, dword ptr [rbp-0x90]
+ vmovaps xmmword ptr [rbp-0xD0], xmm0
+ mov rax, qword ptr [rbp-0xC0]
+ mov qword ptr [rbp-0xE0], rax
+ mov rax, qword ptr [rbp-0xD0]
+ mov qword ptr [rbp-0xE8], rax
+ mov eax, dword ptr [rbp-0xE0]
cdq
- idiv edx:eax, dword ptr [rbp-0x98]
- mov dword ptr [rbp-0x88], eax
- mov eax, dword ptr [rbp-0x8C]
+ idiv edx:eax, dword ptr [rbp-0xE8]
+ mov dword ptr [rbp-0xD8], eax
+ mov eax, dword ptr [rbp-0xDC]
cdq
- idiv edx:eax, dword ptr [rbp-0x94]
- mov dword ptr [rbp-0x84], eax
- mov rcx, qword ptr [rbp-0x88]
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
- mov eax, dword ptr [rbp-0xA8]
+ idiv edx:eax, dword ptr [rbp-0xE4]
+ mov dword ptr [rbp-0xD4], eax
+ mov rdi, qword ptr [rbp-0xD8]
+ mov rax, qword ptr [rbp-0xB8]
+ mov qword ptr [rbp-0xF8], rax
+ mov rax, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0x100], rax
+ mov eax, dword ptr [rbp-0xF8]
cdq
- idiv edx:eax, dword ptr [rbp-0xB0]
- mov dword ptr [rbp-0xA0], eax
- ;; size=251 bbWeight=1 PerfScore 233.00
+ idiv edx:eax, dword ptr [rbp-0x100]
+ mov dword ptr [rbp-0xF0], eax
+ ;; size=314 bbWeight=1 PerfScore 233.00
G_M33333_IG03:
- mov eax, dword ptr [rbp-0xA4]
+ mov eax, dword ptr [rbp-0xF4]
cdq
- idiv edx:eax, dword ptr [rbp-0xAC]
- mov dword ptr [rbp-0x9C], eax
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
+ idiv edx:eax, dword ptr [rbp-0xFC]
+ mov dword ptr [rbp-0xEC], eax
+ mov rsi, qword ptr [rbp-0xF0]
+ mov qword ptr [rbp-0x110], rdi
+ mov qword ptr [rbp-0x108], rsi
+ vinserti128 ymm0, ymm2, xmmword ptr [rbp-0x110], 1
+ vmovups ymmword ptr [rbp-0x30], ymm0
vmovups ymm1, ymmword ptr [rbp+0x30]
+ vmovups ymmword ptr [rbp-0x130], ymm1
vmovups ymm2, ymmword ptr [rbp+0x70]
+ vmovups ymmword ptr [rbp-0x150], ymm2
vmovaps ymm3, ymm1
- vmovaps xmmword ptr [rbp-0xD0], xmm3
+ vmovaps xmmword ptr [rbp-0x170], xmm3
vmovaps ymm3, ymm2
- vmovaps xmmword ptr [rbp-0xE0], xmm3
- mov rax, qword ptr [rbp-0xD0]
- mov qword ptr [rbp-0xF0], rax
- mov rax, qword ptr [rbp-0xE0]
- mov qword ptr [rbp-0xF8], rax
- mov eax, dword ptr [rbp-0xF0]
- cdq
- idiv edx:eax, dword ptr [rbp-0xF8]
- mov dword ptr [rbp-0xE8], eax
- mov eax, dword ptr [rbp-0xEC]
- cdq
- idiv edx:eax, dword ptr [rbp-0xF4]
- mov dword ptr [rbp-0xE4], eax
- mov rcx, qword ptr [rbp-0xE8]
- mov rax, qword ptr [rbp-0xC8]
- mov qword ptr [rbp-0x108], rax
- mov rax, qword ptr [rbp-0xD8]
- mov qword ptr [rbp-0x110], rax
- mov eax, dword ptr [rbp-0x108]
- cdq
- idiv edx:eax, dword ptr [rbp-0x110]
- mov dword ptr [rbp-0x100], eax
- mov eax, dword ptr [rbp-0x104]
- cdq
- idiv edx:eax, dword ptr [rbp-0x10C]
- mov dword ptr [rbp-0xFC], eax
- mov rax, qword ptr [rbp-0x100]
- mov qword ptr [rbp-0x120], rcx
- mov qword ptr [rbp-0x118], rax
- vmovaps xmm3, xmmword ptr [rbp-0x120]
+ vmovaps xmmword ptr [rbp-0x180], xmm3
+ mov rdi, qword ptr [rbp-0x170]
+ mov qword ptr [rbp-0x190], rdi
+ mov rdi, qword ptr [rbp-0x180]
+ mov qword ptr [rbp-0x198], rdi
+ mov edi, dword ptr [rbp-0x190]
+ mov esi, dword ptr [rbp-0x198]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x188], eax
+ mov edi, dword ptr [rbp-0x18C]
+ mov esi, dword ptr [rbp-0x194]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x184], eax
+ mov r15, qword ptr [rbp-0x188]
+ mov rdi, qword ptr [rbp-0x168]
+ mov qword ptr [rbp-0x1A8], rdi
+ mov rdi, qword ptr [rbp-0x178]
+ mov qword ptr [rbp-0x1B0], rdi
+ mov edi, dword ptr [rbp-0x1A8]
+ mov esi, dword ptr [rbp-0x1B0]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x1A0], eax
+ mov edi, dword ptr [rbp-0x1A4]
+ mov esi, dword ptr [rbp-0x1AC]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x19C], eax
+ mov rdi, qword ptr [rbp-0x1A0]
+ mov qword ptr [rbp-0x1C0], r15
+ mov qword ptr [rbp-0x1B8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x1C0]
+ vmovaps xmmword ptr [rbp-0x160], xmm0
+ ;; size=325 bbWeight=1 PerfScore 89.00
+G_M33333_IG04:
+ vmovups ymm1, ymmword ptr [rbp-0x130]
vextractf128 xmm1, ymm1, 1
- vmovaps xmmword ptr [rbp-0x130], xmm1
+ vmovaps xmmword ptr [rbp-0x1D0], xmm1
+ vmovups ymm2, ymmword ptr [rbp-0x150]
vextractf128 xmm1, ymm2, 1
- vmovaps xmmword ptr [rbp-0x140], xmm1
- mov rax, qword ptr [rbp-0x130]
- mov qword ptr [rbp-0x150], rax
- mov rax, qword ptr [rbp-0x140]
- mov qword ptr [rbp-0x158], rax
- mov eax, dword ptr [rbp-0x150]
- cdq
- idiv edx:eax, dword ptr [rbp-0x158]
- mov dword ptr [rbp-0x148], eax
- mov eax, dword ptr [rbp-0x14C]
- cdq
- idiv edx:eax, dword ptr [rbp-0x154]
- mov dword ptr [rbp-0x144], eax
- mov rcx, qword ptr [rbp-0x148]
- ;; size=353 bbWeight=1 PerfScore 236.00
-G_M33333_IG04:
- mov rax, qword ptr [rbp-0x128]
- mov qword ptr [rbp-0x168], rax
- mov rax, qword ptr [rbp-0x138]
- mov qword ptr [rbp-0x170], rax
- mov eax, dword ptr [rbp-0x168]
- cdq
- idiv edx:eax, dword ptr [rbp-0x170]
- mov dword ptr [rbp-0x160], eax
- mov eax, dword ptr [rbp-0x164]
- cdq
- idiv edx:eax, dword ptr [rbp-0x16C]
- mov dword ptr [rbp-0x15C], eax
- mov rax, qword ptr [rbp-0x160]
- mov qword ptr [rbp-0x180], rcx
- mov qword ptr [rbp-0x178], rax
- vinserti128 ymm1, ymm3, xmmword ptr [rbp-0x180], 1
- vmovups ymmword ptr [rdi], ymm0
- vmovups ymmword ptr [rdi+0x20], ymm1
- mov rax, rdi
- ;; size=109 bbWeight=1 PerfScore 70.25
+ vmovaps xmmword ptr [rbp-0x1E0], xmm1
+ mov rdi, qword ptr [rbp-0x1D0]
+ mov qword ptr [rbp-0x1F8], rdi
+ mov rdi, qword ptr [rbp-0x1E0]
+ mov qword ptr [rbp-0x200], rdi
+ mov edi, dword ptr [rbp-0x1F8]
+ mov esi, dword ptr [rbp-0x200]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x1F0], eax
+ mov edi, dword ptr [rbp-0x1F4]
+ mov esi, dword ptr [rbp-0x1FC]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x1EC], eax
+ mov r15, qword ptr [rbp-0x1F0]
+ mov rdi, qword ptr [rbp-0x1C8]
+ mov qword ptr [rsp], rdi
+ mov rdi, qword ptr [rbp-0x1D8]
+ mov qword ptr [rsp+0x08], rdi
+ lea rdi, [rbp-0x1E8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64`1[int]:op_Division(System.Runtime.Intrinsics.Vector64`1[int],System.Runtime.Intrinsics.Vector64`1[int]):System.Runtime.Intrinsics.Vector64`1[int]
+ call [rax]System.Runtime.Intrinsics.Vector64`1[int]:op_Division(System.Runtime.Intrinsics.Vector64`1[int],System.Runtime.Intrinsics.Vector64`1[int]):System.Runtime.Intrinsics.Vector64`1[int]
+ mov qword ptr [rbp-0x210], r15
+ mov rdi, qword ptr [rbp-0x1E8]
+ mov qword ptr [rsp], rdi
+ lea rdi, [rbp-0x210]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[int](byref,System.Runtime.Intrinsics.Vector64`1[int])
+ call [rax]System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[int](byref,System.Runtime.Intrinsics.Vector64`1[int])
+ vmovaps xmm0, xmmword ptr [rbp-0x160]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0x210]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ lea rdi, [rbp-0x50]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[int](System.Runtime.Intrinsics.Vector128`1[int],System.Runtime.Intrinsics.Vector128`1[int]):System.Runtime.Intrinsics.Vector256`1[int]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[int](System.Runtime.Intrinsics.Vector128`1[int],System.Runtime.Intrinsics.Vector128`1[int]):System.Runtime.Intrinsics.Vector256`1[int]
+ vmovups ymm0, ymmword ptr [rbp-0x30]
+ vmovups ymmword ptr [rbp-0x250], ymm0
+ vmovups ymm0, ymmword ptr [rbp-0x50]
+ vmovups ymmword ptr [rsp], ymm0
+ lea rdi, [rbp-0x250]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[int](byref,System.Runtime.Intrinsics.Vector256`1[int])
+ ;; size=301 bbWeight=1 PerfScore 68.50
G_M33333_IG05:
+ call [rax]System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[int](byref,System.Runtime.Intrinsics.Vector256`1[int])
+ vmovdqu ymm0, ymmword ptr [rbp-0x250]
+ vmovdqu ymmword ptr [rbx], ymm0
+ vmovdqu ymm0, ymmword ptr [rbp-0x230]
+ vmovdqu ymmword ptr [rbx+0x20], ymm0
+ mov rax, rbx
+ ;; size=30 bbWeight=1 PerfScore 15.25
+G_M33333_IG06:
vzeroupper
- add rsp, 384
+ add rsp, 608
+ pop rbx
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 741, prolog size 16, PerfScore 543.75, instruction count 146, allocated bytes for code 741 (MethodHash=c5217dca) for method System.Runtime.Intrinsics.Vector512:Divide[int](System.Runtime.Intrinsics.Vector512`1[int],System.Runtime.Intrinsics.Vector512`1[int]):System.Runtime.Intrinsics.Vector512`1[int] (FullOpts)
+; Total bytes of code 1007, prolog size 19, PerfScore 413.50, instruction count 176, allocated bytes for code 1007 (MethodHash=c5217dca) for method System.Runtime.Intrinsics.Vector512:Divide[int](System.Runtime.Intrinsics.Vector512`1[int],System.Runtime.Intrinsics.Vector512`1[int]):System.Runtime.Intrinsics.Vector512`1[int] (FullOpts)237 (27.15 % of base) - System.Runtime.Intrinsics.Vector512:Divide[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):System.Runtime.Intrinsics.Vector512`1[ubyte]
; Assembly listing for method System.Runtime.Intrinsics.Vector512:Divide[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):System.Runtime.Intrinsics.Vector512`1[ubyte] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; fully interruptible
; No PGO data
-; 0 inlinees with PGO data; 43 single block inlinees; 27 inlinees without PGO data
+; 0 inlinees with PGO data; 37 single block inlinees; 21 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T16] ( 5, 5 ) byref -> rdi single-def
+; V00 RetBuf [V00,T14] ( 4, 4 ) byref -> rbx single-def
;* V01 arg0 [V01 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[ubyte]>
;* V02 arg1 [V02 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[ubyte]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T37] ( 2, 4 ) simd32 -> mm0 "impAppendStmt"
-; V05 tmp2 [V05,T38] ( 2, 4 ) simd32 -> mm1 "spilled call-like call argument"
-; V06 tmp3 [V06,T25] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V07 tmp4 [V07,T26] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V08 tmp5 [V08,T39] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T35] ( 2, 4 ) simd32 -> [rbp-0x50] spill-single-def "impAppendStmt"
+; V05 tmp2 [V05,T36] ( 2, 4 ) simd32 -> [rbp-0x70] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument"
+; V06 tmp3 [V06,T23] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V07 tmp4 [V07,T24] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V08 tmp5 [V08,T37] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
;* V09 tmp6 [V09 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V10 tmp7 [V10,T27] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V11 tmp8 [V11,T28] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V10 tmp7 [V10,T25] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V11 tmp8 [V11,T26] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V12 tmp9 [V12 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V13 tmp10 [V13 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V14 tmp11 [V14 ] ( 2, 5 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V15 tmp12 [V15,T00] ( 5, 17 ) int -> rcx "Inline stloc first use temp"
-; V16 tmp13 [V16 ] ( 2, 10 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V17 tmp14 [V17 ] ( 2, 10 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V14 tmp11 [V14 ] ( 2, 5 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V15 tmp12 [V15,T00] ( 5, 17 ) int -> rdi "Inline stloc first use temp"
+; V16 tmp13 [V16 ] ( 2, 10 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V17 tmp14 [V17 ] ( 2, 10 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V18 tmp15 [V18 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V19 tmp16 [V19 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
;* V20 tmp17 [V20 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V21 tmp18 [V21 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V22 tmp19 [V22 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V23 tmp20 [V23 ] ( 2, 5 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V23 tmp20 [V23 ] ( 2, 5 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
; V24 tmp21 [V24,T01] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
-; V25 tmp22 [V25 ] ( 2, 10 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V26 tmp23 [V26 ] ( 2, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V25 tmp22 [V25 ] ( 2, 10 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V26 tmp23 [V26 ] ( 2, 10 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V27 tmp24 [V27 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V28 tmp25 [V28 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
;* V29 tmp26 [V29 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V30 tmp27 [V30 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V31 tmp28 [V31 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V32 tmp29 [V32,T41] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V33 tmp30 [V33,T29] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V34 tmp31 [V34,T30] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V32 tmp29 [V32,T39] ( 3, 3 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V33 tmp30 [V33,T27] ( 3, 6 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V34 tmp31 [V34,T28] ( 3, 6 ) simd16 -> [rbp-0xF0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V35 tmp32 [V35 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V36 tmp33 [V36 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V37 tmp34 [V37 ] ( 2, 5 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V38 tmp35 [V38,T02] ( 5, 17 ) int -> rcx "Inline stloc first use temp"
-; V39 tmp36 [V39 ] ( 2, 10 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V40 tmp37 [V40 ] ( 2, 10 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V37 tmp34 [V37 ] ( 2, 5 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V38 tmp35 [V38,T02] ( 5, 17 ) int -> rdi "Inline stloc first use temp"
+; V39 tmp36 [V39 ] ( 2, 10 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V40 tmp37 [V40 ] ( 2, 10 ) struct ( 8) [rbp-0x108] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V41 tmp38 [V41 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V42 tmp39 [V42 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
;* V43 tmp40 [V43 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V44 tmp41 [V44 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V45 tmp42 [V45 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V46 tmp43 [V46 ] ( 2, 5 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V46 tmp43 [V46 ] ( 2, 5 ) struct ( 8) [rbp-0x110] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
; V47 tmp44 [V47,T03] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
-; V48 tmp45 [V48 ] ( 2, 10 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V49 tmp46 [V49 ] ( 2, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V48 tmp45 [V48 ] ( 2, 10 ) struct ( 8) [rbp-0x118] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V49 tmp46 [V49 ] ( 2, 10 ) struct ( 8) [rbp-0x120] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V50 tmp47 [V50 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V51 tmp48 [V51 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
;* V52 tmp49 [V52 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
;* V53 tmp50 [V53 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
;* V54 tmp51 [V54 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V55 tmp52 [V55,T42] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V55 tmp52 [V55,T40] ( 3, 3 ) simd16 -> [rbp-0x130] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V56 tmp53 [V56 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V57 tmp54 [V57,T31] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V58 tmp55 [V58,T32] ( 3, 6 ) simd32 -> mm2 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V59 tmp56 [V59,T40] ( 2, 4 ) simd16 -> mm3 "impAppendStmt"
+; V57 tmp54 [V57,T29] ( 3, 6 ) simd32 -> [rbp-0x150] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V58 tmp55 [V58,T30] ( 3, 6 ) simd32 -> [rbp-0x170] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V59 tmp56 [V59,T38] ( 2, 4 ) simd16 -> [rbp-0x180] spill-single-def "impAppendStmt"
;* V60 tmp57 [V60 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V61 tmp58 [V61,T33] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V62 tmp59 [V62,T34] ( 3, 6 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V61 tmp58 [V61,T31] ( 3, 6 ) simd16 -> [rbp-0x190] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V62 tmp59 [V62,T32] ( 3, 6 ) simd16 -> [rbp-0x1A0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V63 tmp60 [V63 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V64 tmp61 [V64 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V65 tmp62 [V65 ] ( 2, 5 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V66 tmp63 [V66,T04] ( 5, 17 ) int -> rcx "Inline stloc first use temp"
-; V67 tmp64 [V67 ] ( 2, 10 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V68 tmp65 [V68 ] ( 2, 10 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V65 tmp62 [V65 ] ( 2, 5 ) struct ( 8) [rbp-0x1A8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V66 tmp63 [V66,T04] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V67 tmp64 [V67 ] ( 2, 10 ) struct ( 8) [rbp-0x1B0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V68 tmp65 [V68 ] ( 2, 10 ) struct ( 8) [rbp-0x1B8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V69 tmp66 [V69 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V70 tmp67 [V70 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V71 tmp68 [V71 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V72 tmp69 [V72 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V73 tmp70 [V73 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V74 tmp71 [V74 ] ( 2, 5 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V75 tmp72 [V75,T05] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
-; V76 tmp73 [V76 ] ( 2, 10 ) struct ( 8) [rbp-0x108] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V77 tmp74 [V77 ] ( 2, 10 ) struct ( 8) [rbp-0x110] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V78 tmp75 [V78 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V79 tmp76 [V79 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V80 tmp77 [V80 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V81 tmp78 [V81 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V82 tmp79 [V82 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V83 tmp80 [V83,T43] ( 3, 3 ) simd16 -> [rbp-0x120] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V84 tmp81 [V84,T35] ( 3, 6 ) simd16 -> [rbp-0x130] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V85 tmp82 [V85,T36] ( 3, 6 ) simd16 -> [rbp-0x140] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V86 tmp83 [V86 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V87 tmp84 [V87 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V88 tmp85 [V88 ] ( 2, 5 ) struct ( 8) [rbp-0x148] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V89 tmp86 [V89,T06] ( 5, 17 ) int -> rcx "Inline stloc first use temp"
-; V90 tmp87 [V90 ] ( 2, 10 ) struct ( 8) [rbp-0x150] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V91 tmp88 [V91 ] ( 2, 10 ) struct ( 8) [rbp-0x158] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V92 tmp89 [V92 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V93 tmp90 [V93 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V94 tmp91 [V94 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V95 tmp92 [V95 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V96 tmp93 [V96 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V97 tmp94 [V97 ] ( 2, 5 ) struct ( 8) [rbp-0x160] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V98 tmp95 [V98,T07] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
-; V99 tmp96 [V99 ] ( 2, 10 ) struct ( 8) [rbp-0x168] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V100 tmp97 [V100 ] ( 2, 10 ) struct ( 8) [rbp-0x170] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V101 tmp98 [V101 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V102 tmp99 [V102 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V103 tmp100 [V103 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V104 tmp101 [V104 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V105 tmp102 [V105 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V106 tmp103 [V106,T44] ( 3, 3 ) simd16 -> [rbp-0x180] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V107 tmp104 [V107 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-;* V108 tmp105 [V108 ] ( 0, 0 ) struct (64) zero-ref ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[ubyte]>
-; V109 tmp106 [V109,T47] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
-; V110 tmp107 [V110,T48] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
-; V111 tmp108 [V111,T49] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
-; V112 tmp109 [V112,T50] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
-; V113 tmp110 [V113,T17] ( 2, 2 ) long -> rcx "field V12._00 (fldOffset=0x0)" P-INDEP
-; V114 tmp111 [V114,T18] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
-; V115 tmp112 [V115 ] ( 2, 5 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
-; V116 tmp113 [V116 ] ( 2, 9 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
-; V117 tmp114 [V117 ] ( 2, 9 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V118 tmp115 [V118 ] ( 2, 5 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
-; V119 tmp116 [V119 ] ( 2, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V120 tmp117 [V120 ] ( 2, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
-; V121 tmp118 [V121,T19] ( 2, 2 ) long -> rcx "field V35._00 (fldOffset=0x0)" P-INDEP
-; V122 tmp119 [V122,T20] ( 2, 2 ) long -> rax "field V36._00 (fldOffset=0x0)" P-INDEP
-; V123 tmp120 [V123 ] ( 2, 5 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
-; V124 tmp121 [V124 ] ( 2, 9 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V39._00 (fldOffset=0x0)" P-DEP
-; V125 tmp122 [V125 ] ( 2, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V40._00 (fldOffset=0x0)" P-DEP
-; V126 tmp123 [V126 ] ( 2, 5 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
-; V127 tmp124 [V127 ] ( 2, 9 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V48._00 (fldOffset=0x0)" P-DEP
-; V128 tmp125 [V128 ] ( 2, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
-; V129 tmp126 [V129,T21] ( 2, 2 ) long -> rcx "field V63._00 (fldOffset=0x0)" P-INDEP
-; V130 tmp127 [V130,T22] ( 2, 2 ) long -> rax "field V64._00 (fldOffset=0x0)" P-INDEP
-; V131 tmp128 [V131 ] ( 2, 5 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V65._00 (fldOffset=0x0)" P-DEP
-; V132 tmp129 [V132 ] ( 2, 9 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
-; V133 tmp130 [V133 ] ( 2, 9 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V68._00 (fldOffset=0x0)" P-DEP
-; V134 tmp131 [V134 ] ( 2, 5 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V74._00 (fldOffset=0x0)" P-DEP
-; V135 tmp132 [V135 ] ( 2, 9 ) long -> [rbp-0x108] do-not-enreg[X] addr-exposed "field V76._00 (fldOffset=0x0)" P-DEP
-; V136 tmp133 [V136 ] ( 2, 9 ) long -> [rbp-0x110] do-not-enreg[X] addr-exposed "field V77._00 (fldOffset=0x0)" P-DEP
-; V137 tmp134 [V137,T23] ( 2, 2 ) long -> rcx "field V86._00 (fldOffset=0x0)" P-INDEP
-; V138 tmp135 [V138,T24] ( 2, 2 ) long -> rax "field V87._00 (fldOffset=0x0)" P-INDEP
-; V139 tmp136 [V139 ] ( 2, 5 ) long -> [rbp-0x148] do-not-enreg[X] addr-exposed "field V88._00 (fldOffset=0x0)" P-DEP
-; V140 tmp137 [V140 ] ( 2, 9 ) long -> [rbp-0x150] do-not-enreg[X] addr-exposed "field V90._00 (fldOffset=0x0)" P-DEP
-; V141 tmp138 [V141 ] ( 2, 9 ) long -> [rbp-0x158] do-not-enreg[X] addr-exposed "field V91._00 (fldOffset=0x0)" P-DEP
-; V142 tmp139 [V142 ] ( 2, 5 ) long -> [rbp-0x160] do-not-enreg[X] addr-exposed "field V97._00 (fldOffset=0x0)" P-DEP
-; V143 tmp140 [V143 ] ( 2, 9 ) long -> [rbp-0x168] do-not-enreg[X] addr-exposed "field V99._00 (fldOffset=0x0)" P-DEP
-; V144 tmp141 [V144 ] ( 2, 9 ) long -> [rbp-0x170] do-not-enreg[X] addr-exposed "field V100._00 (fldOffset=0x0)" P-DEP
-; V145 tmp142 [V145,T45] ( 2, 2 ) simd32 -> mm0 "field V108._lower (fldOffset=0x0)" P-INDEP
-; V146 tmp143 [V146,T46] ( 2, 2 ) simd32 -> mm1 "field V108._upper (fldOffset=0x20)" P-INDEP
-; V147 cse0 [V147,T08] ( 4, 16 ) long -> rsi "CSE #01: moderate"
-; V148 cse1 [V148,T09] ( 4, 16 ) long -> r8 "CSE #02: moderate"
-; V149 cse2 [V149,T10] ( 4, 16 ) long -> rsi "CSE #03: moderate"
-; V150 cse3 [V150,T11] ( 4, 16 ) long -> r8 "CSE #04: moderate"
-; V151 cse4 [V151,T12] ( 4, 16 ) long -> rsi "CSE #05: moderate"
-; V152 cse5 [V152,T13] ( 4, 16 ) long -> r8 "CSE #06: moderate"
-; V153 cse6 [V153,T14] ( 4, 16 ) long -> rsi "CSE #07: moderate"
-; V154 cse7 [V154,T15] ( 4, 16 ) long -> r8 "CSE #08: moderate"
+; V71 tmp68 [V71 ] ( 2, 5 ) struct ( 8) [rbp-0x1C0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V72 tmp69 [V72,T05] ( 5, 17 ) int -> r14 "Inline stloc first use temp"
+; V73 tmp70 [V73 ] ( 2, 10 ) struct ( 8) [rbp-0x1C8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V74 tmp71 [V74 ] ( 2, 10 ) struct ( 8) [rbp-0x1D0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V75 tmp72 [V75 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V76 tmp73 [V76 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V77 tmp74 [V77,T41] ( 3, 3 ) simd16 -> [rbp-0x1E0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V78 tmp75 [V78,T33] ( 3, 6 ) simd16 -> [rbp-0x1F0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V79 tmp76 [V79,T34] ( 3, 6 ) simd16 -> [rbp-0x200] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+;* V80 tmp77 [V80 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V81 tmp78 [V81 ] ( 2, 4 ) struct ( 8) [rbp-0x208] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V82 tmp79 [V82 ] ( 2, 5 ) struct ( 8) [rbp-0x210] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V83 tmp80 [V83,T06] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V84 tmp81 [V84 ] ( 2, 10 ) struct ( 8) [rbp-0x218] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V85 tmp82 [V85 ] ( 2, 10 ) struct ( 8) [rbp-0x220] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V86 tmp83 [V86 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V87 tmp84 [V87 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V88 tmp85 [V88 ] ( 3, 3 ) simd16 -> [rbp-0x230] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V89 tmp86 [V89 ] ( 3, 3 ) struct (64) [rbp-0x270] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[ubyte]>
+; V90 tmp87 [V90,T42] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
+; V91 tmp88 [V91,T43] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
+; V92 tmp89 [V92,T44] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
+; V93 tmp90 [V93,T45] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
+; V94 tmp91 [V94,T16] ( 2, 2 ) long -> rdi "field V12._00 (fldOffset=0x0)" P-INDEP
+; V95 tmp92 [V95,T17] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
+; V96 tmp93 [V96 ] ( 2, 5 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
+; V97 tmp94 [V97 ] ( 2, 9 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V98 tmp95 [V98 ] ( 2, 9 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V99 tmp96 [V99 ] ( 2, 5 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V100 tmp97 [V100 ] ( 2, 9 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V101 tmp98 [V101 ] ( 2, 9 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
+; V102 tmp99 [V102,T18] ( 2, 2 ) long -> rdi "field V35._00 (fldOffset=0x0)" P-INDEP
+; V103 tmp100 [V103,T19] ( 2, 2 ) long -> rsi "field V36._00 (fldOffset=0x0)" P-INDEP
+; V104 tmp101 [V104 ] ( 2, 5 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
+; V105 tmp102 [V105 ] ( 2, 9 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V39._00 (fldOffset=0x0)" P-DEP
+; V106 tmp103 [V106 ] ( 2, 9 ) long -> [rbp-0x108] do-not-enreg[X] addr-exposed "field V40._00 (fldOffset=0x0)" P-DEP
+; V107 tmp104 [V107 ] ( 2, 5 ) long -> [rbp-0x110] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
+; V108 tmp105 [V108 ] ( 2, 9 ) long -> [rbp-0x118] do-not-enreg[X] addr-exposed "field V48._00 (fldOffset=0x0)" P-DEP
+; V109 tmp106 [V109 ] ( 2, 9 ) long -> [rbp-0x120] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
+; V110 tmp107 [V110,T20] ( 2, 2 ) long -> r15 "field V63._00 (fldOffset=0x0)" P-INDEP
+; V111 tmp108 [V111,T21] ( 2, 2 ) long -> rdi "field V64._00 (fldOffset=0x0)" P-INDEP
+; V112 tmp109 [V112 ] ( 2, 5 ) long -> [rbp-0x1A8] do-not-enreg[X] addr-exposed "field V65._00 (fldOffset=0x0)" P-DEP
+; V113 tmp110 [V113 ] ( 2, 9 ) long -> [rbp-0x1B0] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
+; V114 tmp111 [V114 ] ( 2, 9 ) long -> [rbp-0x1B8] do-not-enreg[X] addr-exposed "field V68._00 (fldOffset=0x0)" P-DEP
+; V115 tmp112 [V115 ] ( 2, 5 ) long -> [rbp-0x1C0] do-not-enreg[X] addr-exposed "field V71._00 (fldOffset=0x0)" P-DEP
+; V116 tmp113 [V116 ] ( 2, 9 ) long -> [rbp-0x1C8] do-not-enreg[X] addr-exposed "field V73._00 (fldOffset=0x0)" P-DEP
+; V117 tmp114 [V117 ] ( 2, 9 ) long -> [rbp-0x1D0] do-not-enreg[X] addr-exposed "field V74._00 (fldOffset=0x0)" P-DEP
+; V118 tmp115 [V118,T22] ( 2, 2 ) long -> r15 "field V80._00 (fldOffset=0x0)" P-INDEP
+; V119 tmp116 [V119,T15] ( 2, 3 ) long -> [rbp-0x208] do-not-enreg[H] hidden-struct-arg "field V81._00 (fldOffset=0x0)" P-DEP
+; V120 tmp117 [V120 ] ( 2, 5 ) long -> [rbp-0x210] do-not-enreg[X] addr-exposed "field V82._00 (fldOffset=0x0)" P-DEP
+; V121 tmp118 [V121 ] ( 2, 9 ) long -> [rbp-0x218] do-not-enreg[X] addr-exposed "field V84._00 (fldOffset=0x0)" P-DEP
+; V122 tmp119 [V122 ] ( 2, 9 ) long -> [rbp-0x220] do-not-enreg[X] addr-exposed "field V85._00 (fldOffset=0x0)" P-DEP
+; V123 tmp120 [V123 ] ( 3, 3 ) simd32 -> [rbp-0x270] do-not-enreg[XS] addr-exposed "field V89._lower (fldOffset=0x0)" P-DEP
+; V124 tmp121 [V124 ] ( 3, 3 ) simd32 -> [rbp-0x250] do-not-enreg[XS] addr-exposed "field V89._upper (fldOffset=0x20)" P-DEP
+; V125 cse0 [V125,T07] ( 4, 16 ) long -> rsi "CSE #01: aggressive"
+; V126 cse1 [V126,T08] ( 4, 16 ) long -> rcx "CSE #02: aggressive"
+; V127 cse2 [V127,T09] ( 4, 16 ) long -> rsi "CSE #03: aggressive"
+; V128 cse3 [V128,T10] ( 4, 16 ) long -> rcx "CSE #04: aggressive"
+; V129 cse4 [V129,T11] ( 4, 16 ) long -> r14 "CSE #05: aggressive"
+; V130 cse5 [V130,T12] ( 4, 16 ) long -> r13 "CSE #06: aggressive"
+; V131 cse6 [V131,T13] ( 4, 16 ) long -> r14 "CSE #07: aggressive"
;
-; Lcl frame size = 384
+; Lcl frame size = 624
G_M23733_IG01:
push rbp
- sub rsp, 384
- lea rbp, [rsp+0x180]
- ;; size=16 bbWeight=1 PerfScore 1.75
+ push r15
+ push r14
+ push r13
+ push rbx
+ sub rsp, 624
+ lea rbp, [rsp+0x290]
+ mov rbx, rdi
+ ;; size=26 bbWeight=1 PerfScore 6.00
G_M23733_IG02:
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x50]
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x80], xmm2
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- xor ecx, ecx
- align [2 bytes for IG03]
- ;; size=48 bbWeight=1 PerfScore 15.00
+ vmovaps xmmword ptr [rbp-0x90], xmm2
+ mov rax, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0xA0], rax
+ mov rax, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xA8], rax
+ xor edi, edi
+ align [0 bytes for IG03]
+ ;; size=58 bbWeight=1 PerfScore 14.75
G_M23733_IG03:
- lea rax, [rbp-0x30]
- movsxd rsi, ecx
+ lea rax, [rbp-0xA0]
+ movsxd rsi, edi
movzx rax, byte ptr [rax+rsi]
- lea rdx, [rbp-0x38]
- movzx r8, byte ptr [rdx+rsi]
+ lea rdx, [rbp-0xA8]
+ movzx rcx, byte ptr [rdx+rsi]
xor edx, edx
- div edx:eax, r8d
- lea rdx, [rbp-0x28]
+ div edx:eax, ecx
+ lea rdx, [rbp-0x98]
mov byte ptr [rdx+rsi], al
- inc ecx
- cmp ecx, 8
+ inc edi
+ cmp edi, 8
jl SHORT G_M23733_IG03
- ;; size=39 bbWeight=4 PerfScore 134.00
+ ;; size=46 bbWeight=4 PerfScore 134.00
G_M23733_IG04:
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
+ mov rdi, qword ptr [rbp-0x98]
+ mov rax, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xB8], rax
+ mov rax, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0xC0], rax
xor esi, esi
- align [3 bytes for IG05]
- ;; size=25 bbWeight=1 PerfScore 5.50
+ align [0 bytes for IG05]
+ ;; size=34 bbWeight=1 PerfScore 5.25
G_M23733_IG05:
- lea rax, [rbp-0x48]
- movsxd r8, esi
- movzx rax, byte ptr [rax+r8]
- lea rdx, [rbp-0x50]
- movzx r9, byte ptr [rdx+r8]
+ lea rax, [rbp-0xB8]
+ movsxd rcx, esi
+ movzx rax, byte ptr [rax+rcx]
+ lea rdx, [rbp-0xC0]
+ movzx r8, byte ptr [rdx+rcx]
xor edx, edx
- div edx:eax, r9d
- lea rdx, [rbp-0x40]
- mov byte ptr [rdx+r8], al
+ div edx:eax, r8d
+ lea rdx, [rbp-0xB0]
+ mov byte ptr [rdx+rcx], al
inc esi
cmp esi, 8
jl SHORT G_M23733_IG05
- ;; size=41 bbWeight=4 PerfScore 134.00
+ ;; size=48 bbWeight=4 PerfScore 134.00
G_M23733_IG06:
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
+ mov rax, qword ptr [rbp-0xB0]
+ mov qword ptr [rbp-0xD0], rdi
+ mov qword ptr [rbp-0xC8], rax
+ vmovaps xmm2, xmmword ptr [rbp-0xD0]
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
+ vmovaps xmmword ptr [rbp-0xE0], xmm0
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- xor ecx, ecx
+ vmovaps xmmword ptr [rbp-0xF0], xmm0
+ mov rax, qword ptr [rbp-0xE0]
+ mov qword ptr [rbp-0x100], rax
+ mov rax, qword ptr [rbp-0xF0]
+ mov qword ptr [rbp-0x108], rax
+ xor edi, edi
align [0 bytes for IG07]
- ;; size=63 bbWeight=1 PerfScore 16.25
+ ;; size=87 bbWeight=1 PerfScore 16.25
G_M23733_IG07:
- lea rax, [rbp-0x90]
- movsxd rsi, ecx
+ lea rax, [rbp-0x100]
+ movsxd rsi, edi
movzx rax, byte ptr [rax+rsi]
- lea rdx, [rbp-0x98]
- movzx r8, byte ptr [rdx+rsi]
+ lea rdx, [rbp-0x108]
+ movzx rcx, byte ptr [rdx+rsi]
xor edx, edx
- div edx:eax, r8d
- lea rdx, [rbp-0x88]
+ div edx:eax, ecx
+ lea rdx, [rbp-0xF8]
mov byte ptr [rdx+rsi], al
- inc ecx
- cmp ecx, 8
+ inc edi
+ cmp edi, 8
jl SHORT G_M23733_IG07
- ;; size=48 bbWeight=4 PerfScore 134.00
+ ;; size=46 bbWeight=4 PerfScore 134.00
G_M23733_IG08:
- mov rcx, qword ptr [rbp-0x88]
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
+ mov rdi, qword ptr [rbp-0xF8]
+ mov rax, qword ptr [rbp-0xD8]
+ mov qword ptr [rbp-0x118], rax
+ mov rax, qword ptr [rbp-0xE8]
+ mov qword ptr [rbp-0x120], rax
xor esi, esi
- align [0 bytes for IG09]
- ;; size=31 bbWeight=1 PerfScore 5.25
+ align [2 bytes for IG09]
+ ;; size=39 bbWeight=1 PerfScore 5.50
G_M23733_IG09:
- lea rax, [rbp-0xA8]
- movsxd r8, esi
- movzx rax, byte ptr [rax+r8]
- lea rdx, [rbp-0xB0]
- movzx r9, byte ptr [rdx+r8]
+ lea rax, [rbp-0x118]
+ movsxd rcx, esi
+ movzx rax, byte ptr [rax+rcx]
+ lea rdx, [rbp-0x120]
+ movzx r8, byte ptr [rdx+rcx]
xor edx, edx
- div edx:eax, r9d
- lea rdx, [rbp-0xA0]
- mov byte ptr [rdx+r8], al
+ div edx:eax, r8d
+ lea rdx, [rbp-0x110]
+ mov byte ptr [rdx+rcx], al
inc esi
cmp esi, 8
jl SHORT G_M23733_IG09
- ;; size=50 bbWeight=4 PerfScore 134.00
+ ;; size=48 bbWeight=4 PerfScore 134.00
G_M23733_IG10:
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
+ mov rsi, qword ptr [rbp-0x110]
+ mov qword ptr [rbp-0x130], rdi
+ mov qword ptr [rbp-0x128], rsi
+ vinserti128 ymm0, ymm2, xmmword ptr [rbp-0x130], 1
+ vmovups ymmword ptr [rbp-0x50], ymm0
vmovups ymm1, ymmword ptr [rbp+0x30]
+ vmovups ymmword ptr [rbp-0x150], ymm1
vmovups ymm2, ymmword ptr [rbp+0x70]
+ vmovups ymmword ptr [rbp-0x170], ymm2
vmovaps ymm3, ymm1
- vmovaps xmmword ptr [rbp-0xD0], xmm3
+ vmovaps xmmword ptr [rbp-0x190], xmm3
vmovaps ymm3, ymm2
- vmovaps xmmword ptr [rbp-0xE0], xmm3
- mov rax, qword ptr [rbp-0xD0]
- mov qword ptr [rbp-0xF0], rax
- mov rax, qword ptr [rbp-0xE0]
- mov qword ptr [rbp-0xF8], rax
- xor ecx, ecx
- align [0 bytes for IG11]
- ;; size=95 bbWeight=1 PerfScore 21.75
+ vmovaps xmmword ptr [rbp-0x1A0], xmm3
+ mov rdi, qword ptr [rbp-0x190]
+ mov qword ptr [rbp-0x1B0], rdi
+ mov rdi, qword ptr [rbp-0x1A0]
+ mov qword ptr [rbp-0x1B8], rdi
+ xor r15d, r15d
+ ;; size=117 bbWeight=1 PerfScore 24.75
G_M23733_IG11:
- lea rax, [rbp-0xF0]
- movsxd rsi, ecx
- movzx rax, byte ptr [rax+rsi]
- lea rdx, [rbp-0xF8]
- movzx r8, byte ptr [rdx+rsi]
- xor edx, edx
- div edx:eax, r8d
- lea rdx, [rbp-0xE8]
- mov byte ptr [rdx+rsi], al
- inc ecx
- cmp ecx, 8
+ lea rdi, [rbp-0x1B0]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0x1B8]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0x1A8]
+ mov byte ptr [rdi+r14], al
+ inc r15d
+ cmp r15d, 8
jl SHORT G_M23733_IG11
- ;; size=48 bbWeight=4 PerfScore 134.00
+ ;; size=59 bbWeight=4 PerfScore 46.00
G_M23733_IG12:
- mov rcx, qword ptr [rbp-0xE8]
- mov rax, qword ptr [rbp-0xC8]
- mov qword ptr [rbp-0x108], rax
- mov rax, qword ptr [rbp-0xD8]
- mov qword ptr [rbp-0x110], rax
- xor esi, esi
- align [3 bytes for IG13]
- ;; size=40 bbWeight=1 PerfScore 5.50
+ mov r15, qword ptr [rbp-0x1A8]
+ mov rdi, qword ptr [rbp-0x188]
+ mov qword ptr [rbp-0x1C8], rdi
+ mov rdi, qword ptr [rbp-0x198]
+ mov qword ptr [rbp-0x1D0], rdi
+ xor r14d, r14d
+ ;; size=38 bbWeight=1 PerfScore 5.25
G_M23733_IG13:
- lea rax, [rbp-0x108]
- movsxd r8, esi
- movzx rax, byte ptr [rax+r8]
- lea rdx, [rbp-0x110]
- movzx r9, byte ptr [rdx+r8]
- xor edx, edx
- div edx:eax, r9d
- lea rdx, [rbp-0x100]
- mov byte ptr [rdx+r8], al
- inc esi
- cmp esi, 8
+ lea rdi, [rbp-0x1C8]
+ movsxd r13, r14d
+ movzx rdi, byte ptr [rdi+r13]
+ lea rsi, [rbp-0x1D0]
+ movzx rsi, byte ptr [rsi+r13]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0x1C0]
+ mov byte ptr [rdi+r13], al
+ inc r14d
+ cmp r14d, 8
jl SHORT G_M23733_IG13
- ;; size=50 bbWeight=4 PerfScore 134.00
+ ;; size=59 bbWeight=4 PerfScore 46.00
G_M23733_IG14:
- mov rax, qword ptr [rbp-0x100]
- mov qword ptr [rbp-0x120], rcx
- mov qword ptr [rbp-0x118], rax
- vmovaps xmm3, xmmword ptr [rbp-0x120]
+ mov rdi, qword ptr [rbp-0x1C0]
+ mov qword ptr [rbp-0x1E0], r15
+ mov qword ptr [rbp-0x1D8], rdi
+ vmovaps xmm3, xmmword ptr [rbp-0x1E0]
+ vmovaps xmmword ptr [rbp-0x180], xmm3
+ vmovups ymm1, ymmword ptr [rbp-0x150]
vextractf128 xmm1, ymm1, 1
- vmovaps xmmword ptr [rbp-0x130], xmm1
+ vmovaps xmmword ptr [rbp-0x1F0], xmm1
+ vmovups ymm2, ymmword ptr [rbp-0x170]
vextractf128 xmm1, ymm2, 1
- vmovaps xmmword ptr [rbp-0x140], xmm1
- mov rax, qword ptr [rbp-0x130]
- mov qword ptr [rbp-0x150], rax
- mov rax, qword ptr [rbp-0x140]
- mov qword ptr [rbp-0x158], rax
- xor ecx, ecx
- align [0 bytes for IG15]
- ;; size=87 bbWeight=1 PerfScore 16.25
+ vmovaps xmmword ptr [rbp-0x200], xmm1
+ mov rdi, qword ptr [rbp-0x1F0]
+ mov qword ptr [rbp-0x218], rdi
+ mov rdi, qword ptr [rbp-0x200]
+ mov qword ptr [rbp-0x220], rdi
+ xor r15d, r15d
+ ;; size=112 bbWeight=1 PerfScore 25.25
G_M23733_IG15:
- lea rax, [rbp-0x150]
- movsxd rsi, ecx
- movzx rax, byte ptr [rax+rsi]
- lea rdx, [rbp-0x158]
- movzx r8, byte ptr [rdx+rsi]
- xor edx, edx
- div edx:eax, r8d
- lea rdx, [rbp-0x148]
- mov byte ptr [rdx+rsi], al
- inc ecx
- cmp ecx, 8
+ lea rdi, [rbp-0x218]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0x220]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0x210]
+ mov byte ptr [rdi+r14], al
+ inc r15d
+ cmp r15d, 8
jl SHORT G_M23733_IG15
- ;; size=48 bbWeight=4 PerfScore 134.00
+ ;; size=59 bbWeight=4 PerfScore 46.00
G_M23733_IG16:
- mov rcx, qword ptr [rbp-0x148]
- mov rax, qword ptr [rbp-0x128]
- mov qword ptr [rbp-0x168], rax
- mov rax, qword ptr [rbp-0x138]
- mov qword ptr [rbp-0x170], rax
- xor esi, esi
- align [2 bytes for IG17]
- ;; size=39 bbWeight=1 PerfScore 5.50
+ mov r15, qword ptr [rbp-0x210]
+ mov rdi, qword ptr [rbp-0x1E8]
+ mov qword ptr [rsp], rdi
+ mov rdi, qword ptr [rbp-0x1F8]
+ mov qword ptr [rsp+0x08], rdi
+ lea rdi, [rbp-0x208]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64`1[ubyte]:op_Division(System.Runtime.Intrinsics.Vector64`1[ubyte],System.Runtime.Intrinsics.Vector64`1[ubyte]):System.Runtime.Intrinsics.Vector64`1[ubyte]
+ call [rax]System.Runtime.Intrinsics.Vector64`1[ubyte]:op_Division(System.Runtime.Intrinsics.Vector64`1[ubyte],System.Runtime.Intrinsics.Vector64`1[ubyte]):System.Runtime.Intrinsics.Vector64`1[ubyte]
+ mov qword ptr [rbp-0x230], r15
+ mov rdi, qword ptr [rbp-0x208]
+ mov qword ptr [rsp], rdi
+ lea rdi, [rbp-0x230]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[ubyte](byref,System.Runtime.Intrinsics.Vector64`1[ubyte])
+ call [rax]System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[ubyte](byref,System.Runtime.Intrinsics.Vector64`1[ubyte])
+ vmovaps xmm3, xmmword ptr [rbp-0x180]
+ vmovups xmmword ptr [rsp], xmm3
+ vmovaps xmm0, xmmword ptr [rbp-0x230]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ lea rdi, [rbp-0x70]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[ubyte](System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[ubyte](System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte]
+ vmovups ymm0, ymmword ptr [rbp-0x50]
+ vmovups ymmword ptr [rsp], ymm0
+ lea rdi, [rbp-0x270]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector512:SetLowerUnsafe[ubyte](byref,System.Runtime.Intrinsics.Vector256`1[ubyte])
+ call [rax]System.Runtime.Intrinsics.Vector512:SetLowerUnsafe[ubyte](byref,System.Runtime.Intrinsics.Vector256`1[ubyte])
+ vmovups ymm0, ymmword ptr [rbp-0x70]
+ vmovups ymmword ptr [rsp], ymm0
+ lea rdi, [rbp-0x270]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[ubyte](byref,System.Runtime.Intrinsics.Vector256`1[ubyte])
+ call [rax]System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[ubyte](byref,System.Runtime.Intrinsics.Vector256`1[ubyte])
+ vmovdqu ymm0, ymmword ptr [rbp-0x270]
+ vmovdqu ymmword ptr [rbx], ymm0
+ vmovdqu ymm0, ymmword ptr [rbp-0x250]
+ vmovdqu ymmword ptr [rbx+0x20], ymm0
+ mov rax, rbx
+ ;; size=215 bbWeight=1 PerfScore 57.00
G_M23733_IG17:
- lea rax, [rbp-0x168]
- movsxd r8, esi
- movzx rax, byte ptr [rax+r8]
- lea rdx, [rbp-0x170]
- movzx r9, byte ptr [rdx+r8]
- xor edx, edx
- div edx:eax, r9d
- lea rdx, [rbp-0x160]
- mov byte ptr [rdx+r8], al
- inc esi
- cmp esi, 8
- jl SHORT G_M23733_IG17
- ;; size=50 bbWeight=4 PerfScore 134.00
-G_M23733_IG18:
- mov rax, qword ptr [rbp-0x160]
- mov qword ptr [rbp-0x180], rcx
- mov qword ptr [rbp-0x178], rax
- vinserti128 ymm1, ymm3, xmmword ptr [rbp-0x180], 1
- vmovups ymmword ptr [rdi], ymm0
- vmovups ymmword ptr [rdi+0x20], ymm1
- mov rax, rdi
- ;; size=43 bbWeight=1 PerfScore 11.25
-G_M23733_IG19:
vzeroupper
- add rsp, 384
+ add rsp, 624
+ pop rbx
+ pop r13
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=19 bbWeight=1 PerfScore 4.75
-; Total bytes of code 873, prolog size 16, PerfScore 1178.75, instruction count 194, allocated bytes for code 873 (MethodHash=f113a34a) for method System.Runtime.Intrinsics.Vector512:Divide[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):System.Runtime.Intrinsics.Vector512`1[ubyte] (FullOpts)
+; Total bytes of code 1110, prolog size 26, PerfScore 838.75, instruction count 216, allocated bytes for code 1110 (MethodHash=f113a34a) for method System.Runtime.Intrinsics.Vector512:Divide[ubyte](System.Runtime.Intrinsics.Vector512`1[ubyte],System.Runtime.Intrinsics.Vector512`1[ubyte]):System.Runtime.Intrinsics.Vector512`1[ubyte] (FullOpts)233 (36.46 % of base) - System.Runtime.Intrinsics.Vector512:Divide[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):System.Runtime.Intrinsics.Vector512`1[long]
; Assembly listing for method System.Runtime.Intrinsics.Vector512:Divide[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):System.Runtime.Intrinsics.Vector512`1[long] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 43 single block inlinees; 27 inlinees without PGO data
+; 0 inlinees with PGO data; 38 single block inlinees; 21 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T00] ( 5, 5 ) byref -> rdi single-def
+; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rbx single-def
;* V01 arg0 [V01 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[long]>
;* V02 arg1 [V02 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[long]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T29] ( 2, 4 ) simd32 -> mm0 "impAppendStmt"
-; V05 tmp2 [V05,T30] ( 2, 4 ) simd32 -> mm1 "spilled call-like call argument"
-; V06 tmp3 [V06,T17] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V07 tmp4 [V07,T18] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V08 tmp5 [V08,T31] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T28] ( 2, 4 ) simd32 -> [rbp-0x30] spill-single-def "impAppendStmt"
+; V05 tmp2 [V05,T29] ( 2, 4 ) simd32 -> [rbp-0x50] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument"
+; V06 tmp3 [V06,T16] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V07 tmp4 [V07,T17] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V08 tmp5 [V08,T30] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
;* V09 tmp6 [V09 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V10 tmp7 [V10,T19] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V11 tmp8 [V11,T20] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V10 tmp7 [V10,T18] ( 3, 6 ) simd16 -> [rbp-0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V11 tmp8 [V11,T19] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
;* V12 tmp9 [V12 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V13 tmp10 [V13 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V14 tmp11 [V14 ] ( 2, 2 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V14 tmp11 [V14 ] ( 2, 2 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V15 tmp12 [V15,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V16 tmp13 [V16 ] ( 2, 4 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V17 tmp14 [V17 ] ( 2, 4 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V16 tmp13 [V16 ] ( 2, 4 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V17 tmp14 [V17 ] ( 2, 4 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V18 tmp15 [V18 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
;* V19 tmp16 [V19 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
;* V20 tmp17 [V20 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
;* V21 tmp18 [V21 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V22 tmp19 [V22 ] ( 2, 2 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V22 tmp19 [V22 ] ( 2, 2 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V23 tmp20 [V23,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V24 tmp21 [V24 ] ( 2, 4 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V25 tmp22 [V25 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V24 tmp21 [V24 ] ( 2, 4 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V25 tmp22 [V25 ] ( 2, 4 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V26 tmp23 [V26 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
;* V27 tmp24 [V27 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
;* V28 tmp25 [V28 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
;* V29 tmp26 [V29 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V30 tmp27 [V30,T33] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V31 tmp28 [V31,T21] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V32 tmp29 [V32,T22] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V30 tmp27 [V30,T32] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V31 tmp28 [V31,T20] ( 3, 6 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V32 tmp29 [V32,T21] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
;* V33 tmp30 [V33 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V34 tmp31 [V34 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V35 tmp32 [V35 ] ( 2, 2 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V35 tmp32 [V35 ] ( 2, 2 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V36 tmp33 [V36,T11] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V37 tmp34 [V37 ] ( 2, 4 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V38 tmp35 [V38 ] ( 2, 4 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V37 tmp34 [V37 ] ( 2, 4 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V38 tmp35 [V38 ] ( 2, 4 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V39 tmp36 [V39 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
;* V40 tmp37 [V40 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
;* V41 tmp38 [V41 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
;* V42 tmp39 [V42 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V43 tmp40 [V43 ] ( 2, 2 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V43 tmp40 [V43 ] ( 2, 2 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V44 tmp41 [V44,T12] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V45 tmp42 [V45 ] ( 2, 4 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V46 tmp43 [V46 ] ( 2, 4 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V45 tmp42 [V45 ] ( 2, 4 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V46 tmp43 [V46 ] ( 2, 4 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V47 tmp44 [V47 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
;* V48 tmp45 [V48 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
;* V49 tmp46 [V49 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
;* V50 tmp47 [V50 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V51 tmp48 [V51,T34] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V51 tmp48 [V51,T33] ( 3, 3 ) simd16 -> [rbp-0x110] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
;* V52 tmp49 [V52 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V53 tmp50 [V53,T23] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V54 tmp51 [V54,T24] ( 3, 6 ) simd32 -> mm2 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V55 tmp52 [V55,T32] ( 2, 4 ) simd16 -> mm3 "impAppendStmt"
+; V53 tmp50 [V53,T22] ( 3, 6 ) simd32 -> [rbp-0x130] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V54 tmp51 [V54,T23] ( 3, 6 ) simd32 -> [rbp-0x150] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V55 tmp52 [V55,T31] ( 2, 4 ) simd16 -> [rbp-0x160] spill-single-def "impAppendStmt"
;* V56 tmp53 [V56 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V57 tmp54 [V57,T25] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V58 tmp55 [V58,T26] ( 3, 6 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V57 tmp54 [V57,T24] ( 3, 6 ) simd16 -> [rbp-0x170] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V58 tmp55 [V58,T25] ( 3, 6 ) simd16 -> [rbp-0x180] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
;* V59 tmp56 [V59 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V60 tmp57 [V60 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V61 tmp58 [V61 ] ( 2, 2 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V61 tmp58 [V61 ] ( 2, 2 ) struct ( 8) [rbp-0x188] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V62 tmp59 [V62,T13] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V63 tmp60 [V63 ] ( 2, 4 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V64 tmp61 [V64 ] ( 2, 4 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V63 tmp60 [V63 ] ( 2, 4 ) struct ( 8) [rbp-0x190] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V64 tmp61 [V64 ] ( 2, 4 ) struct ( 8) [rbp-0x198] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V65 tmp62 [V65 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
;* V66 tmp63 [V66 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V67 tmp64 [V67 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V68 tmp65 [V68 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V69 tmp66 [V69 ] ( 2, 2 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V70 tmp67 [V70,T14] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V71 tmp68 [V71 ] ( 2, 4 ) struct ( 8) [rbp-0x108] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V72 tmp69 [V72 ] ( 2, 4 ) struct ( 8) [rbp-0x110] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V73 tmp70 [V73 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V74 tmp71 [V74 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V75 tmp72 [V75 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V76 tmp73 [V76 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V77 tmp74 [V77,T35] ( 3, 3 ) simd16 -> [rbp-0x120] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V78 tmp75 [V78,T27] ( 3, 6 ) simd16 -> [rbp-0x130] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V79 tmp76 [V79,T28] ( 3, 6 ) simd16 -> [rbp-0x140] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-;* V80 tmp77 [V80 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V81 tmp78 [V81 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V82 tmp79 [V82 ] ( 2, 2 ) struct ( 8) [rbp-0x148] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V83 tmp80 [V83,T15] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V84 tmp81 [V84 ] ( 2, 4 ) struct ( 8) [rbp-0x150] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V85 tmp82 [V85 ] ( 2, 4 ) struct ( 8) [rbp-0x158] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V86 tmp83 [V86 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V87 tmp84 [V87 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V88 tmp85 [V88 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V89 tmp86 [V89 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V90 tmp87 [V90 ] ( 2, 2 ) struct ( 8) [rbp-0x160] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V91 tmp88 [V91,T16] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V92 tmp89 [V92 ] ( 2, 4 ) struct ( 8) [rbp-0x168] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V93 tmp90 [V93 ] ( 2, 4 ) struct ( 8) [rbp-0x170] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V94 tmp91 [V94 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V95 tmp92 [V95 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V96 tmp93 [V96 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V97 tmp94 [V97 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V98 tmp95 [V98,T36] ( 3, 3 ) simd16 -> [rbp-0x180] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-;* V99 tmp96 [V99 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[long]>
-;* V100 tmp97 [V100 ] ( 0, 0 ) struct (64) zero-ref ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[long]>
-; V101 tmp98 [V101,T39] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
-; V102 tmp99 [V102,T40] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
-; V103 tmp100 [V103,T41] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
-; V104 tmp101 [V104,T42] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
-; V105 tmp102 [V105,T01] ( 2, 2 ) long -> rcx "field V12._00 (fldOffset=0x0)" P-INDEP
-; V106 tmp103 [V106,T02] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
-; V107 tmp104 [V107 ] ( 2, 2 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
-; V108 tmp105 [V108 ] ( 2, 3 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
-; V109 tmp106 [V109 ] ( 2, 3 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V110 tmp107 [V110 ] ( 2, 2 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V22._00 (fldOffset=0x0)" P-DEP
-; V111 tmp108 [V111 ] ( 2, 3 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V24._00 (fldOffset=0x0)" P-DEP
-; V112 tmp109 [V112 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V113 tmp110 [V113,T03] ( 2, 2 ) long -> rcx "field V33._00 (fldOffset=0x0)" P-INDEP
-; V114 tmp111 [V114,T04] ( 2, 2 ) long -> rax "field V34._00 (fldOffset=0x0)" P-INDEP
-; V115 tmp112 [V115 ] ( 2, 2 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
-; V116 tmp113 [V116 ] ( 2, 3 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
-; V117 tmp114 [V117 ] ( 2, 3 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V38._00 (fldOffset=0x0)" P-DEP
-; V118 tmp115 [V118 ] ( 2, 2 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V43._00 (fldOffset=0x0)" P-DEP
-; V119 tmp116 [V119 ] ( 2, 3 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
-; V120 tmp117 [V120 ] ( 2, 3 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
-; V121 tmp118 [V121,T05] ( 2, 2 ) long -> rcx "field V59._00 (fldOffset=0x0)" P-INDEP
-; V122 tmp119 [V122,T06] ( 2, 2 ) long -> rax "field V60._00 (fldOffset=0x0)" P-INDEP
-; V123 tmp120 [V123 ] ( 2, 2 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V61._00 (fldOffset=0x0)" P-DEP
-; V124 tmp121 [V124 ] ( 2, 3 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V63._00 (fldOffset=0x0)" P-DEP
-; V125 tmp122 [V125 ] ( 2, 3 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V64._00 (fldOffset=0x0)" P-DEP
-; V126 tmp123 [V126 ] ( 2, 2 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V69._00 (fldOffset=0x0)" P-DEP
-; V127 tmp124 [V127 ] ( 2, 3 ) long -> [rbp-0x108] do-not-enreg[X] addr-exposed "field V71._00 (fldOffset=0x0)" P-DEP
-; V128 tmp125 [V128 ] ( 2, 3 ) long -> [rbp-0x110] do-not-enreg[X] addr-exposed "field V72._00 (fldOffset=0x0)" P-DEP
-; V129 tmp126 [V129,T07] ( 2, 2 ) long -> rcx "field V80._00 (fldOffset=0x0)" P-INDEP
-; V130 tmp127 [V130,T08] ( 2, 2 ) long -> rax "field V81._00 (fldOffset=0x0)" P-INDEP
-; V131 tmp128 [V131 ] ( 2, 2 ) long -> [rbp-0x148] do-not-enreg[X] addr-exposed "field V82._00 (fldOffset=0x0)" P-DEP
-; V132 tmp129 [V132 ] ( 2, 3 ) long -> [rbp-0x150] do-not-enreg[X] addr-exposed "field V84._00 (fldOffset=0x0)" P-DEP
-; V133 tmp130 [V133 ] ( 2, 3 ) long -> [rbp-0x158] do-not-enreg[X] addr-exposed "field V85._00 (fldOffset=0x0)" P-DEP
-; V134 tmp131 [V134 ] ( 2, 2 ) long -> [rbp-0x160] do-not-enreg[X] addr-exposed "field V90._00 (fldOffset=0x0)" P-DEP
-; V135 tmp132 [V135 ] ( 2, 3 ) long -> [rbp-0x168] do-not-enreg[X] addr-exposed "field V92._00 (fldOffset=0x0)" P-DEP
-; V136 tmp133 [V136 ] ( 2, 3 ) long -> [rbp-0x170] do-not-enreg[X] addr-exposed "field V93._00 (fldOffset=0x0)" P-DEP
-; V137 tmp134 [V137,T37] ( 2, 2 ) simd32 -> mm0 "field V100._lower (fldOffset=0x0)" P-INDEP
-; V138 tmp135 [V138,T38] ( 2, 2 ) simd32 -> mm1 "field V100._upper (fldOffset=0x20)" P-INDEP
+; V67 tmp64 [V67 ] ( 2, 2 ) struct ( 8) [rbp-0x1A0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V68 tmp65 [V68,T14] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V69 tmp66 [V69 ] ( 2, 4 ) struct ( 8) [rbp-0x1A8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V70 tmp67 [V70 ] ( 2, 4 ) struct ( 8) [rbp-0x1B0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V71 tmp68 [V71 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V72 tmp69 [V72 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V73 tmp70 [V73,T34] ( 3, 3 ) simd16 -> [rbp-0x1C0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V74 tmp71 [V74,T26] ( 3, 6 ) simd16 -> [rbp-0x1D0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V75 tmp72 [V75,T27] ( 3, 6 ) simd16 -> [rbp-0x1E0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+;* V76 tmp73 [V76 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V77 tmp74 [V77 ] ( 2, 4 ) struct ( 8) [rbp-0x1E8] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V78 tmp75 [V78 ] ( 2, 2 ) struct ( 8) [rbp-0x1F0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V79 tmp76 [V79,T15] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V80 tmp77 [V80 ] ( 2, 4 ) struct ( 8) [rbp-0x1F8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V81 tmp78 [V81 ] ( 2, 4 ) struct ( 8) [rbp-0x200] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V82 tmp79 [V82 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V83 tmp80 [V83 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V84 tmp81 [V84 ] ( 3, 3 ) simd16 -> [rbp-0x210] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V85 tmp82 [V85 ] ( 3, 3 ) struct (64) [rbp-0x250] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[long]>
+; V86 tmp83 [V86,T35] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
+; V87 tmp84 [V87,T36] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
+; V88 tmp85 [V88,T37] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
+; V89 tmp86 [V89,T38] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
+; V90 tmp87 [V90,T02] ( 2, 2 ) long -> rdi "field V12._00 (fldOffset=0x0)" P-INDEP
+; V91 tmp88 [V91,T03] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
+; V92 tmp89 [V92 ] ( 2, 2 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
+; V93 tmp90 [V93 ] ( 2, 3 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V94 tmp91 [V94 ] ( 2, 3 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V95 tmp92 [V95 ] ( 2, 2 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V22._00 (fldOffset=0x0)" P-DEP
+; V96 tmp93 [V96 ] ( 2, 3 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V24._00 (fldOffset=0x0)" P-DEP
+; V97 tmp94 [V97 ] ( 2, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V98 tmp95 [V98,T04] ( 2, 2 ) long -> rdi "field V33._00 (fldOffset=0x0)" P-INDEP
+; V99 tmp96 [V99,T05] ( 2, 2 ) long -> rsi "field V34._00 (fldOffset=0x0)" P-INDEP
+; V100 tmp97 [V100 ] ( 2, 2 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V101 tmp98 [V101 ] ( 2, 3 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
+; V102 tmp99 [V102 ] ( 2, 3 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V38._00 (fldOffset=0x0)" P-DEP
+; V103 tmp100 [V103 ] ( 2, 2 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V43._00 (fldOffset=0x0)" P-DEP
+; V104 tmp101 [V104 ] ( 2, 3 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
+; V105 tmp102 [V105 ] ( 2, 3 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
+; V106 tmp103 [V106,T06] ( 2, 2 ) long -> r15 "field V59._00 (fldOffset=0x0)" P-INDEP
+; V107 tmp104 [V107,T07] ( 2, 2 ) long -> rdi "field V60._00 (fldOffset=0x0)" P-INDEP
+; V108 tmp105 [V108 ] ( 2, 2 ) long -> [rbp-0x188] do-not-enreg[X] addr-exposed "field V61._00 (fldOffset=0x0)" P-DEP
+; V109 tmp106 [V109 ] ( 2, 3 ) long -> [rbp-0x190] do-not-enreg[X] addr-exposed "field V63._00 (fldOffset=0x0)" P-DEP
+; V110 tmp107 [V110 ] ( 2, 3 ) long -> [rbp-0x198] do-not-enreg[X] addr-exposed "field V64._00 (fldOffset=0x0)" P-DEP
+; V111 tmp108 [V111 ] ( 2, 2 ) long -> [rbp-0x1A0] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
+; V112 tmp109 [V112 ] ( 2, 3 ) long -> [rbp-0x1A8] do-not-enreg[X] addr-exposed "field V69._00 (fldOffset=0x0)" P-DEP
+; V113 tmp110 [V113 ] ( 2, 3 ) long -> [rbp-0x1B0] do-not-enreg[X] addr-exposed "field V70._00 (fldOffset=0x0)" P-DEP
+; V114 tmp111 [V114,T08] ( 2, 2 ) long -> r15 "field V76._00 (fldOffset=0x0)" P-INDEP
+; V115 tmp112 [V115,T01] ( 2, 3 ) long -> [rbp-0x1E8] do-not-enreg[H] hidden-struct-arg "field V77._00 (fldOffset=0x0)" P-DEP
+; V116 tmp113 [V116 ] ( 2, 2 ) long -> [rbp-0x1F0] do-not-enreg[X] addr-exposed "field V78._00 (fldOffset=0x0)" P-DEP
+; V117 tmp114 [V117 ] ( 2, 3 ) long -> [rbp-0x1F8] do-not-enreg[X] addr-exposed "field V80._00 (fldOffset=0x0)" P-DEP
+; V118 tmp115 [V118 ] ( 2, 3 ) long -> [rbp-0x200] do-not-enreg[X] addr-exposed "field V81._00 (fldOffset=0x0)" P-DEP
+; V119 tmp116 [V119 ] ( 3, 3 ) simd32 -> [rbp-0x250] do-not-enreg[XS] addr-exposed "field V85._lower (fldOffset=0x0)" P-DEP
+; V120 tmp117 [V120 ] ( 2, 2 ) simd32 -> [rbp-0x230] do-not-enreg[XS] addr-exposed "field V85._upper (fldOffset=0x20)" P-DEP
;
-; Lcl frame size = 384
+; Lcl frame size = 608
G_M43701_IG01:
push rbp
- sub rsp, 384
- lea rbp, [rsp+0x180]
- ;; size=16 bbWeight=1 PerfScore 1.75
+ push r15
+ push rbx
+ sub rsp, 608
+ lea rbp, [rsp+0x270]
+ mov rbx, rdi
+ ;; size=22 bbWeight=1 PerfScore 4.00
G_M43701_IG02:
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x50]
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x60], xmm2
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- mov rax, qword ptr [rbp-0x30]
- cqo
- idiv rdx:rax, qword ptr [rbp-0x38]
- mov qword ptr [rbp-0x28], rax
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
- mov rax, qword ptr [rbp-0x48]
- cqo
- idiv rdx:rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x40], rax
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
- vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
+ vmovaps xmmword ptr [rbp-0x70], xmm2
+ mov rax, qword ptr [rbp-0x60]
+ mov qword ptr [rbp-0x80], rax
mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
+ mov qword ptr [rbp-0x88], rax
mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- mov rax, qword ptr [rbp-0x90]
cqo
- idiv rdx:rax, qword ptr [rbp-0x98]
- mov qword ptr [rbp-0x88], rax
- mov rcx, qword ptr [rbp-0x88]
+ idiv rdx:rax, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0x78], rax
+ mov rdi, qword ptr [rbp-0x78]
+ mov rax, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x98], rax
mov rax, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0xA0], rax
+ mov rax, qword ptr [rbp-0x98]
+ cqo
+ idiv rdx:rax, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0x90], rax
+ mov rax, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xB0], rdi
mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
- mov rax, qword ptr [rbp-0xA8]
+ vmovaps xmm2, xmmword ptr [rbp-0xB0]
+ vextractf128 xmm0, ymm0, 1
+ vmovaps xmmword ptr [rbp-0xC0], xmm0
+ vextractf128 xmm0, ymm1, 1
+ vmovaps xmmword ptr [rbp-0xD0], xmm0
+ mov rax, qword ptr [rbp-0xC0]
+ mov qword ptr [rbp-0xE0], rax
+ mov rax, qword ptr [rbp-0xD0]
+ mov qword ptr [rbp-0xE8], rax
+ mov rax, qword ptr [rbp-0xE0]
cqo
- idiv rdx:rax, qword ptr [rbp-0xB0]
- mov qword ptr [rbp-0xA0], rax
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
+ idiv rdx:rax, qword ptr [rbp-0xE8]
+ mov qword ptr [rbp-0xD8], rax
+ mov rdi, qword ptr [rbp-0xD8]
+ mov rax, qword ptr [rbp-0xB8]
+ mov qword ptr [rbp-0xF8], rax
+ mov rax, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0x100], rax
+ mov rax, qword ptr [rbp-0xF8]
+ cqo
+ idiv rdx:rax, qword ptr [rbp-0x100]
+ mov qword ptr [rbp-0xF0], rax
+ mov rsi, qword ptr [rbp-0xF0]
+ mov qword ptr [rbp-0x110], rdi
+ mov qword ptr [rbp-0x108], rsi
+ vinserti128 ymm0, ymm2, xmmword ptr [rbp-0x110], 1
+ vmovups ymmword ptr [rbp-0x30], ymm0
vmovups ymm1, ymmword ptr [rbp+0x30]
+ vmovups ymmword ptr [rbp-0x130], ymm1
vmovups ymm2, ymmword ptr [rbp+0x70]
+ vmovups ymmword ptr [rbp-0x150], ymm2
vmovaps ymm3, ymm1
- vmovaps xmmword ptr [rbp-0xD0], xmm3
+ vmovaps xmmword ptr [rbp-0x170], xmm3
vmovaps ymm3, ymm2
- vmovaps xmmword ptr [rbp-0xE0], xmm3
- mov rax, qword ptr [rbp-0xD0]
- mov qword ptr [rbp-0xF0], rax
- mov rax, qword ptr [rbp-0xE0]
- ;; size=314 bbWeight=1 PerfScore 343.00
+ vmovaps xmmword ptr [rbp-0x180], xmm3
+ ;; size=365 bbWeight=1 PerfScore 343.00
G_M43701_IG03:
- mov qword ptr [rbp-0xF8], rax
- mov rax, qword ptr [rbp-0xF0]
- cqo
- idiv rdx:rax, qword ptr [rbp-0xF8]
- mov qword ptr [rbp-0xE8], rax
- mov rcx, qword ptr [rbp-0xE8]
- mov rax, qword ptr [rbp-0xC8]
- mov qword ptr [rbp-0x108], rax
- mov rax, qword ptr [rbp-0xD8]
- mov qword ptr [rbp-0x110], rax
- mov rax, qword ptr [rbp-0x108]
- cqo
- idiv rdx:rax, qword ptr [rbp-0x110]
- mov qword ptr [rbp-0x100], rax
- mov rax, qword ptr [rbp-0x100]
- mov qword ptr [rbp-0x120], rcx
- mov qword ptr [rbp-0x118], rax
- vmovaps xmm3, xmmword ptr [rbp-0x120]
+ mov rdi, qword ptr [rbp-0x170]
+ mov qword ptr [rbp-0x190], rdi
+ mov rdi, qword ptr [rbp-0x180]
+ mov qword ptr [rbp-0x198], rdi
+ mov rdi, qword ptr [rbp-0x190]
+ mov rsi, qword ptr [rbp-0x198]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Divide(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Divide(long,long):long
+ mov qword ptr [rbp-0x188], rax
+ mov r15, qword ptr [rbp-0x188]
+ mov rdi, qword ptr [rbp-0x168]
+ mov qword ptr [rbp-0x1A8], rdi
+ mov rdi, qword ptr [rbp-0x178]
+ mov qword ptr [rbp-0x1B0], rdi
+ mov rdi, qword ptr [rbp-0x1A8]
+ mov rsi, qword ptr [rbp-0x1B0]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Divide(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Divide(long,long):long
+ mov qword ptr [rbp-0x1A0], rax
+ mov rdi, qword ptr [rbp-0x1A0]
+ mov qword ptr [rbp-0x1C0], r15
+ mov qword ptr [rbp-0x1B8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x1C0]
+ vmovaps xmmword ptr [rbp-0x160], xmm0
+ vmovups ymm1, ymmword ptr [rbp-0x130]
vextractf128 xmm1, ymm1, 1
- vmovaps xmmword ptr [rbp-0x130], xmm1
+ vmovaps xmmword ptr [rbp-0x1D0], xmm1
+ vmovups ymm2, ymmword ptr [rbp-0x150]
vextractf128 xmm1, ymm2, 1
- vmovaps xmmword ptr [rbp-0x140], xmm1
- mov rax, qword ptr [rbp-0x130]
- mov qword ptr [rbp-0x150], rax
- mov rax, qword ptr [rbp-0x140]
- mov qword ptr [rbp-0x158], rax
- mov rax, qword ptr [rbp-0x150]
- cqo
- idiv rdx:rax, qword ptr [rbp-0x158]
- mov qword ptr [rbp-0x148], rax
- mov rcx, qword ptr [rbp-0x148]
- mov rax, qword ptr [rbp-0x128]
- mov qword ptr [rbp-0x168], rax
- mov rax, qword ptr [rbp-0x138]
- mov qword ptr [rbp-0x170], rax
- mov rax, qword ptr [rbp-0x168]
- cqo
- idiv rdx:rax, qword ptr [rbp-0x170]
- mov qword ptr [rbp-0x160], rax
- mov rax, qword ptr [rbp-0x160]
- mov qword ptr [rbp-0x180], rcx
- mov qword ptr [rbp-0x178], rax
- vinserti128 ymm1, ymm3, xmmword ptr [rbp-0x180], 1
- vmovups ymmword ptr [rdi], ymm0
- vmovups ymmword ptr [rdi+0x20], ymm1
- mov rax, rdi
- ;; size=297 bbWeight=1 PerfScore 320.25
+ vmovaps xmmword ptr [rbp-0x1E0], xmm1
+ mov rdi, qword ptr [rbp-0x1D0]
+ mov qword ptr [rbp-0x1F8], rdi
+ mov rdi, qword ptr [rbp-0x1E0]
+ mov qword ptr [rbp-0x200], rdi
+ mov rdi, qword ptr [rbp-0x1F8]
+ mov rsi, qword ptr [rbp-0x200]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Divide(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Divide(long,long):long
+ mov qword ptr [rbp-0x1F0], rax
+ mov r15, qword ptr [rbp-0x1F0]
+ mov rdi, qword ptr [rbp-0x1C8]
+ mov qword ptr [rsp], rdi
+ mov rdi, qword ptr [rbp-0x1D8]
+ mov qword ptr [rsp+0x08], rdi
+ lea rdi, [rbp-0x1E8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64`1[long]:op_Division(System.Runtime.Intrinsics.Vector64`1[long],System.Runtime.Intrinsics.Vector64`1[long]):System.Runtime.Intrinsics.Vector64`1[long]
+ call [rax]System.Runtime.Intrinsics.Vector64`1[long]:op_Division(System.Runtime.Intrinsics.Vector64`1[long],System.Runtime.Intrinsics.Vector64`1[long]):System.Runtime.Intrinsics.Vector64`1[long]
+ mov qword ptr [rbp-0x210], r15
+ mov rdi, qword ptr [rbp-0x1E8]
+ mov qword ptr [rsp], rdi
+ lea rdi, [rbp-0x210]
+ ;; size=345 bbWeight=1 PerfScore 65.00
G_M43701_IG04:
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[long](byref,System.Runtime.Intrinsics.Vector64`1[long])
+ call [rax]System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[long](byref,System.Runtime.Intrinsics.Vector64`1[long])
+ vmovaps xmm0, xmmword ptr [rbp-0x160]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0x210]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ lea rdi, [rbp-0x50]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[long](System.Runtime.Intrinsics.Vector128`1[long],System.Runtime.Intrinsics.Vector128`1[long]):System.Runtime.Intrinsics.Vector256`1[long]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[long](System.Runtime.Intrinsics.Vector128`1[long],System.Runtime.Intrinsics.Vector128`1[long]):System.Runtime.Intrinsics.Vector256`1[long]
+ vmovups ymm0, ymmword ptr [rbp-0x30]
+ vmovups ymmword ptr [rbp-0x250], ymm0
+ vmovups ymm0, ymmword ptr [rbp-0x50]
+ vmovups ymmword ptr [rsp], ymm0
+ lea rdi, [rbp-0x250]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[long](byref,System.Runtime.Intrinsics.Vector256`1[long])
+ call [rax]System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[long](byref,System.Runtime.Intrinsics.Vector256`1[long])
+ vmovdqu ymm0, ymmword ptr [rbp-0x250]
+ vmovdqu ymmword ptr [rbx], ymm0
+ vmovdqu ymm0, ymmword ptr [rbp-0x230]
+ vmovdqu ymmword ptr [rbx+0x20], ymm0
+ mov rax, rbx
+ ;; size=125 bbWeight=1 PerfScore 41.00
+G_M43701_IG05:
vzeroupper
- add rsp, 384
+ add rsp, 608
+ pop rbx
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 639, prolog size 16, PerfScore 667.75, instruction count 114, allocated bytes for code 639 (MethodHash=8284554a) for method System.Runtime.Intrinsics.Vector512:Divide[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):System.Runtime.Intrinsics.Vector512`1[long] (FullOpts)
+; Total bytes of code 872, prolog size 19, PerfScore 456.75, instruction count 145, allocated bytes for code 872 (MethodHash=8284554a) for method System.Runtime.Intrinsics.Vector512:Divide[long](System.Runtime.Intrinsics.Vector512`1[long],System.Runtime.Intrinsics.Vector512`1[long]):System.Runtime.Intrinsics.Vector512`1[long] (FullOpts)218 (16.02 % of base) - System.Runtime.Intrinsics.Vector512:Divide[short](System.Runtime.Intrinsics.Vector512`1[short],System.Runtime.Intrinsics.Vector512`1[short]):System.Runtime.Intrinsics.Vector512`1[short]
; Assembly listing for method System.Runtime.Intrinsics.Vector512:Divide[short](System.Runtime.Intrinsics.Vector512`1[short],System.Runtime.Intrinsics.Vector512`1[short]):System.Runtime.Intrinsics.Vector512`1[short] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 43 single block inlinees; 27 inlinees without PGO data
+; 0 inlinees with PGO data; 37 single block inlinees; 21 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T08] ( 5, 5 ) byref -> rdi single-def
+; V00 RetBuf [V00,T04] ( 4, 4 ) byref -> rbx single-def
;* V01 arg0 [V01 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[short]>
;* V02 arg1 [V02 ] ( 0, 0 ) struct (64) zero-ref single-def <System.Runtime.Intrinsics.Vector512`1[short]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T37] ( 2, 4 ) simd32 -> mm0 "impAppendStmt"
-; V05 tmp2 [V05,T38] ( 2, 4 ) simd32 -> mm1 "spilled call-like call argument"
-; V06 tmp3 [V06,T25] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V07 tmp4 [V07,T26] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V08 tmp5 [V08,T39] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T32] ( 2, 4 ) simd32 -> [rbp-0x30] spill-single-def "impAppendStmt"
+; V05 tmp2 [V05,T33] ( 2, 4 ) simd32 -> [rbp-0x50] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument"
+; V06 tmp3 [V06,T20] ( 3, 6 ) simd32 -> mm0 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
+; V07 tmp4 [V07,T21] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
+; V08 tmp5 [V08,T34] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
;* V09 tmp6 [V09 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V10 tmp7 [V10,T27] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V11 tmp8 [V11,T28] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V10 tmp7 [V10,T22] ( 3, 6 ) simd16 -> [rbp-0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V11 tmp8 [V11,T23] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
;* V12 tmp9 [V12 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V13 tmp10 [V13 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V14 tmp11 [V14 ] ( 5, 5 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V15 tmp12 [V15,T17] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V16 tmp13 [V16 ] ( 5, 10 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V17 tmp14 [V17 ] ( 5, 10 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V14 tmp11 [V14 ] ( 5, 5 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V15 tmp12 [V15,T13] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V16 tmp13 [V16 ] ( 5, 10 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V17 tmp14 [V17 ] ( 5, 10 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V18 tmp15 [V18 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V19 tmp16 [V19 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V20 tmp17 [V20,T00] ( 8, 8 ) short -> rcx "Inline return value spill temp"
+; V20 tmp17 [V20,T00] ( 8, 8 ) short -> rdi "Inline return value spill temp"
;* V21 tmp18 [V21 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
;* V22 tmp19 [V22 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V23 tmp20 [V23 ] ( 5, 5 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V24 tmp21 [V24,T18] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V25 tmp22 [V25 ] ( 5, 10 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V26 tmp23 [V26 ] ( 5, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V23 tmp20 [V23 ] ( 5, 5 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V24 tmp21 [V24,T14] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 5, 10 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V26 tmp23 [V26 ] ( 5, 10 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V27 tmp24 [V27 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V28 tmp25 [V28 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
; V29 tmp26 [V29,T01] ( 8, 8 ) short -> rsi "Inline return value spill temp"
;* V30 tmp27 [V30 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
;* V31 tmp28 [V31 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V32 tmp29 [V32,T41] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V33 tmp30 [V33,T29] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V34 tmp31 [V34,T30] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V32 tmp29 [V32,T36] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V33 tmp30 [V33,T24] ( 3, 6 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V34 tmp31 [V34,T25] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
;* V35 tmp32 [V35 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V36 tmp33 [V36 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V37 tmp34 [V37 ] ( 5, 5 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V38 tmp35 [V38,T19] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V39 tmp36 [V39 ] ( 5, 10 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V40 tmp37 [V40 ] ( 5, 10 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V37 tmp34 [V37 ] ( 5, 5 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V38 tmp35 [V38,T15] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V39 tmp36 [V39 ] ( 5, 10 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V40 tmp37 [V40 ] ( 5, 10 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V41 tmp38 [V41 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V42 tmp39 [V42 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V43 tmp40 [V43,T02] ( 8, 8 ) short -> rcx "Inline return value spill temp"
+; V43 tmp40 [V43,T02] ( 8, 8 ) short -> rdi "Inline return value spill temp"
;* V44 tmp41 [V44 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
;* V45 tmp42 [V45 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V46 tmp43 [V46 ] ( 5, 5 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V47 tmp44 [V47,T20] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V48 tmp45 [V48 ] ( 5, 10 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V49 tmp46 [V49 ] ( 5, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V46 tmp43 [V46 ] ( 5, 5 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V47 tmp44 [V47,T16] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V48 tmp45 [V48 ] ( 5, 10 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V49 tmp46 [V49 ] ( 5, 10 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V50 tmp47 [V50 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V51 tmp48 [V51 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
; V52 tmp49 [V52,T03] ( 8, 8 ) short -> rsi "Inline return value spill temp"
;* V53 tmp50 [V53 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
;* V54 tmp51 [V54 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V55 tmp52 [V55,T42] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V55 tmp52 [V55,T37] ( 3, 3 ) simd16 -> [rbp-0x110] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
;* V56 tmp53 [V56 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V57 tmp54 [V57,T31] ( 3, 6 ) simd32 -> mm1 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V58 tmp55 [V58,T32] ( 3, 6 ) simd32 -> mm2 "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V59 tmp56 [V59,T40] ( 2, 4 ) simd16 -> mm3 "impAppendStmt"
+; V57 tmp54 [V57,T26] ( 3, 6 ) simd32 -> [rbp-0x130] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
+; V58 tmp55 [V58,T27] ( 3, 6 ) simd32 -> [rbp-0x150] spill-single-def "Inlining Arg" <System.Runtime.Intrinsics.Vector256`1[short]>
+; V59 tmp56 [V59,T35] ( 2, 4 ) simd16 -> [rbp-0x160] spill-single-def "impAppendStmt"
;* V60 tmp57 [V60 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V61 tmp58 [V61,T33] ( 3, 6 ) simd16 -> [rbp-0xD0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V62 tmp59 [V62,T34] ( 3, 6 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V61 tmp58 [V61,T28] ( 3, 6 ) simd16 -> [rbp-0x170] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V62 tmp59 [V62,T29] ( 3, 6 ) simd16 -> [rbp-0x180] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
;* V63 tmp60 [V63 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V64 tmp61 [V64 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V65 tmp62 [V65 ] ( 5, 5 ) struct ( 8) [rbp-0xE8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V66 tmp63 [V66,T21] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V67 tmp64 [V67 ] ( 5, 10 ) struct ( 8) [rbp-0xF0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V68 tmp65 [V68 ] ( 5, 10 ) struct ( 8) [rbp-0xF8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V65 tmp62 [V65 ] ( 5, 5 ) struct ( 8) [rbp-0x188] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V66 tmp63 [V66,T17] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V67 tmp64 [V67 ] ( 5, 10 ) struct ( 8) [rbp-0x190] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V68 tmp65 [V68 ] ( 5, 10 ) struct ( 8) [rbp-0x198] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V69 tmp66 [V69 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V70 tmp67 [V70 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V71 tmp68 [V71,T04] ( 8, 8 ) short -> rcx "Inline return value spill temp"
-;* V72 tmp69 [V72 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V73 tmp70 [V73 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V74 tmp71 [V74 ] ( 5, 5 ) struct ( 8) [rbp-0x100] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V75 tmp72 [V75,T22] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V76 tmp73 [V76 ] ( 5, 10 ) struct ( 8) [rbp-0x108] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V77 tmp74 [V77 ] ( 5, 10 ) struct ( 8) [rbp-0x110] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V78 tmp75 [V78 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V79 tmp76 [V79 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V80 tmp77 [V80,T05] ( 8, 8 ) short -> rsi "Inline return value spill temp"
-;* V81 tmp78 [V81 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V82 tmp79 [V82 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V83 tmp80 [V83,T43] ( 3, 3 ) simd16 -> [rbp-0x120] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V84 tmp81 [V84,T35] ( 3, 6 ) simd16 -> [rbp-0x130] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V85 tmp82 [V85,T36] ( 3, 6 ) simd16 -> [rbp-0x140] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-;* V86 tmp83 [V86 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V87 tmp84 [V87 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V88 tmp85 [V88 ] ( 5, 5 ) struct ( 8) [rbp-0x148] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V89 tmp86 [V89,T23] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V90 tmp87 [V90 ] ( 5, 10 ) struct ( 8) [rbp-0x150] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V91 tmp88 [V91 ] ( 5, 10 ) struct ( 8) [rbp-0x158] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V92 tmp89 [V92 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V93 tmp90 [V93 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V94 tmp91 [V94,T06] ( 8, 8 ) short -> rcx "Inline return value spill temp"
-;* V95 tmp92 [V95 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V96 tmp93 [V96 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V97 tmp94 [V97 ] ( 5, 5 ) struct ( 8) [rbp-0x160] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V98 tmp95 [V98,T24] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V99 tmp96 [V99 ] ( 5, 10 ) struct ( 8) [rbp-0x168] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V100 tmp97 [V100 ] ( 5, 10 ) struct ( 8) [rbp-0x170] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V101 tmp98 [V101 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V102 tmp99 [V102 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V103 tmp100 [V103,T07] ( 8, 8 ) short -> rsi "Inline return value spill temp"
-;* V104 tmp101 [V104 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V105 tmp102 [V105 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V106 tmp103 [V106,T44] ( 3, 3 ) simd16 -> [rbp-0x180] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-;* V107 tmp104 [V107 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[short]>
-;* V108 tmp105 [V108 ] ( 0, 0 ) struct (64) zero-ref ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[short]>
-; V109 tmp106 [V109,T47] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
-; V110 tmp107 [V110,T48] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
-; V111 tmp108 [V111,T49] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
-; V112 tmp109 [V112,T50] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
-; V113 tmp110 [V113,T09] ( 2, 2 ) long -> rcx "field V12._00 (fldOffset=0x0)" P-INDEP
-; V114 tmp111 [V114,T10] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
-; V115 tmp112 [V115 ] ( 5, 5 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
-; V116 tmp113 [V116 ] ( 5, 9 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
-; V117 tmp114 [V117 ] ( 5, 9 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V118 tmp115 [V118 ] ( 5, 5 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
-; V119 tmp116 [V119 ] ( 5, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V120 tmp117 [V120 ] ( 5, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
-; V121 tmp118 [V121,T11] ( 2, 2 ) long -> rcx "field V35._00 (fldOffset=0x0)" P-INDEP
-; V122 tmp119 [V122,T12] ( 2, 2 ) long -> rax "field V36._00 (fldOffset=0x0)" P-INDEP
-; V123 tmp120 [V123 ] ( 5, 5 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
-; V124 tmp121 [V124 ] ( 5, 9 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V39._00 (fldOffset=0x0)" P-DEP
-; V125 tmp122 [V125 ] ( 5, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V40._00 (fldOffset=0x0)" P-DEP
-; V126 tmp123 [V126 ] ( 5, 5 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
-; V127 tmp124 [V127 ] ( 5, 9 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V48._00 (fldOffset=0x0)" P-DEP
-; V128 tmp125 [V128 ] ( 5, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
-; V129 tmp126 [V129,T13] ( 2, 2 ) long -> rcx "field V63._00 (fldOffset=0x0)" P-INDEP
-; V130 tmp127 [V130,T14] ( 2, 2 ) long -> rax "field V64._00 (fldOffset=0x0)" P-INDEP
-; V131 tmp128 [V131 ] ( 5, 5 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V65._00 (fldOffset=0x0)" P-DEP
-; V132 tmp129 [V132 ] ( 5, 9 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
-; V133 tmp130 [V133 ] ( 5, 9 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V68._00 (fldOffset=0x0)" P-DEP
-; V134 tmp131 [V134 ] ( 5, 5 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V74._00 (fldOffset=0x0)" P-DEP
-; V135 tmp132 [V135 ] ( 5, 9 ) long -> [rbp-0x108] do-not-enreg[X] addr-exposed "field V76._00 (fldOffset=0x0)" P-DEP
-; V136 tmp133 [V136 ] ( 5, 9 ) long -> [rbp-0x110] do-not-enreg[X] addr-exposed "field V77._00 (fldOffset=0x0)" P-DEP
-; V137 tmp134 [V137,T15] ( 2, 2 ) long -> rcx "field V86._00 (fldOffset=0x0)" P-INDEP
-; V138 tmp135 [V138,T16] ( 2, 2 ) long -> rax "field V87._00 (fldOffset=0x0)" P-INDEP
-; V139 tmp136 [V139 ] ( 5, 5 ) long -> [rbp-0x148] do-not-enreg[X] addr-exposed "field V88._00 (fldOffset=0x0)" P-DEP
-; V140 tmp137 [V140 ] ( 5, 9 ) long -> [rbp-0x150] do-not-enreg[X] addr-exposed "field V90._00 (fldOffset=0x0)" P-DEP
-; V141 tmp138 [V141 ] ( 5, 9 ) long -> [rbp-0x158] do-not-enreg[X] addr-exposed "field V91._00 (fldOffset=0x0)" P-DEP
-; V142 tmp139 [V142 ] ( 5, 5 ) long -> [rbp-0x160] do-not-enreg[X] addr-exposed "field V97._00 (fldOffset=0x0)" P-DEP
-; V143 tmp140 [V143 ] ( 5, 9 ) long -> [rbp-0x168] do-not-enreg[X] addr-exposed "field V99._00 (fldOffset=0x0)" P-DEP
-; V144 tmp141 [V144 ] ( 5, 9 ) long -> [rbp-0x170] do-not-enreg[X] addr-exposed "field V100._00 (fldOffset=0x0)" P-DEP
-; V145 tmp142 [V145,T45] ( 2, 2 ) simd32 -> mm0 "field V108._lower (fldOffset=0x0)" P-INDEP
-; V146 tmp143 [V146,T46] ( 2, 2 ) simd32 -> mm1 "field V108._upper (fldOffset=0x20)" P-INDEP
+; V71 tmp68 [V71 ] ( 5, 5 ) struct ( 8) [rbp-0x1A0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V72 tmp69 [V72,T18] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V73 tmp70 [V73 ] ( 5, 10 ) struct ( 8) [rbp-0x1A8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V74 tmp71 [V74 ] ( 5, 10 ) struct ( 8) [rbp-0x1B0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V75 tmp72 [V75 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V76 tmp73 [V76 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
+; V77 tmp74 [V77,T38] ( 3, 3 ) simd16 -> [rbp-0x1C0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V78 tmp75 [V78,T30] ( 3, 6 ) simd16 -> [rbp-0x1D0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V79 tmp76 [V79,T31] ( 3, 6 ) simd16 -> [rbp-0x1E0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+;* V80 tmp77 [V80 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V81 tmp78 [V81 ] ( 2, 4 ) struct ( 8) [rbp-0x1E8] do-not-enreg[HS] hidden-struct-arg "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V82 tmp79 [V82 ] ( 5, 5 ) struct ( 8) [rbp-0x1F0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V83 tmp80 [V83,T19] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V84 tmp81 [V84 ] ( 5, 10 ) struct ( 8) [rbp-0x1F8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V85 tmp82 [V85 ] ( 5, 10 ) struct ( 8) [rbp-0x200] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V86 tmp83 [V86 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V87 tmp84 [V87 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
+; V88 tmp85 [V88 ] ( 3, 3 ) simd16 -> [rbp-0x210] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V89 tmp86 [V89 ] ( 3, 3 ) struct (64) [rbp-0x250] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector512`1[short]>
+; V90 tmp87 [V90,T39] ( 1, 1 ) simd32 -> [rbp+0x10] single-def "field V01._lower (fldOffset=0x0)" P-INDEP
+; V91 tmp88 [V91,T40] ( 1, 1 ) simd32 -> [rbp+0x30] single-def "field V01._upper (fldOffset=0x20)" P-INDEP
+; V92 tmp89 [V92,T41] ( 1, 1 ) simd32 -> [rbp+0x50] single-def "field V02._lower (fldOffset=0x0)" P-INDEP
+; V93 tmp90 [V93,T42] ( 1, 1 ) simd32 -> [rbp+0x70] single-def "field V02._upper (fldOffset=0x20)" P-INDEP
+; V94 tmp91 [V94,T06] ( 2, 2 ) long -> rdi "field V12._00 (fldOffset=0x0)" P-INDEP
+; V95 tmp92 [V95,T07] ( 2, 2 ) long -> rax "field V13._00 (fldOffset=0x0)" P-INDEP
+; V96 tmp93 [V96 ] ( 5, 5 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
+; V97 tmp94 [V97 ] ( 5, 9 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V98 tmp95 [V98 ] ( 5, 9 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V99 tmp96 [V99 ] ( 5, 5 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V100 tmp97 [V100 ] ( 5, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V101 tmp98 [V101 ] ( 5, 9 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
+; V102 tmp99 [V102,T08] ( 2, 2 ) long -> rdi "field V35._00 (fldOffset=0x0)" P-INDEP
+; V103 tmp100 [V103,T09] ( 2, 2 ) long -> rsi "field V36._00 (fldOffset=0x0)" P-INDEP
+; V104 tmp101 [V104 ] ( 5, 5 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
+; V105 tmp102 [V105 ] ( 5, 9 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V39._00 (fldOffset=0x0)" P-DEP
+; V106 tmp103 [V106 ] ( 5, 9 ) long -> [rbp-0xE8] do-not-enreg[X] addr-exposed "field V40._00 (fldOffset=0x0)" P-DEP
+; V107 tmp104 [V107 ] ( 5, 5 ) long -> [rbp-0xF0] do-not-enreg[X] addr-exposed "field V46._00 (fldOffset=0x0)" P-DEP
+; V108 tmp105 [V108 ] ( 5, 9 ) long -> [rbp-0xF8] do-not-enreg[X] addr-exposed "field V48._00 (fldOffset=0x0)" P-DEP
+; V109 tmp106 [V109 ] ( 5, 9 ) long -> [rbp-0x100] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
+; V110 tmp107 [V110,T10] ( 2, 2 ) long -> r15 "field V63._00 (fldOffset=0x0)" P-INDEP
+; V111 tmp108 [V111,T11] ( 2, 2 ) long -> rdi "field V64._00 (fldOffset=0x0)" P-INDEP
+; V112 tmp109 [V112 ] ( 5, 5 ) long -> [rbp-0x188] do-not-enreg[X] addr-exposed "field V65._00 (fldOffset=0x0)" P-DEP
+; V113 tmp110 [V113 ] ( 5, 9 ) long -> [rbp-0x190] do-not-enreg[X] addr-exposed "field V67._00 (fldOffset=0x0)" P-DEP
+; V114 tmp111 [V114 ] ( 5, 9 ) long -> [rbp-0x198] do-not-enreg[X] addr-exposed "field V68._00 (fldOffset=0x0)" P-DEP
+; V115 tmp112 [V115 ] ( 5, 5 ) long -> [rbp-0x1A0] do-not-enreg[X] addr-exposed "field V71._00 (fldOffset=0x0)" P-DEP
+; V116 tmp113 [V116 ] ( 5, 9 ) long -> [rbp-0x1A8] do-not-enreg[X] addr-exposed "field V73._00 (fldOffset=0x0)" P-DEP
+; V117 tmp114 [V117 ] ( 5, 9 ) long -> [rbp-0x1B0] do-not-enreg[X] addr-exposed "field V74._00 (fldOffset=0x0)" P-DEP
+; V118 tmp115 [V118,T12] ( 2, 2 ) long -> r15 "field V80._00 (fldOffset=0x0)" P-INDEP
+; V119 tmp116 [V119,T05] ( 2, 3 ) long -> [rbp-0x1E8] do-not-enreg[H] hidden-struct-arg "field V81._00 (fldOffset=0x0)" P-DEP
+; V120 tmp117 [V120 ] ( 5, 5 ) long -> [rbp-0x1F0] do-not-enreg[X] addr-exposed "field V82._00 (fldOffset=0x0)" P-DEP
+; V121 tmp118 [V121 ] ( 5, 9 ) long -> [rbp-0x1F8] do-not-enreg[X] addr-exposed "field V84._00 (fldOffset=0x0)" P-DEP
+; V122 tmp119 [V122 ] ( 5, 9 ) long -> [rbp-0x200] do-not-enreg[X] addr-exposed "field V85._00 (fldOffset=0x0)" P-DEP
+; V123 tmp120 [V123 ] ( 3, 3 ) simd32 -> [rbp-0x250] do-not-enreg[XS] addr-exposed "field V89._lower (fldOffset=0x0)" P-DEP
+; V124 tmp121 [V124 ] ( 3, 3 ) simd32 -> [rbp-0x230] do-not-enreg[XS] addr-exposed "field V89._upper (fldOffset=0x20)" P-DEP
;
-; Lcl frame size = 384
+; Lcl frame size = 608
G_M40213_IG01:
push rbp
- sub rsp, 384
- lea rbp, [rsp+0x180]
- ;; size=16 bbWeight=1 PerfScore 1.75
+ push r15
+ push rbx
+ sub rsp, 608
+ lea rbp, [rsp+0x270]
+ mov rbx, rdi
+ ;; size=22 bbWeight=1 PerfScore 4.00
G_M40213_IG02:
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x50]
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x60], xmm2
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- movsx rax, word ptr [rbp-0x30]
- movsx rcx, word ptr [rbp-0x38]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x28], cx
- movsx rax, word ptr [rbp-0x2E]
- movsx rcx, word ptr [rbp-0x36]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x26], cx
- movsx rax, word ptr [rbp-0x2C]
- movsx rcx, word ptr [rbp-0x34]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x24], cx
- movsx rax, word ptr [rbp-0x2A]
- movsx rcx, word ptr [rbp-0x32]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x22], cx
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
- movsx rax, word ptr [rbp-0x48]
- movsx rsi, word ptr [rbp-0x50]
+ vmovaps xmmword ptr [rbp-0x70], xmm2
+ mov rax, qword ptr [rbp-0x60]
+ mov qword ptr [rbp-0x80], rax
+ mov rax, qword ptr [rbp-0x70]
+ mov qword ptr [rbp-0x88], rax
+ movsx rax, word ptr [rbp-0x80]
+ movsx rdi, word ptr [rbp-0x88]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0x78], di
+ movsx rax, word ptr [rbp-0x7E]
+ movsx rdi, word ptr [rbp-0x86]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0x76], di
+ movsx rax, word ptr [rbp-0x7C]
+ movsx rdi, word ptr [rbp-0x84]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0x74], di
+ movsx rax, word ptr [rbp-0x7A]
+ movsx rdi, word ptr [rbp-0x82]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0x72], di
+ mov rdi, qword ptr [rbp-0x78]
+ mov rax, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x98], rax
+ mov rax, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0xA0], rax
+ movsx rax, word ptr [rbp-0x98]
+ movsx rsi, word ptr [rbp-0xA0]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x40], si
- movsx rax, word ptr [rbp-0x46]
- movsx rsi, word ptr [rbp-0x4E]
+ mov word ptr [rbp-0x90], si
+ movsx rax, word ptr [rbp-0x96]
+ movsx rsi, word ptr [rbp-0x9E]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x3E], si
- movsx rax, word ptr [rbp-0x44]
- movsx rsi, word ptr [rbp-0x4C]
+ mov word ptr [rbp-0x8E], si
+ movsx rax, word ptr [rbp-0x94]
+ movsx rsi, word ptr [rbp-0x9C]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x3C], si
- movsx rax, word ptr [rbp-0x42]
- movsx rsi, word ptr [rbp-0x4A]
+ mov word ptr [rbp-0x8C], si
+ movsx rax, word ptr [rbp-0x92]
+ movsx rsi, word ptr [rbp-0x9A]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x3A], si
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- ;; size=240 bbWeight=1 PerfScore 283.50
+ mov word ptr [rbp-0x8A], si
+ mov rax, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xB0], rdi
+ ;; size=303 bbWeight=1 PerfScore 283.50
G_M40213_IG03:
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm2, xmmword ptr [rbp-0xB0]
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
+ vmovaps xmmword ptr [rbp-0xC0], xmm0
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- movsx rax, word ptr [rbp-0x90]
- movsx rcx, word ptr [rbp-0x98]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x88], cx
- movsx rax, word ptr [rbp-0x8E]
- movsx rcx, word ptr [rbp-0x96]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x86], cx
- movsx rax, word ptr [rbp-0x8C]
- movsx rcx, word ptr [rbp-0x94]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x84], cx
- movsx rax, word ptr [rbp-0x8A]
- movsx rcx, word ptr [rbp-0x92]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x82], cx
- mov rcx, qword ptr [rbp-0x88]
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
- movsx rax, word ptr [rbp-0xA8]
- movsx rsi, word ptr [rbp-0xB0]
+ vmovaps xmmword ptr [rbp-0xD0], xmm0
+ mov rax, qword ptr [rbp-0xC0]
+ mov qword ptr [rbp-0xE0], rax
+ mov rax, qword ptr [rbp-0xD0]
+ mov qword ptr [rbp-0xE8], rax
+ movsx rax, word ptr [rbp-0xE0]
+ movsx rdi, word ptr [rbp-0xE8]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0xD8], di
+ movsx rax, word ptr [rbp-0xDE]
+ movsx rdi, word ptr [rbp-0xE6]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0xD6], di
+ movsx rax, word ptr [rbp-0xDC]
+ movsx rdi, word ptr [rbp-0xE4]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0xD4], di
+ movsx rax, word ptr [rbp-0xDA]
+ movsx rdi, word ptr [rbp-0xE2]
+ cdq
+ idiv edx:eax, edi
+ movsx rdi, ax
+ mov word ptr [rbp-0xD2], di
+ mov rdi, qword ptr [rbp-0xD8]
+ mov rax, qword ptr [rbp-0xB8]
+ mov qword ptr [rbp-0xF8], rax
+ mov rax, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0x100], rax
+ movsx rax, word ptr [rbp-0xF8]
+ movsx rsi, word ptr [rbp-0x100]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0xA0], si
- movsx rax, word ptr [rbp-0xA6]
- movsx rsi, word ptr [rbp-0xAE]
+ mov word ptr [rbp-0xF0], si
+ movsx rax, word ptr [rbp-0xF6]
+ movsx rsi, word ptr [rbp-0xFE]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x9E], si
- movsx rax, word ptr [rbp-0xA4]
- movsx rsi, word ptr [rbp-0xAC]
+ mov word ptr [rbp-0xEE], si
+ movsx rax, word ptr [rbp-0xF4]
+ movsx rsi, word ptr [rbp-0xFC]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x9C], si
- movsx rax, word ptr [rbp-0xA2]
- movsx rsi, word ptr [rbp-0xAA]
+ mov word ptr [rbp-0xEC], si
+ movsx rax, word ptr [rbp-0xF2]
+ movsx rsi, word ptr [rbp-0xFA]
cdq
idiv edx:eax, esi
movsx rsi, ax
- mov word ptr [rbp-0x9A], si
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- ;; size=336 bbWeight=1 PerfScore 283.00
+ mov word ptr [rbp-0xEA], si
+ mov rsi, qword ptr [rbp-0xF0]
+ mov qword ptr [rbp-0x110], rdi
+ ;; size=360 bbWeight=1 PerfScore 283.00
G_M40213_IG04:
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
+ mov qword ptr [rbp-0x108], rsi
+ vinserti128 ymm0, ymm2, xmmword ptr [rbp-0x110], 1
+ vmovups ymmword ptr [rbp-0x30], ymm0
vmovups ymm1, ymmword ptr [rbp+0x30]
+ vmovups ymmword ptr [rbp-0x130], ymm1
vmovups ymm2, ymmword ptr [rbp+0x70]
+ vmovups ymmword ptr [rbp-0x150], ymm2
vmovaps ymm3, ymm1
- vmovaps xmmword ptr [rbp-0xD0], xmm3
+ vmovaps xmmword ptr [rbp-0x170], xmm3
vmovaps ymm3, ymm2
- vmovaps xmmword ptr [rbp-0xE0], xmm3
- mov rax, qword ptr [rbp-0xD0]
- mov qword ptr [rbp-0xF0], rax
- mov rax, qword ptr [rbp-0xE0]
- mov qword ptr [rbp-0xF8], rax
- movsx rax, word ptr [rbp-0xF0]
- movsx rcx, word ptr [rbp-0xF8]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0xE8], cx
- movsx rax, word ptr [rbp-0xEE]
- movsx rcx, word ptr [rbp-0xF6]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0xE6], cx
- movsx rax, word ptr [rbp-0xEC]
- movsx rcx, word ptr [rbp-0xF4]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0xE4], cx
- movsx rax, word ptr [rbp-0xEA]
- movsx rcx, word ptr [rbp-0xF2]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0xE2], cx
- mov rcx, qword ptr [rbp-0xE8]
- mov rax, qword ptr [rbp-0xC8]
- mov qword ptr [rbp-0x108], rax
- mov rax, qword ptr [rbp-0xD8]
- mov qword ptr [rbp-0x110], rax
- movsx rax, word ptr [rbp-0x108]
- movsx rsi, word ptr [rbp-0x110]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x100], si
- movsx rax, word ptr [rbp-0x106]
- movsx rsi, word ptr [rbp-0x10E]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0xFE], si
- movsx rax, word ptr [rbp-0x104]
- movsx rsi, word ptr [rbp-0x10C]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0xFC], si
- movsx rax, word ptr [rbp-0x102]
- movsx rsi, word ptr [rbp-0x10A]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0xFA], si
- ;; size=354 bbWeight=1 PerfScore 286.50
+ vmovaps xmmword ptr [rbp-0x180], xmm3
+ mov rdi, qword ptr [rbp-0x170]
+ mov qword ptr [rbp-0x190], rdi
+ mov rdi, qword ptr [rbp-0x180]
+ mov qword ptr [rbp-0x198], rdi
+ movsx rdi, word ptr [rbp-0x190]
+ movsx rsi, word ptr [rbp-0x198]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x188], ax
+ movsx rdi, word ptr [rbp-0x18E]
+ movsx rsi, word ptr [rbp-0x196]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x186], ax
+ movsx rdi, word ptr [rbp-0x18C]
+ movsx rsi, word ptr [rbp-0x194]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x184], ax
+ movsx rdi, word ptr [rbp-0x18A]
+ movsx rsi, word ptr [rbp-0x192]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x182], ax
+ mov r15, qword ptr [rbp-0x188]
+ mov rdi, qword ptr [rbp-0x168]
+ mov qword ptr [rbp-0x1A8], rdi
+ mov rdi, qword ptr [rbp-0x178]
+ mov qword ptr [rbp-0x1B0], rdi
+ movsx rdi, word ptr [rbp-0x1A8]
+ movsx rsi, word ptr [rbp-0x1B0]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x1A0], ax
+ movsx rdi, word ptr [rbp-0x1A6]
+ movsx rsi, word ptr [rbp-0x1AE]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ ;; size=336 bbWeight=1 PerfScore 85.00
G_M40213_IG05:
- mov rax, qword ptr [rbp-0x100]
- mov qword ptr [rbp-0x120], rcx
- mov qword ptr [rbp-0x118], rax
- vmovaps xmm3, xmmword ptr [rbp-0x120]
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x19E], ax
+ movsx rdi, word ptr [rbp-0x1A4]
+ movsx rsi, word ptr [rbp-0x1AC]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x19C], ax
+ movsx rdi, word ptr [rbp-0x1A2]
+ movsx rsi, word ptr [rbp-0x1AA]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x19A], ax
+ mov rdi, qword ptr [rbp-0x1A0]
+ mov qword ptr [rbp-0x1C0], r15
+ mov qword ptr [rbp-0x1B8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x1C0]
+ vmovaps xmmword ptr [rbp-0x160], xmm0
+ vmovups ymm1, ymmword ptr [rbp-0x130]
vextractf128 xmm1, ymm1, 1
- vmovaps xmmword ptr [rbp-0x130], xmm1
+ vmovaps xmmword ptr [rbp-0x1D0], xmm1
+ vmovups ymm2, ymmword ptr [rbp-0x150]
vextractf128 xmm1, ymm2, 1
- vmovaps xmmword ptr [rbp-0x140], xmm1
- mov rax, qword ptr [rbp-0x130]
- mov qword ptr [rbp-0x150], rax
- mov rax, qword ptr [rbp-0x140]
- mov qword ptr [rbp-0x158], rax
- movsx rax, word ptr [rbp-0x150]
- movsx rcx, word ptr [rbp-0x158]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x148], cx
- movsx rax, word ptr [rbp-0x14E]
- movsx rcx, word ptr [rbp-0x156]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x146], cx
- movsx rax, word ptr [rbp-0x14C]
- movsx rcx, word ptr [rbp-0x154]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x144], cx
- movsx rax, word ptr [rbp-0x14A]
- movsx rcx, word ptr [rbp-0x152]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x142], cx
- mov rcx, qword ptr [rbp-0x148]
- mov rax, qword ptr [rbp-0x128]
- mov qword ptr [rbp-0x168], rax
- mov rax, qword ptr [rbp-0x138]
- mov qword ptr [rbp-0x170], rax
- movsx rax, word ptr [rbp-0x168]
- movsx rsi, word ptr [rbp-0x170]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x160], si
- movsx rax, word ptr [rbp-0x166]
- movsx rsi, word ptr [rbp-0x16E]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x15E], si
- movsx rax, word ptr [rbp-0x164]
- movsx rsi, word ptr [rbp-0x16C]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x15C], si
- movsx rax, word ptr [rbp-0x162]
- movsx rsi, word ptr [rbp-0x16A]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x15A], si
- ;; size=360 bbWeight=1 PerfScore 283.00
+ vmovaps xmmword ptr [rbp-0x1E0], xmm1
+ mov rdi, qword ptr [rbp-0x1D0]
+ mov qword ptr [rbp-0x1F8], rdi
+ mov rdi, qword ptr [rbp-0x1E0]
+ mov qword ptr [rbp-0x200], rdi
+ movsx rdi, word ptr [rbp-0x1F8]
+ movsx rsi, word ptr [rbp-0x200]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x1F0], ax
+ movsx rdi, word ptr [rbp-0x1F6]
+ movsx rsi, word ptr [rbp-0x1FE]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x1EE], ax
+ movsx rdi, word ptr [rbp-0x1F4]
+ movsx rsi, word ptr [rbp-0x1FC]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x1EC], ax
+ movsx rdi, word ptr [rbp-0x1F2]
+ movsx rsi, word ptr [rbp-0x1FA]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ ;; size=321 bbWeight=1 PerfScore 89.50
G_M40213_IG06:
- mov rax, qword ptr [rbp-0x160]
- mov qword ptr [rbp-0x180], rcx
- mov qword ptr [rbp-0x178], rax
- vinserti128 ymm1, ymm3, xmmword ptr [rbp-0x180], 1
- vmovups ymmword ptr [rdi], ymm0
- vmovups ymmword ptr [rdi+0x20], ymm1
- mov rax, rdi
- ;; size=43 bbWeight=1 PerfScore 11.25
+ mov word ptr [rbp-0x1EA], ax
+ mov r15, qword ptr [rbp-0x1F0]
+ mov rdi, qword ptr [rbp-0x1C8]
+ mov qword ptr [rsp], rdi
+ mov rdi, qword ptr [rbp-0x1D8]
+ mov qword ptr [rsp+0x08], rdi
+ lea rdi, [rbp-0x1E8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector64`1[short]:op_Division(System.Runtime.Intrinsics.Vector64`1[short],System.Runtime.Intrinsics.Vector64`1[short]):System.Runtime.Intrinsics.Vector64`1[short]
+ call [rax]System.Runtime.Intrinsics.Vector64`1[short]:op_Division(System.Runtime.Intrinsics.Vector64`1[short],System.Runtime.Intrinsics.Vector64`1[short]):System.Runtime.Intrinsics.Vector64`1[short]
+ mov qword ptr [rbp-0x210], r15
+ mov rdi, qword ptr [rbp-0x1E8]
+ mov qword ptr [rsp], rdi
+ lea rdi, [rbp-0x210]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[short](byref,System.Runtime.Intrinsics.Vector64`1[short])
+ call [rax]System.Runtime.Intrinsics.Vector128:SetUpperUnsafe[short](byref,System.Runtime.Intrinsics.Vector64`1[short])
+ vmovaps xmm0, xmmword ptr [rbp-0x160]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0x210]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ lea rdi, [rbp-0x50]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[short](System.Runtime.Intrinsics.Vector128`1[short],System.Runtime.Intrinsics.Vector128`1[short]):System.Runtime.Intrinsics.Vector256`1[short]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[short](System.Runtime.Intrinsics.Vector128`1[short],System.Runtime.Intrinsics.Vector128`1[short]):System.Runtime.Intrinsics.Vector256`1[short]
+ vmovups ymm0, ymmword ptr [rbp-0x30]
+ vmovups ymmword ptr [rsp], ymm0
+ lea rdi, [rbp-0x250]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector512:SetLowerUnsafe[short](byref,System.Runtime.Intrinsics.Vector256`1[short])
+ call [rax]System.Runtime.Intrinsics.Vector512:SetLowerUnsafe[short](byref,System.Runtime.Intrinsics.Vector256`1[short])
+ vmovups ymm0, ymmword ptr [rbp-0x50]
+ vmovups ymmword ptr [rsp], ymm0
+ lea rdi, [rbp-0x250]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[short](byref,System.Runtime.Intrinsics.Vector256`1[short])
+ call [rax]System.Runtime.Intrinsics.Vector512:SetUpperUnsafe[short](byref,System.Runtime.Intrinsics.Vector256`1[short])
+ vmovdqu ymm0, ymmword ptr [rbp-0x250]
+ vmovdqu ymmword ptr [rbx], ymm0
+ vmovdqu ymm0, ymmword ptr [rbp-0x230]
+ vmovdqu ymmword ptr [rbx+0x20], ymm0
+ mov rax, rbx
+ ;; size=222 bbWeight=1 PerfScore 58.00
G_M40213_IG07:
vzeroupper
- add rsp, 384
+ add rsp, 608
+ pop rbx
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 1361, prolog size 16, PerfScore 1151.75, instruction count 274, allocated bytes for code 1361 (MethodHash=1f9d62ea) for method System.Runtime.Intrinsics.Vector512:Divide[short](System.Runtime.Intrinsics.Vector512`1[short],System.Runtime.Intrinsics.Vector512`1[short]):System.Runtime.Intrinsics.Vector512`1[short] (FullOpts)
+; Total bytes of code 1579, prolog size 19, PerfScore 806.75, instruction count 273, allocated bytes for code 1579 (MethodHash=1f9d62ea) for method System.Runtime.Intrinsics.Vector512:Divide[short](System.Runtime.Intrinsics.Vector512`1[short],System.Runtime.Intrinsics.Vector512`1[short]):System.Runtime.Intrinsics.Vector512`1[short] (FullOpts)169 (81.25 % of base) - System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],int):System.Runtime.Intrinsics.Vector256`1[int]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],int):System.Runtime.Intrinsics.Vector256`1[int] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 16 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 16 single block inlinees; 9 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T15] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[int]>
-; V02 arg1 [V02,T00] ( 10, 10 ) int -> rsi single-def
+; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> r15 single-def
+; V01 arg0 [V01,T15] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[int]>
+; V02 arg1 [V02,T00] ( 10, 10 ) int -> rbx single-def
;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T12] ( 2, 4 ) simd16 -> mm1 "impAppendStmt"
+; V04 tmp1 [V04,T12] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T10] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V06 tmp3 [V06,T10] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
;* V07 tmp4 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V09 tmp6 [V09 ] ( 3, 3 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V09 tmp6 [V09 ] ( 3, 3 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V10 tmp7 [V10,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V11 tmp8 [V11 ] ( 3, 6 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V11 tmp8 [V11 ] ( 3, 6 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V12 tmp9 [V12 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V13 tmp10 [V13 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V14 tmp11 [V14 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V15 tmp12 [V15 ] ( 3, 3 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V16 tmp13 [V16,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V17 tmp14 [V17 ] ( 3, 6 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V18 tmp15 [V18 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V19 tmp16 [V19 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V20 tmp17 [V20 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V21 tmp18 [V21,T13] ( 3, 3 ) simd16 -> [rbp-0x40] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V22 tmp19 [V22,T11] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V24 tmp21 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V25 tmp22 [V25 ] ( 3, 3 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V26 tmp23 [V26,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V27 tmp24 [V27 ] ( 3, 6 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V13 tmp10 [V13 ] ( 3, 3 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V14 tmp11 [V14,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V15 tmp12 [V15 ] ( 3, 6 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V16 tmp13 [V16 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V17 tmp14 [V17,T13] ( 3, 3 ) simd16 -> [rbp-0x70] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V18 tmp15 [V18,T11] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+;* V19 tmp16 [V19 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V21 tmp18 [V21 ] ( 3, 3 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V22 tmp19 [V22,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V23 tmp20 [V23 ] ( 3, 6 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V24 tmp21 [V24 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 3, 3 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V26 tmp23 [V26,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V27 tmp24 [V27 ] ( 3, 6 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V28 tmp25 [V28 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V29 tmp26 [V29 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V30 tmp27 [V30 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V31 tmp28 [V31 ] ( 3, 3 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V32 tmp29 [V32,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 3, 6 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V34 tmp31 [V34 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V35 tmp32 [V35 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V36 tmp33 [V36 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V37 tmp34 [V37,T14] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-;* V38 tmp35 [V38 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V39 tmp36 [V39,T02] ( 2, 2 ) long -> rcx "field V07._00 (fldOffset=0x0)" P-INDEP
-; V40 tmp37 [V40,T03] ( 2, 2 ) long -> rax "field V08._00 (fldOffset=0x0)" P-INDEP
-; V41 tmp38 [V41 ] ( 3, 3 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V42 tmp39 [V42 ] ( 3, 5 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V43 tmp40 [V43 ] ( 3, 3 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V44 tmp41 [V44 ] ( 3, 5 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V45 tmp42 [V45,T04] ( 2, 2 ) long -> rcx "field V23._00 (fldOffset=0x0)" P-INDEP
-; V46 tmp43 [V46,T05] ( 2, 2 ) long -> rax "field V24._00 (fldOffset=0x0)" P-INDEP
-; V47 tmp44 [V47 ] ( 3, 3 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp45 [V48 ] ( 3, 5 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
-; V49 tmp46 [V49 ] ( 3, 3 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V50 tmp47 [V50 ] ( 3, 5 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V29 tmp26 [V29,T14] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+;* V30 tmp27 [V30 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[int]>
+; V31 tmp28 [V31,T02] ( 2, 2 ) long -> r14 "field V07._00 (fldOffset=0x0)" P-INDEP
+; V32 tmp29 [V32,T03] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+; V33 tmp30 [V33 ] ( 3, 3 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V34 tmp31 [V34 ] ( 3, 5 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
+; V35 tmp32 [V35 ] ( 3, 3 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V36 tmp33 [V36 ] ( 3, 5 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V37 tmp34 [V37,T04] ( 2, 2 ) long -> r14 "field V19._00 (fldOffset=0x0)" P-INDEP
+; V38 tmp35 [V38,T05] ( 2, 2 ) long -> rax "field V20._00 (fldOffset=0x0)" P-INDEP
+; V39 tmp36 [V39 ] ( 3, 3 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
+; V40 tmp37 [V40 ] ( 3, 5 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V41 tmp38 [V41 ] ( 3, 3 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V42 tmp39 [V42 ] ( 3, 5 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
;
-; Lcl frame size = 128
+; Lcl frame size = 152
G_M30697_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
+ push r15
+ push r14
+ push rbx
+ sub rsp, 152
+ lea rbp, [rsp+0xB0]
+ mov r15, rdi
+ mov ebx, esi
vmovups ymm0, ymmword ptr [rbp+0x10]
- ;; size=21 bbWeight=1 PerfScore 5.75
+ ;; size=31 bbWeight=1 PerfScore 9.25
G_M30697_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm1, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm1
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x20], rax
- mov eax, dword ptr [rbp-0x20]
- cdq
- idiv edx:eax, esi
- mov dword ptr [rbp-0x18], eax
- mov eax, dword ptr [rbp-0x1C]
- cdq
- idiv edx:eax, esi
- mov dword ptr [rbp-0x14], eax
- mov rcx, qword ptr [rbp-0x18]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x30], rax
- mov eax, dword ptr [rbp-0x30]
- cdq
- idiv edx:eax, esi
- mov dword ptr [rbp-0x28], eax
- mov eax, dword ptr [rbp-0x2C]
- cdq
- idiv edx:eax, esi
- mov dword ptr [rbp-0x24], eax
- mov rax, qword ptr [rbp-0x28]
- mov qword ptr [rbp-0x40], rcx
- mov qword ptr [rbp-0x38], rax
- vmovaps xmm1, xmmword ptr [rbp-0x40]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- mov rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x60], rax
- mov eax, dword ptr [rbp-0x60]
- cdq
- idiv edx:eax, esi
+ vmovaps xmmword ptr [rbp-0x40], xmm1
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rdi
+ mov edi, dword ptr [rbp-0x50]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x48], eax
+ mov edi, dword ptr [rbp-0x4C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x44], eax
+ mov r14, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x60], rdi
+ mov edi, dword ptr [rbp-0x60]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
mov dword ptr [rbp-0x58], eax
- mov eax, dword ptr [rbp-0x5C]
- cdq
- idiv edx:eax, esi
+ mov edi, dword ptr [rbp-0x5C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
mov dword ptr [rbp-0x54], eax
- mov rcx, qword ptr [rbp-0x58]
- mov rax, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x70], rax
- mov eax, dword ptr [rbp-0x70]
- cdq
- idiv edx:eax, esi
- mov dword ptr [rbp-0x68], eax
- mov eax, dword ptr [rbp-0x6C]
- cdq
- idiv edx:eax, esi
- mov dword ptr [rbp-0x64], eax
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0x80], rcx
- mov qword ptr [rbp-0x78], rax
- vinserti128 ymm0, ymm1, xmmword ptr [rbp-0x80], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=175 bbWeight=1 PerfScore 249.50
+ mov rdi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], r14
+ mov qword ptr [rbp-0x68], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x70]
+ vmovaps xmmword ptr [rbp-0x30], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x80], xmm1
+ mov rdi, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0x90], rdi
+ mov edi, dword ptr [rbp-0x90]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x88], eax
+ mov edi, dword ptr [rbp-0x8C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x84], eax
+ mov r14, qword ptr [rbp-0x88]
+ ;; size=222 bbWeight=1 PerfScore 57.25
G_M30697_IG03:
+ mov rdi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xA0], rdi
+ mov edi, dword ptr [rbp-0xA0]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x98], eax
+ mov edi, dword ptr [rbp-0x9C]
+ mov esi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x94], eax
+ mov rax, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], r14
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm0, xmmword ptr [rbp-0x30]
+ vinserti128 ymm0, ymm0, xmmword ptr [rbp-0xB0], 1
+ vmovups ymmword ptr [r15], ymm0
+ mov rax, r15
+ ;; size=107 bbWeight=1 PerfScore 25.25
+G_M30697_IG04:
vzeroupper
- add rsp, 128
+ add rsp, 152
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 208, prolog size 16, PerfScore 258.00, instruction count 64, allocated bytes for code 208 (MethodHash=d2e48816) for method System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],int):System.Runtime.Intrinsics.Vector256`1[int] (FullOpts)
+; Total bytes of code 377, prolog size 21, PerfScore 96.00, instruction count 84, allocated bytes for code 377 (MethodHash=d2e48816) for method System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],int):System.Runtime.Intrinsics.Vector256`1[int] (FullOpts)165 (49.11 % of base) - System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],System.Runtime.Intrinsics.Vector256`1[int]):System.Runtime.Intrinsics.Vector256`1[int]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],System.Runtime.Intrinsics.Vector256`1[int]):System.Runtime.Intrinsics.Vector256`1[int] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 20 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 20 single block inlinees; 8 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[int]>
-; V02 arg1 [V02,T17] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[int]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rbx single-def
+; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[int]>
+; V02 arg1 [V02,T17] ( 2, 2 ) simd32 -> [rbp+0x30] single-def <System.Runtime.Intrinsics.Vector256`1[int]>
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> [rbp-0x20] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T09] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V07 tmp4 [V07,T10] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V06 tmp3 [V06,T09] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V07 tmp4 [V07,T10] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V09 tmp6 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V10 tmp7 [V10 ] ( 3, 3 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V10 tmp7 [V10 ] ( 3, 3 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V11 tmp8 [V11,T05] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V12 tmp9 [V12 ] ( 3, 6 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V13 tmp10 [V13 ] ( 3, 6 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V12 tmp9 [V12 ] ( 3, 6 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V13 tmp10 [V13 ] ( 3, 6 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
;* V14 tmp11 [V14 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V15 tmp12 [V15 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V16 tmp13 [V16 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V17 tmp14 [V17 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V18 tmp15 [V18 ] ( 3, 3 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V19 tmp16 [V19,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V20 tmp17 [V20 ] ( 3, 6 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V21 tmp18 [V21 ] ( 3, 6 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V22 tmp19 [V22 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V23 tmp20 [V23 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V24 tmp21 [V24 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V25 tmp22 [V25 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V26 tmp23 [V26,T14] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V27 tmp24 [V27,T11] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-; V28 tmp25 [V28,T12] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
-;* V29 tmp26 [V29 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V30 tmp27 [V30 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V31 tmp28 [V31 ] ( 3, 3 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V32 tmp29 [V32,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 3, 6 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V34 tmp31 [V34 ] ( 3, 6 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V35 tmp32 [V35 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V36 tmp33 [V36 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V37 tmp34 [V37 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V38 tmp35 [V38 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V39 tmp36 [V39 ] ( 3, 3 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V40 tmp37 [V40,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V41 tmp38 [V41 ] ( 3, 6 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-; V42 tmp39 [V42 ] ( 3, 6 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
-;* V43 tmp40 [V43 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V44 tmp41 [V44 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-;* V45 tmp42 [V45 ] ( 0, 0 ) int -> zero-ref "Inline return value spill temp"
-;* V46 tmp43 [V46 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V47 tmp44 [V47,T15] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
-;* V48 tmp45 [V48 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[int]>
-; V49 tmp46 [V49,T01] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V50 tmp47 [V50,T02] ( 2, 2 ) long -> rax "field V09._00 (fldOffset=0x0)" P-INDEP
-; V51 tmp48 [V51 ] ( 3, 3 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
-; V52 tmp49 [V52 ] ( 3, 5 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
-; V53 tmp50 [V53 ] ( 3, 5 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
-; V54 tmp51 [V54 ] ( 3, 3 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
-; V55 tmp52 [V55 ] ( 3, 5 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V20._00 (fldOffset=0x0)" P-DEP
-; V56 tmp53 [V56 ] ( 3, 5 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
-; V57 tmp54 [V57,T03] ( 2, 2 ) long -> rcx "field V29._00 (fldOffset=0x0)" P-INDEP
-; V58 tmp55 [V58,T04] ( 2, 2 ) long -> rax "field V30._00 (fldOffset=0x0)" P-INDEP
-; V59 tmp56 [V59 ] ( 3, 3 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V60 tmp57 [V60 ] ( 3, 5 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V61 tmp58 [V61 ] ( 3, 5 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V34._00 (fldOffset=0x0)" P-DEP
-; V62 tmp59 [V62 ] ( 3, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V39._00 (fldOffset=0x0)" P-DEP
-; V63 tmp60 [V63 ] ( 3, 5 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V41._00 (fldOffset=0x0)" P-DEP
-; V64 tmp61 [V64 ] ( 3, 5 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V42._00 (fldOffset=0x0)" P-DEP
+; V16 tmp13 [V16 ] ( 3, 3 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V17 tmp14 [V17,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V18 tmp15 [V18 ] ( 3, 6 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V19 tmp16 [V19 ] ( 3, 6 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V21 tmp18 [V21 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V22 tmp19 [V22,T14] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V23 tmp20 [V23,T11] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V24 tmp21 [V24,T12] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[int]>
+;* V25 tmp22 [V25 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V27 tmp24 [V27 ] ( 3, 3 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V28 tmp25 [V28,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 3, 6 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V30 tmp27 [V30 ] ( 3, 6 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V31 tmp28 [V31 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V32 tmp29 [V32 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V33 tmp30 [V33 ] ( 3, 3 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V34 tmp31 [V34,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V35 tmp32 [V35 ] ( 3, 6 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+; V36 tmp33 [V36 ] ( 3, 6 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[int]>
+;* V37 tmp34 [V37 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V38 tmp35 [V38 ] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V39 tmp36 [V39,T15] ( 3, 3 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[int]>
+; V40 tmp37 [V40,T01] ( 2, 2 ) long -> r15 "field V08._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp38 [V41,T02] ( 2, 2 ) long -> rdi "field V09._00 (fldOffset=0x0)" P-INDEP
+; V42 tmp39 [V42 ] ( 3, 3 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 3, 5 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V44 tmp41 [V44 ] ( 3, 5 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V45 tmp42 [V45 ] ( 3, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V46 tmp43 [V46 ] ( 3, 5 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
+; V47 tmp44 [V47 ] ( 3, 5 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V19._00 (fldOffset=0x0)" P-DEP
+; V48 tmp45 [V48,T03] ( 2, 2 ) long -> r15 "field V25._00 (fldOffset=0x0)" P-INDEP
+; V49 tmp46 [V49,T04] ( 2, 2 ) long -> rdi "field V26._00 (fldOffset=0x0)" P-INDEP
+; V50 tmp47 [V50 ] ( 3, 3 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V51 tmp48 [V51 ] ( 3, 5 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V52 tmp49 [V52 ] ( 3, 5 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V30._00 (fldOffset=0x0)" P-DEP
+; V53 tmp50 [V53 ] ( 3, 3 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V54 tmp51 [V54 ] ( 3, 5 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V55 tmp52 [V55 ] ( 3, 5 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V36._00 (fldOffset=0x0)" P-DEP
;
-; Lcl frame size = 192
+; Lcl frame size = 240
G_M12621_IG01:
push rbp
- sub rsp, 192
- lea rbp, [rsp+0xC0]
+ push r15
+ push rbx
+ sub rsp, 240
+ lea rbp, [rsp+0x100]
+ mov rbx, rdi
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x30]
- ;; size=26 bbWeight=1 PerfScore 9.75
+ ;; size=32 bbWeight=1 PerfScore 12.00
G_M12621_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ vmovups ymmword ptr [rbp+0x30], ymm1
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- mov eax, dword ptr [rbp-0x30]
- cdq
- idiv edx:eax, dword ptr [rbp-0x38]
- mov dword ptr [rbp-0x28], eax
- mov eax, dword ptr [rbp-0x2C]
- cdq
- idiv edx:eax, dword ptr [rbp-0x34]
- mov dword ptr [rbp-0x24], eax
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
- mov eax, dword ptr [rbp-0x48]
- cdq
- idiv edx:eax, dword ptr [rbp-0x50]
- mov dword ptr [rbp-0x40], eax
- mov eax, dword ptr [rbp-0x44]
- cdq
- idiv edx:eax, dword ptr [rbp-0x4C]
- mov dword ptr [rbp-0x3C], eax
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
- vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- mov eax, dword ptr [rbp-0x90]
- cdq
- idiv edx:eax, dword ptr [rbp-0x98]
- mov dword ptr [rbp-0x88], eax
- mov eax, dword ptr [rbp-0x8C]
- cdq
- idiv edx:eax, dword ptr [rbp-0x94]
- mov dword ptr [rbp-0x84], eax
- mov rcx, qword ptr [rbp-0x88]
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
- mov eax, dword ptr [rbp-0xA8]
- cdq
- idiv edx:eax, dword ptr [rbp-0xB0]
- mov dword ptr [rbp-0xA0], eax
- mov eax, dword ptr [rbp-0xA4]
- cdq
- ;; size=248 bbWeight=1 PerfScore 226.50
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ mov rdi, qword ptr [rbp-0x30]
+ mov qword ptr [rbp-0x50], rdi
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x58], rdi
+ mov edi, dword ptr [rbp-0x50]
+ mov esi, dword ptr [rbp-0x58]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x48], eax
+ mov edi, dword ptr [rbp-0x4C]
+ mov esi, dword ptr [rbp-0x54]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x44], eax
+ mov r15, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x28]
+ mov qword ptr [rbp-0x68], rdi
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x70], rdi
+ mov edi, dword ptr [rbp-0x68]
+ mov esi, dword ptr [rbp-0x70]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x60], eax
+ mov edi, dword ptr [rbp-0x64]
+ mov esi, dword ptr [rbp-0x6C]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0x5C], eax
+ mov rdi, qword ptr [rbp-0x60]
+ mov qword ptr [rbp-0x80], r15
+ mov qword ptr [rbp-0x78], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x80]
+ vmovaps xmmword ptr [rbp-0x20], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x90], xmm1
+ vmovups ymm1, ymmword ptr [rbp+0x30]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0xA0], xmm1
+ mov rdi, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xB0], rdi
+ mov rdi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xB8], rdi
+ mov edi, dword ptr [rbp-0xB0]
+ mov esi, dword ptr [rbp-0xB8]
+ ;; size=248 bbWeight=1 PerfScore 65.50
G_M12621_IG03:
- idiv edx:eax, dword ptr [rbp-0xAC]
- mov dword ptr [rbp-0x9C], eax
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=50 bbWeight=1 PerfScore 35.25
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0xA8], eax
+ mov edi, dword ptr [rbp-0xAC]
+ mov esi, dword ptr [rbp-0xB4]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0xA4], eax
+ mov r15, qword ptr [rbp-0xA8]
+ mov rdi, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0xC8], rdi
+ mov rdi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xD0], rdi
+ mov edi, dword ptr [rbp-0xC8]
+ mov esi, dword ptr [rbp-0xD0]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0xC0], eax
+ mov edi, dword ptr [rbp-0xC4]
+ mov esi, dword ptr [rbp-0xCC]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ call [rax]System.Runtime.Intrinsics.Scalar`1[int]:Divide(int,int):int
+ mov dword ptr [rbp-0xBC], eax
+ mov rdi, qword ptr [rbp-0xC0]
+ mov qword ptr [rbp-0xE0], r15
+ mov qword ptr [rbp-0xD8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x20]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0xE0]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[int](System.Runtime.Intrinsics.Vector128`1[int],System.Runtime.Intrinsics.Vector128`1[int]):System.Runtime.Intrinsics.Vector256`1[int]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[int](System.Runtime.Intrinsics.Vector128`1[int],System.Runtime.Intrinsics.Vector128`1[int]):System.Runtime.Intrinsics.Vector256`1[int]
+ mov rax, rbx
+ ;; size=206 bbWeight=1 PerfScore 42.75
G_M12621_IG04:
vzeroupper
- add rsp, 192
+ add rsp, 240
+ pop rbx
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 336, prolog size 16, PerfScore 274.25, instruction count 77, allocated bytes for code 336 (MethodHash=e489ceb2) for method System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],System.Runtime.Intrinsics.Vector256`1[int]):System.Runtime.Intrinsics.Vector256`1[int] (FullOpts)
+; Total bytes of code 501, prolog size 19, PerfScore 124.00, instruction count 100, allocated bytes for code 501 (MethodHash=e489ceb2) for method System.Runtime.Intrinsics.Vector256`1[int]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[int],System.Runtime.Intrinsics.Vector256`1[int]):System.Runtime.Intrinsics.Vector256`1[int] (FullOpts)157 (25.00 % of base) - System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],System.Runtime.Intrinsics.Vector256`1[short]):System.Runtime.Intrinsics.Vector256`1[short]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],System.Runtime.Intrinsics.Vector256`1[short]):System.Runtime.Intrinsics.Vector256`1[short] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 20 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 20 single block inlinees; 8 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T04] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T20] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[short]>
-; V02 arg1 [V02,T21] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[short]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T17] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rbx single-def
+; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[short]>
+; V02 arg1 [V02,T17] ( 2, 2 ) simd32 -> [rbp+0x30] single-def <System.Runtime.Intrinsics.Vector256`1[short]>
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> [rbp-0x20] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T13] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V07 tmp4 [V07,T14] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V06 tmp3 [V06,T09] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V07 tmp4 [V07,T10] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V09 tmp6 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V10 tmp7 [V10 ] ( 5, 5 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V11 tmp8 [V11,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V12 tmp9 [V12 ] ( 5, 10 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V13 tmp10 [V13 ] ( 5, 10 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V10 tmp7 [V10 ] ( 5, 5 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V11 tmp8 [V11,T05] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V12 tmp9 [V12 ] ( 5, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V13 tmp10 [V13 ] ( 5, 10 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V14 tmp11 [V14 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V15 tmp12 [V15 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V16 tmp13 [V16,T00] ( 8, 8 ) short -> rcx "Inline return value spill temp"
-;* V17 tmp14 [V17 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V18 tmp15 [V18 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V19 tmp16 [V19 ] ( 5, 5 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V20 tmp17 [V20,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V21 tmp18 [V21 ] ( 5, 10 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V22 tmp19 [V22 ] ( 5, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V24 tmp21 [V24 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V25 tmp22 [V25,T01] ( 8, 8 ) short -> rsi "Inline return value spill temp"
-;* V26 tmp23 [V26 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V27 tmp24 [V27 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V28 tmp25 [V28,T18] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V29 tmp26 [V29,T15] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-; V30 tmp27 [V30,T16] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
-;* V31 tmp28 [V31 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V32 tmp29 [V32 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V33 tmp30 [V33 ] ( 5, 5 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V34 tmp31 [V34,T11] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V35 tmp32 [V35 ] ( 5, 10 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V36 tmp33 [V36 ] ( 5, 10 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V16 tmp13 [V16 ] ( 5, 5 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V17 tmp14 [V17,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V18 tmp15 [V18 ] ( 5, 10 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V19 tmp16 [V19 ] ( 5, 10 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V21 tmp18 [V21 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
+; V22 tmp19 [V22,T14] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V23 tmp20 [V23,T11] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V24 tmp21 [V24,T12] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[short]>
+;* V25 tmp22 [V25 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V27 tmp24 [V27 ] ( 5, 5 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V28 tmp25 [V28,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 5, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V30 tmp27 [V30 ] ( 5, 10 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V31 tmp28 [V31 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V32 tmp29 [V32 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
+; V33 tmp30 [V33 ] ( 5, 5 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
+;* V34 tmp31 [V34,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V35 tmp32 [V35 ] ( 5, 10 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
+; V36 tmp33 [V36 ] ( 5, 10 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
;* V37 tmp34 [V37 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V38 tmp35 [V38 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V39 tmp36 [V39,T02] ( 8, 8 ) short -> rcx "Inline return value spill temp"
-;* V40 tmp37 [V40 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V41 tmp38 [V41 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V42 tmp39 [V42 ] ( 5, 5 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V43 tmp40 [V43,T12] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V44 tmp41 [V44 ] ( 5, 10 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-; V45 tmp42 [V45 ] ( 5, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[short]>
-;* V46 tmp43 [V46 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V47 tmp44 [V47 ] ( 0, 0 ) short -> zero-ref "Inline stloc first use temp"
-; V48 tmp45 [V48,T03] ( 8, 8 ) short -> rsi "Inline return value spill temp"
-;* V49 tmp46 [V49 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-;* V50 tmp47 [V50 ] ( 0, 0 ) short -> zero-ref "Inlining Arg"
-; V51 tmp48 [V51,T19] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
-;* V52 tmp49 [V52 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[short]>
-; V53 tmp50 [V53,T05] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V54 tmp51 [V54,T06] ( 2, 2 ) long -> rax "field V09._00 (fldOffset=0x0)" P-INDEP
-; V55 tmp52 [V55 ] ( 5, 5 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
-; V56 tmp53 [V56 ] ( 5, 9 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
-; V57 tmp54 [V57 ] ( 5, 9 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
-; V58 tmp55 [V58 ] ( 5, 5 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V19._00 (fldOffset=0x0)" P-DEP
-; V59 tmp56 [V59 ] ( 5, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
-; V60 tmp57 [V60 ] ( 5, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V22._00 (fldOffset=0x0)" P-DEP
-; V61 tmp58 [V61,T07] ( 2, 2 ) long -> rcx "field V31._00 (fldOffset=0x0)" P-INDEP
-; V62 tmp59 [V62,T08] ( 2, 2 ) long -> rax "field V32._00 (fldOffset=0x0)" P-INDEP
-; V63 tmp60 [V63 ] ( 5, 5 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V64 tmp61 [V64 ] ( 5, 9 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
-; V65 tmp62 [V65 ] ( 5, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V36._00 (fldOffset=0x0)" P-DEP
-; V66 tmp63 [V66 ] ( 5, 5 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V42._00 (fldOffset=0x0)" P-DEP
-; V67 tmp64 [V67 ] ( 5, 9 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V44._00 (fldOffset=0x0)" P-DEP
-; V68 tmp65 [V68 ] ( 5, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
+; V39 tmp36 [V39,T15] ( 3, 3 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[short]>
+; V40 tmp37 [V40,T01] ( 2, 2 ) long -> r15 "field V08._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp38 [V41,T02] ( 2, 2 ) long -> rdi "field V09._00 (fldOffset=0x0)" P-INDEP
+; V42 tmp39 [V42 ] ( 5, 5 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 5, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V44 tmp41 [V44 ] ( 5, 9 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V45 tmp42 [V45 ] ( 5, 5 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V46 tmp43 [V46 ] ( 5, 9 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
+; V47 tmp44 [V47 ] ( 5, 9 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V19._00 (fldOffset=0x0)" P-DEP
+; V48 tmp45 [V48,T03] ( 2, 2 ) long -> r15 "field V25._00 (fldOffset=0x0)" P-INDEP
+; V49 tmp46 [V49,T04] ( 2, 2 ) long -> rdi "field V26._00 (fldOffset=0x0)" P-INDEP
+; V50 tmp47 [V50 ] ( 5, 5 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V51 tmp48 [V51 ] ( 5, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V52 tmp49 [V52 ] ( 5, 9 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V30._00 (fldOffset=0x0)" P-DEP
+; V53 tmp50 [V53 ] ( 5, 5 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V54 tmp51 [V54 ] ( 5, 9 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V55 tmp52 [V55 ] ( 5, 9 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V36._00 (fldOffset=0x0)" P-DEP
;
-; Lcl frame size = 192
+; Lcl frame size = 240
G_M52493_IG01:
push rbp
- sub rsp, 192
- lea rbp, [rsp+0xC0]
+ push r15
+ push rbx
+ sub rsp, 240
+ lea rbp, [rsp+0x100]
+ mov rbx, rdi
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x30]
- ;; size=26 bbWeight=1 PerfScore 9.75
+ ;; size=32 bbWeight=1 PerfScore 12.00
G_M52493_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ vmovups ymmword ptr [rbp+0x30], ymm1
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- movsx rax, word ptr [rbp-0x30]
- movsx rcx, word ptr [rbp-0x38]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x28], cx
- movsx rax, word ptr [rbp-0x2E]
- movsx rcx, word ptr [rbp-0x36]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x26], cx
- movsx rax, word ptr [rbp-0x2C]
- movsx rcx, word ptr [rbp-0x34]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x24], cx
- movsx rax, word ptr [rbp-0x2A]
- movsx rcx, word ptr [rbp-0x32]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x22], cx
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
- movsx rax, word ptr [rbp-0x48]
- movsx rsi, word ptr [rbp-0x50]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x40], si
- movsx rax, word ptr [rbp-0x46]
- movsx rsi, word ptr [rbp-0x4E]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x3E], si
- movsx rax, word ptr [rbp-0x44]
- movsx rsi, word ptr [rbp-0x4C]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x3C], si
- movsx rax, word ptr [rbp-0x42]
- movsx rsi, word ptr [rbp-0x4A]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x3A], si
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
- ;; size=239 bbWeight=1 PerfScore 279.50
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ mov rdi, qword ptr [rbp-0x30]
+ mov qword ptr [rbp-0x50], rdi
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x58], rdi
+ movsx rdi, word ptr [rbp-0x50]
+ movsx rsi, word ptr [rbp-0x58]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x48], ax
+ movsx rdi, word ptr [rbp-0x4E]
+ movsx rsi, word ptr [rbp-0x56]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x46], ax
+ movsx rdi, word ptr [rbp-0x4C]
+ movsx rsi, word ptr [rbp-0x54]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x44], ax
+ movsx rdi, word ptr [rbp-0x4A]
+ movsx rsi, word ptr [rbp-0x52]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x42], ax
+ mov r15, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x28]
+ mov qword ptr [rbp-0x68], rdi
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x70], rdi
+ movsx rdi, word ptr [rbp-0x68]
+ movsx rsi, word ptr [rbp-0x70]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x60], ax
+ movsx rdi, word ptr [rbp-0x66]
+ movsx rsi, word ptr [rbp-0x6E]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x5E], ax
+ movsx rdi, word ptr [rbp-0x64]
+ movsx rsi, word ptr [rbp-0x6C]
+ ;; size=230 bbWeight=1 PerfScore 81.00
G_M52493_IG03:
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
- vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- movsx rax, word ptr [rbp-0x90]
- movsx rcx, word ptr [rbp-0x98]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x88], cx
- movsx rax, word ptr [rbp-0x8E]
- movsx rcx, word ptr [rbp-0x96]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x86], cx
- movsx rax, word ptr [rbp-0x8C]
- movsx rcx, word ptr [rbp-0x94]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x84], cx
- movsx rax, word ptr [rbp-0x8A]
- movsx rcx, word ptr [rbp-0x92]
- cdq
- idiv edx:eax, ecx
- movsx rcx, ax
- mov word ptr [rbp-0x82], cx
- mov rcx, qword ptr [rbp-0x88]
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
- movsx rax, word ptr [rbp-0xA8]
- movsx rsi, word ptr [rbp-0xB0]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0xA0], si
- movsx rax, word ptr [rbp-0xA6]
- movsx rsi, word ptr [rbp-0xAE]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x9E], si
- movsx rax, word ptr [rbp-0xA4]
- movsx rsi, word ptr [rbp-0xAC]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x9C], si
- movsx rax, word ptr [rbp-0xA2]
- movsx rsi, word ptr [rbp-0xAA]
- cdq
- idiv edx:eax, esi
- movsx rsi, ax
- mov word ptr [rbp-0x9A], si
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
- ;; size=344 bbWeight=1 PerfScore 284.00
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x5C], ax
+ movsx rdi, word ptr [rbp-0x62]
+ movsx rsi, word ptr [rbp-0x6A]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0x5A], ax
+ mov rdi, qword ptr [rbp-0x60]
+ mov qword ptr [rbp-0x80], r15
+ mov qword ptr [rbp-0x78], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x80]
+ vmovaps xmmword ptr [rbp-0x20], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x90], xmm1
+ vmovups ymm1, ymmword ptr [rbp+0x30]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0xA0], xmm1
+ mov rdi, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xB0], rdi
+ mov rdi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xB8], rdi
+ movsx rdi, word ptr [rbp-0xB0]
+ movsx rsi, word ptr [rbp-0xB8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xA8], ax
+ movsx rdi, word ptr [rbp-0xAE]
+ movsx rsi, word ptr [rbp-0xB6]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xA6], ax
+ movsx rdi, word ptr [rbp-0xAC]
+ movsx rsi, word ptr [rbp-0xB4]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xA4], ax
+ movsx rdi, word ptr [rbp-0xAA]
+ movsx rsi, word ptr [rbp-0xB2]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xA2], ax
+ mov r15, qword ptr [rbp-0xA8]
+ mov rdi, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0xC8], rdi
+ mov rdi, qword ptr [rbp-0x98]
+ ;; size=298 bbWeight=1 PerfScore 84.50
G_M52493_IG04:
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=7 bbWeight=1 PerfScore 2.25
+ mov qword ptr [rbp-0xD0], rdi
+ movsx rdi, word ptr [rbp-0xC8]
+ movsx rsi, word ptr [rbp-0xD0]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xC0], ax
+ movsx rdi, word ptr [rbp-0xC6]
+ movsx rsi, word ptr [rbp-0xCE]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xBE], ax
+ movsx rdi, word ptr [rbp-0xC4]
+ movsx rsi, word ptr [rbp-0xCC]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xBC], ax
+ movsx rdi, word ptr [rbp-0xC2]
+ movsx rsi, word ptr [rbp-0xCA]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ call [rax]System.Runtime.Intrinsics.Scalar`1[short]:Divide(short,short):short
+ mov word ptr [rbp-0xBA], ax
+ mov rdi, qword ptr [rbp-0xC0]
+ mov qword ptr [rbp-0xE0], r15
+ mov qword ptr [rbp-0xD8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x20]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0xE0]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[short](System.Runtime.Intrinsics.Vector128`1[short],System.Runtime.Intrinsics.Vector128`1[short]):System.Runtime.Intrinsics.Vector256`1[short]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[short](System.Runtime.Intrinsics.Vector128`1[short],System.Runtime.Intrinsics.Vector128`1[short]):System.Runtime.Intrinsics.Vector256`1[short]
+ mov rax, rbx
+ ;; size=210 bbWeight=1 PerfScore 56.75
G_M52493_IG05:
vzeroupper
- add rsp, 192
+ add rsp, 240
+ pop rbx
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 628, prolog size 16, PerfScore 578.25, instruction count 141, allocated bytes for code 628 (MethodHash=eb5b32f2) for method System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],System.Runtime.Intrinsics.Vector256`1[short]):System.Runtime.Intrinsics.Vector256`1[short] (FullOpts)
+; Total bytes of code 785, prolog size 19, PerfScore 238.00, instruction count 140, allocated bytes for code 785 (MethodHash=eb5b32f2) for method System.Runtime.Intrinsics.Vector256`1[short]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[short],System.Runtime.Intrinsics.Vector256`1[short]):System.Runtime.Intrinsics.Vector256`1[short] (FullOpts)155 (44.16 % of base) - System.Runtime.Intrinsics.Vector128`1[ubyte]:System.Runtime.Intrinsics.ISimdVector,T>.Dot(System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):ubyte
; Assembly listing for method System.Runtime.Intrinsics.Vector128`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector128<T>,T>.Dot(System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):ubyte (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
-; partially interruptible
+; fully interruptible
; No PGO data
-; 0 inlinees with PGO data; 9 single block inlinees; 11 inlinees without PGO data
+; 0 inlinees with PGO data; 11 single block inlinees; 5 inlinees without PGO data
; Final local variable assignments
;
-; V00 arg0 [V00,T11] ( 2, 2 ) simd16 -> [rbp+0x10] do-not-enreg[SF] single-def <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V01 arg1 [V01,T12] ( 2, 2 ) simd16 -> [rbp+0x20] do-not-enreg[SF] single-def <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V00 arg0 [V00,T13] ( 2, 2 ) simd16 -> [rbp+0x10] do-not-enreg[SF] single-def <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V01 arg1 [V01,T14] ( 2, 2 ) simd16 -> [rbp+0x20] do-not-enreg[SF] single-def <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;# V02 OutArgs [V02 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V03 tmp1 [V03,T06] ( 2, 4 ) int -> rax "impAppendStmt"
-; V04 tmp2 [V04,T02] ( 8, 8 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
-;* V05 tmp3 [V05,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V06 tmp4 [V06 ] ( 9, 18 ) struct ( 8) [rbp-0x08] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V07 tmp5 [V07 ] ( 9, 18 ) struct ( 8) [rbp-0x10] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V08 tmp6 [V08 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V09 tmp7 [V09,T00] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V10 tmp8 [V10 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V11 tmp9 [V11 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V12 tmp10 [V12 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V13 tmp11 [V13 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V14 tmp12 [V14 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V15 tmp13 [V15 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V16 tmp14 [V16,T04] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V17 tmp15 [V17 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V18 tmp16 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V19 tmp17 [V19 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V20 tmp18 [V20 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V21 tmp19 [V21 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V22 tmp20 [V22 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V23 tmp21 [V23,T03] ( 8, 8 ) ubyte -> rcx ld-addr-op "Inline ldloca(s) first use temp"
-;* V24 tmp22 [V24,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V25 tmp23 [V25 ] ( 9, 18 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V26 tmp24 [V26 ] ( 9, 18 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V03 tmp1 [V03,T06] ( 2, 4 ) int -> rbx "impAppendStmt"
+;* V04 tmp2 [V04 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V05 tmp3 [V05 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V06 tmp4 [V06 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V07 tmp5 [V07 ] ( 2, 5 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V08 tmp6 [V08,T02] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
+; V09 tmp7 [V09 ] ( 2, 10 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V10 tmp8 [V10 ] ( 2, 10 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V11 tmp9 [V11 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V12 tmp10 [V12 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+;* V13 tmp11 [V13 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
+;* V14 tmp12 [V14 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+;* V15 tmp13 [V15 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
+; V16 tmp14 [V16,T00] ( 16, 16 ) ubyte -> rbx ld-addr-op "Inline ldloca(s) first use temp"
+;* V17 tmp15 [V17,T11] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V18 tmp16 [V18 ] ( 9, 18 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V19 tmp17 [V19 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V20 tmp18 [V20 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V21 tmp19 [V21 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V22 tmp20 [V22 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V23 tmp21 [V23 ] ( 2, 5 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V24 tmp22 [V24,T03] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V25 tmp23 [V25 ] ( 2, 10 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V26 tmp24 [V26 ] ( 2, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V27 tmp25 [V27 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-; V28 tmp26 [V28,T01] ( 8, 16 ) int -> registers "impAppendStmt"
-;* V29 tmp27 [V29 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V30 tmp28 [V30 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V31 tmp29 [V31 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V32 tmp30 [V32 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V33 tmp31 [V33 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-;* V34 tmp32 [V34 ] ( 0, 0 ) int -> zero-ref "Inlining Arg"
-; V35 tmp33 [V35,T05] ( 8, 8 ) ubyte -> registers "Inline return value spill temp"
-;* V36 tmp34 [V36 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V37 tmp35 [V37 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V38 tmp36 [V38 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V39 tmp37 [V39 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V40 tmp38 [V40 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V41 tmp39 [V41 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V42 tmp40 [V42 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V43 tmp41 [V43 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V44 tmp42 [V44 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V45 tmp43 [V45 ] ( 9, 17 ) long -> [rbp-0x08] do-not-enreg[X] addr-exposed "field V06._00 (fldOffset=0x0)" P-DEP
-; V46 tmp44 [V46 ] ( 9, 17 ) long -> [rbp-0x10] do-not-enreg[X] addr-exposed "field V07._00 (fldOffset=0x0)" P-DEP
-; V47 tmp45 [V47 ] ( 9, 17 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp46 [V48 ] ( 9, 17 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
-; V49 cse0 [V49,T07] ( 2, 2 ) int -> rcx "CSE #01: moderate"
-; V50 cse1 [V50,T08] ( 2, 2 ) int -> rdx "CSE #02: moderate"
+;* V28 tmp26 [V28 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V29 tmp27 [V29,T01] ( 16, 16 ) ubyte -> rax ld-addr-op "Inline ldloca(s) first use temp"
+;* V30 tmp28 [V30,T12] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V31 tmp29 [V31 ] ( 9, 18 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V32 tmp30 [V32 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+; V33 tmp31 [V33,T07] ( 2, 2 ) long -> rsi "field V04._00 (fldOffset=0x0)" P-INDEP
+; V34 tmp32 [V34,T08] ( 2, 2 ) long -> rdi "field V05._00 (fldOffset=0x0)" P-INDEP
+;* V35 tmp33 [V35 ] ( 0, 0 ) long -> zero-ref "field V06._00 (fldOffset=0x0)" P-INDEP
+; V36 tmp34 [V36 ] ( 2, 5 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V07._00 (fldOffset=0x0)" P-DEP
+; V37 tmp35 [V37 ] ( 2, 9 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V38 tmp36 [V38 ] ( 2, 9 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V39 tmp37 [V39 ] ( 9, 17 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
+; V40 tmp38 [V40,T09] ( 2, 2 ) long -> rdi "field V20._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp39 [V41,T10] ( 2, 2 ) long -> rsi "field V21._00 (fldOffset=0x0)" P-INDEP
+;* V42 tmp40 [V42 ] ( 0, 0 ) long -> zero-ref "field V22._00 (fldOffset=0x0)" P-INDEP
+; V43 tmp41 [V43 ] ( 2, 5 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V44 tmp42 [V44 ] ( 2, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V45 tmp43 [V45 ] ( 2, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
+; V46 tmp44 [V46 ] ( 9, 17 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
+; V47 cse0 [V47,T04] ( 4, 16 ) long -> rax "CSE #01: aggressive"
+; V48 cse1 [V48,T05] ( 4, 16 ) long -> r14 "CSE #02: aggressive"
;
-; Lcl frame size = 32
+; Lcl frame size = 72
G_M52421_IG01:
push rbp
- sub rsp, 32
- lea rbp, [rsp+0x20]
- ;; size=10 bbWeight=1 PerfScore 1.75
+ push r15
+ push r14
+ push rbx
+ sub rsp, 72
+ lea rbp, [rsp+0x60]
+ ;; size=15 bbWeight=1 PerfScore 4.75
G_M52421_IG02:
- mov rax, qword ptr [rbp+0x10]
- mov qword ptr [rbp-0x08], rax
- mov rax, qword ptr [rbp+0x20]
- mov qword ptr [rbp-0x10], rax
- movzx rax, byte ptr [rbp-0x08]
- movzx rcx, byte ptr [rbp-0x10]
- imul eax, ecx
- movzx rax, al
- movzx rcx, byte ptr [rbp-0x07]
- movzx rdx, byte ptr [rbp-0x0F]
- imul ecx, edx
- movzx rcx, cl
- add eax, ecx
- movzx rcx, al
- movzx rax, cl
- movzx rcx, byte ptr [rbp-0x06]
- movzx rdx, byte ptr [rbp-0x0E]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x05]
- movzx rdi, byte ptr [rbp-0x0D]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add eax, ecx
- movzx rax, al
- movzx rdx, byte ptr [rbp-0x04]
- movzx rcx, byte ptr [rbp-0x0C]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x03]
- movzx rdi, byte ptr [rbp-0x0B]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add ecx, eax
- movzx rax, cl
- movzx rdx, byte ptr [rbp-0x02]
- movzx rcx, byte ptr [rbp-0x0A]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x01]
- movzx rdi, byte ptr [rbp-0x09]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- add ecx, eax
- movzx rax, cl
- mov rcx, qword ptr [rbp+0x18]
- mov qword ptr [rbp-0x18], rcx
- mov rcx, qword ptr [rbp+0x28]
- mov qword ptr [rbp-0x20], rcx
- movzx rcx, byte ptr [rbp-0x18]
- movzx rdx, byte ptr [rbp-0x20]
- imul ecx, edx
- movzx rcx, cl
- movzx rdx, byte ptr [rbp-0x17]
- movzx rdi, byte ptr [rbp-0x1F]
- imul edx, edi
- movzx rdx, dl
- add ecx, edx
- movzx rdx, cl
- movzx rcx, dl
- movzx rdx, byte ptr [rbp-0x16]
- movzx rdi, byte ptr [rbp-0x1E]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x15]
- movzx rsi, byte ptr [rbp-0x1D]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- ;; size=247 bbWeight=1 PerfScore 63.00
+ mov rsi, qword ptr [rbp+0x10]
+ mov rdi, qword ptr [rbp+0x20]
+ mov qword ptr [rbp-0x28], rsi
+ mov qword ptr [rbp-0x30], rdi
+ xor esi, esi
+ align [0 bytes for IG03]
+ ;; size=18 bbWeight=1 PerfScore 4.25
G_M52421_IG03:
- add ecx, edx
- movzx rcx, cl
- movzx rdi, byte ptr [rbp-0x14]
- movzx rdx, byte ptr [rbp-0x1C]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x13]
- movzx rsi, byte ptr [rbp-0x1B]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- movzx rdi, byte ptr [rbp-0x12]
- movzx rdx, byte ptr [rbp-0x1A]
- imul edx, edi
- movzx rdx, dl
- movzx rdi, byte ptr [rbp-0x11]
- movzx rsi, byte ptr [rbp-0x19]
- imul edi, esi
- movzx rdi, dil
- add edx, edi
- add edx, ecx
- movzx rcx, dl
- add eax, ecx
- movzx rax, al
- ;; size=88 bbWeight=1 PerfScore 19.50
+ lea rdi, [rbp-0x28]
+ movsxd rax, esi
+ movzx rdi, byte ptr [rdi+rax]
+ lea rcx, [rbp-0x30]
+ movzx rcx, byte ptr [rcx+rax]
+ imul edi, ecx
+ lea rcx, [rbp-0x20]
+ mov byte ptr [rcx+rax], dil
+ inc esi
+ cmp esi, 8
+ jl SHORT G_M52421_IG03
+ ;; size=38 bbWeight=4 PerfScore 41.00
G_M52421_IG04:
- add rsp, 32
+ mov rsi, qword ptr [rbp-0x20]
+ mov qword ptr [rbp-0x38], rsi
+ movzx rsi, byte ptr [rbp-0x38]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x37]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x36]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x35]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x34]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x33]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x32]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ movzx rsi, byte ptr [rbp-0x31]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ mov ebx, eax
+ mov rdi, qword ptr [rbp+0x18]
+ mov rsi, qword ptr [rbp+0x28]
+ mov qword ptr [rbp-0x48], rdi
+ mov qword ptr [rbp-0x50], rsi
+ xor r15d, r15d
+ ;; size=195 bbWeight=1 PerfScore 44.25
+G_M52421_IG05:
+ lea rdi, [rbp-0x48]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0x50]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Multiply(ubyte,ubyte):ubyte
+ lea rsi, [rbp-0x40]
+ mov byte ptr [rsi+r14], al
+ inc r15d
+ cmp r15d, 8
+ jl SHORT G_M52421_IG05
+ ;; size=50 bbWeight=4 PerfScore 46.00
+G_M52421_IG06:
+ mov rsi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x58], rsi
+ movzx rsi, byte ptr [rbp-0x58]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x57]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x56]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x55]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x54]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x53]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x52]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rsi, byte ptr [rbp-0x51]
+ mov edi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ movzx rdi, bl
+ mov esi, eax
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Add(ubyte,ubyte):ubyte
+ nop
+ ;; size=179 bbWeight=1 PerfScore 42.00
+G_M52421_IG07:
+ add rsp, 72
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=6 bbWeight=1 PerfScore 1.75
+ ;; size=11 bbWeight=1 PerfScore 3.25
-; Total bytes of code 351, prolog size 10, PerfScore 86.00, instruction count 104, allocated bytes for code 351 (MethodHash=ae36333a) for method System.Runtime.Intrinsics.Vector128`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector128<T>,T>.Dot(System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):ubyte (FullOpts)
+; Total bytes of code 506, prolog size 15, PerfScore 185.50, instruction count 127, allocated bytes for code 506 (MethodHash=ae36333a) for method System.Runtime.Intrinsics.Vector128`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector128<T>,T>.Dot(System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):ubyte (FullOpts)133 (32.36 % of base) - System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector,T>.Divide(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
-; fully interruptible
+; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 20 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 20 single block inlinees; 8 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T08] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T20] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V02 arg1 [V02,T21] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T17] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V00 RetBuf [V00,T08] ( 4, 4 ) byref -> rbx single-def
+; V01 arg0 [V01,T20] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V02 arg1 [V02,T21] ( 2, 2 ) simd32 -> [rbp+0x30] single-def <System.Runtime.Intrinsics.Vector256`1[ubyte]>
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T17] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T13] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V07 tmp4 [V07,T14] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V06 tmp3 [V06,T13] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V07 tmp4 [V07,T14] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V09 tmp6 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V10 tmp7 [V10 ] ( 2, 5 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V11 tmp8 [V11,T00] ( 5, 17 ) int -> rcx "Inline stloc first use temp"
-; V12 tmp9 [V12 ] ( 2, 10 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V13 tmp10 [V13 ] ( 2, 10 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V10 tmp7 [V10 ] ( 2, 5 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V11 tmp8 [V11,T00] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V12 tmp9 [V12 ] ( 2, 10 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V13 tmp10 [V13 ] ( 2, 10 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V14 tmp11 [V14 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V15 tmp12 [V15 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V16 tmp13 [V16 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V17 tmp14 [V17 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V18 tmp15 [V18 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V19 tmp16 [V19 ] ( 2, 5 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V20 tmp17 [V20,T01] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
-; V21 tmp18 [V21 ] ( 2, 10 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V22 tmp19 [V22 ] ( 2, 10 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V24 tmp21 [V24 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V25 tmp22 [V25 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V26 tmp23 [V26 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V27 tmp24 [V27 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V28 tmp25 [V28,T18] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V29 tmp26 [V29,T15] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-; V30 tmp27 [V30,T16] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V31 tmp28 [V31 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V32 tmp29 [V32 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V33 tmp30 [V33 ] ( 2, 5 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V34 tmp31 [V34,T02] ( 5, 17 ) int -> rcx "Inline stloc first use temp"
-; V35 tmp32 [V35 ] ( 2, 10 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V36 tmp33 [V36 ] ( 2, 10 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V16 tmp13 [V16 ] ( 2, 5 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V17 tmp14 [V17,T01] ( 5, 17 ) int -> r14 "Inline stloc first use temp"
+; V18 tmp15 [V18 ] ( 2, 10 ) struct ( 8) [rbp-0x78] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V19 tmp16 [V19 ] ( 2, 10 ) struct ( 8) [rbp-0x80] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V21 tmp18 [V21 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V22 tmp19 [V22,T18] ( 3, 3 ) simd16 -> [rbp-0x90] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V23 tmp20 [V23,T15] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V24 tmp21 [V24,T16] ( 3, 6 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+;* V25 tmp22 [V25 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V27 tmp24 [V27 ] ( 2, 5 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V28 tmp25 [V28,T02] ( 5, 17 ) int -> r15 "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 2, 10 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V30 tmp27 [V30 ] ( 2, 10 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+;* V31 tmp28 [V31 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
+;* V32 tmp29 [V32 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
+; V33 tmp30 [V33 ] ( 2, 5 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V34 tmp31 [V34,T03] ( 5, 17 ) int -> r14 "Inline stloc first use temp"
+; V35 tmp32 [V35 ] ( 2, 10 ) struct ( 8) [rbp-0xD8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
+; V36 tmp33 [V36 ] ( 2, 10 ) struct ( 8) [rbp-0xE0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
;* V37 tmp34 [V37 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
;* V38 tmp35 [V38 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V39 tmp36 [V39 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V40 tmp37 [V40 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V41 tmp38 [V41 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V42 tmp39 [V42 ] ( 2, 5 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V43 tmp40 [V43,T03] ( 5, 17 ) int -> rsi "Inline stloc first use temp"
-; V44 tmp41 [V44 ] ( 2, 10 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-; V45 tmp42 [V45 ] ( 2, 10 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[ubyte]>
-;* V46 tmp43 [V46 ] ( 0, 0 ) int -> zero-ref "impAppendStmt"
-;* V47 tmp44 [V47 ] ( 0, 0 ) ubyte -> zero-ref "Inline stloc first use temp"
-;* V48 tmp45 [V48 ] ( 0, 0 ) ubyte -> zero-ref "Inline return value spill temp"
-;* V49 tmp46 [V49 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-;* V50 tmp47 [V50 ] ( 0, 0 ) ubyte -> zero-ref "Inlining Arg"
-; V51 tmp48 [V51,T19] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
-;* V52 tmp49 [V52 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[ubyte]>
-; V53 tmp50 [V53,T09] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V54 tmp51 [V54,T10] ( 2, 2 ) long -> rax "field V09._00 (fldOffset=0x0)" P-INDEP
-; V55 tmp52 [V55 ] ( 2, 5 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
-; V56 tmp53 [V56 ] ( 2, 9 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
-; V57 tmp54 [V57 ] ( 2, 9 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
-; V58 tmp55 [V58 ] ( 2, 5 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V19._00 (fldOffset=0x0)" P-DEP
-; V59 tmp56 [V59 ] ( 2, 9 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
-; V60 tmp57 [V60 ] ( 2, 9 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V22._00 (fldOffset=0x0)" P-DEP
-; V61 tmp58 [V61,T11] ( 2, 2 ) long -> rcx "field V31._00 (fldOffset=0x0)" P-INDEP
-; V62 tmp59 [V62,T12] ( 2, 2 ) long -> rax "field V32._00 (fldOffset=0x0)" P-INDEP
-; V63 tmp60 [V63 ] ( 2, 5 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V64 tmp61 [V64 ] ( 2, 9 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
-; V65 tmp62 [V65 ] ( 2, 9 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V36._00 (fldOffset=0x0)" P-DEP
-; V66 tmp63 [V66 ] ( 2, 5 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V42._00 (fldOffset=0x0)" P-DEP
-; V67 tmp64 [V67 ] ( 2, 9 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V44._00 (fldOffset=0x0)" P-DEP
-; V68 tmp65 [V68 ] ( 2, 9 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V45._00 (fldOffset=0x0)" P-DEP
-; V69 cse0 [V69,T04] ( 4, 16 ) long -> rsi "CSE #01: aggressive"
-; V70 cse1 [V70,T05] ( 4, 16 ) long -> r8 "CSE #02: aggressive"
-; V71 cse2 [V71,T06] ( 4, 16 ) long -> rsi "CSE #03: aggressive"
-; V72 cse3 [V72,T07] ( 4, 16 ) long -> r8 "CSE #04: aggressive"
+; V39 tmp36 [V39,T19] ( 3, 3 ) simd16 -> [rbp-0xF0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[ubyte]>
+; V40 tmp37 [V40,T09] ( 2, 2 ) long -> r15 "field V08._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp38 [V41,T10] ( 2, 2 ) long -> rdi "field V09._00 (fldOffset=0x0)" P-INDEP
+; V42 tmp39 [V42 ] ( 2, 5 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 2, 9 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V44 tmp41 [V44 ] ( 2, 9 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V45 tmp42 [V45 ] ( 2, 5 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V46 tmp43 [V46 ] ( 2, 9 ) long -> [rbp-0x78] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
+; V47 tmp44 [V47 ] ( 2, 9 ) long -> [rbp-0x80] do-not-enreg[X] addr-exposed "field V19._00 (fldOffset=0x0)" P-DEP
+; V48 tmp45 [V48,T11] ( 2, 2 ) long -> r15 "field V25._00 (fldOffset=0x0)" P-INDEP
+; V49 tmp46 [V49,T12] ( 2, 2 ) long -> rdi "field V26._00 (fldOffset=0x0)" P-INDEP
+; V50 tmp47 [V50 ] ( 2, 5 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V51 tmp48 [V51 ] ( 2, 9 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V52 tmp49 [V52 ] ( 2, 9 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V30._00 (fldOffset=0x0)" P-DEP
+; V53 tmp50 [V53 ] ( 2, 5 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V54 tmp51 [V54 ] ( 2, 9 ) long -> [rbp-0xD8] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V55 tmp52 [V55 ] ( 2, 9 ) long -> [rbp-0xE0] do-not-enreg[X] addr-exposed "field V36._00 (fldOffset=0x0)" P-DEP
+; V56 cse0 [V56,T04] ( 4, 16 ) long -> r14 "CSE #01: aggressive"
+; V57 cse1 [V57,T05] ( 4, 16 ) long -> r13 "CSE #02: aggressive"
+; V58 cse2 [V58,T06] ( 4, 16 ) long -> r14 "CSE #03: aggressive"
+; V59 cse3 [V59,T07] ( 4, 16 ) long -> r13 "CSE #04: aggressive"
;
-; Lcl frame size = 192
+; Lcl frame size = 240
G_M49741_IG01:
push rbp
- sub rsp, 192
- lea rbp, [rsp+0xC0]
+ push r15
+ push r14
+ push r13
+ push rbx
+ sub rsp, 240
+ lea rbp, [rsp+0x110]
+ mov rbx, rdi
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x30]
- ;; size=26 bbWeight=1 PerfScore 9.75
+ ;; size=36 bbWeight=1 PerfScore 14.00
G_M49741_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ vmovups ymmword ptr [rbp+0x30], ymm1
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- xor ecx, ecx
- align [2 bytes for IG03]
- ;; size=38 bbWeight=1 PerfScore 7.00
+ vmovaps xmmword ptr [rbp-0x50], xmm2
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x60], rdi
+ mov rdi, qword ptr [rbp-0x50]
+ mov qword ptr [rbp-0x68], rdi
+ xor r15d, r15d
+ ;; size=47 bbWeight=1 PerfScore 8.75
G_M49741_IG03:
- lea rax, [rbp-0x30]
- movsxd rsi, ecx
- movzx rax, byte ptr [rax+rsi]
- lea rdx, [rbp-0x38]
- movzx r8, byte ptr [rdx+rsi]
- xor edx, edx
- div edx:eax, r8d
- lea rdx, [rbp-0x28]
- mov byte ptr [rdx+rsi], al
- inc ecx
- cmp ecx, 8
+ lea rdi, [rbp-0x60]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0x68]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0x58]
+ mov byte ptr [rdi+r14], al
+ inc r15d
+ cmp r15d, 8
jl SHORT G_M49741_IG03
- ;; size=39 bbWeight=4 PerfScore 134.00
+ ;; size=50 bbWeight=4 PerfScore 46.00
G_M49741_IG04:
- mov rcx, qword ptr [rbp-0x28]
- mov rax, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rax
- mov rax, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rax
- xor esi, esi
- align [3 bytes for IG05]
- ;; size=25 bbWeight=1 PerfScore 5.50
+ mov r15, qword ptr [rbp-0x58]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x78], rdi
+ mov rdi, qword ptr [rbp-0x48]
+ mov qword ptr [rbp-0x80], rdi
+ xor r14d, r14d
+ ;; size=23 bbWeight=1 PerfScore 5.25
G_M49741_IG05:
- lea rax, [rbp-0x48]
- movsxd r8, esi
- movzx rax, byte ptr [rax+r8]
- lea rdx, [rbp-0x50]
- movzx r9, byte ptr [rdx+r8]
- xor edx, edx
- div edx:eax, r9d
- lea rdx, [rbp-0x40]
- mov byte ptr [rdx+r8], al
- inc esi
- cmp esi, 8
+ lea rdi, [rbp-0x78]
+ movsxd r13, r14d
+ movzx rdi, byte ptr [rdi+r13]
+ lea rsi, [rbp-0x80]
+ movzx rsi, byte ptr [rsi+r13]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0x70]
+ mov byte ptr [rdi+r13], al
+ inc r14d
+ cmp r14d, 8
jl SHORT G_M49741_IG05
- ;; size=41 bbWeight=4 PerfScore 134.00
+ ;; size=50 bbWeight=4 PerfScore 46.00
G_M49741_IG06:
- mov rax, qword ptr [rbp-0x40]
- mov qword ptr [rbp-0x60], rcx
- mov qword ptr [rbp-0x58], rax
- vmovaps xmm2, xmmword ptr [rbp-0x60]
+ mov rdi, qword ptr [rbp-0x70]
+ mov qword ptr [rbp-0x90], r15
+ mov qword ptr [rbp-0x88], rdi
+ vmovaps xmm2, xmmword ptr [rbp-0x90]
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ vmovups ymm0, ymmword ptr [rbp+0x10]
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
+ vmovaps xmmword ptr [rbp-0xA0], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x30]
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- xor ecx, ecx
- align [0 bytes for IG07]
- ;; size=63 bbWeight=1 PerfScore 16.25
+ vmovaps xmmword ptr [rbp-0xB0], xmm0
+ mov rdi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xC0], rdi
+ mov rdi, qword ptr [rbp-0xB0]
+ mov qword ptr [rbp-0xC8], rdi
+ xor r15d, r15d
+ ;; size=100 bbWeight=1 PerfScore 25.25
G_M49741_IG07:
- lea rax, [rbp-0x90]
- movsxd rsi, ecx
- movzx rax, byte ptr [rax+rsi]
- lea rdx, [rbp-0x98]
- movzx r8, byte ptr [rdx+rsi]
- xor edx, edx
- div edx:eax, r8d
- lea rdx, [rbp-0x88]
- mov byte ptr [rdx+rsi], al
- inc ecx
- cmp ecx, 8
+ lea rdi, [rbp-0xC0]
+ movsxd r14, r15d
+ movzx rdi, byte ptr [rdi+r14]
+ lea rsi, [rbp-0xC8]
+ movzx rsi, byte ptr [rsi+r14]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0xB8]
+ mov byte ptr [rdi+r14], al
+ inc r15d
+ cmp r15d, 8
jl SHORT G_M49741_IG07
- ;; size=48 bbWeight=4 PerfScore 134.00
+ ;; size=59 bbWeight=4 PerfScore 46.00
G_M49741_IG08:
- mov rcx, qword ptr [rbp-0x88]
- mov rax, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rax
- mov rax, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rax
- xor esi, esi
- align [0 bytes for IG09]
- ;; size=31 bbWeight=1 PerfScore 5.25
+ mov r15, qword ptr [rbp-0xB8]
+ mov rdi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xD8], rdi
+ mov rdi, qword ptr [rbp-0xA8]
+ mov qword ptr [rbp-0xE0], rdi
+ xor r14d, r14d
+ ;; size=38 bbWeight=1 PerfScore 5.25
G_M49741_IG09:
- lea rax, [rbp-0xA8]
- movsxd r8, esi
- movzx rax, byte ptr [rax+r8]
- lea rdx, [rbp-0xB0]
- movzx r9, byte ptr [rdx+r8]
- xor edx, edx
- div edx:eax, r9d
- lea rdx, [rbp-0xA0]
- mov byte ptr [rdx+r8], al
- inc esi
- cmp esi, 8
+ lea rdi, [rbp-0xD8]
+ movsxd r13, r14d
+ movzx rdi, byte ptr [rdi+r13]
+ lea rsi, [rbp-0xE0]
+ movzx rsi, byte ptr [rsi+r13]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ call [rax]System.Runtime.Intrinsics.Scalar`1[ubyte]:Divide(ubyte,ubyte):ubyte
+ lea rdi, [rbp-0xD0]
+ mov byte ptr [rdi+r13], al
+ inc r14d
+ cmp r14d, 8
jl SHORT G_M49741_IG09
- ;; size=50 bbWeight=4 PerfScore 134.00
+ ;; size=59 bbWeight=4 PerfScore 46.00
G_M49741_IG10:
- mov rax, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rcx
- mov qword ptr [rbp-0xB8], rax
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=38 bbWeight=1 PerfScore 9.25
+ mov rdi, qword ptr [rbp-0xD0]
+ mov qword ptr [rbp-0xF0], r15
+ mov qword ptr [rbp-0xE8], rdi
+ vmovaps xmm2, xmmword ptr [rbp-0x30]
+ vmovups xmmword ptr [rsp], xmm2
+ vmovaps xmm0, xmmword ptr [rbp-0xF0]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[ubyte](System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[ubyte](System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte]
+ mov rax, rbx
+ ;; size=63 bbWeight=1 PerfScore 14.75
G_M49741_IG11:
vzeroupper
- add rsp, 192
+ add rsp, 240
+ pop rbx
+ pop r13
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=19 bbWeight=1 PerfScore 4.75
-; Total bytes of code 411, prolog size 26, PerfScore 591.75, instruction count 101, allocated bytes for code 411 (MethodHash=ef7c3db2) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)
+; Total bytes of code 544, prolog size 23, PerfScore 262.00, instruction count 116, allocated bytes for code 544 (MethodHash=ef7c3db2) for method System.Runtime.Intrinsics.Vector256`1[ubyte]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Divide(System.Runtime.Intrinsics.Vector256`1[ubyte],System.Runtime.Intrinsics.Vector256`1[ubyte]):System.Runtime.Intrinsics.Vector256`1[ubyte] (FullOpts)132 (62.56 % of base) - System.Runtime.Intrinsics.Vector256:Dot[long](System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long
; Assembly listing for method System.Runtime.Intrinsics.Vector256:Dot[long](System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
-; rsp based frame
+; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 10 single block inlinees; 11 inlinees without PGO data
+; 0 inlinees with PGO data; 22 single block inlinees; 17 inlinees without PGO data
; Final local variable assignments
;
-; V00 arg0 [V00,T08] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
-; V01 arg1 [V01,T09] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V00 arg0 [V00,T30] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V01 arg1 [V01,T31] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
;# V02 OutArgs [V02 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V03 tmp1 [V03,T00] ( 2, 4 ) long -> rax "impAppendStmt"
-; V04 tmp2 [V04,T04] ( 3, 6 ) simd16 -> [rsp+0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V05 tmp3 [V05,T05] ( 3, 6 ) simd16 -> [rsp+0x60] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V06 tmp4 [V06,T01] ( 2, 4 ) long -> rax "impAppendStmt"
-;* V07 tmp5 [V07 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V08 tmp6 [V08 ] ( 2, 4 ) struct ( 8) [rsp+0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V09 tmp7 [V09 ] ( 2, 4 ) struct ( 8) [rsp+0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V10 tmp8 [V10 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V11 tmp9 [V11 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V12 tmp10 [V12 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V13 tmp11 [V13 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V14 tmp12 [V14 ] ( 2, 4 ) struct ( 8) [rsp+0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V15 tmp13 [V15 ] ( 2, 4 ) struct ( 8) [rsp+0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V16 tmp14 [V16 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V17 tmp15 [V17 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V18 tmp16 [V18 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V19 tmp17 [V19 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V20 tmp18 [V20 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V21 tmp19 [V21,T06] ( 3, 6 ) simd16 -> [rsp+0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V22 tmp20 [V22,T07] ( 3, 6 ) simd16 -> [rsp+0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V23 tmp21 [V23,T02] ( 2, 4 ) long -> rcx "impAppendStmt"
-;* V24 tmp22 [V24 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V25 tmp23 [V25 ] ( 2, 4 ) struct ( 8) [rsp+0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V26 tmp24 [V26 ] ( 2, 4 ) struct ( 8) [rsp+0x10] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V27 tmp25 [V27 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V28 tmp26 [V28 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V29 tmp27 [V29 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V30 tmp28 [V30 ] ( 0, 0 ) long -> zero-ref ld-addr-op "Inline ldloca(s) first use temp"
-; V31 tmp29 [V31 ] ( 2, 4 ) struct ( 8) [rsp+0x08] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V32 tmp30 [V32 ] ( 2, 4 ) struct ( 8) [rsp+0x00] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V33 tmp31 [V33 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V34 tmp32 [V34 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V35 tmp33 [V35 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-;* V36 tmp34 [V36 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V37 tmp35 [V37 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V38 tmp36 [V38,T03] ( 2, 2 ) long -> rax "Inline return value spill temp"
-; V39 tmp37 [V39 ] ( 2, 3 ) long -> [rsp+0x58] do-not-enreg[X] addr-exposed "field V08._00 (fldOffset=0x0)" P-DEP
-; V40 tmp38 [V40 ] ( 2, 3 ) long -> [rsp+0x50] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V41 tmp39 [V41 ] ( 2, 3 ) long -> [rsp+0x48] do-not-enreg[X] addr-exposed "field V14._00 (fldOffset=0x0)" P-DEP
-; V42 tmp40 [V42 ] ( 2, 3 ) long -> [rsp+0x40] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V43 tmp41 [V43 ] ( 2, 3 ) long -> [rsp+0x18] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V44 tmp42 [V44 ] ( 2, 3 ) long -> [rsp+0x10] do-not-enreg[X] addr-exposed "field V26._00 (fldOffset=0x0)" P-DEP
-; V45 tmp43 [V45 ] ( 2, 3 ) long -> [rsp+0x08] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V46 tmp44 [V46 ] ( 2, 3 ) long -> [rsp+0x00] do-not-enreg[X] addr-exposed "field V32._00 (fldOffset=0x0)" P-DEP
+; V03 tmp1 [V03,T00] ( 2, 4 ) long -> rbx "impAppendStmt"
+; V04 tmp2 [V04,T26] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V05 tmp3 [V05,T27] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V06 tmp4 [V06,T01] ( 2, 4 ) long -> rbx "impAppendStmt"
+;* V07 tmp5 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V08 tmp6 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V09 tmp7 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V10 tmp8 [V10 ] ( 2, 2 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V11 tmp9 [V11,T18] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V12 tmp10 [V12 ] ( 2, 4 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V13 tmp11 [V13 ] ( 2, 4 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V14 tmp12 [V14 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V15 tmp13 [V15 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V16 tmp14 [V16 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V17 tmp15 [V17 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V18 tmp16 [V18,T03] ( 2, 2 ) long -> rbx ld-addr-op "Inline ldloca(s) first use temp"
+;* V19 tmp17 [V19,T19] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V20 tmp18 [V20 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V21 tmp19 [V21 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V22 tmp20 [V22 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V23 tmp21 [V23 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+;* V24 tmp22 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V25 tmp23 [V25 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V26 tmp24 [V26 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V27 tmp25 [V27 ] ( 2, 2 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V28 tmp26 [V28,T20] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp27 [V29 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V30 tmp28 [V30 ] ( 2, 4 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V31 tmp29 [V31 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V32 tmp30 [V32 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V33 tmp31 [V33 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V34 tmp32 [V34 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V35 tmp33 [V35,T04] ( 2, 2 ) long -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V36 tmp34 [V36,T21] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V37 tmp35 [V37 ] ( 2, 4 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V38 tmp36 [V38 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V39 tmp37 [V39 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V40 tmp38 [V40 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+;* V41 tmp39 [V41 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V42 tmp40 [V42 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V43 tmp41 [V43,T28] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V44 tmp42 [V44,T29] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V45 tmp43 [V45,T02] ( 2, 4 ) long -> r15 "impAppendStmt"
+;* V46 tmp44 [V46 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V47 tmp45 [V47 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V48 tmp46 [V48 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V49 tmp47 [V49 ] ( 2, 2 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V50 tmp48 [V50,T22] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V51 tmp49 [V51 ] ( 2, 4 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V52 tmp50 [V52 ] ( 2, 4 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V53 tmp51 [V53 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V54 tmp52 [V54 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V55 tmp53 [V55 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V56 tmp54 [V56 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V57 tmp55 [V57,T05] ( 2, 2 ) long -> r15 ld-addr-op "Inline ldloca(s) first use temp"
+;* V58 tmp56 [V58,T23] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V59 tmp57 [V59 ] ( 2, 4 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V60 tmp58 [V60 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V61 tmp59 [V61 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V62 tmp60 [V62 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+;* V63 tmp61 [V63 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V64 tmp62 [V64 ] ( 0, 0 ) struct ( 8) zero-ref "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V65 tmp63 [V65 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V66 tmp64 [V66 ] ( 2, 2 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V67 tmp65 [V67,T24] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V68 tmp66 [V68 ] ( 2, 4 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V69 tmp67 [V69 ] ( 2, 4 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V70 tmp68 [V70 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V71 tmp69 [V71 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+;* V72 tmp70 [V72 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V73 tmp71 [V73 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V74 tmp72 [V74,T06] ( 2, 2 ) long -> rsi ld-addr-op "Inline ldloca(s) first use temp"
+;* V75 tmp73 [V75,T25] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V76 tmp74 [V76 ] ( 2, 4 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V77 tmp75 [V77 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V78 tmp76 [V78 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
+;* V79 tmp77 [V79 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
+; V80 tmp78 [V80,T07] ( 2, 2 ) long -> rsi "field V07._00 (fldOffset=0x0)" P-INDEP
+; V81 tmp79 [V81,T08] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+;* V82 tmp80 [V82 ] ( 0, 0 ) long -> zero-ref "field V09._00 (fldOffset=0x0)" P-INDEP
+; V83 tmp81 [V83 ] ( 2, 2 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V84 tmp82 [V84 ] ( 2, 3 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V85 tmp83 [V85 ] ( 2, 3 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V86 tmp84 [V86 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V20._00 (fldOffset=0x0)" P-DEP
+; V87 tmp85 [V87,T09] ( 2, 2 ) long -> rsi "field V24._00 (fldOffset=0x0)" P-INDEP
+; V88 tmp86 [V88,T10] ( 2, 2 ) long -> rdi "field V25._00 (fldOffset=0x0)" P-INDEP
+;* V89 tmp87 [V89 ] ( 0, 0 ) long -> zero-ref "field V26._00 (fldOffset=0x0)" P-INDEP
+; V90 tmp88 [V90 ] ( 2, 2 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V91 tmp89 [V91 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V92 tmp90 [V92 ] ( 2, 3 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V30._00 (fldOffset=0x0)" P-DEP
+; V93 tmp91 [V93 ] ( 2, 3 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V37._00 (fldOffset=0x0)" P-DEP
+; V94 tmp92 [V94,T11] ( 2, 2 ) long -> rsi "field V46._00 (fldOffset=0x0)" P-INDEP
+; V95 tmp93 [V95,T12] ( 2, 2 ) long -> rdi "field V47._00 (fldOffset=0x0)" P-INDEP
+;* V96 tmp94 [V96 ] ( 0, 0 ) long -> zero-ref "field V48._00 (fldOffset=0x0)" P-INDEP
+; V97 tmp95 [V97 ] ( 2, 2 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V49._00 (fldOffset=0x0)" P-DEP
+; V98 tmp96 [V98 ] ( 2, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V51._00 (fldOffset=0x0)" P-DEP
+; V99 tmp97 [V99 ] ( 2, 3 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V52._00 (fldOffset=0x0)" P-DEP
+; V100 tmp98 [V100 ] ( 2, 3 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V59._00 (fldOffset=0x0)" P-DEP
+; V101 tmp99 [V101,T13] ( 2, 2 ) long -> rsi "field V63._00 (fldOffset=0x0)" P-INDEP
+; V102 tmp100 [V102,T14] ( 2, 2 ) long -> rdi "field V64._00 (fldOffset=0x0)" P-INDEP
+;* V103 tmp101 [V103 ] ( 0, 0 ) long -> zero-ref "field V65._00 (fldOffset=0x0)" P-INDEP
+; V104 tmp102 [V104 ] ( 2, 2 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V66._00 (fldOffset=0x0)" P-DEP
+; V105 tmp103 [V105 ] ( 2, 3 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V68._00 (fldOffset=0x0)" P-DEP
+; V106 tmp104 [V106 ] ( 2, 3 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V69._00 (fldOffset=0x0)" P-DEP
+; V107 tmp105 [V107 ] ( 2, 3 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V76._00 (fldOffset=0x0)" P-DEP
+; V108 cse0 [V108,T15] ( 2, 2 ) long -> rbx "CSE #01: moderate"
+; V109 cse1 [V109,T16] ( 2, 2 ) long -> r15 "CSE #02: moderate"
+; V110 cse2 [V110,T17] ( 2, 2 ) long -> r15 "CSE #03: moderate"
;
-; Lcl frame size = 136
+; Lcl frame size = 192
G_M27773_IG01:
- sub rsp, 136
- vmovups ymm0, ymmword ptr [rsp+0x90]
- vmovups ymm1, ymmword ptr [rsp+0xB0]
- ;; size=25 bbWeight=1 PerfScore 8.25
+ push rbp
+ push r15
+ push rbx
+ sub rsp, 192
+ lea rbp, [rsp+0xD0]
+ vmovups ymm0, ymmword ptr [rbp+0x10]
+ vmovups ymm1, ymmword ptr [rbp+0x30]
+ ;; size=29 bbWeight=1 PerfScore 11.75
G_M27773_IG02:
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rsp+0x70], xmm2
+ vmovaps xmmword ptr [rbp-0x20], xmm2
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rsp+0x60], xmm2
- mov rax, qword ptr [rsp+0x70]
- mov qword ptr [rsp+0x58], rax
- mov rax, qword ptr [rsp+0x60]
- mov qword ptr [rsp+0x50], rax
- mov rax, qword ptr [rsp+0x58]
- imul rax, qword ptr [rsp+0x50]
- mov rcx, qword ptr [rsp+0x78]
- mov qword ptr [rsp+0x48], rcx
- mov rcx, qword ptr [rsp+0x68]
- mov qword ptr [rsp+0x40], rcx
- mov rcx, qword ptr [rsp+0x48]
- imul rcx, qword ptr [rsp+0x40]
- add rax, rcx
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ mov rsi, qword ptr [rbp-0x20]
+ mov rdi, qword ptr [rbp-0x30]
+ mov qword ptr [rbp-0x40], rsi
+ mov qword ptr [rbp-0x48], rdi
+ mov rsi, qword ptr [rbp-0x40]
+ imul rsi, qword ptr [rbp-0x48]
+ mov qword ptr [rbp-0x38], rsi
+ mov rsi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x50], rsi
+ mov rbx, qword ptr [rbp-0x50]
+ mov rsi, qword ptr [rbp-0x18]
+ mov rdi, qword ptr [rbp-0x28]
+ mov qword ptr [rbp-0x60], rsi
+ mov qword ptr [rbp-0x68], rdi
+ mov rsi, qword ptr [rbp-0x60]
+ imul rsi, qword ptr [rbp-0x68]
+ mov qword ptr [rbp-0x58], rsi
+ mov rsi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], rsi
+ mov r15, qword ptr [rbp-0x70]
+ add rbx, r15
vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rsp+0x30], xmm0
+ vmovaps xmmword ptr [rbp-0x80], xmm0
vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rsp+0x20], xmm0
- mov rcx, qword ptr [rsp+0x30]
- mov qword ptr [rsp+0x18], rcx
- mov rcx, qword ptr [rsp+0x20]
- mov qword ptr [rsp+0x10], rcx
- mov rcx, qword ptr [rsp+0x18]
- imul rcx, qword ptr [rsp+0x10]
- mov rdx, qword ptr [rsp+0x38]
- mov qword ptr [rsp+0x08], rdx
- mov rdx, qword ptr [rsp+0x28]
- mov qword ptr [rsp], rdx
- add rax, rcx
- mov rcx, qword ptr [rsp+0x08]
- imul rcx, qword ptr [rsp]
- add rax, rcx
- ;; size=175 bbWeight=1 PerfScore 45.25
+ vmovaps xmmword ptr [rbp-0x90], xmm0
+ mov rsi, qword ptr [rbp-0x80]
+ mov rdi, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xA0], rsi
+ mov qword ptr [rbp-0xA8], rdi
+ mov rsi, qword ptr [rbp-0xA0]
+ imul rsi, qword ptr [rbp-0xA8]
+ mov qword ptr [rbp-0x98], rsi
+ mov rsi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], rsi
+ mov r15, qword ptr [rbp-0xB0]
+ mov rsi, qword ptr [rbp-0x78]
+ mov rdi, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0xC0], rsi
+ mov qword ptr [rbp-0xC8], rdi
+ mov rsi, qword ptr [rbp-0xC0]
+ imul rsi, qword ptr [rbp-0xC8]
+ mov qword ptr [rbp-0xB8], rsi
+ mov rsi, qword ptr [rbp-0xB8]
+ mov qword ptr [rbp-0xD0], rsi
+ mov rsi, qword ptr [rbp-0xD0]
+ xor edi, edi
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ mov rsi, rax
+ mov rdi, r15
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Add(long,long):long
+ add rax, rbx
+ ;; size=299 bbWeight=1 PerfScore 68.25
G_M27773_IG03:
vzeroupper
- add rsp, 136
+ add rsp, 192
+ pop rbx
+ pop r15
+ pop rbp
ret
- ;; size=11 bbWeight=1 PerfScore 2.25
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 211, prolog size 7, PerfScore 55.75, instruction count 41, allocated bytes for code 211 (MethodHash=9cc79382) for method System.Runtime.Intrinsics.Vector256:Dot[long](System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long (FullOpts)
+; Total bytes of code 343, prolog size 19, PerfScore 83.75, instruction count 70, allocated bytes for code 343 (MethodHash=9cc79382) for method System.Runtime.Intrinsics.Vector256:Dot[long](System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):long (FullOpts)121 (41.72 % of base) - System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector,T>.Multiply(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Multiply(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 20 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 20 single block inlinees; 8 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
-; V02 arg1 [V02,T17] ( 2, 2 ) simd32 -> mm1 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> mm2 "impAppendStmt"
+; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rbx single-def
+; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V02 arg1 [V02,T17] ( 2, 2 ) simd32 -> [rbp+0x30] single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> [rbp-0x20] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T09] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V07 tmp4 [V07,T10] ( 3, 6 ) simd16 -> [rbp-0x20] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V06 tmp3 [V06,T09] ( 3, 6 ) simd16 -> [rbp-0x30] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V07 tmp4 [V07,T10] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V09 tmp6 [V09 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V10 tmp7 [V10 ] ( 2, 2 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V10 tmp7 [V10 ] ( 2, 2 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V11 tmp8 [V11,T05] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V12 tmp9 [V12 ] ( 2, 4 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V13 tmp10 [V13 ] ( 2, 4 ) struct ( 8) [rbp-0x38] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V12 tmp9 [V12 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V13 tmp10 [V13 ] ( 2, 4 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V14 tmp11 [V14 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
;* V15 tmp12 [V15 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V16 tmp13 [V16 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V17 tmp14 [V17 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V18 tmp15 [V18 ] ( 2, 2 ) struct ( 8) [rbp-0x40] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V19 tmp16 [V19,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V20 tmp17 [V20 ] ( 2, 4 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V21 tmp18 [V21 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V22 tmp19 [V22 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V23 tmp20 [V23 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V24 tmp21 [V24 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V25 tmp22 [V25 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V26 tmp23 [V26,T14] ( 3, 3 ) simd16 -> [rbp-0x60] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V27 tmp24 [V27,T11] ( 3, 6 ) simd16 -> [rbp-0x70] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V28 tmp25 [V28,T12] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-;* V29 tmp26 [V29 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V30 tmp27 [V30 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V31 tmp28 [V31 ] ( 2, 2 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V32 tmp29 [V32,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 2, 4 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V34 tmp31 [V34 ] ( 2, 4 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V35 tmp32 [V35 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V36 tmp33 [V36 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V37 tmp34 [V37 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V38 tmp35 [V38 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V39 tmp36 [V39 ] ( 2, 2 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V40 tmp37 [V40,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V41 tmp38 [V41 ] ( 2, 4 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V42 tmp39 [V42 ] ( 2, 4 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V43 tmp40 [V43 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
-;* V44 tmp41 [V44 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V45 tmp42 [V45 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V46 tmp43 [V46 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V47 tmp44 [V47,T15] ( 3, 3 ) simd16 -> [rbp-0xC0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-;* V48 tmp45 [V48 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V49 tmp46 [V49,T01] ( 2, 2 ) long -> rax "field V08._00 (fldOffset=0x0)" P-INDEP
-; V50 tmp47 [V50,T02] ( 2, 2 ) long -> rcx "field V09._00 (fldOffset=0x0)" P-INDEP
-; V51 tmp48 [V51 ] ( 2, 2 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
-; V52 tmp49 [V52 ] ( 2, 3 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
-; V53 tmp50 [V53 ] ( 2, 3 ) long -> [rbp-0x38] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
-; V54 tmp51 [V54 ] ( 2, 2 ) long -> [rbp-0x40] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
-; V55 tmp52 [V55 ] ( 2, 3 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V20._00 (fldOffset=0x0)" P-DEP
-; V56 tmp53 [V56 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
-; V57 tmp54 [V57,T03] ( 2, 2 ) long -> rax "field V29._00 (fldOffset=0x0)" P-INDEP
-; V58 tmp55 [V58,T04] ( 2, 2 ) long -> rcx "field V30._00 (fldOffset=0x0)" P-INDEP
-; V59 tmp56 [V59 ] ( 2, 2 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V60 tmp57 [V60 ] ( 2, 3 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V61 tmp58 [V61 ] ( 2, 3 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V34._00 (fldOffset=0x0)" P-DEP
-; V62 tmp59 [V62 ] ( 2, 2 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V39._00 (fldOffset=0x0)" P-DEP
-; V63 tmp60 [V63 ] ( 2, 3 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V41._00 (fldOffset=0x0)" P-DEP
-; V64 tmp61 [V64 ] ( 2, 3 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V42._00 (fldOffset=0x0)" P-DEP
+; V16 tmp13 [V16 ] ( 2, 2 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V17 tmp14 [V17,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V18 tmp15 [V18 ] ( 2, 4 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V19 tmp16 [V19 ] ( 2, 4 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V21 tmp18 [V21 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V22 tmp19 [V22,T14] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V23 tmp20 [V23,T11] ( 3, 6 ) simd16 -> [rbp-0x90] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V24 tmp21 [V24,T12] ( 3, 6 ) simd16 -> [rbp-0xA0] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+;* V25 tmp22 [V25 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V27 tmp24 [V27 ] ( 2, 2 ) struct ( 8) [rbp-0xA8] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V28 tmp25 [V28,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 2, 4 ) struct ( 8) [rbp-0xB0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V30 tmp27 [V30 ] ( 2, 4 ) struct ( 8) [rbp-0xB8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V31 tmp28 [V31 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V32 tmp29 [V32 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V33 tmp30 [V33 ] ( 2, 2 ) struct ( 8) [rbp-0xC0] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V34 tmp31 [V34,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V35 tmp32 [V35 ] ( 2, 4 ) struct ( 8) [rbp-0xC8] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V36 tmp33 [V36 ] ( 2, 4 ) struct ( 8) [rbp-0xD0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V37 tmp34 [V37 ] ( 0, 0 ) long -> zero-ref "impAppendStmt"
+;* V38 tmp35 [V38 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V39 tmp36 [V39,T15] ( 3, 3 ) simd16 -> [rbp-0xE0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V40 tmp37 [V40,T01] ( 2, 2 ) long -> r15 "field V08._00 (fldOffset=0x0)" P-INDEP
+; V41 tmp38 [V41,T02] ( 2, 2 ) long -> rdi "field V09._00 (fldOffset=0x0)" P-INDEP
+; V42 tmp39 [V42 ] ( 2, 2 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V10._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V12._00 (fldOffset=0x0)" P-DEP
+; V44 tmp41 [V44 ] ( 2, 3 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V45 tmp42 [V45 ] ( 2, 2 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V16._00 (fldOffset=0x0)" P-DEP
+; V46 tmp43 [V46 ] ( 2, 3 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V18._00 (fldOffset=0x0)" P-DEP
+; V47 tmp44 [V47 ] ( 2, 3 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V19._00 (fldOffset=0x0)" P-DEP
+; V48 tmp45 [V48,T03] ( 2, 2 ) long -> r15 "field V25._00 (fldOffset=0x0)" P-INDEP
+; V49 tmp46 [V49,T04] ( 2, 2 ) long -> rdi "field V26._00 (fldOffset=0x0)" P-INDEP
+; V50 tmp47 [V50 ] ( 2, 2 ) long -> [rbp-0xA8] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V51 tmp48 [V51 ] ( 2, 3 ) long -> [rbp-0xB0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
+; V52 tmp49 [V52 ] ( 2, 3 ) long -> [rbp-0xB8] do-not-enreg[X] addr-exposed "field V30._00 (fldOffset=0x0)" P-DEP
+; V53 tmp50 [V53 ] ( 2, 2 ) long -> [rbp-0xC0] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V54 tmp51 [V54 ] ( 2, 3 ) long -> [rbp-0xC8] do-not-enreg[X] addr-exposed "field V35._00 (fldOffset=0x0)" P-DEP
+; V55 tmp52 [V55 ] ( 2, 3 ) long -> [rbp-0xD0] do-not-enreg[X] addr-exposed "field V36._00 (fldOffset=0x0)" P-DEP
;
-; Lcl frame size = 192
+; Lcl frame size = 240
G_M16594_IG01:
push rbp
- sub rsp, 192
- lea rbp, [rsp+0xC0]
+ push r15
+ push rbx
+ sub rsp, 240
+ lea rbp, [rsp+0x100]
+ mov rbx, rdi
vmovups ymm0, ymmword ptr [rbp+0x10]
vmovups ymm1, ymmword ptr [rbp+0x30]
- ;; size=26 bbWeight=1 PerfScore 9.75
+ ;; size=32 bbWeight=1 PerfScore 12.00
G_M16594_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm2, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm2
+ vmovaps xmmword ptr [rbp-0x30], xmm2
+ vmovups ymmword ptr [rbp+0x30], ymm1
vmovaps ymm2, ymm1
- vmovaps xmmword ptr [rbp-0x20], xmm2
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x30], rax
- mov rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x38], rax
- mov rax, qword ptr [rbp-0x30]
- imul rax, qword ptr [rbp-0x38]
- mov qword ptr [rbp-0x28], rax
- mov rax, qword ptr [rbp-0x28]
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x48], rcx
- mov rcx, qword ptr [rbp-0x18]
- mov qword ptr [rbp-0x50], rcx
- mov rcx, qword ptr [rbp-0x48]
- imul rcx, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x40], rcx
- mov rcx, qword ptr [rbp-0x40]
+ vmovaps xmmword ptr [rbp-0x40], xmm2
+ mov rdi, qword ptr [rbp-0x30]
+ mov qword ptr [rbp-0x50], rdi
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x58], rdi
+ mov rdi, qword ptr [rbp-0x50]
+ mov rsi, qword ptr [rbp-0x58]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x48], rax
+ mov r15, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x28]
+ mov qword ptr [rbp-0x68], rdi
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x70], rdi
+ mov rdi, qword ptr [rbp-0x68]
+ mov rsi, qword ptr [rbp-0x70]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
mov qword ptr [rbp-0x60], rax
- mov qword ptr [rbp-0x58], rcx
- vmovaps xmm2, xmmword ptr [rbp-0x60]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x70], xmm0
- vextractf128 xmm0, ymm1, 1
- vmovaps xmmword ptr [rbp-0x80], xmm0
- mov rax, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x90], rax
- mov rax, qword ptr [rbp-0x80]
- mov qword ptr [rbp-0x98], rax
- mov rax, qword ptr [rbp-0x90]
- imul rax, qword ptr [rbp-0x98]
- mov qword ptr [rbp-0x88], rax
- mov rax, qword ptr [rbp-0x88]
- mov rcx, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0xA8], rcx
- mov rcx, qword ptr [rbp-0x78]
- mov qword ptr [rbp-0xB0], rcx
- mov rcx, qword ptr [rbp-0xA8]
- imul rcx, qword ptr [rbp-0xB0]
- mov qword ptr [rbp-0xA0], rcx
- mov rcx, qword ptr [rbp-0xA0]
- mov qword ptr [rbp-0xC0], rax
- mov qword ptr [rbp-0xB8], rcx
- vinserti128 ymm0, ymm2, xmmword ptr [rbp-0xC0], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=252 bbWeight=1 PerfScore 65.75
+ mov rdi, qword ptr [rbp-0x60]
+ mov qword ptr [rbp-0x80], r15
+ mov qword ptr [rbp-0x78], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x80]
+ vmovaps xmmword ptr [rbp-0x20], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x90], xmm1
+ vmovups ymm1, ymmword ptr [rbp+0x30]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0xA0], xmm1
+ mov rdi, qword ptr [rbp-0x90]
+ mov qword ptr [rbp-0xB0], rdi
+ mov rdi, qword ptr [rbp-0xA0]
+ mov qword ptr [rbp-0xB8], rdi
+ mov rdi, qword ptr [rbp-0xB0]
+ mov rsi, qword ptr [rbp-0xB8]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0xA8], rax
+ mov r15, qword ptr [rbp-0xA8]
+ mov rdi, qword ptr [rbp-0x88]
+ mov qword ptr [rbp-0xC8], rdi
+ mov rdi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xD0], rdi
+ mov rdi, qword ptr [rbp-0xC8]
+ mov rsi, qword ptr [rbp-0xD0]
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ ;; size=292 bbWeight=1 PerfScore 64.50
G_M16594_IG03:
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0xC0], rax
+ mov rdi, qword ptr [rbp-0xC0]
+ mov qword ptr [rbp-0xE0], r15
+ mov qword ptr [rbp-0xD8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x20]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0xE0]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ mov rdi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[long](System.Runtime.Intrinsics.Vector128`1[long],System.Runtime.Intrinsics.Vector128`1[long]):System.Runtime.Intrinsics.Vector256`1[long]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[long](System.Runtime.Intrinsics.Vector128`1[long],System.Runtime.Intrinsics.Vector128`1[long]):System.Runtime.Intrinsics.Vector256`1[long]
+ mov rax, rbx
+ ;; size=72 bbWeight=1 PerfScore 18.75
+G_M16594_IG04:
vzeroupper
- add rsp, 192
+ add rsp, 240
+ pop rbx
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=15 bbWeight=1 PerfScore 3.75
-; Total bytes of code 290, prolog size 16, PerfScore 78.25, instruction count 57, allocated bytes for code 290 (MethodHash=093fbf2d) for method System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Multiply(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long] (FullOpts)
+; Total bytes of code 411, prolog size 19, PerfScore 99.00, instruction count 80, allocated bytes for code 411 (MethodHash=093fbf2d) for method System.Runtime.Intrinsics.Vector256`1[long]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.Multiply(System.Runtime.Intrinsics.Vector256`1[long],System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long] (FullOpts)118 (61.78 % of base) - System.Runtime.Intrinsics.Vector256`1[double]:System.Runtime.Intrinsics.ISimdVector,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[double],int):System.Runtime.Intrinsics.Vector256`1[double]
; Assembly listing for method System.Runtime.Intrinsics.Vector256`1[double]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[double],int):System.Runtime.Intrinsics.Vector256`1[double] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 16 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 16 single block inlinees; 9 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T00] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T16] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[double]>
-; V02 arg1 [V02,T02] ( 3, 3 ) int -> rsi single-def
-;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T13] ( 2, 4 ) simd16 -> mm1 "impAppendStmt"
+; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> r15 single-def
+; V01 arg0 [V01,T15] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[double]>
+; V02 arg1 [V02,T00] ( 6, 6 ) int -> rbx single-def
+; V03 OutArgs [V03 ] ( 1, 1 ) struct (32) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
+; V04 tmp1 [V04,T12] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T11] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[double]>
+; V06 tmp3 [V06,T10] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[double]>
;* V07 tmp4 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[double]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[double]>
-; V09 tmp6 [V09 ] ( 2, 2 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V10 tmp7 [V10,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V11 tmp8 [V11 ] ( 2, 4 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
+; V09 tmp6 [V09 ] ( 2, 2 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V10 tmp7 [V10,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V11 tmp8 [V11 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
;* V12 tmp9 [V12 ] ( 0, 0 ) double -> zero-ref "Inline stloc first use temp"
;* V13 tmp10 [V13 ] ( 0, 0 ) double -> zero-ref "Inline return value spill temp"
;* V14 tmp11 [V14 ] ( 0, 0 ) double -> zero-ref "Inlining Arg"
-; V15 tmp12 [V15 ] ( 2, 2 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V16 tmp13 [V16,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V17 tmp14 [V17 ] ( 2, 4 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
+; V15 tmp12 [V15 ] ( 2, 2 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V16 tmp13 [V16,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V17 tmp14 [V17 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
;* V18 tmp15 [V18 ] ( 0, 0 ) double -> zero-ref "Inline stloc first use temp"
-;* V19 tmp16 [V19 ] ( 0, 0 ) double -> zero-ref "Inline return value spill temp"
-;* V20 tmp17 [V20 ] ( 0, 0 ) double -> zero-ref "Inlining Arg"
-; V21 tmp18 [V21,T14] ( 3, 3 ) simd16 -> [rbp-0x40] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[double]>
-; V22 tmp19 [V22,T12] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[double]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V24 tmp21 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[double]>
-; V25 tmp22 [V25 ] ( 2, 2 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V26 tmp23 [V26,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V27 tmp24 [V27 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V28 tmp25 [V28 ] ( 0, 0 ) double -> zero-ref "Inline stloc first use temp"
-;* V29 tmp26 [V29 ] ( 0, 0 ) double -> zero-ref "Inline return value spill temp"
-;* V30 tmp27 [V30 ] ( 0, 0 ) double -> zero-ref "Inlining Arg"
-; V31 tmp28 [V31 ] ( 2, 2 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V32 tmp29 [V32,T10] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 2, 4 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
-;* V34 tmp31 [V34 ] ( 0, 0 ) double -> zero-ref "Inline stloc first use temp"
-;* V35 tmp32 [V35 ] ( 0, 0 ) double -> zero-ref "Inline return value spill temp"
-;* V36 tmp33 [V36 ] ( 0, 0 ) double -> zero-ref "Inlining Arg"
-; V37 tmp34 [V37,T15] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[double]>
-;* V38 tmp35 [V38 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[double]>
-; V39 tmp36 [V39,T03] ( 2, 2 ) long -> rax "field V07._00 (fldOffset=0x0)" P-INDEP
-; V40 tmp37 [V40,T04] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V41 tmp38 [V41 ] ( 2, 2 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V42 tmp39 [V42 ] ( 2, 3 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V43 tmp40 [V43 ] ( 2, 2 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V44 tmp41 [V44 ] ( 2, 3 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V45 tmp42 [V45,T05] ( 2, 2 ) long -> rax "field V23._00 (fldOffset=0x0)" P-INDEP
-; V46 tmp43 [V46,T06] ( 2, 2 ) long -> rcx "field V24._00 (fldOffset=0x0)" P-INDEP
-; V47 tmp44 [V47 ] ( 2, 2 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp45 [V48 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
-; V49 tmp46 [V49 ] ( 2, 2 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V50 tmp47 [V50 ] ( 2, 3 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
-; V51 cse0 [V51,T01] ( 5, 5 ) int -> rsi "CSE #01: moderate"
+; V19 tmp16 [V19,T13] ( 3, 3 ) simd16 -> [rbp-0x70] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[double]>
+; V20 tmp17 [V20,T11] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[double]>
+;* V21 tmp18 [V21 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V22 tmp19 [V22 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[double]>
+; V23 tmp20 [V23 ] ( 2, 2 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V24 tmp21 [V24,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 2, 4 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V26 tmp23 [V26 ] ( 0, 0 ) double -> zero-ref "Inline stloc first use temp"
+; V27 tmp24 [V27 ] ( 2, 2 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V28 tmp25 [V28,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V29 tmp26 [V29 ] ( 2, 4 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[double]>
+;* V30 tmp27 [V30 ] ( 0, 0 ) double -> zero-ref "Inline stloc first use temp"
+; V31 tmp28 [V31,T14] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[double]>
+; V32 tmp29 [V32,T02] ( 2, 2 ) long -> r14 "field V07._00 (fldOffset=0x0)" P-INDEP
+; V33 tmp30 [V33,T03] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+; V34 tmp31 [V34 ] ( 2, 2 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V35 tmp32 [V35 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
+; V36 tmp33 [V36 ] ( 2, 2 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V37 tmp34 [V37 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
+; V38 tmp35 [V38,T04] ( 2, 2 ) long -> r14 "field V21._00 (fldOffset=0x0)" P-INDEP
+; V39 tmp36 [V39,T05] ( 2, 2 ) long -> rdi "field V22._00 (fldOffset=0x0)" P-INDEP
+; V40 tmp37 [V40 ] ( 2, 2 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V41 tmp38 [V41 ] ( 2, 3 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V42 tmp39 [V42 ] ( 2, 2 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
+; V43 tmp40 [V43 ] ( 2, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V29._00 (fldOffset=0x0)" P-DEP
;
-; Lcl frame size = 128
+; Lcl frame size = 184
G_M32972_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
- vmovups ymm0, ymmword ptr [rbp+0x10]
- ;; size=21 bbWeight=1 PerfScore 5.75
+ push r15
+ push r14
+ push rbx
+ sub rsp, 184
+ lea rbp, [rsp+0xD0]
+ mov r15, rdi
+ mov ebx, esi
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ ;; size=31 bbWeight=1 PerfScore 9.25
G_M32972_IG02:
- vmovaps ymm1, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm1
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x20], rax
- mov rax, qword ptr [rbp-0x20]
- and esi, 63
- sarx rax, rax, rsi
- mov qword ptr [rbp-0x18], rax
- mov rax, qword ptr [rbp-0x18]
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x30], rcx
- mov rcx, qword ptr [rbp-0x30]
- sarx rcx, rcx, rsi
- mov qword ptr [rbp-0x28], rcx
- mov rcx, qword ptr [rbp-0x28]
- mov qword ptr [rbp-0x40], rax
- mov qword ptr [rbp-0x38], rcx
- vmovaps xmm1, xmmword ptr [rbp-0x40]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- mov rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x60], rax
- mov rax, qword ptr [rbp-0x60]
- sarx rax, rax, rsi
- mov qword ptr [rbp-0x58], rax
- mov rax, qword ptr [rbp-0x58]
- mov rcx, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x70], rcx
- mov rcx, qword ptr [rbp-0x70]
- sarx rcx, rcx, rsi
- mov qword ptr [rbp-0x68], rcx
- mov rcx, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0x80], rax
- mov qword ptr [rbp-0x78], rcx
- vinsertf128 ymm0, ymm1, xmmword ptr [rbp-0x80], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=158 bbWeight=1 PerfScore 39.75
+ vmovups ymmword ptr [rbp+0x10], ymm1
+ vmovaps ymm0, ymm1
+ vmovaps xmmword ptr [rbp-0x40], xmm0
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rdi
+ mov rdi, qword ptr [rbp-0x50]
+ sarx rdi, rdi, rbx
+ mov qword ptr [rbp-0x48], rdi
+ mov r14, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x60], rdi
+ vmovsd xmm0, qword ptr [rbp-0x60]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[double]:ShiftRightArithmetic(double,int):double
+ call [rax]System.Runtime.Intrinsics.Scalar`1[double]:ShiftRightArithmetic(double,int):double
+ vmovsd qword ptr [rbp-0x58], xmm0
+ mov rdi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], r14
+ mov qword ptr [rbp-0x68], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x70]
+ vmovaps xmmword ptr [rbp-0x30], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x80], xmm1
+ mov rdi, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0x90], rdi
+ vmovsd xmm0, qword ptr [rbp-0x90]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[double]:ShiftRightArithmetic(double,int):double
+ call [rax]System.Runtime.Intrinsics.Scalar`1[double]:ShiftRightArithmetic(double,int):double
+ vmovsd qword ptr [rbp-0x88], xmm0
+ mov r14, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xA0], rdi
+ vmovsd xmm0, qword ptr [rbp-0xA0]
+ mov edi, ebx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[double]:ShiftRightArithmetic(double,int):double
+ call [rax]System.Runtime.Intrinsics.Scalar`1[double]:ShiftRightArithmetic(double,int):double
+ vmovsd qword ptr [rbp-0x98], xmm0
+ mov rdi, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], r14
+ mov qword ptr [rbp-0xA8], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x30]
+ vmovups xmmword ptr [rsp], xmm0
+ vmovaps xmm0, xmmword ptr [rbp-0xB0]
+ vmovups xmmword ptr [rsp+0x10], xmm0
+ mov rdi, r15
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Vector256:Create[double](System.Runtime.Intrinsics.Vector128`1[double],System.Runtime.Intrinsics.Vector128`1[double]):System.Runtime.Intrinsics.Vector256`1[double]
+ call [rax]System.Runtime.Intrinsics.Vector256:Create[double](System.Runtime.Intrinsics.Vector128`1[double],System.Runtime.Intrinsics.Vector128`1[double]):System.Runtime.Intrinsics.Vector256`1[double]
+ mov rax, r15
+ ;; size=261 bbWeight=1 PerfScore 66.00
G_M32972_IG03:
vzeroupper
- add rsp, 128
+ add rsp, 184
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 191, prolog size 16, PerfScore 48.25, instruction count 45, allocated bytes for code 191 (MethodHash=35477f33) for method System.Runtime.Intrinsics.Vector256`1[double]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[double],int):System.Runtime.Intrinsics.Vector256`1[double] (FullOpts)
+; Total bytes of code 309, prolog size 21, PerfScore 79.50, instruction count 66, allocated bytes for code 309 (MethodHash=35477f33) for method System.Runtime.Intrinsics.Vector256`1[double]:System.Runtime.Intrinsics.ISimdVector<System.Runtime.Intrinsics.Vector256<T>,T>.ShiftRightArithmetic(System.Runtime.Intrinsics.Vector256`1[double],int):System.Runtime.Intrinsics.Vector256`1[double] (FullOpts)117 (64.64 % of base) - System.Runtime.Intrinsics.Vector256:Multiply[long](long,System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long]
; Assembly listing for method System.Runtime.Intrinsics.Vector256:Multiply[long](long,System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long] (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Unix
; FullOpts code
; optimized code
; rbp based frame
; partially interruptible
; No PGO data
-; 0 inlinees with PGO data; 16 single block inlinees; 13 inlinees without PGO data
+; 0 inlinees with PGO data; 16 single block inlinees; 9 inlinees without PGO data
; Final local variable assignments
;
-; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> rdi single-def
-; V01 arg0 [V01,T00] ( 6, 6 ) long -> rsi single-def
-; V02 arg1 [V02,T15] ( 2, 2 ) simd32 -> mm0 single-def <System.Runtime.Intrinsics.Vector256`1[long]>
+; V00 RetBuf [V00,T01] ( 4, 4 ) byref -> r15 single-def
+; V01 arg0 [V01,T00] ( 6, 6 ) long -> rbx single-def
+; V02 arg1 [V02,T15] ( 2, 2 ) simd32 -> [rbp+0x10] single-def <System.Runtime.Intrinsics.Vector256`1[long]>
;# V03 OutArgs [V03 ] ( 1, 1 ) struct ( 0) [rsp+0x00] do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
-; V04 tmp1 [V04,T12] ( 2, 4 ) simd16 -> mm1 "impAppendStmt"
+; V04 tmp1 [V04,T12] ( 2, 4 ) simd16 -> [rbp-0x30] spill-single-def "impAppendStmt"
;* V05 tmp2 [V05 ] ( 0, 0 ) simd16 -> zero-ref "spilled call-like call argument"
-; V06 tmp3 [V06,T10] ( 3, 6 ) simd16 -> [rbp-0x10] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V06 tmp3 [V06,T10] ( 3, 6 ) simd16 -> [rbp-0x40] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
;* V07 tmp4 [V07 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V08 tmp5 [V08 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V09 tmp6 [V09 ] ( 2, 2 ) struct ( 8) [rbp-0x18] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V09 tmp6 [V09 ] ( 2, 2 ) struct ( 8) [rbp-0x48] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V10 tmp7 [V10,T06] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V11 tmp8 [V11 ] ( 2, 4 ) struct ( 8) [rbp-0x20] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V11 tmp8 [V11 ] ( 2, 4 ) struct ( 8) [rbp-0x50] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V12 tmp9 [V12 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V13 tmp10 [V13 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V14 tmp11 [V14 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V15 tmp12 [V15 ] ( 2, 2 ) struct ( 8) [rbp-0x28] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V16 tmp13 [V16,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V17 tmp14 [V17 ] ( 2, 4 ) struct ( 8) [rbp-0x30] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V18 tmp15 [V18 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V19 tmp16 [V19 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V20 tmp17 [V20 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V21 tmp18 [V21,T13] ( 3, 3 ) simd16 -> [rbp-0x40] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-; V22 tmp19 [V22,T11] ( 3, 6 ) simd16 -> [rbp-0x50] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
-;* V23 tmp20 [V23 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V24 tmp21 [V24 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
-; V25 tmp22 [V25 ] ( 2, 2 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V26 tmp23 [V26,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V27 tmp24 [V27 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V13 tmp10 [V13 ] ( 2, 2 ) struct ( 8) [rbp-0x58] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V14 tmp11 [V14,T07] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V15 tmp12 [V15 ] ( 2, 4 ) struct ( 8) [rbp-0x60] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V16 tmp13 [V16 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V17 tmp14 [V17,T13] ( 3, 3 ) simd16 -> [rbp-0x70] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+; V18 tmp15 [V18,T11] ( 3, 6 ) simd16 -> [rbp-0x80] do-not-enreg[SF] "Inlining Arg" <System.Runtime.Intrinsics.Vector128`1[long]>
+;* V19 tmp16 [V19 ] ( 0, 0 ) struct ( 8) zero-ref "impAppendStmt" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V20 tmp17 [V20 ] ( 0, 0 ) struct ( 8) zero-ref "spilled call-like call argument" <System.Runtime.Intrinsics.Vector64`1[long]>
+; V21 tmp18 [V21 ] ( 2, 2 ) struct ( 8) [rbp-0x88] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V22 tmp19 [V22,T08] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V23 tmp20 [V23 ] ( 2, 4 ) struct ( 8) [rbp-0x90] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V24 tmp21 [V24 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
+; V25 tmp22 [V25 ] ( 2, 2 ) struct ( 8) [rbp-0x98] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
+;* V26 tmp23 [V26,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
+; V27 tmp24 [V27 ] ( 2, 4 ) struct ( 8) [rbp-0xA0] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
;* V28 tmp25 [V28 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V29 tmp26 [V29 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V30 tmp27 [V30 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V31 tmp28 [V31 ] ( 2, 2 ) struct ( 8) [rbp-0x68] do-not-enreg[XS] addr-exposed ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V32 tmp29 [V32,T09] ( 0, 0 ) int -> zero-ref "Inline stloc first use temp"
-; V33 tmp30 [V33 ] ( 2, 4 ) struct ( 8) [rbp-0x70] do-not-enreg[XS] addr-exposed ld-addr-op "Inlining Arg" <System.Runtime.Intrinsics.Vector64`1[long]>
-;* V34 tmp31 [V34 ] ( 0, 0 ) long -> zero-ref "Inline stloc first use temp"
-;* V35 tmp32 [V35 ] ( 0, 0 ) long -> zero-ref "Inline return value spill temp"
-;* V36 tmp33 [V36 ] ( 0, 0 ) long -> zero-ref "Inlining Arg"
-; V37 tmp34 [V37,T14] ( 3, 3 ) simd16 -> [rbp-0x80] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
-;* V38 tmp35 [V38 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[long]>
-; V39 tmp36 [V39,T02] ( 2, 2 ) long -> rax "field V07._00 (fldOffset=0x0)" P-INDEP
-; V40 tmp37 [V40,T03] ( 2, 2 ) long -> rcx "field V08._00 (fldOffset=0x0)" P-INDEP
-; V41 tmp38 [V41 ] ( 2, 2 ) long -> [rbp-0x18] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
-; V42 tmp39 [V42 ] ( 2, 3 ) long -> [rbp-0x20] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
-; V43 tmp40 [V43 ] ( 2, 2 ) long -> [rbp-0x28] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
-; V44 tmp41 [V44 ] ( 2, 3 ) long -> [rbp-0x30] do-not-enreg[X] addr-exposed "field V17._00 (fldOffset=0x0)" P-DEP
-; V45 tmp42 [V45,T04] ( 2, 2 ) long -> rax "field V23._00 (fldOffset=0x0)" P-INDEP
-; V46 tmp43 [V46,T05] ( 2, 2 ) long -> rcx "field V24._00 (fldOffset=0x0)" P-INDEP
-; V47 tmp44 [V47 ] ( 2, 2 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
-; V48 tmp45 [V48 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
-; V49 tmp46 [V49 ] ( 2, 2 ) long -> [rbp-0x68] do-not-enreg[X] addr-exposed "field V31._00 (fldOffset=0x0)" P-DEP
-; V50 tmp47 [V50 ] ( 2, 3 ) long -> [rbp-0x70] do-not-enreg[X] addr-exposed "field V33._00 (fldOffset=0x0)" P-DEP
+; V29 tmp26 [V29,T14] ( 3, 3 ) simd16 -> [rbp-0xB0] do-not-enreg[SF] ld-addr-op "Inline ldloca(s) first use temp" <System.Runtime.Intrinsics.Vector128`1[long]>
+;* V30 tmp27 [V30 ] ( 0, 0 ) simd32 -> zero-ref "Inline return value spill temp" <System.Runtime.Intrinsics.Vector256`1[long]>
+; V31 tmp28 [V31,T02] ( 2, 2 ) long -> r14 "field V07._00 (fldOffset=0x0)" P-INDEP
+; V32 tmp29 [V32,T03] ( 2, 2 ) long -> rdi "field V08._00 (fldOffset=0x0)" P-INDEP
+; V33 tmp30 [V33 ] ( 2, 2 ) long -> [rbp-0x48] do-not-enreg[X] addr-exposed "field V09._00 (fldOffset=0x0)" P-DEP
+; V34 tmp31 [V34 ] ( 2, 3 ) long -> [rbp-0x50] do-not-enreg[X] addr-exposed "field V11._00 (fldOffset=0x0)" P-DEP
+; V35 tmp32 [V35 ] ( 2, 2 ) long -> [rbp-0x58] do-not-enreg[X] addr-exposed "field V13._00 (fldOffset=0x0)" P-DEP
+; V36 tmp33 [V36 ] ( 2, 3 ) long -> [rbp-0x60] do-not-enreg[X] addr-exposed "field V15._00 (fldOffset=0x0)" P-DEP
+; V37 tmp34 [V37,T04] ( 2, 2 ) long -> r14 "field V19._00 (fldOffset=0x0)" P-INDEP
+; V38 tmp35 [V38,T05] ( 2, 2 ) long -> rax "field V20._00 (fldOffset=0x0)" P-INDEP
+; V39 tmp36 [V39 ] ( 2, 2 ) long -> [rbp-0x88] do-not-enreg[X] addr-exposed "field V21._00 (fldOffset=0x0)" P-DEP
+; V40 tmp37 [V40 ] ( 2, 3 ) long -> [rbp-0x90] do-not-enreg[X] addr-exposed "field V23._00 (fldOffset=0x0)" P-DEP
+; V41 tmp38 [V41 ] ( 2, 2 ) long -> [rbp-0x98] do-not-enreg[X] addr-exposed "field V25._00 (fldOffset=0x0)" P-DEP
+; V42 tmp39 [V42 ] ( 2, 3 ) long -> [rbp-0xA0] do-not-enreg[X] addr-exposed "field V27._00 (fldOffset=0x0)" P-DEP
;
-; Lcl frame size = 128
+; Lcl frame size = 152
G_M40782_IG01:
push rbp
- sub rsp, 128
- lea rbp, [rsp+0x80]
+ push r15
+ push r14
+ push rbx
+ sub rsp, 152
+ lea rbp, [rsp+0xB0]
+ mov r15, rdi
+ mov rbx, rsi
vmovups ymm0, ymmword ptr [rbp+0x10]
- ;; size=21 bbWeight=1 PerfScore 5.75
+ ;; size=32 bbWeight=1 PerfScore 9.25
G_M40782_IG02:
+ vmovups ymmword ptr [rbp+0x10], ymm0
vmovaps ymm1, ymm0
- vmovaps xmmword ptr [rbp-0x10], xmm1
- mov rax, qword ptr [rbp-0x10]
- mov qword ptr [rbp-0x20], rax
- mov rax, rsi
- imul rax, qword ptr [rbp-0x20]
- mov qword ptr [rbp-0x18], rax
- mov rax, qword ptr [rbp-0x18]
- mov rcx, qword ptr [rbp-0x08]
- mov qword ptr [rbp-0x30], rcx
- mov rcx, rsi
- imul rcx, qword ptr [rbp-0x30]
- mov qword ptr [rbp-0x28], rcx
- mov rcx, qword ptr [rbp-0x28]
- mov qword ptr [rbp-0x40], rax
- mov qword ptr [rbp-0x38], rcx
- vmovaps xmm1, xmmword ptr [rbp-0x40]
- vextractf128 xmm0, ymm0, 1
- vmovaps xmmword ptr [rbp-0x50], xmm0
- mov rax, qword ptr [rbp-0x50]
- mov qword ptr [rbp-0x60], rax
- mov rax, rsi
- imul rax, qword ptr [rbp-0x60]
+ vmovaps xmmword ptr [rbp-0x40], xmm1
+ mov rdi, qword ptr [rbp-0x40]
+ mov qword ptr [rbp-0x50], rdi
+ mov rdi, qword ptr [rbp-0x50]
+ mov rsi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x48], rax
+ mov r14, qword ptr [rbp-0x48]
+ mov rdi, qword ptr [rbp-0x38]
+ mov qword ptr [rbp-0x60], rdi
+ mov rdi, qword ptr [rbp-0x60]
+ mov rsi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
mov qword ptr [rbp-0x58], rax
- mov rax, qword ptr [rbp-0x58]
- mov rcx, qword ptr [rbp-0x48]
- mov qword ptr [rbp-0x70], rcx
- imul rsi, qword ptr [rbp-0x70]
- mov qword ptr [rbp-0x68], rsi
- mov rcx, qword ptr [rbp-0x68]
- mov qword ptr [rbp-0x80], rax
- mov qword ptr [rbp-0x78], rcx
- vinserti128 ymm0, ymm1, xmmword ptr [rbp-0x80], 1
- vmovups ymmword ptr [rdi], ymm0
- mov rax, rdi
- ;; size=148 bbWeight=1 PerfScore 50.25
+ mov rdi, qword ptr [rbp-0x58]
+ mov qword ptr [rbp-0x70], r14
+ mov qword ptr [rbp-0x68], rdi
+ vmovaps xmm0, xmmword ptr [rbp-0x70]
+ vmovaps xmmword ptr [rbp-0x30], xmm0
+ vmovups ymm1, ymmword ptr [rbp+0x10]
+ vextractf128 xmm1, ymm1, 1
+ vmovaps xmmword ptr [rbp-0x80], xmm1
+ mov rdi, qword ptr [rbp-0x80]
+ mov qword ptr [rbp-0x90], rdi
+ mov rdi, qword ptr [rbp-0x90]
+ mov rsi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x88], rax
+ mov r14, qword ptr [rbp-0x88]
+ mov rdi, qword ptr [rbp-0x78]
+ mov qword ptr [rbp-0xA0], rdi
+ mov rdi, qword ptr [rbp-0xA0]
+ mov rsi, rbx
+ mov rax, 0xD1FFAB1E ; code for System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ call [rax]System.Runtime.Intrinsics.Scalar`1[long]:Multiply(long,long):long
+ mov qword ptr [rbp-0x98], rax
+ mov rax, qword ptr [rbp-0x98]
+ mov qword ptr [rbp-0xB0], r14
+ mov qword ptr [rbp-0xA8], rax
+ vmovaps xmm0, xmmword ptr [rbp-0x30]
+ vinserti128 ymm0, ymm0, xmmword ptr [rbp-0xB0], 1
+ vmovups ymmword ptr [r15], ymm0
+ mov rax, r15
+ ;; size=249 bbWeight=1 PerfScore 60.50
G_M40782_IG03:
vzeroupper
- add rsp, 128
+ add rsp, 152
+ pop rbx
+ pop r14
+ pop r15
pop rbp
ret
- ;; size=12 bbWeight=1 PerfScore 2.75
+ ;; size=17 bbWeight=1 PerfScore 4.25
-; Total bytes of code 181, prolog size 16, PerfScore 58.75, instruction count 43, allocated bytes for code 181 (MethodHash=2fe960b1) for method System.Runtime.Intrinsics.Vector256:Multiply[long](long,System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long] (FullOpts)
+; Total bytes of code 298, prolog size 21, PerfScore 74.00, instruction count 64, allocated bytes for code 298 (MethodHash=2fe960b1) for method System.Runtime.Intrinsics.Vector256:Multiply[long](long,System.Runtime.Intrinsics.Vector256`1[long]):System.Runtime.Intrinsics.Vector256`1[long] (FullOpts)