ASM for coeff_α(int,int) compared in Julia v1.4.2 (LLVM v8), Julia v1.5-beta1 (LLVM v9), and LLVM llc vs 8/9

gistfile1.txt

# Julia v1.4.2 (with LLVM v8):
        .text
        vcvtsi2sd       xmm0, xmm0, rdi
        movabs  rax, 139958522833784
        vcvtsi2sd       xmm1, xmm1, rsi
        vaddsd  xmm2, xmm0, xmm0
        vaddsd  xmm3, xmm2, qword ptr [rax]
        movabs  rax, 139958522833792
        vmulsd  xmm4, xmm0, xmm0
        vaddsd  xmm2, xmm2, qword ptr [rax]
        movabs  rax, 139958522833800
        vmulsd  xmm1, xmm1, xmm1
        vsubsd  xmm1, xmm4, xmm1
        vmulsd  xmm1, xmm2, xmm1
        vdivsd  xmm1, xmm3, xmm1
        vmovsd  xmm3, qword ptr [rax]   # xmm3 = mem[0],zero
        movabs  rax, 139958522833808
        vaddsd  xmm0, xmm0, xmm3
        vmulsd  xmm0, xmm0, xmm0
        vmulsd  xmm0, xmm0, qword ptr [rax]
        vaddsd  xmm0, xmm0, xmm3
        vmulsd  xmm0, xmm1, xmm0
        vsqrtsd xmm0, xmm0, xmm0
        ret
        nop

# Julia v1.5-beta1 (with LLVM v9):
# julia> @code_native syntax=:intel debuginfo=:none coeff_α(1,1)
        .text
        vcvtsi2sd       xmm0, xmm0, rdi
        movabs  rax, offset .rodata.cst8
        vcvtsi2sd       xmm1, xmm1, rsi
        vmulsd  xmm4, xmm0, xmm0
        vaddsd  xmm2, xmm0, xmm0
        vaddsd  xmm3, xmm2, qword ptr [rax]
        movabs  rax, 140666936433288
        vaddsd  xmm2, xmm2, qword ptr [rax]
        movabs  rax, 140666936433296
        vmulsd  xmm1, xmm1, xmm1
        vsubsd  xmm1, xmm4, xmm1
        vmovsd  xmm4, qword ptr [rax]   # xmm4 = mem[0],zero
        movabs  rax, 140666936433304
        vmulsd  xmm1, xmm2, xmm1
        vaddsd  xmm0, xmm0, xmm4
        vmulsd  xmm0, xmm0, xmm0
        vmulsd  xmm0, xmm0, qword ptr [rax]
        vaddsd  xmm0, xmm0, xmm4
        vmulsd  xmm0, xmm0, xmm3
        vdivsd  xmm0, xmm0, xmm1
        vsqrtsd xmm0, xmm0, xmm0
        ret
        nop

# Godbolt llc with LLVM v8 and v9
# llc -march=x86-64 -mcpu=sandybridge -O1
.LCPI0_0:
  .quad 4607182418800017408 # double 1
.LCPI0_1:
  .quad -4609434218613702656 # double -3
.LCPI0_2:
  .quad 4616189618054758400 # double 4
"julia_coeff_α_1367": # @"julia_coeff_\CE\B1_1367"
  push r14
  push rbx
  push rax
  mov r14, rsi
  mov rbx, rdi
  call julia.ptls_states
  vcvtsi2sd xmm0, xmm0, rbx
  vcvtsi2sd xmm1, xmm1, r14
  vaddsd xmm2, xmm0, xmm0
  vmovsd xmm3, qword ptr [rip + .LCPI0_0] # xmm3 = mem[0],zero
  vaddsd xmm4, xmm2, xmm3
  vaddsd xmm2, xmm2, qword ptr [rip + .LCPI0_1]
  vmulsd xmm5, xmm0, xmm0
  vmulsd xmm1, xmm1, xmm1
  vsubsd xmm1, xmm5, xmm1
  vmulsd xmm1, xmm2, xmm1
  vsubsd xmm0, xmm0, xmm3
  vmulsd xmm0, xmm0, xmm0
  vmulsd xmm0, xmm0, qword ptr [rip + .LCPI0_2]
  vdivsd xmm1, xmm4, xmm1
  vsubsd xmm0, xmm0, xmm3
  vmulsd xmm0, xmm1, xmm0
  vsqrtsd xmm0, xmm0, xmm0
  add rsp, 8
  pop rbx
  pop r14
  ret