CPS Call Proposal Examples

1_cps_call.ll

;;;;;
; (1) a complete example
;;;;;

define cc18  {i8**, i64} @foo (i8** %sp, i64 %x) noinline {
  store i8* blockaddress (@foo, %retpt), i8** %sp
  %retV = call cc18 {i8**, i64} @bar(i8** %sp, i64 %x) ; the 'cps' call
  br label %retpt

retpt:
  ; load parameters of this block
  %newSP.x = extractvalue {i8**, i64} %retV, 1
  %y.x = extractvalue {i8**, i64} %retV, 0
  ;;;;;
  %y = ptrtoint i8** %y.x to i64
  %newSP = inttoptr i64 %newSP.x to i8**
  ;;;;;
  ; perform a CPS return from @foo
  %retAddr = load i8*, i8** %newSP
  %retAddr1 = bitcast i8* %retAddr to {i8**, i64} (i8**, i64)*
  %dummy = musttail call cc18  {i8**, i64} %retAddr1 (i8** %newSP, i64 %y)
  ret {i8**, i64} %dummy

}

define cc18  {i8**, i64} @bar (i8** %sp, i64 %y) noinline {
  %y1 = add i64 %y, 1
  ; perform a CPS return from @bar 
  %retAddr = load i8*, i8** %sp
  %retAddr1 = bitcast i8* %retAddr to {i8**, i64} (i8**, i64)*
  ;;;;;
  ; for fun, move SP to another register when returning
  %sp.1 = ptrtoint i8** %sp to i64
  %y1.1 = inttoptr i64 %y1 to i8**
  ;;;;;
  %dummy = musttail call cc18  {i8**, i64} %retAddr1 (i8** %y1.1, i64 %sp.1)
  ret {i8**, i64} %dummy
}

2_isel_idea.ll

+++++++++++++++++++++++++++++++++
+ What ISEL would produce for @foo with a custom lowering of the call.
+ the important parts are blocked off with plus symbols
+ (note: the numbers are messed up but this is the basic idea)
+++++++++++++++++++++++++++++++++

=== foo
Initial selection DAG: BB#0 'foo:'
SelectionDAG has 29 nodes:
  t0: ch = EntryToken
  t2: i64,ch = CopyFromReg t0, Register:i64 %vreg2
  t6: i64 = Constant<0>
  t9: i64 = GlobalAddress<{ i8**, i64 } (i8**, i64)* @bar> 0
      t8: ch = store<ST8[%sp]> t0, BlockAddress:i64<@foo, %retpt> 0, t2, undef:i64
    t11: ch,glue = callseq_start t8, TargetConstant:i64<0>
  t13: ch,glue = CopyToReg t11, Register:i64 %RSI, t2
    t4: i64,ch = CopyFromReg t0, Register:i64 %vreg3
  t15: ch,glue = CopyToReg t13, Register:i64 %R11, t4, t13:1
  ++++++++++++++++++++++++ now a tail call ++++++++++++++++++++++++
  t19: ch,glue = callseq_end t18, TargetConstant:i64<0>, TargetConstant:i64<0>, t18:1
  t18: ch,glue = X86ISD::TC_RETURN t15, TargetGlobalAddress:i64<{ i8**, i64 } (i8**, i64)* @bar> 0, Register:i64 %RSI, Register:i64 %R11, RegisterMask:Untyped, t15:1
  +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
  
  
Initial selection DAG: BB#1 'foo:retpt'
SelectionDAG has 19 nodes:
  t0: ch = EntryToken
  ++++++++ these were sunk from BB#0, appearing after call to @bar ++++++++++
  t20: i64,ch,glue = CopyFromReg t19, Register:i64 %RSI, t19:1
  t21: i64,ch,glue = CopyFromReg t20:1, Register:i64 %R11, t20:2
  t22: i64,i64 = merge_values t20, t21
      t24: ch = CopyToReg t0, Register:i64 %vreg0, t22
      t26: ch = CopyToReg t0, Register:i64 %vreg1, t22:1
    t27: ch = TokenFactor t24, t26
  t28: ch = TokenFactor t27, t21:1
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
  t2: i64,ch = CopyFromReg t0, Register:i64 %vreg0
    t4: i64,ch = CopyFromReg t2:1, Register:i64 %vreg1
  t5: i64,i64 = merge_values t2, t4
  t6: i64 = Constant<0>
  t8: i64,ch = load<LD8[%newSP]> t0, t5:1, undef:i64
    t10: ch,glue = callseq_start t8:1, TargetConstant:i64<0>
  t12: ch,glue = CopyToReg t10, Register:i64 %RSI, t5:1
  t14: ch,glue = CopyToReg t12, Register:i64 %R11, t5, t12:1
  t15: ch,glue = callseq_end t14, TargetConstant:i64<0>, TargetConstant:i64<0>, t14:1
  t18: ch,glue = X86ISD::TC_RETURN t15, t8, Constant:i32<0>, Register:i64 %RSI, Register:i64 %R11, RegisterMask:Untyped, t15:1

3_cps_call.s

######
# (3) this is currently produced by llc (with a custom cc18 convention patch) for the cps_call.ll example.
# I have noted the areas that should fix this output using the isel idea.
######

  .section	__TEXT,__text,regular,pure_instructions
	.macosx_version_min 10, 11
	.globl	_foo
	.align	4, 0x90
_foo:                                   ## @foo
	.cfi_startproc
## BB#0:
	pushq	%rax
Ltmp0:
	.cfi_def_cfa_offset 16
	leaq	Ltmp1(%rip), %rax
	movq	%rax, (%rsi)
	callq	_bar             # <-- will instead lower to:  jmp _bar
	movq	%rsi, %rax	 # <-- will be sunk during isel
	
Ltmp1:                                  ## Block address taken
## BB#1:                                ## %retpt
	movq	(%r11), %rcx
	movq	%r11, %rsi
	movq	%rax, %r11
	popq	%rax
	jmpq	*%rcx                   ## TAILCALL
	.cfi_endproc

	.globl	_bar
	.align	4, 0x90
_bar:                                   ## @bar
	.cfi_startproc
## BB#0:
	pushq	%rax
Ltmp2:
	.cfi_def_cfa_offset 16
	movq	%rsi, %rax
	leaq	1(%r11), %rsi
	movq	(%rax), %rcx
	movq	%rax, %r11
	popq	%rax
	jmpq	*%rcx                   ## TAILCALL
	.cfi_endproc


.subsections_via_symbols