[PATCH] D89578: [CostModel] Return TCC_Expensive for non-speculatable ctlz/cttz.

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

This might be restoring the old behavior, but I want to make sure we are ok with the potential regressions. If not, we should make some adjustments to the x86 model first.

The SLP v4i32 diff with zero-is-undef set would be something like this, and it's hard to justify IMO:

  	bsfl	src32(%rip), %eax
  	bsfl	src32+4(%rip), %ecx
  	bsfl	src32+8(%rip), %edx
  	bsfl	src32+12(%rip), %esi
  	movl	%eax, dst32(%rip)
  	movl	%ecx, dst32+4(%rip)
  	movl	%edx, dst32+8(%rip)
  	movl	%esi, dst32+12(%rip)
  	retq

vs.

  	movdqa	src32(%rip), %xmm0
  	pcmpeqd	%xmm1, %xmm1
  	paddd	%xmm0, %xmm1
  	pandn	%xmm1, %xmm0
  	movdqa	%xmm0, %xmm1
  	psrlw	$1, %xmm1
  	pand	.LCPI0_0(%rip), %xmm1
  	psubb	%xmm1, %xmm0
  	movdqa	.LCPI0_1(%rip), %xmm1           # xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
  	movdqa	%xmm0, %xmm2
  	pand	%xmm1, %xmm2
  	psrlw	$2, %xmm0
  	pand	%xmm1, %xmm0
  	paddb	%xmm2, %xmm0
  	movdqa	%xmm0, %xmm1
  	psrlw	$4, %xmm1
  	paddb	%xmm0, %xmm1
  	pand	.LCPI0_2(%rip), %xmm1
  	pxor	%xmm0, %xmm0
  	movdqa	%xmm1, %xmm2
  	punpckhdq	%xmm0, %xmm2            # xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
  	psadbw	%xmm0, %xmm2
  	punpckldq	%xmm0, %xmm1            # xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
  	psadbw	%xmm0, %xmm1
  	packuswb	%xmm2, %xmm1
  	movdqa	%xmm1, dst32(%rip)
  	retq



================
Comment at: llvm/test/CodeGen/X86/dagcombine-select.ll:438-441
+; NOBMI-NEXT:    testl %edi, %edi
+; NOBMI-NEXT:    je .LBB26_2
+; NOBMI-NEXT:  # %bb.1: # %select.false.sink
+; NOBMI-NEXT:    bsfl %edi, %eax
----------------
I think we would view this as a regression based on https://bugs.llvm.org/PR46203 / 2328cab16ccd8f17fee782c29fb844662c089fbb

Do we need to adjust the isCheapToSpeculateXXXX APIs to acknowledge the zero-is-undef parameter of the intrinsic?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D89578/new/

https://reviews.llvm.org/D89578