[PATCH] D13988: [X86][SSE] Add general memory folding for (V)INSERTPS instruction

[Robert Lougher via llvm-commits]

rob.lougher added a comment.

Hi Simon,

We discovered a bug internally caused by the non-zeroing of the countS bits in the folding of the insertps load.  Although countS bits are ignored when loading from memory on insertps, we need to explicitly set them to 0 as another optimization may later "unfold" the load.  This is demonstrated by the following testcase (the checks are based on the RUN lines from the sse41.ll file).

define <4 x float> @foo(<4 x float>* %v0, <4 x float>* %v1) {
; X32-LABEL: foo:
; X32:       ## BB#0:
; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X32-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
; X32-NEXT:    movaps (%eax), %xmm0
; X32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
; X32-NEXT:    addps %xmm1, %xmm0
; X32-NEXT:    retl
;
; X64-LABEL: foo:
; X64:       ## BB#0:
; X64-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
; X64-NEXT:    movaps (%rdi), %xmm0
; X64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
; X64-NEXT:    addps %xmm1, %xmm0
; X64-NEXT:    retq

  %a = getelementptr inbounds <4 x float>, <4 x float>* %v1, i64 0, i64 1
  %b = load float, float* %a, align 4
  %c = insertelement <4 x float> undef, float %b, i32 0
  %d = load <4 x float>, <4 x float>* %v1, align 16
  %e = load <4 x float>, <4 x float>* %v0, align 16
  %f = shufflevector <4 x float> %e, <4 x float> %d, <4 x i32> <i32 0, i32 1, i32 2, i32 5>
  %g = fadd <4 x float> %c, %f
  ret <4 x float> %g

}

Another minor comment is that your change will do general memory load folding in addition to stack folding, but you've only got tests for stack folding.


Repository:
  rL LLVM

http://reviews.llvm.org/D13988