[X86] Delaying widening results in an unnecessary `vpmovsxwd` copy

[dzaima]

This C code:

__m256i iter(int16_t* src1p) {
    __m256i ten = _mm256_set1_epi32(10);
    __m256i wload = _mm256_cvtepi16_epi32(_mm_loadu_si128((void*)src1p));
    __m256i mask = _mm256_cmpgt_epi32(wload, ten);
    return _mm256_add_epi32(wload, mask);
}

compiled with -O3 -march=haswell, results in:

iter:
        vmovdqu xmm0, xmmword ptr [rdi]
        vpmovsxwd       ymm1, xmm0
        vpcmpgtw        xmm0, xmm0, xmmword ptr [rip + .LCPI0_0]
        vpmovsxwd       ymm0, xmm0
        vpaddd  ymm0, ymm0, ymm1
        ret

but it could be

iter:
        vpmovsxwd       ymm0, xmmword ptr [rdi]
        vpbroadcastd    ymm1, dword ptr [rip + .LCPI0_0]
        vpcmpgtd        ymm1, ymm0, ymm1
        vpaddd  ymm0, ymm1, ymm0
        ret

avoiding having two vpmovsxwds, and allowing the one that's left to have the memory operand inline.

https://godbolt.org/z/Ezrf9YbYn

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: dzaima (dzaima)

This C code: ```c __m256i iter(int16_t* src1p) { __m256i ten = _mm256_set1_epi32(10); __m256i wload = _mm256_cvtepi16_epi32(_mm_loadu_si128((void*)src1p)); __m256i mask = _mm256_cmpgt_epi32(wload, ten); return _mm256_add_epi32(wload, mask); } ``` compiled with `-O3 -march=haswell`, results in: ```asm iter: vmovdqu xmm0, xmmword ptr [rdi] vpmovsxwd ymm1, xmm0 vpcmpgtw xmm0, xmm0, xmmword ptr [rip + .LCPI0_0] vpmovsxwd ymm0, xmm0 vpaddd ymm0, ymm0, ymm1 ret ``` but it could be ```asm iter: vpmovsxwd ymm0, xmmword ptr [rdi] vpbroadcastd ymm1, dword ptr [rip + .LCPI0_0] vpcmpgtd ymm1, ymm0, ymm1 vpaddd ymm0, ymm1, ymm0 ret ``` avoiding having two `vpmovsxwd`s, and allowing the one that's left to have the memory operand inline.

https://godbolt.org/z/Ezrf9YbYn

[RKSimon]

This appears to be from an InstCcombine fold that doesn't account for the other use of the sext:

define noundef <4 x i64> @iter(ptr noundef readonly captures(none) %src1p) 
entry:
  %0 = load <2 x i64>, ptr %src1p, align 1
  %1 = bitcast <2 x i64> %0 to <8 x i16>
  %conv.i = sext <8 x i16> %1 to <8 x i32>
  %cmp.i = icmp sgt <8 x i32> %conv.i, splat (i32 10)
  %sext.i = sext <8 x i1> %cmp.i to <8 x i32>
  %add.i = add nsw <8 x i32> %conv.i, %sext.i
  %2 = bitcast <8 x i32> %add.i to <4 x i64>
  ret <4 x i64> %2
}

-->

define noundef <4 x i64> @iter(ptr noundef readonly captures(none) %src1p) 
entry:
  %0 = load <8 x i16>, ptr %src1p, align 1
  %conv.i = sext <8 x i16> %0 to <8 x i32>
  %cmp.i = icmp sgt <8 x i16> %0, splat (i16 10)
  %sext.i = sext <8 x i1> %cmp.i to <8 x i32>
  %add.i = add nsw <8 x i32> %conv.i, %sext.i
  %1 = bitcast <8 x i32> %add.i to <4 x i64>
  ret <4 x i64> %1
}