[X86] InstCombine fails to replace shuffled _mm_blendv_epi8+icmp with select+icmp

[RKSimon]

InstCombine has x86 folds that can recognise when a _mm_blendv_epi8 intrinsic mask comes from a comparison and converts it to a select, even through all the bitcasts due to the __m128i type:

#include <x86intrin.h>

__m128i select_m128i(__m128i a, __m128i b, __m128i c, __m128i d)
{
    __m128i mask = (__m128i)((__v16qi)c < (__v16qi)d);
  return _mm_blendv_epi8( a, b, mask );
}

define <2 x i64> @select_m128i(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i64> %d) {
entry:
  %0 = bitcast <2 x i64> %c to <16 x i8>
  %1 = bitcast <2 x i64> %d to <16 x i8>
  %cmp = icmp slt <16 x i8> %0, %1
  %2 = bitcast <2 x i64> %a to <16 x i8>
  %3 = bitcast <2 x i64> %b to <16 x i8>
  %4 = select <16 x i1> %cmp, <16 x i8> %3, <16 x i8> %2
  %5 = bitcast <16 x i8> %4 to <2 x i64>
  ret <2 x i64> %5
}

But this fails entirely for AVX1 targets that try to emulate the 256-bit equivalent by splitting/concatenation:

__m256i select_m256i(__m256i a, __m256i b, __m256i c, __m256i d)
{
    __m256i mask = (__m256i)((__v32qi)c < (__v32qi)d);
  __m128i lo = _mm_blendv_epi8( _mm256_extractf128_si256( a, 0 ), _mm256_extractf128_si256( b, 0 ), _mm256_extractf128_si256( mask, 0 ) );
  __m128i hi = _mm_blendv_epi8( _mm256_extractf128_si256( a, 1 ), _mm256_extractf128_si256( b, 1 ), _mm256_extractf128_si256( mask, 1 ) );

#ifdef ALT
    return _mm256_setr_m128i( lo, hi );
#else
  return _mm256_insertf128_si256( _mm256_castsi128_si256( lo ), hi, 1 );
#endif
}

define <4 x i64> @select_m256i(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {
entry:
  %0 = bitcast <4 x i64> %c to <32 x i8>
  %1 = bitcast <4 x i64> %d to <32 x i8>
  %cmp = icmp slt <32 x i8> %0, %1
  %sext = sext <32 x i1> %cmp to <32 x i8>
  %2 = bitcast <4 x i64> %a to <8 x i32>
  %extract = shufflevector <8 x i32> %2, <8 x i32> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %3 = bitcast <4 x i64> %b to <8 x i32>
  %extract1 = shufflevector <8 x i32> %3, <8 x i32> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <32 x i8> %sext to <8 x i32>
  %extract2 = shufflevector <8 x i32> %4, <8 x i32> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %5 = bitcast <4 x i32> %extract to <16 x i8>
  %6 = bitcast <4 x i32> %extract1 to <16 x i8>
  %7 = bitcast <4 x i32> %extract2 to <16 x i8>
  %8 = tail call <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8> %5, <16 x i8> %6, <16 x i8> %7)
  %9 = bitcast <16 x i8> %8 to <2 x i64>
  %extract3 = shufflevector <8 x i32> %2, <8 x i32> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
  %extract4 = shufflevector <8 x i32> %3, <8 x i32> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
  %extract5 = shufflevector <8 x i32> %4, <8 x i32> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
  %10 = bitcast <4 x i32> %extract3 to <16 x i8>
  %11 = bitcast <4 x i32> %extract4 to <16 x i8>
  %12 = bitcast <4 x i32> %extract5 to <16 x i8>
  %13 = tail call <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8> %10, <16 x i8> %11, <16 x i8> %12)
  %shuffle.i = shufflevector <2 x i64> %9, <2 x i64> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
  %14 = bitcast <4 x i64> %shuffle.i to <8 x i32>
  %15 = bitcast <16 x i8> %13 to <4 x i32>
  %widen = shufflevector <4 x i32> %15, <4 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
  %insert = shufflevector <8 x i32> %14, <8 x i32> %widen, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>
  %16 = bitcast <8 x i32> %insert to <4 x i64>
  ret <4 x i64> %16
}
declare <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8>, <16 x i8>, <16 x i8>)

Goldbolt: https://godbolt.org/z/r61dhrqYM

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

define <4 x i64> @select_m256i(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {
entry:
  %0 = bitcast <4 x i64> %c to <32 x i8>
  %1 = bitcast <4 x i64> %d to <32 x i8>
  %cmp = icmp slt <32 x i8> %0, %1
  %sext = sext <32 x i1> %cmp to <32 x i8>
  %2 = bitcast <4 x i64> %a to <32 x i8>
  %3 = shufflevector <32 x i8> %2, <32 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <4 x i64> %b to <32 x i8>
  %5 = shufflevector <32 x i8> %4, <32 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = shufflevector <32 x i8> %sext, <32 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %7 = tail call <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8> %3, <16 x i8> %5, <16 x i8> %6)
  %8 = bitcast <4 x i64> %a to <32 x i8>
  %9 = shufflevector <32 x i8> %8, <32 x i8> poison, <16 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %10 = bitcast <4 x i64> %b to <32 x i8>
  %11 = shufflevector <32 x i8> %10, <32 x i8> poison, <16 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %12 = shufflevector <32 x i8> %sext, <32 x i8> poison, <16 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %13 = tail call <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8> %9, <16 x i8> %11, <16 x i8> %12)
  %14 = bitcast <16 x i8> %7 to <4 x i32>
  %15 = bitcast <16 x i8> %13 to <4 x i32>
  %insert = shufflevector <4 x i32> %14, <4 x i32> %15, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %16 = bitcast <8 x i32> %insert to <4 x i64>
  ret <4 x i64> %16
}