[X86] isShuffleFoldableLoad - only check that the SDValue has one use (#126900)

We don't need the entire load node to have oneuse, just the loaded value - prevents load chains from interfering with shuffle commutation

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -12480,7 +12480,7 @@ static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
 /// This is particularly important because the set of instructions varies
 /// significantly based on whether the operand is a load or not.
 static bool isShuffleFoldableLoad(SDValue V) {
-  return V->hasOneUse() &&
+  return V.hasOneUse() &&
          ISD::isNON_EXTLoad(peekThroughOneUseBitcasts(V).getNode());
 }
 

llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast_from_memory.ll

@@ -1665,10 +1665,9 @@ define void @vec256_i32_widen_to_i128_factor4_broadcast_to_v2i128_factor2(ptr %i
 ;
 ; AVX512F-FAST-LABEL: vec256_i32_widen_to_i128_factor4_broadcast_to_v2i128_factor2:
 ; AVX512F-FAST:       # %bb.0:
-; AVX512F-FAST-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512F-FAST-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [8,1,2,3,8,5,6,7]
-; AVX512F-FAST-NEXT:    vpermi2d (%rdi), %ymm0, %ymm1
-; AVX512F-FAST-NEXT:    vpaddb (%rsi), %ymm1, %ymm0
+; AVX512F-FAST-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,9,10,11,0,13,14,15]
+; AVX512F-FAST-NEXT:    vpermd (%rdi), %zmm0, %zmm0
+; AVX512F-FAST-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
 ; AVX512F-FAST-NEXT:    vmovdqa %ymm0, (%rdx)
 ; AVX512F-FAST-NEXT:    vzeroupper
 ; AVX512F-FAST-NEXT:    retq
@@ -1684,10 +1683,9 @@ define void @vec256_i32_widen_to_i128_factor4_broadcast_to_v2i128_factor2(ptr %i
 ;
 ; AVX512DQ-FAST-LABEL: vec256_i32_widen_to_i128_factor4_broadcast_to_v2i128_factor2:
 ; AVX512DQ-FAST:       # %bb.0:
-; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512DQ-FAST-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [8,1,2,3,8,5,6,7]
-; AVX512DQ-FAST-NEXT:    vpermi2d (%rdi), %ymm0, %ymm1
-; AVX512DQ-FAST-NEXT:    vpaddb (%rsi), %ymm1, %ymm0
+; AVX512DQ-FAST-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,9,10,11,0,13,14,15]
+; AVX512DQ-FAST-NEXT:    vpermd (%rdi), %zmm0, %zmm0
+; AVX512DQ-FAST-NEXT:    vpaddb (%rsi), %ymm0, %ymm0
 ; AVX512DQ-FAST-NEXT:    vmovdqa %ymm0, (%rdx)
 ; AVX512DQ-FAST-NEXT:    vzeroupper
 ; AVX512DQ-FAST-NEXT:    retq
@@ -1703,10 +1701,9 @@ define void @vec256_i32_widen_to_i128_factor4_broadcast_to_v2i128_factor2(ptr %i
 ;
 ; AVX512BW-FAST-LABEL: vec256_i32_widen_to_i128_factor4_broadcast_to_v2i128_factor2:
 ; AVX512BW-FAST:       # %bb.0:
-; AVX512BW-FAST-NEXT:    vmovdqa 32(%rdi), %ymm0
-; AVX512BW-FAST-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [8,1,2,3,8,5,6,7]
-; AVX512BW-FAST-NEXT:    vpermi2d (%rdi), %ymm0, %ymm1
-; AVX512BW-FAST-NEXT:    vpaddb (%rsi), %zmm1, %zmm0
+; AVX512BW-FAST-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,9,10,11,0,13,14,15]
+; AVX512BW-FAST-NEXT:    vpermd (%rdi), %zmm0, %zmm0
+; AVX512BW-FAST-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
 ; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm0, (%rdx)
 ; AVX512BW-FAST-NEXT:    vzeroupper
 ; AVX512BW-FAST-NEXT:    retq

llvm/test/CodeGen/X86/vector-interleaved-load-i32-stride-6.ll

@@ -201,8 +201,8 @@ define void @load_i32_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
 ; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm4
 ; AVX512-FCP-NEXT:    vmovdqa 32(%rdi), %xmm5
 ; AVX512-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm3
-; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [3,5,0,0]
-; AVX512-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm6
+; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [7,1,0,0]
+; AVX512-FCP-NEXT:    vpermi2d %xmm4, %xmm5, %xmm6
 ; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm4 = [4,2,0,0]
 ; AVX512-FCP-NEXT:    vblendps {{.*#+}} ymm1 = mem[0,1,2,3],ymm1[4,5,6,7]
 ; AVX512-FCP-NEXT:    vpermps %ymm1, %ymm4, %ymm4
@@ -260,8 +260,8 @@ define void @load_i32_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
 ; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm4
 ; AVX512DQ-FCP-NEXT:    vmovdqa 32(%rdi), %xmm5
 ; AVX512DQ-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm3
-; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [3,5,0,0]
-; AVX512DQ-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm6
+; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [7,1,0,0]
+; AVX512DQ-FCP-NEXT:    vpermi2d %xmm4, %xmm5, %xmm6
 ; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm4 = [4,2,0,0]
 ; AVX512DQ-FCP-NEXT:    vblendps {{.*#+}} ymm1 = mem[0,1,2,3],ymm1[4,5,6,7]
 ; AVX512DQ-FCP-NEXT:    vpermps %ymm1, %ymm4, %ymm4
@@ -319,8 +319,8 @@ define void @load_i32_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
 ; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm4
 ; AVX512BW-FCP-NEXT:    vmovdqa 32(%rdi), %xmm5
 ; AVX512BW-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm3
-; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [3,5,0,0]
-; AVX512BW-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm6
+; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [7,1,0,0]
+; AVX512BW-FCP-NEXT:    vpermi2d %xmm4, %xmm5, %xmm6
 ; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm4 = [4,2,0,0]
 ; AVX512BW-FCP-NEXT:    vblendps {{.*#+}} ymm1 = mem[0,1,2,3],ymm1[4,5,6,7]
 ; AVX512BW-FCP-NEXT:    vpermps %ymm1, %ymm4, %ymm4
@@ -378,8 +378,8 @@ define void @load_i32_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm4
 ; AVX512DQ-BW-FCP-NEXT:    vmovdqa 32(%rdi), %xmm5
 ; AVX512DQ-BW-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm3
-; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [3,5,0,0]
-; AVX512DQ-BW-FCP-NEXT:    vpermi2d %xmm5, %xmm4, %xmm6
+; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm6 = [7,1,0,0]
+; AVX512DQ-BW-FCP-NEXT:    vpermi2d %xmm4, %xmm5, %xmm6
 ; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} xmm4 = [4,2,0,0]
 ; AVX512DQ-BW-FCP-NEXT:    vblendps {{.*#+}} ymm1 = mem[0,1,2,3],ymm1[4,5,6,7]
 ; AVX512DQ-BW-FCP-NEXT:    vpermps %ymm1, %ymm4, %ymm4

llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll

@@ -403,13 +403,12 @@ define void @PR39483() {
 ;
 ; X86-AVX512-LABEL: PR39483:
 ; X86-AVX512:       # %bb.0: # %entry
-; X86-AVX512-NEXT:    vmovups 0, %zmm0
-; X86-AVX512-NEXT:    vmovups 64, %ymm1
-; X86-AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [2,5,8,11,14,17,20,23]
-; X86-AVX512-NEXT:    vpermi2ps %zmm1, %zmm0, %zmm2
-; X86-AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; X86-AVX512-NEXT:    vmulps %ymm0, %ymm2, %ymm1
-; X86-AVX512-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; X86-AVX512-NEXT:    vmovups 64, %ymm0
+; X86-AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [18,21,24,27,30,1,4,7]
+; X86-AVX512-NEXT:    vpermt2ps 0, %zmm1, %zmm0
+; X86-AVX512-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; X86-AVX512-NEXT:    vmulps %ymm1, %ymm0, %ymm0
+; X86-AVX512-NEXT:    vaddps %ymm1, %ymm0, %ymm0
 ; X86-AVX512-NEXT:    vmovups %ymm0, (%eax)
 ;
 ; X64-AVX1-LABEL: PR39483:
@@ -444,13 +443,12 @@ define void @PR39483() {
 ;
 ; X64-AVX512-LABEL: PR39483:
 ; X64-AVX512:       # %bb.0: # %entry
-; X64-AVX512-NEXT:    vmovups 0, %zmm0
-; X64-AVX512-NEXT:    vmovups 64, %ymm1
-; X64-AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [2,5,8,11,14,17,20,23]
-; X64-AVX512-NEXT:    vpermi2ps %zmm1, %zmm0, %zmm2
-; X64-AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; X64-AVX512-NEXT:    vmulps %ymm0, %ymm2, %ymm1
-; X64-AVX512-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; X64-AVX512-NEXT:    vmovups 64, %ymm0
+; X64-AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [18,21,24,27,30,1,4,7]
+; X64-AVX512-NEXT:    vpermt2ps 0, %zmm1, %zmm0
+; X64-AVX512-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; X64-AVX512-NEXT:    vmulps %ymm1, %ymm0, %ymm0
+; X64-AVX512-NEXT:    vaddps %ymm1, %ymm0, %ymm0
 ; X64-AVX512-NEXT:    vmovups %ymm0, (%rax)
 entry:
   %wide.vec = load <24 x float>, ptr null, align 4

llvm/test/CodeGen/X86/vselect-avx.ll

@@ -377,14 +377,14 @@ define void @vselect_concat_splat() {
 ; AVX512-NEXT:    vmovaps %ymm2, %ymm3
 ; AVX512-NEXT:    vpermi2ps %ymm1, %ymm0, %ymm3
 ; AVX512-NEXT:    vmovups 32, %xmm4
-; AVX512-NEXT:    vmovups 0, %ymm5
-; AVX512-NEXT:    vxorps %xmm6, %xmm6, %xmm6
-; AVX512-NEXT:    vcmpneqps %xmm6, %xmm3, %k0
+; AVX512-NEXT:    vxorps %xmm5, %xmm5, %xmm5
+; AVX512-NEXT:    vcmpneqps %xmm5, %xmm3, %k0
 ; AVX512-NEXT:    kshiftlw $4, %k0, %k1
 ; AVX512-NEXT:    korw %k1, %k0, %k1
-; AVX512-NEXT:    vpermt2ps %ymm4, %ymm2, %ymm5
 ; AVX512-NEXT:    vpermt2ps %ymm1, %ymm2, %ymm0
-; AVX512-NEXT:    vmovaps %ymm5, %ymm0 {%k1}
+; AVX512-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [8,11,14,1,9,12,15,2]
+; AVX512-NEXT:    vpermi2ps 0, %ymm4, %ymm1
+; AVX512-NEXT:    vmovaps %ymm1, %ymm0 {%k1}
 ; AVX512-NEXT:    vmovups %ymm0, (%rax)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq