[WIP][X86] combineX86ShufflesRecursively - attempt to combine shuffles with larger types from EXTRACT_SUBVECTOR nodes

This replaces the rather limited combineX86ShuffleChainWithExtract function with handling for EXTRACT_SUBVECTOR node as we recurse down the shuffle chain, widening the shuffle mask to accommodate the larger value type.

This will mainly help AVX2/AVX512 cases with cross-lane shuffles, but it also helps collapse some cases where the same subvector has gotten reused in multiple lanes.

Exposed missing DemandedElts handling inside ISD::TRUNCATE nodes for ComputeNumSignBits

Author: RKSimon

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -5109,7 +5109,8 @@ unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, const APInt &DemandedElts,
   case ISD::TRUNCATE: {
     // Check if the sign bits of source go down as far as the truncated value.
     unsigned NumSrcBits = Op.getOperand(0).getScalarValueSizeInBits();
-    unsigned NumSrcSignBits = ComputeNumSignBits(Op.getOperand(0), Depth + 1);
+    unsigned NumSrcSignBits =
+        ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
     if (NumSrcSignBits > (NumSrcBits - VTBits))
       return NumSrcSignBits - (NumSrcBits - VTBits);
     break;

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -641,25 +641,16 @@ define void @vec128_i16_widen_to_i32_factor2_broadcast_to_v4i32_factor4(ptr %in.
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
-; AVX512BW-SLOW-LABEL: vec128_i16_widen_to_i32_factor2_broadcast_to_v4i32_factor4:
-; AVX512BW-SLOW:       # %bb.0:
-; AVX512BW-SLOW-NEXT:    vpmovsxbw {{.*#+}} xmm0 = [0,9,0,11,0,13,0,15]
-; AVX512BW-SLOW-NEXT:    vpermw (%rdi), %ymm0, %ymm0
-; AVX512BW-SLOW-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm0, (%rdx)
-; AVX512BW-SLOW-NEXT:    vzeroupper
-; AVX512BW-SLOW-NEXT:    retq
-;
-; AVX512BW-FAST-LABEL: vec128_i16_widen_to_i32_factor2_broadcast_to_v4i32_factor4:
-; AVX512BW-FAST:       # %bb.0:
-; AVX512BW-FAST-NEXT:    vpmovsxbw {{.*#+}} xmm0 = [0,9,0,11,0,13,6,7]
-; AVX512BW-FAST-NEXT:    vpermw (%rdi), %ymm0, %ymm0
-; AVX512BW-FAST-NEXT:    vpinsrw $6, (%rdi), %xmm0, %xmm0
-; AVX512BW-FAST-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6],mem[7]
-; AVX512BW-FAST-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
-; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm0, (%rdx)
-; AVX512BW-FAST-NEXT:    vzeroupper
-; AVX512BW-FAST-NEXT:    retq
+; AVX512BW-LABEL: vec128_i16_widen_to_i32_factor2_broadcast_to_v4i32_factor4:
+; AVX512BW:       # %bb.0:
+; AVX512BW-NEXT:    vpmovsxbw {{.*#+}} xmm0 = [16,25,16,27,16,29,0,23]
+; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm1
+; AVX512BW-NEXT:    vpermt2w (%rdi), %ymm0, %ymm1
+; AVX512BW-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5,6],mem[7]
+; AVX512BW-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
+; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
+; AVX512BW-NEXT:    vzeroupper
+; AVX512BW-NEXT:    retq
   %in.vec = load <64 x i8>, ptr %in.elt.ptr, align 64
   %in.vec.cast = bitcast <64 x i8> %in.vec to <32 x i16>
   %broadcast.of.zextinreg = shufflevector <32 x i16> %in.vec.cast, <32 x i16> poison, <8 x i32> <i32 0, i32 9, i32 0, i32 11, i32 0, i32 13, i32 0, i32 15>
@@ -735,25 +726,18 @@ define void @vec128_i16_widen_to_i64_factor4_broadcast_to_v2i64_factor2(ptr %in.
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
-; AVX512BW-SLOW-LABEL: vec128_i16_widen_to_i64_factor4_broadcast_to_v2i64_factor2:
-; AVX512BW-SLOW:       # %bb.0:
-; AVX512BW-SLOW-NEXT:    vpmovsxbw {{.*#+}} xmm0 = [0,9,10,11,0,13,6,7]
-; AVX512BW-SLOW-NEXT:    vpermw (%rdi), %ymm0, %ymm0
-; AVX512BW-SLOW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],mem[3]
-; AVX512BW-SLOW-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm0, (%rdx)
-; AVX512BW-SLOW-NEXT:    vzeroupper
-; AVX512BW-SLOW-NEXT:    retq
-;
-; AVX512BW-FAST-LABEL: vec128_i16_widen_to_i64_factor4_broadcast_to_v2i64_factor2:
-; AVX512BW-FAST:       # %bb.0:
-; AVX512BW-FAST-NEXT:    vpmovsxbw {{.*#+}} xmm0 = [0,9,10,11,0,5,6,7]
-; AVX512BW-FAST-NEXT:    vpermw (%rdi), %ymm0, %ymm0
-; AVX512BW-FAST-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],mem[5,6,7]
-; AVX512BW-FAST-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
-; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm0, (%rdx)
-; AVX512BW-FAST-NEXT:    vzeroupper
-; AVX512BW-FAST-NEXT:    retq
+; AVX512BW-LABEL: vec128_i16_widen_to_i64_factor4_broadcast_to_v2i64_factor2:
+; AVX512BW:       # %bb.0:
+; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX512BW-NEXT:    vmovdqa 16(%rdi), %xmm1
+; AVX512BW-NEXT:    vmovd %xmm0, %eax
+; AVX512BW-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3],xmm0[4,5,6,7]
+; AVX512BW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
+; AVX512BW-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm1[5,6,7]
+; AVX512BW-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
+; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
+; AVX512BW-NEXT:    vzeroupper
+; AVX512BW-NEXT:    retq
   %in.vec = load <64 x i8>, ptr %in.elt.ptr, align 64
   %in.vec.cast = bitcast <64 x i8> %in.vec to <32 x i16>
   %broadcast.of.zextinreg = shufflevector <32 x i16> %in.vec.cast, <32 x i16> poison, <8 x i32> <i32 0, i32 9, i32 10, i32 11, i32 0, i32 13, i32 14, i32 15>

llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast.ll

@@ -179,11 +179,11 @@ define <4 x i64> @f64to4sl(<4 x double> %a) {
 ; NODQ:       # %bb.0:
 ; NODQ-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; NODQ-NEXT:    vcvttsd2si %xmm1, %rax
-; NODQ-NEXT:    vmovq %rax, %xmm2
-; NODQ-NEXT:    vshufpd {{.*#+}} xmm1 = xmm1[1,0]
-; NODQ-NEXT:    vcvttsd2si %xmm1, %rax
 ; NODQ-NEXT:    vmovq %rax, %xmm1
-; NODQ-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
+; NODQ-NEXT:    vpermpd {{.*#+}} ymm2 = ymm0[3,2,2,3]
+; NODQ-NEXT:    vcvttsd2si %xmm2, %rax
+; NODQ-NEXT:    vmovq %rax, %xmm2
+; NODQ-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
 ; NODQ-NEXT:    vcvttsd2si %xmm0, %rax
 ; NODQ-NEXT:    vmovq %rax, %xmm2
 ; NODQ-NEXT:    vshufpd {{.*#+}} xmm0 = xmm0[1,0]

llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast_from_memory.ll

@@ -221,9 +221,10 @@ define double @fsub_noundef_ee (<8 x double> %x225, <8 x double> %x227) {
 ;
 ; SKX-LABEL: fsub_noundef_ee:
 ; SKX:       # %bb.0:
-; SKX-NEXT:    vextractf32x4 $2, %zmm1, %xmm0
-; SKX-NEXT:    vshufpd {{.*#+}} xmm1 = xmm0[1,0]
-; SKX-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
+; SKX-NEXT:    vmovapd {{.*#+}} xmm0 = [5,4]
+; SKX-NEXT:    vpermpd %zmm1, %zmm0, %zmm0
+; SKX-NEXT:    vextractf32x4 $2, %zmm1, %xmm1
+; SKX-NEXT:    vsubsd %xmm0, %xmm1, %xmm0
 ; SKX-NEXT:    vzeroupper
 ; SKX-NEXT:    retq
   %x226 = shufflevector <8 x double> %x225, <8 x double> %x227, <8 x i32> <i32 0, i32 8, i32 2, i32 10, i32 4, i32 12, i32 6, i32 14>