[X86] getFauxShuffleMask - handle insert_subvector(src, bitcast(extract_subvector(sub))) patterns

Add bitcast handling to the existing insert_subvector(src, extract_subvector(sub)) pattern, and recognise undef src cases to allow us to detect vector widening patterns.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -5646,17 +5646,28 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
     unsigned NumSubElts = SubVT.getVectorNumElements();
     if (!N->isOnlyUserOf(Sub.getNode()))
       return false;
+    SDValue SubBC = peekThroughBitcasts(Sub);
     uint64_t InsertIdx = N.getConstantOperandVal(2);
     // Handle INSERT_SUBVECTOR(SRC0, EXTRACT_SUBVECTOR(SRC1)).
-    if (Sub.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
-        Sub.getOperand(0).getValueType() == VT) {
-      uint64_t ExtractIdx = Sub.getConstantOperandVal(1);
-      for (int i = 0; i != (int)NumElts; ++i)
-        Mask.push_back(i);
+    if (SubBC.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+        SubBC.getOperand(0).getValueSizeInBits() == NumSizeInBits) {
+      uint64_t ExtractIdx = SubBC.getConstantOperandVal(1);
+      SDValue SubBCSrc = SubBC.getOperand(0);
+      unsigned NumSubSrcBCElts = SubBCSrc.getValueType().getVectorNumElements();
+      unsigned MaxElts = std::max(NumElts, NumSubSrcBCElts);
+      assert((MaxElts % NumElts) == 0 && (MaxElts % NumSubSrcBCElts) == 0 &&
+             "Subvector valuetype mismatch");
+      InsertIdx *= (MaxElts / NumElts);
+      ExtractIdx *= (MaxElts / NumSubSrcBCElts);
+      NumSubElts *= (MaxElts / NumElts);
+      bool SrcIsUndef = Src.isUndef();
+      for (int i = 0; i != (int)MaxElts; ++i)
+        Mask.push_back(SrcIsUndef ? SM_SentinelUndef : i);
       for (int i = 0; i != (int)NumSubElts; ++i)
-        Mask[InsertIdx + i] = NumElts + ExtractIdx + i;
-      Ops.push_back(Src);
-      Ops.push_back(Sub.getOperand(0));
+        Mask[InsertIdx + i] = (SrcIsUndef ? 0 : MaxElts) + ExtractIdx + i;
+      if (!SrcIsUndef)
+        Ops.push_back(Src);
+      Ops.push_back(SubBCSrc);
       return true;
     }
     // Handle INSERT_SUBVECTOR(SRC0, SHUFFLE(SRC1)).

llvm/test/CodeGen/X86/vector-trunc-packus.ll

@@ -873,9 +873,8 @@ define <2 x i16> @trunc_packus_v2i64_v2i16(<2 x i64> %a0) {
 ; SSE41-NEXT:    pand %xmm5, %xmm0
 ; SSE41-NEXT:    por %xmm4, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
-; SSE41-NEXT:    packusdw %xmm1, %xmm1
-; SSE41-NEXT:    packusdw %xmm1, %xmm1
-; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
+; SSE41-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_packus_v2i64_v2i16:
@@ -887,21 +886,32 @@ define <2 x i16> @trunc_packus_v2i64_v2i16(<2 x i64> %a0) {
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: trunc_packus_v2i64_v2i16:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [65535,65535]
-; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
-; AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm1
-; AVX2-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX2-NEXT:    retq
+; AVX2-SLOW-LABEL: trunc_packus_v2i64_v2i16:
+; AVX2-SLOW:       # %bb.0:
+; AVX2-SLOW-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [65535,65535]
+; AVX2-SLOW-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
+; AVX2-SLOW-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
+; AVX2-SLOW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-SLOW-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm1
+; AVX2-SLOW-NEXT:    vpand %xmm0, %xmm1, %xmm0
+; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
+; AVX2-SLOW-NEXT:    retq
+;
+; AVX2-FAST-LABEL: trunc_packus_v2i64_v2i16:
+; AVX2-FAST:       # %bb.0:
+; AVX2-FAST-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [65535,65535]
+; AVX2-FAST-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
+; AVX2-FAST-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
+; AVX2-FAST-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-FAST-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm1
+; AVX2-FAST-NEXT:    vpand %xmm0, %xmm1, %xmm0
+; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,1,8,9,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX2-FAST-NEXT:    retq
 ;
 ; AVX512F-LABEL: trunc_packus_v2i64_v2i16:
 ; AVX512F:       # %bb.0: