[X86][AVX] SimplifyDemandedVectorElts - handle extraction from X86ISD::SUBV_BROADCAST source (PR42819)

PR42819 showed an issue that we couldn't handle the case where we demanded a 'sub-sub-vector' of the SUBV_BROADCAST 'sub-vector' source.

This patch recognizes these cases and extracts the sub-sub-vector instead of trying to broadcast to a type smaller than the 'sub-vector' source. 

llvm-svn: 367306

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -34214,14 +34214,16 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
       // Subvector broadcast.
     case X86ISD::SUBV_BROADCAST: {
       SDLoc DL(Op);
-      SDValue Ext = Op.getOperand(0);
-      if (Ext.getValueSizeInBits() != ExtSizeInBits) {
-        MVT ExtSVT = Ext.getSimpleValueType().getScalarType();
-        MVT ExtVT =
-            MVT::getVectorVT(ExtSVT, ExtSizeInBits / ExtSVT.getSizeInBits());
-        Ext = TLO.DAG.getNode(X86ISD::SUBV_BROADCAST, DL, ExtVT, Ext);
-      }
-      return TLO.CombineTo(Op, insertSubVector(TLO.DAG.getUNDEF(VT), Ext, 0,
+      SDValue Src = Op.getOperand(0);
+      if (Src.getValueSizeInBits() > ExtSizeInBits)
+        Src = extractSubVector(Src, 0, TLO.DAG, DL, ExtSizeInBits);
+      else if (Src.getValueSizeInBits() < ExtSizeInBits) {
+        MVT SrcSVT = Src.getSimpleValueType().getScalarType();
+        MVT SrcVT =
+            MVT::getVectorVT(SrcSVT, ExtSizeInBits / SrcSVT.getSizeInBits());
+        Src = TLO.DAG.getNode(X86ISD::SUBV_BROADCAST, DL, SrcVT, Src);
+      }
+      return TLO.CombineTo(Op, insertSubVector(TLO.DAG.getUNDEF(VT), Src, 0,
                                                TLO.DAG, DL, ExtSizeInBits));
     }
       // Byte shifts by immediate.

llvm/test/CodeGen/X86/oddsubvector.ll

@@ -158,3 +158,35 @@ define void @PR40815(%struct.Mat4* nocapture readonly dereferenceable(64), %stru
   store <4 x float> %5, <4 x float>* %13, align 16
   ret void
 }
+
+define <16 x i32> @PR42819(<8 x i32>* %a0) {
+; SSE-LABEL: PR42819:
+; SSE:       # %bb.0:
+; SSE-NEXT:    movdqu (%rdi), %xmm3
+; SSE-NEXT:    pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,xmm3[0,1,2,3,4,5,6,7,8,9,10,11]
+; SSE-NEXT:    xorps %xmm0, %xmm0
+; SSE-NEXT:    xorps %xmm1, %xmm1
+; SSE-NEXT:    xorps %xmm2, %xmm2
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: PR42819:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = mem[0,0,1,2]
+; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm0[5,6,7]
+; AVX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
+;
+; AVX512-LABEL: PR42819:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vmovdqu (%rdi), %xmm0
+; AVX512-NEXT:    movw $-8192, %ax # imm = 0xE000
+; AVX512-NEXT:    kmovw %eax, %k1
+; AVX512-NEXT:    vpexpandd %zmm0, %zmm0 {%k1} {z}
+; AVX512-NEXT:    retq
+  %1 = load <8 x i32>, <8 x i32>* %a0, align 4
+  %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %3 = shufflevector <16 x i32> zeroinitializer, <16 x i32> %2, <16 x i32> <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, i32 16, i32 17, i32 18>
+  ret <16 x i32> %3
+}