[X86] combineConcatVectorOps - IsConcatFree must check extraction index

Identified in the regression reported by @alexfh on rGb5d7beeb9792 - IsConcatFree wasn't ensuring the subvector extraction index matched the position it would be concatenated back into.

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 92ee38e..a124d14 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -53613,11 +53613,16 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
         Subs.push_back(SubOp.getOperand(I));
       return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Subs);
     };
-    auto IsConcatFree = [](MVT VT, ArrayRef<SDValue> SubOps, unsigned I) {
-      return all_of(SubOps, [VT, I](SDValue Sub) {
-        return Sub.getOperand(I).getOpcode() == ISD::EXTRACT_SUBVECTOR &&
-               Sub.getOperand(I).getOperand(0).getValueType() == VT;
-      });
+    auto IsConcatFree = [](MVT VT, ArrayRef<SDValue> SubOps, unsigned Op) {
+      for (unsigned I = 0, E = SubOps.size(); I != E; ++I) {
+        SDValue Sub = SubOps[I].getOperand(Op);
+        unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
+        if (Sub.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
+            Sub.getOperand(0).getValueType() != VT ||
+            Sub.getConstantOperandAPInt(1) != (I * NumSubElts))
+          return false;
+      }
+      return true;
     };
 
     unsigned NumOps = Ops.size();

diff --git a/llvm/test/CodeGen/X86/vselect-avx.ll b/llvm/test/CodeGen/X86/vselect-avx.ll
index 4f0e7b8..f6f4e22 100644 (file)
--- a/llvm/test/CodeGen/X86/vselect-avx.ll
+++ b/llvm/test/CodeGen/X86/vselect-avx.ll
@@ -214,3 +214,66 @@ entry:
   store <8 x i32> %7, ptr undef
   ret void
 }
+
+; Regression test for rGb5d7beeb9792
+define void @vselect_concat_splat() {
+; AVX1-LABEL: vselect_concat_splat:
+; AVX1:       ## %bb.0: ## %entry
+; AVX1-NEXT:    vmovups (%rax), %xmm0
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[0,3,2,1]
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,0,3,2]
+; AVX1-NEXT:    vmovups 16, %xmm2
+; AVX1-NEXT:    vmovups 32, %xmm3
+; AVX1-NEXT:    vblendps {{.*#+}} xmm4 = mem[0],xmm3[1],mem[2,3]
+; AVX1-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm2[2],xmm4[3]
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm4 = xmm4[0,3,2,1]
+; AVX1-NEXT:    vblendps {{.*#+}} xmm3 = mem[0,1],xmm3[2,3]
+; AVX1-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm3[1,2],xmm2[3]
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[1,0,3,2]
+; AVX1-NEXT:    vxorps %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vcmpneqps %xmm3, %xmm1, %xmm3
+; AVX1-NEXT:    vblendvps %xmm3, %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vblendvps %xmm3, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vmovups %xmm0, (%rax)
+; AVX1-NEXT:    vmovups %xmm1, (%rax)
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: vselect_concat_splat:
+; AVX2:       ## %bb.0: ## %entry
+; AVX2-NEXT:    vmovups (%rax), %ymm0
+; AVX2-NEXT:    vmovups (%rax), %xmm1
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm2 = [0,3,6,1]
+; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7]
+; AVX2-NEXT:    vpermps %ymm3, %ymm2, %ymm3
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm4 = [1,4,7,2]
+; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    vpermps %ymm0, %ymm4, %ymm0
+; AVX2-NEXT:    vmovups 0, %ymm1
+; AVX2-NEXT:    vmovups 32, %xmm5
+; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm1[0],ymm5[1],ymm1[2,3,4,5,6,7]
+; AVX2-NEXT:    vpermps %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm5[2,3],ymm1[4,5,6,7]
+; AVX2-NEXT:    vpermps %ymm1, %ymm4, %ymm1
+; AVX2-NEXT:    vxorps %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vcmpneqps %xmm4, %xmm3, %xmm4
+; AVX2-NEXT:    vblendvps %xmm4, %xmm2, %xmm3, %xmm2
+; AVX2-NEXT:    vblendvps %xmm4, %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vmovups %xmm0, (%rax)
+; AVX2-NEXT:    vmovups %xmm2, (%rax)
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+entry:
+  %wide.vec = load <12 x float>, ptr undef, align 1
+  %strided.vec = shufflevector <12 x float> %wide.vec, <12 x float> poison, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
+  %strided.vec29 = shufflevector <12 x float> %wide.vec, <12 x float> poison, <4 x i32> <i32 1, i32 4, i32 7, i32 10>
+  %wide.vec31 = load <12 x float>, ptr null, align 1
+  %strided.vec32 = shufflevector <12 x float> %wide.vec31, <12 x float> poison, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
+  %strided.vec33 = shufflevector <12 x float> %wide.vec31, <12 x float> poison, <4 x i32> <i32 1, i32 4, i32 7, i32 10>
+  %i = select i1 false, <4 x float> zeroinitializer, <4 x float> %strided.vec
+  %i1 = fcmp une <4 x float> %i, zeroinitializer
+  %i2 = select <4 x i1> %i1, <4 x float> %strided.vec32, <4 x float> %strided.vec
+  %.v = select <4 x i1> %i1, <4 x float> %strided.vec33, <4 x float> %strided.vec29
+  %.uncasted = shufflevector <4 x float> %i2, <4 x float> %.v, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  store <8 x float>  %.uncasted, ptr undef, align 1
+  ret void
+}