[X86] combineConcatVectorOps - don't concat(vselect,vselect) if the concatenated...

One of the crash regression tests now exposes an existing issue with SelectionDAG::simplifySelect not folding vselect with constant masks

Fixes #59003

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a58430d..5f4d26c 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -54809,10 +54809,11 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
           IsConcatFree(VT, Ops, 1) && IsConcatFree(VT, Ops, 2)) {
         EVT SelVT = Ops[0].getOperand(0).getValueType();
         SelVT = SelVT.getDoubleNumVectorElementsVT(*DAG.getContext());
-        return DAG.getNode(Op0.getOpcode(), DL, VT,
-                           ConcatSubOperand(SelVT.getSimpleVT(), Ops, 0),
-                           ConcatSubOperand(VT, Ops, 1),
-                           ConcatSubOperand(VT, Ops, 2));
+        if (DAG.getTargetLoweringInfo().isTypeLegal(SelVT))
+          return DAG.getNode(Op0.getOpcode(), DL, VT,
+                             ConcatSubOperand(SelVT.getSimpleVT(), Ops, 0),
+                             ConcatSubOperand(VT, Ops, 1),
+                             ConcatSubOperand(VT, Ops, 2));
       }
       break;
     }

diff --git a/llvm/test/CodeGen/X86/vselect-avx.ll b/llvm/test/CodeGen/X86/vselect-avx.ll
index 69dfc01..8b08679 100644 (file)
--- a/llvm/test/CodeGen/X86/vselect-avx.ll
+++ b/llvm/test/CodeGen/X86/vselect-avx.ll
@@ -189,6 +189,28 @@ define <32 x i8> @PR22706(<32 x i1> %x) {
   ret <32 x i8> %tmp
 }
 
+; Don't concat select/blendv ops if the concatenated mask isn't legal.
+define void @PR59003(<2 x float> %0, <2 x float> %1, <8 x i1> %shuffle108) {
+; AVX-LABEL: PR59003:
+; AVX:       ## %bb.0: ## %entry
+; AVX-NEXT:    .p2align 4, 0x90
+; AVX-NEXT:  LBB4_1: ## %for.body.i
+; AVX-NEXT:    ## =>This Inner Loop Header: Depth=1
+; AVX-NEXT:    jmp LBB4_1
+entry:
+  br label %for.body.i
+
+for.body.i:                                       ; preds = %for.body.i, %entry
+  %2 = phi <8 x float> [ zeroinitializer, %entry ], [ %3, %for.body.i ]
+  %shuffle111 = shufflevector <2 x float> %0, <2 x float> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
+  %shuffle112 = shufflevector <2 x float> %1, <2 x float> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
+  %3 = select <8 x i1> %shuffle108, <8 x float> %shuffle111, <8 x float> %shuffle112
+  %4 = shufflevector <8 x float> zeroinitializer, <8 x float> %2, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>
+  %5 = select <8 x i1> zeroinitializer, <8 x float> zeroinitializer, <8 x float> %2
+  br label %for.body.i
+}
+
+
 ; Split a 256-bit select into two 128-bit selects when the operands are concatenated.
 
 define void @blendv_split(ptr %p, <8 x i32> %cond, <8 x i32> %a, <8 x i32> %x, <8 x i32> %y, <8 x i32> %z, <8 x i32> %w) {
@@ -244,10 +266,25 @@ define void @blendv_split(ptr %p, <8 x i32> %cond, <8 x i32> %a, <8 x i32> %x, <
 }
 
 ; Regression test for rGea8fb3b60196
+; FIXME: Missing fold vselect(zero, T, F) -> F
 define void @vselect_concat() {
-; AVX-LABEL: vselect_concat:
-; AVX:       ## %bb.0: ## %entry
-; AVX-NEXT:    retq
+; AVX1-LABEL: vselect_concat:
+; AVX1:       ## %bb.0: ## %entry
+; AVX1-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
+; AVX1-NEXT:    vmovaps %ymm0, (%rax)
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: vselect_concat:
+; AVX2:       ## %bb.0: ## %entry
+; AVX2-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
+; AVX2-NEXT:    vmovaps %ymm0, (%rax)
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: vselect_concat:
+; AVX512:       ## %bb.0: ## %entry
+; AVX512-NEXT:    retq
 entry:
   %0 = load <8 x i32>, ptr undef
   %1 = shufflevector <8 x i32> zeroinitializer, <8 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>