[VectorCombine] fix insertion point of shuffles
authorSanjay Patel <spatel@rotateright.com>
Fri, 10 Feb 2023 15:53:22 +0000 (10:53 -0500)
committerSanjay Patel <spatel@rotateright.com>
Fri, 10 Feb 2023 15:57:11 +0000 (10:57 -0500)
As shown in issue #60649, the new shuffles were
being inserted before a phi, and that is invalid.

It seems like most test coverage for this fold
(foldSelectShuffle) lives in the AArch64 dir,
but this doesn't repro there for a base target.

llvm/lib/Transforms/Vectorize/VectorCombine.cpp
llvm/test/Transforms/VectorCombine/X86/select-shuffle.ll [new file with mode: 0644]

index 6dc5f5f..d60a1c4 100644 (file)
@@ -1660,16 +1660,16 @@ bool VectorCombine::foldSelectShuffle(Instruction &I, bool FromReduction) {
           return SSV->getOperand(Op);
     return SV->getOperand(Op);
   };
-  Builder.SetInsertPoint(SVI0A->getNextNode());
+  Builder.SetInsertPoint(SVI0A->getInsertionPointAfterDef());
   Value *NSV0A = Builder.CreateShuffleVector(GetShuffleOperand(SVI0A, 0),
                                              GetShuffleOperand(SVI0A, 1), V1A);
-  Builder.SetInsertPoint(SVI0B->getNextNode());
+  Builder.SetInsertPoint(SVI0B->getInsertionPointAfterDef());
   Value *NSV0B = Builder.CreateShuffleVector(GetShuffleOperand(SVI0B, 0),
                                              GetShuffleOperand(SVI0B, 1), V1B);
-  Builder.SetInsertPoint(SVI1A->getNextNode());
+  Builder.SetInsertPoint(SVI1A->getInsertionPointAfterDef());
   Value *NSV1A = Builder.CreateShuffleVector(GetShuffleOperand(SVI1A, 0),
                                              GetShuffleOperand(SVI1A, 1), V2A);
-  Builder.SetInsertPoint(SVI1B->getNextNode());
+  Builder.SetInsertPoint(SVI1B->getInsertionPointAfterDef());
   Value *NSV1B = Builder.CreateShuffleVector(GetShuffleOperand(SVI1B, 0),
                                              GetShuffleOperand(SVI1B, 1), V2B);
   Builder.SetInsertPoint(Op0);
diff --git a/llvm/test/Transforms/VectorCombine/X86/select-shuffle.ll b/llvm/test/Transforms/VectorCombine/X86/select-shuffle.ll
new file mode 100644 (file)
index 0000000..d51ac6a
--- /dev/null
@@ -0,0 +1,38 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- | FileCheck %s
+
+target datalayout = "e-p:64:64-i64:64-f80:128-n8:16:32:64-S128"
+
+; This would insert before a phi instruction which is invalid IR.
+
+define <4 x double> @PR60649() {
+; CHECK-LABEL: @PR60649(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[END:%.*]]
+; CHECK:       unreachable:
+; CHECK-NEXT:    br label [[END]]
+; CHECK:       end:
+; CHECK-NEXT:    [[T0:%.*]] = phi <4 x double> [ zeroinitializer, [[ENTRY:%.*]] ], [ zeroinitializer, [[UNREACHABLE:%.*]] ]
+; CHECK-NEXT:    [[T1:%.*]] = phi <4 x double> [ zeroinitializer, [[ENTRY]] ], [ zeroinitializer, [[UNREACHABLE]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = shufflevector <4 x double> [[T0]], <4 x double> [[T0]], <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x double> [[T0]], <4 x double> [[T0]], <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+; CHECK-NEXT:    [[TMP2:%.*]] = fdiv <4 x double> [[TMP1]], <double 0.000000e+00, double 0.000000e+00, double undef, double undef>
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul <4 x double> [[TMP0]], <double 0.000000e+00, double 0.000000e+00, double undef, double undef>
+; CHECK-NEXT:    [[T5:%.*]] = shufflevector <4 x double> [[TMP2]], <4 x double> [[TMP3]], <4 x i32> <i32 0, i32 1, i32 4, i32 5>
+; CHECK-NEXT:    ret <4 x double> [[T5]]
+;
+entry:
+  br label %end
+
+unreachable:
+  br label %end
+
+end:
+  %t0 = phi <4 x double> [ zeroinitializer, %entry ], [ zeroinitializer, %unreachable ]
+  %t1 = phi <4 x double> [ zeroinitializer, %entry ], [ zeroinitializer, %unreachable ]
+  %t2 = shufflevector <4 x double> zeroinitializer, <4 x double> zeroinitializer, <4 x i32> <i32 0, i32 0, i32 1, i32 1>
+  %t3 = fdiv <4 x double> %t0, %t2
+  %t4 = fmul <4 x double> %t0, %t2
+  %t5 = shufflevector <4 x double> %t3, <4 x double> %t4, <4 x i32> <i32 0, i32 1, i32 6, i32 7>
+  ret <4 x double> %t5
+}