From 790f2eb16a279f4cb4b57397fcafd5cadb49d6b7 Mon Sep 17 00:00:00 2001
From: Alex MacLean <amaclean@nvidia.com>
Date: Tue, 17 Sep 2024 15:31:28 -0700
Subject: [PATCH] [InstCombine] Avoid simplifying bitcast of undef to a
 zeroinitializer vector (#108872)

In some cases, if an undef value is the product of another instcombine
simplification, a bitcast of undef is simplified to a zeroinitializer
vector instead of undef.
---
 .../InstCombine/InstCombineCasts.cpp          |  5 ++++
 llvm/test/Transforms/InstCombine/bitcast.ll   | 23 +++++++++++++++++++
 2 files changed, 28 insertions(+)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index 5c9faa9449f539..ea51d779045718 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -2323,6 +2323,11 @@ static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
   auto *DestVecTy = cast<FixedVectorType>(CI.getType());
   Value *IntInput = CI.getOperand(0);
 
+  // if the int input is just an undef value do not try to optimize to vector
+  // insertions as it will prevent undef propagation
+  if (isa<UndefValue>(IntInput))
+    return nullptr;
+
   SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
   if (!collectInsertionElements(IntInput, 0, Elements,
                                 DestVecTy->getElementType(),
diff --git a/llvm/test/Transforms/InstCombine/bitcast.ll b/llvm/test/Transforms/InstCombine/bitcast.ll
index 4ab24ce7b925dc..79e6370b7242f5 100644
--- a/llvm/test/Transforms/InstCombine/bitcast.ll
+++ b/llvm/test/Transforms/InstCombine/bitcast.ll
@@ -879,3 +879,26 @@ define half @copysign_idiom_constant_wrong_type2(bfloat %x, i16 %mag) {
   %y = bitcast i16 %res to half
   ret half %y
 }
+
+define i16 @bitcast_undef_to_vector() {
+; CHECK-LABEL: @bitcast_undef_to_vector(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[END:%.*]]
+; CHECK:       unreachable:
+; CHECK-NEXT:    br label [[END]]
+; CHECK:       end:
+; CHECK-NEXT:    ret i16 undef
+;
+entry:
+  br label %end
+
+unreachable:                                 ; No predecessors!
+  %0 = extractvalue { i32, i32 } zeroinitializer, 1
+  br label %end
+
+end:                                        ; preds = %unreachable, %entry
+  %1 = phi i32 [ %0, %unreachable ], [ undef, %entry ]
+  %2 = bitcast i32 %1 to <2 x i16>
+  %3 = extractelement <2 x i16> %2, i64 0
+  ret i16 %3
+}