[llvm] r304975 - InferAddressSpaces: Avoid assertion failure with replacing identical
Nirav Dave via llvm-commits
llvm-commits at lists.llvm.org
Thu Jun 8 06:20:56 PDT 2017
Author: niravd
Date: Thu Jun 8 08:20:55 2017
New Revision: 304975
URL: http://llvm.org/viewvc/llvm-project?rev=304975&view=rev
Log:
InferAddressSpaces: Avoid assertion failure with replacing identical
cloned constexpr
Have cloneConstantExprWithNewAddressSpaces return nullptr when
returning initial ConstantExpr.
Reviewers: arsenm
Subscribers: jholewinski, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D33995
Added:
llvm/trunk/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll
Modified:
llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp
Modified: llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp?rev=304975&r1=304974&r2=304975&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp Thu Jun 8 08:20:55 2017
@@ -500,6 +500,7 @@ static Value *cloneConstantExprWithNewAd
}
// Computes the operands of the new constant expression.
+ bool IsNew = false;
SmallVector<Constant *, 4> NewOperands;
for (unsigned Index = 0; Index < CE->getNumOperands(); ++Index) {
Constant *Operand = CE->getOperand(Index);
@@ -509,6 +510,7 @@ static Value *cloneConstantExprWithNewAd
// bitcast, and getelementptr) do not incur cycles in the data flow graph
// and (2) this function is called on constant expressions in postorder.
if (Value *NewOperand = ValueWithNewAddrSpace.lookup(Operand)) {
+ IsNew = true;
NewOperands.push_back(cast<Constant>(NewOperand));
} else {
// Otherwise, reuses the old operand.
@@ -516,6 +518,11 @@ static Value *cloneConstantExprWithNewAd
}
}
+ // If !IsNew, we will replace the Value with itself. However, replaced values
+ // are assumed to wrapped in a addrspace cast later so drop it now.
+ if (!IsNew)
+ return nullptr;
+
if (CE->getOpcode() == Instruction::GetElementPtr) {
// Needs to specify the source type while constructing a getelementptr
// constant expression.
Added: llvm/trunk/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll?rev=304975&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll (added)
+++ llvm/trunk/test/Transforms/InferAddressSpaces/NVPTX/clone_constexpr.ll Thu Jun 8 08:20:55 2017
@@ -0,0 +1,36 @@
+; RUN: opt -S -mtriple=nvptx64-nvidia-cuda -infer-address-spaces %s | FileCheck %s
+
+%struct.S = type { [5 x i32] }
+
+$g1 = comdat any
+
+ at g1 = linkonce_odr addrspace(3) global %struct.S zeroinitializer, comdat, align 4
+
+; CHECK-LABEL: @foo(
+; CHECK: %x0 = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #2
+; CHECK: %idxprom.i = zext i32 %x0 to i64
+; CHECK: %arrayidx.i = getelementptr %struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 %idxprom.i
+; CHECK: tail call void @f1(i32* %arrayidx.i, i32 undef) #0
+; CHECK: %x1 = load i32, i32* getelementptr (%struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 0), align 4
+; CHECK: %L.sroa.0.0.insert.ext.i = zext i32 %x1 to i64
+; CHECK: tail call void @f2(i64* null, i64 %L.sroa.0.0.insert.ext.i) #0
+; CHECK: ret void
+define void @foo() local_unnamed_addr #0 {
+entry:
+ %x0 = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #2
+ %idxprom.i = zext i32 %x0 to i64
+ %arrayidx.i = getelementptr %struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 %idxprom.i
+ tail call void @f1(i32* %arrayidx.i, i32 undef) #0
+ %x1 = load i32, i32* getelementptr (%struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 0), align 4
+ %L.sroa.0.0.insert.ext.i = zext i32 %x1 to i64
+ tail call void @f2(i64* null, i64 %L.sroa.0.0.insert.ext.i) #0
+ ret void
+}
+
+declare void @f1(i32*, i32) local_unnamed_addr #0
+declare void @f2(i64*, i64) local_unnamed_addr #0
+declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() #1
+
+attributes #0 = { convergent nounwind }
+attributes #1 = { nounwind readnone }
+attributes #2 = { nounwind }
More information about the llvm-commits
mailing list