[LLVM] Fix dangling access to aliasee in stripAndAccumulateConstantOf…

[bcardosolopes]

…fsets

Make sure LLVM bails when aliasee aren't present. Found while using mlir-translate --mlir-to-llvm. Added unittest exercises the builder using a target folder, this is the original IR triggering the issue:

@m2 = internal alias { [1 x i64] }, ptr @k0
@k0 = private constant {{ [1 x i64] }} {
  { [1 x i64] }
    {
      [1 x i64]
        [
         i64 ptrtoint (ptr getelementptr inbounds ({ [1 x i64] },
                       ptr @m2, i32 0, i32 0, i32 1) to i64)
        ]
    }
  }, align 4

[llvmbot]

@llvm/pr-subscribers-llvm-ir

Author: Bruno Cardoso Lopes (bcardosolopes)

Changes

…fsets

Make sure LLVM bails when aliasee aren't present. Found while using mlir-translate --mlir-to-llvm. Added unittest exercises the builder using a target folder, this is the original IR triggering the issue:

@<!-- -->m2 = internal alias { [1 x i64] }, ptr @<!-- -->k0
@<!-- -->k0 = private constant {{ [1 x i64] }} {
  { [1 x i64] }
    {
      [1 x i64]
        [
         i64 ptrtoint (ptr getelementptr inbounds ({ [1 x i64] },
                       ptr @<!-- -->m2, i32 0, i32 0, i32 1) to i64)
        ]
    }
  }, align 4

---

Full diff: https://github.com/llvm/llvm-project/pull/159214.diff

2 Files Affected:

• (modified) llvm/lib/IR/Value.cpp (+1-1)
• (modified) llvm/unittests/IR/InstructionsTest.cpp (+31)

diff --git a/llvm/lib/IR/Value.cpp b/llvm/lib/IR/Value.cpp
index 4e8f359481b81..67b80c2825fd2 100644
--- a/llvm/lib/IR/Value.cpp
+++ b/llvm/lib/IR/Value.cpp
@@ -776,7 +776,7 @@ const Value *Value::stripAndAccumulateConstantOffsets(
                Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
       V = cast<Operator>(V)->getOperand(0);
     } else if (auto *GA = dyn_cast<GlobalAlias>(V)) {
-      if (!GA->isInterposable())
+      if (!GA->isInterposable() && GA->getAliasee())
         V = GA->getAliasee();
     } else if (const auto *Call = dyn_cast<CallBase>(V)) {
         if (const Value *RV = Call->getReturnedArgOperand())
diff --git a/llvm/unittests/IR/InstructionsTest.cpp b/llvm/unittests/IR/InstructionsTest.cpp
index 21d45960dce7c..8aa1783bceaaa 100644
--- a/llvm/unittests/IR/InstructionsTest.cpp
+++ b/llvm/unittests/IR/InstructionsTest.cpp
@@ -10,6 +10,7 @@
 #include "llvm-c/Core.h"
 #include "llvm/ADT/CombinationGenerator.h"
 #include "llvm/ADT/STLExtras.h"
+#include "llvm/Analysis/TargetFolder.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/Analysis/VectorUtils.h"
 #include "llvm/AsmParser/Parser.h"
@@ -874,6 +875,36 @@ TEST(InstructionsTest, GEPIndices) {
   delete GEPI;
 }
 
+TEST(InstructionsTest, PtrToIntGEPFolding) {
+  LLVMContext Context;
+  llvm::Module *M = new llvm::Module("PtrToIntGEPFolding", Context);
+  llvm::IRBuilder<llvm::TargetFolder> Builder(
+      Context, llvm::TargetFolder(M->getDataLayout()));
+
+  auto *Int64Ty = Builder.getInt64Ty();
+  auto *ArrTy = llvm::ArrayType::get(Int64Ty, 1);
+  auto *StructTy = llvm::StructType::create(Context, {ArrTy}, "anon_struct");
+
+  // Create alias @m2 to @0
+  llvm::GlobalAlias *AliasM2 = llvm::GlobalAlias::create(
+      StructTy,
+      0, // address space
+      llvm::GlobalValue::InternalLinkage, "m2", (llvm::Constant *)nullptr, M);
+
+  // Build getelementptr for ptrtoint
+  std::vector<llvm::Value *> GEPIdxs = {
+      llvm::ConstantInt::get(llvm::Type::getInt32Ty(Context), 0),
+      llvm::ConstantInt::get(llvm::Type::getInt32Ty(Context), 0),
+      llvm::ConstantInt::get(llvm::Type::getInt32Ty(Context), 1)};
+  llvm::Value *GEP = Builder.CreateGEP(StructTy, AliasM2, GEPIdxs, "",
+                                       llvm::GEPNoWrapFlags::inBounds());
+  // Make sure stripAndAccumulateConstantOffsets can properly handle
+  // GEPs with nullptr global aliasees.
+  llvm::Value *PtrToInt = Builder.CreatePtrToInt(GEP, Int64Ty);
+  EXPECT_TRUE(PtrToInt);
+  delete M;
+}
+
 TEST(InstructionsTest, ZeroIndexGEP) {
   LLVMContext Context;
   DataLayout DL;

[nikic]

Per

llvm-project/llvm/lib/IR/Verifier.cpp

Lines 988 to 989 in 4c7de93

	const Constant *Aliasee = GA.getAliasee();
	Check(Aliasee, "Aliasee cannot be NULL!", &GA);

this is ill-formed IR.

Generally speaking, code is not expected to handle ill-formed IR. This would need some additional justification for why we should do this. Why is MLIR generating this invalid alias and expecting it to work?