[LLVM-C] Allow LLVM{Get,Set}Volatile to work on memory intrinsics

[AMS21]

While working on rust-lang/rust#147474 it was discovered that LLVMSetVolatile does not work with memory intrinsic functions like memset, memcpy and memmove.

To quote @nikic source

LLVMSetVolatile does not support volatile on intrinsics. It only works on the instruction-level volatile flag.

This patch fixes that so that getting and setting the volatile flag now works with these memory intrinsic functions.

[llvmbot]

@llvm/pr-subscribers-llvm-ir

Author: None (AMS21)

Changes

While working on rust-lang/rust#147474 it was discovered that LLVMSetVolatile does not work with memory intrinsic functions like memset, memcpy and memmove.

To quote @nikic source
> LLVMSetVolatile does not support volatile on intrinsics. It only works on the instruction-level volatile flag.

This patch fixes that so that getting and setting the volatile flag now works with these memory intrinsic functions.

---

Full diff: https://github.com/llvm/llvm-project/pull/162810.diff

3 Files Affected:

• (modified) llvm/docs/ReleaseNotes.md (+1)
• (modified) llvm/lib/IR/Core.cpp (+4)
• (modified) llvm/unittests/IR/InstructionsTest.cpp (+107)

diff --git a/llvm/docs/ReleaseNotes.md b/llvm/docs/ReleaseNotes.md
index 79d93d08b8398..80fc20c5d341a 100644
--- a/llvm/docs/ReleaseNotes.md
+++ b/llvm/docs/ReleaseNotes.md
@@ -147,6 +147,7 @@ Changes to the C API
 --------------------
 
 * Add `LLVMGetOrInsertFunction` to get or insert a function, replacing the combination of `LLVMGetNamedFunction` and `LLVMAddFunction`.
+* Allow `LLVMGetVolatile` and `LLVMSetVolatile` to work on memory intrinsics like `memset`, `memcpy`, and `memmove`.
 
 Changes to the CodeGen infrastructure
 -------------------------------------
diff --git a/llvm/lib/IR/Core.cpp b/llvm/lib/IR/Core.cpp
index 3f1cc1e4e6d0e..3c330bb3c476e 100644
--- a/llvm/lib/IR/Core.cpp
+++ b/llvm/lib/IR/Core.cpp
@@ -4106,6 +4106,8 @@ LLVMBool LLVMGetVolatile(LLVMValueRef MemAccessInst) {
     return SI->isVolatile();
   if (AtomicRMWInst *AI = dyn_cast<AtomicRMWInst>(P))
     return AI->isVolatile();
+  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(P))
+    return MI->isVolatile();
   return cast<AtomicCmpXchgInst>(P)->isVolatile();
 }
 
@@ -4117,6 +4119,8 @@ void LLVMSetVolatile(LLVMValueRef MemAccessInst, LLVMBool isVolatile) {
     return SI->setVolatile(isVolatile);
   if (AtomicRMWInst *AI = dyn_cast<AtomicRMWInst>(P))
     return AI->setVolatile(isVolatile);
+  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(P))
+    return MI->setVolatile(ConstantInt::getBool(P->getContext(), isVolatile));
   return cast<AtomicCmpXchgInst>(P)->setVolatile(isVolatile);
 }
 
diff --git a/llvm/unittests/IR/InstructionsTest.cpp b/llvm/unittests/IR/InstructionsTest.cpp
index fe9e7e8228490..020512ba7225a 100644
--- a/llvm/unittests/IR/InstructionsTest.cpp
+++ b/llvm/unittests/IR/InstructionsTest.cpp
@@ -1933,5 +1933,112 @@ TEST(InstructionsTest, StripAndAccumulateConstantOffset) {
   EXPECT_TRUE(Offset.isZero());
 }
 
+TEST(InstructionsTest, LLVMGetSetVolatile) {
+  LLVMContext Ctx;
+  Module M("M", Ctx);
+  FunctionType *FT = FunctionType::get(Type::getVoidTy(Ctx), {}, false);
+  Function *F = Function::Create(FT, Function::ExternalLinkage, "F", M);
+  BasicBlock *BB = BasicBlock::Create(Ctx, "entry", F);
+  IRBuilder<> B(BB);
+
+  Type *Int32Ty = Type::getInt32Ty(Ctx);
+  Value *Val = B.getInt32(1);
+  Value* Ptr = B.CreateAlloca(Int32Ty);
+
+  // LoadInst
+  {
+    LoadInst *Load = B.CreateLoad(Int32Ty, Ptr, "");
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(Load)));
+    LLVMSetVolatile(wrap(Load), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(Load)));
+    LLVMSetVolatile(wrap(Load), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(Load)));
+  }
+
+  // StoreInst
+  {
+    StoreInst *Store = B.CreateStore(Val, Ptr);
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(Store)));
+    LLVMSetVolatile(wrap(Store), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(Store)));
+    LLVMSetVolatile(wrap(Store), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(Store)));
+  }
+
+  // AtomicRMWInst
+  {
+    AtomicRMWInst *RMW =
+        B.CreateAtomicRMW(AtomicRMWInst::Add, Ptr, Val,
+                                MaybeAlign(), AtomicOrdering::Monotonic);
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(RMW)));
+    LLVMSetVolatile(wrap(RMW), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(RMW)));
+    LLVMSetVolatile(wrap(RMW), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(RMW)));
+  }
+
+  // MemCpy
+  {
+    CallInst *MemCpy =
+        B.CreateMemCpy(Ptr, MaybeAlign(), Ptr, MaybeAlign(), 4);
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemCpy)));
+    LLVMSetVolatile(wrap(MemCpy), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(MemCpy)));
+    LLVMSetVolatile(wrap(MemCpy), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemCpy)));
+  }
+
+  // MemMove
+  {
+    CallInst *MemMove =
+        B.CreateMemMove(Ptr, MaybeAlign(), Ptr, MaybeAlign(), 4);
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemMove)));
+    LLVMSetVolatile(wrap(MemMove), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(MemMove)));
+    LLVMSetVolatile(wrap(MemMove), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemMove)));
+  }
+
+  // MemSet
+  {
+    CallInst *MemSet = B.CreateMemSet(Ptr, B.getInt8(0), 1, Align(1));
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemSet)));
+    LLVMSetVolatile(wrap(MemSet), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(MemSet)));
+    LLVMSetVolatile(wrap(MemSet), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemSet)));
+  }
+
+  // MemSetInline
+  {
+    CallInst *MemSetInline =
+        B.CreateMemSetInline(Ptr, Align(1), B.getInt8(0), Val);
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemSetInline)));
+    LLVMSetVolatile(wrap(MemSetInline), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(MemSetInline)));
+    LLVMSetVolatile(wrap(MemSetInline), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemSetInline)));
+  }
+
+  // MemCpyInline
+  {
+    CallInst *MemCpyInline = B.CreateMemCpyInline(
+        Ptr, MaybeAlign(), Val, MaybeAlign(), Val);
+
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemCpyInline)));
+    LLVMSetVolatile(wrap(MemCpyInline), true);
+    EXPECT_TRUE(LLVMGetVolatile(wrap(MemCpyInline)));
+    LLVMSetVolatile(wrap(MemCpyInline), false);
+    EXPECT_FALSE(LLVMGetVolatile(wrap(MemCpyInline)));
+  }
+}
+
 } // end anonymous namespace
 } // end namespace llvm

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[nikic]

We could replace this whole function with cast<Instruction>(P)->isVolatile(). This would just return false for instructions that don't support volatile.

[nikic]

After thinking about this a bit more, I'm not sure we should do this. I am planning to change the volatile representation for memory intrinsics from an argument to an operand bundle. This means that it will no longer be possible to make the intrinsic volatile in-place, it requires creating a new one. So there is a high risk that this support is going to be very short-lived.

[AMS21]

Ah okay. Yes I agree adding such a short lived change is not very useful.

What would be a better solution to this? Adding something like LLVMBuildMemSet2 etc. which would accepts an additional parameter for volatile like the LLVMRustBuildMemSet wrapper do.