[TLSVariableHoist] Do not introduce bitcast on llvm.threadlocal.address argument

[abhishek-kaushik22]

The method AddressingModeMatcher::matchOperationAddr and Verifier::visitIntrinsicCall expect a GlobalValue as the operand of the llvm.threadlocal.address.p0 intrinsic. The hoist TLS pass replaces them with a bitcast instruction, causing a crash.

Fixes #112771

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-llvm-transforms

Author: None (abhishek-kaushik22)

Changes

The method AddressingModeMatcher::matchOperationAddr and Verifier::visitIntrinsicCall expect a GlobalValue as the operand of the llvm.threadlocal.address.p0 intrinsic. The hoist TLS pass replaces them with a bitcast instruction, causing a crash.

Fixes #112771

---

Full diff: https://github.com/llvm/llvm-project/pull/112772.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Scalar/TLSVariableHoist.cpp (+6-1)

diff --git a/llvm/lib/Transforms/Scalar/TLSVariableHoist.cpp b/llvm/lib/Transforms/Scalar/TLSVariableHoist.cpp
index 58ea5b68d5488b..93cd736945aace 100644
--- a/llvm/lib/Transforms/Scalar/TLSVariableHoist.cpp
+++ b/llvm/lib/Transforms/Scalar/TLSVariableHoist.cpp
@@ -238,8 +238,13 @@ bool TLSVariableHoistPass::tryReplaceTLSCandidate(Function &Fn,
   auto *CastInst = genBitCastInst(Fn, GV);
 
   // to replace the uses of TLS Candidate
-  for (auto &User : Cand.Users)
+  for (auto &User : Cand.Users) {
+    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User.Inst)) {
+      if (II->getIntrinsicID() == Intrinsic::threadlocal_address)
+        continue;
+    }
     User.Inst->setOperand(User.OpndIdx, CastInst);
+  }
 
   return true;
 }

[MalaySanghi]

@abhishek-kaushik22
I think we should add a lit test here

[abhishek-kaushik22]

@nikic can you please review?

[nikic]

I don't think this is the right approach to fix this: If llvm.threadlocal.address is used, then we need to hoist that call. Skipping it would defeat the point.

However ... I think this entire pass may just be useless how? Since we switched to using llvm.threadlocal.address, I believe that normal EarlyCSE/GVN and LICM will already perform the optimization this pass is intended to do. Maybe we can just delete this code entirely?

[abhishek-kaushik22]

I've opened #114740 to delete the pass.