[llvm] [TSan] Fix atomicrmw xchg with pointer and floats (PR #85228)

[Nikita Popov via llvm-commits]

================
@@ -752,11 +752,12 @@ bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
     const unsigned ByteSize = 1U << Idx;
     const unsigned BitSize = ByteSize * 8;
     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
-    Value *Args[] = {Addr,
-                     IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
+    Value *Val = RMWI->getValOperand();
+    Value *Args[] = {Addr, IRB.CreateBitOrPointerCast(Val, Ty),
                      createOrdering(&IRB, RMWI->getOrdering())};
-    CallInst *C = CallInst::Create(F, Args);
-    ReplaceInstWithInst(I, C);
+    Value *C = IRB.CreateCall(F, Args);
+    I->replaceAllUsesWith(IRB.CreateBitOrPointerCast(C, Val->getType()));
----------------
nikic wrote:

The reason I changed this is that we now go through IRBuilder, which will insert the instructions itself, so we can no longer use ReplaceInstWithInst. This is also why the cmpxchg code does the same.

https://github.com/llvm/llvm-project/pull/85228