[llvm-branch-commits] [clang] [llvm] [HLSL] Add float overload for `InterlockedExchange` (PR #222163)

[Joshua Batista via llvm-branch-commits]

================
@@ -356,6 +356,34 @@ static bool updateFnegToFsub(Instruction &I,
   return true;
 }
 
+// DXIL has no floating-point atomic operation. A float exchange only moves the
+// bit pattern, so exchange an integer of the same width instead. Opaque
+// pointers keep the pointer operand type-agnostic, so only the value and the
+// result need a cast. This matches what DXC emits for groupshared memory.
+static bool
+legalizeFloatAtomicExchange(Instruction &I,
+                            SmallVectorImpl<Instruction *> &ToRemove,
+                            DenseMap<Value *, Value *> &) {
+  auto *AI = dyn_cast<AtomicRMWInst>(&I);
+  if (!AI || AI->getOperation() != AtomicRMWInst::Xchg)
+    return false;
+
+  Type *ValTy = AI->getValOperand()->getType();
+  if (!ValTy->isFloatingPointTy())
----------------
bob80905 wrote:

SemaHLSL should catch non-conformant types. But you're right that we could catch illegal widths here, so I've added a width check right after this.

https://github.com/llvm/llvm-project/pull/222163