[llvm] [DirectX] Propagate shader flags mask of callees to callers (PR #118306)

[Chris B via llvm-commits]

================
@@ -81,16 +101,31 @@ void ComputedShaderFlags::print(raw_ostream &OS) const {
   OS << ";\n";
 }
 
-/// Return the shader flags mask of the specified function Func.
-const ComputedShaderFlags &
-ModuleShaderFlags::getFunctionFlags(const Function *Func) const {
+auto ModuleShaderFlags::getFunctionShaderFlagInfo(const Function *Func) const {
   const auto Iter = llvm::lower_bound(
       FunctionFlags, Func,
       [](const std::pair<const Function *, ComputedShaderFlags> FSM,
          const Function *FindFunc) { return (FSM.first < FindFunc); });
   assert((Iter != FunctionFlags.end() && Iter->first == Func) &&
          "No Shader Flags Mask exists for function");
-  return Iter->second;
+  return Iter;
+}
+
+/// Merge mask NewSF to that of Func, if different.
+/// Return true if mask of Func is changed, else false.
+bool ModuleShaderFlags::mergeFunctionShaderFlags(
+    const Function *Func, const ComputedShaderFlags NewSF) {
+  const auto FuncSFInfo = getFunctionShaderFlagInfo(Func);
+  if ((FuncSFInfo->second & NewSF) != NewSF) {
+    const_cast<ComputedShaderFlags &>(FuncSFInfo->second).merge(NewSF);
----------------
llvm-beanz wrote:

Wait... I'm clearly not thinking straight because that last comment doesn't fully make sense. I do think we should probably get rid of the implicit conversion operator. Clearly `operator==` isn't the right approach.

For flag merging, we maybe don't even need to check it conditionally though. This should all be really well optimized straight-line code to merge flags, so the branch is probably unnecessary.

https://github.com/llvm/llvm-project/pull/118306