[clang] [llvm] [inlineasm] Add special support for "rm" inline asm constraints (PR #181973)

[Nick Desaulniers via cfe-commits]

================
@@ -207,28 +636,83 @@ static bool processCallBrInst(Function &F, CallBrInst *CBR, DominatorTree *DT) {
   return Changed;
 }
 
-static SmallVector<CallBrInst *, 2> findCallBrs(Function &F) {
-  SmallVector<CallBrInst *, 2> CBRs;
-  for (BasicBlock &BB : F)
-    if (auto *CBR = dyn_cast<CallBrInst>(BB.getTerminator()))
-      if (!CBR->getType()->isVoidTy() && !CBR->use_empty())
-        CBRs.push_back(CBR);
-  return CBRs;
-}
-
-static bool runImpl(Function &F, ArrayRef<CallBrInst *> CBRs,
-                    DominatorTree *DT) {
+static bool runImpl(Function &F, ArrayRef<CallBase *> IAs, DominatorTree *DT,
+                    const TargetMachine *TM) {
   bool Changed = false;
+  bool isOptLevelNone = TM->getOptLevel() == CodeGenOptLevel::None;
----------------
nickdesaulniers wrote:

>  We would still need a way to wedge in loads and stores though

hmm...

The whole point of the CallBrPrepare pass was that SelectionDAGISel could not modify BasicBlocks other than the one it was currently iterating; so we likely can't insert loads in successor blocks for callbr during ISEL.

---
Spitballing other ideas here, since I can focus more time here now...

Perhaps modifying the front end then to only produce "m" when "rm" appear in sources but `TM->getOptLevel() == CodeGenOptLevel::None` is perhaps also a way to avoid transforming the instructions.  Then the middle end could fold redundant loads and kill dead stores.

---
Perhaps ultimately if we _must_ transform from "rm" to "m", we have machinery under utils/ that are well tested, then we make use of those here.

https://github.com/llvm/llvm-project/pull/181973