[llvm] [NVPTX] Improve modeling of inline PTX (PR #130675)

[Alex MacLean via llvm-commits]

================
@@ -115,3 +116,30 @@ ModRefInfo NVPTXAAResult::getModRefInfoMask(const MemoryLocation &Loc,
 
   return ModRefInfo::ModRef;
 }
+
+MemoryEffects NVPTXAAResult::getMemoryEffects(const CallBase *Call,
+                                              AAQueryInfo &AAQI) {
+  // Inline assembly with no side-effect or memory clobbers should not
+  // indirectly access memory in the PTX specification.
+  if (const auto *IA = dyn_cast<InlineAsm>(Call->getCalledOperand())) {
+    // Volatile is translated as side-effects.
+    if (IA->hasSideEffects())
+      return MemoryEffects::unknown();
+
+    for (const InlineAsm::ConstraintInfo &Constraint : IA->ParseConstraints()) {
+      // Indirect constraints (e.g. =*m) are unsupported in inline PTX.
+      if (Constraint.isIndirect)
+        return MemoryEffects::unknown();
----------------
AlexMaclean wrote:

I prefer to conservatively bail out with `MemoryEffects::unknown()`. I don't know of any case where a frontend would generate an indirect constraint or a reason to use one but just in case some IR exists somewhere with these I wouldn't want to start crashing when we can easily handle them. 

https://github.com/llvm/llvm-project/pull/130675