[llvm] [AArch64][SME] Allow inlining when streaming-mode attributes dont match up. (PR #68415)

[Jon Roelofs via llvm-commits]

================
@@ -190,16 +190,49 @@ static cl::opt<bool> EnableFixedwidthAutovecInStreamingMode(
 static cl::opt<bool> EnableScalableAutovecInStreamingMode(
     "enable-scalable-autovec-in-streaming-mode", cl::init(false), cl::Hidden);
 
+static bool isSMEABIRoutineCall(const CallInst &CI) {
+  const auto *F = CI.getCalledFunction();
+  return F && StringSwitch<bool>(F->getName())
+                  .Case("__arm_sme_state", true)
+                  .Case("__arm_tpidr2_save", true)
+                  .Case("__arm_tpidr2_restore", true)
+                  .Case("__arm_za_disable", true)
+                  .Default(false);
+}
+
+/// Returns true if the function has explicit operations that can only be lowered
+/// using incompatible instructions for the selected mode.
+/// This also returns true if the function F may use or modify ZA state.
+static bool hasPossibleIncompatibleOps(const Function *F) {
+  for (const BasicBlock &BB : *F) {
+    for (const Instruction &I : BB) {
+      // Be conservative for now and assume that any call to inline asm or to
+      // intrinsics could could result in non-streaming ops (e.g. calls to
+      // @llvm.aarch64.* or @llvm.gather/scatter intrinsics). We can assume that
+      // all native LLVM instructions can be lowered to compatible instructions.
+      if (isa<CallInst>(I) && !I.isDebugOrPseudoInst() &&
+          (cast<CallInst>(I).isInlineAsm() || isa<IntrinsicInst>(I) ||
+           isSMEABIRoutineCall(cast<CallInst>(I))))
----------------
jroelofs wrote:

This would be a great place for a remark that explains why we can't inline.

https://github.com/llvm/llvm-project/pull/68415