[llvm] [AArch64] Fix using NEON copies in streaming-mode-enable regions. (PR #174738)

[Benjamin Maxwell via llvm-commits]

================
@@ -5389,6 +5389,38 @@ void AArch64InstrInfo::copyGPRRegTuple(MachineBasicBlock &MBB,
   }
 }
 
+/// Returns true if the instruction at I is in a streaming call site region,
+/// within a single basic block.
+/// A "call site streaming region" starts after smstart and ends at smstop
+/// around a call to a streaming function. This walks backward from I.
+static bool isInStreamingCallSiteRegion(MachineBasicBlock &MBB,
+                                        MachineBasicBlock::iterator I) {
+  MachineFunction &MF = *MBB.getParent();
+  AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
+  if (!AFI->hasStreamingModeChanges())
+    return false;
+  // Walk backwards to find smstart/smstop
+  for (MachineInstr &MI : reverse(make_range(MBB.begin(), I))) {
+    unsigned Opc = MI.getOpcode();
+    if (Opc == AArch64::MSRpstatesvcrImm1 || Opc == AArch64::MSRpstatePseudo) {
+      // Check if this is SM change (not ZA)
+      int64_t PState = MI.getOperand(0).getImm();
+      if (PState == AArch64SVCR::SVCRSM || PState == AArch64SVCR::SVCRSMZA) {
+        // Operand 1 is 1 for start, 0 for stop
+        return MI.getOperand(1).getImm() == 1;
+      }
+    }
+  }
+  return false;
+}
+
+/// Returns true if in a streaming call site region without SME-FA64.
+static bool mustAvoidNeonAtMBBI(const AArch64Subtarget &Subtarget,
+                                MachineBasicBlock &MBB,
+                                MachineBasicBlock::iterator I) {
+  return isInStreamingCallSiteRegion(MBB, I) && !Subtarget.hasSMEFA64();
----------------
MacDue wrote:

nit: Do the cheaper check first:
```suggestion
  return !Subtarget.hasSMEFA64() && isInStreamingCallSiteRegion(MBB, I);
```

https://github.com/llvm/llvm-project/pull/174738