[llvm] [AArch64] Add streaming-mode stack hazard optimization remarks (PR #101695)

[David Green via llvm-commits]

================
@@ -5029,3 +5050,171 @@ void AArch64FrameLowering::inlineStackProbe(MachineFunction &MF,
     MI->eraseFromParent();
   }
 }
+
+struct StackAccess {
+  enum AccessType {
+    NotAccessed = 0, // Stack object not accessed by load/store instructions.
+    GPR = 1 << 0,    // A general purpose register.
+    PPR = 1 << 1,    // A predicate register.
+    FPR = 1 << 2,    // A floating point/Neon/SVE register.
+  };
+
+  int Idx;
+  StackOffset Offset;
+  int64_t Size;
+  unsigned AccessTypes;
+
+  StackAccess() : Idx(0), Offset(), Size(0), AccessTypes(NotAccessed) {}
+
+  bool operator<(const StackAccess &Rhs) const {
+    return std::make_tuple(start(), Idx) <
+           std::make_tuple(Rhs.start(), Rhs.Idx);
+  }
+
+  bool isCPU() const {
+    // Predicate register load and store instructions execute on the CPU.
+    return AccessTypes & (AccessType::GPR | AccessType::PPR);
+  }
+  bool isSME() const { return AccessTypes & AccessType::FPR; }
+  bool isMixed() const { return ((AccessTypes & (AccessTypes - 1)) != 0); }
+
+  int64_t start() const { return Offset.getFixed() + Offset.getScalable(); }
+  int64_t end() const { return start() + Size; }
+
+  std::string getTypeString() const {
+    switch (AccessTypes) {
+    case AccessType::FPR:
+      return "FPR";
+    case AccessType::PPR:
+      return "PPR";
+    case AccessType::GPR:
+      return "GPR";
+    case AccessType::NotAccessed:
+      return "NA";
+    default:
+      return "Mixed";
+    }
+  }
+
+  void print(raw_ostream &OS) const {
+    OS << getTypeString() << " stack object at [SP"
+       << (Offset.getFixed() < 0 ? "" : "+") << Offset.getFixed();
+    if (Offset.getScalable())
+      OS << (Offset.getScalable() < 0 ? "" : "+") << Offset.getScalable()
+         << " * vscale";
+    OS << "]";
+  }
+};
+
+static inline raw_ostream &operator<<(raw_ostream &OS, const StackAccess &SA) {
+  SA.print(OS);
+  return OS;
+}
+
+void AArch64FrameLowering::emitRemarks(
+    const MachineFunction &MF, MachineOptimizationRemarkEmitter &ORE) const {
+
+  SMEAttrs Attrs(MF.getFunction());
+  if (Attrs.hasNonStreamingInterfaceAndBody())
+    return;
+
+  const uint64_t HazardSize =
+      (StackHazardSize) ? StackHazardSize : StackHazardRemarkSize;
+
+  if (HazardSize == 0)
+    return;
+
+  const MachineFrameInfo &MFI = MF.getFrameInfo();
+
+  std::vector<StackAccess> StackAccesses(MFI.getNumObjects());
----------------
davemgreen wrote:

I'm not sure what happens if `MFI.getNumObjects()==0`, but it might be worth bailing out early.

https://github.com/llvm/llvm-project/pull/101695