[llvm] [AArch64] SME implementation for agnostic-ZA functions (PR #120150)

[Sander de Smalen via llvm-commits]

================
@@ -3218,6 +3220,39 @@ AArch64TargetLowering::EmitAllocateZABuffer(MachineInstr &MI,
   return BB;
 }
 
+// TODO: Find a way to merge this with EmitAllocateZABuffer.
+MachineBasicBlock *
+AArch64TargetLowering::EmitAllocateSMESaveBuffer(MachineInstr &MI,
+                                                 MachineBasicBlock *BB) const {
+  MachineFunction *MF = BB->getParent();
+  MachineFrameInfo &MFI = MF->getFrameInfo();
+  AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>();
+  assert(!MF->getSubtarget<AArch64Subtarget>().isTargetWindows() &&
+         "Lazy ZA save is not yet supported on Windows");
+
+  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
+  if (FuncInfo->getSMESaveBufferUsed()) {
+    // Allocate a lazy-save buffer object of the size given, normally SVL * SVL
+    auto Size = MI.getOperand(1).getReg();
+    auto Dest = MI.getOperand(0).getReg();
+    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::SUBXrx64), Dest)
+        .addReg(AArch64::SP)
+        .addReg(Size)
+        .addImm(AArch64_AM::getArithExtendImm(AArch64_AM::UXTX, 0));
+    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY),
+            AArch64::SP)
+        .addReg(Dest);
----------------
sdesmalen-arm wrote:

No particular reason. This function was initially copied from `EmitAllocateZABuffer` and I merely updated the `sub` operation to remove the multiplication.

https://github.com/llvm/llvm-project/pull/120150