[AArch64] Remove copy in SVE/SME predicate spill

7dc20ab introduced an extra COPY when spilling a PNR register, which
can't be elided as the input (PNR predicate) and output (PPR predicate)
register classes differ. This patch emits a new ConvertPNRtoPPR
pseudo instruction instead. When this is expanded, it gets erased if the
PNR is a subregister of the PPR, since the conversion is implicit,
otherwise it is lowered to an ORR.

Author: SamTebbs33

llvm/lib/Target/AArch64/AArch64ExpandPseudoInsts.cpp

@@ -1112,6 +1112,23 @@ bool AArch64ExpandPseudo::expandMI(MachineBasicBlock &MBB,
   default:
     break;
 
+  case AArch64::ConvertPNRtoPPR: {
+    auto TRI = MBB.getParent()->getSubtarget().getRegisterInfo();
+    MachineOperand DstMO = MI.getOperand(0);
+    MachineOperand SrcMO = MI.getOperand(1);
+    unsigned SrcReg = SrcMO.getReg();
+    if (!TRI->isSubRegister(DstMO.getReg(), SrcReg)) {
+      unsigned SrcSuperReg = TRI->getMatchingSuperReg(SrcReg, AArch64::psub,
+                                                      &AArch64::PPRRegClass);
+      BuildMI(MBB, MBBI, MI.getDebugLoc(), TII->get(AArch64::ORR_PPzPP))
+          .add(DstMO)
+          .addReg(SrcSuperReg)
+          .addReg(SrcSuperReg)
+          .addReg(SrcSuperReg);
+    }
+    MI.eraseFromParent();
+    return true;
+  }
   case AArch64::BSPv8i8:
   case AArch64::BSPv16i8: {
     Register DstReg = MI.getOperand(0).getReg();

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

@@ -4805,6 +4805,7 @@ void AArch64InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
   bool Offset = true;
   MCRegister PNRReg = MCRegister::NoRegister;
   unsigned StackID = TargetStackID::Default;
+  const TargetInstrInfo *TII = MBB.getParent()->getSubtarget().getInstrInfo();
   switch (TRI->getSpillSize(*RC)) {
   case 1:
     if (AArch64::FPR8RegClass.hasSubClassEq(RC))
@@ -4823,8 +4824,9 @@ void AArch64InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
              "Unexpected register store without SVE2p1 or SME2");
       if (SrcReg.isVirtual()) {
         auto NewSrcReg =
-            MF.getRegInfo().createVirtualRegister(&AArch64::PPRRegClass);
-        BuildMI(MBB, MBBI, DebugLoc(), get(TargetOpcode::COPY), NewSrcReg)
+            MF.getRegInfo().createVirtualRegister(&AArch64::PPR_p8to15RegClass);
+        BuildMI(MBB, MBBI, DebugLoc(), TII->get(AArch64::ConvertPNRtoPPR),
+                NewSrcReg)
             .addReg(SrcReg);
         SrcReg = NewSrcReg;
       } else

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

@@ -2313,6 +2313,8 @@ let Predicates = [HasBF16, HasSVEorSME] in {
   defm BFCVTNT_ZPmZ : sve_bfloat_convert<0b0, "bfcvtnt", int_aarch64_sve_fcvtnt_bf16f32>;
 } // End HasBF16, HasSVEorSME
 
+def ConvertPNRtoPPR : Pseudo<(outs PPRAny:$Pd), (ins PNRAny:$Pm), []>, Sched<[]>;
+
 let Predicates = [HasSVEorSME] in {
   // InstAliases
   def : InstAlias<"mov $Zd, $Zn",

llvm/test/CodeGen/AArch64/spillfill-sve-different-predicate.mir

@@ -0,0 +1,30 @@
+# RUN: llc -mtriple=aarch64-linux-gnu -start-after=virtregrewriter -stop-after=aarch64-expand-pseudo -mattr=+sme2 -verify-machineinstrs -o - %s \
+# RUN: | FileCheck %s
+
+--- |
+  target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+  target triple = "aarch64-unknown-linux-gnu"
+
+  define void @test_convert_different_reg() #0 { entry: unreachable }
+
+  attributes #0 = { "target-features"="+sme2" }
+
+---
+name:            test_convert_different_reg
+tracksRegLiveness: true
+body: |
+  bb.0.entry:
+    ; CHECK-LABEL: name: test_convert_different_reg
+    ; CHECK: renamable $pn8 = WHILEGE_CXX_B undef $x0, undef $x0, 0, implicit-def dead $nzcv
+    ; CHECK-NEXT: renamable $p9 = ORR_PPzPP $p8, $p8, $p8
+    ; CHECK-NEXT: STR_PXI killed renamable $p9, $sp, 7
+    ; CHECK-NEXT: renamable $p0 = LDR_PXI $sp, 7
+    early-clobber $sp = frame-setup STRXpre killed $fp, $sp, -16
+    frame-setup CFI_INSTRUCTION escape 0x0f, 0x0c, 0x8f, 0x00, 0x11, 0x10, 0x22, 0x11, 0x08, 0x92, 0x2e, 0x00, 0x1e, 0x22
+    frame-setup CFI_INSTRUCTION offset $w29, -16
+    renamable $pn8 = WHILEGE_CXX_B undef $x0, undef $x0, 0, implicit-def dead $nzcv
+    renamable $p9 = ConvertPNRtoPPR killed renamable $pn8
+    STR_PXI killed renamable $p9, $sp, 7
+    renamable $p0 = LDR_PXI $sp, 7
+    early-clobber $sp, $fp = frame-destroy LDRXpost $sp, 16
+    RET undef $lr

llvm/test/CodeGen/AArch64/spillfill-sve.mir

@@ -213,8 +213,7 @@ body:             |
 
     ; EXPAND-LABEL: name: spills_fills_stack_id_virtreg_pnr
     ; EXPAND: renamable $pn8 = WHILEGE_CXX_B
-    ; EXPAND: $p0 = ORR_PPzPP $p8, $p8, killed $p8
-    ; EXPAND: STR_PXI killed renamable $p0, $sp, 7
+    ; EXPAND: STR_PXI killed renamable $p8, $sp, 7
     ;
     ; EXPAND: renamable $p0 = LDR_PXI $sp, 7
     ; EXPAND: $p8 = ORR_PPzPP $p0, $p0, killed $p0, implicit-def $pn8