[llvm] [AArch64][SME] Spill p-regs as z-regs when streaming hazards are possible (PR #123752)

[Sander de Smalen via llvm-commits]

================
@@ -4155,8 +4168,297 @@ int64_t AArch64FrameLowering::assignSVEStackObjectOffsets(
                                         true);
 }
 
+/// Attempts to scavenge a register from \p ScavengeableRegs given the used
+/// registers in \p UsedRegs.
+static Register tryScavengeRegister(LiveRegUnits const &UsedRegs,
+                                    BitVector const &ScavengeableRegs) {
+  for (auto Reg : ScavengeableRegs.set_bits()) {
+    if (UsedRegs.available(Reg))
+      return Reg;
+  }
+  return AArch64::NoRegister;
+}
+
+/// Propagates frame-setup/destroy flags from \p SourceMI to all instructions in
+/// \p MachineInstrs.
+static void propagateFrameFlags(MachineInstr &SourceMI,
+                                ArrayRef<MachineInstr *> MachineInstrs) {
+  for (MachineInstr *MI : MachineInstrs) {
+    if (SourceMI.getFlag(MachineInstr::FrameSetup))
+      MI->setFlag(MachineInstr::FrameSetup);
+    if (SourceMI.getFlag(MachineInstr::FrameDestroy))
+      MI->setFlag(MachineInstr::FrameDestroy);
+  }
+}
+
+/// RAII helper class for scavenging or spilling a register. On construction
+/// attempts to find a free register of class \p RC (given \p UsedRegs and \p
+/// AllocatableRegs), if no register can be found spills \p SpillCandidate to \p
+/// MaybeSpillFI to free a register. The free'd register is returned via the \p
+/// FreeReg output parameter. On destruction, if there is a spill, its previous
+/// value is reloaded. The spilling and scavenging is only valid at the
+/// insertion point \p MBBI, this class should _not_ be used in places that
+/// create or manipulate basic blocks, moving the expected insertion point.
+struct ScopedScavengeOrSpill {
+  ScopedScavengeOrSpill(const ScopedScavengeOrSpill &) = delete;
+  ScopedScavengeOrSpill(ScopedScavengeOrSpill &&) = delete;
+
+  ScopedScavengeOrSpill(MachineFunction &MF, MachineBasicBlock &MBB,
+                        MachineBasicBlock::iterator MBBI, Register &FreeReg,
+                        Register SpillCandidate, const TargetRegisterClass &RC,
+                        LiveRegUnits const &UsedRegs,
+                        BitVector const &AllocatableRegs,
+                        std::optional<int> *MaybeSpillFI)
+      : MBB(MBB), MBBI(MBBI), RC(RC), TII(static_cast<const AArch64InstrInfo &>(
+                                          *MF.getSubtarget().getInstrInfo())),
+        TRI(*MF.getSubtarget().getRegisterInfo()) {
+    FreeReg = tryScavengeRegister(UsedRegs, AllocatableRegs);
+    if (FreeReg != AArch64::NoRegister)
+      return;
+    assert(MaybeSpillFI && "Expected emergency spill slot FI information "
+                           "(attempted to spill in prologue/epilogue?)");
+    if (!MaybeSpillFI->has_value()) {
+      MachineFrameInfo &MFI = MF.getFrameInfo();
+      *MaybeSpillFI = MFI.CreateSpillStackObject(TRI.getSpillSize(RC),
+                                                 TRI.getSpillAlign(RC));
+    }
+    FreeReg = SpilledReg = SpillCandidate;
+    SpillFI = MaybeSpillFI->value();
+    TII.storeRegToStackSlot(MBB, MBBI, SpilledReg, false, SpillFI, &RC, &TRI,
+                            Register());
+  }
+
+  bool hasSpilled() const { return SpilledReg != AArch64::NoRegister; }
+
+  ~ScopedScavengeOrSpill() {
+    if (hasSpilled())
+      TII.loadRegFromStackSlot(MBB, MBBI, SpilledReg, SpillFI, &RC, &TRI,
+                               Register());
+  }
+
+private:
+  MachineBasicBlock &MBB;
+  MachineBasicBlock::iterator MBBI;
+  const TargetRegisterClass &RC;
+  const AArch64InstrInfo &TII;
+  const TargetRegisterInfo &TRI;
+  Register SpilledReg = AArch64::NoRegister;
+  int SpillFI = -1;
+};
+
+/// Emergency stack slots for expanding SPILL_PPR_TO_ZPR_SLOT_PSEUDO and
+/// FILL_PPR_FROM_ZPR_SLOT_PSEUDO.
+struct EmergencyStackSlots {
+  std::optional<int> ZPRSpillFI;
+  std::optional<int> PPRSpillFI;
+  std::optional<int> GPRSpillFI;
+};
+
+/// Registers available for scavenging (ZPR, PPR3b, GPR).
+struct ScavengeableRegs {
+  BitVector ZPRRegs;
+  BitVector PPR3bRegs;
+  BitVector GPRRegs;
+};
+
+static bool isInPrologueOrEpilogue(const MachineInstr &MI) {
+  return MI.getFlag(MachineInstr::FrameSetup) ||
+         MI.getFlag(MachineInstr::FrameDestroy);
+}
+
+/// Expands:
+/// ```
+/// SPILL_PPR_TO_ZPR_SLOT_PSEUDO $p0, %stack.0, 0
+/// ```
+/// To:
+/// ```
+/// $z0 = CPY_ZPzI_B $p0, 1, 0
+/// STR_ZXI $z0, $stack.0, 0
+/// ```
+/// While ensuring a ZPR ($z0 in this example) is free for the predicate (
+/// spilling if necessary).
+static void expandSpillPPRToZPRSlotPseudo(MachineBasicBlock &MBB,
+                                          MachineInstr &MI,
+                                          const TargetRegisterInfo &TRI,
+                                          LiveRegUnits const &UsedRegs,
+                                          ScavengeableRegs const &SR,
+                                          EmergencyStackSlots &SpillSlots) {
+  MachineFunction &MF = *MBB.getParent();
+  auto *TII =
+      static_cast<const AArch64InstrInfo *>(MF.getSubtarget().getInstrInfo());
+
+  Register ZPredReg = AArch64::NoRegister;
+  ScopedScavengeOrSpill FindZPRReg(
+      MF, MBB, MachineBasicBlock::iterator(MI), ZPredReg, AArch64::Z0,
+      AArch64::ZPRRegClass, UsedRegs, SR.ZPRRegs,
+      isInPrologueOrEpilogue(MI) ? nullptr : &SpillSlots.ZPRSpillFI);
+
+  SmallVector<MachineInstr *, 2> MachineInstrs;
+  const DebugLoc &DL = MI.getDebugLoc();
+  MachineInstrs.push_back(BuildMI(MBB, MI, DL, TII->get(AArch64::CPY_ZPzI_B))
+                              .addReg(ZPredReg, RegState::Define)
+                              .add(MI.getOperand(0))
+                              .addImm(1)
+                              .addImm(0)
+                              .getInstr());
+  MachineInstrs.push_back(BuildMI(MBB, MI, DL, TII->get(AArch64::STR_ZXI))
+                              .addReg(ZPredReg)
+                              .add(MI.getOperand(1))
+                              .addImm(MI.getOperand(2).getImm())
+                              .setMemRefs(MI.memoperands())
+                              .getInstr());
+  propagateFrameFlags(MI, MachineInstrs);
+}
+
+/// Expands:
+/// ```
+/// $p0 = FILL_PPR_FROM_ZPR_SLOT_PSEUDO %stack.0, 0
+/// ```
+/// To:
+/// ```
+/// $z0 = LDR_ZXI %stack.0, 0
+/// $p0 = PTRUE_B 31, implicit $vg
+/// $p0 = CMPNE_PPzZI_B $p0, $z0, 0, implicit-def $nzcv, implicit-def $nzcv
+/// ```
+/// While ensuring a ZPR ($z0 in this example) is free for the predicate (
+/// spilling if necessary). If the status flags are in use at the point of
+/// expansion they are preserved (by moving them to/from a GPR). This may cause
+/// an additional spill if no GPR is free at the expansion point.
+static bool expandFillPPRFromZPRSlotPseudo(MachineBasicBlock &MBB,
+                                           MachineInstr &MI,
+                                           const TargetRegisterInfo &TRI,
+                                           LiveRegUnits const &UsedRegs,
+                                           ScavengeableRegs const &SR,
+                                           EmergencyStackSlots &SpillSlots) {
+  MachineFunction &MF = *MBB.getParent();
+  auto *TII =
+      static_cast<const AArch64InstrInfo *>(MF.getSubtarget().getInstrInfo());
+
+  Register ZPredReg = AArch64::NoRegister;
+  ScopedScavengeOrSpill FindZPRReg(
----------------
sdesmalen-arm wrote:

nit: Just a thought; you could add a `operator*()` to `ScopedScavengeOrSpill` such that you can query the free(d) register by doing something like:
```
ScopedScavengeOrSpill ZPRReg(...);

  ... SomeUseOf(*ZPRReg) ...
```

That removes the need for an explicit `Register` variable, and makes the code a bit simpler to follow, because at the moment `FindZPPRReg` doesn't have any further uses (and the RAII behaviour is hidden).

https://github.com/llvm/llvm-project/pull/123752