[AArch64][SME] Spill p-regs as z-regs when streaming hazards are possible

This patch adds a new option `-aarch64-enable-zpr-predicate-spills`
(which is disabled by default), this option replaces predicate spills
with vector spills in streaming[-compatible] functions.

For example:

```
str	p8, [sp, #7, mul vl]            // 2-byte Folded Spill
// ...
ldr	p8, [sp, #7, mul vl]            // 2-byte Folded Reload
```

Becomes:

```
mov	z0.b, p8/z, #1
str	z0, [sp]                        // 16-byte Folded Spill
// ...
ldr	z0, [sp]                        // 16-byte Folded Reload
ptrue	p4.b
cmpne	p8.b, p4/z, z0.b, #0
```

This is done to avoid streaming memory hazards between FPR/vector and
predicate spills, which currently occupy the same stack area even when
the `-aarch64-stack-hazard-size` flag is set.

This is implemented with two new pseudos SPILL_PPR_TO_ZPR_SLOT_PSEUDO
and FILL_PPR_FROM_ZPR_SLOT_PSEUDO. The expansion of these pseudos
handles scavenging the required registers (z0 in the above example) and,
in the worst case spilling a register to an emergency stack slot in the
expansion. The condition flags are also preserved around the `cmpne`
in case they are live at the expansion point.

Author: MacDue

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

@@ -81,7 +81,7 @@ static cl::opt<unsigned>
 AArch64InstrInfo::AArch64InstrInfo(const AArch64Subtarget &STI)
     : AArch64GenInstrInfo(AArch64::ADJCALLSTACKDOWN, AArch64::ADJCALLSTACKUP,
                           AArch64::CATCHRET),
-      RI(STI.getTargetTriple()), Subtarget(STI) {}
+      RI(STI.getTargetTriple(), STI.getHwMode()), Subtarget(STI) {}
 
 /// GetInstSize - Return the number of bytes of code the specified
 /// instruction may be.  This returns the maximum number of bytes.
@@ -2438,6 +2438,8 @@ unsigned AArch64InstrInfo::getLoadStoreImmIdx(unsigned Opc) {
   case AArch64::STZ2Gi:
   case AArch64::STZGi:
   case AArch64::TAGPstack:
+  case AArch64::SPILL_PPR_TO_ZPR_SLOT_PSEUDO:
+  case AArch64::FILL_PPR_FROM_ZPR_SLOT_PSEUDO:
     return 2;
   case AArch64::LD1B_D_IMM:
   case AArch64::LD1B_H_IMM:
@@ -4223,6 +4225,8 @@ bool AArch64InstrInfo::getMemOpInfo(unsigned Opcode, TypeSize &Scale,
     MinOffset = -256;
     MaxOffset = 254;
     break;
+  case AArch64::SPILL_PPR_TO_ZPR_SLOT_PSEUDO:
+  case AArch64::FILL_PPR_FROM_ZPR_SLOT_PSEUDO:
   case AArch64::LDR_ZXI:
   case AArch64::STR_ZXI:
     Scale = TypeSize::getScalable(16);
@@ -5354,6 +5358,11 @@ void AArch64InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
              "Unexpected register store without SVE store instructions");
       Opc = AArch64::STR_ZXI;
       StackID = TargetStackID::ScalableVector;
+    } else if (AArch64::PPRRegClass.hasSubClassEq(RC)) {
+      assert(Subtarget.isSVEorStreamingSVEAvailable() &&
+             "Unexpected predicate store without SVE store instructions");
+      Opc = AArch64::SPILL_PPR_TO_ZPR_SLOT_PSEUDO;
+      StackID = TargetStackID::ScalableVector;
     }
     break;
   case 24:
@@ -5528,6 +5537,11 @@ void AArch64InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
              "Unexpected register load without SVE load instructions");
       Opc = AArch64::LDR_ZXI;
       StackID = TargetStackID::ScalableVector;
+    } else if (AArch64::PPRRegClass.hasSubClassEq(RC)) {
+      assert(Subtarget.isSVEorStreamingSVEAvailable() &&
+             "Unexpected predicate load without SVE load instructions");
+      Opc = AArch64::FILL_PPR_FROM_ZPR_SLOT_PSEUDO;
+      StackID = TargetStackID::ScalableVector;
     }
     break;
   case 24:

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

@@ -38,8 +38,8 @@ using namespace llvm;
 #define GET_REGINFO_TARGET_DESC
 #include "AArch64GenRegisterInfo.inc"
 
-AArch64RegisterInfo::AArch64RegisterInfo(const Triple &TT)
-    : AArch64GenRegisterInfo(AArch64::LR), TT(TT) {
+AArch64RegisterInfo::AArch64RegisterInfo(const Triple &TT, unsigned HwMode)
+    : AArch64GenRegisterInfo(AArch64::LR, 0, 0, 0, HwMode), TT(TT) {
   AArch64_MC::initLLVMToCVRegMapping(this);
 }
 

llvm/lib/Target/AArch64/AArch64RegisterInfo.cpp

@@ -27,7 +27,7 @@ class AArch64RegisterInfo final : public AArch64GenRegisterInfo {
   const Triple &TT;
 
 public:
-  AArch64RegisterInfo(const Triple &TT);
+  AArch64RegisterInfo(const Triple &TT, unsigned HwMode);
 
   // FIXME: This should be tablegen'd like getDwarfRegNum is
   int getSEHRegNum(unsigned i) const {

llvm/lib/Target/AArch64/AArch64RegisterInfo.h

@@ -979,10 +979,19 @@ class ZPRRegOp <string Suffix, AsmOperandClass C, ElementSizeEnum Size,
 //******************************************************************************
 
 // SVE predicate register classes.
+
+// Note: This hardware mode is enabled in AArch64Subtarget::getHwModeSet()
+// (without the use of the table-gen'd predicates).
+def SMEWithStreamingMemoryHazards : HwMode<"", [Predicate<"false">]>;
+
+def PPRSpillFillRI : RegInfoByHwMode<
+      [DefaultMode,              SMEWithStreamingMemoryHazards],
+      [RegInfo<16,16,16>,        RegInfo<16,128,128>]>;
+
 class PPRClass<int firstreg, int lastreg, int step = 1> : RegisterClass<"AArch64",
                                   [ nxv16i1, nxv8i1, nxv4i1, nxv2i1, nxv1i1 ], 16,
                                   (sequence "P%u", firstreg, lastreg, step)> {
-  let Size = 16;
+  let RegInfos = PPRSpillFillRI;
 }
 
 def PPR    : PPRClass<0, 15> {

llvm/lib/Target/AArch64/AArch64RegisterInfo.td

@@ -86,6 +86,11 @@ static cl::alias AArch64StreamingStackHazardSize(
     cl::desc("alias for -aarch64-streaming-hazard-size"),
     cl::aliasopt(AArch64StreamingHazardSize));
 
+static cl::opt<bool> EnableZPRPredicateSpills(
+    "aarch64-enable-zpr-predicate-spills", cl::init(false), cl::Hidden,
+    cl::desc(
+        "Enables spilling/reloading SVE predicates as data vectors (ZPRs)"));
+
 // Subreg liveness tracking is disabled by default for now until all issues
 // are ironed out. This option allows the feature to be used in tests.
 static cl::opt<bool>
@@ -397,6 +402,23 @@ AArch64Subtarget::AArch64Subtarget(const Triple &TT, StringRef CPU,
   EnableSubregLiveness = EnableSubregLivenessTracking.getValue();
 }
 
+unsigned AArch64Subtarget::getHwModeSet() const {
+  unsigned Modes = 0;
+
+  // Use a special hardware mode in streaming functions with stack hazards.
+  // This changes the spill size (and alignment) for the predicate register
+  // class.
+  //
+  // FIXME: This overrides the table-gen'd `getHwModeSet()` which only looks at
+  // CPU features.
+  if (EnableZPRPredicateSpills.getValue() &&
+      (isStreaming() || isStreamingCompatible())) {
+    Modes |= (1 << 0);
+  }
+
+  return Modes;
+}
+
 const CallLowering *AArch64Subtarget::getCallLowering() const {
   return CallLoweringInfo.get();
 }

llvm/lib/Target/AArch64/AArch64Subtarget.cpp

@@ -130,6 +130,8 @@ class AArch64Subtarget final : public AArch64GenSubtargetInfo {
                    bool IsStreaming = false, bool IsStreamingCompatible = false,
                    bool HasMinSize = false);
 
+  virtual unsigned getHwModeSet() const override;
+
 // Getters for SubtargetFeatures defined in tablegen
 #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER)                    \
   bool GETTER() const { return ATTRIBUTE; }

llvm/lib/Target/AArch64/AArch64Subtarget.h

@@ -59,6 +59,20 @@ def FORM_TRANSPOSED_REG_TUPLE_X4_PSEUDO :
   let hasPostISelHook = 1;
 }
 
+def SPILL_PPR_TO_ZPR_SLOT_PSEUDO :
+  Pseudo<(outs), (ins PPRorPNRAny:$Pt, GPR64sp:$Rn, simm9:$imm9), []>, Sched<[]>
+{
+  let mayStore = 1;
+  let hasSideEffects = 0;
+}
+
+def FILL_PPR_FROM_ZPR_SLOT_PSEUDO :
+  Pseudo<(outs PPRorPNRAny:$Pt), (ins GPR64sp:$Rn, simm9:$imm9), []>, Sched<[]>
+{
+  let mayLoad = 1;
+  let hasSideEffects = 0;
+}
+
 def SDTZALoadStore : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisPtrTy<1>, SDTCisInt<2>]>;
 def AArch64SMELdr : SDNode<"AArch64ISD::SME_ZA_LDR", SDTZALoadStore,
                              [SDNPHasChain, SDNPSideEffect, SDNPMayLoad]>;