| 1 | //===- SMEPeepholeOpt.cpp - SME peephole optimization pass-----------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // This pass tries to remove back-to-back (smstart, smstop) and |
| 9 | // (smstop, smstart) sequences. The pass is conservative when it cannot |
| 10 | // determine that it is safe to remove these sequences. |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "AArch64InstrInfo.h" |
| 14 | #include "AArch64MachineFunctionInfo.h" |
| 15 | #include "AArch64Subtarget.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 18 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 19 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 20 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 21 | |
| 22 | using namespace llvm; |
| 23 | |
| 24 | #define DEBUG_TYPE "aarch64-sme-peephole-opt" |
| 25 | |
| 26 | namespace { |
| 27 | |
| 28 | struct SMEPeepholeOpt : public MachineFunctionPass { |
| 29 | static char ID; |
| 30 | |
| 31 | SMEPeepholeOpt() : MachineFunctionPass(ID) {} |
| 32 | |
| 33 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 34 | |
| 35 | StringRef getPassName() const override { |
| 36 | return "SME Peephole Optimization pass" ; |
| 37 | } |
| 38 | |
| 39 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 40 | AU.setPreservesCFG(); |
| 41 | MachineFunctionPass::getAnalysisUsage(AU); |
| 42 | } |
| 43 | |
| 44 | bool optimizeStartStopPairs(MachineBasicBlock &MBB, |
| 45 | bool &HasRemovedAllSMChanges) const; |
| 46 | bool visitRegSequence(MachineInstr &MI); |
| 47 | }; |
| 48 | |
| 49 | char SMEPeepholeOpt::ID = 0; |
| 50 | |
| 51 | } // end anonymous namespace |
| 52 | |
| 53 | static bool isConditionalStartStop(const MachineInstr *MI) { |
| 54 | return MI->getOpcode() == AArch64::MSRpstatePseudo; |
| 55 | } |
| 56 | |
| 57 | static bool isMatchingStartStopPair(const MachineInstr *MI1, |
| 58 | const MachineInstr *MI2) { |
| 59 | // We only consider the same type of streaming mode change here, i.e. |
| 60 | // start/stop SM, or start/stop ZA pairs. |
| 61 | if (MI1->getOperand(i: 0).getImm() != MI2->getOperand(i: 0).getImm()) |
| 62 | return false; |
| 63 | |
| 64 | // One must be 'start', the other must be 'stop' |
| 65 | if (MI1->getOperand(i: 1).getImm() == MI2->getOperand(i: 1).getImm()) |
| 66 | return false; |
| 67 | |
| 68 | bool IsConditional = isConditionalStartStop(MI: MI2); |
| 69 | if (isConditionalStartStop(MI: MI1) != IsConditional) |
| 70 | return false; |
| 71 | |
| 72 | if (!IsConditional) |
| 73 | return true; |
| 74 | |
| 75 | // Check to make sure the conditional start/stop pairs are identical. |
| 76 | if (MI1->getOperand(i: 2).getImm() != MI2->getOperand(i: 2).getImm()) |
| 77 | return false; |
| 78 | |
| 79 | // Ensure reg masks are identical. |
| 80 | if (MI1->getOperand(i: 4).getRegMask() != MI2->getOperand(i: 4).getRegMask()) |
| 81 | return false; |
| 82 | |
| 83 | // This optimisation is unlikely to happen in practice for conditional |
| 84 | // smstart/smstop pairs as the virtual registers for pstate.sm will always |
| 85 | // be different. |
| 86 | // TODO: For this optimisation to apply to conditional smstart/smstop, |
| 87 | // this pass will need to do more work to remove redundant calls to |
| 88 | // __arm_sme_state. |
| 89 | |
| 90 | // Only consider conditional start/stop pairs which read the same register |
| 91 | // holding the original value of pstate.sm, as some conditional start/stops |
| 92 | // require the state on entry to the function. |
| 93 | if (MI1->getOperand(i: 3).isReg() && MI2->getOperand(i: 3).isReg()) { |
| 94 | Register Reg1 = MI1->getOperand(i: 3).getReg(); |
| 95 | Register Reg2 = MI2->getOperand(i: 3).getReg(); |
| 96 | if (Reg1.isPhysical() || Reg2.isPhysical() || Reg1 != Reg2) |
| 97 | return false; |
| 98 | } |
| 99 | |
| 100 | return true; |
| 101 | } |
| 102 | |
| 103 | static bool ChangesStreamingMode(const MachineInstr *MI) { |
| 104 | assert((MI->getOpcode() == AArch64::MSRpstatesvcrImm1 || |
| 105 | MI->getOpcode() == AArch64::MSRpstatePseudo) && |
| 106 | "Expected MI to be a smstart/smstop instruction" ); |
| 107 | return MI->getOperand(i: 0).getImm() == AArch64SVCR::SVCRSM || |
| 108 | MI->getOperand(i: 0).getImm() == AArch64SVCR::SVCRSMZA; |
| 109 | } |
| 110 | |
| 111 | static bool isSVERegOp(const TargetRegisterInfo &TRI, |
| 112 | const MachineRegisterInfo &MRI, |
| 113 | const MachineOperand &MO) { |
| 114 | if (!MO.isReg()) |
| 115 | return false; |
| 116 | |
| 117 | Register R = MO.getReg(); |
| 118 | if (R.isPhysical()) |
| 119 | return llvm::any_of(Range: TRI.subregs_inclusive(Reg: R), P: [](const MCPhysReg &SR) { |
| 120 | return AArch64::ZPRRegClass.contains(Reg: SR) || |
| 121 | AArch64::PPRRegClass.contains(Reg: SR); |
| 122 | }); |
| 123 | |
| 124 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: R); |
| 125 | return TRI.getCommonSubClass(A: &AArch64::ZPRRegClass, B: RC) || |
| 126 | TRI.getCommonSubClass(A: &AArch64::PPRRegClass, B: RC); |
| 127 | } |
| 128 | |
| 129 | bool SMEPeepholeOpt::optimizeStartStopPairs( |
| 130 | MachineBasicBlock &MBB, bool &HasRemovedAllSMChanges) const { |
| 131 | const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
| 132 | const TargetRegisterInfo &TRI = |
| 133 | *MBB.getParent()->getSubtarget().getRegisterInfo(); |
| 134 | |
| 135 | bool Changed = false; |
| 136 | MachineInstr *Prev = nullptr; |
| 137 | SmallVector<MachineInstr *, 4> ToBeRemoved; |
| 138 | |
| 139 | // Convenience function to reset the matching of a sequence. |
| 140 | auto Reset = [&]() { |
| 141 | Prev = nullptr; |
| 142 | ToBeRemoved.clear(); |
| 143 | }; |
| 144 | |
| 145 | // Walk through instructions in the block trying to find pairs of smstart |
| 146 | // and smstop nodes that cancel each other out. We only permit a limited |
| 147 | // set of instructions to appear between them, otherwise we reset our |
| 148 | // tracking. |
| 149 | unsigned NumSMChanges = 0; |
| 150 | unsigned NumSMChangesRemoved = 0; |
| 151 | for (MachineInstr &MI : make_early_inc_range(Range&: MBB)) { |
| 152 | switch (MI.getOpcode()) { |
| 153 | case AArch64::MSRpstatesvcrImm1: |
| 154 | case AArch64::MSRpstatePseudo: { |
| 155 | if (ChangesStreamingMode(MI: &MI)) |
| 156 | NumSMChanges++; |
| 157 | |
| 158 | if (!Prev) |
| 159 | Prev = &MI; |
| 160 | else if (isMatchingStartStopPair(MI1: Prev, MI2: &MI)) { |
| 161 | // If they match, we can remove them, and possibly any instructions |
| 162 | // that we marked for deletion in between. |
| 163 | Prev->eraseFromParent(); |
| 164 | MI.eraseFromParent(); |
| 165 | for (MachineInstr *TBR : ToBeRemoved) |
| 166 | TBR->eraseFromParent(); |
| 167 | ToBeRemoved.clear(); |
| 168 | Prev = nullptr; |
| 169 | Changed = true; |
| 170 | NumSMChangesRemoved += 2; |
| 171 | } else { |
| 172 | Reset(); |
| 173 | Prev = &MI; |
| 174 | } |
| 175 | continue; |
| 176 | } |
| 177 | default: |
| 178 | if (!Prev) |
| 179 | // Avoid doing expensive checks when Prev is nullptr. |
| 180 | continue; |
| 181 | break; |
| 182 | } |
| 183 | |
| 184 | // Test if the instructions in between the start/stop sequence are agnostic |
| 185 | // of streaming mode. If not, the algorithm should reset. |
| 186 | switch (MI.getOpcode()) { |
| 187 | default: |
| 188 | Reset(); |
| 189 | break; |
| 190 | case AArch64::COALESCER_BARRIER_FPR16: |
| 191 | case AArch64::COALESCER_BARRIER_FPR32: |
| 192 | case AArch64::COALESCER_BARRIER_FPR64: |
| 193 | case AArch64::COALESCER_BARRIER_FPR128: |
| 194 | case AArch64::COPY: |
| 195 | // These instructions should be safe when executed on their own, but |
| 196 | // the code remains conservative when SVE registers are used. There may |
| 197 | // exist subtle cases where executing a COPY in a different mode results |
| 198 | // in different behaviour, even if we can't yet come up with any |
| 199 | // concrete example/test-case. |
| 200 | if (isSVERegOp(TRI, MRI, MO: MI.getOperand(i: 0)) || |
| 201 | isSVERegOp(TRI, MRI, MO: MI.getOperand(i: 1))) |
| 202 | Reset(); |
| 203 | break; |
| 204 | case AArch64::ADJCALLSTACKDOWN: |
| 205 | case AArch64::ADJCALLSTACKUP: |
| 206 | case AArch64::ANDXri: |
| 207 | case AArch64::ADDXri: |
| 208 | // We permit these as they don't generate SVE/NEON instructions. |
| 209 | break; |
| 210 | case AArch64::VGRestorePseudo: |
| 211 | case AArch64::VGSavePseudo: |
| 212 | // When the smstart/smstop are removed, we should also remove |
| 213 | // the pseudos that save/restore the VG value for CFI info. |
| 214 | ToBeRemoved.push_back(Elt: &MI); |
| 215 | break; |
| 216 | case AArch64::MSRpstatesvcrImm1: |
| 217 | case AArch64::MSRpstatePseudo: |
| 218 | llvm_unreachable("Should have been handled" ); |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | HasRemovedAllSMChanges = |
| 223 | NumSMChanges && (NumSMChanges == NumSMChangesRemoved); |
| 224 | return Changed; |
| 225 | } |
| 226 | |
| 227 | // Using the FORM_TRANSPOSED_REG_TUPLE pseudo can improve register allocation |
| 228 | // of multi-vector intrinsics. However, the pseudo should only be emitted if |
| 229 | // the input registers of the REG_SEQUENCE are copy nodes where the source |
| 230 | // register is in a StridedOrContiguous class. For example: |
| 231 | // |
| 232 | // %3:zpr2stridedorcontiguous = LD1B_2Z_IMM_PSEUDO .. |
| 233 | // %4:zpr = COPY %3.zsub1:zpr2stridedorcontiguous |
| 234 | // %5:zpr = COPY %3.zsub0:zpr2stridedorcontiguous |
| 235 | // %6:zpr2stridedorcontiguous = LD1B_2Z_PSEUDO .. |
| 236 | // %7:zpr = COPY %6.zsub1:zpr2stridedorcontiguous |
| 237 | // %8:zpr = COPY %6.zsub0:zpr2stridedorcontiguous |
| 238 | // %9:zpr2mul2 = REG_SEQUENCE %5:zpr, %subreg.zsub0, %8:zpr, %subreg.zsub1 |
| 239 | // |
| 240 | // -> %9:zpr2mul2 = FORM_TRANSPOSED_REG_TUPLE_X2_PSEUDO %5:zpr, %8:zpr |
| 241 | // |
| 242 | bool SMEPeepholeOpt::visitRegSequence(MachineInstr &MI) { |
| 243 | assert(MI.getMF()->getRegInfo().isSSA() && "Expected to be run on SSA form!" ); |
| 244 | |
| 245 | MachineRegisterInfo &MRI = MI.getMF()->getRegInfo(); |
| 246 | switch (MRI.getRegClass(Reg: MI.getOperand(i: 0).getReg())->getID()) { |
| 247 | case AArch64::ZPR2RegClassID: |
| 248 | case AArch64::ZPR4RegClassID: |
| 249 | case AArch64::ZPR2Mul2RegClassID: |
| 250 | case AArch64::ZPR4Mul4RegClassID: |
| 251 | break; |
| 252 | default: |
| 253 | return false; |
| 254 | } |
| 255 | |
| 256 | // The first operand is the register class created by the REG_SEQUENCE. |
| 257 | // Each operand pair after this consists of a vreg + subreg index, so |
| 258 | // for example a sequence of 2 registers will have a total of 5 operands. |
| 259 | if (MI.getNumOperands() != 5 && MI.getNumOperands() != 9) |
| 260 | return false; |
| 261 | |
| 262 | MCRegister SubReg = MCRegister::NoRegister; |
| 263 | for (unsigned I = 1; I < MI.getNumOperands(); I += 2) { |
| 264 | MachineOperand &MO = MI.getOperand(i: I); |
| 265 | |
| 266 | MachineOperand *Def = MRI.getOneDef(Reg: MO.getReg()); |
| 267 | if (!Def || !Def->getParent()->isCopy()) |
| 268 | return false; |
| 269 | |
| 270 | const MachineOperand &CopySrc = Def->getParent()->getOperand(i: 1); |
| 271 | unsigned OpSubReg = CopySrc.getSubReg(); |
| 272 | if (SubReg == MCRegister::NoRegister) |
| 273 | SubReg = OpSubReg; |
| 274 | |
| 275 | MachineOperand *CopySrcOp = MRI.getOneDef(Reg: CopySrc.getReg()); |
| 276 | if (!CopySrcOp || !CopySrcOp->isReg() || OpSubReg != SubReg || |
| 277 | CopySrcOp->getReg().isPhysical()) |
| 278 | return false; |
| 279 | |
| 280 | const TargetRegisterClass *CopySrcClass = |
| 281 | MRI.getRegClass(Reg: CopySrcOp->getReg()); |
| 282 | if (CopySrcClass != &AArch64::ZPR2StridedOrContiguousRegClass && |
| 283 | CopySrcClass != &AArch64::ZPR4StridedOrContiguousRegClass) |
| 284 | return false; |
| 285 | } |
| 286 | |
| 287 | unsigned Opc = MI.getNumOperands() == 5 |
| 288 | ? AArch64::FORM_TRANSPOSED_REG_TUPLE_X2_PSEUDO |
| 289 | : AArch64::FORM_TRANSPOSED_REG_TUPLE_X4_PSEUDO; |
| 290 | |
| 291 | const TargetInstrInfo *TII = |
| 292 | MI.getMF()->getSubtarget<AArch64Subtarget>().getInstrInfo(); |
| 293 | MachineInstrBuilder MIB = BuildMI(BB&: *MI.getParent(), I&: MI, MIMD: MI.getDebugLoc(), |
| 294 | MCID: TII->get(Opcode: Opc), DestReg: MI.getOperand(i: 0).getReg()); |
| 295 | for (unsigned I = 1; I < MI.getNumOperands(); I += 2) |
| 296 | MIB.addReg(RegNo: MI.getOperand(i: I).getReg()); |
| 297 | |
| 298 | MI.eraseFromParent(); |
| 299 | return true; |
| 300 | } |
| 301 | |
| 302 | INITIALIZE_PASS(SMEPeepholeOpt, "aarch64-sme-peephole-opt" , |
| 303 | "SME Peephole Optimization" , false, false) |
| 304 | |
| 305 | bool SMEPeepholeOpt::runOnMachineFunction(MachineFunction &MF) { |
| 306 | if (skipFunction(F: MF.getFunction())) |
| 307 | return false; |
| 308 | |
| 309 | if (!MF.getSubtarget<AArch64Subtarget>().hasSME()) |
| 310 | return false; |
| 311 | |
| 312 | assert(MF.getRegInfo().isSSA() && "Expected to be run on SSA form!" ); |
| 313 | |
| 314 | bool Changed = false; |
| 315 | bool FunctionHasAllSMChangesRemoved = false; |
| 316 | |
| 317 | // Even if the block lives in a function with no SME attributes attached we |
| 318 | // still have to analyze all the blocks because we may call a streaming |
| 319 | // function that requires smstart/smstop pairs. |
| 320 | for (MachineBasicBlock &MBB : MF) { |
| 321 | bool BlockHasAllSMChangesRemoved; |
| 322 | Changed |= optimizeStartStopPairs(MBB, HasRemovedAllSMChanges&: BlockHasAllSMChangesRemoved); |
| 323 | FunctionHasAllSMChangesRemoved |= BlockHasAllSMChangesRemoved; |
| 324 | |
| 325 | if (MF.getSubtarget<AArch64Subtarget>().isStreaming()) { |
| 326 | for (MachineInstr &MI : make_early_inc_range(Range&: MBB)) |
| 327 | if (MI.getOpcode() == AArch64::REG_SEQUENCE) |
| 328 | Changed |= visitRegSequence(MI); |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>(); |
| 333 | if (FunctionHasAllSMChangesRemoved) |
| 334 | AFI->setHasStreamingModeChanges(false); |
| 335 | |
| 336 | return Changed; |
| 337 | } |
| 338 | |
| 339 | FunctionPass *llvm::createSMEPeepholeOptPass() { return new SMEPeepholeOpt(); } |
| 340 | |