| 1 | //=- AArch64ConditionOptimizer.cpp - Remove useless comparisons for AArch64 -=// |
| 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 | // |
| 9 | // |
| 10 | // This pass tries to make consecutive comparisons of values use the same |
| 11 | // operands to allow the CSE pass to remove duplicate instructions. It adjusts |
| 12 | // comparisons with immediate values by converting between inclusive and |
| 13 | // exclusive forms (GE <-> GT, LE <-> LT) and correcting immediate values to |
| 14 | // make them equal. |
| 15 | // |
| 16 | // The pass handles: |
| 17 | // * Cross-block: SUBS/ADDS followed by conditional branches |
| 18 | // * Intra-block: Select-family conditional instructions |
| 19 | // |
| 20 | // |
| 21 | // Consider the following example in C: |
| 22 | // |
| 23 | // if ((a < 5 && ...) || (a > 5 && ...)) { |
| 24 | // ~~~~~ ~~~~~ |
| 25 | // ^ ^ |
| 26 | // x y |
| 27 | // |
| 28 | // Here both "x" and "y" expressions compare "a" with "5". When "x" evaluates |
| 29 | // to "false", "y" can just check flags set by the first comparison. As a |
| 30 | // result of the canonicalization employed by |
| 31 | // SelectionDAGBuilder::visitSwitchCase, DAGCombine, and other target-specific |
| 32 | // code, assembly ends up in the form that is not CSE friendly: |
| 33 | // |
| 34 | // ... |
| 35 | // cmp w8, #4 |
| 36 | // b.gt .LBB0_3 |
| 37 | // ... |
| 38 | // .LBB0_3: |
| 39 | // cmp w8, #6 |
| 40 | // b.lt .LBB0_6 |
| 41 | // ... |
| 42 | // |
| 43 | // Same assembly after the pass: |
| 44 | // |
| 45 | // ... |
| 46 | // cmp w8, #5 |
| 47 | // b.ge .LBB0_3 |
| 48 | // ... |
| 49 | // .LBB0_3: |
| 50 | // cmp w8, #5 // <-- CSE pass removes this instruction |
| 51 | // b.le .LBB0_6 |
| 52 | // ... |
| 53 | // |
| 54 | // See optimizeCrossBlock() and optimizeIntraBlock() for implementation details. |
| 55 | // |
| 56 | // TODO: maybe handle TBNZ/TBZ the same way as CMP when used instead for "a < 0" |
| 57 | // TODO: For cross-block: |
| 58 | // - allow second branching to be anything if it doesn't require adjusting |
| 59 | // |
| 60 | // Cross-block optimizeCrossBlock() handles four head/true-successor |
| 61 | // combinations: |
| 62 | // Bcc (head) + Bcc (true) -- original case |
| 63 | // Select (head) + Bcc (true) -- head ends with CSEL/CSET/etc. |
| 64 | // Bcc (head) + Select (true) -- true-successor ends with CSEL/CSET/etc. |
| 65 | // Select (head) + Select (true) -- both blocks end with a select |
| 66 | // |
| 67 | //===----------------------------------------------------------------------===// |
| 68 | |
| 69 | #include "AArch64.h" |
| 70 | #include "AArch64Subtarget.h" |
| 71 | #include "MCTargetDesc/AArch64AddressingModes.h" |
| 72 | #include "Utils/AArch64BaseInfo.h" |
| 73 | #include "llvm/ADT/ArrayRef.h" |
| 74 | #include "llvm/ADT/DepthFirstIterator.h" |
| 75 | #include "llvm/ADT/SmallVector.h" |
| 76 | #include "llvm/ADT/Statistic.h" |
| 77 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 78 | #include "llvm/CodeGen/MachineDominators.h" |
| 79 | #include "llvm/CodeGen/MachineFunction.h" |
| 80 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 81 | #include "llvm/CodeGen/MachineInstr.h" |
| 82 | #include "llvm/CodeGen/MachineOperand.h" |
| 83 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 84 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 85 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 86 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 87 | #include "llvm/InitializePasses.h" |
| 88 | #include "llvm/Pass.h" |
| 89 | #include "llvm/Support/Debug.h" |
| 90 | #include "llvm/Support/ErrorHandling.h" |
| 91 | #include "llvm/Support/raw_ostream.h" |
| 92 | #include <cassert> |
| 93 | #include <cstdlib> |
| 94 | |
| 95 | using namespace llvm; |
| 96 | |
| 97 | #define DEBUG_TYPE "aarch64-condopt" |
| 98 | |
| 99 | STATISTIC(NumConditionsAdjusted, "Number of conditions adjusted" ); |
| 100 | |
| 101 | namespace { |
| 102 | |
| 103 | /// Bundles the parameters needed to adjust a comparison instruction. |
| 104 | struct CmpInfo { |
| 105 | int Imm; |
| 106 | unsigned Opc; |
| 107 | AArch64CC::CondCode CC; |
| 108 | }; |
| 109 | |
| 110 | class AArch64ConditionOptimizerImpl { |
| 111 | /// Represents a comparison instruction paired with its consuming |
| 112 | /// conditional instruction |
| 113 | struct CmpCondPair { |
| 114 | MachineInstr *CmpMI; |
| 115 | MachineInstr *CondMI; |
| 116 | AArch64CC::CondCode CC; |
| 117 | |
| 118 | int getImm() const { return CmpMI->getOperand(i: 2).getImm(); } |
| 119 | unsigned getOpc() const { return CmpMI->getOpcode(); } |
| 120 | }; |
| 121 | |
| 122 | const AArch64InstrInfo *TII; |
| 123 | const TargetRegisterInfo *TRI; |
| 124 | MachineDominatorTree *DomTree; |
| 125 | const MachineRegisterInfo *MRI; |
| 126 | |
| 127 | public: |
| 128 | bool run(MachineFunction &MF, MachineDominatorTree &MDT); |
| 129 | |
| 130 | private: |
| 131 | bool canAdjustCmp(MachineInstr &CmpMI); |
| 132 | bool registersMatch(MachineInstr *FirstMI, MachineInstr *SecondMI); |
| 133 | bool nzcvLivesOut(MachineBasicBlock *MBB); |
| 134 | MachineInstr *getBccTerminator(MachineBasicBlock *MBB); |
| 135 | MachineInstr *findAdjustableCmp(MachineInstr *CondMI); |
| 136 | CmpInfo getAdjustedCmpInfo(MachineInstr *CmpMI, AArch64CC::CondCode Cmp); |
| 137 | void updateCmpInstr(MachineInstr *CmpMI, int NewImm, unsigned NewOpc); |
| 138 | void updateCondInstr(MachineInstr *CondMI, AArch64CC::CondCode NewCC); |
| 139 | void applyCmpAdjustment(CmpCondPair &Pair, const CmpInfo &Info); |
| 140 | bool commitPendingPair(std::optional<CmpCondPair> &PendingPair, |
| 141 | SmallDenseMap<Register, CmpCondPair> &PairsByReg); |
| 142 | bool tryOptimizePair(CmpCondPair &First, CmpCondPair &Second); |
| 143 | bool optimizeIntraBlock(MachineBasicBlock &MBB); |
| 144 | bool optimizeCrossBlock(MachineBasicBlock &HBB); |
| 145 | std::pair<MachineInstr *, AArch64CC::CondCode> |
| 146 | findCondConsumer(MachineBasicBlock *MBB); |
| 147 | }; |
| 148 | |
| 149 | class AArch64ConditionOptimizerLegacy : public MachineFunctionPass { |
| 150 | public: |
| 151 | static char ID; |
| 152 | AArch64ConditionOptimizerLegacy() : MachineFunctionPass(ID) {} |
| 153 | |
| 154 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 155 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 156 | |
| 157 | StringRef getPassName() const override { |
| 158 | return "AArch64 Condition Optimizer" ; |
| 159 | } |
| 160 | }; |
| 161 | |
| 162 | } // end anonymous namespace |
| 163 | |
| 164 | char AArch64ConditionOptimizerLegacy::ID = 0; |
| 165 | |
| 166 | INITIALIZE_PASS_BEGIN(AArch64ConditionOptimizerLegacy, "aarch64-condopt" , |
| 167 | "AArch64 CondOpt Pass" , false, false) |
| 168 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
| 169 | INITIALIZE_PASS_END(AArch64ConditionOptimizerLegacy, "aarch64-condopt" , |
| 170 | "AArch64 CondOpt Pass" , false, false) |
| 171 | |
| 172 | FunctionPass *llvm::createAArch64ConditionOptimizerLegacyPass() { |
| 173 | return new AArch64ConditionOptimizerLegacy(); |
| 174 | } |
| 175 | |
| 176 | void AArch64ConditionOptimizerLegacy::getAnalysisUsage( |
| 177 | AnalysisUsage &AU) const { |
| 178 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
| 179 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
| 180 | MachineFunctionPass::getAnalysisUsage(AU); |
| 181 | } |
| 182 | |
| 183 | // Verify that the MI's immediate is adjustable and it only sets flags (pure |
| 184 | // cmp) |
| 185 | bool AArch64ConditionOptimizerImpl::canAdjustCmp(MachineInstr &CmpMI) { |
| 186 | unsigned ShiftAmt = AArch64_AM::getShiftValue(Imm: CmpMI.getOperand(i: 3).getImm()); |
| 187 | if (!CmpMI.getOperand(i: 2).isImm()) { |
| 188 | LLVM_DEBUG(dbgs() << "Immediate of cmp is symbolic, " << CmpMI << '\n'); |
| 189 | return false; |
| 190 | } else if (CmpMI.getOperand(i: 2).getImm() << ShiftAmt >= 0xfff) { |
| 191 | LLVM_DEBUG(dbgs() << "Immediate of cmp may be out of range, " << CmpMI |
| 192 | << '\n'); |
| 193 | return false; |
| 194 | } else if (!MRI->use_nodbg_empty(RegNo: CmpMI.getOperand(i: 0).getReg())) { |
| 195 | LLVM_DEBUG(dbgs() << "Destination of cmp is not dead, " << CmpMI << '\n'); |
| 196 | return false; |
| 197 | } |
| 198 | |
| 199 | return true; |
| 200 | } |
| 201 | |
| 202 | // Ensure both compare MIs use the same register, tracing through copies. |
| 203 | bool AArch64ConditionOptimizerImpl::registersMatch(MachineInstr *FirstMI, |
| 204 | MachineInstr *SecondMI) { |
| 205 | Register FirstReg = FirstMI->getOperand(i: 1).getReg(); |
| 206 | Register SecondReg = SecondMI->getOperand(i: 1).getReg(); |
| 207 | Register FirstCmpReg = |
| 208 | FirstReg.isVirtual() ? TRI->lookThruCopyLike(SrcReg: FirstReg, MRI) : FirstReg; |
| 209 | Register SecondCmpReg = |
| 210 | SecondReg.isVirtual() ? TRI->lookThruCopyLike(SrcReg: SecondReg, MRI) : SecondReg; |
| 211 | if (FirstCmpReg != SecondCmpReg) { |
| 212 | LLVM_DEBUG(dbgs() << "CMPs compare different registers\n" ); |
| 213 | return false; |
| 214 | } |
| 215 | |
| 216 | return true; |
| 217 | } |
| 218 | |
| 219 | // Check if NZCV lives out to any successor block. |
| 220 | bool AArch64ConditionOptimizerImpl::nzcvLivesOut(MachineBasicBlock *MBB) { |
| 221 | for (auto *SuccBB : MBB->successors()) { |
| 222 | if (SuccBB->isLiveIn(Reg: AArch64::NZCV)) { |
| 223 | LLVM_DEBUG(dbgs() << "NZCV live into successor " |
| 224 | << printMBBReference(*SuccBB) << " from " |
| 225 | << printMBBReference(*MBB) << '\n'); |
| 226 | return true; |
| 227 | } |
| 228 | } |
| 229 | return false; |
| 230 | } |
| 231 | |
| 232 | // Returns true if the opcode is a comparison instruction (CMP/CMN). |
| 233 | static bool isCmpInstruction(unsigned Opc) { |
| 234 | switch (Opc) { |
| 235 | // cmp is an alias for SUBS with a dead destination register. |
| 236 | case AArch64::SUBSWri: |
| 237 | case AArch64::SUBSXri: |
| 238 | // cmp is an alias for ADDS with a dead destination register. |
| 239 | case AArch64::ADDSWri: |
| 240 | case AArch64::ADDSXri: |
| 241 | return true; |
| 242 | default: |
| 243 | return false; |
| 244 | } |
| 245 | } |
| 246 | |
| 247 | // Returns the Bcc terminator if present, otherwise nullptr. |
| 248 | MachineInstr * |
| 249 | AArch64ConditionOptimizerImpl::getBccTerminator(MachineBasicBlock *MBB) { |
| 250 | MachineBasicBlock::iterator Term = MBB->getFirstTerminator(); |
| 251 | if (Term == MBB->end()) { |
| 252 | LLVM_DEBUG(dbgs() << "No terminator in " << printMBBReference(*MBB) |
| 253 | << '\n'); |
| 254 | return nullptr; |
| 255 | } |
| 256 | |
| 257 | if (Term->getOpcode() != AArch64::Bcc) { |
| 258 | LLVM_DEBUG(dbgs() << "Non-Bcc terminator in " << printMBBReference(*MBB) |
| 259 | << ": " << *Term); |
| 260 | return nullptr; |
| 261 | } |
| 262 | |
| 263 | return &*Term; |
| 264 | } |
| 265 | |
| 266 | // Find the CMP instruction controlling the given conditional instruction and |
| 267 | // ensure it can be adjusted for CSE optimization. Searches backward from |
| 268 | // CondMI, ensuring no NZCV interference. Returns nullptr if no suitable CMP |
| 269 | // is found or if adjustments are not safe. |
| 270 | MachineInstr * |
| 271 | AArch64ConditionOptimizerImpl::findAdjustableCmp(MachineInstr *CondMI) { |
| 272 | assert(CondMI && "CondMI cannot be null" ); |
| 273 | MachineBasicBlock *MBB = CondMI->getParent(); |
| 274 | |
| 275 | // Search backward from the conditional to find the instruction controlling |
| 276 | // it. |
| 277 | for (MachineBasicBlock::iterator B = MBB->begin(), |
| 278 | It = MachineBasicBlock::iterator(CondMI); |
| 279 | It != B;) { |
| 280 | It = prev_nodbg(It, Begin: B); |
| 281 | MachineInstr &I = *It; |
| 282 | assert(!I.isTerminator() && "Spurious terminator" ); |
| 283 | // Ensure there is no use of NZCV between CMP and conditional. |
| 284 | if (I.readsRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) |
| 285 | return nullptr; |
| 286 | |
| 287 | if (isCmpInstruction(Opc: I.getOpcode())) { |
| 288 | if (!canAdjustCmp(CmpMI&: I)) { |
| 289 | return nullptr; |
| 290 | } |
| 291 | return &I; |
| 292 | } |
| 293 | |
| 294 | if (I.modifiesRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) |
| 295 | return nullptr; |
| 296 | } |
| 297 | LLVM_DEBUG(dbgs() << "Flags not defined in " << printMBBReference(*MBB) |
| 298 | << '\n'); |
| 299 | return nullptr; |
| 300 | } |
| 301 | |
| 302 | // Changes opcode adds <-> subs considering register operand width. |
| 303 | static int getComplementOpc(int Opc) { |
| 304 | switch (Opc) { |
| 305 | case AArch64::ADDSWri: return AArch64::SUBSWri; |
| 306 | case AArch64::ADDSXri: return AArch64::SUBSXri; |
| 307 | case AArch64::SUBSWri: return AArch64::ADDSWri; |
| 308 | case AArch64::SUBSXri: return AArch64::ADDSXri; |
| 309 | default: |
| 310 | llvm_unreachable("Unexpected opcode" ); |
| 311 | } |
| 312 | } |
| 313 | |
| 314 | // Changes form of comparison inclusive <-> exclusive. |
| 315 | static AArch64CC::CondCode getAdjustedCmp(AArch64CC::CondCode Cmp) { |
| 316 | switch (Cmp) { |
| 317 | case AArch64CC::GT: |
| 318 | return AArch64CC::GE; |
| 319 | case AArch64CC::GE: |
| 320 | return AArch64CC::GT; |
| 321 | case AArch64CC::LT: |
| 322 | return AArch64CC::LE; |
| 323 | case AArch64CC::LE: |
| 324 | return AArch64CC::LT; |
| 325 | case AArch64CC::HI: |
| 326 | return AArch64CC::HS; |
| 327 | case AArch64CC::HS: |
| 328 | return AArch64CC::HI; |
| 329 | case AArch64CC::LO: |
| 330 | return AArch64CC::LS; |
| 331 | case AArch64CC::LS: |
| 332 | return AArch64CC::LO; |
| 333 | default: |
| 334 | llvm_unreachable("Unexpected condition code" ); |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | // Returns the adjusted immediate, opcode, and condition code for switching |
| 339 | // between inclusive/exclusive forms (GT <-> GE, LT <-> LE). |
| 340 | CmpInfo |
| 341 | AArch64ConditionOptimizerImpl::getAdjustedCmpInfo(MachineInstr *CmpMI, |
| 342 | AArch64CC::CondCode Cmp) { |
| 343 | unsigned Opc = CmpMI->getOpcode(); |
| 344 | |
| 345 | bool IsSigned = Cmp == AArch64CC::GT || Cmp == AArch64CC::GE || |
| 346 | Cmp == AArch64CC::LT || Cmp == AArch64CC::LE; |
| 347 | |
| 348 | // CMN (compare with negative immediate) is an alias to ADDS (as |
| 349 | // "operand - negative" == "operand + positive") |
| 350 | bool Negative = (Opc == AArch64::ADDSWri || Opc == AArch64::ADDSXri); |
| 351 | |
| 352 | int Correction = (Cmp == AArch64CC::GT || Cmp == AArch64CC::HI) ? 1 : -1; |
| 353 | // Negate Correction value for comparison with negative immediate (CMN). |
| 354 | if (Negative) { |
| 355 | Correction = -Correction; |
| 356 | } |
| 357 | |
| 358 | const int OldImm = (int)CmpMI->getOperand(i: 2).getImm(); |
| 359 | const int NewImm = std::abs(x: OldImm + Correction); |
| 360 | |
| 361 | // Bail out on cmn 0 (ADDS with immediate 0). It is a valid instruction but |
| 362 | // doesn't set flags in a way we can safely transform, so skip optimization. |
| 363 | if (OldImm == 0 && Negative) |
| 364 | return {.Imm: OldImm, .Opc: Opc, .CC: Cmp}; |
| 365 | |
| 366 | if ((OldImm == 1 && Negative && Correction == -1) || |
| 367 | (OldImm == 0 && Correction == -1)) { |
| 368 | // If we change opcodes for unsigned comparisons, this means we did an |
| 369 | // unsigned wrap (e.g., 0 wrapping to 0xFFFFFFFF), so return the old cmp. |
| 370 | // Note: For signed comparisons, opcode changes (cmn 1 ↔ cmp 0) are valid. |
| 371 | if (!IsSigned) |
| 372 | return {.Imm: OldImm, .Opc: Opc, .CC: Cmp}; |
| 373 | Opc = getComplementOpc(Opc); |
| 374 | } |
| 375 | |
| 376 | return {.Imm: NewImm, .Opc: Opc, .CC: getAdjustedCmp(Cmp)}; |
| 377 | } |
| 378 | |
| 379 | // Modifies a comparison instruction's immediate and opcode. |
| 380 | void AArch64ConditionOptimizerImpl::updateCmpInstr(MachineInstr *CmpMI, |
| 381 | int NewImm, |
| 382 | unsigned NewOpc) { |
| 383 | CmpMI->getOperand(i: 2).setImm(NewImm); |
| 384 | CmpMI->setDesc(TII->get(Opcode: NewOpc)); |
| 385 | } |
| 386 | |
| 387 | // Modifies the condition code of a conditional instruction. |
| 388 | void AArch64ConditionOptimizerImpl::updateCondInstr(MachineInstr *CondMI, |
| 389 | AArch64CC::CondCode NewCC) { |
| 390 | int CCOpIdx = |
| 391 | AArch64InstrInfo::findCondCodeUseOperandIdxForBranchOrSelect(Instr: *CondMI); |
| 392 | assert(CCOpIdx >= 0 && "Unsupported conditional instruction" ); |
| 393 | CondMI->getOperand(i: CCOpIdx).setImm(NewCC); |
| 394 | ++NumConditionsAdjusted; |
| 395 | } |
| 396 | |
| 397 | // Applies a comparison adjustment to a cmp/cond instruction pair. |
| 398 | void AArch64ConditionOptimizerImpl::applyCmpAdjustment(CmpCondPair &Pair, |
| 399 | const CmpInfo &Info) { |
| 400 | updateCmpInstr(CmpMI: Pair.CmpMI, NewImm: Info.Imm, NewOpc: Info.Opc); |
| 401 | updateCondInstr(CondMI: Pair.CondMI, NewCC: Info.CC); |
| 402 | Pair.CC = Info.CC; |
| 403 | } |
| 404 | |
| 405 | // Extracts the condition code from the result of analyzeBranch. |
| 406 | // Returns the CondCode or Invalid if the format is not a simple br.cond. |
| 407 | static AArch64CC::CondCode parseCondCode(ArrayRef<MachineOperand> Cond) { |
| 408 | assert(!Cond.empty() && "Expected non-empty condition from analyzeBranch" ); |
| 409 | // A normal br.cond simply has the condition code. |
| 410 | if (Cond[0].getImm() != -1) { |
| 411 | assert(Cond.size() == 1 && "Unknown Cond array format" ); |
| 412 | return (AArch64CC::CondCode)(int)Cond[0].getImm(); |
| 413 | } |
| 414 | return AArch64CC::CondCode::Invalid; |
| 415 | } |
| 416 | |
| 417 | static bool isGreaterThan(AArch64CC::CondCode Cmp) { |
| 418 | return Cmp == AArch64CC::GT || Cmp == AArch64CC::HI; |
| 419 | } |
| 420 | |
| 421 | static bool isLessThan(AArch64CC::CondCode Cmp) { |
| 422 | return Cmp == AArch64CC::LT || Cmp == AArch64CC::LO; |
| 423 | } |
| 424 | |
| 425 | bool AArch64ConditionOptimizerImpl::tryOptimizePair(CmpCondPair &First, |
| 426 | CmpCondPair &Second) { |
| 427 | if (!((isGreaterThan(Cmp: First.CC) || isLessThan(Cmp: First.CC)) && |
| 428 | (isGreaterThan(Cmp: Second.CC) || isLessThan(Cmp: Second.CC)))) |
| 429 | return false; |
| 430 | |
| 431 | int FirstImmTrueValue = First.getImm(); |
| 432 | int SecondImmTrueValue = Second.getImm(); |
| 433 | |
| 434 | // Normalize immediate of CMN (ADDS) instructions |
| 435 | if (First.getOpc() == AArch64::ADDSWri || First.getOpc() == AArch64::ADDSXri) |
| 436 | FirstImmTrueValue = -FirstImmTrueValue; |
| 437 | if (Second.getOpc() == AArch64::ADDSWri || |
| 438 | Second.getOpc() == AArch64::ADDSXri) |
| 439 | SecondImmTrueValue = -SecondImmTrueValue; |
| 440 | |
| 441 | CmpInfo FirstAdj = getAdjustedCmpInfo(CmpMI: First.CmpMI, Cmp: First.CC); |
| 442 | CmpInfo SecondAdj = getAdjustedCmpInfo(CmpMI: Second.CmpMI, Cmp: Second.CC); |
| 443 | |
| 444 | if (((isGreaterThan(Cmp: First.CC) && isLessThan(Cmp: Second.CC)) || |
| 445 | (isLessThan(Cmp: First.CC) && isGreaterThan(Cmp: Second.CC))) && |
| 446 | std::abs(x: SecondImmTrueValue - FirstImmTrueValue) == 2) { |
| 447 | // This branch transforms machine instructions that correspond to |
| 448 | // |
| 449 | // 1) (a > {SecondImm} && ...) || (a < {FirstImm} && ...) |
| 450 | // 2) (a < {SecondImm} && ...) || (a > {FirstImm} && ...) |
| 451 | // |
| 452 | // into |
| 453 | // |
| 454 | // 1) (a >= {NewImm} && ...) || (a <= {NewImm} && ...) |
| 455 | // 2) (a <= {NewImm} && ...) || (a >= {NewImm} && ...) |
| 456 | |
| 457 | // Verify both adjustments converge to identical comparisons (same |
| 458 | // immediate and opcode). This ensures CSE can eliminate the duplicate. |
| 459 | if (FirstAdj.Imm != SecondAdj.Imm || FirstAdj.Opc != SecondAdj.Opc) |
| 460 | return false; |
| 461 | |
| 462 | LLVM_DEBUG(dbgs() << "Optimized (opposite): " |
| 463 | << AArch64CC::getCondCodeName(First.CC) << " #" |
| 464 | << First.getImm() << ", " |
| 465 | << AArch64CC::getCondCodeName(Second.CC) << " #" |
| 466 | << Second.getImm() << " -> " |
| 467 | << AArch64CC::getCondCodeName(FirstAdj.CC) << " #" |
| 468 | << FirstAdj.Imm << ", " |
| 469 | << AArch64CC::getCondCodeName(SecondAdj.CC) << " #" |
| 470 | << SecondAdj.Imm << '\n'); |
| 471 | applyCmpAdjustment(Pair&: First, Info: FirstAdj); |
| 472 | applyCmpAdjustment(Pair&: Second, Info: SecondAdj); |
| 473 | return true; |
| 474 | |
| 475 | } else if (((isGreaterThan(Cmp: First.CC) && isGreaterThan(Cmp: Second.CC)) || |
| 476 | (isLessThan(Cmp: First.CC) && isLessThan(Cmp: Second.CC))) && |
| 477 | std::abs(x: SecondImmTrueValue - FirstImmTrueValue) == 1) { |
| 478 | // This branch transforms machine instructions that correspond to |
| 479 | // |
| 480 | // 1) (a > {SecondImm} && ...) || (a > {FirstImm} && ...) |
| 481 | // 2) (a < {SecondImm} && ...) || (a < {FirstImm} && ...) |
| 482 | // |
| 483 | // into |
| 484 | // |
| 485 | // 1) (a <= {NewImm} && ...) || (a > {NewImm} && ...) |
| 486 | // 2) (a < {NewImm} && ...) || (a >= {NewImm} && ...) |
| 487 | |
| 488 | // GT -> GE transformation increases immediate value, so picking the |
| 489 | // smaller one; LT -> LE decreases immediate value so invert the choice. |
| 490 | bool AdjustFirst = (FirstImmTrueValue < SecondImmTrueValue); |
| 491 | if (isLessThan(Cmp: First.CC)) |
| 492 | AdjustFirst = !AdjustFirst; |
| 493 | |
| 494 | CmpCondPair &Target = AdjustFirst ? Second : First; |
| 495 | CmpCondPair &ToChange = AdjustFirst ? First : Second; |
| 496 | CmpInfo &Adj = AdjustFirst ? FirstAdj : SecondAdj; |
| 497 | |
| 498 | // Verify the adjustment converges to the target's comparison (same |
| 499 | // immediate and opcode). This ensures CSE can eliminate the duplicate. |
| 500 | if (Adj.Imm != Target.getImm() || Adj.Opc != Target.getOpc()) |
| 501 | return false; |
| 502 | |
| 503 | LLVM_DEBUG(dbgs() << "Optimized (same-direction): " |
| 504 | << AArch64CC::getCondCodeName(ToChange.CC) << " #" |
| 505 | << ToChange.getImm() << " -> " |
| 506 | << AArch64CC::getCondCodeName(Adj.CC) << " #" << Adj.Imm |
| 507 | << '\n'); |
| 508 | applyCmpAdjustment(Pair&: ToChange, Info: Adj); |
| 509 | return true; |
| 510 | } |
| 511 | |
| 512 | // Other transformation cases almost never occur due to generation of < or > |
| 513 | // comparisons instead of <= and >=. |
| 514 | return false; |
| 515 | } |
| 516 | |
| 517 | bool AArch64ConditionOptimizerImpl::commitPendingPair( |
| 518 | std::optional<CmpCondPair> &PendingPair, |
| 519 | SmallDenseMap<Register, CmpCondPair> &PairsByReg) { |
| 520 | if (!PendingPair) |
| 521 | return false; |
| 522 | |
| 523 | Register Reg = PendingPair->CmpMI->getOperand(i: 1).getReg(); |
| 524 | Register Key = Reg.isVirtual() ? TRI->lookThruCopyLike(SrcReg: Reg, MRI) : Reg; |
| 525 | |
| 526 | auto MatchingPair = PairsByReg.find(Val: Key); |
| 527 | bool Changed = MatchingPair != PairsByReg.end() && |
| 528 | tryOptimizePair(First&: MatchingPair->second, Second&: *PendingPair); |
| 529 | |
| 530 | PairsByReg[Key] = *PendingPair; |
| 531 | PendingPair = std::nullopt; |
| 532 | return Changed; |
| 533 | } |
| 534 | |
| 535 | // This function transforms cmps and their consuming conditionals (CmpCondPairs) |
| 536 | // 1. Same direction: when both conditions are the same (e.g. GT/GT or LT/LT) |
| 537 | // and immediates differ by 1 |
| 538 | // 2. Opposite direction: when both conditions are adjustable to a common middle |
| 539 | // (e.g., GT/LT) and immediates differ by 2. |
| 540 | // The compare instructions are made to match to enable CSE. |
| 541 | // All cmp/cond pairs within a basic block are examined |
| 542 | // |
| 543 | // Example transformation: |
| 544 | // cmp w8, #10 |
| 545 | // csinc w9, w0, w1, gt ; w9 = (w8 > 10) ? w0 : w1+1 |
| 546 | // cmp w8, #9 |
| 547 | // csinc w10, w0, w1, gt ; w10 = (w8 > 9) ? w0 : w1+1 |
| 548 | // |
| 549 | // Into: |
| 550 | // cmp w8, #10 |
| 551 | // csinc w9, w0, w1, gt ; w9 = (w8 > 10) ? w0 : w1+1 |
| 552 | // cmp w8, #10 ; <- CSE can remove the redundant cmp |
| 553 | // csinc w10, w0, w1, ge ; w10 = (w8 >= 10) ? w0 : w1+1 |
| 554 | // |
| 555 | bool AArch64ConditionOptimizerImpl::optimizeIntraBlock(MachineBasicBlock &MBB) { |
| 556 | SmallDenseMap<Register, CmpCondPair> PairsByReg; |
| 557 | std::optional<CmpCondPair> PendingPair; |
| 558 | MachineInstr *ActiveCmp = nullptr; |
| 559 | bool Changed = false; |
| 560 | |
| 561 | for (MachineInstr &MI : MBB) { |
| 562 | if (MI.isDebugInstr()) |
| 563 | continue; |
| 564 | |
| 565 | if (isCmpInstruction(Opc: MI.getOpcode()) && canAdjustCmp(CmpMI&: MI)) { |
| 566 | Changed |= commitPendingPair(PendingPair, PairsByReg); |
| 567 | ActiveCmp = &MI; |
| 568 | continue; |
| 569 | } |
| 570 | |
| 571 | if (MI.modifiesRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) { |
| 572 | // Non-CMP clobber: commit any pending pair and reset all state, since |
| 573 | // unknown flag state at this point invalidates all prior pairs |
| 574 | Changed |= commitPendingPair(PendingPair, PairsByReg); |
| 575 | ActiveCmp = nullptr; |
| 576 | PairsByReg.clear(); |
| 577 | continue; |
| 578 | } |
| 579 | |
| 580 | if (AArch64InstrInfo::findCondCodeUseOperandIdxForBranchOrSelect(Instr: MI) >= 0 && |
| 581 | !MI.isBranch()) { |
| 582 | if (PendingPair) { |
| 583 | // A second conditional consuming the same CMP would invalidate any |
| 584 | // optimization: modifying the CMP would silently change what both |
| 585 | // consumers compare against. Mark the CMP spent. |
| 586 | PendingPair = std::nullopt; |
| 587 | ActiveCmp = nullptr; |
| 588 | } else if (ActiveCmp) { |
| 589 | int CCOpIdx = |
| 590 | AArch64InstrInfo::findCondCodeUseOperandIdxForBranchOrSelect(Instr: MI); |
| 591 | assert(CCOpIdx >= 0 && "Unsupported conditional instruction" ); |
| 592 | AArch64CC::CondCode CC = |
| 593 | (AArch64CC::CondCode)(int)MI.getOperand(i: CCOpIdx).getImm(); |
| 594 | PendingPair = CmpCondPair{.CmpMI: ActiveCmp, .CondMI: &MI, .CC: CC}; |
| 595 | } |
| 596 | continue; |
| 597 | } |
| 598 | |
| 599 | if (MI.readsRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) { |
| 600 | ActiveCmp = nullptr; |
| 601 | PendingPair = std::nullopt; |
| 602 | continue; |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | // Only commit the final pending pair if NZCV doesn't live out: a cross-block |
| 607 | // consumer would be affected by any CMP adjustment we make. |
| 608 | if (!nzcvLivesOut(MBB: &MBB)) |
| 609 | Changed |= commitPendingPair(PendingPair, PairsByReg); |
| 610 | |
| 611 | return Changed; |
| 612 | } |
| 613 | |
| 614 | // Finds the last valid conditional consumer in MBB and returns it together |
| 615 | // with its condition code. Handles two cases: |
| 616 | // |
| 617 | // 1. Bcc terminator: if the block ends with a Bcc, analyzeBranch extracts |
| 618 | // the condition code directly from the branch operands. |
| 619 | // |
| 620 | // 2. Select-family instruction (CSET/CSEL/CSINC/CSINV/CSNEG): scans |
| 621 | // backward past terminators to find the sole non-branch NZCV consumer, |
| 622 | // verifying there is no interfering NZCV read or write between it and |
| 623 | // the CMP that produces the flags. |
| 624 | // |
| 625 | // Returns {nullptr, Invalid} if no suitable consumer is found or if any |
| 626 | // safety check fails. |
| 627 | std::pair<MachineInstr *, AArch64CC::CondCode> |
| 628 | AArch64ConditionOptimizerImpl::findCondConsumer(MachineBasicBlock *MBB) { |
| 629 | // Case 1: block ends with a Bcc terminator. |
| 630 | if (MachineInstr *BrMI = getBccTerminator(MBB)) { |
| 631 | SmallVector<MachineOperand, 4> CondOperands; |
| 632 | MachineBasicBlock *TBBDest = nullptr, *FBBDest = nullptr; |
| 633 | if (TII->analyzeBranch(MBB&: *MBB, TBB&: TBBDest, FBB&: FBBDest, Cond&: CondOperands)) |
| 634 | return {nullptr, AArch64CC::Invalid}; |
| 635 | AArch64CC::CondCode CC = parseCondCode(Cond: CondOperands); |
| 636 | if (CC == AArch64CC::Invalid) |
| 637 | return {nullptr, AArch64CC::Invalid}; |
| 638 | return {BrMI, CC}; |
| 639 | } |
| 640 | |
| 641 | // Case 2: no Bcc terminator — scan backward for a select-family instruction |
| 642 | // (CSET/CSEL/CSINC/CSINV/CSNEG) that is the sole NZCV consumer in the block. |
| 643 | MachineInstr *Found = nullptr; |
| 644 | AArch64CC::CondCode FoundCC = AArch64CC::Invalid; |
| 645 | |
| 646 | for (MachineInstr &MI : reverse(C&: *MBB)) { |
| 647 | // Skip terminators (e.g. an unconditional branch at the end of the block) |
| 648 | // and debug instructions, which carry no real semantics. |
| 649 | if (MI.isTerminator() || MI.isDebugInstr()) |
| 650 | continue; |
| 651 | |
| 652 | if (!Found) { |
| 653 | // We have not yet found the select. Keep scanning backward. |
| 654 | |
| 655 | // If something writes NZCV before we find a select, the flags at that |
| 656 | // point are not from the CMP we are looking for. Stop searching. |
| 657 | if (MI.modifiesRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) |
| 658 | return {nullptr, AArch64CC::Invalid}; |
| 659 | |
| 660 | // findCondCodeUseOperandIdxForBranchOrSelect returns the operand index |
| 661 | // of the condition code for any branch or select-family instruction, or |
| 662 | // -1 if the instruction does not use a condition code. |
| 663 | // We exclude branches because getBccTerminator already handles those; |
| 664 | // we only want non-branch conditionals: CSET, CSEL, CSINC, CSINV, CSNEG. |
| 665 | int CCOpIdx = |
| 666 | AArch64InstrInfo::findCondCodeUseOperandIdxForBranchOrSelect(Instr: MI); |
| 667 | if (CCOpIdx >= 0 && !MI.isBranch()) { |
| 668 | Found = &MI; |
| 669 | FoundCC = (AArch64CC::CondCode)(int)MI.getOperand(i: CCOpIdx).getImm(); |
| 670 | continue; |
| 671 | } |
| 672 | |
| 673 | // Any other instruction that reads NZCV (but is not a select) means the |
| 674 | // flags are consumed by something we do not understand. Stop searching. |
| 675 | if (MI.readsRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) |
| 676 | return {nullptr, AArch64CC::Invalid}; |
| 677 | |
| 678 | } else { |
| 679 | // We already found a select. Now verify there is no second NZCV reader |
| 680 | // between the found select and the CMP. If there is, the CMP feeds two |
| 681 | // consumers and cannot be safely adjusted. |
| 682 | if (MI.readsRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) |
| 683 | return {nullptr, AArch64CC::Invalid}; |
| 684 | |
| 685 | if (MI.modifiesRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) { |
| 686 | // A CMP instruction is the flag producer we are looking for; stop |
| 687 | // scanning. findAdjustableCmp will locate it from CondMI. |
| 688 | if (isCmpInstruction(Opc: MI.getOpcode())) |
| 689 | break; |
| 690 | // Any other NZCV writer means the select is not reading from the CMP |
| 691 | // we would find further back. |
| 692 | return {nullptr, AArch64CC::Invalid}; |
| 693 | } |
| 694 | } |
| 695 | } |
| 696 | return {Found, FoundCC}; |
| 697 | } |
| 698 | |
| 699 | // Optimizes CMP+conditional pairs across two basic blocks in the dominator |
| 700 | // tree. The conditional consumer in each block may be a Bcc terminator or a |
| 701 | // select-family instruction (CSEL/CSET/CSINC/CSINV/CSNEG). |
| 702 | bool AArch64ConditionOptimizerImpl::optimizeCrossBlock(MachineBasicBlock &HBB) { |
| 703 | SmallVector<MachineOperand, 4> HeadCondOperands; |
| 704 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
| 705 | if (TII->analyzeBranch(MBB&: HBB, TBB, FBB, Cond&: HeadCondOperands)) { |
| 706 | return false; |
| 707 | } |
| 708 | |
| 709 | // Equivalence check is to skip loops. |
| 710 | if (!TBB || TBB == &HBB) { |
| 711 | return false; |
| 712 | } |
| 713 | |
| 714 | // Find the conditional consumer(Bcc or select-family) and its condition |
| 715 | // code in each block. findCondConsumer() handles both cases uniformly. |
| 716 | auto [HeadCondMI, HeadCondCode] = findCondConsumer(MBB: &HBB); |
| 717 | if (!HeadCondMI) |
| 718 | return false; |
| 719 | |
| 720 | auto [TrueCondMI, TrueCondCode] = findCondConsumer(MBB: TBB); |
| 721 | if (!TrueCondMI) |
| 722 | return false; |
| 723 | |
| 724 | // Since we may modify cmps in these blocks, make sure NZCV does not live out. |
| 725 | if (nzcvLivesOut(MBB: &HBB) || nzcvLivesOut(MBB: TBB)) |
| 726 | return false; |
| 727 | |
| 728 | // Find the CMPs controlling each conditional. |
| 729 | MachineInstr *HeadCmpMI = findAdjustableCmp(CondMI: HeadCondMI); |
| 730 | MachineInstr *TrueCmpMI = findAdjustableCmp(CondMI: TrueCondMI); |
| 731 | if (!HeadCmpMI || !TrueCmpMI) |
| 732 | return false; |
| 733 | |
| 734 | if (!registersMatch(FirstMI: HeadCmpMI, SecondMI: TrueCmpMI)) |
| 735 | return false; |
| 736 | |
| 737 | LLVM_DEBUG(dbgs() << "Checking cross-block pair: " |
| 738 | << AArch64CC::getCondCodeName(HeadCondCode) << " #" |
| 739 | << HeadCmpMI->getOperand(2).getImm() << ", " |
| 740 | << AArch64CC::getCondCodeName(TrueCondCode) << " #" |
| 741 | << TrueCmpMI->getOperand(2).getImm() << '\n'); |
| 742 | |
| 743 | CmpCondPair Head{.CmpMI: HeadCmpMI, .CondMI: HeadCondMI, .CC: HeadCondCode}; |
| 744 | CmpCondPair True{.CmpMI: TrueCmpMI, .CondMI: TrueCondMI, .CC: TrueCondCode}; |
| 745 | |
| 746 | return tryOptimizePair(First&: Head, Second&: True); |
| 747 | } |
| 748 | |
| 749 | bool AArch64ConditionOptimizerLegacy::runOnMachineFunction( |
| 750 | MachineFunction &MF) { |
| 751 | if (skipFunction(F: MF.getFunction())) |
| 752 | return false; |
| 753 | MachineDominatorTree &MDT = |
| 754 | getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
| 755 | return AArch64ConditionOptimizerImpl().run(MF, MDT); |
| 756 | } |
| 757 | |
| 758 | bool AArch64ConditionOptimizerImpl::run(MachineFunction &MF, |
| 759 | MachineDominatorTree &MDT) { |
| 760 | LLVM_DEBUG(dbgs() << "********** AArch64 Conditional Compares **********\n" |
| 761 | << "********** Function: " << MF.getName() << '\n'); |
| 762 | |
| 763 | TII = static_cast<const AArch64InstrInfo *>(MF.getSubtarget().getInstrInfo()); |
| 764 | TRI = MF.getSubtarget().getRegisterInfo(); |
| 765 | DomTree = &MDT; |
| 766 | MRI = &MF.getRegInfo(); |
| 767 | |
| 768 | bool Changed = false; |
| 769 | |
| 770 | // Visit blocks in dominator tree pre-order. The pre-order enables multiple |
| 771 | // cmp-conversions from the same head block. |
| 772 | // Note that updateDomTree() modifies the children of the DomTree node |
| 773 | // currently being visited. The df_iterator supports that; it doesn't look at |
| 774 | // child_begin() / child_end() until after a node has been visited. |
| 775 | for (MachineDomTreeNode *I : depth_first(G: DomTree)) { |
| 776 | MachineBasicBlock *HBB = I->getBlock(); |
| 777 | Changed |= optimizeIntraBlock(MBB&: *HBB); |
| 778 | Changed |= optimizeCrossBlock(HBB&: *HBB); |
| 779 | } |
| 780 | |
| 781 | return Changed; |
| 782 | } |
| 783 | |
| 784 | PreservedAnalyses |
| 785 | AArch64ConditionOptimizerPass::run(MachineFunction &MF, |
| 786 | MachineFunctionAnalysisManager &MFAM) { |
| 787 | auto &MDT = MFAM.getResult<MachineDominatorTreeAnalysis>(IR&: MF); |
| 788 | bool Changed = AArch64ConditionOptimizerImpl().run(MF, MDT); |
| 789 | if (!Changed) |
| 790 | return PreservedAnalyses::all(); |
| 791 | PreservedAnalyses PA = getMachineFunctionPassPreservedAnalyses(); |
| 792 | PA.preserveSet<CFGAnalyses>(); |
| 793 | return PA; |
| 794 | } |
| 795 | |