1 //=== AArch64PostLegalizerCombiner.cpp --------------------------*- C++ -*-===// 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 /// \file 10 /// Post-legalization combines on generic MachineInstrs. 11 /// 12 /// The combines here must preserve instruction legality. 13 /// 14 /// Lowering combines (e.g. pseudo matching) should be handled by 15 /// AArch64PostLegalizerLowering. 16 /// 17 /// Combines which don't rely on instruction legality should go in the 18 /// AArch64PreLegalizerCombiner. 19 /// 20 //===----------------------------------------------------------------------===// 21 22 #include "AArch64TargetMachine.h" 23 #include "llvm/CodeGen/GlobalISel/Combiner.h" 24 #include "llvm/CodeGen/GlobalISel/CombinerHelper.h" 25 #include "llvm/CodeGen/GlobalISel/CombinerInfo.h" 26 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h" 27 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h" 28 #include "llvm/CodeGen/GlobalISel/Utils.h" 29 #include "llvm/CodeGen/MachineDominators.h" 30 #include "llvm/CodeGen/MachineFunctionPass.h" 31 #include "llvm/CodeGen/MachineRegisterInfo.h" 32 #include "llvm/CodeGen/TargetOpcodes.h" 33 #include "llvm/CodeGen/TargetPassConfig.h" 34 #include "llvm/Support/Debug.h" 35 36 #define DEBUG_TYPE "aarch64-postlegalizer-combiner" 37 38 using namespace llvm; 39 40 /// This combine tries do what performExtractVectorEltCombine does in SDAG. 41 /// Rewrite for pairwise fadd pattern 42 /// (s32 (g_extract_vector_elt 43 /// (g_fadd (vXs32 Other) 44 /// (g_vector_shuffle (vXs32 Other) undef <1,X,...> )) 0)) 45 /// -> 46 /// (s32 (g_fadd (g_extract_vector_elt (vXs32 Other) 0) 47 /// (g_extract_vector_elt (vXs32 Other) 1)) 48 bool matchExtractVecEltPairwiseAdd( 49 MachineInstr &MI, MachineRegisterInfo &MRI, 50 std::tuple<unsigned, LLT, Register> &MatchInfo) { 51 Register Src1 = MI.getOperand(1).getReg(); 52 Register Src2 = MI.getOperand(2).getReg(); 53 LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); 54 55 auto Cst = getConstantVRegValWithLookThrough(Src2, MRI); 56 if (!Cst || Cst->Value != 0) 57 return false; 58 // SDAG also checks for FullFP16, but this looks to be beneficial anyway. 59 60 // Now check for an fadd operation. TODO: expand this for integer add? 61 auto *FAddMI = getOpcodeDef(TargetOpcode::G_FADD, Src1, MRI); 62 if (!FAddMI) 63 return false; 64 65 // If we add support for integer add, must restrict these types to just s64. 66 unsigned DstSize = DstTy.getSizeInBits(); 67 if (DstSize != 16 && DstSize != 32 && DstSize != 64) 68 return false; 69 70 Register Src1Op1 = FAddMI->getOperand(1).getReg(); 71 Register Src1Op2 = FAddMI->getOperand(2).getReg(); 72 MachineInstr *Shuffle = 73 getOpcodeDef(TargetOpcode::G_SHUFFLE_VECTOR, Src1Op2, MRI); 74 MachineInstr *Other = MRI.getVRegDef(Src1Op1); 75 if (!Shuffle) { 76 Shuffle = getOpcodeDef(TargetOpcode::G_SHUFFLE_VECTOR, Src1Op1, MRI); 77 Other = MRI.getVRegDef(Src1Op2); 78 } 79 80 // We're looking for a shuffle that moves the second element to index 0. 81 if (Shuffle && Shuffle->getOperand(3).getShuffleMask()[0] == 1 && 82 Other == MRI.getVRegDef(Shuffle->getOperand(1).getReg())) { 83 std::get<0>(MatchInfo) = TargetOpcode::G_FADD; 84 std::get<1>(MatchInfo) = DstTy; 85 std::get<2>(MatchInfo) = Other->getOperand(0).getReg(); 86 return true; 87 } 88 return false; 89 } 90 91 bool applyExtractVecEltPairwiseAdd( 92 MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B, 93 std::tuple<unsigned, LLT, Register> &MatchInfo) { 94 unsigned Opc = std::get<0>(MatchInfo); 95 assert(Opc == TargetOpcode::G_FADD && "Unexpected opcode!"); 96 // We want to generate two extracts of elements 0 and 1, and add them. 97 LLT Ty = std::get<1>(MatchInfo); 98 Register Src = std::get<2>(MatchInfo); 99 LLT s64 = LLT::scalar(64); 100 B.setInstrAndDebugLoc(MI); 101 auto Elt0 = B.buildExtractVectorElement(Ty, Src, B.buildConstant(s64, 0)); 102 auto Elt1 = B.buildExtractVectorElement(Ty, Src, B.buildConstant(s64, 1)); 103 B.buildInstr(Opc, {MI.getOperand(0).getReg()}, {Elt0, Elt1}); 104 MI.eraseFromParent(); 105 return true; 106 } 107 108 static bool isSignExtended(Register R, MachineRegisterInfo &MRI) { 109 // TODO: check if extended build vector as well. 110 unsigned Opc = MRI.getVRegDef(R)->getOpcode(); 111 return Opc == TargetOpcode::G_SEXT || Opc == TargetOpcode::G_SEXT_INREG; 112 } 113 114 static bool isZeroExtended(Register R, MachineRegisterInfo &MRI) { 115 // TODO: check if extended build vector as well. 116 return MRI.getVRegDef(R)->getOpcode() == TargetOpcode::G_ZEXT; 117 } 118 119 bool matchAArch64MulConstCombine( 120 MachineInstr &MI, MachineRegisterInfo &MRI, 121 std::function<void(MachineIRBuilder &B, Register DstReg)> &ApplyFn) { 122 assert(MI.getOpcode() == TargetOpcode::G_MUL); 123 Register LHS = MI.getOperand(1).getReg(); 124 Register RHS = MI.getOperand(2).getReg(); 125 Register Dst = MI.getOperand(0).getReg(); 126 const LLT Ty = MRI.getType(LHS); 127 128 // The below optimizations require a constant RHS. 129 auto Const = getConstantVRegValWithLookThrough(RHS, MRI); 130 if (!Const) 131 return false; 132 133 const APInt ConstValue = Const->Value.sextOrSelf(Ty.getSizeInBits()); 134 // The following code is ported from AArch64ISelLowering. 135 // Multiplication of a power of two plus/minus one can be done more 136 // cheaply as as shift+add/sub. For now, this is true unilaterally. If 137 // future CPUs have a cheaper MADD instruction, this may need to be 138 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and 139 // 64-bit is 5 cycles, so this is always a win. 140 // More aggressively, some multiplications N0 * C can be lowered to 141 // shift+add+shift if the constant C = A * B where A = 2^N + 1 and B = 2^M, 142 // e.g. 6=3*2=(2+1)*2. 143 // TODO: consider lowering more cases, e.g. C = 14, -6, -14 or even 45 144 // which equals to (1+2)*16-(1+2). 145 // TrailingZeroes is used to test if the mul can be lowered to 146 // shift+add+shift. 147 unsigned TrailingZeroes = ConstValue.countTrailingZeros(); 148 if (TrailingZeroes) { 149 // Conservatively do not lower to shift+add+shift if the mul might be 150 // folded into smul or umul. 151 if (MRI.hasOneNonDBGUse(LHS) && 152 (isSignExtended(LHS, MRI) || isZeroExtended(LHS, MRI))) 153 return false; 154 // Conservatively do not lower to shift+add+shift if the mul might be 155 // folded into madd or msub. 156 if (MRI.hasOneNonDBGUse(Dst)) { 157 MachineInstr &UseMI = *MRI.use_instr_begin(Dst); 158 unsigned UseOpc = UseMI.getOpcode(); 159 if (UseOpc == TargetOpcode::G_ADD || UseOpc == TargetOpcode::G_PTR_ADD || 160 UseOpc == TargetOpcode::G_SUB) 161 return false; 162 } 163 } 164 // Use ShiftedConstValue instead of ConstValue to support both shift+add/sub 165 // and shift+add+shift. 166 APInt ShiftedConstValue = ConstValue.ashr(TrailingZeroes); 167 168 unsigned ShiftAmt, AddSubOpc; 169 // Is the shifted value the LHS operand of the add/sub? 170 bool ShiftValUseIsLHS = true; 171 // Do we need to negate the result? 172 bool NegateResult = false; 173 174 if (ConstValue.isNonNegative()) { 175 // (mul x, 2^N + 1) => (add (shl x, N), x) 176 // (mul x, 2^N - 1) => (sub (shl x, N), x) 177 // (mul x, (2^N + 1) * 2^M) => (shl (add (shl x, N), x), M) 178 APInt SCVMinus1 = ShiftedConstValue - 1; 179 APInt CVPlus1 = ConstValue + 1; 180 if (SCVMinus1.isPowerOf2()) { 181 ShiftAmt = SCVMinus1.logBase2(); 182 AddSubOpc = TargetOpcode::G_ADD; 183 } else if (CVPlus1.isPowerOf2()) { 184 ShiftAmt = CVPlus1.logBase2(); 185 AddSubOpc = TargetOpcode::G_SUB; 186 } else 187 return false; 188 } else { 189 // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) 190 // (mul x, -(2^N + 1)) => - (add (shl x, N), x) 191 APInt CVNegPlus1 = -ConstValue + 1; 192 APInt CVNegMinus1 = -ConstValue - 1; 193 if (CVNegPlus1.isPowerOf2()) { 194 ShiftAmt = CVNegPlus1.logBase2(); 195 AddSubOpc = TargetOpcode::G_SUB; 196 ShiftValUseIsLHS = false; 197 } else if (CVNegMinus1.isPowerOf2()) { 198 ShiftAmt = CVNegMinus1.logBase2(); 199 AddSubOpc = TargetOpcode::G_ADD; 200 NegateResult = true; 201 } else 202 return false; 203 } 204 205 if (NegateResult && TrailingZeroes) 206 return false; 207 208 ApplyFn = [=](MachineIRBuilder &B, Register DstReg) { 209 auto Shift = B.buildConstant(LLT::scalar(64), ShiftAmt); 210 auto ShiftedVal = B.buildShl(Ty, LHS, Shift); 211 212 Register AddSubLHS = ShiftValUseIsLHS ? ShiftedVal.getReg(0) : LHS; 213 Register AddSubRHS = ShiftValUseIsLHS ? LHS : ShiftedVal.getReg(0); 214 auto Res = B.buildInstr(AddSubOpc, {Ty}, {AddSubLHS, AddSubRHS}); 215 assert(!(NegateResult && TrailingZeroes) && 216 "NegateResult and TrailingZeroes cannot both be true for now."); 217 // Negate the result. 218 if (NegateResult) { 219 B.buildSub(DstReg, B.buildConstant(Ty, 0), Res); 220 return; 221 } 222 // Shift the result. 223 if (TrailingZeroes) { 224 B.buildShl(DstReg, Res, B.buildConstant(LLT::scalar(64), TrailingZeroes)); 225 return; 226 } 227 B.buildCopy(DstReg, Res.getReg(0)); 228 }; 229 return true; 230 } 231 232 bool applyAArch64MulConstCombine( 233 MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B, 234 std::function<void(MachineIRBuilder &B, Register DstReg)> &ApplyFn) { 235 B.setInstrAndDebugLoc(MI); 236 ApplyFn(B, MI.getOperand(0).getReg()); 237 MI.eraseFromParent(); 238 return true; 239 } 240 241 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_DEPS 242 #include "AArch64GenPostLegalizeGICombiner.inc" 243 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_DEPS 244 245 namespace { 246 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_H 247 #include "AArch64GenPostLegalizeGICombiner.inc" 248 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_H 249 250 class AArch64PostLegalizerCombinerInfo : public CombinerInfo { 251 GISelKnownBits *KB; 252 MachineDominatorTree *MDT; 253 254 public: 255 AArch64GenPostLegalizerCombinerHelperRuleConfig GeneratedRuleCfg; 256 257 AArch64PostLegalizerCombinerInfo(bool EnableOpt, bool OptSize, bool MinSize, 258 GISelKnownBits *KB, 259 MachineDominatorTree *MDT) 260 : CombinerInfo(/*AllowIllegalOps*/ true, /*ShouldLegalizeIllegal*/ false, 261 /*LegalizerInfo*/ nullptr, EnableOpt, OptSize, MinSize), 262 KB(KB), MDT(MDT) { 263 if (!GeneratedRuleCfg.parseCommandLineOption()) 264 report_fatal_error("Invalid rule identifier"); 265 } 266 267 virtual bool combine(GISelChangeObserver &Observer, MachineInstr &MI, 268 MachineIRBuilder &B) const override; 269 }; 270 271 bool AArch64PostLegalizerCombinerInfo::combine(GISelChangeObserver &Observer, 272 MachineInstr &MI, 273 MachineIRBuilder &B) const { 274 const auto *LI = 275 MI.getParent()->getParent()->getSubtarget().getLegalizerInfo(); 276 CombinerHelper Helper(Observer, B, KB, MDT, LI); 277 AArch64GenPostLegalizerCombinerHelper Generated(GeneratedRuleCfg); 278 return Generated.tryCombineAll(Observer, MI, B, Helper); 279 } 280 281 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_CPP 282 #include "AArch64GenPostLegalizeGICombiner.inc" 283 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_CPP 284 285 class AArch64PostLegalizerCombiner : public MachineFunctionPass { 286 public: 287 static char ID; 288 289 AArch64PostLegalizerCombiner(bool IsOptNone = false); 290 291 StringRef getPassName() const override { 292 return "AArch64PostLegalizerCombiner"; 293 } 294 295 bool runOnMachineFunction(MachineFunction &MF) override; 296 void getAnalysisUsage(AnalysisUsage &AU) const override; 297 298 private: 299 bool IsOptNone; 300 }; 301 } // end anonymous namespace 302 303 void AArch64PostLegalizerCombiner::getAnalysisUsage(AnalysisUsage &AU) const { 304 AU.addRequired<TargetPassConfig>(); 305 AU.setPreservesCFG(); 306 getSelectionDAGFallbackAnalysisUsage(AU); 307 AU.addRequired<GISelKnownBitsAnalysis>(); 308 AU.addPreserved<GISelKnownBitsAnalysis>(); 309 if (!IsOptNone) { 310 AU.addRequired<MachineDominatorTree>(); 311 AU.addPreserved<MachineDominatorTree>(); 312 } 313 MachineFunctionPass::getAnalysisUsage(AU); 314 } 315 316 AArch64PostLegalizerCombiner::AArch64PostLegalizerCombiner(bool IsOptNone) 317 : MachineFunctionPass(ID), IsOptNone(IsOptNone) { 318 initializeAArch64PostLegalizerCombinerPass(*PassRegistry::getPassRegistry()); 319 } 320 321 bool AArch64PostLegalizerCombiner::runOnMachineFunction(MachineFunction &MF) { 322 if (MF.getProperties().hasProperty( 323 MachineFunctionProperties::Property::FailedISel)) 324 return false; 325 assert(MF.getProperties().hasProperty( 326 MachineFunctionProperties::Property::Legalized) && 327 "Expected a legalized function?"); 328 auto *TPC = &getAnalysis<TargetPassConfig>(); 329 const Function &F = MF.getFunction(); 330 bool EnableOpt = 331 MF.getTarget().getOptLevel() != CodeGenOpt::None && !skipFunction(F); 332 GISelKnownBits *KB = &getAnalysis<GISelKnownBitsAnalysis>().get(MF); 333 MachineDominatorTree *MDT = 334 IsOptNone ? nullptr : &getAnalysis<MachineDominatorTree>(); 335 AArch64PostLegalizerCombinerInfo PCInfo(EnableOpt, F.hasOptSize(), 336 F.hasMinSize(), KB, MDT); 337 Combiner C(PCInfo, TPC); 338 return C.combineMachineInstrs(MF, /*CSEInfo*/ nullptr); 339 } 340 341 char AArch64PostLegalizerCombiner::ID = 0; 342 INITIALIZE_PASS_BEGIN(AArch64PostLegalizerCombiner, DEBUG_TYPE, 343 "Combine AArch64 MachineInstrs after legalization", false, 344 false) 345 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 346 INITIALIZE_PASS_DEPENDENCY(GISelKnownBitsAnalysis) 347 INITIALIZE_PASS_END(AArch64PostLegalizerCombiner, DEBUG_TYPE, 348 "Combine AArch64 MachineInstrs after legalization", false, 349 false) 350 351 namespace llvm { 352 FunctionPass *createAArch64PostLegalizerCombiner(bool IsOptNone) { 353 return new AArch64PostLegalizerCombiner(IsOptNone); 354 } 355 } // end namespace llvm 356