1 //===- lib/CodeGen/GlobalISel/GISelKnownBits.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 /// Provides analysis for querying information about KnownBits during GISel 10 /// passes. 11 // 12 //===------------------ 13 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h" 14 #include "llvm/Analysis/ValueTracking.h" 15 #include "llvm/CodeGen/GlobalISel/Utils.h" 16 #include "llvm/CodeGen/MachineFrameInfo.h" 17 #include "llvm/CodeGen/MachineRegisterInfo.h" 18 #include "llvm/CodeGen/TargetLowering.h" 19 #include "llvm/CodeGen/TargetOpcodes.h" 20 21 #define DEBUG_TYPE "gisel-known-bits" 22 23 using namespace llvm; 24 25 char llvm::GISelKnownBitsAnalysis::ID = 0; 26 27 INITIALIZE_PASS_BEGIN(GISelKnownBitsAnalysis, DEBUG_TYPE, 28 "Analysis for ComputingKnownBits", false, true) 29 INITIALIZE_PASS_END(GISelKnownBitsAnalysis, DEBUG_TYPE, 30 "Analysis for ComputingKnownBits", false, true) 31 32 GISelKnownBits::GISelKnownBits(MachineFunction &MF) 33 : MF(MF), MRI(MF.getRegInfo()), TL(*MF.getSubtarget().getTargetLowering()), 34 DL(MF.getFunction().getParent()->getDataLayout()) {} 35 36 Align GISelKnownBits::inferAlignmentForFrameIdx(int FrameIdx, int Offset, 37 const MachineFunction &MF) { 38 const MachineFrameInfo &MFI = MF.getFrameInfo(); 39 return commonAlignment(Align(MFI.getObjectAlignment(FrameIdx)), Offset); 40 // TODO: How to handle cases with Base + Offset? 41 } 42 43 MaybeAlign GISelKnownBits::inferPtrAlignment(const MachineInstr &MI) { 44 if (MI.getOpcode() == TargetOpcode::G_FRAME_INDEX) { 45 int FrameIdx = MI.getOperand(1).getIndex(); 46 return inferAlignmentForFrameIdx(FrameIdx, 0, *MI.getMF()); 47 } 48 return None; 49 } 50 51 void GISelKnownBits::computeKnownBitsForFrameIndex(Register R, KnownBits &Known, 52 const APInt &DemandedElts, 53 unsigned Depth) { 54 const MachineInstr &MI = *MRI.getVRegDef(R); 55 computeKnownBitsForAlignment(Known, inferPtrAlignment(MI)); 56 } 57 58 void GISelKnownBits::computeKnownBitsForAlignment(KnownBits &Known, 59 MaybeAlign Alignment) { 60 if (Alignment) 61 // The low bits are known zero if the pointer is aligned. 62 Known.Zero.setLowBits(Log2(Alignment)); 63 } 64 65 KnownBits GISelKnownBits::getKnownBits(MachineInstr &MI) { 66 return getKnownBits(MI.getOperand(0).getReg()); 67 } 68 69 KnownBits GISelKnownBits::getKnownBits(Register R) { 70 KnownBits Known; 71 LLT Ty = MRI.getType(R); 72 APInt DemandedElts = 73 Ty.isVector() ? APInt::getAllOnesValue(Ty.getNumElements()) : APInt(1, 1); 74 computeKnownBitsImpl(R, Known, DemandedElts); 75 return Known; 76 } 77 78 bool GISelKnownBits::signBitIsZero(Register R) { 79 LLT Ty = MRI.getType(R); 80 unsigned BitWidth = Ty.getScalarSizeInBits(); 81 return maskedValueIsZero(R, APInt::getSignMask(BitWidth)); 82 } 83 84 APInt GISelKnownBits::getKnownZeroes(Register R) { 85 return getKnownBits(R).Zero; 86 } 87 88 APInt GISelKnownBits::getKnownOnes(Register R) { return getKnownBits(R).One; } 89 90 void GISelKnownBits::computeKnownBitsImpl(Register R, KnownBits &Known, 91 const APInt &DemandedElts, 92 unsigned Depth) { 93 MachineInstr &MI = *MRI.getVRegDef(R); 94 unsigned Opcode = MI.getOpcode(); 95 LLT DstTy = MRI.getType(R); 96 97 // Handle the case where this is called on a register that does not have a 98 // type constraint (i.e. it has a register class constraint instead). This is 99 // unlikely to occur except by looking through copies but it is possible for 100 // the initial register being queried to be in this state. 101 if (!DstTy.isValid()) { 102 Known = KnownBits(); 103 return; 104 } 105 106 unsigned BitWidth = DstTy.getSizeInBits(); 107 Known = KnownBits(BitWidth); // Don't know anything 108 109 if (DstTy.isVector()) 110 return; // TODO: Handle vectors. 111 112 if (Depth == getMaxDepth()) 113 return; 114 115 if (!DemandedElts) 116 return; // No demanded elts, better to assume we don't know anything. 117 118 KnownBits Known2; 119 120 switch (Opcode) { 121 default: 122 TL.computeKnownBitsForTargetInstr(*this, R, Known, DemandedElts, MRI, 123 Depth); 124 break; 125 case TargetOpcode::COPY: { 126 MachineOperand Dst = MI.getOperand(0); 127 MachineOperand Src = MI.getOperand(1); 128 // Look through trivial copies but don't look through trivial copies of the 129 // form `%1:(s32) = OP %0:gpr32` known-bits analysis is currently unable to 130 // determine the bit width of a register class. 131 // 132 // We can't use NoSubRegister by name as it's defined by each target but 133 // it's always defined to be 0 by tablegen. 134 if (Dst.getSubReg() == 0 /*NoSubRegister*/ && Src.getReg().isVirtual() && 135 Src.getSubReg() == 0 /*NoSubRegister*/ && 136 MRI.getType(Src.getReg()).isValid()) { 137 // Don't increment Depth for this one since we didn't do any work. 138 computeKnownBitsImpl(Src.getReg(), Known, DemandedElts, Depth); 139 } 140 break; 141 } 142 case TargetOpcode::G_CONSTANT: { 143 auto CstVal = getConstantVRegVal(R, MRI); 144 if (!CstVal) 145 break; 146 Known.One = *CstVal; 147 Known.Zero = ~Known.One; 148 break; 149 } 150 case TargetOpcode::G_FRAME_INDEX: { 151 computeKnownBitsForFrameIndex(R, Known, DemandedElts); 152 break; 153 } 154 case TargetOpcode::G_SUB: { 155 // If low bits are known to be zero in both operands, then we know they are 156 // going to be 0 in the result. Both addition and complement operations 157 // preserve the low zero bits. 158 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known2, DemandedElts, 159 Depth + 1); 160 unsigned KnownZeroLow = Known2.countMinTrailingZeros(); 161 if (KnownZeroLow == 0) 162 break; 163 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known2, DemandedElts, 164 Depth + 1); 165 KnownZeroLow = std::min(KnownZeroLow, Known2.countMinTrailingZeros()); 166 Known.Zero.setLowBits(KnownZeroLow); 167 break; 168 } 169 case TargetOpcode::G_XOR: { 170 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known, DemandedElts, 171 Depth + 1); 172 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known2, DemandedElts, 173 Depth + 1); 174 175 // Output known-0 bits are known if clear or set in both the LHS & RHS. 176 APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One); 177 // Output known-1 are known to be set if set in only one of the LHS, RHS. 178 Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero); 179 Known.Zero = KnownZeroOut; 180 break; 181 } 182 case TargetOpcode::G_PTR_ADD: { 183 // G_PTR_ADD is like G_ADD. FIXME: Is this true for all targets? 184 LLT Ty = MRI.getType(MI.getOperand(1).getReg()); 185 if (DL.isNonIntegralAddressSpace(Ty.getAddressSpace())) 186 break; 187 LLVM_FALLTHROUGH; 188 } 189 case TargetOpcode::G_ADD: { 190 // Output known-0 bits are known if clear or set in both the low clear bits 191 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the 192 // low 3 bits clear. 193 // Output known-0 bits are also known if the top bits of each input are 194 // known to be clear. For example, if one input has the top 10 bits clear 195 // and the other has the top 8 bits clear, we know the top 7 bits of the 196 // output must be clear. 197 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known2, DemandedElts, 198 Depth + 1); 199 unsigned KnownZeroHigh = Known2.countMinLeadingZeros(); 200 unsigned KnownZeroLow = Known2.countMinTrailingZeros(); 201 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known2, DemandedElts, 202 Depth + 1); 203 KnownZeroHigh = std::min(KnownZeroHigh, Known2.countMinLeadingZeros()); 204 KnownZeroLow = std::min(KnownZeroLow, Known2.countMinTrailingZeros()); 205 Known.Zero.setLowBits(KnownZeroLow); 206 if (KnownZeroHigh > 1) 207 Known.Zero.setHighBits(KnownZeroHigh - 1); 208 break; 209 } 210 case TargetOpcode::G_AND: { 211 // If either the LHS or the RHS are Zero, the result is zero. 212 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known, DemandedElts, 213 Depth + 1); 214 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known2, DemandedElts, 215 Depth + 1); 216 217 // Output known-1 bits are only known if set in both the LHS & RHS. 218 Known.One &= Known2.One; 219 // Output known-0 are known to be clear if zero in either the LHS | RHS. 220 Known.Zero |= Known2.Zero; 221 break; 222 } 223 case TargetOpcode::G_OR: { 224 // If either the LHS or the RHS are Zero, the result is zero. 225 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known, DemandedElts, 226 Depth + 1); 227 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known2, DemandedElts, 228 Depth + 1); 229 230 // Output known-0 bits are only known if clear in both the LHS & RHS. 231 Known.Zero &= Known2.Zero; 232 // Output known-1 are known to be set if set in either the LHS | RHS. 233 Known.One |= Known2.One; 234 break; 235 } 236 case TargetOpcode::G_MUL: { 237 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known, DemandedElts, 238 Depth + 1); 239 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known2, DemandedElts, 240 Depth + 1); 241 // If low bits are zero in either operand, output low known-0 bits. 242 // Also compute a conservative estimate for high known-0 bits. 243 // More trickiness is possible, but this is sufficient for the 244 // interesting case of alignment computation. 245 unsigned TrailZ = 246 Known.countMinTrailingZeros() + Known2.countMinTrailingZeros(); 247 unsigned LeadZ = 248 std::max(Known.countMinLeadingZeros() + Known2.countMinLeadingZeros(), 249 BitWidth) - 250 BitWidth; 251 252 Known.resetAll(); 253 Known.Zero.setLowBits(std::min(TrailZ, BitWidth)); 254 Known.Zero.setHighBits(std::min(LeadZ, BitWidth)); 255 break; 256 } 257 case TargetOpcode::G_SELECT: { 258 computeKnownBitsImpl(MI.getOperand(3).getReg(), Known, DemandedElts, 259 Depth + 1); 260 // If we don't know any bits, early out. 261 if (Known.isUnknown()) 262 break; 263 computeKnownBitsImpl(MI.getOperand(2).getReg(), Known2, DemandedElts, 264 Depth + 1); 265 // Only known if known in both the LHS and RHS. 266 Known.One &= Known2.One; 267 Known.Zero &= Known2.Zero; 268 break; 269 } 270 case TargetOpcode::G_FCMP: 271 case TargetOpcode::G_ICMP: { 272 if (TL.getBooleanContents(DstTy.isVector(), 273 Opcode == TargetOpcode::G_FCMP) == 274 TargetLowering::ZeroOrOneBooleanContent && 275 BitWidth > 1) 276 Known.Zero.setBitsFrom(1); 277 break; 278 } 279 case TargetOpcode::G_SEXT: { 280 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known, DemandedElts, 281 Depth + 1); 282 // If the sign bit is known to be zero or one, then sext will extend 283 // it to the top bits, else it will just zext. 284 Known = Known.sext(BitWidth); 285 break; 286 } 287 case TargetOpcode::G_ANYEXT: { 288 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known, DemandedElts, 289 Depth + 1); 290 Known = Known.zext(BitWidth, true /* ExtendedBitsAreKnownZero */); 291 break; 292 } 293 case TargetOpcode::G_LOAD: { 294 if (MI.hasOneMemOperand()) { 295 const MachineMemOperand *MMO = *MI.memoperands_begin(); 296 if (const MDNode *Ranges = MMO->getRanges()) { 297 computeKnownBitsFromRangeMetadata(*Ranges, Known); 298 } 299 } 300 break; 301 } 302 case TargetOpcode::G_ZEXTLOAD: { 303 // Everything above the retrieved bits is zero 304 if (MI.hasOneMemOperand()) 305 Known.Zero.setBitsFrom((*MI.memoperands_begin())->getSizeInBits()); 306 break; 307 } 308 case TargetOpcode::G_ASHR: 309 case TargetOpcode::G_LSHR: 310 case TargetOpcode::G_SHL: { 311 KnownBits RHSKnown; 312 computeKnownBitsImpl(MI.getOperand(2).getReg(), RHSKnown, DemandedElts, 313 Depth + 1); 314 if (!RHSKnown.isConstant()) { 315 LLVM_DEBUG( 316 MachineInstr *RHSMI = MRI.getVRegDef(MI.getOperand(2).getReg()); 317 dbgs() << '[' << Depth << "] Shift not known constant: " << *RHSMI); 318 break; 319 } 320 uint64_t Shift = RHSKnown.getConstant().getZExtValue(); 321 LLVM_DEBUG(dbgs() << '[' << Depth << "] Shift is " << Shift << '\n'); 322 323 computeKnownBitsImpl(MI.getOperand(1).getReg(), Known, DemandedElts, 324 Depth + 1); 325 326 switch (Opcode) { 327 case TargetOpcode::G_ASHR: 328 Known.Zero = Known.Zero.ashr(Shift); 329 Known.One = Known.One.ashr(Shift); 330 break; 331 case TargetOpcode::G_LSHR: 332 Known.Zero = Known.Zero.lshr(Shift); 333 Known.One = Known.One.lshr(Shift); 334 Known.Zero.setBitsFrom(Known.Zero.getBitWidth() - Shift); 335 break; 336 case TargetOpcode::G_SHL: 337 Known.Zero = Known.Zero.shl(Shift); 338 Known.One = Known.One.shl(Shift); 339 Known.Zero.setBits(0, Shift); 340 break; 341 } 342 break; 343 } 344 case TargetOpcode::G_INTTOPTR: 345 case TargetOpcode::G_PTRTOINT: 346 // Fall through and handle them the same as zext/trunc. 347 LLVM_FALLTHROUGH; 348 case TargetOpcode::G_ZEXT: 349 case TargetOpcode::G_TRUNC: { 350 Register SrcReg = MI.getOperand(1).getReg(); 351 LLT SrcTy = MRI.getType(SrcReg); 352 unsigned SrcBitWidth = SrcTy.isPointer() 353 ? DL.getIndexSizeInBits(SrcTy.getAddressSpace()) 354 : SrcTy.getSizeInBits(); 355 assert(SrcBitWidth && "SrcBitWidth can't be zero"); 356 Known = Known.zextOrTrunc(SrcBitWidth, true); 357 computeKnownBitsImpl(SrcReg, Known, DemandedElts, Depth + 1); 358 Known = Known.zextOrTrunc(BitWidth, true); 359 if (BitWidth > SrcBitWidth) 360 Known.Zero.setBitsFrom(SrcBitWidth); 361 break; 362 } 363 } 364 365 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 366 LLVM_DEBUG(dbgs() << "[" << Depth << "] Compute known bits: " << MI << "[" 367 << Depth << "] Computed for: " << MI << "[" << Depth 368 << "] Known: 0x" 369 << (Known.Zero | Known.One).toString(16, false) << "\n" 370 << "[" << Depth << "] Zero: 0x" 371 << Known.Zero.toString(16, false) << "\n" 372 << "[" << Depth << "] One: 0x" 373 << Known.One.toString(16, false) << "\n"); 374 } 375 376 unsigned GISelKnownBits::computeNumSignBits(Register R, 377 const APInt &DemandedElts, 378 unsigned Depth) { 379 MachineInstr &MI = *MRI.getVRegDef(R); 380 unsigned Opcode = MI.getOpcode(); 381 382 if (Opcode == TargetOpcode::G_CONSTANT) 383 return MI.getOperand(1).getCImm()->getValue().getNumSignBits(); 384 385 if (Depth == getMaxDepth()) 386 return 1; 387 388 if (!DemandedElts) 389 return 1; // No demanded elts, better to assume we don't know anything. 390 391 LLT DstTy = MRI.getType(R); 392 393 // Handle the case where this is called on a register that does not have a 394 // type constraint. This is unlikely to occur except by looking through copies 395 // but it is possible for the initial register being queried to be in this 396 // state. 397 if (!DstTy.isValid()) 398 return 1; 399 400 switch (Opcode) { 401 case TargetOpcode::COPY: { 402 MachineOperand &Src = MI.getOperand(1); 403 if (Src.getReg().isVirtual() && Src.getSubReg() == 0 && 404 MRI.getType(Src.getReg()).isValid()) { 405 // Don't increment Depth for this one since we didn't do any work. 406 return computeNumSignBits(Src.getReg(), DemandedElts, Depth); 407 } 408 409 return 1; 410 } 411 case TargetOpcode::G_SEXT: { 412 Register Src = MI.getOperand(1).getReg(); 413 LLT SrcTy = MRI.getType(Src); 414 unsigned Tmp = DstTy.getScalarSizeInBits() - SrcTy.getScalarSizeInBits(); 415 return computeNumSignBits(Src, DemandedElts, Depth + 1) + Tmp; 416 } 417 case TargetOpcode::G_TRUNC: { 418 Register Src = MI.getOperand(1).getReg(); 419 LLT SrcTy = MRI.getType(Src); 420 421 // Check if the sign bits of source go down as far as the truncated value. 422 unsigned DstTyBits = DstTy.getScalarSizeInBits(); 423 unsigned NumSrcBits = SrcTy.getScalarSizeInBits(); 424 unsigned NumSrcSignBits = computeNumSignBits(Src, DemandedElts, Depth + 1); 425 if (NumSrcSignBits > (NumSrcBits - DstTyBits)) 426 return NumSrcSignBits - (NumSrcBits - DstTyBits); 427 break; 428 } 429 default: 430 break; 431 } 432 433 // TODO: Handle target instructions 434 // TODO: Fall back to known bits 435 return 1; 436 } 437 438 unsigned GISelKnownBits::computeNumSignBits(Register R, unsigned Depth) { 439 LLT Ty = MRI.getType(R); 440 APInt DemandedElts = Ty.isVector() 441 ? APInt::getAllOnesValue(Ty.getNumElements()) 442 : APInt(1, 1); 443 return computeNumSignBits(R, DemandedElts, Depth); 444 } 445 446 void GISelKnownBitsAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 447 AU.setPreservesAll(); 448 MachineFunctionPass::getAnalysisUsage(AU); 449 } 450 451 bool GISelKnownBitsAnalysis::runOnMachineFunction(MachineFunction &MF) { 452 return false; 453 } 454