| 1 | //===-- KnownBits.cpp - Stores known zeros/ones ---------------------------===// |
| 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 | // This file contains a class for representing known zeros and ones used by |
| 10 | // computeKnownBits. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/Support/KnownBits.h" |
| 15 | #include "llvm/ADT/Sequence.h" |
| 16 | #include "llvm/Support/Debug.h" |
| 17 | #include "llvm/Support/raw_ostream.h" |
| 18 | #include <cassert> |
| 19 | |
| 20 | using namespace llvm; |
| 21 | |
| 22 | KnownBits KnownBits::flipSignBit(const KnownBits &Val) { |
| 23 | unsigned SignBitPosition = Val.getBitWidth() - 1; |
| 24 | APInt Zero = Val.Zero; |
| 25 | APInt One = Val.One; |
| 26 | Zero.setBitVal(BitPosition: SignBitPosition, BitValue: Val.One[SignBitPosition]); |
| 27 | One.setBitVal(BitPosition: SignBitPosition, BitValue: Val.Zero[SignBitPosition]); |
| 28 | return KnownBits(Zero, One); |
| 29 | } |
| 30 | |
| 31 | static KnownBits computeForAddCarry(const KnownBits &LHS, const KnownBits &RHS, |
| 32 | bool CarryZero, bool CarryOne) { |
| 33 | |
| 34 | APInt PossibleSumZero = LHS.getMaxValue() + RHS.getMaxValue() + !CarryZero; |
| 35 | APInt PossibleSumOne = LHS.getMinValue() + RHS.getMinValue() + CarryOne; |
| 36 | |
| 37 | // Compute known bits of the carry. |
| 38 | APInt CarryKnownZero = ~(PossibleSumZero ^ LHS.Zero ^ RHS.Zero); |
| 39 | APInt CarryKnownOne = PossibleSumOne ^ LHS.One ^ RHS.One; |
| 40 | |
| 41 | // Compute set of known bits (where all three relevant bits are known). |
| 42 | APInt LHSKnownUnion = LHS.Zero | LHS.One; |
| 43 | APInt RHSKnownUnion = RHS.Zero | RHS.One; |
| 44 | APInt CarryKnownUnion = std::move(CarryKnownZero) | CarryKnownOne; |
| 45 | APInt Known = std::move(LHSKnownUnion) & RHSKnownUnion & CarryKnownUnion; |
| 46 | |
| 47 | // Compute known bits of the result. |
| 48 | KnownBits KnownOut; |
| 49 | KnownOut.Zero = ~std::move(PossibleSumZero) & Known; |
| 50 | KnownOut.One = std::move(PossibleSumOne) & Known; |
| 51 | return KnownOut; |
| 52 | } |
| 53 | |
| 54 | KnownBits KnownBits::computeForAddCarry( |
| 55 | const KnownBits &LHS, const KnownBits &RHS, const KnownBits &Carry) { |
| 56 | assert(Carry.getBitWidth() == 1 && "Carry must be 1-bit" ); |
| 57 | return ::computeForAddCarry( |
| 58 | LHS, RHS, CarryZero: Carry.Zero.getBoolValue(), CarryOne: Carry.One.getBoolValue()); |
| 59 | } |
| 60 | |
| 61 | KnownBits KnownBits::computeForAddSub(bool Add, bool NSW, bool NUW, |
| 62 | const KnownBits &LHS, |
| 63 | const KnownBits &RHS) { |
| 64 | unsigned BitWidth = LHS.getBitWidth(); |
| 65 | KnownBits KnownOut(BitWidth); |
| 66 | // This can be a relatively expensive helper, so optimistically save some |
| 67 | // work. |
| 68 | if (LHS.isUnknown() && RHS.isUnknown()) |
| 69 | return KnownOut; |
| 70 | |
| 71 | if (!LHS.isUnknown() && !RHS.isUnknown()) { |
| 72 | if (Add) { |
| 73 | // Sum = LHS + RHS + 0 |
| 74 | KnownOut = ::computeForAddCarry(LHS, RHS, /*CarryZero=*/true, |
| 75 | /*CarryOne=*/false); |
| 76 | } else { |
| 77 | // Sum = LHS + ~RHS + 1 |
| 78 | KnownBits NotRHS = RHS; |
| 79 | std::swap(a&: NotRHS.Zero, b&: NotRHS.One); |
| 80 | KnownOut = ::computeForAddCarry(LHS, RHS: NotRHS, /*CarryZero=*/false, |
| 81 | /*CarryOne=*/true); |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | // Handle add/sub given nsw and/or nuw. |
| 86 | if (NUW) { |
| 87 | if (Add) { |
| 88 | // (add nuw X, Y) |
| 89 | APInt MinVal = LHS.getMinValue().uadd_sat(RHS: RHS.getMinValue()); |
| 90 | // None of the adds can end up overflowing, so min consecutive highbits |
| 91 | // in minimum possible of X + Y must all remain set. |
| 92 | if (NSW) { |
| 93 | unsigned NumBits = MinVal.trunc(width: BitWidth - 1).countl_one(); |
| 94 | // If we have NSW as well, we also know we can't overflow the signbit so |
| 95 | // can start counting from 1 bit back. |
| 96 | KnownOut.One.setBits(loBit: BitWidth - 1 - NumBits, hiBit: BitWidth - 1); |
| 97 | } |
| 98 | KnownOut.One.setHighBits(MinVal.countl_one()); |
| 99 | } else { |
| 100 | // (sub nuw X, Y) |
| 101 | APInt MaxVal = LHS.getMaxValue().usub_sat(RHS: RHS.getMinValue()); |
| 102 | // None of the subs can overflow at any point, so any common high bits |
| 103 | // will subtract away and result in zeros. |
| 104 | if (NSW) { |
| 105 | // If we have NSW as well, we also know we can't overflow the signbit so |
| 106 | // can start counting from 1 bit back. |
| 107 | unsigned NumBits = MaxVal.trunc(width: BitWidth - 1).countl_zero(); |
| 108 | KnownOut.Zero.setBits(loBit: BitWidth - 1 - NumBits, hiBit: BitWidth - 1); |
| 109 | } |
| 110 | KnownOut.Zero.setHighBits(MaxVal.countl_zero()); |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | if (NSW) { |
| 115 | APInt MinVal; |
| 116 | APInt MaxVal; |
| 117 | if (Add) { |
| 118 | // (add nsw X, Y) |
| 119 | MinVal = LHS.getSignedMinValue().sadd_sat(RHS: RHS.getSignedMinValue()); |
| 120 | MaxVal = LHS.getSignedMaxValue().sadd_sat(RHS: RHS.getSignedMaxValue()); |
| 121 | } else { |
| 122 | // (sub nsw X, Y) |
| 123 | MinVal = LHS.getSignedMinValue().ssub_sat(RHS: RHS.getSignedMaxValue()); |
| 124 | MaxVal = LHS.getSignedMaxValue().ssub_sat(RHS: RHS.getSignedMinValue()); |
| 125 | } |
| 126 | if (MinVal.isNonNegative()) { |
| 127 | // If min is non-negative, result will always be non-neg (can't overflow |
| 128 | // around). |
| 129 | unsigned NumBits = MinVal.trunc(width: BitWidth - 1).countl_one(); |
| 130 | KnownOut.One.setBits(loBit: BitWidth - 1 - NumBits, hiBit: BitWidth - 1); |
| 131 | KnownOut.Zero.setSignBit(); |
| 132 | } |
| 133 | if (MaxVal.isNegative()) { |
| 134 | // If max is negative, result will always be neg (can't overflow around). |
| 135 | unsigned NumBits = MaxVal.trunc(width: BitWidth - 1).countl_zero(); |
| 136 | KnownOut.Zero.setBits(loBit: BitWidth - 1 - NumBits, hiBit: BitWidth - 1); |
| 137 | KnownOut.One.setSignBit(); |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | // Just return 0 if the nsw/nuw is violated and we have poison. |
| 142 | if (KnownOut.hasConflict()) |
| 143 | KnownOut.setAllZero(); |
| 144 | return KnownOut; |
| 145 | } |
| 146 | |
| 147 | KnownBits KnownBits::computeForSubBorrow(const KnownBits &LHS, KnownBits RHS, |
| 148 | const KnownBits &Borrow) { |
| 149 | assert(Borrow.getBitWidth() == 1 && "Borrow must be 1-bit" ); |
| 150 | |
| 151 | // LHS - RHS = LHS + ~RHS + 1 |
| 152 | // Carry 1 - Borrow in ::computeForAddCarry |
| 153 | std::swap(a&: RHS.Zero, b&: RHS.One); |
| 154 | return ::computeForAddCarry(LHS, RHS, |
| 155 | /*CarryZero=*/Borrow.One.getBoolValue(), |
| 156 | /*CarryOne=*/Borrow.Zero.getBoolValue()); |
| 157 | } |
| 158 | |
| 159 | KnownBits KnownBits::truncSSat(unsigned BitWidth) const { |
| 160 | unsigned InputBits = getBitWidth(); |
| 161 | APInt MinInRange = APInt::getSignedMinValue(numBits: BitWidth).sext(width: InputBits); |
| 162 | APInt MaxInRange = APInt::getSignedMaxValue(numBits: BitWidth).sext(width: InputBits); |
| 163 | APInt InputMin = getSignedMinValue(); |
| 164 | APInt InputMax = getSignedMaxValue(); |
| 165 | KnownBits Known(BitWidth); |
| 166 | |
| 167 | // Case 1: All values fit - just truncate |
| 168 | if (InputMin.sge(RHS: MinInRange) && InputMax.sle(RHS: MaxInRange)) { |
| 169 | Known = trunc(BitWidth); |
| 170 | } |
| 171 | // Case 2: All saturate to min |
| 172 | else if (InputMax.slt(RHS: MinInRange)) { |
| 173 | Known = KnownBits::makeConstant(C: APInt::getSignedMinValue(numBits: BitWidth)); |
| 174 | } |
| 175 | // Case 3: All saturate to max |
| 176 | else if (InputMin.sgt(RHS: MaxInRange)) { |
| 177 | Known = KnownBits::makeConstant(C: APInt::getSignedMaxValue(numBits: BitWidth)); |
| 178 | } |
| 179 | // Case 4: All non-negative, some fit, some saturate to max |
| 180 | else if (InputMin.isNonNegative()) { |
| 181 | // Output: truncated OR max saturation |
| 182 | // Max saturation has only sign bit as 0 |
| 183 | Known.Zero = APInt(BitWidth, 0); |
| 184 | Known.Zero.setBit(BitWidth - 1); // Sign bit always 0 |
| 185 | // Max saturation has all lower bits as 1, so preserve InputOneLower |
| 186 | Known.One = One.trunc(width: BitWidth); |
| 187 | Known.One.clearBit(BitPosition: BitWidth - 1); // Sign bit is 0, not 1 |
| 188 | } |
| 189 | // Case 5: All negative, some fit, some saturate to min |
| 190 | else if (InputMax.isNegative()) { |
| 191 | // Output: truncated OR min saturation |
| 192 | // Min saturation has all lower bits as 0, so preserve InputZeroLower |
| 193 | Known.Zero = Zero.trunc(width: BitWidth); |
| 194 | Known.Zero.clearBit(BitPosition: BitWidth - 1); // Sign bit is 1, not 0 |
| 195 | // Min saturation has only sign bit as 1 |
| 196 | Known.One = APInt(BitWidth, 0); |
| 197 | Known.One.setBit(BitWidth - 1); // Sign bit always 1 |
| 198 | } |
| 199 | // Case 6: Mixed positive and negative |
| 200 | else { |
| 201 | // Output: min saturation, truncated, or max saturation |
| 202 | APInt InputZeroLower = Zero.trunc(width: BitWidth); |
| 203 | APInt InputOneLower = One.trunc(width: BitWidth); |
| 204 | APInt MinSat = APInt::getSignedMinValue(numBits: BitWidth); |
| 205 | APInt MaxSat = APInt::getSignedMaxValue(numBits: BitWidth); |
| 206 | KnownBits MinSatKB = KnownBits::makeConstant(C: MinSat); |
| 207 | KnownBits MaxSatKB = KnownBits::makeConstant(C: MaxSat); |
| 208 | if (InputMax.sle(RHS: MaxInRange)) { |
| 209 | // Positive values fit, only negatives saturate to min |
| 210 | Known.Zero = InputZeroLower & MinSatKB.Zero; |
| 211 | Known.One = InputOneLower & MinSatKB.One; |
| 212 | } else if (InputMin.sge(RHS: MinInRange)) { |
| 213 | // Negative values fit, only positives saturate to max |
| 214 | Known.Zero = InputZeroLower & MaxSatKB.Zero; |
| 215 | Known.One = InputOneLower & MaxSatKB.One; |
| 216 | } else { |
| 217 | // Both positive and negative values might saturate |
| 218 | Known.Zero = InputZeroLower & MinSatKB.Zero & MaxSatKB.Zero; |
| 219 | Known.One = InputOneLower & MinSatKB.One & MaxSatKB.One; |
| 220 | } |
| 221 | } |
| 222 | return Known; |
| 223 | } |
| 224 | |
| 225 | KnownBits KnownBits::truncSSatU(unsigned BitWidth) const { |
| 226 | unsigned InputBits = getBitWidth(); |
| 227 | APInt MaxInRange = APInt::getAllOnes(numBits: BitWidth).zext(width: InputBits); |
| 228 | APInt InputMin = getSignedMinValue(); |
| 229 | APInt InputMax = getSignedMaxValue(); |
| 230 | KnownBits Known(BitWidth); |
| 231 | |
| 232 | if (isNegative()) { |
| 233 | Known.setAllZero(); |
| 234 | } else if (InputMin.isNonNegative() && InputMax.ule(RHS: MaxInRange)) { |
| 235 | Known = trunc(BitWidth); |
| 236 | } else if (InputMin.isNonNegative() && InputMin.ugt(RHS: MaxInRange)) { |
| 237 | Known.setAllOnes(); |
| 238 | } else if (InputMin.isNonNegative()) { |
| 239 | // All non-negative but mixed: some fit, some saturate to all-ones |
| 240 | // No common zero bits (saturation is all-ones) |
| 241 | Known.Zero = APInt(BitWidth, 0); |
| 242 | // Common one bits: bits that are 1 in input stay 1 |
| 243 | Known.One = One.trunc(width: BitWidth); |
| 244 | } else { |
| 245 | // Mixed: sign bit unknown |
| 246 | if (InputMax.ule(RHS: MaxInRange)) { |
| 247 | // Positive values all fit (no saturation to all-ones) |
| 248 | // Output: all-zeros (negatives) OR truncated (positives) |
| 249 | Known.Zero = Zero.trunc(width: BitWidth); |
| 250 | Known.One = APInt(BitWidth, 0); |
| 251 | } else { |
| 252 | // Positive values might saturate to all-ones |
| 253 | // Output: all-zeros OR truncated OR all-ones |
| 254 | // No common bits |
| 255 | Known.Zero = APInt(BitWidth, 0); |
| 256 | Known.One = APInt(BitWidth, 0); |
| 257 | } |
| 258 | } |
| 259 | return Known; |
| 260 | } |
| 261 | |
| 262 | KnownBits KnownBits::truncUSat(unsigned BitWidth) const { |
| 263 | unsigned InputBits = getBitWidth(); |
| 264 | APInt MaxInRange = APInt::getLowBitsSet(numBits: InputBits, loBitsSet: BitWidth); |
| 265 | APInt InputMax = getMaxValue(); |
| 266 | APInt InputMin = getMinValue(); |
| 267 | KnownBits Known(BitWidth); |
| 268 | |
| 269 | if (InputMax.ule(RHS: MaxInRange)) { |
| 270 | Known = trunc(BitWidth); |
| 271 | } else if (InputMin.ugt(RHS: MaxInRange)) { |
| 272 | Known.setAllOnes(); |
| 273 | } else { |
| 274 | Known.resetAll(); |
| 275 | Known.One = One.trunc(width: BitWidth); |
| 276 | } |
| 277 | return Known; |
| 278 | } |
| 279 | |
| 280 | KnownBits KnownBits::sextInReg(unsigned SrcBitWidth) const { |
| 281 | unsigned BitWidth = getBitWidth(); |
| 282 | assert(0 < SrcBitWidth && SrcBitWidth <= BitWidth && |
| 283 | "Illegal sext-in-register" ); |
| 284 | |
| 285 | if (SrcBitWidth == BitWidth) |
| 286 | return *this; |
| 287 | |
| 288 | unsigned ExtBits = BitWidth - SrcBitWidth; |
| 289 | KnownBits Result; |
| 290 | Result.One = One << ExtBits; |
| 291 | Result.Zero = Zero << ExtBits; |
| 292 | Result.One.ashrInPlace(ShiftAmt: ExtBits); |
| 293 | Result.Zero.ashrInPlace(ShiftAmt: ExtBits); |
| 294 | return Result; |
| 295 | } |
| 296 | |
| 297 | KnownBits KnownBits::makeGE(const APInt &Val) const { |
| 298 | // Count the number of leading bit positions where our underlying value is |
| 299 | // known to be less than or equal to Val. |
| 300 | unsigned N = (Zero | Val).countl_one(); |
| 301 | |
| 302 | // For each of those bit positions, if Val has a 1 in that bit then our |
| 303 | // underlying value must also have a 1. |
| 304 | APInt MaskedVal(Val); |
| 305 | MaskedVal.clearLowBits(loBits: getBitWidth() - N); |
| 306 | return KnownBits(Zero, One | MaskedVal); |
| 307 | } |
| 308 | |
| 309 | KnownBits KnownBits::umax(const KnownBits &LHS, const KnownBits &RHS) { |
| 310 | // If we can prove that LHS >= RHS then use LHS as the result. Likewise for |
| 311 | // RHS. Ideally our caller would already have spotted these cases and |
| 312 | // optimized away the umax operation, but we handle them here for |
| 313 | // completeness. |
| 314 | if (LHS.getMinValue().uge(RHS: RHS.getMaxValue())) |
| 315 | return LHS; |
| 316 | if (RHS.getMinValue().uge(RHS: LHS.getMaxValue())) |
| 317 | return RHS; |
| 318 | |
| 319 | // If the result of the umax is LHS then it must be greater than or equal to |
| 320 | // the minimum possible value of RHS. Likewise for RHS. Any known bits that |
| 321 | // are common to these two values are also known in the result. |
| 322 | KnownBits L = LHS.makeGE(Val: RHS.getMinValue()); |
| 323 | KnownBits R = RHS.makeGE(Val: LHS.getMinValue()); |
| 324 | return L.intersectWith(RHS: R); |
| 325 | } |
| 326 | |
| 327 | KnownBits KnownBits::umin(const KnownBits &LHS, const KnownBits &RHS) { |
| 328 | // Flip the range of values: [0, 0xFFFFFFFF] <-> [0xFFFFFFFF, 0] |
| 329 | auto Flip = [](const KnownBits &Val) { return KnownBits(Val.One, Val.Zero); }; |
| 330 | return Flip(umax(LHS: Flip(LHS), RHS: Flip(RHS))); |
| 331 | } |
| 332 | |
| 333 | KnownBits KnownBits::smax(const KnownBits &LHS, const KnownBits &RHS) { |
| 334 | return flipSignBit(Val: umax(LHS: flipSignBit(Val: LHS), RHS: flipSignBit(Val: RHS))); |
| 335 | } |
| 336 | |
| 337 | KnownBits KnownBits::smin(const KnownBits &LHS, const KnownBits &RHS) { |
| 338 | return flipSignBit(Val: umin(LHS: flipSignBit(Val: LHS), RHS: flipSignBit(Val: RHS))); |
| 339 | } |
| 340 | |
| 341 | KnownBits KnownBits::abdu(const KnownBits &LHS, const KnownBits &RHS) { |
| 342 | // If we know which argument is larger, return (sub LHS, RHS) or |
| 343 | // (sub RHS, LHS) directly. |
| 344 | if (LHS.getMinValue().uge(RHS: RHS.getMaxValue())) |
| 345 | return computeForAddSub(/*Add=*/false, /*NSW=*/false, /*NUW=*/false, LHS, |
| 346 | RHS); |
| 347 | if (RHS.getMinValue().uge(RHS: LHS.getMaxValue())) |
| 348 | return computeForAddSub(/*Add=*/false, /*NSW=*/false, /*NUW=*/false, LHS: RHS, |
| 349 | RHS: LHS); |
| 350 | |
| 351 | // By construction, the subtraction in abdu never has unsigned overflow. |
| 352 | // Find the common bits between (sub nuw LHS, RHS) and (sub nuw RHS, LHS). |
| 353 | KnownBits Diff0 = |
| 354 | computeForAddSub(/*Add=*/false, /*NSW=*/false, /*NUW=*/true, LHS, RHS); |
| 355 | KnownBits Diff1 = |
| 356 | computeForAddSub(/*Add=*/false, /*NSW=*/false, /*NUW=*/true, LHS: RHS, RHS: LHS); |
| 357 | return Diff0.intersectWith(RHS: Diff1); |
| 358 | } |
| 359 | |
| 360 | KnownBits KnownBits::abds(KnownBits LHS, KnownBits RHS) { |
| 361 | // Shift both arguments from the signed range to the unsigned range, e.g. from |
| 362 | // [-0x80, 0x7F] to [0, 0xFF]. |
| 363 | return abdu(LHS: flipSignBit(Val: LHS), RHS: flipSignBit(Val: RHS)); |
| 364 | } |
| 365 | |
| 366 | static unsigned getMaxShiftAmount(const APInt &MaxValue, unsigned BitWidth) { |
| 367 | if (isPowerOf2_32(Value: BitWidth)) |
| 368 | // Clamp to the shift amount's width: a narrower amount is already |
| 369 | // < BitWidth, so this stays a valid upper bound. |
| 370 | return MaxValue.extractBitsAsZExtValue( |
| 371 | numBits: std::min(a: Log2_32(Value: BitWidth), b: MaxValue.getBitWidth()), bitPosition: 0); |
| 372 | // This is only an approximate upper bound. |
| 373 | return MaxValue.getLimitedValue(Limit: BitWidth - 1); |
| 374 | } |
| 375 | |
| 376 | KnownBits KnownBits::shl(const KnownBits &LHS, const KnownBits &RHS, bool NUW, |
| 377 | bool NSW, bool ShAmtNonZero) { |
| 378 | unsigned BitWidth = LHS.getBitWidth(); |
| 379 | auto ShiftByConst = [&](const KnownBits &LHS, unsigned ShiftAmt) { |
| 380 | KnownBits Known; |
| 381 | bool ShiftedOutZero, ShiftedOutOne; |
| 382 | Known.Zero = LHS.Zero.ushl_ov(Amt: ShiftAmt, Overflow&: ShiftedOutZero); |
| 383 | Known.Zero.setLowBits(ShiftAmt); |
| 384 | Known.One = LHS.One.ushl_ov(Amt: ShiftAmt, Overflow&: ShiftedOutOne); |
| 385 | |
| 386 | // All cases returning poison have been handled by MaxShiftAmount already. |
| 387 | if (NSW) { |
| 388 | if (NUW && ShiftAmt != 0) |
| 389 | // NUW means we can assume anything shifted out was a zero. |
| 390 | ShiftedOutZero = true; |
| 391 | |
| 392 | if (ShiftedOutZero) |
| 393 | Known.makeNonNegative(); |
| 394 | else if (ShiftedOutOne) |
| 395 | Known.makeNegative(); |
| 396 | } |
| 397 | return Known; |
| 398 | }; |
| 399 | |
| 400 | // Fast path for a common case when LHS is completely unknown. |
| 401 | KnownBits Known(BitWidth); |
| 402 | unsigned MinShiftAmount = RHS.getMinValue().getLimitedValue(Limit: BitWidth); |
| 403 | if (MinShiftAmount == 0 && ShAmtNonZero) |
| 404 | MinShiftAmount = 1; |
| 405 | if (LHS.isUnknown()) { |
| 406 | Known.Zero.setLowBits(MinShiftAmount); |
| 407 | if (NUW && NSW && MinShiftAmount != 0) |
| 408 | Known.makeNonNegative(); |
| 409 | return Known; |
| 410 | } |
| 411 | |
| 412 | // Determine maximum shift amount, taking NUW/NSW flags into account. |
| 413 | APInt MaxValue = RHS.getMaxValue(); |
| 414 | unsigned MaxShiftAmount = getMaxShiftAmount(MaxValue, BitWidth); |
| 415 | if (NUW && NSW) |
| 416 | MaxShiftAmount = std::min(a: MaxShiftAmount, b: LHS.countMaxLeadingZeros() - 1); |
| 417 | if (NUW) |
| 418 | MaxShiftAmount = std::min(a: MaxShiftAmount, b: LHS.countMaxLeadingZeros()); |
| 419 | if (NSW) |
| 420 | MaxShiftAmount = std::min( |
| 421 | a: MaxShiftAmount, |
| 422 | b: std::max(a: LHS.countMaxLeadingZeros(), b: LHS.countMaxLeadingOnes()) - 1); |
| 423 | |
| 424 | // Fast path for common case where the shift amount is unknown. |
| 425 | if (MinShiftAmount == 0 && MaxShiftAmount == BitWidth - 1 && |
| 426 | isPowerOf2_32(Value: BitWidth)) { |
| 427 | Known.Zero.setLowBits(LHS.countMinTrailingZeros()); |
| 428 | if (LHS.isAllOnes()) |
| 429 | Known.One.setSignBit(); |
| 430 | if (NSW) { |
| 431 | if (LHS.isNonNegative()) |
| 432 | Known.makeNonNegative(); |
| 433 | if (LHS.isNegative()) |
| 434 | Known.makeNegative(); |
| 435 | } |
| 436 | return Known; |
| 437 | } |
| 438 | |
| 439 | // Find the common bits from all possible shifts. |
| 440 | unsigned ShiftAmtZeroMask = RHS.Zero.zextOrTrunc(width: 32).getZExtValue(); |
| 441 | unsigned ShiftAmtOneMask = RHS.One.zextOrTrunc(width: 32).getZExtValue(); |
| 442 | Known.setAllConflict(); |
| 443 | for (unsigned ShiftAmt = MinShiftAmount; ShiftAmt <= MaxShiftAmount; |
| 444 | ++ShiftAmt) { |
| 445 | // Skip if the shift amount is impossible. |
| 446 | if ((ShiftAmtZeroMask & ShiftAmt) != 0 || |
| 447 | (ShiftAmtOneMask | ShiftAmt) != ShiftAmt) |
| 448 | continue; |
| 449 | Known = Known.intersectWith(RHS: ShiftByConst(LHS, ShiftAmt)); |
| 450 | if (Known.isUnknown()) |
| 451 | break; |
| 452 | } |
| 453 | |
| 454 | // All shift amounts may result in poison. |
| 455 | if (Known.hasConflict()) |
| 456 | Known.setAllZero(); |
| 457 | return Known; |
| 458 | } |
| 459 | |
| 460 | KnownBits KnownBits::lshr(const KnownBits &LHS, const KnownBits &RHS, |
| 461 | bool ShAmtNonZero, bool Exact) { |
| 462 | unsigned BitWidth = LHS.getBitWidth(); |
| 463 | auto ShiftByConst = [&](const KnownBits &LHS, unsigned ShiftAmt) { |
| 464 | KnownBits Known = LHS; |
| 465 | Known >>= ShiftAmt; |
| 466 | // High bits are known zero. |
| 467 | Known.Zero.setHighBits(ShiftAmt); |
| 468 | return Known; |
| 469 | }; |
| 470 | |
| 471 | // Fast path for a common case when LHS is completely unknown. |
| 472 | KnownBits Known(BitWidth); |
| 473 | unsigned MinShiftAmount = RHS.getMinValue().getLimitedValue(Limit: BitWidth); |
| 474 | if (MinShiftAmount == 0 && ShAmtNonZero) |
| 475 | MinShiftAmount = 1; |
| 476 | if (LHS.isUnknown()) { |
| 477 | Known.Zero.setHighBits(MinShiftAmount); |
| 478 | return Known; |
| 479 | } |
| 480 | |
| 481 | // Find the common bits from all possible shifts. |
| 482 | APInt MaxValue = RHS.getMaxValue(); |
| 483 | unsigned MaxShiftAmount = getMaxShiftAmount(MaxValue, BitWidth); |
| 484 | |
| 485 | // If exact, bound MaxShiftAmount to first known 1 in LHS. |
| 486 | if (Exact) { |
| 487 | unsigned FirstOne = LHS.countMaxTrailingZeros(); |
| 488 | if (FirstOne < MinShiftAmount) { |
| 489 | // Always poison. Return zero because we don't like returning conflict. |
| 490 | Known.setAllZero(); |
| 491 | return Known; |
| 492 | } |
| 493 | MaxShiftAmount = std::min(a: MaxShiftAmount, b: FirstOne); |
| 494 | } |
| 495 | |
| 496 | unsigned ShiftAmtZeroMask = RHS.Zero.zextOrTrunc(width: 32).getZExtValue(); |
| 497 | unsigned ShiftAmtOneMask = RHS.One.zextOrTrunc(width: 32).getZExtValue(); |
| 498 | Known.setAllConflict(); |
| 499 | for (unsigned ShiftAmt = MinShiftAmount; ShiftAmt <= MaxShiftAmount; |
| 500 | ++ShiftAmt) { |
| 501 | // Skip if the shift amount is impossible. |
| 502 | if ((ShiftAmtZeroMask & ShiftAmt) != 0 || |
| 503 | (ShiftAmtOneMask | ShiftAmt) != ShiftAmt) |
| 504 | continue; |
| 505 | Known = Known.intersectWith(RHS: ShiftByConst(LHS, ShiftAmt)); |
| 506 | if (Known.isUnknown()) |
| 507 | break; |
| 508 | } |
| 509 | |
| 510 | // All shift amounts may result in poison. |
| 511 | if (Known.hasConflict()) |
| 512 | Known.setAllZero(); |
| 513 | return Known; |
| 514 | } |
| 515 | |
| 516 | KnownBits KnownBits::ashr(const KnownBits &LHS, const KnownBits &RHS, |
| 517 | bool ShAmtNonZero, bool Exact) { |
| 518 | unsigned BitWidth = LHS.getBitWidth(); |
| 519 | auto ShiftByConst = [&](const KnownBits &LHS, unsigned ShiftAmt) { |
| 520 | KnownBits Known = LHS; |
| 521 | Known.Zero.ashrInPlace(ShiftAmt); |
| 522 | Known.One.ashrInPlace(ShiftAmt); |
| 523 | return Known; |
| 524 | }; |
| 525 | |
| 526 | // Fast path for a common case when LHS is completely unknown. |
| 527 | KnownBits Known(BitWidth); |
| 528 | unsigned MinShiftAmount = RHS.getMinValue().getLimitedValue(Limit: BitWidth); |
| 529 | if (MinShiftAmount == 0 && ShAmtNonZero) |
| 530 | MinShiftAmount = 1; |
| 531 | if (LHS.isUnknown()) { |
| 532 | if (MinShiftAmount == BitWidth) { |
| 533 | // Always poison. Return zero because we don't like returning conflict. |
| 534 | Known.setAllZero(); |
| 535 | return Known; |
| 536 | } |
| 537 | return Known; |
| 538 | } |
| 539 | |
| 540 | // Find the common bits from all possible shifts. |
| 541 | APInt MaxValue = RHS.getMaxValue(); |
| 542 | unsigned MaxShiftAmount = getMaxShiftAmount(MaxValue, BitWidth); |
| 543 | |
| 544 | // If exact, bound MaxShiftAmount to first known 1 in LHS. |
| 545 | if (Exact) { |
| 546 | unsigned FirstOne = LHS.countMaxTrailingZeros(); |
| 547 | if (FirstOne < MinShiftAmount) { |
| 548 | // Always poison. Return zero because we don't like returning conflict. |
| 549 | Known.setAllZero(); |
| 550 | return Known; |
| 551 | } |
| 552 | MaxShiftAmount = std::min(a: MaxShiftAmount, b: FirstOne); |
| 553 | } |
| 554 | |
| 555 | unsigned ShiftAmtZeroMask = RHS.Zero.zextOrTrunc(width: 32).getZExtValue(); |
| 556 | unsigned ShiftAmtOneMask = RHS.One.zextOrTrunc(width: 32).getZExtValue(); |
| 557 | Known.setAllConflict(); |
| 558 | for (unsigned ShiftAmt = MinShiftAmount; ShiftAmt <= MaxShiftAmount; |
| 559 | ++ShiftAmt) { |
| 560 | // Skip if the shift amount is impossible. |
| 561 | if ((ShiftAmtZeroMask & ShiftAmt) != 0 || |
| 562 | (ShiftAmtOneMask | ShiftAmt) != ShiftAmt) |
| 563 | continue; |
| 564 | Known = Known.intersectWith(RHS: ShiftByConst(LHS, ShiftAmt)); |
| 565 | if (Known.isUnknown()) |
| 566 | break; |
| 567 | } |
| 568 | |
| 569 | // All shift amounts may result in poison. |
| 570 | if (Known.hasConflict()) |
| 571 | Known.setAllZero(); |
| 572 | return Known; |
| 573 | } |
| 574 | |
| 575 | KnownBits KnownBits::fshl(const KnownBits &LHS, const KnownBits &RHS, |
| 576 | const APInt &Amt) { |
| 577 | return KnownBits(APIntOps::fshl(Hi: LHS.Zero, Lo: RHS.Zero, Shift: Amt), |
| 578 | APIntOps::fshl(Hi: LHS.One, Lo: RHS.One, Shift: Amt)); |
| 579 | } |
| 580 | |
| 581 | KnownBits KnownBits::fshr(const KnownBits &LHS, const KnownBits &RHS, |
| 582 | const APInt &Amt) { |
| 583 | return KnownBits(APIntOps::fshr(Hi: LHS.Zero, Lo: RHS.Zero, Shift: Amt), |
| 584 | APIntOps::fshr(Hi: LHS.One, Lo: RHS.One, Shift: Amt)); |
| 585 | } |
| 586 | |
| 587 | KnownBits KnownBits::clmul(const KnownBits &LHS, const KnownBits &RHS) { |
| 588 | KnownBits Res = |
| 589 | makeConstant(C: APIntOps::clmul(LHS: LHS.getMinValue(), RHS: RHS.getMinValue())); |
| 590 | |
| 591 | // This is the same operation as clmul except it accumulates the result with |
| 592 | // an OR instead of an XOR. |
| 593 | auto ClMulOr = [](const APInt &LHS, const APInt &RHS) { |
| 594 | unsigned BW = LHS.getBitWidth(); |
| 595 | assert(BW == RHS.getBitWidth() && "Operand mismatch" ); |
| 596 | APInt Result(BW, 0); |
| 597 | for (unsigned I : |
| 598 | seq(Size: std::min(a: RHS.getActiveBits(), b: BW - LHS.countr_zero()))) |
| 599 | if (RHS[I]) |
| 600 | Result |= LHS << I; |
| 601 | return Result; |
| 602 | }; |
| 603 | |
| 604 | // Bits in the result are known if, for every corresponding pair of input |
| 605 | // bits, both input bits are known or either input bit is known to be zero. |
| 606 | APInt Known = ~(ClMulOr(~LHS.Zero & ~LHS.One, ~RHS.Zero) | |
| 607 | ClMulOr(~LHS.Zero, ~RHS.Zero & ~RHS.One)); |
| 608 | Res.Zero &= Known; |
| 609 | Res.One &= Known; |
| 610 | |
| 611 | return Res; |
| 612 | } |
| 613 | |
| 614 | KnownBits KnownBits::pext(const KnownBits &Val, const KnownBits &Mask) { |
| 615 | unsigned BitWidth = Val.getBitWidth(); |
| 616 | KnownBits Res(BitWidth); |
| 617 | // We start by asserting that bits cannot be 0 and cannot be 1, then clear |
| 618 | // Res bits where we know a bit could have some value. |
| 619 | Res.setAllConflict(); |
| 620 | |
| 621 | // For each source position I where Mask[I] could be 1, the output position J |
| 622 | // lies in [M0, M1], where M0 and M1 track the range of possible 1-bit counts |
| 623 | // seen so far in Mask. Note that M0=M1 as long as bits in Mask are known; |
| 624 | // otherwise, the range of possible output positions widens. |
| 625 | unsigned M0 = 0, M1 = 0; |
| 626 | for (unsigned I = 0; I < BitWidth; ++I) { |
| 627 | if (!Mask.Zero[I]) { |
| 628 | // Mask[I] could be 1, so we decide what value the Res bits could have. |
| 629 | if (!Val.Zero[I]) |
| 630 | // Val[I] could be 1 => Res[J] for J in [M0, M1] could be 1 |
| 631 | Res.Zero.clearBits(LoBit: M0, HiBit: M1 + 1); |
| 632 | if (!Val.One[I]) |
| 633 | // Val[I] could be 0 => Res[J] for J in [M0, M1] could be 0 |
| 634 | Res.One.clearBits(LoBit: M0, HiBit: M1 + 1); |
| 635 | } |
| 636 | if (Mask.One[I]) |
| 637 | ++M0, ++M1; |
| 638 | else if (!Mask.Zero[I]) |
| 639 | ++M1; |
| 640 | } |
| 641 | |
| 642 | // Output bits at J >= M0 may have no source (popcount(Mask) may be <= J), so |
| 643 | // they may be 0. |
| 644 | Res.One.clearBits(LoBit: M0, HiBit: BitWidth); |
| 645 | return Res; |
| 646 | } |
| 647 | |
| 648 | KnownBits KnownBits::pdep(const KnownBits &Val, const KnownBits &Mask) { |
| 649 | unsigned BitWidth = Val.getBitWidth(); |
| 650 | KnownBits Res(BitWidth); |
| 651 | // We start by asserting that bits cannot be 0 and cannot be 1, then clear |
| 652 | // Res bits where we know a bit could have some value. |
| 653 | Res.setAllConflict(); |
| 654 | |
| 655 | // For each output position I where Mask[I] could be 1, the source position J |
| 656 | // lies in [M0, M1], where M0 and M1 track the range of possible 1-bit counts |
| 657 | // seen so far in Mask. Note that M0=M1 as long as bits in Mask are known; |
| 658 | // otherwise, the range of possible source positions widens. |
| 659 | unsigned M0 = 0, M1 = 0; |
| 660 | for (unsigned I = 0; I < BitWidth; ++I) { |
| 661 | if (!Mask.One[I]) |
| 662 | // Mask[I] could be 0 => Res[I] could be 0 |
| 663 | Res.One.clearBit(BitPosition: I); |
| 664 | if (!Mask.Zero[I]) { |
| 665 | // Mask[I] could be 1, so we check what value the Val bits could have. |
| 666 | APInt Range = APInt::getBitsSet(numBits: BitWidth, loBit: M0, hiBit: M1 + 1); |
| 667 | if (!Range.isSubsetOf(RHS: Val.One)) |
| 668 | // Any Val[J] for J in [M0, M1] could be 0 => Res[I] could be 0 |
| 669 | Res.One.clearBit(BitPosition: I); |
| 670 | if (!Range.isSubsetOf(RHS: Val.Zero)) |
| 671 | // Any Val[J] for J in [M0, M1] could be 1 => Res[I] could be 1 |
| 672 | Res.Zero.clearBit(BitPosition: I); |
| 673 | } |
| 674 | if (Mask.One[I]) |
| 675 | ++M0, ++M1; |
| 676 | else if (!Mask.Zero[I]) |
| 677 | ++M1; |
| 678 | } |
| 679 | return Res; |
| 680 | } |
| 681 | |
| 682 | std::optional<bool> KnownBits::eq(const KnownBits &LHS, const KnownBits &RHS) { |
| 683 | if (LHS.isConstant() && RHS.isConstant()) |
| 684 | return LHS.getConstant() == RHS.getConstant(); |
| 685 | if (LHS.One.intersects(RHS: RHS.Zero) || RHS.One.intersects(RHS: LHS.Zero)) |
| 686 | return false; |
| 687 | return std::nullopt; |
| 688 | } |
| 689 | |
| 690 | std::optional<bool> KnownBits::ne(const KnownBits &LHS, const KnownBits &RHS) { |
| 691 | if (std::optional<bool> KnownEQ = eq(LHS, RHS)) |
| 692 | return !*KnownEQ; |
| 693 | return std::nullopt; |
| 694 | } |
| 695 | |
| 696 | std::optional<bool> KnownBits::ugt(const KnownBits &LHS, const KnownBits &RHS) { |
| 697 | // LHS >u RHS -> false if umax(LHS) <= umax(RHS) |
| 698 | if (LHS.getMaxValue().ule(RHS: RHS.getMinValue())) |
| 699 | return false; |
| 700 | // LHS >u RHS -> true if umin(LHS) > umax(RHS) |
| 701 | if (LHS.getMinValue().ugt(RHS: RHS.getMaxValue())) |
| 702 | return true; |
| 703 | return std::nullopt; |
| 704 | } |
| 705 | |
| 706 | std::optional<bool> KnownBits::uge(const KnownBits &LHS, const KnownBits &RHS) { |
| 707 | if (std::optional<bool> IsUGT = ugt(LHS: RHS, RHS: LHS)) |
| 708 | return !*IsUGT; |
| 709 | return std::nullopt; |
| 710 | } |
| 711 | |
| 712 | std::optional<bool> KnownBits::ult(const KnownBits &LHS, const KnownBits &RHS) { |
| 713 | return ugt(LHS: RHS, RHS: LHS); |
| 714 | } |
| 715 | |
| 716 | std::optional<bool> KnownBits::ule(const KnownBits &LHS, const KnownBits &RHS) { |
| 717 | return uge(LHS: RHS, RHS: LHS); |
| 718 | } |
| 719 | |
| 720 | std::optional<bool> KnownBits::sgt(const KnownBits &LHS, const KnownBits &RHS) { |
| 721 | // LHS >s RHS -> false if smax(LHS) <= smax(RHS) |
| 722 | if (LHS.getSignedMaxValue().sle(RHS: RHS.getSignedMinValue())) |
| 723 | return false; |
| 724 | // LHS >s RHS -> true if smin(LHS) > smax(RHS) |
| 725 | if (LHS.getSignedMinValue().sgt(RHS: RHS.getSignedMaxValue())) |
| 726 | return true; |
| 727 | return std::nullopt; |
| 728 | } |
| 729 | |
| 730 | std::optional<bool> KnownBits::sge(const KnownBits &LHS, const KnownBits &RHS) { |
| 731 | if (std::optional<bool> KnownSGT = sgt(LHS: RHS, RHS: LHS)) |
| 732 | return !*KnownSGT; |
| 733 | return std::nullopt; |
| 734 | } |
| 735 | |
| 736 | std::optional<bool> KnownBits::slt(const KnownBits &LHS, const KnownBits &RHS) { |
| 737 | return sgt(LHS: RHS, RHS: LHS); |
| 738 | } |
| 739 | |
| 740 | std::optional<bool> KnownBits::sle(const KnownBits &LHS, const KnownBits &RHS) { |
| 741 | return sge(LHS: RHS, RHS: LHS); |
| 742 | } |
| 743 | |
| 744 | KnownBits KnownBits::abs(bool IntMinIsPoison) const { |
| 745 | // If the source's MSB is zero then we know the rest of the bits already. |
| 746 | if (isNonNegative()) |
| 747 | return *this; |
| 748 | |
| 749 | // Absolute value preserves trailing zero count. |
| 750 | KnownBits KnownAbs(getBitWidth()); |
| 751 | |
| 752 | // If the input is negative, then abs(x) == -x. |
| 753 | if (isNegative()) { |
| 754 | KnownBits Tmp = *this; |
| 755 | // Special case for IntMinIsPoison. We know the sign bit is set and we know |
| 756 | // all the rest of the bits except one to be zero. Since we have |
| 757 | // IntMinIsPoison, that final bit MUST be a one, as otherwise the input is |
| 758 | // INT_MIN. |
| 759 | if (IntMinIsPoison && (Zero.popcount() + 2) == getBitWidth()) |
| 760 | Tmp.One.setBit(countMinTrailingZeros()); |
| 761 | |
| 762 | KnownAbs = computeForAddSub( |
| 763 | /*Add*/ false, NSW: IntMinIsPoison, /*NUW=*/false, |
| 764 | LHS: KnownBits::makeConstant(C: APInt(getBitWidth(), 0)), RHS: Tmp); |
| 765 | |
| 766 | // One more special case for IntMinIsPoison. If we don't know any ones other |
| 767 | // than the signbit, we know for certain that all the unknowns can't be |
| 768 | // zero. So if we know high zero bits, but have unknown low bits, we know |
| 769 | // for certain those high-zero bits will end up as one. This is because, |
| 770 | // the low bits can't be all zeros, so the +1 in (~x + 1) cannot carry up |
| 771 | // to the high bits. If we know a known INT_MIN input skip this. The result |
| 772 | // is poison anyways. |
| 773 | if (IntMinIsPoison && Tmp.countMinPopulation() == 1 && |
| 774 | Tmp.countMaxPopulation() != 1) { |
| 775 | Tmp.One.clearSignBit(); |
| 776 | Tmp.Zero.setSignBit(); |
| 777 | KnownAbs.One.setBits(loBit: getBitWidth() - Tmp.countMinLeadingZeros(), |
| 778 | hiBit: getBitWidth() - 1); |
| 779 | } |
| 780 | |
| 781 | } else { |
| 782 | unsigned MaxTZ = countMaxTrailingZeros(); |
| 783 | unsigned MinTZ = countMinTrailingZeros(); |
| 784 | |
| 785 | KnownAbs.Zero.setLowBits(MinTZ); |
| 786 | // If we know the lowest set 1, then preserve it. |
| 787 | if (MaxTZ == MinTZ && MaxTZ < getBitWidth()) |
| 788 | KnownAbs.One.setBit(MaxTZ); |
| 789 | |
| 790 | // We only know that the absolute values's MSB will be zero if INT_MIN is |
| 791 | // poison, or there is a set bit that isn't the sign bit (otherwise it could |
| 792 | // be INT_MIN). |
| 793 | if (IntMinIsPoison || (!One.isZero() && !One.isMinSignedValue())) { |
| 794 | KnownAbs.One.clearSignBit(); |
| 795 | KnownAbs.Zero.setSignBit(); |
| 796 | } |
| 797 | } |
| 798 | |
| 799 | return KnownAbs; |
| 800 | } |
| 801 | |
| 802 | KnownBits KnownBits::reduceAdd(unsigned NumElts) const { |
| 803 | if (NumElts == 0) |
| 804 | return KnownBits(getBitWidth()); |
| 805 | |
| 806 | unsigned BitWidth = getBitWidth(); |
| 807 | KnownBits Result(BitWidth); |
| 808 | |
| 809 | if (isConstant()) |
| 810 | // If all elements are the same constant, we can simply compute it |
| 811 | return KnownBits::makeConstant(C: NumElts * getConstant()); |
| 812 | |
| 813 | // The main idea is as follows. |
| 814 | // |
| 815 | // If KnownBits for each element has L leading zeros then |
| 816 | // X_i < 2^(W - L) for every i from [1, N]. |
| 817 | // |
| 818 | // ADD X_i <= ADD max(X_i) = N * max(X_i) |
| 819 | // < N * 2^(W - L) |
| 820 | // < 2^(W - L + ceil(log2(N))) |
| 821 | // |
| 822 | // As the result, we can conclude that |
| 823 | // |
| 824 | // L' = L - ceil(log2(N)) |
| 825 | // |
| 826 | // Similar logic can be applied to leading ones. |
| 827 | unsigned LostBits = Log2_32_Ceil(Value: NumElts); |
| 828 | |
| 829 | if (isNonNegative()) { |
| 830 | unsigned LeadingZeros = countMinLeadingZeros(); |
| 831 | LeadingZeros = LeadingZeros > LostBits ? LeadingZeros - LostBits : 0; |
| 832 | Result.Zero.setHighBits(LeadingZeros); |
| 833 | } else if (isNegative()) { |
| 834 | unsigned LeadingOnes = countMinLeadingOnes(); |
| 835 | LeadingOnes = LeadingOnes > LostBits ? LeadingOnes - LostBits : 0; |
| 836 | Result.One.setHighBits(LeadingOnes); |
| 837 | } |
| 838 | |
| 839 | return Result; |
| 840 | } |
| 841 | |
| 842 | static KnownBits computeForSatAddSub(bool Add, bool Signed, |
| 843 | const KnownBits &LHS, |
| 844 | const KnownBits &RHS) { |
| 845 | // We don't see NSW even for sadd/ssub as we want to check if the result has |
| 846 | // signed overflow. |
| 847 | unsigned BitWidth = LHS.getBitWidth(); |
| 848 | |
| 849 | std::optional<bool> Overflow; |
| 850 | // Even if we can't entirely rule out overflow, we may be able to rule out |
| 851 | // overflow in one direction. This allows us to potentially keep some of the |
| 852 | // add/sub bits. I.e if we can't overflow in the positive direction we won't |
| 853 | // clamp to INT_MAX so we can keep low 0s from the add/sub result. |
| 854 | bool MayNegClamp = true; |
| 855 | bool MayPosClamp = true; |
| 856 | if (Signed) { |
| 857 | // Easy cases we can rule out any overflow. |
| 858 | if (Add && ((LHS.isNegative() && RHS.isNonNegative()) || |
| 859 | (LHS.isNonNegative() && RHS.isNegative()))) |
| 860 | Overflow = false; |
| 861 | else if (!Add && (((LHS.isNegative() && RHS.isNegative()) || |
| 862 | (LHS.isNonNegative() && RHS.isNonNegative())))) |
| 863 | Overflow = false; |
| 864 | else { |
| 865 | // Check if we may overflow. If we can't rule out overflow then check if |
| 866 | // we can rule out a direction at least. |
| 867 | KnownBits UnsignedLHS = LHS; |
| 868 | KnownBits UnsignedRHS = RHS; |
| 869 | // Get version of LHS/RHS with clearer signbit. This allows us to detect |
| 870 | // how the addition/subtraction might overflow into the signbit. Then |
| 871 | // using the actual known signbits of LHS/RHS, we can figure out which |
| 872 | // overflows are/aren't possible. |
| 873 | UnsignedLHS.One.clearSignBit(); |
| 874 | UnsignedLHS.Zero.setSignBit(); |
| 875 | UnsignedRHS.One.clearSignBit(); |
| 876 | UnsignedRHS.Zero.setSignBit(); |
| 877 | KnownBits Res = |
| 878 | KnownBits::computeForAddSub(Add, /*NSW=*/false, |
| 879 | /*NUW=*/false, LHS: UnsignedLHS, RHS: UnsignedRHS); |
| 880 | if (Add) { |
| 881 | if (Res.isNegative()) { |
| 882 | // Only overflow scenario is Pos + Pos. |
| 883 | MayNegClamp = false; |
| 884 | // Pos + Pos will overflow with extra signbit. |
| 885 | if (LHS.isNonNegative() && RHS.isNonNegative()) |
| 886 | Overflow = true; |
| 887 | } else if (Res.isNonNegative()) { |
| 888 | // Only overflow scenario is Neg + Neg |
| 889 | MayPosClamp = false; |
| 890 | // Neg + Neg will overflow without extra signbit. |
| 891 | if (LHS.isNegative() && RHS.isNegative()) |
| 892 | Overflow = true; |
| 893 | } |
| 894 | // We will never clamp to the opposite sign of N-bit result. |
| 895 | if (LHS.isNegative() || RHS.isNegative()) |
| 896 | MayPosClamp = false; |
| 897 | if (LHS.isNonNegative() || RHS.isNonNegative()) |
| 898 | MayNegClamp = false; |
| 899 | } else { |
| 900 | if (Res.isNegative()) { |
| 901 | // Only overflow scenario is Neg - Pos. |
| 902 | MayPosClamp = false; |
| 903 | // Neg - Pos will overflow with extra signbit. |
| 904 | if (LHS.isNegative() && RHS.isNonNegative()) |
| 905 | Overflow = true; |
| 906 | } else if (Res.isNonNegative()) { |
| 907 | // Only overflow scenario is Pos - Neg. |
| 908 | MayNegClamp = false; |
| 909 | // Pos - Neg will overflow without extra signbit. |
| 910 | if (LHS.isNonNegative() && RHS.isNegative()) |
| 911 | Overflow = true; |
| 912 | } |
| 913 | // We will never clamp to the opposite sign of N-bit result. |
| 914 | if (LHS.isNegative() || RHS.isNonNegative()) |
| 915 | MayPosClamp = false; |
| 916 | if (LHS.isNonNegative() || RHS.isNegative()) |
| 917 | MayNegClamp = false; |
| 918 | } |
| 919 | } |
| 920 | // If we have ruled out all clamping, we will never overflow. |
| 921 | if (!MayNegClamp && !MayPosClamp) |
| 922 | Overflow = false; |
| 923 | } else if (Add) { |
| 924 | // uadd.sat |
| 925 | bool Of; |
| 926 | (void)LHS.getMaxValue().uadd_ov(RHS: RHS.getMaxValue(), Overflow&: Of); |
| 927 | if (!Of) { |
| 928 | Overflow = false; |
| 929 | } else { |
| 930 | (void)LHS.getMinValue().uadd_ov(RHS: RHS.getMinValue(), Overflow&: Of); |
| 931 | if (Of) |
| 932 | Overflow = true; |
| 933 | } |
| 934 | } else { |
| 935 | // usub.sat |
| 936 | bool Of; |
| 937 | (void)LHS.getMinValue().usub_ov(RHS: RHS.getMaxValue(), Overflow&: Of); |
| 938 | if (!Of) { |
| 939 | Overflow = false; |
| 940 | } else { |
| 941 | (void)LHS.getMaxValue().usub_ov(RHS: RHS.getMinValue(), Overflow&: Of); |
| 942 | if (Of) |
| 943 | Overflow = true; |
| 944 | } |
| 945 | } |
| 946 | |
| 947 | KnownBits Res = KnownBits::computeForAddSub(Add, /*NSW=*/Signed, |
| 948 | /*NUW=*/!Signed, LHS, RHS); |
| 949 | |
| 950 | if (Overflow) { |
| 951 | // We know whether or not we overflowed. |
| 952 | if (!(*Overflow)) { |
| 953 | // No overflow. |
| 954 | return Res; |
| 955 | } |
| 956 | |
| 957 | // We overflowed |
| 958 | APInt C; |
| 959 | if (Signed) { |
| 960 | // sadd.sat / ssub.sat |
| 961 | assert(!LHS.isSignUnknown() && |
| 962 | "We somehow know overflow without knowing input sign" ); |
| 963 | C = LHS.isNegative() ? APInt::getSignedMinValue(numBits: BitWidth) |
| 964 | : APInt::getSignedMaxValue(numBits: BitWidth); |
| 965 | } else if (Add) { |
| 966 | // uadd.sat |
| 967 | C = APInt::getMaxValue(numBits: BitWidth); |
| 968 | } else { |
| 969 | // uadd.sat |
| 970 | C = APInt::getMinValue(numBits: BitWidth); |
| 971 | } |
| 972 | |
| 973 | Res.One = C; |
| 974 | Res.Zero = ~C; |
| 975 | return Res; |
| 976 | } |
| 977 | |
| 978 | // We don't know if we overflowed. |
| 979 | if (Signed) { |
| 980 | // sadd.sat/ssub.sat |
| 981 | // We can keep our information about the sign bits. |
| 982 | if (MayPosClamp) |
| 983 | Res.Zero.clearLowBits(loBits: BitWidth - 1); |
| 984 | if (MayNegClamp) |
| 985 | Res.One.clearLowBits(loBits: BitWidth - 1); |
| 986 | } else if (Add) { |
| 987 | // uadd.sat |
| 988 | // We need to clear all the known zeros as we can only use the leading ones. |
| 989 | Res.Zero.clearAllBits(); |
| 990 | } else { |
| 991 | // usub.sat |
| 992 | // We need to clear all the known ones as we can only use the leading zero. |
| 993 | Res.One.clearAllBits(); |
| 994 | } |
| 995 | |
| 996 | return Res; |
| 997 | } |
| 998 | |
| 999 | KnownBits KnownBits::sadd_sat(const KnownBits &LHS, const KnownBits &RHS) { |
| 1000 | return computeForSatAddSub(/*Add*/ true, /*Signed*/ true, LHS, RHS); |
| 1001 | } |
| 1002 | KnownBits KnownBits::ssub_sat(const KnownBits &LHS, const KnownBits &RHS) { |
| 1003 | return computeForSatAddSub(/*Add*/ false, /*Signed*/ true, LHS, RHS); |
| 1004 | } |
| 1005 | KnownBits KnownBits::uadd_sat(const KnownBits &LHS, const KnownBits &RHS) { |
| 1006 | return computeForSatAddSub(/*Add*/ true, /*Signed*/ false, LHS, RHS); |
| 1007 | } |
| 1008 | KnownBits KnownBits::usub_sat(const KnownBits &LHS, const KnownBits &RHS) { |
| 1009 | return computeForSatAddSub(/*Add*/ false, /*Signed*/ false, LHS, RHS); |
| 1010 | } |
| 1011 | |
| 1012 | static KnownBits avgComputeU(KnownBits LHS, KnownBits RHS, bool IsCeil) { |
| 1013 | unsigned BitWidth = LHS.getBitWidth(); |
| 1014 | LHS = LHS.zext(BitWidth: BitWidth + 1); |
| 1015 | RHS = RHS.zext(BitWidth: BitWidth + 1); |
| 1016 | LHS = |
| 1017 | computeForAddCarry(LHS, RHS, /*CarryZero*/ !IsCeil, /*CarryOne*/ IsCeil); |
| 1018 | LHS = LHS.extractBits(NumBits: BitWidth, BitPosition: 1); |
| 1019 | return LHS; |
| 1020 | } |
| 1021 | |
| 1022 | KnownBits KnownBits::avgFloorS(const KnownBits &LHS, const KnownBits &RHS) { |
| 1023 | return flipSignBit(Val: avgFloorU(LHS: flipSignBit(Val: LHS), RHS: flipSignBit(Val: RHS))); |
| 1024 | } |
| 1025 | |
| 1026 | KnownBits KnownBits::avgFloorU(const KnownBits &LHS, const KnownBits &RHS) { |
| 1027 | return avgComputeU(LHS, RHS, /*IsCeil=*/false); |
| 1028 | } |
| 1029 | |
| 1030 | KnownBits KnownBits::avgCeilS(const KnownBits &LHS, const KnownBits &RHS) { |
| 1031 | return flipSignBit(Val: avgCeilU(LHS: flipSignBit(Val: LHS), RHS: flipSignBit(Val: RHS))); |
| 1032 | } |
| 1033 | |
| 1034 | KnownBits KnownBits::avgCeilU(const KnownBits &LHS, const KnownBits &RHS) { |
| 1035 | return avgComputeU(LHS, RHS, /*IsCeil=*/true); |
| 1036 | } |
| 1037 | |
| 1038 | KnownBits KnownBits::mul(const KnownBits &LHS, const KnownBits &RHS, |
| 1039 | bool NoUndefSelfMultiply) { |
| 1040 | unsigned BitWidth = LHS.getBitWidth(); |
| 1041 | assert(BitWidth == RHS.getBitWidth() && "Operand mismatch" ); |
| 1042 | assert((!NoUndefSelfMultiply || LHS == RHS) && |
| 1043 | "Self multiplication knownbits mismatch" ); |
| 1044 | |
| 1045 | // Compute the high known-0 bits by multiplying the unsigned max of each side. |
| 1046 | // Conservatively, M active bits * N active bits results in M + N bits in the |
| 1047 | // result. But if we know a value is a power-of-2 for example, then this |
| 1048 | // computes one more leading zero. |
| 1049 | // TODO: This could be generalized to number of sign bits (negative numbers). |
| 1050 | APInt UMaxLHS = LHS.getMaxValue(); |
| 1051 | APInt UMaxRHS = RHS.getMaxValue(); |
| 1052 | |
| 1053 | // For leading zeros in the result to be valid, the unsigned max product must |
| 1054 | // fit in the bitwidth (it must not overflow). |
| 1055 | bool HasOverflow; |
| 1056 | APInt UMaxResult = UMaxLHS.umul_ov(RHS: UMaxRHS, Overflow&: HasOverflow); |
| 1057 | unsigned LeadZ = HasOverflow ? 0 : UMaxResult.countl_zero(); |
| 1058 | |
| 1059 | // The result of the bottom bits of an integer multiply can be |
| 1060 | // inferred by looking at the bottom bits of both operands and |
| 1061 | // multiplying them together. |
| 1062 | // We can infer at least the minimum number of known trailing bits |
| 1063 | // of both operands. Depending on number of trailing zeros, we can |
| 1064 | // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming |
| 1065 | // a and b are divisible by m and n respectively. |
| 1066 | // We then calculate how many of those bits are inferrable and set |
| 1067 | // the output. For example, the i8 mul: |
| 1068 | // a = XXXX1100 (12) |
| 1069 | // b = XXXX1110 (14) |
| 1070 | // We know the bottom 3 bits are zero since the first can be divided by |
| 1071 | // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4). |
| 1072 | // Applying the multiplication to the trimmed arguments gets: |
| 1073 | // XX11 (3) |
| 1074 | // X111 (7) |
| 1075 | // ------- |
| 1076 | // XX11 |
| 1077 | // XX11 |
| 1078 | // XX11 |
| 1079 | // XX11 |
| 1080 | // ------- |
| 1081 | // XXXXX01 |
| 1082 | // Which allows us to infer the 2 LSBs. Since we're multiplying the result |
| 1083 | // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits. |
| 1084 | // The proof for this can be described as: |
| 1085 | // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) && |
| 1086 | // (C7 == (1 << (umin(countTrailingZeros(C1), C5) + |
| 1087 | // umin(countTrailingZeros(C2), C6) + |
| 1088 | // umin(C5 - umin(countTrailingZeros(C1), C5), |
| 1089 | // C6 - umin(countTrailingZeros(C2), C6)))) - 1) |
| 1090 | // %aa = shl i8 %a, C5 |
| 1091 | // %bb = shl i8 %b, C6 |
| 1092 | // %aaa = or i8 %aa, C1 |
| 1093 | // %bbb = or i8 %bb, C2 |
| 1094 | // %mul = mul i8 %aaa, %bbb |
| 1095 | // %mask = and i8 %mul, C7 |
| 1096 | // => |
| 1097 | // %mask = i8 ((C1*C2)&C7) |
| 1098 | // Where C5, C6 describe the known bits of %a, %b |
| 1099 | // C1, C2 describe the known bottom bits of %a, %b. |
| 1100 | // C7 describes the mask of the known bits of the result. |
| 1101 | |
| 1102 | // How many times we'd be able to divide each argument by 2 (shr by 1). |
| 1103 | // This gives us the number of trailing zeros on the multiplication result. |
| 1104 | unsigned TrailBitsKnownLHS = (LHS.Zero | LHS.One).countr_one(); |
| 1105 | unsigned TrailBitsKnownRHS = (RHS.Zero | RHS.One).countr_one(); |
| 1106 | unsigned TrailZeroLHS = LHS.countMinTrailingZeros(); |
| 1107 | unsigned TrailZeroRHS = RHS.countMinTrailingZeros(); |
| 1108 | unsigned TrailZ = TrailZeroLHS + TrailZeroRHS; |
| 1109 | |
| 1110 | // Figure out the fewest known-bits operand. |
| 1111 | unsigned SmallestOperand = std::min(a: TrailBitsKnownLHS - TrailZeroLHS, |
| 1112 | b: TrailBitsKnownRHS - TrailZeroRHS); |
| 1113 | unsigned ResultBitsKnown = std::min(a: SmallestOperand + TrailZ, b: BitWidth); |
| 1114 | |
| 1115 | // The lower ResultBitsKnown bits of this are known. |
| 1116 | APInt BottomKnown = LHS.One * RHS.One; |
| 1117 | |
| 1118 | KnownBits Res(BitWidth); |
| 1119 | Res.Zero.setHighBits(LeadZ); |
| 1120 | Res.Zero |= (~BottomKnown).getLoBits(numBits: ResultBitsKnown); |
| 1121 | Res.One = BottomKnown.getLoBits(numBits: ResultBitsKnown); |
| 1122 | |
| 1123 | if (NoUndefSelfMultiply) { |
| 1124 | // If X has at least TZ trailing zeroes, then bit (2 * TZ + 1) must be zero. |
| 1125 | unsigned TwoTZP1 = 2 * TrailZeroLHS + 1; |
| 1126 | if (TwoTZP1 < BitWidth) |
| 1127 | Res.Zero.setBit(TwoTZP1); |
| 1128 | |
| 1129 | // If X has exactly TZ trailing zeros, then bit (2 * TZ + 2) must also be |
| 1130 | // zero. |
| 1131 | if (TrailZeroLHS < BitWidth && LHS.One[TrailZeroLHS]) { |
| 1132 | unsigned TwoTZP2 = TwoTZP1 + 1; |
| 1133 | if (TwoTZP2 < BitWidth) |
| 1134 | Res.Zero.setBit(TwoTZP2); |
| 1135 | } |
| 1136 | } |
| 1137 | |
| 1138 | return Res; |
| 1139 | } |
| 1140 | |
| 1141 | KnownBits KnownBits::mulhs(const KnownBits &LHS, const KnownBits &RHS) { |
| 1142 | unsigned BitWidth = LHS.getBitWidth(); |
| 1143 | assert(BitWidth == RHS.getBitWidth() && "Operand mismatch" ); |
| 1144 | KnownBits WideLHS = LHS.sext(BitWidth: 2 * BitWidth); |
| 1145 | KnownBits WideRHS = RHS.sext(BitWidth: 2 * BitWidth); |
| 1146 | return mul(LHS: WideLHS, RHS: WideRHS).extractBits(NumBits: BitWidth, BitPosition: BitWidth); |
| 1147 | } |
| 1148 | |
| 1149 | KnownBits KnownBits::mulhu(const KnownBits &LHS, const KnownBits &RHS) { |
| 1150 | unsigned BitWidth = LHS.getBitWidth(); |
| 1151 | assert(BitWidth == RHS.getBitWidth() && "Operand mismatch" ); |
| 1152 | KnownBits WideLHS = LHS.zext(BitWidth: 2 * BitWidth); |
| 1153 | KnownBits WideRHS = RHS.zext(BitWidth: 2 * BitWidth); |
| 1154 | return mul(LHS: WideLHS, RHS: WideRHS).extractBits(NumBits: BitWidth, BitPosition: BitWidth); |
| 1155 | } |
| 1156 | |
| 1157 | static KnownBits divComputeLowBit(KnownBits Known, const KnownBits &LHS, |
| 1158 | const KnownBits &RHS, bool Exact) { |
| 1159 | |
| 1160 | if (!Exact) |
| 1161 | return Known; |
| 1162 | |
| 1163 | // If LHS is Odd, the result is Odd no matter what. |
| 1164 | // Odd / Odd -> Odd |
| 1165 | // Odd / Even -> Impossible (because its exact division) |
| 1166 | if (LHS.One[0]) |
| 1167 | Known.One.setBit(0); |
| 1168 | |
| 1169 | int MinTZ = |
| 1170 | (int)LHS.countMinTrailingZeros() - (int)RHS.countMaxTrailingZeros(); |
| 1171 | int MaxTZ = |
| 1172 | (int)LHS.countMaxTrailingZeros() - (int)RHS.countMinTrailingZeros(); |
| 1173 | if (MinTZ >= 0) { |
| 1174 | // Result has at least MinTZ trailing zeros. |
| 1175 | Known.Zero.setLowBits(MinTZ); |
| 1176 | if (MinTZ == MaxTZ) { |
| 1177 | // Result has exactly MinTZ trailing zeros. |
| 1178 | Known.One.setBit(MinTZ); |
| 1179 | } |
| 1180 | } else if (MaxTZ < 0) { |
| 1181 | // Poison Result |
| 1182 | Known.setAllZero(); |
| 1183 | } |
| 1184 | |
| 1185 | // In the KnownBits exhaustive tests, we have poison inputs for exact values |
| 1186 | // a LOT. If we have a conflict, just return all zeros. |
| 1187 | if (Known.hasConflict()) |
| 1188 | Known.setAllZero(); |
| 1189 | |
| 1190 | return Known; |
| 1191 | } |
| 1192 | |
| 1193 | KnownBits KnownBits::sdiv(const KnownBits &LHS, const KnownBits &RHS, |
| 1194 | bool Exact) { |
| 1195 | // Equivalent of `udiv`. We must have caught this before it was folded. |
| 1196 | if (LHS.isNonNegative() && RHS.isNonNegative()) |
| 1197 | return udiv(LHS, RHS, Exact); |
| 1198 | |
| 1199 | unsigned BitWidth = LHS.getBitWidth(); |
| 1200 | KnownBits Known(BitWidth); |
| 1201 | |
| 1202 | if (LHS.isZero() || RHS.isZero()) { |
| 1203 | // Result is either known Zero or UB. Return Zero either way. |
| 1204 | // Checking this earlier saves us a lot of special cases later on. |
| 1205 | Known.setAllZero(); |
| 1206 | return Known; |
| 1207 | } |
| 1208 | |
| 1209 | std::optional<APInt> Res; |
| 1210 | if (LHS.isNegative() && RHS.isNegative()) { |
| 1211 | // Result non-negative. |
| 1212 | APInt Denom = RHS.getSignedMaxValue(); |
| 1213 | APInt Num = LHS.getSignedMinValue(); |
| 1214 | // INT_MIN/-1 would be a poison result (impossible). Estimate the division |
| 1215 | // as signed max (we will only set sign bit in the result). |
| 1216 | Res = (Num.isMinSignedValue() && Denom.isAllOnes()) |
| 1217 | ? APInt::getSignedMaxValue(numBits: BitWidth) |
| 1218 | : Num.sdiv(RHS: Denom); |
| 1219 | } else if (LHS.isNegative() && RHS.isNonNegative()) { |
| 1220 | // Result is negative if Exact OR -LHS u>= RHS. |
| 1221 | if (Exact || (-LHS.getSignedMaxValue()).uge(RHS: RHS.getSignedMaxValue())) { |
| 1222 | APInt Denom = RHS.getSignedMinValue(); |
| 1223 | APInt Num = LHS.getSignedMinValue(); |
| 1224 | // If the denominator may be zero, use the lowest unknown bit as the |
| 1225 | // minimum non-zero denominator, matching udiv. |
| 1226 | if (Denom.isZero()) |
| 1227 | Denom.setBit(RHS.countMinTrailingZeros()); |
| 1228 | Res = Num.sdiv(RHS: Denom); |
| 1229 | } |
| 1230 | } else if (LHS.isStrictlyPositive() && RHS.isNegative()) { |
| 1231 | // Result is negative if Exact OR LHS u>= -RHS. |
| 1232 | if (Exact || LHS.getSignedMinValue().uge(RHS: -RHS.getSignedMinValue())) { |
| 1233 | APInt Denom = RHS.getSignedMaxValue(); |
| 1234 | APInt Num = LHS.getSignedMaxValue(); |
| 1235 | Res = Num.sdiv(RHS: Denom); |
| 1236 | } |
| 1237 | } |
| 1238 | |
| 1239 | if (Res) { |
| 1240 | if (Res->isNonNegative()) { |
| 1241 | unsigned LeadZ = Res->countLeadingZeros(); |
| 1242 | Known.Zero.setHighBits(LeadZ); |
| 1243 | } else { |
| 1244 | unsigned LeadO = Res->countLeadingOnes(); |
| 1245 | Known.One.setHighBits(LeadO); |
| 1246 | } |
| 1247 | } |
| 1248 | |
| 1249 | Known = divComputeLowBit(Known, LHS, RHS, Exact); |
| 1250 | return Known; |
| 1251 | } |
| 1252 | |
| 1253 | KnownBits KnownBits::udiv(const KnownBits &LHS, const KnownBits &RHS, |
| 1254 | bool Exact) { |
| 1255 | unsigned BitWidth = LHS.getBitWidth(); |
| 1256 | KnownBits Known(BitWidth); |
| 1257 | |
| 1258 | if (LHS.isZero() || RHS.isZero()) { |
| 1259 | // Result is either known Zero or UB. Return Zero either way. |
| 1260 | // Checking this earlier saves us a lot of special cases later on. |
| 1261 | Known.setAllZero(); |
| 1262 | return Known; |
| 1263 | } |
| 1264 | |
| 1265 | // A zero denominator is UB, so the minimum attainable denominator is the |
| 1266 | // smallest non-zero value consistent with the known bits. |
| 1267 | APInt MinDenom = RHS.getMinValue(); |
| 1268 | if (MinDenom.isZero()) |
| 1269 | MinDenom.setBit(RHS.countMinTrailingZeros()); |
| 1270 | |
| 1271 | // The quotient grows when the numerator grows and shrinks when the |
| 1272 | // denominator grows, and all four operand bounds are attainable, so the |
| 1273 | // result is in [MinNum / MaxDenom, MaxNum / MinDenom]. If the division is |
| 1274 | // exact we tighten the bounds on minimum result, so we round up the lower |
| 1275 | // bound to its ceiling. Keep the common leading bits of the two bounds (as |
| 1276 | // ConstantRange::toKnownBits does), which can include leading ones, not |
| 1277 | // just leading zeros. |
| 1278 | APInt MaxDenom = RHS.getMaxValue(); |
| 1279 | APInt MaxRes = LHS.getMaxValue().udiv(RHS: MinDenom); |
| 1280 | APInt MinRes = Exact ? APIntOps::RoundingUDiv(A: LHS.getMinValue(), B: MaxDenom, |
| 1281 | RM: APInt::Rounding::UP) |
| 1282 | : LHS.getMinValue().udiv(RHS: MaxDenom); |
| 1283 | |
| 1284 | Known = KnownBits::makeConstant(C: MinRes); |
| 1285 | if (std::optional<unsigned> DifferentBit = |
| 1286 | APIntOps::GetMostSignificantDifferentBit(A: MinRes, B: MaxRes)) { |
| 1287 | Known.Zero.clearLowBits(loBits: *DifferentBit + 1); |
| 1288 | Known.One.clearLowBits(loBits: *DifferentBit + 1); |
| 1289 | } |
| 1290 | |
| 1291 | Known = divComputeLowBit(Known, LHS, RHS, Exact); |
| 1292 | |
| 1293 | return Known; |
| 1294 | } |
| 1295 | |
| 1296 | KnownBits KnownBits::remGetLowBits(const KnownBits &LHS, const KnownBits &RHS) { |
| 1297 | unsigned BitWidth = LHS.getBitWidth(); |
| 1298 | if (!RHS.isZero() && RHS.Zero[0]) { |
| 1299 | // rem X, Y where Y[0:N] is zero will preserve X[0:N] in the result. |
| 1300 | unsigned RHSZeros = RHS.countMinTrailingZeros(); |
| 1301 | APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: RHSZeros); |
| 1302 | APInt OnesMask = LHS.One & Mask; |
| 1303 | APInt ZerosMask = LHS.Zero & Mask; |
| 1304 | return KnownBits(ZerosMask, OnesMask); |
| 1305 | } |
| 1306 | return KnownBits(BitWidth); |
| 1307 | } |
| 1308 | |
| 1309 | KnownBits KnownBits::urem(const KnownBits &LHS, const KnownBits &RHS) { |
| 1310 | // L % R for L < R is L. (A zero divisor is UB, and if MinR is zero this |
| 1311 | // condition cannot hold anyway.) |
| 1312 | if (LHS.getMaxValue().ult(RHS: RHS.getMinValue())) |
| 1313 | return LHS; |
| 1314 | |
| 1315 | KnownBits Known = remGetLowBits(LHS, RHS); |
| 1316 | |
| 1317 | // L % R is at most L and, since a zero divisor is UB, strictly less than |
| 1318 | // R, so the result has at least as many leading zeros as |
| 1319 | // umin(MaxL, MaxR - 1), as in ConstantRange::urem. For a power-of-two |
| 1320 | // constant R this computes the exact L & (R - 1): remGetLowBits preserved |
| 1321 | // the low bits and this clears the bits from log2(R) up. If R is known to |
| 1322 | // be zero, MaxR - 1 wraps around, which is harmless: every input is UB. |
| 1323 | APInt MaxRes = APIntOps::umin(A: LHS.getMaxValue(), B: RHS.getMaxValue() - 1); |
| 1324 | Known.Zero.setHighBits(MaxRes.countLeadingZeros()); |
| 1325 | return Known; |
| 1326 | } |
| 1327 | |
| 1328 | KnownBits KnownBits::srem(const KnownBits &LHS, const KnownBits &RHS) { |
| 1329 | KnownBits Known = remGetLowBits(LHS, RHS); |
| 1330 | if (RHS.isConstant() && RHS.getConstant().isPowerOf2()) { |
| 1331 | // NB: Low bits are set in `remGetLowBits`. |
| 1332 | APInt LowBits = RHS.getConstant() - 1; |
| 1333 | // If the first operand is non-negative or has all low bits zero, then |
| 1334 | // the upper bits are all zero. |
| 1335 | if (LHS.isNonNegative() || LowBits.isSubsetOf(RHS: LHS.Zero)) |
| 1336 | Known.Zero |= ~LowBits; |
| 1337 | |
| 1338 | // If the first operand is negative and not all low bits are zero, then |
| 1339 | // the upper bits are all one. |
| 1340 | if (LHS.isNegative() && LowBits.intersects(RHS: LHS.One)) |
| 1341 | Known.One |= ~LowBits; |
| 1342 | return Known; |
| 1343 | } |
| 1344 | |
| 1345 | // The sign bit is the LHS's sign bit, except when the result of the |
| 1346 | // remainder is zero. The magnitude of the result should be less than or |
| 1347 | // equal to the magnitude of either operand. |
| 1348 | if (LHS.isNegative() && Known.isNonZero()) |
| 1349 | Known.One.setHighBits( |
| 1350 | std::max(a: LHS.countMinLeadingOnes(), b: RHS.countMinSignBits())); |
| 1351 | else if (LHS.isNonNegative()) |
| 1352 | Known.Zero.setHighBits( |
| 1353 | std::max(a: LHS.countMinLeadingZeros(), b: RHS.countMinSignBits())); |
| 1354 | return Known; |
| 1355 | } |
| 1356 | |
| 1357 | KnownBits &KnownBits::operator&=(const KnownBits &RHS) { |
| 1358 | // Result bit is 0 if either operand bit is 0. |
| 1359 | Zero |= RHS.Zero; |
| 1360 | // Result bit is 1 if both operand bits are 1. |
| 1361 | One &= RHS.One; |
| 1362 | return *this; |
| 1363 | } |
| 1364 | |
| 1365 | KnownBits &KnownBits::operator|=(const KnownBits &RHS) { |
| 1366 | // Result bit is 0 if both operand bits are 0. |
| 1367 | Zero &= RHS.Zero; |
| 1368 | // Result bit is 1 if either operand bit is 1. |
| 1369 | One |= RHS.One; |
| 1370 | return *this; |
| 1371 | } |
| 1372 | |
| 1373 | KnownBits &KnownBits::operator^=(const KnownBits &RHS) { |
| 1374 | // Result bit is 0 if both operand bits are 0 or both are 1. |
| 1375 | APInt Z = (Zero & RHS.Zero) | (One & RHS.One); |
| 1376 | // Result bit is 1 if one operand bit is 0 and the other is 1. |
| 1377 | One = (Zero & RHS.One) | (One & RHS.Zero); |
| 1378 | Zero = std::move(Z); |
| 1379 | return *this; |
| 1380 | } |
| 1381 | |
| 1382 | KnownBits KnownBits::blsi() const { |
| 1383 | unsigned BitWidth = getBitWidth(); |
| 1384 | KnownBits Known(Zero, APInt(BitWidth, 0)); |
| 1385 | unsigned Max = countMaxTrailingZeros(); |
| 1386 | Known.Zero.setBitsFrom(std::min(a: Max + 1, b: BitWidth)); |
| 1387 | unsigned Min = countMinTrailingZeros(); |
| 1388 | if (Max == Min && Max < BitWidth) |
| 1389 | Known.One.setBit(Max); |
| 1390 | return Known; |
| 1391 | } |
| 1392 | |
| 1393 | KnownBits KnownBits::blsmsk() const { |
| 1394 | unsigned BitWidth = getBitWidth(); |
| 1395 | KnownBits Known(BitWidth); |
| 1396 | unsigned Max = countMaxTrailingZeros(); |
| 1397 | Known.Zero.setBitsFrom(std::min(a: Max + 1, b: BitWidth)); |
| 1398 | unsigned Min = countMinTrailingZeros(); |
| 1399 | Known.One.setLowBits(std::min(a: Min + 1, b: BitWidth)); |
| 1400 | return Known; |
| 1401 | } |
| 1402 | |
| 1403 | void KnownBits::print(raw_ostream &OS) const { |
| 1404 | unsigned BitWidth = getBitWidth(); |
| 1405 | for (unsigned I = 0; I < BitWidth; ++I) { |
| 1406 | unsigned N = BitWidth - I - 1; |
| 1407 | if (Zero[N] && One[N]) |
| 1408 | OS << "!" ; |
| 1409 | else if (Zero[N]) |
| 1410 | OS << "0" ; |
| 1411 | else if (One[N]) |
| 1412 | OS << "1" ; |
| 1413 | else |
| 1414 | OS << "?" ; |
| 1415 | } |
| 1416 | } |
| 1417 | |
| 1418 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 1419 | LLVM_DUMP_METHOD void KnownBits::dump() const { |
| 1420 | print(dbgs()); |
| 1421 | dbgs() << "\n" ; |
| 1422 | } |
| 1423 | #endif |
| 1424 | |
| 1425 | unsigned llvm::SignBitsOps::rot(unsigned SrcSignBits, unsigned BitWidth, |
| 1426 | std::optional<APInt> RotAmt, |
| 1427 | bool IsRotateRight) { |
| 1428 | // If we're rotating an 0/-1 value, then it stays an 0/-1 value. |
| 1429 | if (SrcSignBits == BitWidth) |
| 1430 | return BitWidth; |
| 1431 | |
| 1432 | if (!RotAmt) |
| 1433 | return 1; |
| 1434 | |
| 1435 | unsigned Amt = RotAmt->urem(RHS: BitWidth); |
| 1436 | |
| 1437 | // Handle rotate right by N like a rotate left by BitWidth-N. |
| 1438 | if (IsRotateRight) |
| 1439 | Amt = (BitWidth - Amt) % BitWidth; |
| 1440 | |
| 1441 | // If we aren't rotating out all of the known-in sign bits, return the |
| 1442 | // number that are left. This handles rotl(sext(x), 1) for example. |
| 1443 | if (SrcSignBits > Amt + 1) |
| 1444 | return SrcSignBits - Amt; |
| 1445 | return 1; |
| 1446 | } |
| 1447 | |