1 //===- APFixedPoint.cpp - Fixed point constant handling ---------*- 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 /// Defines the implementation for the fixed point number interface. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/ADT/APFixedPoint.h" 15 #include "llvm/ADT/APFloat.h" 16 17 namespace llvm { 18 19 APFixedPoint APFixedPoint::convert(const FixedPointSemantics &DstSema, 20 bool *Overflow) const { 21 APSInt NewVal = Val; 22 unsigned DstWidth = DstSema.getWidth(); 23 unsigned DstScale = DstSema.getScale(); 24 bool Upscaling = DstScale > getScale(); 25 if (Overflow) 26 *Overflow = false; 27 28 if (Upscaling) { 29 NewVal = NewVal.extend(NewVal.getBitWidth() + DstScale - getScale()); 30 NewVal <<= (DstScale - getScale()); 31 } else { 32 NewVal >>= (getScale() - DstScale); 33 } 34 35 auto Mask = APInt::getBitsSetFrom( 36 NewVal.getBitWidth(), 37 std::min(DstScale + DstSema.getIntegralBits(), NewVal.getBitWidth())); 38 APInt Masked(NewVal & Mask); 39 40 // Change in the bits above the sign 41 if (!(Masked == Mask || Masked == 0)) { 42 // Found overflow in the bits above the sign 43 if (DstSema.isSaturated()) 44 NewVal = NewVal.isNegative() ? Mask : ~Mask; 45 else if (Overflow) 46 *Overflow = true; 47 } 48 49 // If the dst semantics are unsigned, but our value is signed and negative, we 50 // clamp to zero. 51 if (!DstSema.isSigned() && NewVal.isSigned() && NewVal.isNegative()) { 52 // Found negative overflow for unsigned result 53 if (DstSema.isSaturated()) 54 NewVal = 0; 55 else if (Overflow) 56 *Overflow = true; 57 } 58 59 NewVal = NewVal.extOrTrunc(DstWidth); 60 NewVal.setIsSigned(DstSema.isSigned()); 61 return APFixedPoint(NewVal, DstSema); 62 } 63 64 int APFixedPoint::compare(const APFixedPoint &Other) const { 65 APSInt ThisVal = getValue(); 66 APSInt OtherVal = Other.getValue(); 67 bool ThisSigned = Val.isSigned(); 68 bool OtherSigned = OtherVal.isSigned(); 69 unsigned OtherScale = Other.getScale(); 70 unsigned OtherWidth = OtherVal.getBitWidth(); 71 72 unsigned CommonWidth = std::max(Val.getBitWidth(), OtherWidth); 73 74 // Prevent overflow in the event the widths are the same but the scales differ 75 CommonWidth += getScale() >= OtherScale ? getScale() - OtherScale 76 : OtherScale - getScale(); 77 78 ThisVal = ThisVal.extOrTrunc(CommonWidth); 79 OtherVal = OtherVal.extOrTrunc(CommonWidth); 80 81 unsigned CommonScale = std::max(getScale(), OtherScale); 82 ThisVal = ThisVal.shl(CommonScale - getScale()); 83 OtherVal = OtherVal.shl(CommonScale - OtherScale); 84 85 if (ThisSigned && OtherSigned) { 86 if (ThisVal.sgt(OtherVal)) 87 return 1; 88 else if (ThisVal.slt(OtherVal)) 89 return -1; 90 } else if (!ThisSigned && !OtherSigned) { 91 if (ThisVal.ugt(OtherVal)) 92 return 1; 93 else if (ThisVal.ult(OtherVal)) 94 return -1; 95 } else if (ThisSigned && !OtherSigned) { 96 if (ThisVal.isSignBitSet()) 97 return -1; 98 else if (ThisVal.ugt(OtherVal)) 99 return 1; 100 else if (ThisVal.ult(OtherVal)) 101 return -1; 102 } else { 103 // !ThisSigned && OtherSigned 104 if (OtherVal.isSignBitSet()) 105 return 1; 106 else if (ThisVal.ugt(OtherVal)) 107 return 1; 108 else if (ThisVal.ult(OtherVal)) 109 return -1; 110 } 111 112 return 0; 113 } 114 115 APFixedPoint APFixedPoint::getMax(const FixedPointSemantics &Sema) { 116 bool IsUnsigned = !Sema.isSigned(); 117 auto Val = APSInt::getMaxValue(Sema.getWidth(), IsUnsigned); 118 if (IsUnsigned && Sema.hasUnsignedPadding()) 119 Val = Val.lshr(1); 120 return APFixedPoint(Val, Sema); 121 } 122 123 APFixedPoint APFixedPoint::getMin(const FixedPointSemantics &Sema) { 124 auto Val = APSInt::getMinValue(Sema.getWidth(), !Sema.isSigned()); 125 return APFixedPoint(Val, Sema); 126 } 127 128 bool FixedPointSemantics::fitsInFloatSemantics( 129 const fltSemantics &FloatSema) const { 130 // A fixed point semantic fits in a floating point semantic if the maximum 131 // and minimum values as integers of the fixed point semantic can fit in the 132 // floating point semantic. 133 134 // If these values do not fit, then a floating point rescaling of the true 135 // maximum/minimum value will not fit either, so the floating point semantic 136 // cannot be used to perform such a rescaling. 137 138 APSInt MaxInt = APFixedPoint::getMax(*this).getValue(); 139 APFloat F(FloatSema); 140 APFloat::opStatus Status = F.convertFromAPInt(MaxInt, MaxInt.isSigned(), 141 APFloat::rmNearestTiesToAway); 142 if ((Status & APFloat::opOverflow) || !isSigned()) 143 return !(Status & APFloat::opOverflow); 144 145 APSInt MinInt = APFixedPoint::getMin(*this).getValue(); 146 Status = F.convertFromAPInt(MinInt, MinInt.isSigned(), 147 APFloat::rmNearestTiesToAway); 148 return !(Status & APFloat::opOverflow); 149 } 150 151 FixedPointSemantics FixedPointSemantics::getCommonSemantics( 152 const FixedPointSemantics &Other) const { 153 unsigned CommonScale = std::max(getScale(), Other.getScale()); 154 unsigned CommonWidth = 155 std::max(getIntegralBits(), Other.getIntegralBits()) + CommonScale; 156 157 bool ResultIsSigned = isSigned() || Other.isSigned(); 158 bool ResultIsSaturated = isSaturated() || Other.isSaturated(); 159 bool ResultHasUnsignedPadding = false; 160 if (!ResultIsSigned) { 161 // Both are unsigned. 162 ResultHasUnsignedPadding = hasUnsignedPadding() && 163 Other.hasUnsignedPadding() && !ResultIsSaturated; 164 } 165 166 // If the result is signed, add an extra bit for the sign. Otherwise, if it is 167 // unsigned and has unsigned padding, we only need to add the extra padding 168 // bit back if we are not saturating. 169 if (ResultIsSigned || ResultHasUnsignedPadding) 170 CommonWidth++; 171 172 return FixedPointSemantics(CommonWidth, CommonScale, ResultIsSigned, 173 ResultIsSaturated, ResultHasUnsignedPadding); 174 } 175 176 APFixedPoint APFixedPoint::add(const APFixedPoint &Other, 177 bool *Overflow) const { 178 auto CommonFXSema = Sema.getCommonSemantics(Other.getSemantics()); 179 APFixedPoint ConvertedThis = convert(CommonFXSema); 180 APFixedPoint ConvertedOther = Other.convert(CommonFXSema); 181 APSInt ThisVal = ConvertedThis.getValue(); 182 APSInt OtherVal = ConvertedOther.getValue(); 183 bool Overflowed = false; 184 185 APSInt Result; 186 if (CommonFXSema.isSaturated()) { 187 Result = CommonFXSema.isSigned() ? ThisVal.sadd_sat(OtherVal) 188 : ThisVal.uadd_sat(OtherVal); 189 } else { 190 Result = ThisVal.isSigned() ? ThisVal.sadd_ov(OtherVal, Overflowed) 191 : ThisVal.uadd_ov(OtherVal, Overflowed); 192 } 193 194 if (Overflow) 195 *Overflow = Overflowed; 196 197 return APFixedPoint(Result, CommonFXSema); 198 } 199 200 APFixedPoint APFixedPoint::sub(const APFixedPoint &Other, 201 bool *Overflow) const { 202 auto CommonFXSema = Sema.getCommonSemantics(Other.getSemantics()); 203 APFixedPoint ConvertedThis = convert(CommonFXSema); 204 APFixedPoint ConvertedOther = Other.convert(CommonFXSema); 205 APSInt ThisVal = ConvertedThis.getValue(); 206 APSInt OtherVal = ConvertedOther.getValue(); 207 bool Overflowed = false; 208 209 APSInt Result; 210 if (CommonFXSema.isSaturated()) { 211 Result = CommonFXSema.isSigned() ? ThisVal.ssub_sat(OtherVal) 212 : ThisVal.usub_sat(OtherVal); 213 } else { 214 Result = ThisVal.isSigned() ? ThisVal.ssub_ov(OtherVal, Overflowed) 215 : ThisVal.usub_ov(OtherVal, Overflowed); 216 } 217 218 if (Overflow) 219 *Overflow = Overflowed; 220 221 return APFixedPoint(Result, CommonFXSema); 222 } 223 224 APFixedPoint APFixedPoint::mul(const APFixedPoint &Other, 225 bool *Overflow) const { 226 auto CommonFXSema = Sema.getCommonSemantics(Other.getSemantics()); 227 APFixedPoint ConvertedThis = convert(CommonFXSema); 228 APFixedPoint ConvertedOther = Other.convert(CommonFXSema); 229 APSInt ThisVal = ConvertedThis.getValue(); 230 APSInt OtherVal = ConvertedOther.getValue(); 231 bool Overflowed = false; 232 233 // Widen the LHS and RHS so we can perform a full multiplication. 234 unsigned Wide = CommonFXSema.getWidth() * 2; 235 if (CommonFXSema.isSigned()) { 236 ThisVal = ThisVal.sextOrSelf(Wide); 237 OtherVal = OtherVal.sextOrSelf(Wide); 238 } else { 239 ThisVal = ThisVal.zextOrSelf(Wide); 240 OtherVal = OtherVal.zextOrSelf(Wide); 241 } 242 243 // Perform the full multiplication and downscale to get the same scale. 244 // 245 // Note that the right shifts here perform an implicit downwards rounding. 246 // This rounding could discard bits that would technically place the result 247 // outside the representable range. We interpret the spec as allowing us to 248 // perform the rounding step first, avoiding the overflow case that would 249 // arise. 250 APSInt Result; 251 if (CommonFXSema.isSigned()) 252 Result = ThisVal.smul_ov(OtherVal, Overflowed) 253 .ashr(CommonFXSema.getScale()); 254 else 255 Result = ThisVal.umul_ov(OtherVal, Overflowed) 256 .lshr(CommonFXSema.getScale()); 257 assert(!Overflowed && "Full multiplication cannot overflow!"); 258 Result.setIsSigned(CommonFXSema.isSigned()); 259 260 // If our result lies outside of the representative range of the common 261 // semantic, we either have overflow or saturation. 262 APSInt Max = APFixedPoint::getMax(CommonFXSema).getValue() 263 .extOrTrunc(Wide); 264 APSInt Min = APFixedPoint::getMin(CommonFXSema).getValue() 265 .extOrTrunc(Wide); 266 if (CommonFXSema.isSaturated()) { 267 if (Result < Min) 268 Result = Min; 269 else if (Result > Max) 270 Result = Max; 271 } else 272 Overflowed = Result < Min || Result > Max; 273 274 if (Overflow) 275 *Overflow = Overflowed; 276 277 return APFixedPoint(Result.sextOrTrunc(CommonFXSema.getWidth()), 278 CommonFXSema); 279 } 280 281 APFixedPoint APFixedPoint::div(const APFixedPoint &Other, 282 bool *Overflow) const { 283 auto CommonFXSema = Sema.getCommonSemantics(Other.getSemantics()); 284 APFixedPoint ConvertedThis = convert(CommonFXSema); 285 APFixedPoint ConvertedOther = Other.convert(CommonFXSema); 286 APSInt ThisVal = ConvertedThis.getValue(); 287 APSInt OtherVal = ConvertedOther.getValue(); 288 bool Overflowed = false; 289 290 // Widen the LHS and RHS so we can perform a full division. 291 unsigned Wide = CommonFXSema.getWidth() * 2; 292 if (CommonFXSema.isSigned()) { 293 ThisVal = ThisVal.sextOrSelf(Wide); 294 OtherVal = OtherVal.sextOrSelf(Wide); 295 } else { 296 ThisVal = ThisVal.zextOrSelf(Wide); 297 OtherVal = OtherVal.zextOrSelf(Wide); 298 } 299 300 // Upscale to compensate for the loss of precision from division, and 301 // perform the full division. 302 ThisVal = ThisVal.shl(CommonFXSema.getScale()); 303 APSInt Result; 304 if (CommonFXSema.isSigned()) { 305 APInt Rem; 306 APInt::sdivrem(ThisVal, OtherVal, Result, Rem); 307 // If the quotient is negative and the remainder is nonzero, round 308 // towards negative infinity by subtracting epsilon from the result. 309 if (ThisVal.isNegative() != OtherVal.isNegative() && !Rem.isZero()) 310 Result = Result - 1; 311 } else 312 Result = ThisVal.udiv(OtherVal); 313 Result.setIsSigned(CommonFXSema.isSigned()); 314 315 // If our result lies outside of the representative range of the common 316 // semantic, we either have overflow or saturation. 317 APSInt Max = APFixedPoint::getMax(CommonFXSema).getValue() 318 .extOrTrunc(Wide); 319 APSInt Min = APFixedPoint::getMin(CommonFXSema).getValue() 320 .extOrTrunc(Wide); 321 if (CommonFXSema.isSaturated()) { 322 if (Result < Min) 323 Result = Min; 324 else if (Result > Max) 325 Result = Max; 326 } else 327 Overflowed = Result < Min || Result > Max; 328 329 if (Overflow) 330 *Overflow = Overflowed; 331 332 return APFixedPoint(Result.sextOrTrunc(CommonFXSema.getWidth()), 333 CommonFXSema); 334 } 335 336 APFixedPoint APFixedPoint::shl(unsigned Amt, bool *Overflow) const { 337 APSInt ThisVal = Val; 338 bool Overflowed = false; 339 340 // Widen the LHS. 341 unsigned Wide = Sema.getWidth() * 2; 342 if (Sema.isSigned()) 343 ThisVal = ThisVal.sextOrSelf(Wide); 344 else 345 ThisVal = ThisVal.zextOrSelf(Wide); 346 347 // Clamp the shift amount at the original width, and perform the shift. 348 Amt = std::min(Amt, ThisVal.getBitWidth()); 349 APSInt Result = ThisVal << Amt; 350 Result.setIsSigned(Sema.isSigned()); 351 352 // If our result lies outside of the representative range of the 353 // semantic, we either have overflow or saturation. 354 APSInt Max = APFixedPoint::getMax(Sema).getValue().extOrTrunc(Wide); 355 APSInt Min = APFixedPoint::getMin(Sema).getValue().extOrTrunc(Wide); 356 if (Sema.isSaturated()) { 357 if (Result < Min) 358 Result = Min; 359 else if (Result > Max) 360 Result = Max; 361 } else 362 Overflowed = Result < Min || Result > Max; 363 364 if (Overflow) 365 *Overflow = Overflowed; 366 367 return APFixedPoint(Result.sextOrTrunc(Sema.getWidth()), Sema); 368 } 369 370 void APFixedPoint::toString(SmallVectorImpl<char> &Str) const { 371 APSInt Val = getValue(); 372 unsigned Scale = getScale(); 373 374 if (Val.isSigned() && Val.isNegative() && Val != -Val) { 375 Val = -Val; 376 Str.push_back('-'); 377 } 378 379 APSInt IntPart = Val >> Scale; 380 381 // Add 4 digits to hold the value after multiplying 10 (the radix) 382 unsigned Width = Val.getBitWidth() + 4; 383 APInt FractPart = Val.zextOrTrunc(Scale).zext(Width); 384 APInt FractPartMask = APInt::getAllOnes(Scale).zext(Width); 385 APInt RadixInt = APInt(Width, 10); 386 387 IntPart.toString(Str, /*Radix=*/10); 388 Str.push_back('.'); 389 do { 390 (FractPart * RadixInt) 391 .lshr(Scale) 392 .toString(Str, /*Radix=*/10, Val.isSigned()); 393 FractPart = (FractPart * RadixInt) & FractPartMask; 394 } while (FractPart != 0); 395 } 396 397 APFixedPoint APFixedPoint::negate(bool *Overflow) const { 398 if (!isSaturated()) { 399 if (Overflow) 400 *Overflow = 401 (!isSigned() && Val != 0) || (isSigned() && Val.isMinSignedValue()); 402 return APFixedPoint(-Val, Sema); 403 } 404 405 // We never overflow for saturation 406 if (Overflow) 407 *Overflow = false; 408 409 if (isSigned()) 410 return Val.isMinSignedValue() ? getMax(Sema) : APFixedPoint(-Val, Sema); 411 else 412 return APFixedPoint(Sema); 413 } 414 415 APSInt APFixedPoint::convertToInt(unsigned DstWidth, bool DstSign, 416 bool *Overflow) const { 417 APSInt Result = getIntPart(); 418 unsigned SrcWidth = getWidth(); 419 420 APSInt DstMin = APSInt::getMinValue(DstWidth, !DstSign); 421 APSInt DstMax = APSInt::getMaxValue(DstWidth, !DstSign); 422 423 if (SrcWidth < DstWidth) { 424 Result = Result.extend(DstWidth); 425 } else if (SrcWidth > DstWidth) { 426 DstMin = DstMin.extend(SrcWidth); 427 DstMax = DstMax.extend(SrcWidth); 428 } 429 430 if (Overflow) { 431 if (Result.isSigned() && !DstSign) { 432 *Overflow = Result.isNegative() || Result.ugt(DstMax); 433 } else if (Result.isUnsigned() && DstSign) { 434 *Overflow = Result.ugt(DstMax); 435 } else { 436 *Overflow = Result < DstMin || Result > DstMax; 437 } 438 } 439 440 Result.setIsSigned(DstSign); 441 return Result.extOrTrunc(DstWidth); 442 } 443 444 const fltSemantics *APFixedPoint::promoteFloatSemantics(const fltSemantics *S) { 445 if (S == &APFloat::BFloat()) 446 return &APFloat::IEEEdouble(); 447 else if (S == &APFloat::IEEEhalf()) 448 return &APFloat::IEEEsingle(); 449 else if (S == &APFloat::IEEEsingle()) 450 return &APFloat::IEEEdouble(); 451 else if (S == &APFloat::IEEEdouble()) 452 return &APFloat::IEEEquad(); 453 llvm_unreachable("Could not promote float type!"); 454 } 455 456 APFloat APFixedPoint::convertToFloat(const fltSemantics &FloatSema) const { 457 // For some operations, rounding mode has an effect on the result, while 458 // other operations are lossless and should never result in rounding. 459 // To signify which these operations are, we define two rounding modes here. 460 APFloat::roundingMode RM = APFloat::rmNearestTiesToEven; 461 APFloat::roundingMode LosslessRM = APFloat::rmTowardZero; 462 463 // Make sure that we are operating in a type that works with this fixed-point 464 // semantic. 465 const fltSemantics *OpSema = &FloatSema; 466 while (!Sema.fitsInFloatSemantics(*OpSema)) 467 OpSema = promoteFloatSemantics(OpSema); 468 469 // Convert the fixed point value bits as an integer. If the floating point 470 // value does not have the required precision, we will round according to the 471 // given mode. 472 APFloat Flt(*OpSema); 473 APFloat::opStatus S = Flt.convertFromAPInt(Val, Sema.isSigned(), RM); 474 475 // If we cared about checking for precision loss, we could look at this 476 // status. 477 (void)S; 478 479 // Scale down the integer value in the float to match the correct scaling 480 // factor. 481 APFloat ScaleFactor(std::pow(2, -(int)Sema.getScale())); 482 bool Ignored; 483 ScaleFactor.convert(*OpSema, LosslessRM, &Ignored); 484 Flt.multiply(ScaleFactor, LosslessRM); 485 486 if (OpSema != &FloatSema) 487 Flt.convert(FloatSema, RM, &Ignored); 488 489 return Flt; 490 } 491 492 APFixedPoint APFixedPoint::getFromIntValue(const APSInt &Value, 493 const FixedPointSemantics &DstFXSema, 494 bool *Overflow) { 495 FixedPointSemantics IntFXSema = FixedPointSemantics::GetIntegerSemantics( 496 Value.getBitWidth(), Value.isSigned()); 497 return APFixedPoint(Value, IntFXSema).convert(DstFXSema, Overflow); 498 } 499 500 APFixedPoint 501 APFixedPoint::getFromFloatValue(const APFloat &Value, 502 const FixedPointSemantics &DstFXSema, 503 bool *Overflow) { 504 // For some operations, rounding mode has an effect on the result, while 505 // other operations are lossless and should never result in rounding. 506 // To signify which these operations are, we define two rounding modes here, 507 // even though they are the same mode. 508 APFloat::roundingMode RM = APFloat::rmTowardZero; 509 APFloat::roundingMode LosslessRM = APFloat::rmTowardZero; 510 511 const fltSemantics &FloatSema = Value.getSemantics(); 512 513 if (Value.isNaN()) { 514 // Handle NaN immediately. 515 if (Overflow) 516 *Overflow = true; 517 return APFixedPoint(DstFXSema); 518 } 519 520 // Make sure that we are operating in a type that works with this fixed-point 521 // semantic. 522 const fltSemantics *OpSema = &FloatSema; 523 while (!DstFXSema.fitsInFloatSemantics(*OpSema)) 524 OpSema = promoteFloatSemantics(OpSema); 525 526 APFloat Val = Value; 527 528 bool Ignored; 529 if (&FloatSema != OpSema) 530 Val.convert(*OpSema, LosslessRM, &Ignored); 531 532 // Scale up the float so that the 'fractional' part of the mantissa ends up in 533 // the integer range instead. Rounding mode is irrelevant here. 534 // It is fine if this overflows to infinity even for saturating types, 535 // since we will use floating point comparisons to check for saturation. 536 APFloat ScaleFactor(std::pow(2, DstFXSema.getScale())); 537 ScaleFactor.convert(*OpSema, LosslessRM, &Ignored); 538 Val.multiply(ScaleFactor, LosslessRM); 539 540 // Convert to the integral representation of the value. This rounding mode 541 // is significant. 542 APSInt Res(DstFXSema.getWidth(), !DstFXSema.isSigned()); 543 Val.convertToInteger(Res, RM, &Ignored); 544 545 // Round the integral value and scale back. This makes the 546 // overflow calculations below work properly. If we do not round here, 547 // we risk checking for overflow with a value that is outside the 548 // representable range of the fixed-point semantic even though no overflow 549 // would occur had we rounded first. 550 ScaleFactor = APFloat(std::pow(2, -(int)DstFXSema.getScale())); 551 ScaleFactor.convert(*OpSema, LosslessRM, &Ignored); 552 Val.roundToIntegral(RM); 553 Val.multiply(ScaleFactor, LosslessRM); 554 555 // Check for overflow/saturation by checking if the floating point value 556 // is outside the range representable by the fixed-point value. 557 APFloat FloatMax = getMax(DstFXSema).convertToFloat(*OpSema); 558 APFloat FloatMin = getMin(DstFXSema).convertToFloat(*OpSema); 559 bool Overflowed = false; 560 if (DstFXSema.isSaturated()) { 561 if (Val > FloatMax) 562 Res = getMax(DstFXSema).getValue(); 563 else if (Val < FloatMin) 564 Res = getMin(DstFXSema).getValue(); 565 } else 566 Overflowed = Val > FloatMax || Val < FloatMin; 567 568 if (Overflow) 569 *Overflow = Overflowed; 570 571 return APFixedPoint(Res, DstFXSema); 572 } 573 574 } // namespace llvm 575