1 | //===- ConstantRange.cpp - ConstantRange implementation -------------------===// |
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 | // Represent a range of possible values that may occur when the program is run |
10 | // for an integral value. This keeps track of a lower and upper bound for the |
11 | // constant, which MAY wrap around the end of the numeric range. To do this, it |
12 | // keeps track of a [lower, upper) bound, which specifies an interval just like |
13 | // STL iterators. When used with boolean values, the following are important |
14 | // ranges (other integral ranges use min/max values for special range values): |
15 | // |
16 | // [F, F) = {} = Empty set |
17 | // [T, F) = {T} |
18 | // [F, T) = {F} |
19 | // [T, T) = {F, T} = Full set |
20 | // |
21 | //===----------------------------------------------------------------------===// |
22 | |
23 | #include "llvm/IR/ConstantRange.h" |
24 | #include "llvm/ADT/APInt.h" |
25 | #include "llvm/Config/llvm-config.h" |
26 | #include "llvm/IR/Constants.h" |
27 | #include "llvm/IR/InstrTypes.h" |
28 | #include "llvm/IR/Instruction.h" |
29 | #include "llvm/IR/Instructions.h" |
30 | #include "llvm/IR/Intrinsics.h" |
31 | #include "llvm/IR/Metadata.h" |
32 | #include "llvm/IR/Operator.h" |
33 | #include "llvm/Support/Compiler.h" |
34 | #include "llvm/Support/Debug.h" |
35 | #include "llvm/Support/ErrorHandling.h" |
36 | #include "llvm/Support/KnownBits.h" |
37 | #include "llvm/Support/raw_ostream.h" |
38 | #include <algorithm> |
39 | #include <cassert> |
40 | #include <cstdint> |
41 | #include <optional> |
42 | |
43 | using namespace llvm; |
44 | |
45 | ConstantRange::ConstantRange(uint32_t BitWidth, bool Full) |
46 | : Lower(Full ? APInt::getMaxValue(numBits: BitWidth) : APInt::getMinValue(numBits: BitWidth)), |
47 | Upper(Lower) {} |
48 | |
49 | ConstantRange::ConstantRange(APInt V) |
50 | : Lower(std::move(V)), Upper(Lower + 1) {} |
51 | |
52 | ConstantRange::ConstantRange(APInt L, APInt U) |
53 | : Lower(std::move(L)), Upper(std::move(U)) { |
54 | assert(Lower.getBitWidth() == Upper.getBitWidth() && |
55 | "ConstantRange with unequal bit widths" ); |
56 | assert((Lower != Upper || (Lower.isMaxValue() || Lower.isMinValue())) && |
57 | "Lower == Upper, but they aren't min or max value!" ); |
58 | } |
59 | |
60 | ConstantRange ConstantRange::fromKnownBits(const KnownBits &Known, |
61 | bool IsSigned) { |
62 | if (Known.hasConflict()) |
63 | return getEmpty(BitWidth: Known.getBitWidth()); |
64 | if (Known.isUnknown()) |
65 | return getFull(BitWidth: Known.getBitWidth()); |
66 | |
67 | // For unsigned ranges, or signed ranges with known sign bit, create a simple |
68 | // range between the smallest and largest possible value. |
69 | if (!IsSigned || Known.isNegative() || Known.isNonNegative()) |
70 | return ConstantRange(Known.getMinValue(), Known.getMaxValue() + 1); |
71 | |
72 | // If we don't know the sign bit, pick the lower bound as a negative number |
73 | // and the upper bound as a non-negative one. |
74 | APInt Lower = Known.getMinValue(), Upper = Known.getMaxValue(); |
75 | Lower.setSignBit(); |
76 | Upper.clearSignBit(); |
77 | return ConstantRange(Lower, Upper + 1); |
78 | } |
79 | |
80 | KnownBits ConstantRange::toKnownBits() const { |
81 | // TODO: We could return conflicting known bits here, but consumers are |
82 | // likely not prepared for that. |
83 | if (isEmptySet()) |
84 | return KnownBits(getBitWidth()); |
85 | |
86 | // We can only retain the top bits that are the same between min and max. |
87 | APInt Min = getUnsignedMin(); |
88 | APInt Max = getUnsignedMax(); |
89 | KnownBits Known = KnownBits::makeConstant(C: Min); |
90 | if (std::optional<unsigned> DifferentBit = |
91 | APIntOps::GetMostSignificantDifferentBit(A: Min, B: Max)) { |
92 | Known.Zero.clearLowBits(loBits: *DifferentBit + 1); |
93 | Known.One.clearLowBits(loBits: *DifferentBit + 1); |
94 | } |
95 | return Known; |
96 | } |
97 | |
98 | std::pair<ConstantRange, ConstantRange> ConstantRange::splitPosNeg() const { |
99 | uint32_t BW = getBitWidth(); |
100 | APInt Zero = APInt::getZero(numBits: BW), One = APInt(BW, 1); |
101 | APInt SignedMin = APInt::getSignedMinValue(numBits: BW); |
102 | // There are no positive 1-bit values. The 1 would get interpreted as -1. |
103 | ConstantRange PosFilter = |
104 | BW == 1 ? getEmpty() : ConstantRange(One, SignedMin); |
105 | ConstantRange NegFilter(SignedMin, Zero); |
106 | return {intersectWith(CR: PosFilter), intersectWith(CR: NegFilter)}; |
107 | } |
108 | |
109 | ConstantRange ConstantRange::makeAllowedICmpRegion(CmpInst::Predicate Pred, |
110 | const ConstantRange &CR) { |
111 | if (CR.isEmptySet()) |
112 | return CR; |
113 | |
114 | uint32_t W = CR.getBitWidth(); |
115 | switch (Pred) { |
116 | default: |
117 | llvm_unreachable("Invalid ICmp predicate to makeAllowedICmpRegion()" ); |
118 | case CmpInst::ICMP_EQ: |
119 | return CR; |
120 | case CmpInst::ICMP_NE: |
121 | if (CR.isSingleElement()) |
122 | return ConstantRange(CR.getUpper(), CR.getLower()); |
123 | return getFull(BitWidth: W); |
124 | case CmpInst::ICMP_ULT: { |
125 | APInt UMax(CR.getUnsignedMax()); |
126 | if (UMax.isMinValue()) |
127 | return getEmpty(BitWidth: W); |
128 | return ConstantRange(APInt::getMinValue(numBits: W), std::move(UMax)); |
129 | } |
130 | case CmpInst::ICMP_SLT: { |
131 | APInt SMax(CR.getSignedMax()); |
132 | if (SMax.isMinSignedValue()) |
133 | return getEmpty(BitWidth: W); |
134 | return ConstantRange(APInt::getSignedMinValue(numBits: W), std::move(SMax)); |
135 | } |
136 | case CmpInst::ICMP_ULE: |
137 | return getNonEmpty(Lower: APInt::getMinValue(numBits: W), Upper: CR.getUnsignedMax() + 1); |
138 | case CmpInst::ICMP_SLE: |
139 | return getNonEmpty(Lower: APInt::getSignedMinValue(numBits: W), Upper: CR.getSignedMax() + 1); |
140 | case CmpInst::ICMP_UGT: { |
141 | APInt UMin(CR.getUnsignedMin()); |
142 | if (UMin.isMaxValue()) |
143 | return getEmpty(BitWidth: W); |
144 | return ConstantRange(std::move(UMin) + 1, APInt::getZero(numBits: W)); |
145 | } |
146 | case CmpInst::ICMP_SGT: { |
147 | APInt SMin(CR.getSignedMin()); |
148 | if (SMin.isMaxSignedValue()) |
149 | return getEmpty(BitWidth: W); |
150 | return ConstantRange(std::move(SMin) + 1, APInt::getSignedMinValue(numBits: W)); |
151 | } |
152 | case CmpInst::ICMP_UGE: |
153 | return getNonEmpty(Lower: CR.getUnsignedMin(), Upper: APInt::getZero(numBits: W)); |
154 | case CmpInst::ICMP_SGE: |
155 | return getNonEmpty(Lower: CR.getSignedMin(), Upper: APInt::getSignedMinValue(numBits: W)); |
156 | } |
157 | } |
158 | |
159 | ConstantRange ConstantRange::makeSatisfyingICmpRegion(CmpInst::Predicate Pred, |
160 | const ConstantRange &CR) { |
161 | // Follows from De-Morgan's laws: |
162 | // |
163 | // ~(~A union ~B) == A intersect B. |
164 | // |
165 | return makeAllowedICmpRegion(Pred: CmpInst::getInversePredicate(pred: Pred), CR) |
166 | .inverse(); |
167 | } |
168 | |
169 | ConstantRange ConstantRange::makeExactICmpRegion(CmpInst::Predicate Pred, |
170 | const APInt &C) { |
171 | // Computes the exact range that is equal to both the constant ranges returned |
172 | // by makeAllowedICmpRegion and makeSatisfyingICmpRegion. This is always true |
173 | // when RHS is a singleton such as an APInt. However for non-singleton RHS, |
174 | // for example ult [2,5) makeAllowedICmpRegion returns [0,4) but |
175 | // makeSatisfyICmpRegion returns [0,2). |
176 | // |
177 | return makeAllowedICmpRegion(Pred, CR: C); |
178 | } |
179 | |
180 | bool ConstantRange::areInsensitiveToSignednessOfICmpPredicate( |
181 | const ConstantRange &CR1, const ConstantRange &CR2) { |
182 | if (CR1.isEmptySet() || CR2.isEmptySet()) |
183 | return true; |
184 | |
185 | return (CR1.isAllNonNegative() && CR2.isAllNonNegative()) || |
186 | (CR1.isAllNegative() && CR2.isAllNegative()); |
187 | } |
188 | |
189 | bool ConstantRange::areInsensitiveToSignednessOfInvertedICmpPredicate( |
190 | const ConstantRange &CR1, const ConstantRange &CR2) { |
191 | if (CR1.isEmptySet() || CR2.isEmptySet()) |
192 | return true; |
193 | |
194 | return (CR1.isAllNonNegative() && CR2.isAllNegative()) || |
195 | (CR1.isAllNegative() && CR2.isAllNonNegative()); |
196 | } |
197 | |
198 | CmpInst::Predicate ConstantRange::getEquivalentPredWithFlippedSignedness( |
199 | CmpInst::Predicate Pred, const ConstantRange &CR1, |
200 | const ConstantRange &CR2) { |
201 | assert(CmpInst::isIntPredicate(Pred) && CmpInst::isRelational(Pred) && |
202 | "Only for relational integer predicates!" ); |
203 | |
204 | CmpInst::Predicate FlippedSignednessPred = |
205 | ICmpInst::getFlippedSignednessPredicate(Pred); |
206 | |
207 | if (areInsensitiveToSignednessOfICmpPredicate(CR1, CR2)) |
208 | return FlippedSignednessPred; |
209 | |
210 | if (areInsensitiveToSignednessOfInvertedICmpPredicate(CR1, CR2)) |
211 | return CmpInst::getInversePredicate(pred: FlippedSignednessPred); |
212 | |
213 | return CmpInst::Predicate::BAD_ICMP_PREDICATE; |
214 | } |
215 | |
216 | void ConstantRange::getEquivalentICmp(CmpInst::Predicate &Pred, |
217 | APInt &RHS, APInt &Offset) const { |
218 | Offset = APInt(getBitWidth(), 0); |
219 | if (isFullSet() || isEmptySet()) { |
220 | Pred = isEmptySet() ? CmpInst::ICMP_ULT : CmpInst::ICMP_UGE; |
221 | RHS = APInt(getBitWidth(), 0); |
222 | } else if (auto *OnlyElt = getSingleElement()) { |
223 | Pred = CmpInst::ICMP_EQ; |
224 | RHS = *OnlyElt; |
225 | } else if (auto *OnlyMissingElt = getSingleMissingElement()) { |
226 | Pred = CmpInst::ICMP_NE; |
227 | RHS = *OnlyMissingElt; |
228 | } else if (getLower().isMinSignedValue() || getLower().isMinValue()) { |
229 | Pred = |
230 | getLower().isMinSignedValue() ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT; |
231 | RHS = getUpper(); |
232 | } else if (getUpper().isMinSignedValue() || getUpper().isMinValue()) { |
233 | Pred = |
234 | getUpper().isMinSignedValue() ? CmpInst::ICMP_SGE : CmpInst::ICMP_UGE; |
235 | RHS = getLower(); |
236 | } else { |
237 | Pred = CmpInst::ICMP_ULT; |
238 | RHS = getUpper() - getLower(); |
239 | Offset = -getLower(); |
240 | } |
241 | |
242 | assert(ConstantRange::makeExactICmpRegion(Pred, RHS) == add(Offset) && |
243 | "Bad result!" ); |
244 | } |
245 | |
246 | bool ConstantRange::getEquivalentICmp(CmpInst::Predicate &Pred, |
247 | APInt &RHS) const { |
248 | APInt Offset; |
249 | getEquivalentICmp(Pred, RHS, Offset); |
250 | return Offset.isZero(); |
251 | } |
252 | |
253 | bool ConstantRange::icmp(CmpInst::Predicate Pred, |
254 | const ConstantRange &Other) const { |
255 | if (isEmptySet() || Other.isEmptySet()) |
256 | return true; |
257 | |
258 | switch (Pred) { |
259 | case CmpInst::ICMP_EQ: |
260 | if (const APInt *L = getSingleElement()) |
261 | if (const APInt *R = Other.getSingleElement()) |
262 | return *L == *R; |
263 | return false; |
264 | case CmpInst::ICMP_NE: |
265 | return inverse().contains(CR: Other); |
266 | case CmpInst::ICMP_ULT: |
267 | return getUnsignedMax().ult(RHS: Other.getUnsignedMin()); |
268 | case CmpInst::ICMP_ULE: |
269 | return getUnsignedMax().ule(RHS: Other.getUnsignedMin()); |
270 | case CmpInst::ICMP_UGT: |
271 | return getUnsignedMin().ugt(RHS: Other.getUnsignedMax()); |
272 | case CmpInst::ICMP_UGE: |
273 | return getUnsignedMin().uge(RHS: Other.getUnsignedMax()); |
274 | case CmpInst::ICMP_SLT: |
275 | return getSignedMax().slt(RHS: Other.getSignedMin()); |
276 | case CmpInst::ICMP_SLE: |
277 | return getSignedMax().sle(RHS: Other.getSignedMin()); |
278 | case CmpInst::ICMP_SGT: |
279 | return getSignedMin().sgt(RHS: Other.getSignedMax()); |
280 | case CmpInst::ICMP_SGE: |
281 | return getSignedMin().sge(RHS: Other.getSignedMax()); |
282 | default: |
283 | llvm_unreachable("Invalid ICmp predicate" ); |
284 | } |
285 | } |
286 | |
287 | /// Exact mul nuw region for single element RHS. |
288 | static ConstantRange makeExactMulNUWRegion(const APInt &V) { |
289 | unsigned BitWidth = V.getBitWidth(); |
290 | if (V == 0) |
291 | return ConstantRange::getFull(BitWidth: V.getBitWidth()); |
292 | |
293 | return ConstantRange::getNonEmpty( |
294 | Lower: APIntOps::RoundingUDiv(A: APInt::getMinValue(numBits: BitWidth), B: V, |
295 | RM: APInt::Rounding::UP), |
296 | Upper: APIntOps::RoundingUDiv(A: APInt::getMaxValue(numBits: BitWidth), B: V, |
297 | RM: APInt::Rounding::DOWN) + 1); |
298 | } |
299 | |
300 | /// Exact mul nsw region for single element RHS. |
301 | static ConstantRange makeExactMulNSWRegion(const APInt &V) { |
302 | // Handle 0 and -1 separately to avoid division by zero or overflow. |
303 | unsigned BitWidth = V.getBitWidth(); |
304 | if (V == 0) |
305 | return ConstantRange::getFull(BitWidth); |
306 | |
307 | APInt MinValue = APInt::getSignedMinValue(numBits: BitWidth); |
308 | APInt MaxValue = APInt::getSignedMaxValue(numBits: BitWidth); |
309 | // e.g. Returning [-127, 127], represented as [-127, -128). |
310 | if (V.isAllOnes()) |
311 | return ConstantRange(-MaxValue, MinValue); |
312 | |
313 | APInt Lower, Upper; |
314 | if (V.isNegative()) { |
315 | Lower = APIntOps::RoundingSDiv(A: MaxValue, B: V, RM: APInt::Rounding::UP); |
316 | Upper = APIntOps::RoundingSDiv(A: MinValue, B: V, RM: APInt::Rounding::DOWN); |
317 | } else { |
318 | Lower = APIntOps::RoundingSDiv(A: MinValue, B: V, RM: APInt::Rounding::UP); |
319 | Upper = APIntOps::RoundingSDiv(A: MaxValue, B: V, RM: APInt::Rounding::DOWN); |
320 | } |
321 | return ConstantRange::getNonEmpty(Lower, Upper: Upper + 1); |
322 | } |
323 | |
324 | ConstantRange |
325 | ConstantRange::makeGuaranteedNoWrapRegion(Instruction::BinaryOps BinOp, |
326 | const ConstantRange &Other, |
327 | unsigned NoWrapKind) { |
328 | using OBO = OverflowingBinaryOperator; |
329 | |
330 | assert(Instruction::isBinaryOp(BinOp) && "Binary operators only!" ); |
331 | |
332 | assert((NoWrapKind == OBO::NoSignedWrap || |
333 | NoWrapKind == OBO::NoUnsignedWrap) && |
334 | "NoWrapKind invalid!" ); |
335 | |
336 | bool Unsigned = NoWrapKind == OBO::NoUnsignedWrap; |
337 | unsigned BitWidth = Other.getBitWidth(); |
338 | |
339 | switch (BinOp) { |
340 | default: |
341 | llvm_unreachable("Unsupported binary op" ); |
342 | |
343 | case Instruction::Add: { |
344 | if (Unsigned) |
345 | return getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), Upper: -Other.getUnsignedMax()); |
346 | |
347 | APInt SignedMinVal = APInt::getSignedMinValue(numBits: BitWidth); |
348 | APInt SMin = Other.getSignedMin(), SMax = Other.getSignedMax(); |
349 | return getNonEmpty( |
350 | Lower: SMin.isNegative() ? SignedMinVal - SMin : SignedMinVal, |
351 | Upper: SMax.isStrictlyPositive() ? SignedMinVal - SMax : SignedMinVal); |
352 | } |
353 | |
354 | case Instruction::Sub: { |
355 | if (Unsigned) |
356 | return getNonEmpty(Lower: Other.getUnsignedMax(), Upper: APInt::getMinValue(numBits: BitWidth)); |
357 | |
358 | APInt SignedMinVal = APInt::getSignedMinValue(numBits: BitWidth); |
359 | APInt SMin = Other.getSignedMin(), SMax = Other.getSignedMax(); |
360 | return getNonEmpty( |
361 | Lower: SMax.isStrictlyPositive() ? SignedMinVal + SMax : SignedMinVal, |
362 | Upper: SMin.isNegative() ? SignedMinVal + SMin : SignedMinVal); |
363 | } |
364 | |
365 | case Instruction::Mul: |
366 | if (Unsigned) |
367 | return makeExactMulNUWRegion(V: Other.getUnsignedMax()); |
368 | |
369 | // Avoid one makeExactMulNSWRegion() call for the common case of constants. |
370 | if (const APInt *C = Other.getSingleElement()) |
371 | return makeExactMulNSWRegion(V: *C); |
372 | |
373 | return makeExactMulNSWRegion(V: Other.getSignedMin()) |
374 | .intersectWith(CR: makeExactMulNSWRegion(V: Other.getSignedMax())); |
375 | |
376 | case Instruction::Shl: { |
377 | // For given range of shift amounts, if we ignore all illegal shift amounts |
378 | // (that always produce poison), what shift amount range is left? |
379 | ConstantRange ShAmt = Other.intersectWith( |
380 | CR: ConstantRange(APInt(BitWidth, 0), APInt(BitWidth, (BitWidth - 1) + 1))); |
381 | if (ShAmt.isEmptySet()) { |
382 | // If the entire range of shift amounts is already poison-producing, |
383 | // then we can freely add more poison-producing flags ontop of that. |
384 | return getFull(BitWidth); |
385 | } |
386 | // There are some legal shift amounts, we can compute conservatively-correct |
387 | // range of no-wrap inputs. Note that by now we have clamped the ShAmtUMax |
388 | // to be at most bitwidth-1, which results in most conservative range. |
389 | APInt ShAmtUMax = ShAmt.getUnsignedMax(); |
390 | if (Unsigned) |
391 | return getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), |
392 | Upper: APInt::getMaxValue(numBits: BitWidth).lshr(ShiftAmt: ShAmtUMax) + 1); |
393 | return getNonEmpty(Lower: APInt::getSignedMinValue(numBits: BitWidth).ashr(ShiftAmt: ShAmtUMax), |
394 | Upper: APInt::getSignedMaxValue(numBits: BitWidth).ashr(ShiftAmt: ShAmtUMax) + 1); |
395 | } |
396 | } |
397 | } |
398 | |
399 | ConstantRange ConstantRange::makeExactNoWrapRegion(Instruction::BinaryOps BinOp, |
400 | const APInt &Other, |
401 | unsigned NoWrapKind) { |
402 | // makeGuaranteedNoWrapRegion() is exact for single-element ranges, as |
403 | // "for all" and "for any" coincide in this case. |
404 | return makeGuaranteedNoWrapRegion(BinOp, Other: ConstantRange(Other), NoWrapKind); |
405 | } |
406 | |
407 | ConstantRange ConstantRange::makeMaskNotEqualRange(const APInt &Mask, |
408 | const APInt &C) { |
409 | unsigned BitWidth = Mask.getBitWidth(); |
410 | |
411 | if ((Mask & C) != C) |
412 | return getFull(BitWidth); |
413 | |
414 | if (Mask.isZero()) |
415 | return getEmpty(BitWidth); |
416 | |
417 | // If (Val & Mask) != C, constrained to the non-equality being |
418 | // satisfiable, then the value must be larger than the lowest set bit of |
419 | // Mask, offset by constant C. |
420 | return ConstantRange::getNonEmpty( |
421 | Lower: APInt::getOneBitSet(numBits: BitWidth, BitNo: Mask.countr_zero()) + C, Upper: C); |
422 | } |
423 | |
424 | bool ConstantRange::isFullSet() const { |
425 | return Lower == Upper && Lower.isMaxValue(); |
426 | } |
427 | |
428 | bool ConstantRange::isEmptySet() const { |
429 | return Lower == Upper && Lower.isMinValue(); |
430 | } |
431 | |
432 | bool ConstantRange::isWrappedSet() const { |
433 | return Lower.ugt(RHS: Upper) && !Upper.isZero(); |
434 | } |
435 | |
436 | bool ConstantRange::isUpperWrapped() const { |
437 | return Lower.ugt(RHS: Upper); |
438 | } |
439 | |
440 | bool ConstantRange::isSignWrappedSet() const { |
441 | return Lower.sgt(RHS: Upper) && !Upper.isMinSignedValue(); |
442 | } |
443 | |
444 | bool ConstantRange::isUpperSignWrapped() const { |
445 | return Lower.sgt(RHS: Upper); |
446 | } |
447 | |
448 | bool |
449 | ConstantRange::isSizeStrictlySmallerThan(const ConstantRange &Other) const { |
450 | assert(getBitWidth() == Other.getBitWidth()); |
451 | if (isFullSet()) |
452 | return false; |
453 | if (Other.isFullSet()) |
454 | return true; |
455 | return (Upper - Lower).ult(RHS: Other.Upper - Other.Lower); |
456 | } |
457 | |
458 | bool |
459 | ConstantRange::isSizeLargerThan(uint64_t MaxSize) const { |
460 | // If this a full set, we need special handling to avoid needing an extra bit |
461 | // to represent the size. |
462 | if (isFullSet()) |
463 | return MaxSize == 0 || APInt::getMaxValue(numBits: getBitWidth()).ugt(RHS: MaxSize - 1); |
464 | |
465 | return (Upper - Lower).ugt(RHS: MaxSize); |
466 | } |
467 | |
468 | bool ConstantRange::isAllNegative() const { |
469 | // Empty set is all negative, full set is not. |
470 | if (isEmptySet()) |
471 | return true; |
472 | if (isFullSet()) |
473 | return false; |
474 | |
475 | return !isUpperSignWrapped() && !Upper.isStrictlyPositive(); |
476 | } |
477 | |
478 | bool ConstantRange::isAllNonNegative() const { |
479 | // Empty and full set are automatically treated correctly. |
480 | return !isSignWrappedSet() && Lower.isNonNegative(); |
481 | } |
482 | |
483 | bool ConstantRange::isAllPositive() const { |
484 | // Empty set is all positive, full set is not. |
485 | if (isEmptySet()) |
486 | return true; |
487 | if (isFullSet()) |
488 | return false; |
489 | |
490 | return !isSignWrappedSet() && Lower.isStrictlyPositive(); |
491 | } |
492 | |
493 | APInt ConstantRange::getUnsignedMax() const { |
494 | if (isFullSet() || isUpperWrapped()) |
495 | return APInt::getMaxValue(numBits: getBitWidth()); |
496 | return getUpper() - 1; |
497 | } |
498 | |
499 | APInt ConstantRange::getUnsignedMin() const { |
500 | if (isFullSet() || isWrappedSet()) |
501 | return APInt::getMinValue(numBits: getBitWidth()); |
502 | return getLower(); |
503 | } |
504 | |
505 | APInt ConstantRange::getSignedMax() const { |
506 | if (isFullSet() || isUpperSignWrapped()) |
507 | return APInt::getSignedMaxValue(numBits: getBitWidth()); |
508 | return getUpper() - 1; |
509 | } |
510 | |
511 | APInt ConstantRange::getSignedMin() const { |
512 | if (isFullSet() || isSignWrappedSet()) |
513 | return APInt::getSignedMinValue(numBits: getBitWidth()); |
514 | return getLower(); |
515 | } |
516 | |
517 | bool ConstantRange::contains(const APInt &V) const { |
518 | if (Lower == Upper) |
519 | return isFullSet(); |
520 | |
521 | if (!isUpperWrapped()) |
522 | return Lower.ule(RHS: V) && V.ult(RHS: Upper); |
523 | return Lower.ule(RHS: V) || V.ult(RHS: Upper); |
524 | } |
525 | |
526 | bool ConstantRange::contains(const ConstantRange &Other) const { |
527 | if (isFullSet() || Other.isEmptySet()) return true; |
528 | if (isEmptySet() || Other.isFullSet()) return false; |
529 | |
530 | if (!isUpperWrapped()) { |
531 | if (Other.isUpperWrapped()) |
532 | return false; |
533 | |
534 | return Lower.ule(RHS: Other.getLower()) && Other.getUpper().ule(RHS: Upper); |
535 | } |
536 | |
537 | if (!Other.isUpperWrapped()) |
538 | return Other.getUpper().ule(RHS: Upper) || |
539 | Lower.ule(RHS: Other.getLower()); |
540 | |
541 | return Other.getUpper().ule(RHS: Upper) && Lower.ule(RHS: Other.getLower()); |
542 | } |
543 | |
544 | unsigned ConstantRange::getActiveBits() const { |
545 | if (isEmptySet()) |
546 | return 0; |
547 | |
548 | return getUnsignedMax().getActiveBits(); |
549 | } |
550 | |
551 | unsigned ConstantRange::getMinSignedBits() const { |
552 | if (isEmptySet()) |
553 | return 0; |
554 | |
555 | return std::max(a: getSignedMin().getSignificantBits(), |
556 | b: getSignedMax().getSignificantBits()); |
557 | } |
558 | |
559 | ConstantRange ConstantRange::subtract(const APInt &Val) const { |
560 | assert(Val.getBitWidth() == getBitWidth() && "Wrong bit width" ); |
561 | // If the set is empty or full, don't modify the endpoints. |
562 | if (Lower == Upper) |
563 | return *this; |
564 | return ConstantRange(Lower - Val, Upper - Val); |
565 | } |
566 | |
567 | ConstantRange ConstantRange::difference(const ConstantRange &CR) const { |
568 | return intersectWith(CR: CR.inverse()); |
569 | } |
570 | |
571 | static ConstantRange getPreferredRange( |
572 | const ConstantRange &CR1, const ConstantRange &CR2, |
573 | ConstantRange::PreferredRangeType Type) { |
574 | if (Type == ConstantRange::Unsigned) { |
575 | if (!CR1.isWrappedSet() && CR2.isWrappedSet()) |
576 | return CR1; |
577 | if (CR1.isWrappedSet() && !CR2.isWrappedSet()) |
578 | return CR2; |
579 | } else if (Type == ConstantRange::Signed) { |
580 | if (!CR1.isSignWrappedSet() && CR2.isSignWrappedSet()) |
581 | return CR1; |
582 | if (CR1.isSignWrappedSet() && !CR2.isSignWrappedSet()) |
583 | return CR2; |
584 | } |
585 | |
586 | if (CR1.isSizeStrictlySmallerThan(Other: CR2)) |
587 | return CR1; |
588 | return CR2; |
589 | } |
590 | |
591 | ConstantRange ConstantRange::intersectWith(const ConstantRange &CR, |
592 | PreferredRangeType Type) const { |
593 | assert(getBitWidth() == CR.getBitWidth() && |
594 | "ConstantRange types don't agree!" ); |
595 | |
596 | // Handle common cases. |
597 | if ( isEmptySet() || CR.isFullSet()) return *this; |
598 | if (CR.isEmptySet() || isFullSet()) return CR; |
599 | |
600 | if (!isUpperWrapped() && CR.isUpperWrapped()) |
601 | return CR.intersectWith(CR: *this, Type); |
602 | |
603 | if (!isUpperWrapped() && !CR.isUpperWrapped()) { |
604 | if (Lower.ult(RHS: CR.Lower)) { |
605 | // L---U : this |
606 | // L---U : CR |
607 | if (Upper.ule(RHS: CR.Lower)) |
608 | return getEmpty(); |
609 | |
610 | // L---U : this |
611 | // L---U : CR |
612 | if (Upper.ult(RHS: CR.Upper)) |
613 | return ConstantRange(CR.Lower, Upper); |
614 | |
615 | // L-------U : this |
616 | // L---U : CR |
617 | return CR; |
618 | } |
619 | // L---U : this |
620 | // L-------U : CR |
621 | if (Upper.ult(RHS: CR.Upper)) |
622 | return *this; |
623 | |
624 | // L-----U : this |
625 | // L-----U : CR |
626 | if (Lower.ult(RHS: CR.Upper)) |
627 | return ConstantRange(Lower, CR.Upper); |
628 | |
629 | // L---U : this |
630 | // L---U : CR |
631 | return getEmpty(); |
632 | } |
633 | |
634 | if (isUpperWrapped() && !CR.isUpperWrapped()) { |
635 | if (CR.Lower.ult(RHS: Upper)) { |
636 | // ------U L--- : this |
637 | // L--U : CR |
638 | if (CR.Upper.ult(RHS: Upper)) |
639 | return CR; |
640 | |
641 | // ------U L--- : this |
642 | // L------U : CR |
643 | if (CR.Upper.ule(RHS: Lower)) |
644 | return ConstantRange(CR.Lower, Upper); |
645 | |
646 | // ------U L--- : this |
647 | // L----------U : CR |
648 | return getPreferredRange(CR1: *this, CR2: CR, Type); |
649 | } |
650 | if (CR.Lower.ult(RHS: Lower)) { |
651 | // --U L---- : this |
652 | // L--U : CR |
653 | if (CR.Upper.ule(RHS: Lower)) |
654 | return getEmpty(); |
655 | |
656 | // --U L---- : this |
657 | // L------U : CR |
658 | return ConstantRange(Lower, CR.Upper); |
659 | } |
660 | |
661 | // --U L------ : this |
662 | // L--U : CR |
663 | return CR; |
664 | } |
665 | |
666 | if (CR.Upper.ult(RHS: Upper)) { |
667 | // ------U L-- : this |
668 | // --U L------ : CR |
669 | if (CR.Lower.ult(RHS: Upper)) |
670 | return getPreferredRange(CR1: *this, CR2: CR, Type); |
671 | |
672 | // ----U L-- : this |
673 | // --U L---- : CR |
674 | if (CR.Lower.ult(RHS: Lower)) |
675 | return ConstantRange(Lower, CR.Upper); |
676 | |
677 | // ----U L---- : this |
678 | // --U L-- : CR |
679 | return CR; |
680 | } |
681 | if (CR.Upper.ule(RHS: Lower)) { |
682 | // --U L-- : this |
683 | // ----U L---- : CR |
684 | if (CR.Lower.ult(RHS: Lower)) |
685 | return *this; |
686 | |
687 | // --U L---- : this |
688 | // ----U L-- : CR |
689 | return ConstantRange(CR.Lower, Upper); |
690 | } |
691 | |
692 | // --U L------ : this |
693 | // ------U L-- : CR |
694 | return getPreferredRange(CR1: *this, CR2: CR, Type); |
695 | } |
696 | |
697 | ConstantRange ConstantRange::unionWith(const ConstantRange &CR, |
698 | PreferredRangeType Type) const { |
699 | assert(getBitWidth() == CR.getBitWidth() && |
700 | "ConstantRange types don't agree!" ); |
701 | |
702 | if ( isFullSet() || CR.isEmptySet()) return *this; |
703 | if (CR.isFullSet() || isEmptySet()) return CR; |
704 | |
705 | if (!isUpperWrapped() && CR.isUpperWrapped()) |
706 | return CR.unionWith(CR: *this, Type); |
707 | |
708 | if (!isUpperWrapped() && !CR.isUpperWrapped()) { |
709 | // L---U and L---U : this |
710 | // L---U L---U : CR |
711 | // result in one of |
712 | // L---------U |
713 | // -----U L----- |
714 | if (CR.Upper.ult(RHS: Lower) || Upper.ult(RHS: CR.Lower)) |
715 | return getPreferredRange( |
716 | CR1: ConstantRange(Lower, CR.Upper), CR2: ConstantRange(CR.Lower, Upper), Type); |
717 | |
718 | APInt L = CR.Lower.ult(RHS: Lower) ? CR.Lower : Lower; |
719 | APInt U = (CR.Upper - 1).ugt(RHS: Upper - 1) ? CR.Upper : Upper; |
720 | |
721 | if (L.isZero() && U.isZero()) |
722 | return getFull(); |
723 | |
724 | return ConstantRange(std::move(L), std::move(U)); |
725 | } |
726 | |
727 | if (!CR.isUpperWrapped()) { |
728 | // ------U L----- and ------U L----- : this |
729 | // L--U L--U : CR |
730 | if (CR.Upper.ule(RHS: Upper) || CR.Lower.uge(RHS: Lower)) |
731 | return *this; |
732 | |
733 | // ------U L----- : this |
734 | // L---------U : CR |
735 | if (CR.Lower.ule(RHS: Upper) && Lower.ule(RHS: CR.Upper)) |
736 | return getFull(); |
737 | |
738 | // ----U L---- : this |
739 | // L---U : CR |
740 | // results in one of |
741 | // ----------U L---- |
742 | // ----U L---------- |
743 | if (Upper.ult(RHS: CR.Lower) && CR.Upper.ult(RHS: Lower)) |
744 | return getPreferredRange( |
745 | CR1: ConstantRange(Lower, CR.Upper), CR2: ConstantRange(CR.Lower, Upper), Type); |
746 | |
747 | // ----U L----- : this |
748 | // L----U : CR |
749 | if (Upper.ult(RHS: CR.Lower) && Lower.ule(RHS: CR.Upper)) |
750 | return ConstantRange(CR.Lower, Upper); |
751 | |
752 | // ------U L---- : this |
753 | // L-----U : CR |
754 | assert(CR.Lower.ule(Upper) && CR.Upper.ult(Lower) && |
755 | "ConstantRange::unionWith missed a case with one range wrapped" ); |
756 | return ConstantRange(Lower, CR.Upper); |
757 | } |
758 | |
759 | // ------U L---- and ------U L---- : this |
760 | // -U L----------- and ------------U L : CR |
761 | if (CR.Lower.ule(RHS: Upper) || Lower.ule(RHS: CR.Upper)) |
762 | return getFull(); |
763 | |
764 | APInt L = CR.Lower.ult(RHS: Lower) ? CR.Lower : Lower; |
765 | APInt U = CR.Upper.ugt(RHS: Upper) ? CR.Upper : Upper; |
766 | |
767 | return ConstantRange(std::move(L), std::move(U)); |
768 | } |
769 | |
770 | std::optional<ConstantRange> |
771 | ConstantRange::exactIntersectWith(const ConstantRange &CR) const { |
772 | // TODO: This can be implemented more efficiently. |
773 | ConstantRange Result = intersectWith(CR); |
774 | if (Result == inverse().unionWith(CR: CR.inverse()).inverse()) |
775 | return Result; |
776 | return std::nullopt; |
777 | } |
778 | |
779 | std::optional<ConstantRange> |
780 | ConstantRange::exactUnionWith(const ConstantRange &CR) const { |
781 | // TODO: This can be implemented more efficiently. |
782 | ConstantRange Result = unionWith(CR); |
783 | if (Result == inverse().intersectWith(CR: CR.inverse()).inverse()) |
784 | return Result; |
785 | return std::nullopt; |
786 | } |
787 | |
788 | ConstantRange ConstantRange::castOp(Instruction::CastOps CastOp, |
789 | uint32_t ResultBitWidth) const { |
790 | switch (CastOp) { |
791 | default: |
792 | llvm_unreachable("unsupported cast type" ); |
793 | case Instruction::Trunc: |
794 | return truncate(BitWidth: ResultBitWidth); |
795 | case Instruction::SExt: |
796 | return signExtend(BitWidth: ResultBitWidth); |
797 | case Instruction::ZExt: |
798 | return zeroExtend(BitWidth: ResultBitWidth); |
799 | case Instruction::BitCast: |
800 | return *this; |
801 | case Instruction::FPToUI: |
802 | case Instruction::FPToSI: |
803 | if (getBitWidth() == ResultBitWidth) |
804 | return *this; |
805 | else |
806 | return getFull(BitWidth: ResultBitWidth); |
807 | case Instruction::UIToFP: { |
808 | // TODO: use input range if available |
809 | auto BW = getBitWidth(); |
810 | APInt Min = APInt::getMinValue(numBits: BW); |
811 | APInt Max = APInt::getMaxValue(numBits: BW); |
812 | if (ResultBitWidth > BW) { |
813 | Min = Min.zext(width: ResultBitWidth); |
814 | Max = Max.zext(width: ResultBitWidth); |
815 | } |
816 | return getNonEmpty(Lower: std::move(Min), Upper: std::move(Max) + 1); |
817 | } |
818 | case Instruction::SIToFP: { |
819 | // TODO: use input range if available |
820 | auto BW = getBitWidth(); |
821 | APInt SMin = APInt::getSignedMinValue(numBits: BW); |
822 | APInt SMax = APInt::getSignedMaxValue(numBits: BW); |
823 | if (ResultBitWidth > BW) { |
824 | SMin = SMin.sext(width: ResultBitWidth); |
825 | SMax = SMax.sext(width: ResultBitWidth); |
826 | } |
827 | return getNonEmpty(Lower: std::move(SMin), Upper: std::move(SMax) + 1); |
828 | } |
829 | case Instruction::FPTrunc: |
830 | case Instruction::FPExt: |
831 | case Instruction::IntToPtr: |
832 | case Instruction::PtrToInt: |
833 | case Instruction::AddrSpaceCast: |
834 | // Conservatively return getFull set. |
835 | return getFull(BitWidth: ResultBitWidth); |
836 | }; |
837 | } |
838 | |
839 | ConstantRange ConstantRange::zeroExtend(uint32_t DstTySize) const { |
840 | if (isEmptySet()) return getEmpty(BitWidth: DstTySize); |
841 | |
842 | unsigned SrcTySize = getBitWidth(); |
843 | assert(SrcTySize < DstTySize && "Not a value extension" ); |
844 | if (isFullSet() || isUpperWrapped()) { |
845 | // Change into [0, 1 << src bit width) |
846 | APInt LowerExt(DstTySize, 0); |
847 | if (!Upper) // special case: [X, 0) -- not really wrapping around |
848 | LowerExt = Lower.zext(width: DstTySize); |
849 | return ConstantRange(std::move(LowerExt), |
850 | APInt::getOneBitSet(numBits: DstTySize, BitNo: SrcTySize)); |
851 | } |
852 | |
853 | return ConstantRange(Lower.zext(width: DstTySize), Upper.zext(width: DstTySize)); |
854 | } |
855 | |
856 | ConstantRange ConstantRange::signExtend(uint32_t DstTySize) const { |
857 | if (isEmptySet()) return getEmpty(BitWidth: DstTySize); |
858 | |
859 | unsigned SrcTySize = getBitWidth(); |
860 | assert(SrcTySize < DstTySize && "Not a value extension" ); |
861 | |
862 | // special case: [X, INT_MIN) -- not really wrapping around |
863 | if (Upper.isMinSignedValue()) |
864 | return ConstantRange(Lower.sext(width: DstTySize), Upper.zext(width: DstTySize)); |
865 | |
866 | if (isFullSet() || isSignWrappedSet()) { |
867 | return ConstantRange(APInt::getHighBitsSet(numBits: DstTySize,hiBitsSet: DstTySize-SrcTySize+1), |
868 | APInt::getLowBitsSet(numBits: DstTySize, loBitsSet: SrcTySize-1) + 1); |
869 | } |
870 | |
871 | return ConstantRange(Lower.sext(width: DstTySize), Upper.sext(width: DstTySize)); |
872 | } |
873 | |
874 | ConstantRange ConstantRange::truncate(uint32_t DstTySize) const { |
875 | assert(getBitWidth() > DstTySize && "Not a value truncation" ); |
876 | if (isEmptySet()) |
877 | return getEmpty(BitWidth: DstTySize); |
878 | if (isFullSet()) |
879 | return getFull(BitWidth: DstTySize); |
880 | |
881 | APInt LowerDiv(Lower), UpperDiv(Upper); |
882 | ConstantRange Union(DstTySize, /*isFullSet=*/false); |
883 | |
884 | // Analyze wrapped sets in their two parts: [0, Upper) \/ [Lower, MaxValue] |
885 | // We use the non-wrapped set code to analyze the [Lower, MaxValue) part, and |
886 | // then we do the union with [MaxValue, Upper) |
887 | if (isUpperWrapped()) { |
888 | // If Upper is greater than or equal to MaxValue(DstTy), it covers the whole |
889 | // truncated range. |
890 | if (Upper.getActiveBits() > DstTySize || Upper.countr_one() == DstTySize) |
891 | return getFull(BitWidth: DstTySize); |
892 | |
893 | Union = ConstantRange(APInt::getMaxValue(numBits: DstTySize),Upper.trunc(width: DstTySize)); |
894 | UpperDiv.setAllBits(); |
895 | |
896 | // Union covers the MaxValue case, so return if the remaining range is just |
897 | // MaxValue(DstTy). |
898 | if (LowerDiv == UpperDiv) |
899 | return Union; |
900 | } |
901 | |
902 | // Chop off the most significant bits that are past the destination bitwidth. |
903 | if (LowerDiv.getActiveBits() > DstTySize) { |
904 | // Mask to just the signficant bits and subtract from LowerDiv/UpperDiv. |
905 | APInt Adjust = LowerDiv & APInt::getBitsSetFrom(numBits: getBitWidth(), loBit: DstTySize); |
906 | LowerDiv -= Adjust; |
907 | UpperDiv -= Adjust; |
908 | } |
909 | |
910 | unsigned UpperDivWidth = UpperDiv.getActiveBits(); |
911 | if (UpperDivWidth <= DstTySize) |
912 | return ConstantRange(LowerDiv.trunc(width: DstTySize), |
913 | UpperDiv.trunc(width: DstTySize)).unionWith(CR: Union); |
914 | |
915 | // The truncated value wraps around. Check if we can do better than fullset. |
916 | if (UpperDivWidth == DstTySize + 1) { |
917 | // Clear the MSB so that UpperDiv wraps around. |
918 | UpperDiv.clearBit(BitPosition: DstTySize); |
919 | if (UpperDiv.ult(RHS: LowerDiv)) |
920 | return ConstantRange(LowerDiv.trunc(width: DstTySize), |
921 | UpperDiv.trunc(width: DstTySize)).unionWith(CR: Union); |
922 | } |
923 | |
924 | return getFull(BitWidth: DstTySize); |
925 | } |
926 | |
927 | ConstantRange ConstantRange::zextOrTrunc(uint32_t DstTySize) const { |
928 | unsigned SrcTySize = getBitWidth(); |
929 | if (SrcTySize > DstTySize) |
930 | return truncate(DstTySize); |
931 | if (SrcTySize < DstTySize) |
932 | return zeroExtend(DstTySize); |
933 | return *this; |
934 | } |
935 | |
936 | ConstantRange ConstantRange::sextOrTrunc(uint32_t DstTySize) const { |
937 | unsigned SrcTySize = getBitWidth(); |
938 | if (SrcTySize > DstTySize) |
939 | return truncate(DstTySize); |
940 | if (SrcTySize < DstTySize) |
941 | return signExtend(DstTySize); |
942 | return *this; |
943 | } |
944 | |
945 | ConstantRange ConstantRange::binaryOp(Instruction::BinaryOps BinOp, |
946 | const ConstantRange &Other) const { |
947 | assert(Instruction::isBinaryOp(BinOp) && "Binary operators only!" ); |
948 | |
949 | switch (BinOp) { |
950 | case Instruction::Add: |
951 | return add(Other); |
952 | case Instruction::Sub: |
953 | return sub(Other); |
954 | case Instruction::Mul: |
955 | return multiply(Other); |
956 | case Instruction::UDiv: |
957 | return udiv(Other); |
958 | case Instruction::SDiv: |
959 | return sdiv(Other); |
960 | case Instruction::URem: |
961 | return urem(Other); |
962 | case Instruction::SRem: |
963 | return srem(Other); |
964 | case Instruction::Shl: |
965 | return shl(Other); |
966 | case Instruction::LShr: |
967 | return lshr(Other); |
968 | case Instruction::AShr: |
969 | return ashr(Other); |
970 | case Instruction::And: |
971 | return binaryAnd(Other); |
972 | case Instruction::Or: |
973 | return binaryOr(Other); |
974 | case Instruction::Xor: |
975 | return binaryXor(Other); |
976 | // Note: floating point operations applied to abstract ranges are just |
977 | // ideal integer operations with a lossy representation |
978 | case Instruction::FAdd: |
979 | return add(Other); |
980 | case Instruction::FSub: |
981 | return sub(Other); |
982 | case Instruction::FMul: |
983 | return multiply(Other); |
984 | default: |
985 | // Conservatively return getFull set. |
986 | return getFull(); |
987 | } |
988 | } |
989 | |
990 | ConstantRange ConstantRange::overflowingBinaryOp(Instruction::BinaryOps BinOp, |
991 | const ConstantRange &Other, |
992 | unsigned NoWrapKind) const { |
993 | assert(Instruction::isBinaryOp(BinOp) && "Binary operators only!" ); |
994 | |
995 | switch (BinOp) { |
996 | case Instruction::Add: |
997 | return addWithNoWrap(Other, NoWrapKind); |
998 | case Instruction::Sub: |
999 | return subWithNoWrap(Other, NoWrapKind); |
1000 | case Instruction::Mul: |
1001 | return multiplyWithNoWrap(Other, NoWrapKind); |
1002 | case Instruction::Shl: |
1003 | return shlWithNoWrap(Other, NoWrapKind); |
1004 | default: |
1005 | // Don't know about this Overflowing Binary Operation. |
1006 | // Conservatively fallback to plain binop handling. |
1007 | return binaryOp(BinOp, Other); |
1008 | } |
1009 | } |
1010 | |
1011 | bool ConstantRange::isIntrinsicSupported(Intrinsic::ID IntrinsicID) { |
1012 | switch (IntrinsicID) { |
1013 | case Intrinsic::uadd_sat: |
1014 | case Intrinsic::usub_sat: |
1015 | case Intrinsic::sadd_sat: |
1016 | case Intrinsic::ssub_sat: |
1017 | case Intrinsic::umin: |
1018 | case Intrinsic::umax: |
1019 | case Intrinsic::smin: |
1020 | case Intrinsic::smax: |
1021 | case Intrinsic::abs: |
1022 | case Intrinsic::ctlz: |
1023 | case Intrinsic::cttz: |
1024 | case Intrinsic::ctpop: |
1025 | return true; |
1026 | default: |
1027 | return false; |
1028 | } |
1029 | } |
1030 | |
1031 | ConstantRange ConstantRange::intrinsic(Intrinsic::ID IntrinsicID, |
1032 | ArrayRef<ConstantRange> Ops) { |
1033 | switch (IntrinsicID) { |
1034 | case Intrinsic::uadd_sat: |
1035 | return Ops[0].uadd_sat(Other: Ops[1]); |
1036 | case Intrinsic::usub_sat: |
1037 | return Ops[0].usub_sat(Other: Ops[1]); |
1038 | case Intrinsic::sadd_sat: |
1039 | return Ops[0].sadd_sat(Other: Ops[1]); |
1040 | case Intrinsic::ssub_sat: |
1041 | return Ops[0].ssub_sat(Other: Ops[1]); |
1042 | case Intrinsic::umin: |
1043 | return Ops[0].umin(Other: Ops[1]); |
1044 | case Intrinsic::umax: |
1045 | return Ops[0].umax(Other: Ops[1]); |
1046 | case Intrinsic::smin: |
1047 | return Ops[0].smin(Other: Ops[1]); |
1048 | case Intrinsic::smax: |
1049 | return Ops[0].smax(Other: Ops[1]); |
1050 | case Intrinsic::abs: { |
1051 | const APInt *IntMinIsPoison = Ops[1].getSingleElement(); |
1052 | assert(IntMinIsPoison && "Must be known (immarg)" ); |
1053 | assert(IntMinIsPoison->getBitWidth() == 1 && "Must be boolean" ); |
1054 | return Ops[0].abs(IntMinIsPoison: IntMinIsPoison->getBoolValue()); |
1055 | } |
1056 | case Intrinsic::ctlz: { |
1057 | const APInt *ZeroIsPoison = Ops[1].getSingleElement(); |
1058 | assert(ZeroIsPoison && "Must be known (immarg)" ); |
1059 | assert(ZeroIsPoison->getBitWidth() == 1 && "Must be boolean" ); |
1060 | return Ops[0].ctlz(ZeroIsPoison: ZeroIsPoison->getBoolValue()); |
1061 | } |
1062 | case Intrinsic::cttz: { |
1063 | const APInt *ZeroIsPoison = Ops[1].getSingleElement(); |
1064 | assert(ZeroIsPoison && "Must be known (immarg)" ); |
1065 | assert(ZeroIsPoison->getBitWidth() == 1 && "Must be boolean" ); |
1066 | return Ops[0].cttz(ZeroIsPoison: ZeroIsPoison->getBoolValue()); |
1067 | } |
1068 | case Intrinsic::ctpop: |
1069 | return Ops[0].ctpop(); |
1070 | default: |
1071 | assert(!isIntrinsicSupported(IntrinsicID) && "Shouldn't be supported" ); |
1072 | llvm_unreachable("Unsupported intrinsic" ); |
1073 | } |
1074 | } |
1075 | |
1076 | ConstantRange |
1077 | ConstantRange::add(const ConstantRange &Other) const { |
1078 | if (isEmptySet() || Other.isEmptySet()) |
1079 | return getEmpty(); |
1080 | if (isFullSet() || Other.isFullSet()) |
1081 | return getFull(); |
1082 | |
1083 | APInt NewLower = getLower() + Other.getLower(); |
1084 | APInt NewUpper = getUpper() + Other.getUpper() - 1; |
1085 | if (NewLower == NewUpper) |
1086 | return getFull(); |
1087 | |
1088 | ConstantRange X = ConstantRange(std::move(NewLower), std::move(NewUpper)); |
1089 | if (X.isSizeStrictlySmallerThan(Other: *this) || |
1090 | X.isSizeStrictlySmallerThan(Other)) |
1091 | // We've wrapped, therefore, full set. |
1092 | return getFull(); |
1093 | return X; |
1094 | } |
1095 | |
1096 | ConstantRange ConstantRange::addWithNoWrap(const ConstantRange &Other, |
1097 | unsigned NoWrapKind, |
1098 | PreferredRangeType RangeType) const { |
1099 | // Calculate the range for "X + Y" which is guaranteed not to wrap(overflow). |
1100 | // (X is from this, and Y is from Other) |
1101 | if (isEmptySet() || Other.isEmptySet()) |
1102 | return getEmpty(); |
1103 | if (isFullSet() && Other.isFullSet()) |
1104 | return getFull(); |
1105 | |
1106 | using OBO = OverflowingBinaryOperator; |
1107 | ConstantRange Result = add(Other); |
1108 | |
1109 | // If an overflow happens for every value pair in these two constant ranges, |
1110 | // we must return Empty set. In this case, we get that for free, because we |
1111 | // get lucky that intersection of add() with uadd_sat()/sadd_sat() results |
1112 | // in an empty set. |
1113 | |
1114 | if (NoWrapKind & OBO::NoSignedWrap) |
1115 | Result = Result.intersectWith(CR: sadd_sat(Other), Type: RangeType); |
1116 | |
1117 | if (NoWrapKind & OBO::NoUnsignedWrap) |
1118 | Result = Result.intersectWith(CR: uadd_sat(Other), Type: RangeType); |
1119 | |
1120 | return Result; |
1121 | } |
1122 | |
1123 | ConstantRange |
1124 | ConstantRange::sub(const ConstantRange &Other) const { |
1125 | if (isEmptySet() || Other.isEmptySet()) |
1126 | return getEmpty(); |
1127 | if (isFullSet() || Other.isFullSet()) |
1128 | return getFull(); |
1129 | |
1130 | APInt NewLower = getLower() - Other.getUpper() + 1; |
1131 | APInt NewUpper = getUpper() - Other.getLower(); |
1132 | if (NewLower == NewUpper) |
1133 | return getFull(); |
1134 | |
1135 | ConstantRange X = ConstantRange(std::move(NewLower), std::move(NewUpper)); |
1136 | if (X.isSizeStrictlySmallerThan(Other: *this) || |
1137 | X.isSizeStrictlySmallerThan(Other)) |
1138 | // We've wrapped, therefore, full set. |
1139 | return getFull(); |
1140 | return X; |
1141 | } |
1142 | |
1143 | ConstantRange ConstantRange::subWithNoWrap(const ConstantRange &Other, |
1144 | unsigned NoWrapKind, |
1145 | PreferredRangeType RangeType) const { |
1146 | // Calculate the range for "X - Y" which is guaranteed not to wrap(overflow). |
1147 | // (X is from this, and Y is from Other) |
1148 | if (isEmptySet() || Other.isEmptySet()) |
1149 | return getEmpty(); |
1150 | if (isFullSet() && Other.isFullSet()) |
1151 | return getFull(); |
1152 | |
1153 | using OBO = OverflowingBinaryOperator; |
1154 | ConstantRange Result = sub(Other); |
1155 | |
1156 | // If an overflow happens for every value pair in these two constant ranges, |
1157 | // we must return Empty set. In signed case, we get that for free, because we |
1158 | // get lucky that intersection of sub() with ssub_sat() results in an |
1159 | // empty set. But for unsigned we must perform the overflow check manually. |
1160 | |
1161 | if (NoWrapKind & OBO::NoSignedWrap) |
1162 | Result = Result.intersectWith(CR: ssub_sat(Other), Type: RangeType); |
1163 | |
1164 | if (NoWrapKind & OBO::NoUnsignedWrap) { |
1165 | if (getUnsignedMax().ult(RHS: Other.getUnsignedMin())) |
1166 | return getEmpty(); // Always overflows. |
1167 | Result = Result.intersectWith(CR: usub_sat(Other), Type: RangeType); |
1168 | } |
1169 | |
1170 | return Result; |
1171 | } |
1172 | |
1173 | ConstantRange |
1174 | ConstantRange::multiply(const ConstantRange &Other) const { |
1175 | // TODO: If either operand is a single element and the multiply is known to |
1176 | // be non-wrapping, round the result min and max value to the appropriate |
1177 | // multiple of that element. If wrapping is possible, at least adjust the |
1178 | // range according to the greatest power-of-two factor of the single element. |
1179 | |
1180 | if (isEmptySet() || Other.isEmptySet()) |
1181 | return getEmpty(); |
1182 | |
1183 | if (const APInt *C = getSingleElement()) { |
1184 | if (C->isOne()) |
1185 | return Other; |
1186 | if (C->isAllOnes()) |
1187 | return ConstantRange(APInt::getZero(numBits: getBitWidth())).sub(Other); |
1188 | } |
1189 | |
1190 | if (const APInt *C = Other.getSingleElement()) { |
1191 | if (C->isOne()) |
1192 | return *this; |
1193 | if (C->isAllOnes()) |
1194 | return ConstantRange(APInt::getZero(numBits: getBitWidth())).sub(Other: *this); |
1195 | } |
1196 | |
1197 | // Multiplication is signedness-independent. However different ranges can be |
1198 | // obtained depending on how the input ranges are treated. These different |
1199 | // ranges are all conservatively correct, but one might be better than the |
1200 | // other. We calculate two ranges; one treating the inputs as unsigned |
1201 | // and the other signed, then return the smallest of these ranges. |
1202 | |
1203 | // Unsigned range first. |
1204 | APInt this_min = getUnsignedMin().zext(width: getBitWidth() * 2); |
1205 | APInt this_max = getUnsignedMax().zext(width: getBitWidth() * 2); |
1206 | APInt Other_min = Other.getUnsignedMin().zext(width: getBitWidth() * 2); |
1207 | APInt Other_max = Other.getUnsignedMax().zext(width: getBitWidth() * 2); |
1208 | |
1209 | ConstantRange Result_zext = ConstantRange(this_min * Other_min, |
1210 | this_max * Other_max + 1); |
1211 | ConstantRange UR = Result_zext.truncate(DstTySize: getBitWidth()); |
1212 | |
1213 | // If the unsigned range doesn't wrap, and isn't negative then it's a range |
1214 | // from one positive number to another which is as good as we can generate. |
1215 | // In this case, skip the extra work of generating signed ranges which aren't |
1216 | // going to be better than this range. |
1217 | if (!UR.isUpperWrapped() && |
1218 | (UR.getUpper().isNonNegative() || UR.getUpper().isMinSignedValue())) |
1219 | return UR; |
1220 | |
1221 | // Now the signed range. Because we could be dealing with negative numbers |
1222 | // here, the lower bound is the smallest of the cartesian product of the |
1223 | // lower and upper ranges; for example: |
1224 | // [-1,4) * [-2,3) = min(-1*-2, -1*2, 3*-2, 3*2) = -6. |
1225 | // Similarly for the upper bound, swapping min for max. |
1226 | |
1227 | this_min = getSignedMin().sext(width: getBitWidth() * 2); |
1228 | this_max = getSignedMax().sext(width: getBitWidth() * 2); |
1229 | Other_min = Other.getSignedMin().sext(width: getBitWidth() * 2); |
1230 | Other_max = Other.getSignedMax().sext(width: getBitWidth() * 2); |
1231 | |
1232 | auto L = {this_min * Other_min, this_min * Other_max, |
1233 | this_max * Other_min, this_max * Other_max}; |
1234 | auto Compare = [](const APInt &A, const APInt &B) { return A.slt(RHS: B); }; |
1235 | ConstantRange Result_sext(std::min(l: L, comp: Compare), std::max(l: L, comp: Compare) + 1); |
1236 | ConstantRange SR = Result_sext.truncate(DstTySize: getBitWidth()); |
1237 | |
1238 | return UR.isSizeStrictlySmallerThan(Other: SR) ? UR : SR; |
1239 | } |
1240 | |
1241 | ConstantRange |
1242 | ConstantRange::multiplyWithNoWrap(const ConstantRange &Other, |
1243 | unsigned NoWrapKind, |
1244 | PreferredRangeType RangeType) const { |
1245 | if (isEmptySet() || Other.isEmptySet()) |
1246 | return getEmpty(); |
1247 | if (isFullSet() && Other.isFullSet()) |
1248 | return getFull(); |
1249 | |
1250 | ConstantRange Result = multiply(Other); |
1251 | |
1252 | if (NoWrapKind & OverflowingBinaryOperator::NoSignedWrap) |
1253 | Result = Result.intersectWith(CR: smul_sat(Other), Type: RangeType); |
1254 | |
1255 | if (NoWrapKind & OverflowingBinaryOperator::NoUnsignedWrap) |
1256 | Result = Result.intersectWith(CR: umul_sat(Other), Type: RangeType); |
1257 | |
1258 | // mul nsw nuw X, Y s>= 0 if X s> 1 or Y s> 1 |
1259 | if ((NoWrapKind == (OverflowingBinaryOperator::NoSignedWrap | |
1260 | OverflowingBinaryOperator::NoUnsignedWrap)) && |
1261 | !Result.isAllNonNegative()) { |
1262 | if (getSignedMin().sgt(RHS: 1) || Other.getSignedMin().sgt(RHS: 1)) |
1263 | Result = Result.intersectWith( |
1264 | CR: getNonEmpty(Lower: APInt::getZero(numBits: getBitWidth()), |
1265 | Upper: APInt::getSignedMinValue(numBits: getBitWidth())), |
1266 | Type: RangeType); |
1267 | } |
1268 | |
1269 | return Result; |
1270 | } |
1271 | |
1272 | ConstantRange ConstantRange::smul_fast(const ConstantRange &Other) const { |
1273 | if (isEmptySet() || Other.isEmptySet()) |
1274 | return getEmpty(); |
1275 | |
1276 | APInt Min = getSignedMin(); |
1277 | APInt Max = getSignedMax(); |
1278 | APInt OtherMin = Other.getSignedMin(); |
1279 | APInt OtherMax = Other.getSignedMax(); |
1280 | |
1281 | bool O1, O2, O3, O4; |
1282 | auto Muls = {Min.smul_ov(RHS: OtherMin, Overflow&: O1), Min.smul_ov(RHS: OtherMax, Overflow&: O2), |
1283 | Max.smul_ov(RHS: OtherMin, Overflow&: O3), Max.smul_ov(RHS: OtherMax, Overflow&: O4)}; |
1284 | if (O1 || O2 || O3 || O4) |
1285 | return getFull(); |
1286 | |
1287 | auto Compare = [](const APInt &A, const APInt &B) { return A.slt(RHS: B); }; |
1288 | return getNonEmpty(Lower: std::min(l: Muls, comp: Compare), Upper: std::max(l: Muls, comp: Compare) + 1); |
1289 | } |
1290 | |
1291 | ConstantRange |
1292 | ConstantRange::smax(const ConstantRange &Other) const { |
1293 | // X smax Y is: range(smax(X_smin, Y_smin), |
1294 | // smax(X_smax, Y_smax)) |
1295 | if (isEmptySet() || Other.isEmptySet()) |
1296 | return getEmpty(); |
1297 | APInt NewL = APIntOps::smax(A: getSignedMin(), B: Other.getSignedMin()); |
1298 | APInt NewU = APIntOps::smax(A: getSignedMax(), B: Other.getSignedMax()) + 1; |
1299 | ConstantRange Res = getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1300 | if (isSignWrappedSet() || Other.isSignWrappedSet()) |
1301 | return Res.intersectWith(CR: unionWith(CR: Other, Type: Signed), Type: Signed); |
1302 | return Res; |
1303 | } |
1304 | |
1305 | ConstantRange |
1306 | ConstantRange::umax(const ConstantRange &Other) const { |
1307 | // X umax Y is: range(umax(X_umin, Y_umin), |
1308 | // umax(X_umax, Y_umax)) |
1309 | if (isEmptySet() || Other.isEmptySet()) |
1310 | return getEmpty(); |
1311 | APInt NewL = APIntOps::umax(A: getUnsignedMin(), B: Other.getUnsignedMin()); |
1312 | APInt NewU = APIntOps::umax(A: getUnsignedMax(), B: Other.getUnsignedMax()) + 1; |
1313 | ConstantRange Res = getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1314 | if (isWrappedSet() || Other.isWrappedSet()) |
1315 | return Res.intersectWith(CR: unionWith(CR: Other, Type: Unsigned), Type: Unsigned); |
1316 | return Res; |
1317 | } |
1318 | |
1319 | ConstantRange |
1320 | ConstantRange::smin(const ConstantRange &Other) const { |
1321 | // X smin Y is: range(smin(X_smin, Y_smin), |
1322 | // smin(X_smax, Y_smax)) |
1323 | if (isEmptySet() || Other.isEmptySet()) |
1324 | return getEmpty(); |
1325 | APInt NewL = APIntOps::smin(A: getSignedMin(), B: Other.getSignedMin()); |
1326 | APInt NewU = APIntOps::smin(A: getSignedMax(), B: Other.getSignedMax()) + 1; |
1327 | ConstantRange Res = getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1328 | if (isSignWrappedSet() || Other.isSignWrappedSet()) |
1329 | return Res.intersectWith(CR: unionWith(CR: Other, Type: Signed), Type: Signed); |
1330 | return Res; |
1331 | } |
1332 | |
1333 | ConstantRange |
1334 | ConstantRange::umin(const ConstantRange &Other) const { |
1335 | // X umin Y is: range(umin(X_umin, Y_umin), |
1336 | // umin(X_umax, Y_umax)) |
1337 | if (isEmptySet() || Other.isEmptySet()) |
1338 | return getEmpty(); |
1339 | APInt NewL = APIntOps::umin(A: getUnsignedMin(), B: Other.getUnsignedMin()); |
1340 | APInt NewU = APIntOps::umin(A: getUnsignedMax(), B: Other.getUnsignedMax()) + 1; |
1341 | ConstantRange Res = getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1342 | if (isWrappedSet() || Other.isWrappedSet()) |
1343 | return Res.intersectWith(CR: unionWith(CR: Other, Type: Unsigned), Type: Unsigned); |
1344 | return Res; |
1345 | } |
1346 | |
1347 | ConstantRange |
1348 | ConstantRange::udiv(const ConstantRange &RHS) const { |
1349 | if (isEmptySet() || RHS.isEmptySet() || RHS.getUnsignedMax().isZero()) |
1350 | return getEmpty(); |
1351 | |
1352 | APInt Lower = getUnsignedMin().udiv(RHS: RHS.getUnsignedMax()); |
1353 | |
1354 | APInt RHS_umin = RHS.getUnsignedMin(); |
1355 | if (RHS_umin.isZero()) { |
1356 | // We want the lowest value in RHS excluding zero. Usually that would be 1 |
1357 | // except for a range in the form of [X, 1) in which case it would be X. |
1358 | if (RHS.getUpper() == 1) |
1359 | RHS_umin = RHS.getLower(); |
1360 | else |
1361 | RHS_umin = 1; |
1362 | } |
1363 | |
1364 | APInt Upper = getUnsignedMax().udiv(RHS: RHS_umin) + 1; |
1365 | return getNonEmpty(Lower: std::move(Lower), Upper: std::move(Upper)); |
1366 | } |
1367 | |
1368 | ConstantRange ConstantRange::sdiv(const ConstantRange &RHS) const { |
1369 | APInt Zero = APInt::getZero(numBits: getBitWidth()); |
1370 | APInt SignedMin = APInt::getSignedMinValue(numBits: getBitWidth()); |
1371 | |
1372 | // We split up the LHS and RHS into positive and negative components |
1373 | // and then also compute the positive and negative components of the result |
1374 | // separately by combining division results with the appropriate signs. |
1375 | auto [PosL, NegL] = splitPosNeg(); |
1376 | auto [PosR, NegR] = RHS.splitPosNeg(); |
1377 | |
1378 | ConstantRange PosRes = getEmpty(); |
1379 | if (!PosL.isEmptySet() && !PosR.isEmptySet()) |
1380 | // pos / pos = pos. |
1381 | PosRes = ConstantRange(PosL.Lower.sdiv(RHS: PosR.Upper - 1), |
1382 | (PosL.Upper - 1).sdiv(RHS: PosR.Lower) + 1); |
1383 | |
1384 | if (!NegL.isEmptySet() && !NegR.isEmptySet()) { |
1385 | // neg / neg = pos. |
1386 | // |
1387 | // We need to deal with one tricky case here: SignedMin / -1 is UB on the |
1388 | // IR level, so we'll want to exclude this case when calculating bounds. |
1389 | // (For APInts the operation is well-defined and yields SignedMin.) We |
1390 | // handle this by dropping either SignedMin from the LHS or -1 from the RHS. |
1391 | APInt Lo = (NegL.Upper - 1).sdiv(RHS: NegR.Lower); |
1392 | if (NegL.Lower.isMinSignedValue() && NegR.Upper.isZero()) { |
1393 | // Remove -1 from the LHS. Skip if it's the only element, as this would |
1394 | // leave us with an empty set. |
1395 | if (!NegR.Lower.isAllOnes()) { |
1396 | APInt AdjNegRUpper; |
1397 | if (RHS.Lower.isAllOnes()) |
1398 | // Negative part of [-1, X] without -1 is [SignedMin, X]. |
1399 | AdjNegRUpper = RHS.Upper; |
1400 | else |
1401 | // [X, -1] without -1 is [X, -2]. |
1402 | AdjNegRUpper = NegR.Upper - 1; |
1403 | |
1404 | PosRes = PosRes.unionWith( |
1405 | CR: ConstantRange(Lo, NegL.Lower.sdiv(RHS: AdjNegRUpper - 1) + 1)); |
1406 | } |
1407 | |
1408 | // Remove SignedMin from the RHS. Skip if it's the only element, as this |
1409 | // would leave us with an empty set. |
1410 | if (NegL.Upper != SignedMin + 1) { |
1411 | APInt AdjNegLLower; |
1412 | if (Upper == SignedMin + 1) |
1413 | // Negative part of [X, SignedMin] without SignedMin is [X, -1]. |
1414 | AdjNegLLower = Lower; |
1415 | else |
1416 | // [SignedMin, X] without SignedMin is [SignedMin + 1, X]. |
1417 | AdjNegLLower = NegL.Lower + 1; |
1418 | |
1419 | PosRes = PosRes.unionWith( |
1420 | CR: ConstantRange(std::move(Lo), |
1421 | AdjNegLLower.sdiv(RHS: NegR.Upper - 1) + 1)); |
1422 | } |
1423 | } else { |
1424 | PosRes = PosRes.unionWith( |
1425 | CR: ConstantRange(std::move(Lo), NegL.Lower.sdiv(RHS: NegR.Upper - 1) + 1)); |
1426 | } |
1427 | } |
1428 | |
1429 | ConstantRange NegRes = getEmpty(); |
1430 | if (!PosL.isEmptySet() && !NegR.isEmptySet()) |
1431 | // pos / neg = neg. |
1432 | NegRes = ConstantRange((PosL.Upper - 1).sdiv(RHS: NegR.Upper - 1), |
1433 | PosL.Lower.sdiv(RHS: NegR.Lower) + 1); |
1434 | |
1435 | if (!NegL.isEmptySet() && !PosR.isEmptySet()) |
1436 | // neg / pos = neg. |
1437 | NegRes = NegRes.unionWith( |
1438 | CR: ConstantRange(NegL.Lower.sdiv(RHS: PosR.Lower), |
1439 | (NegL.Upper - 1).sdiv(RHS: PosR.Upper - 1) + 1)); |
1440 | |
1441 | // Prefer a non-wrapping signed range here. |
1442 | ConstantRange Res = NegRes.unionWith(CR: PosRes, Type: PreferredRangeType::Signed); |
1443 | |
1444 | // Preserve the zero that we dropped when splitting the LHS by sign. |
1445 | if (contains(V: Zero) && (!PosR.isEmptySet() || !NegR.isEmptySet())) |
1446 | Res = Res.unionWith(CR: ConstantRange(Zero)); |
1447 | return Res; |
1448 | } |
1449 | |
1450 | ConstantRange ConstantRange::urem(const ConstantRange &RHS) const { |
1451 | if (isEmptySet() || RHS.isEmptySet() || RHS.getUnsignedMax().isZero()) |
1452 | return getEmpty(); |
1453 | |
1454 | if (const APInt *RHSInt = RHS.getSingleElement()) { |
1455 | // UREM by null is UB. |
1456 | if (RHSInt->isZero()) |
1457 | return getEmpty(); |
1458 | // Use APInt's implementation of UREM for single element ranges. |
1459 | if (const APInt *LHSInt = getSingleElement()) |
1460 | return {LHSInt->urem(RHS: *RHSInt)}; |
1461 | } |
1462 | |
1463 | // L % R for L < R is L. |
1464 | if (getUnsignedMax().ult(RHS: RHS.getUnsignedMin())) |
1465 | return *this; |
1466 | |
1467 | // L % R is <= L and < R. |
1468 | APInt Upper = APIntOps::umin(A: getUnsignedMax(), B: RHS.getUnsignedMax() - 1) + 1; |
1469 | return getNonEmpty(Lower: APInt::getZero(numBits: getBitWidth()), Upper: std::move(Upper)); |
1470 | } |
1471 | |
1472 | ConstantRange ConstantRange::srem(const ConstantRange &RHS) const { |
1473 | if (isEmptySet() || RHS.isEmptySet()) |
1474 | return getEmpty(); |
1475 | |
1476 | if (const APInt *RHSInt = RHS.getSingleElement()) { |
1477 | // SREM by null is UB. |
1478 | if (RHSInt->isZero()) |
1479 | return getEmpty(); |
1480 | // Use APInt's implementation of SREM for single element ranges. |
1481 | if (const APInt *LHSInt = getSingleElement()) |
1482 | return {LHSInt->srem(RHS: *RHSInt)}; |
1483 | } |
1484 | |
1485 | ConstantRange AbsRHS = RHS.abs(); |
1486 | APInt MinAbsRHS = AbsRHS.getUnsignedMin(); |
1487 | APInt MaxAbsRHS = AbsRHS.getUnsignedMax(); |
1488 | |
1489 | // Modulus by zero is UB. |
1490 | if (MaxAbsRHS.isZero()) |
1491 | return getEmpty(); |
1492 | |
1493 | if (MinAbsRHS.isZero()) |
1494 | ++MinAbsRHS; |
1495 | |
1496 | APInt MinLHS = getSignedMin(), MaxLHS = getSignedMax(); |
1497 | |
1498 | if (MinLHS.isNonNegative()) { |
1499 | // L % R for L < R is L. |
1500 | if (MaxLHS.ult(RHS: MinAbsRHS)) |
1501 | return *this; |
1502 | |
1503 | // L % R is <= L and < R. |
1504 | APInt Upper = APIntOps::umin(A: MaxLHS, B: MaxAbsRHS - 1) + 1; |
1505 | return ConstantRange(APInt::getZero(numBits: getBitWidth()), std::move(Upper)); |
1506 | } |
1507 | |
1508 | // Same basic logic as above, but the result is negative. |
1509 | if (MaxLHS.isNegative()) { |
1510 | if (MinLHS.ugt(RHS: -MinAbsRHS)) |
1511 | return *this; |
1512 | |
1513 | APInt Lower = APIntOps::umax(A: MinLHS, B: -MaxAbsRHS + 1); |
1514 | return ConstantRange(std::move(Lower), APInt(getBitWidth(), 1)); |
1515 | } |
1516 | |
1517 | // LHS range crosses zero. |
1518 | APInt Lower = APIntOps::umax(A: MinLHS, B: -MaxAbsRHS + 1); |
1519 | APInt Upper = APIntOps::umin(A: MaxLHS, B: MaxAbsRHS - 1) + 1; |
1520 | return ConstantRange(std::move(Lower), std::move(Upper)); |
1521 | } |
1522 | |
1523 | ConstantRange ConstantRange::binaryNot() const { |
1524 | return ConstantRange(APInt::getAllOnes(numBits: getBitWidth())).sub(Other: *this); |
1525 | } |
1526 | |
1527 | /// Estimate the 'bit-masked AND' operation's lower bound. |
1528 | /// |
1529 | /// E.g., given two ranges as follows (single quotes are separators and |
1530 | /// have no meaning here), |
1531 | /// |
1532 | /// LHS = [10'00101'1, ; LLo |
1533 | /// 10'10000'0] ; LHi |
1534 | /// RHS = [10'11111'0, ; RLo |
1535 | /// 10'11111'1] ; RHi |
1536 | /// |
1537 | /// we know that the higher 2 bits of the result is always 10; and we also |
1538 | /// notice that RHS[1:6] are always 1, so the result[1:6] cannot be less than |
1539 | /// LHS[1:6] (i.e., 00101). Thus, the lower bound is 10'00101'0. |
1540 | /// |
1541 | /// The algorithm is as follows, |
1542 | /// 1. we first calculate a mask to find the higher common bits by |
1543 | /// Mask = ~((LLo ^ LHi) | (RLo ^ RHi) | (LLo ^ RLo)); |
1544 | /// Mask = clear all non-leading-ones bits in Mask; |
1545 | /// in the example, the Mask is set to 11'00000'0; |
1546 | /// 2. calculate a new mask by setting all common leading bits to 1 in RHS, and |
1547 | /// keeping the longest leading ones (i.e., 11'11111'0 in the example); |
1548 | /// 3. return (LLo & new mask) as the lower bound; |
1549 | /// 4. repeat the step 2 and 3 with LHS and RHS swapped, and update the lower |
1550 | /// bound with the larger one. |
1551 | static APInt estimateBitMaskedAndLowerBound(const ConstantRange &LHS, |
1552 | const ConstantRange &RHS) { |
1553 | auto BitWidth = LHS.getBitWidth(); |
1554 | // If either is full set or unsigned wrapped, then the range must contain '0' |
1555 | // which leads the lower bound to 0. |
1556 | if ((LHS.isFullSet() || RHS.isFullSet()) || |
1557 | (LHS.isWrappedSet() || RHS.isWrappedSet())) |
1558 | return APInt::getZero(numBits: BitWidth); |
1559 | |
1560 | auto LLo = LHS.getLower(); |
1561 | auto LHi = LHS.getUpper() - 1; |
1562 | auto RLo = RHS.getLower(); |
1563 | auto RHi = RHS.getUpper() - 1; |
1564 | |
1565 | // Calculate the mask for the higher common bits. |
1566 | auto Mask = ~((LLo ^ LHi) | (RLo ^ RHi) | (LLo ^ RLo)); |
1567 | unsigned LeadingOnes = Mask.countLeadingOnes(); |
1568 | Mask.clearLowBits(loBits: BitWidth - LeadingOnes); |
1569 | |
1570 | auto estimateBound = [BitWidth, &Mask](APInt ALo, const APInt &BLo, |
1571 | const APInt &BHi) { |
1572 | unsigned LeadingOnes = ((BLo & BHi) | Mask).countLeadingOnes(); |
1573 | unsigned StartBit = BitWidth - LeadingOnes; |
1574 | ALo.clearLowBits(loBits: StartBit); |
1575 | return ALo; |
1576 | }; |
1577 | |
1578 | auto LowerBoundByLHS = estimateBound(LLo, RLo, RHi); |
1579 | auto LowerBoundByRHS = estimateBound(RLo, LLo, LHi); |
1580 | |
1581 | return APIntOps::umax(A: LowerBoundByLHS, B: LowerBoundByRHS); |
1582 | } |
1583 | |
1584 | ConstantRange ConstantRange::binaryAnd(const ConstantRange &Other) const { |
1585 | if (isEmptySet() || Other.isEmptySet()) |
1586 | return getEmpty(); |
1587 | |
1588 | ConstantRange KnownBitsRange = |
1589 | fromKnownBits(Known: toKnownBits() & Other.toKnownBits(), IsSigned: false); |
1590 | auto LowerBound = estimateBitMaskedAndLowerBound(LHS: *this, RHS: Other); |
1591 | ConstantRange UMinUMaxRange = getNonEmpty( |
1592 | Lower: LowerBound, Upper: APIntOps::umin(A: Other.getUnsignedMax(), B: getUnsignedMax()) + 1); |
1593 | return KnownBitsRange.intersectWith(CR: UMinUMaxRange); |
1594 | } |
1595 | |
1596 | ConstantRange ConstantRange::binaryOr(const ConstantRange &Other) const { |
1597 | if (isEmptySet() || Other.isEmptySet()) |
1598 | return getEmpty(); |
1599 | |
1600 | ConstantRange KnownBitsRange = |
1601 | fromKnownBits(Known: toKnownBits() | Other.toKnownBits(), IsSigned: false); |
1602 | |
1603 | // ~a & ~b >= x |
1604 | // <=> ~(~a & ~b) <= ~x |
1605 | // <=> a | b <= ~x |
1606 | // <=> a | b < ~x + 1 = -x |
1607 | // thus, UpperBound(a | b) == -LowerBound(~a & ~b) |
1608 | auto UpperBound = |
1609 | -estimateBitMaskedAndLowerBound(LHS: binaryNot(), RHS: Other.binaryNot()); |
1610 | // Upper wrapped range. |
1611 | ConstantRange UMaxUMinRange = getNonEmpty( |
1612 | Lower: APIntOps::umax(A: getUnsignedMin(), B: Other.getUnsignedMin()), Upper: UpperBound); |
1613 | return KnownBitsRange.intersectWith(CR: UMaxUMinRange); |
1614 | } |
1615 | |
1616 | ConstantRange ConstantRange::binaryXor(const ConstantRange &Other) const { |
1617 | if (isEmptySet() || Other.isEmptySet()) |
1618 | return getEmpty(); |
1619 | |
1620 | // Use APInt's implementation of XOR for single element ranges. |
1621 | if (isSingleElement() && Other.isSingleElement()) |
1622 | return {*getSingleElement() ^ *Other.getSingleElement()}; |
1623 | |
1624 | // Special-case binary complement, since we can give a precise answer. |
1625 | if (Other.isSingleElement() && Other.getSingleElement()->isAllOnes()) |
1626 | return binaryNot(); |
1627 | if (isSingleElement() && getSingleElement()->isAllOnes()) |
1628 | return Other.binaryNot(); |
1629 | |
1630 | KnownBits LHSKnown = toKnownBits(); |
1631 | KnownBits RHSKnown = Other.toKnownBits(); |
1632 | KnownBits Known = LHSKnown ^ RHSKnown; |
1633 | ConstantRange CR = fromKnownBits(Known, /*IsSigned*/ false); |
1634 | // Typically the following code doesn't improve the result if BW = 1. |
1635 | if (getBitWidth() == 1) |
1636 | return CR; |
1637 | |
1638 | // If LHS is known to be the subset of RHS, treat LHS ^ RHS as RHS -nuw/nsw |
1639 | // LHS. If RHS is known to be the subset of LHS, treat LHS ^ RHS as LHS |
1640 | // -nuw/nsw RHS. |
1641 | if ((~LHSKnown.Zero).isSubsetOf(RHS: RHSKnown.One)) |
1642 | CR = CR.intersectWith(CR: Other.sub(Other: *this), Type: PreferredRangeType::Unsigned); |
1643 | else if ((~RHSKnown.Zero).isSubsetOf(RHS: LHSKnown.One)) |
1644 | CR = CR.intersectWith(CR: this->sub(Other), Type: PreferredRangeType::Unsigned); |
1645 | return CR; |
1646 | } |
1647 | |
1648 | ConstantRange |
1649 | ConstantRange::shl(const ConstantRange &Other) const { |
1650 | if (isEmptySet() || Other.isEmptySet()) |
1651 | return getEmpty(); |
1652 | |
1653 | APInt Min = getUnsignedMin(); |
1654 | APInt Max = getUnsignedMax(); |
1655 | if (const APInt *RHS = Other.getSingleElement()) { |
1656 | unsigned BW = getBitWidth(); |
1657 | if (RHS->uge(RHS: BW)) |
1658 | return getEmpty(); |
1659 | |
1660 | unsigned EqualLeadingBits = (Min ^ Max).countl_zero(); |
1661 | if (RHS->ule(RHS: EqualLeadingBits)) |
1662 | return getNonEmpty(Lower: Min << *RHS, Upper: (Max << *RHS) + 1); |
1663 | |
1664 | return getNonEmpty(Lower: APInt::getZero(numBits: BW), |
1665 | Upper: APInt::getBitsSetFrom(numBits: BW, loBit: RHS->getZExtValue()) + 1); |
1666 | } |
1667 | |
1668 | APInt OtherMax = Other.getUnsignedMax(); |
1669 | if (isAllNegative() && OtherMax.ule(RHS: Min.countl_one())) { |
1670 | // For negative numbers, if the shift does not overflow in a signed sense, |
1671 | // a larger shift will make the number smaller. |
1672 | Max <<= Other.getUnsignedMin(); |
1673 | Min <<= OtherMax; |
1674 | return ConstantRange::getNonEmpty(Lower: std::move(Min), Upper: std::move(Max) + 1); |
1675 | } |
1676 | |
1677 | // There's overflow! |
1678 | if (OtherMax.ugt(RHS: Max.countl_zero())) |
1679 | return getFull(); |
1680 | |
1681 | // FIXME: implement the other tricky cases |
1682 | |
1683 | Min <<= Other.getUnsignedMin(); |
1684 | Max <<= OtherMax; |
1685 | |
1686 | return ConstantRange::getNonEmpty(Lower: std::move(Min), Upper: std::move(Max) + 1); |
1687 | } |
1688 | |
1689 | static ConstantRange computeShlNUW(const ConstantRange &LHS, |
1690 | const ConstantRange &RHS) { |
1691 | unsigned BitWidth = LHS.getBitWidth(); |
1692 | bool Overflow; |
1693 | APInt LHSMin = LHS.getUnsignedMin(); |
1694 | unsigned RHSMin = RHS.getUnsignedMin().getLimitedValue(Limit: BitWidth); |
1695 | APInt MinShl = LHSMin.ushl_ov(Amt: RHSMin, Overflow); |
1696 | if (Overflow) |
1697 | return ConstantRange::getEmpty(BitWidth); |
1698 | APInt LHSMax = LHS.getUnsignedMax(); |
1699 | unsigned RHSMax = RHS.getUnsignedMax().getLimitedValue(Limit: BitWidth); |
1700 | APInt MaxShl = MinShl; |
1701 | unsigned MaxShAmt = LHSMax.countLeadingZeros(); |
1702 | if (RHSMin <= MaxShAmt) |
1703 | MaxShl = LHSMax << std::min(a: RHSMax, b: MaxShAmt); |
1704 | RHSMin = std::max(a: RHSMin, b: MaxShAmt + 1); |
1705 | RHSMax = std::min(a: RHSMax, b: LHSMin.countLeadingZeros()); |
1706 | if (RHSMin <= RHSMax) |
1707 | MaxShl = APIntOps::umax(A: MaxShl, |
1708 | B: APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - RHSMin)); |
1709 | return ConstantRange::getNonEmpty(Lower: MinShl, Upper: MaxShl + 1); |
1710 | } |
1711 | |
1712 | static ConstantRange computeShlNSWWithNNegLHS(const APInt &LHSMin, |
1713 | const APInt &LHSMax, |
1714 | unsigned RHSMin, |
1715 | unsigned RHSMax) { |
1716 | unsigned BitWidth = LHSMin.getBitWidth(); |
1717 | bool Overflow; |
1718 | APInt MinShl = LHSMin.sshl_ov(Amt: RHSMin, Overflow); |
1719 | if (Overflow) |
1720 | return ConstantRange::getEmpty(BitWidth); |
1721 | APInt MaxShl = MinShl; |
1722 | unsigned MaxShAmt = LHSMax.countLeadingZeros() - 1; |
1723 | if (RHSMin <= MaxShAmt) |
1724 | MaxShl = LHSMax << std::min(a: RHSMax, b: MaxShAmt); |
1725 | RHSMin = std::max(a: RHSMin, b: MaxShAmt + 1); |
1726 | RHSMax = std::min(a: RHSMax, b: LHSMin.countLeadingZeros() - 1); |
1727 | if (RHSMin <= RHSMax) |
1728 | MaxShl = APIntOps::umax(A: MaxShl, |
1729 | B: APInt::getBitsSet(numBits: BitWidth, loBit: RHSMin, hiBit: BitWidth - 1)); |
1730 | return ConstantRange::getNonEmpty(Lower: MinShl, Upper: MaxShl + 1); |
1731 | } |
1732 | |
1733 | static ConstantRange computeShlNSWWithNegLHS(const APInt &LHSMin, |
1734 | const APInt &LHSMax, |
1735 | unsigned RHSMin, unsigned RHSMax) { |
1736 | unsigned BitWidth = LHSMin.getBitWidth(); |
1737 | bool Overflow; |
1738 | APInt MaxShl = LHSMax.sshl_ov(Amt: RHSMin, Overflow); |
1739 | if (Overflow) |
1740 | return ConstantRange::getEmpty(BitWidth); |
1741 | APInt MinShl = MaxShl; |
1742 | unsigned MaxShAmt = LHSMin.countLeadingOnes() - 1; |
1743 | if (RHSMin <= MaxShAmt) |
1744 | MinShl = LHSMin.shl(shiftAmt: std::min(a: RHSMax, b: MaxShAmt)); |
1745 | RHSMin = std::max(a: RHSMin, b: MaxShAmt + 1); |
1746 | RHSMax = std::min(a: RHSMax, b: LHSMax.countLeadingOnes() - 1); |
1747 | if (RHSMin <= RHSMax) |
1748 | MinShl = APInt::getSignMask(BitWidth); |
1749 | return ConstantRange::getNonEmpty(Lower: MinShl, Upper: MaxShl + 1); |
1750 | } |
1751 | |
1752 | static ConstantRange computeShlNSW(const ConstantRange &LHS, |
1753 | const ConstantRange &RHS) { |
1754 | unsigned BitWidth = LHS.getBitWidth(); |
1755 | unsigned RHSMin = RHS.getUnsignedMin().getLimitedValue(Limit: BitWidth); |
1756 | unsigned RHSMax = RHS.getUnsignedMax().getLimitedValue(Limit: BitWidth); |
1757 | APInt LHSMin = LHS.getSignedMin(); |
1758 | APInt LHSMax = LHS.getSignedMax(); |
1759 | if (LHSMin.isNonNegative()) |
1760 | return computeShlNSWWithNNegLHS(LHSMin, LHSMax, RHSMin, RHSMax); |
1761 | else if (LHSMax.isNegative()) |
1762 | return computeShlNSWWithNegLHS(LHSMin, LHSMax, RHSMin, RHSMax); |
1763 | return computeShlNSWWithNNegLHS(LHSMin: APInt::getZero(numBits: BitWidth), LHSMax, RHSMin, |
1764 | RHSMax) |
1765 | .unionWith(CR: computeShlNSWWithNegLHS(LHSMin, LHSMax: APInt::getAllOnes(numBits: BitWidth), |
1766 | RHSMin, RHSMax), |
1767 | Type: ConstantRange::Signed); |
1768 | } |
1769 | |
1770 | ConstantRange ConstantRange::shlWithNoWrap(const ConstantRange &Other, |
1771 | unsigned NoWrapKind, |
1772 | PreferredRangeType RangeType) const { |
1773 | if (isEmptySet() || Other.isEmptySet()) |
1774 | return getEmpty(); |
1775 | |
1776 | switch (NoWrapKind) { |
1777 | case 0: |
1778 | return shl(Other); |
1779 | case OverflowingBinaryOperator::NoSignedWrap: |
1780 | return computeShlNSW(LHS: *this, RHS: Other); |
1781 | case OverflowingBinaryOperator::NoUnsignedWrap: |
1782 | return computeShlNUW(LHS: *this, RHS: Other); |
1783 | case OverflowingBinaryOperator::NoSignedWrap | |
1784 | OverflowingBinaryOperator::NoUnsignedWrap: |
1785 | return computeShlNSW(LHS: *this, RHS: Other) |
1786 | .intersectWith(CR: computeShlNUW(LHS: *this, RHS: Other), Type: RangeType); |
1787 | default: |
1788 | llvm_unreachable("Invalid NoWrapKind" ); |
1789 | } |
1790 | } |
1791 | |
1792 | ConstantRange |
1793 | ConstantRange::lshr(const ConstantRange &Other) const { |
1794 | if (isEmptySet() || Other.isEmptySet()) |
1795 | return getEmpty(); |
1796 | |
1797 | APInt max = getUnsignedMax().lshr(ShiftAmt: Other.getUnsignedMin()) + 1; |
1798 | APInt min = getUnsignedMin().lshr(ShiftAmt: Other.getUnsignedMax()); |
1799 | return getNonEmpty(Lower: std::move(min), Upper: std::move(max)); |
1800 | } |
1801 | |
1802 | ConstantRange |
1803 | ConstantRange::ashr(const ConstantRange &Other) const { |
1804 | if (isEmptySet() || Other.isEmptySet()) |
1805 | return getEmpty(); |
1806 | |
1807 | // May straddle zero, so handle both positive and negative cases. |
1808 | // 'PosMax' is the upper bound of the result of the ashr |
1809 | // operation, when Upper of the LHS of ashr is a non-negative. |
1810 | // number. Since ashr of a non-negative number will result in a |
1811 | // smaller number, the Upper value of LHS is shifted right with |
1812 | // the minimum value of 'Other' instead of the maximum value. |
1813 | APInt PosMax = getSignedMax().ashr(ShiftAmt: Other.getUnsignedMin()) + 1; |
1814 | |
1815 | // 'PosMin' is the lower bound of the result of the ashr |
1816 | // operation, when Lower of the LHS is a non-negative number. |
1817 | // Since ashr of a non-negative number will result in a smaller |
1818 | // number, the Lower value of LHS is shifted right with the |
1819 | // maximum value of 'Other'. |
1820 | APInt PosMin = getSignedMin().ashr(ShiftAmt: Other.getUnsignedMax()); |
1821 | |
1822 | // 'NegMax' is the upper bound of the result of the ashr |
1823 | // operation, when Upper of the LHS of ashr is a negative number. |
1824 | // Since 'ashr' of a negative number will result in a bigger |
1825 | // number, the Upper value of LHS is shifted right with the |
1826 | // maximum value of 'Other'. |
1827 | APInt NegMax = getSignedMax().ashr(ShiftAmt: Other.getUnsignedMax()) + 1; |
1828 | |
1829 | // 'NegMin' is the lower bound of the result of the ashr |
1830 | // operation, when Lower of the LHS of ashr is a negative number. |
1831 | // Since 'ashr' of a negative number will result in a bigger |
1832 | // number, the Lower value of LHS is shifted right with the |
1833 | // minimum value of 'Other'. |
1834 | APInt NegMin = getSignedMin().ashr(ShiftAmt: Other.getUnsignedMin()); |
1835 | |
1836 | APInt max, min; |
1837 | if (getSignedMin().isNonNegative()) { |
1838 | // Upper and Lower of LHS are non-negative. |
1839 | min = PosMin; |
1840 | max = PosMax; |
1841 | } else if (getSignedMax().isNegative()) { |
1842 | // Upper and Lower of LHS are negative. |
1843 | min = NegMin; |
1844 | max = NegMax; |
1845 | } else { |
1846 | // Upper is non-negative and Lower is negative. |
1847 | min = NegMin; |
1848 | max = PosMax; |
1849 | } |
1850 | return getNonEmpty(Lower: std::move(min), Upper: std::move(max)); |
1851 | } |
1852 | |
1853 | ConstantRange ConstantRange::uadd_sat(const ConstantRange &Other) const { |
1854 | if (isEmptySet() || Other.isEmptySet()) |
1855 | return getEmpty(); |
1856 | |
1857 | APInt NewL = getUnsignedMin().uadd_sat(RHS: Other.getUnsignedMin()); |
1858 | APInt NewU = getUnsignedMax().uadd_sat(RHS: Other.getUnsignedMax()) + 1; |
1859 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1860 | } |
1861 | |
1862 | ConstantRange ConstantRange::sadd_sat(const ConstantRange &Other) const { |
1863 | if (isEmptySet() || Other.isEmptySet()) |
1864 | return getEmpty(); |
1865 | |
1866 | APInt NewL = getSignedMin().sadd_sat(RHS: Other.getSignedMin()); |
1867 | APInt NewU = getSignedMax().sadd_sat(RHS: Other.getSignedMax()) + 1; |
1868 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1869 | } |
1870 | |
1871 | ConstantRange ConstantRange::usub_sat(const ConstantRange &Other) const { |
1872 | if (isEmptySet() || Other.isEmptySet()) |
1873 | return getEmpty(); |
1874 | |
1875 | APInt NewL = getUnsignedMin().usub_sat(RHS: Other.getUnsignedMax()); |
1876 | APInt NewU = getUnsignedMax().usub_sat(RHS: Other.getUnsignedMin()) + 1; |
1877 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1878 | } |
1879 | |
1880 | ConstantRange ConstantRange::ssub_sat(const ConstantRange &Other) const { |
1881 | if (isEmptySet() || Other.isEmptySet()) |
1882 | return getEmpty(); |
1883 | |
1884 | APInt NewL = getSignedMin().ssub_sat(RHS: Other.getSignedMax()); |
1885 | APInt NewU = getSignedMax().ssub_sat(RHS: Other.getSignedMin()) + 1; |
1886 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1887 | } |
1888 | |
1889 | ConstantRange ConstantRange::umul_sat(const ConstantRange &Other) const { |
1890 | if (isEmptySet() || Other.isEmptySet()) |
1891 | return getEmpty(); |
1892 | |
1893 | APInt NewL = getUnsignedMin().umul_sat(RHS: Other.getUnsignedMin()); |
1894 | APInt NewU = getUnsignedMax().umul_sat(RHS: Other.getUnsignedMax()) + 1; |
1895 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1896 | } |
1897 | |
1898 | ConstantRange ConstantRange::smul_sat(const ConstantRange &Other) const { |
1899 | if (isEmptySet() || Other.isEmptySet()) |
1900 | return getEmpty(); |
1901 | |
1902 | // Because we could be dealing with negative numbers here, the lower bound is |
1903 | // the smallest of the cartesian product of the lower and upper ranges; |
1904 | // for example: |
1905 | // [-1,4) * [-2,3) = min(-1*-2, -1*2, 3*-2, 3*2) = -6. |
1906 | // Similarly for the upper bound, swapping min for max. |
1907 | |
1908 | APInt Min = getSignedMin(); |
1909 | APInt Max = getSignedMax(); |
1910 | APInt OtherMin = Other.getSignedMin(); |
1911 | APInt OtherMax = Other.getSignedMax(); |
1912 | |
1913 | auto L = {Min.smul_sat(RHS: OtherMin), Min.smul_sat(RHS: OtherMax), |
1914 | Max.smul_sat(RHS: OtherMin), Max.smul_sat(RHS: OtherMax)}; |
1915 | auto Compare = [](const APInt &A, const APInt &B) { return A.slt(RHS: B); }; |
1916 | return getNonEmpty(Lower: std::min(l: L, comp: Compare), Upper: std::max(l: L, comp: Compare) + 1); |
1917 | } |
1918 | |
1919 | ConstantRange ConstantRange::ushl_sat(const ConstantRange &Other) const { |
1920 | if (isEmptySet() || Other.isEmptySet()) |
1921 | return getEmpty(); |
1922 | |
1923 | APInt NewL = getUnsignedMin().ushl_sat(RHS: Other.getUnsignedMin()); |
1924 | APInt NewU = getUnsignedMax().ushl_sat(RHS: Other.getUnsignedMax()) + 1; |
1925 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1926 | } |
1927 | |
1928 | ConstantRange ConstantRange::sshl_sat(const ConstantRange &Other) const { |
1929 | if (isEmptySet() || Other.isEmptySet()) |
1930 | return getEmpty(); |
1931 | |
1932 | APInt Min = getSignedMin(), Max = getSignedMax(); |
1933 | APInt ShAmtMin = Other.getUnsignedMin(), ShAmtMax = Other.getUnsignedMax(); |
1934 | APInt NewL = Min.sshl_sat(RHS: Min.isNonNegative() ? ShAmtMin : ShAmtMax); |
1935 | APInt NewU = Max.sshl_sat(RHS: Max.isNegative() ? ShAmtMin : ShAmtMax) + 1; |
1936 | return getNonEmpty(Lower: std::move(NewL), Upper: std::move(NewU)); |
1937 | } |
1938 | |
1939 | ConstantRange ConstantRange::inverse() const { |
1940 | if (isFullSet()) |
1941 | return getEmpty(); |
1942 | if (isEmptySet()) |
1943 | return getFull(); |
1944 | return ConstantRange(Upper, Lower); |
1945 | } |
1946 | |
1947 | ConstantRange ConstantRange::abs(bool IntMinIsPoison) const { |
1948 | if (isEmptySet()) |
1949 | return getEmpty(); |
1950 | |
1951 | if (isSignWrappedSet()) { |
1952 | APInt Lo; |
1953 | // Check whether the range crosses zero. |
1954 | if (Upper.isStrictlyPositive() || !Lower.isStrictlyPositive()) |
1955 | Lo = APInt::getZero(numBits: getBitWidth()); |
1956 | else |
1957 | Lo = APIntOps::umin(A: Lower, B: -Upper + 1); |
1958 | |
1959 | // If SignedMin is not poison, then it is included in the result range. |
1960 | if (IntMinIsPoison) |
1961 | return ConstantRange(Lo, APInt::getSignedMinValue(numBits: getBitWidth())); |
1962 | else |
1963 | return ConstantRange(Lo, APInt::getSignedMinValue(numBits: getBitWidth()) + 1); |
1964 | } |
1965 | |
1966 | APInt SMin = getSignedMin(), SMax = getSignedMax(); |
1967 | |
1968 | // Skip SignedMin if it is poison. |
1969 | if (IntMinIsPoison && SMin.isMinSignedValue()) { |
1970 | // The range may become empty if it *only* contains SignedMin. |
1971 | if (SMax.isMinSignedValue()) |
1972 | return getEmpty(); |
1973 | ++SMin; |
1974 | } |
1975 | |
1976 | // All non-negative. |
1977 | if (SMin.isNonNegative()) |
1978 | return ConstantRange(SMin, SMax + 1); |
1979 | |
1980 | // All negative. |
1981 | if (SMax.isNegative()) |
1982 | return ConstantRange(-SMax, -SMin + 1); |
1983 | |
1984 | // Range crosses zero. |
1985 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: getBitWidth()), |
1986 | Upper: APIntOps::umax(A: -SMin, B: SMax) + 1); |
1987 | } |
1988 | |
1989 | ConstantRange ConstantRange::ctlz(bool ZeroIsPoison) const { |
1990 | if (isEmptySet()) |
1991 | return getEmpty(); |
1992 | |
1993 | APInt Zero = APInt::getZero(numBits: getBitWidth()); |
1994 | if (ZeroIsPoison && contains(V: Zero)) { |
1995 | // ZeroIsPoison is set, and zero is contained. We discern three cases, in |
1996 | // which a zero can appear: |
1997 | // 1) Lower is zero, handling cases of kind [0, 1), [0, 2), etc. |
1998 | // 2) Upper is zero, wrapped set, handling cases of kind [3, 0], etc. |
1999 | // 3) Zero contained in a wrapped set, e.g., [3, 2), [3, 1), etc. |
2000 | |
2001 | if (getLower().isZero()) { |
2002 | if ((getUpper() - 1).isZero()) { |
2003 | // We have in input interval of kind [0, 1). In this case we cannot |
2004 | // really help but return empty-set. |
2005 | return getEmpty(); |
2006 | } |
2007 | |
2008 | // Compute the resulting range by excluding zero from Lower. |
2009 | return ConstantRange( |
2010 | APInt(getBitWidth(), (getUpper() - 1).countl_zero()), |
2011 | APInt(getBitWidth(), (getLower() + 1).countl_zero() + 1)); |
2012 | } else if ((getUpper() - 1).isZero()) { |
2013 | // Compute the resulting range by excluding zero from Upper. |
2014 | return ConstantRange(Zero, |
2015 | APInt(getBitWidth(), getLower().countl_zero() + 1)); |
2016 | } else { |
2017 | return ConstantRange(Zero, APInt(getBitWidth(), getBitWidth())); |
2018 | } |
2019 | } |
2020 | |
2021 | // Zero is either safe or not in the range. The output range is composed by |
2022 | // the result of countLeadingZero of the two extremes. |
2023 | return getNonEmpty(Lower: APInt(getBitWidth(), getUnsignedMax().countl_zero()), |
2024 | Upper: APInt(getBitWidth(), getUnsignedMin().countl_zero()) + 1); |
2025 | } |
2026 | |
2027 | static ConstantRange getUnsignedCountTrailingZerosRange(const APInt &Lower, |
2028 | const APInt &Upper) { |
2029 | assert(!ConstantRange(Lower, Upper).isWrappedSet() && |
2030 | "Unexpected wrapped set." ); |
2031 | assert(Lower != Upper && "Unexpected empty set." ); |
2032 | unsigned BitWidth = Lower.getBitWidth(); |
2033 | if (Lower + 1 == Upper) |
2034 | return ConstantRange(APInt(BitWidth, Lower.countr_zero())); |
2035 | if (Lower.isZero()) |
2036 | return ConstantRange(APInt::getZero(numBits: BitWidth), |
2037 | APInt(BitWidth, BitWidth + 1)); |
2038 | |
2039 | // Calculate longest common prefix. |
2040 | unsigned LCPLength = (Lower ^ (Upper - 1)).countl_zero(); |
2041 | // If Lower is {LCP, 000...}, the maximum is Lower.countr_zero(). |
2042 | // Otherwise, the maximum is BitWidth - LCPLength - 1 ({LCP, 100...}). |
2043 | return ConstantRange( |
2044 | APInt::getZero(numBits: BitWidth), |
2045 | APInt(BitWidth, |
2046 | std::max(a: BitWidth - LCPLength - 1, b: Lower.countr_zero()) + 1)); |
2047 | } |
2048 | |
2049 | ConstantRange ConstantRange::cttz(bool ZeroIsPoison) const { |
2050 | if (isEmptySet()) |
2051 | return getEmpty(); |
2052 | |
2053 | unsigned BitWidth = getBitWidth(); |
2054 | APInt Zero = APInt::getZero(numBits: BitWidth); |
2055 | if (ZeroIsPoison && contains(V: Zero)) { |
2056 | // ZeroIsPoison is set, and zero is contained. We discern three cases, in |
2057 | // which a zero can appear: |
2058 | // 1) Lower is zero, handling cases of kind [0, 1), [0, 2), etc. |
2059 | // 2) Upper is zero, wrapped set, handling cases of kind [3, 0], etc. |
2060 | // 3) Zero contained in a wrapped set, e.g., [3, 2), [3, 1), etc. |
2061 | |
2062 | if (Lower.isZero()) { |
2063 | if (Upper == 1) { |
2064 | // We have in input interval of kind [0, 1). In this case we cannot |
2065 | // really help but return empty-set. |
2066 | return getEmpty(); |
2067 | } |
2068 | |
2069 | // Compute the resulting range by excluding zero from Lower. |
2070 | return getUnsignedCountTrailingZerosRange(Lower: APInt(BitWidth, 1), Upper); |
2071 | } else if (Upper == 1) { |
2072 | // Compute the resulting range by excluding zero from Upper. |
2073 | return getUnsignedCountTrailingZerosRange(Lower, Upper: Zero); |
2074 | } else { |
2075 | ConstantRange CR1 = getUnsignedCountTrailingZerosRange(Lower, Upper: Zero); |
2076 | ConstantRange CR2 = |
2077 | getUnsignedCountTrailingZerosRange(Lower: APInt(BitWidth, 1), Upper); |
2078 | return CR1.unionWith(CR: CR2); |
2079 | } |
2080 | } |
2081 | |
2082 | if (isFullSet()) |
2083 | return getNonEmpty(Lower: Zero, Upper: APInt(BitWidth, BitWidth) + 1); |
2084 | if (!isWrappedSet()) |
2085 | return getUnsignedCountTrailingZerosRange(Lower, Upper); |
2086 | // The range is wrapped. We decompose it into two ranges, [0, Upper) and |
2087 | // [Lower, 0). |
2088 | // Handle [Lower, 0) |
2089 | ConstantRange CR1 = getUnsignedCountTrailingZerosRange(Lower, Upper: Zero); |
2090 | // Handle [0, Upper) |
2091 | ConstantRange CR2 = getUnsignedCountTrailingZerosRange(Lower: Zero, Upper); |
2092 | return CR1.unionWith(CR: CR2); |
2093 | } |
2094 | |
2095 | static ConstantRange getUnsignedPopCountRange(const APInt &Lower, |
2096 | const APInt &Upper) { |
2097 | assert(!ConstantRange(Lower, Upper).isWrappedSet() && |
2098 | "Unexpected wrapped set." ); |
2099 | assert(Lower != Upper && "Unexpected empty set." ); |
2100 | unsigned BitWidth = Lower.getBitWidth(); |
2101 | if (Lower + 1 == Upper) |
2102 | return ConstantRange(APInt(BitWidth, Lower.popcount())); |
2103 | |
2104 | APInt Max = Upper - 1; |
2105 | // Calculate longest common prefix. |
2106 | unsigned LCPLength = (Lower ^ Max).countl_zero(); |
2107 | unsigned LCPPopCount = Lower.getHiBits(numBits: LCPLength).popcount(); |
2108 | // If Lower is {LCP, 000...}, the minimum is the popcount of LCP. |
2109 | // Otherwise, the minimum is the popcount of LCP + 1. |
2110 | unsigned MinBits = |
2111 | LCPPopCount + (Lower.countr_zero() < BitWidth - LCPLength ? 1 : 0); |
2112 | // If Max is {LCP, 111...}, the maximum is the popcount of LCP + (BitWidth - |
2113 | // length of LCP). |
2114 | // Otherwise, the minimum is the popcount of LCP + (BitWidth - |
2115 | // length of LCP - 1). |
2116 | unsigned MaxBits = LCPPopCount + (BitWidth - LCPLength) - |
2117 | (Max.countr_one() < BitWidth - LCPLength ? 1 : 0); |
2118 | return ConstantRange(APInt(BitWidth, MinBits), APInt(BitWidth, MaxBits + 1)); |
2119 | } |
2120 | |
2121 | ConstantRange ConstantRange::ctpop() const { |
2122 | if (isEmptySet()) |
2123 | return getEmpty(); |
2124 | |
2125 | unsigned BitWidth = getBitWidth(); |
2126 | APInt Zero = APInt::getZero(numBits: BitWidth); |
2127 | if (isFullSet()) |
2128 | return getNonEmpty(Lower: Zero, Upper: APInt(BitWidth, BitWidth) + 1); |
2129 | if (!isWrappedSet()) |
2130 | return getUnsignedPopCountRange(Lower, Upper); |
2131 | // The range is wrapped. We decompose it into two ranges, [0, Upper) and |
2132 | // [Lower, 0). |
2133 | // Handle [Lower, 0) == [Lower, Max] |
2134 | ConstantRange CR1 = ConstantRange(APInt(BitWidth, Lower.countl_one()), |
2135 | APInt(BitWidth, BitWidth + 1)); |
2136 | // Handle [0, Upper) |
2137 | ConstantRange CR2 = getUnsignedPopCountRange(Lower: Zero, Upper); |
2138 | return CR1.unionWith(CR: CR2); |
2139 | } |
2140 | |
2141 | ConstantRange::OverflowResult ConstantRange::unsignedAddMayOverflow( |
2142 | const ConstantRange &Other) const { |
2143 | if (isEmptySet() || Other.isEmptySet()) |
2144 | return OverflowResult::MayOverflow; |
2145 | |
2146 | APInt Min = getUnsignedMin(), Max = getUnsignedMax(); |
2147 | APInt OtherMin = Other.getUnsignedMin(), OtherMax = Other.getUnsignedMax(); |
2148 | |
2149 | // a u+ b overflows high iff a u> ~b. |
2150 | if (Min.ugt(RHS: ~OtherMin)) |
2151 | return OverflowResult::AlwaysOverflowsHigh; |
2152 | if (Max.ugt(RHS: ~OtherMax)) |
2153 | return OverflowResult::MayOverflow; |
2154 | return OverflowResult::NeverOverflows; |
2155 | } |
2156 | |
2157 | ConstantRange::OverflowResult ConstantRange::signedAddMayOverflow( |
2158 | const ConstantRange &Other) const { |
2159 | if (isEmptySet() || Other.isEmptySet()) |
2160 | return OverflowResult::MayOverflow; |
2161 | |
2162 | APInt Min = getSignedMin(), Max = getSignedMax(); |
2163 | APInt OtherMin = Other.getSignedMin(), OtherMax = Other.getSignedMax(); |
2164 | |
2165 | APInt SignedMin = APInt::getSignedMinValue(numBits: getBitWidth()); |
2166 | APInt SignedMax = APInt::getSignedMaxValue(numBits: getBitWidth()); |
2167 | |
2168 | // a s+ b overflows high iff a s>=0 && b s>= 0 && a s> smax - b. |
2169 | // a s+ b overflows low iff a s< 0 && b s< 0 && a s< smin - b. |
2170 | if (Min.isNonNegative() && OtherMin.isNonNegative() && |
2171 | Min.sgt(RHS: SignedMax - OtherMin)) |
2172 | return OverflowResult::AlwaysOverflowsHigh; |
2173 | if (Max.isNegative() && OtherMax.isNegative() && |
2174 | Max.slt(RHS: SignedMin - OtherMax)) |
2175 | return OverflowResult::AlwaysOverflowsLow; |
2176 | |
2177 | if (Max.isNonNegative() && OtherMax.isNonNegative() && |
2178 | Max.sgt(RHS: SignedMax - OtherMax)) |
2179 | return OverflowResult::MayOverflow; |
2180 | if (Min.isNegative() && OtherMin.isNegative() && |
2181 | Min.slt(RHS: SignedMin - OtherMin)) |
2182 | return OverflowResult::MayOverflow; |
2183 | |
2184 | return OverflowResult::NeverOverflows; |
2185 | } |
2186 | |
2187 | ConstantRange::OverflowResult ConstantRange::unsignedSubMayOverflow( |
2188 | const ConstantRange &Other) const { |
2189 | if (isEmptySet() || Other.isEmptySet()) |
2190 | return OverflowResult::MayOverflow; |
2191 | |
2192 | APInt Min = getUnsignedMin(), Max = getUnsignedMax(); |
2193 | APInt OtherMin = Other.getUnsignedMin(), OtherMax = Other.getUnsignedMax(); |
2194 | |
2195 | // a u- b overflows low iff a u< b. |
2196 | if (Max.ult(RHS: OtherMin)) |
2197 | return OverflowResult::AlwaysOverflowsLow; |
2198 | if (Min.ult(RHS: OtherMax)) |
2199 | return OverflowResult::MayOverflow; |
2200 | return OverflowResult::NeverOverflows; |
2201 | } |
2202 | |
2203 | ConstantRange::OverflowResult ConstantRange::signedSubMayOverflow( |
2204 | const ConstantRange &Other) const { |
2205 | if (isEmptySet() || Other.isEmptySet()) |
2206 | return OverflowResult::MayOverflow; |
2207 | |
2208 | APInt Min = getSignedMin(), Max = getSignedMax(); |
2209 | APInt OtherMin = Other.getSignedMin(), OtherMax = Other.getSignedMax(); |
2210 | |
2211 | APInt SignedMin = APInt::getSignedMinValue(numBits: getBitWidth()); |
2212 | APInt SignedMax = APInt::getSignedMaxValue(numBits: getBitWidth()); |
2213 | |
2214 | // a s- b overflows high iff a s>=0 && b s< 0 && a s> smax + b. |
2215 | // a s- b overflows low iff a s< 0 && b s>= 0 && a s< smin + b. |
2216 | if (Min.isNonNegative() && OtherMax.isNegative() && |
2217 | Min.sgt(RHS: SignedMax + OtherMax)) |
2218 | return OverflowResult::AlwaysOverflowsHigh; |
2219 | if (Max.isNegative() && OtherMin.isNonNegative() && |
2220 | Max.slt(RHS: SignedMin + OtherMin)) |
2221 | return OverflowResult::AlwaysOverflowsLow; |
2222 | |
2223 | if (Max.isNonNegative() && OtherMin.isNegative() && |
2224 | Max.sgt(RHS: SignedMax + OtherMin)) |
2225 | return OverflowResult::MayOverflow; |
2226 | if (Min.isNegative() && OtherMax.isNonNegative() && |
2227 | Min.slt(RHS: SignedMin + OtherMax)) |
2228 | return OverflowResult::MayOverflow; |
2229 | |
2230 | return OverflowResult::NeverOverflows; |
2231 | } |
2232 | |
2233 | ConstantRange::OverflowResult ConstantRange::unsignedMulMayOverflow( |
2234 | const ConstantRange &Other) const { |
2235 | if (isEmptySet() || Other.isEmptySet()) |
2236 | return OverflowResult::MayOverflow; |
2237 | |
2238 | APInt Min = getUnsignedMin(), Max = getUnsignedMax(); |
2239 | APInt OtherMin = Other.getUnsignedMin(), OtherMax = Other.getUnsignedMax(); |
2240 | bool Overflow; |
2241 | |
2242 | (void) Min.umul_ov(RHS: OtherMin, Overflow); |
2243 | if (Overflow) |
2244 | return OverflowResult::AlwaysOverflowsHigh; |
2245 | |
2246 | (void) Max.umul_ov(RHS: OtherMax, Overflow); |
2247 | if (Overflow) |
2248 | return OverflowResult::MayOverflow; |
2249 | |
2250 | return OverflowResult::NeverOverflows; |
2251 | } |
2252 | |
2253 | void ConstantRange::print(raw_ostream &OS) const { |
2254 | if (isFullSet()) |
2255 | OS << "full-set" ; |
2256 | else if (isEmptySet()) |
2257 | OS << "empty-set" ; |
2258 | else |
2259 | OS << "[" << Lower << "," << Upper << ")" ; |
2260 | } |
2261 | |
2262 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
2263 | LLVM_DUMP_METHOD void ConstantRange::dump() const { |
2264 | print(dbgs()); |
2265 | } |
2266 | #endif |
2267 | |
2268 | ConstantRange llvm::getConstantRangeFromMetadata(const MDNode &Ranges) { |
2269 | const unsigned NumRanges = Ranges.getNumOperands() / 2; |
2270 | assert(NumRanges >= 1 && "Must have at least one range!" ); |
2271 | assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs" ); |
2272 | |
2273 | auto *FirstLow = mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 0)); |
2274 | auto *FirstHigh = mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 1)); |
2275 | |
2276 | ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue()); |
2277 | |
2278 | for (unsigned i = 1; i < NumRanges; ++i) { |
2279 | auto *Low = mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 2 * i + 0)); |
2280 | auto *High = mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 2 * i + 1)); |
2281 | |
2282 | // Note: unionWith will potentially create a range that contains values not |
2283 | // contained in any of the original N ranges. |
2284 | CR = CR.unionWith(CR: ConstantRange(Low->getValue(), High->getValue())); |
2285 | } |
2286 | |
2287 | return CR; |
2288 | } |
2289 | |