Bug Summary

File:llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
Warning:line 3556, column 7
Called C++ object pointer is uninitialized

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name InstCombineCompares.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-12/lib/clang/12.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-08-06-171148-17323-1 -x c++ /build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

1//===- InstCombineCompares.cpp --------------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visitICmp and visitFCmp functions.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APSInt.h"
15#include "llvm/ADT/SetVector.h"
16#include "llvm/ADT/Statistic.h"
17#include "llvm/Analysis/ConstantFolding.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/TargetLibraryInfo.h"
20#include "llvm/IR/ConstantRange.h"
21#include "llvm/IR/DataLayout.h"
22#include "llvm/IR/GetElementPtrTypeIterator.h"
23#include "llvm/IR/IntrinsicInst.h"
24#include "llvm/IR/PatternMatch.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/KnownBits.h"
27#include "llvm/Transforms/InstCombine/InstCombiner.h"
28
29using namespace llvm;
30using namespace PatternMatch;
31
32#define DEBUG_TYPE"instcombine" "instcombine"
33
34// How many times is a select replaced by one of its operands?
35STATISTIC(NumSel, "Number of select opts")static llvm::Statistic NumSel = {"instcombine", "NumSel", "Number of select opts"
}
;
36
37
38/// Compute Result = In1+In2, returning true if the result overflowed for this
39/// type.
40static bool addWithOverflow(APInt &Result, const APInt &In1,
41 const APInt &In2, bool IsSigned = false) {
42 bool Overflow;
43 if (IsSigned)
44 Result = In1.sadd_ov(In2, Overflow);
45 else
46 Result = In1.uadd_ov(In2, Overflow);
47
48 return Overflow;
49}
50
51/// Compute Result = In1-In2, returning true if the result overflowed for this
52/// type.
53static bool subWithOverflow(APInt &Result, const APInt &In1,
54 const APInt &In2, bool IsSigned = false) {
55 bool Overflow;
56 if (IsSigned)
57 Result = In1.ssub_ov(In2, Overflow);
58 else
59 Result = In1.usub_ov(In2, Overflow);
60
61 return Overflow;
62}
63
64/// Given an icmp instruction, return true if any use of this comparison is a
65/// branch on sign bit comparison.
66static bool hasBranchUse(ICmpInst &I) {
67 for (auto *U : I.users())
68 if (isa<BranchInst>(U))
69 return true;
70 return false;
71}
72
73/// Returns true if the exploded icmp can be expressed as a signed comparison
74/// to zero and updates the predicate accordingly.
75/// The signedness of the comparison is preserved.
76/// TODO: Refactor with decomposeBitTestICmp()?
77static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
78 if (!ICmpInst::isSigned(Pred))
79 return false;
80
81 if (C.isNullValue())
82 return ICmpInst::isRelational(Pred);
83
84 if (C.isOneValue()) {
85 if (Pred == ICmpInst::ICMP_SLT) {
86 Pred = ICmpInst::ICMP_SLE;
87 return true;
88 }
89 } else if (C.isAllOnesValue()) {
90 if (Pred == ICmpInst::ICMP_SGT) {
91 Pred = ICmpInst::ICMP_SGE;
92 return true;
93 }
94 }
95
96 return false;
97}
98
99/// Given a signed integer type and a set of known zero and one bits, compute
100/// the maximum and minimum values that could have the specified known zero and
101/// known one bits, returning them in Min/Max.
102/// TODO: Move to method on KnownBits struct?
103static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
104 APInt &Min, APInt &Max) {
105 assert(Known.getBitWidth() == Min.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 107, __PRETTY_FUNCTION__))
106 Known.getBitWidth() == Max.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 107, __PRETTY_FUNCTION__))
107 "KnownZero, KnownOne and Min, Max must have equal bitwidth.")((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 107, __PRETTY_FUNCTION__))
;
108 APInt UnknownBits = ~(Known.Zero|Known.One);
109
110 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
111 // bit if it is unknown.
112 Min = Known.One;
113 Max = Known.One|UnknownBits;
114
115 if (UnknownBits.isNegative()) { // Sign bit is unknown
116 Min.setSignBit();
117 Max.clearSignBit();
118 }
119}
120
121/// Given an unsigned integer type and a set of known zero and one bits, compute
122/// the maximum and minimum values that could have the specified known zero and
123/// known one bits, returning them in Min/Max.
124/// TODO: Move to method on KnownBits struct?
125static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
126 APInt &Min, APInt &Max) {
127 assert(Known.getBitWidth() == Min.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 129, __PRETTY_FUNCTION__))
128 Known.getBitWidth() == Max.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 129, __PRETTY_FUNCTION__))
129 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.")((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 129, __PRETTY_FUNCTION__))
;
130 APInt UnknownBits = ~(Known.Zero|Known.One);
131
132 // The minimum value is when the unknown bits are all zeros.
133 Min = Known.One;
134 // The maximum value is when the unknown bits are all ones.
135 Max = Known.One|UnknownBits;
136}
137
138/// This is called when we see this pattern:
139/// cmp pred (load (gep GV, ...)), cmpcst
140/// where GV is a global variable with a constant initializer. Try to simplify
141/// this into some simple computation that does not need the load. For example
142/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
143///
144/// If AndCst is non-null, then the loaded value is masked with that constant
145/// before doing the comparison. This handles cases like "A[i]&4 == 0".
146Instruction *
147InstCombinerImpl::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
148 GlobalVariable *GV, CmpInst &ICI,
149 ConstantInt *AndCst) {
150 Constant *Init = GV->getInitializer();
151 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
152 return nullptr;
153
154 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
155 // Don't blow up on huge arrays.
156 if (ArrayElementCount > MaxArraySizeForCombine)
157 return nullptr;
158
159 // There are many forms of this optimization we can handle, for now, just do
160 // the simple index into a single-dimensional array.
161 //
162 // Require: GEP GV, 0, i {{, constant indices}}
163 if (GEP->getNumOperands() < 3 ||
164 !isa<ConstantInt>(GEP->getOperand(1)) ||
165 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
166 isa<Constant>(GEP->getOperand(2)))
167 return nullptr;
168
169 // Check that indices after the variable are constants and in-range for the
170 // type they index. Collect the indices. This is typically for arrays of
171 // structs.
172 SmallVector<unsigned, 4> LaterIndices;
173
174 Type *EltTy = Init->getType()->getArrayElementType();
175 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
176 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
177 if (!Idx) return nullptr; // Variable index.
178
179 uint64_t IdxVal = Idx->getZExtValue();
180 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
181
182 if (StructType *STy = dyn_cast<StructType>(EltTy))
183 EltTy = STy->getElementType(IdxVal);
184 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
185 if (IdxVal >= ATy->getNumElements()) return nullptr;
186 EltTy = ATy->getElementType();
187 } else {
188 return nullptr; // Unknown type.
189 }
190
191 LaterIndices.push_back(IdxVal);
192 }
193
194 enum { Overdefined = -3, Undefined = -2 };
195
196 // Variables for our state machines.
197
198 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
199 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
200 // and 87 is the second (and last) index. FirstTrueElement is -2 when
201 // undefined, otherwise set to the first true element. SecondTrueElement is
202 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
203 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
204
205 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
206 // form "i != 47 & i != 87". Same state transitions as for true elements.
207 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
208
209 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
210 /// define a state machine that triggers for ranges of values that the index
211 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
212 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
213 /// index in the range (inclusive). We use -2 for undefined here because we
214 /// use relative comparisons and don't want 0-1 to match -1.
215 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
216
217 // MagicBitvector - This is a magic bitvector where we set a bit if the
218 // comparison is true for element 'i'. If there are 64 elements or less in
219 // the array, this will fully represent all the comparison results.
220 uint64_t MagicBitvector = 0;
221
222 // Scan the array and see if one of our patterns matches.
223 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
224 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
225 Constant *Elt = Init->getAggregateElement(i);
226 if (!Elt) return nullptr;
227
228 // If this is indexing an array of structures, get the structure element.
229 if (!LaterIndices.empty())
230 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
231
232 // If the element is masked, handle it.
233 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
234
235 // Find out if the comparison would be true or false for the i'th element.
236 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
237 CompareRHS, DL, &TLI);
238 // If the result is undef for this element, ignore it.
239 if (isa<UndefValue>(C)) {
240 // Extend range state machines to cover this element in case there is an
241 // undef in the middle of the range.
242 if (TrueRangeEnd == (int)i-1)
243 TrueRangeEnd = i;
244 if (FalseRangeEnd == (int)i-1)
245 FalseRangeEnd = i;
246 continue;
247 }
248
249 // If we can't compute the result for any of the elements, we have to give
250 // up evaluating the entire conditional.
251 if (!isa<ConstantInt>(C)) return nullptr;
252
253 // Otherwise, we know if the comparison is true or false for this element,
254 // update our state machines.
255 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
256
257 // State machine for single/double/range index comparison.
258 if (IsTrueForElt) {
259 // Update the TrueElement state machine.
260 if (FirstTrueElement == Undefined)
261 FirstTrueElement = TrueRangeEnd = i; // First true element.
262 else {
263 // Update double-compare state machine.
264 if (SecondTrueElement == Undefined)
265 SecondTrueElement = i;
266 else
267 SecondTrueElement = Overdefined;
268
269 // Update range state machine.
270 if (TrueRangeEnd == (int)i-1)
271 TrueRangeEnd = i;
272 else
273 TrueRangeEnd = Overdefined;
274 }
275 } else {
276 // Update the FalseElement state machine.
277 if (FirstFalseElement == Undefined)
278 FirstFalseElement = FalseRangeEnd = i; // First false element.
279 else {
280 // Update double-compare state machine.
281 if (SecondFalseElement == Undefined)
282 SecondFalseElement = i;
283 else
284 SecondFalseElement = Overdefined;
285
286 // Update range state machine.
287 if (FalseRangeEnd == (int)i-1)
288 FalseRangeEnd = i;
289 else
290 FalseRangeEnd = Overdefined;
291 }
292 }
293
294 // If this element is in range, update our magic bitvector.
295 if (i < 64 && IsTrueForElt)
296 MagicBitvector |= 1ULL << i;
297
298 // If all of our states become overdefined, bail out early. Since the
299 // predicate is expensive, only check it every 8 elements. This is only
300 // really useful for really huge arrays.
301 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
302 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
303 FalseRangeEnd == Overdefined)
304 return nullptr;
305 }
306
307 // Now that we've scanned the entire array, emit our new comparison(s). We
308 // order the state machines in complexity of the generated code.
309 Value *Idx = GEP->getOperand(2);
310
311 // If the index is larger than the pointer size of the target, truncate the
312 // index down like the GEP would do implicitly. We don't have to do this for
313 // an inbounds GEP because the index can't be out of range.
314 if (!GEP->isInBounds()) {
315 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
316 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
317 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
318 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
319 }
320
321 // If the comparison is only true for one or two elements, emit direct
322 // comparisons.
323 if (SecondTrueElement != Overdefined) {
324 // None true -> false.
325 if (FirstTrueElement == Undefined)
326 return replaceInstUsesWith(ICI, Builder.getFalse());
327
328 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
329
330 // True for one element -> 'i == 47'.
331 if (SecondTrueElement == Undefined)
332 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
333
334 // True for two elements -> 'i == 47 | i == 72'.
335 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
336 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
337 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
338 return BinaryOperator::CreateOr(C1, C2);
339 }
340
341 // If the comparison is only false for one or two elements, emit direct
342 // comparisons.
343 if (SecondFalseElement != Overdefined) {
344 // None false -> true.
345 if (FirstFalseElement == Undefined)
346 return replaceInstUsesWith(ICI, Builder.getTrue());
347
348 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
349
350 // False for one element -> 'i != 47'.
351 if (SecondFalseElement == Undefined)
352 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
353
354 // False for two elements -> 'i != 47 & i != 72'.
355 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
356 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
357 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
358 return BinaryOperator::CreateAnd(C1, C2);
359 }
360
361 // If the comparison can be replaced with a range comparison for the elements
362 // where it is true, emit the range check.
363 if (TrueRangeEnd != Overdefined) {
364 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare")((TrueRangeEnd != FirstTrueElement && "Should emit single compare"
) ? static_cast<void> (0) : __assert_fail ("TrueRangeEnd != FirstTrueElement && \"Should emit single compare\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 364, __PRETTY_FUNCTION__))
;
365
366 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
367 if (FirstTrueElement) {
368 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
369 Idx = Builder.CreateAdd(Idx, Offs);
370 }
371
372 Value *End = ConstantInt::get(Idx->getType(),
373 TrueRangeEnd-FirstTrueElement+1);
374 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
375 }
376
377 // False range check.
378 if (FalseRangeEnd != Overdefined) {
379 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare")((FalseRangeEnd != FirstFalseElement && "Should emit single compare"
) ? static_cast<void> (0) : __assert_fail ("FalseRangeEnd != FirstFalseElement && \"Should emit single compare\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 379, __PRETTY_FUNCTION__))
;
380 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
381 if (FirstFalseElement) {
382 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
383 Idx = Builder.CreateAdd(Idx, Offs);
384 }
385
386 Value *End = ConstantInt::get(Idx->getType(),
387 FalseRangeEnd-FirstFalseElement);
388 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
389 }
390
391 // If a magic bitvector captures the entire comparison state
392 // of this load, replace it with computation that does:
393 // ((magic_cst >> i) & 1) != 0
394 {
395 Type *Ty = nullptr;
396
397 // Look for an appropriate type:
398 // - The type of Idx if the magic fits
399 // - The smallest fitting legal type
400 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
401 Ty = Idx->getType();
402 else
403 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
404
405 if (Ty) {
406 Value *V = Builder.CreateIntCast(Idx, Ty, false);
407 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
408 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
409 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
410 }
411 }
412
413 return nullptr;
414}
415
416/// Return a value that can be used to compare the *offset* implied by a GEP to
417/// zero. For example, if we have &A[i], we want to return 'i' for
418/// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
419/// are involved. The above expression would also be legal to codegen as
420/// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
421/// This latter form is less amenable to optimization though, and we are allowed
422/// to generate the first by knowing that pointer arithmetic doesn't overflow.
423///
424/// If we can't emit an optimized form for this expression, this returns null.
425///
426static Value *evaluateGEPOffsetExpression(User *GEP, InstCombinerImpl &IC,
427 const DataLayout &DL) {
428 gep_type_iterator GTI = gep_type_begin(GEP);
429
430 // Check to see if this gep only has a single variable index. If so, and if
431 // any constant indices are a multiple of its scale, then we can compute this
432 // in terms of the scale of the variable index. For example, if the GEP
433 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
434 // because the expression will cross zero at the same point.
435 unsigned i, e = GEP->getNumOperands();
436 int64_t Offset = 0;
437 for (i = 1; i != e; ++i, ++GTI) {
438 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
439 // Compute the aggregate offset of constant indices.
440 if (CI->isZero()) continue;
441
442 // Handle a struct index, which adds its field offset to the pointer.
443 if (StructType *STy = GTI.getStructTypeOrNull()) {
444 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
445 } else {
446 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
447 Offset += Size*CI->getSExtValue();
448 }
449 } else {
450 // Found our variable index.
451 break;
452 }
453 }
454
455 // If there are no variable indices, we must have a constant offset, just
456 // evaluate it the general way.
457 if (i == e) return nullptr;
458
459 Value *VariableIdx = GEP->getOperand(i);
460 // Determine the scale factor of the variable element. For example, this is
461 // 4 if the variable index is into an array of i32.
462 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
463
464 // Verify that there are no other variable indices. If so, emit the hard way.
465 for (++i, ++GTI; i != e; ++i, ++GTI) {
466 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
467 if (!CI) return nullptr;
468
469 // Compute the aggregate offset of constant indices.
470 if (CI->isZero()) continue;
471
472 // Handle a struct index, which adds its field offset to the pointer.
473 if (StructType *STy = GTI.getStructTypeOrNull()) {
474 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
475 } else {
476 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
477 Offset += Size*CI->getSExtValue();
478 }
479 }
480
481 // Okay, we know we have a single variable index, which must be a
482 // pointer/array/vector index. If there is no offset, life is simple, return
483 // the index.
484 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
485 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
486 if (Offset == 0) {
487 // Cast to intptrty in case a truncation occurs. If an extension is needed,
488 // we don't need to bother extending: the extension won't affect where the
489 // computation crosses zero.
490 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
491 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
492 }
493 return VariableIdx;
494 }
495
496 // Otherwise, there is an index. The computation we will do will be modulo
497 // the pointer size.
498 Offset = SignExtend64(Offset, IntPtrWidth);
499 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
500
501 // To do this transformation, any constant index must be a multiple of the
502 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
503 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
504 // multiple of the variable scale.
505 int64_t NewOffs = Offset / (int64_t)VariableScale;
506 if (Offset != NewOffs*(int64_t)VariableScale)
507 return nullptr;
508
509 // Okay, we can do this evaluation. Start by converting the index to intptr.
510 if (VariableIdx->getType() != IntPtrTy)
511 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
512 true /*Signed*/);
513 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
514 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
515}
516
517/// Returns true if we can rewrite Start as a GEP with pointer Base
518/// and some integer offset. The nodes that need to be re-written
519/// for this transformation will be added to Explored.
520static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
521 const DataLayout &DL,
522 SetVector<Value *> &Explored) {
523 SmallVector<Value *, 16> WorkList(1, Start);
524 Explored.insert(Base);
525
526 // The following traversal gives us an order which can be used
527 // when doing the final transformation. Since in the final
528 // transformation we create the PHI replacement instructions first,
529 // we don't have to get them in any particular order.
530 //
531 // However, for other instructions we will have to traverse the
532 // operands of an instruction first, which means that we have to
533 // do a post-order traversal.
534 while (!WorkList.empty()) {
535 SetVector<PHINode *> PHIs;
536
537 while (!WorkList.empty()) {
538 if (Explored.size() >= 100)
539 return false;
540
541 Value *V = WorkList.back();
542
543 if (Explored.count(V) != 0) {
544 WorkList.pop_back();
545 continue;
546 }
547
548 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
549 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
550 // We've found some value that we can't explore which is different from
551 // the base. Therefore we can't do this transformation.
552 return false;
553
554 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
555 auto *CI = dyn_cast<CastInst>(V);
556 if (!CI->isNoopCast(DL))
557 return false;
558
559 if (Explored.count(CI->getOperand(0)) == 0)
560 WorkList.push_back(CI->getOperand(0));
561 }
562
563 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
564 // We're limiting the GEP to having one index. This will preserve
565 // the original pointer type. We could handle more cases in the
566 // future.
567 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
568 GEP->getType() != Start->getType())
569 return false;
570
571 if (Explored.count(GEP->getOperand(0)) == 0)
572 WorkList.push_back(GEP->getOperand(0));
573 }
574
575 if (WorkList.back() == V) {
576 WorkList.pop_back();
577 // We've finished visiting this node, mark it as such.
578 Explored.insert(V);
579 }
580
581 if (auto *PN = dyn_cast<PHINode>(V)) {
582 // We cannot transform PHIs on unsplittable basic blocks.
583 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
584 return false;
585 Explored.insert(PN);
586 PHIs.insert(PN);
587 }
588 }
589
590 // Explore the PHI nodes further.
591 for (auto *PN : PHIs)
592 for (Value *Op : PN->incoming_values())
593 if (Explored.count(Op) == 0)
594 WorkList.push_back(Op);
595 }
596
597 // Make sure that we can do this. Since we can't insert GEPs in a basic
598 // block before a PHI node, we can't easily do this transformation if
599 // we have PHI node users of transformed instructions.
600 for (Value *Val : Explored) {
601 for (Value *Use : Val->uses()) {
602
603 auto *PHI = dyn_cast<PHINode>(Use);
604 auto *Inst = dyn_cast<Instruction>(Val);
605
606 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
607 Explored.count(PHI) == 0)
608 continue;
609
610 if (PHI->getParent() == Inst->getParent())
611 return false;
612 }
613 }
614 return true;
615}
616
617// Sets the appropriate insert point on Builder where we can add
618// a replacement Instruction for V (if that is possible).
619static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
620 bool Before = true) {
621 if (auto *PHI = dyn_cast<PHINode>(V)) {
622 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
623 return;
624 }
625 if (auto *I = dyn_cast<Instruction>(V)) {
626 if (!Before)
627 I = &*std::next(I->getIterator());
628 Builder.SetInsertPoint(I);
629 return;
630 }
631 if (auto *A = dyn_cast<Argument>(V)) {
632 // Set the insertion point in the entry block.
633 BasicBlock &Entry = A->getParent()->getEntryBlock();
634 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
635 return;
636 }
637 // Otherwise, this is a constant and we don't need to set a new
638 // insertion point.
639 assert(isa<Constant>(V) && "Setting insertion point for unknown value!")((isa<Constant>(V) && "Setting insertion point for unknown value!"
) ? static_cast<void> (0) : __assert_fail ("isa<Constant>(V) && \"Setting insertion point for unknown value!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 639, __PRETTY_FUNCTION__))
;
640}
641
642/// Returns a re-written value of Start as an indexed GEP using Base as a
643/// pointer.
644static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
645 const DataLayout &DL,
646 SetVector<Value *> &Explored) {
647 // Perform all the substitutions. This is a bit tricky because we can
648 // have cycles in our use-def chains.
649 // 1. Create the PHI nodes without any incoming values.
650 // 2. Create all the other values.
651 // 3. Add the edges for the PHI nodes.
652 // 4. Emit GEPs to get the original pointers.
653 // 5. Remove the original instructions.
654 Type *IndexType = IntegerType::get(
655 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
656
657 DenseMap<Value *, Value *> NewInsts;
658 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
659
660 // Create the new PHI nodes, without adding any incoming values.
661 for (Value *Val : Explored) {
662 if (Val == Base)
663 continue;
664 // Create empty phi nodes. This avoids cyclic dependencies when creating
665 // the remaining instructions.
666 if (auto *PHI = dyn_cast<PHINode>(Val))
667 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
668 PHI->getName() + ".idx", PHI);
669 }
670 IRBuilder<> Builder(Base->getContext());
671
672 // Create all the other instructions.
673 for (Value *Val : Explored) {
674
675 if (NewInsts.find(Val) != NewInsts.end())
676 continue;
677
678 if (auto *CI = dyn_cast<CastInst>(Val)) {
679 // Don't get rid of the intermediate variable here; the store can grow
680 // the map which will invalidate the reference to the input value.
681 Value *V = NewInsts[CI->getOperand(0)];
682 NewInsts[CI] = V;
683 continue;
684 }
685 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
686 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
687 : GEP->getOperand(1);
688 setInsertionPoint(Builder, GEP);
689 // Indices might need to be sign extended. GEPs will magically do
690 // this, but we need to do it ourselves here.
691 if (Index->getType()->getScalarSizeInBits() !=
692 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
693 Index = Builder.CreateSExtOrTrunc(
694 Index, NewInsts[GEP->getOperand(0)]->getType(),
695 GEP->getOperand(0)->getName() + ".sext");
696 }
697
698 auto *Op = NewInsts[GEP->getOperand(0)];
699 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
700 NewInsts[GEP] = Index;
701 else
702 NewInsts[GEP] = Builder.CreateNSWAdd(
703 Op, Index, GEP->getOperand(0)->getName() + ".add");
704 continue;
705 }
706 if (isa<PHINode>(Val))
707 continue;
708
709 llvm_unreachable("Unexpected instruction type")::llvm::llvm_unreachable_internal("Unexpected instruction type"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 709)
;
710 }
711
712 // Add the incoming values to the PHI nodes.
713 for (Value *Val : Explored) {
714 if (Val == Base)
715 continue;
716 // All the instructions have been created, we can now add edges to the
717 // phi nodes.
718 if (auto *PHI = dyn_cast<PHINode>(Val)) {
719 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
720 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
721 Value *NewIncoming = PHI->getIncomingValue(I);
722
723 if (NewInsts.find(NewIncoming) != NewInsts.end())
724 NewIncoming = NewInsts[NewIncoming];
725
726 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
727 }
728 }
729 }
730
731 for (Value *Val : Explored) {
732 if (Val == Base)
733 continue;
734
735 // Depending on the type, for external users we have to emit
736 // a GEP or a GEP + ptrtoint.
737 setInsertionPoint(Builder, Val, false);
738
739 // If required, create an inttoptr instruction for Base.
740 Value *NewBase = Base;
741 if (!Base->getType()->isPointerTy())
742 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
743 Start->getName() + "to.ptr");
744
745 Value *GEP = Builder.CreateInBoundsGEP(
746 Start->getType()->getPointerElementType(), NewBase,
747 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
748
749 if (!Val->getType()->isPointerTy()) {
750 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
751 Val->getName() + ".conv");
752 GEP = Cast;
753 }
754 Val->replaceAllUsesWith(GEP);
755 }
756
757 return NewInsts[Start];
758}
759
760/// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
761/// the input Value as a constant indexed GEP. Returns a pair containing
762/// the GEPs Pointer and Index.
763static std::pair<Value *, Value *>
764getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
765 Type *IndexType = IntegerType::get(V->getContext(),
766 DL.getIndexTypeSizeInBits(V->getType()));
767
768 Constant *Index = ConstantInt::getNullValue(IndexType);
769 while (true) {
770 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
771 // We accept only inbouds GEPs here to exclude the possibility of
772 // overflow.
773 if (!GEP->isInBounds())
774 break;
775 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
776 GEP->getType() == V->getType()) {
777 V = GEP->getOperand(0);
778 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
779 Index = ConstantExpr::getAdd(
780 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
781 continue;
782 }
783 break;
784 }
785 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
786 if (!CI->isNoopCast(DL))
787 break;
788 V = CI->getOperand(0);
789 continue;
790 }
791 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
792 if (!CI->isNoopCast(DL))
793 break;
794 V = CI->getOperand(0);
795 continue;
796 }
797 break;
798 }
799 return {V, Index};
800}
801
802/// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
803/// We can look through PHIs, GEPs and casts in order to determine a common base
804/// between GEPLHS and RHS.
805static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
806 ICmpInst::Predicate Cond,
807 const DataLayout &DL) {
808 // FIXME: Support vector of pointers.
809 if (GEPLHS->getType()->isVectorTy())
810 return nullptr;
811
812 if (!GEPLHS->hasAllConstantIndices())
813 return nullptr;
814
815 // Make sure the pointers have the same type.
816 if (GEPLHS->getType() != RHS->getType())
817 return nullptr;
818
819 Value *PtrBase, *Index;
820 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
821
822 // The set of nodes that will take part in this transformation.
823 SetVector<Value *> Nodes;
824
825 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
826 return nullptr;
827
828 // We know we can re-write this as
829 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
830 // Since we've only looked through inbouds GEPs we know that we
831 // can't have overflow on either side. We can therefore re-write
832 // this as:
833 // OFFSET1 cmp OFFSET2
834 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
835
836 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
837 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
838 // offset. Since Index is the offset of LHS to the base pointer, we will now
839 // compare the offsets instead of comparing the pointers.
840 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
841}
842
843/// Fold comparisons between a GEP instruction and something else. At this point
844/// we know that the GEP is on the LHS of the comparison.
845Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
846 ICmpInst::Predicate Cond,
847 Instruction &I) {
848 // Don't transform signed compares of GEPs into index compares. Even if the
849 // GEP is inbounds, the final add of the base pointer can have signed overflow
850 // and would change the result of the icmp.
851 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
852 // the maximum signed value for the pointer type.
853 if (ICmpInst::isSigned(Cond))
854 return nullptr;
855
856 // Look through bitcasts and addrspacecasts. We do not however want to remove
857 // 0 GEPs.
858 if (!isa<GetElementPtrInst>(RHS))
859 RHS = RHS->stripPointerCasts();
860
861 Value *PtrBase = GEPLHS->getOperand(0);
862 // FIXME: Support vector pointer GEPs.
863 if (PtrBase == RHS && GEPLHS->isInBounds() &&
864 !GEPLHS->getType()->isVectorTy()) {
865 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
866 // This transformation (ignoring the base and scales) is valid because we
867 // know pointers can't overflow since the gep is inbounds. See if we can
868 // output an optimized form.
869 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
870
871 // If not, synthesize the offset the hard way.
872 if (!Offset)
873 Offset = EmitGEPOffset(GEPLHS);
874 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
875 Constant::getNullValue(Offset->getType()));
876 }
877
878 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
879 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
880 !NullPointerIsDefined(I.getFunction(),
881 RHS->getType()->getPointerAddressSpace())) {
882 // For most address spaces, an allocation can't be placed at null, but null
883 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
884 // the only valid inbounds address derived from null, is null itself.
885 // Thus, we have four cases to consider:
886 // 1) Base == nullptr, Offset == 0 -> inbounds, null
887 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
888 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
889 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
890 //
891 // (Note if we're indexing a type of size 0, that simply collapses into one
892 // of the buckets above.)
893 //
894 // In general, we're allowed to make values less poison (i.e. remove
895 // sources of full UB), so in this case, we just select between the two
896 // non-poison cases (1 and 4 above).
897 //
898 // For vectors, we apply the same reasoning on a per-lane basis.
899 auto *Base = GEPLHS->getPointerOperand();
900 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
901 int NumElts = cast<VectorType>(GEPLHS->getType())->getNumElements();
902 Base = Builder.CreateVectorSplat(NumElts, Base);
903 }
904 return new ICmpInst(Cond, Base,
905 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
906 cast<Constant>(RHS), Base->getType()));
907 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
908 // If the base pointers are different, but the indices are the same, just
909 // compare the base pointer.
910 if (PtrBase != GEPRHS->getOperand(0)) {
911 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
912 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
913 GEPRHS->getOperand(0)->getType();
914 if (IndicesTheSame)
915 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
916 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
917 IndicesTheSame = false;
918 break;
919 }
920
921 // If all indices are the same, just compare the base pointers.
922 Type *BaseType = GEPLHS->getOperand(0)->getType();
923 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
924 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
925
926 // If we're comparing GEPs with two base pointers that only differ in type
927 // and both GEPs have only constant indices or just one use, then fold
928 // the compare with the adjusted indices.
929 // FIXME: Support vector of pointers.
930 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
931 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
932 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
933 PtrBase->stripPointerCasts() ==
934 GEPRHS->getOperand(0)->stripPointerCasts() &&
935 !GEPLHS->getType()->isVectorTy()) {
936 Value *LOffset = EmitGEPOffset(GEPLHS);
937 Value *ROffset = EmitGEPOffset(GEPRHS);
938
939 // If we looked through an addrspacecast between different sized address
940 // spaces, the LHS and RHS pointers are different sized
941 // integers. Truncate to the smaller one.
942 Type *LHSIndexTy = LOffset->getType();
943 Type *RHSIndexTy = ROffset->getType();
944 if (LHSIndexTy != RHSIndexTy) {
945 if (LHSIndexTy->getPrimitiveSizeInBits() <
946 RHSIndexTy->getPrimitiveSizeInBits()) {
947 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
948 } else
949 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
950 }
951
952 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
953 LOffset, ROffset);
954 return replaceInstUsesWith(I, Cmp);
955 }
956
957 // Otherwise, the base pointers are different and the indices are
958 // different. Try convert this to an indexed compare by looking through
959 // PHIs/casts.
960 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
961 }
962
963 // If one of the GEPs has all zero indices, recurse.
964 // FIXME: Handle vector of pointers.
965 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
966 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
967 ICmpInst::getSwappedPredicate(Cond), I);
968
969 // If the other GEP has all zero indices, recurse.
970 // FIXME: Handle vector of pointers.
971 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
972 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
973
974 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
975 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
976 // If the GEPs only differ by one index, compare it.
977 unsigned NumDifferences = 0; // Keep track of # differences.
978 unsigned DiffOperand = 0; // The operand that differs.
979 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
980 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
981 Type *LHSType = GEPLHS->getOperand(i)->getType();
982 Type *RHSType = GEPRHS->getOperand(i)->getType();
983 // FIXME: Better support for vector of pointers.
984 if (LHSType->getPrimitiveSizeInBits() !=
985 RHSType->getPrimitiveSizeInBits() ||
986 (GEPLHS->getType()->isVectorTy() &&
987 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
988 // Irreconcilable differences.
989 NumDifferences = 2;
990 break;
991 }
992
993 if (NumDifferences++) break;
994 DiffOperand = i;
995 }
996
997 if (NumDifferences == 0) // SAME GEP?
998 return replaceInstUsesWith(I, // No comparison is needed here.
999 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
1000
1001 else if (NumDifferences == 1 && GEPsInBounds) {
1002 Value *LHSV = GEPLHS->getOperand(DiffOperand);
1003 Value *RHSV = GEPRHS->getOperand(DiffOperand);
1004 // Make sure we do a signed comparison here.
1005 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1006 }
1007 }
1008
1009 // Only lower this if the icmp is the only user of the GEP or if we expect
1010 // the result to fold to a constant!
1011 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1012 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1013 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1014 Value *L = EmitGEPOffset(GEPLHS);
1015 Value *R = EmitGEPOffset(GEPRHS);
1016 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1017 }
1018 }
1019
1020 // Try convert this to an indexed compare by looking through PHIs/casts as a
1021 // last resort.
1022 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1023}
1024
1025Instruction *InstCombinerImpl::foldAllocaCmp(ICmpInst &ICI,
1026 const AllocaInst *Alloca,
1027 const Value *Other) {
1028 assert(ICI.isEquality() && "Cannot fold non-equality comparison.")((ICI.isEquality() && "Cannot fold non-equality comparison."
) ? static_cast<void> (0) : __assert_fail ("ICI.isEquality() && \"Cannot fold non-equality comparison.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1028, __PRETTY_FUNCTION__))
;
1029
1030 // It would be tempting to fold away comparisons between allocas and any
1031 // pointer not based on that alloca (e.g. an argument). However, even
1032 // though such pointers cannot alias, they can still compare equal.
1033 //
1034 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1035 // doesn't escape we can argue that it's impossible to guess its value, and we
1036 // can therefore act as if any such guesses are wrong.
1037 //
1038 // The code below checks that the alloca doesn't escape, and that it's only
1039 // used in a comparison once (the current instruction). The
1040 // single-comparison-use condition ensures that we're trivially folding all
1041 // comparisons against the alloca consistently, and avoids the risk of
1042 // erroneously folding a comparison of the pointer with itself.
1043
1044 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1045
1046 SmallVector<const Use *, 32> Worklist;
1047 for (const Use &U : Alloca->uses()) {
1048 if (Worklist.size() >= MaxIter)
1049 return nullptr;
1050 Worklist.push_back(&U);
1051 }
1052
1053 unsigned NumCmps = 0;
1054 while (!Worklist.empty()) {
1055 assert(Worklist.size() <= MaxIter)((Worklist.size() <= MaxIter) ? static_cast<void> (0
) : __assert_fail ("Worklist.size() <= MaxIter", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1055, __PRETTY_FUNCTION__))
;
1056 const Use *U = Worklist.pop_back_val();
1057 const Value *V = U->getUser();
1058 --MaxIter;
1059
1060 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1061 isa<SelectInst>(V)) {
1062 // Track the uses.
1063 } else if (isa<LoadInst>(V)) {
1064 // Loading from the pointer doesn't escape it.
1065 continue;
1066 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1067 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1068 if (SI->getValueOperand() == U->get())
1069 return nullptr;
1070 continue;
1071 } else if (isa<ICmpInst>(V)) {
1072 if (NumCmps++)
1073 return nullptr; // Found more than one cmp.
1074 continue;
1075 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1076 switch (Intrin->getIntrinsicID()) {
1077 // These intrinsics don't escape or compare the pointer. Memset is safe
1078 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1079 // we don't allow stores, so src cannot point to V.
1080 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1081 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1082 continue;
1083 default:
1084 return nullptr;
1085 }
1086 } else {
1087 return nullptr;
1088 }
1089 for (const Use &U : V->uses()) {
1090 if (Worklist.size() >= MaxIter)
1091 return nullptr;
1092 Worklist.push_back(&U);
1093 }
1094 }
1095
1096 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1097 return replaceInstUsesWith(
1098 ICI,
1099 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1100}
1101
1102/// Fold "icmp pred (X+C), X".
1103Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
1104 ICmpInst::Predicate Pred) {
1105 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1106 // so the values can never be equal. Similarly for all other "or equals"
1107 // operators.
1108 assert(!!C && "C should not be zero!")((!!C && "C should not be zero!") ? static_cast<void
> (0) : __assert_fail ("!!C && \"C should not be zero!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1108, __PRETTY_FUNCTION__))
;
1109
1110 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1111 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1112 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1113 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1114 Constant *R = ConstantInt::get(X->getType(),
1115 APInt::getMaxValue(C.getBitWidth()) - C);
1116 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1117 }
1118
1119 // (X+1) >u X --> X <u (0-1) --> X != 255
1120 // (X+2) >u X --> X <u (0-2) --> X <u 254
1121 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1122 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1123 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1124 ConstantInt::get(X->getType(), -C));
1125
1126 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1127
1128 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1129 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1130 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1131 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1132 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1133 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1134 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1135 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1136 ConstantInt::get(X->getType(), SMax - C));
1137
1138 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1139 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1140 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1141 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1142 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1143 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1144
1145 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) ?
static_cast<void> (0) : __assert_fail ("Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1145, __PRETTY_FUNCTION__))
;
1146 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1147 ConstantInt::get(X->getType(), SMax - (C - 1)));
1148}
1149
1150/// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1151/// (icmp eq/ne A, Log2(AP2/AP1)) ->
1152/// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1153Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
1154 const APInt &AP1,
1155 const APInt &AP2) {
1156 assert(I.isEquality() && "Cannot fold icmp gt/lt")((I.isEquality() && "Cannot fold icmp gt/lt") ? static_cast
<void> (0) : __assert_fail ("I.isEquality() && \"Cannot fold icmp gt/lt\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1156, __PRETTY_FUNCTION__))
;
1157
1158 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1159 if (I.getPredicate() == I.ICMP_NE)
1160 Pred = CmpInst::getInversePredicate(Pred);
1161 return new ICmpInst(Pred, LHS, RHS);
1162 };
1163
1164 // Don't bother doing any work for cases which InstSimplify handles.
1165 if (AP2.isNullValue())
1166 return nullptr;
1167
1168 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1169 if (IsAShr) {
1170 if (AP2.isAllOnesValue())
1171 return nullptr;
1172 if (AP2.isNegative() != AP1.isNegative())
1173 return nullptr;
1174 if (AP2.sgt(AP1))
1175 return nullptr;
1176 }
1177
1178 if (!AP1)
1179 // 'A' must be large enough to shift out the highest set bit.
1180 return getICmp(I.ICMP_UGT, A,
1181 ConstantInt::get(A->getType(), AP2.logBase2()));
1182
1183 if (AP1 == AP2)
1184 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1185
1186 int Shift;
1187 if (IsAShr && AP1.isNegative())
1188 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1189 else
1190 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1191
1192 if (Shift > 0) {
1193 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1194 // There are multiple solutions if we are comparing against -1 and the LHS
1195 // of the ashr is not a power of two.
1196 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1197 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1198 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1199 } else if (AP1 == AP2.lshr(Shift)) {
1200 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1201 }
1202 }
1203
1204 // Shifting const2 will never be equal to const1.
1205 // FIXME: This should always be handled by InstSimplify?
1206 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1207 return replaceInstUsesWith(I, TorF);
1208}
1209
1210/// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1211/// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1212Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1213 const APInt &AP1,
1214 const APInt &AP2) {
1215 assert(I.isEquality() && "Cannot fold icmp gt/lt")((I.isEquality() && "Cannot fold icmp gt/lt") ? static_cast
<void> (0) : __assert_fail ("I.isEquality() && \"Cannot fold icmp gt/lt\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1215, __PRETTY_FUNCTION__))
;
1216
1217 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1218 if (I.getPredicate() == I.ICMP_NE)
1219 Pred = CmpInst::getInversePredicate(Pred);
1220 return new ICmpInst(Pred, LHS, RHS);
1221 };
1222
1223 // Don't bother doing any work for cases which InstSimplify handles.
1224 if (AP2.isNullValue())
1225 return nullptr;
1226
1227 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1228
1229 if (!AP1 && AP2TrailingZeros != 0)
1230 return getICmp(
1231 I.ICMP_UGE, A,
1232 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1233
1234 if (AP1 == AP2)
1235 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1236
1237 // Get the distance between the lowest bits that are set.
1238 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1239
1240 if (Shift > 0 && AP2.shl(Shift) == AP1)
1241 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1242
1243 // Shifting const2 will never be equal to const1.
1244 // FIXME: This should always be handled by InstSimplify?
1245 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1246 return replaceInstUsesWith(I, TorF);
1247}
1248
1249/// The caller has matched a pattern of the form:
1250/// I = icmp ugt (add (add A, B), CI2), CI1
1251/// If this is of the form:
1252/// sum = a + b
1253/// if (sum+128 >u 255)
1254/// Then replace it with llvm.sadd.with.overflow.i8.
1255///
1256static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1257 ConstantInt *CI2, ConstantInt *CI1,
1258 InstCombinerImpl &IC) {
1259 // The transformation we're trying to do here is to transform this into an
1260 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1261 // with a narrower add, and discard the add-with-constant that is part of the
1262 // range check (if we can't eliminate it, this isn't profitable).
1263
1264 // In order to eliminate the add-with-constant, the compare can be its only
1265 // use.
1266 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1267 if (!AddWithCst->hasOneUse())
1268 return nullptr;
1269
1270 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1271 if (!CI2->getValue().isPowerOf2())
1272 return nullptr;
1273 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1274 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1275 return nullptr;
1276
1277 // The width of the new add formed is 1 more than the bias.
1278 ++NewWidth;
1279
1280 // Check to see that CI1 is an all-ones value with NewWidth bits.
1281 if (CI1->getBitWidth() == NewWidth ||
1282 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1283 return nullptr;
1284
1285 // This is only really a signed overflow check if the inputs have been
1286 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1287 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1288 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1289 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1290 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1291 return nullptr;
1292
1293 // In order to replace the original add with a narrower
1294 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1295 // and truncates that discard the high bits of the add. Verify that this is
1296 // the case.
1297 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1298 for (User *U : OrigAdd->users()) {
1299 if (U == AddWithCst)
1300 continue;
1301
1302 // Only accept truncates for now. We would really like a nice recursive
1303 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1304 // chain to see which bits of a value are actually demanded. If the
1305 // original add had another add which was then immediately truncated, we
1306 // could still do the transformation.
1307 TruncInst *TI = dyn_cast<TruncInst>(U);
1308 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1309 return nullptr;
1310 }
1311
1312 // If the pattern matches, truncate the inputs to the narrower type and
1313 // use the sadd_with_overflow intrinsic to efficiently compute both the
1314 // result and the overflow bit.
1315 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1316 Function *F = Intrinsic::getDeclaration(
1317 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1318
1319 InstCombiner::BuilderTy &Builder = IC.Builder;
1320
1321 // Put the new code above the original add, in case there are any uses of the
1322 // add between the add and the compare.
1323 Builder.SetInsertPoint(OrigAdd);
1324
1325 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1326 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1327 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1328 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1329 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1330
1331 // The inner add was the result of the narrow add, zero extended to the
1332 // wider type. Replace it with the result computed by the intrinsic.
1333 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1334 IC.eraseInstFromFunction(*OrigAdd);
1335
1336 // The original icmp gets replaced with the overflow value.
1337 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1338}
1339
1340/// If we have:
1341/// icmp eq/ne (urem/srem %x, %y), 0
1342/// iff %y is a power-of-two, we can replace this with a bit test:
1343/// icmp eq/ne (and %x, (add %y, -1)), 0
1344Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1345 // This fold is only valid for equality predicates.
1346 if (!I.isEquality())
1347 return nullptr;
1348 ICmpInst::Predicate Pred;
1349 Value *X, *Y, *Zero;
1350 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1351 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1352 return nullptr;
1353 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1354 return nullptr;
1355 // This may increase instruction count, we don't enforce that Y is a constant.
1356 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1357 Value *Masked = Builder.CreateAnd(X, Mask);
1358 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1359}
1360
1361/// Fold equality-comparison between zero and any (maybe truncated) right-shift
1362/// by one-less-than-bitwidth into a sign test on the original value.
1363Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1364 Instruction *Val;
1365 ICmpInst::Predicate Pred;
1366 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1367 return nullptr;
1368
1369 Value *X;
1370 Type *XTy;
1371
1372 Constant *C;
1373 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1374 XTy = X->getType();
1375 unsigned XBitWidth = XTy->getScalarSizeInBits();
1376 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1377 APInt(XBitWidth, XBitWidth - 1))))
1378 return nullptr;
1379 } else if (isa<BinaryOperator>(Val) &&
1380 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1381 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1382 /*AnalyzeForSignBitExtraction=*/true))) {
1383 XTy = X->getType();
1384 } else
1385 return nullptr;
1386
1387 return ICmpInst::Create(Instruction::ICmp,
1388 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1389 : ICmpInst::ICMP_SLT,
1390 X, ConstantInt::getNullValue(XTy));
1391}
1392
1393// Handle icmp pred X, 0
1394Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1395 CmpInst::Predicate Pred = Cmp.getPredicate();
1396 if (!match(Cmp.getOperand(1), m_Zero()))
1397 return nullptr;
1398
1399 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1400 if (Pred == ICmpInst::ICMP_SGT) {
1401 Value *A, *B;
1402 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1403 if (SPR.Flavor == SPF_SMIN) {
1404 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1405 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1406 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1407 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1408 }
1409 }
1410
1411 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1412 return New;
1413
1414 // Given:
1415 // icmp eq/ne (urem %x, %y), 0
1416 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1417 // icmp eq/ne %x, 0
1418 Value *X, *Y;
1419 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1420 ICmpInst::isEquality(Pred)) {
1421 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1422 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1423 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1424 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1425 }
1426
1427 return nullptr;
1428}
1429
1430/// Fold icmp Pred X, C.
1431/// TODO: This code structure does not make sense. The saturating add fold
1432/// should be moved to some other helper and extended as noted below (it is also
1433/// possible that code has been made unnecessary - do we canonicalize IR to
1434/// overflow/saturating intrinsics or not?).
1435Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1436 // Match the following pattern, which is a common idiom when writing
1437 // overflow-safe integer arithmetic functions. The source performs an addition
1438 // in wider type and explicitly checks for overflow using comparisons against
1439 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1440 //
1441 // TODO: This could probably be generalized to handle other overflow-safe
1442 // operations if we worked out the formulas to compute the appropriate magic
1443 // constants.
1444 //
1445 // sum = a + b
1446 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1447 CmpInst::Predicate Pred = Cmp.getPredicate();
1448 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1449 Value *A, *B;
1450 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1451 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1452 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1453 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1454 return Res;
1455
1456 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1457 Constant *C = dyn_cast<Constant>(Op1);
1458 if (!C)
1459 return nullptr;
1460
1461 if (auto *Phi = dyn_cast<PHINode>(Op0))
1462 if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1463 Type *Ty = Cmp.getType();
1464 Builder.SetInsertPoint(Phi);
1465 PHINode *NewPhi =
1466 Builder.CreatePHI(Ty, Phi->getNumOperands());
1467 for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1468 auto *Input =
1469 cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1470 auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1471 NewPhi->addIncoming(BoolInput, Predecessor);
1472 }
1473 NewPhi->takeName(&Cmp);
1474 return replaceInstUsesWith(Cmp, NewPhi);
1475 }
1476
1477 return nullptr;
1478}
1479
1480/// Canonicalize icmp instructions based on dominating conditions.
1481Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1482 // This is a cheap/incomplete check for dominance - just match a single
1483 // predecessor with a conditional branch.
1484 BasicBlock *CmpBB = Cmp.getParent();
1485 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1486 if (!DomBB)
1487 return nullptr;
1488
1489 Value *DomCond;
1490 BasicBlock *TrueBB, *FalseBB;
1491 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1492 return nullptr;
1493
1494 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&(((TrueBB == CmpBB || FalseBB == CmpBB) && "Predecessor block does not point to successor?"
) ? static_cast<void> (0) : __assert_fail ("(TrueBB == CmpBB || FalseBB == CmpBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1495, __PRETTY_FUNCTION__))
1495 "Predecessor block does not point to successor?")(((TrueBB == CmpBB || FalseBB == CmpBB) && "Predecessor block does not point to successor?"
) ? static_cast<void> (0) : __assert_fail ("(TrueBB == CmpBB || FalseBB == CmpBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1495, __PRETTY_FUNCTION__))
;
1496
1497 // The branch should get simplified. Don't bother simplifying this condition.
1498 if (TrueBB == FalseBB)
1499 return nullptr;
1500
1501 // Try to simplify this compare to T/F based on the dominating condition.
1502 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1503 if (Imp)
1504 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1505
1506 CmpInst::Predicate Pred = Cmp.getPredicate();
1507 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1508 ICmpInst::Predicate DomPred;
1509 const APInt *C, *DomC;
1510 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1511 match(Y, m_APInt(C))) {
1512 // We have 2 compares of a variable with constants. Calculate the constant
1513 // ranges of those compares to see if we can transform the 2nd compare:
1514 // DomBB:
1515 // DomCond = icmp DomPred X, DomC
1516 // br DomCond, CmpBB, FalseBB
1517 // CmpBB:
1518 // Cmp = icmp Pred X, C
1519 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1520 ConstantRange DominatingCR =
1521 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1522 : ConstantRange::makeExactICmpRegion(
1523 CmpInst::getInversePredicate(DomPred), *DomC);
1524 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1525 ConstantRange Difference = DominatingCR.difference(CR);
1526 if (Intersection.isEmptySet())
1527 return replaceInstUsesWith(Cmp, Builder.getFalse());
1528 if (Difference.isEmptySet())
1529 return replaceInstUsesWith(Cmp, Builder.getTrue());
1530
1531 // Canonicalizing a sign bit comparison that gets used in a branch,
1532 // pessimizes codegen by generating branch on zero instruction instead
1533 // of a test and branch. So we avoid canonicalizing in such situations
1534 // because test and branch instruction has better branch displacement
1535 // than compare and branch instruction.
1536 bool UnusedBit;
1537 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1538 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1539 return nullptr;
1540
1541 if (const APInt *EqC = Intersection.getSingleElement())
1542 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1543 if (const APInt *NeC = Difference.getSingleElement())
1544 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1545 }
1546
1547 return nullptr;
1548}
1549
1550/// Fold icmp (trunc X, Y), C.
1551Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1552 TruncInst *Trunc,
1553 const APInt &C) {
1554 ICmpInst::Predicate Pred = Cmp.getPredicate();
1555 Value *X = Trunc->getOperand(0);
1556 if (C.isOneValue() && C.getBitWidth() > 1) {
1557 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1558 Value *V = nullptr;
1559 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1560 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1561 ConstantInt::get(V->getType(), 1));
1562 }
1563
1564 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1565 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1566 // of the high bits truncated out of x are known.
1567 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1568 SrcBits = X->getType()->getScalarSizeInBits();
1569 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1570
1571 // If all the high bits are known, we can do this xform.
1572 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1573 // Pull in the high bits from known-ones set.
1574 APInt NewRHS = C.zext(SrcBits);
1575 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1576 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1577 }
1578 }
1579
1580 return nullptr;
1581}
1582
1583/// Fold icmp (xor X, Y), C.
1584Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1585 BinaryOperator *Xor,
1586 const APInt &C) {
1587 Value *X = Xor->getOperand(0);
1588 Value *Y = Xor->getOperand(1);
1589 const APInt *XorC;
1590 if (!match(Y, m_APInt(XorC)))
1591 return nullptr;
1592
1593 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1594 // fold the xor.
1595 ICmpInst::Predicate Pred = Cmp.getPredicate();
1596 bool TrueIfSigned = false;
1597 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1598
1599 // If the sign bit of the XorCst is not set, there is no change to
1600 // the operation, just stop using the Xor.
1601 if (!XorC->isNegative())
1602 return replaceOperand(Cmp, 0, X);
1603
1604 // Emit the opposite comparison.
1605 if (TrueIfSigned)
1606 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1607 ConstantInt::getAllOnesValue(X->getType()));
1608 else
1609 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1610 ConstantInt::getNullValue(X->getType()));
1611 }
1612
1613 if (Xor->hasOneUse()) {
1614 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1615 if (!Cmp.isEquality() && XorC->isSignMask()) {
1616 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1617 : Cmp.getSignedPredicate();
1618 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1619 }
1620
1621 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1622 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1623 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1624 : Cmp.getSignedPredicate();
1625 Pred = Cmp.getSwappedPredicate(Pred);
1626 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1627 }
1628 }
1629
1630 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1631 if (Pred == ICmpInst::ICMP_UGT) {
1632 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1633 if (*XorC == ~C && (C + 1).isPowerOf2())
1634 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1635 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1636 if (*XorC == C && (C + 1).isPowerOf2())
1637 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1638 }
1639 if (Pred == ICmpInst::ICMP_ULT) {
1640 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1641 if (*XorC == -C && C.isPowerOf2())
1642 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1643 ConstantInt::get(X->getType(), ~C));
1644 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1645 if (*XorC == C && (-C).isPowerOf2())
1646 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1647 ConstantInt::get(X->getType(), ~C));
1648 }
1649 return nullptr;
1650}
1651
1652/// Fold icmp (and (sh X, Y), C2), C1.
1653Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1654 BinaryOperator *And,
1655 const APInt &C1,
1656 const APInt &C2) {
1657 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1658 if (!Shift || !Shift->isShift())
1659 return nullptr;
1660
1661 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1662 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1663 // code produced by the clang front-end, for bitfield access.
1664 // This seemingly simple opportunity to fold away a shift turns out to be
1665 // rather complicated. See PR17827 for details.
1666 unsigned ShiftOpcode = Shift->getOpcode();
1667 bool IsShl = ShiftOpcode == Instruction::Shl;
1668 const APInt *C3;
1669 if (match(Shift->getOperand(1), m_APInt(C3))) {
1670 APInt NewAndCst, NewCmpCst;
1671 bool AnyCmpCstBitsShiftedOut;
1672 if (ShiftOpcode == Instruction::Shl) {
1673 // For a left shift, we can fold if the comparison is not signed. We can
1674 // also fold a signed comparison if the mask value and comparison value
1675 // are not negative. These constraints may not be obvious, but we can
1676 // prove that they are correct using an SMT solver.
1677 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1678 return nullptr;
1679
1680 NewCmpCst = C1.lshr(*C3);
1681 NewAndCst = C2.lshr(*C3);
1682 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1683 } else if (ShiftOpcode == Instruction::LShr) {
1684 // For a logical right shift, we can fold if the comparison is not signed.
1685 // We can also fold a signed comparison if the shifted mask value and the
1686 // shifted comparison value are not negative. These constraints may not be
1687 // obvious, but we can prove that they are correct using an SMT solver.
1688 NewCmpCst = C1.shl(*C3);
1689 NewAndCst = C2.shl(*C3);
1690 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1691 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1692 return nullptr;
1693 } else {
1694 // For an arithmetic shift, check that both constants don't use (in a
1695 // signed sense) the top bits being shifted out.
1696 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode")((ShiftOpcode == Instruction::AShr && "Unknown shift opcode"
) ? static_cast<void> (0) : __assert_fail ("ShiftOpcode == Instruction::AShr && \"Unknown shift opcode\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1696, __PRETTY_FUNCTION__))
;
1697 NewCmpCst = C1.shl(*C3);
1698 NewAndCst = C2.shl(*C3);
1699 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1700 if (NewAndCst.ashr(*C3) != C2)
1701 return nullptr;
1702 }
1703
1704 if (AnyCmpCstBitsShiftedOut) {
1705 // If we shifted bits out, the fold is not going to work out. As a
1706 // special case, check to see if this means that the result is always
1707 // true or false now.
1708 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1709 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1710 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1711 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1712 } else {
1713 Value *NewAnd = Builder.CreateAnd(
1714 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1715 return new ICmpInst(Cmp.getPredicate(),
1716 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1717 }
1718 }
1719
1720 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1721 // preferable because it allows the C2 << Y expression to be hoisted out of a
1722 // loop if Y is invariant and X is not.
1723 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1724 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1725 // Compute C2 << Y.
1726 Value *NewShift =
1727 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1728 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1729
1730 // Compute X & (C2 << Y).
1731 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1732 return replaceOperand(Cmp, 0, NewAnd);
1733 }
1734
1735 return nullptr;
1736}
1737
1738/// Fold icmp (and X, C2), C1.
1739Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1740 BinaryOperator *And,
1741 const APInt &C1) {
1742 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1743
1744 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1745 // TODO: We canonicalize to the longer form for scalars because we have
1746 // better analysis/folds for icmp, and codegen may be better with icmp.
1747 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1748 match(And->getOperand(1), m_One()))
1749 return new TruncInst(And->getOperand(0), Cmp.getType());
1750
1751 const APInt *C2;
1752 Value *X;
1753 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1754 return nullptr;
1755
1756 // Don't perform the following transforms if the AND has multiple uses
1757 if (!And->hasOneUse())
1758 return nullptr;
1759
1760 if (Cmp.isEquality() && C1.isNullValue()) {
1761 // Restrict this fold to single-use 'and' (PR10267).
1762 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1763 if (C2->isSignMask()) {
1764 Constant *Zero = Constant::getNullValue(X->getType());
1765 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1766 return new ICmpInst(NewPred, X, Zero);
1767 }
1768
1769 // Restrict this fold only for single-use 'and' (PR10267).
1770 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1771 if ((~(*C2) + 1).isPowerOf2()) {
1772 Constant *NegBOC =
1773 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1774 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1775 return new ICmpInst(NewPred, X, NegBOC);
1776 }
1777 }
1778
1779 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1780 // the input width without changing the value produced, eliminate the cast:
1781 //
1782 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1783 //
1784 // We can do this transformation if the constants do not have their sign bits
1785 // set or if it is an equality comparison. Extending a relational comparison
1786 // when we're checking the sign bit would not work.
1787 Value *W;
1788 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1789 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1790 // TODO: Is this a good transform for vectors? Wider types may reduce
1791 // throughput. Should this transform be limited (even for scalars) by using
1792 // shouldChangeType()?
1793 if (!Cmp.getType()->isVectorTy()) {
1794 Type *WideType = W->getType();
1795 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1796 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1797 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1798 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1799 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1800 }
1801 }
1802
1803 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1804 return I;
1805
1806 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1807 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1808 //
1809 // iff pred isn't signed
1810 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1811 match(And->getOperand(1), m_One())) {
1812 Constant *One = cast<Constant>(And->getOperand(1));
1813 Value *Or = And->getOperand(0);
1814 Value *A, *B, *LShr;
1815 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1816 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1817 unsigned UsesRemoved = 0;
1818 if (And->hasOneUse())
1819 ++UsesRemoved;
1820 if (Or->hasOneUse())
1821 ++UsesRemoved;
1822 if (LShr->hasOneUse())
1823 ++UsesRemoved;
1824
1825 // Compute A & ((1 << B) | 1)
1826 Value *NewOr = nullptr;
1827 if (auto *C = dyn_cast<Constant>(B)) {
1828 if (UsesRemoved >= 1)
1829 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1830 } else {
1831 if (UsesRemoved >= 3)
1832 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1833 /*HasNUW=*/true),
1834 One, Or->getName());
1835 }
1836 if (NewOr) {
1837 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1838 return replaceOperand(Cmp, 0, NewAnd);
1839 }
1840 }
1841 }
1842
1843 return nullptr;
1844}
1845
1846/// Fold icmp (and X, Y), C.
1847Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1848 BinaryOperator *And,
1849 const APInt &C) {
1850 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1851 return I;
1852
1853 // TODO: These all require that Y is constant too, so refactor with the above.
1854
1855 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1856 Value *X = And->getOperand(0);
1857 Value *Y = And->getOperand(1);
1858 if (auto *LI = dyn_cast<LoadInst>(X))
1859 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1860 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1861 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1862 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1863 ConstantInt *C2 = cast<ConstantInt>(Y);
1864 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1865 return Res;
1866 }
1867
1868 if (!Cmp.isEquality())
1869 return nullptr;
1870
1871 // X & -C == -C -> X > u ~C
1872 // X & -C != -C -> X <= u ~C
1873 // iff C is a power of 2
1874 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1875 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1876 : CmpInst::ICMP_ULE;
1877 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1878 }
1879
1880 // (X & C2) == 0 -> (trunc X) >= 0
1881 // (X & C2) != 0 -> (trunc X) < 0
1882 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1883 const APInt *C2;
1884 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1885 int32_t ExactLogBase2 = C2->exactLogBase2();
1886 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1887 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1888 if (auto *AndVTy = dyn_cast<VectorType>(And->getType()))
1889 NTy = FixedVectorType::get(NTy, AndVTy->getNumElements());
1890 Value *Trunc = Builder.CreateTrunc(X, NTy);
1891 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1892 : CmpInst::ICMP_SLT;
1893 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1894 }
1895 }
1896
1897 return nullptr;
1898}
1899
1900/// Fold icmp (or X, Y), C.
1901Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
1902 BinaryOperator *Or,
1903 const APInt &C) {
1904 ICmpInst::Predicate Pred = Cmp.getPredicate();
1905 if (C.isOneValue()) {
1906 // icmp slt signum(V) 1 --> icmp slt V, 1
1907 Value *V = nullptr;
1908 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1909 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1910 ConstantInt::get(V->getType(), 1));
1911 }
1912
1913 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1914 const APInt *MaskC;
1915 if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1916 if (*MaskC == C && (C + 1).isPowerOf2()) {
1917 // X | C == C --> X <=u C
1918 // X | C != C --> X >u C
1919 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1920 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1921 return new ICmpInst(Pred, OrOp0, OrOp1);
1922 }
1923
1924 // More general: canonicalize 'equality with set bits mask' to
1925 // 'equality with clear bits mask'.
1926 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1927 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1928 if (Or->hasOneUse()) {
1929 Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1930 Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1931 return new ICmpInst(Pred, And, NewC);
1932 }
1933 }
1934
1935 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1936 return nullptr;
1937
1938 Value *P, *Q;
1939 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1940 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1941 // -> and (icmp eq P, null), (icmp eq Q, null).
1942 Value *CmpP =
1943 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1944 Value *CmpQ =
1945 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1946 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1947 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1948 }
1949
1950 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1951 // a shorter form that has more potential to be folded even further.
1952 Value *X1, *X2, *X3, *X4;
1953 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1954 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1955 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1956 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1957 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1958 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1959 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1960 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1961 }
1962
1963 return nullptr;
1964}
1965
1966/// Fold icmp (mul X, Y), C.
1967Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
1968 BinaryOperator *Mul,
1969 const APInt &C) {
1970 const APInt *MulC;
1971 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1972 return nullptr;
1973
1974 // If this is a test of the sign bit and the multiply is sign-preserving with
1975 // a constant operand, use the multiply LHS operand instead.
1976 ICmpInst::Predicate Pred = Cmp.getPredicate();
1977 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1978 if (MulC->isNegative())
1979 Pred = ICmpInst::getSwappedPredicate(Pred);
1980 return new ICmpInst(Pred, Mul->getOperand(0),
1981 Constant::getNullValue(Mul->getType()));
1982 }
1983
1984 // If the multiply does not wrap, try to divide the compare constant by the
1985 // multiplication factor.
1986 if (Cmp.isEquality() && !MulC->isNullValue()) {
1987 // (mul nsw X, MulC) == C --> X == C /s MulC
1988 if (Mul->hasNoSignedWrap() && C.srem(*MulC).isNullValue()) {
1989 Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
1990 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1991 }
1992 // (mul nuw X, MulC) == C --> X == C /u MulC
1993 if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isNullValue()) {
1994 Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
1995 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1996 }
1997 }
1998
1999 return nullptr;
2000}
2001
2002/// Fold icmp (shl 1, Y), C.
2003static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
2004 const APInt &C) {
2005 Value *Y;
2006 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2007 return nullptr;
2008
2009 Type *ShiftType = Shl->getType();
2010 unsigned TypeBits = C.getBitWidth();
2011 bool CIsPowerOf2 = C.isPowerOf2();
2012 ICmpInst::Predicate Pred = Cmp.getPredicate();
2013 if (Cmp.isUnsigned()) {
2014 // (1 << Y) pred C -> Y pred Log2(C)
2015 if (!CIsPowerOf2) {
2016 // (1 << Y) < 30 -> Y <= 4
2017 // (1 << Y) <= 30 -> Y <= 4
2018 // (1 << Y) >= 30 -> Y > 4
2019 // (1 << Y) > 30 -> Y > 4
2020 if (Pred == ICmpInst::ICMP_ULT)
2021 Pred = ICmpInst::ICMP_ULE;
2022 else if (Pred == ICmpInst::ICMP_UGE)
2023 Pred = ICmpInst::ICMP_UGT;
2024 }
2025
2026 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2027 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2028 unsigned CLog2 = C.logBase2();
2029 if (CLog2 == TypeBits - 1) {
2030 if (Pred == ICmpInst::ICMP_UGE)
2031 Pred = ICmpInst::ICMP_EQ;
2032 else if (Pred == ICmpInst::ICMP_ULT)
2033 Pred = ICmpInst::ICMP_NE;
2034 }
2035 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2036 } else if (Cmp.isSigned()) {
2037 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2038 if (C.isAllOnesValue()) {
2039 // (1 << Y) <= -1 -> Y == 31
2040 if (Pred == ICmpInst::ICMP_SLE)
2041 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2042
2043 // (1 << Y) > -1 -> Y != 31
2044 if (Pred == ICmpInst::ICMP_SGT)
2045 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2046 } else if (!C) {
2047 // (1 << Y) < 0 -> Y == 31
2048 // (1 << Y) <= 0 -> Y == 31
2049 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2050 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2051
2052 // (1 << Y) >= 0 -> Y != 31
2053 // (1 << Y) > 0 -> Y != 31
2054 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2055 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2056 }
2057 } else if (Cmp.isEquality() && CIsPowerOf2) {
2058 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2059 }
2060
2061 return nullptr;
2062}
2063
2064/// Fold icmp (shl X, Y), C.
2065Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2066 BinaryOperator *Shl,
2067 const APInt &C) {
2068 const APInt *ShiftVal;
2069 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2070 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2071
2072 const APInt *ShiftAmt;
2073 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2074 return foldICmpShlOne(Cmp, Shl, C);
2075
2076 // Check that the shift amount is in range. If not, don't perform undefined
2077 // shifts. When the shift is visited, it will be simplified.
2078 unsigned TypeBits = C.getBitWidth();
2079 if (ShiftAmt->uge(TypeBits))
2080 return nullptr;
2081
2082 ICmpInst::Predicate Pred = Cmp.getPredicate();
2083 Value *X = Shl->getOperand(0);
2084 Type *ShType = Shl->getType();
2085
2086 // NSW guarantees that we are only shifting out sign bits from the high bits,
2087 // so we can ASHR the compare constant without needing a mask and eliminate
2088 // the shift.
2089 if (Shl->hasNoSignedWrap()) {
2090 if (Pred == ICmpInst::ICMP_SGT) {
2091 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2092 APInt ShiftedC = C.ashr(*ShiftAmt);
2093 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2094 }
2095 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2096 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2097 APInt ShiftedC = C.ashr(*ShiftAmt);
2098 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2099 }
2100 if (Pred == ICmpInst::ICMP_SLT) {
2101 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2102 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2103 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2104 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2105 assert(!C.isMinSignedValue() && "Unexpected icmp slt")((!C.isMinSignedValue() && "Unexpected icmp slt") ? static_cast
<void> (0) : __assert_fail ("!C.isMinSignedValue() && \"Unexpected icmp slt\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2105, __PRETTY_FUNCTION__))
;
2106 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2107 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2108 }
2109 // If this is a signed comparison to 0 and the shift is sign preserving,
2110 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2111 // do that if we're sure to not continue on in this function.
2112 if (isSignTest(Pred, C))
2113 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2114 }
2115
2116 // NUW guarantees that we are only shifting out zero bits from the high bits,
2117 // so we can LSHR the compare constant without needing a mask and eliminate
2118 // the shift.
2119 if (Shl->hasNoUnsignedWrap()) {
2120 if (Pred == ICmpInst::ICMP_UGT) {
2121 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2122 APInt ShiftedC = C.lshr(*ShiftAmt);
2123 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2124 }
2125 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2126 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2127 APInt ShiftedC = C.lshr(*ShiftAmt);
2128 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2129 }
2130 if (Pred == ICmpInst::ICMP_ULT) {
2131 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2132 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2133 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2134 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2135 assert(C.ugt(0) && "ult 0 should have been eliminated")((C.ugt(0) && "ult 0 should have been eliminated") ? static_cast
<void> (0) : __assert_fail ("C.ugt(0) && \"ult 0 should have been eliminated\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2135, __PRETTY_FUNCTION__))
;
2136 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2137 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2138 }
2139 }
2140
2141 if (Cmp.isEquality() && Shl->hasOneUse()) {
2142 // Strength-reduce the shift into an 'and'.
2143 Constant *Mask = ConstantInt::get(
2144 ShType,
2145 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2146 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2147 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2148 return new ICmpInst(Pred, And, LShrC);
2149 }
2150
2151 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2152 bool TrueIfSigned = false;
2153 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2154 // (X << 31) <s 0 --> (X & 1) != 0
2155 Constant *Mask = ConstantInt::get(
2156 ShType,
2157 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2158 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2159 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2160 And, Constant::getNullValue(ShType));
2161 }
2162
2163 // Simplify 'shl' inequality test into 'and' equality test.
2164 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2165 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2166 if ((C + 1).isPowerOf2() &&
2167 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2168 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2169 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2170 : ICmpInst::ICMP_NE,
2171 And, Constant::getNullValue(ShType));
2172 }
2173 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2174 if (C.isPowerOf2() &&
2175 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2176 Value *And =
2177 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2178 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2179 : ICmpInst::ICMP_NE,
2180 And, Constant::getNullValue(ShType));
2181 }
2182 }
2183
2184 // Transform (icmp pred iM (shl iM %v, N), C)
2185 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2186 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2187 // This enables us to get rid of the shift in favor of a trunc that may be
2188 // free on the target. It has the additional benefit of comparing to a
2189 // smaller constant that may be more target-friendly.
2190 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2191 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2192 DL.isLegalInteger(TypeBits - Amt)) {
2193 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2194 if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2195 TruncTy = FixedVectorType::get(TruncTy, ShVTy->getNumElements());
2196 Constant *NewC =
2197 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2198 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2199 }
2200
2201 return nullptr;
2202}
2203
2204/// Fold icmp ({al}shr X, Y), C.
2205Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2206 BinaryOperator *Shr,
2207 const APInt &C) {
2208 // An exact shr only shifts out zero bits, so:
2209 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2210 Value *X = Shr->getOperand(0);
2211 CmpInst::Predicate Pred = Cmp.getPredicate();
2212 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2213 C.isNullValue())
2214 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2215
2216 const APInt *ShiftVal;
2217 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2218 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2219
2220 const APInt *ShiftAmt;
2221 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2222 return nullptr;
2223
2224 // Check that the shift amount is in range. If not, don't perform undefined
2225 // shifts. When the shift is visited it will be simplified.
2226 unsigned TypeBits = C.getBitWidth();
2227 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2228 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2229 return nullptr;
2230
2231 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2232 bool IsExact = Shr->isExact();
2233 Type *ShrTy = Shr->getType();
2234 // TODO: If we could guarantee that InstSimplify would handle all of the
2235 // constant-value-based preconditions in the folds below, then we could assert
2236 // those conditions rather than checking them. This is difficult because of
2237 // undef/poison (PR34838).
2238 if (IsAShr) {
2239 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2240 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2241 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2242 APInt ShiftedC = C.shl(ShAmtVal);
2243 if (ShiftedC.ashr(ShAmtVal) == C)
2244 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2245 }
2246 if (Pred == CmpInst::ICMP_SGT) {
2247 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2248 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2249 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2250 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2251 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2252 }
2253 } else {
2254 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2255 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2256 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2257 APInt ShiftedC = C.shl(ShAmtVal);
2258 if (ShiftedC.lshr(ShAmtVal) == C)
2259 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2260 }
2261 if (Pred == CmpInst::ICMP_UGT) {
2262 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2263 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2264 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2265 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2266 }
2267 }
2268
2269 if (!Cmp.isEquality())
2270 return nullptr;
2271
2272 // Handle equality comparisons of shift-by-constant.
2273
2274 // If the comparison constant changes with the shift, the comparison cannot
2275 // succeed (bits of the comparison constant cannot match the shifted value).
2276 // This should be known by InstSimplify and already be folded to true/false.
2277 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||((((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
(!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
"Expected icmp+shr simplify did not occur.") ? static_cast<
void> (0) : __assert_fail ("((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && \"Expected icmp+shr simplify did not occur.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2279, __PRETTY_FUNCTION__))
2278 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&((((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
(!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
"Expected icmp+shr simplify did not occur.") ? static_cast<
void> (0) : __assert_fail ("((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && \"Expected icmp+shr simplify did not occur.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2279, __PRETTY_FUNCTION__))
2279 "Expected icmp+shr simplify did not occur.")((((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
(!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
"Expected icmp+shr simplify did not occur.") ? static_cast<
void> (0) : __assert_fail ("((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && \"Expected icmp+shr simplify did not occur.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2279, __PRETTY_FUNCTION__))
;
2280
2281 // If the bits shifted out are known zero, compare the unshifted value:
2282 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2283 if (Shr->isExact())
2284 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2285
2286 if (Shr->hasOneUse()) {
2287 // Canonicalize the shift into an 'and':
2288 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2289 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2290 Constant *Mask = ConstantInt::get(ShrTy, Val);
2291 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2292 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2293 }
2294
2295 return nullptr;
2296}
2297
2298Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2299 BinaryOperator *SRem,
2300 const APInt &C) {
2301 // Match an 'is positive' or 'is negative' comparison of remainder by a
2302 // constant power-of-2 value:
2303 // (X % pow2C) sgt/slt 0
2304 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2305 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2306 return nullptr;
2307
2308 // TODO: The one-use check is standard because we do not typically want to
2309 // create longer instruction sequences, but this might be a special-case
2310 // because srem is not good for analysis or codegen.
2311 if (!SRem->hasOneUse())
2312 return nullptr;
2313
2314 const APInt *DivisorC;
2315 if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2316 return nullptr;
2317
2318 // Mask off the sign bit and the modulo bits (low-bits).
2319 Type *Ty = SRem->getType();
2320 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2321 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2322 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2323
2324 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2325 // bit is set. Example:
2326 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2327 if (Pred == ICmpInst::ICMP_SGT)
2328 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2329
2330 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2331 // bit is set. Example:
2332 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2333 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2334}
2335
2336/// Fold icmp (udiv X, Y), C.
2337Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2338 BinaryOperator *UDiv,
2339 const APInt &C) {
2340 const APInt *C2;
2341 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2342 return nullptr;
2343
2344 assert(*C2 != 0 && "udiv 0, X should have been simplified already.")((*C2 != 0 && "udiv 0, X should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("*C2 != 0 && \"udiv 0, X should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2344, __PRETTY_FUNCTION__))
;
2345
2346 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2347 Value *Y = UDiv->getOperand(1);
2348 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2349 assert(!C.isMaxValue() &&((!C.isMaxValue() && "icmp ugt X, UINT_MAX should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("!C.isMaxValue() && \"icmp ugt X, UINT_MAX should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2350, __PRETTY_FUNCTION__))
2350 "icmp ugt X, UINT_MAX should have been simplified already.")((!C.isMaxValue() && "icmp ugt X, UINT_MAX should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("!C.isMaxValue() && \"icmp ugt X, UINT_MAX should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2350, __PRETTY_FUNCTION__))
;
2351 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2352 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2353 }
2354
2355 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2356 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2357 assert(C != 0 && "icmp ult X, 0 should have been simplified already.")((C != 0 && "icmp ult X, 0 should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("C != 0 && \"icmp ult X, 0 should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2357, __PRETTY_FUNCTION__))
;
2358 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2359 ConstantInt::get(Y->getType(), C2->udiv(C)));
2360 }
2361
2362 return nullptr;
2363}
2364
2365/// Fold icmp ({su}div X, Y), C.
2366Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2367 BinaryOperator *Div,
2368 const APInt &C) {
2369 // Fold: icmp pred ([us]div X, C2), C -> range test
2370 // Fold this div into the comparison, producing a range check.
2371 // Determine, based on the divide type, what the range is being
2372 // checked. If there is an overflow on the low or high side, remember
2373 // it, otherwise compute the range [low, hi) bounding the new value.
2374 // See: InsertRangeTest above for the kinds of replacements possible.
2375 const APInt *C2;
2376 if (!match(Div->getOperand(1), m_APInt(C2)))
2377 return nullptr;
2378
2379 // FIXME: If the operand types don't match the type of the divide
2380 // then don't attempt this transform. The code below doesn't have the
2381 // logic to deal with a signed divide and an unsigned compare (and
2382 // vice versa). This is because (x /s C2) <s C produces different
2383 // results than (x /s C2) <u C or (x /u C2) <s C or even
2384 // (x /u C2) <u C. Simply casting the operands and result won't
2385 // work. :( The if statement below tests that condition and bails
2386 // if it finds it.
2387 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2388 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2389 return nullptr;
2390
2391 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2392 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2393 // division-by-constant cases should be present, we can not assert that they
2394 // have happened before we reach this icmp instruction.
2395 if (C2->isNullValue() || C2->isOneValue() ||
2396 (DivIsSigned && C2->isAllOnesValue()))
2397 return nullptr;
2398
2399 // Compute Prod = C * C2. We are essentially solving an equation of
2400 // form X / C2 = C. We solve for X by multiplying C2 and C.
2401 // By solving for X, we can turn this into a range check instead of computing
2402 // a divide.
2403 APInt Prod = C * *C2;
2404
2405 // Determine if the product overflows by seeing if the product is not equal to
2406 // the divide. Make sure we do the same kind of divide as in the LHS
2407 // instruction that we're folding.
2408 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2409
2410 ICmpInst::Predicate Pred = Cmp.getPredicate();
2411
2412 // If the division is known to be exact, then there is no remainder from the
2413 // divide, so the covered range size is unit, otherwise it is the divisor.
2414 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2415
2416 // Figure out the interval that is being checked. For example, a comparison
2417 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2418 // Compute this interval based on the constants involved and the signedness of
2419 // the compare/divide. This computes a half-open interval, keeping track of
2420 // whether either value in the interval overflows. After analysis each
2421 // overflow variable is set to 0 if it's corresponding bound variable is valid
2422 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2423 int LoOverflow = 0, HiOverflow = 0;
2424 APInt LoBound, HiBound;
2425
2426 if (!DivIsSigned) { // udiv
2427 // e.g. X/5 op 3 --> [15, 20)
2428 LoBound = Prod;
2429 HiOverflow = LoOverflow = ProdOV;
2430 if (!HiOverflow) {
2431 // If this is not an exact divide, then many values in the range collapse
2432 // to the same result value.
2433 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2434 }
2435 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2436 if (C.isNullValue()) { // (X / pos) op 0
2437 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2438 LoBound = -(RangeSize - 1);
2439 HiBound = RangeSize;
2440 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2441 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2442 HiOverflow = LoOverflow = ProdOV;
2443 if (!HiOverflow)
2444 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2445 } else { // (X / pos) op neg
2446 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2447 HiBound = Prod + 1;
2448 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2449 if (!LoOverflow) {
2450 APInt DivNeg = -RangeSize;
2451 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2452 }
2453 }
2454 } else if (C2->isNegative()) { // Divisor is < 0.
2455 if (Div->isExact())
2456 RangeSize.negate();
2457 if (C.isNullValue()) { // (X / neg) op 0
2458 // e.g. X/-5 op 0 --> [-4, 5)
2459 LoBound = RangeSize + 1;
2460 HiBound = -RangeSize;
2461 if (HiBound == *C2) { // -INTMIN = INTMIN
2462 HiOverflow = 1; // [INTMIN+1, overflow)
2463 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2464 }
2465 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2466 // e.g. X/-5 op 3 --> [-19, -14)
2467 HiBound = Prod + 1;
2468 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2469 if (!LoOverflow)
2470 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2471 } else { // (X / neg) op neg
2472 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2473 LoOverflow = HiOverflow = ProdOV;
2474 if (!HiOverflow)
2475 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2476 }
2477
2478 // Dividing by a negative swaps the condition. LT <-> GT
2479 Pred = ICmpInst::getSwappedPredicate(Pred);
2480 }
2481
2482 Value *X = Div->getOperand(0);
2483 switch (Pred) {
2484 default: llvm_unreachable("Unhandled icmp opcode!")::llvm::llvm_unreachable_internal("Unhandled icmp opcode!", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2484)
;
2485 case ICmpInst::ICMP_EQ:
2486 if (LoOverflow && HiOverflow)
2487 return replaceInstUsesWith(Cmp, Builder.getFalse());
2488 if (HiOverflow)
2489 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2490 ICmpInst::ICMP_UGE, X,
2491 ConstantInt::get(Div->getType(), LoBound));
2492 if (LoOverflow)
2493 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2494 ICmpInst::ICMP_ULT, X,
2495 ConstantInt::get(Div->getType(), HiBound));
2496 return replaceInstUsesWith(
2497 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2498 case ICmpInst::ICMP_NE:
2499 if (LoOverflow && HiOverflow)
2500 return replaceInstUsesWith(Cmp, Builder.getTrue());
2501 if (HiOverflow)
2502 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2503 ICmpInst::ICMP_ULT, X,
2504 ConstantInt::get(Div->getType(), LoBound));
2505 if (LoOverflow)
2506 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2507 ICmpInst::ICMP_UGE, X,
2508 ConstantInt::get(Div->getType(), HiBound));
2509 return replaceInstUsesWith(Cmp,
2510 insertRangeTest(X, LoBound, HiBound,
2511 DivIsSigned, false));
2512 case ICmpInst::ICMP_ULT:
2513 case ICmpInst::ICMP_SLT:
2514 if (LoOverflow == +1) // Low bound is greater than input range.
2515 return replaceInstUsesWith(Cmp, Builder.getTrue());
2516 if (LoOverflow == -1) // Low bound is less than input range.
2517 return replaceInstUsesWith(Cmp, Builder.getFalse());
2518 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2519 case ICmpInst::ICMP_UGT:
2520 case ICmpInst::ICMP_SGT:
2521 if (HiOverflow == +1) // High bound greater than input range.
2522 return replaceInstUsesWith(Cmp, Builder.getFalse());
2523 if (HiOverflow == -1) // High bound less than input range.
2524 return replaceInstUsesWith(Cmp, Builder.getTrue());
2525 if (Pred == ICmpInst::ICMP_UGT)
2526 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2527 ConstantInt::get(Div->getType(), HiBound));
2528 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2529 ConstantInt::get(Div->getType(), HiBound));
2530 }
2531
2532 return nullptr;
2533}
2534
2535/// Fold icmp (sub X, Y), C.
2536Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2537 BinaryOperator *Sub,
2538 const APInt &C) {
2539 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2540 ICmpInst::Predicate Pred = Cmp.getPredicate();
2541 const APInt *C2;
2542 APInt SubResult;
2543
2544 // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2545 if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2546 return new ICmpInst(Cmp.getPredicate(), Y,
2547 ConstantInt::get(Y->getType(), 0));
2548
2549 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2550 if (match(X, m_APInt(C2)) &&
2551 ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2552 (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2553 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2554 return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2555 ConstantInt::get(Y->getType(), SubResult));
2556
2557 // The following transforms are only worth it if the only user of the subtract
2558 // is the icmp.
2559 if (!Sub->hasOneUse())
2560 return nullptr;
2561
2562 if (Sub->hasNoSignedWrap()) {
2563 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2564 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2565 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2566
2567 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2568 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2569 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2570
2571 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2572 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2573 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2574
2575 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2576 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2577 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2578 }
2579
2580 if (!match(X, m_APInt(C2)))
2581 return nullptr;
2582
2583 // C2 - Y <u C -> (Y | (C - 1)) == C2
2584 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2585 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2586 (*C2 & (C - 1)) == (C - 1))
2587 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2588
2589 // C2 - Y >u C -> (Y | C) != C2
2590 // iff C2 & C == C and C + 1 is a power of 2
2591 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2592 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2593
2594 return nullptr;
2595}
2596
2597/// Fold icmp (add X, Y), C.
2598Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
2599 BinaryOperator *Add,
2600 const APInt &C) {
2601 Value *Y = Add->getOperand(1);
2602 const APInt *C2;
2603 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2604 return nullptr;
2605
2606 // Fold icmp pred (add X, C2), C.
2607 Value *X = Add->getOperand(0);
2608 Type *Ty = Add->getType();
2609 CmpInst::Predicate Pred = Cmp.getPredicate();
2610
2611 // If the add does not wrap, we can always adjust the compare by subtracting
2612 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2613 // are canonicalized to SGT/SLT/UGT/ULT.
2614 if ((Add->hasNoSignedWrap() &&
2615 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2616 (Add->hasNoUnsignedWrap() &&
2617 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2618 bool Overflow;
2619 APInt NewC =
2620 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2621 // If there is overflow, the result must be true or false.
2622 // TODO: Can we assert there is no overflow because InstSimplify always
2623 // handles those cases?
2624 if (!Overflow)
2625 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2626 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2627 }
2628
2629 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2630 const APInt &Upper = CR.getUpper();
2631 const APInt &Lower = CR.getLower();
2632 if (Cmp.isSigned()) {
2633 if (Lower.isSignMask())
2634 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2635 if (Upper.isSignMask())
2636 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2637 } else {
2638 if (Lower.isMinValue())
2639 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2640 if (Upper.isMinValue())
2641 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2642 }
2643
2644 if (!Add->hasOneUse())
2645 return nullptr;
2646
2647 // X+C <u C2 -> (X & -C2) == C
2648 // iff C & (C2-1) == 0
2649 // C2 is a power of 2
2650 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2651 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2652 ConstantExpr::getNeg(cast<Constant>(Y)));
2653
2654 // X+C >u C2 -> (X & ~C2) != C
2655 // iff C & C2 == 0
2656 // C2+1 is a power of 2
2657 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2658 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2659 ConstantExpr::getNeg(cast<Constant>(Y)));
2660
2661 return nullptr;
2662}
2663
2664bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2665 Value *&RHS, ConstantInt *&Less,
2666 ConstantInt *&Equal,
2667 ConstantInt *&Greater) {
2668 // TODO: Generalize this to work with other comparison idioms or ensure
2669 // they get canonicalized into this form.
2670
2671 // select i1 (a == b),
2672 // i32 Equal,
2673 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2674 // where Equal, Less and Greater are placeholders for any three constants.
2675 ICmpInst::Predicate PredA;
2676 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2677 !ICmpInst::isEquality(PredA))
2678 return false;
2679 Value *EqualVal = SI->getTrueValue();
2680 Value *UnequalVal = SI->getFalseValue();
2681 // We still can get non-canonical predicate here, so canonicalize.
2682 if (PredA == ICmpInst::ICMP_NE)
2683 std::swap(EqualVal, UnequalVal);
2684 if (!match(EqualVal, m_ConstantInt(Equal)))
2685 return false;
2686 ICmpInst::Predicate PredB;
2687 Value *LHS2, *RHS2;
2688 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2689 m_ConstantInt(Less), m_ConstantInt(Greater))))
2690 return false;
2691 // We can get predicate mismatch here, so canonicalize if possible:
2692 // First, ensure that 'LHS' match.
2693 if (LHS2 != LHS) {
2694 // x sgt y <--> y slt x
2695 std::swap(LHS2, RHS2);
2696 PredB = ICmpInst::getSwappedPredicate(PredB);
2697 }
2698 if (LHS2 != LHS)
2699 return false;
2700 // We also need to canonicalize 'RHS'.
2701 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2702 // x sgt C-1 <--> x sge C <--> not(x slt C)
2703 auto FlippedStrictness =
2704 InstCombiner::getFlippedStrictnessPredicateAndConstant(
2705 PredB, cast<Constant>(RHS2));
2706 if (!FlippedStrictness)
2707 return false;
2708 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check")((FlippedStrictness->first == ICmpInst::ICMP_SGE &&
"Sanity check") ? static_cast<void> (0) : __assert_fail
("FlippedStrictness->first == ICmpInst::ICMP_SGE && \"Sanity check\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2708, __PRETTY_FUNCTION__))
;
2709 RHS2 = FlippedStrictness->second;
2710 // And kind-of perform the result swap.
2711 std::swap(Less, Greater);
2712 PredB = ICmpInst::ICMP_SLT;
2713 }
2714 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2715}
2716
2717Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
2718 SelectInst *Select,
2719 ConstantInt *C) {
2720
2721 assert(C && "Cmp RHS should be a constant int!")((C && "Cmp RHS should be a constant int!") ? static_cast
<void> (0) : __assert_fail ("C && \"Cmp RHS should be a constant int!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2721, __PRETTY_FUNCTION__))
;
2722 // If we're testing a constant value against the result of a three way
2723 // comparison, the result can be expressed directly in terms of the
2724 // original values being compared. Note: We could possibly be more
2725 // aggressive here and remove the hasOneUse test. The original select is
2726 // really likely to simplify or sink when we remove a test of the result.
2727 Value *OrigLHS, *OrigRHS;
2728 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2729 if (Cmp.hasOneUse() &&
2730 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2731 C3GreaterThan)) {
2732 assert(C1LessThan && C2Equal && C3GreaterThan)((C1LessThan && C2Equal && C3GreaterThan) ? static_cast
<void> (0) : __assert_fail ("C1LessThan && C2Equal && C3GreaterThan"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2732, __PRETTY_FUNCTION__))
;
2733
2734 bool TrueWhenLessThan =
2735 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2736 ->isAllOnesValue();
2737 bool TrueWhenEqual =
2738 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2739 ->isAllOnesValue();
2740 bool TrueWhenGreaterThan =
2741 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2742 ->isAllOnesValue();
2743
2744 // This generates the new instruction that will replace the original Cmp
2745 // Instruction. Instead of enumerating the various combinations when
2746 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2747 // false, we rely on chaining of ORs and future passes of InstCombine to
2748 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2749
2750 // When none of the three constants satisfy the predicate for the RHS (C),
2751 // the entire original Cmp can be simplified to a false.
2752 Value *Cond = Builder.getFalse();
2753 if (TrueWhenLessThan)
2754 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2755 OrigLHS, OrigRHS));
2756 if (TrueWhenEqual)
2757 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2758 OrigLHS, OrigRHS));
2759 if (TrueWhenGreaterThan)
2760 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2761 OrigLHS, OrigRHS));
2762
2763 return replaceInstUsesWith(Cmp, Cond);
2764 }
2765 return nullptr;
2766}
2767
2768static Instruction *foldICmpBitCast(ICmpInst &Cmp,
2769 InstCombiner::BuilderTy &Builder) {
2770 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2771 if (!Bitcast)
2772 return nullptr;
2773
2774 ICmpInst::Predicate Pred = Cmp.getPredicate();
2775 Value *Op1 = Cmp.getOperand(1);
2776 Value *BCSrcOp = Bitcast->getOperand(0);
2777
2778 // Make sure the bitcast doesn't change the number of vector elements.
2779 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2780 Bitcast->getDestTy()->getScalarSizeInBits()) {
2781 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2782 Value *X;
2783 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2784 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2785 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2786 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2787 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2788 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2789 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2790 match(Op1, m_Zero()))
2791 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2792
2793 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2794 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2795 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2796
2797 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2798 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2799 return new ICmpInst(Pred, X,
2800 ConstantInt::getAllOnesValue(X->getType()));
2801 }
2802
2803 // Zero-equality checks are preserved through unsigned floating-point casts:
2804 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2805 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2806 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2807 if (Cmp.isEquality() && match(Op1, m_Zero()))
2808 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2809
2810 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2811 // the FP extend/truncate because that cast does not change the sign-bit.
2812 // This is true for all standard IEEE-754 types and the X86 80-bit type.
2813 // The sign-bit is always the most significant bit in those types.
2814 const APInt *C;
2815 bool TrueIfSigned;
2816 if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2817 InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2818 if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2819 match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2820 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2821 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2822 Type *XType = X->getType();
2823
2824 // We can't currently handle Power style floating point operations here.
2825 if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2826
2827 Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2828 if (auto *XVTy = dyn_cast<VectorType>(XType))
2829 NewType = FixedVectorType::get(NewType, XVTy->getNumElements());
2830 Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2831 if (TrueIfSigned)
2832 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2833 ConstantInt::getNullValue(NewType));
2834 else
2835 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2836 ConstantInt::getAllOnesValue(NewType));
2837 }
2838 }
2839 }
2840 }
2841
2842 // Test to see if the operands of the icmp are casted versions of other
2843 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2844 if (Bitcast->getType()->isPointerTy() &&
2845 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2846 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2847 // so eliminate it as well.
2848 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2849 Op1 = BC2->getOperand(0);
2850
2851 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2852 return new ICmpInst(Pred, BCSrcOp, Op1);
2853 }
2854
2855 // Folding: icmp <pred> iN X, C
2856 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2857 // and C is a splat of a K-bit pattern
2858 // and SC is a constant vector = <C', C', C', ..., C'>
2859 // Into:
2860 // %E = extractelement <M x iK> %vec, i32 C'
2861 // icmp <pred> iK %E, trunc(C)
2862 const APInt *C;
2863 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2864 !Bitcast->getType()->isIntegerTy() ||
2865 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2866 return nullptr;
2867
2868 Value *Vec;
2869 ArrayRef<int> Mask;
2870 if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2871 // Check whether every element of Mask is the same constant
2872 if (is_splat(Mask)) {
2873 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2874 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2875 if (C->isSplat(EltTy->getBitWidth())) {
2876 // Fold the icmp based on the value of C
2877 // If C is M copies of an iK sized bit pattern,
2878 // then:
2879 // => %E = extractelement <N x iK> %vec, i32 Elem
2880 // icmp <pred> iK %SplatVal, <pattern>
2881 Value *Elem = Builder.getInt32(Mask[0]);
2882 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2883 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2884 return new ICmpInst(Pred, Extract, NewC);
2885 }
2886 }
2887 }
2888 return nullptr;
2889}
2890
2891/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2892/// where X is some kind of instruction.
2893Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
2894 const APInt *C;
2895 if (!match(Cmp.getOperand(1), m_APInt(C)))
2896 return nullptr;
2897
2898 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2899 switch (BO->getOpcode()) {
2900 case Instruction::Xor:
2901 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2902 return I;
2903 break;
2904 case Instruction::And:
2905 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2906 return I;
2907 break;
2908 case Instruction::Or:
2909 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2910 return I;
2911 break;
2912 case Instruction::Mul:
2913 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2914 return I;
2915 break;
2916 case Instruction::Shl:
2917 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2918 return I;
2919 break;
2920 case Instruction::LShr:
2921 case Instruction::AShr:
2922 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2923 return I;
2924 break;
2925 case Instruction::SRem:
2926 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2927 return I;
2928 break;
2929 case Instruction::UDiv:
2930 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2931 return I;
2932 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2933 case Instruction::SDiv:
2934 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2935 return I;
2936 break;
2937 case Instruction::Sub:
2938 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2939 return I;
2940 break;
2941 case Instruction::Add:
2942 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2943 return I;
2944 break;
2945 default:
2946 break;
2947 }
2948 // TODO: These folds could be refactored to be part of the above calls.
2949 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2950 return I;
2951 }
2952
2953 // Match against CmpInst LHS being instructions other than binary operators.
2954
2955 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2956 // For now, we only support constant integers while folding the
2957 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2958 // similar to the cases handled by binary ops above.
2959 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2960 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2961 return I;
2962 }
2963
2964 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2965 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2966 return I;
2967 }
2968
2969 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2970 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2971 return I;
2972
2973 return nullptr;
2974}
2975
2976/// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2977/// icmp eq/ne BO, C.
2978Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
2979 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
2980 // TODO: Some of these folds could work with arbitrary constants, but this
2981 // function is limited to scalar and vector splat constants.
2982 if (!Cmp.isEquality())
2983 return nullptr;
2984
2985 ICmpInst::Predicate Pred = Cmp.getPredicate();
2986 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2987 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2988 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2989
2990 switch (BO->getOpcode()) {
2991 case Instruction::SRem:
2992 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2993 if (C.isNullValue() && BO->hasOneUse()) {
2994 const APInt *BOC;
2995 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2996 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2997 return new ICmpInst(Pred, NewRem,
2998 Constant::getNullValue(BO->getType()));
2999 }
3000 }
3001 break;
3002 case Instruction::Add: {
3003 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3004 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3005 if (BO->hasOneUse())
3006 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3007 } else if (C.isNullValue()) {
3008 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3009 // efficiently invertible, or if the add has just this one use.
3010 if (Value *NegVal = dyn_castNegVal(BOp1))
3011 return new ICmpInst(Pred, BOp0, NegVal);
3012 if (Value *NegVal = dyn_castNegVal(BOp0))
3013 return new ICmpInst(Pred, NegVal, BOp1);
3014 if (BO->hasOneUse()) {
3015 Value *Neg = Builder.CreateNeg(BOp1);
3016 Neg->takeName(BO);
3017 return new ICmpInst(Pred, BOp0, Neg);
3018 }
3019 }
3020 break;
3021 }
3022 case Instruction::Xor:
3023 if (BO->hasOneUse()) {
3024 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3025 // For the xor case, we can xor two constants together, eliminating
3026 // the explicit xor.
3027 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3028 } else if (C.isNullValue()) {
3029 // Replace ((xor A, B) != 0) with (A != B)
3030 return new ICmpInst(Pred, BOp0, BOp1);
3031 }
3032 }
3033 break;
3034 case Instruction::Sub:
3035 if (BO->hasOneUse()) {
3036 // Only check for constant LHS here, as constant RHS will be canonicalized
3037 // to add and use the fold above.
3038 if (Constant *BOC = dyn_cast<Constant>(BOp0)) {
3039 // Replace ((sub BOC, B) != C) with (B != BOC-C).
3040 return new ICmpInst(Pred, BOp1, ConstantExpr::getSub(BOC, RHS));
3041 } else if (C.isNullValue()) {
3042 // Replace ((sub A, B) != 0) with (A != B).
3043 return new ICmpInst(Pred, BOp0, BOp1);
3044 }
3045 }
3046 break;
3047 case Instruction::Or: {
3048 const APInt *BOC;
3049 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3050 // Comparing if all bits outside of a constant mask are set?
3051 // Replace (X | C) == -1 with (X & ~C) == ~C.
3052 // This removes the -1 constant.
3053 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3054 Value *And = Builder.CreateAnd(BOp0, NotBOC);
3055 return new ICmpInst(Pred, And, NotBOC);
3056 }
3057 break;
3058 }
3059 case Instruction::And: {
3060 const APInt *BOC;
3061 if (match(BOp1, m_APInt(BOC))) {
3062 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3063 if (C == *BOC && C.isPowerOf2())
3064 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3065 BO, Constant::getNullValue(RHS->getType()));
3066 }
3067 break;
3068 }
3069 case Instruction::UDiv:
3070 if (C.isNullValue()) {
3071 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3072 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3073 return new ICmpInst(NewPred, BOp1, BOp0);
3074 }
3075 break;
3076 default:
3077 break;
3078 }
3079 return nullptr;
3080}
3081
3082/// Fold an equality icmp with LLVM intrinsic and constant operand.
3083Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3084 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3085 Type *Ty = II->getType();
3086 unsigned BitWidth = C.getBitWidth();
3087 switch (II->getIntrinsicID()) {
3088 case Intrinsic::bswap:
3089 // bswap(A) == C -> A == bswap(C)
3090 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3091 ConstantInt::get(Ty, C.byteSwap()));
3092
3093 case Intrinsic::ctlz:
3094 case Intrinsic::cttz: {
3095 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3096 if (C == BitWidth)
3097 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3098 ConstantInt::getNullValue(Ty));
3099
3100 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3101 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3102 // Limit to one use to ensure we don't increase instruction count.
3103 unsigned Num = C.getLimitedValue(BitWidth);
3104 if (Num != BitWidth && II->hasOneUse()) {
3105 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3106 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3107 : APInt::getHighBitsSet(BitWidth, Num + 1);
3108 APInt Mask2 = IsTrailing
3109 ? APInt::getOneBitSet(BitWidth, Num)
3110 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3111 return new ICmpInst(Cmp.getPredicate(),
3112 Builder.CreateAnd(II->getArgOperand(0), Mask1),
3113 ConstantInt::get(Ty, Mask2));
3114 }
3115 break;
3116 }
3117
3118 case Intrinsic::ctpop: {
3119 // popcount(A) == 0 -> A == 0 and likewise for !=
3120 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3121 bool IsZero = C.isNullValue();
3122 if (IsZero || C == BitWidth)
3123 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3124 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty));
3125
3126 break;
3127 }
3128
3129 case Intrinsic::uadd_sat: {
3130 // uadd.sat(a, b) == 0 -> (a | b) == 0
3131 if (C.isNullValue()) {
3132 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3133 return new ICmpInst(Cmp.getPredicate(), Or, Constant::getNullValue(Ty));
3134 }
3135 break;
3136 }
3137
3138 case Intrinsic::usub_sat: {
3139 // usub.sat(a, b) == 0 -> a <= b
3140 if (C.isNullValue()) {
3141 ICmpInst::Predicate NewPred = Cmp.getPredicate() == ICmpInst::ICMP_EQ
3142 ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3143 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3144 }
3145 break;
3146 }
3147 default:
3148 break;
3149 }
3150
3151 return nullptr;
3152}
3153
3154/// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3155Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3156 IntrinsicInst *II,
3157 const APInt &C) {
3158 if (Cmp.isEquality())
3159 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3160
3161 Type *Ty = II->getType();
3162 unsigned BitWidth = C.getBitWidth();
3163 switch (II->getIntrinsicID()) {
3164 case Intrinsic::ctlz: {
3165 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3166 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3167 unsigned Num = C.getLimitedValue();
3168 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3169 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3170 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3171 }
3172
3173 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3174 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3175 C.uge(1) && C.ule(BitWidth)) {
3176 unsigned Num = C.getLimitedValue();
3177 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3178 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3179 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3180 }
3181 break;
3182 }
3183 case Intrinsic::cttz: {
3184 // Limit to one use to ensure we don't increase instruction count.
3185 if (!II->hasOneUse())
3186 return nullptr;
3187
3188 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3189 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3190 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3191 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3192 Builder.CreateAnd(II->getArgOperand(0), Mask),
3193 ConstantInt::getNullValue(Ty));
3194 }
3195
3196 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3197 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3198 C.uge(1) && C.ule(BitWidth)) {
3199 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3200 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3201 Builder.CreateAnd(II->getArgOperand(0), Mask),
3202 ConstantInt::getNullValue(Ty));
3203 }
3204 break;
3205 }
3206 default:
3207 break;
3208 }
3209
3210 return nullptr;
3211}
3212
3213/// Handle icmp with constant (but not simple integer constant) RHS.
3214Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3215 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3216 Constant *RHSC = dyn_cast<Constant>(Op1);
3217 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3218 if (!RHSC || !LHSI)
3219 return nullptr;
3220
3221 switch (LHSI->getOpcode()) {
3222 case Instruction::GetElementPtr:
3223 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3224 if (RHSC->isNullValue() &&
3225 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3226 return new ICmpInst(
3227 I.getPredicate(), LHSI->getOperand(0),
3228 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3229 break;
3230 case Instruction::PHI:
3231 // Only fold icmp into the PHI if the phi and icmp are in the same
3232 // block. If in the same block, we're encouraging jump threading. If
3233 // not, we are just pessimizing the code by making an i1 phi.
3234 if (LHSI->getParent() == I.getParent())
3235 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3236 return NV;
3237 break;
3238 case Instruction::Select: {
3239 // If either operand of the select is a constant, we can fold the
3240 // comparison into the select arms, which will cause one to be
3241 // constant folded and the select turned into a bitwise or.
3242 Value *Op1 = nullptr, *Op2 = nullptr;
3243 ConstantInt *CI = nullptr;
3244 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3245 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3246 CI = dyn_cast<ConstantInt>(Op1);
3247 }
3248 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3249 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3250 CI = dyn_cast<ConstantInt>(Op2);
3251 }
3252
3253 // We only want to perform this transformation if it will not lead to
3254 // additional code. This is true if either both sides of the select
3255 // fold to a constant (in which case the icmp is replaced with a select
3256 // which will usually simplify) or this is the only user of the
3257 // select (in which case we are trading a select+icmp for a simpler
3258 // select+icmp) or all uses of the select can be replaced based on
3259 // dominance information ("Global cases").
3260 bool Transform = false;
3261 if (Op1 && Op2)
3262 Transform = true;
3263 else if (Op1 || Op2) {
3264 // Local case
3265 if (LHSI->hasOneUse())
3266 Transform = true;
3267 // Global cases
3268 else if (CI && !CI->isZero())
3269 // When Op1 is constant try replacing select with second operand.
3270 // Otherwise Op2 is constant and try replacing select with first
3271 // operand.
3272 Transform =
3273 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3274 }
3275 if (Transform) {
3276 if (!Op1)
3277 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3278 I.getName());
3279 if (!Op2)
3280 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3281 I.getName());
3282 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3283 }
3284 break;
3285 }
3286 case Instruction::IntToPtr:
3287 // icmp pred inttoptr(X), null -> icmp pred X, 0
3288 if (RHSC->isNullValue() &&
3289 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3290 return new ICmpInst(
3291 I.getPredicate(), LHSI->getOperand(0),
3292 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3293 break;
3294
3295 case Instruction::Load:
3296 // Try to optimize things like "A[i] > 4" to index computations.
3297 if (GetElementPtrInst *GEP =
3298 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3299 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3300 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3301 !cast<LoadInst>(LHSI)->isVolatile())
3302 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3303 return Res;
3304 }
3305 break;
3306 }
3307
3308 return nullptr;
3309}
3310
3311/// Some comparisons can be simplified.
3312/// In this case, we are looking for comparisons that look like
3313/// a check for a lossy truncation.
3314/// Folds:
3315/// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3316/// Where Mask is some pattern that produces all-ones in low bits:
3317/// (-1 >> y)
3318/// ((-1 << y) >> y) <- non-canonical, has extra uses
3319/// ~(-1 << y)
3320/// ((1 << y) + (-1)) <- non-canonical, has extra uses
3321/// The Mask can be a constant, too.
3322/// For some predicates, the operands are commutative.
3323/// For others, x can only be on a specific side.
3324static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3325 InstCombiner::BuilderTy &Builder) {
3326 ICmpInst::Predicate SrcPred;
3327 Value *X, *M, *Y;
3328 auto m_VariableMask = m_CombineOr(
3329 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3330 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3331 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3332 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3333 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3334 if (!match(&I, m_c_ICmp(SrcPred,
3335 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3336 m_Deferred(X))))
3337 return nullptr;
3338
3339 ICmpInst::Predicate DstPred;
3340 switch (SrcPred) {
3341 case ICmpInst::Predicate::ICMP_EQ:
3342 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3343 DstPred = ICmpInst::Predicate::ICMP_ULE;
3344 break;
3345 case ICmpInst::Predicate::ICMP_NE:
3346 // x & (-1 >> y) != x -> x u> (-1 >> y)
3347 DstPred = ICmpInst::Predicate::ICMP_UGT;
3348 break;
3349 case ICmpInst::Predicate::ICMP_ULT:
3350 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3351 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3352 DstPred = ICmpInst::Predicate::ICMP_UGT;
3353 break;
3354 case ICmpInst::Predicate::ICMP_UGE:
3355 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3356 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3357 DstPred = ICmpInst::Predicate::ICMP_ULE;
3358 break;
3359 case ICmpInst::Predicate::ICMP_SLT:
3360 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3361 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3362 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3363 return nullptr;
3364 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3365 return nullptr;
3366 DstPred = ICmpInst::Predicate::ICMP_SGT;
3367 break;
3368 case ICmpInst::Predicate::ICMP_SGE:
3369 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3370 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3371 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3372 return nullptr;
3373 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3374 return nullptr;
3375 DstPred = ICmpInst::Predicate::ICMP_SLE;
3376 break;
3377 case ICmpInst::Predicate::ICMP_SGT:
3378 case ICmpInst::Predicate::ICMP_SLE:
3379 return nullptr;
3380 case ICmpInst::Predicate::ICMP_UGT:
3381 case ICmpInst::Predicate::ICMP_ULE:
3382 llvm_unreachable("Instsimplify took care of commut. variant")::llvm::llvm_unreachable_internal("Instsimplify took care of commut. variant"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3382)
;
3383 break;
3384 default:
3385 llvm_unreachable("All possible folds are handled.")::llvm::llvm_unreachable_internal("All possible folds are handled."
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3385)
;
3386 }
3387
3388 // The mask value may be a vector constant that has undefined elements. But it
3389 // may not be safe to propagate those undefs into the new compare, so replace
3390 // those elements by copying an existing, defined, and safe scalar constant.
3391 Type *OpTy = M->getType();
3392 auto *VecC = dyn_cast<Constant>(M);
3393 if (OpTy->isVectorTy() && VecC && VecC->containsUndefElement()) {
3394 auto *OpVTy = cast<VectorType>(OpTy);
3395 Constant *SafeReplacementConstant = nullptr;
3396 for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3397 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3398 SafeReplacementConstant = VecC->getAggregateElement(i);
3399 break;
3400 }
3401 }
3402 assert(SafeReplacementConstant && "Failed to find undef replacement")((SafeReplacementConstant && "Failed to find undef replacement"
) ? static_cast<void> (0) : __assert_fail ("SafeReplacementConstant && \"Failed to find undef replacement\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3402, __PRETTY_FUNCTION__))
;
3403 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3404 }
3405
3406 return Builder.CreateICmp(DstPred, X, M);
3407}
3408
3409/// Some comparisons can be simplified.
3410/// In this case, we are looking for comparisons that look like
3411/// a check for a lossy signed truncation.
3412/// Folds: (MaskedBits is a constant.)
3413/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3414/// Into:
3415/// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3416/// Where KeptBits = bitwidth(%x) - MaskedBits
3417static Value *
3418foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3419 InstCombiner::BuilderTy &Builder) {
3420 ICmpInst::Predicate SrcPred;
3421 Value *X;
3422 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3423 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3424 if (!match(&I, m_c_ICmp(SrcPred,
3425 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3426 m_APInt(C1))),
3427 m_Deferred(X))))
3428 return nullptr;
3429
3430 // Potential handling of non-splats: for each element:
3431 // * if both are undef, replace with constant 0.
3432 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3433 // * if both are not undef, and are different, bailout.
3434 // * else, only one is undef, then pick the non-undef one.
3435
3436 // The shift amount must be equal.
3437 if (*C0 != *C1)
3438 return nullptr;
3439 const APInt &MaskedBits = *C0;
3440 assert(MaskedBits != 0 && "shift by zero should be folded away already.")((MaskedBits != 0 && "shift by zero should be folded away already."
) ? static_cast<void> (0) : __assert_fail ("MaskedBits != 0 && \"shift by zero should be folded away already.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3440, __PRETTY_FUNCTION__))
;
3441
3442 ICmpInst::Predicate DstPred;
3443 switch (SrcPred) {
3444 case ICmpInst::Predicate::ICMP_EQ:
3445 // ((%x << MaskedBits) a>> MaskedBits) == %x
3446 // =>
3447 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3448 DstPred = ICmpInst::Predicate::ICMP_ULT;
3449 break;
3450 case ICmpInst::Predicate::ICMP_NE:
3451 // ((%x << MaskedBits) a>> MaskedBits) != %x
3452 // =>
3453 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3454 DstPred = ICmpInst::Predicate::ICMP_UGE;
3455 break;
3456 // FIXME: are more folds possible?
3457 default:
3458 return nullptr;
3459 }
3460
3461 auto *XType = X->getType();
3462 const unsigned XBitWidth = XType->getScalarSizeInBits();
3463 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3464 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched")((BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched"
) ? static_cast<void> (0) : __assert_fail ("BitWidth.ugt(MaskedBits) && \"shifts should leave some bits untouched\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3464, __PRETTY_FUNCTION__))
;
3465
3466 // KeptBits = bitwidth(%x) - MaskedBits
3467 const APInt KeptBits = BitWidth - MaskedBits;
3468 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable")((KeptBits.ugt(0) && KeptBits.ult(BitWidth) &&
"unreachable") ? static_cast<void> (0) : __assert_fail
("KeptBits.ugt(0) && KeptBits.ult(BitWidth) && \"unreachable\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3468, __PRETTY_FUNCTION__))
;
3469 // ICmpCst = (1 << KeptBits)
3470 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3471 assert(ICmpCst.isPowerOf2())((ICmpCst.isPowerOf2()) ? static_cast<void> (0) : __assert_fail
("ICmpCst.isPowerOf2()", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3471, __PRETTY_FUNCTION__))
;
3472 // AddCst = (1 << (KeptBits-1))
3473 const APInt AddCst = ICmpCst.lshr(1);
3474 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2())((AddCst.ult(ICmpCst) && AddCst.isPowerOf2()) ? static_cast
<void> (0) : __assert_fail ("AddCst.ult(ICmpCst) && AddCst.isPowerOf2()"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3474, __PRETTY_FUNCTION__))
;
3475
3476 // T0 = add %x, AddCst
3477 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3478 // T1 = T0 DstPred ICmpCst
3479 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3480
3481 return T1;
3482}
3483
3484// Given pattern:
3485// icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3486// we should move shifts to the same hand of 'and', i.e. rewrite as
3487// icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3488// We are only interested in opposite logical shifts here.
3489// One of the shifts can be truncated.
3490// If we can, we want to end up creating 'lshr' shift.
3491static Value *
3492foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3493 InstCombiner::BuilderTy &Builder) {
3494 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
31
Assuming the condition is false
33
Taking false branch
3495 !I.getOperand(0)->hasOneUse())
32
Assuming the condition is false
3496 return nullptr;
3497
3498 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3499
3500 // Look for an 'and' of two logical shifts, one of which may be truncated.
3501 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3502 Instruction *XShift, *MaybeTruncation, *YShift;
3503 if (!match(
34
Calling 'match<llvm::Value, llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 28, true>>'
42
Returning from 'match<llvm::Value, llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 28, true>>'
43
Taking false branch
3504 I.getOperand(0),
3505 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3506 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3507 m_AnyLogicalShift, m_Instruction(YShift))),
3508 m_Instruction(MaybeTruncation)))))
3509 return nullptr;
3510
3511 // We potentially looked past 'trunc', but only when matching YShift,
3512 // therefore YShift must have the widest type.
3513 Instruction *WidestShift = YShift;
3514 // Therefore XShift must have the shallowest type.
3515 // Or they both have identical types if there was no truncation.
3516 Instruction *NarrowestShift = XShift;
3517
3518 Type *WidestTy = WidestShift->getType();
3519 Type *NarrowestTy = NarrowestShift->getType();
3520 assert(NarrowestTy == I.getOperand(0)->getType() &&((NarrowestTy == I.getOperand(0)->getType() && "We did not look past any shifts while matching XShift though."
) ? static_cast<void> (0) : __assert_fail ("NarrowestTy == I.getOperand(0)->getType() && \"We did not look past any shifts while matching XShift though.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3521, __PRETTY_FUNCTION__))
44
Assuming the condition is true
45
'?' condition is true
3521 "We did not look past any shifts while matching XShift though.")((NarrowestTy == I.getOperand(0)->getType() && "We did not look past any shifts while matching XShift though."
) ? static_cast<void> (0) : __assert_fail ("NarrowestTy == I.getOperand(0)->getType() && \"We did not look past any shifts while matching XShift though.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3521, __PRETTY_FUNCTION__))
;
3522 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
46
Assuming the condition is false
3523
3524 // If YShift is a 'lshr', swap the shifts around.
3525 if (match(YShift, m_LShr(m_Value(), m_Value())))
47
Taking false branch
3526 std::swap(XShift, YShift);
3527
3528 // The shifts must be in opposite directions.
3529 auto XShiftOpcode = XShift->getOpcode();
3530 if (XShiftOpcode == YShift->getOpcode())
48
Assuming the condition is false
49
Taking false branch
3531 return nullptr; // Do not care about same-direction shifts here.
3532
3533 Value *X, *XShAmt, *Y, *YShAmt;
50
'XShAmt' declared without an initial value
3534 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
51
Calling 'm_Value'
55
Returning from 'm_Value'
56
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
64
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
65
Calling 'm_BinOp<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'
67
Returning from 'm_BinOp<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'
3535 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3536
3537 // If one of the values being shifted is a constant, then we will end with
3538 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3539 // however, we will need to ensure that we won't increase instruction count.
3540 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
68
Assuming 'X' is a 'Constant'
3541 // At least one of the hands of the 'and' should be one-use shift.
3542 if (!match(I.getOperand(0),
3543 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3544 return nullptr;
3545 if (HadTrunc) {
3546 // Due to the 'trunc', we will need to widen X. For that either the old
3547 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3548 if (!MaybeTruncation->hasOneUse() &&
3549 !NarrowestShift->getOperand(1)->hasOneUse())
3550 return nullptr;
3551 }
3552 }
3553
3554 // We have two shift amounts from two different shifts. The types of those
3555 // shift amounts may not match. If that's the case let's bailout now.
3556 if (XShAmt->getType() != YShAmt->getType())
69
Called C++ object pointer is uninitialized
3557 return nullptr;
3558
3559 // As input, we have the following pattern:
3560 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3561 // We want to rewrite that as:
3562 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3563 // While we know that originally (Q+K) would not overflow
3564 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3565 // shift amounts. so it may now overflow in smaller bitwidth.
3566 // To ensure that does not happen, we need to ensure that the total maximal
3567 // shift amount is still representable in that smaller bit width.
3568 unsigned MaximalPossibleTotalShiftAmount =
3569 (WidestTy->getScalarSizeInBits() - 1) +
3570 (NarrowestTy->getScalarSizeInBits() - 1);
3571 APInt MaximalRepresentableShiftAmount =
3572 APInt::getAllOnesValue(XShAmt->getType()->getScalarSizeInBits());
3573 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3574 return nullptr;
3575
3576 // Can we fold (XShAmt+YShAmt) ?
3577 auto *NewShAmt = dyn_cast_or_null<Constant>(
3578 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3579 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3580 if (!NewShAmt)
3581 return nullptr;
3582 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3583 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3584
3585 // Is the new shift amount smaller than the bit width?
3586 // FIXME: could also rely on ConstantRange.
3587 if (!match(NewShAmt,
3588 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3589 APInt(WidestBitWidth, WidestBitWidth))))
3590 return nullptr;
3591
3592 // An extra legality check is needed if we had trunc-of-lshr.
3593 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3594 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3595 WidestShift]() {
3596 // It isn't obvious whether it's worth it to analyze non-constants here.
3597 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3598 // If *any* of these preconditions matches we can perform the fold.
3599 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3600 ? NewShAmt->getSplatValue()
3601 : NewShAmt;
3602 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3603 if (NewShAmtSplat &&
3604 (NewShAmtSplat->isNullValue() ||
3605 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3606 return true;
3607 // We consider *min* leading zeros so a single outlier
3608 // blocks the transform as opposed to allowing it.
3609 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3610 KnownBits Known = computeKnownBits(C, SQ.DL);
3611 unsigned MinLeadZero = Known.countMinLeadingZeros();
3612 // If the value being shifted has at most lowest bit set we can fold.
3613 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3614 if (MaxActiveBits <= 1)
3615 return true;
3616 // Precondition: NewShAmt u<= countLeadingZeros(C)
3617 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3618 return true;
3619 }
3620 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3621 KnownBits Known = computeKnownBits(C, SQ.DL);
3622 unsigned MinLeadZero = Known.countMinLeadingZeros();
3623 // If the value being shifted has at most lowest bit set we can fold.
3624 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3625 if (MaxActiveBits <= 1)
3626 return true;
3627 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3628 if (NewShAmtSplat) {
3629 APInt AdjNewShAmt =
3630 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3631 if (AdjNewShAmt.ule(MinLeadZero))
3632 return true;
3633 }
3634 }
3635 return false; // Can't tell if it's ok.
3636 };
3637 if (!CanFold())
3638 return nullptr;
3639 }
3640
3641 // All good, we can do this fold.
3642 X = Builder.CreateZExt(X, WidestTy);
3643 Y = Builder.CreateZExt(Y, WidestTy);
3644 // The shift is the same that was for X.
3645 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3646 ? Builder.CreateLShr(X, NewShAmt)
3647 : Builder.CreateShl(X, NewShAmt);
3648 Value *T1 = Builder.CreateAnd(T0, Y);
3649 return Builder.CreateICmp(I.getPredicate(), T1,
3650 Constant::getNullValue(WidestTy));
3651}
3652
3653/// Fold
3654/// (-1 u/ x) u< y
3655/// ((x * y) u/ x) != y
3656/// to
3657/// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3658/// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3659/// will mean that we are looking for the opposite answer.
3660Value *InstCombinerImpl::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3661 ICmpInst::Predicate Pred;
3662 Value *X, *Y;
3663 Instruction *Mul;
3664 bool NeedNegation;
3665 // Look for: (-1 u/ x) u</u>= y
3666 if (!I.isEquality() &&
3667 match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3668 m_Value(Y)))) {
3669 Mul = nullptr;
3670
3671 // Are we checking that overflow does not happen, or does happen?
3672 switch (Pred) {
3673 case ICmpInst::Predicate::ICMP_ULT:
3674 NeedNegation = false;
3675 break; // OK
3676 case ICmpInst::Predicate::ICMP_UGE:
3677 NeedNegation = true;
3678 break; // OK
3679 default:
3680 return nullptr; // Wrong predicate.
3681 }
3682 } else // Look for: ((x * y) u/ x) !=/== y
3683 if (I.isEquality() &&
3684 match(&I, m_c_ICmp(Pred, m_Value(Y),
3685 m_OneUse(m_UDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3686 m_Value(X)),
3687 m_Instruction(Mul)),
3688 m_Deferred(X)))))) {
3689 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3690 } else
3691 return nullptr;
3692
3693 BuilderTy::InsertPointGuard Guard(Builder);
3694 // If the pattern included (x * y), we'll want to insert new instructions
3695 // right before that original multiplication so that we can replace it.
3696 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3697 if (MulHadOtherUses)
3698 Builder.SetInsertPoint(Mul);
3699
3700 Function *F = Intrinsic::getDeclaration(
3701 I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3702 CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3703
3704 // If the multiplication was used elsewhere, to ensure that we don't leave
3705 // "duplicate" instructions, replace uses of that original multiplication
3706 // with the multiplication result from the with.overflow intrinsic.
3707 if (MulHadOtherUses)
3708 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3709
3710 Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3711 if (NeedNegation) // This technically increases instruction count.
3712 Res = Builder.CreateNot(Res, "umul.not.ov");
3713
3714 // If we replaced the mul, erase it. Do this after all uses of Builder,
3715 // as the mul is used as insertion point.
3716 if (MulHadOtherUses)
3717 eraseInstFromFunction(*Mul);
3718
3719 return Res;
3720}
3721
3722Instruction *foldICmpXNegX(ICmpInst &I) {
3723 CmpInst::Predicate Pred;
3724 Value *X;
3725 if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3726 return nullptr;
3727
3728 if (ICmpInst::isSigned(Pred))
3729 Pred = ICmpInst::getSwappedPredicate(Pred);
3730 else if (ICmpInst::isUnsigned(Pred))
3731 Pred = ICmpInst::getSignedPredicate(Pred);
3732 // else for equality-comparisons just keep the predicate.
3733
3734 return ICmpInst::Create(Instruction::ICmp, Pred, X,
3735 Constant::getNullValue(X->getType()), I.getName());
3736}
3737
3738/// Try to fold icmp (binop), X or icmp X, (binop).
3739/// TODO: A large part of this logic is duplicated in InstSimplify's
3740/// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3741/// duplication.
3742Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
3743 const SimplifyQuery &SQ) {
3744 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3745 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3746
3747 // Special logic for binary operators.
3748 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
1
Assuming 'Op0' is a 'BinaryOperator'
3749 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2
Assuming 'Op1' is not a 'BinaryOperator'
3750 if (!BO0
2.1
'BO0' is non-null
2.1
'BO0' is non-null
&& !BO1)
3751 return nullptr;
3752
3753 if (Instruction *NewICmp
2.2
'NewICmp' is null
2.2
'NewICmp' is null
= foldICmpXNegX(I))
3
Taking false branch
3754 return NewICmp;
3755
3756 const CmpInst::Predicate Pred = I.getPredicate();
3757 Value *X;
3758
3759 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3760 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3761 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
4
Taking false branch
3762 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3763 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3764 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3765 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
5
Taking false branch
3766 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3767 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3768
3769 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3770 if (BO0
5.1
'BO0' is non-null
5.1
'BO0' is non-null
&& isa<OverflowingBinaryOperator>(BO0))
6
Assuming 'BO0' is not a 'OverflowingBinaryOperator'
7
Taking false branch
3771 NoOp0WrapProblem =
3772 ICmpInst::isEquality(Pred) ||
3773 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3774 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3775 if (BO1
7.1
'BO1' is null
7.1
'BO1' is null
&& isa<OverflowingBinaryOperator>(BO1))
3776 NoOp1WrapProblem =
3777 ICmpInst::isEquality(Pred) ||
3778 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3779 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3780
3781 // Analyze the case when either Op0 or Op1 is an add instruction.
3782 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3783 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3784 if (BO0
7.2
'BO0' is non-null
7.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Add) {
8
Assuming the condition is false
9
Taking false branch
3785 A = BO0->getOperand(0);
3786 B = BO0->getOperand(1);
3787 }
3788 if (BO1
9.1
'BO1' is null
9.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Add) {
3789 C = BO1->getOperand(0);
3790 D = BO1->getOperand(1);
3791 }
3792
3793 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3794 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3795 if ((A
9.2
'A' is not equal to 'Op1'
9.2
'A' is not equal to 'Op1'
== Op1 || B
9.3
'B' is not equal to 'Op1'
9.3
'B' is not equal to 'Op1'
== Op1) && NoOp0WrapProblem)
3796 return new ICmpInst(Pred, A == Op1 ? B : A,
3797 Constant::getNullValue(Op1->getType()));
3798
3799 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3800 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3801 if ((C
9.4
'C' is not equal to 'Op0'
9.4
'C' is not equal to 'Op0'
== Op0 || D
9.5
'D' is not equal to 'Op0'
9.5
'D' is not equal to 'Op0'
== Op0) && NoOp1WrapProblem)
3802 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3803 C == Op0 ? D : C);
3804
3805 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3806 if (A
9.6
'A' is null
9.6
'A' is null
&& C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3807 NoOp1WrapProblem) {
3808 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3809 Value *Y, *Z;
3810 if (A == C) {
3811 // C + B == C + D -> B == D
3812 Y = B;
3813 Z = D;
3814 } else if (A == D) {
3815 // D + B == C + D -> B == C
3816 Y = B;
3817 Z = C;
3818 } else if (B == C) {
3819 // A + C == C + D -> A == D
3820 Y = A;
3821 Z = D;
3822 } else {
3823 assert(B == D)((B == D) ? static_cast<void> (0) : __assert_fail ("B == D"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3823, __PRETTY_FUNCTION__))
;
3824 // A + D == C + D -> A == C
3825 Y = A;
3826 Z = C;
3827 }
3828 return new ICmpInst(Pred, Y, Z);
3829 }
3830
3831 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3832 if (A
9.7
'A' is null
9.7
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
10
Taking false branch
3833 match(B, m_AllOnes()))
3834 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3835
3836 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3837 if (A
10.1
'A' is null
10.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
11
Taking false branch
3838 match(B, m_AllOnes()))
3839 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3840
3841 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3842 if (A
11.1
'A' is null
11.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
12
Taking false branch
3843 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3844
3845 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3846 if (A
12.1
'A' is null
12.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
13
Taking false branch
3847 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3848
3849 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3850 if (C
13.1
'C' is null
13.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
14
Taking false branch
3851 match(D, m_AllOnes()))
3852 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3853
3854 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3855 if (C
14.1
'C' is null
14.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
15
Taking false branch
3856 match(D, m_AllOnes()))
3857 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3858
3859 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3860 if (C
15.1
'C' is null
15.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
16
Taking false branch
3861 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3862
3863 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3864 if (C
16.1
'C' is null
16.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
17
Taking false branch
3865 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3866
3867 // TODO: The subtraction-related identities shown below also hold, but
3868 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3869 // wouldn't happen even if they were implemented.
3870 //
3871 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3872 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3873 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3874 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3875
3876 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3877 if (A
17.1
'A' is null
17.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
18
Taking false branch
3878 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3879
3880 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3881 if (A
18.1
'A' is null
18.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
19
Taking false branch
3882 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3883
3884 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3885 if (C
19.1
'C' is null
19.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
20
Taking false branch
3886 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3887
3888 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3889 if (C
20.1
'C' is null
20.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
21
Taking false branch
3890 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3891
3892 // if C1 has greater magnitude than C2:
3893 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3894 // s.t. C3 = C1 - C2
3895 //
3896 // if C2 has greater magnitude than C1:
3897 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3898 // s.t. C3 = C2 - C1
3899 if (A
21.1
'A' is null
21.1
'A' is null
&& C && NoOp0WrapProblem && NoOp1WrapProblem &&
3900 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3901 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3902 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3903 const APInt &AP1 = C1->getValue();
3904 const APInt &AP2 = C2->getValue();
3905 if (AP1.isNegative() == AP2.isNegative()) {
3906 APInt AP1Abs = C1->getValue().abs();
3907 APInt AP2Abs = C2->getValue().abs();
3908 if (AP1Abs.uge(AP2Abs)) {
3909 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3910 Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3911 return new ICmpInst(Pred, NewAdd, C);
3912 } else {
3913 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3914 Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3915 return new ICmpInst(Pred, A, NewAdd);
3916 }
3917 }
3918 }
3919
3920 // Analyze the case when either Op0 or Op1 is a sub instruction.
3921 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3922 A = nullptr;
3923 B = nullptr;
3924 C = nullptr;
3925 D = nullptr;
3926 if (BO0
21.2
'BO0' is non-null
21.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Sub) {
22
Assuming the condition is false
23
Taking false branch
3927 A = BO0->getOperand(0);
3928 B = BO0->getOperand(1);
3929 }
3930 if (BO1
23.1
'BO1' is null
23.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Sub) {
3931 C = BO1->getOperand(0);
3932 D = BO1->getOperand(1);
3933 }
3934
3935 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3936 if (A
23.2
'A' is not equal to 'Op1'
23.2
'A' is not equal to 'Op1'
== Op1 && NoOp0WrapProblem)
3937 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3938 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
3939 if (C
23.3
'C' is not equal to 'Op0'
23.3
'C' is not equal to 'Op0'
== Op0 && NoOp1WrapProblem)
3940 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3941
3942 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3943 // (A - B) u>/u<= A --> B u>/u<= A
3944 if (A
23.4
'A' is not equal to 'Op1'
23.4
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3945 return new ICmpInst(Pred, B, A);
3946 // C u</u>= (C - D) --> C u</u>= D
3947 if (C
23.5
'C' is not equal to 'Op0'
23.5
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3948 return new ICmpInst(Pred, C, D);
3949 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
3950 if (A
23.6
'A' is not equal to 'Op1'
23.6
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
3951 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3952 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
3953 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
3954 if (C
23.7
'C' is not equal to 'Op0'
23.7
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
3955 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3956 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
3957
3958 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
3959 if (B
23.8
'B' is null
23.8
'B' is null
&& D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
3960 return new ICmpInst(Pred, A, C);
3961
3962 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
3963 if (A
23.9
'A' is null
23.9
'A' is null
&& C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
3964 return new ICmpInst(Pred, D, B);
3965
3966 // icmp (0-X) < cst --> x > -cst
3967 if (NoOp0WrapProblem
23.10
'NoOp0WrapProblem' is false
23.10
'NoOp0WrapProblem' is false
&& ICmpInst::isSigned(Pred)) {
3968 Value *X;
3969 if (match(BO0, m_Neg(m_Value(X))))
3970 if (Constant *RHSC = dyn_cast<Constant>(Op1))
3971 if (RHSC->isNotMinSignedValue())
3972 return new ICmpInst(I.getSwappedPredicate(), X,
3973 ConstantExpr::getNeg(RHSC));
3974 }
3975
3976 BinaryOperator *SRem = nullptr;
3977 // icmp (srem X, Y), Y
3978 if (BO0
23.11
'BO0' is non-null
23.11
'BO0' is non-null
&& BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
24
Assuming the condition is false
3979 SRem = BO0;
3980 // icmp Y, (srem X, Y)
3981 else if (BO1
24.1
'BO1' is null
24.1
'BO1' is null
&& BO1->getOpcode() == Instruction::SRem &&
3982 Op0 == BO1->getOperand(1))
3983 SRem = BO1;
3984 if (SRem
24.2
'SRem' is null
24.2
'SRem' is null
) {
25
Taking false branch
3985 // We don't check hasOneUse to avoid increasing register pressure because
3986 // the value we use is the same value this instruction was already using.
3987 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3988 default:
3989 break;
3990 case ICmpInst::ICMP_EQ:
3991 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3992 case ICmpInst::ICMP_NE:
3993 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3994 case ICmpInst::ICMP_SGT:
3995 case ICmpInst::ICMP_SGE:
3996 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3997 Constant::getAllOnesValue(SRem->getType()));
3998 case ICmpInst::ICMP_SLT:
3999 case ICmpInst::ICMP_SLE:
4000 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4001 Constant::getNullValue(SRem->getType()));
4002 }
4003 }
4004
4005 if (BO0
25.1
'BO0' is non-null
25.1
'BO0' is non-null
&& BO1
25.2
'BO1' is null
25.2
'BO1' is null
&& BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4006 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4007 switch (BO0->getOpcode()) {
4008 default:
4009 break;
4010 case Instruction::Add:
4011 case Instruction::Sub:
4012 case Instruction::Xor: {
4013 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4014 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4015
4016 const APInt *C;
4017 if (match(BO0->getOperand(1), m_APInt(C))) {
4018 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4019 if (C->isSignMask()) {
4020 ICmpInst::Predicate NewPred =
4021 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
4022 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4023 }
4024
4025 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4026 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4027 ICmpInst::Predicate NewPred =
4028 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
4029 NewPred = I.getSwappedPredicate(NewPred);
4030 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4031 }
4032 }
4033 break;
4034 }
4035 case Instruction::Mul: {
4036 if (!I.isEquality())
4037 break;
4038
4039 const APInt *C;
4040 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
4041 !C->isOneValue()) {
4042 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4043 // Mask = -1 >> count-trailing-zeros(C).
4044 if (unsigned TZs = C->countTrailingZeros()) {
4045 Constant *Mask = ConstantInt::get(
4046 BO0->getType(),
4047 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4048 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4049 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4050 return new ICmpInst(Pred, And1, And2);
4051 }
4052 // If there are no trailing zeros in the multiplier, just eliminate
4053 // the multiplies (no masking is needed):
4054 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
4055 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4056 }
4057 break;
4058 }
4059 case Instruction::UDiv:
4060 case Instruction::LShr:
4061 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4062 break;
4063 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4064
4065 case Instruction::SDiv:
4066 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4067 break;
4068 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4069
4070 case Instruction::AShr:
4071 if (!BO0->isExact() || !BO1->isExact())
4072 break;
4073 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4074
4075 case Instruction::Shl: {
4076 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4077 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4078 if (!NUW && !NSW)
4079 break;
4080 if (!NSW && I.isSigned())
4081 break;
4082 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4083 }
4084 }
4085 }
4086
4087 if (BO0
25.3
'BO0' is non-null
25.3
'BO0' is non-null
) {
26
Taking true branch
4088 // Transform A & (L - 1) `ult` L --> L != 0
4089 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4090 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4091
4092 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4093 auto *Zero = Constant::getNullValue(BO0->getType());
4094 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4095 }
4096 }
4097
4098 if (Value *V
26.1
'V' is null
26.1
'V' is null
= foldUnsignedMultiplicationOverflowCheck(I))
27
Taking false branch
4099 return replaceInstUsesWith(I, V);
4100
4101 if (Value *V
27.1
'V' is null
27.1
'V' is null
= foldICmpWithLowBitMaskedVal(I, Builder))
28
Taking false branch
4102 return replaceInstUsesWith(I, V);
4103
4104 if (Value *V
28.1
'V' is null
28.1
'V' is null
= foldICmpWithTruncSignExtendedVal(I, Builder))
29
Taking false branch
4105 return replaceInstUsesWith(I, V);
4106
4107 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
30
Calling 'foldShiftIntoShiftInAnotherHandOfAndInICmp'
4108 return replaceInstUsesWith(I, V);
4109
4110 return nullptr;
4111}
4112
4113/// Fold icmp Pred min|max(X, Y), X.
4114static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4115 ICmpInst::Predicate Pred = Cmp.getPredicate();
4116 Value *Op0 = Cmp.getOperand(0);
4117 Value *X = Cmp.getOperand(1);
4118
4119 // Canonicalize minimum or maximum operand to LHS of the icmp.
4120 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4121 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4122 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4123 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4124 std::swap(Op0, X);
4125 Pred = Cmp.getSwappedPredicate();
4126 }
4127
4128 Value *Y;
4129 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4130 // smin(X, Y) == X --> X s<= Y
4131 // smin(X, Y) s>= X --> X s<= Y
4132 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4133 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4134
4135 // smin(X, Y) != X --> X s> Y
4136 // smin(X, Y) s< X --> X s> Y
4137 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4138 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4139
4140 // These cases should be handled in InstSimplify:
4141 // smin(X, Y) s<= X --> true
4142 // smin(X, Y) s> X --> false
4143 return nullptr;
4144 }
4145
4146 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4147 // smax(X, Y) == X --> X s>= Y
4148 // smax(X, Y) s<= X --> X s>= Y
4149 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4150 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4151
4152 // smax(X, Y) != X --> X s< Y
4153 // smax(X, Y) s> X --> X s< Y
4154 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4155 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4156
4157 // These cases should be handled in InstSimplify:
4158 // smax(X, Y) s>= X --> true
4159 // smax(X, Y) s< X --> false
4160 return nullptr;
4161 }
4162
4163 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4164 // umin(X, Y) == X --> X u<= Y
4165 // umin(X, Y) u>= X --> X u<= Y
4166 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4167 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4168
4169 // umin(X, Y) != X --> X u> Y
4170 // umin(X, Y) u< X --> X u> Y
4171 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4172 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4173
4174 // These cases should be handled in InstSimplify:
4175 // umin(X, Y) u<= X --> true
4176 // umin(X, Y) u> X --> false
4177 return nullptr;
4178 }
4179
4180 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4181 // umax(X, Y) == X --> X u>= Y
4182 // umax(X, Y) u<= X --> X u>= Y
4183 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4184 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4185
4186 // umax(X, Y) != X --> X u< Y
4187 // umax(X, Y) u> X --> X u< Y
4188 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4189 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4190
4191 // These cases should be handled in InstSimplify:
4192 // umax(X, Y) u>= X --> true
4193 // umax(X, Y) u< X --> false
4194 return nullptr;
4195 }
4196
4197 return nullptr;
4198}
4199
4200Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
4201 if (!I.isEquality())
4202 return nullptr;
4203
4204 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4205 const CmpInst::Predicate Pred = I.getPredicate();
4206 Value *A, *B, *C, *D;
4207 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4208 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4209 Value *OtherVal = A == Op1 ? B : A;
4210 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4211 }
4212
4213 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4214 // A^c1 == C^c2 --> A == C^(c1^c2)
4215 ConstantInt *C1, *C2;
4216 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4217 Op1->hasOneUse()) {
4218 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4219 Value *Xor = Builder.CreateXor(C, NC);
4220 return new ICmpInst(Pred, A, Xor);
4221 }
4222
4223 // A^B == A^D -> B == D
4224 if (A == C)
4225 return new ICmpInst(Pred, B, D);
4226 if (A == D)
4227 return new ICmpInst(Pred, B, C);
4228 if (B == C)
4229 return new ICmpInst(Pred, A, D);
4230 if (B == D)
4231 return new ICmpInst(Pred, A, C);
4232 }
4233 }
4234
4235 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4236 // A == (A^B) -> B == 0
4237 Value *OtherVal = A == Op0 ? B : A;
4238 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4239 }
4240
4241 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4242 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4243 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4244 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4245
4246 if (A == C) {
4247 X = B;
4248 Y = D;
4249 Z = A;
4250 } else if (A == D) {
4251 X = B;
4252 Y = C;
4253 Z = A;
4254 } else if (B == C) {
4255 X = A;
4256 Y = D;
4257 Z = B;
4258 } else if (B == D) {
4259 X = A;
4260 Y = C;
4261 Z = B;
4262 }
4263
4264 if (X) { // Build (X^Y) & Z
4265 Op1 = Builder.CreateXor(X, Y);
4266 Op1 = Builder.CreateAnd(Op1, Z);
4267 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4268 }
4269 }
4270
4271 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4272 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4273 ConstantInt *Cst1;
4274 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4275 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4276 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4277 match(Op1, m_ZExt(m_Value(A))))) {
4278 APInt Pow2 = Cst1->getValue() + 1;
4279 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4280 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4281 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4282 }
4283
4284 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4285 // For lshr and ashr pairs.
4286 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4287 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4288 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4289 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4290 unsigned TypeBits = Cst1->getBitWidth();
4291 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4292 if (ShAmt < TypeBits && ShAmt != 0) {
4293 ICmpInst::Predicate NewPred =
4294 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4295 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4296 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4297 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4298 }
4299 }
4300
4301 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4302 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4303 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4304 unsigned TypeBits = Cst1->getBitWidth();
4305 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4306 if (ShAmt < TypeBits && ShAmt != 0) {
4307 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4308 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4309 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4310 I.getName() + ".mask");
4311 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4312 }
4313 }
4314
4315 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4316 // "icmp (and X, mask), cst"
4317 uint64_t ShAmt = 0;
4318 if (Op0->hasOneUse() &&
4319 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4320 match(Op1, m_ConstantInt(Cst1)) &&
4321 // Only do this when A has multiple uses. This is most important to do
4322 // when it exposes other optimizations.
4323 !A->hasOneUse()) {
4324 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4325
4326 if (ShAmt < ASize) {
4327 APInt MaskV =
4328 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4329 MaskV <<= ShAmt;
4330
4331 APInt CmpV = Cst1->getValue().zext(ASize);
4332 CmpV <<= ShAmt;
4333
4334 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4335 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4336 }
4337 }
4338
4339 // If both operands are byte-swapped or bit-reversed, just compare the
4340 // original values.
4341 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4342 // and handle more intrinsics.
4343 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4344 (match(Op0, m_BitReverse(m_Value(A))) &&
4345 match(Op1, m_BitReverse(m_Value(B)))))
4346 return new ICmpInst(Pred, A, B);
4347
4348 // Canonicalize checking for a power-of-2-or-zero value:
4349 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4350 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4351 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4352 m_Deferred(A)))) ||
4353 !match(Op1, m_ZeroInt()))
4354 A = nullptr;
4355
4356 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4357 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4358 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4359 A = Op1;
4360 else if (match(Op1,
4361 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4362 A = Op0;
4363
4364 if (A) {
4365 Type *Ty = A->getType();
4366 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4367 return Pred == ICmpInst::ICMP_EQ
4368 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4369 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4370 }
4371
4372 return nullptr;
4373}
4374
4375static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4376 InstCombiner::BuilderTy &Builder) {
4377 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0")((isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0"
) ? static_cast<void> (0) : __assert_fail ("isa<CastInst>(ICmp.getOperand(0)) && \"Expected cast for operand 0\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4377, __PRETTY_FUNCTION__))
;
4378 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4379 Value *X;
4380 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4381 return nullptr;
4382
4383 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4384 bool IsSignedCmp = ICmp.isSigned();
4385 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4386 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4387 // and the other is a zext), then we can't handle this.
4388 // TODO: This is too strict. We can handle some predicates (equality?).
4389 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4390 return nullptr;
4391
4392 // Not an extension from the same type?
4393 Value *Y = CastOp1->getOperand(0);
4394 Type *XTy = X->getType(), *YTy = Y->getType();
4395 if (XTy != YTy) {
4396 // One of the casts must have one use because we are creating a new cast.
4397 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4398 return nullptr;
4399 // Extend the narrower operand to the type of the wider operand.
4400 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4401 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4402 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4403 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4404 else
4405 return nullptr;
4406 }
4407
4408 // (zext X) == (zext Y) --> X == Y
4409 // (sext X) == (sext Y) --> X == Y
4410 if (ICmp.isEquality())
4411 return new ICmpInst(ICmp.getPredicate(), X, Y);
4412
4413 // A signed comparison of sign extended values simplifies into a
4414 // signed comparison.
4415 if (IsSignedCmp && IsSignedExt)
4416 return new ICmpInst(ICmp.getPredicate(), X, Y);
4417
4418 // The other three cases all fold into an unsigned comparison.
4419 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4420 }
4421
4422 // Below here, we are only folding a compare with constant.
4423 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4424 if (!C)
4425 return nullptr;
4426
4427 // Compute the constant that would happen if we truncated to SrcTy then
4428 // re-extended to DestTy.
4429 Type *SrcTy = CastOp0->getSrcTy();
4430 Type *DestTy = CastOp0->getDestTy();
4431 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4432 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4433
4434 // If the re-extended constant didn't change...
4435 if (Res2 == C) {
4436 if (ICmp.isEquality())
4437 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4438
4439 // A signed comparison of sign extended values simplifies into a
4440 // signed comparison.
4441 if (IsSignedExt && IsSignedCmp)
4442 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4443
4444 // The other three cases all fold into an unsigned comparison.
4445 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4446 }
4447
4448 // The re-extended constant changed, partly changed (in the case of a vector),
4449 // or could not be determined to be equal (in the case of a constant
4450 // expression), so the constant cannot be represented in the shorter type.
4451 // All the cases that fold to true or false will have already been handled
4452 // by SimplifyICmpInst, so only deal with the tricky case.
4453 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4454 return nullptr;
4455
4456 // Is source op positive?
4457 // icmp ult (sext X), C --> icmp sgt X, -1
4458 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4459 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4460
4461 // Is source op negative?
4462 // icmp ugt (sext X), C --> icmp slt X, 0
4463 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!")((ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"
) ? static_cast<void> (0) : __assert_fail ("ICmp.getPredicate() == ICmpInst::ICMP_UGT && \"ICmp should be folded!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4463, __PRETTY_FUNCTION__))
;
4464 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4465}
4466
4467/// Handle icmp (cast x), (cast or constant).
4468Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
4469 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4470 if (!CastOp0)
4471 return nullptr;
4472 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4473 return nullptr;
4474
4475 Value *Op0Src = CastOp0->getOperand(0);
4476 Type *SrcTy = CastOp0->getSrcTy();
4477 Type *DestTy = CastOp0->getDestTy();
4478
4479 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4480 // integer type is the same size as the pointer type.
4481 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4482 if (isa<VectorType>(SrcTy)) {
4483 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4484 DestTy = cast<VectorType>(DestTy)->getElementType();
4485 }
4486 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4487 };
4488 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4489 CompatibleSizes(SrcTy, DestTy)) {
4490 Value *NewOp1 = nullptr;
4491 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4492 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4493 if (PtrSrc->getType()->getPointerAddressSpace() ==
4494 Op0Src->getType()->getPointerAddressSpace()) {
4495 NewOp1 = PtrToIntOp1->getOperand(0);
4496 // If the pointer types don't match, insert a bitcast.
4497 if (Op0Src->getType() != NewOp1->getType())
4498 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4499 }
4500 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4501 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4502 }
4503
4504 if (NewOp1)
4505 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4506 }
4507
4508 return foldICmpWithZextOrSext(ICmp, Builder);
4509}
4510
4511static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4512 switch (BinaryOp) {
4513 default:
4514 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4514)
;
4515 case Instruction::Add:
4516 case Instruction::Sub:
4517 return match(RHS, m_Zero());
4518 case Instruction::Mul:
4519 return match(RHS, m_One());
4520 }
4521}
4522
4523OverflowResult
4524InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
4525 bool IsSigned, Value *LHS, Value *RHS,
4526 Instruction *CxtI) const {
4527 switch (BinaryOp) {
4528 default:
4529 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4529)
;
4530 case Instruction::Add:
4531 if (IsSigned)
4532 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4533 else
4534 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4535 case Instruction::Sub:
4536 if (IsSigned)
4537 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4538 else
4539 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4540 case Instruction::Mul:
4541 if (IsSigned)
4542 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4543 else
4544 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4545 }
4546}
4547
4548bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
4549 bool IsSigned, Value *LHS,
4550 Value *RHS, Instruction &OrigI,
4551 Value *&Result,
4552 Constant *&Overflow) {
4553 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4554 std::swap(LHS, RHS);
4555
4556 // If the overflow check was an add followed by a compare, the insertion point
4557 // may be pointing to the compare. We want to insert the new instructions
4558 // before the add in case there are uses of the add between the add and the
4559 // compare.
4560 Builder.SetInsertPoint(&OrigI);
4561
4562 if (isNeutralValue(BinaryOp, RHS)) {
4563 Result = LHS;
4564 Overflow = Builder.getFalse();
4565 return true;
4566 }
4567
4568 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4569 case OverflowResult::MayOverflow:
4570 return false;
4571 case OverflowResult::AlwaysOverflowsLow:
4572 case OverflowResult::AlwaysOverflowsHigh:
4573 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4574 Result->takeName(&OrigI);
4575 Overflow = Builder.getTrue();
4576 return true;
4577 case OverflowResult::NeverOverflows:
4578 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4579 Result->takeName(&OrigI);
4580 Overflow = Builder.getFalse();
4581 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4582 if (IsSigned)
4583 Inst->setHasNoSignedWrap();
4584 else
4585 Inst->setHasNoUnsignedWrap();
4586 }
4587 return true;
4588 }
4589
4590 llvm_unreachable("Unexpected overflow result")::llvm::llvm_unreachable_internal("Unexpected overflow result"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4590)
;
4591}
4592
4593/// Recognize and process idiom involving test for multiplication
4594/// overflow.
4595///
4596/// The caller has matched a pattern of the form:
4597/// I = cmp u (mul(zext A, zext B), V
4598/// The function checks if this is a test for overflow and if so replaces
4599/// multiplication with call to 'mul.with.overflow' intrinsic.
4600///
4601/// \param I Compare instruction.
4602/// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4603/// the compare instruction. Must be of integer type.
4604/// \param OtherVal The other argument of compare instruction.
4605/// \returns Instruction which must replace the compare instruction, NULL if no
4606/// replacement required.
4607static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4608 Value *OtherVal,
4609 InstCombinerImpl &IC) {
4610 // Don't bother doing this transformation for pointers, don't do it for
4611 // vectors.
4612 if (!isa<IntegerType>(MulVal->getType()))
4613 return nullptr;
4614
4615 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal)((I.getOperand(0) == MulVal || I.getOperand(1) == MulVal) ? static_cast
<void> (0) : __assert_fail ("I.getOperand(0) == MulVal || I.getOperand(1) == MulVal"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4615, __PRETTY_FUNCTION__))
;
4616 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal)((I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal)
? static_cast<void> (0) : __assert_fail ("I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4616, __PRETTY_FUNCTION__))
;
4617 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4618 if (!MulInstr)
4619 return nullptr;
4620 assert(MulInstr->getOpcode() == Instruction::Mul)((MulInstr->getOpcode() == Instruction::Mul) ? static_cast
<void> (0) : __assert_fail ("MulInstr->getOpcode() == Instruction::Mul"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4620, __PRETTY_FUNCTION__))
;
4621
4622 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4623 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4624 assert(LHS->getOpcode() == Instruction::ZExt)((LHS->getOpcode() == Instruction::ZExt) ? static_cast<
void> (0) : __assert_fail ("LHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4624, __PRETTY_FUNCTION__))
;
4625 assert(RHS->getOpcode() == Instruction::ZExt)((RHS->getOpcode() == Instruction::ZExt) ? static_cast<
void> (0) : __assert_fail ("RHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4625, __PRETTY_FUNCTION__))
;
4626 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4627
4628 // Calculate type and width of the result produced by mul.with.overflow.
4629 Type *TyA = A->getType(), *TyB = B->getType();
4630 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4631 WidthB = TyB->getPrimitiveSizeInBits();
4632 unsigned MulWidth;
4633 Type *MulType;
4634 if (WidthB > WidthA) {
4635 MulWidth = WidthB;
4636 MulType = TyB;
4637 } else {
4638 MulWidth = WidthA;
4639 MulType = TyA;
4640 }
4641
4642 // In order to replace the original mul with a narrower mul.with.overflow,
4643 // all uses must ignore upper bits of the product. The number of used low
4644 // bits must be not greater than the width of mul.with.overflow.
4645 if (MulVal->hasNUsesOrMore(2))
4646 for (User *U : MulVal->users()) {
4647 if (U == &I)
4648 continue;
4649 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4650 // Check if truncation ignores bits above MulWidth.
4651 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4652 if (TruncWidth > MulWidth)
4653 return nullptr;
4654 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4655 // Check if AND ignores bits above MulWidth.
4656 if (BO->getOpcode() != Instruction::And)
4657 return nullptr;
4658 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4659 const APInt &CVal = CI->getValue();
4660 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4661 return nullptr;
4662 } else {
4663 // In this case we could have the operand of the binary operation
4664 // being defined in another block, and performing the replacement
4665 // could break the dominance relation.
4666 return nullptr;
4667 }
4668 } else {
4669 // Other uses prohibit this transformation.
4670 return nullptr;
4671 }
4672 }
4673
4674 // Recognize patterns
4675 switch (I.getPredicate()) {
4676 case ICmpInst::ICMP_EQ:
4677 case ICmpInst::ICMP_NE:
4678 // Recognize pattern:
4679 // mulval = mul(zext A, zext B)
4680 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4681 ConstantInt *CI;
4682 Value *ValToMask;
4683 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4684 if (ValToMask != MulVal)
4685 return nullptr;
4686 const APInt &CVal = CI->getValue() + 1;
4687 if (CVal.isPowerOf2()) {
4688 unsigned MaskWidth = CVal.logBase2();
4689 if (MaskWidth == MulWidth)
4690 break; // Recognized
4691 }
4692 }
4693 return nullptr;
4694
4695 case ICmpInst::ICMP_UGT:
4696 // Recognize pattern:
4697 // mulval = mul(zext A, zext B)
4698 // cmp ugt mulval, max
4699 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4700 APInt MaxVal = APInt::getMaxValue(MulWidth);
4701 MaxVal = MaxVal.zext(CI->getBitWidth());
4702 if (MaxVal.eq(CI->getValue()))
4703 break; // Recognized
4704 }
4705 return nullptr;
4706
4707 case ICmpInst::ICMP_UGE:
4708 // Recognize pattern:
4709 // mulval = mul(zext A, zext B)
4710 // cmp uge mulval, max+1
4711 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4712 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4713 if (MaxVal.eq(CI->getValue()))
4714 break; // Recognized
4715 }
4716 return nullptr;
4717
4718 case ICmpInst::ICMP_ULE:
4719 // Recognize pattern:
4720 // mulval = mul(zext A, zext B)
4721 // cmp ule mulval, max
4722 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4723 APInt MaxVal = APInt::getMaxValue(MulWidth);
4724 MaxVal = MaxVal.zext(CI->getBitWidth());
4725 if (MaxVal.eq(CI->getValue()))
4726 break; // Recognized
4727 }
4728 return nullptr;
4729
4730 case ICmpInst::ICMP_ULT:
4731 // Recognize pattern:
4732 // mulval = mul(zext A, zext B)
4733 // cmp ule mulval, max + 1
4734 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4735 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4736 if (MaxVal.eq(CI->getValue()))
4737 break; // Recognized
4738 }
4739 return nullptr;
4740
4741 default:
4742 return nullptr;
4743 }
4744
4745 InstCombiner::BuilderTy &Builder = IC.Builder;
4746 Builder.SetInsertPoint(MulInstr);
4747
4748 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4749 Value *MulA = A, *MulB = B;
4750 if (WidthA < MulWidth)
4751 MulA = Builder.CreateZExt(A, MulType);
4752 if (WidthB < MulWidth)
4753 MulB = Builder.CreateZExt(B, MulType);
4754 Function *F = Intrinsic::getDeclaration(
4755 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4756 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4757 IC.addToWorklist(MulInstr);
4758
4759 // If there are uses of mul result other than the comparison, we know that
4760 // they are truncation or binary AND. Change them to use result of
4761 // mul.with.overflow and adjust properly mask/size.
4762 if (MulVal->hasNUsesOrMore(2)) {
4763 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4764 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4765 User *U = *UI++;
4766 if (U == &I || U == OtherVal)
4767 continue;
4768 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4769 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4770 IC.replaceInstUsesWith(*TI, Mul);
4771 else
4772 TI->setOperand(0, Mul);
4773 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4774 assert(BO->getOpcode() == Instruction::And)((BO->getOpcode() == Instruction::And) ? static_cast<void
> (0) : __assert_fail ("BO->getOpcode() == Instruction::And"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4774, __PRETTY_FUNCTION__))
;
4775 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4776 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4777 APInt ShortMask = CI->getValue().trunc(MulWidth);
4778 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4779 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
4780 IC.replaceInstUsesWith(*BO, Zext);
4781 } else {
4782 llvm_unreachable("Unexpected Binary operation")::llvm::llvm_unreachable_internal("Unexpected Binary operation"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4782)
;
4783 }
4784 IC.addToWorklist(cast<Instruction>(U));
4785 }
4786 }
4787 if (isa<Instruction>(OtherVal))
4788 IC.addToWorklist(cast<Instruction>(OtherVal));
4789
4790 // The original icmp gets replaced with the overflow value, maybe inverted
4791 // depending on predicate.
4792 bool Inverse = false;
4793 switch (I.getPredicate()) {
4794 case ICmpInst::ICMP_NE:
4795 break;
4796 case ICmpInst::ICMP_EQ:
4797 Inverse = true;
4798 break;
4799 case ICmpInst::ICMP_UGT:
4800 case ICmpInst::ICMP_UGE:
4801 if (I.getOperand(0) == MulVal)
4802 break;
4803 Inverse = true;
4804 break;
4805 case ICmpInst::ICMP_ULT:
4806 case ICmpInst::ICMP_ULE:
4807 if (I.getOperand(1) == MulVal)
4808 break;
4809 Inverse = true;
4810 break;
4811 default:
4812 llvm_unreachable("Unexpected predicate")::llvm::llvm_unreachable_internal("Unexpected predicate", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4812)
;
4813 }
4814 if (Inverse) {
4815 Value *Res = Builder.CreateExtractValue(Call, 1);
4816 return BinaryOperator::CreateNot(Res);
4817 }
4818
4819 return ExtractValueInst::Create(Call, 1);
4820}
4821
4822/// When performing a comparison against a constant, it is possible that not all
4823/// the bits in the LHS are demanded. This helper method computes the mask that
4824/// IS demanded.
4825static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4826 const APInt *RHS;
4827 if (!match(I.getOperand(1), m_APInt(RHS)))
4828 return APInt::getAllOnesValue(BitWidth);
4829
4830 // If this is a normal comparison, it demands all bits. If it is a sign bit
4831 // comparison, it only demands the sign bit.
4832 bool UnusedBit;
4833 if (InstCombiner::isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4834 return APInt::getSignMask(BitWidth);
4835
4836 switch (I.getPredicate()) {
4837 // For a UGT comparison, we don't care about any bits that
4838 // correspond to the trailing ones of the comparand. The value of these
4839 // bits doesn't impact the outcome of the comparison, because any value
4840 // greater than the RHS must differ in a bit higher than these due to carry.
4841 case ICmpInst::ICMP_UGT:
4842 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4843
4844 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4845 // Any value less than the RHS must differ in a higher bit because of carries.
4846 case ICmpInst::ICMP_ULT:
4847 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4848
4849 default:
4850 return APInt::getAllOnesValue(BitWidth);
4851 }
4852}
4853
4854/// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4855/// should be swapped.
4856/// The decision is based on how many times these two operands are reused
4857/// as subtract operands and their positions in those instructions.
4858/// The rationale is that several architectures use the same instruction for
4859/// both subtract and cmp. Thus, it is better if the order of those operands
4860/// match.
4861/// \return true if Op0 and Op1 should be swapped.
4862static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4863 // Filter out pointer values as those cannot appear directly in subtract.
4864 // FIXME: we may want to go through inttoptrs or bitcasts.
4865 if (Op0->getType()->isPointerTy())
4866 return false;
4867 // If a subtract already has the same operands as a compare, swapping would be
4868 // bad. If a subtract has the same operands as a compare but in reverse order,
4869 // then swapping is good.
4870 int GoodToSwap = 0;
4871 for (const User *U : Op0->users()) {
4872 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4873 GoodToSwap++;
4874 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4875 GoodToSwap--;
4876 }
4877 return GoodToSwap > 0;
4878}
4879
4880/// Check that one use is in the same block as the definition and all
4881/// other uses are in blocks dominated by a given block.
4882///
4883/// \param DI Definition
4884/// \param UI Use
4885/// \param DB Block that must dominate all uses of \p DI outside
4886/// the parent block
4887/// \return true when \p UI is the only use of \p DI in the parent block
4888/// and all other uses of \p DI are in blocks dominated by \p DB.
4889///
4890bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
4891 const Instruction *UI,
4892 const BasicBlock *DB) const {
4893 assert(DI && UI && "Instruction not defined\n")((DI && UI && "Instruction not defined\n") ? static_cast
<void> (0) : __assert_fail ("DI && UI && \"Instruction not defined\\n\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4893, __PRETTY_FUNCTION__))
;
4894 // Ignore incomplete definitions.
4895 if (!DI->getParent())
4896 return false;
4897 // DI and UI must be in the same block.
4898 if (DI->getParent() != UI->getParent())
4899 return false;
4900 // Protect from self-referencing blocks.
4901 if (DI->getParent() == DB)
4902 return false;
4903 for (const User *U : DI->users()) {
4904 auto *Usr = cast<Instruction>(U);
4905 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4906 return false;
4907 }
4908 return true;
4909}
4910
4911/// Return true when the instruction sequence within a block is select-cmp-br.
4912static bool isChainSelectCmpBranch(const SelectInst *SI) {
4913 const BasicBlock *BB = SI->getParent();
4914 if (!BB)
4915 return false;
4916 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4917 if (!BI || BI->getNumSuccessors() != 2)
4918 return false;
4919 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4920 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4921 return false;
4922 return true;
4923}
4924
4925/// True when a select result is replaced by one of its operands
4926/// in select-icmp sequence. This will eventually result in the elimination
4927/// of the select.
4928///
4929/// \param SI Select instruction
4930/// \param Icmp Compare instruction
4931/// \param SIOpd Operand that replaces the select
4932///
4933/// Notes:
4934/// - The replacement is global and requires dominator information
4935/// - The caller is responsible for the actual replacement
4936///
4937/// Example:
4938///
4939/// entry:
4940/// %4 = select i1 %3, %C* %0, %C* null
4941/// %5 = icmp eq %C* %4, null
4942/// br i1 %5, label %9, label %7
4943/// ...
4944/// ; <label>:7 ; preds = %entry
4945/// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4946/// ...
4947///
4948/// can be transformed to
4949///
4950/// %5 = icmp eq %C* %0, null
4951/// %6 = select i1 %3, i1 %5, i1 true
4952/// br i1 %6, label %9, label %7
4953/// ...
4954/// ; <label>:7 ; preds = %entry
4955/// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4956///
4957/// Similar when the first operand of the select is a constant or/and
4958/// the compare is for not equal rather than equal.
4959///
4960/// NOTE: The function is only called when the select and compare constants
4961/// are equal, the optimization can work only for EQ predicates. This is not a
4962/// major restriction since a NE compare should be 'normalized' to an equal
4963/// compare, which usually happens in the combiner and test case
4964/// select-cmp-br.ll checks for it.
4965bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
4966 const ICmpInst *Icmp,
4967 const unsigned SIOpd) {
4968 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!")(((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!"
) ? static_cast<void> (0) : __assert_fail ("(SIOpd == 1 || SIOpd == 2) && \"Invalid select operand!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4968, __PRETTY_FUNCTION__))
;
4969 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4970 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4971 // The check for the single predecessor is not the best that can be
4972 // done. But it protects efficiently against cases like when SI's
4973 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4974 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4975 // replaced can be reached on either path. So the uniqueness check
4976 // guarantees that the path all uses of SI (outside SI's parent) are on
4977 // is disjoint from all other paths out of SI. But that information
4978 // is more expensive to compute, and the trade-off here is in favor
4979 // of compile-time. It should also be noticed that we check for a single
4980 // predecessor and not only uniqueness. This to handle the situation when
4981 // Succ and Succ1 points to the same basic block.
4982 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4983 NumSel++;
4984 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4985 return true;
4986 }
4987 }
4988 return false;
4989}
4990
4991/// Try to fold the comparison based on range information we can get by checking
4992/// whether bits are known to be zero or one in the inputs.
4993Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
4994 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4995 Type *Ty = Op0->getType();
4996 ICmpInst::Predicate Pred = I.getPredicate();
4997
4998 // Get scalar or pointer size.
4999 unsigned BitWidth = Ty->isIntOrIntVectorTy()
5000 ? Ty->getScalarSizeInBits()
5001 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
5002
5003 if (!BitWidth)
5004 return nullptr;
5005
5006 KnownBits Op0Known(BitWidth);
5007 KnownBits Op1Known(BitWidth);
5008
5009 if (SimplifyDemandedBits(&I, 0,
5010 getDemandedBitsLHSMask(I, BitWidth),
5011 Op0Known, 0))
5012 return &I;
5013
5014 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
5015 Op1Known, 0))
5016 return &I;
5017
5018 // Given the known and unknown bits, compute a range that the LHS could be
5019 // in. Compute the Min, Max and RHS values based on the known bits. For the
5020 // EQ and NE we use unsigned values.
5021 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
5022 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
5023 if (I.isSigned()) {
5024 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
5025 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
5026 } else {
5027 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
5028 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
5029 }
5030
5031 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
5032 // out that the LHS or RHS is a constant. Constant fold this now, so that
5033 // code below can assume that Min != Max.
5034 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
5035 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
5036 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
5037 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
5038
5039 // Based on the range information we know about the LHS, see if we can
5040 // simplify this comparison. For example, (x&4) < 8 is always true.
5041 switch (Pred) {
5042 default:
5043 llvm_unreachable("Unknown icmp opcode!")::llvm::llvm_unreachable_internal("Unknown icmp opcode!", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5043)
;
5044 case ICmpInst::ICMP_EQ:
5045 case ICmpInst::ICMP_NE: {
5046 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
5047 return Pred == CmpInst::ICMP_EQ
5048 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
5049 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5050 }
5051
5052 // If all bits are known zero except for one, then we know at most one bit
5053 // is set. If the comparison is against zero, then this is a check to see if
5054 // *that* bit is set.
5055 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
5056 if (Op1Known.isZero()) {
5057 // If the LHS is an AND with the same constant, look through it.
5058 Value *LHS = nullptr;
5059 const APInt *LHSC;
5060 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
5061 *LHSC != Op0KnownZeroInverted)
5062 LHS = Op0;
5063
5064 Value *X;
5065 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
5066 APInt ValToCheck = Op0KnownZeroInverted;
5067 Type *XTy = X->getType();
5068 if (ValToCheck.isPowerOf2()) {
5069 // ((1 << X) & 8) == 0 -> X != 3
5070 // ((1 << X) & 8) != 0 -> X == 3
5071 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5072 auto NewPred = ICmpInst::getInversePredicate(Pred);
5073 return new ICmpInst(NewPred, X, CmpC);
5074 } else if ((++ValToCheck).isPowerOf2()) {
5075 // ((1 << X) & 7) == 0 -> X >= 3
5076 // ((1 << X) & 7) != 0 -> X < 3
5077 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5078 auto NewPred =
5079 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
5080 return new ICmpInst(NewPred, X, CmpC);
5081 }
5082 }
5083
5084 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5085 const APInt *CI;
5086 if (Op0KnownZeroInverted.isOneValue() &&
5087 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5088 // ((8 >>u X) & 1) == 0 -> X != 3
5089 // ((8 >>u X) & 1) != 0 -> X == 3
5090 unsigned CmpVal = CI->countTrailingZeros();
5091 auto NewPred = ICmpInst::getInversePredicate(Pred);
5092 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5093 }
5094 }
5095 break;
5096 }
5097 case ICmpInst::ICMP_ULT: {
5098 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5099 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5100 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5101 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5102 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5103 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5104
5105 const APInt *CmpC;
5106 if (match(Op1, m_APInt(CmpC))) {
5107 // A <u C -> A == C-1 if min(A)+1 == C
5108 if (*CmpC == Op0Min + 1)
5109 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5110 ConstantInt::get(Op1->getType(), *CmpC - 1));
5111 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5112 // exceeds the log2 of C.
5113 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5114 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5115 Constant::getNullValue(Op1->getType()));
5116 }
5117 break;
5118 }
5119 case ICmpInst::ICMP_UGT: {
5120 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5121 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5122 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5123 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5124 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5125 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5126
5127 const APInt *CmpC;
5128 if (match(Op1, m_APInt(CmpC))) {
5129 // A >u C -> A == C+1 if max(a)-1 == C
5130 if (*CmpC == Op0Max - 1)
5131 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5132 ConstantInt::get(Op1->getType(), *CmpC + 1));
5133 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5134 // exceeds the log2 of C.
5135 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5136 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5137 Constant::getNullValue(Op1->getType()));
5138 }
5139 break;
5140 }
5141 case ICmpInst::ICMP_SLT: {
5142 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5143 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5144 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5145 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5146 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5147 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5148 const APInt *CmpC;
5149 if (match(Op1, m_APInt(CmpC))) {
5150 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5151 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5152 ConstantInt::get(Op1->getType(), *CmpC - 1));
5153 }
5154 break;
5155 }
5156 case ICmpInst::ICMP_SGT: {
5157 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5158 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5159 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5160 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5161 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5162 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5163 const APInt *CmpC;
5164 if (match(Op1, m_APInt(CmpC))) {
5165 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5166 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5167 ConstantInt::get(Op1->getType(), *CmpC + 1));
5168 }
5169 break;
5170 }
5171 case ICmpInst::ICMP_SGE:
5172 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_SGE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5172, __PRETTY_FUNCTION__))
;
5173 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5174 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5175 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5176 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5177 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5178 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5179 break;
5180 case ICmpInst::ICMP_SLE:
5181 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_SLE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5181, __PRETTY_FUNCTION__))
;
5182 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5183 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5184 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5185 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5186 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5187 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5188 break;
5189 case ICmpInst::ICMP_UGE:
5190 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_UGE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5190, __PRETTY_FUNCTION__))
;
5191 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5192 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5193 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5194 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5195 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5196 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5197 break;
5198 case ICmpInst::ICMP_ULE:
5199 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_ULE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5199, __PRETTY_FUNCTION__))
;
5200 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5201 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5202 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5203 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5204 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5205 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5206 break;
5207 }
5208
5209 // Turn a signed comparison into an unsigned one if both operands are known to
5210 // have the same sign.
5211 if (I.isSigned() &&
5212 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5213 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5214 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5215
5216 return nullptr;
5217}
5218
5219llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
5220InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5221 Constant *C) {
5222 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&((ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate
(Pred) && "Only for relational integer predicates.") ?
static_cast<void> (0) : __assert_fail ("ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && \"Only for relational integer predicates.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5223, __PRETTY_FUNCTION__))
5223 "Only for relational integer predicates.")((ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate
(Pred) && "Only for relational integer predicates.") ?
static_cast<void> (0) : __assert_fail ("ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && \"Only for relational integer predicates.\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5223, __PRETTY_FUNCTION__))
;
5224
5225 Type *Type = C->getType();
5226 bool IsSigned = ICmpInst::isSigned(Pred);
5227
5228 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5229 bool WillIncrement =
5230 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5231
5232 // Check if the constant operand can be safely incremented/decremented
5233 // without overflowing/underflowing.
5234 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5235 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5236 };
5237
5238 Constant *SafeReplacementConstant = nullptr;
5239 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5240 // Bail out if the constant can't be safely incremented/decremented.
5241 if (!ConstantIsOk(CI))
5242 return llvm::None;
5243 } else if (auto *VTy = dyn_cast<VectorType>(Type)) {
5244 unsigned NumElts = VTy->getNumElements();
5245 for (unsigned i = 0; i != NumElts; ++i) {
5246 Constant *Elt = C->getAggregateElement(i);
5247 if (!Elt)
5248 return llvm::None;
5249
5250 if (isa<UndefValue>(Elt))
5251 continue;
5252
5253 // Bail out if we can't determine if this constant is min/max or if we
5254 // know that this constant is min/max.
5255 auto *CI = dyn_cast<ConstantInt>(Elt);
5256 if (!CI || !ConstantIsOk(CI))
5257 return llvm::None;
5258
5259 if (!SafeReplacementConstant)
5260 SafeReplacementConstant = CI;
5261 }
5262 } else {
5263 // ConstantExpr?
5264 return llvm::None;
5265 }
5266
5267 // It may not be safe to change a compare predicate in the presence of
5268 // undefined elements, so replace those elements with the first safe constant
5269 // that we found.
5270 if (C->containsUndefElement()) {
5271 assert(SafeReplacementConstant && "Replacement constant not set")((SafeReplacementConstant && "Replacement constant not set"
) ? static_cast<void> (0) : __assert_fail ("SafeReplacementConstant && \"Replacement constant not set\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5271, __PRETTY_FUNCTION__))
;
5272 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5273 }
5274
5275 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5276
5277 // Increment or decrement the constant.
5278 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5279 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5280
5281 return std::make_pair(NewPred, NewC);
5282}
5283
5284/// If we have an icmp le or icmp ge instruction with a constant operand, turn
5285/// it into the appropriate icmp lt or icmp gt instruction. This transform
5286/// allows them to be folded in visitICmpInst.
5287static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5288 ICmpInst::Predicate Pred = I.getPredicate();
5289 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5290 InstCombiner::isCanonicalPredicate(Pred))
5291 return nullptr;
5292
5293 Value *Op0 = I.getOperand(0);
5294 Value *Op1 = I.getOperand(1);
5295 auto *Op1C = dyn_cast<Constant>(Op1);
5296 if (!Op1C)
5297 return nullptr;
5298
5299 auto FlippedStrictness =
5300 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5301 if (!FlippedStrictness)
5302 return nullptr;
5303
5304 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5305}
5306
5307/// If we have a comparison with a non-canonical predicate, if we can update
5308/// all the users, invert the predicate and adjust all the users.
5309static CmpInst *canonicalizeICmpPredicate(CmpInst &I) {
5310 // Is the predicate already canonical?
5311 CmpInst::Predicate Pred = I.getPredicate();
5312 if (InstCombiner::isCanonicalPredicate(Pred))
5313 return nullptr;
5314
5315 // Can all users be adjusted to predicate inversion?
5316 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
5317 return nullptr;
5318
5319 // Ok, we can canonicalize comparison!
5320 // Let's first invert the comparison's predicate.
5321 I.setPredicate(CmpInst::getInversePredicate(Pred));
5322 I.setName(I.getName() + ".not");
5323
5324 // And now let's adjust every user.
5325 for (User *U : I.users()) {
5326 switch (cast<Instruction>(U)->getOpcode()) {
5327 case Instruction::Select: {
5328 auto *SI = cast<SelectInst>(U);
5329 SI->swapValues();
5330 SI->swapProfMetadata();
5331 break;
5332 }
5333 case Instruction::Br:
5334 cast<BranchInst>(U)->swapSuccessors(); // swaps prof metadata too
5335 break;
5336 case Instruction::Xor:
5337 U->replaceAllUsesWith(&I);
5338 break;
5339 default:
5340 llvm_unreachable("Got unexpected user - out of sync with "::llvm::llvm_unreachable_internal("Got unexpected user - out of sync with "
"canFreelyInvertAllUsersOf() ?", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5341)
5341 "canFreelyInvertAllUsersOf() ?")::llvm::llvm_unreachable_internal("Got unexpected user - out of sync with "
"canFreelyInvertAllUsersOf() ?", "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5341)
;
5342 }
5343 }
5344
5345 return &I;
5346}
5347
5348/// Integer compare with boolean values can always be turned into bitwise ops.
5349static Instruction *canonicalizeICmpBool(ICmpInst &I,
5350 InstCombiner::BuilderTy &Builder) {
5351 Value *A = I.getOperand(0), *B = I.getOperand(1);
5352 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only")((A->getType()->isIntOrIntVectorTy(1) && "Bools only"
) ? static_cast<void> (0) : __assert_fail ("A->getType()->isIntOrIntVectorTy(1) && \"Bools only\""
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5352, __PRETTY_FUNCTION__))
;
5353
5354 // A boolean compared to true/false can be simplified to Op0/true/false in
5355 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5356 // Cases not handled by InstSimplify are always 'not' of Op0.
5357 if (match(B, m_Zero())) {
5358 switch (I.getPredicate()) {
5359 case CmpInst::ICMP_EQ: // A == 0 -> !A
5360 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5361 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5362 return BinaryOperator::CreateNot(A);
5363 default:
5364 llvm_unreachable("ICmp i1 X, C not simplified as expected.")::llvm::llvm_unreachable_internal("ICmp i1 X, C not simplified as expected."
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5364)
;
5365 }
5366 } else if (match(B, m_One())) {
5367 switch (I.getPredicate()) {
5368 case CmpInst::ICMP_NE: // A != 1 -> !A
5369 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5370 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5371 return BinaryOperator::CreateNot(A);
5372 default:
5373 llvm_unreachable("ICmp i1 X, C not simplified as expected.")::llvm::llvm_unreachable_internal("ICmp i1 X, C not simplified as expected."
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5373)
;
5374 }
5375 }
5376
5377 switch (I.getPredicate()) {
5378 default:
5379 llvm_unreachable("Invalid icmp instruction!")::llvm::llvm_unreachable_internal("Invalid icmp instruction!"
, "/build/llvm-toolchain-snapshot-12~++20200806111125+5446ec85070/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5379)
;
5380 case ICmpInst::ICMP_EQ:
5381 // icmp eq i1 A, B -> ~(A ^ B)
5382 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5383
5384 case ICmpInst::ICMP_NE:
5385 // icmp ne i1 A, B -> A ^ B
5386 return BinaryOperator::CreateXor(A, B);
5387
5388 case ICmpInst::ICMP_UGT:
5389 // icmp ugt -> icmp ult
5390 std::swap(A, B);
5391 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5392 case ICmpInst::ICMP_ULT:
5393 // icmp ult i1 A, B -> ~A & B
5394 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5395
5396 case ICmpInst::ICMP_SGT:
5397 // icmp sgt -> icmp slt
5398 std::swap(A, B);
5399 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5400 case ICmpInst::ICMP_SLT:
5401 // icmp slt i1 A, B -> A & ~B
5402 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5403
5404 case ICmpInst::ICMP_UGE:
5405 // icmp uge -> icmp ule
5406 std::swap(A, B);
5407 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5408 case ICmpInst::ICMP_ULE:
5409 // icmp ule i1 A, B -> ~A | B
5410 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5411
5412 case ICmpInst::ICMP_SGE:
5413 // icmp sge -> icmp sle
5414 std::swap(A, B);
5415 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5416 case ICmpInst::ICMP_SLE:
5417 // icmp sle i1 A, B -> A | ~B
5418 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5419 }
5420}
5421
5422// Transform pattern like:
5423// (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5424// (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5425// Into:
5426// (X l>> Y) != 0
5427// (X l>> Y) == 0
5428static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5429 InstCombiner::BuilderTy &Builder) {
5430 ICmpInst::Predicate Pred, NewPred;
5431 Value *X, *Y;
5432 if (match(&Cmp,
5433 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5434 switch (Pred) {
5435 case ICmpInst::ICMP_ULE:
5436 NewPred = ICmpInst::ICMP_NE;
5437 break;
5438 case ICmpInst::ICMP_UGT:
5439 NewPred = ICmpInst::ICMP_EQ;
5440 break;
5441 default:
5442 return nullptr;
5443 }
5444 } else if (match(&Cmp, m_c_ICmp(Pred,
5445 m_OneUse(m_CombineOr(
5446 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5447 m_Add(m_Shl(m_One(), m_Value(Y)),
5448 m_AllOnes()))),
5449 m_Value(X)))) {
5450 // The variant with 'add' is not canonical, (the variant with 'not' is)
5451 // we only get it because it has extra uses, and can't be canonicalized,
5452
5453 switch (Pred) {
5454 case ICmpInst::ICMP_ULT:
5455 NewPred = ICmpInst::ICMP_NE;
5456 break;
5457 case ICmpInst::ICMP_UGE:
5458 NewPred = ICmpInst::ICMP_EQ;
5459 break;
5460 default:
5461 return nullptr;
5462 }
5463 } else
5464 return nullptr;
5465
5466 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5467 Constant *Zero = Constant::getNullValue(NewX->getType());
5468 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5469}
5470
5471static Instruction *foldVectorCmp(CmpInst &Cmp,
5472 InstCombiner::BuilderTy &Builder) {
5473 const CmpInst::Predicate Pred = Cmp.getPredicate();
5474 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5475 Value *V1, *V2;
5476 ArrayRef<int> M;
5477 if (!match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(M))))
5478 return nullptr;
5479
5480 // If both arguments of the cmp are shuffles that use the same mask and
5481 // shuffle within a single vector, move the shuffle after the cmp:
5482 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5483 Type *V1Ty = V1->getType();
5484 if (match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(M))) &&
5485 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
5486 Value *NewCmp = Builder.CreateCmp(Pred, V1, V2);
5487 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5488 }
5489
5490 // Try to canonicalize compare with splatted operand and splat constant.
5491 // TODO: We could generalize this for more than splats. See/use the code in
5492 // InstCombiner::foldVectorBinop().
5493 Constant *C;
5494 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
5495 return nullptr;
5496
5497 // Length-changing splats are ok, so adjust the constants as needed:
5498 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
5499 Constant *ScalarC = C->getSplatValue(/* AllowUndefs */ true);
5500 int MaskSplatIndex;
5501 if (ScalarC && match(M, m_SplatOrUndefMask(MaskSplatIndex))) {
5502 // We allow undefs in matching, but this transform removes those for safety.
5503 // Demanded elements analysis should be able to recover some/all of that.
5504 C = ConstantVector::getSplat(cast<VectorType>(V1Ty)->getElementCount(),
5505 ScalarC);
5506 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
5507 Value *NewCmp = Builder.CreateCmp(Pred, V1, C);
5508 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()),
5509 NewM);
5510 }
5511
5512 return nullptr;
5513}
5514
5515// extract(uadd.with.overflow(A, B), 0) ult A
5516// -> extract(uadd.with.overflow(A, B), 1)
5517static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5518 CmpInst::Predicate Pred = I.getPredicate();
5519 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5520
5521 Value *UAddOv;
5522 Value *A, *B;
5523 auto UAddOvResultPat = m_ExtractValue<0>(
5524 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5525 if (match(Op0, UAddOvResultPat) &&
5526 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5527 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5528 (match(A, m_One()) || match(B, m_One()))) ||
5529 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5530 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5531 // extract(uadd.with.overflow(A, B), 0) < A
5532 // extract(uadd.with.overflow(A, 1), 0) == 0
5533 // extract(uadd.with.overflow(A, -1), 0) != -1
5534 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5535 else if (match(Op1, UAddOvResultPat) &&
5536 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5537 // A > extract(uadd.with.overflow(A, B), 0)
5538 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5539 else
5540 return nullptr;
5541
5542 return ExtractValueInst::Create(UAddOv, 1);
5543}
5544
5545Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
5546 bool Changed = false;
5547 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5548 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5549 unsigned Op0Cplxity = getComplexity(Op0);
5550 unsigned Op1Cplxity = getComplexity(Op1);
5551
5552 /// Orders the operands of the compare so that they are listed from most
5553 /// complex to least complex. This puts constants before unary operators,
5554 /// before binary operators.
5555 if (Op0Cplxity < Op1Cplxity ||
5556 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5557 I.swapOperands();
5558 std::swap(Op0, Op1);
5559 Changed = true;
5560 }
5561
5562 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, Q))
5563 return replaceInstUsesWith(I, V);
5564
5565 // Comparing -val or val with non-zero is the same as just comparing val
5566 // ie, abs(val) != 0 -> val != 0
5567 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5568 Value *Cond, *SelectTrue, *SelectFalse;
5569 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5570 m_Value(SelectFalse)))) {
5571 if (Value *V = dyn_castNegVal(SelectTrue)) {
5572 if (V == SelectFalse)
5573 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5574 }
5575 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5576 if (V == SelectTrue)
5577 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5578 }
5579 }
5580 }
5581
5582 if (Op0->getType()->isIntOrIntVectorTy(1))
5583 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
5584 return Res;
5585
5586 if (Instruction *Res = canonicalizeCmpWithConstant(I))
5587 return Res;
5588
5589 if (Instruction *Res = canonicalizeICmpPredicate(I))
5590 return Res;
5591
5592 if (Instruction *Res = foldICmpWithConstant(I))
5593 return Res;
5594
5595 if (Instruction *Res = foldICmpWithDominatingICmp(I))
5596 return Res;
5597
5598 if (Instruction *Res = foldICmpBinOp(I, Q))
5599 return Res;
5600
5601 if (Instruction *Res = foldICmpUsingKnownBits(I))
5602 return Res;
5603