Bug Summary

File:llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
Warning:line 3491, column 7
Called C++ object pointer is uninitialized

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -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 -mthread-model posix -mframe-pointer=none -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -target-cpu x86-64 -dwarf-column-info -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-11/lib/clang/11.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/build-llvm/include -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/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-11/lib/clang/11.0.0/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-11~++20200309111110+2c36c23f347/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-03-09-184146-41876-1 -x c++ /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/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
28using namespace llvm;
29using namespace PatternMatch;
30
31#define DEBUG_TYPE"instcombine" "instcombine"
32
33// How many times is a select replaced by one of its operands?
34STATISTIC(NumSel, "Number of select opts")static llvm::Statistic NumSel = {"instcombine", "NumSel", "Number of select opts"
}
;
35
36
37/// Compute Result = In1+In2, returning true if the result overflowed for this
38/// type.
39static bool addWithOverflow(APInt &Result, const APInt &In1,
40 const APInt &In2, bool IsSigned = false) {
41 bool Overflow;
42 if (IsSigned)
43 Result = In1.sadd_ov(In2, Overflow);
44 else
45 Result = In1.uadd_ov(In2, Overflow);
46
47 return Overflow;
48}
49
50/// Compute Result = In1-In2, returning true if the result overflowed for this
51/// type.
52static bool subWithOverflow(APInt &Result, const APInt &In1,
53 const APInt &In2, bool IsSigned = false) {
54 bool Overflow;
55 if (IsSigned)
56 Result = In1.ssub_ov(In2, Overflow);
57 else
58 Result = In1.usub_ov(In2, Overflow);
59
60 return Overflow;
61}
62
63/// Given an icmp instruction, return true if any use of this comparison is a
64/// branch on sign bit comparison.
65static bool hasBranchUse(ICmpInst &I) {
66 for (auto *U : I.users())
67 if (isa<BranchInst>(U))
68 return true;
69 return false;
70}
71
72/// Returns true if the exploded icmp can be expressed as a signed comparison
73/// to zero and updates the predicate accordingly.
74/// The signedness of the comparison is preserved.
75/// TODO: Refactor with decomposeBitTestICmp()?
76static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
77 if (!ICmpInst::isSigned(Pred))
78 return false;
79
80 if (C.isNullValue())
81 return ICmpInst::isRelational(Pred);
82
83 if (C.isOneValue()) {
84 if (Pred == ICmpInst::ICMP_SLT) {
85 Pred = ICmpInst::ICMP_SLE;
86 return true;
87 }
88 } else if (C.isAllOnesValue()) {
89 if (Pred == ICmpInst::ICMP_SGT) {
90 Pred = ICmpInst::ICMP_SGE;
91 return true;
92 }
93 }
94
95 return false;
96}
97
98/// Given a signed integer type and a set of known zero and one bits, compute
99/// the maximum and minimum values that could have the specified known zero and
100/// known one bits, returning them in Min/Max.
101/// TODO: Move to method on KnownBits struct?
102static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
103 APInt &Min, APInt &Max) {
104 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 106, __PRETTY_FUNCTION__))
105 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 106, __PRETTY_FUNCTION__))
106 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 106, __PRETTY_FUNCTION__))
;
107 APInt UnknownBits = ~(Known.Zero|Known.One);
108
109 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
110 // bit if it is unknown.
111 Min = Known.One;
112 Max = Known.One|UnknownBits;
113
114 if (UnknownBits.isNegative()) { // Sign bit is unknown
115 Min.setSignBit();
116 Max.clearSignBit();
117 }
118}
119
120/// Given an unsigned integer type and a set of known zero and one bits, compute
121/// the maximum and minimum values that could have the specified known zero and
122/// known one bits, returning them in Min/Max.
123/// TODO: Move to method on KnownBits struct?
124static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
125 APInt &Min, APInt &Max) {
126 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 128, __PRETTY_FUNCTION__))
127 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 128, __PRETTY_FUNCTION__))
128 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 128, __PRETTY_FUNCTION__))
;
129 APInt UnknownBits = ~(Known.Zero|Known.One);
130
131 // The minimum value is when the unknown bits are all zeros.
132 Min = Known.One;
133 // The maximum value is when the unknown bits are all ones.
134 Max = Known.One|UnknownBits;
135}
136
137/// This is called when we see this pattern:
138/// cmp pred (load (gep GV, ...)), cmpcst
139/// where GV is a global variable with a constant initializer. Try to simplify
140/// this into some simple computation that does not need the load. For example
141/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
142///
143/// If AndCst is non-null, then the loaded value is masked with that constant
144/// before doing the comparison. This handles cases like "A[i]&4 == 0".
145Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
146 GlobalVariable *GV,
147 CmpInst &ICI,
148 ConstantInt *AndCst) {
149 Constant *Init = GV->getInitializer();
150 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
151 return nullptr;
152
153 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
154 // Don't blow up on huge arrays.
155 if (ArrayElementCount > MaxArraySizeForCombine)
156 return nullptr;
157
158 // There are many forms of this optimization we can handle, for now, just do
159 // the simple index into a single-dimensional array.
160 //
161 // Require: GEP GV, 0, i {{, constant indices}}
162 if (GEP->getNumOperands() < 3 ||
163 !isa<ConstantInt>(GEP->getOperand(1)) ||
164 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
165 isa<Constant>(GEP->getOperand(2)))
166 return nullptr;
167
168 // Check that indices after the variable are constants and in-range for the
169 // type they index. Collect the indices. This is typically for arrays of
170 // structs.
171 SmallVector<unsigned, 4> LaterIndices;
172
173 Type *EltTy = Init->getType()->getArrayElementType();
174 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
175 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
176 if (!Idx) return nullptr; // Variable index.
177
178 uint64_t IdxVal = Idx->getZExtValue();
179 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
180
181 if (StructType *STy = dyn_cast<StructType>(EltTy))
182 EltTy = STy->getElementType(IdxVal);
183 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
184 if (IdxVal >= ATy->getNumElements()) return nullptr;
185 EltTy = ATy->getElementType();
186 } else {
187 return nullptr; // Unknown type.
188 }
189
190 LaterIndices.push_back(IdxVal);
191 }
192
193 enum { Overdefined = -3, Undefined = -2 };
194
195 // Variables for our state machines.
196
197 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
198 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
199 // and 87 is the second (and last) index. FirstTrueElement is -2 when
200 // undefined, otherwise set to the first true element. SecondTrueElement is
201 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
202 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
203
204 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
205 // form "i != 47 & i != 87". Same state transitions as for true elements.
206 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
207
208 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
209 /// define a state machine that triggers for ranges of values that the index
210 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
211 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
212 /// index in the range (inclusive). We use -2 for undefined here because we
213 /// use relative comparisons and don't want 0-1 to match -1.
214 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
215
216 // MagicBitvector - This is a magic bitvector where we set a bit if the
217 // comparison is true for element 'i'. If there are 64 elements or less in
218 // the array, this will fully represent all the comparison results.
219 uint64_t MagicBitvector = 0;
220
221 // Scan the array and see if one of our patterns matches.
222 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
223 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
224 Constant *Elt = Init->getAggregateElement(i);
225 if (!Elt) return nullptr;
226
227 // If this is indexing an array of structures, get the structure element.
228 if (!LaterIndices.empty())
229 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
230
231 // If the element is masked, handle it.
232 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
233
234 // Find out if the comparison would be true or false for the i'th element.
235 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
236 CompareRHS, DL, &TLI);
237 // If the result is undef for this element, ignore it.
238 if (isa<UndefValue>(C)) {
239 // Extend range state machines to cover this element in case there is an
240 // undef in the middle of the range.
241 if (TrueRangeEnd == (int)i-1)
242 TrueRangeEnd = i;
243 if (FalseRangeEnd == (int)i-1)
244 FalseRangeEnd = i;
245 continue;
246 }
247
248 // If we can't compute the result for any of the elements, we have to give
249 // up evaluating the entire conditional.
250 if (!isa<ConstantInt>(C)) return nullptr;
251
252 // Otherwise, we know if the comparison is true or false for this element,
253 // update our state machines.
254 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
255
256 // State machine for single/double/range index comparison.
257 if (IsTrueForElt) {
258 // Update the TrueElement state machine.
259 if (FirstTrueElement == Undefined)
260 FirstTrueElement = TrueRangeEnd = i; // First true element.
261 else {
262 // Update double-compare state machine.
263 if (SecondTrueElement == Undefined)
264 SecondTrueElement = i;
265 else
266 SecondTrueElement = Overdefined;
267
268 // Update range state machine.
269 if (TrueRangeEnd == (int)i-1)
270 TrueRangeEnd = i;
271 else
272 TrueRangeEnd = Overdefined;
273 }
274 } else {
275 // Update the FalseElement state machine.
276 if (FirstFalseElement == Undefined)
277 FirstFalseElement = FalseRangeEnd = i; // First false element.
278 else {
279 // Update double-compare state machine.
280 if (SecondFalseElement == Undefined)
281 SecondFalseElement = i;
282 else
283 SecondFalseElement = Overdefined;
284
285 // Update range state machine.
286 if (FalseRangeEnd == (int)i-1)
287 FalseRangeEnd = i;
288 else
289 FalseRangeEnd = Overdefined;
290 }
291 }
292
293 // If this element is in range, update our magic bitvector.
294 if (i < 64 && IsTrueForElt)
295 MagicBitvector |= 1ULL << i;
296
297 // If all of our states become overdefined, bail out early. Since the
298 // predicate is expensive, only check it every 8 elements. This is only
299 // really useful for really huge arrays.
300 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
301 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
302 FalseRangeEnd == Overdefined)
303 return nullptr;
304 }
305
306 // Now that we've scanned the entire array, emit our new comparison(s). We
307 // order the state machines in complexity of the generated code.
308 Value *Idx = GEP->getOperand(2);
309
310 // If the index is larger than the pointer size of the target, truncate the
311 // index down like the GEP would do implicitly. We don't have to do this for
312 // an inbounds GEP because the index can't be out of range.
313 if (!GEP->isInBounds()) {
314 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
315 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
316 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
317 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
318 }
319
320 // If the comparison is only true for one or two elements, emit direct
321 // comparisons.
322 if (SecondTrueElement != Overdefined) {
323 // None true -> false.
324 if (FirstTrueElement == Undefined)
325 return replaceInstUsesWith(ICI, Builder.getFalse());
326
327 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
328
329 // True for one element -> 'i == 47'.
330 if (SecondTrueElement == Undefined)
331 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
332
333 // True for two elements -> 'i == 47 | i == 72'.
334 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
335 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
336 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
337 return BinaryOperator::CreateOr(C1, C2);
338 }
339
340 // If the comparison is only false for one or two elements, emit direct
341 // comparisons.
342 if (SecondFalseElement != Overdefined) {
343 // None false -> true.
344 if (FirstFalseElement == Undefined)
345 return replaceInstUsesWith(ICI, Builder.getTrue());
346
347 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
348
349 // False for one element -> 'i != 47'.
350 if (SecondFalseElement == Undefined)
351 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
352
353 // False for two elements -> 'i != 47 & i != 72'.
354 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
355 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
356 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
357 return BinaryOperator::CreateAnd(C1, C2);
358 }
359
360 // If the comparison can be replaced with a range comparison for the elements
361 // where it is true, emit the range check.
362 if (TrueRangeEnd != Overdefined) {
363 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 363, __PRETTY_FUNCTION__))
;
364
365 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
366 if (FirstTrueElement) {
367 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
368 Idx = Builder.CreateAdd(Idx, Offs);
369 }
370
371 Value *End = ConstantInt::get(Idx->getType(),
372 TrueRangeEnd-FirstTrueElement+1);
373 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
374 }
375
376 // False range check.
377 if (FalseRangeEnd != Overdefined) {
378 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 378, __PRETTY_FUNCTION__))
;
379 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
380 if (FirstFalseElement) {
381 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
382 Idx = Builder.CreateAdd(Idx, Offs);
383 }
384
385 Value *End = ConstantInt::get(Idx->getType(),
386 FalseRangeEnd-FirstFalseElement);
387 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
388 }
389
390 // If a magic bitvector captures the entire comparison state
391 // of this load, replace it with computation that does:
392 // ((magic_cst >> i) & 1) != 0
393 {
394 Type *Ty = nullptr;
395
396 // Look for an appropriate type:
397 // - The type of Idx if the magic fits
398 // - The smallest fitting legal type
399 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
400 Ty = Idx->getType();
401 else
402 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
403
404 if (Ty) {
405 Value *V = Builder.CreateIntCast(Idx, Ty, false);
406 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
407 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
408 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
409 }
410 }
411
412 return nullptr;
413}
414
415/// Return a value that can be used to compare the *offset* implied by a GEP to
416/// zero. For example, if we have &A[i], we want to return 'i' for
417/// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
418/// are involved. The above expression would also be legal to codegen as
419/// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
420/// This latter form is less amenable to optimization though, and we are allowed
421/// to generate the first by knowing that pointer arithmetic doesn't overflow.
422///
423/// If we can't emit an optimized form for this expression, this returns null.
424///
425static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
426 const DataLayout &DL) {
427 gep_type_iterator GTI = gep_type_begin(GEP);
428
429 // Check to see if this gep only has a single variable index. If so, and if
430 // any constant indices are a multiple of its scale, then we can compute this
431 // in terms of the scale of the variable index. For example, if the GEP
432 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
433 // because the expression will cross zero at the same point.
434 unsigned i, e = GEP->getNumOperands();
435 int64_t Offset = 0;
436 for (i = 1; i != e; ++i, ++GTI) {
437 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
438 // Compute the aggregate offset of constant indices.
439 if (CI->isZero()) continue;
440
441 // Handle a struct index, which adds its field offset to the pointer.
442 if (StructType *STy = GTI.getStructTypeOrNull()) {
443 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
444 } else {
445 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
446 Offset += Size*CI->getSExtValue();
447 }
448 } else {
449 // Found our variable index.
450 break;
451 }
452 }
453
454 // If there are no variable indices, we must have a constant offset, just
455 // evaluate it the general way.
456 if (i == e) return nullptr;
457
458 Value *VariableIdx = GEP->getOperand(i);
459 // Determine the scale factor of the variable element. For example, this is
460 // 4 if the variable index is into an array of i32.
461 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
462
463 // Verify that there are no other variable indices. If so, emit the hard way.
464 for (++i, ++GTI; i != e; ++i, ++GTI) {
465 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
466 if (!CI) return nullptr;
467
468 // Compute the aggregate offset of constant indices.
469 if (CI->isZero()) continue;
470
471 // Handle a struct index, which adds its field offset to the pointer.
472 if (StructType *STy = GTI.getStructTypeOrNull()) {
473 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
474 } else {
475 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
476 Offset += Size*CI->getSExtValue();
477 }
478 }
479
480 // Okay, we know we have a single variable index, which must be a
481 // pointer/array/vector index. If there is no offset, life is simple, return
482 // the index.
483 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
484 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
485 if (Offset == 0) {
486 // Cast to intptrty in case a truncation occurs. If an extension is needed,
487 // we don't need to bother extending: the extension won't affect where the
488 // computation crosses zero.
489 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
490 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
491 }
492 return VariableIdx;
493 }
494
495 // Otherwise, there is an index. The computation we will do will be modulo
496 // the pointer size.
497 Offset = SignExtend64(Offset, IntPtrWidth);
498 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
499
500 // To do this transformation, any constant index must be a multiple of the
501 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
502 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
503 // multiple of the variable scale.
504 int64_t NewOffs = Offset / (int64_t)VariableScale;
505 if (Offset != NewOffs*(int64_t)VariableScale)
506 return nullptr;
507
508 // Okay, we can do this evaluation. Start by converting the index to intptr.
509 if (VariableIdx->getType() != IntPtrTy)
510 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
511 true /*Signed*/);
512 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
513 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
514}
515
516/// Returns true if we can rewrite Start as a GEP with pointer Base
517/// and some integer offset. The nodes that need to be re-written
518/// for this transformation will be added to Explored.
519static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
520 const DataLayout &DL,
521 SetVector<Value *> &Explored) {
522 SmallVector<Value *, 16> WorkList(1, Start);
523 Explored.insert(Base);
524
525 // The following traversal gives us an order which can be used
526 // when doing the final transformation. Since in the final
527 // transformation we create the PHI replacement instructions first,
528 // we don't have to get them in any particular order.
529 //
530 // However, for other instructions we will have to traverse the
531 // operands of an instruction first, which means that we have to
532 // do a post-order traversal.
533 while (!WorkList.empty()) {
534 SetVector<PHINode *> PHIs;
535
536 while (!WorkList.empty()) {
537 if (Explored.size() >= 100)
538 return false;
539
540 Value *V = WorkList.back();
541
542 if (Explored.count(V) != 0) {
543 WorkList.pop_back();
544 continue;
545 }
546
547 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
548 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
549 // We've found some value that we can't explore which is different from
550 // the base. Therefore we can't do this transformation.
551 return false;
552
553 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
554 auto *CI = dyn_cast<CastInst>(V);
555 if (!CI->isNoopCast(DL))
556 return false;
557
558 if (Explored.count(CI->getOperand(0)) == 0)
559 WorkList.push_back(CI->getOperand(0));
560 }
561
562 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
563 // We're limiting the GEP to having one index. This will preserve
564 // the original pointer type. We could handle more cases in the
565 // future.
566 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
567 GEP->getType() != Start->getType())
568 return false;
569
570 if (Explored.count(GEP->getOperand(0)) == 0)
571 WorkList.push_back(GEP->getOperand(0));
572 }
573
574 if (WorkList.back() == V) {
575 WorkList.pop_back();
576 // We've finished visiting this node, mark it as such.
577 Explored.insert(V);
578 }
579
580 if (auto *PN = dyn_cast<PHINode>(V)) {
581 // We cannot transform PHIs on unsplittable basic blocks.
582 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
583 return false;
584 Explored.insert(PN);
585 PHIs.insert(PN);
586 }
587 }
588
589 // Explore the PHI nodes further.
590 for (auto *PN : PHIs)
591 for (Value *Op : PN->incoming_values())
592 if (Explored.count(Op) == 0)
593 WorkList.push_back(Op);
594 }
595
596 // Make sure that we can do this. Since we can't insert GEPs in a basic
597 // block before a PHI node, we can't easily do this transformation if
598 // we have PHI node users of transformed instructions.
599 for (Value *Val : Explored) {
600 for (Value *Use : Val->uses()) {
601
602 auto *PHI = dyn_cast<PHINode>(Use);
603 auto *Inst = dyn_cast<Instruction>(Val);
604
605 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
606 Explored.count(PHI) == 0)
607 continue;
608
609 if (PHI->getParent() == Inst->getParent())
610 return false;
611 }
612 }
613 return true;
614}
615
616// Sets the appropriate insert point on Builder where we can add
617// a replacement Instruction for V (if that is possible).
618static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
619 bool Before = true) {
620 if (auto *PHI = dyn_cast<PHINode>(V)) {
621 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
622 return;
623 }
624 if (auto *I = dyn_cast<Instruction>(V)) {
625 if (!Before)
626 I = &*std::next(I->getIterator());
627 Builder.SetInsertPoint(I);
628 return;
629 }
630 if (auto *A = dyn_cast<Argument>(V)) {
631 // Set the insertion point in the entry block.
632 BasicBlock &Entry = A->getParent()->getEntryBlock();
633 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
634 return;
635 }
636 // Otherwise, this is a constant and we don't need to set a new
637 // insertion point.
638 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 638, __PRETTY_FUNCTION__))
;
639}
640
641/// Returns a re-written value of Start as an indexed GEP using Base as a
642/// pointer.
643static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
644 const DataLayout &DL,
645 SetVector<Value *> &Explored) {
646 // Perform all the substitutions. This is a bit tricky because we can
647 // have cycles in our use-def chains.
648 // 1. Create the PHI nodes without any incoming values.
649 // 2. Create all the other values.
650 // 3. Add the edges for the PHI nodes.
651 // 4. Emit GEPs to get the original pointers.
652 // 5. Remove the original instructions.
653 Type *IndexType = IntegerType::get(
654 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
655
656 DenseMap<Value *, Value *> NewInsts;
657 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
658
659 // Create the new PHI nodes, without adding any incoming values.
660 for (Value *Val : Explored) {
661 if (Val == Base)
662 continue;
663 // Create empty phi nodes. This avoids cyclic dependencies when creating
664 // the remaining instructions.
665 if (auto *PHI = dyn_cast<PHINode>(Val))
666 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
667 PHI->getName() + ".idx", PHI);
668 }
669 IRBuilder<> Builder(Base->getContext());
670
671 // Create all the other instructions.
672 for (Value *Val : Explored) {
673
674 if (NewInsts.find(Val) != NewInsts.end())
675 continue;
676
677 if (auto *CI = dyn_cast<CastInst>(Val)) {
678 // Don't get rid of the intermediate variable here; the store can grow
679 // the map which will invalidate the reference to the input value.
680 Value *V = NewInsts[CI->getOperand(0)];
681 NewInsts[CI] = V;
682 continue;
683 }
684 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
685 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
686 : GEP->getOperand(1);
687 setInsertionPoint(Builder, GEP);
688 // Indices might need to be sign extended. GEPs will magically do
689 // this, but we need to do it ourselves here.
690 if (Index->getType()->getScalarSizeInBits() !=
691 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
692 Index = Builder.CreateSExtOrTrunc(
693 Index, NewInsts[GEP->getOperand(0)]->getType(),
694 GEP->getOperand(0)->getName() + ".sext");
695 }
696
697 auto *Op = NewInsts[GEP->getOperand(0)];
698 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
699 NewInsts[GEP] = Index;
700 else
701 NewInsts[GEP] = Builder.CreateNSWAdd(
702 Op, Index, GEP->getOperand(0)->getName() + ".add");
703 continue;
704 }
705 if (isa<PHINode>(Val))
706 continue;
707
708 llvm_unreachable("Unexpected instruction type")::llvm::llvm_unreachable_internal("Unexpected instruction type"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 708)
;
709 }
710
711 // Add the incoming values to the PHI nodes.
712 for (Value *Val : Explored) {
713 if (Val == Base)
714 continue;
715 // All the instructions have been created, we can now add edges to the
716 // phi nodes.
717 if (auto *PHI = dyn_cast<PHINode>(Val)) {
718 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
719 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
720 Value *NewIncoming = PHI->getIncomingValue(I);
721
722 if (NewInsts.find(NewIncoming) != NewInsts.end())
723 NewIncoming = NewInsts[NewIncoming];
724
725 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
726 }
727 }
728 }
729
730 for (Value *Val : Explored) {
731 if (Val == Base)
732 continue;
733
734 // Depending on the type, for external users we have to emit
735 // a GEP or a GEP + ptrtoint.
736 setInsertionPoint(Builder, Val, false);
737
738 // If required, create an inttoptr instruction for Base.
739 Value *NewBase = Base;
740 if (!Base->getType()->isPointerTy())
741 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
742 Start->getName() + "to.ptr");
743
744 Value *GEP = Builder.CreateInBoundsGEP(
745 Start->getType()->getPointerElementType(), NewBase,
746 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
747
748 if (!Val->getType()->isPointerTy()) {
749 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
750 Val->getName() + ".conv");
751 GEP = Cast;
752 }
753 Val->replaceAllUsesWith(GEP);
754 }
755
756 return NewInsts[Start];
757}
758
759/// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
760/// the input Value as a constant indexed GEP. Returns a pair containing
761/// the GEPs Pointer and Index.
762static std::pair<Value *, Value *>
763getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
764 Type *IndexType = IntegerType::get(V->getContext(),
765 DL.getIndexTypeSizeInBits(V->getType()));
766
767 Constant *Index = ConstantInt::getNullValue(IndexType);
768 while (true) {
769 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
770 // We accept only inbouds GEPs here to exclude the possibility of
771 // overflow.
772 if (!GEP->isInBounds())
773 break;
774 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
775 GEP->getType() == V->getType()) {
776 V = GEP->getOperand(0);
777 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
778 Index = ConstantExpr::getAdd(
779 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
780 continue;
781 }
782 break;
783 }
784 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
785 if (!CI->isNoopCast(DL))
786 break;
787 V = CI->getOperand(0);
788 continue;
789 }
790 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
791 if (!CI->isNoopCast(DL))
792 break;
793 V = CI->getOperand(0);
794 continue;
795 }
796 break;
797 }
798 return {V, Index};
799}
800
801/// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
802/// We can look through PHIs, GEPs and casts in order to determine a common base
803/// between GEPLHS and RHS.
804static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
805 ICmpInst::Predicate Cond,
806 const DataLayout &DL) {
807 // FIXME: Support vector of pointers.
808 if (GEPLHS->getType()->isVectorTy())
809 return nullptr;
810
811 if (!GEPLHS->hasAllConstantIndices())
812 return nullptr;
813
814 // Make sure the pointers have the same type.
815 if (GEPLHS->getType() != RHS->getType())
816 return nullptr;
817
818 Value *PtrBase, *Index;
819 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
820
821 // The set of nodes that will take part in this transformation.
822 SetVector<Value *> Nodes;
823
824 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
825 return nullptr;
826
827 // We know we can re-write this as
828 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
829 // Since we've only looked through inbouds GEPs we know that we
830 // can't have overflow on either side. We can therefore re-write
831 // this as:
832 // OFFSET1 cmp OFFSET2
833 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
834
835 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
836 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
837 // offset. Since Index is the offset of LHS to the base pointer, we will now
838 // compare the offsets instead of comparing the pointers.
839 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
840}
841
842/// Fold comparisons between a GEP instruction and something else. At this point
843/// we know that the GEP is on the LHS of the comparison.
844Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
845 ICmpInst::Predicate Cond,
846 Instruction &I) {
847 // Don't transform signed compares of GEPs into index compares. Even if the
848 // GEP is inbounds, the final add of the base pointer can have signed overflow
849 // and would change the result of the icmp.
850 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
851 // the maximum signed value for the pointer type.
852 if (ICmpInst::isSigned(Cond))
853 return nullptr;
854
855 // Look through bitcasts and addrspacecasts. We do not however want to remove
856 // 0 GEPs.
857 if (!isa<GetElementPtrInst>(RHS))
858 RHS = RHS->stripPointerCasts();
859
860 Value *PtrBase = GEPLHS->getOperand(0);
861 // FIXME: Support vector pointer GEPs.
862 if (PtrBase == RHS && GEPLHS->isInBounds() &&
863 !GEPLHS->getType()->isVectorTy()) {
864 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
865 // This transformation (ignoring the base and scales) is valid because we
866 // know pointers can't overflow since the gep is inbounds. See if we can
867 // output an optimized form.
868 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
869
870 // If not, synthesize the offset the hard way.
871 if (!Offset)
872 Offset = EmitGEPOffset(GEPLHS);
873 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
874 Constant::getNullValue(Offset->getType()));
875 }
876
877 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
878 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
879 !NullPointerIsDefined(I.getFunction(),
880 RHS->getType()->getPointerAddressSpace())) {
881 // For most address spaces, an allocation can't be placed at null, but null
882 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
883 // the only valid inbounds address derived from null, is null itself.
884 // Thus, we have four cases to consider:
885 // 1) Base == nullptr, Offset == 0 -> inbounds, null
886 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
887 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
888 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
889 //
890 // (Note if we're indexing a type of size 0, that simply collapses into one
891 // of the buckets above.)
892 //
893 // In general, we're allowed to make values less poison (i.e. remove
894 // sources of full UB), so in this case, we just select between the two
895 // non-poison cases (1 and 4 above).
896 //
897 // For vectors, we apply the same reasoning on a per-lane basis.
898 auto *Base = GEPLHS->getPointerOperand();
899 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
900 int NumElts = GEPLHS->getType()->getVectorNumElements();
901 Base = Builder.CreateVectorSplat(NumElts, Base);
902 }
903 return new ICmpInst(Cond, Base,
904 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
905 cast<Constant>(RHS), Base->getType()));
906 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
907 // If the base pointers are different, but the indices are the same, just
908 // compare the base pointer.
909 if (PtrBase != GEPRHS->getOperand(0)) {
910 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
911 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
912 GEPRHS->getOperand(0)->getType();
913 if (IndicesTheSame)
914 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
915 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
916 IndicesTheSame = false;
917 break;
918 }
919
920 // If all indices are the same, just compare the base pointers.
921 Type *BaseType = GEPLHS->getOperand(0)->getType();
922 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
923 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
924
925 // If we're comparing GEPs with two base pointers that only differ in type
926 // and both GEPs have only constant indices or just one use, then fold
927 // the compare with the adjusted indices.
928 // FIXME: Support vector of pointers.
929 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
930 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
931 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
932 PtrBase->stripPointerCasts() ==
933 GEPRHS->getOperand(0)->stripPointerCasts() &&
934 !GEPLHS->getType()->isVectorTy()) {
935 Value *LOffset = EmitGEPOffset(GEPLHS);
936 Value *ROffset = EmitGEPOffset(GEPRHS);
937
938 // If we looked through an addrspacecast between different sized address
939 // spaces, the LHS and RHS pointers are different sized
940 // integers. Truncate to the smaller one.
941 Type *LHSIndexTy = LOffset->getType();
942 Type *RHSIndexTy = ROffset->getType();
943 if (LHSIndexTy != RHSIndexTy) {
944 if (LHSIndexTy->getPrimitiveSizeInBits() <
945 RHSIndexTy->getPrimitiveSizeInBits()) {
946 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
947 } else
948 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
949 }
950
951 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
952 LOffset, ROffset);
953 return replaceInstUsesWith(I, Cmp);
954 }
955
956 // Otherwise, the base pointers are different and the indices are
957 // different. Try convert this to an indexed compare by looking through
958 // PHIs/casts.
959 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
960 }
961
962 // If one of the GEPs has all zero indices, recurse.
963 // FIXME: Handle vector of pointers.
964 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
965 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
966 ICmpInst::getSwappedPredicate(Cond), I);
967
968 // If the other GEP has all zero indices, recurse.
969 // FIXME: Handle vector of pointers.
970 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
971 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
972
973 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
974 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
975 // If the GEPs only differ by one index, compare it.
976 unsigned NumDifferences = 0; // Keep track of # differences.
977 unsigned DiffOperand = 0; // The operand that differs.
978 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
979 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
980 Type *LHSType = GEPLHS->getOperand(i)->getType();
981 Type *RHSType = GEPRHS->getOperand(i)->getType();
982 // FIXME: Better support for vector of pointers.
983 if (LHSType->getPrimitiveSizeInBits() !=
984 RHSType->getPrimitiveSizeInBits() ||
985 (GEPLHS->getType()->isVectorTy() &&
986 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
987 // Irreconcilable differences.
988 NumDifferences = 2;
989 break;
990 }
991
992 if (NumDifferences++) break;
993 DiffOperand = i;
994 }
995
996 if (NumDifferences == 0) // SAME GEP?
997 return replaceInstUsesWith(I, // No comparison is needed here.
998 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
999
1000 else if (NumDifferences == 1 && GEPsInBounds) {
1001 Value *LHSV = GEPLHS->getOperand(DiffOperand);
1002 Value *RHSV = GEPRHS->getOperand(DiffOperand);
1003 // Make sure we do a signed comparison here.
1004 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1005 }
1006 }
1007
1008 // Only lower this if the icmp is the only user of the GEP or if we expect
1009 // the result to fold to a constant!
1010 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1011 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1012 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1013 Value *L = EmitGEPOffset(GEPLHS);
1014 Value *R = EmitGEPOffset(GEPRHS);
1015 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1016 }
1017 }
1018
1019 // Try convert this to an indexed compare by looking through PHIs/casts as a
1020 // last resort.
1021 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1022}
1023
1024Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1025 const AllocaInst *Alloca,
1026 const Value *Other) {
1027 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1027, __PRETTY_FUNCTION__))
;
1028
1029 // It would be tempting to fold away comparisons between allocas and any
1030 // pointer not based on that alloca (e.g. an argument). However, even
1031 // though such pointers cannot alias, they can still compare equal.
1032 //
1033 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1034 // doesn't escape we can argue that it's impossible to guess its value, and we
1035 // can therefore act as if any such guesses are wrong.
1036 //
1037 // The code below checks that the alloca doesn't escape, and that it's only
1038 // used in a comparison once (the current instruction). The
1039 // single-comparison-use condition ensures that we're trivially folding all
1040 // comparisons against the alloca consistently, and avoids the risk of
1041 // erroneously folding a comparison of the pointer with itself.
1042
1043 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1044
1045 SmallVector<const Use *, 32> Worklist;
1046 for (const Use &U : Alloca->uses()) {
1047 if (Worklist.size() >= MaxIter)
1048 return nullptr;
1049 Worklist.push_back(&U);
1050 }
1051
1052 unsigned NumCmps = 0;
1053 while (!Worklist.empty()) {
1054 assert(Worklist.size() <= MaxIter)((Worklist.size() <= MaxIter) ? static_cast<void> (0
) : __assert_fail ("Worklist.size() <= MaxIter", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1054, __PRETTY_FUNCTION__))
;
1055 const Use *U = Worklist.pop_back_val();
1056 const Value *V = U->getUser();
1057 --MaxIter;
1058
1059 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1060 isa<SelectInst>(V)) {
1061 // Track the uses.
1062 } else if (isa<LoadInst>(V)) {
1063 // Loading from the pointer doesn't escape it.
1064 continue;
1065 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1066 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1067 if (SI->getValueOperand() == U->get())
1068 return nullptr;
1069 continue;
1070 } else if (isa<ICmpInst>(V)) {
1071 if (NumCmps++)
1072 return nullptr; // Found more than one cmp.
1073 continue;
1074 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1075 switch (Intrin->getIntrinsicID()) {
1076 // These intrinsics don't escape or compare the pointer. Memset is safe
1077 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1078 // we don't allow stores, so src cannot point to V.
1079 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1080 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1081 continue;
1082 default:
1083 return nullptr;
1084 }
1085 } else {
1086 return nullptr;
1087 }
1088 for (const Use &U : V->uses()) {
1089 if (Worklist.size() >= MaxIter)
1090 return nullptr;
1091 Worklist.push_back(&U);
1092 }
1093 }
1094
1095 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1096 return replaceInstUsesWith(
1097 ICI,
1098 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1099}
1100
1101/// Fold "icmp pred (X+C), X".
1102Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1103 ICmpInst::Predicate Pred) {
1104 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1105 // so the values can never be equal. Similarly for all other "or equals"
1106 // operators.
1107 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1107, __PRETTY_FUNCTION__))
;
1108
1109 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1110 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1111 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1112 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1113 Constant *R = ConstantInt::get(X->getType(),
1114 APInt::getMaxValue(C.getBitWidth()) - C);
1115 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1116 }
1117
1118 // (X+1) >u X --> X <u (0-1) --> X != 255
1119 // (X+2) >u X --> X <u (0-2) --> X <u 254
1120 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1121 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1122 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1123 ConstantInt::get(X->getType(), -C));
1124
1125 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1126
1127 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1128 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1129 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1130 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1131 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1132 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1133 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1134 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1135 ConstantInt::get(X->getType(), SMax - C));
1136
1137 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1138 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1139 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1140 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1141 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1142 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1143
1144 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1144, __PRETTY_FUNCTION__))
;
1145 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1146 ConstantInt::get(X->getType(), SMax - (C - 1)));
1147}
1148
1149/// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1150/// (icmp eq/ne A, Log2(AP2/AP1)) ->
1151/// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1152Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1153 const APInt &AP1,
1154 const APInt &AP2) {
1155 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1155, __PRETTY_FUNCTION__))
;
1156
1157 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1158 if (I.getPredicate() == I.ICMP_NE)
1159 Pred = CmpInst::getInversePredicate(Pred);
1160 return new ICmpInst(Pred, LHS, RHS);
1161 };
1162
1163 // Don't bother doing any work for cases which InstSimplify handles.
1164 if (AP2.isNullValue())
1165 return nullptr;
1166
1167 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1168 if (IsAShr) {
1169 if (AP2.isAllOnesValue())
1170 return nullptr;
1171 if (AP2.isNegative() != AP1.isNegative())
1172 return nullptr;
1173 if (AP2.sgt(AP1))
1174 return nullptr;
1175 }
1176
1177 if (!AP1)
1178 // 'A' must be large enough to shift out the highest set bit.
1179 return getICmp(I.ICMP_UGT, A,
1180 ConstantInt::get(A->getType(), AP2.logBase2()));
1181
1182 if (AP1 == AP2)
1183 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1184
1185 int Shift;
1186 if (IsAShr && AP1.isNegative())
1187 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1188 else
1189 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1190
1191 if (Shift > 0) {
1192 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1193 // There are multiple solutions if we are comparing against -1 and the LHS
1194 // of the ashr is not a power of two.
1195 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1196 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1197 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1198 } else if (AP1 == AP2.lshr(Shift)) {
1199 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1200 }
1201 }
1202
1203 // Shifting const2 will never be equal to const1.
1204 // FIXME: This should always be handled by InstSimplify?
1205 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1206 return replaceInstUsesWith(I, TorF);
1207}
1208
1209/// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1210/// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1211Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1212 const APInt &AP1,
1213 const APInt &AP2) {
1214 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1214, __PRETTY_FUNCTION__))
;
1215
1216 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1217 if (I.getPredicate() == I.ICMP_NE)
1218 Pred = CmpInst::getInversePredicate(Pred);
1219 return new ICmpInst(Pred, LHS, RHS);
1220 };
1221
1222 // Don't bother doing any work for cases which InstSimplify handles.
1223 if (AP2.isNullValue())
1224 return nullptr;
1225
1226 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1227
1228 if (!AP1 && AP2TrailingZeros != 0)
1229 return getICmp(
1230 I.ICMP_UGE, A,
1231 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1232
1233 if (AP1 == AP2)
1234 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1235
1236 // Get the distance between the lowest bits that are set.
1237 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1238
1239 if (Shift > 0 && AP2.shl(Shift) == AP1)
1240 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1241
1242 // Shifting const2 will never be equal to const1.
1243 // FIXME: This should always be handled by InstSimplify?
1244 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1245 return replaceInstUsesWith(I, TorF);
1246}
1247
1248/// The caller has matched a pattern of the form:
1249/// I = icmp ugt (add (add A, B), CI2), CI1
1250/// If this is of the form:
1251/// sum = a + b
1252/// if (sum+128 >u 255)
1253/// Then replace it with llvm.sadd.with.overflow.i8.
1254///
1255static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1256 ConstantInt *CI2, ConstantInt *CI1,
1257 InstCombiner &IC) {
1258 // The transformation we're trying to do here is to transform this into an
1259 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1260 // with a narrower add, and discard the add-with-constant that is part of the
1261 // range check (if we can't eliminate it, this isn't profitable).
1262
1263 // In order to eliminate the add-with-constant, the compare can be its only
1264 // use.
1265 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1266 if (!AddWithCst->hasOneUse())
1267 return nullptr;
1268
1269 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1270 if (!CI2->getValue().isPowerOf2())
1271 return nullptr;
1272 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1273 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1274 return nullptr;
1275
1276 // The width of the new add formed is 1 more than the bias.
1277 ++NewWidth;
1278
1279 // Check to see that CI1 is an all-ones value with NewWidth bits.
1280 if (CI1->getBitWidth() == NewWidth ||
1281 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1282 return nullptr;
1283
1284 // This is only really a signed overflow check if the inputs have been
1285 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1286 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1287 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1288 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1289 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1290 return nullptr;
1291
1292 // In order to replace the original add with a narrower
1293 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1294 // and truncates that discard the high bits of the add. Verify that this is
1295 // the case.
1296 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1297 for (User *U : OrigAdd->users()) {
1298 if (U == AddWithCst)
1299 continue;
1300
1301 // Only accept truncates for now. We would really like a nice recursive
1302 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1303 // chain to see which bits of a value are actually demanded. If the
1304 // original add had another add which was then immediately truncated, we
1305 // could still do the transformation.
1306 TruncInst *TI = dyn_cast<TruncInst>(U);
1307 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1308 return nullptr;
1309 }
1310
1311 // If the pattern matches, truncate the inputs to the narrower type and
1312 // use the sadd_with_overflow intrinsic to efficiently compute both the
1313 // result and the overflow bit.
1314 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1315 Function *F = Intrinsic::getDeclaration(
1316 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1317
1318 InstCombiner::BuilderTy &Builder = IC.Builder;
1319
1320 // Put the new code above the original add, in case there are any uses of the
1321 // add between the add and the compare.
1322 Builder.SetInsertPoint(OrigAdd);
1323
1324 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1325 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1326 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1327 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1328 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1329
1330 // The inner add was the result of the narrow add, zero extended to the
1331 // wider type. Replace it with the result computed by the intrinsic.
1332 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1333
1334 // The original icmp gets replaced with the overflow value.
1335 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1336}
1337
1338/// If we have:
1339/// icmp eq/ne (urem/srem %x, %y), 0
1340/// iff %y is a power-of-two, we can replace this with a bit test:
1341/// icmp eq/ne (and %x, (add %y, -1)), 0
1342Instruction *InstCombiner::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1343 // This fold is only valid for equality predicates.
1344 if (!I.isEquality())
1345 return nullptr;
1346 ICmpInst::Predicate Pred;
1347 Value *X, *Y, *Zero;
1348 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1349 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1350 return nullptr;
1351 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1352 return nullptr;
1353 // This may increase instruction count, we don't enforce that Y is a constant.
1354 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1355 Value *Masked = Builder.CreateAnd(X, Mask);
1356 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1357}
1358
1359/// Fold equality-comparison between zero and any (maybe truncated) right-shift
1360/// by one-less-than-bitwidth into a sign test on the original value.
1361Instruction *InstCombiner::foldSignBitTest(ICmpInst &I) {
1362 Instruction *Val;
1363 ICmpInst::Predicate Pred;
1364 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1365 return nullptr;
1366
1367 Value *X;
1368 Type *XTy;
1369
1370 Constant *C;
1371 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1372 XTy = X->getType();
1373 unsigned XBitWidth = XTy->getScalarSizeInBits();
1374 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1375 APInt(XBitWidth, XBitWidth - 1))))
1376 return nullptr;
1377 } else if (isa<BinaryOperator>(Val) &&
1378 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1379 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1380 /*AnalyzeForSignBitExtraction=*/true))) {
1381 XTy = X->getType();
1382 } else
1383 return nullptr;
1384
1385 return ICmpInst::Create(Instruction::ICmp,
1386 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1387 : ICmpInst::ICMP_SLT,
1388 X, ConstantInt::getNullValue(XTy));
1389}
1390
1391// Handle icmp pred X, 0
1392Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1393 CmpInst::Predicate Pred = Cmp.getPredicate();
1394 if (!match(Cmp.getOperand(1), m_Zero()))
1395 return nullptr;
1396
1397 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1398 if (Pred == ICmpInst::ICMP_SGT) {
1399 Value *A, *B;
1400 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1401 if (SPR.Flavor == SPF_SMIN) {
1402 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1403 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1404 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1405 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1406 }
1407 }
1408
1409 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1410 return New;
1411
1412 // Given:
1413 // icmp eq/ne (urem %x, %y), 0
1414 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1415 // icmp eq/ne %x, 0
1416 Value *X, *Y;
1417 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1418 ICmpInst::isEquality(Pred)) {
1419 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1420 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1421 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1422 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1423 }
1424
1425 return nullptr;
1426}
1427
1428/// Fold icmp Pred X, C.
1429/// TODO: This code structure does not make sense. The saturating add fold
1430/// should be moved to some other helper and extended as noted below (it is also
1431/// possible that code has been made unnecessary - do we canonicalize IR to
1432/// overflow/saturating intrinsics or not?).
1433Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1434 // Match the following pattern, which is a common idiom when writing
1435 // overflow-safe integer arithmetic functions. The source performs an addition
1436 // in wider type and explicitly checks for overflow using comparisons against
1437 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1438 //
1439 // TODO: This could probably be generalized to handle other overflow-safe
1440 // operations if we worked out the formulas to compute the appropriate magic
1441 // constants.
1442 //
1443 // sum = a + b
1444 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1445 CmpInst::Predicate Pred = Cmp.getPredicate();
1446 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1447 Value *A, *B;
1448 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1449 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1450 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1451 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1452 return Res;
1453
1454 return nullptr;
1455}
1456
1457/// Canonicalize icmp instructions based on dominating conditions.
1458Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1459 // This is a cheap/incomplete check for dominance - just match a single
1460 // predecessor with a conditional branch.
1461 BasicBlock *CmpBB = Cmp.getParent();
1462 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1463 if (!DomBB)
1464 return nullptr;
1465
1466 Value *DomCond;
1467 BasicBlock *TrueBB, *FalseBB;
1468 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1469 return nullptr;
1470
1471 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1472, __PRETTY_FUNCTION__))
1472 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1472, __PRETTY_FUNCTION__))
;
1473
1474 // The branch should get simplified. Don't bother simplifying this condition.
1475 if (TrueBB == FalseBB)
1476 return nullptr;
1477
1478 // Try to simplify this compare to T/F based on the dominating condition.
1479 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1480 if (Imp)
1481 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1482
1483 CmpInst::Predicate Pred = Cmp.getPredicate();
1484 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1485 ICmpInst::Predicate DomPred;
1486 const APInt *C, *DomC;
1487 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1488 match(Y, m_APInt(C))) {
1489 // We have 2 compares of a variable with constants. Calculate the constant
1490 // ranges of those compares to see if we can transform the 2nd compare:
1491 // DomBB:
1492 // DomCond = icmp DomPred X, DomC
1493 // br DomCond, CmpBB, FalseBB
1494 // CmpBB:
1495 // Cmp = icmp Pred X, C
1496 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1497 ConstantRange DominatingCR =
1498 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1499 : ConstantRange::makeExactICmpRegion(
1500 CmpInst::getInversePredicate(DomPred), *DomC);
1501 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1502 ConstantRange Difference = DominatingCR.difference(CR);
1503 if (Intersection.isEmptySet())
1504 return replaceInstUsesWith(Cmp, Builder.getFalse());
1505 if (Difference.isEmptySet())
1506 return replaceInstUsesWith(Cmp, Builder.getTrue());
1507
1508 // Canonicalizing a sign bit comparison that gets used in a branch,
1509 // pessimizes codegen by generating branch on zero instruction instead
1510 // of a test and branch. So we avoid canonicalizing in such situations
1511 // because test and branch instruction has better branch displacement
1512 // than compare and branch instruction.
1513 bool UnusedBit;
1514 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1515 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1516 return nullptr;
1517
1518 if (const APInt *EqC = Intersection.getSingleElement())
1519 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1520 if (const APInt *NeC = Difference.getSingleElement())
1521 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1522 }
1523
1524 return nullptr;
1525}
1526
1527/// Fold icmp (trunc X, Y), C.
1528Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1529 TruncInst *Trunc,
1530 const APInt &C) {
1531 ICmpInst::Predicate Pred = Cmp.getPredicate();
1532 Value *X = Trunc->getOperand(0);
1533 if (C.isOneValue() && C.getBitWidth() > 1) {
1534 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1535 Value *V = nullptr;
1536 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1537 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1538 ConstantInt::get(V->getType(), 1));
1539 }
1540
1541 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1542 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1543 // of the high bits truncated out of x are known.
1544 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1545 SrcBits = X->getType()->getScalarSizeInBits();
1546 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1547
1548 // If all the high bits are known, we can do this xform.
1549 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1550 // Pull in the high bits from known-ones set.
1551 APInt NewRHS = C.zext(SrcBits);
1552 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1553 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1554 }
1555 }
1556
1557 return nullptr;
1558}
1559
1560/// Fold icmp (xor X, Y), C.
1561Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1562 BinaryOperator *Xor,
1563 const APInt &C) {
1564 Value *X = Xor->getOperand(0);
1565 Value *Y = Xor->getOperand(1);
1566 const APInt *XorC;
1567 if (!match(Y, m_APInt(XorC)))
1568 return nullptr;
1569
1570 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1571 // fold the xor.
1572 ICmpInst::Predicate Pred = Cmp.getPredicate();
1573 bool TrueIfSigned = false;
1574 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1575
1576 // If the sign bit of the XorCst is not set, there is no change to
1577 // the operation, just stop using the Xor.
1578 if (!XorC->isNegative())
1579 return replaceOperand(Cmp, 0, X);
1580
1581 // Emit the opposite comparison.
1582 if (TrueIfSigned)
1583 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1584 ConstantInt::getAllOnesValue(X->getType()));
1585 else
1586 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1587 ConstantInt::getNullValue(X->getType()));
1588 }
1589
1590 if (Xor->hasOneUse()) {
1591 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1592 if (!Cmp.isEquality() && XorC->isSignMask()) {
1593 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1594 : Cmp.getSignedPredicate();
1595 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1596 }
1597
1598 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1599 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1600 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1601 : Cmp.getSignedPredicate();
1602 Pred = Cmp.getSwappedPredicate(Pred);
1603 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1604 }
1605 }
1606
1607 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1608 if (Pred == ICmpInst::ICMP_UGT) {
1609 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1610 if (*XorC == ~C && (C + 1).isPowerOf2())
1611 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1612 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1613 if (*XorC == C && (C + 1).isPowerOf2())
1614 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1615 }
1616 if (Pred == ICmpInst::ICMP_ULT) {
1617 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1618 if (*XorC == -C && C.isPowerOf2())
1619 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1620 ConstantInt::get(X->getType(), ~C));
1621 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1622 if (*XorC == C && (-C).isPowerOf2())
1623 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1624 ConstantInt::get(X->getType(), ~C));
1625 }
1626 return nullptr;
1627}
1628
1629/// Fold icmp (and (sh X, Y), C2), C1.
1630Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1631 const APInt &C1, const APInt &C2) {
1632 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1633 if (!Shift || !Shift->isShift())
1634 return nullptr;
1635
1636 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1637 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1638 // code produced by the clang front-end, for bitfield access.
1639 // This seemingly simple opportunity to fold away a shift turns out to be
1640 // rather complicated. See PR17827 for details.
1641 unsigned ShiftOpcode = Shift->getOpcode();
1642 bool IsShl = ShiftOpcode == Instruction::Shl;
1643 const APInt *C3;
1644 if (match(Shift->getOperand(1), m_APInt(C3))) {
1645 APInt NewAndCst, NewCmpCst;
1646 bool AnyCmpCstBitsShiftedOut;
1647 if (ShiftOpcode == Instruction::Shl) {
1648 // For a left shift, we can fold if the comparison is not signed. We can
1649 // also fold a signed comparison if the mask value and comparison value
1650 // are not negative. These constraints may not be obvious, but we can
1651 // prove that they are correct using an SMT solver.
1652 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1653 return nullptr;
1654
1655 NewCmpCst = C1.lshr(*C3);
1656 NewAndCst = C2.lshr(*C3);
1657 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1658 } else if (ShiftOpcode == Instruction::LShr) {
1659 // For a logical right shift, we can fold if the comparison is not signed.
1660 // We can also fold a signed comparison if the shifted mask value and the
1661 // shifted comparison value are not negative. These constraints may not be
1662 // obvious, but we can prove that they are correct using an SMT solver.
1663 NewCmpCst = C1.shl(*C3);
1664 NewAndCst = C2.shl(*C3);
1665 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1666 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1667 return nullptr;
1668 } else {
1669 // For an arithmetic shift, check that both constants don't use (in a
1670 // signed sense) the top bits being shifted out.
1671 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1671, __PRETTY_FUNCTION__))
;
1672 NewCmpCst = C1.shl(*C3);
1673 NewAndCst = C2.shl(*C3);
1674 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1675 if (NewAndCst.ashr(*C3) != C2)
1676 return nullptr;
1677 }
1678
1679 if (AnyCmpCstBitsShiftedOut) {
1680 // If we shifted bits out, the fold is not going to work out. As a
1681 // special case, check to see if this means that the result is always
1682 // true or false now.
1683 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1684 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1685 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1686 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1687 } else {
1688 Value *NewAnd = Builder.CreateAnd(
1689 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1690 return new ICmpInst(Cmp.getPredicate(),
1691 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1692 }
1693 }
1694
1695 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1696 // preferable because it allows the C2 << Y expression to be hoisted out of a
1697 // loop if Y is invariant and X is not.
1698 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1699 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1700 // Compute C2 << Y.
1701 Value *NewShift =
1702 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1703 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1704
1705 // Compute X & (C2 << Y).
1706 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1707 return replaceOperand(Cmp, 0, NewAnd);
1708 }
1709
1710 return nullptr;
1711}
1712
1713/// Fold icmp (and X, C2), C1.
1714Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1715 BinaryOperator *And,
1716 const APInt &C1) {
1717 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1718
1719 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1720 // TODO: We canonicalize to the longer form for scalars because we have
1721 // better analysis/folds for icmp, and codegen may be better with icmp.
1722 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1723 match(And->getOperand(1), m_One()))
1724 return new TruncInst(And->getOperand(0), Cmp.getType());
1725
1726 const APInt *C2;
1727 Value *X;
1728 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1729 return nullptr;
1730
1731 // Don't perform the following transforms if the AND has multiple uses
1732 if (!And->hasOneUse())
1733 return nullptr;
1734
1735 if (Cmp.isEquality() && C1.isNullValue()) {
1736 // Restrict this fold to single-use 'and' (PR10267).
1737 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1738 if (C2->isSignMask()) {
1739 Constant *Zero = Constant::getNullValue(X->getType());
1740 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1741 return new ICmpInst(NewPred, X, Zero);
1742 }
1743
1744 // Restrict this fold only for single-use 'and' (PR10267).
1745 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1746 if ((~(*C2) + 1).isPowerOf2()) {
1747 Constant *NegBOC =
1748 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1749 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1750 return new ICmpInst(NewPred, X, NegBOC);
1751 }
1752 }
1753
1754 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1755 // the input width without changing the value produced, eliminate the cast:
1756 //
1757 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1758 //
1759 // We can do this transformation if the constants do not have their sign bits
1760 // set or if it is an equality comparison. Extending a relational comparison
1761 // when we're checking the sign bit would not work.
1762 Value *W;
1763 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1764 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1765 // TODO: Is this a good transform for vectors? Wider types may reduce
1766 // throughput. Should this transform be limited (even for scalars) by using
1767 // shouldChangeType()?
1768 if (!Cmp.getType()->isVectorTy()) {
1769 Type *WideType = W->getType();
1770 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1771 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1772 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1773 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1774 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1775 }
1776 }
1777
1778 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1779 return I;
1780
1781 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1782 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1783 //
1784 // iff pred isn't signed
1785 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1786 match(And->getOperand(1), m_One())) {
1787 Constant *One = cast<Constant>(And->getOperand(1));
1788 Value *Or = And->getOperand(0);
1789 Value *A, *B, *LShr;
1790 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1791 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1792 unsigned UsesRemoved = 0;
1793 if (And->hasOneUse())
1794 ++UsesRemoved;
1795 if (Or->hasOneUse())
1796 ++UsesRemoved;
1797 if (LShr->hasOneUse())
1798 ++UsesRemoved;
1799
1800 // Compute A & ((1 << B) | 1)
1801 Value *NewOr = nullptr;
1802 if (auto *C = dyn_cast<Constant>(B)) {
1803 if (UsesRemoved >= 1)
1804 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1805 } else {
1806 if (UsesRemoved >= 3)
1807 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1808 /*HasNUW=*/true),
1809 One, Or->getName());
1810 }
1811 if (NewOr) {
1812 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1813 return replaceOperand(Cmp, 0, NewAnd);
1814 }
1815 }
1816 }
1817
1818 return nullptr;
1819}
1820
1821/// Fold icmp (and X, Y), C.
1822Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1823 BinaryOperator *And,
1824 const APInt &C) {
1825 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1826 return I;
1827
1828 // TODO: These all require that Y is constant too, so refactor with the above.
1829
1830 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1831 Value *X = And->getOperand(0);
1832 Value *Y = And->getOperand(1);
1833 if (auto *LI = dyn_cast<LoadInst>(X))
1834 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1835 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1836 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1837 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1838 ConstantInt *C2 = cast<ConstantInt>(Y);
1839 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1840 return Res;
1841 }
1842
1843 if (!Cmp.isEquality())
1844 return nullptr;
1845
1846 // X & -C == -C -> X > u ~C
1847 // X & -C != -C -> X <= u ~C
1848 // iff C is a power of 2
1849 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1850 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1851 : CmpInst::ICMP_ULE;
1852 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1853 }
1854
1855 // (X & C2) == 0 -> (trunc X) >= 0
1856 // (X & C2) != 0 -> (trunc X) < 0
1857 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1858 const APInt *C2;
1859 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1860 int32_t ExactLogBase2 = C2->exactLogBase2();
1861 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1862 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1863 if (And->getType()->isVectorTy())
1864 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1865 Value *Trunc = Builder.CreateTrunc(X, NTy);
1866 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1867 : CmpInst::ICMP_SLT;
1868 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1869 }
1870 }
1871
1872 return nullptr;
1873}
1874
1875/// Fold icmp (or X, Y), C.
1876Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1877 const APInt &C) {
1878 ICmpInst::Predicate Pred = Cmp.getPredicate();
1879 if (C.isOneValue()) {
1880 // icmp slt signum(V) 1 --> icmp slt V, 1
1881 Value *V = nullptr;
1882 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1883 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1884 ConstantInt::get(V->getType(), 1));
1885 }
1886
1887 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1888 if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1889 // X | C == C --> X <=u C
1890 // X | C != C --> X >u C
1891 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1892 if ((C + 1).isPowerOf2()) {
1893 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1894 return new ICmpInst(Pred, OrOp0, OrOp1);
1895 }
1896 // More general: are all bits outside of a mask constant set or not set?
1897 // X | C == C --> (X & ~C) == 0
1898 // X | C != C --> (X & ~C) != 0
1899 if (Or->hasOneUse()) {
1900 Value *A = Builder.CreateAnd(OrOp0, ~C);
1901 return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1902 }
1903 }
1904
1905 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1906 return nullptr;
1907
1908 Value *P, *Q;
1909 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1910 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1911 // -> and (icmp eq P, null), (icmp eq Q, null).
1912 Value *CmpP =
1913 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1914 Value *CmpQ =
1915 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1916 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1917 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1918 }
1919
1920 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1921 // a shorter form that has more potential to be folded even further.
1922 Value *X1, *X2, *X3, *X4;
1923 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1924 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1925 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1926 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1927 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1928 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1929 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1930 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1931 }
1932
1933 return nullptr;
1934}
1935
1936/// Fold icmp (mul X, Y), C.
1937Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1938 BinaryOperator *Mul,
1939 const APInt &C) {
1940 const APInt *MulC;
1941 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1942 return nullptr;
1943
1944 // If this is a test of the sign bit and the multiply is sign-preserving with
1945 // a constant operand, use the multiply LHS operand instead.
1946 ICmpInst::Predicate Pred = Cmp.getPredicate();
1947 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1948 if (MulC->isNegative())
1949 Pred = ICmpInst::getSwappedPredicate(Pred);
1950 return new ICmpInst(Pred, Mul->getOperand(0),
1951 Constant::getNullValue(Mul->getType()));
1952 }
1953
1954 return nullptr;
1955}
1956
1957/// Fold icmp (shl 1, Y), C.
1958static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1959 const APInt &C) {
1960 Value *Y;
1961 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1962 return nullptr;
1963
1964 Type *ShiftType = Shl->getType();
1965 unsigned TypeBits = C.getBitWidth();
1966 bool CIsPowerOf2 = C.isPowerOf2();
1967 ICmpInst::Predicate Pred = Cmp.getPredicate();
1968 if (Cmp.isUnsigned()) {
1969 // (1 << Y) pred C -> Y pred Log2(C)
1970 if (!CIsPowerOf2) {
1971 // (1 << Y) < 30 -> Y <= 4
1972 // (1 << Y) <= 30 -> Y <= 4
1973 // (1 << Y) >= 30 -> Y > 4
1974 // (1 << Y) > 30 -> Y > 4
1975 if (Pred == ICmpInst::ICMP_ULT)
1976 Pred = ICmpInst::ICMP_ULE;
1977 else if (Pred == ICmpInst::ICMP_UGE)
1978 Pred = ICmpInst::ICMP_UGT;
1979 }
1980
1981 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1982 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1983 unsigned CLog2 = C.logBase2();
1984 if (CLog2 == TypeBits - 1) {
1985 if (Pred == ICmpInst::ICMP_UGE)
1986 Pred = ICmpInst::ICMP_EQ;
1987 else if (Pred == ICmpInst::ICMP_ULT)
1988 Pred = ICmpInst::ICMP_NE;
1989 }
1990 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1991 } else if (Cmp.isSigned()) {
1992 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1993 if (C.isAllOnesValue()) {
1994 // (1 << Y) <= -1 -> Y == 31
1995 if (Pred == ICmpInst::ICMP_SLE)
1996 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1997
1998 // (1 << Y) > -1 -> Y != 31
1999 if (Pred == ICmpInst::ICMP_SGT)
2000 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2001 } else if (!C) {
2002 // (1 << Y) < 0 -> Y == 31
2003 // (1 << Y) <= 0 -> Y == 31
2004 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2005 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2006
2007 // (1 << Y) >= 0 -> Y != 31
2008 // (1 << Y) > 0 -> Y != 31
2009 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2010 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2011 }
2012 } else if (Cmp.isEquality() && CIsPowerOf2) {
2013 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2014 }
2015
2016 return nullptr;
2017}
2018
2019/// Fold icmp (shl X, Y), C.
2020Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
2021 BinaryOperator *Shl,
2022 const APInt &C) {
2023 const APInt *ShiftVal;
2024 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2025 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2026
2027 const APInt *ShiftAmt;
2028 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2029 return foldICmpShlOne(Cmp, Shl, C);
2030
2031 // Check that the shift amount is in range. If not, don't perform undefined
2032 // shifts. When the shift is visited, it will be simplified.
2033 unsigned TypeBits = C.getBitWidth();
2034 if (ShiftAmt->uge(TypeBits))
2035 return nullptr;
2036
2037 ICmpInst::Predicate Pred = Cmp.getPredicate();
2038 Value *X = Shl->getOperand(0);
2039 Type *ShType = Shl->getType();
2040
2041 // NSW guarantees that we are only shifting out sign bits from the high bits,
2042 // so we can ASHR the compare constant without needing a mask and eliminate
2043 // the shift.
2044 if (Shl->hasNoSignedWrap()) {
2045 if (Pred == ICmpInst::ICMP_SGT) {
2046 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2047 APInt ShiftedC = C.ashr(*ShiftAmt);
2048 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2049 }
2050 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2051 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2052 APInt ShiftedC = C.ashr(*ShiftAmt);
2053 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2054 }
2055 if (Pred == ICmpInst::ICMP_SLT) {
2056 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2057 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2058 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2059 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2060 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2060, __PRETTY_FUNCTION__))
;
2061 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2062 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2063 }
2064 // If this is a signed comparison to 0 and the shift is sign preserving,
2065 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2066 // do that if we're sure to not continue on in this function.
2067 if (isSignTest(Pred, C))
2068 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2069 }
2070
2071 // NUW guarantees that we are only shifting out zero bits from the high bits,
2072 // so we can LSHR the compare constant without needing a mask and eliminate
2073 // the shift.
2074 if (Shl->hasNoUnsignedWrap()) {
2075 if (Pred == ICmpInst::ICMP_UGT) {
2076 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2077 APInt ShiftedC = C.lshr(*ShiftAmt);
2078 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2079 }
2080 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2081 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2082 APInt ShiftedC = C.lshr(*ShiftAmt);
2083 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2084 }
2085 if (Pred == ICmpInst::ICMP_ULT) {
2086 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2087 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2088 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2089 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2090 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2090, __PRETTY_FUNCTION__))
;
2091 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2092 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2093 }
2094 }
2095
2096 if (Cmp.isEquality() && Shl->hasOneUse()) {
2097 // Strength-reduce the shift into an 'and'.
2098 Constant *Mask = ConstantInt::get(
2099 ShType,
2100 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2101 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2102 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2103 return new ICmpInst(Pred, And, LShrC);
2104 }
2105
2106 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2107 bool TrueIfSigned = false;
2108 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2109 // (X << 31) <s 0 --> (X & 1) != 0
2110 Constant *Mask = ConstantInt::get(
2111 ShType,
2112 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2113 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2114 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2115 And, Constant::getNullValue(ShType));
2116 }
2117
2118 // Simplify 'shl' inequality test into 'and' equality test.
2119 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2120 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2121 if ((C + 1).isPowerOf2() &&
2122 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2123 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2124 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2125 : ICmpInst::ICMP_NE,
2126 And, Constant::getNullValue(ShType));
2127 }
2128 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2129 if (C.isPowerOf2() &&
2130 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2131 Value *And =
2132 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2133 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2134 : ICmpInst::ICMP_NE,
2135 And, Constant::getNullValue(ShType));
2136 }
2137 }
2138
2139 // Transform (icmp pred iM (shl iM %v, N), C)
2140 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2141 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2142 // This enables us to get rid of the shift in favor of a trunc that may be
2143 // free on the target. It has the additional benefit of comparing to a
2144 // smaller constant that may be more target-friendly.
2145 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2146 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2147 DL.isLegalInteger(TypeBits - Amt)) {
2148 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2149 if (ShType->isVectorTy())
2150 TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2151 Constant *NewC =
2152 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2153 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2154 }
2155
2156 return nullptr;
2157}
2158
2159/// Fold icmp ({al}shr X, Y), C.
2160Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2161 BinaryOperator *Shr,
2162 const APInt &C) {
2163 // An exact shr only shifts out zero bits, so:
2164 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2165 Value *X = Shr->getOperand(0);
2166 CmpInst::Predicate Pred = Cmp.getPredicate();
2167 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2168 C.isNullValue())
2169 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2170
2171 const APInt *ShiftVal;
2172 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2173 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2174
2175 const APInt *ShiftAmt;
2176 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2177 return nullptr;
2178
2179 // Check that the shift amount is in range. If not, don't perform undefined
2180 // shifts. When the shift is visited it will be simplified.
2181 unsigned TypeBits = C.getBitWidth();
2182 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2183 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2184 return nullptr;
2185
2186 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2187 bool IsExact = Shr->isExact();
2188 Type *ShrTy = Shr->getType();
2189 // TODO: If we could guarantee that InstSimplify would handle all of the
2190 // constant-value-based preconditions in the folds below, then we could assert
2191 // those conditions rather than checking them. This is difficult because of
2192 // undef/poison (PR34838).
2193 if (IsAShr) {
2194 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2195 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2196 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2197 APInt ShiftedC = C.shl(ShAmtVal);
2198 if (ShiftedC.ashr(ShAmtVal) == C)
2199 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2200 }
2201 if (Pred == CmpInst::ICMP_SGT) {
2202 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2203 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2204 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2205 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2206 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2207 }
2208 } else {
2209 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2210 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2211 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2212 APInt ShiftedC = C.shl(ShAmtVal);
2213 if (ShiftedC.lshr(ShAmtVal) == C)
2214 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2215 }
2216 if (Pred == CmpInst::ICMP_UGT) {
2217 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2218 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2219 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2220 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2221 }
2222 }
2223
2224 if (!Cmp.isEquality())
2225 return nullptr;
2226
2227 // Handle equality comparisons of shift-by-constant.
2228
2229 // If the comparison constant changes with the shift, the comparison cannot
2230 // succeed (bits of the comparison constant cannot match the shifted value).
2231 // This should be known by InstSimplify and already be folded to true/false.
2232 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2234, __PRETTY_FUNCTION__))
2233 (!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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2234, __PRETTY_FUNCTION__))
2234 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2234, __PRETTY_FUNCTION__))
;
2235
2236 // If the bits shifted out are known zero, compare the unshifted value:
2237 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2238 if (Shr->isExact())
2239 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2240
2241 if (Shr->hasOneUse()) {
2242 // Canonicalize the shift into an 'and':
2243 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2244 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2245 Constant *Mask = ConstantInt::get(ShrTy, Val);
2246 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2247 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2248 }
2249
2250 return nullptr;
2251}
2252
2253Instruction *InstCombiner::foldICmpSRemConstant(ICmpInst &Cmp,
2254 BinaryOperator *SRem,
2255 const APInt &C) {
2256 // Match an 'is positive' or 'is negative' comparison of remainder by a
2257 // constant power-of-2 value:
2258 // (X % pow2C) sgt/slt 0
2259 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2260 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2261 return nullptr;
2262
2263 // TODO: The one-use check is standard because we do not typically want to
2264 // create longer instruction sequences, but this might be a special-case
2265 // because srem is not good for analysis or codegen.
2266 if (!SRem->hasOneUse())
2267 return nullptr;
2268
2269 const APInt *DivisorC;
2270 if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2271 return nullptr;
2272
2273 // Mask off the sign bit and the modulo bits (low-bits).
2274 Type *Ty = SRem->getType();
2275 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2276 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2277 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2278
2279 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2280 // bit is set. Example:
2281 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2282 if (Pred == ICmpInst::ICMP_SGT)
2283 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2284
2285 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2286 // bit is set. Example:
2287 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2288 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2289}
2290
2291/// Fold icmp (udiv X, Y), C.
2292Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2293 BinaryOperator *UDiv,
2294 const APInt &C) {
2295 const APInt *C2;
2296 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2297 return nullptr;
2298
2299 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2299, __PRETTY_FUNCTION__))
;
2300
2301 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2302 Value *Y = UDiv->getOperand(1);
2303 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2304 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2305, __PRETTY_FUNCTION__))
2305 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2305, __PRETTY_FUNCTION__))
;
2306 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2307 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2308 }
2309
2310 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2311 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2312 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2312, __PRETTY_FUNCTION__))
;
2313 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2314 ConstantInt::get(Y->getType(), C2->udiv(C)));
2315 }
2316
2317 return nullptr;
2318}
2319
2320/// Fold icmp ({su}div X, Y), C.
2321Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2322 BinaryOperator *Div,
2323 const APInt &C) {
2324 // Fold: icmp pred ([us]div X, C2), C -> range test
2325 // Fold this div into the comparison, producing a range check.
2326 // Determine, based on the divide type, what the range is being
2327 // checked. If there is an overflow on the low or high side, remember
2328 // it, otherwise compute the range [low, hi) bounding the new value.
2329 // See: InsertRangeTest above for the kinds of replacements possible.
2330 const APInt *C2;
2331 if (!match(Div->getOperand(1), m_APInt(C2)))
2332 return nullptr;
2333
2334 // FIXME: If the operand types don't match the type of the divide
2335 // then don't attempt this transform. The code below doesn't have the
2336 // logic to deal with a signed divide and an unsigned compare (and
2337 // vice versa). This is because (x /s C2) <s C produces different
2338 // results than (x /s C2) <u C or (x /u C2) <s C or even
2339 // (x /u C2) <u C. Simply casting the operands and result won't
2340 // work. :( The if statement below tests that condition and bails
2341 // if it finds it.
2342 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2343 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2344 return nullptr;
2345
2346 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2347 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2348 // division-by-constant cases should be present, we can not assert that they
2349 // have happened before we reach this icmp instruction.
2350 if (C2->isNullValue() || C2->isOneValue() ||
2351 (DivIsSigned && C2->isAllOnesValue()))
2352 return nullptr;
2353
2354 // Compute Prod = C * C2. We are essentially solving an equation of
2355 // form X / C2 = C. We solve for X by multiplying C2 and C.
2356 // By solving for X, we can turn this into a range check instead of computing
2357 // a divide.
2358 APInt Prod = C * *C2;
2359
2360 // Determine if the product overflows by seeing if the product is not equal to
2361 // the divide. Make sure we do the same kind of divide as in the LHS
2362 // instruction that we're folding.
2363 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2364
2365 ICmpInst::Predicate Pred = Cmp.getPredicate();
2366
2367 // If the division is known to be exact, then there is no remainder from the
2368 // divide, so the covered range size is unit, otherwise it is the divisor.
2369 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2370
2371 // Figure out the interval that is being checked. For example, a comparison
2372 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2373 // Compute this interval based on the constants involved and the signedness of
2374 // the compare/divide. This computes a half-open interval, keeping track of
2375 // whether either value in the interval overflows. After analysis each
2376 // overflow variable is set to 0 if it's corresponding bound variable is valid
2377 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2378 int LoOverflow = 0, HiOverflow = 0;
2379 APInt LoBound, HiBound;
2380
2381 if (!DivIsSigned) { // udiv
2382 // e.g. X/5 op 3 --> [15, 20)
2383 LoBound = Prod;
2384 HiOverflow = LoOverflow = ProdOV;
2385 if (!HiOverflow) {
2386 // If this is not an exact divide, then many values in the range collapse
2387 // to the same result value.
2388 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2389 }
2390 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2391 if (C.isNullValue()) { // (X / pos) op 0
2392 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2393 LoBound = -(RangeSize - 1);
2394 HiBound = RangeSize;
2395 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2396 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2397 HiOverflow = LoOverflow = ProdOV;
2398 if (!HiOverflow)
2399 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2400 } else { // (X / pos) op neg
2401 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2402 HiBound = Prod + 1;
2403 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2404 if (!LoOverflow) {
2405 APInt DivNeg = -RangeSize;
2406 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2407 }
2408 }
2409 } else if (C2->isNegative()) { // Divisor is < 0.
2410 if (Div->isExact())
2411 RangeSize.negate();
2412 if (C.isNullValue()) { // (X / neg) op 0
2413 // e.g. X/-5 op 0 --> [-4, 5)
2414 LoBound = RangeSize + 1;
2415 HiBound = -RangeSize;
2416 if (HiBound == *C2) { // -INTMIN = INTMIN
2417 HiOverflow = 1; // [INTMIN+1, overflow)
2418 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2419 }
2420 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2421 // e.g. X/-5 op 3 --> [-19, -14)
2422 HiBound = Prod + 1;
2423 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2424 if (!LoOverflow)
2425 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2426 } else { // (X / neg) op neg
2427 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2428 LoOverflow = HiOverflow = ProdOV;
2429 if (!HiOverflow)
2430 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2431 }
2432
2433 // Dividing by a negative swaps the condition. LT <-> GT
2434 Pred = ICmpInst::getSwappedPredicate(Pred);
2435 }
2436
2437 Value *X = Div->getOperand(0);
2438 switch (Pred) {
2439 default: llvm_unreachable("Unhandled icmp opcode!")::llvm::llvm_unreachable_internal("Unhandled icmp opcode!", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2439)
;
2440 case ICmpInst::ICMP_EQ:
2441 if (LoOverflow && HiOverflow)
2442 return replaceInstUsesWith(Cmp, Builder.getFalse());
2443 if (HiOverflow)
2444 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2445 ICmpInst::ICMP_UGE, X,
2446 ConstantInt::get(Div->getType(), LoBound));
2447 if (LoOverflow)
2448 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2449 ICmpInst::ICMP_ULT, X,
2450 ConstantInt::get(Div->getType(), HiBound));
2451 return replaceInstUsesWith(
2452 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2453 case ICmpInst::ICMP_NE:
2454 if (LoOverflow && HiOverflow)
2455 return replaceInstUsesWith(Cmp, Builder.getTrue());
2456 if (HiOverflow)
2457 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2458 ICmpInst::ICMP_ULT, X,
2459 ConstantInt::get(Div->getType(), LoBound));
2460 if (LoOverflow)
2461 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2462 ICmpInst::ICMP_UGE, X,
2463 ConstantInt::get(Div->getType(), HiBound));
2464 return replaceInstUsesWith(Cmp,
2465 insertRangeTest(X, LoBound, HiBound,
2466 DivIsSigned, false));
2467 case ICmpInst::ICMP_ULT:
2468 case ICmpInst::ICMP_SLT:
2469 if (LoOverflow == +1) // Low bound is greater than input range.
2470 return replaceInstUsesWith(Cmp, Builder.getTrue());
2471 if (LoOverflow == -1) // Low bound is less than input range.
2472 return replaceInstUsesWith(Cmp, Builder.getFalse());
2473 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2474 case ICmpInst::ICMP_UGT:
2475 case ICmpInst::ICMP_SGT:
2476 if (HiOverflow == +1) // High bound greater than input range.
2477 return replaceInstUsesWith(Cmp, Builder.getFalse());
2478 if (HiOverflow == -1) // High bound less than input range.
2479 return replaceInstUsesWith(Cmp, Builder.getTrue());
2480 if (Pred == ICmpInst::ICMP_UGT)
2481 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2482 ConstantInt::get(Div->getType(), HiBound));
2483 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2484 ConstantInt::get(Div->getType(), HiBound));
2485 }
2486
2487 return nullptr;
2488}
2489
2490/// Fold icmp (sub X, Y), C.
2491Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2492 BinaryOperator *Sub,
2493 const APInt &C) {
2494 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2495 ICmpInst::Predicate Pred = Cmp.getPredicate();
2496 const APInt *C2;
2497 APInt SubResult;
2498
2499 // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2500 if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2501 return new ICmpInst(Cmp.getPredicate(), Y,
2502 ConstantInt::get(Y->getType(), 0));
2503
2504 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2505 if (match(X, m_APInt(C2)) &&
2506 ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2507 (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2508 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2509 return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2510 ConstantInt::get(Y->getType(), SubResult));
2511
2512 // The following transforms are only worth it if the only user of the subtract
2513 // is the icmp.
2514 if (!Sub->hasOneUse())
2515 return nullptr;
2516
2517 if (Sub->hasNoSignedWrap()) {
2518 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2519 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2520 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2521
2522 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2523 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2524 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2525
2526 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2527 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2528 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2529
2530 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2531 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2532 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2533 }
2534
2535 if (!match(X, m_APInt(C2)))
2536 return nullptr;
2537
2538 // C2 - Y <u C -> (Y | (C - 1)) == C2
2539 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2540 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2541 (*C2 & (C - 1)) == (C - 1))
2542 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2543
2544 // C2 - Y >u C -> (Y | C) != C2
2545 // iff C2 & C == C and C + 1 is a power of 2
2546 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2547 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2548
2549 return nullptr;
2550}
2551
2552/// Fold icmp (add X, Y), C.
2553Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2554 BinaryOperator *Add,
2555 const APInt &C) {
2556 Value *Y = Add->getOperand(1);
2557 const APInt *C2;
2558 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2559 return nullptr;
2560
2561 // Fold icmp pred (add X, C2), C.
2562 Value *X = Add->getOperand(0);
2563 Type *Ty = Add->getType();
2564 CmpInst::Predicate Pred = Cmp.getPredicate();
2565
2566 // If the add does not wrap, we can always adjust the compare by subtracting
2567 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2568 // are canonicalized to SGT/SLT/UGT/ULT.
2569 if ((Add->hasNoSignedWrap() &&
2570 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2571 (Add->hasNoUnsignedWrap() &&
2572 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2573 bool Overflow;
2574 APInt NewC =
2575 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2576 // If there is overflow, the result must be true or false.
2577 // TODO: Can we assert there is no overflow because InstSimplify always
2578 // handles those cases?
2579 if (!Overflow)
2580 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2581 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2582 }
2583
2584 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2585 const APInt &Upper = CR.getUpper();
2586 const APInt &Lower = CR.getLower();
2587 if (Cmp.isSigned()) {
2588 if (Lower.isSignMask())
2589 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2590 if (Upper.isSignMask())
2591 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2592 } else {
2593 if (Lower.isMinValue())
2594 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2595 if (Upper.isMinValue())
2596 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2597 }
2598
2599 if (!Add->hasOneUse())
2600 return nullptr;
2601
2602 // X+C <u C2 -> (X & -C2) == C
2603 // iff C & (C2-1) == 0
2604 // C2 is a power of 2
2605 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2606 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2607 ConstantExpr::getNeg(cast<Constant>(Y)));
2608
2609 // X+C >u C2 -> (X & ~C2) != C
2610 // iff C & C2 == 0
2611 // C2+1 is a power of 2
2612 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2613 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2614 ConstantExpr::getNeg(cast<Constant>(Y)));
2615
2616 return nullptr;
2617}
2618
2619bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2620 Value *&RHS, ConstantInt *&Less,
2621 ConstantInt *&Equal,
2622 ConstantInt *&Greater) {
2623 // TODO: Generalize this to work with other comparison idioms or ensure
2624 // they get canonicalized into this form.
2625
2626 // select i1 (a == b),
2627 // i32 Equal,
2628 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2629 // where Equal, Less and Greater are placeholders for any three constants.
2630 ICmpInst::Predicate PredA;
2631 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2632 !ICmpInst::isEquality(PredA))
2633 return false;
2634 Value *EqualVal = SI->getTrueValue();
2635 Value *UnequalVal = SI->getFalseValue();
2636 // We still can get non-canonical predicate here, so canonicalize.
2637 if (PredA == ICmpInst::ICMP_NE)
2638 std::swap(EqualVal, UnequalVal);
2639 if (!match(EqualVal, m_ConstantInt(Equal)))
2640 return false;
2641 ICmpInst::Predicate PredB;
2642 Value *LHS2, *RHS2;
2643 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2644 m_ConstantInt(Less), m_ConstantInt(Greater))))
2645 return false;
2646 // We can get predicate mismatch here, so canonicalize if possible:
2647 // First, ensure that 'LHS' match.
2648 if (LHS2 != LHS) {
2649 // x sgt y <--> y slt x
2650 std::swap(LHS2, RHS2);
2651 PredB = ICmpInst::getSwappedPredicate(PredB);
2652 }
2653 if (LHS2 != LHS)
2654 return false;
2655 // We also need to canonicalize 'RHS'.
2656 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2657 // x sgt C-1 <--> x sge C <--> not(x slt C)
2658 auto FlippedStrictness =
2659 getFlippedStrictnessPredicateAndConstant(PredB, cast<Constant>(RHS2));
2660 if (!FlippedStrictness)
2661 return false;
2662 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2662, __PRETTY_FUNCTION__))
;
2663 RHS2 = FlippedStrictness->second;
2664 // And kind-of perform the result swap.
2665 std::swap(Less, Greater);
2666 PredB = ICmpInst::ICMP_SLT;
2667 }
2668 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2669}
2670
2671Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2672 SelectInst *Select,
2673 ConstantInt *C) {
2674
2675 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2675, __PRETTY_FUNCTION__))
;
2676 // If we're testing a constant value against the result of a three way
2677 // comparison, the result can be expressed directly in terms of the
2678 // original values being compared. Note: We could possibly be more
2679 // aggressive here and remove the hasOneUse test. The original select is
2680 // really likely to simplify or sink when we remove a test of the result.
2681 Value *OrigLHS, *OrigRHS;
2682 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2683 if (Cmp.hasOneUse() &&
2684 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2685 C3GreaterThan)) {
2686 assert(C1LessThan && C2Equal && C3GreaterThan)((C1LessThan && C2Equal && C3GreaterThan) ? static_cast
<void> (0) : __assert_fail ("C1LessThan && C2Equal && C3GreaterThan"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2686, __PRETTY_FUNCTION__))
;
2687
2688 bool TrueWhenLessThan =
2689 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2690 ->isAllOnesValue();
2691 bool TrueWhenEqual =
2692 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2693 ->isAllOnesValue();
2694 bool TrueWhenGreaterThan =
2695 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2696 ->isAllOnesValue();
2697
2698 // This generates the new instruction that will replace the original Cmp
2699 // Instruction. Instead of enumerating the various combinations when
2700 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2701 // false, we rely on chaining of ORs and future passes of InstCombine to
2702 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2703
2704 // When none of the three constants satisfy the predicate for the RHS (C),
2705 // the entire original Cmp can be simplified to a false.
2706 Value *Cond = Builder.getFalse();
2707 if (TrueWhenLessThan)
2708 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2709 OrigLHS, OrigRHS));
2710 if (TrueWhenEqual)
2711 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2712 OrigLHS, OrigRHS));
2713 if (TrueWhenGreaterThan)
2714 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2715 OrigLHS, OrigRHS));
2716
2717 return replaceInstUsesWith(Cmp, Cond);
2718 }
2719 return nullptr;
2720}
2721
2722static Instruction *foldICmpBitCast(ICmpInst &Cmp,
2723 InstCombiner::BuilderTy &Builder) {
2724 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2725 if (!Bitcast)
2726 return nullptr;
2727
2728 ICmpInst::Predicate Pred = Cmp.getPredicate();
2729 Value *Op1 = Cmp.getOperand(1);
2730 Value *BCSrcOp = Bitcast->getOperand(0);
2731
2732 // Make sure the bitcast doesn't change the number of vector elements.
2733 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2734 Bitcast->getDestTy()->getScalarSizeInBits()) {
2735 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2736 Value *X;
2737 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2738 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2739 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2740 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2741 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2742 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2743 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2744 match(Op1, m_Zero()))
2745 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2746
2747 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2748 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2749 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2750
2751 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2752 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2753 return new ICmpInst(Pred, X,
2754 ConstantInt::getAllOnesValue(X->getType()));
2755 }
2756
2757 // Zero-equality checks are preserved through unsigned floating-point casts:
2758 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2759 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2760 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2761 if (Cmp.isEquality() && match(Op1, m_Zero()))
2762 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2763 }
2764
2765 // Test to see if the operands of the icmp are casted versions of other
2766 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2767 if (Bitcast->getType()->isPointerTy() &&
2768 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2769 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2770 // so eliminate it as well.
2771 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2772 Op1 = BC2->getOperand(0);
2773
2774 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2775 return new ICmpInst(Pred, BCSrcOp, Op1);
2776 }
2777
2778 // Folding: icmp <pred> iN X, C
2779 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2780 // and C is a splat of a K-bit pattern
2781 // and SC is a constant vector = <C', C', C', ..., C'>
2782 // Into:
2783 // %E = extractelement <M x iK> %vec, i32 C'
2784 // icmp <pred> iK %E, trunc(C)
2785 const APInt *C;
2786 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2787 !Bitcast->getType()->isIntegerTy() ||
2788 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2789 return nullptr;
2790
2791 Value *Vec;
2792 Constant *Mask;
2793 if (match(BCSrcOp,
2794 m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2795 // Check whether every element of Mask is the same constant
2796 if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2797 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2798 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2799 if (C->isSplat(EltTy->getBitWidth())) {
2800 // Fold the icmp based on the value of C
2801 // If C is M copies of an iK sized bit pattern,
2802 // then:
2803 // => %E = extractelement <N x iK> %vec, i32 Elem
2804 // icmp <pred> iK %SplatVal, <pattern>
2805 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2806 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2807 return new ICmpInst(Pred, Extract, NewC);
2808 }
2809 }
2810 }
2811 return nullptr;
2812}
2813
2814/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2815/// where X is some kind of instruction.
2816Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2817 const APInt *C;
2818 if (!match(Cmp.getOperand(1), m_APInt(C)))
2819 return nullptr;
2820
2821 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2822 switch (BO->getOpcode()) {
2823 case Instruction::Xor:
2824 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2825 return I;
2826 break;
2827 case Instruction::And:
2828 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2829 return I;
2830 break;
2831 case Instruction::Or:
2832 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2833 return I;
2834 break;
2835 case Instruction::Mul:
2836 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2837 return I;
2838 break;
2839 case Instruction::Shl:
2840 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2841 return I;
2842 break;
2843 case Instruction::LShr:
2844 case Instruction::AShr:
2845 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2846 return I;
2847 break;
2848 case Instruction::SRem:
2849 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2850 return I;
2851 break;
2852 case Instruction::UDiv:
2853 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2854 return I;
2855 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2856 case Instruction::SDiv:
2857 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2858 return I;
2859 break;
2860 case Instruction::Sub:
2861 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2862 return I;
2863 break;
2864 case Instruction::Add:
2865 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2866 return I;
2867 break;
2868 default:
2869 break;
2870 }
2871 // TODO: These folds could be refactored to be part of the above calls.
2872 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2873 return I;
2874 }
2875
2876 // Match against CmpInst LHS being instructions other than binary operators.
2877
2878 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2879 // For now, we only support constant integers while folding the
2880 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2881 // similar to the cases handled by binary ops above.
2882 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2883 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2884 return I;
2885 }
2886
2887 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2888 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2889 return I;
2890 }
2891
2892 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2893 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2894 return I;
2895
2896 return nullptr;
2897}
2898
2899/// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2900/// icmp eq/ne BO, C.
2901Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2902 BinaryOperator *BO,
2903 const APInt &C) {
2904 // TODO: Some of these folds could work with arbitrary constants, but this
2905 // function is limited to scalar and vector splat constants.
2906 if (!Cmp.isEquality())
2907 return nullptr;
2908
2909 ICmpInst::Predicate Pred = Cmp.getPredicate();
2910 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2911 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2912 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2913
2914 switch (BO->getOpcode()) {
2915 case Instruction::SRem:
2916 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2917 if (C.isNullValue() && BO->hasOneUse()) {
2918 const APInt *BOC;
2919 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2920 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2921 return new ICmpInst(Pred, NewRem,
2922 Constant::getNullValue(BO->getType()));
2923 }
2924 }
2925 break;
2926 case Instruction::Add: {
2927 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2928 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2929 if (BO->hasOneUse())
2930 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
2931 } else if (C.isNullValue()) {
2932 // Replace ((add A, B) != 0) with (A != -B) if A or B is
2933 // efficiently invertible, or if the add has just this one use.
2934 if (Value *NegVal = dyn_castNegVal(BOp1))
2935 return new ICmpInst(Pred, BOp0, NegVal);
2936 if (Value *NegVal = dyn_castNegVal(BOp0))
2937 return new ICmpInst(Pred, NegVal, BOp1);
2938 if (BO->hasOneUse()) {
2939 Value *Neg = Builder.CreateNeg(BOp1);
2940 Neg->takeName(BO);
2941 return new ICmpInst(Pred, BOp0, Neg);
2942 }
2943 }
2944 break;
2945 }
2946 case Instruction::Xor:
2947 if (BO->hasOneUse()) {
2948 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2949 // For the xor case, we can xor two constants together, eliminating
2950 // the explicit xor.
2951 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2952 } else if (C.isNullValue()) {
2953 // Replace ((xor A, B) != 0) with (A != B)
2954 return new ICmpInst(Pred, BOp0, BOp1);
2955 }
2956 }
2957 break;
2958 case Instruction::Sub:
2959 if (BO->hasOneUse()) {
2960 // Only check for constant LHS here, as constant RHS will be canonicalized
2961 // to add and use the fold above.
2962 if (Constant *BOC = dyn_cast<Constant>(BOp0)) {
2963 // Replace ((sub BOC, B) != C) with (B != BOC-C).
2964 return new ICmpInst(Pred, BOp1, ConstantExpr::getSub(BOC, RHS));
2965 } else if (C.isNullValue()) {
2966 // Replace ((sub A, B) != 0) with (A != B).
2967 return new ICmpInst(Pred, BOp0, BOp1);
2968 }
2969 }
2970 break;
2971 case Instruction::Or: {
2972 const APInt *BOC;
2973 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2974 // Comparing if all bits outside of a constant mask are set?
2975 // Replace (X | C) == -1 with (X & ~C) == ~C.
2976 // This removes the -1 constant.
2977 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2978 Value *And = Builder.CreateAnd(BOp0, NotBOC);
2979 return new ICmpInst(Pred, And, NotBOC);
2980 }
2981 break;
2982 }
2983 case Instruction::And: {
2984 const APInt *BOC;
2985 if (match(BOp1, m_APInt(BOC))) {
2986 // If we have ((X & C) == C), turn it into ((X & C) != 0).
2987 if (C == *BOC && C.isPowerOf2())
2988 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2989 BO, Constant::getNullValue(RHS->getType()));
2990 }
2991 break;
2992 }
2993 case Instruction::Mul:
2994 if (C.isNullValue() && BO->hasNoSignedWrap()) {
2995 const APInt *BOC;
2996 if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2997 // The trivial case (mul X, 0) is handled by InstSimplify.
2998 // General case : (mul X, C) != 0 iff X != 0
2999 // (mul X, C) == 0 iff X == 0
3000 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
3001 }
3002 }
3003 break;
3004 case Instruction::UDiv:
3005 if (C.isNullValue()) {
3006 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3007 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3008 return new ICmpInst(NewPred, BOp1, BOp0);
3009 }
3010 break;
3011 default:
3012 break;
3013 }
3014 return nullptr;
3015}
3016
3017/// Fold an equality icmp with LLVM intrinsic and constant operand.
3018Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
3019 IntrinsicInst *II,
3020 const APInt &C) {
3021 Type *Ty = II->getType();
3022 unsigned BitWidth = C.getBitWidth();
3023 switch (II->getIntrinsicID()) {
3024 case Intrinsic::bswap:
3025 // bswap(A) == C -> A == bswap(C)
3026 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3027 ConstantInt::get(Ty, C.byteSwap()));
3028
3029 case Intrinsic::ctlz:
3030 case Intrinsic::cttz: {
3031 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3032 if (C == BitWidth)
3033 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3034 ConstantInt::getNullValue(Ty));
3035
3036 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3037 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3038 // Limit to one use to ensure we don't increase instruction count.
3039 unsigned Num = C.getLimitedValue(BitWidth);
3040 if (Num != BitWidth && II->hasOneUse()) {
3041 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3042 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3043 : APInt::getHighBitsSet(BitWidth, Num + 1);
3044 APInt Mask2 = IsTrailing
3045 ? APInt::getOneBitSet(BitWidth, Num)
3046 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3047 return new ICmpInst(Cmp.getPredicate(),
3048 Builder.CreateAnd(II->getArgOperand(0), Mask1),
3049 ConstantInt::get(Ty, Mask2));
3050 }
3051 break;
3052 }
3053
3054 case Intrinsic::ctpop: {
3055 // popcount(A) == 0 -> A == 0 and likewise for !=
3056 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3057 bool IsZero = C.isNullValue();
3058 if (IsZero || C == BitWidth)
3059 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3060 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty));
3061
3062 break;
3063 }
3064
3065 case Intrinsic::uadd_sat: {
3066 // uadd.sat(a, b) == 0 -> (a | b) == 0
3067 if (C.isNullValue()) {
3068 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3069 return new ICmpInst(Cmp.getPredicate(), Or, Constant::getNullValue(Ty));
3070 }
3071 break;
3072 }
3073
3074 case Intrinsic::usub_sat: {
3075 // usub.sat(a, b) == 0 -> a <= b
3076 if (C.isNullValue()) {
3077 ICmpInst::Predicate NewPred = Cmp.getPredicate() == ICmpInst::ICMP_EQ
3078 ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3079 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3080 }
3081 break;
3082 }
3083 default:
3084 break;
3085 }
3086
3087 return nullptr;
3088}
3089
3090/// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3091Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3092 IntrinsicInst *II,
3093 const APInt &C) {
3094 if (Cmp.isEquality())
3095 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3096
3097 Type *Ty = II->getType();
3098 unsigned BitWidth = C.getBitWidth();
3099 switch (II->getIntrinsicID()) {
3100 case Intrinsic::ctlz: {
3101 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3102 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3103 unsigned Num = C.getLimitedValue();
3104 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3105 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3106 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3107 }
3108
3109 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3110 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3111 C.uge(1) && C.ule(BitWidth)) {
3112 unsigned Num = C.getLimitedValue();
3113 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3114 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3115 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3116 }
3117 break;
3118 }
3119 case Intrinsic::cttz: {
3120 // Limit to one use to ensure we don't increase instruction count.
3121 if (!II->hasOneUse())
3122 return nullptr;
3123
3124 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3125 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3126 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3127 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3128 Builder.CreateAnd(II->getArgOperand(0), Mask),
3129 ConstantInt::getNullValue(Ty));
3130 }
3131
3132 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3133 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3134 C.uge(1) && C.ule(BitWidth)) {
3135 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3136 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3137 Builder.CreateAnd(II->getArgOperand(0), Mask),
3138 ConstantInt::getNullValue(Ty));
3139 }
3140 break;
3141 }
3142 default:
3143 break;
3144 }
3145
3146 return nullptr;
3147}
3148
3149/// Handle icmp with constant (but not simple integer constant) RHS.
3150Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3151 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3152 Constant *RHSC = dyn_cast<Constant>(Op1);
3153 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3154 if (!RHSC || !LHSI)
3155 return nullptr;
3156
3157 switch (LHSI->getOpcode()) {
3158 case Instruction::GetElementPtr:
3159 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3160 if (RHSC->isNullValue() &&
3161 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3162 return new ICmpInst(
3163 I.getPredicate(), LHSI->getOperand(0),
3164 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3165 break;
3166 case Instruction::PHI:
3167 // Only fold icmp into the PHI if the phi and icmp are in the same
3168 // block. If in the same block, we're encouraging jump threading. If
3169 // not, we are just pessimizing the code by making an i1 phi.
3170 if (LHSI->getParent() == I.getParent())
3171 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3172 return NV;
3173 break;
3174 case Instruction::Select: {
3175 // If either operand of the select is a constant, we can fold the
3176 // comparison into the select arms, which will cause one to be
3177 // constant folded and the select turned into a bitwise or.
3178 Value *Op1 = nullptr, *Op2 = nullptr;
3179 ConstantInt *CI = nullptr;
3180 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3181 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3182 CI = dyn_cast<ConstantInt>(Op1);
3183 }
3184 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3185 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3186 CI = dyn_cast<ConstantInt>(Op2);
3187 }
3188
3189 // We only want to perform this transformation if it will not lead to
3190 // additional code. This is true if either both sides of the select
3191 // fold to a constant (in which case the icmp is replaced with a select
3192 // which will usually simplify) or this is the only user of the
3193 // select (in which case we are trading a select+icmp for a simpler
3194 // select+icmp) or all uses of the select can be replaced based on
3195 // dominance information ("Global cases").
3196 bool Transform = false;
3197 if (Op1 && Op2)
3198 Transform = true;
3199 else if (Op1 || Op2) {
3200 // Local case
3201 if (LHSI->hasOneUse())
3202 Transform = true;
3203 // Global cases
3204 else if (CI && !CI->isZero())
3205 // When Op1 is constant try replacing select with second operand.
3206 // Otherwise Op2 is constant and try replacing select with first
3207 // operand.
3208 Transform =
3209 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3210 }
3211 if (Transform) {
3212 if (!Op1)
3213 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3214 I.getName());
3215 if (!Op2)
3216 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3217 I.getName());
3218 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3219 }
3220 break;
3221 }
3222 case Instruction::IntToPtr:
3223 // icmp pred inttoptr(X), null -> icmp pred X, 0
3224 if (RHSC->isNullValue() &&
3225 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3226 return new ICmpInst(
3227 I.getPredicate(), LHSI->getOperand(0),
3228 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3229 break;
3230
3231 case Instruction::Load:
3232 // Try to optimize things like "A[i] > 4" to index computations.
3233 if (GetElementPtrInst *GEP =
3234 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3235 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3236 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3237 !cast<LoadInst>(LHSI)->isVolatile())
3238 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3239 return Res;
3240 }
3241 break;
3242 }
3243
3244 return nullptr;
3245}
3246
3247/// Some comparisons can be simplified.
3248/// In this case, we are looking for comparisons that look like
3249/// a check for a lossy truncation.
3250/// Folds:
3251/// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3252/// Where Mask is some pattern that produces all-ones in low bits:
3253/// (-1 >> y)
3254/// ((-1 << y) >> y) <- non-canonical, has extra uses
3255/// ~(-1 << y)
3256/// ((1 << y) + (-1)) <- non-canonical, has extra uses
3257/// The Mask can be a constant, too.
3258/// For some predicates, the operands are commutative.
3259/// For others, x can only be on a specific side.
3260static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3261 InstCombiner::BuilderTy &Builder) {
3262 ICmpInst::Predicate SrcPred;
3263 Value *X, *M, *Y;
3264 auto m_VariableMask = m_CombineOr(
3265 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3266 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3267 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3268 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3269 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3270 if (!match(&I, m_c_ICmp(SrcPred,
3271 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3272 m_Deferred(X))))
3273 return nullptr;
3274
3275 ICmpInst::Predicate DstPred;
3276 switch (SrcPred) {
3277 case ICmpInst::Predicate::ICMP_EQ:
3278 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3279 DstPred = ICmpInst::Predicate::ICMP_ULE;
3280 break;
3281 case ICmpInst::Predicate::ICMP_NE:
3282 // x & (-1 >> y) != x -> x u> (-1 >> y)
3283 DstPred = ICmpInst::Predicate::ICMP_UGT;
3284 break;
3285 case ICmpInst::Predicate::ICMP_ULT:
3286 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3287 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3288 DstPred = ICmpInst::Predicate::ICMP_UGT;
3289 break;
3290 case ICmpInst::Predicate::ICMP_UGE:
3291 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3292 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3293 DstPred = ICmpInst::Predicate::ICMP_ULE;
3294 break;
3295 case ICmpInst::Predicate::ICMP_SLT:
3296 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3297 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3298 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3299 return nullptr;
3300 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3301 return nullptr;
3302 DstPred = ICmpInst::Predicate::ICMP_SGT;
3303 break;
3304 case ICmpInst::Predicate::ICMP_SGE:
3305 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3306 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3307 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3308 return nullptr;
3309 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3310 return nullptr;
3311 DstPred = ICmpInst::Predicate::ICMP_SLE;
3312 break;
3313 case ICmpInst::Predicate::ICMP_SGT:
3314 case ICmpInst::Predicate::ICMP_SLE:
3315 return nullptr;
3316 case ICmpInst::Predicate::ICMP_UGT:
3317 case ICmpInst::Predicate::ICMP_ULE:
3318 llvm_unreachable("Instsimplify took care of commut. variant")::llvm::llvm_unreachable_internal("Instsimplify took care of commut. variant"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3318)
;
3319 break;
3320 default:
3321 llvm_unreachable("All possible folds are handled.")::llvm::llvm_unreachable_internal("All possible folds are handled."
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3321)
;
3322 }
3323
3324 // The mask value may be a vector constant that has undefined elements. But it
3325 // may not be safe to propagate those undefs into the new compare, so replace
3326 // those elements by copying an existing, defined, and safe scalar constant.
3327 Type *OpTy = M->getType();
3328 auto *VecC = dyn_cast<Constant>(M);
3329 if (OpTy->isVectorTy() && VecC && VecC->containsUndefElement()) {
3330 Constant *SafeReplacementConstant = nullptr;
3331 for (unsigned i = 0, e = OpTy->getVectorNumElements(); i != e; ++i) {
3332 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3333 SafeReplacementConstant = VecC->getAggregateElement(i);
3334 break;
3335 }
3336 }
3337 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3337, __PRETTY_FUNCTION__))
;
3338 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3339 }
3340
3341 return Builder.CreateICmp(DstPred, X, M);
3342}
3343
3344/// Some comparisons can be simplified.
3345/// In this case, we are looking for comparisons that look like
3346/// a check for a lossy signed truncation.
3347/// Folds: (MaskedBits is a constant.)
3348/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3349/// Into:
3350/// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3351/// Where KeptBits = bitwidth(%x) - MaskedBits
3352static Value *
3353foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3354 InstCombiner::BuilderTy &Builder) {
3355 ICmpInst::Predicate SrcPred;
3356 Value *X;
3357 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3358 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3359 if (!match(&I, m_c_ICmp(SrcPred,
3360 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3361 m_APInt(C1))),
3362 m_Deferred(X))))
3363 return nullptr;
3364
3365 // Potential handling of non-splats: for each element:
3366 // * if both are undef, replace with constant 0.
3367 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3368 // * if both are not undef, and are different, bailout.
3369 // * else, only one is undef, then pick the non-undef one.
3370
3371 // The shift amount must be equal.
3372 if (*C0 != *C1)
3373 return nullptr;
3374 const APInt &MaskedBits = *C0;
3375 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3375, __PRETTY_FUNCTION__))
;
3376
3377 ICmpInst::Predicate DstPred;
3378 switch (SrcPred) {
3379 case ICmpInst::Predicate::ICMP_EQ:
3380 // ((%x << MaskedBits) a>> MaskedBits) == %x
3381 // =>
3382 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3383 DstPred = ICmpInst::Predicate::ICMP_ULT;
3384 break;
3385 case ICmpInst::Predicate::ICMP_NE:
3386 // ((%x << MaskedBits) a>> MaskedBits) != %x
3387 // =>
3388 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3389 DstPred = ICmpInst::Predicate::ICMP_UGE;
3390 break;
3391 // FIXME: are more folds possible?
3392 default:
3393 return nullptr;
3394 }
3395
3396 auto *XType = X->getType();
3397 const unsigned XBitWidth = XType->getScalarSizeInBits();
3398 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3399 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3399, __PRETTY_FUNCTION__))
;
3400
3401 // KeptBits = bitwidth(%x) - MaskedBits
3402 const APInt KeptBits = BitWidth - MaskedBits;
3403 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3403, __PRETTY_FUNCTION__))
;
3404 // ICmpCst = (1 << KeptBits)
3405 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3406 assert(ICmpCst.isPowerOf2())((ICmpCst.isPowerOf2()) ? static_cast<void> (0) : __assert_fail
("ICmpCst.isPowerOf2()", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3406, __PRETTY_FUNCTION__))
;
3407 // AddCst = (1 << (KeptBits-1))
3408 const APInt AddCst = ICmpCst.lshr(1);
3409 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3409, __PRETTY_FUNCTION__))
;
3410
3411 // T0 = add %x, AddCst
3412 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3413 // T1 = T0 DstPred ICmpCst
3414 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3415
3416 return T1;
3417}
3418
3419// Given pattern:
3420// icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3421// we should move shifts to the same hand of 'and', i.e. rewrite as
3422// icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3423// We are only interested in opposite logical shifts here.
3424// One of the shifts can be truncated.
3425// If we can, we want to end up creating 'lshr' shift.
3426static Value *
3427foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3428 InstCombiner::BuilderTy &Builder) {
3429 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
30
Assuming the condition is false
32
Taking false branch
3430 !I.getOperand(0)->hasOneUse())
31
Assuming the condition is false
3431 return nullptr;
3432
3433 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3434
3435 // Look for an 'and' of two logical shifts, one of which may be truncated.
3436 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3437 Instruction *XShift, *MaybeTruncation, *YShift;
3438 if (!match(
33
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>>'
41
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>>'
42
Taking false branch
3439 I.getOperand(0),
3440 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3441 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3442 m_AnyLogicalShift, m_Instruction(YShift))),
3443 m_Instruction(MaybeTruncation)))))
3444 return nullptr;
3445
3446 // We potentially looked past 'trunc', but only when matching YShift,
3447 // therefore YShift must have the widest type.
3448 Instruction *WidestShift = YShift;
3449 // Therefore XShift must have the shallowest type.
3450 // Or they both have identical types if there was no truncation.
3451 Instruction *NarrowestShift = XShift;
3452
3453 Type *WidestTy = WidestShift->getType();
3454 Type *NarrowestTy = NarrowestShift->getType();
3455 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3456, __PRETTY_FUNCTION__))
43
Assuming the condition is true
44
'?' condition is true
3456 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3456, __PRETTY_FUNCTION__))
;
3457 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
45
Assuming the condition is false
3458
3459 // If YShift is a 'lshr', swap the shifts around.
3460 if (match(YShift, m_LShr(m_Value(), m_Value())))
46
Taking false branch
3461 std::swap(XShift, YShift);
3462
3463 // The shifts must be in opposite directions.
3464 auto XShiftOpcode = XShift->getOpcode();
3465 if (XShiftOpcode == YShift->getOpcode())
47
Assuming the condition is false
48
Taking false branch
3466 return nullptr; // Do not care about same-direction shifts here.
3467
3468 Value *X, *XShAmt, *Y, *YShAmt;
49
'XShAmt' declared without an initial value
3469 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
50
Calling 'm_Value'
54
Returning from 'm_Value'
55
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
63
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
64
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> >>'
66
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> >>'
3470 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3471
3472 // If one of the values being shifted is a constant, then we will end with
3473 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3474 // however, we will need to ensure that we won't increase instruction count.
3475 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
67
Assuming 'X' is a 'Constant'
3476 // At least one of the hands of the 'and' should be one-use shift.
3477 if (!match(I.getOperand(0),
3478 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3479 return nullptr;
3480 if (HadTrunc) {
3481 // Due to the 'trunc', we will need to widen X. For that either the old
3482 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3483 if (!MaybeTruncation->hasOneUse() &&
3484 !NarrowestShift->getOperand(1)->hasOneUse())
3485 return nullptr;
3486 }
3487 }
3488
3489 // We have two shift amounts from two different shifts. The types of those
3490 // shift amounts may not match. If that's the case let's bailout now.
3491 if (XShAmt->getType() != YShAmt->getType())
68
Called C++ object pointer is uninitialized
3492 return nullptr;
3493
3494 // As input, we have the following pattern:
3495 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3496 // We want to rewrite that as:
3497 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3498 // While we know that originally (Q+K) would not overflow
3499 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3500 // shift amounts. so it may now overflow in smaller bitwidth.
3501 // To ensure that does not happen, we need to ensure that the total maximal
3502 // shift amount is still representable in that smaller bit width.
3503 unsigned MaximalPossibleTotalShiftAmount =
3504 (WidestTy->getScalarSizeInBits() - 1) +
3505 (NarrowestTy->getScalarSizeInBits() - 1);
3506 APInt MaximalRepresentableShiftAmount =
3507 APInt::getAllOnesValue(XShAmt->getType()->getScalarSizeInBits());
3508 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3509 return nullptr;
3510
3511 // Can we fold (XShAmt+YShAmt) ?
3512 auto *NewShAmt = dyn_cast_or_null<Constant>(
3513 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3514 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3515 if (!NewShAmt)
3516 return nullptr;
3517 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3518 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3519
3520 // Is the new shift amount smaller than the bit width?
3521 // FIXME: could also rely on ConstantRange.
3522 if (!match(NewShAmt,
3523 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3524 APInt(WidestBitWidth, WidestBitWidth))))
3525 return nullptr;
3526
3527 // An extra legality check is needed if we had trunc-of-lshr.
3528 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3529 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3530 WidestShift]() {
3531 // It isn't obvious whether it's worth it to analyze non-constants here.
3532 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3533 // If *any* of these preconditions matches we can perform the fold.
3534 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3535 ? NewShAmt->getSplatValue()
3536 : NewShAmt;
3537 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3538 if (NewShAmtSplat &&
3539 (NewShAmtSplat->isNullValue() ||
3540 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3541 return true;
3542 // We consider *min* leading zeros so a single outlier
3543 // blocks the transform as opposed to allowing it.
3544 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3545 KnownBits Known = computeKnownBits(C, SQ.DL);
3546 unsigned MinLeadZero = Known.countMinLeadingZeros();
3547 // If the value being shifted has at most lowest bit set we can fold.
3548 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3549 if (MaxActiveBits <= 1)
3550 return true;
3551 // Precondition: NewShAmt u<= countLeadingZeros(C)
3552 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3553 return true;
3554 }
3555 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3556 KnownBits Known = computeKnownBits(C, SQ.DL);
3557 unsigned MinLeadZero = Known.countMinLeadingZeros();
3558 // If the value being shifted has at most lowest bit set we can fold.
3559 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3560 if (MaxActiveBits <= 1)
3561 return true;
3562 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3563 if (NewShAmtSplat) {
3564 APInt AdjNewShAmt =
3565 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3566 if (AdjNewShAmt.ule(MinLeadZero))
3567 return true;
3568 }
3569 }
3570 return false; // Can't tell if it's ok.
3571 };
3572 if (!CanFold())
3573 return nullptr;
3574 }
3575
3576 // All good, we can do this fold.
3577 X = Builder.CreateZExt(X, WidestTy);
3578 Y = Builder.CreateZExt(Y, WidestTy);
3579 // The shift is the same that was for X.
3580 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3581 ? Builder.CreateLShr(X, NewShAmt)
3582 : Builder.CreateShl(X, NewShAmt);
3583 Value *T1 = Builder.CreateAnd(T0, Y);
3584 return Builder.CreateICmp(I.getPredicate(), T1,
3585 Constant::getNullValue(WidestTy));
3586}
3587
3588/// Fold
3589/// (-1 u/ x) u< y
3590/// ((x * y) u/ x) != y
3591/// to
3592/// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3593/// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3594/// will mean that we are looking for the opposite answer.
3595Value *InstCombiner::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3596 ICmpInst::Predicate Pred;
3597 Value *X, *Y;
3598 Instruction *Mul;
3599 bool NeedNegation;
3600 // Look for: (-1 u/ x) u</u>= y
3601 if (!I.isEquality() &&
3602 match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3603 m_Value(Y)))) {
3604 Mul = nullptr;
3605
3606 // Are we checking that overflow does not happen, or does happen?
3607 switch (Pred) {
3608 case ICmpInst::Predicate::ICMP_ULT:
3609 NeedNegation = false;
3610 break; // OK
3611 case ICmpInst::Predicate::ICMP_UGE:
3612 NeedNegation = true;
3613 break; // OK
3614 default:
3615 return nullptr; // Wrong predicate.
3616 }
3617 } else // Look for: ((x * y) u/ x) !=/== y
3618 if (I.isEquality() &&
3619 match(&I, m_c_ICmp(Pred, m_Value(Y),
3620 m_OneUse(m_UDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3621 m_Value(X)),
3622 m_Instruction(Mul)),
3623 m_Deferred(X)))))) {
3624 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3625 } else
3626 return nullptr;
3627
3628 BuilderTy::InsertPointGuard Guard(Builder);
3629 // If the pattern included (x * y), we'll want to insert new instructions
3630 // right before that original multiplication so that we can replace it.
3631 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3632 if (MulHadOtherUses)
3633 Builder.SetInsertPoint(Mul);
3634
3635 Function *F = Intrinsic::getDeclaration(
3636 I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3637 CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3638
3639 // If the multiplication was used elsewhere, to ensure that we don't leave
3640 // "duplicate" instructions, replace uses of that original multiplication
3641 // with the multiplication result from the with.overflow intrinsic.
3642 if (MulHadOtherUses)
3643 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3644
3645 Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3646 if (NeedNegation) // This technically increases instruction count.
3647 Res = Builder.CreateNot(Res, "umul.not.ov");
3648
3649 return Res;
3650}
3651
3652/// Try to fold icmp (binop), X or icmp X, (binop).
3653/// TODO: A large part of this logic is duplicated in InstSimplify's
3654/// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3655/// duplication.
3656Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I, const SimplifyQuery &SQ) {
3657 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3658 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3659
3660 // Special logic for binary operators.
3661 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
1
Assuming 'Op0' is a 'BinaryOperator'
3662 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2
Assuming 'Op1' is not a 'BinaryOperator'
3663 if (!BO0
2.1
'BO0' is non-null
2.1
'BO0' is non-null
&& !BO1)
3664 return nullptr;
3665
3666 const CmpInst::Predicate Pred = I.getPredicate();
3667 Value *X;
3668
3669 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3670 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3671 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3
Taking false branch
3672 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3673 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3674 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3675 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
4
Taking false branch
3676 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3677 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3678
3679 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3680 if (BO0
4.1
'BO0' is non-null
4.1
'BO0' is non-null
&& isa<OverflowingBinaryOperator>(BO0))
5
Assuming 'BO0' is not a 'OverflowingBinaryOperator'
6
Taking false branch
3681 NoOp0WrapProblem =
3682 ICmpInst::isEquality(Pred) ||
3683 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3684 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3685 if (BO1
6.1
'BO1' is null
6.1
'BO1' is null
&& isa<OverflowingBinaryOperator>(BO1))
3686 NoOp1WrapProblem =
3687 ICmpInst::isEquality(Pred) ||
3688 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3689 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3690
3691 // Analyze the case when either Op0 or Op1 is an add instruction.
3692 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3693 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3694 if (BO0
6.2
'BO0' is non-null
6.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Add) {
7
Assuming the condition is false
8
Taking false branch
3695 A = BO0->getOperand(0);
3696 B = BO0->getOperand(1);
3697 }
3698 if (BO1
8.1
'BO1' is null
8.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Add) {
3699 C = BO1->getOperand(0);
3700 D = BO1->getOperand(1);
3701 }
3702
3703 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3704 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3705 if ((A
8.2
'A' is not equal to 'Op1'
8.2
'A' is not equal to 'Op1'
== Op1 || B
8.3
'B' is not equal to 'Op1'
8.3
'B' is not equal to 'Op1'
== Op1) && NoOp0WrapProblem)
3706 return new ICmpInst(Pred, A == Op1 ? B : A,
3707 Constant::getNullValue(Op1->getType()));
3708
3709 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3710 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3711 if ((C
8.4
'C' is not equal to 'Op0'
8.4
'C' is not equal to 'Op0'
== Op0 || D
8.5
'D' is not equal to 'Op0'
8.5
'D' is not equal to 'Op0'
== Op0) && NoOp1WrapProblem)
3712 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3713 C == Op0 ? D : C);
3714
3715 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3716 if (A
8.6
'A' is null
8.6
'A' is null
&& C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3717 NoOp1WrapProblem) {
3718 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3719 Value *Y, *Z;
3720 if (A == C) {
3721 // C + B == C + D -> B == D
3722 Y = B;
3723 Z = D;
3724 } else if (A == D) {
3725 // D + B == C + D -> B == C
3726 Y = B;
3727 Z = C;
3728 } else if (B == C) {
3729 // A + C == C + D -> A == D
3730 Y = A;
3731 Z = D;
3732 } else {
3733 assert(B == D)((B == D) ? static_cast<void> (0) : __assert_fail ("B == D"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3733, __PRETTY_FUNCTION__))
;
3734 // A + D == C + D -> A == C
3735 Y = A;
3736 Z = C;
3737 }
3738 return new ICmpInst(Pred, Y, Z);
3739 }
3740
3741 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3742 if (A
8.7
'A' is null
8.7
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
9
Taking false branch
3743 match(B, m_AllOnes()))
3744 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3745
3746 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3747 if (A
9.1
'A' is null
9.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
10
Taking false branch
3748 match(B, m_AllOnes()))
3749 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3750
3751 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3752 if (A
10.1
'A' is null
10.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
11
Taking false branch
3753 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3754
3755 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3756 if (A
11.1
'A' is null
11.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
12
Taking false branch
3757 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3758
3759 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3760 if (C
12.1
'C' is null
12.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
13
Taking false branch
3761 match(D, m_AllOnes()))
3762 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3763
3764 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3765 if (C
13.1
'C' is null
13.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
14
Taking false branch
3766 match(D, m_AllOnes()))
3767 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3768
3769 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3770 if (C
14.1
'C' is null
14.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
15
Taking false branch
3771 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3772
3773 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3774 if (C
15.1
'C' is null
15.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
16
Taking false branch
3775 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3776
3777 // TODO: The subtraction-related identities shown below also hold, but
3778 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3779 // wouldn't happen even if they were implemented.
3780 //
3781 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3782 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3783 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3784 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3785
3786 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3787 if (A
16.1
'A' is null
16.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
17
Taking false branch
3788 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3789
3790 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3791 if (A
17.1
'A' is null
17.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
18
Taking false branch
3792 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3793
3794 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3795 if (C
18.1
'C' is null
18.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
19
Taking false branch
3796 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3797
3798 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3799 if (C
19.1
'C' is null
19.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
20
Taking false branch
3800 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3801
3802 // if C1 has greater magnitude than C2:
3803 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3804 // s.t. C3 = C1 - C2
3805 //
3806 // if C2 has greater magnitude than C1:
3807 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3808 // s.t. C3 = C2 - C1
3809 if (A
20.1
'A' is null
20.1
'A' is null
&& C && NoOp0WrapProblem && NoOp1WrapProblem &&
3810 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3811 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3812 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3813 const APInt &AP1 = C1->getValue();
3814 const APInt &AP2 = C2->getValue();
3815 if (AP1.isNegative() == AP2.isNegative()) {
3816 APInt AP1Abs = C1->getValue().abs();
3817 APInt AP2Abs = C2->getValue().abs();
3818 if (AP1Abs.uge(AP2Abs)) {
3819 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3820 Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3821 return new ICmpInst(Pred, NewAdd, C);
3822 } else {
3823 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3824 Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3825 return new ICmpInst(Pred, A, NewAdd);
3826 }
3827 }
3828 }
3829
3830 // Analyze the case when either Op0 or Op1 is a sub instruction.
3831 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3832 A = nullptr;
3833 B = nullptr;
3834 C = nullptr;
3835 D = nullptr;
3836 if (BO0
20.2
'BO0' is non-null
20.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Sub) {
21
Assuming the condition is false
22
Taking false branch
3837 A = BO0->getOperand(0);
3838 B = BO0->getOperand(1);
3839 }
3840 if (BO1
22.1
'BO1' is null
22.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Sub) {
3841 C = BO1->getOperand(0);
3842 D = BO1->getOperand(1);
3843 }
3844
3845 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3846 if (A
22.2
'A' is not equal to 'Op1'
22.2
'A' is not equal to 'Op1'
== Op1 && NoOp0WrapProblem)
3847 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3848 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
3849 if (C
22.3
'C' is not equal to 'Op0'
22.3
'C' is not equal to 'Op0'
== Op0 && NoOp1WrapProblem)
3850 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3851
3852 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3853 // (A - B) u>/u<= A --> B u>/u<= A
3854 if (A
22.4
'A' is not equal to 'Op1'
22.4
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3855 return new ICmpInst(Pred, B, A);
3856 // C u</u>= (C - D) --> C u</u>= D
3857 if (C
22.5
'C' is not equal to 'Op0'
22.5
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3858 return new ICmpInst(Pred, C, D);
3859 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
3860 if (A
22.6
'A' is not equal to 'Op1'
22.6
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
3861 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3862 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
3863 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
3864 if (C
22.7
'C' is not equal to 'Op0'
22.7
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
3865 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3866 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
3867
3868 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
3869 if (B
22.8
'B' is null
22.8
'B' is null
&& D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
3870 return new ICmpInst(Pred, A, C);
3871
3872 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
3873 if (A
22.9
'A' is null
22.9
'A' is null
&& C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
3874 return new ICmpInst(Pred, D, B);
3875
3876 // icmp (0-X) < cst --> x > -cst
3877 if (NoOp0WrapProblem
22.10
'NoOp0WrapProblem' is false
22.10
'NoOp0WrapProblem' is false
&& ICmpInst::isSigned(Pred)) {
3878 Value *X;
3879 if (match(BO0, m_Neg(m_Value(X))))
3880 if (Constant *RHSC = dyn_cast<Constant>(Op1))
3881 if (RHSC->isNotMinSignedValue())
3882 return new ICmpInst(I.getSwappedPredicate(), X,
3883 ConstantExpr::getNeg(RHSC));
3884 }
3885
3886 BinaryOperator *SRem = nullptr;
3887 // icmp (srem X, Y), Y
3888 if (BO0
22.11
'BO0' is non-null
22.11
'BO0' is non-null
&& BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
23
Assuming the condition is false
3889 SRem = BO0;
3890 // icmp Y, (srem X, Y)
3891 else if (BO1
23.1
'BO1' is null
23.1
'BO1' is null
&& BO1->getOpcode() == Instruction::SRem &&
3892 Op0 == BO1->getOperand(1))
3893 SRem = BO1;
3894 if (SRem
23.2
'SRem' is null
23.2
'SRem' is null
) {
24
Taking false branch
3895 // We don't check hasOneUse to avoid increasing register pressure because
3896 // the value we use is the same value this instruction was already using.
3897 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3898 default:
3899 break;
3900 case ICmpInst::ICMP_EQ:
3901 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3902 case ICmpInst::ICMP_NE:
3903 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3904 case ICmpInst::ICMP_SGT:
3905 case ICmpInst::ICMP_SGE:
3906 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3907 Constant::getAllOnesValue(SRem->getType()));
3908 case ICmpInst::ICMP_SLT:
3909 case ICmpInst::ICMP_SLE:
3910 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3911 Constant::getNullValue(SRem->getType()));
3912 }
3913 }
3914
3915 if (BO0
24.1
'BO0' is non-null
24.1
'BO0' is non-null
&& BO1
24.2
'BO1' is null
24.2
'BO1' is null
&& BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3916 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3917 switch (BO0->getOpcode()) {
3918 default:
3919 break;
3920 case Instruction::Add:
3921 case Instruction::Sub:
3922 case Instruction::Xor: {
3923 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3924 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3925
3926 const APInt *C;
3927 if (match(BO0->getOperand(1), m_APInt(C))) {
3928 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3929 if (C->isSignMask()) {
3930 ICmpInst::Predicate NewPred =
3931 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3932 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3933 }
3934
3935 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3936 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3937 ICmpInst::Predicate NewPred =
3938 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3939 NewPred = I.getSwappedPredicate(NewPred);
3940 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3941 }
3942 }
3943 break;
3944 }
3945 case Instruction::Mul: {
3946 if (!I.isEquality())
3947 break;
3948
3949 const APInt *C;
3950 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3951 !C->isOneValue()) {
3952 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3953 // Mask = -1 >> count-trailing-zeros(C).
3954 if (unsigned TZs = C->countTrailingZeros()) {
3955 Constant *Mask = ConstantInt::get(
3956 BO0->getType(),
3957 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3958 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3959 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3960 return new ICmpInst(Pred, And1, And2);
3961 }
3962 // If there are no trailing zeros in the multiplier, just eliminate
3963 // the multiplies (no masking is needed):
3964 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3965 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3966 }
3967 break;
3968 }
3969 case Instruction::UDiv:
3970 case Instruction::LShr:
3971 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3972 break;
3973 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3974
3975 case Instruction::SDiv:
3976 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3977 break;
3978 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3979
3980 case Instruction::AShr:
3981 if (!BO0->isExact() || !BO1->isExact())
3982 break;
3983 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3984
3985 case Instruction::Shl: {
3986 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3987 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3988 if (!NUW && !NSW)
3989 break;
3990 if (!NSW && I.isSigned())
3991 break;
3992 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3993 }
3994 }
3995 }
3996
3997 if (BO0
24.3
'BO0' is non-null
24.3
'BO0' is non-null
) {
25
Taking true branch
3998 // Transform A & (L - 1) `ult` L --> L != 0
3999 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4000 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4001
4002 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4003 auto *Zero = Constant::getNullValue(BO0->getType());
4004 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4005 }
4006 }
4007
4008 if (Value *V
25.1
'V' is null
25.1
'V' is null
= foldUnsignedMultiplicationOverflowCheck(I))
26
Taking false branch
4009 return replaceInstUsesWith(I, V);
4010
4011 if (Value *V
26.1
'V' is null
26.1
'V' is null
= foldICmpWithLowBitMaskedVal(I, Builder))
27
Taking false branch
4012 return replaceInstUsesWith(I, V);
4013
4014 if (Value *V
27.1
'V' is null
27.1
'V' is null
= foldICmpWithTruncSignExtendedVal(I, Builder))
28
Taking false branch
4015 return replaceInstUsesWith(I, V);
4016
4017 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
29
Calling 'foldShiftIntoShiftInAnotherHandOfAndInICmp'
4018 return replaceInstUsesWith(I, V);
4019
4020 return nullptr;
4021}
4022
4023/// Fold icmp Pred min|max(X, Y), X.
4024static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4025 ICmpInst::Predicate Pred = Cmp.getPredicate();
4026 Value *Op0 = Cmp.getOperand(0);
4027 Value *X = Cmp.getOperand(1);
4028
4029 // Canonicalize minimum or maximum operand to LHS of the icmp.
4030 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4031 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4032 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4033 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4034 std::swap(Op0, X);
4035 Pred = Cmp.getSwappedPredicate();
4036 }
4037
4038 Value *Y;
4039 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4040 // smin(X, Y) == X --> X s<= Y
4041 // smin(X, Y) s>= X --> X s<= Y
4042 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4043 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4044
4045 // smin(X, Y) != X --> X s> Y
4046 // smin(X, Y) s< X --> X s> Y
4047 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4048 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4049
4050 // These cases should be handled in InstSimplify:
4051 // smin(X, Y) s<= X --> true
4052 // smin(X, Y) s> X --> false
4053 return nullptr;
4054 }
4055
4056 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4057 // smax(X, Y) == X --> X s>= Y
4058 // smax(X, Y) s<= X --> X s>= Y
4059 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4060 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4061
4062 // smax(X, Y) != X --> X s< Y
4063 // smax(X, Y) s> X --> X s< Y
4064 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4065 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4066
4067 // These cases should be handled in InstSimplify:
4068 // smax(X, Y) s>= X --> true
4069 // smax(X, Y) s< X --> false
4070 return nullptr;
4071 }
4072
4073 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4074 // umin(X, Y) == X --> X u<= Y
4075 // umin(X, Y) u>= X --> X u<= Y
4076 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4077 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4078
4079 // umin(X, Y) != X --> X u> Y
4080 // umin(X, Y) u< X --> X u> Y
4081 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4082 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4083
4084 // These cases should be handled in InstSimplify:
4085 // umin(X, Y) u<= X --> true
4086 // umin(X, Y) u> X --> false
4087 return nullptr;
4088 }
4089
4090 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4091 // umax(X, Y) == X --> X u>= Y
4092 // umax(X, Y) u<= X --> X u>= Y
4093 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4094 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4095
4096 // umax(X, Y) != X --> X u< Y
4097 // umax(X, Y) u> X --> X u< Y
4098 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4099 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4100
4101 // These cases should be handled in InstSimplify:
4102 // umax(X, Y) u>= X --> true
4103 // umax(X, Y) u< X --> false
4104 return nullptr;
4105 }
4106
4107 return nullptr;
4108}
4109
4110Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
4111 if (!I.isEquality())
4112 return nullptr;
4113
4114 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4115 const CmpInst::Predicate Pred = I.getPredicate();
4116 Value *A, *B, *C, *D;
4117 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4118 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4119 Value *OtherVal = A == Op1 ? B : A;
4120 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4121 }
4122
4123 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4124 // A^c1 == C^c2 --> A == C^(c1^c2)
4125 ConstantInt *C1, *C2;
4126 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4127 Op1->hasOneUse()) {
4128 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4129 Value *Xor = Builder.CreateXor(C, NC);
4130 return new ICmpInst(Pred, A, Xor);
4131 }
4132
4133 // A^B == A^D -> B == D
4134 if (A == C)
4135 return new ICmpInst(Pred, B, D);
4136 if (A == D)
4137 return new ICmpInst(Pred, B, C);
4138 if (B == C)
4139 return new ICmpInst(Pred, A, D);
4140 if (B == D)
4141 return new ICmpInst(Pred, A, C);
4142 }
4143 }
4144
4145 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4146 // A == (A^B) -> B == 0
4147 Value *OtherVal = A == Op0 ? B : A;
4148 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4149 }
4150
4151 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4152 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4153 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4154 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4155
4156 if (A == C) {
4157 X = B;
4158 Y = D;
4159 Z = A;
4160 } else if (A == D) {
4161 X = B;
4162 Y = C;
4163 Z = A;
4164 } else if (B == C) {
4165 X = A;
4166 Y = D;
4167 Z = B;
4168 } else if (B == D) {
4169 X = A;
4170 Y = C;
4171 Z = B;
4172 }
4173
4174 if (X) { // Build (X^Y) & Z
4175 Op1 = Builder.CreateXor(X, Y);
4176 Op1 = Builder.CreateAnd(Op1, Z);
4177 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4178 }
4179 }
4180
4181 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4182 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4183 ConstantInt *Cst1;
4184 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4185 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4186 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4187 match(Op1, m_ZExt(m_Value(A))))) {
4188 APInt Pow2 = Cst1->getValue() + 1;
4189 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4190 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4191 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4192 }
4193
4194 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4195 // For lshr and ashr pairs.
4196 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4197 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4198 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4199 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4200 unsigned TypeBits = Cst1->getBitWidth();
4201 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4202 if (ShAmt < TypeBits && ShAmt != 0) {
4203 ICmpInst::Predicate NewPred =
4204 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4205 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4206 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4207 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4208 }
4209 }
4210
4211 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4212 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4213 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4214 unsigned TypeBits = Cst1->getBitWidth();
4215 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4216 if (ShAmt < TypeBits && ShAmt != 0) {
4217 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4218 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4219 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4220 I.getName() + ".mask");
4221 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4222 }
4223 }
4224
4225 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4226 // "icmp (and X, mask), cst"
4227 uint64_t ShAmt = 0;
4228 if (Op0->hasOneUse() &&
4229 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4230 match(Op1, m_ConstantInt(Cst1)) &&
4231 // Only do this when A has multiple uses. This is most important to do
4232 // when it exposes other optimizations.
4233 !A->hasOneUse()) {
4234 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4235
4236 if (ShAmt < ASize) {
4237 APInt MaskV =
4238 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4239 MaskV <<= ShAmt;
4240
4241 APInt CmpV = Cst1->getValue().zext(ASize);
4242 CmpV <<= ShAmt;
4243
4244 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4245 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4246 }
4247 }
4248
4249 // If both operands are byte-swapped or bit-reversed, just compare the
4250 // original values.
4251 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4252 // and handle more intrinsics.
4253 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4254 (match(Op0, m_BitReverse(m_Value(A))) &&
4255 match(Op1, m_BitReverse(m_Value(B)))))
4256 return new ICmpInst(Pred, A, B);
4257
4258 // Canonicalize checking for a power-of-2-or-zero value:
4259 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4260 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4261 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4262 m_Deferred(A)))) ||
4263 !match(Op1, m_ZeroInt()))
4264 A = nullptr;
4265
4266 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4267 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4268 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4269 A = Op1;
4270 else if (match(Op1,
4271 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4272 A = Op0;
4273
4274 if (A) {
4275 Type *Ty = A->getType();
4276 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4277 return Pred == ICmpInst::ICMP_EQ
4278 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4279 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4280 }
4281
4282 return nullptr;
4283}
4284
4285static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4286 InstCombiner::BuilderTy &Builder) {
4287 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4287, __PRETTY_FUNCTION__))
;
4288 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4289 Value *X;
4290 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4291 return nullptr;
4292
4293 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4294 bool IsSignedCmp = ICmp.isSigned();
4295 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4296 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4297 // and the other is a zext), then we can't handle this.
4298 // TODO: This is too strict. We can handle some predicates (equality?).
4299 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4300 return nullptr;
4301
4302 // Not an extension from the same type?
4303 Value *Y = CastOp1->getOperand(0);
4304 Type *XTy = X->getType(), *YTy = Y->getType();
4305 if (XTy != YTy) {
4306 // One of the casts must have one use because we are creating a new cast.
4307 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4308 return nullptr;
4309 // Extend the narrower operand to the type of the wider operand.
4310 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4311 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4312 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4313 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4314 else
4315 return nullptr;
4316 }
4317
4318 // (zext X) == (zext Y) --> X == Y
4319 // (sext X) == (sext Y) --> X == Y
4320 if (ICmp.isEquality())
4321 return new ICmpInst(ICmp.getPredicate(), X, Y);
4322
4323 // A signed comparison of sign extended values simplifies into a
4324 // signed comparison.
4325 if (IsSignedCmp && IsSignedExt)
4326 return new ICmpInst(ICmp.getPredicate(), X, Y);
4327
4328 // The other three cases all fold into an unsigned comparison.
4329 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4330 }
4331
4332 // Below here, we are only folding a compare with constant.
4333 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4334 if (!C)
4335 return nullptr;
4336
4337 // Compute the constant that would happen if we truncated to SrcTy then
4338 // re-extended to DestTy.
4339 Type *SrcTy = CastOp0->getSrcTy();
4340 Type *DestTy = CastOp0->getDestTy();
4341 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4342 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4343
4344 // If the re-extended constant didn't change...
4345 if (Res2 == C) {
4346 if (ICmp.isEquality())
4347 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4348
4349 // A signed comparison of sign extended values simplifies into a
4350 // signed comparison.
4351 if (IsSignedExt && IsSignedCmp)
4352 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4353
4354 // The other three cases all fold into an unsigned comparison.
4355 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4356 }
4357
4358 // The re-extended constant changed, partly changed (in the case of a vector),
4359 // or could not be determined to be equal (in the case of a constant
4360 // expression), so the constant cannot be represented in the shorter type.
4361 // All the cases that fold to true or false will have already been handled
4362 // by SimplifyICmpInst, so only deal with the tricky case.
4363 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4364 return nullptr;
4365
4366 // Is source op positive?
4367 // icmp ult (sext X), C --> icmp sgt X, -1
4368 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4369 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4370
4371 // Is source op negative?
4372 // icmp ugt (sext X), C --> icmp slt X, 0
4373 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4373, __PRETTY_FUNCTION__))
;
4374 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4375}
4376
4377/// Handle icmp (cast x), (cast or constant).
4378Instruction *InstCombiner::foldICmpWithCastOp(ICmpInst &ICmp) {
4379 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4380 if (!CastOp0)
4381 return nullptr;
4382 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4383 return nullptr;
4384
4385 Value *Op0Src = CastOp0->getOperand(0);
4386 Type *SrcTy = CastOp0->getSrcTy();
4387 Type *DestTy = CastOp0->getDestTy();
4388
4389 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4390 // integer type is the same size as the pointer type.
4391 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4392 if (isa<VectorType>(SrcTy)) {
4393 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4394 DestTy = cast<VectorType>(DestTy)->getElementType();
4395 }
4396 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4397 };
4398 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4399 CompatibleSizes(SrcTy, DestTy)) {
4400 Value *NewOp1 = nullptr;
4401 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4402 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4403 if (PtrSrc->getType()->getPointerAddressSpace() ==
4404 Op0Src->getType()->getPointerAddressSpace()) {
4405 NewOp1 = PtrToIntOp1->getOperand(0);
4406 // If the pointer types don't match, insert a bitcast.
4407 if (Op0Src->getType() != NewOp1->getType())
4408 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4409 }
4410 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4411 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4412 }
4413
4414 if (NewOp1)
4415 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4416 }
4417
4418 return foldICmpWithZextOrSext(ICmp, Builder);
4419}
4420
4421static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4422 switch (BinaryOp) {
4423 default:
4424 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4424)
;
4425 case Instruction::Add:
4426 case Instruction::Sub:
4427 return match(RHS, m_Zero());
4428 case Instruction::Mul:
4429 return match(RHS, m_One());
4430 }
4431}
4432
4433OverflowResult InstCombiner::computeOverflow(
4434 Instruction::BinaryOps BinaryOp, bool IsSigned,
4435 Value *LHS, Value *RHS, Instruction *CxtI) const {
4436 switch (BinaryOp) {
4437 default:
4438 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4438)
;
4439 case Instruction::Add:
4440 if (IsSigned)
4441 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4442 else
4443 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4444 case Instruction::Sub:
4445 if (IsSigned)
4446 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4447 else
4448 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4449 case Instruction::Mul:
4450 if (IsSigned)
4451 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4452 else
4453 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4454 }
4455}
4456
4457bool InstCombiner::OptimizeOverflowCheck(
4458 Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS,
4459 Instruction &OrigI, Value *&Result, Constant *&Overflow) {
4460 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4461 std::swap(LHS, RHS);
4462
4463 // If the overflow check was an add followed by a compare, the insertion point
4464 // may be pointing to the compare. We want to insert the new instructions
4465 // before the add in case there are uses of the add between the add and the
4466 // compare.
4467 Builder.SetInsertPoint(&OrigI);
4468
4469 if (isNeutralValue(BinaryOp, RHS)) {
4470 Result = LHS;
4471 Overflow = Builder.getFalse();
4472 return true;
4473 }
4474
4475 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4476 case OverflowResult::MayOverflow:
4477 return false;
4478 case OverflowResult::AlwaysOverflowsLow:
4479 case OverflowResult::AlwaysOverflowsHigh:
4480 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4481 Result->takeName(&OrigI);
4482 Overflow = Builder.getTrue();
4483 return true;
4484 case OverflowResult::NeverOverflows:
4485 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4486 Result->takeName(&OrigI);
4487 Overflow = Builder.getFalse();
4488 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4489 if (IsSigned)
4490 Inst->setHasNoSignedWrap();
4491 else
4492 Inst->setHasNoUnsignedWrap();
4493 }
4494 return true;
4495 }
4496
4497 llvm_unreachable("Unexpected overflow result")::llvm::llvm_unreachable_internal("Unexpected overflow result"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4497)
;
4498}
4499
4500/// Recognize and process idiom involving test for multiplication
4501/// overflow.
4502///
4503/// The caller has matched a pattern of the form:
4504/// I = cmp u (mul(zext A, zext B), V
4505/// The function checks if this is a test for overflow and if so replaces
4506/// multiplication with call to 'mul.with.overflow' intrinsic.
4507///
4508/// \param I Compare instruction.
4509/// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4510/// the compare instruction. Must be of integer type.
4511/// \param OtherVal The other argument of compare instruction.
4512/// \returns Instruction which must replace the compare instruction, NULL if no
4513/// replacement required.
4514static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4515 Value *OtherVal, InstCombiner &IC) {
4516 // Don't bother doing this transformation for pointers, don't do it for
4517 // vectors.
4518 if (!isa<IntegerType>(MulVal->getType()))
4519 return nullptr;
4520
4521 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4521, __PRETTY_FUNCTION__))
;
4522 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4522, __PRETTY_FUNCTION__))
;
4523 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4524 if (!MulInstr)
4525 return nullptr;
4526 assert(MulInstr->getOpcode() == Instruction::Mul)((MulInstr->getOpcode() == Instruction::Mul) ? static_cast
<void> (0) : __assert_fail ("MulInstr->getOpcode() == Instruction::Mul"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4526, __PRETTY_FUNCTION__))
;
4527
4528 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4529 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4530 assert(LHS->getOpcode() == Instruction::ZExt)((LHS->getOpcode() == Instruction::ZExt) ? static_cast<
void> (0) : __assert_fail ("LHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4530, __PRETTY_FUNCTION__))
;
4531 assert(RHS->getOpcode() == Instruction::ZExt)((RHS->getOpcode() == Instruction::ZExt) ? static_cast<
void> (0) : __assert_fail ("RHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4531, __PRETTY_FUNCTION__))
;
4532 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4533
4534 // Calculate type and width of the result produced by mul.with.overflow.
4535 Type *TyA = A->getType(), *TyB = B->getType();
4536 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4537 WidthB = TyB->getPrimitiveSizeInBits();
4538 unsigned MulWidth;
4539 Type *MulType;
4540 if (WidthB > WidthA) {
4541 MulWidth = WidthB;
4542 MulType = TyB;
4543 } else {
4544 MulWidth = WidthA;
4545 MulType = TyA;
4546 }
4547
4548 // In order to replace the original mul with a narrower mul.with.overflow,
4549 // all uses must ignore upper bits of the product. The number of used low
4550 // bits must be not greater than the width of mul.with.overflow.
4551 if (MulVal->hasNUsesOrMore(2))
4552 for (User *U : MulVal->users()) {
4553 if (U == &I)
4554 continue;
4555 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4556 // Check if truncation ignores bits above MulWidth.
4557 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4558 if (TruncWidth > MulWidth)
4559 return nullptr;
4560 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4561 // Check if AND ignores bits above MulWidth.
4562 if (BO->getOpcode() != Instruction::And)
4563 return nullptr;
4564 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4565 const APInt &CVal = CI->getValue();
4566 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4567 return nullptr;
4568 } else {
4569 // In this case we could have the operand of the binary operation
4570 // being defined in another block, and performing the replacement
4571 // could break the dominance relation.
4572 return nullptr;
4573 }
4574 } else {
4575 // Other uses prohibit this transformation.
4576 return nullptr;
4577 }
4578 }
4579
4580 // Recognize patterns
4581 switch (I.getPredicate()) {
4582 case ICmpInst::ICMP_EQ:
4583 case ICmpInst::ICMP_NE:
4584 // Recognize pattern:
4585 // mulval = mul(zext A, zext B)
4586 // cmp eq/neq mulval, zext trunc mulval
4587 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4588 if (Zext->hasOneUse()) {
4589 Value *ZextArg = Zext->getOperand(0);
4590 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4591 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4592 break; //Recognized
4593 }
4594
4595 // Recognize pattern:
4596 // mulval = mul(zext A, zext B)
4597 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4598 ConstantInt *CI;
4599 Value *ValToMask;
4600 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4601 if (ValToMask != MulVal)
4602 return nullptr;
4603 const APInt &CVal = CI->getValue() + 1;
4604 if (CVal.isPowerOf2()) {
4605 unsigned MaskWidth = CVal.logBase2();
4606 if (MaskWidth == MulWidth)
4607 break; // Recognized
4608 }
4609 }
4610 return nullptr;
4611
4612 case ICmpInst::ICMP_UGT:
4613 // Recognize pattern:
4614 // mulval = mul(zext A, zext B)
4615 // cmp ugt mulval, max
4616 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4617 APInt MaxVal = APInt::getMaxValue(MulWidth);
4618 MaxVal = MaxVal.zext(CI->getBitWidth());
4619 if (MaxVal.eq(CI->getValue()))
4620 break; // Recognized
4621 }
4622 return nullptr;
4623
4624 case ICmpInst::ICMP_UGE:
4625 // Recognize pattern:
4626 // mulval = mul(zext A, zext B)
4627 // cmp uge mulval, max+1
4628 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4629 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4630 if (MaxVal.eq(CI->getValue()))
4631 break; // Recognized
4632 }
4633 return nullptr;
4634
4635 case ICmpInst::ICMP_ULE:
4636 // Recognize pattern:
4637 // mulval = mul(zext A, zext B)
4638 // cmp ule mulval, max
4639 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4640 APInt MaxVal = APInt::getMaxValue(MulWidth);
4641 MaxVal = MaxVal.zext(CI->getBitWidth());
4642 if (MaxVal.eq(CI->getValue()))
4643 break; // Recognized
4644 }
4645 return nullptr;
4646
4647 case ICmpInst::ICMP_ULT:
4648 // Recognize pattern:
4649 // mulval = mul(zext A, zext B)
4650 // cmp ule mulval, max + 1
4651 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4652 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4653 if (MaxVal.eq(CI->getValue()))
4654 break; // Recognized
4655 }
4656 return nullptr;
4657
4658 default:
4659 return nullptr;
4660 }
4661
4662 InstCombiner::BuilderTy &Builder = IC.Builder;
4663 Builder.SetInsertPoint(MulInstr);
4664
4665 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4666 Value *MulA = A, *MulB = B;
4667 if (WidthA < MulWidth)
4668 MulA = Builder.CreateZExt(A, MulType);
4669 if (WidthB < MulWidth)
4670 MulB = Builder.CreateZExt(B, MulType);
4671 Function *F = Intrinsic::getDeclaration(
4672 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4673 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4674 IC.Worklist.push(MulInstr);
4675
4676 // If there are uses of mul result other than the comparison, we know that
4677 // they are truncation or binary AND. Change them to use result of
4678 // mul.with.overflow and adjust properly mask/size.
4679 if (MulVal->hasNUsesOrMore(2)) {
4680 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4681 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4682 User *U = *UI++;
4683 if (U == &I || U == OtherVal)
4684 continue;
4685 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4686 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4687 IC.replaceInstUsesWith(*TI, Mul);
4688 else
4689 TI->setOperand(0, Mul);
4690 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4691 assert(BO->getOpcode() == Instruction::And)((BO->getOpcode() == Instruction::And) ? static_cast<void
> (0) : __assert_fail ("BO->getOpcode() == Instruction::And"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4691, __PRETTY_FUNCTION__))
;
4692 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4693 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4694 APInt ShortMask = CI->getValue().trunc(MulWidth);
4695 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4696 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
4697 IC.replaceInstUsesWith(*BO, Zext);
4698 } else {
4699 llvm_unreachable("Unexpected Binary operation")::llvm::llvm_unreachable_internal("Unexpected Binary operation"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4699)
;
4700 }
4701 IC.Worklist.push(cast<Instruction>(U));
4702 }
4703 }
4704 if (isa<Instruction>(OtherVal))
4705 IC.Worklist.push(cast<Instruction>(OtherVal));
4706
4707 // The original icmp gets replaced with the overflow value, maybe inverted
4708 // depending on predicate.
4709 bool Inverse = false;
4710 switch (I.getPredicate()) {
4711 case ICmpInst::ICMP_NE:
4712 break;
4713 case ICmpInst::ICMP_EQ:
4714 Inverse = true;
4715 break;
4716 case ICmpInst::ICMP_UGT:
4717 case ICmpInst::ICMP_UGE:
4718 if (I.getOperand(0) == MulVal)
4719 break;
4720 Inverse = true;
4721 break;
4722 case ICmpInst::ICMP_ULT:
4723 case ICmpInst::ICMP_ULE:
4724 if (I.getOperand(1) == MulVal)
4725 break;
4726 Inverse = true;
4727 break;
4728 default:
4729 llvm_unreachable("Unexpected predicate")::llvm::llvm_unreachable_internal("Unexpected predicate", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4729)
;
4730 }
4731 if (Inverse) {
4732 Value *Res = Builder.CreateExtractValue(Call, 1);
4733 return BinaryOperator::CreateNot(Res);
4734 }
4735
4736 return ExtractValueInst::Create(Call, 1);
4737}
4738
4739/// When performing a comparison against a constant, it is possible that not all
4740/// the bits in the LHS are demanded. This helper method computes the mask that
4741/// IS demanded.
4742static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4743 const APInt *RHS;
4744 if (!match(I.getOperand(1), m_APInt(RHS)))
4745 return APInt::getAllOnesValue(BitWidth);
4746
4747 // If this is a normal comparison, it demands all bits. If it is a sign bit
4748 // comparison, it only demands the sign bit.
4749 bool UnusedBit;
4750 if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4751 return APInt::getSignMask(BitWidth);
4752
4753 switch (I.getPredicate()) {
4754 // For a UGT comparison, we don't care about any bits that
4755 // correspond to the trailing ones of the comparand. The value of these
4756 // bits doesn't impact the outcome of the comparison, because any value
4757 // greater than the RHS must differ in a bit higher than these due to carry.
4758 case ICmpInst::ICMP_UGT:
4759 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4760
4761 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4762 // Any value less than the RHS must differ in a higher bit because of carries.
4763 case ICmpInst::ICMP_ULT:
4764 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4765
4766 default:
4767 return APInt::getAllOnesValue(BitWidth);
4768 }
4769}
4770
4771/// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4772/// should be swapped.
4773/// The decision is based on how many times these two operands are reused
4774/// as subtract operands and their positions in those instructions.
4775/// The rationale is that several architectures use the same instruction for
4776/// both subtract and cmp. Thus, it is better if the order of those operands
4777/// match.
4778/// \return true if Op0 and Op1 should be swapped.
4779static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4780 // Filter out pointer values as those cannot appear directly in subtract.
4781 // FIXME: we may want to go through inttoptrs or bitcasts.
4782 if (Op0->getType()->isPointerTy())
4783 return false;
4784 // If a subtract already has the same operands as a compare, swapping would be
4785 // bad. If a subtract has the same operands as a compare but in reverse order,
4786 // then swapping is good.
4787 int GoodToSwap = 0;
4788 for (const User *U : Op0->users()) {
4789 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4790 GoodToSwap++;
4791 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4792 GoodToSwap--;
4793 }
4794 return GoodToSwap > 0;
4795}
4796
4797/// Check that one use is in the same block as the definition and all
4798/// other uses are in blocks dominated by a given block.
4799///
4800/// \param DI Definition
4801/// \param UI Use
4802/// \param DB Block that must dominate all uses of \p DI outside
4803/// the parent block
4804/// \return true when \p UI is the only use of \p DI in the parent block
4805/// and all other uses of \p DI are in blocks dominated by \p DB.
4806///
4807bool InstCombiner::dominatesAllUses(const Instruction *DI,
4808 const Instruction *UI,
4809 const BasicBlock *DB) const {
4810 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4810, __PRETTY_FUNCTION__))
;
4811 // Ignore incomplete definitions.
4812 if (!DI->getParent())
4813 return false;
4814 // DI and UI must be in the same block.
4815 if (DI->getParent() != UI->getParent())
4816 return false;
4817 // Protect from self-referencing blocks.
4818 if (DI->getParent() == DB)
4819 return false;
4820 for (const User *U : DI->users()) {
4821 auto *Usr = cast<Instruction>(U);
4822 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4823 return false;
4824 }
4825 return true;
4826}
4827
4828/// Return true when the instruction sequence within a block is select-cmp-br.
4829static bool isChainSelectCmpBranch(const SelectInst *SI) {
4830 const BasicBlock *BB = SI->getParent();
4831 if (!BB)
4832 return false;
4833 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4834 if (!BI || BI->getNumSuccessors() != 2)
4835 return false;
4836 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4837 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4838 return false;
4839 return true;
4840}
4841
4842/// True when a select result is replaced by one of its operands
4843/// in select-icmp sequence. This will eventually result in the elimination
4844/// of the select.
4845///
4846/// \param SI Select instruction
4847/// \param Icmp Compare instruction
4848/// \param SIOpd Operand that replaces the select
4849///
4850/// Notes:
4851/// - The replacement is global and requires dominator information
4852/// - The caller is responsible for the actual replacement
4853///
4854/// Example:
4855///
4856/// entry:
4857/// %4 = select i1 %3, %C* %0, %C* null
4858/// %5 = icmp eq %C* %4, null
4859/// br i1 %5, label %9, label %7
4860/// ...
4861/// ; <label>:7 ; preds = %entry
4862/// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4863/// ...
4864///
4865/// can be transformed to
4866///
4867/// %5 = icmp eq %C* %0, null
4868/// %6 = select i1 %3, i1 %5, i1 true
4869/// br i1 %6, label %9, label %7
4870/// ...
4871/// ; <label>:7 ; preds = %entry
4872/// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4873///
4874/// Similar when the first operand of the select is a constant or/and
4875/// the compare is for not equal rather than equal.
4876///
4877/// NOTE: The function is only called when the select and compare constants
4878/// are equal, the optimization can work only for EQ predicates. This is not a
4879/// major restriction since a NE compare should be 'normalized' to an equal
4880/// compare, which usually happens in the combiner and test case
4881/// select-cmp-br.ll checks for it.
4882bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
4883 const ICmpInst *Icmp,
4884 const unsigned SIOpd) {
4885 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4885, __PRETTY_FUNCTION__))
;
4886 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4887 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4888 // The check for the single predecessor is not the best that can be
4889 // done. But it protects efficiently against cases like when SI's
4890 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4891 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4892 // replaced can be reached on either path. So the uniqueness check
4893 // guarantees that the path all uses of SI (outside SI's parent) are on
4894 // is disjoint from all other paths out of SI. But that information
4895 // is more expensive to compute, and the trade-off here is in favor
4896 // of compile-time. It should also be noticed that we check for a single
4897 // predecessor and not only uniqueness. This to handle the situation when
4898 // Succ and Succ1 points to the same basic block.
4899 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4900 NumSel++;
4901 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4902 return true;
4903 }
4904 }
4905 return false;
4906}
4907
4908/// Try to fold the comparison based on range information we can get by checking
4909/// whether bits are known to be zero or one in the inputs.
4910Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4911 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4912 Type *Ty = Op0->getType();
4913 ICmpInst::Predicate Pred = I.getPredicate();
4914
4915 // Get scalar or pointer size.
4916 unsigned BitWidth = Ty->isIntOrIntVectorTy()
4917 ? Ty->getScalarSizeInBits()
4918 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
4919
4920 if (!BitWidth)
4921 return nullptr;
4922
4923 KnownBits Op0Known(BitWidth);
4924 KnownBits Op1Known(BitWidth);
4925
4926 if (SimplifyDemandedBits(&I, 0,
4927 getDemandedBitsLHSMask(I, BitWidth),
4928 Op0Known, 0))
4929 return &I;
4930
4931 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4932 Op1Known, 0))
4933 return &I;
4934
4935 // Given the known and unknown bits, compute a range that the LHS could be
4936 // in. Compute the Min, Max and RHS values based on the known bits. For the
4937 // EQ and NE we use unsigned values.
4938 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4939 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4940 if (I.isSigned()) {
4941 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4942 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4943 } else {
4944 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4945 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4946 }
4947
4948 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4949 // out that the LHS or RHS is a constant. Constant fold this now, so that
4950 // code below can assume that Min != Max.
4951 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4952 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4953 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4954 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4955
4956 // Based on the range information we know about the LHS, see if we can
4957 // simplify this comparison. For example, (x&4) < 8 is always true.
4958 switch (Pred) {
4959 default:
4960 llvm_unreachable("Unknown icmp opcode!")::llvm::llvm_unreachable_internal("Unknown icmp opcode!", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4960)
;
4961 case ICmpInst::ICMP_EQ:
4962 case ICmpInst::ICMP_NE: {
4963 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4964 return Pred == CmpInst::ICMP_EQ
4965 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4966 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4967 }
4968
4969 // If all bits are known zero except for one, then we know at most one bit
4970 // is set. If the comparison is against zero, then this is a check to see if
4971 // *that* bit is set.
4972 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4973 if (Op1Known.isZero()) {
4974 // If the LHS is an AND with the same constant, look through it.
4975 Value *LHS = nullptr;
4976 const APInt *LHSC;
4977 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4978 *LHSC != Op0KnownZeroInverted)
4979 LHS = Op0;
4980
4981 Value *X;
4982 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4983 APInt ValToCheck = Op0KnownZeroInverted;
4984 Type *XTy = X->getType();
4985 if (ValToCheck.isPowerOf2()) {
4986 // ((1 << X) & 8) == 0 -> X != 3
4987 // ((1 << X) & 8) != 0 -> X == 3
4988 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4989 auto NewPred = ICmpInst::getInversePredicate(Pred);
4990 return new ICmpInst(NewPred, X, CmpC);
4991 } else if ((++ValToCheck).isPowerOf2()) {
4992 // ((1 << X) & 7) == 0 -> X >= 3
4993 // ((1 << X) & 7) != 0 -> X < 3
4994 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4995 auto NewPred =
4996 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
4997 return new ICmpInst(NewPred, X, CmpC);
4998 }
4999 }
5000
5001 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5002 const APInt *CI;
5003 if (Op0KnownZeroInverted.isOneValue() &&
5004 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5005 // ((8 >>u X) & 1) == 0 -> X != 3
5006 // ((8 >>u X) & 1) != 0 -> X == 3
5007 unsigned CmpVal = CI->countTrailingZeros();
5008 auto NewPred = ICmpInst::getInversePredicate(Pred);
5009 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5010 }
5011 }
5012 break;
5013 }
5014 case ICmpInst::ICMP_ULT: {
5015 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5016 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5017 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5018 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5019 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5020 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5021
5022 const APInt *CmpC;
5023 if (match(Op1, m_APInt(CmpC))) {
5024 // A <u C -> A == C-1 if min(A)+1 == C
5025 if (*CmpC == Op0Min + 1)
5026 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5027 ConstantInt::get(Op1->getType(), *CmpC - 1));
5028 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5029 // exceeds the log2 of C.
5030 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5031 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5032 Constant::getNullValue(Op1->getType()));
5033 }
5034 break;
5035 }
5036 case ICmpInst::ICMP_UGT: {
5037 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5038 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5039 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5040 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5041 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5042 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5043
5044 const APInt *CmpC;
5045 if (match(Op1, m_APInt(CmpC))) {
5046 // A >u C -> A == C+1 if max(a)-1 == C
5047 if (*CmpC == Op0Max - 1)
5048 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5049 ConstantInt::get(Op1->getType(), *CmpC + 1));
5050 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5051 // exceeds the log2 of C.
5052 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5053 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5054 Constant::getNullValue(Op1->getType()));
5055 }
5056 break;
5057 }
5058 case ICmpInst::ICMP_SLT: {
5059 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5060 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5061 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5062 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5063 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5064 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5065 const APInt *CmpC;
5066 if (match(Op1, m_APInt(CmpC))) {
5067 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5068 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5069 ConstantInt::get(Op1->getType(), *CmpC - 1));
5070 }
5071 break;
5072 }
5073 case ICmpInst::ICMP_SGT: {
5074 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5075 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5076 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5077 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5078 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5079 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5080 const APInt *CmpC;
5081 if (match(Op1, m_APInt(CmpC))) {
5082 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5083 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5084 ConstantInt::get(Op1->getType(), *CmpC + 1));
5085 }
5086 break;
5087 }
5088 case ICmpInst::ICMP_SGE:
5089 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5089, __PRETTY_FUNCTION__))
;
5090 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5091 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5092 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5093 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5094 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5095 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5096 break;
5097 case ICmpInst::ICMP_SLE:
5098 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5098, __PRETTY_FUNCTION__))
;
5099 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5100 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5101 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5102 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5103 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5104 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5105 break;
5106 case ICmpInst::ICMP_UGE:
5107 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5107, __PRETTY_FUNCTION__))
;
5108 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5109 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5110 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5111 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5112 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5113 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5114 break;
5115 case ICmpInst::ICMP_ULE:
5116 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5116, __PRETTY_FUNCTION__))
;
5117 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5118 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5119 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5120 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5121 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5122 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5123 break;
5124 }
5125
5126 // Turn a signed comparison into an unsigned one if both operands are known to
5127 // have the same sign.
5128 if (I.isSigned() &&
5129 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5130 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5131 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5132
5133 return nullptr;
5134}
5135
5136llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
5137llvm::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5138 Constant *C) {
5139 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5140, __PRETTY_FUNCTION__))
5140 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5140, __PRETTY_FUNCTION__))
;
5141
5142 Type *Type = C->getType();
5143 bool IsSigned = ICmpInst::isSigned(Pred);
5144
5145 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5146 bool WillIncrement =
5147 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5148
5149 // Check if the constant operand can be safely incremented/decremented
5150 // without overflowing/underflowing.
5151 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5152 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5153 };
5154
5155 Constant *SafeReplacementConstant = nullptr;
5156 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5157 // Bail out if the constant can't be safely incremented/decremented.
5158 if (!ConstantIsOk(CI))
5159 return llvm::None;
5160 } else if (Type->isVectorTy()) {
5161 unsigned NumElts = Type->getVectorNumElements();
5162 for (unsigned i = 0; i != NumElts; ++i) {
5163 Constant *Elt = C->getAggregateElement(i);
5164 if (!Elt)
5165 return llvm::None;
5166
5167 if (isa<UndefValue>(Elt))
5168 continue;
5169
5170 // Bail out if we can't determine if this constant is min/max or if we
5171 // know that this constant is min/max.
5172 auto *CI = dyn_cast<ConstantInt>(Elt);
5173 if (!CI || !ConstantIsOk(CI))
5174 return llvm::None;
5175
5176 if (!SafeReplacementConstant)
5177 SafeReplacementConstant = CI;
5178 }
5179 } else {
5180 // ConstantExpr?
5181 return llvm::None;
5182 }
5183
5184 // It may not be safe to change a compare predicate in the presence of
5185 // undefined elements, so replace those elements with the first safe constant
5186 // that we found.
5187 if (C->containsUndefElement()) {
5188 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5188, __PRETTY_FUNCTION__))
;
5189 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5190 }
5191
5192 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5193
5194 // Increment or decrement the constant.
5195 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5196 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5197
5198 return std::make_pair(NewPred, NewC);
5199}
5200
5201/// If we have an icmp le or icmp ge instruction with a constant operand, turn
5202/// it into the appropriate icmp lt or icmp gt instruction. This transform
5203/// allows them to be folded in visitICmpInst.
5204static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5205 ICmpInst::Predicate Pred = I.getPredicate();
5206 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5207 isCanonicalPredicate(Pred))
5208 return nullptr;
5209
5210 Value *Op0 = I.getOperand(0);
5211 Value *Op1 = I.getOperand(1);
5212 auto *Op1C = dyn_cast<Constant>(Op1);
5213 if (!Op1C)
5214 return nullptr;
5215
5216 auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5217 if (!FlippedStrictness)
5218 return nullptr;
5219
5220 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5221}
5222
5223/// Integer compare with boolean values can always be turned into bitwise ops.
5224static Instruction *canonicalizeICmpBool(ICmpInst &I,
5225 InstCombiner::BuilderTy &Builder) {
5226 Value *A = I.getOperand(0), *B = I.getOperand(1);
5227 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5227, __PRETTY_FUNCTION__))
;
5228
5229 // A boolean compared to true/false can be simplified to Op0/true/false in
5230 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5231 // Cases not handled by InstSimplify are always 'not' of Op0.
5232 if (match(B, m_Zero())) {
5233 switch (I.getPredicate()) {
5234 case CmpInst::ICMP_EQ: // A == 0 -> !A
5235 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5236 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5237 return BinaryOperator::CreateNot(A);
5238 default:
5239 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5239)
;
5240 }
5241 } else if (match(B, m_One())) {
5242 switch (I.getPredicate()) {
5243 case CmpInst::ICMP_NE: // A != 1 -> !A
5244 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5245 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5246 return BinaryOperator::CreateNot(A);
5247 default:
5248 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5248)
;
5249 }
5250 }
5251
5252 switch (I.getPredicate()) {
5253 default:
5254 llvm_unreachable("Invalid icmp instruction!")::llvm::llvm_unreachable_internal("Invalid icmp instruction!"
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5254)
;
5255 case ICmpInst::ICMP_EQ:
5256 // icmp eq i1 A, B -> ~(A ^ B)
5257 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5258
5259 case ICmpInst::ICMP_NE:
5260 // icmp ne i1 A, B -> A ^ B
5261 return BinaryOperator::CreateXor(A, B);
5262
5263 case ICmpInst::ICMP_UGT:
5264 // icmp ugt -> icmp ult
5265 std::swap(A, B);
5266 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5267 case ICmpInst::ICMP_ULT:
5268 // icmp ult i1 A, B -> ~A & B
5269 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5270
5271 case ICmpInst::ICMP_SGT:
5272 // icmp sgt -> icmp slt
5273 std::swap(A, B);
5274 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5275 case ICmpInst::ICMP_SLT:
5276 // icmp slt i1 A, B -> A & ~B
5277 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5278
5279 case ICmpInst::ICMP_UGE:
5280 // icmp uge -> icmp ule
5281 std::swap(A, B);
5282 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5283 case ICmpInst::ICMP_ULE:
5284 // icmp ule i1 A, B -> ~A | B
5285 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5286
5287 case ICmpInst::ICMP_SGE:
5288 // icmp sge -> icmp sle
5289 std::swap(A, B);
5290 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5291 case ICmpInst::ICMP_SLE:
5292 // icmp sle i1 A, B -> A | ~B
5293 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5294 }
5295}
5296
5297// Transform pattern like:
5298// (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5299// (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5300// Into:
5301// (X l>> Y) != 0
5302// (X l>> Y) == 0
5303static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5304 InstCombiner::BuilderTy &Builder) {
5305 ICmpInst::Predicate Pred, NewPred;
5306 Value *X, *Y;
5307 if (match(&Cmp,
5308 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5309 switch (Pred) {
5310 case ICmpInst::ICMP_ULE:
5311 NewPred = ICmpInst::ICMP_NE;
5312 break;
5313 case ICmpInst::ICMP_UGT:
5314 NewPred = ICmpInst::ICMP_EQ;
5315 break;
5316 default:
5317 return nullptr;
5318 }
5319 } else if (match(&Cmp, m_c_ICmp(Pred,
5320 m_OneUse(m_CombineOr(
5321 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5322 m_Add(m_Shl(m_One(), m_Value(Y)),
5323 m_AllOnes()))),
5324 m_Value(X)))) {
5325 // The variant with 'add' is not canonical, (the variant with 'not' is)
5326 // we only get it because it has extra uses, and can't be canonicalized,
5327
5328 switch (Pred) {
5329 case ICmpInst::ICMP_ULT:
5330 NewPred = ICmpInst::ICMP_NE;
5331 break;
5332 case ICmpInst::ICMP_UGE:
5333 NewPred = ICmpInst::ICMP_EQ;
5334 break;
5335 default:
5336 return nullptr;
5337 }
5338 } else
5339 return nullptr;
5340
5341 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5342 Constant *Zero = Constant::getNullValue(NewX->getType());
5343 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5344}
5345
5346static Instruction *foldVectorCmp(CmpInst &Cmp,
5347 InstCombiner::BuilderTy &Builder) {
5348 const CmpInst::Predicate Pred = Cmp.getPredicate();
5349 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5350 bool IsFP = isa<FCmpInst>(Cmp);
5351
5352 Value *V1, *V2;
5353 Constant *M;
5354 if (!match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))))
5355 return nullptr;
5356
5357 // If both arguments of the cmp are shuffles that use the same mask and
5358 // shuffle within a single vector, move the shuffle after the cmp:
5359 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5360 Type *V1Ty = V1->getType();
5361 if (match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
5362 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
5363 Value *NewCmp = IsFP ? Builder.CreateFCmp(Pred, V1, V2)
5364 : Builder.CreateICmp(Pred, V1, V2);
5365 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5366 }
5367
5368 // Try to canonicalize compare with splatted operand and splat constant.
5369 // TODO: We could generalize this for more than splats. See/use the code in
5370 // InstCombiner::foldVectorBinop().
5371 Constant *C;
5372 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
5373 return nullptr;
5374
5375 // Length-changing splats are ok, so adjust the constants as needed:
5376 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
5377 Constant *ScalarC = C->getSplatValue(/* AllowUndefs */ true);
5378 Constant *ScalarM = M->getSplatValue(/* AllowUndefs */ true);
5379 if (ScalarC && ScalarM) {
5380 // We allow undefs in matching, but this transform removes those for safety.
5381 // Demanded elements analysis should be able to recover some/all of that.
5382 C = ConstantVector::getSplat(V1Ty->getVectorNumElements(), ScalarC);
5383 M = ConstantVector::getSplat(M->getType()->getVectorNumElements(), ScalarM);
5384 Value *NewCmp = IsFP ? Builder.CreateFCmp(Pred, V1, C)
5385 : Builder.CreateICmp(Pred, V1, C);
5386 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5387 }
5388
5389 return nullptr;
5390}
5391
5392// extract(uadd.with.overflow(A, B), 0) ult A
5393// -> extract(uadd.with.overflow(A, B), 1)
5394static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5395 CmpInst::Predicate Pred = I.getPredicate();
5396 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5397
5398 Value *UAddOv;
5399 Value *A, *B;
5400 auto UAddOvResultPat = m_ExtractValue<0>(
5401 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5402 if (match(Op0, UAddOvResultPat) &&
5403 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5404 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5405 (match(A, m_One()) || match(B, m_One()))) ||
5406 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5407 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5408 // extract(uadd.with.overflow(A, B), 0) < A
5409 // extract(uadd.with.overflow(A, 1), 0) == 0
5410 // extract(uadd.with.overflow(A, -1), 0) != -1
5411 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5412 else if (match(Op1, UAddOvResultPat) &&
5413 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5414 // A > extract(uadd.with.overflow(A, B), 0)
5415 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5416 else
5417 return nullptr;
5418
5419 return ExtractValueInst::Create(UAddOv, 1);
5420}
5421
5422Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5423 bool Changed = false;
5424 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5425 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5426 unsigned Op0Cplxity = getComplexity(Op0);
5427 unsigned Op1Cplxity = getComplexity(Op1);
5428
5429 /// Orders the operands of the compare so that they are listed from most
5430 /// complex to least complex. This puts constants before unary operators,
5431 /// before binary operators.
5432 if (Op0Cplxity < Op1Cplxity ||
5433 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5434 I.swapOperands();
5435 std::swap(Op0, Op1);
5436 Changed = true;
5437 }
5438
5439 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, Q))
5440 return replaceInstUsesWith(I, V);
5441
5442 // Comparing -val or val with non-zero is the same as just comparing val
5443 // ie, abs(val) != 0 -> val != 0
5444 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5445 Value *Cond, *SelectTrue, *SelectFalse;
5446 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5447 m_Value(SelectFalse)))) {
5448 if (Value *V = dyn_castNegVal(SelectTrue)) {
5449 if (V == SelectFalse)
5450 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5451 }
5452 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5453 if (V == SelectTrue)
5454 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5455 }
5456 }
5457 }
5458
5459 if (Op0->getType()->isIntOrIntVectorTy(1))
5460 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
5461 return Res;
5462
5463 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
5464 return NewICmp;
5465
5466 if (Instruction *Res = foldICmpWithConstant(I))
5467 return Res;
5468
5469 if (Instruction *Res = foldICmpWithDominatingICmp(I))
5470 return Res;
5471
5472 if (Instruction *Res = foldICmpBinOp(I, Q))
5473 return Res;
5474
5475 if (Instruction *Res = foldICmpUsingKnownBits(I))
5476 return Res;
5477
5478 // Test if the ICmpInst instruction is used exclusively by a select as
5479 // part of a minimum or maximum operation. If so, refrain from doing
5480 // any other folding. This helps out other analyses which understand
5481 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5482 // and CodeGen. And in this case, at least one of the comparison
5483 // operands has at least one user besides the compare (the select),
5484 // which would often largely negate the benefit of folding anyway.
5485 //
5486 // Do the same for the other patterns recognized by matchSelectPattern.
5487 if (I.hasOneUse())
5488 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5489 Value *A, *B;
5490 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5491 if (SPR.Flavor != SPF_UNKNOWN)
5492 return nullptr;
5493 }
5494
5495 // Do this after checking for min/max to prevent infinite looping.
5496 if (Instruction *Res = foldICmpWithZero(I))
5497 return Res;
5498
5499 // FIXME: We only do this after checking for min/max to prevent infinite
5500 // looping caused by a reverse canonicalization of these patterns for min/max.
5501 // FIXME: The organization of folds is a mess. These would naturally go into
5502 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
5503 // down here after the min/max restriction.
5504 ICmpInst::Predicate Pred = I.getPredicate();
5505 const APInt *C;
5506 if (match(Op1, m_APInt(C))) {
5507 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
5508 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
5509 Constant *Zero = Constant::getNullValue(Op0->getType());
5510 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
5511 }
5512
5513 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
5514 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
5515 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
5516 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
5517 }
5518 }
5519
5520 if (Instruction *Res = foldICmpInstWithConstant(I))
5521 return Res;
5522
5523 // Try to match comparison as a sign bit test. Intentionally do this after
5524 // foldICmpInstWithConstant() to potentially let other folds to happen first.
5525 if (Instruction *New = foldSignBitTest(I))
5526 return New;
5527
5528 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
5529 return Res;
5530
5531 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5532 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
5533 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
5534 return NI;
5535 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
5536 if (Instruction *NI = foldGEPICmp(GEP, Op0,
5537 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5538 return NI;
5539
5540 // Try to optimize equality comparisons against alloca-based pointers.
5541 if (Op0->getType()->isPointerTy() && I.isEquality()) {
5542 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?")((Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?"
) ? static_cast<void> (0) : __assert_fail ("Op1->getType()->isPointerTy() && \"Comparing pointer with non-pointer?\""
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5542, __PRETTY_FUNCTION__))
;
5543 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
5544 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
5545 return New;
5546 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
5547 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
5548 return New;
5549 }
5550
5551 if (Instruction *Res = foldICmpBitCast(I, Builder))
5552 return Res;
5553
5554 if (Instruction *R = foldICmpWithCastOp(I))
5555 return R;
5556
5557 if (Instruction *Res = foldICmpWithMinMax(I))
5558 return Res;
5559
5560 {
5561 Value *A, *B;
5562 // Transform (A & ~B) == 0 --> (A & B) != 0
5563 // and (A & ~B) != 0 --> (A & B) == 0
5564 // if A is a power of 2.
5565 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
5566 match(Op1, m_Zero()) &&
5567 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
5568 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
5569 Op1);
5570
5571 // ~X < ~Y --> Y < X
5572 // ~X < C --> X > ~C
5573 if (match(Op0, m_Not(m_Value(A)))) {
5574 if (match(Op1, m_Not(m_Value(B))))
5575 return new ICmpInst(I.getPredicate(), B, A);
5576
5577 const APInt *C;
5578 if (match(Op1, m_APInt(C)))
5579 return new ICmpInst(I.getSwappedPredicate(), A,
5580 ConstantInt::get(Op1->getType(), ~(*C)));
5581 }
5582
5583 Instruction *AddI = nullptr;
5584 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
5585 m_Instruction(AddI))) &&
5586 isa<IntegerType>(A->getType())) {
5587 Value *Result;
5588 Constant *Overflow;
5589 // m_UAddWithOverflow can match patterns that do not include an explicit
5590 // "add" instruction, so check the opcode of the matched op.
5591 if (AddI->getOpcode() == Instruction::Add &&
5592 OptimizeOverflowCheck(Instruction::Add, /*Signed*/ false, A, B, *AddI,
5593 Result, Overflow)) {
5594 replaceInstUsesWith(*AddI, Result);
5595 return replaceInstUsesWith(I, Overflow);
5596 }
5597 }
5598
5599 // (zext a) * (zext b) --> llvm.umul.with.overflow.
5600 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5601 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
5602 return R;
5603 }
5604 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5605 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
5606 return R;
5607 }