Bug Summary

File:llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
Warning:line 3615, column 7
Called C++ object pointer is uninitialized

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name InstCombineCompares.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/build-llvm/lib/Transforms/InstCombine -resource-dir /usr/lib/llvm-13/lib/clang/13.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/build-llvm/include -I /build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/include -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-13/lib/clang/13.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/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-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-06-21-164211-33944-1 -x c++ /build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

1//===- InstCombineCompares.cpp --------------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visitICmp and visitFCmp functions.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APSInt.h"
15#include "llvm/ADT/SetVector.h"
16#include "llvm/ADT/Statistic.h"
17#include "llvm/Analysis/ConstantFolding.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/TargetLibraryInfo.h"
20#include "llvm/IR/ConstantRange.h"
21#include "llvm/IR/DataLayout.h"
22#include "llvm/IR/GetElementPtrTypeIterator.h"
23#include "llvm/IR/IntrinsicInst.h"
24#include "llvm/IR/PatternMatch.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/KnownBits.h"
27#include "llvm/Transforms/InstCombine/InstCombiner.h"
28
29using namespace llvm;
30using namespace PatternMatch;
31
32#define DEBUG_TYPE"instcombine" "instcombine"
33
34// How many times is a select replaced by one of its operands?
35STATISTIC(NumSel, "Number of select opts")static llvm::Statistic NumSel = {"instcombine", "NumSel", "Number of select opts"
}
;
36
37
38/// Compute Result = In1+In2, returning true if the result overflowed for this
39/// type.
40static bool addWithOverflow(APInt &Result, const APInt &In1,
41 const APInt &In2, bool IsSigned = false) {
42 bool Overflow;
43 if (IsSigned)
44 Result = In1.sadd_ov(In2, Overflow);
45 else
46 Result = In1.uadd_ov(In2, Overflow);
47
48 return Overflow;
49}
50
51/// Compute Result = In1-In2, returning true if the result overflowed for this
52/// type.
53static bool subWithOverflow(APInt &Result, const APInt &In1,
54 const APInt &In2, bool IsSigned = false) {
55 bool Overflow;
56 if (IsSigned)
57 Result = In1.ssub_ov(In2, Overflow);
58 else
59 Result = In1.usub_ov(In2, Overflow);
60
61 return Overflow;
62}
63
64/// Given an icmp instruction, return true if any use of this comparison is a
65/// branch on sign bit comparison.
66static bool hasBranchUse(ICmpInst &I) {
67 for (auto *U : I.users())
68 if (isa<BranchInst>(U))
69 return true;
70 return false;
71}
72
73/// Returns true if the exploded icmp can be expressed as a signed comparison
74/// to zero and updates the predicate accordingly.
75/// The signedness of the comparison is preserved.
76/// TODO: Refactor with decomposeBitTestICmp()?
77static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
78 if (!ICmpInst::isSigned(Pred))
79 return false;
80
81 if (C.isNullValue())
82 return ICmpInst::isRelational(Pred);
83
84 if (C.isOneValue()) {
85 if (Pred == ICmpInst::ICMP_SLT) {
86 Pred = ICmpInst::ICMP_SLE;
87 return true;
88 }
89 } else if (C.isAllOnesValue()) {
90 if (Pred == ICmpInst::ICMP_SGT) {
91 Pred = ICmpInst::ICMP_SGE;
92 return true;
93 }
94 }
95
96 return false;
97}
98
99/// This is called when we see this pattern:
100/// cmp pred (load (gep GV, ...)), cmpcst
101/// where GV is a global variable with a constant initializer. Try to simplify
102/// this into some simple computation that does not need the load. For example
103/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
104///
105/// If AndCst is non-null, then the loaded value is masked with that constant
106/// before doing the comparison. This handles cases like "A[i]&4 == 0".
107Instruction *
108InstCombinerImpl::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
109 GlobalVariable *GV, CmpInst &ICI,
110 ConstantInt *AndCst) {
111 Constant *Init = GV->getInitializer();
112 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
113 return nullptr;
114
115 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
116 // Don't blow up on huge arrays.
117 if (ArrayElementCount > MaxArraySizeForCombine)
118 return nullptr;
119
120 // There are many forms of this optimization we can handle, for now, just do
121 // the simple index into a single-dimensional array.
122 //
123 // Require: GEP GV, 0, i {{, constant indices}}
124 if (GEP->getNumOperands() < 3 ||
125 !isa<ConstantInt>(GEP->getOperand(1)) ||
126 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
127 isa<Constant>(GEP->getOperand(2)))
128 return nullptr;
129
130 // Check that indices after the variable are constants and in-range for the
131 // type they index. Collect the indices. This is typically for arrays of
132 // structs.
133 SmallVector<unsigned, 4> LaterIndices;
134
135 Type *EltTy = Init->getType()->getArrayElementType();
136 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
137 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
138 if (!Idx) return nullptr; // Variable index.
139
140 uint64_t IdxVal = Idx->getZExtValue();
141 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
142
143 if (StructType *STy = dyn_cast<StructType>(EltTy))
144 EltTy = STy->getElementType(IdxVal);
145 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
146 if (IdxVal >= ATy->getNumElements()) return nullptr;
147 EltTy = ATy->getElementType();
148 } else {
149 return nullptr; // Unknown type.
150 }
151
152 LaterIndices.push_back(IdxVal);
153 }
154
155 enum { Overdefined = -3, Undefined = -2 };
156
157 // Variables for our state machines.
158
159 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
160 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
161 // and 87 is the second (and last) index. FirstTrueElement is -2 when
162 // undefined, otherwise set to the first true element. SecondTrueElement is
163 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
164 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
165
166 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
167 // form "i != 47 & i != 87". Same state transitions as for true elements.
168 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
169
170 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
171 /// define a state machine that triggers for ranges of values that the index
172 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
173 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
174 /// index in the range (inclusive). We use -2 for undefined here because we
175 /// use relative comparisons and don't want 0-1 to match -1.
176 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
177
178 // MagicBitvector - This is a magic bitvector where we set a bit if the
179 // comparison is true for element 'i'. If there are 64 elements or less in
180 // the array, this will fully represent all the comparison results.
181 uint64_t MagicBitvector = 0;
182
183 // Scan the array and see if one of our patterns matches.
184 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
185 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
186 Constant *Elt = Init->getAggregateElement(i);
187 if (!Elt) return nullptr;
188
189 // If this is indexing an array of structures, get the structure element.
190 if (!LaterIndices.empty())
191 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
192
193 // If the element is masked, handle it.
194 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
195
196 // Find out if the comparison would be true or false for the i'th element.
197 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
198 CompareRHS, DL, &TLI);
199 // If the result is undef for this element, ignore it.
200 if (isa<UndefValue>(C)) {
201 // Extend range state machines to cover this element in case there is an
202 // undef in the middle of the range.
203 if (TrueRangeEnd == (int)i-1)
204 TrueRangeEnd = i;
205 if (FalseRangeEnd == (int)i-1)
206 FalseRangeEnd = i;
207 continue;
208 }
209
210 // If we can't compute the result for any of the elements, we have to give
211 // up evaluating the entire conditional.
212 if (!isa<ConstantInt>(C)) return nullptr;
213
214 // Otherwise, we know if the comparison is true or false for this element,
215 // update our state machines.
216 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
217
218 // State machine for single/double/range index comparison.
219 if (IsTrueForElt) {
220 // Update the TrueElement state machine.
221 if (FirstTrueElement == Undefined)
222 FirstTrueElement = TrueRangeEnd = i; // First true element.
223 else {
224 // Update double-compare state machine.
225 if (SecondTrueElement == Undefined)
226 SecondTrueElement = i;
227 else
228 SecondTrueElement = Overdefined;
229
230 // Update range state machine.
231 if (TrueRangeEnd == (int)i-1)
232 TrueRangeEnd = i;
233 else
234 TrueRangeEnd = Overdefined;
235 }
236 } else {
237 // Update the FalseElement state machine.
238 if (FirstFalseElement == Undefined)
239 FirstFalseElement = FalseRangeEnd = i; // First false element.
240 else {
241 // Update double-compare state machine.
242 if (SecondFalseElement == Undefined)
243 SecondFalseElement = i;
244 else
245 SecondFalseElement = Overdefined;
246
247 // Update range state machine.
248 if (FalseRangeEnd == (int)i-1)
249 FalseRangeEnd = i;
250 else
251 FalseRangeEnd = Overdefined;
252 }
253 }
254
255 // If this element is in range, update our magic bitvector.
256 if (i < 64 && IsTrueForElt)
257 MagicBitvector |= 1ULL << i;
258
259 // If all of our states become overdefined, bail out early. Since the
260 // predicate is expensive, only check it every 8 elements. This is only
261 // really useful for really huge arrays.
262 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
263 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
264 FalseRangeEnd == Overdefined)
265 return nullptr;
266 }
267
268 // Now that we've scanned the entire array, emit our new comparison(s). We
269 // order the state machines in complexity of the generated code.
270 Value *Idx = GEP->getOperand(2);
271
272 // If the index is larger than the pointer size of the target, truncate the
273 // index down like the GEP would do implicitly. We don't have to do this for
274 // an inbounds GEP because the index can't be out of range.
275 if (!GEP->isInBounds()) {
276 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
277 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
278 if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
279 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
280 }
281
282 // If inbounds keyword is not present, Idx * ElementSize can overflow.
283 // Let's assume that ElementSize is 2 and the wanted value is at offset 0.
284 // Then, there are two possible values for Idx to match offset 0:
285 // 0x00..00, 0x80..00.
286 // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
287 // comparison is false if Idx was 0x80..00.
288 // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
289 unsigned ElementSize =
290 DL.getTypeAllocSize(Init->getType()->getArrayElementType());
291 auto MaskIdx = [&](Value* Idx){
292 if (!GEP->isInBounds() && countTrailingZeros(ElementSize) != 0) {
293 Value *Mask = ConstantInt::get(Idx->getType(), -1);
294 Mask = Builder.CreateLShr(Mask, countTrailingZeros(ElementSize));
295 Idx = Builder.CreateAnd(Idx, Mask);
296 }
297 return Idx;
298 };
299
300 // If the comparison is only true for one or two elements, emit direct
301 // comparisons.
302 if (SecondTrueElement != Overdefined) {
303 Idx = MaskIdx(Idx);
304 // None true -> false.
305 if (FirstTrueElement == Undefined)
306 return replaceInstUsesWith(ICI, Builder.getFalse());
307
308 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
309
310 // True for one element -> 'i == 47'.
311 if (SecondTrueElement == Undefined)
312 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
313
314 // True for two elements -> 'i == 47 | i == 72'.
315 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
316 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
317 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
318 return BinaryOperator::CreateOr(C1, C2);
319 }
320
321 // If the comparison is only false for one or two elements, emit direct
322 // comparisons.
323 if (SecondFalseElement != Overdefined) {
324 Idx = MaskIdx(Idx);
325 // None false -> true.
326 if (FirstFalseElement == Undefined)
327 return replaceInstUsesWith(ICI, Builder.getTrue());
328
329 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
330
331 // False for one element -> 'i != 47'.
332 if (SecondFalseElement == Undefined)
333 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
334
335 // False for two elements -> 'i != 47 & i != 72'.
336 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
337 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
338 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
339 return BinaryOperator::CreateAnd(C1, C2);
340 }
341
342 // If the comparison can be replaced with a range comparison for the elements
343 // where it is true, emit the range check.
344 if (TrueRangeEnd != Overdefined) {
345 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare")(static_cast <bool> (TrueRangeEnd != FirstTrueElement &&
"Should emit single compare") ? void (0) : __assert_fail ("TrueRangeEnd != FirstTrueElement && \"Should emit single compare\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 345, __extension__ __PRETTY_FUNCTION__))
;
346 Idx = MaskIdx(Idx);
347
348 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
349 if (FirstTrueElement) {
350 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
351 Idx = Builder.CreateAdd(Idx, Offs);
352 }
353
354 Value *End = ConstantInt::get(Idx->getType(),
355 TrueRangeEnd-FirstTrueElement+1);
356 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
357 }
358
359 // False range check.
360 if (FalseRangeEnd != Overdefined) {
361 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare")(static_cast <bool> (FalseRangeEnd != FirstFalseElement
&& "Should emit single compare") ? void (0) : __assert_fail
("FalseRangeEnd != FirstFalseElement && \"Should emit single compare\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 361, __extension__ __PRETTY_FUNCTION__))
;
362 Idx = MaskIdx(Idx);
363 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
364 if (FirstFalseElement) {
365 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
366 Idx = Builder.CreateAdd(Idx, Offs);
367 }
368
369 Value *End = ConstantInt::get(Idx->getType(),
370 FalseRangeEnd-FirstFalseElement);
371 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
372 }
373
374 // If a magic bitvector captures the entire comparison state
375 // of this load, replace it with computation that does:
376 // ((magic_cst >> i) & 1) != 0
377 {
378 Type *Ty = nullptr;
379
380 // Look for an appropriate type:
381 // - The type of Idx if the magic fits
382 // - The smallest fitting legal type
383 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
384 Ty = Idx->getType();
385 else
386 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
387
388 if (Ty) {
389 Idx = MaskIdx(Idx);
390 Value *V = Builder.CreateIntCast(Idx, Ty, false);
391 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
392 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
393 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
394 }
395 }
396
397 return nullptr;
398}
399
400/// Return a value that can be used to compare the *offset* implied by a GEP to
401/// zero. For example, if we have &A[i], we want to return 'i' for
402/// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
403/// are involved. The above expression would also be legal to codegen as
404/// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
405/// This latter form is less amenable to optimization though, and we are allowed
406/// to generate the first by knowing that pointer arithmetic doesn't overflow.
407///
408/// If we can't emit an optimized form for this expression, this returns null.
409///
410static Value *evaluateGEPOffsetExpression(User *GEP, InstCombinerImpl &IC,
411 const DataLayout &DL) {
412 gep_type_iterator GTI = gep_type_begin(GEP);
413
414 // Check to see if this gep only has a single variable index. If so, and if
415 // any constant indices are a multiple of its scale, then we can compute this
416 // in terms of the scale of the variable index. For example, if the GEP
417 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
418 // because the expression will cross zero at the same point.
419 unsigned i, e = GEP->getNumOperands();
420 int64_t Offset = 0;
421 for (i = 1; i != e; ++i, ++GTI) {
422 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
423 // Compute the aggregate offset of constant indices.
424 if (CI->isZero()) continue;
425
426 // Handle a struct index, which adds its field offset to the pointer.
427 if (StructType *STy = GTI.getStructTypeOrNull()) {
428 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
429 } else {
430 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
431 Offset += Size*CI->getSExtValue();
432 }
433 } else {
434 // Found our variable index.
435 break;
436 }
437 }
438
439 // If there are no variable indices, we must have a constant offset, just
440 // evaluate it the general way.
441 if (i == e) return nullptr;
442
443 Value *VariableIdx = GEP->getOperand(i);
444 // Determine the scale factor of the variable element. For example, this is
445 // 4 if the variable index is into an array of i32.
446 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
447
448 // Verify that there are no other variable indices. If so, emit the hard way.
449 for (++i, ++GTI; i != e; ++i, ++GTI) {
450 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
451 if (!CI) return nullptr;
452
453 // Compute the aggregate offset of constant indices.
454 if (CI->isZero()) continue;
455
456 // Handle a struct index, which adds its field offset to the pointer.
457 if (StructType *STy = GTI.getStructTypeOrNull()) {
458 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
459 } else {
460 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
461 Offset += Size*CI->getSExtValue();
462 }
463 }
464
465 // Okay, we know we have a single variable index, which must be a
466 // pointer/array/vector index. If there is no offset, life is simple, return
467 // the index.
468 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
469 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
470 if (Offset == 0) {
471 // Cast to intptrty in case a truncation occurs. If an extension is needed,
472 // we don't need to bother extending: the extension won't affect where the
473 // computation crosses zero.
474 if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
475 IntPtrWidth) {
476 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
477 }
478 return VariableIdx;
479 }
480
481 // Otherwise, there is an index. The computation we will do will be modulo
482 // the pointer size.
483 Offset = SignExtend64(Offset, IntPtrWidth);
484 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
485
486 // To do this transformation, any constant index must be a multiple of the
487 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
488 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
489 // multiple of the variable scale.
490 int64_t NewOffs = Offset / (int64_t)VariableScale;
491 if (Offset != NewOffs*(int64_t)VariableScale)
492 return nullptr;
493
494 // Okay, we can do this evaluation. Start by converting the index to intptr.
495 if (VariableIdx->getType() != IntPtrTy)
496 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
497 true /*Signed*/);
498 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
499 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
500}
501
502/// Returns true if we can rewrite Start as a GEP with pointer Base
503/// and some integer offset. The nodes that need to be re-written
504/// for this transformation will be added to Explored.
505static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
506 const DataLayout &DL,
507 SetVector<Value *> &Explored) {
508 SmallVector<Value *, 16> WorkList(1, Start);
509 Explored.insert(Base);
510
511 // The following traversal gives us an order which can be used
512 // when doing the final transformation. Since in the final
513 // transformation we create the PHI replacement instructions first,
514 // we don't have to get them in any particular order.
515 //
516 // However, for other instructions we will have to traverse the
517 // operands of an instruction first, which means that we have to
518 // do a post-order traversal.
519 while (!WorkList.empty()) {
520 SetVector<PHINode *> PHIs;
521
522 while (!WorkList.empty()) {
523 if (Explored.size() >= 100)
524 return false;
525
526 Value *V = WorkList.back();
527
528 if (Explored.contains(V)) {
529 WorkList.pop_back();
530 continue;
531 }
532
533 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
534 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
535 // We've found some value that we can't explore which is different from
536 // the base. Therefore we can't do this transformation.
537 return false;
538
539 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
540 auto *CI = cast<CastInst>(V);
541 if (!CI->isNoopCast(DL))
542 return false;
543
544 if (Explored.count(CI->getOperand(0)) == 0)
545 WorkList.push_back(CI->getOperand(0));
546 }
547
548 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
549 // We're limiting the GEP to having one index. This will preserve
550 // the original pointer type. We could handle more cases in the
551 // future.
552 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
553 GEP->getType() != Start->getType())
554 return false;
555
556 if (Explored.count(GEP->getOperand(0)) == 0)
557 WorkList.push_back(GEP->getOperand(0));
558 }
559
560 if (WorkList.back() == V) {
561 WorkList.pop_back();
562 // We've finished visiting this node, mark it as such.
563 Explored.insert(V);
564 }
565
566 if (auto *PN = dyn_cast<PHINode>(V)) {
567 // We cannot transform PHIs on unsplittable basic blocks.
568 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
569 return false;
570 Explored.insert(PN);
571 PHIs.insert(PN);
572 }
573 }
574
575 // Explore the PHI nodes further.
576 for (auto *PN : PHIs)
577 for (Value *Op : PN->incoming_values())
578 if (Explored.count(Op) == 0)
579 WorkList.push_back(Op);
580 }
581
582 // Make sure that we can do this. Since we can't insert GEPs in a basic
583 // block before a PHI node, we can't easily do this transformation if
584 // we have PHI node users of transformed instructions.
585 for (Value *Val : Explored) {
586 for (Value *Use : Val->uses()) {
587
588 auto *PHI = dyn_cast<PHINode>(Use);
589 auto *Inst = dyn_cast<Instruction>(Val);
590
591 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
592 Explored.count(PHI) == 0)
593 continue;
594
595 if (PHI->getParent() == Inst->getParent())
596 return false;
597 }
598 }
599 return true;
600}
601
602// Sets the appropriate insert point on Builder where we can add
603// a replacement Instruction for V (if that is possible).
604static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
605 bool Before = true) {
606 if (auto *PHI = dyn_cast<PHINode>(V)) {
607 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
608 return;
609 }
610 if (auto *I = dyn_cast<Instruction>(V)) {
611 if (!Before)
612 I = &*std::next(I->getIterator());
613 Builder.SetInsertPoint(I);
614 return;
615 }
616 if (auto *A = dyn_cast<Argument>(V)) {
617 // Set the insertion point in the entry block.
618 BasicBlock &Entry = A->getParent()->getEntryBlock();
619 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
620 return;
621 }
622 // Otherwise, this is a constant and we don't need to set a new
623 // insertion point.
624 assert(isa<Constant>(V) && "Setting insertion point for unknown value!")(static_cast <bool> (isa<Constant>(V) && "Setting insertion point for unknown value!"
) ? void (0) : __assert_fail ("isa<Constant>(V) && \"Setting insertion point for unknown value!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 624, __extension__ __PRETTY_FUNCTION__))
;
625}
626
627/// Returns a re-written value of Start as an indexed GEP using Base as a
628/// pointer.
629static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
630 const DataLayout &DL,
631 SetVector<Value *> &Explored) {
632 // Perform all the substitutions. This is a bit tricky because we can
633 // have cycles in our use-def chains.
634 // 1. Create the PHI nodes without any incoming values.
635 // 2. Create all the other values.
636 // 3. Add the edges for the PHI nodes.
637 // 4. Emit GEPs to get the original pointers.
638 // 5. Remove the original instructions.
639 Type *IndexType = IntegerType::get(
640 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
641
642 DenseMap<Value *, Value *> NewInsts;
643 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
644
645 // Create the new PHI nodes, without adding any incoming values.
646 for (Value *Val : Explored) {
647 if (Val == Base)
648 continue;
649 // Create empty phi nodes. This avoids cyclic dependencies when creating
650 // the remaining instructions.
651 if (auto *PHI = dyn_cast<PHINode>(Val))
652 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
653 PHI->getName() + ".idx", PHI);
654 }
655 IRBuilder<> Builder(Base->getContext());
656
657 // Create all the other instructions.
658 for (Value *Val : Explored) {
659
660 if (NewInsts.find(Val) != NewInsts.end())
661 continue;
662
663 if (auto *CI = dyn_cast<CastInst>(Val)) {
664 // Don't get rid of the intermediate variable here; the store can grow
665 // the map which will invalidate the reference to the input value.
666 Value *V = NewInsts[CI->getOperand(0)];
667 NewInsts[CI] = V;
668 continue;
669 }
670 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
671 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
672 : GEP->getOperand(1);
673 setInsertionPoint(Builder, GEP);
674 // Indices might need to be sign extended. GEPs will magically do
675 // this, but we need to do it ourselves here.
676 if (Index->getType()->getScalarSizeInBits() !=
677 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
678 Index = Builder.CreateSExtOrTrunc(
679 Index, NewInsts[GEP->getOperand(0)]->getType(),
680 GEP->getOperand(0)->getName() + ".sext");
681 }
682
683 auto *Op = NewInsts[GEP->getOperand(0)];
684 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
685 NewInsts[GEP] = Index;
686 else
687 NewInsts[GEP] = Builder.CreateNSWAdd(
688 Op, Index, GEP->getOperand(0)->getName() + ".add");
689 continue;
690 }
691 if (isa<PHINode>(Val))
692 continue;
693
694 llvm_unreachable("Unexpected instruction type")::llvm::llvm_unreachable_internal("Unexpected instruction type"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 694)
;
695 }
696
697 // Add the incoming values to the PHI nodes.
698 for (Value *Val : Explored) {
699 if (Val == Base)
700 continue;
701 // All the instructions have been created, we can now add edges to the
702 // phi nodes.
703 if (auto *PHI = dyn_cast<PHINode>(Val)) {
704 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
705 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
706 Value *NewIncoming = PHI->getIncomingValue(I);
707
708 if (NewInsts.find(NewIncoming) != NewInsts.end())
709 NewIncoming = NewInsts[NewIncoming];
710
711 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
712 }
713 }
714 }
715
716 for (Value *Val : Explored) {
717 if (Val == Base)
718 continue;
719
720 // Depending on the type, for external users we have to emit
721 // a GEP or a GEP + ptrtoint.
722 setInsertionPoint(Builder, Val, false);
723
724 // If required, create an inttoptr instruction for Base.
725 Value *NewBase = Base;
726 if (!Base->getType()->isPointerTy())
727 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
728 Start->getName() + "to.ptr");
729
730 Value *GEP = Builder.CreateInBoundsGEP(
731 Start->getType()->getPointerElementType(), NewBase,
732 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
733
734 if (!Val->getType()->isPointerTy()) {
735 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
736 Val->getName() + ".conv");
737 GEP = Cast;
738 }
739 Val->replaceAllUsesWith(GEP);
740 }
741
742 return NewInsts[Start];
743}
744
745/// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
746/// the input Value as a constant indexed GEP. Returns a pair containing
747/// the GEPs Pointer and Index.
748static std::pair<Value *, Value *>
749getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
750 Type *IndexType = IntegerType::get(V->getContext(),
751 DL.getIndexTypeSizeInBits(V->getType()));
752
753 Constant *Index = ConstantInt::getNullValue(IndexType);
754 while (true) {
755 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
756 // We accept only inbouds GEPs here to exclude the possibility of
757 // overflow.
758 if (!GEP->isInBounds())
759 break;
760 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
761 GEP->getType() == V->getType()) {
762 V = GEP->getOperand(0);
763 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
764 Index = ConstantExpr::getAdd(
765 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
766 continue;
767 }
768 break;
769 }
770 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
771 if (!CI->isNoopCast(DL))
772 break;
773 V = CI->getOperand(0);
774 continue;
775 }
776 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
777 if (!CI->isNoopCast(DL))
778 break;
779 V = CI->getOperand(0);
780 continue;
781 }
782 break;
783 }
784 return {V, Index};
785}
786
787/// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
788/// We can look through PHIs, GEPs and casts in order to determine a common base
789/// between GEPLHS and RHS.
790static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
791 ICmpInst::Predicate Cond,
792 const DataLayout &DL) {
793 // FIXME: Support vector of pointers.
794 if (GEPLHS->getType()->isVectorTy())
795 return nullptr;
796
797 if (!GEPLHS->hasAllConstantIndices())
798 return nullptr;
799
800 // Make sure the pointers have the same type.
801 if (GEPLHS->getType() != RHS->getType())
802 return nullptr;
803
804 Value *PtrBase, *Index;
805 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
806
807 // The set of nodes that will take part in this transformation.
808 SetVector<Value *> Nodes;
809
810 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
811 return nullptr;
812
813 // We know we can re-write this as
814 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
815 // Since we've only looked through inbouds GEPs we know that we
816 // can't have overflow on either side. We can therefore re-write
817 // this as:
818 // OFFSET1 cmp OFFSET2
819 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
820
821 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
822 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
823 // offset. Since Index is the offset of LHS to the base pointer, we will now
824 // compare the offsets instead of comparing the pointers.
825 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
826}
827
828/// Fold comparisons between a GEP instruction and something else. At this point
829/// we know that the GEP is on the LHS of the comparison.
830Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
831 ICmpInst::Predicate Cond,
832 Instruction &I) {
833 // Don't transform signed compares of GEPs into index compares. Even if the
834 // GEP is inbounds, the final add of the base pointer can have signed overflow
835 // and would change the result of the icmp.
836 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
837 // the maximum signed value for the pointer type.
838 if (ICmpInst::isSigned(Cond))
839 return nullptr;
840
841 // Look through bitcasts and addrspacecasts. We do not however want to remove
842 // 0 GEPs.
843 if (!isa<GetElementPtrInst>(RHS))
844 RHS = RHS->stripPointerCasts();
845
846 Value *PtrBase = GEPLHS->getOperand(0);
847 // FIXME: Support vector pointer GEPs.
848 if (PtrBase == RHS && GEPLHS->isInBounds() &&
849 !GEPLHS->getType()->isVectorTy()) {
850 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
851 // This transformation (ignoring the base and scales) is valid because we
852 // know pointers can't overflow since the gep is inbounds. See if we can
853 // output an optimized form.
854 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
855
856 // If not, synthesize the offset the hard way.
857 if (!Offset)
858 Offset = EmitGEPOffset(GEPLHS);
859 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
860 Constant::getNullValue(Offset->getType()));
861 }
862
863 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
864 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
865 !NullPointerIsDefined(I.getFunction(),
866 RHS->getType()->getPointerAddressSpace())) {
867 // For most address spaces, an allocation can't be placed at null, but null
868 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
869 // the only valid inbounds address derived from null, is null itself.
870 // Thus, we have four cases to consider:
871 // 1) Base == nullptr, Offset == 0 -> inbounds, null
872 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
873 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
874 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
875 //
876 // (Note if we're indexing a type of size 0, that simply collapses into one
877 // of the buckets above.)
878 //
879 // In general, we're allowed to make values less poison (i.e. remove
880 // sources of full UB), so in this case, we just select between the two
881 // non-poison cases (1 and 4 above).
882 //
883 // For vectors, we apply the same reasoning on a per-lane basis.
884 auto *Base = GEPLHS->getPointerOperand();
885 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
886 auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
887 Base = Builder.CreateVectorSplat(EC, Base);
888 }
889 return new ICmpInst(Cond, Base,
890 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
891 cast<Constant>(RHS), Base->getType()));
892 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
893 // If the base pointers are different, but the indices are the same, just
894 // compare the base pointer.
895 if (PtrBase != GEPRHS->getOperand(0)) {
896 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
897 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
898 GEPRHS->getOperand(0)->getType();
899 if (IndicesTheSame)
900 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
901 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
902 IndicesTheSame = false;
903 break;
904 }
905
906 // If all indices are the same, just compare the base pointers.
907 Type *BaseType = GEPLHS->getOperand(0)->getType();
908 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
909 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
910
911 // If we're comparing GEPs with two base pointers that only differ in type
912 // and both GEPs have only constant indices or just one use, then fold
913 // the compare with the adjusted indices.
914 // FIXME: Support vector of pointers.
915 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
916 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
917 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
918 PtrBase->stripPointerCasts() ==
919 GEPRHS->getOperand(0)->stripPointerCasts() &&
920 !GEPLHS->getType()->isVectorTy()) {
921 Value *LOffset = EmitGEPOffset(GEPLHS);
922 Value *ROffset = EmitGEPOffset(GEPRHS);
923
924 // If we looked through an addrspacecast between different sized address
925 // spaces, the LHS and RHS pointers are different sized
926 // integers. Truncate to the smaller one.
927 Type *LHSIndexTy = LOffset->getType();
928 Type *RHSIndexTy = ROffset->getType();
929 if (LHSIndexTy != RHSIndexTy) {
930 if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
931 RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
932 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
933 } else
934 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
935 }
936
937 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
938 LOffset, ROffset);
939 return replaceInstUsesWith(I, Cmp);
940 }
941
942 // Otherwise, the base pointers are different and the indices are
943 // different. Try convert this to an indexed compare by looking through
944 // PHIs/casts.
945 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
946 }
947
948 // If one of the GEPs has all zero indices, recurse.
949 // FIXME: Handle vector of pointers.
950 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
951 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
952 ICmpInst::getSwappedPredicate(Cond), I);
953
954 // If the other GEP has all zero indices, recurse.
955 // FIXME: Handle vector of pointers.
956 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
957 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
958
959 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
960 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
961 // If the GEPs only differ by one index, compare it.
962 unsigned NumDifferences = 0; // Keep track of # differences.
963 unsigned DiffOperand = 0; // The operand that differs.
964 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
965 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
966 Type *LHSType = GEPLHS->getOperand(i)->getType();
967 Type *RHSType = GEPRHS->getOperand(i)->getType();
968 // FIXME: Better support for vector of pointers.
969 if (LHSType->getPrimitiveSizeInBits() !=
970 RHSType->getPrimitiveSizeInBits() ||
971 (GEPLHS->getType()->isVectorTy() &&
972 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
973 // Irreconcilable differences.
974 NumDifferences = 2;
975 break;
976 }
977
978 if (NumDifferences++) break;
979 DiffOperand = i;
980 }
981
982 if (NumDifferences == 0) // SAME GEP?
983 return replaceInstUsesWith(I, // No comparison is needed here.
984 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
985
986 else if (NumDifferences == 1 && GEPsInBounds) {
987 Value *LHSV = GEPLHS->getOperand(DiffOperand);
988 Value *RHSV = GEPRHS->getOperand(DiffOperand);
989 // Make sure we do a signed comparison here.
990 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
991 }
992 }
993
994 // Only lower this if the icmp is the only user of the GEP or if we expect
995 // the result to fold to a constant!
996 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
997 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
998 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
999 Value *L = EmitGEPOffset(GEPLHS);
1000 Value *R = EmitGEPOffset(GEPRHS);
1001 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1002 }
1003 }
1004
1005 // Try convert this to an indexed compare by looking through PHIs/casts as a
1006 // last resort.
1007 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1008}
1009
1010Instruction *InstCombinerImpl::foldAllocaCmp(ICmpInst &ICI,
1011 const AllocaInst *Alloca,
1012 const Value *Other) {
1013 assert(ICI.isEquality() && "Cannot fold non-equality comparison.")(static_cast <bool> (ICI.isEquality() && "Cannot fold non-equality comparison."
) ? void (0) : __assert_fail ("ICI.isEquality() && \"Cannot fold non-equality comparison.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1013, __extension__ __PRETTY_FUNCTION__))
;
1014
1015 // It would be tempting to fold away comparisons between allocas and any
1016 // pointer not based on that alloca (e.g. an argument). However, even
1017 // though such pointers cannot alias, they can still compare equal.
1018 //
1019 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1020 // doesn't escape we can argue that it's impossible to guess its value, and we
1021 // can therefore act as if any such guesses are wrong.
1022 //
1023 // The code below checks that the alloca doesn't escape, and that it's only
1024 // used in a comparison once (the current instruction). The
1025 // single-comparison-use condition ensures that we're trivially folding all
1026 // comparisons against the alloca consistently, and avoids the risk of
1027 // erroneously folding a comparison of the pointer with itself.
1028
1029 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1030
1031 SmallVector<const Use *, 32> Worklist;
1032 for (const Use &U : Alloca->uses()) {
1033 if (Worklist.size() >= MaxIter)
1034 return nullptr;
1035 Worklist.push_back(&U);
1036 }
1037
1038 unsigned NumCmps = 0;
1039 while (!Worklist.empty()) {
1040 assert(Worklist.size() <= MaxIter)(static_cast <bool> (Worklist.size() <= MaxIter) ? void
(0) : __assert_fail ("Worklist.size() <= MaxIter", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1040, __extension__ __PRETTY_FUNCTION__))
;
1041 const Use *U = Worklist.pop_back_val();
1042 const Value *V = U->getUser();
1043 --MaxIter;
1044
1045 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1046 isa<SelectInst>(V)) {
1047 // Track the uses.
1048 } else if (isa<LoadInst>(V)) {
1049 // Loading from the pointer doesn't escape it.
1050 continue;
1051 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1052 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1053 if (SI->getValueOperand() == U->get())
1054 return nullptr;
1055 continue;
1056 } else if (isa<ICmpInst>(V)) {
1057 if (NumCmps++)
1058 return nullptr; // Found more than one cmp.
1059 continue;
1060 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1061 switch (Intrin->getIntrinsicID()) {
1062 // These intrinsics don't escape or compare the pointer. Memset is safe
1063 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1064 // we don't allow stores, so src cannot point to V.
1065 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1066 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1067 continue;
1068 default:
1069 return nullptr;
1070 }
1071 } else {
1072 return nullptr;
1073 }
1074 for (const Use &U : V->uses()) {
1075 if (Worklist.size() >= MaxIter)
1076 return nullptr;
1077 Worklist.push_back(&U);
1078 }
1079 }
1080
1081 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1082 return replaceInstUsesWith(
1083 ICI,
1084 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1085}
1086
1087/// Fold "icmp pred (X+C), X".
1088Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
1089 ICmpInst::Predicate Pred) {
1090 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1091 // so the values can never be equal. Similarly for all other "or equals"
1092 // operators.
1093 assert(!!C && "C should not be zero!")(static_cast <bool> (!!C && "C should not be zero!"
) ? void (0) : __assert_fail ("!!C && \"C should not be zero!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1093, __extension__ __PRETTY_FUNCTION__))
;
1094
1095 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1096 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1097 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1098 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1099 Constant *R = ConstantInt::get(X->getType(),
1100 APInt::getMaxValue(C.getBitWidth()) - C);
1101 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1102 }
1103
1104 // (X+1) >u X --> X <u (0-1) --> X != 255
1105 // (X+2) >u X --> X <u (0-2) --> X <u 254
1106 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1107 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1108 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1109 ConstantInt::get(X->getType(), -C));
1110
1111 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1112
1113 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1114 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1115 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1116 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1117 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1118 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1119 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1120 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1121 ConstantInt::get(X->getType(), SMax - C));
1122
1123 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1124 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1125 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1126 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1127 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1128 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1129
1130 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)(static_cast <bool> (Pred == ICmpInst::ICMP_SGT || Pred
== ICmpInst::ICMP_SGE) ? void (0) : __assert_fail ("Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1130, __extension__ __PRETTY_FUNCTION__))
;
1131 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1132 ConstantInt::get(X->getType(), SMax - (C - 1)));
1133}
1134
1135/// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1136/// (icmp eq/ne A, Log2(AP2/AP1)) ->
1137/// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1138Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
1139 const APInt &AP1,
1140 const APInt &AP2) {
1141 assert(I.isEquality() && "Cannot fold icmp gt/lt")(static_cast <bool> (I.isEquality() && "Cannot fold icmp gt/lt"
) ? void (0) : __assert_fail ("I.isEquality() && \"Cannot fold icmp gt/lt\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1141, __extension__ __PRETTY_FUNCTION__))
;
1142
1143 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1144 if (I.getPredicate() == I.ICMP_NE)
1145 Pred = CmpInst::getInversePredicate(Pred);
1146 return new ICmpInst(Pred, LHS, RHS);
1147 };
1148
1149 // Don't bother doing any work for cases which InstSimplify handles.
1150 if (AP2.isNullValue())
1151 return nullptr;
1152
1153 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1154 if (IsAShr) {
1155 if (AP2.isAllOnesValue())
1156 return nullptr;
1157 if (AP2.isNegative() != AP1.isNegative())
1158 return nullptr;
1159 if (AP2.sgt(AP1))
1160 return nullptr;
1161 }
1162
1163 if (!AP1)
1164 // 'A' must be large enough to shift out the highest set bit.
1165 return getICmp(I.ICMP_UGT, A,
1166 ConstantInt::get(A->getType(), AP2.logBase2()));
1167
1168 if (AP1 == AP2)
1169 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1170
1171 int Shift;
1172 if (IsAShr && AP1.isNegative())
1173 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1174 else
1175 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1176
1177 if (Shift > 0) {
1178 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1179 // There are multiple solutions if we are comparing against -1 and the LHS
1180 // of the ashr is not a power of two.
1181 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1182 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1183 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1184 } else if (AP1 == AP2.lshr(Shift)) {
1185 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1186 }
1187 }
1188
1189 // Shifting const2 will never be equal to const1.
1190 // FIXME: This should always be handled by InstSimplify?
1191 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1192 return replaceInstUsesWith(I, TorF);
1193}
1194
1195/// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1196/// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1197Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1198 const APInt &AP1,
1199 const APInt &AP2) {
1200 assert(I.isEquality() && "Cannot fold icmp gt/lt")(static_cast <bool> (I.isEquality() && "Cannot fold icmp gt/lt"
) ? void (0) : __assert_fail ("I.isEquality() && \"Cannot fold icmp gt/lt\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1200, __extension__ __PRETTY_FUNCTION__))
;
1201
1202 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1203 if (I.getPredicate() == I.ICMP_NE)
1204 Pred = CmpInst::getInversePredicate(Pred);
1205 return new ICmpInst(Pred, LHS, RHS);
1206 };
1207
1208 // Don't bother doing any work for cases which InstSimplify handles.
1209 if (AP2.isNullValue())
1210 return nullptr;
1211
1212 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1213
1214 if (!AP1 && AP2TrailingZeros != 0)
1215 return getICmp(
1216 I.ICMP_UGE, A,
1217 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1218
1219 if (AP1 == AP2)
1220 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1221
1222 // Get the distance between the lowest bits that are set.
1223 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1224
1225 if (Shift > 0 && AP2.shl(Shift) == AP1)
1226 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1227
1228 // Shifting const2 will never be equal to const1.
1229 // FIXME: This should always be handled by InstSimplify?
1230 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1231 return replaceInstUsesWith(I, TorF);
1232}
1233
1234/// The caller has matched a pattern of the form:
1235/// I = icmp ugt (add (add A, B), CI2), CI1
1236/// If this is of the form:
1237/// sum = a + b
1238/// if (sum+128 >u 255)
1239/// Then replace it with llvm.sadd.with.overflow.i8.
1240///
1241static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1242 ConstantInt *CI2, ConstantInt *CI1,
1243 InstCombinerImpl &IC) {
1244 // The transformation we're trying to do here is to transform this into an
1245 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1246 // with a narrower add, and discard the add-with-constant that is part of the
1247 // range check (if we can't eliminate it, this isn't profitable).
1248
1249 // In order to eliminate the add-with-constant, the compare can be its only
1250 // use.
1251 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1252 if (!AddWithCst->hasOneUse())
1253 return nullptr;
1254
1255 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1256 if (!CI2->getValue().isPowerOf2())
1257 return nullptr;
1258 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1259 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1260 return nullptr;
1261
1262 // The width of the new add formed is 1 more than the bias.
1263 ++NewWidth;
1264
1265 // Check to see that CI1 is an all-ones value with NewWidth bits.
1266 if (CI1->getBitWidth() == NewWidth ||
1267 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1268 return nullptr;
1269
1270 // This is only really a signed overflow check if the inputs have been
1271 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1272 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1273 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1274 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1275 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1276 return nullptr;
1277
1278 // In order to replace the original add with a narrower
1279 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1280 // and truncates that discard the high bits of the add. Verify that this is
1281 // the case.
1282 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1283 for (User *U : OrigAdd->users()) {
1284 if (U == AddWithCst)
1285 continue;
1286
1287 // Only accept truncates for now. We would really like a nice recursive
1288 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1289 // chain to see which bits of a value are actually demanded. If the
1290 // original add had another add which was then immediately truncated, we
1291 // could still do the transformation.
1292 TruncInst *TI = dyn_cast<TruncInst>(U);
1293 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1294 return nullptr;
1295 }
1296
1297 // If the pattern matches, truncate the inputs to the narrower type and
1298 // use the sadd_with_overflow intrinsic to efficiently compute both the
1299 // result and the overflow bit.
1300 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1301 Function *F = Intrinsic::getDeclaration(
1302 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1303
1304 InstCombiner::BuilderTy &Builder = IC.Builder;
1305
1306 // Put the new code above the original add, in case there are any uses of the
1307 // add between the add and the compare.
1308 Builder.SetInsertPoint(OrigAdd);
1309
1310 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1311 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1312 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1313 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1314 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1315
1316 // The inner add was the result of the narrow add, zero extended to the
1317 // wider type. Replace it with the result computed by the intrinsic.
1318 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1319 IC.eraseInstFromFunction(*OrigAdd);
1320
1321 // The original icmp gets replaced with the overflow value.
1322 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1323}
1324
1325/// If we have:
1326/// icmp eq/ne (urem/srem %x, %y), 0
1327/// iff %y is a power-of-two, we can replace this with a bit test:
1328/// icmp eq/ne (and %x, (add %y, -1)), 0
1329Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1330 // This fold is only valid for equality predicates.
1331 if (!I.isEquality())
1332 return nullptr;
1333 ICmpInst::Predicate Pred;
1334 Value *X, *Y, *Zero;
1335 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1336 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1337 return nullptr;
1338 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1339 return nullptr;
1340 // This may increase instruction count, we don't enforce that Y is a constant.
1341 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1342 Value *Masked = Builder.CreateAnd(X, Mask);
1343 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1344}
1345
1346/// Fold equality-comparison between zero and any (maybe truncated) right-shift
1347/// by one-less-than-bitwidth into a sign test on the original value.
1348Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1349 Instruction *Val;
1350 ICmpInst::Predicate Pred;
1351 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1352 return nullptr;
1353
1354 Value *X;
1355 Type *XTy;
1356
1357 Constant *C;
1358 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1359 XTy = X->getType();
1360 unsigned XBitWidth = XTy->getScalarSizeInBits();
1361 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1362 APInt(XBitWidth, XBitWidth - 1))))
1363 return nullptr;
1364 } else if (isa<BinaryOperator>(Val) &&
1365 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1366 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1367 /*AnalyzeForSignBitExtraction=*/true))) {
1368 XTy = X->getType();
1369 } else
1370 return nullptr;
1371
1372 return ICmpInst::Create(Instruction::ICmp,
1373 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1374 : ICmpInst::ICMP_SLT,
1375 X, ConstantInt::getNullValue(XTy));
1376}
1377
1378// Handle icmp pred X, 0
1379Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1380 CmpInst::Predicate Pred = Cmp.getPredicate();
1381 if (!match(Cmp.getOperand(1), m_Zero()))
1382 return nullptr;
1383
1384 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1385 if (Pred == ICmpInst::ICMP_SGT) {
1386 Value *A, *B;
1387 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1388 if (SPR.Flavor == SPF_SMIN) {
1389 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1390 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1391 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1392 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1393 }
1394 }
1395
1396 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1397 return New;
1398
1399 // Given:
1400 // icmp eq/ne (urem %x, %y), 0
1401 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1402 // icmp eq/ne %x, 0
1403 Value *X, *Y;
1404 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1405 ICmpInst::isEquality(Pred)) {
1406 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1407 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1408 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1409 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1410 }
1411
1412 return nullptr;
1413}
1414
1415/// Fold icmp Pred X, C.
1416/// TODO: This code structure does not make sense. The saturating add fold
1417/// should be moved to some other helper and extended as noted below (it is also
1418/// possible that code has been made unnecessary - do we canonicalize IR to
1419/// overflow/saturating intrinsics or not?).
1420Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1421 // Match the following pattern, which is a common idiom when writing
1422 // overflow-safe integer arithmetic functions. The source performs an addition
1423 // in wider type and explicitly checks for overflow using comparisons against
1424 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1425 //
1426 // TODO: This could probably be generalized to handle other overflow-safe
1427 // operations if we worked out the formulas to compute the appropriate magic
1428 // constants.
1429 //
1430 // sum = a + b
1431 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1432 CmpInst::Predicate Pred = Cmp.getPredicate();
1433 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1434 Value *A, *B;
1435 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1436 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1437 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1438 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1439 return Res;
1440
1441 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1442 Constant *C = dyn_cast<Constant>(Op1);
1443 if (!C)
1444 return nullptr;
1445
1446 if (auto *Phi = dyn_cast<PHINode>(Op0))
1447 if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1448 Type *Ty = Cmp.getType();
1449 Builder.SetInsertPoint(Phi);
1450 PHINode *NewPhi =
1451 Builder.CreatePHI(Ty, Phi->getNumOperands());
1452 for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1453 auto *Input =
1454 cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1455 auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1456 NewPhi->addIncoming(BoolInput, Predecessor);
1457 }
1458 NewPhi->takeName(&Cmp);
1459 return replaceInstUsesWith(Cmp, NewPhi);
1460 }
1461
1462 return nullptr;
1463}
1464
1465/// Canonicalize icmp instructions based on dominating conditions.
1466Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1467 // This is a cheap/incomplete check for dominance - just match a single
1468 // predecessor with a conditional branch.
1469 BasicBlock *CmpBB = Cmp.getParent();
1470 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1471 if (!DomBB)
1472 return nullptr;
1473
1474 Value *DomCond;
1475 BasicBlock *TrueBB, *FalseBB;
1476 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1477 return nullptr;
1478
1479 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&(static_cast <bool> ((TrueBB == CmpBB || FalseBB == CmpBB
) && "Predecessor block does not point to successor?"
) ? void (0) : __assert_fail ("(TrueBB == CmpBB || FalseBB == CmpBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1480, __extension__ __PRETTY_FUNCTION__))
1480 "Predecessor block does not point to successor?")(static_cast <bool> ((TrueBB == CmpBB || FalseBB == CmpBB
) && "Predecessor block does not point to successor?"
) ? void (0) : __assert_fail ("(TrueBB == CmpBB || FalseBB == CmpBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1480, __extension__ __PRETTY_FUNCTION__))
;
1481
1482 // The branch should get simplified. Don't bother simplifying this condition.
1483 if (TrueBB == FalseBB)
1484 return nullptr;
1485
1486 // Try to simplify this compare to T/F based on the dominating condition.
1487 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1488 if (Imp)
1489 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1490
1491 CmpInst::Predicate Pred = Cmp.getPredicate();
1492 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1493 ICmpInst::Predicate DomPred;
1494 const APInt *C, *DomC;
1495 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1496 match(Y, m_APInt(C))) {
1497 // We have 2 compares of a variable with constants. Calculate the constant
1498 // ranges of those compares to see if we can transform the 2nd compare:
1499 // DomBB:
1500 // DomCond = icmp DomPred X, DomC
1501 // br DomCond, CmpBB, FalseBB
1502 // CmpBB:
1503 // Cmp = icmp Pred X, C
1504 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1505 ConstantRange DominatingCR =
1506 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1507 : ConstantRange::makeExactICmpRegion(
1508 CmpInst::getInversePredicate(DomPred), *DomC);
1509 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1510 ConstantRange Difference = DominatingCR.difference(CR);
1511 if (Intersection.isEmptySet())
1512 return replaceInstUsesWith(Cmp, Builder.getFalse());
1513 if (Difference.isEmptySet())
1514 return replaceInstUsesWith(Cmp, Builder.getTrue());
1515
1516 // Canonicalizing a sign bit comparison that gets used in a branch,
1517 // pessimizes codegen by generating branch on zero instruction instead
1518 // of a test and branch. So we avoid canonicalizing in such situations
1519 // because test and branch instruction has better branch displacement
1520 // than compare and branch instruction.
1521 bool UnusedBit;
1522 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1523 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1524 return nullptr;
1525
1526 // Avoid an infinite loop with min/max canonicalization.
1527 // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1528 if (Cmp.hasOneUse() &&
1529 match(Cmp.user_back(), m_MaxOrMin(m_Value(), m_Value())))
1530 return nullptr;
1531
1532 if (const APInt *EqC = Intersection.getSingleElement())
1533 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1534 if (const APInt *NeC = Difference.getSingleElement())
1535 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1536 }
1537
1538 return nullptr;
1539}
1540
1541/// Fold icmp (trunc X, Y), C.
1542Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1543 TruncInst *Trunc,
1544 const APInt &C) {
1545 ICmpInst::Predicate Pred = Cmp.getPredicate();
1546 Value *X = Trunc->getOperand(0);
1547 if (C.isOneValue() && C.getBitWidth() > 1) {
1548 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1549 Value *V = nullptr;
1550 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1551 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1552 ConstantInt::get(V->getType(), 1));
1553 }
1554
1555 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1556 SrcBits = X->getType()->getScalarSizeInBits();
1557 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1558 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1559 // of the high bits truncated out of x are known.
1560 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1561
1562 // If all the high bits are known, we can do this xform.
1563 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1564 // Pull in the high bits from known-ones set.
1565 APInt NewRHS = C.zext(SrcBits);
1566 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1567 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1568 }
1569 }
1570
1571 // Look through truncated right-shift of the sign-bit for a sign-bit check:
1572 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1573 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1574 Value *ShOp;
1575 const APInt *ShAmtC;
1576 bool TrueIfSigned;
1577 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1578 match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1579 DstBits == SrcBits - ShAmtC->getZExtValue()) {
1580 return TrueIfSigned
1581 ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1582 ConstantInt::getNullValue(X->getType()))
1583 : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1584 ConstantInt::getAllOnesValue(X->getType()));
1585 }
1586
1587 return nullptr;
1588}
1589
1590/// Fold icmp (xor X, Y), C.
1591Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1592 BinaryOperator *Xor,
1593 const APInt &C) {
1594 Value *X = Xor->getOperand(0);
1595 Value *Y = Xor->getOperand(1);
1596 const APInt *XorC;
1597 if (!match(Y, m_APInt(XorC)))
1598 return nullptr;
1599
1600 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1601 // fold the xor.
1602 ICmpInst::Predicate Pred = Cmp.getPredicate();
1603 bool TrueIfSigned = false;
1604 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1605
1606 // If the sign bit of the XorCst is not set, there is no change to
1607 // the operation, just stop using the Xor.
1608 if (!XorC->isNegative())
1609 return replaceOperand(Cmp, 0, X);
1610
1611 // Emit the opposite comparison.
1612 if (TrueIfSigned)
1613 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1614 ConstantInt::getAllOnesValue(X->getType()));
1615 else
1616 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1617 ConstantInt::getNullValue(X->getType()));
1618 }
1619
1620 if (Xor->hasOneUse()) {
1621 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1622 if (!Cmp.isEquality() && XorC->isSignMask()) {
1623 Pred = Cmp.getFlippedSignednessPredicate();
1624 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1625 }
1626
1627 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1628 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1629 Pred = Cmp.getFlippedSignednessPredicate();
1630 Pred = Cmp.getSwappedPredicate(Pred);
1631 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1632 }
1633 }
1634
1635 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1636 if (Pred == ICmpInst::ICMP_UGT) {
1637 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1638 if (*XorC == ~C && (C + 1).isPowerOf2())
1639 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1640 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1641 if (*XorC == C && (C + 1).isPowerOf2())
1642 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1643 }
1644 if (Pred == ICmpInst::ICMP_ULT) {
1645 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1646 if (*XorC == -C && C.isPowerOf2())
1647 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1648 ConstantInt::get(X->getType(), ~C));
1649 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1650 if (*XorC == C && (-C).isPowerOf2())
1651 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1652 ConstantInt::get(X->getType(), ~C));
1653 }
1654 return nullptr;
1655}
1656
1657/// Fold icmp (and (sh X, Y), C2), C1.
1658Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1659 BinaryOperator *And,
1660 const APInt &C1,
1661 const APInt &C2) {
1662 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1663 if (!Shift || !Shift->isShift())
1664 return nullptr;
1665
1666 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1667 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1668 // code produced by the clang front-end, for bitfield access.
1669 // This seemingly simple opportunity to fold away a shift turns out to be
1670 // rather complicated. See PR17827 for details.
1671 unsigned ShiftOpcode = Shift->getOpcode();
1672 bool IsShl = ShiftOpcode == Instruction::Shl;
1673 const APInt *C3;
1674 if (match(Shift->getOperand(1), m_APInt(C3))) {
1675 APInt NewAndCst, NewCmpCst;
1676 bool AnyCmpCstBitsShiftedOut;
1677 if (ShiftOpcode == Instruction::Shl) {
1678 // For a left shift, we can fold if the comparison is not signed. We can
1679 // also fold a signed comparison if the mask value and comparison value
1680 // are not negative. These constraints may not be obvious, but we can
1681 // prove that they are correct using an SMT solver.
1682 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1683 return nullptr;
1684
1685 NewCmpCst = C1.lshr(*C3);
1686 NewAndCst = C2.lshr(*C3);
1687 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1688 } else if (ShiftOpcode == Instruction::LShr) {
1689 // For a logical right shift, we can fold if the comparison is not signed.
1690 // We can also fold a signed comparison if the shifted mask value and the
1691 // shifted comparison value are not negative. These constraints may not be
1692 // obvious, but we can prove that they are correct using an SMT solver.
1693 NewCmpCst = C1.shl(*C3);
1694 NewAndCst = C2.shl(*C3);
1695 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1696 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1697 return nullptr;
1698 } else {
1699 // For an arithmetic shift, check that both constants don't use (in a
1700 // signed sense) the top bits being shifted out.
1701 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode")(static_cast <bool> (ShiftOpcode == Instruction::AShr &&
"Unknown shift opcode") ? void (0) : __assert_fail ("ShiftOpcode == Instruction::AShr && \"Unknown shift opcode\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1701, __extension__ __PRETTY_FUNCTION__))
;
1702 NewCmpCst = C1.shl(*C3);
1703 NewAndCst = C2.shl(*C3);
1704 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1705 if (NewAndCst.ashr(*C3) != C2)
1706 return nullptr;
1707 }
1708
1709 if (AnyCmpCstBitsShiftedOut) {
1710 // If we shifted bits out, the fold is not going to work out. As a
1711 // special case, check to see if this means that the result is always
1712 // true or false now.
1713 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1714 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1715 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1716 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1717 } else {
1718 Value *NewAnd = Builder.CreateAnd(
1719 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1720 return new ICmpInst(Cmp.getPredicate(),
1721 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1722 }
1723 }
1724
1725 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1726 // preferable because it allows the C2 << Y expression to be hoisted out of a
1727 // loop if Y is invariant and X is not.
1728 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1729 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1730 // Compute C2 << Y.
1731 Value *NewShift =
1732 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1733 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1734
1735 // Compute X & (C2 << Y).
1736 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1737 return replaceOperand(Cmp, 0, NewAnd);
1738 }
1739
1740 return nullptr;
1741}
1742
1743/// Fold icmp (and X, C2), C1.
1744Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1745 BinaryOperator *And,
1746 const APInt &C1) {
1747 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1748
1749 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1750 // TODO: We canonicalize to the longer form for scalars because we have
1751 // better analysis/folds for icmp, and codegen may be better with icmp.
1752 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1753 match(And->getOperand(1), m_One()))
1754 return new TruncInst(And->getOperand(0), Cmp.getType());
1755
1756 const APInt *C2;
1757 Value *X;
1758 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1759 return nullptr;
1760
1761 // Don't perform the following transforms if the AND has multiple uses
1762 if (!And->hasOneUse())
1763 return nullptr;
1764
1765 if (Cmp.isEquality() && C1.isNullValue()) {
1766 // Restrict this fold to single-use 'and' (PR10267).
1767 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1768 if (C2->isSignMask()) {
1769 Constant *Zero = Constant::getNullValue(X->getType());
1770 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1771 return new ICmpInst(NewPred, X, Zero);
1772 }
1773
1774 // Restrict this fold only for single-use 'and' (PR10267).
1775 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1776 if ((~(*C2) + 1).isPowerOf2()) {
1777 Constant *NegBOC =
1778 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1779 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1780 return new ICmpInst(NewPred, X, NegBOC);
1781 }
1782 }
1783
1784 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1785 // the input width without changing the value produced, eliminate the cast:
1786 //
1787 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1788 //
1789 // We can do this transformation if the constants do not have their sign bits
1790 // set or if it is an equality comparison. Extending a relational comparison
1791 // when we're checking the sign bit would not work.
1792 Value *W;
1793 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1794 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1795 // TODO: Is this a good transform for vectors? Wider types may reduce
1796 // throughput. Should this transform be limited (even for scalars) by using
1797 // shouldChangeType()?
1798 if (!Cmp.getType()->isVectorTy()) {
1799 Type *WideType = W->getType();
1800 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1801 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1802 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1803 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1804 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1805 }
1806 }
1807
1808 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1809 return I;
1810
1811 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1812 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1813 //
1814 // iff pred isn't signed
1815 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1816 match(And->getOperand(1), m_One())) {
1817 Constant *One = cast<Constant>(And->getOperand(1));
1818 Value *Or = And->getOperand(0);
1819 Value *A, *B, *LShr;
1820 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1821 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1822 unsigned UsesRemoved = 0;
1823 if (And->hasOneUse())
1824 ++UsesRemoved;
1825 if (Or->hasOneUse())
1826 ++UsesRemoved;
1827 if (LShr->hasOneUse())
1828 ++UsesRemoved;
1829
1830 // Compute A & ((1 << B) | 1)
1831 Value *NewOr = nullptr;
1832 if (auto *C = dyn_cast<Constant>(B)) {
1833 if (UsesRemoved >= 1)
1834 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1835 } else {
1836 if (UsesRemoved >= 3)
1837 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1838 /*HasNUW=*/true),
1839 One, Or->getName());
1840 }
1841 if (NewOr) {
1842 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1843 return replaceOperand(Cmp, 0, NewAnd);
1844 }
1845 }
1846 }
1847
1848 return nullptr;
1849}
1850
1851/// Fold icmp (and X, Y), C.
1852Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1853 BinaryOperator *And,
1854 const APInt &C) {
1855 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1856 return I;
1857
1858 // TODO: These all require that Y is constant too, so refactor with the above.
1859
1860 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1861 Value *X = And->getOperand(0);
1862 Value *Y = And->getOperand(1);
1863 if (auto *LI = dyn_cast<LoadInst>(X))
1864 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1865 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1866 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1867 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1868 ConstantInt *C2 = cast<ConstantInt>(Y);
1869 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1870 return Res;
1871 }
1872
1873 if (!Cmp.isEquality())
1874 return nullptr;
1875
1876 // X & -C == -C -> X > u ~C
1877 // X & -C != -C -> X <= u ~C
1878 // iff C is a power of 2
1879 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1880 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1881 : CmpInst::ICMP_ULE;
1882 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1883 }
1884
1885 // (X & C2) == 0 -> (trunc X) >= 0
1886 // (X & C2) != 0 -> (trunc X) < 0
1887 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1888 const APInt *C2;
1889 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1890 int32_t ExactLogBase2 = C2->exactLogBase2();
1891 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1892 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1893 if (auto *AndVTy = dyn_cast<VectorType>(And->getType()))
1894 NTy = VectorType::get(NTy, AndVTy->getElementCount());
1895 Value *Trunc = Builder.CreateTrunc(X, NTy);
1896 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1897 : CmpInst::ICMP_SLT;
1898 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1899 }
1900 }
1901
1902 return nullptr;
1903}
1904
1905/// Fold icmp (or X, Y), C.
1906Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
1907 BinaryOperator *Or,
1908 const APInt &C) {
1909 ICmpInst::Predicate Pred = Cmp.getPredicate();
1910 if (C.isOneValue()) {
1911 // icmp slt signum(V) 1 --> icmp slt V, 1
1912 Value *V = nullptr;
1913 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1914 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1915 ConstantInt::get(V->getType(), 1));
1916 }
1917
1918 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1919 const APInt *MaskC;
1920 if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1921 if (*MaskC == C && (C + 1).isPowerOf2()) {
1922 // X | C == C --> X <=u C
1923 // X | C != C --> X >u C
1924 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1925 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1926 return new ICmpInst(Pred, OrOp0, OrOp1);
1927 }
1928
1929 // More general: canonicalize 'equality with set bits mask' to
1930 // 'equality with clear bits mask'.
1931 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1932 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1933 if (Or->hasOneUse()) {
1934 Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1935 Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1936 return new ICmpInst(Pred, And, NewC);
1937 }
1938 }
1939
1940 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1941 return nullptr;
1942
1943 Value *P, *Q;
1944 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1945 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1946 // -> and (icmp eq P, null), (icmp eq Q, null).
1947 Value *CmpP =
1948 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1949 Value *CmpQ =
1950 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1951 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1952 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1953 }
1954
1955 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1956 // a shorter form that has more potential to be folded even further.
1957 Value *X1, *X2, *X3, *X4;
1958 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1959 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1960 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1961 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1962 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1963 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1964 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1965 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1966 }
1967
1968 return nullptr;
1969}
1970
1971/// Fold icmp (mul X, Y), C.
1972Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
1973 BinaryOperator *Mul,
1974 const APInt &C) {
1975 const APInt *MulC;
1976 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1977 return nullptr;
1978
1979 // If this is a test of the sign bit and the multiply is sign-preserving with
1980 // a constant operand, use the multiply LHS operand instead.
1981 ICmpInst::Predicate Pred = Cmp.getPredicate();
1982 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1983 if (MulC->isNegative())
1984 Pred = ICmpInst::getSwappedPredicate(Pred);
1985 return new ICmpInst(Pred, Mul->getOperand(0),
1986 Constant::getNullValue(Mul->getType()));
1987 }
1988
1989 // If the multiply does not wrap, try to divide the compare constant by the
1990 // multiplication factor.
1991 if (Cmp.isEquality() && !MulC->isNullValue()) {
1992 // (mul nsw X, MulC) == C --> X == C /s MulC
1993 if (Mul->hasNoSignedWrap() && C.srem(*MulC).isNullValue()) {
1994 Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
1995 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1996 }
1997 // (mul nuw X, MulC) == C --> X == C /u MulC
1998 if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isNullValue()) {
1999 Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
2000 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2001 }
2002 }
2003
2004 return nullptr;
2005}
2006
2007/// Fold icmp (shl 1, Y), C.
2008static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
2009 const APInt &C) {
2010 Value *Y;
2011 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2012 return nullptr;
2013
2014 Type *ShiftType = Shl->getType();
2015 unsigned TypeBits = C.getBitWidth();
2016 bool CIsPowerOf2 = C.isPowerOf2();
2017 ICmpInst::Predicate Pred = Cmp.getPredicate();
2018 if (Cmp.isUnsigned()) {
2019 // (1 << Y) pred C -> Y pred Log2(C)
2020 if (!CIsPowerOf2) {
2021 // (1 << Y) < 30 -> Y <= 4
2022 // (1 << Y) <= 30 -> Y <= 4
2023 // (1 << Y) >= 30 -> Y > 4
2024 // (1 << Y) > 30 -> Y > 4
2025 if (Pred == ICmpInst::ICMP_ULT)
2026 Pred = ICmpInst::ICMP_ULE;
2027 else if (Pred == ICmpInst::ICMP_UGE)
2028 Pred = ICmpInst::ICMP_UGT;
2029 }
2030
2031 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2032 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2033 unsigned CLog2 = C.logBase2();
2034 if (CLog2 == TypeBits - 1) {
2035 if (Pred == ICmpInst::ICMP_UGE)
2036 Pred = ICmpInst::ICMP_EQ;
2037 else if (Pred == ICmpInst::ICMP_ULT)
2038 Pred = ICmpInst::ICMP_NE;
2039 }
2040 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2041 } else if (Cmp.isSigned()) {
2042 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2043 if (C.isAllOnesValue()) {
2044 // (1 << Y) <= -1 -> Y == 31
2045 if (Pred == ICmpInst::ICMP_SLE)
2046 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2047
2048 // (1 << Y) > -1 -> Y != 31
2049 if (Pred == ICmpInst::ICMP_SGT)
2050 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2051 } else if (!C) {
2052 // (1 << Y) < 0 -> Y == 31
2053 // (1 << Y) <= 0 -> Y == 31
2054 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2055 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2056
2057 // (1 << Y) >= 0 -> Y != 31
2058 // (1 << Y) > 0 -> Y != 31
2059 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2060 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2061 }
2062 } else if (Cmp.isEquality() && CIsPowerOf2) {
2063 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2064 }
2065
2066 return nullptr;
2067}
2068
2069/// Fold icmp (shl X, Y), C.
2070Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2071 BinaryOperator *Shl,
2072 const APInt &C) {
2073 const APInt *ShiftVal;
2074 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2075 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2076
2077 const APInt *ShiftAmt;
2078 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2079 return foldICmpShlOne(Cmp, Shl, C);
2080
2081 // Check that the shift amount is in range. If not, don't perform undefined
2082 // shifts. When the shift is visited, it will be simplified.
2083 unsigned TypeBits = C.getBitWidth();
2084 if (ShiftAmt->uge(TypeBits))
2085 return nullptr;
2086
2087 ICmpInst::Predicate Pred = Cmp.getPredicate();
2088 Value *X = Shl->getOperand(0);
2089 Type *ShType = Shl->getType();
2090
2091 // NSW guarantees that we are only shifting out sign bits from the high bits,
2092 // so we can ASHR the compare constant without needing a mask and eliminate
2093 // the shift.
2094 if (Shl->hasNoSignedWrap()) {
2095 if (Pred == ICmpInst::ICMP_SGT) {
2096 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2097 APInt ShiftedC = C.ashr(*ShiftAmt);
2098 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2099 }
2100 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2101 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2102 APInt ShiftedC = C.ashr(*ShiftAmt);
2103 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2104 }
2105 if (Pred == ICmpInst::ICMP_SLT) {
2106 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2107 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2108 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2109 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2110 assert(!C.isMinSignedValue() && "Unexpected icmp slt")(static_cast <bool> (!C.isMinSignedValue() && "Unexpected icmp slt"
) ? void (0) : __assert_fail ("!C.isMinSignedValue() && \"Unexpected icmp slt\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2110, __extension__ __PRETTY_FUNCTION__))
;
2111 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2112 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2113 }
2114 // If this is a signed comparison to 0 and the shift is sign preserving,
2115 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2116 // do that if we're sure to not continue on in this function.
2117 if (isSignTest(Pred, C))
2118 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2119 }
2120
2121 // NUW guarantees that we are only shifting out zero bits from the high bits,
2122 // so we can LSHR the compare constant without needing a mask and eliminate
2123 // the shift.
2124 if (Shl->hasNoUnsignedWrap()) {
2125 if (Pred == ICmpInst::ICMP_UGT) {
2126 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2127 APInt ShiftedC = C.lshr(*ShiftAmt);
2128 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2129 }
2130 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2131 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2132 APInt ShiftedC = C.lshr(*ShiftAmt);
2133 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2134 }
2135 if (Pred == ICmpInst::ICMP_ULT) {
2136 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2137 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2138 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2139 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2140 assert(C.ugt(0) && "ult 0 should have been eliminated")(static_cast <bool> (C.ugt(0) && "ult 0 should have been eliminated"
) ? void (0) : __assert_fail ("C.ugt(0) && \"ult 0 should have been eliminated\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2140, __extension__ __PRETTY_FUNCTION__))
;
2141 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2142 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2143 }
2144 }
2145
2146 if (Cmp.isEquality() && Shl->hasOneUse()) {
2147 // Strength-reduce the shift into an 'and'.
2148 Constant *Mask = ConstantInt::get(
2149 ShType,
2150 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2151 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2152 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2153 return new ICmpInst(Pred, And, LShrC);
2154 }
2155
2156 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2157 bool TrueIfSigned = false;
2158 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2159 // (X << 31) <s 0 --> (X & 1) != 0
2160 Constant *Mask = ConstantInt::get(
2161 ShType,
2162 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2163 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2164 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2165 And, Constant::getNullValue(ShType));
2166 }
2167
2168 // Simplify 'shl' inequality test into 'and' equality test.
2169 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2170 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2171 if ((C + 1).isPowerOf2() &&
2172 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2173 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2174 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2175 : ICmpInst::ICMP_NE,
2176 And, Constant::getNullValue(ShType));
2177 }
2178 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2179 if (C.isPowerOf2() &&
2180 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2181 Value *And =
2182 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2183 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2184 : ICmpInst::ICMP_NE,
2185 And, Constant::getNullValue(ShType));
2186 }
2187 }
2188
2189 // Transform (icmp pred iM (shl iM %v, N), C)
2190 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2191 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2192 // This enables us to get rid of the shift in favor of a trunc that may be
2193 // free on the target. It has the additional benefit of comparing to a
2194 // smaller constant that may be more target-friendly.
2195 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2196 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2197 DL.isLegalInteger(TypeBits - Amt)) {
2198 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2199 if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2200 TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
2201 Constant *NewC =
2202 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2203 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2204 }
2205
2206 return nullptr;
2207}
2208
2209/// Fold icmp ({al}shr X, Y), C.
2210Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2211 BinaryOperator *Shr,
2212 const APInt &C) {
2213 // An exact shr only shifts out zero bits, so:
2214 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2215 Value *X = Shr->getOperand(0);
2216 CmpInst::Predicate Pred = Cmp.getPredicate();
2217 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2218 C.isNullValue())
2219 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2220
2221 const APInt *ShiftVal;
2222 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2223 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2224
2225 const APInt *ShiftAmt;
2226 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2227 return nullptr;
2228
2229 // Check that the shift amount is in range. If not, don't perform undefined
2230 // shifts. When the shift is visited it will be simplified.
2231 unsigned TypeBits = C.getBitWidth();
2232 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2233 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2234 return nullptr;
2235
2236 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2237 bool IsExact = Shr->isExact();
2238 Type *ShrTy = Shr->getType();
2239 // TODO: If we could guarantee that InstSimplify would handle all of the
2240 // constant-value-based preconditions in the folds below, then we could assert
2241 // those conditions rather than checking them. This is difficult because of
2242 // undef/poison (PR34838).
2243 if (IsAShr) {
2244 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2245 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2246 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2247 APInt ShiftedC = C.shl(ShAmtVal);
2248 if (ShiftedC.ashr(ShAmtVal) == C)
2249 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2250 }
2251 if (Pred == CmpInst::ICMP_SGT) {
2252 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2253 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2254 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2255 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2256 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2257 }
2258
2259 // If the compare constant has significant bits above the lowest sign-bit,
2260 // then convert an unsigned cmp to a test of the sign-bit:
2261 // (ashr X, ShiftC) u> C --> X s< 0
2262 // (ashr X, ShiftC) u< C --> X s> -1
2263 if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2264 if (Pred == CmpInst::ICMP_UGT) {
2265 return new ICmpInst(CmpInst::ICMP_SLT, X,
2266 ConstantInt::getNullValue(ShrTy));
2267 }
2268 if (Pred == CmpInst::ICMP_ULT) {
2269 return new ICmpInst(CmpInst::ICMP_SGT, X,
2270 ConstantInt::getAllOnesValue(ShrTy));
2271 }
2272 }
2273 } else {
2274 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2275 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2276 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2277 APInt ShiftedC = C.shl(ShAmtVal);
2278 if (ShiftedC.lshr(ShAmtVal) == C)
2279 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2280 }
2281 if (Pred == CmpInst::ICMP_UGT) {
2282 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2283 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2284 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2285 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2286 }
2287 }
2288
2289 if (!Cmp.isEquality())
2290 return nullptr;
2291
2292 // Handle equality comparisons of shift-by-constant.
2293
2294 // If the comparison constant changes with the shift, the comparison cannot
2295 // succeed (bits of the comparison constant cannot match the shifted value).
2296 // This should be known by InstSimplify and already be folded to true/false.
2297 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||(static_cast <bool> (((IsAShr && C.shl(ShAmtVal
).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal)
.lshr(ShAmtVal) == C)) && "Expected icmp+shr simplify did not occur."
) ? 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-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2299, __extension__ __PRETTY_FUNCTION__))
2298 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&(static_cast <bool> (((IsAShr && C.shl(ShAmtVal
).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal)
.lshr(ShAmtVal) == C)) && "Expected icmp+shr simplify did not occur."
) ? 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-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2299, __extension__ __PRETTY_FUNCTION__))
2299 "Expected icmp+shr simplify did not occur.")(static_cast <bool> (((IsAShr && C.shl(ShAmtVal
).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal)
.lshr(ShAmtVal) == C)) && "Expected icmp+shr simplify did not occur."
) ? 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-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2299, __extension__ __PRETTY_FUNCTION__))
;
2300
2301 // If the bits shifted out are known zero, compare the unshifted value:
2302 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2303 if (Shr->isExact())
2304 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2305
2306 if (C.isNullValue()) {
2307 // == 0 is u< 1.
2308 if (Pred == CmpInst::ICMP_EQ)
2309 return new ICmpInst(CmpInst::ICMP_ULT, X,
2310 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2311 else
2312 return new ICmpInst(CmpInst::ICMP_UGT, X,
2313 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2314 }
2315
2316 if (Shr->hasOneUse()) {
2317 // Canonicalize the shift into an 'and':
2318 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2319 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2320 Constant *Mask = ConstantInt::get(ShrTy, Val);
2321 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2322 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2323 }
2324
2325 return nullptr;
2326}
2327
2328Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2329 BinaryOperator *SRem,
2330 const APInt &C) {
2331 // Match an 'is positive' or 'is negative' comparison of remainder by a
2332 // constant power-of-2 value:
2333 // (X % pow2C) sgt/slt 0
2334 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2335 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2336 return nullptr;
2337
2338 // TODO: The one-use check is standard because we do not typically want to
2339 // create longer instruction sequences, but this might be a special-case
2340 // because srem is not good for analysis or codegen.
2341 if (!SRem->hasOneUse())
2342 return nullptr;
2343
2344 const APInt *DivisorC;
2345 if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2346 return nullptr;
2347
2348 // Mask off the sign bit and the modulo bits (low-bits).
2349 Type *Ty = SRem->getType();
2350 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2351 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2352 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2353
2354 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2355 // bit is set. Example:
2356 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2357 if (Pred == ICmpInst::ICMP_SGT)
2358 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2359
2360 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2361 // bit is set. Example:
2362 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2363 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2364}
2365
2366/// Fold icmp (udiv X, Y), C.
2367Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2368 BinaryOperator *UDiv,
2369 const APInt &C) {
2370 const APInt *C2;
2371 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2372 return nullptr;
2373
2374 assert(*C2 != 0 && "udiv 0, X should have been simplified already.")(static_cast <bool> (*C2 != 0 && "udiv 0, X should have been simplified already."
) ? void (0) : __assert_fail ("*C2 != 0 && \"udiv 0, X should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2374, __extension__ __PRETTY_FUNCTION__))
;
2375
2376 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2377 Value *Y = UDiv->getOperand(1);
2378 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2379 assert(!C.isMaxValue() &&(static_cast <bool> (!C.isMaxValue() && "icmp ugt X, UINT_MAX should have been simplified already."
) ? void (0) : __assert_fail ("!C.isMaxValue() && \"icmp ugt X, UINT_MAX should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2380, __extension__ __PRETTY_FUNCTION__))
2380 "icmp ugt X, UINT_MAX should have been simplified already.")(static_cast <bool> (!C.isMaxValue() && "icmp ugt X, UINT_MAX should have been simplified already."
) ? void (0) : __assert_fail ("!C.isMaxValue() && \"icmp ugt X, UINT_MAX should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2380, __extension__ __PRETTY_FUNCTION__))
;
2381 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2382 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2383 }
2384
2385 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2386 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2387 assert(C != 0 && "icmp ult X, 0 should have been simplified already.")(static_cast <bool> (C != 0 && "icmp ult X, 0 should have been simplified already."
) ? void (0) : __assert_fail ("C != 0 && \"icmp ult X, 0 should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2387, __extension__ __PRETTY_FUNCTION__))
;
2388 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2389 ConstantInt::get(Y->getType(), C2->udiv(C)));
2390 }
2391
2392 return nullptr;
2393}
2394
2395/// Fold icmp ({su}div X, Y), C.
2396Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2397 BinaryOperator *Div,
2398 const APInt &C) {
2399 // Fold: icmp pred ([us]div X, C2), C -> range test
2400 // Fold this div into the comparison, producing a range check.
2401 // Determine, based on the divide type, what the range is being
2402 // checked. If there is an overflow on the low or high side, remember
2403 // it, otherwise compute the range [low, hi) bounding the new value.
2404 // See: InsertRangeTest above for the kinds of replacements possible.
2405 const APInt *C2;
2406 if (!match(Div->getOperand(1), m_APInt(C2)))
2407 return nullptr;
2408
2409 // FIXME: If the operand types don't match the type of the divide
2410 // then don't attempt this transform. The code below doesn't have the
2411 // logic to deal with a signed divide and an unsigned compare (and
2412 // vice versa). This is because (x /s C2) <s C produces different
2413 // results than (x /s C2) <u C or (x /u C2) <s C or even
2414 // (x /u C2) <u C. Simply casting the operands and result won't
2415 // work. :( The if statement below tests that condition and bails
2416 // if it finds it.
2417 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2418 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2419 return nullptr;
2420
2421 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2422 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2423 // division-by-constant cases should be present, we can not assert that they
2424 // have happened before we reach this icmp instruction.
2425 if (C2->isNullValue() || C2->isOneValue() ||
2426 (DivIsSigned && C2->isAllOnesValue()))
2427 return nullptr;
2428
2429 // Compute Prod = C * C2. We are essentially solving an equation of
2430 // form X / C2 = C. We solve for X by multiplying C2 and C.
2431 // By solving for X, we can turn this into a range check instead of computing
2432 // a divide.
2433 APInt Prod = C * *C2;
2434
2435 // Determine if the product overflows by seeing if the product is not equal to
2436 // the divide. Make sure we do the same kind of divide as in the LHS
2437 // instruction that we're folding.
2438 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2439
2440 ICmpInst::Predicate Pred = Cmp.getPredicate();
2441
2442 // If the division is known to be exact, then there is no remainder from the
2443 // divide, so the covered range size is unit, otherwise it is the divisor.
2444 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2445
2446 // Figure out the interval that is being checked. For example, a comparison
2447 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2448 // Compute this interval based on the constants involved and the signedness of
2449 // the compare/divide. This computes a half-open interval, keeping track of
2450 // whether either value in the interval overflows. After analysis each
2451 // overflow variable is set to 0 if it's corresponding bound variable is valid
2452 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2453 int LoOverflow = 0, HiOverflow = 0;
2454 APInt LoBound, HiBound;
2455
2456 if (!DivIsSigned) { // udiv
2457 // e.g. X/5 op 3 --> [15, 20)
2458 LoBound = Prod;
2459 HiOverflow = LoOverflow = ProdOV;
2460 if (!HiOverflow) {
2461 // If this is not an exact divide, then many values in the range collapse
2462 // to the same result value.
2463 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2464 }
2465 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2466 if (C.isNullValue()) { // (X / pos) op 0
2467 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2468 LoBound = -(RangeSize - 1);
2469 HiBound = RangeSize;
2470 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2471 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2472 HiOverflow = LoOverflow = ProdOV;
2473 if (!HiOverflow)
2474 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2475 } else { // (X / pos) op neg
2476 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2477 HiBound = Prod + 1;
2478 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2479 if (!LoOverflow) {
2480 APInt DivNeg = -RangeSize;
2481 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2482 }
2483 }
2484 } else if (C2->isNegative()) { // Divisor is < 0.
2485 if (Div->isExact())
2486 RangeSize.negate();
2487 if (C.isNullValue()) { // (X / neg) op 0
2488 // e.g. X/-5 op 0 --> [-4, 5)
2489 LoBound = RangeSize + 1;
2490 HiBound = -RangeSize;
2491 if (HiBound == *C2) { // -INTMIN = INTMIN
2492 HiOverflow = 1; // [INTMIN+1, overflow)
2493 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2494 }
2495 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2496 // e.g. X/-5 op 3 --> [-19, -14)
2497 HiBound = Prod + 1;
2498 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2499 if (!LoOverflow)
2500 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2501 } else { // (X / neg) op neg
2502 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2503 LoOverflow = HiOverflow = ProdOV;
2504 if (!HiOverflow)
2505 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2506 }
2507
2508 // Dividing by a negative swaps the condition. LT <-> GT
2509 Pred = ICmpInst::getSwappedPredicate(Pred);
2510 }
2511
2512 Value *X = Div->getOperand(0);
2513 switch (Pred) {
2514 default: llvm_unreachable("Unhandled icmp opcode!")::llvm::llvm_unreachable_internal("Unhandled icmp opcode!", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2514)
;
2515 case ICmpInst::ICMP_EQ:
2516 if (LoOverflow && HiOverflow)
2517 return replaceInstUsesWith(Cmp, Builder.getFalse());
2518 if (HiOverflow)
2519 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2520 ICmpInst::ICMP_UGE, X,
2521 ConstantInt::get(Div->getType(), LoBound));
2522 if (LoOverflow)
2523 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2524 ICmpInst::ICMP_ULT, X,
2525 ConstantInt::get(Div->getType(), HiBound));
2526 return replaceInstUsesWith(
2527 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2528 case ICmpInst::ICMP_NE:
2529 if (LoOverflow && HiOverflow)
2530 return replaceInstUsesWith(Cmp, Builder.getTrue());
2531 if (HiOverflow)
2532 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2533 ICmpInst::ICMP_ULT, X,
2534 ConstantInt::get(Div->getType(), LoBound));
2535 if (LoOverflow)
2536 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2537 ICmpInst::ICMP_UGE, X,
2538 ConstantInt::get(Div->getType(), HiBound));
2539 return replaceInstUsesWith(Cmp,
2540 insertRangeTest(X, LoBound, HiBound,
2541 DivIsSigned, false));
2542 case ICmpInst::ICMP_ULT:
2543 case ICmpInst::ICMP_SLT:
2544 if (LoOverflow == +1) // Low bound is greater than input range.
2545 return replaceInstUsesWith(Cmp, Builder.getTrue());
2546 if (LoOverflow == -1) // Low bound is less than input range.
2547 return replaceInstUsesWith(Cmp, Builder.getFalse());
2548 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2549 case ICmpInst::ICMP_UGT:
2550 case ICmpInst::ICMP_SGT:
2551 if (HiOverflow == +1) // High bound greater than input range.
2552 return replaceInstUsesWith(Cmp, Builder.getFalse());
2553 if (HiOverflow == -1) // High bound less than input range.
2554 return replaceInstUsesWith(Cmp, Builder.getTrue());
2555 if (Pred == ICmpInst::ICMP_UGT)
2556 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2557 ConstantInt::get(Div->getType(), HiBound));
2558 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2559 ConstantInt::get(Div->getType(), HiBound));
2560 }
2561
2562 return nullptr;
2563}
2564
2565/// Fold icmp (sub X, Y), C.
2566Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2567 BinaryOperator *Sub,
2568 const APInt &C) {
2569 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2570 ICmpInst::Predicate Pred = Cmp.getPredicate();
2571 const APInt *C2;
2572 APInt SubResult;
2573
2574 // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2575 if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2576 return new ICmpInst(Cmp.getPredicate(), Y,
2577 ConstantInt::get(Y->getType(), 0));
2578
2579 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2580 if (match(X, m_APInt(C2)) &&
2581 ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2582 (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2583 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2584 return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2585 ConstantInt::get(Y->getType(), SubResult));
2586
2587 // The following transforms are only worth it if the only user of the subtract
2588 // is the icmp.
2589 if (!Sub->hasOneUse())
2590 return nullptr;
2591
2592 if (Sub->hasNoSignedWrap()) {
2593 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2594 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2595 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2596
2597 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2598 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2599 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2600
2601 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2602 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2603 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2604
2605 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2606 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2607 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2608 }
2609
2610 if (!match(X, m_APInt(C2)))
2611 return nullptr;
2612
2613 // C2 - Y <u C -> (Y | (C - 1)) == C2
2614 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2615 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2616 (*C2 & (C - 1)) == (C - 1))
2617 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2618
2619 // C2 - Y >u C -> (Y | C) != C2
2620 // iff C2 & C == C and C + 1 is a power of 2
2621 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2622 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2623
2624 return nullptr;
2625}
2626
2627/// Fold icmp (add X, Y), C.
2628Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
2629 BinaryOperator *Add,
2630 const APInt &C) {
2631 Value *Y = Add->getOperand(1);
2632 const APInt *C2;
2633 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2634 return nullptr;
2635
2636 // Fold icmp pred (add X, C2), C.
2637 Value *X = Add->getOperand(0);
2638 Type *Ty = Add->getType();
2639 CmpInst::Predicate Pred = Cmp.getPredicate();
2640
2641 // If the add does not wrap, we can always adjust the compare by subtracting
2642 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2643 // are canonicalized to SGT/SLT/UGT/ULT.
2644 if ((Add->hasNoSignedWrap() &&
2645 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2646 (Add->hasNoUnsignedWrap() &&
2647 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2648 bool Overflow;
2649 APInt NewC =
2650 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2651 // If there is overflow, the result must be true or false.
2652 // TODO: Can we assert there is no overflow because InstSimplify always
2653 // handles those cases?
2654 if (!Overflow)
2655 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2656 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2657 }
2658
2659 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2660 const APInt &Upper = CR.getUpper();
2661 const APInt &Lower = CR.getLower();
2662 if (Cmp.isSigned()) {
2663 if (Lower.isSignMask())
2664 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2665 if (Upper.isSignMask())
2666 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2667 } else {
2668 if (Lower.isMinValue())
2669 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2670 if (Upper.isMinValue())
2671 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2672 }
2673
2674 if (!Add->hasOneUse())
2675 return nullptr;
2676
2677 // X+C <u C2 -> (X & -C2) == C
2678 // iff C & (C2-1) == 0
2679 // C2 is a power of 2
2680 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2681 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2682 ConstantExpr::getNeg(cast<Constant>(Y)));
2683
2684 // X+C >u C2 -> (X & ~C2) != C
2685 // iff C & C2 == 0
2686 // C2+1 is a power of 2
2687 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2688 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2689 ConstantExpr::getNeg(cast<Constant>(Y)));
2690
2691 return nullptr;
2692}
2693
2694bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2695 Value *&RHS, ConstantInt *&Less,
2696 ConstantInt *&Equal,
2697 ConstantInt *&Greater) {
2698 // TODO: Generalize this to work with other comparison idioms or ensure
2699 // they get canonicalized into this form.
2700
2701 // select i1 (a == b),
2702 // i32 Equal,
2703 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2704 // where Equal, Less and Greater are placeholders for any three constants.
2705 ICmpInst::Predicate PredA;
2706 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2707 !ICmpInst::isEquality(PredA))
2708 return false;
2709 Value *EqualVal = SI->getTrueValue();
2710 Value *UnequalVal = SI->getFalseValue();
2711 // We still can get non-canonical predicate here, so canonicalize.
2712 if (PredA == ICmpInst::ICMP_NE)
2713 std::swap(EqualVal, UnequalVal);
2714 if (!match(EqualVal, m_ConstantInt(Equal)))
2715 return false;
2716 ICmpInst::Predicate PredB;
2717 Value *LHS2, *RHS2;
2718 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2719 m_ConstantInt(Less), m_ConstantInt(Greater))))
2720 return false;
2721 // We can get predicate mismatch here, so canonicalize if possible:
2722 // First, ensure that 'LHS' match.
2723 if (LHS2 != LHS) {
2724 // x sgt y <--> y slt x
2725 std::swap(LHS2, RHS2);
2726 PredB = ICmpInst::getSwappedPredicate(PredB);
2727 }
2728 if (LHS2 != LHS)
2729 return false;
2730 // We also need to canonicalize 'RHS'.
2731 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2732 // x sgt C-1 <--> x sge C <--> not(x slt C)
2733 auto FlippedStrictness =
2734 InstCombiner::getFlippedStrictnessPredicateAndConstant(
2735 PredB, cast<Constant>(RHS2));
2736 if (!FlippedStrictness)
2737 return false;
2738 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check")(static_cast <bool> (FlippedStrictness->first == ICmpInst
::ICMP_SGE && "Sanity check") ? void (0) : __assert_fail
("FlippedStrictness->first == ICmpInst::ICMP_SGE && \"Sanity check\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2738, __extension__ __PRETTY_FUNCTION__))
;
2739 RHS2 = FlippedStrictness->second;
2740 // And kind-of perform the result swap.
2741 std::swap(Less, Greater);
2742 PredB = ICmpInst::ICMP_SLT;
2743 }
2744 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2745}
2746
2747Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
2748 SelectInst *Select,
2749 ConstantInt *C) {
2750
2751 assert(C && "Cmp RHS should be a constant int!")(static_cast <bool> (C && "Cmp RHS should be a constant int!"
) ? void (0) : __assert_fail ("C && \"Cmp RHS should be a constant int!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2751, __extension__ __PRETTY_FUNCTION__))
;
2752 // If we're testing a constant value against the result of a three way
2753 // comparison, the result can be expressed directly in terms of the
2754 // original values being compared. Note: We could possibly be more
2755 // aggressive here and remove the hasOneUse test. The original select is
2756 // really likely to simplify or sink when we remove a test of the result.
2757 Value *OrigLHS, *OrigRHS;
2758 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2759 if (Cmp.hasOneUse() &&
2760 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2761 C3GreaterThan)) {
2762 assert(C1LessThan && C2Equal && C3GreaterThan)(static_cast <bool> (C1LessThan && C2Equal &&
C3GreaterThan) ? void (0) : __assert_fail ("C1LessThan && C2Equal && C3GreaterThan"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2762, __extension__ __PRETTY_FUNCTION__))
;
2763
2764 bool TrueWhenLessThan =
2765 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2766 ->isAllOnesValue();
2767 bool TrueWhenEqual =
2768 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2769 ->isAllOnesValue();
2770 bool TrueWhenGreaterThan =
2771 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2772 ->isAllOnesValue();
2773
2774 // This generates the new instruction that will replace the original Cmp
2775 // Instruction. Instead of enumerating the various combinations when
2776 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2777 // false, we rely on chaining of ORs and future passes of InstCombine to
2778 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2779
2780 // When none of the three constants satisfy the predicate for the RHS (C),
2781 // the entire original Cmp can be simplified to a false.
2782 Value *Cond = Builder.getFalse();
2783 if (TrueWhenLessThan)
2784 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2785 OrigLHS, OrigRHS));
2786 if (TrueWhenEqual)
2787 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2788 OrigLHS, OrigRHS));
2789 if (TrueWhenGreaterThan)
2790 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2791 OrigLHS, OrigRHS));
2792
2793 return replaceInstUsesWith(Cmp, Cond);
2794 }
2795 return nullptr;
2796}
2797
2798static Instruction *foldICmpBitCast(ICmpInst &Cmp,
2799 InstCombiner::BuilderTy &Builder) {
2800 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2801 if (!Bitcast)
2802 return nullptr;
2803
2804 ICmpInst::Predicate Pred = Cmp.getPredicate();
2805 Value *Op1 = Cmp.getOperand(1);
2806 Value *BCSrcOp = Bitcast->getOperand(0);
2807
2808 // Make sure the bitcast doesn't change the number of vector elements.
2809 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2810 Bitcast->getDestTy()->getScalarSizeInBits()) {
2811 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2812 Value *X;
2813 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2814 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2815 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2816 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2817 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2818 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2819 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2820 match(Op1, m_Zero()))
2821 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2822
2823 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2824 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2825 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2826
2827 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2828 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2829 return new ICmpInst(Pred, X,
2830 ConstantInt::getAllOnesValue(X->getType()));
2831 }
2832
2833 // Zero-equality checks are preserved through unsigned floating-point casts:
2834 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2835 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2836 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2837 if (Cmp.isEquality() && match(Op1, m_Zero()))
2838 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2839
2840 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2841 // the FP extend/truncate because that cast does not change the sign-bit.
2842 // This is true for all standard IEEE-754 types and the X86 80-bit type.
2843 // The sign-bit is always the most significant bit in those types.
2844 const APInt *C;
2845 bool TrueIfSigned;
2846 if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2847 InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2848 if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2849 match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2850 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2851 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2852 Type *XType = X->getType();
2853
2854 // We can't currently handle Power style floating point operations here.
2855 if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2856
2857 Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2858 if (auto *XVTy = dyn_cast<VectorType>(XType))
2859 NewType = VectorType::get(NewType, XVTy->getElementCount());
2860 Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2861 if (TrueIfSigned)
2862 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2863 ConstantInt::getNullValue(NewType));
2864 else
2865 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2866 ConstantInt::getAllOnesValue(NewType));
2867 }
2868 }
2869 }
2870 }
2871
2872 // Test to see if the operands of the icmp are casted versions of other
2873 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2874 if (Bitcast->getType()->isPointerTy() &&
2875 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2876 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2877 // so eliminate it as well.
2878 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2879 Op1 = BC2->getOperand(0);
2880
2881 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2882 return new ICmpInst(Pred, BCSrcOp, Op1);
2883 }
2884
2885 // Folding: icmp <pred> iN X, C
2886 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2887 // and C is a splat of a K-bit pattern
2888 // and SC is a constant vector = <C', C', C', ..., C'>
2889 // Into:
2890 // %E = extractelement <M x iK> %vec, i32 C'
2891 // icmp <pred> iK %E, trunc(C)
2892 const APInt *C;
2893 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2894 !Bitcast->getType()->isIntegerTy() ||
2895 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2896 return nullptr;
2897
2898 Value *Vec;
2899 ArrayRef<int> Mask;
2900 if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2901 // Check whether every element of Mask is the same constant
2902 if (is_splat(Mask)) {
2903 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2904 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2905 if (C->isSplat(EltTy->getBitWidth())) {
2906 // Fold the icmp based on the value of C
2907 // If C is M copies of an iK sized bit pattern,
2908 // then:
2909 // => %E = extractelement <N x iK> %vec, i32 Elem
2910 // icmp <pred> iK %SplatVal, <pattern>
2911 Value *Elem = Builder.getInt32(Mask[0]);
2912 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2913 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2914 return new ICmpInst(Pred, Extract, NewC);
2915 }
2916 }
2917 }
2918 return nullptr;
2919}
2920
2921/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2922/// where X is some kind of instruction.
2923Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
2924 const APInt *C;
2925 if (!match(Cmp.getOperand(1), m_APInt(C)))
2926 return nullptr;
2927
2928 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2929 switch (BO->getOpcode()) {
2930 case Instruction::Xor:
2931 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2932 return I;
2933 break;
2934 case Instruction::And:
2935 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2936 return I;
2937 break;
2938 case Instruction::Or:
2939 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2940 return I;
2941 break;
2942 case Instruction::Mul:
2943 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2944 return I;
2945 break;
2946 case Instruction::Shl:
2947 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2948 return I;
2949 break;
2950 case Instruction::LShr:
2951 case Instruction::AShr:
2952 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2953 return I;
2954 break;
2955 case Instruction::SRem:
2956 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2957 return I;
2958 break;
2959 case Instruction::UDiv:
2960 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2961 return I;
2962 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2963 case Instruction::SDiv:
2964 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2965 return I;
2966 break;
2967 case Instruction::Sub:
2968 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2969 return I;
2970 break;
2971 case Instruction::Add:
2972 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2973 return I;
2974 break;
2975 default:
2976 break;
2977 }
2978 // TODO: These folds could be refactored to be part of the above calls.
2979 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2980 return I;
2981 }
2982
2983 // Match against CmpInst LHS being instructions other than binary operators.
2984
2985 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2986 // For now, we only support constant integers while folding the
2987 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2988 // similar to the cases handled by binary ops above.
2989 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2990 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2991 return I;
2992 }
2993
2994 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2995 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2996 return I;
2997 }
2998
2999 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
3000 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
3001 return I;
3002
3003 return nullptr;
3004}
3005
3006/// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3007/// icmp eq/ne BO, C.
3008Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
3009 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3010 // TODO: Some of these folds could work with arbitrary constants, but this
3011 // function is limited to scalar and vector splat constants.
3012 if (!Cmp.isEquality())
3013 return nullptr;
3014
3015 ICmpInst::Predicate Pred = Cmp.getPredicate();
3016 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3017 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
3018 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3019
3020 switch (BO->getOpcode()) {
3021 case Instruction::SRem:
3022 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3023 if (C.isNullValue() && BO->hasOneUse()) {
3024 const APInt *BOC;
3025 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3026 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3027 return new ICmpInst(Pred, NewRem,
3028 Constant::getNullValue(BO->getType()));
3029 }
3030 }
3031 break;
3032 case Instruction::Add: {
3033 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3034 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3035 if (BO->hasOneUse())
3036 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3037 } else if (C.isNullValue()) {
3038 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3039 // efficiently invertible, or if the add has just this one use.
3040 if (Value *NegVal = dyn_castNegVal(BOp1))
3041 return new ICmpInst(Pred, BOp0, NegVal);
3042 if (Value *NegVal = dyn_castNegVal(BOp0))
3043 return new ICmpInst(Pred, NegVal, BOp1);
3044 if (BO->hasOneUse()) {
3045 Value *Neg = Builder.CreateNeg(BOp1);
3046 Neg->takeName(BO);
3047 return new ICmpInst(Pred, BOp0, Neg);
3048 }
3049 }
3050 break;
3051 }
3052 case Instruction::Xor:
3053 if (BO->hasOneUse()) {
3054 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3055 // For the xor case, we can xor two constants together, eliminating
3056 // the explicit xor.
3057 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3058 } else if (C.isNullValue()) {
3059 // Replace ((xor A, B) != 0) with (A != B)
3060 return new ICmpInst(Pred, BOp0, BOp1);
3061 }
3062 }
3063 break;
3064 case Instruction::Sub:
3065 if (BO->hasOneUse()) {
3066 // Only check for constant LHS here, as constant RHS will be canonicalized
3067 // to add and use the fold above.
3068 if (Constant *BOC = dyn_cast<Constant>(BOp0)) {
3069 // Replace ((sub BOC, B) != C) with (B != BOC-C).
3070 return new ICmpInst(Pred, BOp1, ConstantExpr::getSub(BOC, RHS));
3071 } else if (C.isNullValue()) {
3072 // Replace ((sub A, B) != 0) with (A != B).
3073 return new ICmpInst(Pred, BOp0, BOp1);
3074 }
3075 }
3076 break;
3077 case Instruction::Or: {
3078 const APInt *BOC;
3079 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3080 // Comparing if all bits outside of a constant mask are set?
3081 // Replace (X | C) == -1 with (X & ~C) == ~C.
3082 // This removes the -1 constant.
3083 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3084 Value *And = Builder.CreateAnd(BOp0, NotBOC);
3085 return new ICmpInst(Pred, And, NotBOC);
3086 }
3087 break;
3088 }
3089 case Instruction::And: {
3090 const APInt *BOC;
3091 if (match(BOp1, m_APInt(BOC))) {
3092 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3093 if (C == *BOC && C.isPowerOf2())
3094 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3095 BO, Constant::getNullValue(RHS->getType()));
3096 }
3097 break;
3098 }
3099 case Instruction::UDiv:
3100 if (C.isNullValue()) {
3101 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3102 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3103 return new ICmpInst(NewPred, BOp1, BOp0);
3104 }
3105 break;
3106 default:
3107 break;
3108 }
3109 return nullptr;
3110}
3111
3112/// Fold an equality icmp with LLVM intrinsic and constant operand.
3113Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3114 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3115 Type *Ty = II->getType();
3116 unsigned BitWidth = C.getBitWidth();
3117 switch (II->getIntrinsicID()) {
3118 case Intrinsic::abs:
3119 // abs(A) == 0 -> A == 0
3120 // abs(A) == INT_MIN -> A == INT_MIN
3121 if (C.isNullValue() || C.isMinSignedValue())
3122 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3123 ConstantInt::get(Ty, C));
3124 break;
3125
3126 case Intrinsic::bswap:
3127 // bswap(A) == C -> A == bswap(C)
3128 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3129 ConstantInt::get(Ty, C.byteSwap()));
3130
3131 case Intrinsic::ctlz:
3132 case Intrinsic::cttz: {
3133 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3134 if (C == BitWidth)
3135 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3136 ConstantInt::getNullValue(Ty));
3137
3138 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3139 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3140 // Limit to one use to ensure we don't increase instruction count.
3141 unsigned Num = C.getLimitedValue(BitWidth);
3142 if (Num != BitWidth && II->hasOneUse()) {
3143 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3144 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3145 : APInt::getHighBitsSet(BitWidth, Num + 1);
3146 APInt Mask2 = IsTrailing
3147 ? APInt::getOneBitSet(BitWidth, Num)
3148 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3149 return new ICmpInst(Cmp.getPredicate(),
3150 Builder.CreateAnd(II->getArgOperand(0), Mask1),
3151 ConstantInt::get(Ty, Mask2));
3152 }
3153 break;
3154 }
3155
3156 case Intrinsic::ctpop: {
3157 // popcount(A) == 0 -> A == 0 and likewise for !=
3158 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3159 bool IsZero = C.isNullValue();
3160 if (IsZero || C == BitWidth)
3161 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3162 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty));
3163
3164 break;
3165 }
3166
3167 case Intrinsic::uadd_sat: {
3168 // uadd.sat(a, b) == 0 -> (a | b) == 0
3169 if (C.isNullValue()) {
3170 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3171 return new ICmpInst(Cmp.getPredicate(), Or, Constant::getNullValue(Ty));
3172 }
3173 break;
3174 }
3175
3176 case Intrinsic::usub_sat: {
3177 // usub.sat(a, b) == 0 -> a <= b
3178 if (C.isNullValue()) {
3179 ICmpInst::Predicate NewPred = Cmp.getPredicate() == ICmpInst::ICMP_EQ
3180 ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3181 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3182 }
3183 break;
3184 }
3185 default:
3186 break;
3187 }
3188
3189 return nullptr;
3190}
3191
3192/// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3193Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3194 IntrinsicInst *II,
3195 const APInt &C) {
3196 if (Cmp.isEquality())
3197 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3198
3199 Type *Ty = II->getType();
3200 unsigned BitWidth = C.getBitWidth();
3201 ICmpInst::Predicate Pred = Cmp.getPredicate();
3202 switch (II->getIntrinsicID()) {
3203 case Intrinsic::ctpop: {
3204 // (ctpop X > BitWidth - 1) --> X == -1
3205 Value *X = II->getArgOperand(0);
3206 if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
3207 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
3208 ConstantInt::getAllOnesValue(Ty));
3209 // (ctpop X < BitWidth) --> X != -1
3210 if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
3211 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
3212 ConstantInt::getAllOnesValue(Ty));
3213 break;
3214 }
3215 case Intrinsic::ctlz: {
3216 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3217 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3218 unsigned Num = C.getLimitedValue();
3219 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3220 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3221 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3222 }
3223
3224 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3225 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3226 unsigned Num = C.getLimitedValue();
3227 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3228 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3229 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3230 }
3231 break;
3232 }
3233 case Intrinsic::cttz: {
3234 // Limit to one use to ensure we don't increase instruction count.
3235 if (!II->hasOneUse())
3236 return nullptr;
3237
3238 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3239 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3240 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3241 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3242 Builder.CreateAnd(II->getArgOperand(0), Mask),
3243 ConstantInt::getNullValue(Ty));
3244 }
3245
3246 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3247 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3248 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3249 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3250 Builder.CreateAnd(II->getArgOperand(0), Mask),
3251 ConstantInt::getNullValue(Ty));
3252 }
3253 break;
3254 }
3255 default:
3256 break;
3257 }
3258
3259 return nullptr;
3260}
3261
3262/// Handle icmp with constant (but not simple integer constant) RHS.
3263Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3264 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3265 Constant *RHSC = dyn_cast<Constant>(Op1);
3266 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3267 if (!RHSC || !LHSI)
3268 return nullptr;
3269
3270 switch (LHSI->getOpcode()) {
3271 case Instruction::GetElementPtr:
3272 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3273 if (RHSC->isNullValue() &&
3274 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3275 return new ICmpInst(
3276 I.getPredicate(), LHSI->getOperand(0),
3277 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3278 break;
3279 case Instruction::PHI:
3280 // Only fold icmp into the PHI if the phi and icmp are in the same
3281 // block. If in the same block, we're encouraging jump threading. If
3282 // not, we are just pessimizing the code by making an i1 phi.
3283 if (LHSI->getParent() == I.getParent())
3284 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3285 return NV;
3286 break;
3287 case Instruction::Select: {
3288 // If either operand of the select is a constant, we can fold the
3289 // comparison into the select arms, which will cause one to be
3290 // constant folded and the select turned into a bitwise or.
3291 Value *Op1 = nullptr, *Op2 = nullptr;
3292 ConstantInt *CI = nullptr;
3293
3294 auto SimplifyOp = [&](Value *V) {
3295 Value *Op = nullptr;
3296 if (Constant *C = dyn_cast<Constant>(V)) {
3297 Op = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3298 } else if (RHSC->isNullValue()) {
3299 // If null is being compared, check if it can be further simplified.
3300 Op = SimplifyICmpInst(I.getPredicate(), V, RHSC, SQ);
3301 }
3302 return Op;
3303 };
3304 Op1 = SimplifyOp(LHSI->getOperand(1));
3305 if (Op1)
3306 CI = dyn_cast<ConstantInt>(Op1);
3307
3308 Op2 = SimplifyOp(LHSI->getOperand(2));
3309 if (Op2)
3310 CI = dyn_cast<ConstantInt>(Op2);
3311
3312 // We only want to perform this transformation if it will not lead to
3313 // additional code. This is true if either both sides of the select
3314 // fold to a constant (in which case the icmp is replaced with a select
3315 // which will usually simplify) or this is the only user of the
3316 // select (in which case we are trading a select+icmp for a simpler
3317 // select+icmp) or all uses of the select can be replaced based on
3318 // dominance information ("Global cases").
3319 bool Transform = false;
3320 if (Op1 && Op2)
3321 Transform = true;
3322 else if (Op1 || Op2) {
3323 // Local case
3324 if (LHSI->hasOneUse())
3325 Transform = true;
3326 // Global cases
3327 else if (CI && !CI->isZero())
3328 // When Op1 is constant try replacing select with second operand.
3329 // Otherwise Op2 is constant and try replacing select with first
3330 // operand.
3331 Transform =
3332 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3333 }
3334 if (Transform) {
3335 if (!Op1)
3336 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3337 I.getName());
3338 if (!Op2)
3339 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3340 I.getName());
3341 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3342 }
3343 break;
3344 }
3345 case Instruction::IntToPtr:
3346 // icmp pred inttoptr(X), null -> icmp pred X, 0
3347 if (RHSC->isNullValue() &&
3348 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3349 return new ICmpInst(
3350 I.getPredicate(), LHSI->getOperand(0),
3351 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3352 break;
3353
3354 case Instruction::Load:
3355 // Try to optimize things like "A[i] > 4" to index computations.
3356 if (GetElementPtrInst *GEP =
3357 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3358 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3359 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3360 !cast<LoadInst>(LHSI)->isVolatile())
3361 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3362 return Res;
3363 }
3364 break;
3365 }
3366
3367 return nullptr;
3368}
3369
3370/// Some comparisons can be simplified.
3371/// In this case, we are looking for comparisons that look like
3372/// a check for a lossy truncation.
3373/// Folds:
3374/// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3375/// Where Mask is some pattern that produces all-ones in low bits:
3376/// (-1 >> y)
3377/// ((-1 << y) >> y) <- non-canonical, has extra uses
3378/// ~(-1 << y)
3379/// ((1 << y) + (-1)) <- non-canonical, has extra uses
3380/// The Mask can be a constant, too.
3381/// For some predicates, the operands are commutative.
3382/// For others, x can only be on a specific side.
3383static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3384 InstCombiner::BuilderTy &Builder) {
3385 ICmpInst::Predicate SrcPred;
3386 Value *X, *M, *Y;
3387 auto m_VariableMask = m_CombineOr(
3388 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3389 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3390 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3391 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3392 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3393 if (!match(&I, m_c_ICmp(SrcPred,
3394 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3395 m_Deferred(X))))
3396 return nullptr;
3397
3398 ICmpInst::Predicate DstPred;
3399 switch (SrcPred) {
3400 case ICmpInst::Predicate::ICMP_EQ:
3401 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3402 DstPred = ICmpInst::Predicate::ICMP_ULE;
3403 break;
3404 case ICmpInst::Predicate::ICMP_NE:
3405 // x & (-1 >> y) != x -> x u> (-1 >> y)
3406 DstPred = ICmpInst::Predicate::ICMP_UGT;
3407 break;
3408 case ICmpInst::Predicate::ICMP_ULT:
3409 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3410 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3411 DstPred = ICmpInst::Predicate::ICMP_UGT;
3412 break;
3413 case ICmpInst::Predicate::ICMP_UGE:
3414 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3415 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3416 DstPred = ICmpInst::Predicate::ICMP_ULE;
3417 break;
3418 case ICmpInst::Predicate::ICMP_SLT:
3419 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3420 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3421 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3422 return nullptr;
3423 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3424 return nullptr;
3425 DstPred = ICmpInst::Predicate::ICMP_SGT;
3426 break;
3427 case ICmpInst::Predicate::ICMP_SGE:
3428 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3429 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3430 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3431 return nullptr;
3432 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3433 return nullptr;
3434 DstPred = ICmpInst::Predicate::ICMP_SLE;
3435 break;
3436 case ICmpInst::Predicate::ICMP_SGT:
3437 case ICmpInst::Predicate::ICMP_SLE:
3438 return nullptr;
3439 case ICmpInst::Predicate::ICMP_UGT:
3440 case ICmpInst::Predicate::ICMP_ULE:
3441 llvm_unreachable("Instsimplify took care of commut. variant")::llvm::llvm_unreachable_internal("Instsimplify took care of commut. variant"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3441)
;
3442 break;
3443 default:
3444 llvm_unreachable("All possible folds are handled.")::llvm::llvm_unreachable_internal("All possible folds are handled."
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3444)
;
3445 }
3446
3447 // The mask value may be a vector constant that has undefined elements. But it
3448 // may not be safe to propagate those undefs into the new compare, so replace
3449 // those elements by copying an existing, defined, and safe scalar constant.
3450 Type *OpTy = M->getType();
3451 auto *VecC = dyn_cast<Constant>(M);
3452 auto *OpVTy = dyn_cast<FixedVectorType>(OpTy);
3453 if (OpVTy && VecC && VecC->containsUndefOrPoisonElement()) {
3454 Constant *SafeReplacementConstant = nullptr;
3455 for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3456 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3457 SafeReplacementConstant = VecC->getAggregateElement(i);
3458 break;
3459 }
3460 }
3461 assert(SafeReplacementConstant && "Failed to find undef replacement")(static_cast <bool> (SafeReplacementConstant &&
"Failed to find undef replacement") ? void (0) : __assert_fail
("SafeReplacementConstant && \"Failed to find undef replacement\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3461, __extension__ __PRETTY_FUNCTION__))
;
3462 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3463 }
3464
3465 return Builder.CreateICmp(DstPred, X, M);
3466}
3467
3468/// Some comparisons can be simplified.
3469/// In this case, we are looking for comparisons that look like
3470/// a check for a lossy signed truncation.
3471/// Folds: (MaskedBits is a constant.)
3472/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3473/// Into:
3474/// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3475/// Where KeptBits = bitwidth(%x) - MaskedBits
3476static Value *
3477foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3478 InstCombiner::BuilderTy &Builder) {
3479 ICmpInst::Predicate SrcPred;
3480 Value *X;
3481 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3482 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3483 if (!match(&I, m_c_ICmp(SrcPred,
3484 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3485 m_APInt(C1))),
3486 m_Deferred(X))))
3487 return nullptr;
3488
3489 // Potential handling of non-splats: for each element:
3490 // * if both are undef, replace with constant 0.
3491 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3492 // * if both are not undef, and are different, bailout.
3493 // * else, only one is undef, then pick the non-undef one.
3494
3495 // The shift amount must be equal.
3496 if (*C0 != *C1)
3497 return nullptr;
3498 const APInt &MaskedBits = *C0;
3499 assert(MaskedBits != 0 && "shift by zero should be folded away already.")(static_cast <bool> (MaskedBits != 0 && "shift by zero should be folded away already."
) ? void (0) : __assert_fail ("MaskedBits != 0 && \"shift by zero should be folded away already.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3499, __extension__ __PRETTY_FUNCTION__))
;
3500
3501 ICmpInst::Predicate DstPred;
3502 switch (SrcPred) {
3503 case ICmpInst::Predicate::ICMP_EQ:
3504 // ((%x << MaskedBits) a>> MaskedBits) == %x
3505 // =>
3506 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3507 DstPred = ICmpInst::Predicate::ICMP_ULT;
3508 break;
3509 case ICmpInst::Predicate::ICMP_NE:
3510 // ((%x << MaskedBits) a>> MaskedBits) != %x
3511 // =>
3512 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3513 DstPred = ICmpInst::Predicate::ICMP_UGE;
3514 break;
3515 // FIXME: are more folds possible?
3516 default:
3517 return nullptr;
3518 }
3519
3520 auto *XType = X->getType();
3521 const unsigned XBitWidth = XType->getScalarSizeInBits();
3522 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3523 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched")(static_cast <bool> (BitWidth.ugt(MaskedBits) &&
"shifts should leave some bits untouched") ? void (0) : __assert_fail
("BitWidth.ugt(MaskedBits) && \"shifts should leave some bits untouched\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3523, __extension__ __PRETTY_FUNCTION__))
;
3524
3525 // KeptBits = bitwidth(%x) - MaskedBits
3526 const APInt KeptBits = BitWidth - MaskedBits;
3527 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable")(static_cast <bool> (KeptBits.ugt(0) && KeptBits
.ult(BitWidth) && "unreachable") ? void (0) : __assert_fail
("KeptBits.ugt(0) && KeptBits.ult(BitWidth) && \"unreachable\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3527, __extension__ __PRETTY_FUNCTION__))
;
3528 // ICmpCst = (1 << KeptBits)
3529 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3530 assert(ICmpCst.isPowerOf2())(static_cast <bool> (ICmpCst.isPowerOf2()) ? void (0) :
__assert_fail ("ICmpCst.isPowerOf2()", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3530, __extension__ __PRETTY_FUNCTION__))
;
3531 // AddCst = (1 << (KeptBits-1))
3532 const APInt AddCst = ICmpCst.lshr(1);
3533 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2())(static_cast <bool> (AddCst.ult(ICmpCst) && AddCst
.isPowerOf2()) ? void (0) : __assert_fail ("AddCst.ult(ICmpCst) && AddCst.isPowerOf2()"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3533, __extension__ __PRETTY_FUNCTION__))
;
3534
3535 // T0 = add %x, AddCst
3536 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3537 // T1 = T0 DstPred ICmpCst
3538 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3539
3540 return T1;
3541}
3542
3543// Given pattern:
3544// icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3545// we should move shifts to the same hand of 'and', i.e. rewrite as
3546// icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3547// We are only interested in opposite logical shifts here.
3548// One of the shifts can be truncated.
3549// If we can, we want to end up creating 'lshr' shift.
3550static Value *
3551foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3552 InstCombiner::BuilderTy &Builder) {
3553 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
32
Assuming the condition is false
34
Taking false branch
3554 !I.getOperand(0)->hasOneUse())
33
Assuming the condition is false
3555 return nullptr;
3556
3557 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3558
3559 // Look for an 'and' of two logical shifts, one of which may be truncated.
3560 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3561 Instruction *XShift, *MaybeTruncation, *YShift;
3562 if (!match(
35
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>>'
43
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>>'
44
Taking false branch
3563 I.getOperand(0),
3564 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3565 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3566 m_AnyLogicalShift, m_Instruction(YShift))),
3567 m_Instruction(MaybeTruncation)))))
3568 return nullptr;
3569
3570 // We potentially looked past 'trunc', but only when matching YShift,
3571 // therefore YShift must have the widest type.
3572 Instruction *WidestShift = YShift;
3573 // Therefore XShift must have the shallowest type.
3574 // Or they both have identical types if there was no truncation.
3575 Instruction *NarrowestShift = XShift;
3576
3577 Type *WidestTy = WidestShift->getType();
3578 Type *NarrowestTy = NarrowestShift->getType();
3579 assert(NarrowestTy == I.getOperand(0)->getType() &&(static_cast <bool> (NarrowestTy == I.getOperand(0)->
getType() && "We did not look past any shifts while matching XShift though."
) ? void (0) : __assert_fail ("NarrowestTy == I.getOperand(0)->getType() && \"We did not look past any shifts while matching XShift though.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3580, __extension__ __PRETTY_FUNCTION__))
45
Assuming the condition is true
46
'?' condition is true
3580 "We did not look past any shifts while matching XShift though.")(static_cast <bool> (NarrowestTy == I.getOperand(0)->
getType() && "We did not look past any shifts while matching XShift though."
) ? void (0) : __assert_fail ("NarrowestTy == I.getOperand(0)->getType() && \"We did not look past any shifts while matching XShift though.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3580, __extension__ __PRETTY_FUNCTION__))
;
3581 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
47
Assuming the condition is false
3582
3583 // If YShift is a 'lshr', swap the shifts around.
3584 if (match(YShift, m_LShr(m_Value(), m_Value())))
48
Taking false branch
3585 std::swap(XShift, YShift);
3586
3587 // The shifts must be in opposite directions.
3588 auto XShiftOpcode = XShift->getOpcode();
3589 if (XShiftOpcode == YShift->getOpcode())
49
Assuming the condition is false
50
Taking false branch
3590 return nullptr; // Do not care about same-direction shifts here.
3591
3592 Value *X, *XShAmt, *Y, *YShAmt;
51
'XShAmt' declared without an initial value
3593 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
52
Calling 'm_Value'
56
Returning from 'm_Value'
57
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
65
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
66
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>>>'
68
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>>>'
3594 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3595
3596 // If one of the values being shifted is a constant, then we will end with
3597 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3598 // however, we will need to ensure that we won't increase instruction count.
3599 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
69
Assuming 'X' is a 'Constant'
3600 // At least one of the hands of the 'and' should be one-use shift.
3601 if (!match(I.getOperand(0),
3602 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3603 return nullptr;
3604 if (HadTrunc) {
3605 // Due to the 'trunc', we will need to widen X. For that either the old
3606 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3607 if (!MaybeTruncation->hasOneUse() &&
3608 !NarrowestShift->getOperand(1)->hasOneUse())
3609 return nullptr;
3610 }
3611 }
3612
3613 // We have two shift amounts from two different shifts. The types of those
3614 // shift amounts may not match. If that's the case let's bailout now.
3615 if (XShAmt->getType() != YShAmt->getType())
70
Called C++ object pointer is uninitialized
3616 return nullptr;
3617
3618 // As input, we have the following pattern:
3619 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3620 // We want to rewrite that as:
3621 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3622 // While we know that originally (Q+K) would not overflow
3623 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3624 // shift amounts. so it may now overflow in smaller bitwidth.
3625 // To ensure that does not happen, we need to ensure that the total maximal
3626 // shift amount is still representable in that smaller bit width.
3627 unsigned MaximalPossibleTotalShiftAmount =
3628 (WidestTy->getScalarSizeInBits() - 1) +
3629 (NarrowestTy->getScalarSizeInBits() - 1);
3630 APInt MaximalRepresentableShiftAmount =
3631 APInt::getAllOnesValue(XShAmt->getType()->getScalarSizeInBits());
3632 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3633 return nullptr;
3634
3635 // Can we fold (XShAmt+YShAmt) ?
3636 auto *NewShAmt = dyn_cast_or_null<Constant>(
3637 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3638 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3639 if (!NewShAmt)
3640 return nullptr;
3641 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3642 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3643
3644 // Is the new shift amount smaller than the bit width?
3645 // FIXME: could also rely on ConstantRange.
3646 if (!match(NewShAmt,
3647 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3648 APInt(WidestBitWidth, WidestBitWidth))))
3649 return nullptr;
3650
3651 // An extra legality check is needed if we had trunc-of-lshr.
3652 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3653 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3654 WidestShift]() {
3655 // It isn't obvious whether it's worth it to analyze non-constants here.
3656 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3657 // If *any* of these preconditions matches we can perform the fold.
3658 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3659 ? NewShAmt->getSplatValue()
3660 : NewShAmt;
3661 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3662 if (NewShAmtSplat &&
3663 (NewShAmtSplat->isNullValue() ||
3664 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3665 return true;
3666 // We consider *min* leading zeros so a single outlier
3667 // blocks the transform as opposed to allowing it.
3668 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3669 KnownBits Known = computeKnownBits(C, SQ.DL);
3670 unsigned MinLeadZero = Known.countMinLeadingZeros();
3671 // If the value being shifted has at most lowest bit set we can fold.
3672 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3673 if (MaxActiveBits <= 1)
3674 return true;
3675 // Precondition: NewShAmt u<= countLeadingZeros(C)
3676 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3677 return true;
3678 }
3679 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3680 KnownBits Known = computeKnownBits(C, SQ.DL);
3681 unsigned MinLeadZero = Known.countMinLeadingZeros();
3682 // If the value being shifted has at most lowest bit set we can fold.
3683 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3684 if (MaxActiveBits <= 1)
3685 return true;
3686 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3687 if (NewShAmtSplat) {
3688 APInt AdjNewShAmt =
3689 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3690 if (AdjNewShAmt.ule(MinLeadZero))
3691 return true;
3692 }
3693 }
3694 return false; // Can't tell if it's ok.
3695 };
3696 if (!CanFold())
3697 return nullptr;
3698 }
3699
3700 // All good, we can do this fold.
3701 X = Builder.CreateZExt(X, WidestTy);
3702 Y = Builder.CreateZExt(Y, WidestTy);
3703 // The shift is the same that was for X.
3704 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3705 ? Builder.CreateLShr(X, NewShAmt)
3706 : Builder.CreateShl(X, NewShAmt);
3707 Value *T1 = Builder.CreateAnd(T0, Y);
3708 return Builder.CreateICmp(I.getPredicate(), T1,
3709 Constant::getNullValue(WidestTy));
3710}
3711
3712/// Fold
3713/// (-1 u/ x) u< y
3714/// ((x * y) u/ x) != y
3715/// to
3716/// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3717/// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3718/// will mean that we are looking for the opposite answer.
3719Value *InstCombinerImpl::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3720 ICmpInst::Predicate Pred;
3721 Value *X, *Y;
3722 Instruction *Mul;
3723 bool NeedNegation;
3724 // Look for: (-1 u/ x) u</u>= y
3725 if (!I.isEquality() &&
3726 match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3727 m_Value(Y)))) {
3728 Mul = nullptr;
3729
3730 // Are we checking that overflow does not happen, or does happen?
3731 switch (Pred) {
3732 case ICmpInst::Predicate::ICMP_ULT:
3733 NeedNegation = false;
3734 break; // OK
3735 case ICmpInst::Predicate::ICMP_UGE:
3736 NeedNegation = true;
3737 break; // OK
3738 default:
3739 return nullptr; // Wrong predicate.
3740 }
3741 } else // Look for: ((x * y) u/ x) !=/== y
3742 if (I.isEquality() &&
3743 match(&I, m_c_ICmp(Pred, m_Value(Y),
3744 m_OneUse(m_UDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3745 m_Value(X)),
3746 m_Instruction(Mul)),
3747 m_Deferred(X)))))) {
3748 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3749 } else
3750 return nullptr;
3751
3752 BuilderTy::InsertPointGuard Guard(Builder);
3753 // If the pattern included (x * y), we'll want to insert new instructions
3754 // right before that original multiplication so that we can replace it.
3755 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3756 if (MulHadOtherUses)
3757 Builder.SetInsertPoint(Mul);
3758
3759 Function *F = Intrinsic::getDeclaration(
3760 I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3761 CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3762
3763 // If the multiplication was used elsewhere, to ensure that we don't leave
3764 // "duplicate" instructions, replace uses of that original multiplication
3765 // with the multiplication result from the with.overflow intrinsic.
3766 if (MulHadOtherUses)
3767 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3768
3769 Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3770 if (NeedNegation) // This technically increases instruction count.
3771 Res = Builder.CreateNot(Res, "umul.not.ov");
3772
3773 // If we replaced the mul, erase it. Do this after all uses of Builder,
3774 // as the mul is used as insertion point.
3775 if (MulHadOtherUses)
3776 eraseInstFromFunction(*Mul);
3777
3778 return Res;
3779}
3780
3781static Instruction *foldICmpXNegX(ICmpInst &I) {
3782 CmpInst::Predicate Pred;
3783 Value *X;
3784 if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3785 return nullptr;
3786
3787 if (ICmpInst::isSigned(Pred))
3788 Pred = ICmpInst::getSwappedPredicate(Pred);
3789 else if (ICmpInst::isUnsigned(Pred))
3790 Pred = ICmpInst::getSignedPredicate(Pred);
3791 // else for equality-comparisons just keep the predicate.
3792
3793 return ICmpInst::Create(Instruction::ICmp, Pred, X,
3794 Constant::getNullValue(X->getType()), I.getName());
3795}
3796
3797/// Try to fold icmp (binop), X or icmp X, (binop).
3798/// TODO: A large part of this logic is duplicated in InstSimplify's
3799/// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3800/// duplication.
3801Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
3802 const SimplifyQuery &SQ) {
3803 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3804 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3805
3806 // Special logic for binary operators.
3807 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
1
Assuming 'Op0' is a 'BinaryOperator'
3808 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2
Assuming 'Op1' is not a 'BinaryOperator'
3809 if (!BO0
2.1
'BO0' is non-null
2.1
'BO0' is non-null
&& !BO1)
3810 return nullptr;
3811
3812 if (Instruction *NewICmp
2.2
'NewICmp' is null
2.2
'NewICmp' is null
= foldICmpXNegX(I))
3
Taking false branch
3813 return NewICmp;
3814
3815 const CmpInst::Predicate Pred = I.getPredicate();
3816 Value *X;
3817
3818 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3819 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3820 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
4
Taking false branch
3821 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3822 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3823 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3824 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
5
Taking false branch
3825 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3826 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3827
3828 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3829 if (BO0
5.1
'BO0' is non-null
5.1
'BO0' is non-null
&& isa<OverflowingBinaryOperator>(BO0))
6
Assuming 'BO0' is not a 'OverflowingBinaryOperator'
7
Taking false branch
3830 NoOp0WrapProblem =
3831 ICmpInst::isEquality(Pred) ||
3832 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3833 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3834 if (BO1
7.1
'BO1' is null
7.1
'BO1' is null
&& isa<OverflowingBinaryOperator>(BO1))
3835 NoOp1WrapProblem =
3836 ICmpInst::isEquality(Pred) ||
3837 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3838 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3839
3840 // Analyze the case when either Op0 or Op1 is an add instruction.
3841 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3842 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3843 if (BO0
7.2
'BO0' is non-null
7.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Add) {
8
Assuming the condition is false
9
Taking false branch
3844 A = BO0->getOperand(0);
3845 B = BO0->getOperand(1);
3846 }
3847 if (BO1
9.1
'BO1' is null
9.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Add) {
3848 C = BO1->getOperand(0);
3849 D = BO1->getOperand(1);
3850 }
3851
3852 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3853 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3854 if ((A
9.2
'A' is not equal to 'Op1'
9.2
'A' is not equal to 'Op1'
== Op1 || B
9.3
'B' is not equal to 'Op1'
9.3
'B' is not equal to 'Op1'
== Op1) && NoOp0WrapProblem)
3855 return new ICmpInst(Pred, A == Op1 ? B : A,
3856 Constant::getNullValue(Op1->getType()));
3857
3858 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3859 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3860 if ((C
9.4
'C' is not equal to 'Op0'
9.4
'C' is not equal to 'Op0'
== Op0 || D
9.5
'D' is not equal to 'Op0'
9.5
'D' is not equal to 'Op0'
== Op0) && NoOp1WrapProblem)
3861 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3862 C == Op0 ? D : C);
3863
3864 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3865 if (A
9.6
'A' is null
9.6
'A' is null
&& C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3866 NoOp1WrapProblem) {
3867 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3868 Value *Y, *Z;
3869 if (A == C) {
3870 // C + B == C + D -> B == D
3871 Y = B;
3872 Z = D;
3873 } else if (A == D) {
3874 // D + B == C + D -> B == C
3875 Y = B;
3876 Z = C;
3877 } else if (B == C) {
3878 // A + C == C + D -> A == D
3879 Y = A;
3880 Z = D;
3881 } else {
3882 assert(B == D)(static_cast <bool> (B == D) ? void (0) : __assert_fail
("B == D", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3882, __extension__ __PRETTY_FUNCTION__))
;
3883 // A + D == C + D -> A == C
3884 Y = A;
3885 Z = C;
3886 }
3887 return new ICmpInst(Pred, Y, Z);
3888 }
3889
3890 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3891 if (A
9.7
'A' is null
9.7
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
10
Taking false branch
3892 match(B, m_AllOnes()))
3893 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3894
3895 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3896 if (A
10.1
'A' is null
10.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
11
Taking false branch
3897 match(B, m_AllOnes()))
3898 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3899
3900 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3901 if (A
11.1
'A' is null
11.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
12
Taking false branch
3902 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3903
3904 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3905 if (A
12.1
'A' is null
12.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
13
Taking false branch
3906 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3907
3908 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3909 if (C
13.1
'C' is null
13.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
14
Taking false branch
3910 match(D, m_AllOnes()))
3911 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3912
3913 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3914 if (C
14.1
'C' is null
14.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
15
Taking false branch
3915 match(D, m_AllOnes()))
3916 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3917
3918 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3919 if (C
15.1
'C' is null
15.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
16
Taking false branch
3920 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3921
3922 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3923 if (C
16.1
'C' is null
16.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
17
Taking false branch
3924 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3925
3926 // TODO: The subtraction-related identities shown below also hold, but
3927 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3928 // wouldn't happen even if they were implemented.
3929 //
3930 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3931 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3932 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3933 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3934
3935 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3936 if (A
17.1
'A' is null
17.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
18
Taking false branch
3937 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3938
3939 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3940 if (A
18.1
'A' is null
18.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
19
Taking false branch
3941 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3942
3943 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3944 if (C
19.1
'C' is null
19.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
20
Taking false branch
3945 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3946
3947 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3948 if (C
20.1
'C' is null
20.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
21
Taking false branch
3949 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3950
3951 // if C1 has greater magnitude than C2:
3952 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3953 // s.t. C3 = C1 - C2
3954 //
3955 // if C2 has greater magnitude than C1:
3956 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3957 // s.t. C3 = C2 - C1
3958 if (A
21.1
'A' is null
21.1
'A' is null
&& C && NoOp0WrapProblem && NoOp1WrapProblem &&
3959 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3960 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3961 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3962 const APInt &AP1 = C1->getValue();
3963 const APInt &AP2 = C2->getValue();
3964 if (AP1.isNegative() == AP2.isNegative()) {
3965 APInt AP1Abs = C1->getValue().abs();
3966 APInt AP2Abs = C2->getValue().abs();
3967 if (AP1Abs.uge(AP2Abs)) {
3968 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3969 bool HasNUW = BO0->hasNoUnsignedWrap() && C3->getValue().ule(AP1);
3970 bool HasNSW = BO0->hasNoSignedWrap();
3971 Value *NewAdd = Builder.CreateAdd(A, C3, "", HasNUW, HasNSW);
3972 return new ICmpInst(Pred, NewAdd, C);
3973 } else {
3974 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3975 bool HasNUW = BO1->hasNoUnsignedWrap() && C3->getValue().ule(AP2);
3976 bool HasNSW = BO1->hasNoSignedWrap();
3977 Value *NewAdd = Builder.CreateAdd(C, C3, "", HasNUW, HasNSW);
3978 return new ICmpInst(Pred, A, NewAdd);
3979 }
3980 }
3981 }
3982
3983 // Analyze the case when either Op0 or Op1 is a sub instruction.
3984 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3985 A = nullptr;
3986 B = nullptr;
3987 C = nullptr;
3988 D = nullptr;
3989 if (BO0
21.2
'BO0' is non-null
21.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Sub) {
22
Assuming the condition is false
23
Taking false branch
3990 A = BO0->getOperand(0);
3991 B = BO0->getOperand(1);
3992 }
3993 if (BO1
23.1
'BO1' is null
23.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Sub) {
3994 C = BO1->getOperand(0);
3995 D = BO1->getOperand(1);
3996 }
3997
3998 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3999 if (A
23.2
'A' is not equal to 'Op1'
23.2
'A' is not equal to 'Op1'
== Op1 && NoOp0WrapProblem)
4000 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
4001 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
4002 if (C
23.3
'C' is not equal to 'Op0'
23.3
'C' is not equal to 'Op0'
== Op0 && NoOp1WrapProblem)
4003 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
4004
4005 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
4006 // (A - B) u>/u<= A --> B u>/u<= A
4007 if (A
23.4
'A' is not equal to 'Op1'
23.4
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
4008 return new ICmpInst(Pred, B, A);
4009 // C u</u>= (C - D) --> C u</u>= D
4010 if (C
23.5
'C' is not equal to 'Op0'
23.5
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
4011 return new ICmpInst(Pred, C, D);
4012 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
4013 if (A
23.6
'A' is not equal to 'Op1'
23.6
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
4014 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4015 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
4016 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
4017 if (C
23.7
'C' is not equal to 'Op0'
23.7
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
4018 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4019 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
4020
4021 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
4022 if (B
23.8
'B' is null
23.8
'B' is null
&& D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
4023 return new ICmpInst(Pred, A, C);
4024
4025 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
4026 if (A
23.9
'A' is null
23.9
'A' is null
&& C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
4027 return new ICmpInst(Pred, D, B);
4028
4029 // icmp (0-X) < cst --> x > -cst
4030 if (NoOp0WrapProblem
23.10
'NoOp0WrapProblem' is false
23.10
'NoOp0WrapProblem' is false
&& ICmpInst::isSigned(Pred)) {
4031 Value *X;
4032 if (match(BO0, m_Neg(m_Value(X))))
4033 if (Constant *RHSC = dyn_cast<Constant>(Op1))
4034 if (RHSC->isNotMinSignedValue())
4035 return new ICmpInst(I.getSwappedPredicate(), X,
4036 ConstantExpr::getNeg(RHSC));
4037 }
4038
4039 {
4040 // Try to remove shared constant multiplier from equality comparison:
4041 // X * C == Y * C (with no overflowing/aliasing) --> X == Y
4042 Value *X, *Y;
4043 const APInt *C;
4044 if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
24
Taking false branch
4045 match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
4046 if (!C->countTrailingZeros() ||
4047 (BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
4048 (BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
4049 return new ICmpInst(Pred, X, Y);
4050 }
4051
4052 BinaryOperator *SRem = nullptr;
4053 // icmp (srem X, Y), Y
4054 if (BO0
24.1
'BO0' is non-null
24.1
'BO0' is non-null
&& BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
25
Assuming the condition is false
4055 SRem = BO0;
4056 // icmp Y, (srem X, Y)
4057 else if (BO1
25.1
'BO1' is null
25.1
'BO1' is null
&& BO1->getOpcode() == Instruction::SRem &&
4058 Op0 == BO1->getOperand(1))
4059 SRem = BO1;
4060 if (SRem
25.2
'SRem' is null
25.2
'SRem' is null
) {
26
Taking false branch
4061 // We don't check hasOneUse to avoid increasing register pressure because
4062 // the value we use is the same value this instruction was already using.
4063 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4064 default:
4065 break;
4066 case ICmpInst::ICMP_EQ:
4067 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4068 case ICmpInst::ICMP_NE:
4069 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4070 case ICmpInst::ICMP_SGT:
4071 case ICmpInst::ICMP_SGE:
4072 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4073 Constant::getAllOnesValue(SRem->getType()));
4074 case ICmpInst::ICMP_SLT:
4075 case ICmpInst::ICMP_SLE:
4076 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4077 Constant::getNullValue(SRem->getType()));
4078 }
4079 }
4080
4081 if (BO0
26.1
'BO0' is non-null
26.1
'BO0' is non-null
&& BO1
26.2
'BO1' is null
26.2
'BO1' is null
&& BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4082 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4083 switch (BO0->getOpcode()) {
4084 default:
4085 break;
4086 case Instruction::Add:
4087 case Instruction::Sub:
4088 case Instruction::Xor: {
4089 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4090 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4091
4092 const APInt *C;
4093 if (match(BO0->getOperand(1), m_APInt(C))) {
4094 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4095 if (C->isSignMask()) {
4096 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4097 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4098 }
4099
4100 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4101 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4102 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4103 NewPred = I.getSwappedPredicate(NewPred);
4104 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4105 }
4106 }
4107 break;
4108 }
4109 case Instruction::Mul: {
4110 if (!I.isEquality())
4111 break;
4112
4113 const APInt *C;
4114 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
4115 !C->isOneValue()) {
4116 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4117 // Mask = -1 >> count-trailing-zeros(C).
4118 if (unsigned TZs = C->countTrailingZeros()) {
4119 Constant *Mask = ConstantInt::get(
4120 BO0->getType(),
4121 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4122 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4123 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4124 return new ICmpInst(Pred, And1, And2);
4125 }
4126 }
4127 break;
4128 }
4129 case Instruction::UDiv:
4130 case Instruction::LShr:
4131 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4132 break;
4133 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4134
4135 case Instruction::SDiv:
4136 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4137 break;
4138 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4139
4140 case Instruction::AShr:
4141 if (!BO0->isExact() || !BO1->isExact())
4142 break;
4143 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4144
4145 case Instruction::Shl: {
4146 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4147 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4148 if (!NUW && !NSW)
4149 break;
4150 if (!NSW && I.isSigned())
4151 break;
4152 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4153 }
4154 }
4155 }
4156
4157 if (BO0
26.3
'BO0' is non-null
26.3
'BO0' is non-null
) {
27
Taking true branch
4158 // Transform A & (L - 1) `ult` L --> L != 0
4159 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4160 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4161
4162 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4163 auto *Zero = Constant::getNullValue(BO0->getType());
4164 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4165 }
4166 }
4167
4168 if (Value *V
27.1
'V' is null
27.1
'V' is null
= foldUnsignedMultiplicationOverflowCheck(I))
28
Taking false branch
4169 return replaceInstUsesWith(I, V);
4170
4171 if (Value *V
28.1
'V' is null
28.1
'V' is null
= foldICmpWithLowBitMaskedVal(I, Builder))
29
Taking false branch
4172 return replaceInstUsesWith(I, V);
4173
4174 if (Value *V
29.1
'V' is null
29.1
'V' is null
= foldICmpWithTruncSignExtendedVal(I, Builder))
30
Taking false branch
4175 return replaceInstUsesWith(I, V);
4176
4177 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
31
Calling 'foldShiftIntoShiftInAnotherHandOfAndInICmp'
4178 return replaceInstUsesWith(I, V);
4179
4180 return nullptr;
4181}
4182
4183/// Fold icmp Pred min|max(X, Y), X.
4184static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4185 ICmpInst::Predicate Pred = Cmp.getPredicate();
4186 Value *Op0 = Cmp.getOperand(0);
4187 Value *X = Cmp.getOperand(1);
4188
4189 // Canonicalize minimum or maximum operand to LHS of the icmp.
4190 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4191 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4192 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4193 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4194 std::swap(Op0, X);
4195 Pred = Cmp.getSwappedPredicate();
4196 }
4197
4198 Value *Y;
4199 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4200 // smin(X, Y) == X --> X s<= Y
4201 // smin(X, Y) s>= X --> X s<= Y
4202 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4203 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4204
4205 // smin(X, Y) != X --> X s> Y
4206 // smin(X, Y) s< X --> X s> Y
4207 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4208 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4209
4210 // These cases should be handled in InstSimplify:
4211 // smin(X, Y) s<= X --> true
4212 // smin(X, Y) s> X --> false
4213 return nullptr;
4214 }
4215
4216 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4217 // smax(X, Y) == X --> X s>= Y
4218 // smax(X, Y) s<= X --> X s>= Y
4219 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4220 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4221
4222 // smax(X, Y) != X --> X s< Y
4223 // smax(X, Y) s> X --> X s< Y
4224 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4225 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4226
4227 // These cases should be handled in InstSimplify:
4228 // smax(X, Y) s>= X --> true
4229 // smax(X, Y) s< X --> false
4230 return nullptr;
4231 }
4232
4233 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4234 // umin(X, Y) == X --> X u<= Y
4235 // umin(X, Y) u>= X --> X u<= Y
4236 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4237 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4238
4239 // umin(X, Y) != X --> X u> Y
4240 // umin(X, Y) u< X --> X u> Y
4241 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4242 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4243
4244 // These cases should be handled in InstSimplify:
4245 // umin(X, Y) u<= X --> true
4246 // umin(X, Y) u> X --> false
4247 return nullptr;
4248 }
4249
4250 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4251 // umax(X, Y) == X --> X u>= Y
4252 // umax(X, Y) u<= X --> X u>= Y
4253 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4254 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4255
4256 // umax(X, Y) != X --> X u< Y
4257 // umax(X, Y) u> X --> X u< Y
4258 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4259 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4260
4261 // These cases should be handled in InstSimplify:
4262 // umax(X, Y) u>= X --> true
4263 // umax(X, Y) u< X --> false
4264 return nullptr;
4265 }
4266
4267 return nullptr;
4268}
4269
4270Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
4271 if (!I.isEquality())
4272 return nullptr;
4273
4274 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4275 const CmpInst::Predicate Pred = I.getPredicate();
4276 Value *A, *B, *C, *D;
4277 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4278 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4279 Value *OtherVal = A == Op1 ? B : A;
4280 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4281 }
4282
4283 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4284 // A^c1 == C^c2 --> A == C^(c1^c2)
4285 ConstantInt *C1, *C2;
4286 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4287 Op1->hasOneUse()) {
4288 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4289 Value *Xor = Builder.CreateXor(C, NC);
4290 return new ICmpInst(Pred, A, Xor);
4291 }
4292
4293 // A^B == A^D -> B == D
4294 if (A == C)
4295 return new ICmpInst(Pred, B, D);
4296 if (A == D)
4297 return new ICmpInst(Pred, B, C);
4298 if (B == C)
4299 return new ICmpInst(Pred, A, D);
4300 if (B == D)
4301 return new ICmpInst(Pred, A, C);
4302 }
4303 }
4304
4305 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4306 // A == (A^B) -> B == 0
4307 Value *OtherVal = A == Op0 ? B : A;
4308 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4309 }
4310
4311 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4312 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4313 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4314 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4315
4316 if (A == C) {
4317 X = B;
4318 Y = D;
4319 Z = A;
4320 } else if (A == D) {
4321 X = B;
4322 Y = C;
4323 Z = A;
4324 } else if (B == C) {
4325 X = A;
4326 Y = D;
4327 Z = B;
4328 } else if (B == D) {
4329 X = A;
4330 Y = C;
4331 Z = B;
4332 }
4333
4334 if (X) { // Build (X^Y) & Z
4335 Op1 = Builder.CreateXor(X, Y);
4336 Op1 = Builder.CreateAnd(Op1, Z);
4337 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4338 }
4339 }
4340
4341 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4342 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4343 ConstantInt *Cst1;
4344 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4345 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4346 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4347 match(Op1, m_ZExt(m_Value(A))))) {
4348 APInt Pow2 = Cst1->getValue() + 1;
4349 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4350 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4351 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4352 }
4353
4354 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4355 // For lshr and ashr pairs.
4356 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4357 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4358 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4359 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4360 unsigned TypeBits = Cst1->getBitWidth();
4361 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4362 if (ShAmt < TypeBits && ShAmt != 0) {
4363 ICmpInst::Predicate NewPred =
4364 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4365 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4366 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4367 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4368 }
4369 }
4370
4371 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4372 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4373 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4374 unsigned TypeBits = Cst1->getBitWidth();
4375 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4376 if (ShAmt < TypeBits && ShAmt != 0) {
4377 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4378 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4379 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4380 I.getName() + ".mask");
4381 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4382 }
4383 }
4384
4385 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4386 // "icmp (and X, mask), cst"
4387 uint64_t ShAmt = 0;
4388 if (Op0->hasOneUse() &&
4389 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4390 match(Op1, m_ConstantInt(Cst1)) &&
4391 // Only do this when A has multiple uses. This is most important to do
4392 // when it exposes other optimizations.
4393 !A->hasOneUse()) {
4394 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4395
4396 if (ShAmt < ASize) {
4397 APInt MaskV =
4398 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4399 MaskV <<= ShAmt;
4400
4401 APInt CmpV = Cst1->getValue().zext(ASize);
4402 CmpV <<= ShAmt;
4403
4404 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4405 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4406 }
4407 }
4408
4409 // If both operands are byte-swapped or bit-reversed, just compare the
4410 // original values.
4411 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4412 // and handle more intrinsics.
4413 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4414 (match(Op0, m_BitReverse(m_Value(A))) &&
4415 match(Op1, m_BitReverse(m_Value(B)))))
4416 return new ICmpInst(Pred, A, B);
4417
4418 // Canonicalize checking for a power-of-2-or-zero value:
4419 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4420 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4421 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4422 m_Deferred(A)))) ||
4423 !match(Op1, m_ZeroInt()))
4424 A = nullptr;
4425
4426 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4427 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4428 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4429 A = Op1;
4430 else if (match(Op1,
4431 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4432 A = Op0;
4433
4434 if (A) {
4435 Type *Ty = A->getType();
4436 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4437 return Pred == ICmpInst::ICMP_EQ
4438 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4439 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4440 }
4441
4442 return nullptr;
4443}
4444
4445static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4446 InstCombiner::BuilderTy &Builder) {
4447 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0")(static_cast <bool> (isa<CastInst>(ICmp.getOperand
(0)) && "Expected cast for operand 0") ? void (0) : __assert_fail
("isa<CastInst>(ICmp.getOperand(0)) && \"Expected cast for operand 0\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4447, __extension__ __PRETTY_FUNCTION__))
;
4448 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4449 Value *X;
4450 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4451 return nullptr;
4452
4453 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4454 bool IsSignedCmp = ICmp.isSigned();
4455 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4456 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4457 // and the other is a zext), then we can't handle this.
4458 // TODO: This is too strict. We can handle some predicates (equality?).
4459 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4460 return nullptr;
4461
4462 // Not an extension from the same type?
4463 Value *Y = CastOp1->getOperand(0);
4464 Type *XTy = X->getType(), *YTy = Y->getType();
4465 if (XTy != YTy) {
4466 // One of the casts must have one use because we are creating a new cast.
4467 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4468 return nullptr;
4469 // Extend the narrower operand to the type of the wider operand.
4470 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4471 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4472 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4473 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4474 else
4475 return nullptr;
4476 }
4477
4478 // (zext X) == (zext Y) --> X == Y
4479 // (sext X) == (sext Y) --> X == Y
4480 if (ICmp.isEquality())
4481 return new ICmpInst(ICmp.getPredicate(), X, Y);
4482
4483 // A signed comparison of sign extended values simplifies into a
4484 // signed comparison.
4485 if (IsSignedCmp && IsSignedExt)
4486 return new ICmpInst(ICmp.getPredicate(), X, Y);
4487
4488 // The other three cases all fold into an unsigned comparison.
4489 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4490 }
4491
4492 // Below here, we are only folding a compare with constant.
4493 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4494 if (!C)
4495 return nullptr;
4496
4497 // Compute the constant that would happen if we truncated to SrcTy then
4498 // re-extended to DestTy.
4499 Type *SrcTy = CastOp0->getSrcTy();
4500 Type *DestTy = CastOp0->getDestTy();
4501 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4502 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4503
4504 // If the re-extended constant didn't change...
4505 if (Res2 == C) {
4506 if (ICmp.isEquality())
4507 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4508
4509 // A signed comparison of sign extended values simplifies into a
4510 // signed comparison.
4511 if (IsSignedExt && IsSignedCmp)
4512 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4513
4514 // The other three cases all fold into an unsigned comparison.
4515 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4516 }
4517
4518 // The re-extended constant changed, partly changed (in the case of a vector),
4519 // or could not be determined to be equal (in the case of a constant
4520 // expression), so the constant cannot be represented in the shorter type.
4521 // All the cases that fold to true or false will have already been handled
4522 // by SimplifyICmpInst, so only deal with the tricky case.
4523 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4524 return nullptr;
4525
4526 // Is source op positive?
4527 // icmp ult (sext X), C --> icmp sgt X, -1
4528 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4529 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4530
4531 // Is source op negative?
4532 // icmp ugt (sext X), C --> icmp slt X, 0
4533 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!")(static_cast <bool> (ICmp.getPredicate() == ICmpInst::ICMP_UGT
&& "ICmp should be folded!") ? void (0) : __assert_fail
("ICmp.getPredicate() == ICmpInst::ICMP_UGT && \"ICmp should be folded!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4533, __extension__ __PRETTY_FUNCTION__))
;
4534 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4535}
4536
4537/// Handle icmp (cast x), (cast or constant).
4538Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
4539 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4540 if (!CastOp0)
4541 return nullptr;
4542 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4543 return nullptr;
4544
4545 Value *Op0Src = CastOp0->getOperand(0);
4546 Type *SrcTy = CastOp0->getSrcTy();
4547 Type *DestTy = CastOp0->getDestTy();
4548
4549 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4550 // integer type is the same size as the pointer type.
4551 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4552 if (isa<VectorType>(SrcTy)) {
4553 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4554 DestTy = cast<VectorType>(DestTy)->getElementType();
4555 }
4556 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4557 };
4558 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4559 CompatibleSizes(SrcTy, DestTy)) {
4560 Value *NewOp1 = nullptr;
4561 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4562 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4563 if (PtrSrc->getType()->getPointerAddressSpace() ==
4564 Op0Src->getType()->getPointerAddressSpace()) {
4565 NewOp1 = PtrToIntOp1->getOperand(0);
4566 // If the pointer types don't match, insert a bitcast.
4567 if (Op0Src->getType() != NewOp1->getType())
4568 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4569 }
4570 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4571 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4572 }
4573
4574 if (NewOp1)
4575 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4576 }
4577
4578 return foldICmpWithZextOrSext(ICmp, Builder);
4579}
4580
4581static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4582 switch (BinaryOp) {
4583 default:
4584 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4584)
;
4585 case Instruction::Add:
4586 case Instruction::Sub:
4587 return match(RHS, m_Zero());
4588 case Instruction::Mul:
4589 return match(RHS, m_One());
4590 }
4591}
4592
4593OverflowResult
4594InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
4595 bool IsSigned, Value *LHS, Value *RHS,
4596 Instruction *CxtI) const {
4597 switch (BinaryOp) {
4598 default:
4599 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4599)
;
4600 case Instruction::Add:
4601 if (IsSigned)
4602 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4603 else
4604 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4605 case Instruction::Sub:
4606 if (IsSigned)
4607 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4608 else
4609 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4610 case Instruction::Mul:
4611 if (IsSigned)
4612 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4613 else
4614 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4615 }
4616}
4617
4618bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
4619 bool IsSigned, Value *LHS,
4620 Value *RHS, Instruction &OrigI,
4621 Value *&Result,
4622 Constant *&Overflow) {
4623 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4624 std::swap(LHS, RHS);
4625
4626 // If the overflow check was an add followed by a compare, the insertion point
4627 // may be pointing to the compare. We want to insert the new instructions
4628 // before the add in case there are uses of the add between the add and the
4629 // compare.
4630 Builder.SetInsertPoint(&OrigI);
4631
4632 Type *OverflowTy = Type::getInt1Ty(LHS->getContext());
4633 if (auto *LHSTy = dyn_cast<VectorType>(LHS->getType()))
4634 OverflowTy = VectorType::get(OverflowTy, LHSTy->getElementCount());
4635
4636 if (isNeutralValue(BinaryOp, RHS)) {
4637 Result = LHS;
4638 Overflow = ConstantInt::getFalse(OverflowTy);
4639 return true;
4640 }
4641
4642 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4643 case OverflowResult::MayOverflow:
4644 return false;
4645 case OverflowResult::AlwaysOverflowsLow:
4646 case OverflowResult::AlwaysOverflowsHigh:
4647 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4648 Result->takeName(&OrigI);
4649 Overflow = ConstantInt::getTrue(OverflowTy);
4650 return true;
4651 case OverflowResult::NeverOverflows:
4652 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4653 Result->takeName(&OrigI);
4654 Overflow = ConstantInt::getFalse(OverflowTy);
4655 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4656 if (IsSigned)
4657 Inst->setHasNoSignedWrap();
4658 else
4659 Inst->setHasNoUnsignedWrap();
4660 }
4661 return true;
4662 }
4663
4664 llvm_unreachable("Unexpected overflow result")::llvm::llvm_unreachable_internal("Unexpected overflow result"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4664)
;
4665}
4666
4667/// Recognize and process idiom involving test for multiplication
4668/// overflow.
4669///
4670/// The caller has matched a pattern of the form:
4671/// I = cmp u (mul(zext A, zext B), V
4672/// The function checks if this is a test for overflow and if so replaces
4673/// multiplication with call to 'mul.with.overflow' intrinsic.
4674///
4675/// \param I Compare instruction.
4676/// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4677/// the compare instruction. Must be of integer type.
4678/// \param OtherVal The other argument of compare instruction.
4679/// \returns Instruction which must replace the compare instruction, NULL if no
4680/// replacement required.
4681static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4682 Value *OtherVal,
4683 InstCombinerImpl &IC) {
4684 // Don't bother doing this transformation for pointers, don't do it for
4685 // vectors.
4686 if (!isa<IntegerType>(MulVal->getType()))
4687 return nullptr;
4688
4689 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal)(static_cast <bool> (I.getOperand(0) == MulVal || I.getOperand
(1) == MulVal) ? void (0) : __assert_fail ("I.getOperand(0) == MulVal || I.getOperand(1) == MulVal"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4689, __extension__ __PRETTY_FUNCTION__))
;
4690 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal)(static_cast <bool> (I.getOperand(0) == OtherVal || I.getOperand
(1) == OtherVal) ? void (0) : __assert_fail ("I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4690, __extension__ __PRETTY_FUNCTION__))
;
4691 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4692 if (!MulInstr)
4693 return nullptr;
4694 assert(MulInstr->getOpcode() == Instruction::Mul)(static_cast <bool> (MulInstr->getOpcode() == Instruction
::Mul) ? void (0) : __assert_fail ("MulInstr->getOpcode() == Instruction::Mul"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4694, __extension__ __PRETTY_FUNCTION__))
;
4695
4696 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4697 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4698 assert(LHS->getOpcode() == Instruction::ZExt)(static_cast <bool> (LHS->getOpcode() == Instruction
::ZExt) ? void (0) : __assert_fail ("LHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4698, __extension__ __PRETTY_FUNCTION__))
;
4699 assert(RHS->getOpcode() == Instruction::ZExt)(static_cast <bool> (RHS->getOpcode() == Instruction
::ZExt) ? void (0) : __assert_fail ("RHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4699, __extension__ __PRETTY_FUNCTION__))
;
4700 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4701
4702 // Calculate type and width of the result produced by mul.with.overflow.
4703 Type *TyA = A->getType(), *TyB = B->getType();
4704 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4705 WidthB = TyB->getPrimitiveSizeInBits();
4706 unsigned MulWidth;
4707 Type *MulType;
4708 if (WidthB > WidthA) {
4709 MulWidth = WidthB;
4710 MulType = TyB;
4711 } else {
4712 MulWidth = WidthA;
4713 MulType = TyA;
4714 }
4715
4716 // In order to replace the original mul with a narrower mul.with.overflow,
4717 // all uses must ignore upper bits of the product. The number of used low
4718 // bits must be not greater than the width of mul.with.overflow.
4719 if (MulVal->hasNUsesOrMore(2))
4720 for (User *U : MulVal->users()) {
4721 if (U == &I)
4722 continue;
4723 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4724 // Check if truncation ignores bits above MulWidth.
4725 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4726 if (TruncWidth > MulWidth)
4727 return nullptr;
4728 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4729 // Check if AND ignores bits above MulWidth.
4730 if (BO->getOpcode() != Instruction::And)
4731 return nullptr;
4732 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4733 const APInt &CVal = CI->getValue();
4734 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4735 return nullptr;
4736 } else {
4737 // In this case we could have the operand of the binary operation
4738 // being defined in another block, and performing the replacement
4739 // could break the dominance relation.
4740 return nullptr;
4741 }
4742 } else {
4743 // Other uses prohibit this transformation.
4744 return nullptr;
4745 }
4746 }
4747
4748 // Recognize patterns
4749 switch (I.getPredicate()) {
4750 case ICmpInst::ICMP_EQ:
4751 case ICmpInst::ICMP_NE:
4752 // Recognize pattern:
4753 // mulval = mul(zext A, zext B)
4754 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4755 ConstantInt *CI;
4756 Value *ValToMask;
4757 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4758 if (ValToMask != MulVal)
4759 return nullptr;
4760 const APInt &CVal = CI->getValue() + 1;
4761 if (CVal.isPowerOf2()) {
4762 unsigned MaskWidth = CVal.logBase2();
4763 if (MaskWidth == MulWidth)
4764 break; // Recognized
4765 }
4766 }
4767 return nullptr;
4768
4769 case ICmpInst::ICMP_UGT:
4770 // Recognize pattern:
4771 // mulval = mul(zext A, zext B)
4772 // cmp ugt mulval, max
4773 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4774 APInt MaxVal = APInt::getMaxValue(MulWidth);
4775 MaxVal = MaxVal.zext(CI->getBitWidth());
4776 if (MaxVal.eq(CI->getValue()))
4777 break; // Recognized
4778 }
4779 return nullptr;
4780
4781 case ICmpInst::ICMP_UGE:
4782 // Recognize pattern:
4783 // mulval = mul(zext A, zext B)
4784 // cmp uge mulval, max+1
4785 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4786 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4787 if (MaxVal.eq(CI->getValue()))
4788 break; // Recognized
4789 }
4790 return nullptr;
4791
4792 case ICmpInst::ICMP_ULE:
4793 // Recognize pattern:
4794 // mulval = mul(zext A, zext B)
4795 // cmp ule mulval, max
4796 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4797 APInt MaxVal = APInt::getMaxValue(MulWidth);
4798 MaxVal = MaxVal.zext(CI->getBitWidth());
4799 if (MaxVal.eq(CI->getValue()))
4800 break; // Recognized
4801 }
4802 return nullptr;
4803
4804 case ICmpInst::ICMP_ULT:
4805 // Recognize pattern:
4806 // mulval = mul(zext A, zext B)
4807 // cmp ule mulval, max + 1
4808 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4809 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4810 if (MaxVal.eq(CI->getValue()))
4811 break; // Recognized
4812 }
4813 return nullptr;
4814
4815 default:
4816 return nullptr;
4817 }
4818
4819 InstCombiner::BuilderTy &Builder = IC.Builder;
4820 Builder.SetInsertPoint(MulInstr);
4821
4822 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4823 Value *MulA = A, *MulB = B;
4824 if (WidthA < MulWidth)
4825 MulA = Builder.CreateZExt(A, MulType);
4826 if (WidthB < MulWidth)
4827 MulB = Builder.CreateZExt(B, MulType);
4828 Function *F = Intrinsic::getDeclaration(
4829 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4830 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4831 IC.addToWorklist(MulInstr);
4832
4833 // If there are uses of mul result other than the comparison, we know that
4834 // they are truncation or binary AND. Change them to use result of
4835 // mul.with.overflow and adjust properly mask/size.
4836 if (MulVal->hasNUsesOrMore(2)) {
4837 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4838 for (User *U : make_early_inc_range(MulVal->users())) {
4839 if (U == &I || U == OtherVal)
4840 continue;
4841 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4842 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4843 IC.replaceInstUsesWith(*TI, Mul);
4844 else
4845 TI->setOperand(0, Mul);
4846 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4847 assert(BO->getOpcode() == Instruction::And)(static_cast <bool> (BO->getOpcode() == Instruction::
And) ? void (0) : __assert_fail ("BO->getOpcode() == Instruction::And"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4847, __extension__ __PRETTY_FUNCTION__))
;
4848 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4849 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4850 APInt ShortMask = CI->getValue().trunc(MulWidth);
4851 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4852 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
4853 IC.replaceInstUsesWith(*BO, Zext);
4854 } else {
4855 llvm_unreachable("Unexpected Binary operation")::llvm::llvm_unreachable_internal("Unexpected Binary operation"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4855)
;
4856 }
4857 IC.addToWorklist(cast<Instruction>(U));
4858 }
4859 }
4860 if (isa<Instruction>(OtherVal))
4861 IC.addToWorklist(cast<Instruction>(OtherVal));
4862
4863 // The original icmp gets replaced with the overflow value, maybe inverted
4864 // depending on predicate.
4865 bool Inverse = false;
4866 switch (I.getPredicate()) {
4867 case ICmpInst::ICMP_NE:
4868 break;
4869 case ICmpInst::ICMP_EQ:
4870 Inverse = true;
4871 break;
4872 case ICmpInst::ICMP_UGT:
4873 case ICmpInst::ICMP_UGE:
4874 if (I.getOperand(0) == MulVal)
4875 break;
4876 Inverse = true;
4877 break;
4878 case ICmpInst::ICMP_ULT:
4879 case ICmpInst::ICMP_ULE:
4880 if (I.getOperand(1) == MulVal)
4881 break;
4882 Inverse = true;
4883 break;
4884 default:
4885 llvm_unreachable("Unexpected predicate")::llvm::llvm_unreachable_internal("Unexpected predicate", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4885)
;
4886 }
4887 if (Inverse) {
4888 Value *Res = Builder.CreateExtractValue(Call, 1);
4889 return BinaryOperator::CreateNot(Res);
4890 }
4891
4892 return ExtractValueInst::Create(Call, 1);
4893}
4894
4895/// When performing a comparison against a constant, it is possible that not all
4896/// the bits in the LHS are demanded. This helper method computes the mask that
4897/// IS demanded.
4898static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4899 const APInt *RHS;
4900 if (!match(I.getOperand(1), m_APInt(RHS)))
4901 return APInt::getAllOnesValue(BitWidth);
4902
4903 // If this is a normal comparison, it demands all bits. If it is a sign bit
4904 // comparison, it only demands the sign bit.
4905 bool UnusedBit;
4906 if (InstCombiner::isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4907 return APInt::getSignMask(BitWidth);
4908
4909 switch (I.getPredicate()) {
4910 // For a UGT comparison, we don't care about any bits that
4911 // correspond to the trailing ones of the comparand. The value of these
4912 // bits doesn't impact the outcome of the comparison, because any value
4913 // greater than the RHS must differ in a bit higher than these due to carry.
4914 case ICmpInst::ICMP_UGT:
4915 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4916
4917 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4918 // Any value less than the RHS must differ in a higher bit because of carries.
4919 case ICmpInst::ICMP_ULT:
4920 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4921
4922 default:
4923 return APInt::getAllOnesValue(BitWidth);
4924 }
4925}
4926
4927/// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4928/// should be swapped.
4929/// The decision is based on how many times these two operands are reused
4930/// as subtract operands and their positions in those instructions.
4931/// The rationale is that several architectures use the same instruction for
4932/// both subtract and cmp. Thus, it is better if the order of those operands
4933/// match.
4934/// \return true if Op0 and Op1 should be swapped.
4935static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4936 // Filter out pointer values as those cannot appear directly in subtract.
4937 // FIXME: we may want to go through inttoptrs or bitcasts.
4938 if (Op0->getType()->isPointerTy())
4939 return false;
4940 // If a subtract already has the same operands as a compare, swapping would be
4941 // bad. If a subtract has the same operands as a compare but in reverse order,
4942 // then swapping is good.
4943 int GoodToSwap = 0;
4944 for (const User *U : Op0->users()) {
4945 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4946 GoodToSwap++;
4947 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4948 GoodToSwap--;
4949 }
4950 return GoodToSwap > 0;
4951}
4952
4953/// Check that one use is in the same block as the definition and all
4954/// other uses are in blocks dominated by a given block.
4955///
4956/// \param DI Definition
4957/// \param UI Use
4958/// \param DB Block that must dominate all uses of \p DI outside
4959/// the parent block
4960/// \return true when \p UI is the only use of \p DI in the parent block
4961/// and all other uses of \p DI are in blocks dominated by \p DB.
4962///
4963bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
4964 const Instruction *UI,
4965 const BasicBlock *DB) const {
4966 assert(DI && UI && "Instruction not defined\n")(static_cast <bool> (DI && UI && "Instruction not defined\n"
) ? void (0) : __assert_fail ("DI && UI && \"Instruction not defined\\n\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4966, __extension__ __PRETTY_FUNCTION__))
;
4967 // Ignore incomplete definitions.
4968 if (!DI->getParent())
4969 return false;
4970 // DI and UI must be in the same block.
4971 if (DI->getParent() != UI->getParent())
4972 return false;
4973 // Protect from self-referencing blocks.
4974 if (DI->getParent() == DB)
4975 return false;
4976 for (const User *U : DI->users()) {
4977 auto *Usr = cast<Instruction>(U);
4978 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4979 return false;
4980 }
4981 return true;
4982}
4983
4984/// Return true when the instruction sequence within a block is select-cmp-br.
4985static bool isChainSelectCmpBranch(const SelectInst *SI) {
4986 const BasicBlock *BB = SI->getParent();
4987 if (!BB)
4988 return false;
4989 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4990 if (!BI || BI->getNumSuccessors() != 2)
4991 return false;
4992 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4993 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4994 return false;
4995 return true;
4996}
4997
4998/// True when a select result is replaced by one of its operands
4999/// in select-icmp sequence. This will eventually result in the elimination
5000/// of the select.
5001///
5002/// \param SI Select instruction
5003/// \param Icmp Compare instruction
5004/// \param SIOpd Operand that replaces the select
5005///
5006/// Notes:
5007/// - The replacement is global and requires dominator information
5008/// - The caller is responsible for the actual replacement
5009///
5010/// Example:
5011///
5012/// entry:
5013/// %4 = select i1 %3, %C* %0, %C* null
5014/// %5 = icmp eq %C* %4, null
5015/// br i1 %5, label %9, label %7
5016/// ...
5017/// ; <label>:7 ; preds = %entry
5018/// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
5019/// ...
5020///
5021/// can be transformed to
5022///
5023/// %5 = icmp eq %C* %0, null
5024/// %6 = select i1 %3, i1 %5, i1 true
5025/// br i1 %6, label %9, label %7
5026/// ...
5027/// ; <label>:7 ; preds = %entry
5028/// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
5029///
5030/// Similar when the first operand of the select is a constant or/and
5031/// the compare is for not equal rather than equal.
5032///
5033/// NOTE: The function is only called when the select and compare constants
5034/// are equal, the optimization can work only for EQ predicates. This is not a
5035/// major restriction since a NE compare should be 'normalized' to an equal
5036/// compare, which usually happens in the combiner and test case
5037/// select-cmp-br.ll checks for it.
5038bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
5039 const ICmpInst *Icmp,
5040 const unsigned SIOpd) {
5041 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!")(static_cast <bool> ((SIOpd == 1 || SIOpd == 2) &&
"Invalid select operand!") ? void (0) : __assert_fail ("(SIOpd == 1 || SIOpd == 2) && \"Invalid select operand!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5041, __extension__ __PRETTY_FUNCTION__))
;
5042 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
5043 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
5044 // The check for the single predecessor is not the best that can be
5045 // done. But it protects efficiently against cases like when SI's
5046 // home block has two successors, Succ and Succ1, and Succ1 predecessor
5047 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
5048 // replaced can be reached on either path. So the uniqueness check
5049 // guarantees that the path all uses of SI (outside SI's parent) are on
5050 // is disjoint from all other paths out of SI. But that information
5051 // is more expensive to compute, and the trade-off here is in favor
5052 // of compile-time. It should also be noticed that we check for a single
5053 // predecessor and not only uniqueness. This to handle the situation when
5054 // Succ and Succ1 points to the same basic block.
5055 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
5056 NumSel++;
5057 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
5058 return true;
5059 }
5060 }
5061 return false;
5062}
5063
5064/// Try to fold the comparison based on range information we can get by checking
5065/// whether bits are known to be zero or one in the inputs.
5066Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
5067 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5068 Type *Ty = Op0->getType();
5069 ICmpInst::Predicate Pred = I.getPredicate();
5070
5071 // Get scalar or pointer size.
5072 unsigned BitWidth = Ty->isIntOrIntVectorTy()
5073 ? Ty->getScalarSizeInBits()
5074 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
5075
5076 if (!BitWidth)
5077 return nullptr;
5078
5079 KnownBits Op0Known(BitWidth);
5080 KnownBits Op1Known(BitWidth);
5081
5082 if (SimplifyDemandedBits(&I, 0,
5083 getDemandedBitsLHSMask(I, BitWidth),
5084 Op0Known, 0))
5085 return &I;
5086
5087 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
5088 Op1Known, 0))
5089 return &I;
5090
5091 // Given the known and unknown bits, compute a range that the LHS could be
5092 // in. Compute the Min, Max and RHS values based on the known bits. For the
5093 // EQ and NE we use unsigned values.
5094 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
5095 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
5096 if (I.isSigned()) {
5097 Op0Min = Op0Known.getSignedMinValue();
5098 Op0Max = Op0Known.getSignedMaxValue();
5099 Op1Min = Op1Known.getSignedMinValue();
5100 Op1Max = Op1Known.getSignedMaxValue();
5101 } else {
5102 Op0Min = Op0Known.getMinValue();
5103 Op0Max = Op0Known.getMaxValue();
5104 Op1Min = Op1Known.getMinValue();
5105 Op1Max = Op1Known.getMaxValue();
5106 }
5107
5108 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
5109 // out that the LHS or RHS is a constant. Constant fold this now, so that
5110 // code below can assume that Min != Max.
5111 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
5112 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
5113 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
5114 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
5115
5116 // Based on the range information we know about the LHS, see if we can
5117 // simplify this comparison. For example, (x&4) < 8 is always true.
5118 switch (Pred) {
5119 default:
5120 llvm_unreachable("Unknown icmp opcode!")::llvm::llvm_unreachable_internal("Unknown icmp opcode!", "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5120)
;
5121 case ICmpInst::ICMP_EQ:
5122 case ICmpInst::ICMP_NE: {
5123 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
5124 return replaceInstUsesWith(
5125 I, ConstantInt::getBool(I.getType(), Pred == CmpInst::ICMP_NE));
5126
5127 // If all bits are known zero except for one, then we know at most one bit
5128 // is set. If the comparison is against zero, then this is a check to see if
5129 // *that* bit is set.
5130 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
5131 if (Op1Known.isZero()) {
5132 // If the LHS is an AND with the same constant, look through it.
5133 Value *LHS = nullptr;
5134 const APInt *LHSC;
5135 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
5136 *LHSC != Op0KnownZeroInverted)
5137 LHS = Op0;
5138
5139 Value *X;
5140 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
5141 APInt ValToCheck = Op0KnownZeroInverted;
5142 Type *XTy = X->getType();
5143 if (ValToCheck.isPowerOf2()) {
5144 // ((1 << X) & 8) == 0 -> X != 3
5145 // ((1 << X) & 8) != 0 -> X == 3
5146 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5147 auto NewPred = ICmpInst::getInversePredicate(Pred);
5148 return new ICmpInst(NewPred, X, CmpC);
5149 } else if ((++ValToCheck).isPowerOf2()) {
5150 // ((1 << X) & 7) == 0 -> X >= 3
5151 // ((1 << X) & 7) != 0 -> X < 3
5152 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5153 auto NewPred =
5154 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
5155 return new ICmpInst(NewPred, X, CmpC);
5156 }
5157 }
5158
5159 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5160 const APInt *CI;
5161 if (Op0KnownZeroInverted.isOneValue() &&
5162 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5163 // ((8 >>u X) & 1) == 0 -> X != 3
5164 // ((8 >>u X) & 1) != 0 -> X == 3
5165 unsigned CmpVal = CI->countTrailingZeros();
5166 auto NewPred = ICmpInst::getInversePredicate(Pred);
5167 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5168 }
5169 }
5170 break;
5171 }
5172 case ICmpInst::ICMP_ULT: {
5173 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5174 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5175 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5176 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5177 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5178 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5179
5180 const APInt *CmpC;
5181 if (match(Op1, m_APInt(CmpC))) {
5182 // A <u C -> A == C-1 if min(A)+1 == C
5183 if (*CmpC == Op0Min + 1)
5184 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5185 ConstantInt::get(Op1->getType(), *CmpC - 1));
5186 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5187 // exceeds the log2 of C.
5188 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5189 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5190 Constant::getNullValue(Op1->getType()));
5191 }
5192 break;
5193 }
5194 case ICmpInst::ICMP_UGT: {
5195 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5196 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5197 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5198 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5199 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5200 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5201
5202 const APInt *CmpC;
5203 if (match(Op1, m_APInt(CmpC))) {
5204 // A >u C -> A == C+1 if max(a)-1 == C
5205 if (*CmpC == Op0Max - 1)
5206 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5207 ConstantInt::get(Op1->getType(), *CmpC + 1));
5208 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5209 // exceeds the log2 of C.
5210 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5211 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5212 Constant::getNullValue(Op1->getType()));
5213 }
5214 break;
5215 }
5216 case ICmpInst::ICMP_SLT: {
5217 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5218 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5219 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5220 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5221 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5222 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5223 const APInt *CmpC;
5224 if (match(Op1, m_APInt(CmpC))) {
5225 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5226 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5227 ConstantInt::get(Op1->getType(), *CmpC - 1));
5228 }
5229 break;
5230 }
5231 case ICmpInst::ICMP_SGT: {
5232 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5233 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5234 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5235 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5236 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5237 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5238 const APInt *CmpC;
5239 if (match(Op1, m_APInt(CmpC))) {
5240 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5241 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5242 ConstantInt::get(Op1->getType(), *CmpC + 1));
5243 }
5244 break;
5245 }
5246 case ICmpInst::ICMP_SGE:
5247 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!")(static_cast <bool> (!isa<ConstantInt>(Op1) &&
"ICMP_SGE with ConstantInt not folded!") ? void (0) : __assert_fail
("!isa<ConstantInt>(Op1) && \"ICMP_SGE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5247, __extension__ __PRETTY_FUNCTION__))
;
5248 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5249 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5250 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5251 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5252 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5253 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5254 break;
5255 case ICmpInst::ICMP_SLE:
5256 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!")(static_cast <bool> (!isa<ConstantInt>(Op1) &&
"ICMP_SLE with ConstantInt not folded!") ? void (0) : __assert_fail
("!isa<ConstantInt>(Op1) && \"ICMP_SLE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5256, __extension__ __PRETTY_FUNCTION__))
;
5257 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5258 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5259 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5260 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5261 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5262 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5263 break;
5264 case ICmpInst::ICMP_UGE:
5265 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!")(static_cast <bool> (!isa<ConstantInt>(Op1) &&
"ICMP_UGE with ConstantInt not folded!") ? void (0) : __assert_fail
("!isa<ConstantInt>(Op1) && \"ICMP_UGE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5265, __extension__ __PRETTY_FUNCTION__))
;
5266 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5267 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5268 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5269 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5270 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5271 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5272 break;
5273 case ICmpInst::ICMP_ULE:
5274 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!")(static_cast <bool> (!isa<ConstantInt>(Op1) &&
"ICMP_ULE with ConstantInt not folded!") ? void (0) : __assert_fail
("!isa<ConstantInt>(Op1) && \"ICMP_ULE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5274, __extension__ __PRETTY_FUNCTION__))
;
5275 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5276 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5277 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5278 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5279 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5280 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5281 break;
5282 }
5283
5284 // Turn a signed comparison into an unsigned one if both operands are known to
5285 // have the same sign.
5286 if (I.isSigned() &&
5287 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5288 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5289 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5290
5291 return nullptr;
5292}
5293
5294llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
5295InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5296 Constant *C) {
5297 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&(static_cast <bool> (ICmpInst::isRelational(Pred) &&
ICmpInst::isIntPredicate(Pred) && "Only for relational integer predicates."
) ? void (0) : __assert_fail ("ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && \"Only for relational integer predicates.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5298, __extension__ __PRETTY_FUNCTION__))
5298 "Only for relational integer predicates.")(static_cast <bool> (ICmpInst::isRelational(Pred) &&
ICmpInst::isIntPredicate(Pred) && "Only for relational integer predicates."
) ? void (0) : __assert_fail ("ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && \"Only for relational integer predicates.\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5298, __extension__ __PRETTY_FUNCTION__))
;
5299
5300 Type *Type = C->getType();
5301 bool IsSigned = ICmpInst::isSigned(Pred);
5302
5303 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5304 bool WillIncrement =
5305 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5306
5307 // Check if the constant operand can be safely incremented/decremented
5308 // without overflowing/underflowing.
5309 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5310 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5311 };
5312
5313 Constant *SafeReplacementConstant = nullptr;
5314 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5315 // Bail out if the constant can't be safely incremented/decremented.
5316 if (!ConstantIsOk(CI))
5317 return llvm::None;
5318 } else if (auto *FVTy = dyn_cast<FixedVectorType>(Type)) {
5319 unsigned NumElts = FVTy->getNumElements();
5320 for (unsigned i = 0; i != NumElts; ++i) {
5321 Constant *Elt = C->getAggregateElement(i);
5322 if (!Elt)
5323 return llvm::None;
5324
5325 if (isa<UndefValue>(Elt))
5326 continue;
5327
5328 // Bail out if we can't determine if this constant is min/max or if we
5329 // know that this constant is min/max.
5330 auto *CI = dyn_cast<ConstantInt>(Elt);
5331 if (!CI || !ConstantIsOk(CI))
5332 return llvm::None;
5333
5334 if (!SafeReplacementConstant)
5335 SafeReplacementConstant = CI;
5336 }
5337 } else {
5338 // ConstantExpr?
5339 return llvm::None;
5340 }
5341
5342 // It may not be safe to change a compare predicate in the presence of
5343 // undefined elements, so replace those elements with the first safe constant
5344 // that we found.
5345 // TODO: in case of poison, it is safe; let's replace undefs only.
5346 if (C->containsUndefOrPoisonElement()) {
5347 assert(SafeReplacementConstant && "Replacement constant not set")(static_cast <bool> (SafeReplacementConstant &&
"Replacement constant not set") ? void (0) : __assert_fail (
"SafeReplacementConstant && \"Replacement constant not set\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5347, __extension__ __PRETTY_FUNCTION__))
;
5348 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5349 }
5350
5351 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5352
5353 // Increment or decrement the constant.
5354 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5355 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5356
5357 return std::make_pair(NewPred, NewC);
5358}
5359
5360/// If we have an icmp le or icmp ge instruction with a constant operand, turn
5361/// it into the appropriate icmp lt or icmp gt instruction. This transform
5362/// allows them to be folded in visitICmpInst.
5363static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5364 ICmpInst::Predicate Pred = I.getPredicate();
5365 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5366 InstCombiner::isCanonicalPredicate(Pred))
5367 return nullptr;
5368
5369 Value *Op0 = I.getOperand(0);
5370 Value *Op1 = I.getOperand(1);
5371 auto *Op1C = dyn_cast<Constant>(Op1);
5372 if (!Op1C)
5373 return nullptr;
5374
5375 auto FlippedStrictness =
5376 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5377 if (!FlippedStrictness)
5378 return nullptr;
5379
5380 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5381}
5382
5383/// If we have a comparison with a non-canonical predicate, if we can update
5384/// all the users, invert the predicate and adjust all the users.
5385CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
5386 // Is the predicate already canonical?
5387 CmpInst::Predicate Pred = I.getPredicate();
5388 if (InstCombiner::isCanonicalPredicate(Pred))
5389 return nullptr;
5390
5391 // Can all users be adjusted to predicate inversion?
5392 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
5393 return nullptr;
5394
5395 // Ok, we can canonicalize comparison!
5396 // Let's first invert the comparison's predicate.
5397 I.setPredicate(CmpInst::getInversePredicate(Pred));
5398 I.setName(I.getName() + ".not");
5399
5400 // And, adapt users.
5401 freelyInvertAllUsersOf(&I);
5402
5403 return &I;
5404}
5405
5406/// Integer compare with boolean values can always be turned into bitwise ops.
5407static Instruction *canonicalizeICmpBool(ICmpInst &I,
5408 InstCombiner::BuilderTy &Builder) {
5409 Value *A = I.getOperand(0), *B = I.getOperand(1);
5410 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only")(static_cast <bool> (A->getType()->isIntOrIntVectorTy
(1) && "Bools only") ? void (0) : __assert_fail ("A->getType()->isIntOrIntVectorTy(1) && \"Bools only\""
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5410, __extension__ __PRETTY_FUNCTION__))
;
5411
5412 // A boolean compared to true/false can be simplified to Op0/true/false in
5413 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5414 // Cases not handled by InstSimplify are always 'not' of Op0.
5415 if (match(B, m_Zero())) {
5416 switch (I.getPredicate()) {
5417 case CmpInst::ICMP_EQ: // A == 0 -> !A
5418 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5419 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5420 return BinaryOperator::CreateNot(A);
5421 default:
5422 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-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5422)
;
5423 }
5424 } else if (match(B, m_One())) {
5425 switch (I.getPredicate()) {
5426 case CmpInst::ICMP_NE: // A != 1 -> !A
5427 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5428 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5429 return BinaryOperator::CreateNot(A);
5430 default:
5431 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-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5431)
;
5432 }
5433 }
5434
5435 switch (I.getPredicate()) {
5436 default:
5437 llvm_unreachable("Invalid icmp instruction!")::llvm::llvm_unreachable_internal("Invalid icmp instruction!"
, "/build/llvm-toolchain-snapshot-13~++20210621111111+acefe0eaaf82/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5437)
;
5438 case ICmpInst::ICMP_EQ:
5439 // icmp eq i1 A, B -> ~(A ^ B)
5440 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5441
5442 case ICmpInst::ICMP_NE:
5443 // icmp ne i1 A, B -> A ^ B
5444 return BinaryOperator::CreateXor(A, B);
5445
5446 case ICmpInst::ICMP_UGT:
5447 // icmp ugt -> icmp ult
5448 std::swap(A, B);
5449 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5450 case ICmpInst::ICMP_ULT:
5451 // icmp ult i1 A, B -> ~A & B
5452 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5453
5454 case ICmpInst::ICMP_SGT:
5455 // icmp sgt -> icmp slt
5456 std::swap(A, B);
5457 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5458 case ICmpInst::ICMP_SLT:
5459 // icmp slt i1 A, B -> A & ~B
5460 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5461
5462 case ICmpInst::ICMP_UGE:
5463 // icmp uge -> icmp ule
5464 std::swap(A, B);
5465 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5466 case ICmpInst::ICMP_ULE:
5467 // icmp ule i1 A, B -> ~A | B
5468 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5469
5470 case ICmpInst::ICMP_SGE:
5471 // icmp sge -> icmp sle
5472 std::swap(A, B);
5473 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5474 case ICmpInst::ICMP_SLE:
5475 // icmp sle i1 A, B -> A | ~B
5476 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5477 }
5478}
5479
5480// Transform pattern like:
5481// (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5482// (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5483// Into:
5484// (X l>> Y) != 0
5485// (X l>> Y) == 0
5486static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5487 InstCombiner::BuilderTy &Builder) {
5488 ICmpInst::Predicate Pred, NewPred;
5489 Value *X, *Y;
5490 if (match(&Cmp,
5491 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5492 switch (Pred) {
5493 case ICmpInst::ICMP_ULE:
5494 NewPred = ICmpInst::ICMP_NE;
5495 break;
5496 case ICmpInst::ICMP_UGT:
5497 NewPred = ICmpInst::ICMP_EQ;
5498 break;
5499 default:
5500 return nullptr;
5501 }
5502 } else if (match(&Cmp, m_c_ICmp(Pred,
5503 m_OneUse(m_CombineOr(
5504 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5505 m_Add(m_Shl(m_One(), m_Value(Y)),
5506 m_AllOnes()))),
5507 m_Value(X)))) {
5508 // The variant with 'add' is not canonical, (the variant with 'not' is)
5509 // we only get it because it has extra uses, and can't be canonicalized,
5510
5511 switch (Pred) {
5512 case ICmpInst::ICMP_ULT:
5513 NewPred = ICmpInst::ICMP_NE;
5514 break;
5515 case ICmpInst::ICMP_UGE:
5516 NewPred = ICmpInst::ICMP_EQ;
5517 break;
5518 default:
5519 return nullptr;
5520 }
5521 } else
5522 return nullptr;
5523
5524 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5525 Constant *Zero = Constant::getNullValue(NewX->getType());
5526 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5527}
5528
5529static Instruction *foldVectorCmp(CmpInst &Cmp,
5530 InstCombiner::BuilderTy &Builder) {
5531 const CmpInst::Predicate Pred = Cmp.getPredicate();
5532 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5533 Value *V1, *V2;
5534 ArrayRef<int> M;
5535 if (!match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(M))))
5536 return nullptr;
5537
5538 // If both arguments of the cmp are shuffles that use the same mask and
5539 // shuffle within a single vector, move the shuffle after the cmp:
5540 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5541 Type *V1Ty = V1->getType();
5542 if (match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(M))) &&
5543 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
5544 Value *NewCmp = Builder.CreateCmp(Pred, V1, V2);
5545 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5546 }
5547
5548 // Try to canonicalize compare with splatted operand and splat constant.
5549 // TODO: We could generalize this for more than splats. See/use the code in
5550 // InstCombiner::foldVectorBinop().
5551 Constant *C;
5552 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
5553 return nullptr;
5554
5555 // Length-changing splats are ok, so adjust the constants as needed:
5556 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
5557 Constant *ScalarC = C->getSplatValue(/* AllowUndefs */ true);
5558 int MaskSplatIndex;
5559 if (ScalarC && match(M, m_SplatOrUndefMask(MaskSplatIndex))) {
5560 // We allow undefs in matching, but this transform removes those for safety.
5561 // Demanded elements analysis should be able to recover some/all of that.
5562 C = ConstantVector::getSplat(cast<VectorType>(V1Ty)->getElementCount(),
5563 ScalarC);
5564 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
5565 Value *NewCmp = Builder.CreateCmp(Pred, V1, C);
5566 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()),
5567 NewM);
5568 }
5569
5570 return nullptr;
5571}
5572
5573// extract(uadd.with.overflow(A, B), 0) ult A
5574// -> extract(uadd.with.overflow(A, B), 1)
5575static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5576 CmpInst::Predicate Pred = I.getPredicate();
5577 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5578
5579 Value *UAddOv;
5580 Value *A, *B;
5581 auto UAddOvResultPat = m_ExtractValue<0>(
5582 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5583 if (match(Op0, UAddOvResultPat) &&
5584 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5585 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5586 (match(A, m_One()) || match(B, m_One()))) ||
5587 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5588 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5589 // extract(uadd.with.overflow(A, B), 0) < A
5590 // extract(uadd.with.overflow(A, 1), 0) == 0
5591 // extract(uadd.with.overflow(A, -1), 0) != -1
5592 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5593 else if (match(Op1, UAddOvResultPat) &&
5594 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5595 // A > extract(uadd.with.overflow(A, B), 0)
5596 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5597 else
5598 return nullptr;
5599
5600 return ExtractValueInst::Create(UAddOv, 1);
5601}
5602
5603Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
5604 bool Changed = false;
5605 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5606 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5607 unsigned Op0Cplxity = getComplexity(Op0);
5608 unsigned Op1Cplxity = getComplexity(Op1);
5609
5610 /// Orders the operands of the compare so that they are listed from most
5611 /// complex to least complex. This puts constants before unary operators,