1 //===- SeparateConstOffsetFromGEP.cpp -------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Loop unrolling may create many similar GEPs for array accesses.
11 // e.g., a 2-level loop
13 // float a[32][32]; // global variable
15 // for (int i = 0; i < 2; ++i) {
16 // for (int j = 0; j < 2; ++j) {
18 // ... = a[x + i][y + j];
23 // will probably be unrolled to:
25 // gep %a, 0, %x, %y; load
26 // gep %a, 0, %x, %y + 1; load
27 // gep %a, 0, %x + 1, %y; load
28 // gep %a, 0, %x + 1, %y + 1; load
30 // LLVM's GVN does not use partial redundancy elimination yet, and is thus
31 // unable to reuse (gep %a, 0, %x, %y). As a result, this misoptimization incurs
32 // significant slowdown in targets with limited addressing modes. For instance,
33 // because the PTX target does not support the reg+reg addressing mode, the
34 // NVPTX backend emits PTX code that literally computes the pointer address of
35 // each GEP, wasting tons of registers. It emits the following PTX for the
36 // first load and similar PTX for other loads.
40 // mul.wide.u32 %rl2, %r1, 128;
42 // add.s64 %rl4, %rl3, %rl2;
43 // mul.wide.u32 %rl5, %r2, 4;
44 // add.s64 %rl6, %rl4, %rl5;
45 // ld.global.f32 %f1, [%rl6];
47 // To reduce the register pressure, the optimization implemented in this file
48 // merges the common part of a group of GEPs, so we can compute each pointer
49 // address by adding a simple offset to the common part, saving many registers.
51 // It works by splitting each GEP into a variadic base and a constant offset.
52 // The variadic base can be computed once and reused by multiple GEPs, and the
53 // constant offsets can be nicely folded into the reg+immediate addressing mode
54 // (supported by most targets) without using any extra register.
56 // For instance, we transform the four GEPs and four loads in the above example
59 // base = gep a, 0, x, y
61 // laod base + 1 * sizeof(float)
62 // load base + 32 * sizeof(float)
63 // load base + 33 * sizeof(float)
65 // Given the transformed IR, a backend that supports the reg+immediate
66 // addressing mode can easily fold the pointer arithmetics into the loads. For
67 // example, the NVPTX backend can easily fold the pointer arithmetics into the
68 // ld.global.f32 instructions, and the resultant PTX uses much fewer registers.
70 // mov.u32 %r1, %tid.x;
71 // mov.u32 %r2, %tid.y;
72 // mul.wide.u32 %rl2, %r1, 128;
74 // add.s64 %rl4, %rl3, %rl2;
75 // mul.wide.u32 %rl5, %r2, 4;
76 // add.s64 %rl6, %rl4, %rl5;
77 // ld.global.f32 %f1, [%rl6]; // so far the same as unoptimized PTX
78 // ld.global.f32 %f2, [%rl6+4]; // much better
79 // ld.global.f32 %f3, [%rl6+128]; // much better
80 // ld.global.f32 %f4, [%rl6+132]; // much better
82 // Another improvement enabled by the LowerGEP flag is to lower a GEP with
83 // multiple indices to either multiple GEPs with a single index or arithmetic
84 // operations (depending on whether the target uses alias analysis in codegen).
85 // Such transformation can have following benefits:
86 // (1) It can always extract constants in the indices of structure type.
87 // (2) After such Lowering, there are more optimization opportunities such as
90 // E.g. The following GEPs have multiple indices:
92 // %p = getelementptr [10 x %struct]* %ptr, i64 %i, i64 %j1, i32 3
96 // %p2 = getelementptr [10 x %struct]* %ptr, i64 %i, i64 %j1, i32 2
100 // We can not do CSE to the common part related to index "i64 %i". Lowering
101 // GEPs can achieve such goals.
102 // If the target does not use alias analysis in codegen, this pass will
103 // lower a GEP with multiple indices into arithmetic operations:
105 // %1 = ptrtoint [10 x %struct]* %ptr to i64 ; CSE opportunity
106 // %2 = mul i64 %i, length_of_10xstruct ; CSE opportunity
107 // %3 = add i64 %1, %2 ; CSE opportunity
108 // %4 = mul i64 %j1, length_of_struct
109 // %5 = add i64 %3, %4
110 // %6 = add i64 %3, struct_field_3 ; Constant offset
111 // %p = inttoptr i64 %6 to i32*
115 // %7 = ptrtoint [10 x %struct]* %ptr to i64 ; CSE opportunity
116 // %8 = mul i64 %i, length_of_10xstruct ; CSE opportunity
117 // %9 = add i64 %7, %8 ; CSE opportunity
118 // %10 = mul i64 %j2, length_of_struct
119 // %11 = add i64 %9, %10
120 // %12 = add i64 %11, struct_field_2 ; Constant offset
121 // %p = inttoptr i64 %12 to i32*
125 // If the target uses alias analysis in codegen, this pass will lower a GEP
126 // with multiple indices into multiple GEPs with a single index:
128 // %1 = bitcast [10 x %struct]* %ptr to i8* ; CSE opportunity
129 // %2 = mul i64 %i, length_of_10xstruct ; CSE opportunity
130 // %3 = getelementptr i8* %1, i64 %2 ; CSE opportunity
131 // %4 = mul i64 %j1, length_of_struct
132 // %5 = getelementptr i8* %3, i64 %4
133 // %6 = getelementptr i8* %5, struct_field_3 ; Constant offset
134 // %p = bitcast i8* %6 to i32*
138 // %7 = bitcast [10 x %struct]* %ptr to i8* ; CSE opportunity
139 // %8 = mul i64 %i, length_of_10xstruct ; CSE opportunity
140 // %9 = getelementptr i8* %7, i64 %8 ; CSE opportunity
141 // %10 = mul i64 %j2, length_of_struct
142 // %11 = getelementptr i8* %9, i64 %10
143 // %12 = getelementptr i8* %11, struct_field_2 ; Constant offset
144 // %p2 = bitcast i8* %12 to i32*
148 // Lowering GEPs can also benefit other passes such as LICM and CGP.
149 // LICM (Loop Invariant Code Motion) can not hoist/sink a GEP of multiple
150 // indices if one of the index is variant. If we lower such GEP into invariant
151 // parts and variant parts, LICM can hoist/sink those invariant parts.
152 // CGP (CodeGen Prepare) tries to sink address calculations that match the
153 // target's addressing modes. A GEP with multiple indices may not match and will
154 // not be sunk. If we lower such GEP into smaller parts, CGP may sink some of
155 // them. So we end up with a better addressing mode.
157 //===----------------------------------------------------------------------===//
159 #include "llvm/ADT/APInt.h"
160 #include "llvm/ADT/DenseMap.h"
161 #include "llvm/ADT/DepthFirstIterator.h"
162 #include "llvm/ADT/SmallVector.h"
163 #include "llvm/Analysis/LoopInfo.h"
164 #include "llvm/Analysis/MemoryBuiltins.h"
165 #include "llvm/Analysis/ScalarEvolution.h"
166 #include "llvm/Analysis/TargetLibraryInfo.h"
167 #include "llvm/Analysis/TargetTransformInfo.h"
168 #include "llvm/Analysis/ValueTracking.h"
169 #include "llvm/CodeGen/TargetSubtargetInfo.h"
170 #include "llvm/IR/BasicBlock.h"
171 #include "llvm/IR/Constant.h"
172 #include "llvm/IR/Constants.h"
173 #include "llvm/IR/DataLayout.h"
174 #include "llvm/IR/DerivedTypes.h"
175 #include "llvm/IR/Dominators.h"
176 #include "llvm/IR/Function.h"
177 #include "llvm/IR/GetElementPtrTypeIterator.h"
178 #include "llvm/IR/IRBuilder.h"
179 #include "llvm/IR/Instruction.h"
180 #include "llvm/IR/Instructions.h"
181 #include "llvm/IR/Module.h"
182 #include "llvm/IR/PatternMatch.h"
183 #include "llvm/IR/Type.h"
184 #include "llvm/IR/User.h"
185 #include "llvm/IR/Value.h"
186 #include "llvm/Pass.h"
187 #include "llvm/Support/Casting.h"
188 #include "llvm/Support/CommandLine.h"
189 #include "llvm/Support/ErrorHandling.h"
190 #include "llvm/Support/raw_ostream.h"
191 #include "llvm/Target/TargetMachine.h"
192 #include "llvm/Transforms/Scalar.h"
193 #include "llvm/Transforms/Utils/Local.h"
198 using namespace llvm;
199 using namespace llvm::PatternMatch;
201 static cl::opt<bool> DisableSeparateConstOffsetFromGEP(
202 "disable-separate-const-offset-from-gep", cl::init(false),
203 cl::desc("Do not separate the constant offset from a GEP instruction"),
206 // Setting this flag may emit false positives when the input module already
207 // contains dead instructions. Therefore, we set it only in unit tests that are
208 // free of dead code.
210 VerifyNoDeadCode("reassociate-geps-verify-no-dead-code", cl::init(false),
211 cl::desc("Verify this pass produces no dead code"),
216 /// \brief A helper class for separating a constant offset from a GEP index.
218 /// In real programs, a GEP index may be more complicated than a simple addition
219 /// of something and a constant integer which can be trivially splitted. For
220 /// example, to split ((a << 3) | 5) + b, we need to search deeper for the
221 /// constant offset, so that we can separate the index to (a << 3) + b and 5.
223 /// Therefore, this class looks into the expression that computes a given GEP
224 /// index, and tries to find a constant integer that can be hoisted to the
225 /// outermost level of the expression as an addition. Not every constant in an
226 /// expression can jump out. e.g., we cannot transform (b * (a + 5)) to (b * a +
227 /// 5); nor can we transform (3 * (a + 5)) to (3 * a + 5), however in this case,
228 /// -instcombine probably already optimized (3 * (a + 5)) to (3 * a + 15).
229 class ConstantOffsetExtractor {
231 /// Extracts a constant offset from the given GEP index. It returns the
232 /// new index representing the remainder (equal to the original index minus
233 /// the constant offset), or nullptr if we cannot extract a constant offset.
234 /// \p Idx The given GEP index
235 /// \p GEP The given GEP
236 /// \p UserChainTail Outputs the tail of UserChain so that we can
237 /// garbage-collect unused instructions in UserChain.
238 static Value *Extract(Value *Idx, GetElementPtrInst *GEP,
239 User *&UserChainTail, const DominatorTree *DT);
241 /// Looks for a constant offset from the given GEP index without extracting
242 /// it. It returns the numeric value of the extracted constant offset (0 if
243 /// failed). The meaning of the arguments are the same as Extract.
244 static int64_t Find(Value *Idx, GetElementPtrInst *GEP,
245 const DominatorTree *DT);
248 ConstantOffsetExtractor(Instruction *InsertionPt, const DominatorTree *DT)
249 : IP(InsertionPt), DL(InsertionPt->getModule()->getDataLayout()), DT(DT) {
252 /// Searches the expression that computes V for a non-zero constant C s.t.
253 /// V can be reassociated into the form V' + C. If the searching is
254 /// successful, returns C and update UserChain as a def-use chain from C to V;
255 /// otherwise, UserChain is empty.
257 /// \p V The given expression
258 /// \p SignExtended Whether V will be sign-extended in the computation of the
260 /// \p ZeroExtended Whether V will be zero-extended in the computation of the
262 /// \p NonNegative Whether V is guaranteed to be non-negative. For example,
263 /// an index of an inbounds GEP is guaranteed to be
264 /// non-negative. Levaraging this, we can better split
266 APInt find(Value *V, bool SignExtended, bool ZeroExtended, bool NonNegative);
268 /// A helper function to look into both operands of a binary operator.
269 APInt findInEitherOperand(BinaryOperator *BO, bool SignExtended,
272 /// After finding the constant offset C from the GEP index I, we build a new
273 /// index I' s.t. I' + C = I. This function builds and returns the new
274 /// index I' according to UserChain produced by function "find".
276 /// The building conceptually takes two steps:
277 /// 1) iteratively distribute s/zext towards the leaves of the expression tree
279 /// 2) reassociate the expression tree to the form I' + C.
281 /// For example, to extract the 5 from sext(a + (b + 5)), we first distribute
282 /// sext to a, b and 5 so that we have
283 /// sext(a) + (sext(b) + 5).
284 /// Then, we reassociate it to
285 /// (sext(a) + sext(b)) + 5.
286 /// Given this form, we know I' is sext(a) + sext(b).
287 Value *rebuildWithoutConstOffset();
289 /// After the first step of rebuilding the GEP index without the constant
290 /// offset, distribute s/zext to the operands of all operators in UserChain.
291 /// e.g., zext(sext(a + (b + 5)) (assuming no overflow) =>
292 /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))).
294 /// The function also updates UserChain to point to new subexpressions after
295 /// distributing s/zext. e.g., the old UserChain of the above example is
296 /// 5 -> b + 5 -> a + (b + 5) -> sext(...) -> zext(sext(...)),
297 /// and the new UserChain is
298 /// zext(sext(5)) -> zext(sext(b)) + zext(sext(5)) ->
299 /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))
301 /// \p ChainIndex The index to UserChain. ChainIndex is initially
302 /// UserChain.size() - 1, and is decremented during
304 Value *distributeExtsAndCloneChain(unsigned ChainIndex);
306 /// Reassociates the GEP index to the form I' + C and returns I'.
307 Value *removeConstOffset(unsigned ChainIndex);
309 /// A helper function to apply ExtInsts, a list of s/zext, to value V.
310 /// e.g., if ExtInsts = [sext i32 to i64, zext i16 to i32], this function
311 /// returns "sext i32 (zext i16 V to i32) to i64".
312 Value *applyExts(Value *V);
314 /// A helper function that returns whether we can trace into the operands
315 /// of binary operator BO for a constant offset.
317 /// \p SignExtended Whether BO is surrounded by sext
318 /// \p ZeroExtended Whether BO is surrounded by zext
319 /// \p NonNegative Whether BO is known to be non-negative, e.g., an in-bound
321 bool CanTraceInto(bool SignExtended, bool ZeroExtended, BinaryOperator *BO,
324 /// The path from the constant offset to the old GEP index. e.g., if the GEP
325 /// index is "a * b + (c + 5)". After running function find, UserChain[0] will
326 /// be the constant 5, UserChain[1] will be the subexpression "c + 5", and
327 /// UserChain[2] will be the entire expression "a * b + (c + 5)".
329 /// This path helps to rebuild the new GEP index.
330 SmallVector<User *, 8> UserChain;
332 /// A data structure used in rebuildWithoutConstOffset. Contains all
333 /// sext/zext instructions along UserChain.
334 SmallVector<CastInst *, 16> ExtInsts;
336 /// Insertion position of cloned instructions.
339 const DataLayout &DL;
340 const DominatorTree *DT;
343 /// \brief A pass that tries to split every GEP in the function into a variadic
344 /// base and a constant offset. It is a FunctionPass because searching for the
345 /// constant offset may inspect other basic blocks.
346 class SeparateConstOffsetFromGEP : public FunctionPass {
350 SeparateConstOffsetFromGEP(const TargetMachine *TM = nullptr,
351 bool LowerGEP = false)
352 : FunctionPass(ID), TM(TM), LowerGEP(LowerGEP) {
353 initializeSeparateConstOffsetFromGEPPass(*PassRegistry::getPassRegistry());
356 void getAnalysisUsage(AnalysisUsage &AU) const override {
357 AU.addRequired<DominatorTreeWrapperPass>();
358 AU.addRequired<ScalarEvolutionWrapperPass>();
359 AU.addRequired<TargetTransformInfoWrapperPass>();
360 AU.addRequired<LoopInfoWrapperPass>();
361 AU.setPreservesCFG();
362 AU.addRequired<TargetLibraryInfoWrapperPass>();
365 bool doInitialization(Module &M) override {
366 DL = &M.getDataLayout();
370 bool runOnFunction(Function &F) override;
373 /// Tries to split the given GEP into a variadic base and a constant offset,
374 /// and returns true if the splitting succeeds.
375 bool splitGEP(GetElementPtrInst *GEP);
377 /// Lower a GEP with multiple indices into multiple GEPs with a single index.
378 /// Function splitGEP already split the original GEP into a variadic part and
379 /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
380 /// variadic part into a set of GEPs with a single index and applies
381 /// AccumulativeByteOffset to it.
382 /// \p Variadic The variadic part of the original GEP.
383 /// \p AccumulativeByteOffset The constant offset.
384 void lowerToSingleIndexGEPs(GetElementPtrInst *Variadic,
385 int64_t AccumulativeByteOffset);
387 /// Lower a GEP with multiple indices into ptrtoint+arithmetics+inttoptr form.
388 /// Function splitGEP already split the original GEP into a variadic part and
389 /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
390 /// variadic part into a set of arithmetic operations and applies
391 /// AccumulativeByteOffset to it.
392 /// \p Variadic The variadic part of the original GEP.
393 /// \p AccumulativeByteOffset The constant offset.
394 void lowerToArithmetics(GetElementPtrInst *Variadic,
395 int64_t AccumulativeByteOffset);
397 /// Finds the constant offset within each index and accumulates them. If
398 /// LowerGEP is true, it finds in indices of both sequential and structure
399 /// types, otherwise it only finds in sequential indices. The output
400 /// NeedsExtraction indicates whether we successfully find a non-zero constant
402 int64_t accumulateByteOffset(GetElementPtrInst *GEP, bool &NeedsExtraction);
404 /// Canonicalize array indices to pointer-size integers. This helps to
405 /// simplify the logic of splitting a GEP. For example, if a + b is a
406 /// pointer-size integer, we have
407 /// gep base, a + b = gep (gep base, a), b
408 /// However, this equality may not hold if the size of a + b is smaller than
409 /// the pointer size, because LLVM conceptually sign-extends GEP indices to
410 /// pointer size before computing the address
411 /// (http://llvm.org/docs/LangRef.html#id181).
413 /// This canonicalization is very likely already done in clang and
414 /// instcombine. Therefore, the program will probably remain the same.
416 /// Returns true if the module changes.
418 /// Verified in @i32_add in split-gep.ll
419 bool canonicalizeArrayIndicesToPointerSize(GetElementPtrInst *GEP);
421 /// Optimize sext(a)+sext(b) to sext(a+b) when a+b can't sign overflow.
422 /// SeparateConstOffsetFromGEP distributes a sext to leaves before extracting
423 /// the constant offset. After extraction, it becomes desirable to reunion the
424 /// distributed sexts. For example,
426 /// &a[sext(i +nsw (j +nsw 5)]
427 /// => distribute &a[sext(i) +nsw (sext(j) +nsw 5)]
428 /// => constant extraction &a[sext(i) + sext(j)] + 5
429 /// => reunion &a[sext(i +nsw j)] + 5
430 bool reuniteExts(Function &F);
432 /// A helper that reunites sexts in an instruction.
433 bool reuniteExts(Instruction *I);
435 /// Find the closest dominator of <Dominatee> that is equivalent to <Key>.
436 Instruction *findClosestMatchingDominator(const SCEV *Key,
437 Instruction *Dominatee);
438 /// Verify F is free of dead code.
439 void verifyNoDeadCode(Function &F);
441 bool hasMoreThanOneUseInLoop(Value *v, Loop *L);
443 // Swap the index operand of two GEP.
444 void swapGEPOperand(GetElementPtrInst *First, GetElementPtrInst *Second);
446 // Check if it is safe to swap operand of two GEP.
447 bool isLegalToSwapOperand(GetElementPtrInst *First, GetElementPtrInst *Second,
450 const DataLayout *DL = nullptr;
451 DominatorTree *DT = nullptr;
453 const TargetMachine *TM;
456 TargetLibraryInfo *TLI;
458 /// Whether to lower a GEP with multiple indices into arithmetic operations or
459 /// multiple GEPs with a single index.
462 DenseMap<const SCEV *, SmallVector<Instruction *, 2>> DominatingExprs;
465 } // end anonymous namespace
467 char SeparateConstOffsetFromGEP::ID = 0;
469 INITIALIZE_PASS_BEGIN(
470 SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
471 "Split GEPs to a variadic base and a constant offset for better CSE", false,
473 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
474 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
475 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
476 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
477 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
479 SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
480 "Split GEPs to a variadic base and a constant offset for better CSE", false,
484 llvm::createSeparateConstOffsetFromGEPPass(const TargetMachine *TM,
486 return new SeparateConstOffsetFromGEP(TM, LowerGEP);
489 bool ConstantOffsetExtractor::CanTraceInto(bool SignExtended,
493 // We only consider ADD, SUB and OR, because a non-zero constant found in
494 // expressions composed of these operations can be easily hoisted as a
495 // constant offset by reassociation.
496 if (BO->getOpcode() != Instruction::Add &&
497 BO->getOpcode() != Instruction::Sub &&
498 BO->getOpcode() != Instruction::Or) {
502 Value *LHS = BO->getOperand(0), *RHS = BO->getOperand(1);
503 // Do not trace into "or" unless it is equivalent to "add". If LHS and RHS
504 // don't have common bits, (LHS | RHS) is equivalent to (LHS + RHS).
505 if (BO->getOpcode() == Instruction::Or &&
506 !haveNoCommonBitsSet(LHS, RHS, DL, nullptr, BO, DT))
509 // In addition, tracing into BO requires that its surrounding s/zext (if
510 // any) is distributable to both operands.
512 // Suppose BO = A op B.
513 // SignExtended | ZeroExtended | Distributable?
514 // --------------+--------------+----------------------------------
515 // 0 | 0 | true because no s/zext exists
516 // 0 | 1 | zext(BO) == zext(A) op zext(B)
517 // 1 | 0 | sext(BO) == sext(A) op sext(B)
518 // 1 | 1 | zext(sext(BO)) ==
519 // | | zext(sext(A)) op zext(sext(B))
520 if (BO->getOpcode() == Instruction::Add && !ZeroExtended && NonNegative) {
521 // If a + b >= 0 and (a >= 0 or b >= 0), then
522 // sext(a + b) = sext(a) + sext(b)
523 // even if the addition is not marked nsw.
525 // Leveraging this invarient, we can trace into an sext'ed inbound GEP
526 // index if the constant offset is non-negative.
528 // Verified in @sext_add in split-gep.ll.
529 if (ConstantInt *ConstLHS = dyn_cast<ConstantInt>(LHS)) {
530 if (!ConstLHS->isNegative())
533 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
534 if (!ConstRHS->isNegative())
539 // sext (add/sub nsw A, B) == add/sub nsw (sext A), (sext B)
540 // zext (add/sub nuw A, B) == add/sub nuw (zext A), (zext B)
541 if (BO->getOpcode() == Instruction::Add ||
542 BO->getOpcode() == Instruction::Sub) {
543 if (SignExtended && !BO->hasNoSignedWrap())
545 if (ZeroExtended && !BO->hasNoUnsignedWrap())
552 APInt ConstantOffsetExtractor::findInEitherOperand(BinaryOperator *BO,
555 // BO being non-negative does not shed light on whether its operands are
556 // non-negative. Clear the NonNegative flag here.
557 APInt ConstantOffset = find(BO->getOperand(0), SignExtended, ZeroExtended,
558 /* NonNegative */ false);
559 // If we found a constant offset in the left operand, stop and return that.
560 // This shortcut might cause us to miss opportunities of combining the
561 // constant offsets in both operands, e.g., (a + 4) + (b + 5) => (a + b) + 9.
562 // However, such cases are probably already handled by -instcombine,
563 // given this pass runs after the standard optimizations.
564 if (ConstantOffset != 0) return ConstantOffset;
565 ConstantOffset = find(BO->getOperand(1), SignExtended, ZeroExtended,
566 /* NonNegative */ false);
567 // If U is a sub operator, negate the constant offset found in the right
569 if (BO->getOpcode() == Instruction::Sub)
570 ConstantOffset = -ConstantOffset;
571 return ConstantOffset;
574 APInt ConstantOffsetExtractor::find(Value *V, bool SignExtended,
575 bool ZeroExtended, bool NonNegative) {
576 // TODO(jingyue): We could trace into integer/pointer casts, such as
577 // inttoptr, ptrtoint, bitcast, and addrspacecast. We choose to handle only
578 // integers because it gives good enough results for our benchmarks.
579 unsigned BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
581 // We cannot do much with Values that are not a User, such as an Argument.
582 User *U = dyn_cast<User>(V);
583 if (U == nullptr) return APInt(BitWidth, 0);
585 APInt ConstantOffset(BitWidth, 0);
586 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
587 // Hooray, we found it!
588 ConstantOffset = CI->getValue();
589 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) {
590 // Trace into subexpressions for more hoisting opportunities.
591 if (CanTraceInto(SignExtended, ZeroExtended, BO, NonNegative))
592 ConstantOffset = findInEitherOperand(BO, SignExtended, ZeroExtended);
593 } else if (isa<SExtInst>(V)) {
594 ConstantOffset = find(U->getOperand(0), /* SignExtended */ true,
595 ZeroExtended, NonNegative).sext(BitWidth);
596 } else if (isa<ZExtInst>(V)) {
597 // As an optimization, we can clear the SignExtended flag because
598 // sext(zext(a)) = zext(a). Verified in @sext_zext in split-gep.ll.
600 // Clear the NonNegative flag, because zext(a) >= 0 does not imply a >= 0.
602 find(U->getOperand(0), /* SignExtended */ false,
603 /* ZeroExtended */ true, /* NonNegative */ false).zext(BitWidth);
606 // If we found a non-zero constant offset, add it to the path for
607 // rebuildWithoutConstOffset. Zero is a valid constant offset, but doesn't
608 // help this optimization.
609 if (ConstantOffset != 0)
610 UserChain.push_back(U);
611 return ConstantOffset;
614 Value *ConstantOffsetExtractor::applyExts(Value *V) {
616 // ExtInsts is built in the use-def order. Therefore, we apply them to V
617 // in the reversed order.
618 for (auto I = ExtInsts.rbegin(), E = ExtInsts.rend(); I != E; ++I) {
619 if (Constant *C = dyn_cast<Constant>(Current)) {
620 // If Current is a constant, apply s/zext using ConstantExpr::getCast.
621 // ConstantExpr::getCast emits a ConstantInt if C is a ConstantInt.
622 Current = ConstantExpr::getCast((*I)->getOpcode(), C, (*I)->getType());
624 Instruction *Ext = (*I)->clone();
625 Ext->setOperand(0, Current);
626 Ext->insertBefore(IP);
633 Value *ConstantOffsetExtractor::rebuildWithoutConstOffset() {
634 distributeExtsAndCloneChain(UserChain.size() - 1);
635 // Remove all nullptrs (used to be s/zext) from UserChain.
636 unsigned NewSize = 0;
637 for (User *I : UserChain) {
639 UserChain[NewSize] = I;
643 UserChain.resize(NewSize);
644 return removeConstOffset(UserChain.size() - 1);
648 ConstantOffsetExtractor::distributeExtsAndCloneChain(unsigned ChainIndex) {
649 User *U = UserChain[ChainIndex];
650 if (ChainIndex == 0) {
651 assert(isa<ConstantInt>(U));
652 // If U is a ConstantInt, applyExts will return a ConstantInt as well.
653 return UserChain[ChainIndex] = cast<ConstantInt>(applyExts(U));
656 if (CastInst *Cast = dyn_cast<CastInst>(U)) {
657 assert((isa<SExtInst>(Cast) || isa<ZExtInst>(Cast)) &&
658 "We only traced into two types of CastInst: sext and zext");
659 ExtInsts.push_back(Cast);
660 UserChain[ChainIndex] = nullptr;
661 return distributeExtsAndCloneChain(ChainIndex - 1);
664 // Function find only trace into BinaryOperator and CastInst.
665 BinaryOperator *BO = cast<BinaryOperator>(U);
666 // OpNo = which operand of BO is UserChain[ChainIndex - 1]
667 unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
668 Value *TheOther = applyExts(BO->getOperand(1 - OpNo));
669 Value *NextInChain = distributeExtsAndCloneChain(ChainIndex - 1);
671 BinaryOperator *NewBO = nullptr;
673 NewBO = BinaryOperator::Create(BO->getOpcode(), NextInChain, TheOther,
676 NewBO = BinaryOperator::Create(BO->getOpcode(), TheOther, NextInChain,
679 return UserChain[ChainIndex] = NewBO;
682 Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
683 if (ChainIndex == 0) {
684 assert(isa<ConstantInt>(UserChain[ChainIndex]));
685 return ConstantInt::getNullValue(UserChain[ChainIndex]->getType());
688 BinaryOperator *BO = cast<BinaryOperator>(UserChain[ChainIndex]);
689 assert(BO->getNumUses() <= 1 &&
690 "distributeExtsAndCloneChain clones each BinaryOperator in "
691 "UserChain, so no one should be used more than "
694 unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
695 assert(BO->getOperand(OpNo) == UserChain[ChainIndex - 1]);
696 Value *NextInChain = removeConstOffset(ChainIndex - 1);
697 Value *TheOther = BO->getOperand(1 - OpNo);
699 // If NextInChain is 0 and not the LHS of a sub, we can simplify the
700 // sub-expression to be just TheOther.
701 if (ConstantInt *CI = dyn_cast<ConstantInt>(NextInChain)) {
702 if (CI->isZero() && !(BO->getOpcode() == Instruction::Sub && OpNo == 0))
706 BinaryOperator::BinaryOps NewOp = BO->getOpcode();
707 if (BO->getOpcode() == Instruction::Or) {
708 // Rebuild "or" as "add", because "or" may be invalid for the new
711 // For instance, given
712 // a | (b + 5) where a and b + 5 have no common bits,
713 // we can extract 5 as the constant offset.
715 // However, reusing the "or" in the new index would give us
717 // which does not equal a | (b + 5).
719 // Replacing the "or" with "add" is fine, because
720 // a | (b + 5) = a + (b + 5) = (a + b) + 5
721 NewOp = Instruction::Add;
724 BinaryOperator *NewBO;
726 NewBO = BinaryOperator::Create(NewOp, NextInChain, TheOther, "", IP);
728 NewBO = BinaryOperator::Create(NewOp, TheOther, NextInChain, "", IP);
734 Value *ConstantOffsetExtractor::Extract(Value *Idx, GetElementPtrInst *GEP,
735 User *&UserChainTail,
736 const DominatorTree *DT) {
737 ConstantOffsetExtractor Extractor(GEP, DT);
738 // Find a non-zero constant offset first.
739 APInt ConstantOffset =
740 Extractor.find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
742 if (ConstantOffset == 0) {
743 UserChainTail = nullptr;
746 // Separates the constant offset from the GEP index.
747 Value *IdxWithoutConstOffset = Extractor.rebuildWithoutConstOffset();
748 UserChainTail = Extractor.UserChain.back();
749 return IdxWithoutConstOffset;
752 int64_t ConstantOffsetExtractor::Find(Value *Idx, GetElementPtrInst *GEP,
753 const DominatorTree *DT) {
754 // If Idx is an index of an inbound GEP, Idx is guaranteed to be non-negative.
755 return ConstantOffsetExtractor(GEP, DT)
756 .find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
761 bool SeparateConstOffsetFromGEP::canonicalizeArrayIndicesToPointerSize(
762 GetElementPtrInst *GEP) {
763 bool Changed = false;
764 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
765 gep_type_iterator GTI = gep_type_begin(*GEP);
766 for (User::op_iterator I = GEP->op_begin() + 1, E = GEP->op_end();
767 I != E; ++I, ++GTI) {
768 // Skip struct member indices which must be i32.
769 if (GTI.isSequential()) {
770 if ((*I)->getType() != IntPtrTy) {
771 *I = CastInst::CreateIntegerCast(*I, IntPtrTy, true, "idxprom", GEP);
780 SeparateConstOffsetFromGEP::accumulateByteOffset(GetElementPtrInst *GEP,
781 bool &NeedsExtraction) {
782 NeedsExtraction = false;
783 int64_t AccumulativeByteOffset = 0;
784 gep_type_iterator GTI = gep_type_begin(*GEP);
785 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
786 if (GTI.isSequential()) {
787 // Tries to extract a constant offset from this GEP index.
788 int64_t ConstantOffset =
789 ConstantOffsetExtractor::Find(GEP->getOperand(I), GEP, DT);
790 if (ConstantOffset != 0) {
791 NeedsExtraction = true;
792 // A GEP may have multiple indices. We accumulate the extracted
793 // constant offset to a byte offset, and later offset the remainder of
794 // the original GEP with this byte offset.
795 AccumulativeByteOffset +=
796 ConstantOffset * DL->getTypeAllocSize(GTI.getIndexedType());
798 } else if (LowerGEP) {
799 StructType *StTy = GTI.getStructType();
800 uint64_t Field = cast<ConstantInt>(GEP->getOperand(I))->getZExtValue();
801 // Skip field 0 as the offset is always 0.
803 NeedsExtraction = true;
804 AccumulativeByteOffset +=
805 DL->getStructLayout(StTy)->getElementOffset(Field);
809 return AccumulativeByteOffset;
812 void SeparateConstOffsetFromGEP::lowerToSingleIndexGEPs(
813 GetElementPtrInst *Variadic, int64_t AccumulativeByteOffset) {
814 IRBuilder<> Builder(Variadic);
815 Type *IntPtrTy = DL->getIntPtrType(Variadic->getType());
818 Builder.getInt8PtrTy(Variadic->getType()->getPointerAddressSpace());
819 Value *ResultPtr = Variadic->getOperand(0);
820 Loop *L = LI->getLoopFor(Variadic->getParent());
821 // Check if the base is not loop invariant or used more than once.
822 bool isSwapCandidate =
823 L && L->isLoopInvariant(ResultPtr) &&
824 !hasMoreThanOneUseInLoop(ResultPtr, L);
825 Value *FirstResult = nullptr;
827 if (ResultPtr->getType() != I8PtrTy)
828 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
830 gep_type_iterator GTI = gep_type_begin(*Variadic);
831 // Create an ugly GEP for each sequential index. We don't create GEPs for
832 // structure indices, as they are accumulated in the constant offset index.
833 for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
834 if (GTI.isSequential()) {
835 Value *Idx = Variadic->getOperand(I);
836 // Skip zero indices.
837 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
841 APInt ElementSize = APInt(IntPtrTy->getIntegerBitWidth(),
842 DL->getTypeAllocSize(GTI.getIndexedType()));
843 // Scale the index by element size.
844 if (ElementSize != 1) {
845 if (ElementSize.isPowerOf2()) {
846 Idx = Builder.CreateShl(
847 Idx, ConstantInt::get(IntPtrTy, ElementSize.logBase2()));
849 Idx = Builder.CreateMul(Idx, ConstantInt::get(IntPtrTy, ElementSize));
852 // Create an ugly GEP with a single index for each index.
854 Builder.CreateGEP(Builder.getInt8Ty(), ResultPtr, Idx, "uglygep");
855 if (FirstResult == nullptr)
856 FirstResult = ResultPtr;
860 // Create a GEP with the constant offset index.
861 if (AccumulativeByteOffset != 0) {
862 Value *Offset = ConstantInt::get(IntPtrTy, AccumulativeByteOffset);
864 Builder.CreateGEP(Builder.getInt8Ty(), ResultPtr, Offset, "uglygep");
866 isSwapCandidate = false;
868 // If we created a GEP with constant index, and the base is loop invariant,
869 // then we swap the first one with it, so LICM can move constant GEP out
871 GetElementPtrInst *FirstGEP = dyn_cast_or_null<GetElementPtrInst>(FirstResult);
872 GetElementPtrInst *SecondGEP = dyn_cast_or_null<GetElementPtrInst>(ResultPtr);
873 if (isSwapCandidate && isLegalToSwapOperand(FirstGEP, SecondGEP, L))
874 swapGEPOperand(FirstGEP, SecondGEP);
876 if (ResultPtr->getType() != Variadic->getType())
877 ResultPtr = Builder.CreateBitCast(ResultPtr, Variadic->getType());
879 Variadic->replaceAllUsesWith(ResultPtr);
880 Variadic->eraseFromParent();
884 SeparateConstOffsetFromGEP::lowerToArithmetics(GetElementPtrInst *Variadic,
885 int64_t AccumulativeByteOffset) {
886 IRBuilder<> Builder(Variadic);
887 Type *IntPtrTy = DL->getIntPtrType(Variadic->getType());
889 Value *ResultPtr = Builder.CreatePtrToInt(Variadic->getOperand(0), IntPtrTy);
890 gep_type_iterator GTI = gep_type_begin(*Variadic);
891 // Create ADD/SHL/MUL arithmetic operations for each sequential indices. We
892 // don't create arithmetics for structure indices, as they are accumulated
893 // in the constant offset index.
894 for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
895 if (GTI.isSequential()) {
896 Value *Idx = Variadic->getOperand(I);
897 // Skip zero indices.
898 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
902 APInt ElementSize = APInt(IntPtrTy->getIntegerBitWidth(),
903 DL->getTypeAllocSize(GTI.getIndexedType()));
904 // Scale the index by element size.
905 if (ElementSize != 1) {
906 if (ElementSize.isPowerOf2()) {
907 Idx = Builder.CreateShl(
908 Idx, ConstantInt::get(IntPtrTy, ElementSize.logBase2()));
910 Idx = Builder.CreateMul(Idx, ConstantInt::get(IntPtrTy, ElementSize));
913 // Create an ADD for each index.
914 ResultPtr = Builder.CreateAdd(ResultPtr, Idx);
918 // Create an ADD for the constant offset index.
919 if (AccumulativeByteOffset != 0) {
920 ResultPtr = Builder.CreateAdd(
921 ResultPtr, ConstantInt::get(IntPtrTy, AccumulativeByteOffset));
924 ResultPtr = Builder.CreateIntToPtr(ResultPtr, Variadic->getType());
925 Variadic->replaceAllUsesWith(ResultPtr);
926 Variadic->eraseFromParent();
929 bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
931 if (GEP->getType()->isVectorTy())
934 // The backend can already nicely handle the case where all indices are
936 if (GEP->hasAllConstantIndices())
939 bool Changed = canonicalizeArrayIndicesToPointerSize(GEP);
941 bool NeedsExtraction;
942 int64_t AccumulativeByteOffset = accumulateByteOffset(GEP, NeedsExtraction);
944 if (!NeedsExtraction)
946 // If LowerGEP is disabled, before really splitting the GEP, check whether the
947 // backend supports the addressing mode we are about to produce. If no, this
948 // splitting probably won't be beneficial.
949 // If LowerGEP is enabled, even the extracted constant offset can not match
950 // the addressing mode, we can still do optimizations to other lowered parts
951 // of variable indices. Therefore, we don't check for addressing modes in that
954 TargetTransformInfo &TTI =
955 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
956 *GEP->getParent()->getParent());
957 unsigned AddrSpace = GEP->getPointerAddressSpace();
958 if (!TTI.isLegalAddressingMode(GEP->getResultElementType(),
959 /*BaseGV=*/nullptr, AccumulativeByteOffset,
960 /*HasBaseReg=*/true, /*Scale=*/0,
966 // Remove the constant offset in each sequential index. The resultant GEP
967 // computes the variadic base.
968 // Notice that we don't remove struct field indices here. If LowerGEP is
969 // disabled, a structure index is not accumulated and we still use the old
970 // one. If LowerGEP is enabled, a structure index is accumulated in the
971 // constant offset. LowerToSingleIndexGEPs or lowerToArithmetics will later
972 // handle the constant offset and won't need a new structure index.
973 gep_type_iterator GTI = gep_type_begin(*GEP);
974 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
975 if (GTI.isSequential()) {
976 // Splits this GEP index into a variadic part and a constant offset, and
977 // uses the variadic part as the new index.
978 Value *OldIdx = GEP->getOperand(I);
981 ConstantOffsetExtractor::Extract(OldIdx, GEP, UserChainTail, DT);
982 if (NewIdx != nullptr) {
983 // Switches to the index with the constant offset removed.
984 GEP->setOperand(I, NewIdx);
985 // After switching to the new index, we can garbage-collect UserChain
986 // and the old index if they are not used.
987 RecursivelyDeleteTriviallyDeadInstructions(UserChainTail);
988 RecursivelyDeleteTriviallyDeadInstructions(OldIdx);
993 // Clear the inbounds attribute because the new index may be off-bound.
997 // addr = gep inbounds float, float* p, i64 b
999 // is transformed to:
1001 // addr2 = gep float, float* p, i64 a ; inbounds removed
1002 // addr = gep inbounds float, float* addr2, i64 5
1004 // If a is -4, although the old index b is in bounds, the new index a is
1005 // off-bound. http://llvm.org/docs/LangRef.html#id181 says "if the
1006 // inbounds keyword is not present, the offsets are added to the base
1007 // address with silently-wrapping two's complement arithmetic".
1008 // Therefore, the final code will be a semantically equivalent.
1010 // TODO(jingyue): do some range analysis to keep as many inbounds as
1011 // possible. GEPs with inbounds are more friendly to alias analysis.
1012 bool GEPWasInBounds = GEP->isInBounds();
1013 GEP->setIsInBounds(false);
1015 // Lowers a GEP to either GEPs with a single index or arithmetic operations.
1017 // As currently BasicAA does not analyze ptrtoint/inttoptr, do not lower to
1018 // arithmetic operations if the target uses alias analysis in codegen.
1019 if (TM && TM->getSubtargetImpl(*GEP->getParent()->getParent())->useAA())
1020 lowerToSingleIndexGEPs(GEP, AccumulativeByteOffset);
1022 lowerToArithmetics(GEP, AccumulativeByteOffset);
1026 // No need to create another GEP if the accumulative byte offset is 0.
1027 if (AccumulativeByteOffset == 0)
1030 // Offsets the base with the accumulative byte offset.
1035 // => add the offset
1037 // %gep2 ; clone of %gep
1038 // %new.gep = gep %gep2, <offset / sizeof(*%gep)>
1039 // %gep ; will be removed
1042 // => replace all uses of %gep with %new.gep and remove %gep
1044 // %gep2 ; clone of %gep
1045 // %new.gep = gep %gep2, <offset / sizeof(*%gep)>
1048 // If AccumulativeByteOffset is not a multiple of sizeof(*%gep), we emit an
1049 // uglygep (http://llvm.org/docs/GetElementPtr.html#what-s-an-uglygep):
1050 // bitcast %gep2 to i8*, add the offset, and bitcast the result back to the
1053 // %gep2 ; clone of %gep
1054 // %0 = bitcast %gep2 to i8*
1055 // %uglygep = gep %0, <offset>
1056 // %new.gep = bitcast %uglygep to <type of %gep>
1058 Instruction *NewGEP = GEP->clone();
1059 NewGEP->insertBefore(GEP);
1061 // Per ANSI C standard, signed / unsigned = unsigned and signed % unsigned =
1062 // unsigned.. Therefore, we cast ElementTypeSizeOfGEP to signed because it is
1063 // used with unsigned integers later.
1064 int64_t ElementTypeSizeOfGEP = static_cast<int64_t>(
1065 DL->getTypeAllocSize(GEP->getResultElementType()));
1066 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
1067 if (AccumulativeByteOffset % ElementTypeSizeOfGEP == 0) {
1068 // Very likely. As long as %gep is natually aligned, the byte offset we
1069 // extracted should be a multiple of sizeof(*%gep).
1070 int64_t Index = AccumulativeByteOffset / ElementTypeSizeOfGEP;
1071 NewGEP = GetElementPtrInst::Create(GEP->getResultElementType(), NewGEP,
1072 ConstantInt::get(IntPtrTy, Index, true),
1073 GEP->getName(), GEP);
1074 // Inherit the inbounds attribute of the original GEP.
1075 cast<GetElementPtrInst>(NewGEP)->setIsInBounds(GEPWasInBounds);
1077 // Unlikely but possible. For example,
1085 // Suppose the gep before extraction is &s[i + 1].b[j + 3]. After
1086 // extraction, it becomes &s[i].b[j] and AccumulativeByteOffset is
1087 // sizeof(S) + 3 * sizeof(int64) = 100, which is not a multiple of
1090 // Emit an uglygep in this case.
1091 Type *I8PtrTy = Type::getInt8PtrTy(GEP->getContext(),
1092 GEP->getPointerAddressSpace());
1093 NewGEP = new BitCastInst(NewGEP, I8PtrTy, "", GEP);
1094 NewGEP = GetElementPtrInst::Create(
1095 Type::getInt8Ty(GEP->getContext()), NewGEP,
1096 ConstantInt::get(IntPtrTy, AccumulativeByteOffset, true), "uglygep",
1098 // Inherit the inbounds attribute of the original GEP.
1099 cast<GetElementPtrInst>(NewGEP)->setIsInBounds(GEPWasInBounds);
1100 if (GEP->getType() != I8PtrTy)
1101 NewGEP = new BitCastInst(NewGEP, GEP->getType(), GEP->getName(), GEP);
1104 GEP->replaceAllUsesWith(NewGEP);
1105 GEP->eraseFromParent();
1110 bool SeparateConstOffsetFromGEP::runOnFunction(Function &F) {
1111 if (skipFunction(F))
1114 if (DisableSeparateConstOffsetFromGEP)
1117 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1118 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1119 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1120 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1121 bool Changed = false;
1122 for (BasicBlock &B : F) {
1123 for (BasicBlock::iterator I = B.begin(), IE = B.end(); I != IE;)
1124 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I++))
1125 Changed |= splitGEP(GEP);
1126 // No need to split GEP ConstantExprs because all its indices are constant
1130 Changed |= reuniteExts(F);
1132 if (VerifyNoDeadCode)
1133 verifyNoDeadCode(F);
1138 Instruction *SeparateConstOffsetFromGEP::findClosestMatchingDominator(
1139 const SCEV *Key, Instruction *Dominatee) {
1140 auto Pos = DominatingExprs.find(Key);
1141 if (Pos == DominatingExprs.end())
1144 auto &Candidates = Pos->second;
1145 // Because we process the basic blocks in pre-order of the dominator tree, a
1146 // candidate that doesn't dominate the current instruction won't dominate any
1147 // future instruction either. Therefore, we pop it out of the stack. This
1148 // optimization makes the algorithm O(n).
1149 while (!Candidates.empty()) {
1150 Instruction *Candidate = Candidates.back();
1151 if (DT->dominates(Candidate, Dominatee))
1153 Candidates.pop_back();
1158 bool SeparateConstOffsetFromGEP::reuniteExts(Instruction *I) {
1159 if (!SE->isSCEVable(I->getType()))
1163 // I: sext(LHS)+sext(RHS)
1164 // If Dom can't sign overflow and Dom dominates I, optimize I to sext(Dom).
1165 // TODO: handle zext
1166 Value *LHS = nullptr, *RHS = nullptr;
1167 if (match(I, m_Add(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS)))) ||
1168 match(I, m_Sub(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
1169 if (LHS->getType() == RHS->getType()) {
1171 SE->getAddExpr(SE->getUnknown(LHS), SE->getUnknown(RHS));
1172 if (auto *Dom = findClosestMatchingDominator(Key, I)) {
1173 Instruction *NewSExt = new SExtInst(Dom, I->getType(), "", I);
1174 NewSExt->takeName(I);
1175 I->replaceAllUsesWith(NewSExt);
1176 RecursivelyDeleteTriviallyDeadInstructions(I);
1182 // Add I to DominatingExprs if it's an add/sub that can't sign overflow.
1183 if (match(I, m_NSWAdd(m_Value(LHS), m_Value(RHS))) ||
1184 match(I, m_NSWSub(m_Value(LHS), m_Value(RHS)))) {
1185 if (programUndefinedIfFullPoison(I)) {
1187 SE->getAddExpr(SE->getUnknown(LHS), SE->getUnknown(RHS));
1188 DominatingExprs[Key].push_back(I);
1194 bool SeparateConstOffsetFromGEP::reuniteExts(Function &F) {
1195 bool Changed = false;
1196 DominatingExprs.clear();
1197 for (const auto Node : depth_first(DT)) {
1198 BasicBlock *BB = Node->getBlock();
1199 for (auto I = BB->begin(); I != BB->end(); ) {
1200 Instruction *Cur = &*I++;
1201 Changed |= reuniteExts(Cur);
1207 void SeparateConstOffsetFromGEP::verifyNoDeadCode(Function &F) {
1208 for (BasicBlock &B : F) {
1209 for (Instruction &I : B) {
1210 if (isInstructionTriviallyDead(&I)) {
1211 std::string ErrMessage;
1212 raw_string_ostream RSO(ErrMessage);
1213 RSO << "Dead instruction detected!\n" << I << "\n";
1214 llvm_unreachable(RSO.str().c_str());
1220 bool SeparateConstOffsetFromGEP::isLegalToSwapOperand(
1221 GetElementPtrInst *FirstGEP, GetElementPtrInst *SecondGEP, Loop *CurLoop) {
1222 if (!FirstGEP || !FirstGEP->hasOneUse())
1225 if (!SecondGEP || FirstGEP->getParent() != SecondGEP->getParent())
1228 if (FirstGEP == SecondGEP)
1231 unsigned FirstNum = FirstGEP->getNumOperands();
1232 unsigned SecondNum = SecondGEP->getNumOperands();
1233 // Give up if the number of operands are not 2.
1234 if (FirstNum != SecondNum || FirstNum != 2)
1237 Value *FirstBase = FirstGEP->getOperand(0);
1238 Value *SecondBase = SecondGEP->getOperand(0);
1239 Value *FirstOffset = FirstGEP->getOperand(1);
1240 // Give up if the index of the first GEP is loop invariant.
1241 if (CurLoop->isLoopInvariant(FirstOffset))
1244 // Give up if base doesn't have same type.
1245 if (FirstBase->getType() != SecondBase->getType())
1248 Instruction *FirstOffsetDef = dyn_cast<Instruction>(FirstOffset);
1250 // Check if the second operand of first GEP has constant coefficient.
1251 // For an example, for the following code, we won't gain anything by
1252 // hoisting the second GEP out because the second GEP can be folded away.
1253 // %scevgep.sum.ur159 = add i64 %idxprom48.ur, 256
1254 // %67 = shl i64 %scevgep.sum.ur159, 2
1255 // %uglygep160 = getelementptr i8* %65, i64 %67
1256 // %uglygep161 = getelementptr i8* %uglygep160, i64 -1024
1258 // Skip constant shift instruction which may be generated by Splitting GEPs.
1259 if (FirstOffsetDef && FirstOffsetDef->isShift() &&
1260 isa<ConstantInt>(FirstOffsetDef->getOperand(1)))
1261 FirstOffsetDef = dyn_cast<Instruction>(FirstOffsetDef->getOperand(0));
1263 // Give up if FirstOffsetDef is an Add or Sub with constant.
1264 // Because it may not profitable at all due to constant folding.
1266 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FirstOffsetDef)) {
1267 unsigned opc = BO->getOpcode();
1268 if ((opc == Instruction::Add || opc == Instruction::Sub) &&
1269 (isa<ConstantInt>(BO->getOperand(0)) ||
1270 isa<ConstantInt>(BO->getOperand(1))))
1276 bool SeparateConstOffsetFromGEP::hasMoreThanOneUseInLoop(Value *V, Loop *L) {
1278 for (User *U : V->users()) {
1279 if (Instruction *User = dyn_cast<Instruction>(U))
1280 if (L->contains(User))
1281 if (++UsesInLoop > 1)
1287 void SeparateConstOffsetFromGEP::swapGEPOperand(GetElementPtrInst *First,
1288 GetElementPtrInst *Second) {
1289 Value *Offset1 = First->getOperand(1);
1290 Value *Offset2 = Second->getOperand(1);
1291 First->setOperand(1, Offset2);
1292 Second->setOperand(1, Offset1);
1294 // We changed p+o+c to p+c+o, p+c may not be inbound anymore.
1295 const DataLayout &DAL = First->getModule()->getDataLayout();
1296 APInt Offset(DAL.getPointerSizeInBits(
1297 cast<PointerType>(First->getType())->getAddressSpace()),
1300 First->stripAndAccumulateInBoundsConstantOffsets(DAL, Offset);
1301 uint64_t ObjectSize;
1302 if (!getObjectSize(NewBase, ObjectSize, DAL, TLI) ||
1303 Offset.ugt(ObjectSize)) {
1304 First->setIsInBounds(false);
1305 Second->setIsInBounds(false);
1307 First->setIsInBounds(true);