1 //===-- SeparateConstOffsetFromGEP.cpp - ------------------------*- C++ -*-===//
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 for 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/Analysis/ScalarEvolution.h"
160 #include "llvm/Analysis/TargetTransformInfo.h"
161 #include "llvm/Analysis/ValueTracking.h"
162 #include "llvm/IR/Constants.h"
163 #include "llvm/IR/DataLayout.h"
164 #include "llvm/IR/Dominators.h"
165 #include "llvm/IR/Instructions.h"
166 #include "llvm/IR/LLVMContext.h"
167 #include "llvm/IR/Module.h"
168 #include "llvm/IR/PatternMatch.h"
169 #include "llvm/IR/Operator.h"
170 #include "llvm/Support/CommandLine.h"
171 #include "llvm/Support/raw_ostream.h"
172 #include "llvm/Transforms/Scalar.h"
173 #include "llvm/Transforms/Utils/Local.h"
174 #include "llvm/Target/TargetMachine.h"
175 #include "llvm/Target/TargetSubtargetInfo.h"
176 #include "llvm/IR/IRBuilder.h"
178 using namespace llvm;
179 using namespace llvm::PatternMatch;
181 static cl::opt<bool> DisableSeparateConstOffsetFromGEP(
182 "disable-separate-const-offset-from-gep", cl::init(false),
183 cl::desc("Do not separate the constant offset from a GEP instruction"),
185 // Setting this flag may emit false positives when the input module already
186 // contains dead instructions. Therefore, we set it only in unit tests that are
187 // free of dead code.
189 VerifyNoDeadCode("reassociate-geps-verify-no-dead-code", cl::init(false),
190 cl::desc("Verify this pass produces no dead code"),
195 /// \brief A helper class for separating a constant offset from a GEP index.
197 /// In real programs, a GEP index may be more complicated than a simple addition
198 /// of something and a constant integer which can be trivially splitted. For
199 /// example, to split ((a << 3) | 5) + b, we need to search deeper for the
200 /// constant offset, so that we can separate the index to (a << 3) + b and 5.
202 /// Therefore, this class looks into the expression that computes a given GEP
203 /// index, and tries to find a constant integer that can be hoisted to the
204 /// outermost level of the expression as an addition. Not every constant in an
205 /// expression can jump out. e.g., we cannot transform (b * (a + 5)) to (b * a +
206 /// 5); nor can we transform (3 * (a + 5)) to (3 * a + 5), however in this case,
207 /// -instcombine probably already optimized (3 * (a + 5)) to (3 * a + 15).
208 class ConstantOffsetExtractor {
210 /// Extracts a constant offset from the given GEP index. It returns the
211 /// new index representing the remainder (equal to the original index minus
212 /// the constant offset), or nullptr if we cannot extract a constant offset.
213 /// \p Idx The given GEP index
214 /// \p GEP The given GEP
215 /// \p UserChainTail Outputs the tail of UserChain so that we can
216 /// garbage-collect unused instructions in UserChain.
217 static Value *Extract(Value *Idx, GetElementPtrInst *GEP,
218 User *&UserChainTail, const DominatorTree *DT);
219 /// Looks for a constant offset from the given GEP index without extracting
220 /// it. It returns the numeric value of the extracted constant offset (0 if
221 /// failed). The meaning of the arguments are the same as Extract.
222 static int64_t Find(Value *Idx, GetElementPtrInst *GEP,
223 const DominatorTree *DT);
226 ConstantOffsetExtractor(Instruction *InsertionPt, const DominatorTree *DT)
227 : IP(InsertionPt), DL(InsertionPt->getModule()->getDataLayout()), DT(DT) {
229 /// Searches the expression that computes V for a non-zero constant C s.t.
230 /// V can be reassociated into the form V' + C. If the searching is
231 /// successful, returns C and update UserChain as a def-use chain from C to V;
232 /// otherwise, UserChain is empty.
234 /// \p V The given expression
235 /// \p SignExtended Whether V will be sign-extended in the computation of the
237 /// \p ZeroExtended Whether V will be zero-extended in the computation of the
239 /// \p NonNegative Whether V is guaranteed to be non-negative. For example,
240 /// an index of an inbounds GEP is guaranteed to be
241 /// non-negative. Levaraging this, we can better split
243 APInt find(Value *V, bool SignExtended, bool ZeroExtended, bool NonNegative);
244 /// A helper function to look into both operands of a binary operator.
245 APInt findInEitherOperand(BinaryOperator *BO, bool SignExtended,
247 /// After finding the constant offset C from the GEP index I, we build a new
248 /// index I' s.t. I' + C = I. This function builds and returns the new
249 /// index I' according to UserChain produced by function "find".
251 /// The building conceptually takes two steps:
252 /// 1) iteratively distribute s/zext towards the leaves of the expression tree
254 /// 2) reassociate the expression tree to the form I' + C.
256 /// For example, to extract the 5 from sext(a + (b + 5)), we first distribute
257 /// sext to a, b and 5 so that we have
258 /// sext(a) + (sext(b) + 5).
259 /// Then, we reassociate it to
260 /// (sext(a) + sext(b)) + 5.
261 /// Given this form, we know I' is sext(a) + sext(b).
262 Value *rebuildWithoutConstOffset();
263 /// After the first step of rebuilding the GEP index without the constant
264 /// offset, distribute s/zext to the operands of all operators in UserChain.
265 /// e.g., zext(sext(a + (b + 5)) (assuming no overflow) =>
266 /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))).
268 /// The function also updates UserChain to point to new subexpressions after
269 /// distributing s/zext. e.g., the old UserChain of the above example is
270 /// 5 -> b + 5 -> a + (b + 5) -> sext(...) -> zext(sext(...)),
271 /// and the new UserChain is
272 /// zext(sext(5)) -> zext(sext(b)) + zext(sext(5)) ->
273 /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))
275 /// \p ChainIndex The index to UserChain. ChainIndex is initially
276 /// UserChain.size() - 1, and is decremented during
278 Value *distributeExtsAndCloneChain(unsigned ChainIndex);
279 /// Reassociates the GEP index to the form I' + C and returns I'.
280 Value *removeConstOffset(unsigned ChainIndex);
281 /// A helper function to apply ExtInsts, a list of s/zext, to value V.
282 /// e.g., if ExtInsts = [sext i32 to i64, zext i16 to i32], this function
283 /// returns "sext i32 (zext i16 V to i32) to i64".
284 Value *applyExts(Value *V);
286 /// A helper function that returns whether we can trace into the operands
287 /// of binary operator BO for a constant offset.
289 /// \p SignExtended Whether BO is surrounded by sext
290 /// \p ZeroExtended Whether BO is surrounded by zext
291 /// \p NonNegative Whether BO is known to be non-negative, e.g., an in-bound
293 bool CanTraceInto(bool SignExtended, bool ZeroExtended, BinaryOperator *BO,
296 /// The path from the constant offset to the old GEP index. e.g., if the GEP
297 /// index is "a * b + (c + 5)". After running function find, UserChain[0] will
298 /// be the constant 5, UserChain[1] will be the subexpression "c + 5", and
299 /// UserChain[2] will be the entire expression "a * b + (c + 5)".
301 /// This path helps to rebuild the new GEP index.
302 SmallVector<User *, 8> UserChain;
303 /// A data structure used in rebuildWithoutConstOffset. Contains all
304 /// sext/zext instructions along UserChain.
305 SmallVector<CastInst *, 16> ExtInsts;
306 Instruction *IP; /// Insertion position of cloned instructions.
307 const DataLayout &DL;
308 const DominatorTree *DT;
311 /// \brief A pass that tries to split every GEP in the function into a variadic
312 /// base and a constant offset. It is a FunctionPass because searching for the
313 /// constant offset may inspect other basic blocks.
314 class SeparateConstOffsetFromGEP : public FunctionPass {
317 SeparateConstOffsetFromGEP(const TargetMachine *TM = nullptr,
318 bool LowerGEP = false)
319 : FunctionPass(ID), DL(nullptr), DT(nullptr), TM(TM), LowerGEP(LowerGEP) {
320 initializeSeparateConstOffsetFromGEPPass(*PassRegistry::getPassRegistry());
323 void getAnalysisUsage(AnalysisUsage &AU) const override {
324 AU.addRequired<DominatorTreeWrapperPass>();
325 AU.addRequired<ScalarEvolutionWrapperPass>();
326 AU.addRequired<TargetTransformInfoWrapperPass>();
327 AU.setPreservesCFG();
330 bool doInitialization(Module &M) override {
331 DL = &M.getDataLayout();
334 bool runOnFunction(Function &F) override;
337 /// Tries to split the given GEP into a variadic base and a constant offset,
338 /// and returns true if the splitting succeeds.
339 bool splitGEP(GetElementPtrInst *GEP);
340 /// Lower a GEP with multiple indices into multiple GEPs with a single index.
341 /// Function splitGEP already split the original GEP into a variadic part and
342 /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
343 /// variadic part into a set of GEPs with a single index and applies
344 /// AccumulativeByteOffset to it.
345 /// \p Variadic The variadic part of the original GEP.
346 /// \p AccumulativeByteOffset The constant offset.
347 void lowerToSingleIndexGEPs(GetElementPtrInst *Variadic,
348 int64_t AccumulativeByteOffset);
349 /// Lower a GEP with multiple indices into ptrtoint+arithmetics+inttoptr form.
350 /// Function splitGEP already split the original GEP into a variadic part and
351 /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
352 /// variadic part into a set of arithmetic operations and applies
353 /// AccumulativeByteOffset to it.
354 /// \p Variadic The variadic part of the original GEP.
355 /// \p AccumulativeByteOffset The constant offset.
356 void lowerToArithmetics(GetElementPtrInst *Variadic,
357 int64_t AccumulativeByteOffset);
358 /// Finds the constant offset within each index and accumulates them. If
359 /// LowerGEP is true, it finds in indices of both sequential and structure
360 /// types, otherwise it only finds in sequential indices. The output
361 /// NeedsExtraction indicates whether we successfully find a non-zero constant
363 int64_t accumulateByteOffset(GetElementPtrInst *GEP, bool &NeedsExtraction);
364 /// Canonicalize array indices to pointer-size integers. This helps to
365 /// simplify the logic of splitting a GEP. For example, if a + b is a
366 /// pointer-size integer, we have
367 /// gep base, a + b = gep (gep base, a), b
368 /// However, this equality may not hold if the size of a + b is smaller than
369 /// the pointer size, because LLVM conceptually sign-extends GEP indices to
370 /// pointer size before computing the address
371 /// (http://llvm.org/docs/LangRef.html#id181).
373 /// This canonicalization is very likely already done in clang and
374 /// instcombine. Therefore, the program will probably remain the same.
376 /// Returns true if the module changes.
378 /// Verified in @i32_add in split-gep.ll
379 bool canonicalizeArrayIndicesToPointerSize(GetElementPtrInst *GEP);
380 /// Optimize sext(a)+sext(b) to sext(a+b) when a+b can't sign overflow.
381 /// SeparateConstOffsetFromGEP distributes a sext to leaves before extracting
382 /// the constant offset. After extraction, it becomes desirable to reunion the
383 /// distributed sexts. For example,
385 /// &a[sext(i +nsw (j +nsw 5)]
386 /// => distribute &a[sext(i) +nsw (sext(j) +nsw 5)]
387 /// => constant extraction &a[sext(i) + sext(j)] + 5
388 /// => reunion &a[sext(i +nsw j)] + 5
389 bool reuniteExts(Function &F);
390 /// A helper that reunites sexts in an instruction.
391 bool reuniteExts(Instruction *I);
392 /// Find the closest dominator of <Dominatee> that is equivalent to <Key>.
393 Instruction *findClosestMatchingDominator(const SCEV *Key,
394 Instruction *Dominatee);
395 /// Verify F is free of dead code.
396 void verifyNoDeadCode(Function &F);
398 const DataLayout *DL;
401 const TargetMachine *TM;
402 /// Whether to lower a GEP with multiple indices into arithmetic operations or
403 /// multiple GEPs with a single index.
405 DenseMap<const SCEV *, SmallVector<Instruction *, 2>> DominatingExprs;
407 } // anonymous namespace
409 char SeparateConstOffsetFromGEP::ID = 0;
410 INITIALIZE_PASS_BEGIN(
411 SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
412 "Split GEPs to a variadic base and a constant offset for better CSE", false,
414 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
415 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
416 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
418 SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
419 "Split GEPs to a variadic base and a constant offset for better CSE", false,
423 llvm::createSeparateConstOffsetFromGEPPass(const TargetMachine *TM,
425 return new SeparateConstOffsetFromGEP(TM, LowerGEP);
428 bool ConstantOffsetExtractor::CanTraceInto(bool SignExtended,
432 // We only consider ADD, SUB and OR, because a non-zero constant found in
433 // expressions composed of these operations can be easily hoisted as a
434 // constant offset by reassociation.
435 if (BO->getOpcode() != Instruction::Add &&
436 BO->getOpcode() != Instruction::Sub &&
437 BO->getOpcode() != Instruction::Or) {
441 Value *LHS = BO->getOperand(0), *RHS = BO->getOperand(1);
442 // Do not trace into "or" unless it is equivalent to "add". If LHS and RHS
443 // don't have common bits, (LHS | RHS) is equivalent to (LHS + RHS).
444 if (BO->getOpcode() == Instruction::Or &&
445 !haveNoCommonBitsSet(LHS, RHS, DL, nullptr, BO, DT))
448 // In addition, tracing into BO requires that its surrounding s/zext (if
449 // any) is distributable to both operands.
451 // Suppose BO = A op B.
452 // SignExtended | ZeroExtended | Distributable?
453 // --------------+--------------+----------------------------------
454 // 0 | 0 | true because no s/zext exists
455 // 0 | 1 | zext(BO) == zext(A) op zext(B)
456 // 1 | 0 | sext(BO) == sext(A) op sext(B)
457 // 1 | 1 | zext(sext(BO)) ==
458 // | | zext(sext(A)) op zext(sext(B))
459 if (BO->getOpcode() == Instruction::Add && !ZeroExtended && NonNegative) {
460 // If a + b >= 0 and (a >= 0 or b >= 0), then
461 // sext(a + b) = sext(a) + sext(b)
462 // even if the addition is not marked nsw.
464 // Leveraging this invarient, we can trace into an sext'ed inbound GEP
465 // index if the constant offset is non-negative.
467 // Verified in @sext_add in split-gep.ll.
468 if (ConstantInt *ConstLHS = dyn_cast<ConstantInt>(LHS)) {
469 if (!ConstLHS->isNegative())
472 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
473 if (!ConstRHS->isNegative())
478 // sext (add/sub nsw A, B) == add/sub nsw (sext A), (sext B)
479 // zext (add/sub nuw A, B) == add/sub nuw (zext A), (zext B)
480 if (BO->getOpcode() == Instruction::Add ||
481 BO->getOpcode() == Instruction::Sub) {
482 if (SignExtended && !BO->hasNoSignedWrap())
484 if (ZeroExtended && !BO->hasNoUnsignedWrap())
491 APInt ConstantOffsetExtractor::findInEitherOperand(BinaryOperator *BO,
494 // BO being non-negative does not shed light on whether its operands are
495 // non-negative. Clear the NonNegative flag here.
496 APInt ConstantOffset = find(BO->getOperand(0), SignExtended, ZeroExtended,
497 /* NonNegative */ false);
498 // If we found a constant offset in the left operand, stop and return that.
499 // This shortcut might cause us to miss opportunities of combining the
500 // constant offsets in both operands, e.g., (a + 4) + (b + 5) => (a + b) + 9.
501 // However, such cases are probably already handled by -instcombine,
502 // given this pass runs after the standard optimizations.
503 if (ConstantOffset != 0) return ConstantOffset;
504 ConstantOffset = find(BO->getOperand(1), SignExtended, ZeroExtended,
505 /* NonNegative */ false);
506 // If U is a sub operator, negate the constant offset found in the right
508 if (BO->getOpcode() == Instruction::Sub)
509 ConstantOffset = -ConstantOffset;
510 return ConstantOffset;
513 APInt ConstantOffsetExtractor::find(Value *V, bool SignExtended,
514 bool ZeroExtended, bool NonNegative) {
515 // TODO(jingyue): We could trace into integer/pointer casts, such as
516 // inttoptr, ptrtoint, bitcast, and addrspacecast. We choose to handle only
517 // integers because it gives good enough results for our benchmarks.
518 unsigned BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
520 // We cannot do much with Values that are not a User, such as an Argument.
521 User *U = dyn_cast<User>(V);
522 if (U == nullptr) return APInt(BitWidth, 0);
524 APInt ConstantOffset(BitWidth, 0);
525 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
526 // Hooray, we found it!
527 ConstantOffset = CI->getValue();
528 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) {
529 // Trace into subexpressions for more hoisting opportunities.
530 if (CanTraceInto(SignExtended, ZeroExtended, BO, NonNegative))
531 ConstantOffset = findInEitherOperand(BO, SignExtended, ZeroExtended);
532 } else if (isa<SExtInst>(V)) {
533 ConstantOffset = find(U->getOperand(0), /* SignExtended */ true,
534 ZeroExtended, NonNegative).sext(BitWidth);
535 } else if (isa<ZExtInst>(V)) {
536 // As an optimization, we can clear the SignExtended flag because
537 // sext(zext(a)) = zext(a). Verified in @sext_zext in split-gep.ll.
539 // Clear the NonNegative flag, because zext(a) >= 0 does not imply a >= 0.
541 find(U->getOperand(0), /* SignExtended */ false,
542 /* ZeroExtended */ true, /* NonNegative */ false).zext(BitWidth);
545 // If we found a non-zero constant offset, add it to the path for
546 // rebuildWithoutConstOffset. Zero is a valid constant offset, but doesn't
547 // help this optimization.
548 if (ConstantOffset != 0)
549 UserChain.push_back(U);
550 return ConstantOffset;
553 Value *ConstantOffsetExtractor::applyExts(Value *V) {
555 // ExtInsts is built in the use-def order. Therefore, we apply them to V
556 // in the reversed order.
557 for (auto I = ExtInsts.rbegin(), E = ExtInsts.rend(); I != E; ++I) {
558 if (Constant *C = dyn_cast<Constant>(Current)) {
559 // If Current is a constant, apply s/zext using ConstantExpr::getCast.
560 // ConstantExpr::getCast emits a ConstantInt if C is a ConstantInt.
561 Current = ConstantExpr::getCast((*I)->getOpcode(), C, (*I)->getType());
563 Instruction *Ext = (*I)->clone();
564 Ext->setOperand(0, Current);
565 Ext->insertBefore(IP);
572 Value *ConstantOffsetExtractor::rebuildWithoutConstOffset() {
573 distributeExtsAndCloneChain(UserChain.size() - 1);
574 // Remove all nullptrs (used to be s/zext) from UserChain.
575 unsigned NewSize = 0;
576 for (auto I = UserChain.begin(), E = UserChain.end(); I != E; ++I) {
578 UserChain[NewSize] = *I;
582 UserChain.resize(NewSize);
583 return removeConstOffset(UserChain.size() - 1);
587 ConstantOffsetExtractor::distributeExtsAndCloneChain(unsigned ChainIndex) {
588 User *U = UserChain[ChainIndex];
589 if (ChainIndex == 0) {
590 assert(isa<ConstantInt>(U));
591 // If U is a ConstantInt, applyExts will return a ConstantInt as well.
592 return UserChain[ChainIndex] = cast<ConstantInt>(applyExts(U));
595 if (CastInst *Cast = dyn_cast<CastInst>(U)) {
596 assert((isa<SExtInst>(Cast) || isa<ZExtInst>(Cast)) &&
597 "We only traced into two types of CastInst: sext and zext");
598 ExtInsts.push_back(Cast);
599 UserChain[ChainIndex] = nullptr;
600 return distributeExtsAndCloneChain(ChainIndex - 1);
603 // Function find only trace into BinaryOperator and CastInst.
604 BinaryOperator *BO = cast<BinaryOperator>(U);
605 // OpNo = which operand of BO is UserChain[ChainIndex - 1]
606 unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
607 Value *TheOther = applyExts(BO->getOperand(1 - OpNo));
608 Value *NextInChain = distributeExtsAndCloneChain(ChainIndex - 1);
610 BinaryOperator *NewBO = nullptr;
612 NewBO = BinaryOperator::Create(BO->getOpcode(), NextInChain, TheOther,
615 NewBO = BinaryOperator::Create(BO->getOpcode(), TheOther, NextInChain,
618 return UserChain[ChainIndex] = NewBO;
621 Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
622 if (ChainIndex == 0) {
623 assert(isa<ConstantInt>(UserChain[ChainIndex]));
624 return ConstantInt::getNullValue(UserChain[ChainIndex]->getType());
627 BinaryOperator *BO = cast<BinaryOperator>(UserChain[ChainIndex]);
628 assert(BO->getNumUses() <= 1 &&
629 "distributeExtsAndCloneChain clones each BinaryOperator in "
630 "UserChain, so no one should be used more than "
633 unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
634 assert(BO->getOperand(OpNo) == UserChain[ChainIndex - 1]);
635 Value *NextInChain = removeConstOffset(ChainIndex - 1);
636 Value *TheOther = BO->getOperand(1 - OpNo);
638 // If NextInChain is 0 and not the LHS of a sub, we can simplify the
639 // sub-expression to be just TheOther.
640 if (ConstantInt *CI = dyn_cast<ConstantInt>(NextInChain)) {
641 if (CI->isZero() && !(BO->getOpcode() == Instruction::Sub && OpNo == 0))
645 BinaryOperator::BinaryOps NewOp = BO->getOpcode();
646 if (BO->getOpcode() == Instruction::Or) {
647 // Rebuild "or" as "add", because "or" may be invalid for the new
650 // For instance, given
651 // a | (b + 5) where a and b + 5 have no common bits,
652 // we can extract 5 as the constant offset.
654 // However, reusing the "or" in the new index would give us
656 // which does not equal a | (b + 5).
658 // Replacing the "or" with "add" is fine, because
659 // a | (b + 5) = a + (b + 5) = (a + b) + 5
660 NewOp = Instruction::Add;
663 BinaryOperator *NewBO;
665 NewBO = BinaryOperator::Create(NewOp, NextInChain, TheOther, "", IP);
667 NewBO = BinaryOperator::Create(NewOp, TheOther, NextInChain, "", IP);
673 Value *ConstantOffsetExtractor::Extract(Value *Idx, GetElementPtrInst *GEP,
674 User *&UserChainTail,
675 const DominatorTree *DT) {
676 ConstantOffsetExtractor Extractor(GEP, DT);
677 // Find a non-zero constant offset first.
678 APInt ConstantOffset =
679 Extractor.find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
681 if (ConstantOffset == 0) {
682 UserChainTail = nullptr;
685 // Separates the constant offset from the GEP index.
686 Value *IdxWithoutConstOffset = Extractor.rebuildWithoutConstOffset();
687 UserChainTail = Extractor.UserChain.back();
688 return IdxWithoutConstOffset;
691 int64_t ConstantOffsetExtractor::Find(Value *Idx, GetElementPtrInst *GEP,
692 const DominatorTree *DT) {
693 // If Idx is an index of an inbound GEP, Idx is guaranteed to be non-negative.
694 return ConstantOffsetExtractor(GEP, DT)
695 .find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
700 bool SeparateConstOffsetFromGEP::canonicalizeArrayIndicesToPointerSize(
701 GetElementPtrInst *GEP) {
702 bool Changed = false;
703 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
704 gep_type_iterator GTI = gep_type_begin(*GEP);
705 for (User::op_iterator I = GEP->op_begin() + 1, E = GEP->op_end();
706 I != E; ++I, ++GTI) {
707 // Skip struct member indices which must be i32.
708 if (isa<SequentialType>(*GTI)) {
709 if ((*I)->getType() != IntPtrTy) {
710 *I = CastInst::CreateIntegerCast(*I, IntPtrTy, true, "idxprom", GEP);
719 SeparateConstOffsetFromGEP::accumulateByteOffset(GetElementPtrInst *GEP,
720 bool &NeedsExtraction) {
721 NeedsExtraction = false;
722 int64_t AccumulativeByteOffset = 0;
723 gep_type_iterator GTI = gep_type_begin(*GEP);
724 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
725 if (isa<SequentialType>(*GTI)) {
726 // Tries to extract a constant offset from this GEP index.
727 int64_t ConstantOffset =
728 ConstantOffsetExtractor::Find(GEP->getOperand(I), GEP, DT);
729 if (ConstantOffset != 0) {
730 NeedsExtraction = true;
731 // A GEP may have multiple indices. We accumulate the extracted
732 // constant offset to a byte offset, and later offset the remainder of
733 // the original GEP with this byte offset.
734 AccumulativeByteOffset +=
735 ConstantOffset * DL->getTypeAllocSize(GTI.getIndexedType());
737 } else if (LowerGEP) {
738 StructType *StTy = cast<StructType>(*GTI);
739 uint64_t Field = cast<ConstantInt>(GEP->getOperand(I))->getZExtValue();
740 // Skip field 0 as the offset is always 0.
742 NeedsExtraction = true;
743 AccumulativeByteOffset +=
744 DL->getStructLayout(StTy)->getElementOffset(Field);
748 return AccumulativeByteOffset;
751 void SeparateConstOffsetFromGEP::lowerToSingleIndexGEPs(
752 GetElementPtrInst *Variadic, int64_t AccumulativeByteOffset) {
753 IRBuilder<> Builder(Variadic);
754 Type *IntPtrTy = DL->getIntPtrType(Variadic->getType());
757 Builder.getInt8PtrTy(Variadic->getType()->getPointerAddressSpace());
758 Value *ResultPtr = Variadic->getOperand(0);
759 if (ResultPtr->getType() != I8PtrTy)
760 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
762 gep_type_iterator GTI = gep_type_begin(*Variadic);
763 // Create an ugly GEP for each sequential index. We don't create GEPs for
764 // structure indices, as they are accumulated in the constant offset index.
765 for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
766 if (isa<SequentialType>(*GTI)) {
767 Value *Idx = Variadic->getOperand(I);
768 // Skip zero indices.
769 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
773 APInt ElementSize = APInt(IntPtrTy->getIntegerBitWidth(),
774 DL->getTypeAllocSize(GTI.getIndexedType()));
775 // Scale the index by element size.
776 if (ElementSize != 1) {
777 if (ElementSize.isPowerOf2()) {
778 Idx = Builder.CreateShl(
779 Idx, ConstantInt::get(IntPtrTy, ElementSize.logBase2()));
781 Idx = Builder.CreateMul(Idx, ConstantInt::get(IntPtrTy, ElementSize));
784 // Create an ugly GEP with a single index for each index.
786 Builder.CreateGEP(Builder.getInt8Ty(), ResultPtr, Idx, "uglygep");
790 // Create a GEP with the constant offset index.
791 if (AccumulativeByteOffset != 0) {
792 Value *Offset = ConstantInt::get(IntPtrTy, AccumulativeByteOffset);
794 Builder.CreateGEP(Builder.getInt8Ty(), ResultPtr, Offset, "uglygep");
796 if (ResultPtr->getType() != Variadic->getType())
797 ResultPtr = Builder.CreateBitCast(ResultPtr, Variadic->getType());
799 Variadic->replaceAllUsesWith(ResultPtr);
800 Variadic->eraseFromParent();
804 SeparateConstOffsetFromGEP::lowerToArithmetics(GetElementPtrInst *Variadic,
805 int64_t AccumulativeByteOffset) {
806 IRBuilder<> Builder(Variadic);
807 Type *IntPtrTy = DL->getIntPtrType(Variadic->getType());
809 Value *ResultPtr = Builder.CreatePtrToInt(Variadic->getOperand(0), IntPtrTy);
810 gep_type_iterator GTI = gep_type_begin(*Variadic);
811 // Create ADD/SHL/MUL arithmetic operations for each sequential indices. We
812 // don't create arithmetics for structure indices, as they are accumulated
813 // in the constant offset index.
814 for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
815 if (isa<SequentialType>(*GTI)) {
816 Value *Idx = Variadic->getOperand(I);
817 // Skip zero indices.
818 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
822 APInt ElementSize = APInt(IntPtrTy->getIntegerBitWidth(),
823 DL->getTypeAllocSize(GTI.getIndexedType()));
824 // Scale the index by element size.
825 if (ElementSize != 1) {
826 if (ElementSize.isPowerOf2()) {
827 Idx = Builder.CreateShl(
828 Idx, ConstantInt::get(IntPtrTy, ElementSize.logBase2()));
830 Idx = Builder.CreateMul(Idx, ConstantInt::get(IntPtrTy, ElementSize));
833 // Create an ADD for each index.
834 ResultPtr = Builder.CreateAdd(ResultPtr, Idx);
838 // Create an ADD for the constant offset index.
839 if (AccumulativeByteOffset != 0) {
840 ResultPtr = Builder.CreateAdd(
841 ResultPtr, ConstantInt::get(IntPtrTy, AccumulativeByteOffset));
844 ResultPtr = Builder.CreateIntToPtr(ResultPtr, Variadic->getType());
845 Variadic->replaceAllUsesWith(ResultPtr);
846 Variadic->eraseFromParent();
849 bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
851 if (GEP->getType()->isVectorTy())
854 // The backend can already nicely handle the case where all indices are
856 if (GEP->hasAllConstantIndices())
859 bool Changed = canonicalizeArrayIndicesToPointerSize(GEP);
861 bool NeedsExtraction;
862 int64_t AccumulativeByteOffset = accumulateByteOffset(GEP, NeedsExtraction);
864 if (!NeedsExtraction)
866 // If LowerGEP is disabled, before really splitting the GEP, check whether the
867 // backend supports the addressing mode we are about to produce. If no, this
868 // splitting probably won't be beneficial.
869 // If LowerGEP is enabled, even the extracted constant offset can not match
870 // the addressing mode, we can still do optimizations to other lowered parts
871 // of variable indices. Therefore, we don't check for addressing modes in that
874 TargetTransformInfo &TTI =
875 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
876 *GEP->getParent()->getParent());
877 unsigned AddrSpace = GEP->getPointerAddressSpace();
878 if (!TTI.isLegalAddressingMode(GEP->getType()->getElementType(),
879 /*BaseGV=*/nullptr, AccumulativeByteOffset,
880 /*HasBaseReg=*/true, /*Scale=*/0,
886 // Remove the constant offset in each sequential index. The resultant GEP
887 // computes the variadic base.
888 // Notice that we don't remove struct field indices here. If LowerGEP is
889 // disabled, a structure index is not accumulated and we still use the old
890 // one. If LowerGEP is enabled, a structure index is accumulated in the
891 // constant offset. LowerToSingleIndexGEPs or lowerToArithmetics will later
892 // handle the constant offset and won't need a new structure index.
893 gep_type_iterator GTI = gep_type_begin(*GEP);
894 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
895 if (isa<SequentialType>(*GTI)) {
896 // Splits this GEP index into a variadic part and a constant offset, and
897 // uses the variadic part as the new index.
898 Value *OldIdx = GEP->getOperand(I);
901 ConstantOffsetExtractor::Extract(OldIdx, GEP, UserChainTail, DT);
902 if (NewIdx != nullptr) {
903 // Switches to the index with the constant offset removed.
904 GEP->setOperand(I, NewIdx);
905 // After switching to the new index, we can garbage-collect UserChain
906 // and the old index if they are not used.
907 RecursivelyDeleteTriviallyDeadInstructions(UserChainTail);
908 RecursivelyDeleteTriviallyDeadInstructions(OldIdx);
913 // Clear the inbounds attribute because the new index may be off-bound.
917 // addr = gep inbounds float, float* p, i64 b
919 // is transformed to:
921 // addr2 = gep float, float* p, i64 a ; inbounds removed
922 // addr = gep inbounds float, float* addr2, i64 5
924 // If a is -4, although the old index b is in bounds, the new index a is
925 // off-bound. http://llvm.org/docs/LangRef.html#id181 says "if the
926 // inbounds keyword is not present, the offsets are added to the base
927 // address with silently-wrapping two's complement arithmetic".
928 // Therefore, the final code will be a semantically equivalent.
930 // TODO(jingyue): do some range analysis to keep as many inbounds as
931 // possible. GEPs with inbounds are more friendly to alias analysis.
932 bool GEPWasInBounds = GEP->isInBounds();
933 GEP->setIsInBounds(false);
935 // Lowers a GEP to either GEPs with a single index or arithmetic operations.
937 // As currently BasicAA does not analyze ptrtoint/inttoptr, do not lower to
938 // arithmetic operations if the target uses alias analysis in codegen.
939 if (TM && TM->getSubtargetImpl(*GEP->getParent()->getParent())->useAA())
940 lowerToSingleIndexGEPs(GEP, AccumulativeByteOffset);
942 lowerToArithmetics(GEP, AccumulativeByteOffset);
946 // No need to create another GEP if the accumulative byte offset is 0.
947 if (AccumulativeByteOffset == 0)
950 // Offsets the base with the accumulative byte offset.
957 // %gep2 ; clone of %gep
958 // %new.gep = gep %gep2, <offset / sizeof(*%gep)>
959 // %gep ; will be removed
962 // => replace all uses of %gep with %new.gep and remove %gep
964 // %gep2 ; clone of %gep
965 // %new.gep = gep %gep2, <offset / sizeof(*%gep)>
968 // If AccumulativeByteOffset is not a multiple of sizeof(*%gep), we emit an
969 // uglygep (http://llvm.org/docs/GetElementPtr.html#what-s-an-uglygep):
970 // bitcast %gep2 to i8*, add the offset, and bitcast the result back to the
973 // %gep2 ; clone of %gep
974 // %0 = bitcast %gep2 to i8*
975 // %uglygep = gep %0, <offset>
976 // %new.gep = bitcast %uglygep to <type of %gep>
978 Instruction *NewGEP = GEP->clone();
979 NewGEP->insertBefore(GEP);
981 // Per ANSI C standard, signed / unsigned = unsigned and signed % unsigned =
982 // unsigned.. Therefore, we cast ElementTypeSizeOfGEP to signed because it is
983 // used with unsigned integers later.
984 int64_t ElementTypeSizeOfGEP = static_cast<int64_t>(
985 DL->getTypeAllocSize(GEP->getType()->getElementType()));
986 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
987 if (AccumulativeByteOffset % ElementTypeSizeOfGEP == 0) {
988 // Very likely. As long as %gep is natually aligned, the byte offset we
989 // extracted should be a multiple of sizeof(*%gep).
990 int64_t Index = AccumulativeByteOffset / ElementTypeSizeOfGEP;
991 NewGEP = GetElementPtrInst::Create(GEP->getResultElementType(), NewGEP,
992 ConstantInt::get(IntPtrTy, Index, true),
993 GEP->getName(), GEP);
994 // Inherit the inbounds attribute of the original GEP.
995 cast<GetElementPtrInst>(NewGEP)->setIsInBounds(GEPWasInBounds);
997 // Unlikely but possible. For example,
1005 // Suppose the gep before extraction is &s[i + 1].b[j + 3]. After
1006 // extraction, it becomes &s[i].b[j] and AccumulativeByteOffset is
1007 // sizeof(S) + 3 * sizeof(int64) = 100, which is not a multiple of
1010 // Emit an uglygep in this case.
1011 Type *I8PtrTy = Type::getInt8PtrTy(GEP->getContext(),
1012 GEP->getPointerAddressSpace());
1013 NewGEP = new BitCastInst(NewGEP, I8PtrTy, "", GEP);
1014 NewGEP = GetElementPtrInst::Create(
1015 Type::getInt8Ty(GEP->getContext()), NewGEP,
1016 ConstantInt::get(IntPtrTy, AccumulativeByteOffset, true), "uglygep",
1018 // Inherit the inbounds attribute of the original GEP.
1019 cast<GetElementPtrInst>(NewGEP)->setIsInBounds(GEPWasInBounds);
1020 if (GEP->getType() != I8PtrTy)
1021 NewGEP = new BitCastInst(NewGEP, GEP->getType(), GEP->getName(), GEP);
1024 GEP->replaceAllUsesWith(NewGEP);
1025 GEP->eraseFromParent();
1030 bool SeparateConstOffsetFromGEP::runOnFunction(Function &F) {
1031 if (skipOptnoneFunction(F))
1034 if (DisableSeparateConstOffsetFromGEP)
1037 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1038 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1040 bool Changed = false;
1041 for (Function::iterator B = F.begin(), BE = F.end(); B != BE; ++B) {
1042 for (BasicBlock::iterator I = B->begin(), IE = B->end(); I != IE; ) {
1043 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I++)) {
1044 Changed |= splitGEP(GEP);
1046 // No need to split GEP ConstantExprs because all its indices are constant
1051 Changed |= reuniteExts(F);
1053 if (VerifyNoDeadCode)
1054 verifyNoDeadCode(F);
1059 Instruction *SeparateConstOffsetFromGEP::findClosestMatchingDominator(
1060 const SCEV *Key, Instruction *Dominatee) {
1061 auto Pos = DominatingExprs.find(Key);
1062 if (Pos == DominatingExprs.end())
1065 auto &Candidates = Pos->second;
1066 // Because we process the basic blocks in pre-order of the dominator tree, a
1067 // candidate that doesn't dominate the current instruction won't dominate any
1068 // future instruction either. Therefore, we pop it out of the stack. This
1069 // optimization makes the algorithm O(n).
1070 while (!Candidates.empty()) {
1071 Instruction *Candidate = Candidates.back();
1072 if (DT->dominates(Candidate, Dominatee))
1074 Candidates.pop_back();
1079 bool SeparateConstOffsetFromGEP::reuniteExts(Instruction *I) {
1080 if (!SE->isSCEVable(I->getType()))
1084 // I: sext(LHS)+sext(RHS)
1085 // If Dom can't sign overflow and Dom dominates I, optimize I to sext(Dom).
1086 // TODO: handle zext
1087 Value *LHS = nullptr, *RHS = nullptr;
1088 if (match(I, m_Add(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS)))) ||
1089 match(I, m_Sub(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
1090 if (LHS->getType() == RHS->getType()) {
1092 SE->getAddExpr(SE->getUnknown(LHS), SE->getUnknown(RHS));
1093 if (auto *Dom = findClosestMatchingDominator(Key, I)) {
1094 Instruction *NewSExt = new SExtInst(Dom, I->getType(), "", I);
1095 NewSExt->takeName(I);
1096 I->replaceAllUsesWith(NewSExt);
1097 RecursivelyDeleteTriviallyDeadInstructions(I);
1103 // Add I to DominatingExprs if it's an add/sub that can't sign overflow.
1104 if (match(I, m_NSWAdd(m_Value(LHS), m_Value(RHS))) ||
1105 match(I, m_NSWSub(m_Value(LHS), m_Value(RHS)))) {
1106 if (isKnownNotFullPoison(I)) {
1108 SE->getAddExpr(SE->getUnknown(LHS), SE->getUnknown(RHS));
1109 DominatingExprs[Key].push_back(I);
1115 bool SeparateConstOffsetFromGEP::reuniteExts(Function &F) {
1116 bool Changed = false;
1117 DominatingExprs.clear();
1118 for (auto Node = GraphTraits<DominatorTree *>::nodes_begin(DT);
1119 Node != GraphTraits<DominatorTree *>::nodes_end(DT); ++Node) {
1120 BasicBlock *BB = Node->getBlock();
1121 for (auto I = BB->begin(); I != BB->end(); ) {
1122 Instruction *Cur = I++;
1123 Changed |= reuniteExts(Cur);
1129 void SeparateConstOffsetFromGEP::verifyNoDeadCode(Function &F) {
1132 if (isInstructionTriviallyDead(&I)) {
1133 std::string ErrMessage;
1134 raw_string_ostream RSO(ErrMessage);
1135 RSO << "Dead instruction detected!\n" << I << "\n";
1136 llvm_unreachable(RSO.str().c_str());