1//===- SeparateConstOffsetFromGEP.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// Loop unrolling may create many similar GEPs for array accesses.
10// e.g., a 2-level loop
11//
12// float a[32][32]; // global variable
13//
14// for (int i = 0; i < 2; ++i) {
15// for (int j = 0; j < 2; ++j) {
16// ...
17// ... = a[x + i][y + j];
18// ...
19// }
20// }
21//
22// will probably be unrolled to:
23//
24// gep %a, 0, %x, %y; load
25// gep %a, 0, %x, %y + 1; load
26// gep %a, 0, %x + 1, %y; load
27// gep %a, 0, %x + 1, %y + 1; load
28//
29// LLVM's GVN does not use partial redundancy elimination yet, and is thus
30// unable to reuse (gep %a, 0, %x, %y). As a result, this misoptimization incurs
31// significant slowdown in targets with limited addressing modes. For instance,
32// because the PTX target does not support the reg+reg addressing mode, the
33// NVPTX backend emits PTX code that literally computes the pointer address of
34// each GEP, wasting tons of registers. It emits the following PTX for the
35// first load and similar PTX for other loads.
36//
37// mov.u32 %r1, %x;
38// mov.u32 %r2, %y;
39// mul.wide.u32 %rl2, %r1, 128;
40// mov.u64 %rl3, a;
41// add.s64 %rl4, %rl3, %rl2;
42// mul.wide.u32 %rl5, %r2, 4;
43// add.s64 %rl6, %rl4, %rl5;
44// ld.global.f32 %f1, [%rl6];
45//
46// To reduce the register pressure, the optimization implemented in this file
47// merges the common part of a group of GEPs, so we can compute each pointer
48// address by adding a simple offset to the common part, saving many registers.
49//
50// It works by splitting each GEP into a variadic base and a constant offset.
51// The variadic base can be computed once and reused by multiple GEPs, and the
52// constant offsets can be nicely folded into the reg+immediate addressing mode
53// (supported by most targets) without using any extra register.
54//
55// For instance, we transform the four GEPs and four loads in the above example
56// into:
57//
58// base = gep a, 0, x, y
59// load base
60// laod base + 1 * sizeof(float)
61// load base + 32 * sizeof(float)
62// load base + 33 * sizeof(float)
63//
64// Given the transformed IR, a backend that supports the reg+immediate
65// addressing mode can easily fold the pointer arithmetics into the loads. For
66// example, the NVPTX backend can easily fold the pointer arithmetics into the
67// ld.global.f32 instructions, and the resultant PTX uses much fewer registers.
68//
69// mov.u32 %r1, %tid.x;
70// mov.u32 %r2, %tid.y;
71// mul.wide.u32 %rl2, %r1, 128;
72// mov.u64 %rl3, a;
73// add.s64 %rl4, %rl3, %rl2;
74// mul.wide.u32 %rl5, %r2, 4;
75// add.s64 %rl6, %rl4, %rl5;
76// ld.global.f32 %f1, [%rl6]; // so far the same as unoptimized PTX
77// ld.global.f32 %f2, [%rl6+4]; // much better
78// ld.global.f32 %f3, [%rl6+128]; // much better
79// ld.global.f32 %f4, [%rl6+132]; // much better
80//
81// Another improvement enabled by the LowerGEP flag is to lower a GEP with
82// multiple indices to either multiple GEPs with a single index or arithmetic
83// operations (depending on whether the target uses alias analysis in codegen).
84// Such transformation can have following benefits:
85// (1) It can always extract constants in the indices of structure type.
86// (2) After such Lowering, there are more optimization opportunities such as
87// CSE, LICM and CGP.
88//
89// E.g. The following GEPs have multiple indices:
90// BB1:
91// %p = getelementptr [10 x %struct]* %ptr, i64 %i, i64 %j1, i32 3
92// load %p
93// ...
94// BB2:
95// %p2 = getelementptr [10 x %struct]* %ptr, i64 %i, i64 %j1, i32 2
96// load %p2
97// ...
98//
99// We can not do CSE to the common part related to index "i64 %i". Lowering
100// GEPs can achieve such goals.
101// If the target does not use alias analysis in codegen, this pass will
102// lower a GEP with multiple indices into arithmetic operations:
103// BB1:
104// %1 = ptrtoint [10 x %struct]* %ptr to i64 ; CSE opportunity
105// %2 = mul i64 %i, length_of_10xstruct ; CSE opportunity
106// %3 = add i64 %1, %2 ; CSE opportunity
107// %4 = mul i64 %j1, length_of_struct
108// %5 = add i64 %3, %4
109// %6 = add i64 %3, struct_field_3 ; Constant offset
110// %p = inttoptr i64 %6 to i32*
111// load %p
112// ...
113// BB2:
114// %7 = ptrtoint [10 x %struct]* %ptr to i64 ; CSE opportunity
115// %8 = mul i64 %i, length_of_10xstruct ; CSE opportunity
116// %9 = add i64 %7, %8 ; CSE opportunity
117// %10 = mul i64 %j2, length_of_struct
118// %11 = add i64 %9, %10
119// %12 = add i64 %11, struct_field_2 ; Constant offset
120// %p = inttoptr i64 %12 to i32*
121// load %p2
122// ...
123//
124// If the target uses alias analysis in codegen, this pass will lower a GEP
125// with multiple indices into multiple GEPs with a single index:
126// BB1:
127// %1 = bitcast [10 x %struct]* %ptr to i8* ; CSE opportunity
128// %2 = mul i64 %i, length_of_10xstruct ; CSE opportunity
129// %3 = getelementptr i8* %1, i64 %2 ; CSE opportunity
130// %4 = mul i64 %j1, length_of_struct
131// %5 = getelementptr i8* %3, i64 %4
132// %6 = getelementptr i8* %5, struct_field_3 ; Constant offset
133// %p = bitcast i8* %6 to i32*
134// load %p
135// ...
136// BB2:
137// %7 = bitcast [10 x %struct]* %ptr to i8* ; CSE opportunity
138// %8 = mul i64 %i, length_of_10xstruct ; CSE opportunity
139// %9 = getelementptr i8* %7, i64 %8 ; CSE opportunity
140// %10 = mul i64 %j2, length_of_struct
141// %11 = getelementptr i8* %9, i64 %10
142// %12 = getelementptr i8* %11, struct_field_2 ; Constant offset
143// %p2 = bitcast i8* %12 to i32*
144// load %p2
145// ...
146//
147// Lowering GEPs can also benefit other passes such as LICM and CGP.
148// LICM (Loop Invariant Code Motion) can not hoist/sink a GEP of multiple
149// indices if one of the index is variant. If we lower such GEP into invariant
150// parts and variant parts, LICM can hoist/sink those invariant parts.
151// CGP (CodeGen Prepare) tries to sink address calculations that match the
152// target's addressing modes. A GEP with multiple indices may not match and will
153// not be sunk. If we lower such GEP into smaller parts, CGP may sink some of
154// them. So we end up with a better addressing mode.
155//
156//===----------------------------------------------------------------------===//
157
158#include "llvm/Transforms/Scalar/SeparateConstOffsetFromGEP.h"
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/TargetLibraryInfo.h"
166#include "llvm/Analysis/TargetTransformInfo.h"
167#include "llvm/Analysis/ValueTracking.h"
168#include "llvm/IR/BasicBlock.h"
169#include "llvm/IR/Constant.h"
170#include "llvm/IR/Constants.h"
171#include "llvm/IR/DataLayout.h"
172#include "llvm/IR/DerivedTypes.h"
173#include "llvm/IR/Dominators.h"
174#include "llvm/IR/Function.h"
175#include "llvm/IR/GetElementPtrTypeIterator.h"
176#include "llvm/IR/IRBuilder.h"
177#include "llvm/IR/InstrTypes.h"
178#include "llvm/IR/Instruction.h"
179#include "llvm/IR/Instructions.h"
180#include "llvm/IR/Module.h"
181#include "llvm/IR/PassManager.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/InitializePasses.h"
187#include "llvm/Pass.h"
188#include "llvm/Support/Casting.h"
189#include "llvm/Support/CommandLine.h"
190#include "llvm/Support/ErrorHandling.h"
191#include "llvm/Support/raw_ostream.h"
192#include "llvm/Transforms/Scalar.h"
193#include "llvm/Transforms/Utils/Local.h"
194#include <cassert>
195#include <cstdint>
196#include <string>
197
198using namespace llvm;
199using namespace llvm::PatternMatch;
200
201static cl::opt<bool> DisableSeparateConstOffsetFromGEP(
202 "disable-separate-const-offset-from-gep", cl::init(Val: false),
203 cl::desc("Do not separate the constant offset from a GEP instruction"),
204 cl::Hidden);
205
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.
209static cl::opt<bool>
210 VerifyNoDeadCode("reassociate-geps-verify-no-dead-code", cl::init(Val: false),
211 cl::desc("Verify this pass produces no dead code"),
212 cl::Hidden);
213
214namespace {
215
216/// A helper class for separating a constant offset from a GEP index.
217///
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.
222///
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).
229class ConstantOffsetExtractor {
230public:
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);
240
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
246private:
247 ConstantOffsetExtractor(BasicBlock::iterator InsertionPt)
248 : IP(InsertionPt), DL(InsertionPt->getModule()->getDataLayout()) {}
249
250 /// Searches the expression that computes V for a non-zero constant C s.t.
251 /// V can be reassociated into the form V' + C. If the searching is
252 /// successful, returns C and update UserChain as a def-use chain from C to V;
253 /// otherwise, UserChain is empty.
254 ///
255 /// \p V The given expression
256 /// \p SignExtended Whether V will be sign-extended in the computation of the
257 /// GEP index
258 /// \p ZeroExtended Whether V will be zero-extended in the computation of the
259 /// GEP index
260 /// \p NonNegative Whether V is guaranteed to be non-negative. For example,
261 /// an index of an inbounds GEP is guaranteed to be
262 /// non-negative. Levaraging this, we can better split
263 /// inbounds GEPs.
264 APInt find(Value *V, bool SignExtended, bool ZeroExtended, bool NonNegative);
265
266 /// A helper function to look into both operands of a binary operator.
267 APInt findInEitherOperand(BinaryOperator *BO, bool SignExtended,
268 bool ZeroExtended);
269
270 /// After finding the constant offset C from the GEP index I, we build a new
271 /// index I' s.t. I' + C = I. This function builds and returns the new
272 /// index I' according to UserChain produced by function "find".
273 ///
274 /// The building conceptually takes two steps:
275 /// 1) iteratively distribute s/zext towards the leaves of the expression tree
276 /// that computes I
277 /// 2) reassociate the expression tree to the form I' + C.
278 ///
279 /// For example, to extract the 5 from sext(a + (b + 5)), we first distribute
280 /// sext to a, b and 5 so that we have
281 /// sext(a) + (sext(b) + 5).
282 /// Then, we reassociate it to
283 /// (sext(a) + sext(b)) + 5.
284 /// Given this form, we know I' is sext(a) + sext(b).
285 Value *rebuildWithoutConstOffset();
286
287 /// After the first step of rebuilding the GEP index without the constant
288 /// offset, distribute s/zext to the operands of all operators in UserChain.
289 /// e.g., zext(sext(a + (b + 5)) (assuming no overflow) =>
290 /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))).
291 ///
292 /// The function also updates UserChain to point to new subexpressions after
293 /// distributing s/zext. e.g., the old UserChain of the above example is
294 /// 5 -> b + 5 -> a + (b + 5) -> sext(...) -> zext(sext(...)),
295 /// and the new UserChain is
296 /// zext(sext(5)) -> zext(sext(b)) + zext(sext(5)) ->
297 /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))
298 ///
299 /// \p ChainIndex The index to UserChain. ChainIndex is initially
300 /// UserChain.size() - 1, and is decremented during
301 /// the recursion.
302 Value *distributeExtsAndCloneChain(unsigned ChainIndex);
303
304 /// Reassociates the GEP index to the form I' + C and returns I'.
305 Value *removeConstOffset(unsigned ChainIndex);
306
307 /// A helper function to apply ExtInsts, a list of s/zext, to value V.
308 /// e.g., if ExtInsts = [sext i32 to i64, zext i16 to i32], this function
309 /// returns "sext i32 (zext i16 V to i32) to i64".
310 Value *applyExts(Value *V);
311
312 /// A helper function that returns whether we can trace into the operands
313 /// of binary operator BO for a constant offset.
314 ///
315 /// \p SignExtended Whether BO is surrounded by sext
316 /// \p ZeroExtended Whether BO is surrounded by zext
317 /// \p NonNegative Whether BO is known to be non-negative, e.g., an in-bound
318 /// array index.
319 bool CanTraceInto(bool SignExtended, bool ZeroExtended, BinaryOperator *BO,
320 bool NonNegative);
321
322 /// The path from the constant offset to the old GEP index. e.g., if the GEP
323 /// index is "a * b + (c + 5)". After running function find, UserChain[0] will
324 /// be the constant 5, UserChain[1] will be the subexpression "c + 5", and
325 /// UserChain[2] will be the entire expression "a * b + (c + 5)".
326 ///
327 /// This path helps to rebuild the new GEP index.
328 SmallVector<User *, 8> UserChain;
329
330 /// A data structure used in rebuildWithoutConstOffset. Contains all
331 /// sext/zext instructions along UserChain.
332 SmallVector<CastInst *, 16> ExtInsts;
333
334 /// Insertion position of cloned instructions.
335 BasicBlock::iterator IP;
336
337 const DataLayout &DL;
338};
339
340/// A pass that tries to split every GEP in the function into a variadic
341/// base and a constant offset. It is a FunctionPass because searching for the
342/// constant offset may inspect other basic blocks.
343class SeparateConstOffsetFromGEPLegacyPass : public FunctionPass {
344public:
345 static char ID;
346
347 SeparateConstOffsetFromGEPLegacyPass(bool LowerGEP = false)
348 : FunctionPass(ID), LowerGEP(LowerGEP) {
349 initializeSeparateConstOffsetFromGEPLegacyPassPass(
350 *PassRegistry::getPassRegistry());
351 }
352
353 void getAnalysisUsage(AnalysisUsage &AU) const override {
354 AU.addRequired<DominatorTreeWrapperPass>();
355 AU.addRequired<TargetTransformInfoWrapperPass>();
356 AU.addRequired<LoopInfoWrapperPass>();
357 AU.setPreservesCFG();
358 AU.addRequired<TargetLibraryInfoWrapperPass>();
359 }
360
361 bool runOnFunction(Function &F) override;
362
363private:
364 bool LowerGEP;
365};
366
367/// A pass that tries to split every GEP in the function into a variadic
368/// base and a constant offset. It is a FunctionPass because searching for the
369/// constant offset may inspect other basic blocks.
370class SeparateConstOffsetFromGEP {
371public:
372 SeparateConstOffsetFromGEP(
373 DominatorTree *DT, LoopInfo *LI, TargetLibraryInfo *TLI,
374 function_ref<TargetTransformInfo &(Function &)> GetTTI, bool LowerGEP)
375 : DT(DT), LI(LI), TLI(TLI), GetTTI(GetTTI), LowerGEP(LowerGEP) {}
376
377 bool run(Function &F);
378
379private:
380 /// Track the operands of an add or sub.
381 using ExprKey = std::pair<Value *, Value *>;
382
383 /// Create a pair for use as a map key for a commutable operation.
384 static ExprKey createNormalizedCommutablePair(Value *A, Value *B) {
385 if (A < B)
386 return {A, B};
387 return {B, A};
388 }
389
390 /// Tries to split the given GEP into a variadic base and a constant offset,
391 /// and returns true if the splitting succeeds.
392 bool splitGEP(GetElementPtrInst *GEP);
393
394 /// Tries to reorder the given GEP with the GEP that produces the base if
395 /// doing so results in producing a constant offset as the outermost
396 /// index.
397 bool reorderGEP(GetElementPtrInst *GEP, TargetTransformInfo &TTI);
398
399 /// Lower a GEP with multiple indices into multiple GEPs with a single index.
400 /// Function splitGEP already split the original GEP into a variadic part and
401 /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
402 /// variadic part into a set of GEPs with a single index and applies
403 /// AccumulativeByteOffset to it.
404 /// \p Variadic The variadic part of the original GEP.
405 /// \p AccumulativeByteOffset The constant offset.
406 void lowerToSingleIndexGEPs(GetElementPtrInst *Variadic,
407 int64_t AccumulativeByteOffset);
408
409 /// Lower a GEP with multiple indices into ptrtoint+arithmetics+inttoptr form.
410 /// Function splitGEP already split the original GEP into a variadic part and
411 /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
412 /// variadic part into a set of arithmetic operations and applies
413 /// AccumulativeByteOffset to it.
414 /// \p Variadic The variadic part of the original GEP.
415 /// \p AccumulativeByteOffset The constant offset.
416 void lowerToArithmetics(GetElementPtrInst *Variadic,
417 int64_t AccumulativeByteOffset);
418
419 /// Finds the constant offset within each index and accumulates them. If
420 /// LowerGEP is true, it finds in indices of both sequential and structure
421 /// types, otherwise it only finds in sequential indices. The output
422 /// NeedsExtraction indicates whether we successfully find a non-zero constant
423 /// offset.
424 int64_t accumulateByteOffset(GetElementPtrInst *GEP, bool &NeedsExtraction);
425
426 /// Canonicalize array indices to pointer-size integers. This helps to
427 /// simplify the logic of splitting a GEP. For example, if a + b is a
428 /// pointer-size integer, we have
429 /// gep base, a + b = gep (gep base, a), b
430 /// However, this equality may not hold if the size of a + b is smaller than
431 /// the pointer size, because LLVM conceptually sign-extends GEP indices to
432 /// pointer size before computing the address
433 /// (http://llvm.org/docs/LangRef.html#id181).
434 ///
435 /// This canonicalization is very likely already done in clang and
436 /// instcombine. Therefore, the program will probably remain the same.
437 ///
438 /// Returns true if the module changes.
439 ///
440 /// Verified in @i32_add in split-gep.ll
441 bool canonicalizeArrayIndicesToIndexSize(GetElementPtrInst *GEP);
442
443 /// Optimize sext(a)+sext(b) to sext(a+b) when a+b can't sign overflow.
444 /// SeparateConstOffsetFromGEP distributes a sext to leaves before extracting
445 /// the constant offset. After extraction, it becomes desirable to reunion the
446 /// distributed sexts. For example,
447 ///
448 /// &a[sext(i +nsw (j +nsw 5)]
449 /// => distribute &a[sext(i) +nsw (sext(j) +nsw 5)]
450 /// => constant extraction &a[sext(i) + sext(j)] + 5
451 /// => reunion &a[sext(i +nsw j)] + 5
452 bool reuniteExts(Function &F);
453
454 /// A helper that reunites sexts in an instruction.
455 bool reuniteExts(Instruction *I);
456
457 /// Find the closest dominator of <Dominatee> that is equivalent to <Key>.
458 Instruction *findClosestMatchingDominator(
459 ExprKey Key, Instruction *Dominatee,
460 DenseMap<ExprKey, SmallVector<Instruction *, 2>> &DominatingExprs);
461
462 /// Verify F is free of dead code.
463 void verifyNoDeadCode(Function &F);
464
465 bool hasMoreThanOneUseInLoop(Value *v, Loop *L);
466
467 // Swap the index operand of two GEP.
468 void swapGEPOperand(GetElementPtrInst *First, GetElementPtrInst *Second);
469
470 // Check if it is safe to swap operand of two GEP.
471 bool isLegalToSwapOperand(GetElementPtrInst *First, GetElementPtrInst *Second,
472 Loop *CurLoop);
473
474 const DataLayout *DL = nullptr;
475 DominatorTree *DT = nullptr;
476 LoopInfo *LI;
477 TargetLibraryInfo *TLI;
478 // Retrieved lazily since not always used.
479 function_ref<TargetTransformInfo &(Function &)> GetTTI;
480
481 /// Whether to lower a GEP with multiple indices into arithmetic operations or
482 /// multiple GEPs with a single index.
483 bool LowerGEP;
484
485 DenseMap<ExprKey, SmallVector<Instruction *, 2>> DominatingAdds;
486 DenseMap<ExprKey, SmallVector<Instruction *, 2>> DominatingSubs;
487};
488
489} // end anonymous namespace
490
491char SeparateConstOffsetFromGEPLegacyPass::ID = 0;
492
493INITIALIZE_PASS_BEGIN(
494 SeparateConstOffsetFromGEPLegacyPass, "separate-const-offset-from-gep",
495 "Split GEPs to a variadic base and a constant offset for better CSE", false,
496 false)
497INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
498INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
499INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
500INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
501INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
502INITIALIZE_PASS_END(
503 SeparateConstOffsetFromGEPLegacyPass, "separate-const-offset-from-gep",
504 "Split GEPs to a variadic base and a constant offset for better CSE", false,
505 false)
506
507FunctionPass *llvm::createSeparateConstOffsetFromGEPPass(bool LowerGEP) {
508 return new SeparateConstOffsetFromGEPLegacyPass(LowerGEP);
509}
510
511bool ConstantOffsetExtractor::CanTraceInto(bool SignExtended,
512 bool ZeroExtended,
513 BinaryOperator *BO,
514 bool NonNegative) {
515 // We only consider ADD, SUB and OR, because a non-zero constant found in
516 // expressions composed of these operations can be easily hoisted as a
517 // constant offset by reassociation.
518 if (BO->getOpcode() != Instruction::Add &&
519 BO->getOpcode() != Instruction::Sub &&
520 BO->getOpcode() != Instruction::Or) {
521 return false;
522 }
523
524 Value *LHS = BO->getOperand(i_nocapture: 0), *RHS = BO->getOperand(i_nocapture: 1);
525 // Do not trace into "or" unless it is equivalent to "add".
526 // This is the case if the or's disjoint flag is set.
527 if (BO->getOpcode() == Instruction::Or &&
528 !cast<PossiblyDisjointInst>(Val: BO)->isDisjoint())
529 return false;
530
531 // FIXME: We don't currently support constants from the RHS of subs,
532 // when we are zero-extended, because we need a way to zero-extended
533 // them before they are negated.
534 if (ZeroExtended && !SignExtended && BO->getOpcode() == Instruction::Sub)
535 return false;
536
537 // In addition, tracing into BO requires that its surrounding s/zext (if
538 // any) is distributable to both operands.
539 //
540 // Suppose BO = A op B.
541 // SignExtended | ZeroExtended | Distributable?
542 // --------------+--------------+----------------------------------
543 // 0 | 0 | true because no s/zext exists
544 // 0 | 1 | zext(BO) == zext(A) op zext(B)
545 // 1 | 0 | sext(BO) == sext(A) op sext(B)
546 // 1 | 1 | zext(sext(BO)) ==
547 // | | zext(sext(A)) op zext(sext(B))
548 if (BO->getOpcode() == Instruction::Add && !ZeroExtended && NonNegative) {
549 // If a + b >= 0 and (a >= 0 or b >= 0), then
550 // sext(a + b) = sext(a) + sext(b)
551 // even if the addition is not marked nsw.
552 //
553 // Leveraging this invariant, we can trace into an sext'ed inbound GEP
554 // index if the constant offset is non-negative.
555 //
556 // Verified in @sext_add in split-gep.ll.
557 if (ConstantInt *ConstLHS = dyn_cast<ConstantInt>(Val: LHS)) {
558 if (!ConstLHS->isNegative())
559 return true;
560 }
561 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Val: RHS)) {
562 if (!ConstRHS->isNegative())
563 return true;
564 }
565 }
566
567 // sext (add/sub nsw A, B) == add/sub nsw (sext A), (sext B)
568 // zext (add/sub nuw A, B) == add/sub nuw (zext A), (zext B)
569 if (BO->getOpcode() == Instruction::Add ||
570 BO->getOpcode() == Instruction::Sub) {
571 if (SignExtended && !BO->hasNoSignedWrap())
572 return false;
573 if (ZeroExtended && !BO->hasNoUnsignedWrap())
574 return false;
575 }
576
577 return true;
578}
579
580APInt ConstantOffsetExtractor::findInEitherOperand(BinaryOperator *BO,
581 bool SignExtended,
582 bool ZeroExtended) {
583 // Save off the current height of the chain, in case we need to restore it.
584 size_t ChainLength = UserChain.size();
585
586 // BO being non-negative does not shed light on whether its operands are
587 // non-negative. Clear the NonNegative flag here.
588 APInt ConstantOffset = find(V: BO->getOperand(i_nocapture: 0), SignExtended, ZeroExtended,
589 /* NonNegative */ false);
590 // If we found a constant offset in the left operand, stop and return that.
591 // This shortcut might cause us to miss opportunities of combining the
592 // constant offsets in both operands, e.g., (a + 4) + (b + 5) => (a + b) + 9.
593 // However, such cases are probably already handled by -instcombine,
594 // given this pass runs after the standard optimizations.
595 if (ConstantOffset != 0) return ConstantOffset;
596
597 // Reset the chain back to where it was when we started exploring this node,
598 // since visiting the LHS didn't pan out.
599 UserChain.resize(N: ChainLength);
600
601 ConstantOffset = find(V: BO->getOperand(i_nocapture: 1), SignExtended, ZeroExtended,
602 /* NonNegative */ false);
603 // If U is a sub operator, negate the constant offset found in the right
604 // operand.
605 if (BO->getOpcode() == Instruction::Sub)
606 ConstantOffset = -ConstantOffset;
607
608 // If RHS wasn't a suitable candidate either, reset the chain again.
609 if (ConstantOffset == 0)
610 UserChain.resize(N: ChainLength);
611
612 return ConstantOffset;
613}
614
615APInt ConstantOffsetExtractor::find(Value *V, bool SignExtended,
616 bool ZeroExtended, bool NonNegative) {
617 // TODO(jingyue): We could trace into integer/pointer casts, such as
618 // inttoptr, ptrtoint, bitcast, and addrspacecast. We choose to handle only
619 // integers because it gives good enough results for our benchmarks.
620 unsigned BitWidth = cast<IntegerType>(Val: V->getType())->getBitWidth();
621
622 // We cannot do much with Values that are not a User, such as an Argument.
623 User *U = dyn_cast<User>(Val: V);
624 if (U == nullptr) return APInt(BitWidth, 0);
625
626 APInt ConstantOffset(BitWidth, 0);
627 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) {
628 // Hooray, we found it!
629 ConstantOffset = CI->getValue();
630 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: V)) {
631 // Trace into subexpressions for more hoisting opportunities.
632 if (CanTraceInto(SignExtended, ZeroExtended, BO, NonNegative))
633 ConstantOffset = findInEitherOperand(BO, SignExtended, ZeroExtended);
634 } else if (isa<TruncInst>(Val: V)) {
635 ConstantOffset =
636 find(V: U->getOperand(i: 0), SignExtended, ZeroExtended, NonNegative)
637 .trunc(width: BitWidth);
638 } else if (isa<SExtInst>(Val: V)) {
639 ConstantOffset = find(V: U->getOperand(i: 0), /* SignExtended */ true,
640 ZeroExtended, NonNegative).sext(width: BitWidth);
641 } else if (isa<ZExtInst>(Val: V)) {
642 // As an optimization, we can clear the SignExtended flag because
643 // sext(zext(a)) = zext(a). Verified in @sext_zext in split-gep.ll.
644 //
645 // Clear the NonNegative flag, because zext(a) >= 0 does not imply a >= 0.
646 ConstantOffset =
647 find(V: U->getOperand(i: 0), /* SignExtended */ false,
648 /* ZeroExtended */ true, /* NonNegative */ false).zext(width: BitWidth);
649 }
650
651 // If we found a non-zero constant offset, add it to the path for
652 // rebuildWithoutConstOffset. Zero is a valid constant offset, but doesn't
653 // help this optimization.
654 if (ConstantOffset != 0)
655 UserChain.push_back(Elt: U);
656 return ConstantOffset;
657}
658
659Value *ConstantOffsetExtractor::applyExts(Value *V) {
660 Value *Current = V;
661 // ExtInsts is built in the use-def order. Therefore, we apply them to V
662 // in the reversed order.
663 for (CastInst *I : llvm::reverse(C&: ExtInsts)) {
664 if (Constant *C = dyn_cast<Constant>(Val: Current)) {
665 // Try to constant fold the cast.
666 Current = ConstantFoldCastOperand(Opcode: I->getOpcode(), C, DestTy: I->getType(), DL);
667 if (Current)
668 continue;
669 }
670
671 Instruction *Ext = I->clone();
672 Ext->setOperand(i: 0, Val: Current);
673 Ext->insertBefore(BB&: *IP->getParent(), InsertPos: IP);
674 Current = Ext;
675 }
676 return Current;
677}
678
679Value *ConstantOffsetExtractor::rebuildWithoutConstOffset() {
680 distributeExtsAndCloneChain(ChainIndex: UserChain.size() - 1);
681 // Remove all nullptrs (used to be s/zext) from UserChain.
682 unsigned NewSize = 0;
683 for (User *I : UserChain) {
684 if (I != nullptr) {
685 UserChain[NewSize] = I;
686 NewSize++;
687 }
688 }
689 UserChain.resize(N: NewSize);
690 return removeConstOffset(ChainIndex: UserChain.size() - 1);
691}
692
693Value *
694ConstantOffsetExtractor::distributeExtsAndCloneChain(unsigned ChainIndex) {
695 User *U = UserChain[ChainIndex];
696 if (ChainIndex == 0) {
697 assert(isa<ConstantInt>(U));
698 // If U is a ConstantInt, applyExts will return a ConstantInt as well.
699 return UserChain[ChainIndex] = cast<ConstantInt>(Val: applyExts(V: U));
700 }
701
702 if (CastInst *Cast = dyn_cast<CastInst>(Val: U)) {
703 assert(
704 (isa<SExtInst>(Cast) || isa<ZExtInst>(Cast) || isa<TruncInst>(Cast)) &&
705 "Only following instructions can be traced: sext, zext & trunc");
706 ExtInsts.push_back(Elt: Cast);
707 UserChain[ChainIndex] = nullptr;
708 return distributeExtsAndCloneChain(ChainIndex: ChainIndex - 1);
709 }
710
711 // Function find only trace into BinaryOperator and CastInst.
712 BinaryOperator *BO = cast<BinaryOperator>(Val: U);
713 // OpNo = which operand of BO is UserChain[ChainIndex - 1]
714 unsigned OpNo = (BO->getOperand(i_nocapture: 0) == UserChain[ChainIndex - 1] ? 0 : 1);
715 Value *TheOther = applyExts(V: BO->getOperand(i_nocapture: 1 - OpNo));
716 Value *NextInChain = distributeExtsAndCloneChain(ChainIndex: ChainIndex - 1);
717
718 BinaryOperator *NewBO = nullptr;
719 if (OpNo == 0) {
720 NewBO = BinaryOperator::Create(Op: BO->getOpcode(), S1: NextInChain, S2: TheOther,
721 Name: BO->getName(), InsertBefore: IP);
722 } else {
723 NewBO = BinaryOperator::Create(Op: BO->getOpcode(), S1: TheOther, S2: NextInChain,
724 Name: BO->getName(), InsertBefore: IP);
725 }
726 return UserChain[ChainIndex] = NewBO;
727}
728
729Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
730 if (ChainIndex == 0) {
731 assert(isa<ConstantInt>(UserChain[ChainIndex]));
732 return ConstantInt::getNullValue(Ty: UserChain[ChainIndex]->getType());
733 }
734
735 BinaryOperator *BO = cast<BinaryOperator>(Val: UserChain[ChainIndex]);
736 assert((BO->use_empty() || BO->hasOneUse()) &&
737 "distributeExtsAndCloneChain clones each BinaryOperator in "
738 "UserChain, so no one should be used more than "
739 "once");
740
741 unsigned OpNo = (BO->getOperand(i_nocapture: 0) == UserChain[ChainIndex - 1] ? 0 : 1);
742 assert(BO->getOperand(OpNo) == UserChain[ChainIndex - 1]);
743 Value *NextInChain = removeConstOffset(ChainIndex: ChainIndex - 1);
744 Value *TheOther = BO->getOperand(i_nocapture: 1 - OpNo);
745
746 // If NextInChain is 0 and not the LHS of a sub, we can simplify the
747 // sub-expression to be just TheOther.
748 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: NextInChain)) {
749 if (CI->isZero() && !(BO->getOpcode() == Instruction::Sub && OpNo == 0))
750 return TheOther;
751 }
752
753 BinaryOperator::BinaryOps NewOp = BO->getOpcode();
754 if (BO->getOpcode() == Instruction::Or) {
755 // Rebuild "or" as "add", because "or" may be invalid for the new
756 // expression.
757 //
758 // For instance, given
759 // a | (b + 5) where a and b + 5 have no common bits,
760 // we can extract 5 as the constant offset.
761 //
762 // However, reusing the "or" in the new index would give us
763 // (a | b) + 5
764 // which does not equal a | (b + 5).
765 //
766 // Replacing the "or" with "add" is fine, because
767 // a | (b + 5) = a + (b + 5) = (a + b) + 5
768 NewOp = Instruction::Add;
769 }
770
771 BinaryOperator *NewBO;
772 if (OpNo == 0) {
773 NewBO = BinaryOperator::Create(Op: NewOp, S1: NextInChain, S2: TheOther, Name: "", InsertBefore: IP);
774 } else {
775 NewBO = BinaryOperator::Create(Op: NewOp, S1: TheOther, S2: NextInChain, Name: "", InsertBefore: IP);
776 }
777 NewBO->takeName(V: BO);
778 return NewBO;
779}
780
781Value *ConstantOffsetExtractor::Extract(Value *Idx, GetElementPtrInst *GEP,
782 User *&UserChainTail) {
783 ConstantOffsetExtractor Extractor(GEP->getIterator());
784 // Find a non-zero constant offset first.
785 APInt ConstantOffset =
786 Extractor.find(V: Idx, /* SignExtended */ false, /* ZeroExtended */ false,
787 NonNegative: GEP->isInBounds());
788 if (ConstantOffset == 0) {
789 UserChainTail = nullptr;
790 return nullptr;
791 }
792 // Separates the constant offset from the GEP index.
793 Value *IdxWithoutConstOffset = Extractor.rebuildWithoutConstOffset();
794 UserChainTail = Extractor.UserChain.back();
795 return IdxWithoutConstOffset;
796}
797
798int64_t ConstantOffsetExtractor::Find(Value *Idx, GetElementPtrInst *GEP) {
799 // If Idx is an index of an inbound GEP, Idx is guaranteed to be non-negative.
800 return ConstantOffsetExtractor(GEP->getIterator())
801 .find(V: Idx, /* SignExtended */ false, /* ZeroExtended */ false,
802 NonNegative: GEP->isInBounds())
803 .getSExtValue();
804}
805
806bool SeparateConstOffsetFromGEP::canonicalizeArrayIndicesToIndexSize(
807 GetElementPtrInst *GEP) {
808 bool Changed = false;
809 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
810 gep_type_iterator GTI = gep_type_begin(GEP: *GEP);
811 for (User::op_iterator I = GEP->op_begin() + 1, E = GEP->op_end();
812 I != E; ++I, ++GTI) {
813 // Skip struct member indices which must be i32.
814 if (GTI.isSequential()) {
815 if ((*I)->getType() != PtrIdxTy) {
816 *I = CastInst::CreateIntegerCast(S: *I, Ty: PtrIdxTy, isSigned: true, Name: "idxprom",
817 InsertBefore: GEP->getIterator());
818 Changed = true;
819 }
820 }
821 }
822 return Changed;
823}
824
825int64_t
826SeparateConstOffsetFromGEP::accumulateByteOffset(GetElementPtrInst *GEP,
827 bool &NeedsExtraction) {
828 NeedsExtraction = false;
829 int64_t AccumulativeByteOffset = 0;
830 gep_type_iterator GTI = gep_type_begin(GEP: *GEP);
831 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
832 if (GTI.isSequential()) {
833 // Constant offsets of scalable types are not really constant.
834 if (GTI.getIndexedType()->isScalableTy())
835 continue;
836
837 // Tries to extract a constant offset from this GEP index.
838 int64_t ConstantOffset =
839 ConstantOffsetExtractor::Find(Idx: GEP->getOperand(i_nocapture: I), GEP);
840 if (ConstantOffset != 0) {
841 NeedsExtraction = true;
842 // A GEP may have multiple indices. We accumulate the extracted
843 // constant offset to a byte offset, and later offset the remainder of
844 // the original GEP with this byte offset.
845 AccumulativeByteOffset +=
846 ConstantOffset * GTI.getSequentialElementStride(DL: *DL);
847 }
848 } else if (LowerGEP) {
849 StructType *StTy = GTI.getStructType();
850 uint64_t Field = cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: I))->getZExtValue();
851 // Skip field 0 as the offset is always 0.
852 if (Field != 0) {
853 NeedsExtraction = true;
854 AccumulativeByteOffset +=
855 DL->getStructLayout(Ty: StTy)->getElementOffset(Idx: Field);
856 }
857 }
858 }
859 return AccumulativeByteOffset;
860}
861
862void SeparateConstOffsetFromGEP::lowerToSingleIndexGEPs(
863 GetElementPtrInst *Variadic, int64_t AccumulativeByteOffset) {
864 IRBuilder<> Builder(Variadic);
865 Type *PtrIndexTy = DL->getIndexType(PtrTy: Variadic->getType());
866
867 Value *ResultPtr = Variadic->getOperand(i_nocapture: 0);
868 Loop *L = LI->getLoopFor(BB: Variadic->getParent());
869 // Check if the base is not loop invariant or used more than once.
870 bool isSwapCandidate =
871 L && L->isLoopInvariant(V: ResultPtr) &&
872 !hasMoreThanOneUseInLoop(v: ResultPtr, L);
873 Value *FirstResult = nullptr;
874
875 gep_type_iterator GTI = gep_type_begin(GEP: *Variadic);
876 // Create an ugly GEP for each sequential index. We don't create GEPs for
877 // structure indices, as they are accumulated in the constant offset index.
878 for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
879 if (GTI.isSequential()) {
880 Value *Idx = Variadic->getOperand(i_nocapture: I);
881 // Skip zero indices.
882 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Idx))
883 if (CI->isZero())
884 continue;
885
886 APInt ElementSize = APInt(PtrIndexTy->getIntegerBitWidth(),
887 GTI.getSequentialElementStride(DL: *DL));
888 // Scale the index by element size.
889 if (ElementSize != 1) {
890 if (ElementSize.isPowerOf2()) {
891 Idx = Builder.CreateShl(
892 LHS: Idx, RHS: ConstantInt::get(Ty: PtrIndexTy, V: ElementSize.logBase2()));
893 } else {
894 Idx =
895 Builder.CreateMul(LHS: Idx, RHS: ConstantInt::get(Ty: PtrIndexTy, V: ElementSize));
896 }
897 }
898 // Create an ugly GEP with a single index for each index.
899 ResultPtr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: Idx, Name: "uglygep");
900 if (FirstResult == nullptr)
901 FirstResult = ResultPtr;
902 }
903 }
904
905 // Create a GEP with the constant offset index.
906 if (AccumulativeByteOffset != 0) {
907 Value *Offset = ConstantInt::get(Ty: PtrIndexTy, V: AccumulativeByteOffset);
908 ResultPtr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset, Name: "uglygep");
909 } else
910 isSwapCandidate = false;
911
912 // If we created a GEP with constant index, and the base is loop invariant,
913 // then we swap the first one with it, so LICM can move constant GEP out
914 // later.
915 auto *FirstGEP = dyn_cast_or_null<GetElementPtrInst>(Val: FirstResult);
916 auto *SecondGEP = dyn_cast<GetElementPtrInst>(Val: ResultPtr);
917 if (isSwapCandidate && isLegalToSwapOperand(First: FirstGEP, Second: SecondGEP, CurLoop: L))
918 swapGEPOperand(First: FirstGEP, Second: SecondGEP);
919
920 Variadic->replaceAllUsesWith(V: ResultPtr);
921 Variadic->eraseFromParent();
922}
923
924void
925SeparateConstOffsetFromGEP::lowerToArithmetics(GetElementPtrInst *Variadic,
926 int64_t AccumulativeByteOffset) {
927 IRBuilder<> Builder(Variadic);
928 Type *IntPtrTy = DL->getIntPtrType(Variadic->getType());
929 assert(IntPtrTy == DL->getIndexType(Variadic->getType()) &&
930 "Pointer type must match index type for arithmetic-based lowering of "
931 "split GEPs");
932
933 Value *ResultPtr = Builder.CreatePtrToInt(V: Variadic->getOperand(i_nocapture: 0), DestTy: IntPtrTy);
934 gep_type_iterator GTI = gep_type_begin(GEP: *Variadic);
935 // Create ADD/SHL/MUL arithmetic operations for each sequential indices. We
936 // don't create arithmetics for structure indices, as they are accumulated
937 // in the constant offset index.
938 for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
939 if (GTI.isSequential()) {
940 Value *Idx = Variadic->getOperand(i_nocapture: I);
941 // Skip zero indices.
942 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Idx))
943 if (CI->isZero())
944 continue;
945
946 APInt ElementSize = APInt(IntPtrTy->getIntegerBitWidth(),
947 GTI.getSequentialElementStride(DL: *DL));
948 // Scale the index by element size.
949 if (ElementSize != 1) {
950 if (ElementSize.isPowerOf2()) {
951 Idx = Builder.CreateShl(
952 LHS: Idx, RHS: ConstantInt::get(Ty: IntPtrTy, V: ElementSize.logBase2()));
953 } else {
954 Idx = Builder.CreateMul(LHS: Idx, RHS: ConstantInt::get(Ty: IntPtrTy, V: ElementSize));
955 }
956 }
957 // Create an ADD for each index.
958 ResultPtr = Builder.CreateAdd(LHS: ResultPtr, RHS: Idx);
959 }
960 }
961
962 // Create an ADD for the constant offset index.
963 if (AccumulativeByteOffset != 0) {
964 ResultPtr = Builder.CreateAdd(
965 LHS: ResultPtr, RHS: ConstantInt::get(Ty: IntPtrTy, V: AccumulativeByteOffset));
966 }
967
968 ResultPtr = Builder.CreateIntToPtr(V: ResultPtr, DestTy: Variadic->getType());
969 Variadic->replaceAllUsesWith(V: ResultPtr);
970 Variadic->eraseFromParent();
971}
972
973bool SeparateConstOffsetFromGEP::reorderGEP(GetElementPtrInst *GEP,
974 TargetTransformInfo &TTI) {
975 Type *GEPType = GEP->getResultElementType();
976 // TODO: support reordering for non-trivial GEP chains
977 if (GEPType->isAggregateType() || GEP->getNumIndices() != 1)
978 return false;
979
980 auto PtrGEP = dyn_cast<GetElementPtrInst>(Val: GEP->getPointerOperand());
981 if (!PtrGEP)
982 return false;
983 Type *PtrGEPType = PtrGEP->getResultElementType();
984 // TODO: support reordering for non-trivial GEP chains
985 if (PtrGEPType->isAggregateType() || PtrGEP->getNumIndices() != 1)
986 return false;
987
988 // TODO: support reordering for non-trivial GEP chains
989 if (PtrGEPType != GEPType ||
990 PtrGEP->getSourceElementType() != GEP->getSourceElementType())
991 return false;
992
993 bool NestedNeedsExtraction;
994 int64_t NestedByteOffset =
995 accumulateByteOffset(GEP: PtrGEP, NeedsExtraction&: NestedNeedsExtraction);
996 if (!NestedNeedsExtraction)
997 return false;
998
999 unsigned AddrSpace = PtrGEP->getPointerAddressSpace();
1000 if (!TTI.isLegalAddressingMode(Ty: GEP->getResultElementType(),
1001 /*BaseGV=*/nullptr, BaseOffset: NestedByteOffset,
1002 /*HasBaseReg=*/true, /*Scale=*/0, AddrSpace))
1003 return false;
1004
1005 IRBuilder<> Builder(GEP);
1006 Builder.SetCurrentDebugLocation(GEP->getDebugLoc());
1007 bool GEPInBounds = GEP->isInBounds();
1008 bool PtrGEPInBounds = PtrGEP->isInBounds();
1009 bool IsChainInBounds = GEPInBounds && PtrGEPInBounds;
1010 if (IsChainInBounds) {
1011 auto GEPIdx = GEP->indices().begin();
1012 auto KnownGEPIdx = computeKnownBits(V: GEPIdx->get(), DL: *DL);
1013 IsChainInBounds &= KnownGEPIdx.isNonNegative();
1014 if (IsChainInBounds) {
1015 auto PtrGEPIdx = GEP->indices().begin();
1016 auto KnownPtrGEPIdx = computeKnownBits(V: PtrGEPIdx->get(), DL: *DL);
1017 IsChainInBounds &= KnownPtrGEPIdx.isNonNegative();
1018 }
1019 }
1020
1021 // For trivial GEP chains, we can swap the indicies.
1022 auto NewSrc = Builder.CreateGEP(Ty: PtrGEPType, Ptr: PtrGEP->getPointerOperand(),
1023 IdxList: SmallVector<Value *, 4>(GEP->indices()));
1024 cast<GetElementPtrInst>(Val: NewSrc)->setIsInBounds(IsChainInBounds);
1025 auto NewGEP = Builder.CreateGEP(Ty: GEPType, Ptr: NewSrc,
1026 IdxList: SmallVector<Value *, 4>(PtrGEP->indices()));
1027 cast<GetElementPtrInst>(Val: NewGEP)->setIsInBounds(IsChainInBounds);
1028 GEP->replaceAllUsesWith(V: NewGEP);
1029 RecursivelyDeleteTriviallyDeadInstructions(V: GEP);
1030 return true;
1031}
1032
1033bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
1034 // Skip vector GEPs.
1035 if (GEP->getType()->isVectorTy())
1036 return false;
1037
1038 // The backend can already nicely handle the case where all indices are
1039 // constant.
1040 if (GEP->hasAllConstantIndices())
1041 return false;
1042
1043 bool Changed = canonicalizeArrayIndicesToIndexSize(GEP);
1044
1045 bool NeedsExtraction;
1046 int64_t AccumulativeByteOffset = accumulateByteOffset(GEP, NeedsExtraction);
1047
1048 TargetTransformInfo &TTI = GetTTI(*GEP->getFunction());
1049
1050 if (!NeedsExtraction) {
1051 Changed |= reorderGEP(GEP, TTI);
1052 return Changed;
1053 }
1054
1055 // If LowerGEP is disabled, before really splitting the GEP, check whether the
1056 // backend supports the addressing mode we are about to produce. If no, this
1057 // splitting probably won't be beneficial.
1058 // If LowerGEP is enabled, even the extracted constant offset can not match
1059 // the addressing mode, we can still do optimizations to other lowered parts
1060 // of variable indices. Therefore, we don't check for addressing modes in that
1061 // case.
1062 if (!LowerGEP) {
1063 unsigned AddrSpace = GEP->getPointerAddressSpace();
1064 if (!TTI.isLegalAddressingMode(Ty: GEP->getResultElementType(),
1065 /*BaseGV=*/nullptr, BaseOffset: AccumulativeByteOffset,
1066 /*HasBaseReg=*/true, /*Scale=*/0,
1067 AddrSpace)) {
1068 return Changed;
1069 }
1070 }
1071
1072 // Remove the constant offset in each sequential index. The resultant GEP
1073 // computes the variadic base.
1074 // Notice that we don't remove struct field indices here. If LowerGEP is
1075 // disabled, a structure index is not accumulated and we still use the old
1076 // one. If LowerGEP is enabled, a structure index is accumulated in the
1077 // constant offset. LowerToSingleIndexGEPs or lowerToArithmetics will later
1078 // handle the constant offset and won't need a new structure index.
1079 gep_type_iterator GTI = gep_type_begin(GEP: *GEP);
1080 for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
1081 if (GTI.isSequential()) {
1082 // Constant offsets of scalable types are not really constant.
1083 if (GTI.getIndexedType()->isScalableTy())
1084 continue;
1085
1086 // Splits this GEP index into a variadic part and a constant offset, and
1087 // uses the variadic part as the new index.
1088 Value *OldIdx = GEP->getOperand(i_nocapture: I);
1089 User *UserChainTail;
1090 Value *NewIdx =
1091 ConstantOffsetExtractor::Extract(Idx: OldIdx, GEP, UserChainTail);
1092 if (NewIdx != nullptr) {
1093 // Switches to the index with the constant offset removed.
1094 GEP->setOperand(i_nocapture: I, Val_nocapture: NewIdx);
1095 // After switching to the new index, we can garbage-collect UserChain
1096 // and the old index if they are not used.
1097 RecursivelyDeleteTriviallyDeadInstructions(V: UserChainTail);
1098 RecursivelyDeleteTriviallyDeadInstructions(V: OldIdx);
1099 }
1100 }
1101 }
1102
1103 // Clear the inbounds attribute because the new index may be off-bound.
1104 // e.g.,
1105 //
1106 // b = add i64 a, 5
1107 // addr = gep inbounds float, float* p, i64 b
1108 //
1109 // is transformed to:
1110 //
1111 // addr2 = gep float, float* p, i64 a ; inbounds removed
1112 // addr = gep inbounds float, float* addr2, i64 5
1113 //
1114 // If a is -4, although the old index b is in bounds, the new index a is
1115 // off-bound. http://llvm.org/docs/LangRef.html#id181 says "if the
1116 // inbounds keyword is not present, the offsets are added to the base
1117 // address with silently-wrapping two's complement arithmetic".
1118 // Therefore, the final code will be a semantically equivalent.
1119 //
1120 // TODO(jingyue): do some range analysis to keep as many inbounds as
1121 // possible. GEPs with inbounds are more friendly to alias analysis.
1122 bool GEPWasInBounds = GEP->isInBounds();
1123 GEP->setIsInBounds(false);
1124
1125 // Lowers a GEP to either GEPs with a single index or arithmetic operations.
1126 if (LowerGEP) {
1127 // As currently BasicAA does not analyze ptrtoint/inttoptr, do not lower to
1128 // arithmetic operations if the target uses alias analysis in codegen.
1129 // Additionally, pointers that aren't integral (and so can't be safely
1130 // converted to integers) or those whose offset size is different from their
1131 // pointer size (which means that doing integer arithmetic on them could
1132 // affect that data) can't be lowered in this way.
1133 unsigned AddrSpace = GEP->getPointerAddressSpace();
1134 bool PointerHasExtraData = DL->getPointerSizeInBits(AS: AddrSpace) !=
1135 DL->getIndexSizeInBits(AS: AddrSpace);
1136 if (TTI.useAA() || DL->isNonIntegralAddressSpace(AddrSpace) ||
1137 PointerHasExtraData)
1138 lowerToSingleIndexGEPs(Variadic: GEP, AccumulativeByteOffset);
1139 else
1140 lowerToArithmetics(Variadic: GEP, AccumulativeByteOffset);
1141 return true;
1142 }
1143
1144 // No need to create another GEP if the accumulative byte offset is 0.
1145 if (AccumulativeByteOffset == 0)
1146 return true;
1147
1148 // Offsets the base with the accumulative byte offset.
1149 //
1150 // %gep ; the base
1151 // ... %gep ...
1152 //
1153 // => add the offset
1154 //
1155 // %gep2 ; clone of %gep
1156 // %new.gep = gep i8, %gep2, %offset
1157 // %gep ; will be removed
1158 // ... %gep ...
1159 //
1160 // => replace all uses of %gep with %new.gep and remove %gep
1161 //
1162 // %gep2 ; clone of %gep
1163 // %new.gep = gep i8, %gep2, %offset
1164 // ... %new.gep ...
1165 Instruction *NewGEP = GEP->clone();
1166 NewGEP->insertBefore(InsertPos: GEP);
1167
1168 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
1169 IRBuilder<> Builder(GEP);
1170 NewGEP = cast<Instruction>(Val: Builder.CreatePtrAdd(
1171 Ptr: NewGEP, Offset: ConstantInt::get(Ty: PtrIdxTy, V: AccumulativeByteOffset, IsSigned: true),
1172 Name: GEP->getName(), IsInBounds: GEPWasInBounds));
1173 NewGEP->copyMetadata(SrcInst: *GEP);
1174
1175 GEP->replaceAllUsesWith(V: NewGEP);
1176 GEP->eraseFromParent();
1177
1178 return true;
1179}
1180
1181bool SeparateConstOffsetFromGEPLegacyPass::runOnFunction(Function &F) {
1182 if (skipFunction(F))
1183 return false;
1184 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1185 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1186 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1187 auto GetTTI = [this](Function &F) -> TargetTransformInfo & {
1188 return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1189 };
1190 SeparateConstOffsetFromGEP Impl(DT, LI, TLI, GetTTI, LowerGEP);
1191 return Impl.run(F);
1192}
1193
1194bool SeparateConstOffsetFromGEP::run(Function &F) {
1195 if (DisableSeparateConstOffsetFromGEP)
1196 return false;
1197
1198 DL = &F.getParent()->getDataLayout();
1199 bool Changed = false;
1200 for (BasicBlock &B : F) {
1201 if (!DT->isReachableFromEntry(A: &B))
1202 continue;
1203
1204 for (Instruction &I : llvm::make_early_inc_range(Range&: B))
1205 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: &I))
1206 Changed |= splitGEP(GEP);
1207 // No need to split GEP ConstantExprs because all its indices are constant
1208 // already.
1209 }
1210
1211 Changed |= reuniteExts(F);
1212
1213 if (VerifyNoDeadCode)
1214 verifyNoDeadCode(F);
1215
1216 return Changed;
1217}
1218
1219Instruction *SeparateConstOffsetFromGEP::findClosestMatchingDominator(
1220 ExprKey Key, Instruction *Dominatee,
1221 DenseMap<ExprKey, SmallVector<Instruction *, 2>> &DominatingExprs) {
1222 auto Pos = DominatingExprs.find(Val: Key);
1223 if (Pos == DominatingExprs.end())
1224 return nullptr;
1225
1226 auto &Candidates = Pos->second;
1227 // Because we process the basic blocks in pre-order of the dominator tree, a
1228 // candidate that doesn't dominate the current instruction won't dominate any
1229 // future instruction either. Therefore, we pop it out of the stack. This
1230 // optimization makes the algorithm O(n).
1231 while (!Candidates.empty()) {
1232 Instruction *Candidate = Candidates.back();
1233 if (DT->dominates(Def: Candidate, User: Dominatee))
1234 return Candidate;
1235 Candidates.pop_back();
1236 }
1237 return nullptr;
1238}
1239
1240bool SeparateConstOffsetFromGEP::reuniteExts(Instruction *I) {
1241 if (!I->getType()->isIntOrIntVectorTy())
1242 return false;
1243
1244 // Dom: LHS+RHS
1245 // I: sext(LHS)+sext(RHS)
1246 // If Dom can't sign overflow and Dom dominates I, optimize I to sext(Dom).
1247 // TODO: handle zext
1248 Value *LHS = nullptr, *RHS = nullptr;
1249 if (match(V: I, P: m_Add(L: m_SExt(Op: m_Value(V&: LHS)), R: m_SExt(Op: m_Value(V&: RHS))))) {
1250 if (LHS->getType() == RHS->getType()) {
1251 ExprKey Key = createNormalizedCommutablePair(A: LHS, B: RHS);
1252 if (auto *Dom = findClosestMatchingDominator(Key, Dominatee: I, DominatingExprs&: DominatingAdds)) {
1253 Instruction *NewSExt =
1254 new SExtInst(Dom, I->getType(), "", I->getIterator());
1255 NewSExt->takeName(V: I);
1256 I->replaceAllUsesWith(V: NewSExt);
1257 RecursivelyDeleteTriviallyDeadInstructions(V: I);
1258 return true;
1259 }
1260 }
1261 } else if (match(V: I, P: m_Sub(L: m_SExt(Op: m_Value(V&: LHS)), R: m_SExt(Op: m_Value(V&: RHS))))) {
1262 if (LHS->getType() == RHS->getType()) {
1263 if (auto *Dom =
1264 findClosestMatchingDominator(Key: {LHS, RHS}, Dominatee: I, DominatingExprs&: DominatingSubs)) {
1265 Instruction *NewSExt =
1266 new SExtInst(Dom, I->getType(), "", I->getIterator());
1267 NewSExt->takeName(V: I);
1268 I->replaceAllUsesWith(V: NewSExt);
1269 RecursivelyDeleteTriviallyDeadInstructions(V: I);
1270 return true;
1271 }
1272 }
1273 }
1274
1275 // Add I to DominatingExprs if it's an add/sub that can't sign overflow.
1276 if (match(V: I, P: m_NSWAdd(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) {
1277 if (programUndefinedIfPoison(Inst: I)) {
1278 ExprKey Key = createNormalizedCommutablePair(A: LHS, B: RHS);
1279 DominatingAdds[Key].push_back(Elt: I);
1280 }
1281 } else if (match(V: I, P: m_NSWSub(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) {
1282 if (programUndefinedIfPoison(Inst: I))
1283 DominatingSubs[{LHS, RHS}].push_back(Elt: I);
1284 }
1285 return false;
1286}
1287
1288bool SeparateConstOffsetFromGEP::reuniteExts(Function &F) {
1289 bool Changed = false;
1290 DominatingAdds.clear();
1291 DominatingSubs.clear();
1292 for (const auto Node : depth_first(G: DT)) {
1293 BasicBlock *BB = Node->getBlock();
1294 for (Instruction &I : llvm::make_early_inc_range(Range&: *BB))
1295 Changed |= reuniteExts(I: &I);
1296 }
1297 return Changed;
1298}
1299
1300void SeparateConstOffsetFromGEP::verifyNoDeadCode(Function &F) {
1301 for (BasicBlock &B : F) {
1302 for (Instruction &I : B) {
1303 if (isInstructionTriviallyDead(I: &I)) {
1304 std::string ErrMessage;
1305 raw_string_ostream RSO(ErrMessage);
1306 RSO << "Dead instruction detected!\n" << I << "\n";
1307 llvm_unreachable(RSO.str().c_str());
1308 }
1309 }
1310 }
1311}
1312
1313bool SeparateConstOffsetFromGEP::isLegalToSwapOperand(
1314 GetElementPtrInst *FirstGEP, GetElementPtrInst *SecondGEP, Loop *CurLoop) {
1315 if (!FirstGEP || !FirstGEP->hasOneUse())
1316 return false;
1317
1318 if (!SecondGEP || FirstGEP->getParent() != SecondGEP->getParent())
1319 return false;
1320
1321 if (FirstGEP == SecondGEP)
1322 return false;
1323
1324 unsigned FirstNum = FirstGEP->getNumOperands();
1325 unsigned SecondNum = SecondGEP->getNumOperands();
1326 // Give up if the number of operands are not 2.
1327 if (FirstNum != SecondNum || FirstNum != 2)
1328 return false;
1329
1330 Value *FirstBase = FirstGEP->getOperand(i_nocapture: 0);
1331 Value *SecondBase = SecondGEP->getOperand(i_nocapture: 0);
1332 Value *FirstOffset = FirstGEP->getOperand(i_nocapture: 1);
1333 // Give up if the index of the first GEP is loop invariant.
1334 if (CurLoop->isLoopInvariant(V: FirstOffset))
1335 return false;
1336
1337 // Give up if base doesn't have same type.
1338 if (FirstBase->getType() != SecondBase->getType())
1339 return false;
1340
1341 Instruction *FirstOffsetDef = dyn_cast<Instruction>(Val: FirstOffset);
1342
1343 // Check if the second operand of first GEP has constant coefficient.
1344 // For an example, for the following code, we won't gain anything by
1345 // hoisting the second GEP out because the second GEP can be folded away.
1346 // %scevgep.sum.ur159 = add i64 %idxprom48.ur, 256
1347 // %67 = shl i64 %scevgep.sum.ur159, 2
1348 // %uglygep160 = getelementptr i8* %65, i64 %67
1349 // %uglygep161 = getelementptr i8* %uglygep160, i64 -1024
1350
1351 // Skip constant shift instruction which may be generated by Splitting GEPs.
1352 if (FirstOffsetDef && FirstOffsetDef->isShift() &&
1353 isa<ConstantInt>(Val: FirstOffsetDef->getOperand(i: 1)))
1354 FirstOffsetDef = dyn_cast<Instruction>(Val: FirstOffsetDef->getOperand(i: 0));
1355
1356 // Give up if FirstOffsetDef is an Add or Sub with constant.
1357 // Because it may not profitable at all due to constant folding.
1358 if (FirstOffsetDef)
1359 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: FirstOffsetDef)) {
1360 unsigned opc = BO->getOpcode();
1361 if ((opc == Instruction::Add || opc == Instruction::Sub) &&
1362 (isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 0)) ||
1363 isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))))
1364 return false;
1365 }
1366 return true;
1367}
1368
1369bool SeparateConstOffsetFromGEP::hasMoreThanOneUseInLoop(Value *V, Loop *L) {
1370 int UsesInLoop = 0;
1371 for (User *U : V->users()) {
1372 if (Instruction *User = dyn_cast<Instruction>(Val: U))
1373 if (L->contains(Inst: User))
1374 if (++UsesInLoop > 1)
1375 return true;
1376 }
1377 return false;
1378}
1379
1380void SeparateConstOffsetFromGEP::swapGEPOperand(GetElementPtrInst *First,
1381 GetElementPtrInst *Second) {
1382 Value *Offset1 = First->getOperand(i_nocapture: 1);
1383 Value *Offset2 = Second->getOperand(i_nocapture: 1);
1384 First->setOperand(i_nocapture: 1, Val_nocapture: Offset2);
1385 Second->setOperand(i_nocapture: 1, Val_nocapture: Offset1);
1386
1387 // We changed p+o+c to p+c+o, p+c may not be inbound anymore.
1388 const DataLayout &DAL = First->getModule()->getDataLayout();
1389 APInt Offset(DAL.getIndexSizeInBits(
1390 AS: cast<PointerType>(Val: First->getType())->getAddressSpace()),
1391 0);
1392 Value *NewBase =
1393 First->stripAndAccumulateInBoundsConstantOffsets(DL: DAL, Offset);
1394 uint64_t ObjectSize;
1395 if (!getObjectSize(Ptr: NewBase, Size&: ObjectSize, DL: DAL, TLI) ||
1396 Offset.ugt(RHS: ObjectSize)) {
1397 First->setIsInBounds(false);
1398 Second->setIsInBounds(false);
1399 } else
1400 First->setIsInBounds(true);
1401}
1402
1403void SeparateConstOffsetFromGEPPass::printPipeline(
1404 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1405 static_cast<PassInfoMixin<SeparateConstOffsetFromGEPPass> *>(this)
1406 ->printPipeline(OS, MapClassName2PassName);
1407 OS << '<';
1408 if (LowerGEP)
1409 OS << "lower-gep";
1410 OS << '>';
1411}
1412
1413PreservedAnalyses
1414SeparateConstOffsetFromGEPPass::run(Function &F, FunctionAnalysisManager &AM) {
1415 auto *DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F);
1416 auto *LI = &AM.getResult<LoopAnalysis>(IR&: F);
1417 auto *TLI = &AM.getResult<TargetLibraryAnalysis>(IR&: F);
1418 auto GetTTI = [&AM](Function &F) -> TargetTransformInfo & {
1419 return AM.getResult<TargetIRAnalysis>(IR&: F);
1420 };
1421 SeparateConstOffsetFromGEP Impl(DT, LI, TLI, GetTTI, LowerGEP);
1422 if (!Impl.run(F))
1423 return PreservedAnalyses::all();
1424 PreservedAnalyses PA;
1425 PA.preserveSet<CFGAnalyses>();
1426 return PA;
1427}
1428

source code of llvm/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp