1 | //===- MergeICmps.cpp - Optimize chains of integer comparisons ------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This pass turns chains of integer comparisons into memcmp (the memcmp is |
10 | // later typically inlined as a chain of efficient hardware comparisons). This |
11 | // typically benefits c++ member or nonmember operator==(). |
12 | // |
13 | // The basic idea is to replace a longer chain of integer comparisons loaded |
14 | // from contiguous memory locations into a shorter chain of larger integer |
15 | // comparisons. Benefits are double: |
16 | // - There are less jumps, and therefore less opportunities for mispredictions |
17 | // and I-cache misses. |
18 | // - Code size is smaller, both because jumps are removed and because the |
19 | // encoding of a 2*n byte compare is smaller than that of two n-byte |
20 | // compares. |
21 | // |
22 | // Example: |
23 | // |
24 | // struct S { |
25 | // int a; |
26 | // char b; |
27 | // char c; |
28 | // uint16_t d; |
29 | // bool operator==(const S& o) const { |
30 | // return a == o.a && b == o.b && c == o.c && d == o.d; |
31 | // } |
32 | // }; |
33 | // |
34 | // Is optimized as : |
35 | // |
36 | // bool S::operator==(const S& o) const { |
37 | // return memcmp(this, &o, 8) == 0; |
38 | // } |
39 | // |
40 | // Which will later be expanded (ExpandMemCmp) as a single 8-bytes icmp. |
41 | // |
42 | //===----------------------------------------------------------------------===// |
43 | |
44 | #include "llvm/Transforms/Scalar/MergeICmps.h" |
45 | #include "llvm/ADT/SmallString.h" |
46 | #include "llvm/Analysis/DomTreeUpdater.h" |
47 | #include "llvm/Analysis/GlobalsModRef.h" |
48 | #include "llvm/Analysis/Loads.h" |
49 | #include "llvm/Analysis/TargetLibraryInfo.h" |
50 | #include "llvm/Analysis/TargetTransformInfo.h" |
51 | #include "llvm/IR/Dominators.h" |
52 | #include "llvm/IR/Function.h" |
53 | #include "llvm/IR/Instruction.h" |
54 | #include "llvm/IR/IRBuilder.h" |
55 | #include "llvm/InitializePasses.h" |
56 | #include "llvm/Pass.h" |
57 | #include "llvm/Transforms/Scalar.h" |
58 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
59 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
60 | #include <algorithm> |
61 | #include <numeric> |
62 | #include <utility> |
63 | #include <vector> |
64 | |
65 | using namespace llvm; |
66 | |
67 | namespace { |
68 | |
69 | #define DEBUG_TYPE "mergeicmps" |
70 | |
71 | // A BCE atom "Binary Compare Expression Atom" represents an integer load |
72 | // that is a constant offset from a base value, e.g. `a` or `o.c` in the example |
73 | // at the top. |
74 | struct BCEAtom { |
75 | BCEAtom() = default; |
76 | BCEAtom(GetElementPtrInst *GEP, LoadInst *LoadI, int BaseId, APInt Offset) |
77 | : GEP(GEP), LoadI(LoadI), BaseId(BaseId), Offset(std::move(Offset)) {} |
78 | |
79 | BCEAtom(const BCEAtom &) = delete; |
80 | BCEAtom &operator=(const BCEAtom &) = delete; |
81 | |
82 | BCEAtom(BCEAtom &&that) = default; |
83 | BCEAtom &operator=(BCEAtom &&that) { |
84 | if (this == &that) |
85 | return *this; |
86 | GEP = that.GEP; |
87 | LoadI = that.LoadI; |
88 | BaseId = that.BaseId; |
89 | Offset = std::move(that.Offset); |
90 | return *this; |
91 | } |
92 | |
93 | // We want to order BCEAtoms by (Base, Offset). However we cannot use |
94 | // the pointer values for Base because these are non-deterministic. |
95 | // To make sure that the sort order is stable, we first assign to each atom |
96 | // base value an index based on its order of appearance in the chain of |
97 | // comparisons. We call this index `BaseOrdering`. For example, for: |
98 | // b[3] == c[2] && a[1] == d[1] && b[4] == c[3] |
99 | // | block 1 | | block 2 | | block 3 | |
100 | // b gets assigned index 0 and a index 1, because b appears as LHS in block 1, |
101 | // which is before block 2. |
102 | // We then sort by (BaseOrdering[LHS.Base()], LHS.Offset), which is stable. |
103 | bool operator<(const BCEAtom &O) const { |
104 | return BaseId != O.BaseId ? BaseId < O.BaseId : Offset.slt(RHS: O.Offset); |
105 | } |
106 | |
107 | GetElementPtrInst *GEP = nullptr; |
108 | LoadInst *LoadI = nullptr; |
109 | unsigned BaseId = 0; |
110 | APInt Offset; |
111 | }; |
112 | |
113 | // A class that assigns increasing ids to values in the order in which they are |
114 | // seen. See comment in `BCEAtom::operator<()``. |
115 | class BaseIdentifier { |
116 | public: |
117 | // Returns the id for value `Base`, after assigning one if `Base` has not been |
118 | // seen before. |
119 | int getBaseId(const Value *Base) { |
120 | assert(Base && "invalid base" ); |
121 | const auto Insertion = BaseToIndex.try_emplace(Key: Base, Args&: Order); |
122 | if (Insertion.second) |
123 | ++Order; |
124 | return Insertion.first->second; |
125 | } |
126 | |
127 | private: |
128 | unsigned Order = 1; |
129 | DenseMap<const Value*, int> BaseToIndex; |
130 | }; |
131 | |
132 | // If this value is a load from a constant offset w.r.t. a base address, and |
133 | // there are no other users of the load or address, returns the base address and |
134 | // the offset. |
135 | BCEAtom visitICmpLoadOperand(Value *const Val, BaseIdentifier &BaseId) { |
136 | auto *const LoadI = dyn_cast<LoadInst>(Val); |
137 | if (!LoadI) |
138 | return {}; |
139 | LLVM_DEBUG(dbgs() << "load\n" ); |
140 | if (LoadI->isUsedOutsideOfBlock(BB: LoadI->getParent())) { |
141 | LLVM_DEBUG(dbgs() << "used outside of block\n" ); |
142 | return {}; |
143 | } |
144 | // Do not optimize atomic loads to non-atomic memcmp |
145 | if (!LoadI->isSimple()) { |
146 | LLVM_DEBUG(dbgs() << "volatile or atomic\n" ); |
147 | return {}; |
148 | } |
149 | Value *Addr = LoadI->getOperand(i_nocapture: 0); |
150 | if (Addr->getType()->getPointerAddressSpace() != 0) { |
151 | LLVM_DEBUG(dbgs() << "from non-zero AddressSpace\n" ); |
152 | return {}; |
153 | } |
154 | const auto &DL = LoadI->getDataLayout(); |
155 | if (!isDereferenceablePointer(V: Addr, Ty: LoadI->getType(), DL)) { |
156 | LLVM_DEBUG(dbgs() << "not dereferenceable\n" ); |
157 | // We need to make sure that we can do comparison in any order, so we |
158 | // require memory to be unconditionally dereferenceable. |
159 | return {}; |
160 | } |
161 | |
162 | APInt Offset = APInt(DL.getIndexTypeSizeInBits(Ty: Addr->getType()), 0); |
163 | Value *Base = Addr; |
164 | auto *GEP = dyn_cast<GetElementPtrInst>(Val: Addr); |
165 | if (GEP) { |
166 | LLVM_DEBUG(dbgs() << "GEP\n" ); |
167 | if (GEP->isUsedOutsideOfBlock(BB: LoadI->getParent())) { |
168 | LLVM_DEBUG(dbgs() << "used outside of block\n" ); |
169 | return {}; |
170 | } |
171 | if (!GEP->accumulateConstantOffset(DL, Offset)) |
172 | return {}; |
173 | Base = GEP->getPointerOperand(); |
174 | } |
175 | return BCEAtom(GEP, LoadI, BaseId.getBaseId(Base), Offset); |
176 | } |
177 | |
178 | // A comparison between two BCE atoms, e.g. `a == o.a` in the example at the |
179 | // top. |
180 | // Note: the terminology is misleading: the comparison is symmetric, so there |
181 | // is no real {l/r}hs. What we want though is to have the same base on the |
182 | // left (resp. right), so that we can detect consecutive loads. To ensure this |
183 | // we put the smallest atom on the left. |
184 | struct BCECmp { |
185 | BCEAtom Lhs; |
186 | BCEAtom Rhs; |
187 | int SizeBits; |
188 | const ICmpInst *CmpI; |
189 | |
190 | BCECmp(BCEAtom L, BCEAtom R, int SizeBits, const ICmpInst *CmpI) |
191 | : Lhs(std::move(L)), Rhs(std::move(R)), SizeBits(SizeBits), CmpI(CmpI) { |
192 | if (Rhs < Lhs) std::swap(a&: Rhs, b&: Lhs); |
193 | } |
194 | }; |
195 | |
196 | // A basic block with a comparison between two BCE atoms. |
197 | // The block might do extra work besides the atom comparison, in which case |
198 | // doesOtherWork() returns true. Under some conditions, the block can be |
199 | // split into the atom comparison part and the "other work" part |
200 | // (see canSplit()). |
201 | class BCECmpBlock { |
202 | public: |
203 | typedef SmallDenseSet<const Instruction *, 8> InstructionSet; |
204 | |
205 | BCECmpBlock(BCECmp Cmp, BasicBlock *BB, InstructionSet BlockInsts) |
206 | : BB(BB), BlockInsts(std::move(BlockInsts)), Cmp(std::move(Cmp)) {} |
207 | |
208 | const BCEAtom &Lhs() const { return Cmp.Lhs; } |
209 | const BCEAtom &Rhs() const { return Cmp.Rhs; } |
210 | int SizeBits() const { return Cmp.SizeBits; } |
211 | |
212 | // Returns true if the block does other works besides comparison. |
213 | bool doesOtherWork() const; |
214 | |
215 | // Returns true if the non-BCE-cmp instructions can be separated from BCE-cmp |
216 | // instructions in the block. |
217 | bool canSplit(AliasAnalysis &AA) const; |
218 | |
219 | // Return true if this all the relevant instructions in the BCE-cmp-block can |
220 | // be sunk below this instruction. By doing this, we know we can separate the |
221 | // BCE-cmp-block instructions from the non-BCE-cmp-block instructions in the |
222 | // block. |
223 | bool canSinkBCECmpInst(const Instruction *, AliasAnalysis &AA) const; |
224 | |
225 | // We can separate the BCE-cmp-block instructions and the non-BCE-cmp-block |
226 | // instructions. Split the old block and move all non-BCE-cmp-insts into the |
227 | // new parent block. |
228 | void split(BasicBlock *NewParent, AliasAnalysis &AA) const; |
229 | |
230 | // The basic block where this comparison happens. |
231 | BasicBlock *BB; |
232 | // Instructions relating to the BCECmp and branch. |
233 | InstructionSet BlockInsts; |
234 | // The block requires splitting. |
235 | bool RequireSplit = false; |
236 | // Original order of this block in the chain. |
237 | unsigned OrigOrder = 0; |
238 | |
239 | private: |
240 | BCECmp Cmp; |
241 | }; |
242 | |
243 | bool BCECmpBlock::canSinkBCECmpInst(const Instruction *Inst, |
244 | AliasAnalysis &AA) const { |
245 | // If this instruction may clobber the loads and is in middle of the BCE cmp |
246 | // block instructions, then bail for now. |
247 | if (Inst->mayWriteToMemory()) { |
248 | auto MayClobber = [&](LoadInst *LI) { |
249 | // If a potentially clobbering instruction comes before the load, |
250 | // we can still safely sink the load. |
251 | return (Inst->getParent() != LI->getParent() || !Inst->comesBefore(Other: LI)) && |
252 | isModSet(MRI: AA.getModRefInfo(I: Inst, OptLoc: MemoryLocation::get(LI))); |
253 | }; |
254 | if (MayClobber(Cmp.Lhs.LoadI) || MayClobber(Cmp.Rhs.LoadI)) |
255 | return false; |
256 | } |
257 | // Make sure this instruction does not use any of the BCE cmp block |
258 | // instructions as operand. |
259 | return llvm::none_of(Range: Inst->operands(), P: [&](const Value *Op) { |
260 | const Instruction *OpI = dyn_cast<Instruction>(Val: Op); |
261 | return OpI && BlockInsts.contains(V: OpI); |
262 | }); |
263 | } |
264 | |
265 | void BCECmpBlock::split(BasicBlock *NewParent, AliasAnalysis &AA) const { |
266 | llvm::SmallVector<Instruction *, 4> OtherInsts; |
267 | for (Instruction &Inst : *BB) { |
268 | if (BlockInsts.count(V: &Inst)) |
269 | continue; |
270 | assert(canSinkBCECmpInst(&Inst, AA) && "Split unsplittable block" ); |
271 | // This is a non-BCE-cmp-block instruction. And it can be separated |
272 | // from the BCE-cmp-block instruction. |
273 | OtherInsts.push_back(Elt: &Inst); |
274 | } |
275 | |
276 | // Do the actual spliting. |
277 | for (Instruction *Inst : reverse(C&: OtherInsts)) |
278 | Inst->moveBeforePreserving(BB&: *NewParent, I: NewParent->begin()); |
279 | } |
280 | |
281 | bool BCECmpBlock::canSplit(AliasAnalysis &AA) const { |
282 | for (Instruction &Inst : *BB) { |
283 | if (!BlockInsts.count(V: &Inst)) { |
284 | if (!canSinkBCECmpInst(Inst: &Inst, AA)) |
285 | return false; |
286 | } |
287 | } |
288 | return true; |
289 | } |
290 | |
291 | bool BCECmpBlock::doesOtherWork() const { |
292 | // TODO(courbet): Can we allow some other things ? This is very conservative. |
293 | // We might be able to get away with anything does not have any side |
294 | // effects outside of the basic block. |
295 | // Note: The GEPs and/or loads are not necessarily in the same block. |
296 | for (const Instruction &Inst : *BB) { |
297 | if (!BlockInsts.count(V: &Inst)) |
298 | return true; |
299 | } |
300 | return false; |
301 | } |
302 | |
303 | // Visit the given comparison. If this is a comparison between two valid |
304 | // BCE atoms, returns the comparison. |
305 | std::optional<BCECmp> visitICmp(const ICmpInst *const CmpI, |
306 | const ICmpInst::Predicate ExpectedPredicate, |
307 | BaseIdentifier &BaseId) { |
308 | // The comparison can only be used once: |
309 | // - For intermediate blocks, as a branch condition. |
310 | // - For the final block, as an incoming value for the Phi. |
311 | // If there are any other uses of the comparison, we cannot merge it with |
312 | // other comparisons as we would create an orphan use of the value. |
313 | if (!CmpI->hasOneUse()) { |
314 | LLVM_DEBUG(dbgs() << "cmp has several uses\n" ); |
315 | return std::nullopt; |
316 | } |
317 | if (CmpI->getPredicate() != ExpectedPredicate) |
318 | return std::nullopt; |
319 | LLVM_DEBUG(dbgs() << "cmp " |
320 | << (ExpectedPredicate == ICmpInst::ICMP_EQ ? "eq" : "ne" ) |
321 | << "\n" ); |
322 | auto Lhs = visitICmpLoadOperand(Val: CmpI->getOperand(i_nocapture: 0), BaseId); |
323 | if (!Lhs.BaseId) |
324 | return std::nullopt; |
325 | auto Rhs = visitICmpLoadOperand(Val: CmpI->getOperand(i_nocapture: 1), BaseId); |
326 | if (!Rhs.BaseId) |
327 | return std::nullopt; |
328 | const auto &DL = CmpI->getDataLayout(); |
329 | return BCECmp(std::move(Lhs), std::move(Rhs), |
330 | DL.getTypeSizeInBits(Ty: CmpI->getOperand(i_nocapture: 0)->getType()), CmpI); |
331 | } |
332 | |
333 | // Visit the given comparison block. If this is a comparison between two valid |
334 | // BCE atoms, returns the comparison. |
335 | std::optional<BCECmpBlock> visitCmpBlock(Value *const Val, |
336 | BasicBlock *const Block, |
337 | const BasicBlock *const PhiBlock, |
338 | BaseIdentifier &BaseId) { |
339 | if (Block->empty()) |
340 | return std::nullopt; |
341 | auto *const BranchI = dyn_cast<BranchInst>(Val: Block->getTerminator()); |
342 | if (!BranchI) |
343 | return std::nullopt; |
344 | LLVM_DEBUG(dbgs() << "branch\n" ); |
345 | Value *Cond; |
346 | ICmpInst::Predicate ExpectedPredicate; |
347 | if (BranchI->isUnconditional()) { |
348 | // In this case, we expect an incoming value which is the result of the |
349 | // comparison. This is the last link in the chain of comparisons (note |
350 | // that this does not mean that this is the last incoming value, blocks |
351 | // can be reordered). |
352 | Cond = Val; |
353 | ExpectedPredicate = ICmpInst::ICMP_EQ; |
354 | } else { |
355 | // In this case, we expect a constant incoming value (the comparison is |
356 | // chained). |
357 | const auto *const Const = cast<ConstantInt>(Val); |
358 | LLVM_DEBUG(dbgs() << "const\n" ); |
359 | if (!Const->isZero()) |
360 | return std::nullopt; |
361 | LLVM_DEBUG(dbgs() << "false\n" ); |
362 | assert(BranchI->getNumSuccessors() == 2 && "expecting a cond branch" ); |
363 | BasicBlock *const FalseBlock = BranchI->getSuccessor(i: 1); |
364 | Cond = BranchI->getCondition(); |
365 | ExpectedPredicate = |
366 | FalseBlock == PhiBlock ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; |
367 | } |
368 | |
369 | auto *CmpI = dyn_cast<ICmpInst>(Val: Cond); |
370 | if (!CmpI) |
371 | return std::nullopt; |
372 | LLVM_DEBUG(dbgs() << "icmp\n" ); |
373 | |
374 | std::optional<BCECmp> Result = visitICmp(CmpI, ExpectedPredicate, BaseId); |
375 | if (!Result) |
376 | return std::nullopt; |
377 | |
378 | BCECmpBlock::InstructionSet BlockInsts( |
379 | {Result->Lhs.LoadI, Result->Rhs.LoadI, Result->CmpI, BranchI}); |
380 | if (Result->Lhs.GEP) |
381 | BlockInsts.insert(V: Result->Lhs.GEP); |
382 | if (Result->Rhs.GEP) |
383 | BlockInsts.insert(V: Result->Rhs.GEP); |
384 | return BCECmpBlock(std::move(*Result), Block, BlockInsts); |
385 | } |
386 | |
387 | static inline void enqueueBlock(std::vector<BCECmpBlock> &Comparisons, |
388 | BCECmpBlock &&Comparison) { |
389 | LLVM_DEBUG(dbgs() << "Block '" << Comparison.BB->getName() |
390 | << "': Found cmp of " << Comparison.SizeBits() |
391 | << " bits between " << Comparison.Lhs().BaseId << " + " |
392 | << Comparison.Lhs().Offset << " and " |
393 | << Comparison.Rhs().BaseId << " + " |
394 | << Comparison.Rhs().Offset << "\n" ); |
395 | LLVM_DEBUG(dbgs() << "\n" ); |
396 | Comparison.OrigOrder = Comparisons.size(); |
397 | Comparisons.push_back(x: std::move(Comparison)); |
398 | } |
399 | |
400 | // A chain of comparisons. |
401 | class BCECmpChain { |
402 | public: |
403 | using ContiguousBlocks = std::vector<BCECmpBlock>; |
404 | |
405 | BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi, |
406 | AliasAnalysis &AA); |
407 | |
408 | bool simplify(const TargetLibraryInfo &TLI, AliasAnalysis &AA, |
409 | DomTreeUpdater &DTU); |
410 | |
411 | bool atLeastOneMerged() const { |
412 | return any_of(Range: MergedBlocks_, |
413 | P: [](const auto &Blocks) { return Blocks.size() > 1; }); |
414 | } |
415 | |
416 | private: |
417 | PHINode &Phi_; |
418 | // The list of all blocks in the chain, grouped by contiguity. |
419 | std::vector<ContiguousBlocks> MergedBlocks_; |
420 | // The original entry block (before sorting); |
421 | BasicBlock *EntryBlock_; |
422 | }; |
423 | |
424 | static bool areContiguous(const BCECmpBlock &First, const BCECmpBlock &Second) { |
425 | return First.Lhs().BaseId == Second.Lhs().BaseId && |
426 | First.Rhs().BaseId == Second.Rhs().BaseId && |
427 | First.Lhs().Offset + First.SizeBits() / 8 == Second.Lhs().Offset && |
428 | First.Rhs().Offset + First.SizeBits() / 8 == Second.Rhs().Offset; |
429 | } |
430 | |
431 | static unsigned getMinOrigOrder(const BCECmpChain::ContiguousBlocks &Blocks) { |
432 | unsigned MinOrigOrder = std::numeric_limits<unsigned>::max(); |
433 | for (const BCECmpBlock &Block : Blocks) |
434 | MinOrigOrder = std::min(a: MinOrigOrder, b: Block.OrigOrder); |
435 | return MinOrigOrder; |
436 | } |
437 | |
438 | /// Given a chain of comparison blocks, groups the blocks into contiguous |
439 | /// ranges that can be merged together into a single comparison. |
440 | static std::vector<BCECmpChain::ContiguousBlocks> |
441 | mergeBlocks(std::vector<BCECmpBlock> &&Blocks) { |
442 | std::vector<BCECmpChain::ContiguousBlocks> MergedBlocks; |
443 | |
444 | // Sort to detect continuous offsets. |
445 | llvm::sort(C&: Blocks, |
446 | Comp: [](const BCECmpBlock &LhsBlock, const BCECmpBlock &RhsBlock) { |
447 | return std::tie(args: LhsBlock.Lhs(), args: LhsBlock.Rhs()) < |
448 | std::tie(args: RhsBlock.Lhs(), args: RhsBlock.Rhs()); |
449 | }); |
450 | |
451 | BCECmpChain::ContiguousBlocks *LastMergedBlock = nullptr; |
452 | for (BCECmpBlock &Block : Blocks) { |
453 | if (!LastMergedBlock || !areContiguous(First: LastMergedBlock->back(), Second: Block)) { |
454 | MergedBlocks.emplace_back(); |
455 | LastMergedBlock = &MergedBlocks.back(); |
456 | } else { |
457 | LLVM_DEBUG(dbgs() << "Merging block " << Block.BB->getName() << " into " |
458 | << LastMergedBlock->back().BB->getName() << "\n" ); |
459 | } |
460 | LastMergedBlock->push_back(x: std::move(Block)); |
461 | } |
462 | |
463 | // While we allow reordering for merging, do not reorder unmerged comparisons. |
464 | // Doing so may introduce branch on poison. |
465 | llvm::sort(C&: MergedBlocks, Comp: [](const BCECmpChain::ContiguousBlocks &LhsBlocks, |
466 | const BCECmpChain::ContiguousBlocks &RhsBlocks) { |
467 | return getMinOrigOrder(Blocks: LhsBlocks) < getMinOrigOrder(Blocks: RhsBlocks); |
468 | }); |
469 | |
470 | return MergedBlocks; |
471 | } |
472 | |
473 | BCECmpChain::BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi, |
474 | AliasAnalysis &AA) |
475 | : Phi_(Phi) { |
476 | assert(!Blocks.empty() && "a chain should have at least one block" ); |
477 | // Now look inside blocks to check for BCE comparisons. |
478 | std::vector<BCECmpBlock> Comparisons; |
479 | BaseIdentifier BaseId; |
480 | for (BasicBlock *const Block : Blocks) { |
481 | assert(Block && "invalid block" ); |
482 | std::optional<BCECmpBlock> Comparison = visitCmpBlock( |
483 | Val: Phi.getIncomingValueForBlock(BB: Block), Block, PhiBlock: Phi.getParent(), BaseId); |
484 | if (!Comparison) { |
485 | LLVM_DEBUG(dbgs() << "chain with invalid BCECmpBlock, no merge.\n" ); |
486 | return; |
487 | } |
488 | if (Comparison->doesOtherWork()) { |
489 | LLVM_DEBUG(dbgs() << "block '" << Comparison->BB->getName() |
490 | << "' does extra work besides compare\n" ); |
491 | if (Comparisons.empty()) { |
492 | // This is the initial block in the chain, in case this block does other |
493 | // work, we can try to split the block and move the irrelevant |
494 | // instructions to the predecessor. |
495 | // |
496 | // If this is not the initial block in the chain, splitting it wont |
497 | // work. |
498 | // |
499 | // As once split, there will still be instructions before the BCE cmp |
500 | // instructions that do other work in program order, i.e. within the |
501 | // chain before sorting. Unless we can abort the chain at this point |
502 | // and start anew. |
503 | // |
504 | // NOTE: we only handle blocks a with single predecessor for now. |
505 | if (Comparison->canSplit(AA)) { |
506 | LLVM_DEBUG(dbgs() |
507 | << "Split initial block '" << Comparison->BB->getName() |
508 | << "' that does extra work besides compare\n" ); |
509 | Comparison->RequireSplit = true; |
510 | enqueueBlock(Comparisons, Comparison: std::move(*Comparison)); |
511 | } else { |
512 | LLVM_DEBUG(dbgs() |
513 | << "ignoring initial block '" << Comparison->BB->getName() |
514 | << "' that does extra work besides compare\n" ); |
515 | } |
516 | continue; |
517 | } |
518 | // TODO(courbet): Right now we abort the whole chain. We could be |
519 | // merging only the blocks that don't do other work and resume the |
520 | // chain from there. For example: |
521 | // if (a[0] == b[0]) { // bb1 |
522 | // if (a[1] == b[1]) { // bb2 |
523 | // some_value = 3; //bb3 |
524 | // if (a[2] == b[2]) { //bb3 |
525 | // do a ton of stuff //bb4 |
526 | // } |
527 | // } |
528 | // } |
529 | // |
530 | // This is: |
531 | // |
532 | // bb1 --eq--> bb2 --eq--> bb3* -eq--> bb4 --+ |
533 | // \ \ \ \ |
534 | // ne ne ne \ |
535 | // \ \ \ v |
536 | // +------------+-----------+----------> bb_phi |
537 | // |
538 | // We can only merge the first two comparisons, because bb3* does |
539 | // "other work" (setting some_value to 3). |
540 | // We could still merge bb1 and bb2 though. |
541 | return; |
542 | } |
543 | enqueueBlock(Comparisons, Comparison: std::move(*Comparison)); |
544 | } |
545 | |
546 | // It is possible we have no suitable comparison to merge. |
547 | if (Comparisons.empty()) { |
548 | LLVM_DEBUG(dbgs() << "chain with no BCE basic blocks, no merge\n" ); |
549 | return; |
550 | } |
551 | EntryBlock_ = Comparisons[0].BB; |
552 | MergedBlocks_ = mergeBlocks(Blocks: std::move(Comparisons)); |
553 | } |
554 | |
555 | namespace { |
556 | |
557 | // A class to compute the name of a set of merged basic blocks. |
558 | // This is optimized for the common case of no block names. |
559 | class MergedBlockName { |
560 | // Storage for the uncommon case of several named blocks. |
561 | SmallString<16> Scratch; |
562 | |
563 | public: |
564 | explicit MergedBlockName(ArrayRef<BCECmpBlock> Comparisons) |
565 | : Name(makeName(Comparisons)) {} |
566 | const StringRef Name; |
567 | |
568 | private: |
569 | StringRef makeName(ArrayRef<BCECmpBlock> Comparisons) { |
570 | assert(!Comparisons.empty() && "no basic block" ); |
571 | // Fast path: only one block, or no names at all. |
572 | if (Comparisons.size() == 1) |
573 | return Comparisons[0].BB->getName(); |
574 | const int size = std::accumulate(first: Comparisons.begin(), last: Comparisons.end(), init: 0, |
575 | binary_op: [](int i, const BCECmpBlock &Cmp) { |
576 | return i + Cmp.BB->getName().size(); |
577 | }); |
578 | if (size == 0) |
579 | return StringRef("" , 0); |
580 | |
581 | // Slow path: at least two blocks, at least one block with a name. |
582 | Scratch.clear(); |
583 | // We'll have `size` bytes for name and `Comparisons.size() - 1` bytes for |
584 | // separators. |
585 | Scratch.reserve(N: size + Comparisons.size() - 1); |
586 | const auto append = [this](StringRef str) { |
587 | Scratch.append(in_start: str.begin(), in_end: str.end()); |
588 | }; |
589 | append(Comparisons[0].BB->getName()); |
590 | for (int I = 1, E = Comparisons.size(); I < E; ++I) { |
591 | const BasicBlock *const BB = Comparisons[I].BB; |
592 | if (!BB->getName().empty()) { |
593 | append("+" ); |
594 | append(BB->getName()); |
595 | } |
596 | } |
597 | return Scratch.str(); |
598 | } |
599 | }; |
600 | } // namespace |
601 | |
602 | // Merges the given contiguous comparison blocks into one memcmp block. |
603 | static BasicBlock *mergeComparisons(ArrayRef<BCECmpBlock> Comparisons, |
604 | BasicBlock *const InsertBefore, |
605 | BasicBlock *const NextCmpBlock, |
606 | PHINode &Phi, const TargetLibraryInfo &TLI, |
607 | AliasAnalysis &AA, DomTreeUpdater &DTU) { |
608 | assert(!Comparisons.empty() && "merging zero comparisons" ); |
609 | LLVMContext &Context = NextCmpBlock->getContext(); |
610 | const BCECmpBlock &FirstCmp = Comparisons[0]; |
611 | |
612 | // Create a new cmp block before next cmp block. |
613 | BasicBlock *const BB = |
614 | BasicBlock::Create(Context, Name: MergedBlockName(Comparisons).Name, |
615 | Parent: NextCmpBlock->getParent(), InsertBefore); |
616 | IRBuilder<> Builder(BB); |
617 | // Add the GEPs from the first BCECmpBlock. |
618 | Value *Lhs, *Rhs; |
619 | if (FirstCmp.Lhs().GEP) |
620 | Lhs = Builder.Insert(I: FirstCmp.Lhs().GEP->clone()); |
621 | else |
622 | Lhs = FirstCmp.Lhs().LoadI->getPointerOperand(); |
623 | if (FirstCmp.Rhs().GEP) |
624 | Rhs = Builder.Insert(I: FirstCmp.Rhs().GEP->clone()); |
625 | else |
626 | Rhs = FirstCmp.Rhs().LoadI->getPointerOperand(); |
627 | |
628 | Value *IsEqual = nullptr; |
629 | LLVM_DEBUG(dbgs() << "Merging " << Comparisons.size() << " comparisons -> " |
630 | << BB->getName() << "\n" ); |
631 | |
632 | // If there is one block that requires splitting, we do it now, i.e. |
633 | // just before we know we will collapse the chain. The instructions |
634 | // can be executed before any of the instructions in the chain. |
635 | const auto ToSplit = llvm::find_if( |
636 | Range&: Comparisons, P: [](const BCECmpBlock &B) { return B.RequireSplit; }); |
637 | if (ToSplit != Comparisons.end()) { |
638 | LLVM_DEBUG(dbgs() << "Splitting non_BCE work to header\n" ); |
639 | ToSplit->split(NewParent: BB, AA); |
640 | } |
641 | |
642 | if (Comparisons.size() == 1) { |
643 | LLVM_DEBUG(dbgs() << "Only one comparison, updating branches\n" ); |
644 | // Use clone to keep the metadata |
645 | Instruction *const LhsLoad = Builder.Insert(I: FirstCmp.Lhs().LoadI->clone()); |
646 | Instruction *const RhsLoad = Builder.Insert(I: FirstCmp.Rhs().LoadI->clone()); |
647 | LhsLoad->replaceUsesOfWith(From: LhsLoad->getOperand(i: 0), To: Lhs); |
648 | RhsLoad->replaceUsesOfWith(From: RhsLoad->getOperand(i: 0), To: Rhs); |
649 | // There are no blocks to merge, just do the comparison. |
650 | IsEqual = Builder.CreateICmpEQ(LHS: LhsLoad, RHS: RhsLoad); |
651 | } else { |
652 | const unsigned TotalSizeBits = std::accumulate( |
653 | first: Comparisons.begin(), last: Comparisons.end(), init: 0u, |
654 | binary_op: [](int Size, const BCECmpBlock &C) { return Size + C.SizeBits(); }); |
655 | |
656 | // memcmp expects a 'size_t' argument and returns 'int'. |
657 | unsigned SizeTBits = TLI.getSizeTSize(M: *Phi.getModule()); |
658 | unsigned IntBits = TLI.getIntSize(); |
659 | |
660 | // Create memcmp() == 0. |
661 | const auto &DL = Phi.getDataLayout(); |
662 | Value *const MemCmpCall = emitMemCmp( |
663 | Ptr1: Lhs, Ptr2: Rhs, |
664 | Len: ConstantInt::get(Ty: Builder.getIntNTy(N: SizeTBits), V: TotalSizeBits / 8), |
665 | B&: Builder, DL, TLI: &TLI); |
666 | IsEqual = Builder.CreateICmpEQ( |
667 | LHS: MemCmpCall, RHS: ConstantInt::get(Ty: Builder.getIntNTy(N: IntBits), V: 0)); |
668 | } |
669 | |
670 | BasicBlock *const PhiBB = Phi.getParent(); |
671 | // Add a branch to the next basic block in the chain. |
672 | if (NextCmpBlock == PhiBB) { |
673 | // Continue to phi, passing it the comparison result. |
674 | Builder.CreateBr(Dest: PhiBB); |
675 | Phi.addIncoming(V: IsEqual, BB); |
676 | DTU.applyUpdates(Updates: {{DominatorTree::Insert, BB, PhiBB}}); |
677 | } else { |
678 | // Continue to next block if equal, exit to phi else. |
679 | Builder.CreateCondBr(Cond: IsEqual, True: NextCmpBlock, False: PhiBB); |
680 | Phi.addIncoming(V: ConstantInt::getFalse(Context), BB); |
681 | DTU.applyUpdates(Updates: {{DominatorTree::Insert, BB, NextCmpBlock}, |
682 | {DominatorTree::Insert, BB, PhiBB}}); |
683 | } |
684 | return BB; |
685 | } |
686 | |
687 | bool BCECmpChain::simplify(const TargetLibraryInfo &TLI, AliasAnalysis &AA, |
688 | DomTreeUpdater &DTU) { |
689 | assert(atLeastOneMerged() && "simplifying trivial BCECmpChain" ); |
690 | LLVM_DEBUG(dbgs() << "Simplifying comparison chain starting at block " |
691 | << EntryBlock_->getName() << "\n" ); |
692 | |
693 | // Effectively merge blocks. We go in the reverse direction from the phi block |
694 | // so that the next block is always available to branch to. |
695 | BasicBlock *InsertBefore = EntryBlock_; |
696 | BasicBlock *NextCmpBlock = Phi_.getParent(); |
697 | for (const auto &Blocks : reverse(C&: MergedBlocks_)) { |
698 | InsertBefore = NextCmpBlock = mergeComparisons( |
699 | Comparisons: Blocks, InsertBefore, NextCmpBlock, Phi&: Phi_, TLI, AA, DTU); |
700 | } |
701 | |
702 | // Replace the original cmp chain with the new cmp chain by pointing all |
703 | // predecessors of EntryBlock_ to NextCmpBlock instead. This makes all cmp |
704 | // blocks in the old chain unreachable. |
705 | while (!pred_empty(BB: EntryBlock_)) { |
706 | BasicBlock* const Pred = *pred_begin(BB: EntryBlock_); |
707 | LLVM_DEBUG(dbgs() << "Updating jump into old chain from " << Pred->getName() |
708 | << "\n" ); |
709 | Pred->getTerminator()->replaceUsesOfWith(From: EntryBlock_, To: NextCmpBlock); |
710 | DTU.applyUpdates(Updates: {{DominatorTree::Delete, Pred, EntryBlock_}, |
711 | {DominatorTree::Insert, Pred, NextCmpBlock}}); |
712 | } |
713 | |
714 | // If the old cmp chain was the function entry, we need to update the function |
715 | // entry. |
716 | const bool ChainEntryIsFnEntry = EntryBlock_->isEntryBlock(); |
717 | if (ChainEntryIsFnEntry && DTU.hasDomTree()) { |
718 | LLVM_DEBUG(dbgs() << "Changing function entry from " |
719 | << EntryBlock_->getName() << " to " |
720 | << NextCmpBlock->getName() << "\n" ); |
721 | DTU.getDomTree().setNewRoot(NextCmpBlock); |
722 | DTU.applyUpdates(Updates: {{DominatorTree::Delete, NextCmpBlock, EntryBlock_}}); |
723 | } |
724 | EntryBlock_ = nullptr; |
725 | |
726 | // Delete merged blocks. This also removes incoming values in phi. |
727 | SmallVector<BasicBlock *, 16> DeadBlocks; |
728 | for (const auto &Blocks : MergedBlocks_) { |
729 | for (const BCECmpBlock &Block : Blocks) { |
730 | LLVM_DEBUG(dbgs() << "Deleting merged block " << Block.BB->getName() |
731 | << "\n" ); |
732 | DeadBlocks.push_back(Elt: Block.BB); |
733 | } |
734 | } |
735 | DeleteDeadBlocks(BBs: DeadBlocks, DTU: &DTU); |
736 | |
737 | MergedBlocks_.clear(); |
738 | return true; |
739 | } |
740 | |
741 | std::vector<BasicBlock *> getOrderedBlocks(PHINode &Phi, |
742 | BasicBlock *const LastBlock, |
743 | int NumBlocks) { |
744 | // Walk up from the last block to find other blocks. |
745 | std::vector<BasicBlock *> Blocks(NumBlocks); |
746 | assert(LastBlock && "invalid last block" ); |
747 | BasicBlock *CurBlock = LastBlock; |
748 | for (int BlockIndex = NumBlocks - 1; BlockIndex > 0; --BlockIndex) { |
749 | if (CurBlock->hasAddressTaken()) { |
750 | // Somebody is jumping to the block through an address, all bets are |
751 | // off. |
752 | LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex |
753 | << " has its address taken\n" ); |
754 | return {}; |
755 | } |
756 | Blocks[BlockIndex] = CurBlock; |
757 | auto *SinglePredecessor = CurBlock->getSinglePredecessor(); |
758 | if (!SinglePredecessor) { |
759 | // The block has two or more predecessors. |
760 | LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex |
761 | << " has two or more predecessors\n" ); |
762 | return {}; |
763 | } |
764 | if (Phi.getBasicBlockIndex(BB: SinglePredecessor) < 0) { |
765 | // The block does not link back to the phi. |
766 | LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex |
767 | << " does not link back to the phi\n" ); |
768 | return {}; |
769 | } |
770 | CurBlock = SinglePredecessor; |
771 | } |
772 | Blocks[0] = CurBlock; |
773 | return Blocks; |
774 | } |
775 | |
776 | bool processPhi(PHINode &Phi, const TargetLibraryInfo &TLI, AliasAnalysis &AA, |
777 | DomTreeUpdater &DTU) { |
778 | LLVM_DEBUG(dbgs() << "processPhi()\n" ); |
779 | if (Phi.getNumIncomingValues() <= 1) { |
780 | LLVM_DEBUG(dbgs() << "skip: only one incoming value in phi\n" ); |
781 | return false; |
782 | } |
783 | // We are looking for something that has the following structure: |
784 | // bb1 --eq--> bb2 --eq--> bb3 --eq--> bb4 --+ |
785 | // \ \ \ \ |
786 | // ne ne ne \ |
787 | // \ \ \ v |
788 | // +------------+-----------+----------> bb_phi |
789 | // |
790 | // - The last basic block (bb4 here) must branch unconditionally to bb_phi. |
791 | // It's the only block that contributes a non-constant value to the Phi. |
792 | // - All other blocks (b1, b2, b3) must have exactly two successors, one of |
793 | // them being the phi block. |
794 | // - All intermediate blocks (bb2, bb3) must have only one predecessor. |
795 | // - Blocks cannot do other work besides the comparison, see doesOtherWork() |
796 | |
797 | // The blocks are not necessarily ordered in the phi, so we start from the |
798 | // last block and reconstruct the order. |
799 | BasicBlock *LastBlock = nullptr; |
800 | for (unsigned I = 0; I < Phi.getNumIncomingValues(); ++I) { |
801 | if (isa<ConstantInt>(Val: Phi.getIncomingValue(i: I))) continue; |
802 | if (LastBlock) { |
803 | // There are several non-constant values. |
804 | LLVM_DEBUG(dbgs() << "skip: several non-constant values\n" ); |
805 | return false; |
806 | } |
807 | if (!isa<ICmpInst>(Val: Phi.getIncomingValue(i: I)) || |
808 | cast<ICmpInst>(Val: Phi.getIncomingValue(i: I))->getParent() != |
809 | Phi.getIncomingBlock(i: I)) { |
810 | // Non-constant incoming value is not from a cmp instruction or not |
811 | // produced by the last block. We could end up processing the value |
812 | // producing block more than once. |
813 | // |
814 | // This is an uncommon case, so we bail. |
815 | LLVM_DEBUG( |
816 | dbgs() |
817 | << "skip: non-constant value not from cmp or not from last block.\n" ); |
818 | return false; |
819 | } |
820 | LastBlock = Phi.getIncomingBlock(i: I); |
821 | } |
822 | if (!LastBlock) { |
823 | // There is no non-constant block. |
824 | LLVM_DEBUG(dbgs() << "skip: no non-constant block\n" ); |
825 | return false; |
826 | } |
827 | if (LastBlock->getSingleSuccessor() != Phi.getParent()) { |
828 | LLVM_DEBUG(dbgs() << "skip: last block non-phi successor\n" ); |
829 | return false; |
830 | } |
831 | |
832 | const auto Blocks = |
833 | getOrderedBlocks(Phi, LastBlock, NumBlocks: Phi.getNumIncomingValues()); |
834 | if (Blocks.empty()) return false; |
835 | BCECmpChain CmpChain(Blocks, Phi, AA); |
836 | |
837 | if (!CmpChain.atLeastOneMerged()) { |
838 | LLVM_DEBUG(dbgs() << "skip: nothing merged\n" ); |
839 | return false; |
840 | } |
841 | |
842 | return CmpChain.simplify(TLI, AA, DTU); |
843 | } |
844 | |
845 | static bool runImpl(Function &F, const TargetLibraryInfo &TLI, |
846 | const TargetTransformInfo &TTI, AliasAnalysis &AA, |
847 | DominatorTree *DT) { |
848 | LLVM_DEBUG(dbgs() << "MergeICmpsLegacyPass: " << F.getName() << "\n" ); |
849 | |
850 | // We only try merging comparisons if the target wants to expand memcmp later. |
851 | // The rationale is to avoid turning small chains into memcmp calls. |
852 | if (!TTI.enableMemCmpExpansion(OptSize: F.hasOptSize(), IsZeroCmp: true)) |
853 | return false; |
854 | |
855 | // If we don't have memcmp avaiable we can't emit calls to it. |
856 | if (!TLI.has(F: LibFunc_memcmp)) |
857 | return false; |
858 | |
859 | DomTreeUpdater DTU(DT, /*PostDominatorTree*/ nullptr, |
860 | DomTreeUpdater::UpdateStrategy::Eager); |
861 | |
862 | bool MadeChange = false; |
863 | |
864 | for (BasicBlock &BB : llvm::drop_begin(RangeOrContainer&: F)) { |
865 | // A Phi operation is always first in a basic block. |
866 | if (auto *const Phi = dyn_cast<PHINode>(Val: &*BB.begin())) |
867 | MadeChange |= processPhi(Phi&: *Phi, TLI, AA, DTU); |
868 | } |
869 | |
870 | return MadeChange; |
871 | } |
872 | |
873 | class MergeICmpsLegacyPass : public FunctionPass { |
874 | public: |
875 | static char ID; |
876 | |
877 | MergeICmpsLegacyPass() : FunctionPass(ID) { |
878 | initializeMergeICmpsLegacyPassPass(*PassRegistry::getPassRegistry()); |
879 | } |
880 | |
881 | bool runOnFunction(Function &F) override { |
882 | if (skipFunction(F)) return false; |
883 | const auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
884 | const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
885 | // MergeICmps does not need the DominatorTree, but we update it if it's |
886 | // already available. |
887 | auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>(); |
888 | auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
889 | return runImpl(F, TLI, TTI, AA, DT: DTWP ? &DTWP->getDomTree() : nullptr); |
890 | } |
891 | |
892 | private: |
893 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
894 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
895 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
896 | AU.addRequired<AAResultsWrapperPass>(); |
897 | AU.addPreserved<GlobalsAAWrapperPass>(); |
898 | AU.addPreserved<DominatorTreeWrapperPass>(); |
899 | } |
900 | }; |
901 | |
902 | } // namespace |
903 | |
904 | char MergeICmpsLegacyPass::ID = 0; |
905 | INITIALIZE_PASS_BEGIN(MergeICmpsLegacyPass, "mergeicmps" , |
906 | "Merge contiguous icmps into a memcmp" , false, false) |
907 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
908 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
909 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
910 | INITIALIZE_PASS_END(MergeICmpsLegacyPass, "mergeicmps" , |
911 | "Merge contiguous icmps into a memcmp" , false, false) |
912 | |
913 | Pass *llvm::createMergeICmpsLegacyPass() { return new MergeICmpsLegacyPass(); } |
914 | |
915 | PreservedAnalyses MergeICmpsPass::run(Function &F, |
916 | FunctionAnalysisManager &AM) { |
917 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
918 | auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F); |
919 | auto &AA = AM.getResult<AAManager>(IR&: F); |
920 | auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(IR&: F); |
921 | const bool MadeChanges = runImpl(F, TLI, TTI, AA, DT); |
922 | if (!MadeChanges) |
923 | return PreservedAnalyses::all(); |
924 | PreservedAnalyses PA; |
925 | PA.preserve<DominatorTreeAnalysis>(); |
926 | return PA; |
927 | } |
928 | |