1 | //===- HexagonLoopIdiomRecognition.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 | #include "HexagonLoopIdiomRecognition.h" |
10 | #include "llvm/ADT/APInt.h" |
11 | #include "llvm/ADT/DenseMap.h" |
12 | #include "llvm/ADT/SetVector.h" |
13 | #include "llvm/ADT/SmallPtrSet.h" |
14 | #include "llvm/ADT/SmallSet.h" |
15 | #include "llvm/ADT/SmallVector.h" |
16 | #include "llvm/ADT/StringRef.h" |
17 | #include "llvm/Analysis/AliasAnalysis.h" |
18 | #include "llvm/Analysis/InstructionSimplify.h" |
19 | #include "llvm/Analysis/LoopAnalysisManager.h" |
20 | #include "llvm/Analysis/LoopInfo.h" |
21 | #include "llvm/Analysis/LoopPass.h" |
22 | #include "llvm/Analysis/MemoryLocation.h" |
23 | #include "llvm/Analysis/ScalarEvolution.h" |
24 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
25 | #include "llvm/Analysis/TargetLibraryInfo.h" |
26 | #include "llvm/Analysis/ValueTracking.h" |
27 | #include "llvm/IR/Attributes.h" |
28 | #include "llvm/IR/BasicBlock.h" |
29 | #include "llvm/IR/Constant.h" |
30 | #include "llvm/IR/Constants.h" |
31 | #include "llvm/IR/DataLayout.h" |
32 | #include "llvm/IR/DebugLoc.h" |
33 | #include "llvm/IR/DerivedTypes.h" |
34 | #include "llvm/IR/Dominators.h" |
35 | #include "llvm/IR/Function.h" |
36 | #include "llvm/IR/IRBuilder.h" |
37 | #include "llvm/IR/InstrTypes.h" |
38 | #include "llvm/IR/Instruction.h" |
39 | #include "llvm/IR/Instructions.h" |
40 | #include "llvm/IR/IntrinsicInst.h" |
41 | #include "llvm/IR/Intrinsics.h" |
42 | #include "llvm/IR/IntrinsicsHexagon.h" |
43 | #include "llvm/IR/Module.h" |
44 | #include "llvm/IR/PassManager.h" |
45 | #include "llvm/IR/PatternMatch.h" |
46 | #include "llvm/IR/Type.h" |
47 | #include "llvm/IR/User.h" |
48 | #include "llvm/IR/Value.h" |
49 | #include "llvm/InitializePasses.h" |
50 | #include "llvm/Pass.h" |
51 | #include "llvm/Support/Casting.h" |
52 | #include "llvm/Support/CommandLine.h" |
53 | #include "llvm/Support/Compiler.h" |
54 | #include "llvm/Support/Debug.h" |
55 | #include "llvm/Support/ErrorHandling.h" |
56 | #include "llvm/Support/KnownBits.h" |
57 | #include "llvm/Support/raw_ostream.h" |
58 | #include "llvm/TargetParser/Triple.h" |
59 | #include "llvm/Transforms/Scalar.h" |
60 | #include "llvm/Transforms/Utils.h" |
61 | #include "llvm/Transforms/Utils/Local.h" |
62 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
63 | #include <algorithm> |
64 | #include <array> |
65 | #include <cassert> |
66 | #include <cstdint> |
67 | #include <cstdlib> |
68 | #include <deque> |
69 | #include <functional> |
70 | #include <iterator> |
71 | #include <map> |
72 | #include <set> |
73 | #include <utility> |
74 | #include <vector> |
75 | |
76 | #define DEBUG_TYPE "hexagon-lir" |
77 | |
78 | using namespace llvm; |
79 | |
80 | static cl::opt<bool> DisableMemcpyIdiom("disable-memcpy-idiom" , |
81 | cl::Hidden, cl::init(Val: false), |
82 | cl::desc("Disable generation of memcpy in loop idiom recognition" )); |
83 | |
84 | static cl::opt<bool> DisableMemmoveIdiom("disable-memmove-idiom" , |
85 | cl::Hidden, cl::init(Val: false), |
86 | cl::desc("Disable generation of memmove in loop idiom recognition" )); |
87 | |
88 | static cl::opt<unsigned> RuntimeMemSizeThreshold("runtime-mem-idiom-threshold" , |
89 | cl::Hidden, cl::init(Val: 0), cl::desc("Threshold (in bytes) for the runtime " |
90 | "check guarding the memmove." )); |
91 | |
92 | static cl::opt<unsigned> CompileTimeMemSizeThreshold( |
93 | "compile-time-mem-idiom-threshold" , cl::Hidden, cl::init(Val: 64), |
94 | cl::desc("Threshold (in bytes) to perform the transformation, if the " |
95 | "runtime loop count (mem transfer size) is known at compile-time." )); |
96 | |
97 | static cl::opt<bool> OnlyNonNestedMemmove("only-nonnested-memmove-idiom" , |
98 | cl::Hidden, cl::init(Val: true), |
99 | cl::desc("Only enable generating memmove in non-nested loops" )); |
100 | |
101 | static cl::opt<bool> HexagonVolatileMemcpy( |
102 | "disable-hexagon-volatile-memcpy" , cl::Hidden, cl::init(Val: false), |
103 | cl::desc("Enable Hexagon-specific memcpy for volatile destination." )); |
104 | |
105 | static cl::opt<unsigned> SimplifyLimit("hlir-simplify-limit" , cl::init(Val: 10000), |
106 | cl::Hidden, cl::desc("Maximum number of simplification steps in HLIR" )); |
107 | |
108 | static const char *HexagonVolatileMemcpyName |
109 | = "hexagon_memcpy_forward_vp4cp4n2" ; |
110 | |
111 | |
112 | namespace llvm { |
113 | |
114 | void initializeHexagonLoopIdiomRecognizeLegacyPassPass(PassRegistry &); |
115 | Pass *createHexagonLoopIdiomPass(); |
116 | |
117 | } // end namespace llvm |
118 | |
119 | namespace { |
120 | |
121 | class HexagonLoopIdiomRecognize { |
122 | public: |
123 | explicit HexagonLoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, |
124 | LoopInfo *LF, const TargetLibraryInfo *TLI, |
125 | ScalarEvolution *SE) |
126 | : AA(AA), DT(DT), LF(LF), TLI(TLI), SE(SE) {} |
127 | |
128 | bool run(Loop *L); |
129 | |
130 | private: |
131 | int getSCEVStride(const SCEVAddRecExpr *StoreEv); |
132 | bool isLegalStore(Loop *CurLoop, StoreInst *SI); |
133 | void collectStores(Loop *CurLoop, BasicBlock *BB, |
134 | SmallVectorImpl<StoreInst *> &Stores); |
135 | bool processCopyingStore(Loop *CurLoop, StoreInst *SI, const SCEV *BECount); |
136 | bool coverLoop(Loop *L, SmallVectorImpl<Instruction *> &Insts) const; |
137 | bool runOnLoopBlock(Loop *CurLoop, BasicBlock *BB, const SCEV *BECount, |
138 | SmallVectorImpl<BasicBlock *> &ExitBlocks); |
139 | bool runOnCountableLoop(Loop *L); |
140 | |
141 | AliasAnalysis *AA; |
142 | const DataLayout *DL; |
143 | DominatorTree *DT; |
144 | LoopInfo *LF; |
145 | const TargetLibraryInfo *TLI; |
146 | ScalarEvolution *SE; |
147 | bool HasMemcpy, HasMemmove; |
148 | }; |
149 | |
150 | class HexagonLoopIdiomRecognizeLegacyPass : public LoopPass { |
151 | public: |
152 | static char ID; |
153 | |
154 | explicit HexagonLoopIdiomRecognizeLegacyPass() : LoopPass(ID) { |
155 | initializeHexagonLoopIdiomRecognizeLegacyPassPass( |
156 | *PassRegistry::getPassRegistry()); |
157 | } |
158 | |
159 | StringRef getPassName() const override { |
160 | return "Recognize Hexagon-specific loop idioms" ; |
161 | } |
162 | |
163 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
164 | AU.addRequired<LoopInfoWrapperPass>(); |
165 | AU.addRequiredID(ID&: LoopSimplifyID); |
166 | AU.addRequiredID(ID&: LCSSAID); |
167 | AU.addRequired<AAResultsWrapperPass>(); |
168 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
169 | AU.addRequired<DominatorTreeWrapperPass>(); |
170 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
171 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
172 | } |
173 | |
174 | bool runOnLoop(Loop *L, LPPassManager &LPM) override; |
175 | }; |
176 | |
177 | struct Simplifier { |
178 | struct Rule { |
179 | using FuncType = std::function<Value *(Instruction *, LLVMContext &)>; |
180 | Rule(StringRef N, FuncType F) : Name(N), Fn(F) {} |
181 | StringRef Name; // For debugging. |
182 | FuncType Fn; |
183 | }; |
184 | |
185 | void addRule(StringRef N, const Rule::FuncType &F) { |
186 | Rules.push_back(x: Rule(N, F)); |
187 | } |
188 | |
189 | private: |
190 | struct WorkListType { |
191 | WorkListType() = default; |
192 | |
193 | void push_back(Value *V) { |
194 | // Do not push back duplicates. |
195 | if (S.insert(x: V).second) |
196 | Q.push_back(x: V); |
197 | } |
198 | |
199 | Value *pop_front_val() { |
200 | Value *V = Q.front(); |
201 | Q.pop_front(); |
202 | S.erase(x: V); |
203 | return V; |
204 | } |
205 | |
206 | bool empty() const { return Q.empty(); } |
207 | |
208 | private: |
209 | std::deque<Value *> Q; |
210 | std::set<Value *> S; |
211 | }; |
212 | |
213 | using ValueSetType = std::set<Value *>; |
214 | |
215 | std::vector<Rule> Rules; |
216 | |
217 | public: |
218 | struct Context { |
219 | using ValueMapType = DenseMap<Value *, Value *>; |
220 | |
221 | Value *Root; |
222 | ValueSetType Used; // The set of all cloned values used by Root. |
223 | ValueSetType Clones; // The set of all cloned values. |
224 | LLVMContext &Ctx; |
225 | |
226 | Context(Instruction *Exp) |
227 | : Ctx(Exp->getParent()->getParent()->getContext()) { |
228 | initialize(Exp); |
229 | } |
230 | |
231 | ~Context() { cleanup(); } |
232 | |
233 | void print(raw_ostream &OS, const Value *V) const; |
234 | Value *materialize(BasicBlock *B, BasicBlock::iterator At); |
235 | |
236 | private: |
237 | friend struct Simplifier; |
238 | |
239 | void initialize(Instruction *Exp); |
240 | void cleanup(); |
241 | |
242 | template <typename FuncT> void traverse(Value *V, FuncT F); |
243 | void record(Value *V); |
244 | void use(Value *V); |
245 | void unuse(Value *V); |
246 | |
247 | bool equal(const Instruction *I, const Instruction *J) const; |
248 | Value *find(Value *Tree, Value *Sub) const; |
249 | Value *subst(Value *Tree, Value *OldV, Value *NewV); |
250 | void replace(Value *OldV, Value *NewV); |
251 | void link(Instruction *I, BasicBlock *B, BasicBlock::iterator At); |
252 | }; |
253 | |
254 | Value *simplify(Context &C); |
255 | }; |
256 | |
257 | struct PE { |
258 | PE(const Simplifier::Context &c, Value *v = nullptr) : C(c), V(v) {} |
259 | |
260 | const Simplifier::Context &C; |
261 | const Value *V; |
262 | }; |
263 | |
264 | LLVM_ATTRIBUTE_USED |
265 | raw_ostream &operator<<(raw_ostream &OS, const PE &P) { |
266 | P.C.print(OS, V: P.V ? P.V : P.C.Root); |
267 | return OS; |
268 | } |
269 | |
270 | } // end anonymous namespace |
271 | |
272 | char HexagonLoopIdiomRecognizeLegacyPass::ID = 0; |
273 | |
274 | INITIALIZE_PASS_BEGIN(HexagonLoopIdiomRecognizeLegacyPass, "hexagon-loop-idiom" , |
275 | "Recognize Hexagon-specific loop idioms" , false, false) |
276 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) |
277 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify) |
278 | INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass) |
279 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) |
280 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
281 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
282 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
283 | INITIALIZE_PASS_END(HexagonLoopIdiomRecognizeLegacyPass, "hexagon-loop-idiom" , |
284 | "Recognize Hexagon-specific loop idioms" , false, false) |
285 | |
286 | template <typename FuncT> |
287 | void Simplifier::Context::traverse(Value *V, FuncT F) { |
288 | WorkListType Q; |
289 | Q.push_back(V); |
290 | |
291 | while (!Q.empty()) { |
292 | Instruction *U = dyn_cast<Instruction>(Val: Q.pop_front_val()); |
293 | if (!U || U->getParent()) |
294 | continue; |
295 | if (!F(U)) |
296 | continue; |
297 | for (Value *Op : U->operands()) |
298 | Q.push_back(V: Op); |
299 | } |
300 | } |
301 | |
302 | void Simplifier::Context::print(raw_ostream &OS, const Value *V) const { |
303 | const auto *U = dyn_cast<const Instruction>(Val: V); |
304 | if (!U) { |
305 | OS << V << '(' << *V << ')'; |
306 | return; |
307 | } |
308 | |
309 | if (U->getParent()) { |
310 | OS << U << '('; |
311 | U->printAsOperand(O&: OS, PrintType: true); |
312 | OS << ')'; |
313 | return; |
314 | } |
315 | |
316 | unsigned N = U->getNumOperands(); |
317 | if (N != 0) |
318 | OS << U << '('; |
319 | OS << U->getOpcodeName(); |
320 | for (const Value *Op : U->operands()) { |
321 | OS << ' '; |
322 | print(OS, V: Op); |
323 | } |
324 | if (N != 0) |
325 | OS << ')'; |
326 | } |
327 | |
328 | void Simplifier::Context::initialize(Instruction *Exp) { |
329 | // Perform a deep clone of the expression, set Root to the root |
330 | // of the clone, and build a map from the cloned values to the |
331 | // original ones. |
332 | ValueMapType M; |
333 | BasicBlock *Block = Exp->getParent(); |
334 | WorkListType Q; |
335 | Q.push_back(V: Exp); |
336 | |
337 | while (!Q.empty()) { |
338 | Value *V = Q.pop_front_val(); |
339 | if (M.contains(Val: V)) |
340 | continue; |
341 | if (Instruction *U = dyn_cast<Instruction>(Val: V)) { |
342 | if (isa<PHINode>(Val: U) || U->getParent() != Block) |
343 | continue; |
344 | for (Value *Op : U->operands()) |
345 | Q.push_back(V: Op); |
346 | M.insert(KV: {U, U->clone()}); |
347 | } |
348 | } |
349 | |
350 | for (std::pair<Value*,Value*> P : M) { |
351 | Instruction *U = cast<Instruction>(Val: P.second); |
352 | for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i) { |
353 | auto F = M.find(Val: U->getOperand(i)); |
354 | if (F != M.end()) |
355 | U->setOperand(i, Val: F->second); |
356 | } |
357 | } |
358 | |
359 | auto R = M.find(Val: Exp); |
360 | assert(R != M.end()); |
361 | Root = R->second; |
362 | |
363 | record(V: Root); |
364 | use(V: Root); |
365 | } |
366 | |
367 | void Simplifier::Context::record(Value *V) { |
368 | auto Record = [this](Instruction *U) -> bool { |
369 | Clones.insert(x: U); |
370 | return true; |
371 | }; |
372 | traverse(V, F: Record); |
373 | } |
374 | |
375 | void Simplifier::Context::use(Value *V) { |
376 | auto Use = [this](Instruction *U) -> bool { |
377 | Used.insert(x: U); |
378 | return true; |
379 | }; |
380 | traverse(V, F: Use); |
381 | } |
382 | |
383 | void Simplifier::Context::unuse(Value *V) { |
384 | if (!isa<Instruction>(Val: V) || cast<Instruction>(Val: V)->getParent() != nullptr) |
385 | return; |
386 | |
387 | auto Unuse = [this](Instruction *U) -> bool { |
388 | if (!U->use_empty()) |
389 | return false; |
390 | Used.erase(x: U); |
391 | return true; |
392 | }; |
393 | traverse(V, F: Unuse); |
394 | } |
395 | |
396 | Value *Simplifier::Context::subst(Value *Tree, Value *OldV, Value *NewV) { |
397 | if (Tree == OldV) |
398 | return NewV; |
399 | if (OldV == NewV) |
400 | return Tree; |
401 | |
402 | WorkListType Q; |
403 | Q.push_back(V: Tree); |
404 | while (!Q.empty()) { |
405 | Instruction *U = dyn_cast<Instruction>(Val: Q.pop_front_val()); |
406 | // If U is not an instruction, or it's not a clone, skip it. |
407 | if (!U || U->getParent()) |
408 | continue; |
409 | for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i) { |
410 | Value *Op = U->getOperand(i); |
411 | if (Op == OldV) { |
412 | U->setOperand(i, Val: NewV); |
413 | unuse(V: OldV); |
414 | } else { |
415 | Q.push_back(V: Op); |
416 | } |
417 | } |
418 | } |
419 | return Tree; |
420 | } |
421 | |
422 | void Simplifier::Context::replace(Value *OldV, Value *NewV) { |
423 | if (Root == OldV) { |
424 | Root = NewV; |
425 | use(V: Root); |
426 | return; |
427 | } |
428 | |
429 | // NewV may be a complex tree that has just been created by one of the |
430 | // transformation rules. We need to make sure that it is commoned with |
431 | // the existing Root to the maximum extent possible. |
432 | // Identify all subtrees of NewV (including NewV itself) that have |
433 | // equivalent counterparts in Root, and replace those subtrees with |
434 | // these counterparts. |
435 | WorkListType Q; |
436 | Q.push_back(V: NewV); |
437 | while (!Q.empty()) { |
438 | Value *V = Q.pop_front_val(); |
439 | Instruction *U = dyn_cast<Instruction>(Val: V); |
440 | if (!U || U->getParent()) |
441 | continue; |
442 | if (Value *DupV = find(Tree: Root, Sub: V)) { |
443 | if (DupV != V) |
444 | NewV = subst(Tree: NewV, OldV: V, NewV: DupV); |
445 | } else { |
446 | for (Value *Op : U->operands()) |
447 | Q.push_back(V: Op); |
448 | } |
449 | } |
450 | |
451 | // Now, simply replace OldV with NewV in Root. |
452 | Root = subst(Tree: Root, OldV, NewV); |
453 | use(V: Root); |
454 | } |
455 | |
456 | void Simplifier::Context::cleanup() { |
457 | for (Value *V : Clones) { |
458 | Instruction *U = cast<Instruction>(Val: V); |
459 | if (!U->getParent()) |
460 | U->dropAllReferences(); |
461 | } |
462 | |
463 | for (Value *V : Clones) { |
464 | Instruction *U = cast<Instruction>(Val: V); |
465 | if (!U->getParent()) |
466 | U->deleteValue(); |
467 | } |
468 | } |
469 | |
470 | bool Simplifier::Context::equal(const Instruction *I, |
471 | const Instruction *J) const { |
472 | if (I == J) |
473 | return true; |
474 | if (!I->isSameOperationAs(I: J)) |
475 | return false; |
476 | if (isa<PHINode>(Val: I)) |
477 | return I->isIdenticalTo(I: J); |
478 | |
479 | for (unsigned i = 0, n = I->getNumOperands(); i != n; ++i) { |
480 | Value *OpI = I->getOperand(i), *OpJ = J->getOperand(i); |
481 | if (OpI == OpJ) |
482 | continue; |
483 | auto *InI = dyn_cast<const Instruction>(Val: OpI); |
484 | auto *InJ = dyn_cast<const Instruction>(Val: OpJ); |
485 | if (InI && InJ) { |
486 | if (!equal(I: InI, J: InJ)) |
487 | return false; |
488 | } else if (InI != InJ || !InI) |
489 | return false; |
490 | } |
491 | return true; |
492 | } |
493 | |
494 | Value *Simplifier::Context::find(Value *Tree, Value *Sub) const { |
495 | Instruction *SubI = dyn_cast<Instruction>(Val: Sub); |
496 | WorkListType Q; |
497 | Q.push_back(V: Tree); |
498 | |
499 | while (!Q.empty()) { |
500 | Value *V = Q.pop_front_val(); |
501 | if (V == Sub) |
502 | return V; |
503 | Instruction *U = dyn_cast<Instruction>(Val: V); |
504 | if (!U || U->getParent()) |
505 | continue; |
506 | if (SubI && equal(I: SubI, J: U)) |
507 | return U; |
508 | assert(!isa<PHINode>(U)); |
509 | for (Value *Op : U->operands()) |
510 | Q.push_back(V: Op); |
511 | } |
512 | return nullptr; |
513 | } |
514 | |
515 | void Simplifier::Context::link(Instruction *I, BasicBlock *B, |
516 | BasicBlock::iterator At) { |
517 | if (I->getParent()) |
518 | return; |
519 | |
520 | for (Value *Op : I->operands()) { |
521 | if (Instruction *OpI = dyn_cast<Instruction>(Val: Op)) |
522 | link(I: OpI, B, At); |
523 | } |
524 | |
525 | I->insertInto(ParentBB: B, It: At); |
526 | } |
527 | |
528 | Value *Simplifier::Context::materialize(BasicBlock *B, |
529 | BasicBlock::iterator At) { |
530 | if (Instruction *RootI = dyn_cast<Instruction>(Val: Root)) |
531 | link(I: RootI, B, At); |
532 | return Root; |
533 | } |
534 | |
535 | Value *Simplifier::simplify(Context &C) { |
536 | WorkListType Q; |
537 | Q.push_back(V: C.Root); |
538 | unsigned Count = 0; |
539 | const unsigned Limit = SimplifyLimit; |
540 | |
541 | while (!Q.empty()) { |
542 | if (Count++ >= Limit) |
543 | break; |
544 | Instruction *U = dyn_cast<Instruction>(Val: Q.pop_front_val()); |
545 | if (!U || U->getParent() || !C.Used.count(x: U)) |
546 | continue; |
547 | bool Changed = false; |
548 | for (Rule &R : Rules) { |
549 | Value *W = R.Fn(U, C.Ctx); |
550 | if (!W) |
551 | continue; |
552 | Changed = true; |
553 | C.record(V: W); |
554 | C.replace(OldV: U, NewV: W); |
555 | Q.push_back(V: C.Root); |
556 | break; |
557 | } |
558 | if (!Changed) { |
559 | for (Value *Op : U->operands()) |
560 | Q.push_back(V: Op); |
561 | } |
562 | } |
563 | return Count < Limit ? C.Root : nullptr; |
564 | } |
565 | |
566 | //===----------------------------------------------------------------------===// |
567 | // |
568 | // Implementation of PolynomialMultiplyRecognize |
569 | // |
570 | //===----------------------------------------------------------------------===// |
571 | |
572 | namespace { |
573 | |
574 | class PolynomialMultiplyRecognize { |
575 | public: |
576 | explicit PolynomialMultiplyRecognize(Loop *loop, const DataLayout &dl, |
577 | const DominatorTree &dt, const TargetLibraryInfo &tli, |
578 | ScalarEvolution &se) |
579 | : CurLoop(loop), DL(dl), DT(dt), TLI(tli), SE(se) {} |
580 | |
581 | bool recognize(); |
582 | |
583 | private: |
584 | using ValueSeq = SetVector<Value *>; |
585 | |
586 | IntegerType *getPmpyType() const { |
587 | LLVMContext &Ctx = CurLoop->getHeader()->getParent()->getContext(); |
588 | return IntegerType::get(C&: Ctx, NumBits: 32); |
589 | } |
590 | |
591 | bool isPromotableTo(Value *V, IntegerType *Ty); |
592 | void promoteTo(Instruction *In, IntegerType *DestTy, BasicBlock *LoopB); |
593 | bool promoteTypes(BasicBlock *LoopB, BasicBlock *ExitB); |
594 | |
595 | Value *getCountIV(BasicBlock *BB); |
596 | bool findCycle(Value *Out, Value *In, ValueSeq &Cycle); |
597 | void classifyCycle(Instruction *DivI, ValueSeq &Cycle, ValueSeq &Early, |
598 | ValueSeq &Late); |
599 | bool classifyInst(Instruction *UseI, ValueSeq &Early, ValueSeq &Late); |
600 | bool commutesWithShift(Instruction *I); |
601 | bool highBitsAreZero(Value *V, unsigned IterCount); |
602 | bool keepsHighBitsZero(Value *V, unsigned IterCount); |
603 | bool isOperandShifted(Instruction *I, Value *Op); |
604 | bool convertShiftsToLeft(BasicBlock *LoopB, BasicBlock *ExitB, |
605 | unsigned IterCount); |
606 | void cleanupLoopBody(BasicBlock *LoopB); |
607 | |
608 | struct ParsedValues { |
609 | ParsedValues() = default; |
610 | |
611 | Value *M = nullptr; |
612 | Value *P = nullptr; |
613 | Value *Q = nullptr; |
614 | Value *R = nullptr; |
615 | Value *X = nullptr; |
616 | Instruction *Res = nullptr; |
617 | unsigned IterCount = 0; |
618 | bool Left = false; |
619 | bool Inv = false; |
620 | }; |
621 | |
622 | bool matchLeftShift(SelectInst *SelI, Value *CIV, ParsedValues &PV); |
623 | bool matchRightShift(SelectInst *SelI, ParsedValues &PV); |
624 | bool scanSelect(SelectInst *SI, BasicBlock *LoopB, BasicBlock *PrehB, |
625 | Value *CIV, ParsedValues &PV, bool PreScan); |
626 | unsigned getInverseMxN(unsigned QP); |
627 | Value *generate(BasicBlock::iterator At, ParsedValues &PV); |
628 | |
629 | void setupPreSimplifier(Simplifier &S); |
630 | void setupPostSimplifier(Simplifier &S); |
631 | |
632 | Loop *CurLoop; |
633 | const DataLayout &DL; |
634 | const DominatorTree &DT; |
635 | const TargetLibraryInfo &TLI; |
636 | ScalarEvolution &SE; |
637 | }; |
638 | |
639 | } // end anonymous namespace |
640 | |
641 | Value *PolynomialMultiplyRecognize::getCountIV(BasicBlock *BB) { |
642 | pred_iterator PI = pred_begin(BB), PE = pred_end(BB); |
643 | if (std::distance(first: PI, last: PE) != 2) |
644 | return nullptr; |
645 | BasicBlock *PB = (*PI == BB) ? *std::next(x: PI) : *PI; |
646 | |
647 | for (auto I = BB->begin(), E = BB->end(); I != E && isa<PHINode>(Val: I); ++I) { |
648 | auto *PN = cast<PHINode>(Val&: I); |
649 | Value *InitV = PN->getIncomingValueForBlock(BB: PB); |
650 | if (!isa<ConstantInt>(Val: InitV) || !cast<ConstantInt>(Val: InitV)->isZero()) |
651 | continue; |
652 | Value *IterV = PN->getIncomingValueForBlock(BB); |
653 | auto *BO = dyn_cast<BinaryOperator>(Val: IterV); |
654 | if (!BO) |
655 | continue; |
656 | if (BO->getOpcode() != Instruction::Add) |
657 | continue; |
658 | Value *IncV = nullptr; |
659 | if (BO->getOperand(i_nocapture: 0) == PN) |
660 | IncV = BO->getOperand(i_nocapture: 1); |
661 | else if (BO->getOperand(i_nocapture: 1) == PN) |
662 | IncV = BO->getOperand(i_nocapture: 0); |
663 | if (IncV == nullptr) |
664 | continue; |
665 | |
666 | if (auto *T = dyn_cast<ConstantInt>(Val: IncV)) |
667 | if (T->isOne()) |
668 | return PN; |
669 | } |
670 | return nullptr; |
671 | } |
672 | |
673 | static void replaceAllUsesOfWithIn(Value *I, Value *J, BasicBlock *BB) { |
674 | for (auto UI = I->user_begin(), UE = I->user_end(); UI != UE;) { |
675 | Use &TheUse = UI.getUse(); |
676 | ++UI; |
677 | if (auto *II = dyn_cast<Instruction>(Val: TheUse.getUser())) |
678 | if (BB == II->getParent()) |
679 | II->replaceUsesOfWith(From: I, To: J); |
680 | } |
681 | } |
682 | |
683 | bool PolynomialMultiplyRecognize::matchLeftShift(SelectInst *SelI, |
684 | Value *CIV, ParsedValues &PV) { |
685 | // Match the following: |
686 | // select (X & (1 << i)) != 0 ? R ^ (Q << i) : R |
687 | // select (X & (1 << i)) == 0 ? R : R ^ (Q << i) |
688 | // The condition may also check for equality with the masked value, i.e |
689 | // select (X & (1 << i)) == (1 << i) ? R ^ (Q << i) : R |
690 | // select (X & (1 << i)) != (1 << i) ? R : R ^ (Q << i); |
691 | |
692 | Value *CondV = SelI->getCondition(); |
693 | Value *TrueV = SelI->getTrueValue(); |
694 | Value *FalseV = SelI->getFalseValue(); |
695 | |
696 | using namespace PatternMatch; |
697 | |
698 | CmpInst::Predicate P; |
699 | Value *A = nullptr, *B = nullptr, *C = nullptr; |
700 | |
701 | if (!match(V: CondV, P: m_ICmp(Pred&: P, L: m_And(L: m_Value(V&: A), R: m_Value(V&: B)), R: m_Value(V&: C))) && |
702 | !match(V: CondV, P: m_ICmp(Pred&: P, L: m_Value(V&: C), R: m_And(L: m_Value(V&: A), R: m_Value(V&: B))))) |
703 | return false; |
704 | if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE) |
705 | return false; |
706 | // Matched: select (A & B) == C ? ... : ... |
707 | // select (A & B) != C ? ... : ... |
708 | |
709 | Value *X = nullptr, *Sh1 = nullptr; |
710 | // Check (A & B) for (X & (1 << i)): |
711 | if (match(V: A, P: m_Shl(L: m_One(), R: m_Specific(V: CIV)))) { |
712 | Sh1 = A; |
713 | X = B; |
714 | } else if (match(V: B, P: m_Shl(L: m_One(), R: m_Specific(V: CIV)))) { |
715 | Sh1 = B; |
716 | X = A; |
717 | } else { |
718 | // TODO: Could also check for an induction variable containing single |
719 | // bit shifted left by 1 in each iteration. |
720 | return false; |
721 | } |
722 | |
723 | bool TrueIfZero; |
724 | |
725 | // Check C against the possible values for comparison: 0 and (1 << i): |
726 | if (match(V: C, P: m_Zero())) |
727 | TrueIfZero = (P == CmpInst::ICMP_EQ); |
728 | else if (C == Sh1) |
729 | TrueIfZero = (P == CmpInst::ICMP_NE); |
730 | else |
731 | return false; |
732 | |
733 | // So far, matched: |
734 | // select (X & (1 << i)) ? ... : ... |
735 | // including variations of the check against zero/non-zero value. |
736 | |
737 | Value *ShouldSameV = nullptr, *ShouldXoredV = nullptr; |
738 | if (TrueIfZero) { |
739 | ShouldSameV = TrueV; |
740 | ShouldXoredV = FalseV; |
741 | } else { |
742 | ShouldSameV = FalseV; |
743 | ShouldXoredV = TrueV; |
744 | } |
745 | |
746 | Value *Q = nullptr, *R = nullptr, *Y = nullptr, *Z = nullptr; |
747 | Value *T = nullptr; |
748 | if (match(V: ShouldXoredV, P: m_Xor(L: m_Value(V&: Y), R: m_Value(V&: Z)))) { |
749 | // Matched: select +++ ? ... : Y ^ Z |
750 | // select +++ ? Y ^ Z : ... |
751 | // where +++ denotes previously checked matches. |
752 | if (ShouldSameV == Y) |
753 | T = Z; |
754 | else if (ShouldSameV == Z) |
755 | T = Y; |
756 | else |
757 | return false; |
758 | R = ShouldSameV; |
759 | // Matched: select +++ ? R : R ^ T |
760 | // select +++ ? R ^ T : R |
761 | // depending on TrueIfZero. |
762 | |
763 | } else if (match(V: ShouldSameV, P: m_Zero())) { |
764 | // Matched: select +++ ? 0 : ... |
765 | // select +++ ? ... : 0 |
766 | if (!SelI->hasOneUse()) |
767 | return false; |
768 | T = ShouldXoredV; |
769 | // Matched: select +++ ? 0 : T |
770 | // select +++ ? T : 0 |
771 | |
772 | Value *U = *SelI->user_begin(); |
773 | if (!match(V: U, P: m_c_Xor(L: m_Specific(V: SelI), R: m_Value(V&: R)))) |
774 | return false; |
775 | // Matched: xor (select +++ ? 0 : T), R |
776 | // xor (select +++ ? T : 0), R |
777 | } else |
778 | return false; |
779 | |
780 | // The xor input value T is isolated into its own match so that it could |
781 | // be checked against an induction variable containing a shifted bit |
782 | // (todo). |
783 | // For now, check against (Q << i). |
784 | if (!match(V: T, P: m_Shl(L: m_Value(V&: Q), R: m_Specific(V: CIV))) && |
785 | !match(V: T, P: m_Shl(L: m_ZExt(Op: m_Value(V&: Q)), R: m_ZExt(Op: m_Specific(V: CIV))))) |
786 | return false; |
787 | // Matched: select +++ ? R : R ^ (Q << i) |
788 | // select +++ ? R ^ (Q << i) : R |
789 | |
790 | PV.X = X; |
791 | PV.Q = Q; |
792 | PV.R = R; |
793 | PV.Left = true; |
794 | return true; |
795 | } |
796 | |
797 | bool PolynomialMultiplyRecognize::matchRightShift(SelectInst *SelI, |
798 | ParsedValues &PV) { |
799 | // Match the following: |
800 | // select (X & 1) != 0 ? (R >> 1) ^ Q : (R >> 1) |
801 | // select (X & 1) == 0 ? (R >> 1) : (R >> 1) ^ Q |
802 | // The condition may also check for equality with the masked value, i.e |
803 | // select (X & 1) == 1 ? (R >> 1) ^ Q : (R >> 1) |
804 | // select (X & 1) != 1 ? (R >> 1) : (R >> 1) ^ Q |
805 | |
806 | Value *CondV = SelI->getCondition(); |
807 | Value *TrueV = SelI->getTrueValue(); |
808 | Value *FalseV = SelI->getFalseValue(); |
809 | |
810 | using namespace PatternMatch; |
811 | |
812 | Value *C = nullptr; |
813 | CmpInst::Predicate P; |
814 | bool TrueIfZero; |
815 | |
816 | if (match(V: CondV, P: m_c_ICmp(Pred&: P, L: m_Value(V&: C), R: m_Zero()))) { |
817 | if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE) |
818 | return false; |
819 | // Matched: select C == 0 ? ... : ... |
820 | // select C != 0 ? ... : ... |
821 | TrueIfZero = (P == CmpInst::ICMP_EQ); |
822 | } else if (match(V: CondV, P: m_c_ICmp(Pred&: P, L: m_Value(V&: C), R: m_One()))) { |
823 | if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE) |
824 | return false; |
825 | // Matched: select C == 1 ? ... : ... |
826 | // select C != 1 ? ... : ... |
827 | TrueIfZero = (P == CmpInst::ICMP_NE); |
828 | } else |
829 | return false; |
830 | |
831 | Value *X = nullptr; |
832 | if (!match(V: C, P: m_And(L: m_Value(V&: X), R: m_One()))) |
833 | return false; |
834 | // Matched: select (X & 1) == +++ ? ... : ... |
835 | // select (X & 1) != +++ ? ... : ... |
836 | |
837 | Value *R = nullptr, *Q = nullptr; |
838 | if (TrueIfZero) { |
839 | // The select's condition is true if the tested bit is 0. |
840 | // TrueV must be the shift, FalseV must be the xor. |
841 | if (!match(V: TrueV, P: m_LShr(L: m_Value(V&: R), R: m_One()))) |
842 | return false; |
843 | // Matched: select +++ ? (R >> 1) : ... |
844 | if (!match(V: FalseV, P: m_c_Xor(L: m_Specific(V: TrueV), R: m_Value(V&: Q)))) |
845 | return false; |
846 | // Matched: select +++ ? (R >> 1) : (R >> 1) ^ Q |
847 | // with commuting ^. |
848 | } else { |
849 | // The select's condition is true if the tested bit is 1. |
850 | // TrueV must be the xor, FalseV must be the shift. |
851 | if (!match(V: FalseV, P: m_LShr(L: m_Value(V&: R), R: m_One()))) |
852 | return false; |
853 | // Matched: select +++ ? ... : (R >> 1) |
854 | if (!match(V: TrueV, P: m_c_Xor(L: m_Specific(V: FalseV), R: m_Value(V&: Q)))) |
855 | return false; |
856 | // Matched: select +++ ? (R >> 1) ^ Q : (R >> 1) |
857 | // with commuting ^. |
858 | } |
859 | |
860 | PV.X = X; |
861 | PV.Q = Q; |
862 | PV.R = R; |
863 | PV.Left = false; |
864 | return true; |
865 | } |
866 | |
867 | bool PolynomialMultiplyRecognize::scanSelect(SelectInst *SelI, |
868 | BasicBlock *LoopB, BasicBlock *PrehB, Value *CIV, ParsedValues &PV, |
869 | bool PreScan) { |
870 | using namespace PatternMatch; |
871 | |
872 | // The basic pattern for R = P.Q is: |
873 | // for i = 0..31 |
874 | // R = phi (0, R') |
875 | // if (P & (1 << i)) ; test-bit(P, i) |
876 | // R' = R ^ (Q << i) |
877 | // |
878 | // Similarly, the basic pattern for R = (P/Q).Q - P |
879 | // for i = 0..31 |
880 | // R = phi(P, R') |
881 | // if (R & (1 << i)) |
882 | // R' = R ^ (Q << i) |
883 | |
884 | // There exist idioms, where instead of Q being shifted left, P is shifted |
885 | // right. This produces a result that is shifted right by 32 bits (the |
886 | // non-shifted result is 64-bit). |
887 | // |
888 | // For R = P.Q, this would be: |
889 | // for i = 0..31 |
890 | // R = phi (0, R') |
891 | // if ((P >> i) & 1) |
892 | // R' = (R >> 1) ^ Q ; R is cycled through the loop, so it must |
893 | // else ; be shifted by 1, not i. |
894 | // R' = R >> 1 |
895 | // |
896 | // And for the inverse: |
897 | // for i = 0..31 |
898 | // R = phi (P, R') |
899 | // if (R & 1) |
900 | // R' = (R >> 1) ^ Q |
901 | // else |
902 | // R' = R >> 1 |
903 | |
904 | // The left-shifting idioms share the same pattern: |
905 | // select (X & (1 << i)) ? R ^ (Q << i) : R |
906 | // Similarly for right-shifting idioms: |
907 | // select (X & 1) ? (R >> 1) ^ Q |
908 | |
909 | if (matchLeftShift(SelI, CIV, PV)) { |
910 | // If this is a pre-scan, getting this far is sufficient. |
911 | if (PreScan) |
912 | return true; |
913 | |
914 | // Need to make sure that the SelI goes back into R. |
915 | auto *RPhi = dyn_cast<PHINode>(Val: PV.R); |
916 | if (!RPhi) |
917 | return false; |
918 | if (SelI != RPhi->getIncomingValueForBlock(BB: LoopB)) |
919 | return false; |
920 | PV.Res = SelI; |
921 | |
922 | // If X is loop invariant, it must be the input polynomial, and the |
923 | // idiom is the basic polynomial multiply. |
924 | if (CurLoop->isLoopInvariant(V: PV.X)) { |
925 | PV.P = PV.X; |
926 | PV.Inv = false; |
927 | } else { |
928 | // X is not loop invariant. If X == R, this is the inverse pmpy. |
929 | // Otherwise, check for an xor with an invariant value. If the |
930 | // variable argument to the xor is R, then this is still a valid |
931 | // inverse pmpy. |
932 | PV.Inv = true; |
933 | if (PV.X != PV.R) { |
934 | Value *Var = nullptr, *Inv = nullptr, *X1 = nullptr, *X2 = nullptr; |
935 | if (!match(V: PV.X, P: m_Xor(L: m_Value(V&: X1), R: m_Value(V&: X2)))) |
936 | return false; |
937 | auto *I1 = dyn_cast<Instruction>(Val: X1); |
938 | auto *I2 = dyn_cast<Instruction>(Val: X2); |
939 | if (!I1 || I1->getParent() != LoopB) { |
940 | Var = X2; |
941 | Inv = X1; |
942 | } else if (!I2 || I2->getParent() != LoopB) { |
943 | Var = X1; |
944 | Inv = X2; |
945 | } else |
946 | return false; |
947 | if (Var != PV.R) |
948 | return false; |
949 | PV.M = Inv; |
950 | } |
951 | // The input polynomial P still needs to be determined. It will be |
952 | // the entry value of R. |
953 | Value *EntryP = RPhi->getIncomingValueForBlock(BB: PrehB); |
954 | PV.P = EntryP; |
955 | } |
956 | |
957 | return true; |
958 | } |
959 | |
960 | if (matchRightShift(SelI, PV)) { |
961 | // If this is an inverse pattern, the Q polynomial must be known at |
962 | // compile time. |
963 | if (PV.Inv && !isa<ConstantInt>(Val: PV.Q)) |
964 | return false; |
965 | if (PreScan) |
966 | return true; |
967 | // There is no exact matching of right-shift pmpy. |
968 | return false; |
969 | } |
970 | |
971 | return false; |
972 | } |
973 | |
974 | bool PolynomialMultiplyRecognize::isPromotableTo(Value *Val, |
975 | IntegerType *DestTy) { |
976 | IntegerType *T = dyn_cast<IntegerType>(Val: Val->getType()); |
977 | if (!T || T->getBitWidth() > DestTy->getBitWidth()) |
978 | return false; |
979 | if (T->getBitWidth() == DestTy->getBitWidth()) |
980 | return true; |
981 | // Non-instructions are promotable. The reason why an instruction may not |
982 | // be promotable is that it may produce a different result if its operands |
983 | // and the result are promoted, for example, it may produce more non-zero |
984 | // bits. While it would still be possible to represent the proper result |
985 | // in a wider type, it may require adding additional instructions (which |
986 | // we don't want to do). |
987 | Instruction *In = dyn_cast<Instruction>(Val); |
988 | if (!In) |
989 | return true; |
990 | // The bitwidth of the source type is smaller than the destination. |
991 | // Check if the individual operation can be promoted. |
992 | switch (In->getOpcode()) { |
993 | case Instruction::PHI: |
994 | case Instruction::ZExt: |
995 | case Instruction::And: |
996 | case Instruction::Or: |
997 | case Instruction::Xor: |
998 | case Instruction::LShr: // Shift right is ok. |
999 | case Instruction::Select: |
1000 | case Instruction::Trunc: |
1001 | return true; |
1002 | case Instruction::ICmp: |
1003 | if (CmpInst *CI = cast<CmpInst>(Val: In)) |
1004 | return CI->isEquality() || CI->isUnsigned(); |
1005 | llvm_unreachable("Cast failed unexpectedly" ); |
1006 | case Instruction::Add: |
1007 | return In->hasNoSignedWrap() && In->hasNoUnsignedWrap(); |
1008 | } |
1009 | return false; |
1010 | } |
1011 | |
1012 | void PolynomialMultiplyRecognize::promoteTo(Instruction *In, |
1013 | IntegerType *DestTy, BasicBlock *LoopB) { |
1014 | Type *OrigTy = In->getType(); |
1015 | assert(!OrigTy->isVoidTy() && "Invalid instruction to promote" ); |
1016 | |
1017 | // Leave boolean values alone. |
1018 | if (!In->getType()->isIntegerTy(Bitwidth: 1)) |
1019 | In->mutateType(Ty: DestTy); |
1020 | unsigned DestBW = DestTy->getBitWidth(); |
1021 | |
1022 | // Handle PHIs. |
1023 | if (PHINode *P = dyn_cast<PHINode>(Val: In)) { |
1024 | unsigned N = P->getNumIncomingValues(); |
1025 | for (unsigned i = 0; i != N; ++i) { |
1026 | BasicBlock *InB = P->getIncomingBlock(i); |
1027 | if (InB == LoopB) |
1028 | continue; |
1029 | Value *InV = P->getIncomingValue(i); |
1030 | IntegerType *Ty = cast<IntegerType>(Val: InV->getType()); |
1031 | // Do not promote values in PHI nodes of type i1. |
1032 | if (Ty != P->getType()) { |
1033 | // If the value type does not match the PHI type, the PHI type |
1034 | // must have been promoted. |
1035 | assert(Ty->getBitWidth() < DestBW); |
1036 | InV = IRBuilder<>(InB->getTerminator()).CreateZExt(V: InV, DestTy); |
1037 | P->setIncomingValue(i, V: InV); |
1038 | } |
1039 | } |
1040 | } else if (ZExtInst *Z = dyn_cast<ZExtInst>(Val: In)) { |
1041 | Value *Op = Z->getOperand(i_nocapture: 0); |
1042 | if (Op->getType() == Z->getType()) |
1043 | Z->replaceAllUsesWith(V: Op); |
1044 | Z->eraseFromParent(); |
1045 | return; |
1046 | } |
1047 | if (TruncInst *T = dyn_cast<TruncInst>(Val: In)) { |
1048 | IntegerType *TruncTy = cast<IntegerType>(Val: OrigTy); |
1049 | Value *Mask = ConstantInt::get(Ty: DestTy, V: (1u << TruncTy->getBitWidth()) - 1); |
1050 | Value *And = IRBuilder<>(In).CreateAnd(LHS: T->getOperand(i_nocapture: 0), RHS: Mask); |
1051 | T->replaceAllUsesWith(V: And); |
1052 | T->eraseFromParent(); |
1053 | return; |
1054 | } |
1055 | |
1056 | // Promote immediates. |
1057 | for (unsigned i = 0, n = In->getNumOperands(); i != n; ++i) { |
1058 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: In->getOperand(i))) |
1059 | if (CI->getBitWidth() < DestBW) |
1060 | In->setOperand(i, Val: ConstantInt::get(Ty: DestTy, V: CI->getZExtValue())); |
1061 | } |
1062 | } |
1063 | |
1064 | bool PolynomialMultiplyRecognize::promoteTypes(BasicBlock *LoopB, |
1065 | BasicBlock *ExitB) { |
1066 | assert(LoopB); |
1067 | // Skip loops where the exit block has more than one predecessor. The values |
1068 | // coming from the loop block will be promoted to another type, and so the |
1069 | // values coming into the exit block from other predecessors would also have |
1070 | // to be promoted. |
1071 | if (!ExitB || (ExitB->getSinglePredecessor() != LoopB)) |
1072 | return false; |
1073 | IntegerType *DestTy = getPmpyType(); |
1074 | // Check if the exit values have types that are no wider than the type |
1075 | // that we want to promote to. |
1076 | unsigned DestBW = DestTy->getBitWidth(); |
1077 | for (PHINode &P : ExitB->phis()) { |
1078 | if (P.getNumIncomingValues() != 1) |
1079 | return false; |
1080 | assert(P.getIncomingBlock(0) == LoopB); |
1081 | IntegerType *T = dyn_cast<IntegerType>(Val: P.getType()); |
1082 | if (!T || T->getBitWidth() > DestBW) |
1083 | return false; |
1084 | } |
1085 | |
1086 | // Check all instructions in the loop. |
1087 | for (Instruction &In : *LoopB) |
1088 | if (!In.isTerminator() && !isPromotableTo(Val: &In, DestTy)) |
1089 | return false; |
1090 | |
1091 | // Perform the promotion. |
1092 | std::vector<Instruction*> LoopIns; |
1093 | std::transform(first: LoopB->begin(), last: LoopB->end(), result: std::back_inserter(x&: LoopIns), |
1094 | unary_op: [](Instruction &In) { return &In; }); |
1095 | for (Instruction *In : LoopIns) |
1096 | if (!In->isTerminator()) |
1097 | promoteTo(In, DestTy, LoopB); |
1098 | |
1099 | // Fix up the PHI nodes in the exit block. |
1100 | Instruction *EndI = ExitB->getFirstNonPHI(); |
1101 | BasicBlock::iterator End = EndI ? EndI->getIterator() : ExitB->end(); |
1102 | for (auto I = ExitB->begin(); I != End; ++I) { |
1103 | PHINode *P = dyn_cast<PHINode>(Val&: I); |
1104 | if (!P) |
1105 | break; |
1106 | Type *Ty0 = P->getIncomingValue(i: 0)->getType(); |
1107 | Type *PTy = P->getType(); |
1108 | if (PTy != Ty0) { |
1109 | assert(Ty0 == DestTy); |
1110 | // In order to create the trunc, P must have the promoted type. |
1111 | P->mutateType(Ty: Ty0); |
1112 | Value *T = IRBuilder<>(ExitB, End).CreateTrunc(V: P, DestTy: PTy); |
1113 | // In order for the RAUW to work, the types of P and T must match. |
1114 | P->mutateType(Ty: PTy); |
1115 | P->replaceAllUsesWith(V: T); |
1116 | // Final update of the P's type. |
1117 | P->mutateType(Ty: Ty0); |
1118 | cast<Instruction>(Val: T)->setOperand(i: 0, Val: P); |
1119 | } |
1120 | } |
1121 | |
1122 | return true; |
1123 | } |
1124 | |
1125 | bool PolynomialMultiplyRecognize::findCycle(Value *Out, Value *In, |
1126 | ValueSeq &Cycle) { |
1127 | // Out = ..., In, ... |
1128 | if (Out == In) |
1129 | return true; |
1130 | |
1131 | auto *BB = cast<Instruction>(Val: Out)->getParent(); |
1132 | bool HadPhi = false; |
1133 | |
1134 | for (auto *U : Out->users()) { |
1135 | auto *I = dyn_cast<Instruction>(Val: &*U); |
1136 | if (I == nullptr || I->getParent() != BB) |
1137 | continue; |
1138 | // Make sure that there are no multi-iteration cycles, e.g. |
1139 | // p1 = phi(p2) |
1140 | // p2 = phi(p1) |
1141 | // The cycle p1->p2->p1 would span two loop iterations. |
1142 | // Check that there is only one phi in the cycle. |
1143 | bool IsPhi = isa<PHINode>(Val: I); |
1144 | if (IsPhi && HadPhi) |
1145 | return false; |
1146 | HadPhi |= IsPhi; |
1147 | if (!Cycle.insert(X: I)) |
1148 | return false; |
1149 | if (findCycle(Out: I, In, Cycle)) |
1150 | break; |
1151 | Cycle.remove(X: I); |
1152 | } |
1153 | return !Cycle.empty(); |
1154 | } |
1155 | |
1156 | void PolynomialMultiplyRecognize::classifyCycle(Instruction *DivI, |
1157 | ValueSeq &Cycle, ValueSeq &Early, ValueSeq &Late) { |
1158 | // All the values in the cycle that are between the phi node and the |
1159 | // divider instruction will be classified as "early", all other values |
1160 | // will be "late". |
1161 | |
1162 | bool IsE = true; |
1163 | unsigned I, N = Cycle.size(); |
1164 | for (I = 0; I < N; ++I) { |
1165 | Value *V = Cycle[I]; |
1166 | if (DivI == V) |
1167 | IsE = false; |
1168 | else if (!isa<PHINode>(Val: V)) |
1169 | continue; |
1170 | // Stop if found either. |
1171 | break; |
1172 | } |
1173 | // "I" is the index of either DivI or the phi node, whichever was first. |
1174 | // "E" is "false" or "true" respectively. |
1175 | ValueSeq &First = !IsE ? Early : Late; |
1176 | for (unsigned J = 0; J < I; ++J) |
1177 | First.insert(X: Cycle[J]); |
1178 | |
1179 | ValueSeq &Second = IsE ? Early : Late; |
1180 | Second.insert(X: Cycle[I]); |
1181 | for (++I; I < N; ++I) { |
1182 | Value *V = Cycle[I]; |
1183 | if (DivI == V || isa<PHINode>(Val: V)) |
1184 | break; |
1185 | Second.insert(X: V); |
1186 | } |
1187 | |
1188 | for (; I < N; ++I) |
1189 | First.insert(X: Cycle[I]); |
1190 | } |
1191 | |
1192 | bool PolynomialMultiplyRecognize::classifyInst(Instruction *UseI, |
1193 | ValueSeq &Early, ValueSeq &Late) { |
1194 | // Select is an exception, since the condition value does not have to be |
1195 | // classified in the same way as the true/false values. The true/false |
1196 | // values do have to be both early or both late. |
1197 | if (UseI->getOpcode() == Instruction::Select) { |
1198 | Value *TV = UseI->getOperand(i: 1), *FV = UseI->getOperand(i: 2); |
1199 | if (Early.count(key: TV) || Early.count(key: FV)) { |
1200 | if (Late.count(key: TV) || Late.count(key: FV)) |
1201 | return false; |
1202 | Early.insert(X: UseI); |
1203 | } else if (Late.count(key: TV) || Late.count(key: FV)) { |
1204 | if (Early.count(key: TV) || Early.count(key: FV)) |
1205 | return false; |
1206 | Late.insert(X: UseI); |
1207 | } |
1208 | return true; |
1209 | } |
1210 | |
1211 | // Not sure what would be the example of this, but the code below relies |
1212 | // on having at least one operand. |
1213 | if (UseI->getNumOperands() == 0) |
1214 | return true; |
1215 | |
1216 | bool AE = true, AL = true; |
1217 | for (auto &I : UseI->operands()) { |
1218 | if (Early.count(key: &*I)) |
1219 | AL = false; |
1220 | else if (Late.count(key: &*I)) |
1221 | AE = false; |
1222 | } |
1223 | // If the operands appear "all early" and "all late" at the same time, |
1224 | // then it means that none of them are actually classified as either. |
1225 | // This is harmless. |
1226 | if (AE && AL) |
1227 | return true; |
1228 | // Conversely, if they are neither "all early" nor "all late", then |
1229 | // we have a mixture of early and late operands that is not a known |
1230 | // exception. |
1231 | if (!AE && !AL) |
1232 | return false; |
1233 | |
1234 | // Check that we have covered the two special cases. |
1235 | assert(AE != AL); |
1236 | |
1237 | if (AE) |
1238 | Early.insert(X: UseI); |
1239 | else |
1240 | Late.insert(X: UseI); |
1241 | return true; |
1242 | } |
1243 | |
1244 | bool PolynomialMultiplyRecognize::commutesWithShift(Instruction *I) { |
1245 | switch (I->getOpcode()) { |
1246 | case Instruction::And: |
1247 | case Instruction::Or: |
1248 | case Instruction::Xor: |
1249 | case Instruction::LShr: |
1250 | case Instruction::Shl: |
1251 | case Instruction::Select: |
1252 | case Instruction::ICmp: |
1253 | case Instruction::PHI: |
1254 | break; |
1255 | default: |
1256 | return false; |
1257 | } |
1258 | return true; |
1259 | } |
1260 | |
1261 | bool PolynomialMultiplyRecognize::highBitsAreZero(Value *V, |
1262 | unsigned IterCount) { |
1263 | auto *T = dyn_cast<IntegerType>(Val: V->getType()); |
1264 | if (!T) |
1265 | return false; |
1266 | |
1267 | KnownBits Known(T->getBitWidth()); |
1268 | computeKnownBits(V, Known, DL); |
1269 | return Known.countMinLeadingZeros() >= IterCount; |
1270 | } |
1271 | |
1272 | bool PolynomialMultiplyRecognize::keepsHighBitsZero(Value *V, |
1273 | unsigned IterCount) { |
1274 | // Assume that all inputs to the value have the high bits zero. |
1275 | // Check if the value itself preserves the zeros in the high bits. |
1276 | if (auto *C = dyn_cast<ConstantInt>(Val: V)) |
1277 | return C->getValue().countl_zero() >= IterCount; |
1278 | |
1279 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
1280 | switch (I->getOpcode()) { |
1281 | case Instruction::And: |
1282 | case Instruction::Or: |
1283 | case Instruction::Xor: |
1284 | case Instruction::LShr: |
1285 | case Instruction::Select: |
1286 | case Instruction::ICmp: |
1287 | case Instruction::PHI: |
1288 | case Instruction::ZExt: |
1289 | return true; |
1290 | } |
1291 | } |
1292 | |
1293 | return false; |
1294 | } |
1295 | |
1296 | bool PolynomialMultiplyRecognize::isOperandShifted(Instruction *I, Value *Op) { |
1297 | unsigned Opc = I->getOpcode(); |
1298 | if (Opc == Instruction::Shl || Opc == Instruction::LShr) |
1299 | return Op != I->getOperand(i: 1); |
1300 | return true; |
1301 | } |
1302 | |
1303 | bool PolynomialMultiplyRecognize::convertShiftsToLeft(BasicBlock *LoopB, |
1304 | BasicBlock *ExitB, unsigned IterCount) { |
1305 | Value *CIV = getCountIV(BB: LoopB); |
1306 | if (CIV == nullptr) |
1307 | return false; |
1308 | auto *CIVTy = dyn_cast<IntegerType>(Val: CIV->getType()); |
1309 | if (CIVTy == nullptr) |
1310 | return false; |
1311 | |
1312 | ValueSeq RShifts; |
1313 | ValueSeq Early, Late, Cycled; |
1314 | |
1315 | // Find all value cycles that contain logical right shifts by 1. |
1316 | for (Instruction &I : *LoopB) { |
1317 | using namespace PatternMatch; |
1318 | |
1319 | Value *V = nullptr; |
1320 | if (!match(V: &I, P: m_LShr(L: m_Value(V), R: m_One()))) |
1321 | continue; |
1322 | ValueSeq C; |
1323 | if (!findCycle(Out: &I, In: V, Cycle&: C)) |
1324 | continue; |
1325 | |
1326 | // Found a cycle. |
1327 | C.insert(X: &I); |
1328 | classifyCycle(DivI: &I, Cycle&: C, Early, Late); |
1329 | Cycled.insert(Start: C.begin(), End: C.end()); |
1330 | RShifts.insert(X: &I); |
1331 | } |
1332 | |
1333 | // Find the set of all values affected by the shift cycles, i.e. all |
1334 | // cycled values, and (recursively) all their users. |
1335 | ValueSeq Users(Cycled.begin(), Cycled.end()); |
1336 | for (unsigned i = 0; i < Users.size(); ++i) { |
1337 | Value *V = Users[i]; |
1338 | if (!isa<IntegerType>(Val: V->getType())) |
1339 | return false; |
1340 | auto *R = cast<Instruction>(Val: V); |
1341 | // If the instruction does not commute with shifts, the loop cannot |
1342 | // be unshifted. |
1343 | if (!commutesWithShift(I: R)) |
1344 | return false; |
1345 | for (User *U : R->users()) { |
1346 | auto *T = cast<Instruction>(Val: U); |
1347 | // Skip users from outside of the loop. They will be handled later. |
1348 | // Also, skip the right-shifts and phi nodes, since they mix early |
1349 | // and late values. |
1350 | if (T->getParent() != LoopB || RShifts.count(key: T) || isa<PHINode>(Val: T)) |
1351 | continue; |
1352 | |
1353 | Users.insert(X: T); |
1354 | if (!classifyInst(UseI: T, Early, Late)) |
1355 | return false; |
1356 | } |
1357 | } |
1358 | |
1359 | if (Users.empty()) |
1360 | return false; |
1361 | |
1362 | // Verify that high bits remain zero. |
1363 | ValueSeq Internal(Users.begin(), Users.end()); |
1364 | ValueSeq Inputs; |
1365 | for (unsigned i = 0; i < Internal.size(); ++i) { |
1366 | auto *R = dyn_cast<Instruction>(Val: Internal[i]); |
1367 | if (!R) |
1368 | continue; |
1369 | for (Value *Op : R->operands()) { |
1370 | auto *T = dyn_cast<Instruction>(Val: Op); |
1371 | if (T && T->getParent() != LoopB) |
1372 | Inputs.insert(X: Op); |
1373 | else |
1374 | Internal.insert(X: Op); |
1375 | } |
1376 | } |
1377 | for (Value *V : Inputs) |
1378 | if (!highBitsAreZero(V, IterCount)) |
1379 | return false; |
1380 | for (Value *V : Internal) |
1381 | if (!keepsHighBitsZero(V, IterCount)) |
1382 | return false; |
1383 | |
1384 | // Finally, the work can be done. Unshift each user. |
1385 | IRBuilder<> IRB(LoopB); |
1386 | std::map<Value*,Value*> ShiftMap; |
1387 | |
1388 | using CastMapType = std::map<std::pair<Value *, Type *>, Value *>; |
1389 | |
1390 | CastMapType CastMap; |
1391 | |
1392 | auto upcast = [] (CastMapType &CM, IRBuilder<> &IRB, Value *V, |
1393 | IntegerType *Ty) -> Value* { |
1394 | auto H = CM.find(x: std::make_pair(x&: V, y&: Ty)); |
1395 | if (H != CM.end()) |
1396 | return H->second; |
1397 | Value *CV = IRB.CreateIntCast(V, DestTy: Ty, isSigned: false); |
1398 | CM.insert(x: std::make_pair(x: std::make_pair(x&: V, y&: Ty), y&: CV)); |
1399 | return CV; |
1400 | }; |
1401 | |
1402 | for (auto I = LoopB->begin(), E = LoopB->end(); I != E; ++I) { |
1403 | using namespace PatternMatch; |
1404 | |
1405 | if (isa<PHINode>(Val: I) || !Users.count(key: &*I)) |
1406 | continue; |
1407 | |
1408 | // Match lshr x, 1. |
1409 | Value *V = nullptr; |
1410 | if (match(V: &*I, P: m_LShr(L: m_Value(V), R: m_One()))) { |
1411 | replaceAllUsesOfWithIn(I: &*I, J: V, BB: LoopB); |
1412 | continue; |
1413 | } |
1414 | // For each non-cycled operand, replace it with the corresponding |
1415 | // value shifted left. |
1416 | for (auto &J : I->operands()) { |
1417 | Value *Op = J.get(); |
1418 | if (!isOperandShifted(I: &*I, Op)) |
1419 | continue; |
1420 | if (Users.count(key: Op)) |
1421 | continue; |
1422 | // Skip shifting zeros. |
1423 | if (isa<ConstantInt>(Val: Op) && cast<ConstantInt>(Val: Op)->isZero()) |
1424 | continue; |
1425 | // Check if we have already generated a shift for this value. |
1426 | auto F = ShiftMap.find(x: Op); |
1427 | Value *W = (F != ShiftMap.end()) ? F->second : nullptr; |
1428 | if (W == nullptr) { |
1429 | IRB.SetInsertPoint(&*I); |
1430 | // First, the shift amount will be CIV or CIV+1, depending on |
1431 | // whether the value is early or late. Instead of creating CIV+1, |
1432 | // do a single shift of the value. |
1433 | Value *ShAmt = CIV, *ShVal = Op; |
1434 | auto *VTy = cast<IntegerType>(Val: ShVal->getType()); |
1435 | auto *ATy = cast<IntegerType>(Val: ShAmt->getType()); |
1436 | if (Late.count(key: &*I)) |
1437 | ShVal = IRB.CreateShl(LHS: Op, RHS: ConstantInt::get(Ty: VTy, V: 1)); |
1438 | // Second, the types of the shifted value and the shift amount |
1439 | // must match. |
1440 | if (VTy != ATy) { |
1441 | if (VTy->getBitWidth() < ATy->getBitWidth()) |
1442 | ShVal = upcast(CastMap, IRB, ShVal, ATy); |
1443 | else |
1444 | ShAmt = upcast(CastMap, IRB, ShAmt, VTy); |
1445 | } |
1446 | // Ready to generate the shift and memoize it. |
1447 | W = IRB.CreateShl(LHS: ShVal, RHS: ShAmt); |
1448 | ShiftMap.insert(x: std::make_pair(x&: Op, y&: W)); |
1449 | } |
1450 | I->replaceUsesOfWith(From: Op, To: W); |
1451 | } |
1452 | } |
1453 | |
1454 | // Update the users outside of the loop to account for having left |
1455 | // shifts. They would normally be shifted right in the loop, so shift |
1456 | // them right after the loop exit. |
1457 | // Take advantage of the loop-closed SSA form, which has all the post- |
1458 | // loop values in phi nodes. |
1459 | IRB.SetInsertPoint(TheBB: ExitB, IP: ExitB->getFirstInsertionPt()); |
1460 | for (auto P = ExitB->begin(), Q = ExitB->end(); P != Q; ++P) { |
1461 | if (!isa<PHINode>(Val: P)) |
1462 | break; |
1463 | auto *PN = cast<PHINode>(Val&: P); |
1464 | Value *U = PN->getIncomingValueForBlock(BB: LoopB); |
1465 | if (!Users.count(key: U)) |
1466 | continue; |
1467 | Value *S = IRB.CreateLShr(LHS: PN, RHS: ConstantInt::get(Ty: PN->getType(), V: IterCount)); |
1468 | PN->replaceAllUsesWith(V: S); |
1469 | // The above RAUW will create |
1470 | // S = lshr S, IterCount |
1471 | // so we need to fix it back into |
1472 | // S = lshr PN, IterCount |
1473 | cast<User>(Val: S)->replaceUsesOfWith(From: S, To: PN); |
1474 | } |
1475 | |
1476 | return true; |
1477 | } |
1478 | |
1479 | void PolynomialMultiplyRecognize::cleanupLoopBody(BasicBlock *LoopB) { |
1480 | for (auto &I : *LoopB) |
1481 | if (Value *SV = simplifyInstruction(I: &I, Q: {DL, &TLI, &DT})) |
1482 | I.replaceAllUsesWith(V: SV); |
1483 | |
1484 | for (Instruction &I : llvm::make_early_inc_range(Range&: *LoopB)) |
1485 | RecursivelyDeleteTriviallyDeadInstructions(V: &I, TLI: &TLI); |
1486 | } |
1487 | |
1488 | unsigned PolynomialMultiplyRecognize::getInverseMxN(unsigned QP) { |
1489 | // Arrays of coefficients of Q and the inverse, C. |
1490 | // Q[i] = coefficient at x^i. |
1491 | std::array<char,32> Q, C; |
1492 | |
1493 | for (unsigned i = 0; i < 32; ++i) { |
1494 | Q[i] = QP & 1; |
1495 | QP >>= 1; |
1496 | } |
1497 | assert(Q[0] == 1); |
1498 | |
1499 | // Find C, such that |
1500 | // (Q[n]*x^n + ... + Q[1]*x + Q[0]) * (C[n]*x^n + ... + C[1]*x + C[0]) = 1 |
1501 | // |
1502 | // For it to have a solution, Q[0] must be 1. Since this is Z2[x], the |
1503 | // operations * and + are & and ^ respectively. |
1504 | // |
1505 | // Find C[i] recursively, by comparing i-th coefficient in the product |
1506 | // with 0 (or 1 for i=0). |
1507 | // |
1508 | // C[0] = 1, since C[0] = Q[0], and Q[0] = 1. |
1509 | C[0] = 1; |
1510 | for (unsigned i = 1; i < 32; ++i) { |
1511 | // Solve for C[i] in: |
1512 | // C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i]Q[0] = 0 |
1513 | // This is equivalent to |
1514 | // C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i] = 0 |
1515 | // which is |
1516 | // C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] = C[i] |
1517 | unsigned T = 0; |
1518 | for (unsigned j = 0; j < i; ++j) |
1519 | T = T ^ (C[j] & Q[i-j]); |
1520 | C[i] = T; |
1521 | } |
1522 | |
1523 | unsigned QV = 0; |
1524 | for (unsigned i = 0; i < 32; ++i) |
1525 | if (C[i]) |
1526 | QV |= (1 << i); |
1527 | |
1528 | return QV; |
1529 | } |
1530 | |
1531 | Value *PolynomialMultiplyRecognize::generate(BasicBlock::iterator At, |
1532 | ParsedValues &PV) { |
1533 | IRBuilder<> B(&*At); |
1534 | Module *M = At->getParent()->getParent()->getParent(); |
1535 | Function *PMF = Intrinsic::getDeclaration(M, id: Intrinsic::hexagon_M4_pmpyw); |
1536 | |
1537 | Value *P = PV.P, *Q = PV.Q, *P0 = P; |
1538 | unsigned IC = PV.IterCount; |
1539 | |
1540 | if (PV.M != nullptr) |
1541 | P0 = P = B.CreateXor(LHS: P, RHS: PV.M); |
1542 | |
1543 | // Create a bit mask to clear the high bits beyond IterCount. |
1544 | auto *BMI = ConstantInt::get(Ty: P->getType(), V: APInt::getLowBitsSet(numBits: 32, loBitsSet: IC)); |
1545 | |
1546 | if (PV.IterCount != 32) |
1547 | P = B.CreateAnd(LHS: P, RHS: BMI); |
1548 | |
1549 | if (PV.Inv) { |
1550 | auto *QI = dyn_cast<ConstantInt>(Val: PV.Q); |
1551 | assert(QI && QI->getBitWidth() <= 32); |
1552 | |
1553 | // Again, clearing bits beyond IterCount. |
1554 | unsigned M = (1 << PV.IterCount) - 1; |
1555 | unsigned Tmp = (QI->getZExtValue() | 1) & M; |
1556 | unsigned QV = getInverseMxN(QP: Tmp) & M; |
1557 | auto *QVI = ConstantInt::get(Ty: QI->getType(), V: QV); |
1558 | P = B.CreateCall(Callee: PMF, Args: {P, QVI}); |
1559 | P = B.CreateTrunc(V: P, DestTy: QI->getType()); |
1560 | if (IC != 32) |
1561 | P = B.CreateAnd(LHS: P, RHS: BMI); |
1562 | } |
1563 | |
1564 | Value *R = B.CreateCall(Callee: PMF, Args: {P, Q}); |
1565 | |
1566 | if (PV.M != nullptr) |
1567 | R = B.CreateXor(LHS: R, RHS: B.CreateIntCast(V: P0, DestTy: R->getType(), isSigned: false)); |
1568 | |
1569 | return R; |
1570 | } |
1571 | |
1572 | static bool hasZeroSignBit(const Value *V) { |
1573 | if (const auto *CI = dyn_cast<const ConstantInt>(Val: V)) |
1574 | return CI->getValue().isNonNegative(); |
1575 | const Instruction *I = dyn_cast<const Instruction>(Val: V); |
1576 | if (!I) |
1577 | return false; |
1578 | switch (I->getOpcode()) { |
1579 | case Instruction::LShr: |
1580 | if (const auto SI = dyn_cast<const ConstantInt>(Val: I->getOperand(i: 1))) |
1581 | return SI->getZExtValue() > 0; |
1582 | return false; |
1583 | case Instruction::Or: |
1584 | case Instruction::Xor: |
1585 | return hasZeroSignBit(V: I->getOperand(i: 0)) && |
1586 | hasZeroSignBit(V: I->getOperand(i: 1)); |
1587 | case Instruction::And: |
1588 | return hasZeroSignBit(V: I->getOperand(i: 0)) || |
1589 | hasZeroSignBit(V: I->getOperand(i: 1)); |
1590 | } |
1591 | return false; |
1592 | } |
1593 | |
1594 | void PolynomialMultiplyRecognize::setupPreSimplifier(Simplifier &S) { |
1595 | S.addRule(N: "sink-zext" , |
1596 | // Sink zext past bitwise operations. |
1597 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1598 | if (I->getOpcode() != Instruction::ZExt) |
1599 | return nullptr; |
1600 | Instruction *T = dyn_cast<Instruction>(Val: I->getOperand(i: 0)); |
1601 | if (!T) |
1602 | return nullptr; |
1603 | switch (T->getOpcode()) { |
1604 | case Instruction::And: |
1605 | case Instruction::Or: |
1606 | case Instruction::Xor: |
1607 | break; |
1608 | default: |
1609 | return nullptr; |
1610 | } |
1611 | IRBuilder<> B(Ctx); |
1612 | return B.CreateBinOp(Opc: cast<BinaryOperator>(Val: T)->getOpcode(), |
1613 | LHS: B.CreateZExt(V: T->getOperand(i: 0), DestTy: I->getType()), |
1614 | RHS: B.CreateZExt(V: T->getOperand(i: 1), DestTy: I->getType())); |
1615 | }); |
1616 | S.addRule(N: "xor/and -> and/xor" , |
1617 | // (xor (and x a) (and y a)) -> (and (xor x y) a) |
1618 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1619 | if (I->getOpcode() != Instruction::Xor) |
1620 | return nullptr; |
1621 | Instruction *And0 = dyn_cast<Instruction>(Val: I->getOperand(i: 0)); |
1622 | Instruction *And1 = dyn_cast<Instruction>(Val: I->getOperand(i: 1)); |
1623 | if (!And0 || !And1) |
1624 | return nullptr; |
1625 | if (And0->getOpcode() != Instruction::And || |
1626 | And1->getOpcode() != Instruction::And) |
1627 | return nullptr; |
1628 | if (And0->getOperand(i: 1) != And1->getOperand(i: 1)) |
1629 | return nullptr; |
1630 | IRBuilder<> B(Ctx); |
1631 | return B.CreateAnd(LHS: B.CreateXor(LHS: And0->getOperand(i: 0), RHS: And1->getOperand(i: 0)), |
1632 | RHS: And0->getOperand(i: 1)); |
1633 | }); |
1634 | S.addRule(N: "sink binop into select" , |
1635 | // (Op (select c x y) z) -> (select c (Op x z) (Op y z)) |
1636 | // (Op x (select c y z)) -> (select c (Op x y) (Op x z)) |
1637 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1638 | BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I); |
1639 | if (!BO) |
1640 | return nullptr; |
1641 | Instruction::BinaryOps Op = BO->getOpcode(); |
1642 | if (SelectInst *Sel = dyn_cast<SelectInst>(Val: BO->getOperand(i_nocapture: 0))) { |
1643 | IRBuilder<> B(Ctx); |
1644 | Value *X = Sel->getTrueValue(), *Y = Sel->getFalseValue(); |
1645 | Value *Z = BO->getOperand(i_nocapture: 1); |
1646 | return B.CreateSelect(C: Sel->getCondition(), |
1647 | True: B.CreateBinOp(Opc: Op, LHS: X, RHS: Z), |
1648 | False: B.CreateBinOp(Opc: Op, LHS: Y, RHS: Z)); |
1649 | } |
1650 | if (SelectInst *Sel = dyn_cast<SelectInst>(Val: BO->getOperand(i_nocapture: 1))) { |
1651 | IRBuilder<> B(Ctx); |
1652 | Value *X = BO->getOperand(i_nocapture: 0); |
1653 | Value *Y = Sel->getTrueValue(), *Z = Sel->getFalseValue(); |
1654 | return B.CreateSelect(C: Sel->getCondition(), |
1655 | True: B.CreateBinOp(Opc: Op, LHS: X, RHS: Y), |
1656 | False: B.CreateBinOp(Opc: Op, LHS: X, RHS: Z)); |
1657 | } |
1658 | return nullptr; |
1659 | }); |
1660 | S.addRule(N: "fold select-select" , |
1661 | // (select c (select c x y) z) -> (select c x z) |
1662 | // (select c x (select c y z)) -> (select c x z) |
1663 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1664 | SelectInst *Sel = dyn_cast<SelectInst>(Val: I); |
1665 | if (!Sel) |
1666 | return nullptr; |
1667 | IRBuilder<> B(Ctx); |
1668 | Value *C = Sel->getCondition(); |
1669 | if (SelectInst *Sel0 = dyn_cast<SelectInst>(Val: Sel->getTrueValue())) { |
1670 | if (Sel0->getCondition() == C) |
1671 | return B.CreateSelect(C, True: Sel0->getTrueValue(), False: Sel->getFalseValue()); |
1672 | } |
1673 | if (SelectInst *Sel1 = dyn_cast<SelectInst>(Val: Sel->getFalseValue())) { |
1674 | if (Sel1->getCondition() == C) |
1675 | return B.CreateSelect(C, True: Sel->getTrueValue(), False: Sel1->getFalseValue()); |
1676 | } |
1677 | return nullptr; |
1678 | }); |
1679 | S.addRule(N: "or-signbit -> xor-signbit" , |
1680 | // (or (lshr x 1) 0x800.0) -> (xor (lshr x 1) 0x800.0) |
1681 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1682 | if (I->getOpcode() != Instruction::Or) |
1683 | return nullptr; |
1684 | ConstantInt *Msb = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1)); |
1685 | if (!Msb || !Msb->getValue().isSignMask()) |
1686 | return nullptr; |
1687 | if (!hasZeroSignBit(V: I->getOperand(i: 0))) |
1688 | return nullptr; |
1689 | return IRBuilder<>(Ctx).CreateXor(LHS: I->getOperand(i: 0), RHS: Msb); |
1690 | }); |
1691 | S.addRule(N: "sink lshr into binop" , |
1692 | // (lshr (BitOp x y) c) -> (BitOp (lshr x c) (lshr y c)) |
1693 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1694 | if (I->getOpcode() != Instruction::LShr) |
1695 | return nullptr; |
1696 | BinaryOperator *BitOp = dyn_cast<BinaryOperator>(Val: I->getOperand(i: 0)); |
1697 | if (!BitOp) |
1698 | return nullptr; |
1699 | switch (BitOp->getOpcode()) { |
1700 | case Instruction::And: |
1701 | case Instruction::Or: |
1702 | case Instruction::Xor: |
1703 | break; |
1704 | default: |
1705 | return nullptr; |
1706 | } |
1707 | IRBuilder<> B(Ctx); |
1708 | Value *S = I->getOperand(i: 1); |
1709 | return B.CreateBinOp(Opc: BitOp->getOpcode(), |
1710 | LHS: B.CreateLShr(LHS: BitOp->getOperand(i_nocapture: 0), RHS: S), |
1711 | RHS: B.CreateLShr(LHS: BitOp->getOperand(i_nocapture: 1), RHS: S)); |
1712 | }); |
1713 | S.addRule(N: "expose bitop-const" , |
1714 | // (BitOp1 (BitOp2 x a) b) -> (BitOp2 x (BitOp1 a b)) |
1715 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1716 | auto IsBitOp = [](unsigned Op) -> bool { |
1717 | switch (Op) { |
1718 | case Instruction::And: |
1719 | case Instruction::Or: |
1720 | case Instruction::Xor: |
1721 | return true; |
1722 | } |
1723 | return false; |
1724 | }; |
1725 | BinaryOperator *BitOp1 = dyn_cast<BinaryOperator>(Val: I); |
1726 | if (!BitOp1 || !IsBitOp(BitOp1->getOpcode())) |
1727 | return nullptr; |
1728 | BinaryOperator *BitOp2 = dyn_cast<BinaryOperator>(Val: BitOp1->getOperand(i_nocapture: 0)); |
1729 | if (!BitOp2 || !IsBitOp(BitOp2->getOpcode())) |
1730 | return nullptr; |
1731 | ConstantInt *CA = dyn_cast<ConstantInt>(Val: BitOp2->getOperand(i_nocapture: 1)); |
1732 | ConstantInt *CB = dyn_cast<ConstantInt>(Val: BitOp1->getOperand(i_nocapture: 1)); |
1733 | if (!CA || !CB) |
1734 | return nullptr; |
1735 | IRBuilder<> B(Ctx); |
1736 | Value *X = BitOp2->getOperand(i_nocapture: 0); |
1737 | return B.CreateBinOp(Opc: BitOp2->getOpcode(), LHS: X, |
1738 | RHS: B.CreateBinOp(Opc: BitOp1->getOpcode(), LHS: CA, RHS: CB)); |
1739 | }); |
1740 | } |
1741 | |
1742 | void PolynomialMultiplyRecognize::setupPostSimplifier(Simplifier &S) { |
1743 | S.addRule(N: "(and (xor (and x a) y) b) -> (and (xor x y) b), if b == b&a" , |
1744 | F: [](Instruction *I, LLVMContext &Ctx) -> Value* { |
1745 | if (I->getOpcode() != Instruction::And) |
1746 | return nullptr; |
1747 | Instruction *Xor = dyn_cast<Instruction>(Val: I->getOperand(i: 0)); |
1748 | ConstantInt *C0 = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1)); |
1749 | if (!Xor || !C0) |
1750 | return nullptr; |
1751 | if (Xor->getOpcode() != Instruction::Xor) |
1752 | return nullptr; |
1753 | Instruction *And0 = dyn_cast<Instruction>(Val: Xor->getOperand(i: 0)); |
1754 | Instruction *And1 = dyn_cast<Instruction>(Val: Xor->getOperand(i: 1)); |
1755 | // Pick the first non-null and. |
1756 | if (!And0 || And0->getOpcode() != Instruction::And) |
1757 | std::swap(a&: And0, b&: And1); |
1758 | ConstantInt *C1 = dyn_cast<ConstantInt>(Val: And0->getOperand(i: 1)); |
1759 | if (!C1) |
1760 | return nullptr; |
1761 | uint32_t V0 = C0->getZExtValue(); |
1762 | uint32_t V1 = C1->getZExtValue(); |
1763 | if (V0 != (V0 & V1)) |
1764 | return nullptr; |
1765 | IRBuilder<> B(Ctx); |
1766 | return B.CreateAnd(LHS: B.CreateXor(LHS: And0->getOperand(i: 0), RHS: And1), RHS: C0); |
1767 | }); |
1768 | } |
1769 | |
1770 | bool PolynomialMultiplyRecognize::recognize() { |
1771 | LLVM_DEBUG(dbgs() << "Starting PolynomialMultiplyRecognize on loop\n" |
1772 | << *CurLoop << '\n'); |
1773 | // Restrictions: |
1774 | // - The loop must consist of a single block. |
1775 | // - The iteration count must be known at compile-time. |
1776 | // - The loop must have an induction variable starting from 0, and |
1777 | // incremented in each iteration of the loop. |
1778 | BasicBlock *LoopB = CurLoop->getHeader(); |
1779 | LLVM_DEBUG(dbgs() << "Loop header:\n" << *LoopB); |
1780 | |
1781 | if (LoopB != CurLoop->getLoopLatch()) |
1782 | return false; |
1783 | BasicBlock *ExitB = CurLoop->getExitBlock(); |
1784 | if (ExitB == nullptr) |
1785 | return false; |
1786 | BasicBlock *EntryB = CurLoop->getLoopPreheader(); |
1787 | if (EntryB == nullptr) |
1788 | return false; |
1789 | |
1790 | unsigned IterCount = 0; |
1791 | const SCEV *CT = SE.getBackedgeTakenCount(L: CurLoop); |
1792 | if (isa<SCEVCouldNotCompute>(Val: CT)) |
1793 | return false; |
1794 | if (auto *CV = dyn_cast<SCEVConstant>(Val: CT)) |
1795 | IterCount = CV->getValue()->getZExtValue() + 1; |
1796 | |
1797 | Value *CIV = getCountIV(BB: LoopB); |
1798 | ParsedValues PV; |
1799 | Simplifier PreSimp; |
1800 | PV.IterCount = IterCount; |
1801 | LLVM_DEBUG(dbgs() << "Loop IV: " << *CIV << "\nIterCount: " << IterCount |
1802 | << '\n'); |
1803 | |
1804 | setupPreSimplifier(PreSimp); |
1805 | |
1806 | // Perform a preliminary scan of select instructions to see if any of them |
1807 | // looks like a generator of the polynomial multiply steps. Assume that a |
1808 | // loop can only contain a single transformable operation, so stop the |
1809 | // traversal after the first reasonable candidate was found. |
1810 | // XXX: Currently this approach can modify the loop before being 100% sure |
1811 | // that the transformation can be carried out. |
1812 | bool FoundPreScan = false; |
1813 | auto FeedsPHI = [LoopB](const Value *V) -> bool { |
1814 | for (const Value *U : V->users()) { |
1815 | if (const auto *P = dyn_cast<const PHINode>(Val: U)) |
1816 | if (P->getParent() == LoopB) |
1817 | return true; |
1818 | } |
1819 | return false; |
1820 | }; |
1821 | for (Instruction &In : *LoopB) { |
1822 | SelectInst *SI = dyn_cast<SelectInst>(Val: &In); |
1823 | if (!SI || !FeedsPHI(SI)) |
1824 | continue; |
1825 | |
1826 | Simplifier::Context C(SI); |
1827 | Value *T = PreSimp.simplify(C); |
1828 | SelectInst *SelI = (T && isa<SelectInst>(Val: T)) ? cast<SelectInst>(Val: T) : SI; |
1829 | LLVM_DEBUG(dbgs() << "scanSelect(pre-scan): " << PE(C, SelI) << '\n'); |
1830 | if (scanSelect(SelI, LoopB, PrehB: EntryB, CIV, PV, PreScan: true)) { |
1831 | FoundPreScan = true; |
1832 | if (SelI != SI) { |
1833 | Value *NewSel = C.materialize(B: LoopB, At: SI->getIterator()); |
1834 | SI->replaceAllUsesWith(V: NewSel); |
1835 | RecursivelyDeleteTriviallyDeadInstructions(V: SI, TLI: &TLI); |
1836 | } |
1837 | break; |
1838 | } |
1839 | } |
1840 | |
1841 | if (!FoundPreScan) { |
1842 | LLVM_DEBUG(dbgs() << "Have not found candidates for pmpy\n" ); |
1843 | return false; |
1844 | } |
1845 | |
1846 | if (!PV.Left) { |
1847 | // The right shift version actually only returns the higher bits of |
1848 | // the result (each iteration discards the LSB). If we want to convert it |
1849 | // to a left-shifting loop, the working data type must be at least as |
1850 | // wide as the target's pmpy instruction. |
1851 | if (!promoteTypes(LoopB, ExitB)) |
1852 | return false; |
1853 | // Run post-promotion simplifications. |
1854 | Simplifier PostSimp; |
1855 | setupPostSimplifier(PostSimp); |
1856 | for (Instruction &In : *LoopB) { |
1857 | SelectInst *SI = dyn_cast<SelectInst>(Val: &In); |
1858 | if (!SI || !FeedsPHI(SI)) |
1859 | continue; |
1860 | Simplifier::Context C(SI); |
1861 | Value *T = PostSimp.simplify(C); |
1862 | SelectInst *SelI = dyn_cast_or_null<SelectInst>(Val: T); |
1863 | if (SelI != SI) { |
1864 | Value *NewSel = C.materialize(B: LoopB, At: SI->getIterator()); |
1865 | SI->replaceAllUsesWith(V: NewSel); |
1866 | RecursivelyDeleteTriviallyDeadInstructions(V: SI, TLI: &TLI); |
1867 | } |
1868 | break; |
1869 | } |
1870 | |
1871 | if (!convertShiftsToLeft(LoopB, ExitB, IterCount)) |
1872 | return false; |
1873 | cleanupLoopBody(LoopB); |
1874 | } |
1875 | |
1876 | // Scan the loop again, find the generating select instruction. |
1877 | bool FoundScan = false; |
1878 | for (Instruction &In : *LoopB) { |
1879 | SelectInst *SelI = dyn_cast<SelectInst>(Val: &In); |
1880 | if (!SelI) |
1881 | continue; |
1882 | LLVM_DEBUG(dbgs() << "scanSelect: " << *SelI << '\n'); |
1883 | FoundScan = scanSelect(SelI, LoopB, PrehB: EntryB, CIV, PV, PreScan: false); |
1884 | if (FoundScan) |
1885 | break; |
1886 | } |
1887 | assert(FoundScan); |
1888 | |
1889 | LLVM_DEBUG({ |
1890 | StringRef PP = (PV.M ? "(P+M)" : "P" ); |
1891 | if (!PV.Inv) |
1892 | dbgs() << "Found pmpy idiom: R = " << PP << ".Q\n" ; |
1893 | else |
1894 | dbgs() << "Found inverse pmpy idiom: R = (" << PP << "/Q).Q) + " |
1895 | << PP << "\n" ; |
1896 | dbgs() << " Res:" << *PV.Res << "\n P:" << *PV.P << "\n" ; |
1897 | if (PV.M) |
1898 | dbgs() << " M:" << *PV.M << "\n" ; |
1899 | dbgs() << " Q:" << *PV.Q << "\n" ; |
1900 | dbgs() << " Iteration count:" << PV.IterCount << "\n" ; |
1901 | }); |
1902 | |
1903 | BasicBlock::iterator At(EntryB->getTerminator()); |
1904 | Value *PM = generate(At, PV); |
1905 | if (PM == nullptr) |
1906 | return false; |
1907 | |
1908 | if (PM->getType() != PV.Res->getType()) |
1909 | PM = IRBuilder<>(&*At).CreateIntCast(V: PM, DestTy: PV.Res->getType(), isSigned: false); |
1910 | |
1911 | PV.Res->replaceAllUsesWith(V: PM); |
1912 | PV.Res->eraseFromParent(); |
1913 | return true; |
1914 | } |
1915 | |
1916 | int HexagonLoopIdiomRecognize::getSCEVStride(const SCEVAddRecExpr *S) { |
1917 | if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Val: S->getOperand(i: 1))) |
1918 | return SC->getAPInt().getSExtValue(); |
1919 | return 0; |
1920 | } |
1921 | |
1922 | bool HexagonLoopIdiomRecognize::isLegalStore(Loop *CurLoop, StoreInst *SI) { |
1923 | // Allow volatile stores if HexagonVolatileMemcpy is enabled. |
1924 | if (!(SI->isVolatile() && HexagonVolatileMemcpy) && !SI->isSimple()) |
1925 | return false; |
1926 | |
1927 | Value *StoredVal = SI->getValueOperand(); |
1928 | Value *StorePtr = SI->getPointerOperand(); |
1929 | |
1930 | // Reject stores that are so large that they overflow an unsigned. |
1931 | uint64_t SizeInBits = DL->getTypeSizeInBits(Ty: StoredVal->getType()); |
1932 | if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) |
1933 | return false; |
1934 | |
1935 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
1936 | // loop, which indicates a strided store. If we have something else, it's a |
1937 | // random store we can't handle. |
1938 | auto *StoreEv = dyn_cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: StorePtr)); |
1939 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
1940 | return false; |
1941 | |
1942 | // Check to see if the stride matches the size of the store. If so, then we |
1943 | // know that every byte is touched in the loop. |
1944 | int Stride = getSCEVStride(S: StoreEv); |
1945 | if (Stride == 0) |
1946 | return false; |
1947 | unsigned StoreSize = DL->getTypeStoreSize(Ty: SI->getValueOperand()->getType()); |
1948 | if (StoreSize != unsigned(std::abs(x: Stride))) |
1949 | return false; |
1950 | |
1951 | // The store must be feeding a non-volatile load. |
1952 | LoadInst *LI = dyn_cast<LoadInst>(Val: SI->getValueOperand()); |
1953 | if (!LI || !LI->isSimple()) |
1954 | return false; |
1955 | |
1956 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
1957 | // loop, which indicates a strided load. If we have something else, it's a |
1958 | // random load we can't handle. |
1959 | Value *LoadPtr = LI->getPointerOperand(); |
1960 | auto *LoadEv = dyn_cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: LoadPtr)); |
1961 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
1962 | return false; |
1963 | |
1964 | // The store and load must share the same stride. |
1965 | if (StoreEv->getOperand(i: 1) != LoadEv->getOperand(i: 1)) |
1966 | return false; |
1967 | |
1968 | // Success. This store can be converted into a memcpy. |
1969 | return true; |
1970 | } |
1971 | |
1972 | /// mayLoopAccessLocation - Return true if the specified loop might access the |
1973 | /// specified pointer location, which is a loop-strided access. The 'Access' |
1974 | /// argument specifies what the verboten forms of access are (read or write). |
1975 | static bool |
1976 | mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, |
1977 | const SCEV *BECount, unsigned StoreSize, |
1978 | AliasAnalysis &AA, |
1979 | SmallPtrSetImpl<Instruction *> &Ignored) { |
1980 | // Get the location that may be stored across the loop. Since the access |
1981 | // is strided positively through memory, we say that the modified location |
1982 | // starts at the pointer and has infinite size. |
1983 | LocationSize AccessSize = LocationSize::afterPointer(); |
1984 | |
1985 | // If the loop iterates a fixed number of times, we can refine the access |
1986 | // size to be exactly the size of the memset, which is (BECount+1)*StoreSize |
1987 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(Val: BECount)) |
1988 | AccessSize = LocationSize::precise(Value: (BECst->getValue()->getZExtValue() + 1) * |
1989 | StoreSize); |
1990 | |
1991 | // TODO: For this to be really effective, we have to dive into the pointer |
1992 | // operand in the store. Store to &A[i] of 100 will always return may alias |
1993 | // with store of &A[100], we need to StoreLoc to be "A" with size of 100, |
1994 | // which will then no-alias a store to &A[100]. |
1995 | MemoryLocation StoreLoc(Ptr, AccessSize); |
1996 | |
1997 | for (auto *B : L->blocks()) |
1998 | for (auto &I : *B) |
1999 | if (Ignored.count(Ptr: &I) == 0 && |
2000 | isModOrRefSet(MRI: AA.getModRefInfo(I: &I, OptLoc: StoreLoc) & Access)) |
2001 | return true; |
2002 | |
2003 | return false; |
2004 | } |
2005 | |
2006 | void HexagonLoopIdiomRecognize::collectStores(Loop *CurLoop, BasicBlock *BB, |
2007 | SmallVectorImpl<StoreInst*> &Stores) { |
2008 | Stores.clear(); |
2009 | for (Instruction &I : *BB) |
2010 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: &I)) |
2011 | if (isLegalStore(CurLoop, SI)) |
2012 | Stores.push_back(Elt: SI); |
2013 | } |
2014 | |
2015 | bool HexagonLoopIdiomRecognize::processCopyingStore(Loop *CurLoop, |
2016 | StoreInst *SI, const SCEV *BECount) { |
2017 | assert((SI->isSimple() || (SI->isVolatile() && HexagonVolatileMemcpy)) && |
2018 | "Expected only non-volatile stores, or Hexagon-specific memcpy" |
2019 | "to volatile destination." ); |
2020 | |
2021 | Value *StorePtr = SI->getPointerOperand(); |
2022 | auto *StoreEv = cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: StorePtr)); |
2023 | unsigned Stride = getSCEVStride(S: StoreEv); |
2024 | unsigned StoreSize = DL->getTypeStoreSize(Ty: SI->getValueOperand()->getType()); |
2025 | if (Stride != StoreSize) |
2026 | return false; |
2027 | |
2028 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
2029 | // loop, which indicates a strided load. If we have something else, it's a |
2030 | // random load we can't handle. |
2031 | auto *LI = cast<LoadInst>(Val: SI->getValueOperand()); |
2032 | auto *LoadEv = cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: LI->getPointerOperand())); |
2033 | |
2034 | // The trip count of the loop and the base pointer of the addrec SCEV is |
2035 | // guaranteed to be loop invariant, which means that it should dominate the |
2036 | // header. This allows us to insert code for it in the preheader. |
2037 | BasicBlock * = CurLoop->getLoopPreheader(); |
2038 | Instruction *ExpPt = Preheader->getTerminator(); |
2039 | IRBuilder<> Builder(ExpPt); |
2040 | SCEVExpander Expander(*SE, *DL, "hexagon-loop-idiom" ); |
2041 | |
2042 | Type *IntPtrTy = Builder.getIntPtrTy(DL: *DL, AddrSpace: SI->getPointerAddressSpace()); |
2043 | |
2044 | // Okay, we have a strided store "p[i]" of a loaded value. We can turn |
2045 | // this into a memcpy/memmove in the loop preheader now if we want. However, |
2046 | // this would be unsafe to do if there is anything else in the loop that may |
2047 | // read or write the memory region we're storing to. For memcpy, this |
2048 | // includes the load that feeds the stores. Check for an alias by generating |
2049 | // the base address and checking everything. |
2050 | Value *StoreBasePtr = Expander.expandCodeFor(SH: StoreEv->getStart(), |
2051 | Ty: Builder.getPtrTy(AddrSpace: SI->getPointerAddressSpace()), I: ExpPt); |
2052 | Value *LoadBasePtr = nullptr; |
2053 | |
2054 | bool Overlap = false; |
2055 | bool DestVolatile = SI->isVolatile(); |
2056 | Type *BECountTy = BECount->getType(); |
2057 | |
2058 | if (DestVolatile) { |
2059 | // The trip count must fit in i32, since it is the type of the "num_words" |
2060 | // argument to hexagon_memcpy_forward_vp4cp4n2. |
2061 | if (StoreSize != 4 || DL->getTypeSizeInBits(Ty: BECountTy) > 32) { |
2062 | CleanupAndExit: |
2063 | // If we generated new code for the base pointer, clean up. |
2064 | Expander.clear(); |
2065 | if (StoreBasePtr && (LoadBasePtr != StoreBasePtr)) { |
2066 | RecursivelyDeleteTriviallyDeadInstructions(V: StoreBasePtr, TLI); |
2067 | StoreBasePtr = nullptr; |
2068 | } |
2069 | if (LoadBasePtr) { |
2070 | RecursivelyDeleteTriviallyDeadInstructions(V: LoadBasePtr, TLI); |
2071 | LoadBasePtr = nullptr; |
2072 | } |
2073 | return false; |
2074 | } |
2075 | } |
2076 | |
2077 | SmallPtrSet<Instruction*, 2> Ignore1; |
2078 | Ignore1.insert(Ptr: SI); |
2079 | if (mayLoopAccessLocation(Ptr: StoreBasePtr, Access: ModRefInfo::ModRef, L: CurLoop, BECount, |
2080 | StoreSize, AA&: *AA, Ignored&: Ignore1)) { |
2081 | // Check if the load is the offending instruction. |
2082 | Ignore1.insert(Ptr: LI); |
2083 | if (mayLoopAccessLocation(Ptr: StoreBasePtr, Access: ModRefInfo::ModRef, L: CurLoop, |
2084 | BECount, StoreSize, AA&: *AA, Ignored&: Ignore1)) { |
2085 | // Still bad. Nothing we can do. |
2086 | goto CleanupAndExit; |
2087 | } |
2088 | // It worked with the load ignored. |
2089 | Overlap = true; |
2090 | } |
2091 | |
2092 | if (!Overlap) { |
2093 | if (DisableMemcpyIdiom || !HasMemcpy) |
2094 | goto CleanupAndExit; |
2095 | } else { |
2096 | // Don't generate memmove if this function will be inlined. This is |
2097 | // because the caller will undergo this transformation after inlining. |
2098 | Function *Func = CurLoop->getHeader()->getParent(); |
2099 | if (Func->hasFnAttribute(Kind: Attribute::AlwaysInline)) |
2100 | goto CleanupAndExit; |
2101 | |
2102 | // In case of a memmove, the call to memmove will be executed instead |
2103 | // of the loop, so we need to make sure that there is nothing else in |
2104 | // the loop than the load, store and instructions that these two depend |
2105 | // on. |
2106 | SmallVector<Instruction*,2> Insts; |
2107 | Insts.push_back(Elt: SI); |
2108 | Insts.push_back(Elt: LI); |
2109 | if (!coverLoop(L: CurLoop, Insts)) |
2110 | goto CleanupAndExit; |
2111 | |
2112 | if (DisableMemmoveIdiom || !HasMemmove) |
2113 | goto CleanupAndExit; |
2114 | bool IsNested = CurLoop->getParentLoop() != nullptr; |
2115 | if (IsNested && OnlyNonNestedMemmove) |
2116 | goto CleanupAndExit; |
2117 | } |
2118 | |
2119 | // For a memcpy, we have to make sure that the input array is not being |
2120 | // mutated by the loop. |
2121 | LoadBasePtr = Expander.expandCodeFor(SH: LoadEv->getStart(), |
2122 | Ty: Builder.getPtrTy(AddrSpace: LI->getPointerAddressSpace()), I: ExpPt); |
2123 | |
2124 | SmallPtrSet<Instruction*, 2> Ignore2; |
2125 | Ignore2.insert(Ptr: SI); |
2126 | if (mayLoopAccessLocation(Ptr: LoadBasePtr, Access: ModRefInfo::Mod, L: CurLoop, BECount, |
2127 | StoreSize, AA&: *AA, Ignored&: Ignore2)) |
2128 | goto CleanupAndExit; |
2129 | |
2130 | // Check the stride. |
2131 | bool StridePos = getSCEVStride(S: LoadEv) >= 0; |
2132 | |
2133 | // Currently, the volatile memcpy only emulates traversing memory forward. |
2134 | if (!StridePos && DestVolatile) |
2135 | goto CleanupAndExit; |
2136 | |
2137 | bool RuntimeCheck = (Overlap || DestVolatile); |
2138 | |
2139 | BasicBlock *ExitB; |
2140 | if (RuntimeCheck) { |
2141 | // The runtime check needs a single exit block. |
2142 | SmallVector<BasicBlock*, 8> ExitBlocks; |
2143 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
2144 | if (ExitBlocks.size() != 1) |
2145 | goto CleanupAndExit; |
2146 | ExitB = ExitBlocks[0]; |
2147 | } |
2148 | |
2149 | // The # stored bytes is (BECount+1)*Size. Expand the trip count out to |
2150 | // pointer size if it isn't already. |
2151 | LLVMContext &Ctx = SI->getContext(); |
2152 | BECount = SE->getTruncateOrZeroExtend(V: BECount, Ty: IntPtrTy); |
2153 | DebugLoc DLoc = SI->getDebugLoc(); |
2154 | |
2155 | const SCEV *NumBytesS = |
2156 | SE->getAddExpr(LHS: BECount, RHS: SE->getOne(Ty: IntPtrTy), Flags: SCEV::FlagNUW); |
2157 | if (StoreSize != 1) |
2158 | NumBytesS = SE->getMulExpr(LHS: NumBytesS, RHS: SE->getConstant(Ty: IntPtrTy, V: StoreSize), |
2159 | Flags: SCEV::FlagNUW); |
2160 | Value *NumBytes = Expander.expandCodeFor(SH: NumBytesS, Ty: IntPtrTy, I: ExpPt); |
2161 | if (Instruction *In = dyn_cast<Instruction>(Val: NumBytes)) |
2162 | if (Value *Simp = simplifyInstruction(I: In, Q: {*DL, TLI, DT})) |
2163 | NumBytes = Simp; |
2164 | |
2165 | CallInst *NewCall; |
2166 | |
2167 | if (RuntimeCheck) { |
2168 | unsigned Threshold = RuntimeMemSizeThreshold; |
2169 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: NumBytes)) { |
2170 | uint64_t C = CI->getZExtValue(); |
2171 | if (Threshold != 0 && C < Threshold) |
2172 | goto CleanupAndExit; |
2173 | if (C < CompileTimeMemSizeThreshold) |
2174 | goto CleanupAndExit; |
2175 | } |
2176 | |
2177 | BasicBlock * = CurLoop->getHeader(); |
2178 | Function *Func = Header->getParent(); |
2179 | Loop *ParentL = LF->getLoopFor(BB: Preheader); |
2180 | StringRef = Header->getName(); |
2181 | |
2182 | // Create a new (empty) preheader, and update the PHI nodes in the |
2183 | // header to use the new preheader. |
2184 | BasicBlock * = BasicBlock::Create(Context&: Ctx, Name: HeaderName+".rtli.ph" , |
2185 | Parent: Func, InsertBefore: Header); |
2186 | if (ParentL) |
2187 | ParentL->addBasicBlockToLoop(NewBB: NewPreheader, LI&: *LF); |
2188 | IRBuilder<>(NewPreheader).CreateBr(Dest: Header); |
2189 | for (auto &In : *Header) { |
2190 | PHINode *PN = dyn_cast<PHINode>(Val: &In); |
2191 | if (!PN) |
2192 | break; |
2193 | int bx = PN->getBasicBlockIndex(BB: Preheader); |
2194 | if (bx >= 0) |
2195 | PN->setIncomingBlock(i: bx, BB: NewPreheader); |
2196 | } |
2197 | DT->addNewBlock(BB: NewPreheader, DomBB: Preheader); |
2198 | DT->changeImmediateDominator(BB: Header, NewBB: NewPreheader); |
2199 | |
2200 | // Check for safe conditions to execute memmove. |
2201 | // If stride is positive, copying things from higher to lower addresses |
2202 | // is equivalent to memmove. For negative stride, it's the other way |
2203 | // around. Copying forward in memory with positive stride may not be |
2204 | // same as memmove since we may be copying values that we just stored |
2205 | // in some previous iteration. |
2206 | Value *LA = Builder.CreatePtrToInt(V: LoadBasePtr, DestTy: IntPtrTy); |
2207 | Value *SA = Builder.CreatePtrToInt(V: StoreBasePtr, DestTy: IntPtrTy); |
2208 | Value *LowA = StridePos ? SA : LA; |
2209 | Value *HighA = StridePos ? LA : SA; |
2210 | Value *CmpA = Builder.CreateICmpULT(LHS: LowA, RHS: HighA); |
2211 | Value *Cond = CmpA; |
2212 | |
2213 | // Check for distance between pointers. Since the case LowA < HighA |
2214 | // is checked for above, assume LowA >= HighA. |
2215 | Value *Dist = Builder.CreateSub(LHS: LowA, RHS: HighA); |
2216 | Value *CmpD = Builder.CreateICmpSLE(LHS: NumBytes, RHS: Dist); |
2217 | Value *CmpEither = Builder.CreateOr(LHS: Cond, RHS: CmpD); |
2218 | Cond = CmpEither; |
2219 | |
2220 | if (Threshold != 0) { |
2221 | Type *Ty = NumBytes->getType(); |
2222 | Value *Thr = ConstantInt::get(Ty, V: Threshold); |
2223 | Value *CmpB = Builder.CreateICmpULT(LHS: Thr, RHS: NumBytes); |
2224 | Value *CmpBoth = Builder.CreateAnd(LHS: Cond, RHS: CmpB); |
2225 | Cond = CmpBoth; |
2226 | } |
2227 | BasicBlock *MemmoveB = BasicBlock::Create(Context&: Ctx, Name: Header->getName()+".rtli" , |
2228 | Parent: Func, InsertBefore: NewPreheader); |
2229 | if (ParentL) |
2230 | ParentL->addBasicBlockToLoop(NewBB: MemmoveB, LI&: *LF); |
2231 | Instruction *OldT = Preheader->getTerminator(); |
2232 | Builder.CreateCondBr(Cond, True: MemmoveB, False: NewPreheader); |
2233 | OldT->eraseFromParent(); |
2234 | Preheader->setName(Preheader->getName()+".old" ); |
2235 | DT->addNewBlock(BB: MemmoveB, DomBB: Preheader); |
2236 | // Find the new immediate dominator of the exit block. |
2237 | BasicBlock *ExitD = Preheader; |
2238 | for (BasicBlock *PB : predecessors(BB: ExitB)) { |
2239 | ExitD = DT->findNearestCommonDominator(A: ExitD, B: PB); |
2240 | if (!ExitD) |
2241 | break; |
2242 | } |
2243 | // If the prior immediate dominator of ExitB was dominated by the |
2244 | // old preheader, then the old preheader becomes the new immediate |
2245 | // dominator. Otherwise don't change anything (because the newly |
2246 | // added blocks are dominated by the old preheader). |
2247 | if (ExitD && DT->dominates(A: Preheader, B: ExitD)) { |
2248 | DomTreeNode *BN = DT->getNode(BB: ExitB); |
2249 | DomTreeNode *DN = DT->getNode(BB: ExitD); |
2250 | BN->setIDom(DN); |
2251 | } |
2252 | |
2253 | // Add a call to memmove to the conditional block. |
2254 | IRBuilder<> CondBuilder(MemmoveB); |
2255 | CondBuilder.CreateBr(Dest: ExitB); |
2256 | CondBuilder.SetInsertPoint(MemmoveB->getTerminator()); |
2257 | |
2258 | if (DestVolatile) { |
2259 | Type *Int32Ty = Type::getInt32Ty(C&: Ctx); |
2260 | Type *PtrTy = PointerType::get(C&: Ctx, AddressSpace: 0); |
2261 | Type *VoidTy = Type::getVoidTy(C&: Ctx); |
2262 | Module *M = Func->getParent(); |
2263 | FunctionCallee Fn = M->getOrInsertFunction( |
2264 | Name: HexagonVolatileMemcpyName, RetTy: VoidTy, Args: PtrTy, Args: PtrTy, Args: Int32Ty); |
2265 | |
2266 | const SCEV *OneS = SE->getConstant(Ty: Int32Ty, V: 1); |
2267 | const SCEV *BECount32 = SE->getTruncateOrZeroExtend(V: BECount, Ty: Int32Ty); |
2268 | const SCEV *NumWordsS = SE->getAddExpr(LHS: BECount32, RHS: OneS, Flags: SCEV::FlagNUW); |
2269 | Value *NumWords = Expander.expandCodeFor(SH: NumWordsS, Ty: Int32Ty, |
2270 | I: MemmoveB->getTerminator()); |
2271 | if (Instruction *In = dyn_cast<Instruction>(Val: NumWords)) |
2272 | if (Value *Simp = simplifyInstruction(I: In, Q: {*DL, TLI, DT})) |
2273 | NumWords = Simp; |
2274 | |
2275 | NewCall = CondBuilder.CreateCall(Callee: Fn, |
2276 | Args: {StoreBasePtr, LoadBasePtr, NumWords}); |
2277 | } else { |
2278 | NewCall = CondBuilder.CreateMemMove( |
2279 | Dst: StoreBasePtr, DstAlign: SI->getAlign(), Src: LoadBasePtr, SrcAlign: LI->getAlign(), Size: NumBytes); |
2280 | } |
2281 | } else { |
2282 | NewCall = Builder.CreateMemCpy(Dst: StoreBasePtr, DstAlign: SI->getAlign(), Src: LoadBasePtr, |
2283 | SrcAlign: LI->getAlign(), Size: NumBytes); |
2284 | // Okay, the memcpy has been formed. Zap the original store and |
2285 | // anything that feeds into it. |
2286 | RecursivelyDeleteTriviallyDeadInstructions(V: SI, TLI); |
2287 | } |
2288 | |
2289 | NewCall->setDebugLoc(DLoc); |
2290 | |
2291 | LLVM_DEBUG(dbgs() << " Formed " << (Overlap ? "memmove: " : "memcpy: " ) |
2292 | << *NewCall << "\n" |
2293 | << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" |
2294 | << " from store ptr=" << *StoreEv << " at: " << *SI |
2295 | << "\n" ); |
2296 | |
2297 | return true; |
2298 | } |
2299 | |
2300 | // Check if the instructions in Insts, together with their dependencies |
2301 | // cover the loop in the sense that the loop could be safely eliminated once |
2302 | // the instructions in Insts are removed. |
2303 | bool HexagonLoopIdiomRecognize::coverLoop(Loop *L, |
2304 | SmallVectorImpl<Instruction*> &Insts) const { |
2305 | SmallSet<BasicBlock*,8> LoopBlocks; |
2306 | for (auto *B : L->blocks()) |
2307 | LoopBlocks.insert(Ptr: B); |
2308 | |
2309 | SetVector<Instruction*> Worklist(Insts.begin(), Insts.end()); |
2310 | |
2311 | // Collect all instructions from the loop that the instructions in Insts |
2312 | // depend on (plus their dependencies, etc.). These instructions will |
2313 | // constitute the expression trees that feed those in Insts, but the trees |
2314 | // will be limited only to instructions contained in the loop. |
2315 | for (unsigned i = 0; i < Worklist.size(); ++i) { |
2316 | Instruction *In = Worklist[i]; |
2317 | for (auto I = In->op_begin(), E = In->op_end(); I != E; ++I) { |
2318 | Instruction *OpI = dyn_cast<Instruction>(Val: I); |
2319 | if (!OpI) |
2320 | continue; |
2321 | BasicBlock *PB = OpI->getParent(); |
2322 | if (!LoopBlocks.count(Ptr: PB)) |
2323 | continue; |
2324 | Worklist.insert(X: OpI); |
2325 | } |
2326 | } |
2327 | |
2328 | // Scan all instructions in the loop, if any of them have a user outside |
2329 | // of the loop, or outside of the expressions collected above, then either |
2330 | // the loop has a side-effect visible outside of it, or there are |
2331 | // instructions in it that are not involved in the original set Insts. |
2332 | for (auto *B : L->blocks()) { |
2333 | for (auto &In : *B) { |
2334 | if (isa<BranchInst>(Val: In) || isa<DbgInfoIntrinsic>(Val: In)) |
2335 | continue; |
2336 | if (!Worklist.count(key: &In) && In.mayHaveSideEffects()) |
2337 | return false; |
2338 | for (auto *K : In.users()) { |
2339 | Instruction *UseI = dyn_cast<Instruction>(Val: K); |
2340 | if (!UseI) |
2341 | continue; |
2342 | BasicBlock *UseB = UseI->getParent(); |
2343 | if (LF->getLoopFor(BB: UseB) != L) |
2344 | return false; |
2345 | } |
2346 | } |
2347 | } |
2348 | |
2349 | return true; |
2350 | } |
2351 | |
2352 | /// runOnLoopBlock - Process the specified block, which lives in a counted loop |
2353 | /// with the specified backedge count. This block is known to be in the current |
2354 | /// loop and not in any subloops. |
2355 | bool HexagonLoopIdiomRecognize::runOnLoopBlock(Loop *CurLoop, BasicBlock *BB, |
2356 | const SCEV *BECount, SmallVectorImpl<BasicBlock*> &ExitBlocks) { |
2357 | // We can only promote stores in this block if they are unconditionally |
2358 | // executed in the loop. For a block to be unconditionally executed, it has |
2359 | // to dominate all the exit blocks of the loop. Verify this now. |
2360 | auto DominatedByBB = [this,BB] (BasicBlock *EB) -> bool { |
2361 | return DT->dominates(A: BB, B: EB); |
2362 | }; |
2363 | if (!all_of(Range&: ExitBlocks, P: DominatedByBB)) |
2364 | return false; |
2365 | |
2366 | bool MadeChange = false; |
2367 | // Look for store instructions, which may be optimized to memset/memcpy. |
2368 | SmallVector<StoreInst*,8> Stores; |
2369 | collectStores(CurLoop, BB, Stores); |
2370 | |
2371 | // Optimize the store into a memcpy, if it feeds an similarly strided load. |
2372 | for (auto &SI : Stores) |
2373 | MadeChange |= processCopyingStore(CurLoop, SI, BECount); |
2374 | |
2375 | return MadeChange; |
2376 | } |
2377 | |
2378 | bool HexagonLoopIdiomRecognize::runOnCountableLoop(Loop *L) { |
2379 | PolynomialMultiplyRecognize PMR(L, *DL, *DT, *TLI, *SE); |
2380 | if (PMR.recognize()) |
2381 | return true; |
2382 | |
2383 | if (!HasMemcpy && !HasMemmove) |
2384 | return false; |
2385 | |
2386 | const SCEV *BECount = SE->getBackedgeTakenCount(L); |
2387 | assert(!isa<SCEVCouldNotCompute>(BECount) && |
2388 | "runOnCountableLoop() called on a loop without a predictable" |
2389 | "backedge-taken count" ); |
2390 | |
2391 | SmallVector<BasicBlock *, 8> ExitBlocks; |
2392 | L->getUniqueExitBlocks(ExitBlocks); |
2393 | |
2394 | bool Changed = false; |
2395 | |
2396 | // Scan all the blocks in the loop that are not in subloops. |
2397 | for (auto *BB : L->getBlocks()) { |
2398 | // Ignore blocks in subloops. |
2399 | if (LF->getLoopFor(BB) != L) |
2400 | continue; |
2401 | Changed |= runOnLoopBlock(CurLoop: L, BB, BECount, ExitBlocks); |
2402 | } |
2403 | |
2404 | return Changed; |
2405 | } |
2406 | |
2407 | bool HexagonLoopIdiomRecognize::run(Loop *L) { |
2408 | const Module &M = *L->getHeader()->getParent()->getParent(); |
2409 | if (Triple(M.getTargetTriple()).getArch() != Triple::hexagon) |
2410 | return false; |
2411 | |
2412 | // If the loop could not be converted to canonical form, it must have an |
2413 | // indirectbr in it, just give up. |
2414 | if (!L->getLoopPreheader()) |
2415 | return false; |
2416 | |
2417 | // Disable loop idiom recognition if the function's name is a common idiom. |
2418 | StringRef Name = L->getHeader()->getParent()->getName(); |
2419 | if (Name == "memset" || Name == "memcpy" || Name == "memmove" ) |
2420 | return false; |
2421 | |
2422 | DL = &L->getHeader()->getDataLayout(); |
2423 | |
2424 | HasMemcpy = TLI->has(F: LibFunc_memcpy); |
2425 | HasMemmove = TLI->has(F: LibFunc_memmove); |
2426 | |
2427 | if (SE->hasLoopInvariantBackedgeTakenCount(L)) |
2428 | return runOnCountableLoop(L); |
2429 | return false; |
2430 | } |
2431 | |
2432 | bool HexagonLoopIdiomRecognizeLegacyPass::runOnLoop(Loop *L, |
2433 | LPPassManager &LPM) { |
2434 | if (skipLoop(L)) |
2435 | return false; |
2436 | |
2437 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
2438 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
2439 | auto *LF = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
2440 | auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( |
2441 | F: *L->getHeader()->getParent()); |
2442 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
2443 | return HexagonLoopIdiomRecognize(AA, DT, LF, TLI, SE).run(L); |
2444 | } |
2445 | |
2446 | Pass *llvm::createHexagonLoopIdiomPass() { |
2447 | return new HexagonLoopIdiomRecognizeLegacyPass(); |
2448 | } |
2449 | |
2450 | PreservedAnalyses |
2451 | HexagonLoopIdiomRecognitionPass::run(Loop &L, LoopAnalysisManager &AM, |
2452 | LoopStandardAnalysisResults &AR, |
2453 | LPMUpdater &U) { |
2454 | return HexagonLoopIdiomRecognize(&AR.AA, &AR.DT, &AR.LI, &AR.TLI, &AR.SE) |
2455 | .run(L: &L) |
2456 | ? getLoopPassPreservedAnalyses() |
2457 | : PreservedAnalyses::all(); |
2458 | } |
2459 | |