1 //===---- llvm/Analysis/ScalarEvolutionExpander.h - SCEV Exprs --*- C++ -*-===//
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
7 //===----------------------------------------------------------------------===//
9 // This file defines the classes used to generate code from scalar expressions.
11 //===----------------------------------------------------------------------===//
13 #ifndef LLVM_TRANSFORMS_UTILS_SCALAREVOLUTIONEXPANDER_H
14 #define LLVM_TRANSFORMS_UTILS_SCALAREVOLUTIONEXPANDER_H
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/Optional.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
22 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
23 #include "llvm/Analysis/TargetFolder.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/ValueHandle.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/InstructionCost.h"
31 extern cl::opt<unsigned> SCEVCheapExpansionBudget;
33 /// Return true if the given expression is safe to expand in the sense that
34 /// all materialized values are safe to speculate anywhere their operands are
36 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE);
38 /// Return true if the given expression is safe to expand in the sense that
39 /// all materialized values are defined and safe to speculate at the specified
40 /// location and their operands are defined at this location.
41 bool isSafeToExpandAt(const SCEV *S, const Instruction *InsertionPoint,
44 /// struct for holding enough information to help calculate the cost of the
45 /// given SCEV when expanded into IR.
47 explicit SCEVOperand(unsigned Opc, int Idx, const SCEV *S) :
48 ParentOpcode(Opc), OperandIdx(Idx), S(S) { }
49 /// LLVM instruction opcode that uses the operand.
50 unsigned ParentOpcode;
51 /// The use index of an expanded instruction.
53 /// The SCEV operand to be costed.
57 /// This class uses information about analyze scalars to rewrite expressions
58 /// in canonical form.
60 /// Clients should create an instance of this class when rewriting is needed,
61 /// and destroy it when finished to allow the release of the associated
63 class SCEVExpander : public SCEVVisitor<SCEVExpander, Value *> {
67 // New instructions receive a name to identify them with the current pass.
70 /// Indicates whether LCSSA phis should be created for inserted values.
73 // InsertedExpressions caches Values for reuse, so must track RAUW.
74 DenseMap<std::pair<const SCEV *, Instruction *>, TrackingVH<Value>>
77 // InsertedValues only flags inserted instructions so needs no RAUW.
78 DenseSet<AssertingVH<Value>> InsertedValues;
79 DenseSet<AssertingVH<Value>> InsertedPostIncValues;
81 /// Keep track of the existing IR values re-used during expansion.
82 /// FIXME: Ideally re-used instructions would not be added to
83 /// InsertedValues/InsertedPostIncValues.
84 SmallPtrSet<Value *, 16> ReusedValues;
86 // The induction variables generated.
87 SmallVector<WeakVH, 2> InsertedIVs;
89 /// A memoization of the "relevant" loop for a given SCEV.
90 DenseMap<const SCEV *, const Loop *> RelevantLoops;
92 /// Addrecs referring to any of the given loops are expanded in post-inc
93 /// mode. For example, expanding {1,+,1}<L> in post-inc mode returns the add
94 /// instruction that adds one to the phi for {0,+,1}<L>, as opposed to a new
95 /// phi starting at 1. This is only supported in non-canonical mode.
96 PostIncLoopSet PostIncLoops;
98 /// When this is non-null, addrecs expanded in the loop it indicates should
99 /// be inserted with increments at IVIncInsertPos.
100 const Loop *IVIncInsertLoop;
102 /// When expanding addrecs in the IVIncInsertLoop loop, insert the IV
103 /// increment at this position.
104 Instruction *IVIncInsertPos;
106 /// Phis that complete an IV chain. Reuse
107 DenseSet<AssertingVH<PHINode>> ChainedPhis;
109 /// When true, SCEVExpander tries to expand expressions in "canonical" form.
110 /// When false, expressions are expanded in a more literal form.
112 /// In "canonical" form addrecs are expanded as arithmetic based on a
113 /// canonical induction variable. Note that CanonicalMode doesn't guarantee
114 /// that all expressions are expanded in "canonical" form. For some
115 /// expressions literal mode can be preferred.
118 /// When invoked from LSR, the expander is in "strength reduction" mode. The
119 /// only difference is that phi's are only reused if they are already in
123 typedef IRBuilder<TargetFolder, IRBuilderCallbackInserter> BuilderType;
126 // RAII object that stores the current insertion point and restores it when
127 // the object is destroyed. This includes the debug location. Duplicated
128 // from InsertPointGuard to add SetInsertPoint() which is used to updated
129 // InsertPointGuards stack when insert points are moved during SCEV
131 class SCEVInsertPointGuard {
132 IRBuilderBase &Builder;
133 AssertingVH<BasicBlock> Block;
134 BasicBlock::iterator Point;
138 SCEVInsertPointGuard(const SCEVInsertPointGuard &) = delete;
139 SCEVInsertPointGuard &operator=(const SCEVInsertPointGuard &) = delete;
142 SCEVInsertPointGuard(IRBuilderBase &B, SCEVExpander *SE)
143 : Builder(B), Block(B.GetInsertBlock()), Point(B.GetInsertPoint()),
144 DbgLoc(B.getCurrentDebugLocation()), SE(SE) {
145 SE->InsertPointGuards.push_back(this);
148 ~SCEVInsertPointGuard() {
149 // These guards should always created/destroyed in FIFO order since they
150 // are used to guard lexically scoped blocks of code in
151 // ScalarEvolutionExpander.
152 assert(SE->InsertPointGuards.back() == this);
153 SE->InsertPointGuards.pop_back();
154 Builder.restoreIP(IRBuilderBase::InsertPoint(Block, Point));
155 Builder.SetCurrentDebugLocation(DbgLoc);
158 BasicBlock::iterator GetInsertPoint() const { return Point; }
159 void SetInsertPoint(BasicBlock::iterator I) { Point = I; }
162 /// Stack of pointers to saved insert points, used to keep insert points
163 /// consistent when instructions are moved.
164 SmallVector<SCEVInsertPointGuard *, 8> InsertPointGuards;
166 #ifdef LLVM_ENABLE_ABI_BREAKING_CHECKS
167 const char *DebugType;
170 friend struct SCEVVisitor<SCEVExpander, Value *>;
173 /// Construct a SCEVExpander in "canonical" mode.
174 explicit SCEVExpander(ScalarEvolution &se, const DataLayout &DL,
175 const char *name, bool PreserveLCSSA = true)
176 : SE(se), DL(DL), IVName(name), PreserveLCSSA(PreserveLCSSA),
177 IVIncInsertLoop(nullptr), IVIncInsertPos(nullptr), CanonicalMode(true),
179 Builder(se.getContext(), TargetFolder(DL),
180 IRBuilderCallbackInserter(
181 [this](Instruction *I) { rememberInstruction(I); })) {
182 #ifdef LLVM_ENABLE_ABI_BREAKING_CHECKS
188 // Make sure the insert point guard stack is consistent.
189 assert(InsertPointGuards.empty());
192 #ifdef LLVM_ENABLE_ABI_BREAKING_CHECKS
193 void setDebugType(const char *s) { DebugType = s; }
196 /// Erase the contents of the InsertedExpressions map so that users trying
197 /// to expand the same expression into multiple BasicBlocks or different
198 /// places within the same BasicBlock can do so.
200 InsertedExpressions.clear();
201 InsertedValues.clear();
202 InsertedPostIncValues.clear();
203 ReusedValues.clear();
208 ScalarEvolution *getSE() { return &SE; }
209 const SmallVectorImpl<WeakVH> &getInsertedIVs() const { return InsertedIVs; }
211 /// Return a vector containing all instructions inserted during expansion.
212 SmallVector<Instruction *, 32> getAllInsertedInstructions() const {
213 SmallVector<Instruction *, 32> Result;
214 for (auto &VH : InsertedValues) {
216 if (ReusedValues.contains(V))
218 if (auto *Inst = dyn_cast<Instruction>(V))
219 Result.push_back(Inst);
221 for (auto &VH : InsertedPostIncValues) {
223 if (ReusedValues.contains(V))
225 if (auto *Inst = dyn_cast<Instruction>(V))
226 Result.push_back(Inst);
232 /// Return true for expressions that can't be evaluated at runtime
233 /// within given \b Budget.
235 /// At is a parameter which specifies point in code where user is going to
236 /// expand this expression. Sometimes this knowledge can lead to
237 /// a less pessimistic cost estimation.
238 bool isHighCostExpansion(const SCEV *Expr, Loop *L, unsigned Budget,
239 const TargetTransformInfo *TTI,
240 const Instruction *At) {
241 assert(TTI && "This function requires TTI to be provided.");
242 assert(At && "This function requires At instruction to be provided.");
243 if (!TTI) // In assert-less builds, avoid crashing
244 return true; // by always claiming to be high-cost.
245 SmallVector<SCEVOperand, 8> Worklist;
246 SmallPtrSet<const SCEV *, 8> Processed;
247 InstructionCost Cost = 0;
248 unsigned ScaledBudget = Budget * TargetTransformInfo::TCC_Basic;
249 Worklist.emplace_back(-1, -1, Expr);
250 while (!Worklist.empty()) {
251 const SCEVOperand WorkItem = Worklist.pop_back_val();
252 if (isHighCostExpansionHelper(WorkItem, L, *At, Cost, ScaledBudget, *TTI,
253 Processed, Worklist))
256 assert(Cost <= ScaledBudget && "Should have returned from inner loop.");
260 /// Return the induction variable increment's IV operand.
261 Instruction *getIVIncOperand(Instruction *IncV, Instruction *InsertPos,
264 /// Utility for hoisting an IV increment.
265 bool hoistIVInc(Instruction *IncV, Instruction *InsertPos);
267 /// replace congruent phis with their most canonical representative. Return
268 /// the number of phis eliminated.
269 unsigned replaceCongruentIVs(Loop *L, const DominatorTree *DT,
270 SmallVectorImpl<WeakTrackingVH> &DeadInsts,
271 const TargetTransformInfo *TTI = nullptr);
273 /// Insert code to directly compute the specified SCEV expression into the
274 /// program. The code is inserted into the specified block.
275 Value *expandCodeFor(const SCEV *SH, Type *Ty, Instruction *I) {
276 return expandCodeForImpl(SH, Ty, I, true);
279 /// Insert code to directly compute the specified SCEV expression into the
280 /// program. The code is inserted into the SCEVExpander's current
281 /// insertion point. If a type is specified, the result will be expanded to
282 /// have that type, with a cast if necessary.
283 Value *expandCodeFor(const SCEV *SH, Type *Ty = nullptr) {
284 return expandCodeForImpl(SH, Ty, true);
287 /// Generates a code sequence that evaluates this predicate. The inserted
288 /// instructions will be at position \p Loc. The result will be of type i1
289 /// and will have a value of 0 when the predicate is false and 1 otherwise.
290 Value *expandCodeForPredicate(const SCEVPredicate *Pred, Instruction *Loc);
292 /// A specialized variant of expandCodeForPredicate, handling the case when
293 /// we are expanding code for a SCEVEqualPredicate.
294 Value *expandEqualPredicate(const SCEVEqualPredicate *Pred, Instruction *Loc);
296 /// Generates code that evaluates if the \p AR expression will overflow.
297 Value *generateOverflowCheck(const SCEVAddRecExpr *AR, Instruction *Loc,
300 /// A specialized variant of expandCodeForPredicate, handling the case when
301 /// we are expanding code for a SCEVWrapPredicate.
302 Value *expandWrapPredicate(const SCEVWrapPredicate *P, Instruction *Loc);
304 /// A specialized variant of expandCodeForPredicate, handling the case when
305 /// we are expanding code for a SCEVUnionPredicate.
306 Value *expandUnionPredicate(const SCEVUnionPredicate *Pred, Instruction *Loc);
308 /// Set the current IV increment loop and position.
309 void setIVIncInsertPos(const Loop *L, Instruction *Pos) {
310 assert(!CanonicalMode &&
311 "IV increment positions are not supported in CanonicalMode");
313 IVIncInsertPos = Pos;
316 /// Enable post-inc expansion for addrecs referring to the given
317 /// loops. Post-inc expansion is only supported in non-canonical mode.
318 void setPostInc(const PostIncLoopSet &L) {
319 assert(!CanonicalMode &&
320 "Post-inc expansion is not supported in CanonicalMode");
324 /// Disable all post-inc expansion.
325 void clearPostInc() {
326 PostIncLoops.clear();
328 // When we change the post-inc loop set, cached expansions may no
330 InsertedPostIncValues.clear();
333 /// Disable the behavior of expanding expressions in canonical form rather
334 /// than in a more literal form. Non-canonical mode is useful for late
335 /// optimization passes.
336 void disableCanonicalMode() { CanonicalMode = false; }
338 void enableLSRMode() { LSRMode = true; }
340 /// Set the current insertion point. This is useful if multiple calls to
341 /// expandCodeFor() are going to be made with the same insert point and the
342 /// insert point may be moved during one of the expansions (e.g. if the
343 /// insert point is not a block terminator).
344 void setInsertPoint(Instruction *IP) {
346 Builder.SetInsertPoint(IP);
349 /// Clear the current insertion point. This is useful if the instruction
350 /// that had been serving as the insertion point may have been deleted.
351 void clearInsertPoint() { Builder.ClearInsertionPoint(); }
353 /// Set location information used by debugging information.
354 void SetCurrentDebugLocation(DebugLoc L) {
355 Builder.SetCurrentDebugLocation(std::move(L));
358 /// Get location information used by debugging information.
359 DebugLoc getCurrentDebugLocation() const {
360 return Builder.getCurrentDebugLocation();
363 /// Return true if the specified instruction was inserted by the code
364 /// rewriter. If so, the client should not modify the instruction. Note that
365 /// this also includes instructions re-used during expansion.
366 bool isInsertedInstruction(Instruction *I) const {
367 return InsertedValues.count(I) || InsertedPostIncValues.count(I);
370 void setChainedPhi(PHINode *PN) { ChainedPhis.insert(PN); }
372 /// Try to find the ValueOffsetPair for S. The function is mainly used to
373 /// check whether S can be expanded cheaply. If this returns a non-None
374 /// value, we know we can codegen the `ValueOffsetPair` into a suitable
375 /// expansion identical with S so that S can be expanded cheaply.
377 /// L is a hint which tells in which loop to look for the suitable value.
378 /// On success return value which is equivalent to the expanded S at point
379 /// At. Return nullptr if value was not found.
381 /// Note that this function does not perform an exhaustive search. I.e if it
382 /// didn't find any value it does not mean that there is no such value.
384 Optional<ScalarEvolution::ValueOffsetPair>
385 getRelatedExistingExpansion(const SCEV *S, const Instruction *At, Loop *L);
387 /// Returns a suitable insert point after \p I, that dominates \p
388 /// MustDominate. Skips instructions inserted by the expander.
389 BasicBlock::iterator findInsertPointAfter(Instruction *I,
390 Instruction *MustDominate) const;
393 LLVMContext &getContext() const { return SE.getContext(); }
395 /// Insert code to directly compute the specified SCEV expression into the
396 /// program. The code is inserted into the SCEVExpander's current
397 /// insertion point. If a type is specified, the result will be expanded to
398 /// have that type, with a cast if necessary. If \p Root is true, this
399 /// indicates that \p SH is the top-level expression to expand passed from
400 /// an external client call.
401 Value *expandCodeForImpl(const SCEV *SH, Type *Ty, bool Root);
403 /// Insert code to directly compute the specified SCEV expression into the
404 /// program. The code is inserted into the specified block. If \p
405 /// Root is true, this indicates that \p SH is the top-level expression to
406 /// expand passed from an external client call.
407 Value *expandCodeForImpl(const SCEV *SH, Type *Ty, Instruction *I, bool Root);
409 /// Recursive helper function for isHighCostExpansion.
410 bool isHighCostExpansionHelper(const SCEVOperand &WorkItem, Loop *L,
411 const Instruction &At, InstructionCost &Cost,
413 const TargetTransformInfo &TTI,
414 SmallPtrSetImpl<const SCEV *> &Processed,
415 SmallVectorImpl<SCEVOperand> &Worklist);
417 /// Insert the specified binary operator, doing a small amount of work to
418 /// avoid inserting an obviously redundant operation, and hoisting to an
419 /// outer loop when the opportunity is there and it is safe.
420 Value *InsertBinop(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
421 SCEV::NoWrapFlags Flags, bool IsSafeToHoist);
423 /// We want to cast \p V. What would be the best place for such a cast?
424 BasicBlock::iterator GetOptimalInsertionPointForCastOf(Value *V) const;
426 /// Arrange for there to be a cast of V to Ty at IP, reusing an existing
427 /// cast if a suitable one exists, moving an existing cast if a suitable one
428 /// exists but isn't in the right place, or creating a new one.
429 Value *ReuseOrCreateCast(Value *V, Type *Ty, Instruction::CastOps Op,
430 BasicBlock::iterator IP);
432 /// Insert a cast of V to the specified type, which must be possible with a
433 /// noop cast, doing what we can to share the casts.
434 Value *InsertNoopCastOfTo(Value *V, Type *Ty);
436 /// Expand a SCEVAddExpr with a pointer type into a GEP instead of using
437 /// ptrtoint+arithmetic+inttoptr.
438 Value *expandAddToGEP(const SCEV *const *op_begin, const SCEV *const *op_end,
439 PointerType *PTy, Type *Ty, Value *V);
440 Value *expandAddToGEP(const SCEV *Op, PointerType *PTy, Type *Ty, Value *V);
442 /// Find a previous Value in ExprValueMap for expand.
443 ScalarEvolution::ValueOffsetPair
444 FindValueInExprValueMap(const SCEV *S, const Instruction *InsertPt);
446 Value *expand(const SCEV *S);
448 /// Determine the most "relevant" loop for the given SCEV.
449 const Loop *getRelevantLoop(const SCEV *);
451 Value *visitConstant(const SCEVConstant *S) { return S->getValue(); }
453 Value *visitPtrToIntExpr(const SCEVPtrToIntExpr *S);
455 Value *visitTruncateExpr(const SCEVTruncateExpr *S);
457 Value *visitZeroExtendExpr(const SCEVZeroExtendExpr *S);
459 Value *visitSignExtendExpr(const SCEVSignExtendExpr *S);
461 Value *visitAddExpr(const SCEVAddExpr *S);
463 Value *visitMulExpr(const SCEVMulExpr *S);
465 Value *visitUDivExpr(const SCEVUDivExpr *S);
467 Value *visitAddRecExpr(const SCEVAddRecExpr *S);
469 Value *visitSMaxExpr(const SCEVSMaxExpr *S);
471 Value *visitUMaxExpr(const SCEVUMaxExpr *S);
473 Value *visitSMinExpr(const SCEVSMinExpr *S);
475 Value *visitUMinExpr(const SCEVUMinExpr *S);
477 Value *visitUnknown(const SCEVUnknown *S) { return S->getValue(); }
479 void rememberInstruction(Value *I);
481 bool isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, const Loop *L);
483 bool isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, const Loop *L);
485 Value *expandAddRecExprLiterally(const SCEVAddRecExpr *);
486 PHINode *getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
487 const Loop *L, Type *ExpandTy, Type *IntTy,
488 Type *&TruncTy, bool &InvertStep);
489 Value *expandIVInc(PHINode *PN, Value *StepV, const Loop *L, Type *ExpandTy,
490 Type *IntTy, bool useSubtract);
492 void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
493 Instruction *Pos, PHINode *LoopPhi);
495 void fixupInsertPoints(Instruction *I);
497 /// If required, create LCSSA PHIs for \p Users' operand \p OpIdx. If new
498 /// LCSSA PHIs have been created, return the LCSSA PHI available at \p User.
499 /// If no PHIs have been created, return the unchanged operand \p OpIdx.
500 Value *fixupLCSSAFormFor(Instruction *User, unsigned OpIdx);
503 /// Helper to remove instructions inserted during SCEV expansion, unless they
504 /// are marked as used.
505 class SCEVExpanderCleaner {
506 SCEVExpander &Expander;
510 /// Indicates whether the result of the expansion is used. If false, the
511 /// instructions added during expansion are removed.
515 SCEVExpanderCleaner(SCEVExpander &Expander, DominatorTree &DT)
516 : Expander(Expander), DT(DT), ResultUsed(false) {}
518 ~SCEVExpanderCleaner() { cleanup(); }
520 /// Indicate that the result of the expansion is used.
521 void markResultUsed() { ResultUsed = true; }