1 //===---- llvm/Analysis/ScalarEvolutionExpander.h - SCEV Exprs --*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the classes used to generate code from scalar expressions.
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_ANALYSIS_SCALAREVOLUTIONEXPANDER_H
15 #define LLVM_ANALYSIS_SCALAREVOLUTIONEXPANDER_H
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
21 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
22 #include "llvm/Analysis/TargetFolder.h"
23 #include "llvm/IR/IRBuilder.h"
24 #include "llvm/IR/ValueHandle.h"
27 class TargetTransformInfo;
29 /// Return true if the given expression is safe to expand in the sense that
30 /// all materialized values are safe to speculate anywhere their operands are
32 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE);
34 /// Return true if the given expression is safe to expand in the sense that
35 /// all materialized values are defined and safe to speculate at the specified
36 /// location and their operands are defined at this location.
37 bool isSafeToExpandAt(const SCEV *S, const Instruction *InsertionPoint,
40 /// This class uses information about analyze scalars to rewrite expressions
41 /// in canonical form.
43 /// Clients should create an instance of this class when rewriting is needed,
44 /// and destroy it when finished to allow the release of the associated
46 class SCEVExpander : public SCEVVisitor<SCEVExpander, Value*> {
50 // New instructions receive a name to identify them with the current pass.
53 // InsertedExpressions caches Values for reuse, so must track RAUW.
54 DenseMap<std::pair<const SCEV *, Instruction *>, TrackingVH<Value>>
57 // InsertedValues only flags inserted instructions so needs no RAUW.
58 DenseSet<AssertingVH<Value>> InsertedValues;
59 DenseSet<AssertingVH<Value>> InsertedPostIncValues;
61 /// A memoization of the "relevant" loop for a given SCEV.
62 DenseMap<const SCEV *, const Loop *> RelevantLoops;
64 /// Addrecs referring to any of the given loops are expanded in post-inc
65 /// mode. For example, expanding {1,+,1}<L> in post-inc mode returns the add
66 /// instruction that adds one to the phi for {0,+,1}<L>, as opposed to a new
67 /// phi starting at 1. This is only supported in non-canonical mode.
68 PostIncLoopSet PostIncLoops;
70 /// When this is non-null, addrecs expanded in the loop it indicates should
71 /// be inserted with increments at IVIncInsertPos.
72 const Loop *IVIncInsertLoop;
74 /// When expanding addrecs in the IVIncInsertLoop loop, insert the IV
75 /// increment at this position.
76 Instruction *IVIncInsertPos;
78 /// Phis that complete an IV chain. Reuse
79 DenseSet<AssertingVH<PHINode>> ChainedPhis;
81 /// When true, expressions are expanded in "canonical" form. In particular,
82 /// addrecs are expanded as arithmetic based on a canonical induction
83 /// variable. When false, expression are expanded in a more literal form.
86 /// When invoked from LSR, the expander is in "strength reduction" mode. The
87 /// only difference is that phi's are only reused if they are already in
91 typedef IRBuilder<TargetFolder> BuilderType;
94 // RAII object that stores the current insertion point and restores it when
95 // the object is destroyed. This includes the debug location. Duplicated
96 // from InsertPointGuard to add SetInsertPoint() which is used to updated
97 // InsertPointGuards stack when insert points are moved during SCEV
99 class SCEVInsertPointGuard {
100 IRBuilderBase &Builder;
101 AssertingVH<BasicBlock> Block;
102 BasicBlock::iterator Point;
106 SCEVInsertPointGuard(const SCEVInsertPointGuard &) = delete;
107 SCEVInsertPointGuard &operator=(const SCEVInsertPointGuard &) = delete;
110 SCEVInsertPointGuard(IRBuilderBase &B, SCEVExpander *SE)
111 : Builder(B), Block(B.GetInsertBlock()), Point(B.GetInsertPoint()),
112 DbgLoc(B.getCurrentDebugLocation()), SE(SE) {
113 SE->InsertPointGuards.push_back(this);
116 ~SCEVInsertPointGuard() {
117 // These guards should always created/destroyed in FIFO order since they
118 // are used to guard lexically scoped blocks of code in
119 // ScalarEvolutionExpander.
120 assert(SE->InsertPointGuards.back() == this);
121 SE->InsertPointGuards.pop_back();
122 Builder.restoreIP(IRBuilderBase::InsertPoint(Block, Point));
123 Builder.SetCurrentDebugLocation(DbgLoc);
126 BasicBlock::iterator GetInsertPoint() const { return Point; }
127 void SetInsertPoint(BasicBlock::iterator I) { Point = I; }
130 /// Stack of pointers to saved insert points, used to keep insert points
131 /// consistent when instructions are moved.
132 SmallVector<SCEVInsertPointGuard *, 8> InsertPointGuards;
135 const char *DebugType;
138 friend struct SCEVVisitor<SCEVExpander, Value*>;
141 /// Construct a SCEVExpander in "canonical" mode.
142 explicit SCEVExpander(ScalarEvolution &se, const DataLayout &DL,
144 : SE(se), DL(DL), IVName(name), IVIncInsertLoop(nullptr),
145 IVIncInsertPos(nullptr), CanonicalMode(true), LSRMode(false),
146 Builder(se.getContext(), TargetFolder(DL)) {
153 // Make sure the insert point guard stack is consistent.
154 assert(InsertPointGuards.empty());
158 void setDebugType(const char* s) { DebugType = s; }
161 /// Erase the contents of the InsertedExpressions map so that users trying
162 /// to expand the same expression into multiple BasicBlocks or different
163 /// places within the same BasicBlock can do so.
165 InsertedExpressions.clear();
166 InsertedValues.clear();
167 InsertedPostIncValues.clear();
171 /// Return true for expressions that may incur non-trivial cost to evaluate
174 /// At is an optional parameter which specifies point in code where user is
175 /// going to expand this expression. Sometimes this knowledge can lead to a
176 /// more accurate cost estimation.
177 bool isHighCostExpansion(const SCEV *Expr, Loop *L,
178 const Instruction *At = nullptr) {
179 SmallPtrSet<const SCEV *, 8> Processed;
180 return isHighCostExpansionHelper(Expr, L, At, Processed);
183 /// This method returns the canonical induction variable of the specified
184 /// type for the specified loop (inserting one if there is none). A
185 /// canonical induction variable starts at zero and steps by one on each
187 PHINode *getOrInsertCanonicalInductionVariable(const Loop *L, Type *Ty);
189 /// Return the induction variable increment's IV operand.
190 Instruction *getIVIncOperand(Instruction *IncV, Instruction *InsertPos,
193 /// Utility for hoisting an IV increment.
194 bool hoistIVInc(Instruction *IncV, Instruction *InsertPos);
196 /// replace congruent phis with their most canonical representative. Return
197 /// the number of phis eliminated.
198 unsigned replaceCongruentIVs(Loop *L, const DominatorTree *DT,
199 SmallVectorImpl<WeakTrackingVH> &DeadInsts,
200 const TargetTransformInfo *TTI = nullptr);
202 /// Insert code to directly compute the specified SCEV expression into the
203 /// program. The inserted code is inserted into the specified block.
204 Value *expandCodeFor(const SCEV *SH, Type *Ty, Instruction *I);
206 /// Insert code to directly compute the specified SCEV expression into the
207 /// program. The inserted code is inserted into the SCEVExpander's current
208 /// insertion point. If a type is specified, the result will be expanded to
209 /// have that type, with a cast if necessary.
210 Value *expandCodeFor(const SCEV *SH, Type *Ty = nullptr);
213 /// Generates a code sequence that evaluates this predicate. The inserted
214 /// instructions will be at position \p Loc. The result will be of type i1
215 /// and will have a value of 0 when the predicate is false and 1 otherwise.
216 Value *expandCodeForPredicate(const SCEVPredicate *Pred, Instruction *Loc);
218 /// A specialized variant of expandCodeForPredicate, handling the case when
219 /// we are expanding code for a SCEVEqualPredicate.
220 Value *expandEqualPredicate(const SCEVEqualPredicate *Pred,
223 /// Generates code that evaluates if the \p AR expression will overflow.
224 Value *generateOverflowCheck(const SCEVAddRecExpr *AR, Instruction *Loc,
227 /// A specialized variant of expandCodeForPredicate, handling the case when
228 /// we are expanding code for a SCEVWrapPredicate.
229 Value *expandWrapPredicate(const SCEVWrapPredicate *P, Instruction *Loc);
231 /// A specialized variant of expandCodeForPredicate, handling the case when
232 /// we are expanding code for a SCEVUnionPredicate.
233 Value *expandUnionPredicate(const SCEVUnionPredicate *Pred,
236 /// Set the current IV increment loop and position.
237 void setIVIncInsertPos(const Loop *L, Instruction *Pos) {
238 assert(!CanonicalMode &&
239 "IV increment positions are not supported in CanonicalMode");
241 IVIncInsertPos = Pos;
244 /// Enable post-inc expansion for addrecs referring to the given
245 /// loops. Post-inc expansion is only supported in non-canonical mode.
246 void setPostInc(const PostIncLoopSet &L) {
247 assert(!CanonicalMode &&
248 "Post-inc expansion is not supported in CanonicalMode");
252 /// Disable all post-inc expansion.
253 void clearPostInc() {
254 PostIncLoops.clear();
256 // When we change the post-inc loop set, cached expansions may no
258 InsertedPostIncValues.clear();
261 /// Disable the behavior of expanding expressions in canonical form rather
262 /// than in a more literal form. Non-canonical mode is useful for late
263 /// optimization passes.
264 void disableCanonicalMode() { CanonicalMode = false; }
266 void enableLSRMode() { LSRMode = true; }
268 /// Set the current insertion point. This is useful if multiple calls to
269 /// expandCodeFor() are going to be made with the same insert point and the
270 /// insert point may be moved during one of the expansions (e.g. if the
271 /// insert point is not a block terminator).
272 void setInsertPoint(Instruction *IP) {
274 Builder.SetInsertPoint(IP);
277 /// Clear the current insertion point. This is useful if the instruction
278 /// that had been serving as the insertion point may have been deleted.
279 void clearInsertPoint() {
280 Builder.ClearInsertionPoint();
283 /// Return true if the specified instruction was inserted by the code
284 /// rewriter. If so, the client should not modify the instruction.
285 bool isInsertedInstruction(Instruction *I) const {
286 return InsertedValues.count(I) || InsertedPostIncValues.count(I);
289 void setChainedPhi(PHINode *PN) { ChainedPhis.insert(PN); }
291 /// Try to find existing LLVM IR value for S available at the point At.
292 Value *getExactExistingExpansion(const SCEV *S, const Instruction *At,
295 /// Try to find the ValueOffsetPair for S. The function is mainly used to
296 /// check whether S can be expanded cheaply. If this returns a non-None
297 /// value, we know we can codegen the `ValueOffsetPair` into a suitable
298 /// expansion identical with S so that S can be expanded cheaply.
300 /// L is a hint which tells in which loop to look for the suitable value.
301 /// On success return value which is equivalent to the expanded S at point
302 /// At. Return nullptr if value was not found.
304 /// Note that this function does not perform an exhaustive search. I.e if it
305 /// didn't find any value it does not mean that there is no such value.
307 Optional<ScalarEvolution::ValueOffsetPair>
308 getRelatedExistingExpansion(const SCEV *S, const Instruction *At, Loop *L);
311 LLVMContext &getContext() const { return SE.getContext(); }
313 /// Recursive helper function for isHighCostExpansion.
314 bool isHighCostExpansionHelper(const SCEV *S, Loop *L,
315 const Instruction *At,
316 SmallPtrSetImpl<const SCEV *> &Processed);
318 /// Insert the specified binary operator, doing a small amount of work to
319 /// avoid inserting an obviously redundant operation.
320 Value *InsertBinop(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS);
322 /// Arrange for there to be a cast of V to Ty at IP, reusing an existing
323 /// cast if a suitable one exists, moving an existing cast if a suitable one
324 /// exists but isn't in the right place, or or creating a new one.
325 Value *ReuseOrCreateCast(Value *V, Type *Ty,
326 Instruction::CastOps Op,
327 BasicBlock::iterator IP);
329 /// Insert a cast of V to the specified type, which must be possible with a
330 /// noop cast, doing what we can to share the casts.
331 Value *InsertNoopCastOfTo(Value *V, Type *Ty);
333 /// Expand a SCEVAddExpr with a pointer type into a GEP instead of using
334 /// ptrtoint+arithmetic+inttoptr.
335 Value *expandAddToGEP(const SCEV *const *op_begin,
336 const SCEV *const *op_end,
337 PointerType *PTy, Type *Ty, Value *V);
339 /// Find a previous Value in ExprValueMap for expand.
340 ScalarEvolution::ValueOffsetPair
341 FindValueInExprValueMap(const SCEV *S, const Instruction *InsertPt);
343 Value *expand(const SCEV *S);
345 /// Determine the most "relevant" loop for the given SCEV.
346 const Loop *getRelevantLoop(const SCEV *);
348 Value *visitConstant(const SCEVConstant *S) {
349 return S->getValue();
352 Value *visitTruncateExpr(const SCEVTruncateExpr *S);
354 Value *visitZeroExtendExpr(const SCEVZeroExtendExpr *S);
356 Value *visitSignExtendExpr(const SCEVSignExtendExpr *S);
358 Value *visitAddExpr(const SCEVAddExpr *S);
360 Value *visitMulExpr(const SCEVMulExpr *S);
362 Value *visitUDivExpr(const SCEVUDivExpr *S);
364 Value *visitAddRecExpr(const SCEVAddRecExpr *S);
366 Value *visitSMaxExpr(const SCEVSMaxExpr *S);
368 Value *visitUMaxExpr(const SCEVUMaxExpr *S);
370 Value *visitUnknown(const SCEVUnknown *S) {
371 return S->getValue();
374 void rememberInstruction(Value *I);
376 bool isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, const Loop *L);
378 bool isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, const Loop *L);
380 Value *expandAddRecExprLiterally(const SCEVAddRecExpr *);
381 PHINode *getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
387 Value *expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
388 Type *ExpandTy, Type *IntTy, bool useSubtract);
390 void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
391 Instruction *Pos, PHINode *LoopPhi);
393 void fixupInsertPoints(Instruction *I);