Bug Summary

File:lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 3050, column 16
Called C++ object pointer is null

Annotated Source Code

1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This transformation analyzes and transforms the induction variables (and
11// computations derived from them) into forms suitable for efficient execution
12// on the target.
13//
14// This pass performs a strength reduction on array references inside loops that
15// have as one or more of their components the loop induction variable, it
16// rewrites expressions to take advantage of scaled-index addressing modes
17// available on the target, and it performs a variety of other optimizations
18// related to loop induction variables.
19//
20// Terminology note: this code has a lot of handling for "post-increment" or
21// "post-inc" users. This is not talking about post-increment addressing modes;
22// it is instead talking about code like this:
23//
24// %i = phi [ 0, %entry ], [ %i.next, %latch ]
25// ...
26// %i.next = add %i, 1
27// %c = icmp eq %i.next, %n
28//
29// The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30// it's useful to think about these as the same register, with some uses using
31// the value of the register before the add and some using it after. In this
32// example, the icmp is a post-increment user, since it uses %i.next, which is
33// the value of the induction variable after the increment. The other common
34// case of post-increment users is users outside the loop.
35//
36// TODO: More sophistication in the way Formulae are generated and filtered.
37//
38// TODO: Handle multiple loops at a time.
39//
40// TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
41// of a GlobalValue?
42//
43// TODO: When truncation is free, truncate ICmp users' operands to make it a
44// smaller encoding (on x86 at least).
45//
46// TODO: When a negated register is used by an add (such as in a list of
47// multiple base registers, or as the increment expression in an addrec),
48// we may not actually need both reg and (-1 * reg) in registers; the
49// negation can be implemented by using a sub instead of an add. The
50// lack of support for taking this into consideration when making
51// register pressure decisions is partly worked around by the "Special"
52// use kind.
53//
54//===----------------------------------------------------------------------===//
55
56#include "llvm/Transforms/Scalar/LoopStrengthReduce.h"
57#include "llvm/ADT/APInt.h"
58#include "llvm/ADT/DenseMap.h"
59#include "llvm/ADT/DenseSet.h"
60#include "llvm/ADT/Hashing.h"
61#include "llvm/ADT/PointerIntPair.h"
62#include "llvm/ADT/STLExtras.h"
63#include "llvm/ADT/SetVector.h"
64#include "llvm/ADT/SmallBitVector.h"
65#include "llvm/ADT/SmallPtrSet.h"
66#include "llvm/ADT/SmallSet.h"
67#include "llvm/ADT/SmallVector.h"
68#include "llvm/Analysis/IVUsers.h"
69#include "llvm/Analysis/LoopInfo.h"
70#include "llvm/Analysis/LoopPass.h"
71#include "llvm/Analysis/ScalarEvolution.h"
72#include "llvm/Analysis/ScalarEvolutionExpander.h"
73#include "llvm/Analysis/ScalarEvolutionExpressions.h"
74#include "llvm/Analysis/ScalarEvolutionNormalization.h"
75#include "llvm/Analysis/TargetTransformInfo.h"
76#include "llvm/IR/BasicBlock.h"
77#include "llvm/IR/Constant.h"
78#include "llvm/IR/Constants.h"
79#include "llvm/IR/DerivedTypes.h"
80#include "llvm/IR/Dominators.h"
81#include "llvm/IR/GlobalValue.h"
82#include "llvm/IR/IRBuilder.h"
83#include "llvm/IR/Instruction.h"
84#include "llvm/IR/Instructions.h"
85#include "llvm/IR/IntrinsicInst.h"
86#include "llvm/IR/Module.h"
87#include "llvm/IR/OperandTraits.h"
88#include "llvm/IR/Operator.h"
89#include "llvm/IR/Type.h"
90#include "llvm/IR/Value.h"
91#include "llvm/IR/ValueHandle.h"
92#include "llvm/Pass.h"
93#include "llvm/Support/Casting.h"
94#include "llvm/Support/CommandLine.h"
95#include "llvm/Support/Compiler.h"
96#include "llvm/Support/Debug.h"
97#include "llvm/Support/ErrorHandling.h"
98#include "llvm/Support/MathExtras.h"
99#include "llvm/Support/raw_ostream.h"
100#include "llvm/Transforms/Scalar.h"
101#include "llvm/Transforms/Scalar/LoopPassManager.h"
102#include "llvm/Transforms/Utils/BasicBlockUtils.h"
103#include "llvm/Transforms/Utils/Local.h"
104#include <algorithm>
105#include <cassert>
106#include <cstddef>
107#include <cstdint>
108#include <cstdlib>
109#include <iterator>
110#include <map>
111#include <tuple>
112#include <utility>
113
114using namespace llvm;
115
116#define DEBUG_TYPE"loop-reduce" "loop-reduce"
117
118/// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
119/// bail out. This threshold is far beyond the number of users that LSR can
120/// conceivably solve, so it should not affect generated code, but catches the
121/// worst cases before LSR burns too much compile time and stack space.
122static const unsigned MaxIVUsers = 200;
123
124// Temporary flag to cleanup congruent phis after LSR phi expansion.
125// It's currently disabled until we can determine whether it's truly useful or
126// not. The flag should be removed after the v3.0 release.
127// This is now needed for ivchains.
128static cl::opt<bool> EnablePhiElim(
129 "enable-lsr-phielim", cl::Hidden, cl::init(true),
130 cl::desc("Enable LSR phi elimination"));
131
132// The flag adds instruction count to solutions cost comparision.
133static cl::opt<bool> InsnsCost(
134 "lsr-insns-cost", cl::Hidden, cl::init(false),
135 cl::desc("Add instruction count to a LSR cost model"));
136
137// Flag to choose how to narrow complex lsr solution
138static cl::opt<bool> LSRExpNarrow(
139 "lsr-exp-narrow", cl::Hidden, cl::init(true),
140 cl::desc("Narrow LSR complex solution using"
141 " expectation of registers number"));
142
143#ifndef NDEBUG
144// Stress test IV chain generation.
145static cl::opt<bool> StressIVChain(
146 "stress-ivchain", cl::Hidden, cl::init(false),
147 cl::desc("Stress test LSR IV chains"));
148#else
149static bool StressIVChain = false;
150#endif
151
152namespace {
153
154struct MemAccessTy {
155 /// Used in situations where the accessed memory type is unknown.
156 static const unsigned UnknownAddressSpace = ~0u;
157
158 Type *MemTy;
159 unsigned AddrSpace;
160
161 MemAccessTy() : MemTy(nullptr), AddrSpace(UnknownAddressSpace) {}
162
163 MemAccessTy(Type *Ty, unsigned AS) :
164 MemTy(Ty), AddrSpace(AS) {}
165
166 bool operator==(MemAccessTy Other) const {
167 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
168 }
169
170 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
171
172 static MemAccessTy getUnknown(LLVMContext &Ctx,
173 unsigned AS = UnknownAddressSpace) {
174 return MemAccessTy(Type::getVoidTy(Ctx), AS);
175 }
176};
177
178/// This class holds data which is used to order reuse candidates.
179class RegSortData {
180public:
181 /// This represents the set of LSRUse indices which reference
182 /// a particular register.
183 SmallBitVector UsedByIndices;
184
185 void print(raw_ostream &OS) const;
186 void dump() const;
187};
188
189} // end anonymous namespace
190
191void RegSortData::print(raw_ostream &OS) const {
192 OS << "[NumUses=" << UsedByIndices.count() << ']';
193}
194
195#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
196LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
197 print(errs()); errs() << '\n';
198}
199#endif
200
201namespace {
202
203/// Map register candidates to information about how they are used.
204class RegUseTracker {
205 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
206
207 RegUsesTy RegUsesMap;
208 SmallVector<const SCEV *, 16> RegSequence;
209
210public:
211 void countRegister(const SCEV *Reg, size_t LUIdx);
212 void dropRegister(const SCEV *Reg, size_t LUIdx);
213 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
214
215 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
216
217 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
218
219 void clear();
220
221 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
222 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
223 iterator begin() { return RegSequence.begin(); }
224 iterator end() { return RegSequence.end(); }
225 const_iterator begin() const { return RegSequence.begin(); }
226 const_iterator end() const { return RegSequence.end(); }
227};
228
229} // end anonymous namespace
230
231void
232RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
233 std::pair<RegUsesTy::iterator, bool> Pair =
234 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
235 RegSortData &RSD = Pair.first->second;
236 if (Pair.second)
237 RegSequence.push_back(Reg);
238 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
239 RSD.UsedByIndices.set(LUIdx);
240}
241
242void
243RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
244 RegUsesTy::iterator It = RegUsesMap.find(Reg);
245 assert(It != RegUsesMap.end())((It != RegUsesMap.end()) ? static_cast<void> (0) : __assert_fail
("It != RegUsesMap.end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 245, __PRETTY_FUNCTION__))
;
246 RegSortData &RSD = It->second;
247 assert(RSD.UsedByIndices.size() > LUIdx)((RSD.UsedByIndices.size() > LUIdx) ? static_cast<void>
(0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 247, __PRETTY_FUNCTION__))
;
248 RSD.UsedByIndices.reset(LUIdx);
249}
250
251void
252RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
253 assert(LUIdx <= LastLUIdx)((LUIdx <= LastLUIdx) ? static_cast<void> (0) : __assert_fail
("LUIdx <= LastLUIdx", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 253, __PRETTY_FUNCTION__))
;
254
255 // Update RegUses. The data structure is not optimized for this purpose;
256 // we must iterate through it and update each of the bit vectors.
257 for (auto &Pair : RegUsesMap) {
258 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
259 if (LUIdx < UsedByIndices.size())
260 UsedByIndices[LUIdx] =
261 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
262 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
263 }
264}
265
266bool
267RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
268 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
269 if (I == RegUsesMap.end())
270 return false;
271 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
272 int i = UsedByIndices.find_first();
273 if (i == -1) return false;
274 if ((size_t)i != LUIdx) return true;
275 return UsedByIndices.find_next(i) != -1;
276}
277
278const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
279 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
280 assert(I != RegUsesMap.end() && "Unknown register!")((I != RegUsesMap.end() && "Unknown register!") ? static_cast
<void> (0) : __assert_fail ("I != RegUsesMap.end() && \"Unknown register!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 280, __PRETTY_FUNCTION__))
;
281 return I->second.UsedByIndices;
282}
283
284void RegUseTracker::clear() {
285 RegUsesMap.clear();
286 RegSequence.clear();
287}
288
289namespace {
290
291/// This class holds information that describes a formula for computing
292/// satisfying a use. It may include broken-out immediates and scaled registers.
293struct Formula {
294 /// Global base address used for complex addressing.
295 GlobalValue *BaseGV;
296
297 /// Base offset for complex addressing.
298 int64_t BaseOffset;
299
300 /// Whether any complex addressing has a base register.
301 bool HasBaseReg;
302
303 /// The scale of any complex addressing.
304 int64_t Scale;
305
306 /// The list of "base" registers for this use. When this is non-empty. The
307 /// canonical representation of a formula is
308 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
309 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
310 /// 3. The reg containing recurrent expr related with currect loop in the
311 /// formula should be put in the ScaledReg.
312 /// #1 enforces that the scaled register is always used when at least two
313 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
314 /// #2 enforces that 1 * reg is reg.
315 /// #3 ensures invariant regs with respect to current loop can be combined
316 /// together in LSR codegen.
317 /// This invariant can be temporarly broken while building a formula.
318 /// However, every formula inserted into the LSRInstance must be in canonical
319 /// form.
320 SmallVector<const SCEV *, 4> BaseRegs;
321
322 /// The 'scaled' register for this use. This should be non-null when Scale is
323 /// not zero.
324 const SCEV *ScaledReg;
325
326 /// An additional constant offset which added near the use. This requires a
327 /// temporary register, but the offset itself can live in an add immediate
328 /// field rather than a register.
329 int64_t UnfoldedOffset;
330
331 Formula()
332 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
333 ScaledReg(nullptr), UnfoldedOffset(0) {}
334
335 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
336
337 bool isCanonical(const Loop &L) const;
338
339 void canonicalize(const Loop &L);
340
341 bool unscale();
342
343 bool hasZeroEnd() const;
344
345 size_t getNumRegs() const;
346 Type *getType() const;
347
348 void deleteBaseReg(const SCEV *&S);
349
350 bool referencesReg(const SCEV *S) const;
351 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
352 const RegUseTracker &RegUses) const;
353
354 void print(raw_ostream &OS) const;
355 void dump() const;
356};
357
358} // end anonymous namespace
359
360/// Recursion helper for initialMatch.
361static void DoInitialMatch(const SCEV *S, Loop *L,
362 SmallVectorImpl<const SCEV *> &Good,
363 SmallVectorImpl<const SCEV *> &Bad,
364 ScalarEvolution &SE) {
365 // Collect expressions which properly dominate the loop header.
366 if (SE.properlyDominates(S, L->getHeader())) {
367 Good.push_back(S);
368 return;
369 }
370
371 // Look at add operands.
372 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
373 for (const SCEV *S : Add->operands())
374 DoInitialMatch(S, L, Good, Bad, SE);
375 return;
376 }
377
378 // Look at addrec operands.
379 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
380 if (!AR->getStart()->isZero() && AR->isAffine()) {
381 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
382 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
383 AR->getStepRecurrence(SE),
384 // FIXME: AR->getNoWrapFlags()
385 AR->getLoop(), SCEV::FlagAnyWrap),
386 L, Good, Bad, SE);
387 return;
388 }
389
390 // Handle a multiplication by -1 (negation) if it didn't fold.
391 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
392 if (Mul->getOperand(0)->isAllOnesValue()) {
393 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
394 const SCEV *NewMul = SE.getMulExpr(Ops);
395
396 SmallVector<const SCEV *, 4> MyGood;
397 SmallVector<const SCEV *, 4> MyBad;
398 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
399 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
400 SE.getEffectiveSCEVType(NewMul->getType())));
401 for (const SCEV *S : MyGood)
402 Good.push_back(SE.getMulExpr(NegOne, S));
403 for (const SCEV *S : MyBad)
404 Bad.push_back(SE.getMulExpr(NegOne, S));
405 return;
406 }
407
408 // Ok, we can't do anything interesting. Just stuff the whole thing into a
409 // register and hope for the best.
410 Bad.push_back(S);
411}
412
413/// Incorporate loop-variant parts of S into this Formula, attempting to keep
414/// all loop-invariant and loop-computable values in a single base register.
415void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
416 SmallVector<const SCEV *, 4> Good;
417 SmallVector<const SCEV *, 4> Bad;
418 DoInitialMatch(S, L, Good, Bad, SE);
419 if (!Good.empty()) {
420 const SCEV *Sum = SE.getAddExpr(Good);
421 if (!Sum->isZero())
422 BaseRegs.push_back(Sum);
423 HasBaseReg = true;
424 }
425 if (!Bad.empty()) {
426 const SCEV *Sum = SE.getAddExpr(Bad);
427 if (!Sum->isZero())
428 BaseRegs.push_back(Sum);
429 HasBaseReg = true;
430 }
431 canonicalize(*L);
432}
433
434/// \brief Check whether or not this formula statisfies the canonical
435/// representation.
436/// \see Formula::BaseRegs.
437bool Formula::isCanonical(const Loop &L) const {
438 if (!ScaledReg)
439 return BaseRegs.size() <= 1;
440
441 if (Scale != 1)
442 return true;
443
444 if (Scale == 1 && BaseRegs.empty())
445 return false;
446
447 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
448 if (SAR && SAR->getLoop() == &L)
449 return true;
450
451 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
452 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
453 // loop, we want to swap the reg in BaseRegs with ScaledReg.
454 auto I =
455 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
456 return isa<const SCEVAddRecExpr>(S) &&
457 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
458 });
459 return I == BaseRegs.end();
460}
461
462/// \brief Helper method to morph a formula into its canonical representation.
463/// \see Formula::BaseRegs.
464/// Every formula having more than one base register, must use the ScaledReg
465/// field. Otherwise, we would have to do special cases everywhere in LSR
466/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
467/// On the other hand, 1*reg should be canonicalized into reg.
468void Formula::canonicalize(const Loop &L) {
469 if (isCanonical(L))
470 return;
471 // So far we did not need this case. This is easy to implement but it is
472 // useless to maintain dead code. Beside it could hurt compile time.
473 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.")((!BaseRegs.empty() && "1*reg => reg, should not be needed."
) ? static_cast<void> (0) : __assert_fail ("!BaseRegs.empty() && \"1*reg => reg, should not be needed.\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 473, __PRETTY_FUNCTION__))
;
474
475 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
476 if (!ScaledReg) {
477 ScaledReg = BaseRegs.back();
478 BaseRegs.pop_back();
479 Scale = 1;
480 }
481
482 // If ScaledReg is an invariant with respect to L, find the reg from
483 // BaseRegs containing the recurrent expr related with Loop L. Swap the
484 // reg with ScaledReg.
485 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
486 if (!SAR || SAR->getLoop() != &L) {
487 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
488 [&](const SCEV *S) {
489 return isa<const SCEVAddRecExpr>(S) &&
490 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
491 });
492 if (I != BaseRegs.end())
493 std::swap(ScaledReg, *I);
494 }
495}
496
497/// \brief Get rid of the scale in the formula.
498/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
499/// \return true if it was possible to get rid of the scale, false otherwise.
500/// \note After this operation the formula may not be in the canonical form.
501bool Formula::unscale() {
502 if (Scale != 1)
503 return false;
504 Scale = 0;
505 BaseRegs.push_back(ScaledReg);
506 ScaledReg = nullptr;
507 return true;
508}
509
510bool Formula::hasZeroEnd() const {
511 if (UnfoldedOffset || BaseOffset)
512 return false;
513 if (BaseRegs.size() != 1 || ScaledReg)
514 return false;
515 return true;
516}
517
518/// Return the total number of register operands used by this formula. This does
519/// not include register uses implied by non-constant addrec strides.
520size_t Formula::getNumRegs() const {
521 return !!ScaledReg + BaseRegs.size();
522}
523
524/// Return the type of this formula, if it has one, or null otherwise. This type
525/// is meaningless except for the bit size.
526Type *Formula::getType() const {
527 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
528 ScaledReg ? ScaledReg->getType() :
529 BaseGV ? BaseGV->getType() :
530 nullptr;
531}
532
533/// Delete the given base reg from the BaseRegs list.
534void Formula::deleteBaseReg(const SCEV *&S) {
535 if (&S != &BaseRegs.back())
536 std::swap(S, BaseRegs.back());
537 BaseRegs.pop_back();
538}
539
540/// Test if this formula references the given register.
541bool Formula::referencesReg(const SCEV *S) const {
542 return S == ScaledReg || is_contained(BaseRegs, S);
543}
544
545/// Test whether this formula uses registers which are used by uses other than
546/// the use with the given index.
547bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
548 const RegUseTracker &RegUses) const {
549 if (ScaledReg)
550 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
551 return true;
552 for (const SCEV *BaseReg : BaseRegs)
553 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
554 return true;
555 return false;
556}
557
558void Formula::print(raw_ostream &OS) const {
559 bool First = true;
560 if (BaseGV) {
561 if (!First) OS << " + "; else First = false;
562 BaseGV->printAsOperand(OS, /*PrintType=*/false);
563 }
564 if (BaseOffset != 0) {
565 if (!First) OS << " + "; else First = false;
566 OS << BaseOffset;
567 }
568 for (const SCEV *BaseReg : BaseRegs) {
569 if (!First) OS << " + "; else First = false;
570 OS << "reg(" << *BaseReg << ')';
571 }
572 if (HasBaseReg && BaseRegs.empty()) {
573 if (!First) OS << " + "; else First = false;
574 OS << "**error: HasBaseReg**";
575 } else if (!HasBaseReg && !BaseRegs.empty()) {
576 if (!First) OS << " + "; else First = false;
577 OS << "**error: !HasBaseReg**";
578 }
579 if (Scale != 0) {
580 if (!First) OS << " + "; else First = false;
581 OS << Scale << "*reg(";
582 if (ScaledReg)
583 OS << *ScaledReg;
584 else
585 OS << "<unknown>";
586 OS << ')';
587 }
588 if (UnfoldedOffset != 0) {
589 if (!First) OS << " + ";
590 OS << "imm(" << UnfoldedOffset << ')';
591 }
592}
593
594#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
595LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
596 print(errs()); errs() << '\n';
597}
598#endif
599
600/// Return true if the given addrec can be sign-extended without changing its
601/// value.
602static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
603 Type *WideTy =
604 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
605 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
606}
607
608/// Return true if the given add can be sign-extended without changing its
609/// value.
610static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
611 Type *WideTy =
612 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
613 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
614}
615
616/// Return true if the given mul can be sign-extended without changing its
617/// value.
618static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
619 Type *WideTy =
620 IntegerType::get(SE.getContext(),
621 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
622 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
623}
624
625/// Return an expression for LHS /s RHS, if it can be determined and if the
626/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
627/// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
628/// the multiplication may overflow, which is useful when the result will be
629/// used in a context where the most significant bits are ignored.
630static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
631 ScalarEvolution &SE,
632 bool IgnoreSignificantBits = false) {
633 // Handle the trivial case, which works for any SCEV type.
634 if (LHS == RHS)
635 return SE.getConstant(LHS->getType(), 1);
636
637 // Handle a few RHS special cases.
638 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
639 if (RC) {
640 const APInt &RA = RC->getAPInt();
641 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
642 // some folding.
643 if (RA.isAllOnesValue())
644 return SE.getMulExpr(LHS, RC);
645 // Handle x /s 1 as x.
646 if (RA == 1)
647 return LHS;
648 }
649
650 // Check for a division of a constant by a constant.
651 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
652 if (!RC)
653 return nullptr;
654 const APInt &LA = C->getAPInt();
655 const APInt &RA = RC->getAPInt();
656 if (LA.srem(RA) != 0)
657 return nullptr;
658 return SE.getConstant(LA.sdiv(RA));
659 }
660
661 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
662 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
663 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
664 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
665 IgnoreSignificantBits);
666 if (!Step) return nullptr;
667 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
668 IgnoreSignificantBits);
669 if (!Start) return nullptr;
670 // FlagNW is independent of the start value, step direction, and is
671 // preserved with smaller magnitude steps.
672 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
673 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
674 }
675 return nullptr;
676 }
677
678 // Distribute the sdiv over add operands, if the add doesn't overflow.
679 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
680 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
681 SmallVector<const SCEV *, 8> Ops;
682 for (const SCEV *S : Add->operands()) {
683 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
684 if (!Op) return nullptr;
685 Ops.push_back(Op);
686 }
687 return SE.getAddExpr(Ops);
688 }
689 return nullptr;
690 }
691
692 // Check for a multiply operand that we can pull RHS out of.
693 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
694 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
695 SmallVector<const SCEV *, 4> Ops;
696 bool Found = false;
697 for (const SCEV *S : Mul->operands()) {
698 if (!Found)
699 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
700 IgnoreSignificantBits)) {
701 S = Q;
702 Found = true;
703 }
704 Ops.push_back(S);
705 }
706 return Found ? SE.getMulExpr(Ops) : nullptr;
707 }
708 return nullptr;
709 }
710
711 // Otherwise we don't know.
712 return nullptr;
713}
714
715/// If S involves the addition of a constant integer value, return that integer
716/// value, and mutate S to point to a new SCEV with that value excluded.
717static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
718 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
719 if (C->getAPInt().getMinSignedBits() <= 64) {
720 S = SE.getConstant(C->getType(), 0);
721 return C->getValue()->getSExtValue();
722 }
723 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
724 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
725 int64_t Result = ExtractImmediate(NewOps.front(), SE);
726 if (Result != 0)
727 S = SE.getAddExpr(NewOps);
728 return Result;
729 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
730 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
731 int64_t Result = ExtractImmediate(NewOps.front(), SE);
732 if (Result != 0)
733 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
734 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
735 SCEV::FlagAnyWrap);
736 return Result;
737 }
738 return 0;
739}
740
741/// If S involves the addition of a GlobalValue address, return that symbol, and
742/// mutate S to point to a new SCEV with that value excluded.
743static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
744 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
745 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
746 S = SE.getConstant(GV->getType(), 0);
747 return GV;
748 }
749 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
750 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
751 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
752 if (Result)
753 S = SE.getAddExpr(NewOps);
754 return Result;
755 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
756 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
757 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
758 if (Result)
759 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
760 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
761 SCEV::FlagAnyWrap);
762 return Result;
763 }
764 return nullptr;
765}
766
767/// Returns true if the specified instruction is using the specified value as an
768/// address.
769static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
770 bool isAddress = isa<LoadInst>(Inst);
771 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
772 if (SI->getPointerOperand() == OperandVal)
773 isAddress = true;
774 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
775 // Addressing modes can also be folded into prefetches and a variety
776 // of intrinsics.
777 switch (II->getIntrinsicID()) {
778 default: break;
779 case Intrinsic::prefetch:
780 if (II->getArgOperand(0) == OperandVal)
781 isAddress = true;
782 break;
783 }
784 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
785 if (RMW->getPointerOperand() == OperandVal)
786 isAddress = true;
787 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
788 if (CmpX->getPointerOperand() == OperandVal)
789 isAddress = true;
790 }
791 return isAddress;
792}
793
794/// Return the type of the memory being accessed.
795static MemAccessTy getAccessType(const Instruction *Inst) {
796 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
797 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
798 AccessTy.MemTy = SI->getOperand(0)->getType();
799 AccessTy.AddrSpace = SI->getPointerAddressSpace();
800 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
801 AccessTy.AddrSpace = LI->getPointerAddressSpace();
802 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
803 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
804 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
805 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
806 }
807
808 // All pointers have the same requirements, so canonicalize them to an
809 // arbitrary pointer type to minimize variation.
810 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
811 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
812 PTy->getAddressSpace());
813
814 return AccessTy;
815}
816
817/// Return true if this AddRec is already a phi in its loop.
818static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
819 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
820 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
821 if (SE.isSCEVable(PN->getType()) &&
822 (SE.getEffectiveSCEVType(PN->getType()) ==
823 SE.getEffectiveSCEVType(AR->getType())) &&
824 SE.getSCEV(PN) == AR)
825 return true;
826 }
827 return false;
828}
829
830/// Check if expanding this expression is likely to incur significant cost. This
831/// is tricky because SCEV doesn't track which expressions are actually computed
832/// by the current IR.
833///
834/// We currently allow expansion of IV increments that involve adds,
835/// multiplication by constants, and AddRecs from existing phis.
836///
837/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
838/// obvious multiple of the UDivExpr.
839static bool isHighCostExpansion(const SCEV *S,
840 SmallPtrSetImpl<const SCEV*> &Processed,
841 ScalarEvolution &SE) {
842 // Zero/One operand expressions
843 switch (S->getSCEVType()) {
844 case scUnknown:
845 case scConstant:
846 return false;
847 case scTruncate:
848 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
849 Processed, SE);
850 case scZeroExtend:
851 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
852 Processed, SE);
853 case scSignExtend:
854 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
855 Processed, SE);
856 }
857
858 if (!Processed.insert(S).second)
859 return false;
860
861 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
862 for (const SCEV *S : Add->operands()) {
863 if (isHighCostExpansion(S, Processed, SE))
864 return true;
865 }
866 return false;
867 }
868
869 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
870 if (Mul->getNumOperands() == 2) {
871 // Multiplication by a constant is ok
872 if (isa<SCEVConstant>(Mul->getOperand(0)))
873 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
874
875 // If we have the value of one operand, check if an existing
876 // multiplication already generates this expression.
877 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
878 Value *UVal = U->getValue();
879 for (User *UR : UVal->users()) {
880 // If U is a constant, it may be used by a ConstantExpr.
881 Instruction *UI = dyn_cast<Instruction>(UR);
882 if (UI && UI->getOpcode() == Instruction::Mul &&
883 SE.isSCEVable(UI->getType())) {
884 return SE.getSCEV(UI) == Mul;
885 }
886 }
887 }
888 }
889 }
890
891 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
892 if (isExistingPhi(AR, SE))
893 return false;
894 }
895
896 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
897 return true;
898}
899
900/// If any of the instructions is the specified set are trivially dead, delete
901/// them and see if this makes any of their operands subsequently dead.
902static bool
903DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
904 bool Changed = false;
905
906 while (!DeadInsts.empty()) {
907 Value *V = DeadInsts.pop_back_val();
908 Instruction *I = dyn_cast_or_null<Instruction>(V);
909
910 if (!I || !isInstructionTriviallyDead(I))
911 continue;
912
913 for (Use &O : I->operands())
914 if (Instruction *U = dyn_cast<Instruction>(O)) {
915 O = nullptr;
916 if (U->use_empty())
917 DeadInsts.emplace_back(U);
918 }
919
920 I->eraseFromParent();
921 Changed = true;
922 }
923
924 return Changed;
925}
926
927namespace {
928
929class LSRUse;
930
931} // end anonymous namespace
932
933/// \brief Check if the addressing mode defined by \p F is completely
934/// folded in \p LU at isel time.
935/// This includes address-mode folding and special icmp tricks.
936/// This function returns true if \p LU can accommodate what \p F
937/// defines and up to 1 base + 1 scaled + offset.
938/// In other words, if \p F has several base registers, this function may
939/// still return true. Therefore, users still need to account for
940/// additional base registers and/or unfolded offsets to derive an
941/// accurate cost model.
942static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
943 const LSRUse &LU, const Formula &F);
944// Get the cost of the scaling factor used in F for LU.
945static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
946 const LSRUse &LU, const Formula &F,
947 const Loop &L);
948
949namespace {
950
951/// This class is used to measure and compare candidate formulae.
952class Cost {
953 /// TODO: Some of these could be merged. Also, a lexical ordering
954 /// isn't always optimal.
955 unsigned Insns;
956 unsigned NumRegs;
957 unsigned AddRecCost;
958 unsigned NumIVMuls;
959 unsigned NumBaseAdds;
960 unsigned ImmCost;
961 unsigned SetupCost;
962 unsigned ScaleCost;
963
964public:
965 Cost()
966 : Insns(0), NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0),
967 ImmCost(0), SetupCost(0), ScaleCost(0) {}
968
969 bool operator<(const Cost &Other) const;
970
971 void Lose();
972
973#ifndef NDEBUG
974 // Once any of the metrics loses, they must all remain losers.
975 bool isValid() {
976 return ((Insns | NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
977 | ImmCost | SetupCost | ScaleCost) != ~0u)
978 || ((Insns & NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
979 & ImmCost & SetupCost & ScaleCost) == ~0u);
980 }
981#endif
982
983 bool isLoser() {
984 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 984, __PRETTY_FUNCTION__))
;
985 return NumRegs == ~0u;
986 }
987
988 void RateFormula(const TargetTransformInfo &TTI,
989 const Formula &F,
990 SmallPtrSetImpl<const SCEV *> &Regs,
991 const DenseSet<const SCEV *> &VisitedRegs,
992 const Loop *L,
993 ScalarEvolution &SE, DominatorTree &DT,
994 const LSRUse &LU,
995 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
996
997 void print(raw_ostream &OS) const;
998 void dump() const;
999
1000private:
1001 void RateRegister(const SCEV *Reg,
1002 SmallPtrSetImpl<const SCEV *> &Regs,
1003 const Loop *L,
1004 ScalarEvolution &SE, DominatorTree &DT);
1005 void RatePrimaryRegister(const SCEV *Reg,
1006 SmallPtrSetImpl<const SCEV *> &Regs,
1007 const Loop *L,
1008 ScalarEvolution &SE, DominatorTree &DT,
1009 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1010};
1011
1012/// An operand value in an instruction which is to be replaced with some
1013/// equivalent, possibly strength-reduced, replacement.
1014struct LSRFixup {
1015 /// The instruction which will be updated.
1016 Instruction *UserInst;
1017
1018 /// The operand of the instruction which will be replaced. The operand may be
1019 /// used more than once; every instance will be replaced.
1020 Value *OperandValToReplace;
1021
1022 /// If this user is to use the post-incremented value of an induction
1023 /// variable, this variable is non-null and holds the loop associated with the
1024 /// induction variable.
1025 PostIncLoopSet PostIncLoops;
1026
1027 /// A constant offset to be added to the LSRUse expression. This allows
1028 /// multiple fixups to share the same LSRUse with different offsets, for
1029 /// example in an unrolled loop.
1030 int64_t Offset;
1031
1032 bool isUseFullyOutsideLoop(const Loop *L) const;
1033
1034 LSRFixup();
1035
1036 void print(raw_ostream &OS) const;
1037 void dump() const;
1038};
1039
1040/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1041/// SmallVectors of const SCEV*.
1042struct UniquifierDenseMapInfo {
1043 static SmallVector<const SCEV *, 4> getEmptyKey() {
1044 SmallVector<const SCEV *, 4> V;
1045 V.push_back(reinterpret_cast<const SCEV *>(-1));
1046 return V;
1047 }
1048
1049 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1050 SmallVector<const SCEV *, 4> V;
1051 V.push_back(reinterpret_cast<const SCEV *>(-2));
1052 return V;
1053 }
1054
1055 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1056 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1057 }
1058
1059 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1060 const SmallVector<const SCEV *, 4> &RHS) {
1061 return LHS == RHS;
1062 }
1063};
1064
1065/// This class holds the state that LSR keeps for each use in IVUsers, as well
1066/// as uses invented by LSR itself. It includes information about what kinds of
1067/// things can be folded into the user, information about the user itself, and
1068/// information about how the use may be satisfied. TODO: Represent multiple
1069/// users of the same expression in common?
1070class LSRUse {
1071 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1072
1073public:
1074 /// An enum for a kind of use, indicating what types of scaled and immediate
1075 /// operands it might support.
1076 enum KindType {
1077 Basic, ///< A normal use, with no folding.
1078 Special, ///< A special case of basic, allowing -1 scales.
1079 Address, ///< An address use; folding according to TargetLowering
1080 ICmpZero ///< An equality icmp with both operands folded into one.
1081 // TODO: Add a generic icmp too?
1082 };
1083
1084 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1085
1086 KindType Kind;
1087 MemAccessTy AccessTy;
1088
1089 /// The list of operands which are to be replaced.
1090 SmallVector<LSRFixup, 8> Fixups;
1091
1092 /// Keep track of the min and max offsets of the fixups.
1093 int64_t MinOffset;
1094 int64_t MaxOffset;
1095
1096 /// This records whether all of the fixups using this LSRUse are outside of
1097 /// the loop, in which case some special-case heuristics may be used.
1098 bool AllFixupsOutsideLoop;
1099
1100 /// RigidFormula is set to true to guarantee that this use will be associated
1101 /// with a single formula--the one that initially matched. Some SCEV
1102 /// expressions cannot be expanded. This allows LSR to consider the registers
1103 /// used by those expressions without the need to expand them later after
1104 /// changing the formula.
1105 bool RigidFormula;
1106
1107 /// This records the widest use type for any fixup using this
1108 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1109 /// fixup widths to be equivalent, because the narrower one may be relying on
1110 /// the implicit truncation to truncate away bogus bits.
1111 Type *WidestFixupType;
1112
1113 /// A list of ways to build a value that can satisfy this user. After the
1114 /// list is populated, one of these is selected heuristically and used to
1115 /// formulate a replacement for OperandValToReplace in UserInst.
1116 SmallVector<Formula, 12> Formulae;
1117
1118 /// The set of register candidates used by all formulae in this LSRUse.
1119 SmallPtrSet<const SCEV *, 4> Regs;
1120
1121 LSRUse(KindType K, MemAccessTy AT)
1122 : Kind(K), AccessTy(AT), MinOffset(INT64_MAX(9223372036854775807L)), MaxOffset(INT64_MIN(-9223372036854775807L -1)),
1123 AllFixupsOutsideLoop(true), RigidFormula(false),
1124 WidestFixupType(nullptr) {}
1125
1126 LSRFixup &getNewFixup() {
1127 Fixups.push_back(LSRFixup());
1128 return Fixups.back();
1129 }
1130
1131 void pushFixup(LSRFixup &f) {
1132 Fixups.push_back(f);
1133 if (f.Offset > MaxOffset)
1134 MaxOffset = f.Offset;
1135 if (f.Offset < MinOffset)
1136 MinOffset = f.Offset;
1137 }
1138
1139 bool HasFormulaWithSameRegs(const Formula &F) const;
1140 float getNotSelectedProbability(const SCEV *Reg) const;
1141 bool InsertFormula(const Formula &F, const Loop &L);
1142 void DeleteFormula(Formula &F);
1143 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1144
1145 void print(raw_ostream &OS) const;
1146 void dump() const;
1147};
1148
1149} // end anonymous namespace
1150
1151/// Tally up interesting quantities from the given register.
1152void Cost::RateRegister(const SCEV *Reg,
1153 SmallPtrSetImpl<const SCEV *> &Regs,
1154 const Loop *L,
1155 ScalarEvolution &SE, DominatorTree &DT) {
1156 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1157 // If this is an addrec for another loop, it should be an invariant
1158 // with respect to L since L is the innermost loop (at least
1159 // for now LSR only handles innermost loops).
1160 if (AR->getLoop() != L) {
1161 // If the AddRec exists, consider it's register free and leave it alone.
1162 if (isExistingPhi(AR, SE))
1163 return;
1164
1165 // It is bad to allow LSR for current loop to add induction variables
1166 // for its sibling loops.
1167 if (!AR->getLoop()->contains(L)) {
1168 Lose();
1169 return;
1170 }
1171
1172 // Otherwise, it will be an invariant with respect to Loop L.
1173 ++NumRegs;
1174 return;
1175 }
1176 AddRecCost += 1; /// TODO: This should be a function of the stride.
1177
1178 // Add the step value register, if it needs one.
1179 // TODO: The non-affine case isn't precisely modeled here.
1180 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1181 if (!Regs.count(AR->getOperand(1))) {
1182 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
1183 if (isLoser())
1184 return;
1185 }
1186 }
1187 }
1188 ++NumRegs;
1189
1190 // Rough heuristic; favor registers which don't require extra setup
1191 // instructions in the preheader.
1192 if (!isa<SCEVUnknown>(Reg) &&
1193 !isa<SCEVConstant>(Reg) &&
1194 !(isa<SCEVAddRecExpr>(Reg) &&
1195 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1196 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1197 ++SetupCost;
1198
1199 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1200 SE.hasComputableLoopEvolution(Reg, L);
1201}
1202
1203/// Record this register in the set. If we haven't seen it before, rate
1204/// it. Optional LoserRegs provides a way to declare any formula that refers to
1205/// one of those regs an instant loser.
1206void Cost::RatePrimaryRegister(const SCEV *Reg,
1207 SmallPtrSetImpl<const SCEV *> &Regs,
1208 const Loop *L,
1209 ScalarEvolution &SE, DominatorTree &DT,
1210 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1211 if (LoserRegs && LoserRegs->count(Reg)) {
1212 Lose();
1213 return;
1214 }
1215 if (Regs.insert(Reg).second) {
1216 RateRegister(Reg, Regs, L, SE, DT);
1217 if (LoserRegs && isLoser())
1218 LoserRegs->insert(Reg);
1219 }
1220}
1221
1222void Cost::RateFormula(const TargetTransformInfo &TTI,
1223 const Formula &F,
1224 SmallPtrSetImpl<const SCEV *> &Regs,
1225 const DenseSet<const SCEV *> &VisitedRegs,
1226 const Loop *L,
1227 ScalarEvolution &SE, DominatorTree &DT,
1228 const LSRUse &LU,
1229 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1230 assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula")((F.isCanonical(*L) && "Cost is accurate only for canonical formula"
) ? static_cast<void> (0) : __assert_fail ("F.isCanonical(*L) && \"Cost is accurate only for canonical formula\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1230, __PRETTY_FUNCTION__))
;
1231 // Tally up the registers.
1232 unsigned PrevAddRecCost = AddRecCost;
1233 unsigned PrevNumRegs = NumRegs;
1234 unsigned PrevNumBaseAdds = NumBaseAdds;
1235 if (const SCEV *ScaledReg = F.ScaledReg) {
1236 if (VisitedRegs.count(ScaledReg)) {
1237 Lose();
1238 return;
1239 }
1240 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1241 if (isLoser())
1242 return;
1243 }
1244 for (const SCEV *BaseReg : F.BaseRegs) {
1245 if (VisitedRegs.count(BaseReg)) {
1246 Lose();
1247 return;
1248 }
1249 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1250 if (isLoser())
1251 return;
1252 }
1253
1254 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1255 // additional instruction (at least fill).
1256 unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1257 if (NumRegs > TTIRegNum) {
1258 // Cost already exceeded TTIRegNum, then only newly added register can add
1259 // new instructions.
1260 if (PrevNumRegs > TTIRegNum)
1261 Insns += (NumRegs - PrevNumRegs);
1262 else
1263 Insns += (NumRegs - TTIRegNum);
1264 }
1265
1266 // Determine how many (unfolded) adds we'll need inside the loop.
1267 size_t NumBaseParts = F.getNumRegs();
1268 if (NumBaseParts > 1)
1269 // Do not count the base and a possible second register if the target
1270 // allows to fold 2 registers.
1271 NumBaseAdds +=
1272 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1273 NumBaseAdds += (F.UnfoldedOffset != 0);
1274
1275 // Accumulate non-free scaling amounts.
1276 ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1277
1278 // Tally up the non-zero immediates.
1279 for (const LSRFixup &Fixup : LU.Fixups) {
1280 int64_t O = Fixup.Offset;
1281 int64_t Offset = (uint64_t)O + F.BaseOffset;
1282 if (F.BaseGV)
1283 ImmCost += 64; // Handle symbolic values conservatively.
1284 // TODO: This should probably be the pointer size.
1285 else if (Offset != 0)
1286 ImmCost += APInt(64, Offset, true).getMinSignedBits();
1287
1288 // Check with target if this offset with this instruction is
1289 // specifically not supported.
1290 if ((isa<LoadInst>(Fixup.UserInst) || isa<StoreInst>(Fixup.UserInst)) &&
1291 !TTI.isFoldableMemAccessOffset(Fixup.UserInst, Offset))
1292 NumBaseAdds++;
1293 }
1294
1295 // If ICmpZero formula ends with not 0, it could not be replaced by
1296 // just add or sub. We'll need to compare final result of AddRec.
1297 // That means we'll need an additional instruction.
1298 // For -10 + {0, +, 1}:
1299 // i = i + 1;
1300 // cmp i, 10
1301 //
1302 // For {-10, +, 1}:
1303 // i = i + 1;
1304 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd())
1305 Insns++;
1306 // Each new AddRec adds 1 instruction to calculation.
1307 Insns += (AddRecCost - PrevAddRecCost);
1308
1309 // BaseAdds adds instructions for unfolded registers.
1310 if (LU.Kind != LSRUse::ICmpZero)
1311 Insns += NumBaseAdds - PrevNumBaseAdds;
1312 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1312, __PRETTY_FUNCTION__))
;
1313}
1314
1315/// Set this cost to a losing value.
1316void Cost::Lose() {
1317 Insns = ~0u;
1318 NumRegs = ~0u;
1319 AddRecCost = ~0u;
1320 NumIVMuls = ~0u;
1321 NumBaseAdds = ~0u;
1322 ImmCost = ~0u;
1323 SetupCost = ~0u;
1324 ScaleCost = ~0u;
1325}
1326
1327/// Choose the lower cost.
1328bool Cost::operator<(const Cost &Other) const {
1329 if (InsnsCost && Insns != Other.Insns)
1330 return Insns < Other.Insns;
1331 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
1332 ImmCost, SetupCost) <
1333 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
1334 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
1335 Other.SetupCost);
1336}
1337
1338void Cost::print(raw_ostream &OS) const {
1339 OS << Insns << " instruction" << (Insns == 1 ? " " : "s ");
1340 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1341 if (AddRecCost != 0)
1342 OS << ", with addrec cost " << AddRecCost;
1343 if (NumIVMuls != 0)
1344 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1345 if (NumBaseAdds != 0)
1346 OS << ", plus " << NumBaseAdds << " base add"
1347 << (NumBaseAdds == 1 ? "" : "s");
1348 if (ScaleCost != 0)
1349 OS << ", plus " << ScaleCost << " scale cost";
1350 if (ImmCost != 0)
1351 OS << ", plus " << ImmCost << " imm cost";
1352 if (SetupCost != 0)
1353 OS << ", plus " << SetupCost << " setup cost";
1354}
1355
1356#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1357LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1358 print(errs()); errs() << '\n';
1359}
1360#endif
1361
1362LSRFixup::LSRFixup()
1363 : UserInst(nullptr), OperandValToReplace(nullptr),
1364 Offset(0) {}
1365
1366/// Test whether this fixup always uses its value outside of the given loop.
1367bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1368 // PHI nodes use their value in their incoming blocks.
1369 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1370 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1371 if (PN->getIncomingValue(i) == OperandValToReplace &&
1372 L->contains(PN->getIncomingBlock(i)))
1373 return false;
1374 return true;
1375 }
1376
1377 return !L->contains(UserInst);
1378}
1379
1380void LSRFixup::print(raw_ostream &OS) const {
1381 OS << "UserInst=";
1382 // Store is common and interesting enough to be worth special-casing.
1383 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1384 OS << "store ";
1385 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1386 } else if (UserInst->getType()->isVoidTy())
1387 OS << UserInst->getOpcodeName();
1388 else
1389 UserInst->printAsOperand(OS, /*PrintType=*/false);
1390
1391 OS << ", OperandValToReplace=";
1392 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1393
1394 for (const Loop *PIL : PostIncLoops) {
1395 OS << ", PostIncLoop=";
1396 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1397 }
1398
1399 if (Offset != 0)
1400 OS << ", Offset=" << Offset;
1401}
1402
1403#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1404LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1405 print(errs()); errs() << '\n';
1406}
1407#endif
1408
1409/// Test whether this use as a formula which has the same registers as the given
1410/// formula.
1411bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1412 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1413 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1414 // Unstable sort by host order ok, because this is only used for uniquifying.
1415 std::sort(Key.begin(), Key.end());
1416 return Uniquifier.count(Key);
1417}
1418
1419/// The function returns a probability of selecting formula without Reg.
1420float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1421 unsigned FNum = 0;
1422 for (const Formula &F : Formulae)
1423 if (F.referencesReg(Reg))
1424 FNum++;
1425 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1426}
1427
1428/// If the given formula has not yet been inserted, add it to the list, and
1429/// return true. Return false otherwise. The formula must be in canonical form.
1430bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1431 assert(F.isCanonical(L) && "Invalid canonical representation")((F.isCanonical(L) && "Invalid canonical representation"
) ? static_cast<void> (0) : __assert_fail ("F.isCanonical(L) && \"Invalid canonical representation\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1431, __PRETTY_FUNCTION__))
;
1432
1433 if (!Formulae.empty() && RigidFormula)
1434 return false;
1435
1436 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1437 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1438 // Unstable sort by host order ok, because this is only used for uniquifying.
1439 std::sort(Key.begin(), Key.end());
1440
1441 if (!Uniquifier.insert(Key).second)
1442 return false;
1443
1444 // Using a register to hold the value of 0 is not profitable.
1445 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&(((!F.ScaledReg || !F.ScaledReg->isZero()) && "Zero allocated in a scaled register!"
) ? static_cast<void> (0) : __assert_fail ("(!F.ScaledReg || !F.ScaledReg->isZero()) && \"Zero allocated in a scaled register!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1446, __PRETTY_FUNCTION__))
1446 "Zero allocated in a scaled register!")(((!F.ScaledReg || !F.ScaledReg->isZero()) && "Zero allocated in a scaled register!"
) ? static_cast<void> (0) : __assert_fail ("(!F.ScaledReg || !F.ScaledReg->isZero()) && \"Zero allocated in a scaled register!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1446, __PRETTY_FUNCTION__))
;
1447#ifndef NDEBUG
1448 for (const SCEV *BaseReg : F.BaseRegs)
1449 assert(!BaseReg->isZero() && "Zero allocated in a base register!")((!BaseReg->isZero() && "Zero allocated in a base register!"
) ? static_cast<void> (0) : __assert_fail ("!BaseReg->isZero() && \"Zero allocated in a base register!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1449, __PRETTY_FUNCTION__))
;
1450#endif
1451
1452 // Add the formula to the list.
1453 Formulae.push_back(F);
1454
1455 // Record registers now being used by this use.
1456 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1457 if (F.ScaledReg)
1458 Regs.insert(F.ScaledReg);
1459
1460 return true;
1461}
1462
1463/// Remove the given formula from this use's list.
1464void LSRUse::DeleteFormula(Formula &F) {
1465 if (&F != &Formulae.back())
1466 std::swap(F, Formulae.back());
1467 Formulae.pop_back();
1468}
1469
1470/// Recompute the Regs field, and update RegUses.
1471void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1472 // Now that we've filtered out some formulae, recompute the Regs set.
1473 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1474 Regs.clear();
1475 for (const Formula &F : Formulae) {
1476 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1477 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1478 }
1479
1480 // Update the RegTracker.
1481 for (const SCEV *S : OldRegs)
1482 if (!Regs.count(S))
1483 RegUses.dropRegister(S, LUIdx);
1484}
1485
1486void LSRUse::print(raw_ostream &OS) const {
1487 OS << "LSR Use: Kind=";
1488 switch (Kind) {
1489 case Basic: OS << "Basic"; break;
1490 case Special: OS << "Special"; break;
1491 case ICmpZero: OS << "ICmpZero"; break;
1492 case Address:
1493 OS << "Address of ";
1494 if (AccessTy.MemTy->isPointerTy())
1495 OS << "pointer"; // the full pointer type could be really verbose
1496 else {
1497 OS << *AccessTy.MemTy;
1498 }
1499
1500 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1501 }
1502
1503 OS << ", Offsets={";
1504 bool NeedComma = false;
1505 for (const LSRFixup &Fixup : Fixups) {
1506 if (NeedComma) OS << ',';
1507 OS << Fixup.Offset;
1508 NeedComma = true;
1509 }
1510 OS << '}';
1511
1512 if (AllFixupsOutsideLoop)
1513 OS << ", all-fixups-outside-loop";
1514
1515 if (WidestFixupType)
1516 OS << ", widest fixup type: " << *WidestFixupType;
1517}
1518
1519#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1520LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1521 print(errs()); errs() << '\n';
1522}
1523#endif
1524
1525static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1526 LSRUse::KindType Kind, MemAccessTy AccessTy,
1527 GlobalValue *BaseGV, int64_t BaseOffset,
1528 bool HasBaseReg, int64_t Scale) {
1529 switch (Kind) {
1530 case LSRUse::Address:
1531 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1532 HasBaseReg, Scale, AccessTy.AddrSpace);
1533
1534 case LSRUse::ICmpZero:
1535 // There's not even a target hook for querying whether it would be legal to
1536 // fold a GV into an ICmp.
1537 if (BaseGV)
1538 return false;
1539
1540 // ICmp only has two operands; don't allow more than two non-trivial parts.
1541 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1542 return false;
1543
1544 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1545 // putting the scaled register in the other operand of the icmp.
1546 if (Scale != 0 && Scale != -1)
1547 return false;
1548
1549 // If we have low-level target information, ask the target if it can fold an
1550 // integer immediate on an icmp.
1551 if (BaseOffset != 0) {
1552 // We have one of:
1553 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1554 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1555 // Offs is the ICmp immediate.
1556 if (Scale == 0)
1557 // The cast does the right thing with INT64_MIN.
1558 BaseOffset = -(uint64_t)BaseOffset;
1559 return TTI.isLegalICmpImmediate(BaseOffset);
1560 }
1561
1562 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1563 return true;
1564
1565 case LSRUse::Basic:
1566 // Only handle single-register values.
1567 return !BaseGV && Scale == 0 && BaseOffset == 0;
1568
1569 case LSRUse::Special:
1570 // Special case Basic to handle -1 scales.
1571 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1572 }
1573
1574 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1574)
;
1575}
1576
1577static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1578 int64_t MinOffset, int64_t MaxOffset,
1579 LSRUse::KindType Kind, MemAccessTy AccessTy,
1580 GlobalValue *BaseGV, int64_t BaseOffset,
1581 bool HasBaseReg, int64_t Scale) {
1582 // Check for overflow.
1583 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1584 (MinOffset > 0))
1585 return false;
1586 MinOffset = (uint64_t)BaseOffset + MinOffset;
1587 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1588 (MaxOffset > 0))
1589 return false;
1590 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1591
1592 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1593 HasBaseReg, Scale) &&
1594 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1595 HasBaseReg, Scale);
1596}
1597
1598static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1599 int64_t MinOffset, int64_t MaxOffset,
1600 LSRUse::KindType Kind, MemAccessTy AccessTy,
1601 const Formula &F, const Loop &L) {
1602 // For the purpose of isAMCompletelyFolded either having a canonical formula
1603 // or a scale not equal to zero is correct.
1604 // Problems may arise from non canonical formulae having a scale == 0.
1605 // Strictly speaking it would best to just rely on canonical formulae.
1606 // However, when we generate the scaled formulae, we first check that the
1607 // scaling factor is profitable before computing the actual ScaledReg for
1608 // compile time sake.
1609 assert((F.isCanonical(L) || F.Scale != 0))(((F.isCanonical(L) || F.Scale != 0)) ? static_cast<void>
(0) : __assert_fail ("(F.isCanonical(L) || F.Scale != 0)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1609, __PRETTY_FUNCTION__))
;
1610 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1611 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1612}
1613
1614/// Test whether we know how to expand the current formula.
1615static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1616 int64_t MaxOffset, LSRUse::KindType Kind,
1617 MemAccessTy AccessTy, GlobalValue *BaseGV,
1618 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1619 // We know how to expand completely foldable formulae.
1620 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1621 BaseOffset, HasBaseReg, Scale) ||
1622 // Or formulae that use a base register produced by a sum of base
1623 // registers.
1624 (Scale == 1 &&
1625 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1626 BaseGV, BaseOffset, true, 0));
1627}
1628
1629static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1630 int64_t MaxOffset, LSRUse::KindType Kind,
1631 MemAccessTy AccessTy, const Formula &F) {
1632 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1633 F.BaseOffset, F.HasBaseReg, F.Scale);
1634}
1635
1636static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1637 const LSRUse &LU, const Formula &F) {
1638 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1639 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1640 F.Scale);
1641}
1642
1643static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1644 const LSRUse &LU, const Formula &F,
1645 const Loop &L) {
1646 if (!F.Scale)
1647 return 0;
1648
1649 // If the use is not completely folded in that instruction, we will have to
1650 // pay an extra cost only for scale != 1.
1651 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1652 LU.AccessTy, F, L))
1653 return F.Scale != 1;
1654
1655 switch (LU.Kind) {
1656 case LSRUse::Address: {
1657 // Check the scaling factor cost with both the min and max offsets.
1658 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1659 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1660 F.Scale, LU.AccessTy.AddrSpace);
1661 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1662 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1663 F.Scale, LU.AccessTy.AddrSpace);
1664
1665 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&((ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >=
0 && "Legal addressing mode has an illegal cost!") ?
static_cast<void> (0) : __assert_fail ("ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 && \"Legal addressing mode has an illegal cost!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1666, __PRETTY_FUNCTION__))
1666 "Legal addressing mode has an illegal cost!")((ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >=
0 && "Legal addressing mode has an illegal cost!") ?
static_cast<void> (0) : __assert_fail ("ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 && \"Legal addressing mode has an illegal cost!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1666, __PRETTY_FUNCTION__))
;
1667 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1668 }
1669 case LSRUse::ICmpZero:
1670 case LSRUse::Basic:
1671 case LSRUse::Special:
1672 // The use is completely folded, i.e., everything is folded into the
1673 // instruction.
1674 return 0;
1675 }
1676
1677 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1677)
;
1678}
1679
1680static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1681 LSRUse::KindType Kind, MemAccessTy AccessTy,
1682 GlobalValue *BaseGV, int64_t BaseOffset,
1683 bool HasBaseReg) {
1684 // Fast-path: zero is always foldable.
1685 if (BaseOffset == 0 && !BaseGV) return true;
1686
1687 // Conservatively, create an address with an immediate and a
1688 // base and a scale.
1689 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1690
1691 // Canonicalize a scale of 1 to a base register if the formula doesn't
1692 // already have a base register.
1693 if (!HasBaseReg && Scale == 1) {
1694 Scale = 0;
1695 HasBaseReg = true;
1696 }
1697
1698 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1699 HasBaseReg, Scale);
1700}
1701
1702static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1703 ScalarEvolution &SE, int64_t MinOffset,
1704 int64_t MaxOffset, LSRUse::KindType Kind,
1705 MemAccessTy AccessTy, const SCEV *S,
1706 bool HasBaseReg) {
1707 // Fast-path: zero is always foldable.
1708 if (S->isZero()) return true;
1709
1710 // Conservatively, create an address with an immediate and a
1711 // base and a scale.
1712 int64_t BaseOffset = ExtractImmediate(S, SE);
1713 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1714
1715 // If there's anything else involved, it's not foldable.
1716 if (!S->isZero()) return false;
1717
1718 // Fast-path: zero is always foldable.
1719 if (BaseOffset == 0 && !BaseGV) return true;
1720
1721 // Conservatively, create an address with an immediate and a
1722 // base and a scale.
1723 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1724
1725 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1726 BaseOffset, HasBaseReg, Scale);
1727}
1728
1729namespace {
1730
1731/// An individual increment in a Chain of IV increments. Relate an IV user to
1732/// an expression that computes the IV it uses from the IV used by the previous
1733/// link in the Chain.
1734///
1735/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1736/// original IVOperand. The head of the chain's IVOperand is only valid during
1737/// chain collection, before LSR replaces IV users. During chain generation,
1738/// IncExpr can be used to find the new IVOperand that computes the same
1739/// expression.
1740struct IVInc {
1741 Instruction *UserInst;
1742 Value* IVOperand;
1743 const SCEV *IncExpr;
1744
1745 IVInc(Instruction *U, Value *O, const SCEV *E):
1746 UserInst(U), IVOperand(O), IncExpr(E) {}
1747};
1748
1749// The list of IV increments in program order. We typically add the head of a
1750// chain without finding subsequent links.
1751struct IVChain {
1752 SmallVector<IVInc,1> Incs;
1753 const SCEV *ExprBase;
1754
1755 IVChain() : ExprBase(nullptr) {}
1756
1757 IVChain(const IVInc &Head, const SCEV *Base)
1758 : Incs(1, Head), ExprBase(Base) {}
1759
1760 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1761
1762 // Return the first increment in the chain.
1763 const_iterator begin() const {
1764 assert(!Incs.empty())((!Incs.empty()) ? static_cast<void> (0) : __assert_fail
("!Incs.empty()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1764, __PRETTY_FUNCTION__))
;
1765 return std::next(Incs.begin());
1766 }
1767 const_iterator end() const {
1768 return Incs.end();
1769 }
1770
1771 // Returns true if this chain contains any increments.
1772 bool hasIncs() const { return Incs.size() >= 2; }
1773
1774 // Add an IVInc to the end of this chain.
1775 void add(const IVInc &X) { Incs.push_back(X); }
1776
1777 // Returns the last UserInst in the chain.
1778 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1779
1780 // Returns true if IncExpr can be profitably added to this chain.
1781 bool isProfitableIncrement(const SCEV *OperExpr,
1782 const SCEV *IncExpr,
1783 ScalarEvolution&);
1784};
1785
1786/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1787/// between FarUsers that definitely cross IV increments and NearUsers that may
1788/// be used between IV increments.
1789struct ChainUsers {
1790 SmallPtrSet<Instruction*, 4> FarUsers;
1791 SmallPtrSet<Instruction*, 4> NearUsers;
1792};
1793
1794/// This class holds state for the main loop strength reduction logic.
1795class LSRInstance {
1796 IVUsers &IU;
1797 ScalarEvolution &SE;
1798 DominatorTree &DT;
1799 LoopInfo &LI;
1800 const TargetTransformInfo &TTI;
1801 Loop *const L;
1802 bool Changed;
1803
1804 /// This is the insert position that the current loop's induction variable
1805 /// increment should be placed. In simple loops, this is the latch block's
1806 /// terminator. But in more complicated cases, this is a position which will
1807 /// dominate all the in-loop post-increment users.
1808 Instruction *IVIncInsertPos;
1809
1810 /// Interesting factors between use strides.
1811 ///
1812 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1813 /// default, a SmallDenseSet, because we need to use the full range of
1814 /// int64_ts, and there's currently no good way of doing that with
1815 /// SmallDenseSet.
1816 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1817
1818 /// Interesting use types, to facilitate truncation reuse.
1819 SmallSetVector<Type *, 4> Types;
1820
1821 /// The list of interesting uses.
1822 SmallVector<LSRUse, 16> Uses;
1823
1824 /// Track which uses use which register candidates.
1825 RegUseTracker RegUses;
1826
1827 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1828 // have more than a few IV increment chains in a loop. Missing a Chain falls
1829 // back to normal LSR behavior for those uses.
1830 static const unsigned MaxChains = 8;
1831
1832 /// IV users can form a chain of IV increments.
1833 SmallVector<IVChain, MaxChains> IVChainVec;
1834
1835 /// IV users that belong to profitable IVChains.
1836 SmallPtrSet<Use*, MaxChains> IVIncSet;
1837
1838 void OptimizeShadowIV();
1839 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1840 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1841 void OptimizeLoopTermCond();
1842
1843 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1844 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1845 void FinalizeChain(IVChain &Chain);
1846 void CollectChains();
1847 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1848 SmallVectorImpl<WeakVH> &DeadInsts);
1849
1850 void CollectInterestingTypesAndFactors();
1851 void CollectFixupsAndInitialFormulae();
1852
1853 // Support for sharing of LSRUses between LSRFixups.
1854 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1855 UseMapTy UseMap;
1856
1857 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1858 LSRUse::KindType Kind, MemAccessTy AccessTy);
1859
1860 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1861 MemAccessTy AccessTy);
1862
1863 void DeleteUse(LSRUse &LU, size_t LUIdx);
1864
1865 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1866
1867 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1868 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1869 void CountRegisters(const Formula &F, size_t LUIdx);
1870 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1871
1872 void CollectLoopInvariantFixupsAndFormulae();
1873
1874 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1875 unsigned Depth = 0);
1876
1877 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1878 const Formula &Base, unsigned Depth,
1879 size_t Idx, bool IsScaledReg = false);
1880 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1881 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1882 const Formula &Base, size_t Idx,
1883 bool IsScaledReg = false);
1884 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1885 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1886 const Formula &Base,
1887 const SmallVectorImpl<int64_t> &Worklist,
1888 size_t Idx, bool IsScaledReg = false);
1889 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1890 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1891 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1892 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1893 void GenerateCrossUseConstantOffsets();
1894 void GenerateAllReuseFormulae();
1895
1896 void FilterOutUndesirableDedicatedRegisters();
1897
1898 size_t EstimateSearchSpaceComplexity() const;
1899 void NarrowSearchSpaceByDetectingSupersets();
1900 void NarrowSearchSpaceByCollapsingUnrolledCode();
1901 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1902 void NarrowSearchSpaceByDeletingCostlyFormulas();
1903 void NarrowSearchSpaceByPickingWinnerRegs();
1904 void NarrowSearchSpaceUsingHeuristics();
1905
1906 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1907 Cost &SolutionCost,
1908 SmallVectorImpl<const Formula *> &Workspace,
1909 const Cost &CurCost,
1910 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1911 DenseSet<const SCEV *> &VisitedRegs) const;
1912 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1913
1914 BasicBlock::iterator
1915 HoistInsertPosition(BasicBlock::iterator IP,
1916 const SmallVectorImpl<Instruction *> &Inputs) const;
1917 BasicBlock::iterator
1918 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1919 const LSRFixup &LF,
1920 const LSRUse &LU,
1921 SCEVExpander &Rewriter) const;
1922
1923 Value *Expand(const LSRUse &LU, const LSRFixup &LF,
1924 const Formula &F,
1925 BasicBlock::iterator IP,
1926 SCEVExpander &Rewriter,
1927 SmallVectorImpl<WeakVH> &DeadInsts) const;
1928 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
1929 const Formula &F,
1930 SCEVExpander &Rewriter,
1931 SmallVectorImpl<WeakVH> &DeadInsts) const;
1932 void Rewrite(const LSRUse &LU, const LSRFixup &LF,
1933 const Formula &F,
1934 SCEVExpander &Rewriter,
1935 SmallVectorImpl<WeakVH> &DeadInsts) const;
1936 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1937
1938public:
1939 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1940 LoopInfo &LI, const TargetTransformInfo &TTI);
1941
1942 bool getChanged() const { return Changed; }
1943
1944 void print_factors_and_types(raw_ostream &OS) const;
1945 void print_fixups(raw_ostream &OS) const;
1946 void print_uses(raw_ostream &OS) const;
1947 void print(raw_ostream &OS) const;
1948 void dump() const;
1949};
1950
1951} // end anonymous namespace
1952
1953/// If IV is used in a int-to-float cast inside the loop then try to eliminate
1954/// the cast operation.
1955void LSRInstance::OptimizeShadowIV() {
1956 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1957 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1958 return;
1959
1960 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1961 UI != E; /* empty */) {
1962 IVUsers::const_iterator CandidateUI = UI;
1963 ++UI;
1964 Instruction *ShadowUse = CandidateUI->getUser();
1965 Type *DestTy = nullptr;
1966 bool IsSigned = false;
1967
1968 /* If shadow use is a int->float cast then insert a second IV
1969 to eliminate this cast.
1970
1971 for (unsigned i = 0; i < n; ++i)
1972 foo((double)i);
1973
1974 is transformed into
1975
1976 double d = 0.0;
1977 for (unsigned i = 0; i < n; ++i, ++d)
1978 foo(d);
1979 */
1980 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1981 IsSigned = false;
1982 DestTy = UCast->getDestTy();
1983 }
1984 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1985 IsSigned = true;
1986 DestTy = SCast->getDestTy();
1987 }
1988 if (!DestTy) continue;
1989
1990 // If target does not support DestTy natively then do not apply
1991 // this transformation.
1992 if (!TTI.isTypeLegal(DestTy)) continue;
1993
1994 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1995 if (!PH) continue;
1996 if (PH->getNumIncomingValues() != 2) continue;
1997
1998 Type *SrcTy = PH->getType();
1999 int Mantissa = DestTy->getFPMantissaWidth();
2000 if (Mantissa == -1) continue;
2001 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2002 continue;
2003
2004 unsigned Entry, Latch;
2005 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2006 Entry = 0;
2007 Latch = 1;
2008 } else {
2009 Entry = 1;
2010 Latch = 0;
2011 }
2012
2013 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2014 if (!Init) continue;
2015 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2016 (double)Init->getSExtValue() :
2017 (double)Init->getZExtValue());
2018
2019 BinaryOperator *Incr =
2020 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2021 if (!Incr) continue;
2022 if (Incr->getOpcode() != Instruction::Add
2023 && Incr->getOpcode() != Instruction::Sub)
2024 continue;
2025
2026 /* Initialize new IV, double d = 0.0 in above example. */
2027 ConstantInt *C = nullptr;
2028 if (Incr->getOperand(0) == PH)
2029 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2030 else if (Incr->getOperand(1) == PH)
2031 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2032 else
2033 continue;
2034
2035 if (!C) continue;
2036
2037 // Ignore negative constants, as the code below doesn't handle them
2038 // correctly. TODO: Remove this restriction.
2039 if (!C->getValue().isStrictlyPositive()) continue;
2040
2041 /* Add new PHINode. */
2042 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2043
2044 /* create new increment. '++d' in above example. */
2045 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2046 BinaryOperator *NewIncr =
2047 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2048 Instruction::FAdd : Instruction::FSub,
2049 NewPH, CFP, "IV.S.next.", Incr);
2050
2051 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2052 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2053
2054 /* Remove cast operation */
2055 ShadowUse->replaceAllUsesWith(NewPH);
2056 ShadowUse->eraseFromParent();
2057 Changed = true;
2058 break;
2059 }
2060}
2061
2062/// If Cond has an operand that is an expression of an IV, set the IV user and
2063/// stride information and return true, otherwise return false.
2064bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2065 for (IVStrideUse &U : IU)
2066 if (U.getUser() == Cond) {
2067 // NOTE: we could handle setcc instructions with multiple uses here, but
2068 // InstCombine does it as well for simple uses, it's not clear that it
2069 // occurs enough in real life to handle.
2070 CondUse = &U;
2071 return true;
2072 }
2073 return false;
2074}
2075
2076/// Rewrite the loop's terminating condition if it uses a max computation.
2077///
2078/// This is a narrow solution to a specific, but acute, problem. For loops
2079/// like this:
2080///
2081/// i = 0;
2082/// do {
2083/// p[i] = 0.0;
2084/// } while (++i < n);
2085///
2086/// the trip count isn't just 'n', because 'n' might not be positive. And
2087/// unfortunately this can come up even for loops where the user didn't use
2088/// a C do-while loop. For example, seemingly well-behaved top-test loops
2089/// will commonly be lowered like this:
2090//
2091/// if (n > 0) {
2092/// i = 0;
2093/// do {
2094/// p[i] = 0.0;
2095/// } while (++i < n);
2096/// }
2097///
2098/// and then it's possible for subsequent optimization to obscure the if
2099/// test in such a way that indvars can't find it.
2100///
2101/// When indvars can't find the if test in loops like this, it creates a
2102/// max expression, which allows it to give the loop a canonical
2103/// induction variable:
2104///
2105/// i = 0;
2106/// max = n < 1 ? 1 : n;
2107/// do {
2108/// p[i] = 0.0;
2109/// } while (++i != max);
2110///
2111/// Canonical induction variables are necessary because the loop passes
2112/// are designed around them. The most obvious example of this is the
2113/// LoopInfo analysis, which doesn't remember trip count values. It
2114/// expects to be able to rediscover the trip count each time it is
2115/// needed, and it does this using a simple analysis that only succeeds if
2116/// the loop has a canonical induction variable.
2117///
2118/// However, when it comes time to generate code, the maximum operation
2119/// can be quite costly, especially if it's inside of an outer loop.
2120///
2121/// This function solves this problem by detecting this type of loop and
2122/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2123/// the instructions for the maximum computation.
2124///
2125ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2126 // Check that the loop matches the pattern we're looking for.
2127 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2128 Cond->getPredicate() != CmpInst::ICMP_NE)
2129 return Cond;
2130
2131 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2132 if (!Sel || !Sel->hasOneUse()) return Cond;
2133
2134 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2135 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2136 return Cond;
2137 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2138
2139 // Add one to the backedge-taken count to get the trip count.
2140 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2141 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2142
2143 // Check for a max calculation that matches the pattern. There's no check
2144 // for ICMP_ULE here because the comparison would be with zero, which
2145 // isn't interesting.
2146 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2147 const SCEVNAryExpr *Max = nullptr;
2148 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2149 Pred = ICmpInst::ICMP_SLE;
2150 Max = S;
2151 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2152 Pred = ICmpInst::ICMP_SLT;
2153 Max = S;
2154 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2155 Pred = ICmpInst::ICMP_ULT;
2156 Max = U;
2157 } else {
2158 // No match; bail.
2159 return Cond;
2160 }
2161
2162 // To handle a max with more than two operands, this optimization would
2163 // require additional checking and setup.
2164 if (Max->getNumOperands() != 2)
2165 return Cond;
2166
2167 const SCEV *MaxLHS = Max->getOperand(0);
2168 const SCEV *MaxRHS = Max->getOperand(1);
2169
2170 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2171 // for a comparison with 1. For <= and >=, a comparison with zero.
2172 if (!MaxLHS ||
2173 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2174 return Cond;
2175
2176 // Check the relevant induction variable for conformance to
2177 // the pattern.
2178 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2179 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2180 if (!AR || !AR->isAffine() ||
2181 AR->getStart() != One ||
2182 AR->getStepRecurrence(SE) != One)
2183 return Cond;
2184
2185 assert(AR->getLoop() == L &&((AR->getLoop() == L && "Loop condition operand is an addrec in a different loop!"
) ? static_cast<void> (0) : __assert_fail ("AR->getLoop() == L && \"Loop condition operand is an addrec in a different loop!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2186, __PRETTY_FUNCTION__))
2186 "Loop condition operand is an addrec in a different loop!")((AR->getLoop() == L && "Loop condition operand is an addrec in a different loop!"
) ? static_cast<void> (0) : __assert_fail ("AR->getLoop() == L && \"Loop condition operand is an addrec in a different loop!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2186, __PRETTY_FUNCTION__))
;
2187
2188 // Check the right operand of the select, and remember it, as it will
2189 // be used in the new comparison instruction.
2190 Value *NewRHS = nullptr;
2191 if (ICmpInst::isTrueWhenEqual(Pred)) {
2192 // Look for n+1, and grab n.
2193 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2194 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2195 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2196 NewRHS = BO->getOperand(0);
2197 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2198 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2199 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2200 NewRHS = BO->getOperand(0);
2201 if (!NewRHS)
2202 return Cond;
2203 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2204 NewRHS = Sel->getOperand(1);
2205 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2206 NewRHS = Sel->getOperand(2);
2207 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2208 NewRHS = SU->getValue();
2209 else
2210 // Max doesn't match expected pattern.
2211 return Cond;
2212
2213 // Determine the new comparison opcode. It may be signed or unsigned,
2214 // and the original comparison may be either equality or inequality.
2215 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2216 Pred = CmpInst::getInversePredicate(Pred);
2217
2218 // Ok, everything looks ok to change the condition into an SLT or SGE and
2219 // delete the max calculation.
2220 ICmpInst *NewCond =
2221 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2222
2223 // Delete the max calculation instructions.
2224 Cond->replaceAllUsesWith(NewCond);
2225 CondUse->setUser(NewCond);
2226 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2227 Cond->eraseFromParent();
2228 Sel->eraseFromParent();
2229 if (Cmp->use_empty())
2230 Cmp->eraseFromParent();
2231 return NewCond;
2232}
2233
2234/// Change loop terminating condition to use the postinc iv when possible.
2235void
2236LSRInstance::OptimizeLoopTermCond() {
2237 SmallPtrSet<Instruction *, 4> PostIncs;
2238
2239 // We need a different set of heuristics for rotated and non-rotated loops.
2240 // If a loop is rotated then the latch is also the backedge, so inserting
2241 // post-inc expressions just before the latch is ideal. To reduce live ranges
2242 // it also makes sense to rewrite terminating conditions to use post-inc
2243 // expressions.
2244 //
2245 // If the loop is not rotated then the latch is not a backedge; the latch
2246 // check is done in the loop head. Adding post-inc expressions before the
2247 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2248 // in the loop body. In this case we do *not* want to use post-inc expressions
2249 // in the latch check, and we want to insert post-inc expressions before
2250 // the backedge.
2251 BasicBlock *LatchBlock = L->getLoopLatch();
2252 SmallVector<BasicBlock*, 8> ExitingBlocks;
2253 L->getExitingBlocks(ExitingBlocks);
2254 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2255 return LatchBlock != BB;
2256 })) {
2257 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2258 IVIncInsertPos = LatchBlock->getTerminator();
2259 return;
2260 }
2261
2262 // Otherwise treat this as a rotated loop.
2263 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2264
2265 // Get the terminating condition for the loop if possible. If we
2266 // can, we want to change it to use a post-incremented version of its
2267 // induction variable, to allow coalescing the live ranges for the IV into
2268 // one register value.
2269
2270 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2271 if (!TermBr)
2272 continue;
2273 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2274 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2275 continue;
2276
2277 // Search IVUsesByStride to find Cond's IVUse if there is one.
2278 IVStrideUse *CondUse = nullptr;
2279 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2280 if (!FindIVUserForCond(Cond, CondUse))
2281 continue;
2282
2283 // If the trip count is computed in terms of a max (due to ScalarEvolution
2284 // being unable to find a sufficient guard, for example), change the loop
2285 // comparison to use SLT or ULT instead of NE.
2286 // One consequence of doing this now is that it disrupts the count-down
2287 // optimization. That's not always a bad thing though, because in such
2288 // cases it may still be worthwhile to avoid a max.
2289 Cond = OptimizeMax(Cond, CondUse);
2290
2291 // If this exiting block dominates the latch block, it may also use
2292 // the post-inc value if it won't be shared with other uses.
2293 // Check for dominance.
2294 if (!DT.dominates(ExitingBlock, LatchBlock))
2295 continue;
2296
2297 // Conservatively avoid trying to use the post-inc value in non-latch
2298 // exits if there may be pre-inc users in intervening blocks.
2299 if (LatchBlock != ExitingBlock)
2300 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2301 // Test if the use is reachable from the exiting block. This dominator
2302 // query is a conservative approximation of reachability.
2303 if (&*UI != CondUse &&
2304 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2305 // Conservatively assume there may be reuse if the quotient of their
2306 // strides could be a legal scale.
2307 const SCEV *A = IU.getStride(*CondUse, L);
2308 const SCEV *B = IU.getStride(*UI, L);
2309 if (!A || !B) continue;
2310 if (SE.getTypeSizeInBits(A->getType()) !=
2311 SE.getTypeSizeInBits(B->getType())) {
2312 if (SE.getTypeSizeInBits(A->getType()) >
2313 SE.getTypeSizeInBits(B->getType()))
2314 B = SE.getSignExtendExpr(B, A->getType());
2315 else
2316 A = SE.getSignExtendExpr(A, B->getType());
2317 }
2318 if (const SCEVConstant *D =
2319 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2320 const ConstantInt *C = D->getValue();
2321 // Stride of one or negative one can have reuse with non-addresses.
2322 if (C->isOne() || C->isAllOnesValue())
2323 goto decline_post_inc;
2324 // Avoid weird situations.
2325 if (C->getValue().getMinSignedBits() >= 64 ||
2326 C->getValue().isMinSignedValue())
2327 goto decline_post_inc;
2328 // Check for possible scaled-address reuse.
2329 MemAccessTy AccessTy = getAccessType(UI->getUser());
2330 int64_t Scale = C->getSExtValue();
2331 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2332 /*BaseOffset=*/0,
2333 /*HasBaseReg=*/false, Scale,
2334 AccessTy.AddrSpace))
2335 goto decline_post_inc;
2336 Scale = -Scale;
2337 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2338 /*BaseOffset=*/0,
2339 /*HasBaseReg=*/false, Scale,
2340 AccessTy.AddrSpace))
2341 goto decline_post_inc;
2342 }
2343 }
2344
2345 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
2346 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2347
2348 // It's possible for the setcc instruction to be anywhere in the loop, and
2349 // possible for it to have multiple users. If it is not immediately before
2350 // the exiting block branch, move it.
2351 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2352 if (Cond->hasOneUse()) {
2353 Cond->moveBefore(TermBr);
2354 } else {
2355 // Clone the terminating condition and insert into the loopend.
2356 ICmpInst *OldCond = Cond;
2357 Cond = cast<ICmpInst>(Cond->clone());
2358 Cond->setName(L->getHeader()->getName() + ".termcond");
2359 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2360
2361 // Clone the IVUse, as the old use still exists!
2362 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2363 TermBr->replaceUsesOfWith(OldCond, Cond);
2364 }
2365 }
2366
2367 // If we get to here, we know that we can transform the setcc instruction to
2368 // use the post-incremented version of the IV, allowing us to coalesce the
2369 // live ranges for the IV correctly.
2370 CondUse->transformToPostInc(L);
2371 Changed = true;
2372
2373 PostIncs.insert(Cond);
2374 decline_post_inc:;
2375 }
2376
2377 // Determine an insertion point for the loop induction variable increment. It
2378 // must dominate all the post-inc comparisons we just set up, and it must
2379 // dominate the loop latch edge.
2380 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2381 for (Instruction *Inst : PostIncs) {
2382 BasicBlock *BB =
2383 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2384 Inst->getParent());
2385 if (BB == Inst->getParent())
2386 IVIncInsertPos = Inst;
2387 else if (BB != IVIncInsertPos->getParent())
2388 IVIncInsertPos = BB->getTerminator();
2389 }
2390}
2391
2392/// Determine if the given use can accommodate a fixup at the given offset and
2393/// other details. If so, update the use and return true.
2394bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2395 bool HasBaseReg, LSRUse::KindType Kind,
2396 MemAccessTy AccessTy) {
2397 int64_t NewMinOffset = LU.MinOffset;
2398 int64_t NewMaxOffset = LU.MaxOffset;
2399 MemAccessTy NewAccessTy = AccessTy;
2400
2401 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2402 // something conservative, however this can pessimize in the case that one of
2403 // the uses will have all its uses outside the loop, for example.
2404 if (LU.Kind != Kind)
2405 return false;
2406
2407 // Check for a mismatched access type, and fall back conservatively as needed.
2408 // TODO: Be less conservative when the type is similar and can use the same
2409 // addressing modes.
2410 if (Kind == LSRUse::Address) {
2411 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2412 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2413 AccessTy.AddrSpace);
2414 }
2415 }
2416
2417 // Conservatively assume HasBaseReg is true for now.
2418 if (NewOffset < LU.MinOffset) {
2419 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2420 LU.MaxOffset - NewOffset, HasBaseReg))
2421 return false;
2422 NewMinOffset = NewOffset;
2423 } else if (NewOffset > LU.MaxOffset) {
2424 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2425 NewOffset - LU.MinOffset, HasBaseReg))
2426 return false;
2427 NewMaxOffset = NewOffset;
2428 }
2429
2430 // Update the use.
2431 LU.MinOffset = NewMinOffset;
2432 LU.MaxOffset = NewMaxOffset;
2433 LU.AccessTy = NewAccessTy;
2434 return true;
2435}
2436
2437/// Return an LSRUse index and an offset value for a fixup which needs the given
2438/// expression, with the given kind and optional access type. Either reuse an
2439/// existing use or create a new one, as needed.
2440std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2441 LSRUse::KindType Kind,
2442 MemAccessTy AccessTy) {
2443 const SCEV *Copy = Expr;
2444 int64_t Offset = ExtractImmediate(Expr, SE);
2445
2446 // Basic uses can't accept any offset, for example.
2447 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2448 Offset, /*HasBaseReg=*/ true)) {
2449 Expr = Copy;
2450 Offset = 0;
2451 }
2452
2453 std::pair<UseMapTy::iterator, bool> P =
2454 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2455 if (!P.second) {
2456 // A use already existed with this base.
2457 size_t LUIdx = P.first->second;
2458 LSRUse &LU = Uses[LUIdx];
2459 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2460 // Reuse this use.
2461 return std::make_pair(LUIdx, Offset);
2462 }
2463
2464 // Create a new use.
2465 size_t LUIdx = Uses.size();
2466 P.first->second = LUIdx;
2467 Uses.push_back(LSRUse(Kind, AccessTy));
2468 LSRUse &LU = Uses[LUIdx];
2469
2470 LU.MinOffset = Offset;
2471 LU.MaxOffset = Offset;
2472 return std::make_pair(LUIdx, Offset);
2473}
2474
2475/// Delete the given use from the Uses list.
2476void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2477 if (&LU != &Uses.back())
2478 std::swap(LU, Uses.back());
2479 Uses.pop_back();
2480
2481 // Update RegUses.
2482 RegUses.swapAndDropUse(LUIdx, Uses.size());
2483}
2484
2485/// Look for a use distinct from OrigLU which is has a formula that has the same
2486/// registers as the given formula.
2487LSRUse *
2488LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2489 const LSRUse &OrigLU) {
2490 // Search all uses for the formula. This could be more clever.
2491 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2492 LSRUse &LU = Uses[LUIdx];
2493 // Check whether this use is close enough to OrigLU, to see whether it's
2494 // worthwhile looking through its formulae.
2495 // Ignore ICmpZero uses because they may contain formulae generated by
2496 // GenerateICmpZeroScales, in which case adding fixup offsets may
2497 // be invalid.
2498 if (&LU != &OrigLU &&
2499 LU.Kind != LSRUse::ICmpZero &&
2500 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2501 LU.WidestFixupType == OrigLU.WidestFixupType &&
2502 LU.HasFormulaWithSameRegs(OrigF)) {
2503 // Scan through this use's formulae.
2504 for (const Formula &F : LU.Formulae) {
2505 // Check to see if this formula has the same registers and symbols
2506 // as OrigF.
2507 if (F.BaseRegs == OrigF.BaseRegs &&
2508 F.ScaledReg == OrigF.ScaledReg &&
2509 F.BaseGV == OrigF.BaseGV &&
2510 F.Scale == OrigF.Scale &&
2511 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2512 if (F.BaseOffset == 0)
2513 return &LU;
2514 // This is the formula where all the registers and symbols matched;
2515 // there aren't going to be any others. Since we declined it, we
2516 // can skip the rest of the formulae and proceed to the next LSRUse.
2517 break;
2518 }
2519 }
2520 }
2521 }
2522
2523 // Nothing looked good.
2524 return nullptr;
2525}
2526
2527void LSRInstance::CollectInterestingTypesAndFactors() {
2528 SmallSetVector<const SCEV *, 4> Strides;
2529
2530 // Collect interesting types and strides.
2531 SmallVector<const SCEV *, 4> Worklist;
2532 for (const IVStrideUse &U : IU) {
2533 const SCEV *Expr = IU.getExpr(U);
2534
2535 // Collect interesting types.
2536 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2537
2538 // Add strides for mentioned loops.
2539 Worklist.push_back(Expr);
2540 do {
2541 const SCEV *S = Worklist.pop_back_val();
2542 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2543 if (AR->getLoop() == L)
2544 Strides.insert(AR->getStepRecurrence(SE));
2545 Worklist.push_back(AR->getStart());
2546 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2547 Worklist.append(Add->op_begin(), Add->op_end());
2548 }
2549 } while (!Worklist.empty());
2550 }
2551
2552 // Compute interesting factors from the set of interesting strides.
2553 for (SmallSetVector<const SCEV *, 4>::const_iterator
2554 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2555 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2556 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2557 const SCEV *OldStride = *I;
2558 const SCEV *NewStride = *NewStrideIter;
2559
2560 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2561 SE.getTypeSizeInBits(NewStride->getType())) {
2562 if (SE.getTypeSizeInBits(OldStride->getType()) >
2563 SE.getTypeSizeInBits(NewStride->getType()))
2564 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2565 else
2566 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2567 }
2568 if (const SCEVConstant *Factor =
2569 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2570 SE, true))) {
2571 if (Factor->getAPInt().getMinSignedBits() <= 64)
2572 Factors.insert(Factor->getAPInt().getSExtValue());
2573 } else if (const SCEVConstant *Factor =
2574 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2575 NewStride,
2576 SE, true))) {
2577 if (Factor->getAPInt().getMinSignedBits() <= 64)
2578 Factors.insert(Factor->getAPInt().getSExtValue());
2579 }
2580 }
2581
2582 // If all uses use the same type, don't bother looking for truncation-based
2583 // reuse.
2584 if (Types.size() == 1)
2585 Types.clear();
2586
2587 DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2588}
2589
2590/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2591/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2592/// IVStrideUses, we could partially skip this.
2593static User::op_iterator
2594findIVOperand(User::op_iterator OI, User::op_iterator OE,
2595 Loop *L, ScalarEvolution &SE) {
2596 for(; OI != OE; ++OI) {
2597 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2598 if (!SE.isSCEVable(Oper->getType()))
2599 continue;
2600
2601 if (const SCEVAddRecExpr *AR =
2602 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2603 if (AR->getLoop() == L)
2604 break;
2605 }
2606 }
2607 }
2608 return OI;
2609}
2610
2611/// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2612/// a convenient helper.
2613static Value *getWideOperand(Value *Oper) {
2614 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2615 return Trunc->getOperand(0);
2616 return Oper;
2617}
2618
2619/// Return true if we allow an IV chain to include both types.
2620static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2621 Type *LType = LVal->getType();
2622 Type *RType = RVal->getType();
2623 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2624 // Different address spaces means (possibly)
2625 // different types of the pointer implementation,
2626 // e.g. i16 vs i32 so disallow that.
2627 (LType->getPointerAddressSpace() ==
2628 RType->getPointerAddressSpace()));
2629}
2630
2631/// Return an approximation of this SCEV expression's "base", or NULL for any
2632/// constant. Returning the expression itself is conservative. Returning a
2633/// deeper subexpression is more precise and valid as long as it isn't less
2634/// complex than another subexpression. For expressions involving multiple
2635/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2636/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2637/// IVInc==b-a.
2638///
2639/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2640/// SCEVUnknown, we simply return the rightmost SCEV operand.
2641static const SCEV *getExprBase(const SCEV *S) {
2642 switch (S->getSCEVType()) {
2643 default: // uncluding scUnknown.
2644 return S;
2645 case scConstant:
2646 return nullptr;
2647 case scTruncate:
2648 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2649 case scZeroExtend:
2650 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2651 case scSignExtend:
2652 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2653 case scAddExpr: {
2654 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2655 // there's nothing more complex.
2656 // FIXME: not sure if we want to recognize negation.
2657 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2658 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2659 E(Add->op_begin()); I != E; ++I) {
2660 const SCEV *SubExpr = *I;
2661 if (SubExpr->getSCEVType() == scAddExpr)
2662 return getExprBase(SubExpr);
2663
2664 if (SubExpr->getSCEVType() != scMulExpr)
2665 return SubExpr;
2666 }
2667 return S; // all operands are scaled, be conservative.
2668 }
2669 case scAddRecExpr:
2670 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2671 }
2672}
2673
2674/// Return true if the chain increment is profitable to expand into a loop
2675/// invariant value, which may require its own register. A profitable chain
2676/// increment will be an offset relative to the same base. We allow such offsets
2677/// to potentially be used as chain increment as long as it's not obviously
2678/// expensive to expand using real instructions.
2679bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2680 const SCEV *IncExpr,
2681 ScalarEvolution &SE) {
2682 // Aggressively form chains when -stress-ivchain.
2683 if (StressIVChain)
2684 return true;
2685
2686 // Do not replace a constant offset from IV head with a nonconstant IV
2687 // increment.
2688 if (!isa<SCEVConstant>(IncExpr)) {
2689 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2690 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2691 return false;
2692 }
2693
2694 SmallPtrSet<const SCEV*, 8> Processed;
2695 return !isHighCostExpansion(IncExpr, Processed, SE);
2696}
2697
2698/// Return true if the number of registers needed for the chain is estimated to
2699/// be less than the number required for the individual IV users. First prohibit
2700/// any IV users that keep the IV live across increments (the Users set should
2701/// be empty). Next count the number and type of increments in the chain.
2702///
2703/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2704/// effectively use postinc addressing modes. Only consider it profitable it the
2705/// increments can be computed in fewer registers when chained.
2706///
2707/// TODO: Consider IVInc free if it's already used in another chains.
2708static bool
2709isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2710 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2711 if (StressIVChain)
2712 return true;
2713
2714 if (!Chain.hasIncs())
2715 return false;
2716
2717 if (!Users.empty()) {
2718 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
: Users) { dbgs() << " " << *Inst << "\n"
; }; } } while (false)
2719 for (Instruction *Inst : Users) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
: Users) { dbgs() << " " << *Inst << "\n"
; }; } } while (false)
2720 dbgs() << " " << *Inst << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
: Users) { dbgs() << " " << *Inst << "\n"
; }; } } while (false)
2721 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
: Users) { dbgs() << " " << *Inst << "\n"
; }; } } while (false)
;
2722 return false;
2723 }
2724 assert(!Chain.Incs.empty() && "empty IV chains are not allowed")((!Chain.Incs.empty() && "empty IV chains are not allowed"
) ? static_cast<void> (0) : __assert_fail ("!Chain.Incs.empty() && \"empty IV chains are not allowed\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2724, __PRETTY_FUNCTION__))
;
2725
2726 // The chain itself may require a register, so intialize cost to 1.
2727 int cost = 1;
2728
2729 // A complete chain likely eliminates the need for keeping the original IV in
2730 // a register. LSR does not currently know how to form a complete chain unless
2731 // the header phi already exists.
2732 if (isa<PHINode>(Chain.tailUserInst())
2733 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2734 --cost;
2735 }
2736 const SCEV *LastIncExpr = nullptr;
2737 unsigned NumConstIncrements = 0;
2738 unsigned NumVarIncrements = 0;
2739 unsigned NumReusedIncrements = 0;
2740 for (const IVInc &Inc : Chain) {
2741 if (Inc.IncExpr->isZero())
2742 continue;
2743
2744 // Incrementing by zero or some constant is neutral. We assume constants can
2745 // be folded into an addressing mode or an add's immediate operand.
2746 if (isa<SCEVConstant>(Inc.IncExpr)) {
2747 ++NumConstIncrements;
2748 continue;
2749 }
2750
2751 if (Inc.IncExpr == LastIncExpr)
2752 ++NumReusedIncrements;
2753 else
2754 ++NumVarIncrements;
2755
2756 LastIncExpr = Inc.IncExpr;
2757 }
2758 // An IV chain with a single increment is handled by LSR's postinc
2759 // uses. However, a chain with multiple increments requires keeping the IV's
2760 // value live longer than it needs to be if chained.
2761 if (NumConstIncrements > 1)
2762 --cost;
2763
2764 // Materializing increment expressions in the preheader that didn't exist in
2765 // the original code may cost a register. For example, sign-extended array
2766 // indices can produce ridiculous increments like this:
2767 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2768 cost += NumVarIncrements;
2769
2770 // Reusing variable increments likely saves a register to hold the multiple of
2771 // the stride.
2772 cost -= NumReusedIncrements;
2773
2774 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << costdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
2775 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2776
2777 return cost < 0;
2778}
2779
2780/// Add this IV user to an existing chain or make it the head of a new chain.
2781void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2782 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2783 // When IVs are used as types of varying widths, they are generally converted
2784 // to a wider type with some uses remaining narrow under a (free) trunc.
2785 Value *const NextIV = getWideOperand(IVOper);
2786 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2787 const SCEV *const OperExprBase = getExprBase(OperExpr);
2788
2789 // Visit all existing chains. Check if its IVOper can be computed as a
2790 // profitable loop invariant increment from the last link in the Chain.
2791 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2792 const SCEV *LastIncExpr = nullptr;
2793 for (; ChainIdx < NChains; ++ChainIdx) {
2794 IVChain &Chain = IVChainVec[ChainIdx];
2795
2796 // Prune the solution space aggressively by checking that both IV operands
2797 // are expressions that operate on the same unscaled SCEVUnknown. This
2798 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2799 // first avoids creating extra SCEV expressions.
2800 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2801 continue;
2802
2803 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2804 if (!isCompatibleIVType(PrevIV, NextIV))
2805 continue;
2806
2807 // A phi node terminates a chain.
2808 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2809 continue;
2810
2811 // The increment must be loop-invariant so it can be kept in a register.
2812 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2813 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2814 if (!SE.isLoopInvariant(IncExpr, L))
2815 continue;
2816
2817 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2818 LastIncExpr = IncExpr;
2819 break;
2820 }
2821 }
2822 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2823 // bother for phi nodes, because they must be last in the chain.
2824 if (ChainIdx == NChains) {
2825 if (isa<PHINode>(UserInst))
2826 return;
2827 if (NChains >= MaxChains && !StressIVChain) {
2828 DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2829 return;
2830 }
2831 LastIncExpr = OperExpr;
2832 // IVUsers may have skipped over sign/zero extensions. We don't currently
2833 // attempt to form chains involving extensions unless they can be hoisted
2834 // into this loop's AddRec.
2835 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2836 return;
2837 ++NChains;
2838 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2839 OperExprBase));
2840 ChainUsersVec.resize(NChains);
2841 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
2842 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
2843 } else {
2844 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
2845 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
2846 // Add this IV user to the end of the chain.
2847 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2848 }
2849 IVChain &Chain = IVChainVec[ChainIdx];
2850
2851 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2852 // This chain's NearUsers become FarUsers.
2853 if (!LastIncExpr->isZero()) {
2854 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2855 NearUsers.end());
2856 NearUsers.clear();
2857 }
2858
2859 // All other uses of IVOperand become near uses of the chain.
2860 // We currently ignore intermediate values within SCEV expressions, assuming
2861 // they will eventually be used be the current chain, or can be computed
2862 // from one of the chain increments. To be more precise we could
2863 // transitively follow its user and only add leaf IV users to the set.
2864 for (User *U : IVOper->users()) {
2865 Instruction *OtherUse = dyn_cast<Instruction>(U);
2866 if (!OtherUse)
2867 continue;
2868 // Uses in the chain will no longer be uses if the chain is formed.
2869 // Include the head of the chain in this iteration (not Chain.begin()).
2870 IVChain::const_iterator IncIter = Chain.Incs.begin();
2871 IVChain::const_iterator IncEnd = Chain.Incs.end();
2872 for( ; IncIter != IncEnd; ++IncIter) {
2873 if (IncIter->UserInst == OtherUse)
2874 break;
2875 }
2876 if (IncIter != IncEnd)
2877 continue;
2878
2879 if (SE.isSCEVable(OtherUse->getType())
2880 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2881 && IU.isIVUserOrOperand(OtherUse)) {
2882 continue;
2883 }
2884 NearUsers.insert(OtherUse);
2885 }
2886
2887 // Since this user is part of the chain, it's no longer considered a use
2888 // of the chain.
2889 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2890}
2891
2892/// Populate the vector of Chains.
2893///
2894/// This decreases ILP at the architecture level. Targets with ample registers,
2895/// multiple memory ports, and no register renaming probably don't want
2896/// this. However, such targets should probably disable LSR altogether.
2897///
2898/// The job of LSR is to make a reasonable choice of induction variables across
2899/// the loop. Subsequent passes can easily "unchain" computation exposing more
2900/// ILP *within the loop* if the target wants it.
2901///
2902/// Finding the best IV chain is potentially a scheduling problem. Since LSR
2903/// will not reorder memory operations, it will recognize this as a chain, but
2904/// will generate redundant IV increments. Ideally this would be corrected later
2905/// by a smart scheduler:
2906/// = A[i]
2907/// = A[i+x]
2908/// A[i] =
2909/// A[i+x] =
2910///
2911/// TODO: Walk the entire domtree within this loop, not just the path to the
2912/// loop latch. This will discover chains on side paths, but requires
2913/// maintaining multiple copies of the Chains state.
2914void LSRInstance::CollectChains() {
2915 DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
2916 SmallVector<ChainUsers, 8> ChainUsersVec;
2917
2918 SmallVector<BasicBlock *,8> LatchPath;
2919 BasicBlock *LoopHeader = L->getHeader();
2920 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2921 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2922 LatchPath.push_back(Rung->getBlock());
2923 }
2924 LatchPath.push_back(LoopHeader);
2925
2926 // Walk the instruction stream from the loop header to the loop latch.
2927 for (BasicBlock *BB : reverse(LatchPath)) {
2928 for (Instruction &I : *BB) {
2929 // Skip instructions that weren't seen by IVUsers analysis.
2930 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2931 continue;
2932
2933 // Ignore users that are part of a SCEV expression. This way we only
2934 // consider leaf IV Users. This effectively rediscovers a portion of
2935 // IVUsers analysis but in program order this time.
2936 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2937 continue;
2938
2939 // Remove this instruction from any NearUsers set it may be in.
2940 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2941 ChainIdx < NChains; ++ChainIdx) {
2942 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
2943 }
2944 // Search for operands that can be chained.
2945 SmallPtrSet<Instruction*, 4> UniqueOperands;
2946 User::op_iterator IVOpEnd = I.op_end();
2947 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
2948 while (IVOpIter != IVOpEnd) {
2949 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2950 if (UniqueOperands.insert(IVOpInst).second)
2951 ChainInstruction(&I, IVOpInst, ChainUsersVec);
2952 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2953 }
2954 } // Continue walking down the instructions.
2955 } // Continue walking down the domtree.
2956 // Visit phi backedges to determine if the chain can generate the IV postinc.
2957 for (BasicBlock::iterator I = L->getHeader()->begin();
2958 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2959 if (!SE.isSCEVable(PN->getType()))
2960 continue;
2961
2962 Instruction *IncV =
2963 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2964 if (IncV)
2965 ChainInstruction(PN, IncV, ChainUsersVec);
2966 }
2967 // Remove any unprofitable chains.
2968 unsigned ChainIdx = 0;
2969 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2970 UsersIdx < NChains; ++UsersIdx) {
2971 if (!isProfitableChain(IVChainVec[UsersIdx],
2972 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2973 continue;
2974 // Preserve the chain at UsesIdx.
2975 if (ChainIdx != UsersIdx)
2976 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2977 FinalizeChain(IVChainVec[ChainIdx]);
2978 ++ChainIdx;
2979 }
2980 IVChainVec.resize(ChainIdx);
2981}
2982
2983void LSRInstance::FinalizeChain(IVChain &Chain) {
2984 assert(!Chain.Incs.empty() && "empty IV chains are not allowed")((!Chain.Incs.empty() && "empty IV chains are not allowed"
) ? static_cast<void> (0) : __assert_fail ("!Chain.Incs.empty() && \"empty IV chains are not allowed\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2984, __PRETTY_FUNCTION__))
;
2985 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Final Chain: " << *
Chain.Incs[0].UserInst << "\n"; } } while (false)
;
2986
2987 for (const IVInc &Inc : Chain) {
2988 DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
2989 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
2990 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand")((UseI != Inc.UserInst->op_end() && "cannot find IV operand"
) ? static_cast<void> (0) : __assert_fail ("UseI != Inc.UserInst->op_end() && \"cannot find IV operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2990, __PRETTY_FUNCTION__))
;
2991 IVIncSet.insert(UseI);
2992 }
2993}
2994
2995/// Return true if the IVInc can be folded into an addressing mode.
2996static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2997 Value *Operand, const TargetTransformInfo &TTI) {
2998 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2999 if (!IncConst || !isAddressUse(UserInst, Operand))
3000 return false;
3001
3002 if (IncConst->getAPInt().getMinSignedBits() > 64)
3003 return false;
3004
3005 MemAccessTy AccessTy = getAccessType(UserInst);
3006 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3007 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3008 IncOffset, /*HaseBaseReg=*/false))
3009 return false;
3010
3011 return true;
3012}
3013
3014/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3015/// user's operand from the previous IV user's operand.
3016void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3017 SmallVectorImpl<WeakVH> &DeadInsts) {
3018 // Find the new IVOperand for the head of the chain. It may have been replaced
3019 // by LSR.
3020 const IVInc &Head = Chain.Incs[0];
3021 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3022 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3023 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3024 IVOpEnd, L, SE);
3025 Value *IVSrc = nullptr;
1
'IVSrc' initialized to a null pointer value
3026 while (IVOpIter != IVOpEnd) {
2
Loop condition is false. Execution continues on line 3043
3027 IVSrc = getWideOperand(*IVOpIter);
3028
3029 // If this operand computes the expression that the chain needs, we may use
3030 // it. (Check this after setting IVSrc which is used below.)
3031 //
3032 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3033 // narrow for the chain, so we can no longer use it. We do allow using a
3034 // wider phi, assuming the LSR checked for free truncation. In that case we
3035 // should already have a truncate on this operand such that
3036 // getSCEV(IVSrc) == IncExpr.
3037 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3038 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3039 break;
3040 }
3041 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3042 }
3043 if (IVOpIter == IVOpEnd) {
3
Taking false branch
3044 // Gracefully give up on this chain.
3045 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Concealed chain head: " <<
*Head.UserInst << "\n"; } } while (false)
;
3046 return;
3047 }
3048
3049 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3050 Type *IVTy = IVSrc->getType();
4
Called C++ object pointer is null
3051 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3052 const SCEV *LeftOverExpr = nullptr;
3053 for (const IVInc &Inc : Chain) {
3054 Instruction *InsertPt = Inc.UserInst;
3055 if (isa<PHINode>(InsertPt))
3056 InsertPt = L->getLoopLatch()->getTerminator();
3057
3058 // IVOper will replace the current IV User's operand. IVSrc is the IV
3059 // value currently held in a register.
3060 Value *IVOper = IVSrc;
3061 if (!Inc.IncExpr->isZero()) {
3062 // IncExpr was the result of subtraction of two narrow values, so must
3063 // be signed.
3064 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3065 LeftOverExpr = LeftOverExpr ?
3066 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3067 }
3068 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3069 // Expand the IV increment.
3070 Rewriter.clearPostInc();
3071 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3072 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3073 SE.getUnknown(IncV));
3074 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3075
3076 // If an IV increment can't be folded, use it as the next IV value.
3077 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3078 assert(IVTy == IVOper->getType() && "inconsistent IV increment type")((IVTy == IVOper->getType() && "inconsistent IV increment type"
) ? static_cast<void> (0) : __assert_fail ("IVTy == IVOper->getType() && \"inconsistent IV increment type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3078, __PRETTY_FUNCTION__))
;
3079 IVSrc = IVOper;
3080 LeftOverExpr = nullptr;
3081 }
3082 }
3083 Type *OperTy = Inc.IVOperand->getType();
3084 if (IVTy != OperTy) {
3085 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&((SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy
) && "cannot extend a chained IV") ? static_cast<void
> (0) : __assert_fail ("SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && \"cannot extend a chained IV\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3086, __PRETTY_FUNCTION__))
3086 "cannot extend a chained IV")((SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy
) && "cannot extend a chained IV") ? static_cast<void
> (0) : __assert_fail ("SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && \"cannot extend a chained IV\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3086, __PRETTY_FUNCTION__))
;
3087 IRBuilder<> Builder(InsertPt);
3088 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3089 }
3090 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3091 DeadInsts.emplace_back(Inc.IVOperand);
3092 }
3093 // If LSR created a new, wider phi, we may also replace its postinc. We only
3094 // do this if we also found a wide value for the head of the chain.
3095 if (isa<PHINode>(Chain.tailUserInst())) {
3096 for (BasicBlock::iterator I = L->getHeader()->begin();
3097 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
3098 if (!isCompatibleIVType(Phi, IVSrc))
3099 continue;
3100 Instruction *PostIncV = dyn_cast<Instruction>(
3101 Phi->getIncomingValueForBlock(L->getLoopLatch()));
3102 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3103 continue;
3104 Value *IVOper = IVSrc;
3105 Type *PostIncTy = PostIncV->getType();
3106 if (IVTy != PostIncTy) {
3107 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types")((PostIncTy->isPointerTy() && "mixing int/ptr IV types"
) ? static_cast<void> (0) : __assert_fail ("PostIncTy->isPointerTy() && \"mixing int/ptr IV types\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3107, __PRETTY_FUNCTION__))
;
3108 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3109 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3110 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3111 }
3112 Phi->replaceUsesOfWith(PostIncV, IVOper);
3113 DeadInsts.emplace_back(PostIncV);
3114 }
3115 }
3116}
3117
3118void LSRInstance::CollectFixupsAndInitialFormulae() {
3119 for (const IVStrideUse &U : IU) {
3120 Instruction *UserInst = U.getUser();
3121 // Skip IV users that are part of profitable IV Chains.
3122 User::op_iterator UseI =
3123 find(UserInst->operands(), U.getOperandValToReplace());
3124 assert(UseI != UserInst->op_end() && "cannot find IV operand")((UseI != UserInst->op_end() && "cannot find IV operand"
) ? static_cast<void> (0) : __assert_fail ("UseI != UserInst->op_end() && \"cannot find IV operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3124, __PRETTY_FUNCTION__))
;
3125 if (IVIncSet.count(UseI)) {
3126 DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Use is in profitable chain: "
<< **UseI << '\n'; } } while (false)
;
3127 continue;
3128 }
3129
3130 LSRUse::KindType Kind = LSRUse::Basic;
3131 MemAccessTy AccessTy;
3132 if (isAddressUse(UserInst, U.getOperandValToReplace())) {
3133 Kind = LSRUse::Address;
3134 AccessTy = getAccessType(UserInst);
3135 }
3136
3137 const SCEV *S = IU.getExpr(U);
3138 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3139
3140 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3141 // (N - i == 0), and this allows (N - i) to be the expression that we work
3142 // with rather than just N or i, so we can consider the register
3143 // requirements for both N and i at the same time. Limiting this code to
3144 // equality icmps is not a problem because all interesting loops use
3145 // equality icmps, thanks to IndVarSimplify.
3146 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3147 if (CI->isEquality()) {
3148 // Swap the operands if needed to put the OperandValToReplace on the
3149 // left, for consistency.
3150 Value *NV = CI->getOperand(1);
3151 if (NV == U.getOperandValToReplace()) {
3152 CI->setOperand(1, CI->getOperand(0));
3153 CI->setOperand(0, NV);
3154 NV = CI->getOperand(1);
3155 Changed = true;
3156 }
3157
3158 // x == y --> x - y == 0
3159 const SCEV *N = SE.getSCEV(NV);
3160 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3161 // S is normalized, so normalize N before folding it into S
3162 // to keep the result normalized.
3163 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
3164 TmpPostIncLoops, SE, DT);
3165 Kind = LSRUse::ICmpZero;
3166 S = SE.getMinusSCEV(N, S);
3167 }
3168
3169 // -1 and the negations of all interesting strides (except the negation
3170 // of -1) are now also interesting.
3171 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3172 if (Factors[i] != -1)
3173 Factors.insert(-(uint64_t)Factors[i]);
3174 Factors.insert(-1);
3175 }
3176
3177 // Get or create an LSRUse.
3178 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3179 size_t LUIdx = P.first;
3180 int64_t Offset = P.second;
3181 LSRUse &LU = Uses[LUIdx];
3182
3183 // Record the fixup.
3184 LSRFixup &LF = LU.getNewFixup();
3185 LF.UserInst = UserInst;
3186 LF.OperandValToReplace = U.getOperandValToReplace();
3187 LF.PostIncLoops = TmpPostIncLoops;
3188 LF.Offset = Offset;
3189 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3190
3191 if (!LU.WidestFixupType ||
3192 SE.getTypeSizeInBits(LU.WidestFixupType) <
3193 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3194 LU.WidestFixupType = LF.OperandValToReplace->getType();
3195
3196 // If this is the first use of this LSRUse, give it a formula.
3197 if (LU.Formulae.empty()) {
3198 InsertInitialFormula(S, LU, LUIdx);
3199 CountRegisters(LU.Formulae.back(), LUIdx);
3200 }
3201 }
3202
3203 DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3204}
3205
3206/// Insert a formula for the given expression into the given use, separating out
3207/// loop-variant portions from loop-invariant and loop-computable portions.
3208void
3209LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3210 // Mark uses whose expressions cannot be expanded.
3211 if (!isSafeToExpand(S, SE))
3212 LU.RigidFormula = true;
3213
3214 Formula F;
3215 F.initialMatch(S, L, SE);
3216 bool Inserted = InsertFormula(LU, LUIdx, F);
3217 assert(Inserted && "Initial formula already exists!")((Inserted && "Initial formula already exists!") ? static_cast
<void> (0) : __assert_fail ("Inserted && \"Initial formula already exists!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3217, __PRETTY_FUNCTION__))
; (void)Inserted;
3218}
3219
3220/// Insert a simple single-register formula for the given expression into the
3221/// given use.
3222void
3223LSRInstance::InsertSupplementalFormula(const SCEV *S,
3224 LSRUse &LU, size_t LUIdx) {
3225 Formula F;
3226 F.BaseRegs.push_back(S);
3227 F.HasBaseReg = true;
3228 bool Inserted = InsertFormula(LU, LUIdx, F);
3229 assert(Inserted && "Supplemental formula already exists!")((Inserted && "Supplemental formula already exists!")
? static_cast<void> (0) : __assert_fail ("Inserted && \"Supplemental formula already exists!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3229, __PRETTY_FUNCTION__))
; (void)Inserted;
3230}
3231
3232/// Note which registers are used by the given formula, updating RegUses.
3233void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3234 if (F.ScaledReg)
3235 RegUses.countRegister(F.ScaledReg, LUIdx);
3236 for (const SCEV *BaseReg : F.BaseRegs)
3237 RegUses.countRegister(BaseReg, LUIdx);
3238}
3239
3240/// If the given formula has not yet been inserted, add it to the list, and
3241/// return true. Return false otherwise.
3242bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3243 // Do not insert formula that we will not be able to expand.
3244 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&((isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy
, F) && "Formula is illegal") ? static_cast<void>
(0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Formula is illegal\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3245, __PRETTY_FUNCTION__))
3245 "Formula is illegal")((isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy
, F) && "Formula is illegal") ? static_cast<void>
(0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Formula is illegal\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3245, __PRETTY_FUNCTION__))
;
3246
3247 if (!LU.InsertFormula(F, *L))
3248 return false;
3249
3250 CountRegisters(F, LUIdx);
3251 return true;
3252}
3253
3254/// Check for other uses of loop-invariant values which we're tracking. These
3255/// other uses will pin these values in registers, making them less profitable
3256/// for elimination.
3257/// TODO: This currently misses non-constant addrec step registers.
3258/// TODO: Should this give more weight to users inside the loop?
3259void
3260LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3261 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3262 SmallPtrSet<const SCEV *, 32> Visited;
3263
3264 while (!Worklist.empty()) {
3265 const SCEV *S = Worklist.pop_back_val();
3266
3267 // Don't process the same SCEV twice
3268 if (!Visited.insert(S).second)
3269 continue;
3270
3271 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3272 Worklist.append(N->op_begin(), N->op_end());
3273 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3274 Worklist.push_back(C->getOperand());
3275 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3276 Worklist.push_back(D->getLHS());
3277 Worklist.push_back(D->getRHS());
3278 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3279 const Value *V = US->getValue();
3280 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3281 // Look for instructions defined outside the loop.
3282 if (L->contains(Inst)) continue;
3283 } else if (isa<UndefValue>(V))
3284 // Undef doesn't have a live range, so it doesn't matter.
3285 continue;
3286 for (const Use &U : V->uses()) {
3287 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3288 // Ignore non-instructions.
3289 if (!UserInst)
3290 continue;
3291 // Ignore instructions in other functions (as can happen with
3292 // Constants).
3293 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3294 continue;
3295 // Ignore instructions not dominated by the loop.
3296 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3297 UserInst->getParent() :
3298 cast<PHINode>(UserInst)->getIncomingBlock(
3299 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3300 if (!DT.dominates(L->getHeader(), UseBB))
3301 continue;
3302 // Don't bother if the instruction is in a BB which ends in an EHPad.
3303 if (UseBB->getTerminator()->isEHPad())
3304 continue;
3305 // Don't bother rewriting PHIs in catchswitch blocks.
3306 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3307 continue;
3308 // Ignore uses which are part of other SCEV expressions, to avoid
3309 // analyzing them multiple times.
3310 if (SE.isSCEVable(UserInst->getType())) {
3311 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3312 // If the user is a no-op, look through to its uses.
3313 if (!isa<SCEVUnknown>(UserS))
3314 continue;
3315 if (UserS == US) {
3316 Worklist.push_back(
3317 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3318 continue;
3319 }
3320 }
3321 // Ignore icmp instructions which are already being analyzed.
3322 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3323 unsigned OtherIdx = !U.getOperandNo();
3324 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3325 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3326 continue;
3327 }
3328
3329 std::pair<size_t, int64_t> P = getUse(
3330 S, LSRUse::Basic, MemAccessTy());
3331 size_t LUIdx = P.first;
3332 int64_t Offset = P.second;
3333 LSRUse &LU = Uses[LUIdx];
3334 LSRFixup &LF = LU.getNewFixup();
3335 LF.UserInst = const_cast<Instruction *>(UserInst);
3336 LF.OperandValToReplace = U;
3337 LF.Offset = Offset;
3338 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3339 if (!LU.WidestFixupType ||
3340 SE.getTypeSizeInBits(LU.WidestFixupType) <
3341 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3342 LU.WidestFixupType = LF.OperandValToReplace->getType();
3343 InsertSupplementalFormula(US, LU, LUIdx);
3344 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3345 break;
3346 }
3347 }
3348 }
3349}
3350
3351/// Split S into subexpressions which can be pulled out into separate
3352/// registers. If C is non-null, multiply each subexpression by C.
3353///
3354/// Return remainder expression after factoring the subexpressions captured by
3355/// Ops. If Ops is complete, return NULL.
3356static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3357 SmallVectorImpl<const SCEV *> &Ops,
3358 const Loop *L,
3359 ScalarEvolution &SE,
3360 unsigned Depth = 0) {
3361 // Arbitrarily cap recursion to protect compile time.
3362 if (Depth >= 3)
3363 return S;
3364
3365 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3366 // Break out add operands.
3367 for (const SCEV *S : Add->operands()) {
3368 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3369 if (Remainder)
3370 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3371 }
3372 return nullptr;
3373 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3374 // Split a non-zero base out of an addrec.
3375 if (AR->getStart()->isZero() || !AR->isAffine())
3376 return S;
3377
3378 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3379 C, Ops, L, SE, Depth+1);
3380 // Split the non-zero AddRec unless it is part of a nested recurrence that
3381 // does not pertain to this loop.
3382 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3383 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3384 Remainder = nullptr;
3385 }
3386 if (Remainder != AR->getStart()) {
3387 if (!Remainder)
3388 Remainder = SE.getConstant(AR->getType(), 0);
3389 return SE.getAddRecExpr(Remainder,
3390 AR->getStepRecurrence(SE),
3391 AR->getLoop(),
3392 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3393 SCEV::FlagAnyWrap);
3394 }
3395 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3396 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3397 if (Mul->getNumOperands() != 2)
3398 return S;
3399 if (const SCEVConstant *Op0 =
3400 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3401 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3402 const SCEV *Remainder =
3403 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3404 if (Remainder)
3405 Ops.push_back(SE.getMulExpr(C, Remainder));
3406 return nullptr;
3407 }
3408 }
3409 return S;
3410}
3411
3412/// \brief Helper function for LSRInstance::GenerateReassociations.
3413void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3414 const Formula &Base,
3415 unsigned Depth, size_t Idx,
3416 bool IsScaledReg) {
3417 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3418 SmallVector<const SCEV *, 8> AddOps;
3419 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3420 if (Remainder)
3421 AddOps.push_back(Remainder);
3422
3423 if (AddOps.size() == 1)
3424 return;
3425
3426 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3427 JE = AddOps.end();
3428 J != JE; ++J) {
3429
3430 // Loop-variant "unknown" values are uninteresting; we won't be able to
3431 // do anything meaningful with them.
3432 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3433 continue;
3434
3435 // Don't pull a constant into a register if the constant could be folded
3436 // into an immediate field.
3437 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3438 LU.AccessTy, *J, Base.getNumRegs() > 1))
3439 continue;
3440
3441 // Collect all operands except *J.
3442 SmallVector<const SCEV *, 8> InnerAddOps(
3443 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3444 InnerAddOps.append(std::next(J),
3445 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3446
3447 // Don't leave just a constant behind in a register if the constant could
3448 // be folded into an immediate field.
3449 if (InnerAddOps.size() == 1 &&
3450 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3451 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3452 continue;
3453
3454 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3455 if (InnerSum->isZero())
3456 continue;
3457 Formula F = Base;
3458
3459 // Add the remaining pieces of the add back into the new formula.
3460 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3461 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3462 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3463 InnerSumSC->getValue()->getZExtValue())) {
3464 F.UnfoldedOffset =
3465 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3466 if (IsScaledReg)
3467 F.ScaledReg = nullptr;
3468 else
3469 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3470 } else if (IsScaledReg)
3471 F.ScaledReg = InnerSum;
3472 else
3473 F.BaseRegs[Idx] = InnerSum;
3474
3475 // Add J as its own register, or an unfolded immediate.
3476 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3477 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3478 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3479 SC->getValue()->getZExtValue()))
3480 F.UnfoldedOffset =
3481 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3482 else
3483 F.BaseRegs.push_back(*J);
3484 // We may have changed the number of register in base regs, adjust the
3485 // formula accordingly.
3486 F.canonicalize(*L);
3487
3488 if (InsertFormula(LU, LUIdx, F))
3489 // If that formula hadn't been seen before, recurse to find more like
3490 // it.
3491 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3492 }
3493}
3494
3495/// Split out subexpressions from adds and the bases of addrecs.
3496void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3497 Formula Base, unsigned Depth) {
3498 assert(Base.isCanonical(*L) && "Input must be in the canonical form")((Base.isCanonical(*L) && "Input must be in the canonical form"
) ? static_cast<void> (0) : __assert_fail ("Base.isCanonical(*L) && \"Input must be in the canonical form\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3498, __PRETTY_FUNCTION__))
;
3499 // Arbitrarily cap recursion to protect compile time.
3500 if (Depth >= 3)
3501 return;
3502
3503 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3504 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3505
3506 if (Base.Scale == 1)
3507 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3508 /* Idx */ -1, /* IsScaledReg */ true);
3509}
3510
3511/// Generate a formula consisting of all of the loop-dominating registers added
3512/// into a single register.
3513void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3514 Formula Base) {
3515 // This method is only interesting on a plurality of registers.
3516 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3517 return;
3518
3519 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3520 // processing the formula.
3521 Base.unscale();
3522 Formula F = Base;
3523 F.BaseRegs.clear();
3524 SmallVector<const SCEV *, 4> Ops;
3525 for (const SCEV *BaseReg : Base.BaseRegs) {
3526 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3527 !SE.hasComputableLoopEvolution(BaseReg, L))
3528 Ops.push_back(BaseReg);
3529 else
3530 F.BaseRegs.push_back(BaseReg);
3531 }
3532 if (Ops.size() > 1) {
3533 const SCEV *Sum = SE.getAddExpr(Ops);
3534 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3535 // opportunity to fold something. For now, just ignore such cases
3536 // rather than proceed with zero in a register.
3537 if (!Sum->isZero()) {
3538 F.BaseRegs.push_back(Sum);
3539 F.canonicalize(*L);
3540 (void)InsertFormula(LU, LUIdx, F);
3541 }
3542 }
3543}
3544
3545/// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3546void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3547 const Formula &Base, size_t Idx,
3548 bool IsScaledReg) {
3549 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3550 GlobalValue *GV = ExtractSymbol(G, SE);
3551 if (G->isZero() || !GV)
3552 return;
3553 Formula F = Base;
3554 F.BaseGV = GV;
3555 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3556 return;
3557 if (IsScaledReg)
3558 F.ScaledReg = G;
3559 else
3560 F.BaseRegs[Idx] = G;
3561 (void)InsertFormula(LU, LUIdx, F);
3562}
3563
3564/// Generate reuse formulae using symbolic offsets.
3565void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3566 Formula Base) {
3567 // We can't add a symbolic offset if the address already contains one.
3568 if (Base.BaseGV) return;
3569
3570 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3571 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3572 if (Base.Scale == 1)
3573 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3574 /* IsScaledReg */ true);
3575}
3576
3577/// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3578void LSRInstance::GenerateConstantOffsetsImpl(
3579 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3580 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3581 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3582 for (int64_t Offset : Worklist) {
3583 Formula F = Base;
3584 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3585 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3586 LU.AccessTy, F)) {
3587 // Add the offset to the base register.
3588 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3589 // If it cancelled out, drop the base register, otherwise update it.
3590 if (NewG->isZero()) {
3591 if (IsScaledReg) {
3592 F.Scale = 0;
3593 F.ScaledReg = nullptr;
3594 } else
3595 F.deleteBaseReg(F.BaseRegs[Idx]);
3596 F.canonicalize(*L);
3597 } else if (IsScaledReg)
3598 F.ScaledReg = NewG;
3599 else
3600 F.BaseRegs[Idx] = NewG;
3601
3602 (void)InsertFormula(LU, LUIdx, F);
3603 }
3604 }
3605
3606 int64_t Imm = ExtractImmediate(G, SE);
3607 if (G->isZero() || Imm == 0)
3608 return;
3609 Formula F = Base;
3610 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3611 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3612 return;
3613 if (IsScaledReg)
3614 F.ScaledReg = G;
3615 else
3616 F.BaseRegs[Idx] = G;
3617 (void)InsertFormula(LU, LUIdx, F);
3618}
3619
3620/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3621void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3622 Formula Base) {
3623 // TODO: For now, just add the min and max offset, because it usually isn't
3624 // worthwhile looking at everything inbetween.
3625 SmallVector<int64_t, 2> Worklist;
3626 Worklist.push_back(LU.MinOffset);
3627 if (LU.MaxOffset != LU.MinOffset)
3628 Worklist.push_back(LU.MaxOffset);
3629
3630 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3631 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3632 if (Base.Scale == 1)
3633 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3634 /* IsScaledReg */ true);
3635}
3636
3637/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3638/// == y -> x*c == y*c.
3639void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3640 Formula Base) {
3641 if (LU.Kind != LSRUse::ICmpZero) return;
3642
3643 // Determine the integer type for the base formula.
3644 Type *IntTy = Base.getType();
3645 if (!IntTy) return;
3646 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3647
3648 // Don't do this if there is more than one offset.
3649 if (LU.MinOffset != LU.MaxOffset) return;
3650
3651 assert(!Base.BaseGV && "ICmpZero use is not legal!")((!Base.BaseGV && "ICmpZero use is not legal!") ? static_cast
<void> (0) : __assert_fail ("!Base.BaseGV && \"ICmpZero use is not legal!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3651, __PRETTY_FUNCTION__))
;
3652
3653 // Check each interesting stride.
3654 for (int64_t Factor : Factors) {
3655 // Check that the multiplication doesn't overflow.
3656 if (Base.BaseOffset == INT64_MIN(-9223372036854775807L -1) && Factor == -1)
3657 continue;
3658 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3659 if (NewBaseOffset / Factor != Base.BaseOffset)
3660 continue;
3661 // If the offset will be truncated at this use, check that it is in bounds.
3662 if (!IntTy->isPointerTy() &&
3663 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3664 continue;
3665
3666 // Check that multiplying with the use offset doesn't overflow.
3667 int64_t Offset = LU.MinOffset;
3668 if (Offset == INT64_MIN(-9223372036854775807L -1) && Factor == -1)
3669 continue;
3670 Offset = (uint64_t)Offset * Factor;
3671 if (Offset / Factor != LU.MinOffset)
3672 continue;
3673 // If the offset will be truncated at this use, check that it is in bounds.
3674 if (!IntTy->isPointerTy() &&
3675 !ConstantInt::isValueValidForType(IntTy, Offset))
3676 continue;
3677
3678 Formula F = Base;
3679 F.BaseOffset = NewBaseOffset;
3680
3681 // Check that this scale is legal.
3682 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3683 continue;
3684
3685 // Compensate for the use having MinOffset built into it.
3686 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3687
3688 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3689
3690 // Check that multiplying with each base register doesn't overflow.
3691 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3692 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3693 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3694 goto next;
3695 }
3696
3697 // Check that multiplying with the scaled register doesn't overflow.
3698 if (F.ScaledReg) {
3699 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3700 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3701 continue;
3702 }
3703
3704 // Check that multiplying with the unfolded offset doesn't overflow.
3705 if (F.UnfoldedOffset != 0) {
3706 if (F.UnfoldedOffset == INT64_MIN(-9223372036854775807L -1) && Factor == -1)
3707 continue;
3708 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3709 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3710 continue;
3711 // If the offset will be truncated, check that it is in bounds.
3712 if (!IntTy->isPointerTy() &&
3713 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3714 continue;
3715 }
3716
3717 // If we make it here and it's legal, add it.
3718 (void)InsertFormula(LU, LUIdx, F);
3719 next:;
3720 }
3721}
3722
3723/// Generate stride factor reuse formulae by making use of scaled-offset address
3724/// modes, for example.
3725void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3726 // Determine the integer type for the base formula.
3727 Type *IntTy = Base.getType();
3728 if (!IntTy) return;
3729
3730 // If this Formula already has a scaled register, we can't add another one.
3731 // Try to unscale the formula to generate a better scale.
3732 if (Base.Scale != 0 && !Base.unscale())
3733 return;
3734
3735 assert(Base.Scale == 0 && "unscale did not did its job!")((Base.Scale == 0 && "unscale did not did its job!") ?
static_cast<void> (0) : __assert_fail ("Base.Scale == 0 && \"unscale did not did its job!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3735, __PRETTY_FUNCTION__))
;
3736
3737 // Check each interesting stride.
3738 for (int64_t Factor : Factors) {
3739 Base.Scale = Factor;
3740 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3741 // Check whether this scale is going to be legal.
3742 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3743 Base)) {
3744 // As a special-case, handle special out-of-loop Basic users specially.
3745 // TODO: Reconsider this special case.
3746 if (LU.Kind == LSRUse::Basic &&
3747 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3748 LU.AccessTy, Base) &&
3749 LU.AllFixupsOutsideLoop)
3750 LU.Kind = LSRUse::Special;
3751 else
3752 continue;
3753 }
3754 // For an ICmpZero, negating a solitary base register won't lead to
3755 // new solutions.
3756 if (LU.Kind == LSRUse::ICmpZero &&
3757 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3758 continue;
3759 // For each addrec base reg, if its loop is current loop, apply the scale.
3760 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3761 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3762 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3763 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3764 if (FactorS->isZero())
3765 continue;
3766 // Divide out the factor, ignoring high bits, since we'll be
3767 // scaling the value back up in the end.
3768 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3769 // TODO: This could be optimized to avoid all the copying.
3770 Formula F = Base;
3771 F.ScaledReg = Quotient;
3772 F.deleteBaseReg(F.BaseRegs[i]);
3773 // The canonical representation of 1*reg is reg, which is already in
3774 // Base. In that case, do not try to insert the formula, it will be
3775 // rejected anyway.
3776 if (F.Scale == 1 && (F.BaseRegs.empty() ||
3777 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3778 continue;
3779 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3780 // non canonical Formula with ScaledReg's loop not being L.
3781 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3782 F.canonicalize(*L);
3783 (void)InsertFormula(LU, LUIdx, F);
3784 }
3785 }
3786 }
3787 }
3788}
3789
3790/// Generate reuse formulae from different IV types.
3791void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3792 // Don't bother truncating symbolic values.
3793 if (Base.BaseGV) return;
3794
3795 // Determine the integer type for the base formula.
3796 Type *DstTy = Base.getType();
3797 if (!DstTy) return;
3798 DstTy = SE.getEffectiveSCEVType(DstTy);
3799
3800 for (Type *SrcTy : Types) {
3801 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3802 Formula F = Base;
3803
3804 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3805 for (const SCEV *&BaseReg : F.BaseRegs)
3806 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3807
3808 // TODO: This assumes we've done basic processing on all uses and
3809 // have an idea what the register usage is.
3810 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3811 continue;
3812
3813 (void)InsertFormula(LU, LUIdx, F);
3814 }
3815 }
3816}
3817
3818namespace {
3819
3820/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3821/// modifications so that the search phase doesn't have to worry about the data
3822/// structures moving underneath it.
3823struct WorkItem {
3824 size_t LUIdx;
3825 int64_t Imm;
3826 const SCEV *OrigReg;
3827
3828 WorkItem(size_t LI, int64_t I, const SCEV *R)
3829 : LUIdx(LI), Imm(I), OrigReg(R) {}
3830
3831 void print(raw_ostream &OS) const;
3832 void dump() const;
3833};
3834
3835} // end anonymous namespace
3836
3837void WorkItem::print(raw_ostream &OS) const {
3838 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3839 << " , add offset " << Imm;
3840}
3841
3842#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3843LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
3844 print(errs()); errs() << '\n';
3845}
3846#endif
3847
3848/// Look for registers which are a constant distance apart and try to form reuse
3849/// opportunities between them.
3850void LSRInstance::GenerateCrossUseConstantOffsets() {
3851 // Group the registers by their value without any added constant offset.
3852 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3853 DenseMap<const SCEV *, ImmMapTy> Map;
3854 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3855 SmallVector<const SCEV *, 8> Sequence;
3856 for (const SCEV *Use : RegUses) {
3857 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3858 int64_t Imm = ExtractImmediate(Reg, SE);
3859 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3860 if (Pair.second)
3861 Sequence.push_back(Reg);
3862 Pair.first->second.insert(std::make_pair(Imm, Use));
3863 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3864 }
3865
3866 // Now examine each set of registers with the same base value. Build up
3867 // a list of work to do and do the work in a separate step so that we're
3868 // not adding formulae and register counts while we're searching.
3869 SmallVector<WorkItem, 32> WorkItems;
3870 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3871 for (const SCEV *Reg : Sequence) {
3872 const ImmMapTy &Imms = Map.find(Reg)->second;
3873
3874 // It's not worthwhile looking for reuse if there's only one offset.
3875 if (Imms.size() == 1)
3876 continue;
3877
3878 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
<< *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
3879 for (const auto &Entry : Imms)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
<< *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
3880 dbgs() << ' ' << Entry.first;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
<< *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
3881 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
<< *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
;
3882
3883 // Examine each offset.
3884 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3885 J != JE; ++J) {
3886 const SCEV *OrigReg = J->second;
3887
3888 int64_t JImm = J->first;
3889 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3890
3891 if (!isa<SCEVConstant>(OrigReg) &&
3892 UsedByIndicesMap[Reg].count() == 1) {
3893 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
3894 continue;
3895 }
3896
3897 // Conservatively examine offsets between this orig reg a few selected
3898 // other orig regs.
3899 ImmMapTy::const_iterator OtherImms[] = {
3900 Imms.begin(), std::prev(Imms.end()),
3901 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3902 2)
3903 };
3904 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3905 ImmMapTy::const_iterator M = OtherImms[i];
3906 if (M == J || M == JE) continue;
3907
3908 // Compute the difference between the two.
3909 int64_t Imm = (uint64_t)JImm - M->first;
3910 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3911 LUIdx = UsedByIndices.find_next(LUIdx))
3912 // Make a memo of this use, offset, and register tuple.
3913 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3914 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3915 }
3916 }
3917 }
3918
3919 Map.clear();
3920 Sequence.clear();
3921 UsedByIndicesMap.clear();
3922 UniqueItems.clear();
3923
3924 // Now iterate through the worklist and add new formulae.
3925 for (const WorkItem &WI : WorkItems) {
3926 size_t LUIdx = WI.LUIdx;
3927 LSRUse &LU = Uses[LUIdx];
3928 int64_t Imm = WI.Imm;
3929 const SCEV *OrigReg = WI.OrigReg;
3930
3931 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3932 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3933 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3934
3935 // TODO: Use a more targeted data structure.
3936 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3937 Formula F = LU.Formulae[L];
3938 // FIXME: The code for the scaled and unscaled registers looks
3939 // very similar but slightly different. Investigate if they
3940 // could be merged. That way, we would not have to unscale the
3941 // Formula.
3942 F.unscale();
3943 // Use the immediate in the scaled register.
3944 if (F.ScaledReg == OrigReg) {
3945 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3946 // Don't create 50 + reg(-50).
3947 if (F.referencesReg(SE.getSCEV(
3948 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3949 continue;
3950 Formula NewF = F;
3951 NewF.BaseOffset = Offset;
3952 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3953 NewF))
3954 continue;
3955 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3956
3957 // If the new scale is a constant in a register, and adding the constant
3958 // value to the immediate would produce a value closer to zero than the
3959 // immediate itself, then the formula isn't worthwhile.
3960 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3961 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
3962 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
3963 .ule(std::abs(NewF.BaseOffset)))
3964 continue;
3965
3966 // OK, looks good.
3967 NewF.canonicalize(*this->L);
3968 (void)InsertFormula(LU, LUIdx, NewF);
3969 } else {
3970 // Use the immediate in a base register.
3971 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3972 const SCEV *BaseReg = F.BaseRegs[N];
3973 if (BaseReg != OrigReg)
3974 continue;
3975 Formula NewF = F;
3976 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3977 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3978 LU.Kind, LU.AccessTy, NewF)) {
3979 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3980 continue;
3981 NewF = F;
3982 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3983 }
3984 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3985
3986 // If the new formula has a constant in a register, and adding the
3987 // constant value to the immediate would produce a value closer to
3988 // zero than the immediate itself, then the formula isn't worthwhile.
3989 for (const SCEV *NewReg : NewF.BaseRegs)
3990 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
3991 if ((C->getAPInt() + NewF.BaseOffset)
3992 .abs()
3993 .slt(std::abs(NewF.BaseOffset)) &&
3994 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
3995 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3996 goto skip_formula;
3997
3998 // Ok, looks good.
3999 NewF.canonicalize(*this->L);
4000 (void)InsertFormula(LU, LUIdx, NewF);
4001 break;
4002 skip_formula:;
4003 }
4004 }
4005 }
4006 }
4007}
4008
4009/// Generate formulae for each use.
4010void
4011LSRInstance::GenerateAllReuseFormulae() {
4012 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4013 // queries are more precise.
4014 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4015 LSRUse &LU = Uses[LUIdx];
4016 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4017 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4018 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4019 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4020 }
4021 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4022 LSRUse &LU = Uses[LUIdx];
4023 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4024 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4025 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4026 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4027 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4028 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4029 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4030 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4031 }
4032 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4033 LSRUse &LU = Uses[LUIdx];
4034 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4035 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4036 }
4037
4038 GenerateCrossUseConstantOffsets();
4039
4040 DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4041 "After generating reuse formulae:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4042 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4043}
4044
4045/// If there are multiple formulae with the same set of registers used
4046/// by other uses, pick the best one and delete the others.
4047void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4048 DenseSet<const SCEV *> VisitedRegs;
4049 SmallPtrSet<const SCEV *, 16> Regs;
4050 SmallPtrSet<const SCEV *, 16> LoserRegs;
4051#ifndef NDEBUG
4052 bool ChangedFormulae = false;
4053#endif
4054
4055 // Collect the best formula for each unique set of shared registers. This
4056 // is reset for each use.
4057 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
4058 BestFormulaeTy;
4059 BestFormulaeTy BestFormulae;
4060
4061 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4062 LSRUse &LU = Uses[LUIdx];
4063 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4064
4065 bool Any = false;
4066 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4067 FIdx != NumForms; ++FIdx) {
4068 Formula &F = LU.Formulae[FIdx];
4069
4070 // Some formulas are instant losers. For example, they may depend on
4071 // nonexistent AddRecs from other loops. These need to be filtered
4072 // immediately, otherwise heuristics could choose them over others leading
4073 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4074 // avoids the need to recompute this information across formulae using the
4075 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4076 // the corresponding bad register from the Regs set.
4077 Cost CostF;
4078 Regs.clear();
4079 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4080 if (CostF.isLoser()) {
4081 // During initial formula generation, undesirable formulae are generated
4082 // by uses within other loops that have some non-trivial address mode or
4083 // use the postinc form of the IV. LSR needs to provide these formulae
4084 // as the basis of rediscovering the desired formula that uses an AddRec
4085 // corresponding to the existing phi. Once all formulae have been
4086 // generated, these initial losers may be pruned.
4087 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
4088 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4089 }
4090 else {
4091 SmallVector<const SCEV *, 4> Key;
4092 for (const SCEV *Reg : F.BaseRegs) {
4093 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4094 Key.push_back(Reg);
4095 }
4096 if (F.ScaledReg &&
4097 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4098 Key.push_back(F.ScaledReg);
4099 // Unstable sort by host order ok, because this is only used for
4100 // uniquifying.
4101 std::sort(Key.begin(), Key.end());
4102
4103 std::pair<BestFormulaeTy::const_iterator, bool> P =
4104 BestFormulae.insert(std::make_pair(Key, FIdx));
4105 if (P.second)
4106 continue;
4107
4108 Formula &Best = LU.Formulae[P.first->second];
4109
4110 Cost CostBest;
4111 Regs.clear();
4112 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4113 if (CostF < CostBest)
4114 std::swap(F, Best);
4115 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" " in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
4116 dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" " in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
4117 " in favor of formula "; Best.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" " in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
4118 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" " in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
;
4119 }
4120#ifndef NDEBUG
4121 ChangedFormulae = true;
4122#endif
4123 LU.DeleteFormula(F);
4124 --FIdx;
4125 --NumForms;
4126 Any = true;
4127 }
4128
4129 // Now that we've filtered out some formulae, recompute the Regs set.
4130 if (Any)
4131 LU.RecomputeRegs(LUIdx, RegUses);
4132
4133 // Reset this to prepare for the next use.
4134 BestFormulae.clear();
4135 }
4136
4137 DEBUG(if (ChangedFormulae) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
4138 dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
4139 "After filtering out undesirable candidates:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
4140 print_uses(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
4141 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4142}
4143
4144// This is a rough guess that seems to work fairly well.
4145static const size_t ComplexityLimit = UINT16_MAX(65535);
4146
4147/// Estimate the worst-case number of solutions the solver might have to
4148/// consider. It almost never considers this many solutions because it prune the
4149/// search space, but the pruning isn't always sufficient.
4150size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4151 size_t Power = 1;
4152 for (const LSRUse &LU : Uses) {
4153 size_t FSize = LU.Formulae.size();
4154 if (FSize >= ComplexityLimit) {
4155 Power = ComplexityLimit;
4156 break;
4157 }
4158 Power *= FSize;
4159 if (Power >= ComplexityLimit)
4160 break;
4161 }
4162 return Power;
4163}
4164
4165/// When one formula uses a superset of the registers of another formula, it
4166/// won't help reduce register pressure (though it may not necessarily hurt
4167/// register pressure); remove it to simplify the system.
4168void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4169 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4170 DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false)
;
4171
4172 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by eliminating formulae "
"which use a superset of registers used by other " "formulae.\n"
; } } while (false)
4173 "which use a superset of registers used by other "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by eliminating formulae "
"which use a superset of registers used by other " "formulae.\n"
; } } while (false)
4174 "formulae.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by eliminating formulae "
"which use a superset of registers used by other " "formulae.\n"
; } } while (false)
;
4175
4176 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4177 LSRUse &LU = Uses[LUIdx];
4178 bool Any = false;
4179 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4180 Formula &F = LU.Formulae[i];
4181 // Look for a formula with a constant or GV in a register. If the use
4182 // also has a formula with that same value in an immediate field,
4183 // delete the one that uses a register.
4184 for (SmallVectorImpl<const SCEV *>::const_iterator
4185 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4186 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4187 Formula NewF = F;
4188 NewF.BaseOffset += C->getValue()->getSExtValue();
4189 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4190 (I - F.BaseRegs.begin()));
4191 if (LU.HasFormulaWithSameRegs(NewF)) {
4192 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4193 LU.DeleteFormula(F);
4194 --i;
4195 --e;
4196 Any = true;
4197 break;
4198 }
4199 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4200 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4201 if (!F.BaseGV) {
4202 Formula NewF = F;
4203 NewF.BaseGV = GV;
4204 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4205 (I - F.BaseRegs.begin()));
4206 if (LU.HasFormulaWithSameRegs(NewF)) {
4207 DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4208 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4209 LU.DeleteFormula(F);
4210 --i;
4211 --e;
4212 Any = true;
4213 break;
4214 }
4215 }
4216 }
4217 }
4218 }
4219 if (Any)
4220 LU.RecomputeRegs(LUIdx, RegUses);
4221 }
4222
4223 DEBUG(dbgs() << "After pre-selection:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
4224 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
;
4225 }
4226}
4227
4228/// When there are many registers for expressions like A, A+1, A+2, etc.,
4229/// allocate a single register for them.
4230void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4231 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4232 return;
4233
4234 DEBUG(dbgs() << "The search space is too complex.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by assuming that uses separated "
"by a constant offset will use the same registers.\n"; } } while
(false)
4235 "Narrowing the search space by assuming that uses separated "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by assuming that uses separated "
"by a constant offset will use the same registers.\n"; } } while
(false)
4236 "by a constant offset will use the same registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by assuming that uses separated "
"by a constant offset will use the same registers.\n"; } } while
(false)
;
4237
4238 // This is especially useful for unrolled loops.
4239
4240 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4241 LSRUse &LU = Uses[LUIdx];
4242 for (const Formula &F : LU.Formulae) {
4243 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4244 continue;
4245
4246 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4247 if (!LUThatHas)
4248 continue;
4249
4250 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4251 LU.Kind, LU.AccessTy))
4252 continue;
4253
4254 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4255
4256 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4257
4258 // Transfer the fixups of LU to LUThatHas.
4259 for (LSRFixup &Fixup : LU.Fixups) {
4260 Fixup.Offset += F.BaseOffset;
4261 LUThatHas->pushFixup(Fixup);
4262 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New fixup has offset " <<
Fixup.Offset << '\n'; } } while (false)
;
4263 }
4264
4265 // Delete formulae from the new use which are no longer legal.
4266 bool Any = false;
4267 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4268 Formula &F = LUThatHas->Formulae[i];
4269 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4270 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4271 DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4272 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4273 LUThatHas->DeleteFormula(F);
4274 --i;
4275 --e;
4276 Any = true;
4277 }
4278 }
4279
4280 if (Any)
4281 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4282
4283 // Delete the old use.
4284 DeleteUse(LU, LUIdx);
4285 --LUIdx;
4286 --NumUses;
4287 break;
4288 }
4289 }
4290
4291 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
;
4292}
4293
4294/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4295/// we've done more filtering, as it may be able to find more formulae to
4296/// eliminate.
4297void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4298 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4299 DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false)
;
4300
4301 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by re-filtering out "
"undesirable dedicated registers.\n"; } } while (false)
4302 "undesirable dedicated registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by re-filtering out "
"undesirable dedicated registers.\n"; } } while (false)
;
4303
4304 FilterOutUndesirableDedicatedRegisters();
4305
4306 DEBUG(dbgs() << "After pre-selection:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
4307 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
;
4308 }
4309}
4310
4311/// The function delete formulas with high registers number expectation.
4312/// Assuming we don't know the value of each formula (already delete
4313/// all inefficient), generate probability of not selecting for each
4314/// register.
4315/// For example,
4316/// Use1:
4317/// reg(a) + reg({0,+,1})
4318/// reg(a) + reg({-1,+,1}) + 1
4319/// reg({a,+,1})
4320/// Use2:
4321/// reg(b) + reg({0,+,1})
4322/// reg(b) + reg({-1,+,1}) + 1
4323/// reg({b,+,1})
4324/// Use3:
4325/// reg(c) + reg(b) + reg({0,+,1})
4326/// reg(c) + reg({b,+,1})
4327///
4328/// Probability of not selecting
4329/// Use1 Use2 Use3
4330/// reg(a) (1/3) * 1 * 1
4331/// reg(b) 1 * (1/3) * (1/2)
4332/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4333/// reg({-1,+,1}) (2/3) * (2/3) * 1
4334/// reg({a,+,1}) (2/3) * 1 * 1
4335/// reg({b,+,1}) 1 * (2/3) * (2/3)
4336/// reg(c) 1 * 1 * 0
4337///
4338/// Now count registers number mathematical expectation for each formula:
4339/// Note that for each use we exclude probability if not selecting for the use.
4340/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4341/// probabilty 1/3 of not selecting for Use1).
4342/// Use1:
4343/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4344/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4345/// reg({a,+,1}) 1
4346/// Use2:
4347/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4348/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4349/// reg({b,+,1}) 2/3
4350/// Use3:
4351/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4352/// reg(c) + reg({b,+,1}) 1 + 2/3
4353
4354void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4355 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4356 return;
4357 // Ok, we have too many of formulae on our hands to conveniently handle.
4358 // Use a rough heuristic to thin out the list.
4359
4360 // Set of Regs wich will be 100% used in final solution.
4361 // Used in each formula of a solution (in example above this is reg(c)).
4362 // We can skip them in calculations.
4363 SmallPtrSet<const SCEV *, 4> UniqRegs;
4364 DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false)
;
4365
4366 // Map each register to probability of not selecting
4367 DenseMap <const SCEV *, float> RegNumMap;
4368 for (const SCEV *Reg : RegUses) {
4369 if (UniqRegs.count(Reg))
4370 continue;
4371 float PNotSel = 1;
4372 for (const LSRUse &LU : Uses) {
4373 if (!LU.Regs.count(Reg))
4374 continue;
4375 float P = LU.getNotSelectedProbability(Reg);
4376 if (P != 0.0)
4377 PNotSel *= P;
4378 else
4379 UniqRegs.insert(Reg);
4380 }
4381 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4382 }
4383
4384 DEBUG(dbgs() << "Narrowing the search space by deleting costly formulas\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
;
4385
4386 // Delete formulas where registers number expectation is high.
4387 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4388 LSRUse &LU = Uses[LUIdx];
4389 // If nothing to delete - continue.
4390 if (LU.Formulae.size() < 2)
4391 continue;
4392 // This is temporary solution to test performance. Float should be
4393 // replaced with round independent type (based on integers) to avoid
4394 // different results for different target builds.
4395 float FMinRegNum = LU.Formulae[0].getNumRegs();
4396 float FMinARegNum = LU.Formulae[0].getNumRegs();
4397 size_t MinIdx = 0;
4398 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4399 Formula &F = LU.Formulae[i];
4400 float FRegNum = 0;
4401 float FARegNum = 0;
4402 for (const SCEV *BaseReg : F.BaseRegs) {
4403 if (UniqRegs.count(BaseReg))
4404 continue;
4405 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4406 if (isa<SCEVAddRecExpr>(BaseReg))
4407 FARegNum +=
4408 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4409 }
4410 if (const SCEV *ScaledReg = F.ScaledReg) {
4411 if (!UniqRegs.count(ScaledReg)) {
4412 FRegNum +=
4413 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4414 if (isa<SCEVAddRecExpr>(ScaledReg))
4415 FARegNum +=
4416 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4417 }
4418 }
4419 if (FMinRegNum > FRegNum ||
4420 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4421 FMinRegNum = FRegNum;
4422 FMinARegNum = FARegNum;
4423 MinIdx = i;
4424 }
4425 }
4426 DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " The formula "; LU.Formulae
[MinIdx].print(dbgs()); dbgs() << " with min reg num " <<
FMinRegNum << '\n'; } } while (false)
4427 dbgs() << " with min reg num " << FMinRegNum << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " The formula "; LU.Formulae
[MinIdx].print(dbgs()); dbgs() << " with min reg num " <<
FMinRegNum << '\n'; } } while (false)
;
4428 if (MinIdx != 0)
4429 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4430 while (LU.Formulae.size() != 1) {
4431 DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
4432 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4433 LU.Formulae.pop_back();
4434 }
4435 LU.RecomputeRegs(LUIdx, RegUses);
4436 assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula")((LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula"
) ? static_cast<void> (0) : __assert_fail ("LU.Formulae.size() == 1 && \"Should be exactly 1 min regs formula\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4436, __PRETTY_FUNCTION__))
;
4437 Formula &F = LU.Formulae[0];
4438 DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Leaving only "; F.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4439 // When we choose the formula, the regs become unique.
4440 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4441 if (F.ScaledReg)
4442 UniqRegs.insert(F.ScaledReg);
4443 }
4444 DEBUG(dbgs() << "After pre-selection:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
4445 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
;
4446}
4447
4448
4449/// Pick a register which seems likely to be profitable, and then in any use
4450/// which has any reference to that register, delete all formulae which do not
4451/// reference that register.
4452void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4453 // With all other options exhausted, loop until the system is simple
4454 // enough to handle.
4455 SmallPtrSet<const SCEV *, 4> Taken;
4456 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4457 // Ok, we have too many of formulae on our hands to conveniently handle.
4458 // Use a rough heuristic to thin out the list.
4459 DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false)
;
4460
4461 // Pick the register which is used by the most LSRUses, which is likely
4462 // to be a good reuse register candidate.
4463 const SCEV *Best = nullptr;
4464 unsigned BestNum = 0;
4465 for (const SCEV *Reg : RegUses) {
4466 if (Taken.count(Reg))
4467 continue;
4468 if (!Best) {
4469 Best = Reg;
4470 BestNum = RegUses.getUsedByIndices(Reg).count();
4471 } else {
4472 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4473 if (Count > BestNum) {
4474 Best = Reg;
4475 BestNum = Count;
4476 }
4477 }
4478 }
4479
4480 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Bestdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by assuming "
<< *Best << " will yield profitable reuse.\n"; }
} while (false)
4481 << " will yield profitable reuse.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by assuming "
<< *Best << " will yield profitable reuse.\n"; }
} while (false)
;
4482 Taken.insert(Best);
4483
4484 // In any use with formulae which references this register, delete formulae
4485 // which don't reference it.
4486 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4487 LSRUse &LU = Uses[LUIdx];
4488 if (!LU.Regs.count(Best)) continue;
4489
4490 bool Any = false;
4491 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4492 Formula &F = LU.Formulae[i];
4493 if (!F.referencesReg(Best)) {
4494 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4495 LU.DeleteFormula(F);
4496 --e;
4497 --i;
4498 Any = true;
4499 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?")((e != 0 && "Use has no formulae left! Is Regs inconsistent?"
) ? static_cast<void> (0) : __assert_fail ("e != 0 && \"Use has no formulae left! Is Regs inconsistent?\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4499, __PRETTY_FUNCTION__))
;
4500 continue;
4501 }
4502 }
4503
4504 if (Any)
4505 LU.RecomputeRegs(LUIdx, RegUses);
4506 }
4507
4508 DEBUG(dbgs() << "After pre-selection:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
4509 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false)
;
4510 }
4511}
4512
4513/// If there are an extraordinary number of formulae to choose from, use some
4514/// rough heuristics to prune down the number of formulae. This keeps the main
4515/// solver from taking an extraordinary amount of time in some worst-case
4516/// scenarios.
4517void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4518 NarrowSearchSpaceByDetectingSupersets();
4519 NarrowSearchSpaceByCollapsingUnrolledCode();
4520 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4521 if (LSRExpNarrow)
4522 NarrowSearchSpaceByDeletingCostlyFormulas();
4523 else
4524 NarrowSearchSpaceByPickingWinnerRegs();
4525}
4526
4527/// This is the recursive solver.
4528void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4529 Cost &SolutionCost,
4530 SmallVectorImpl<const Formula *> &Workspace,
4531 const Cost &CurCost,
4532 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4533 DenseSet<const SCEV *> &VisitedRegs) const {
4534 // Some ideas:
4535 // - prune more:
4536 // - use more aggressive filtering
4537 // - sort the formula so that the most profitable solutions are found first
4538 // - sort the uses too
4539 // - search faster:
4540 // - don't compute a cost, and then compare. compare while computing a cost
4541 // and bail early.
4542 // - track register sets with SmallBitVector
4543
4544 const LSRUse &LU = Uses[Workspace.size()];
4545
4546 // If this use references any register that's already a part of the
4547 // in-progress solution, consider it a requirement that a formula must
4548 // reference that register in order to be considered. This prunes out
4549 // unprofitable searching.
4550 SmallSetVector<const SCEV *, 4> ReqRegs;
4551 for (const SCEV *S : CurRegs)
4552 if (LU.Regs.count(S))
4553 ReqRegs.insert(S);
4554
4555 SmallPtrSet<const SCEV *, 16> NewRegs;
4556 Cost NewCost;
4557 for (const Formula &F : LU.Formulae) {
4558 // Ignore formulae which may not be ideal in terms of register reuse of
4559 // ReqRegs. The formula should use all required registers before
4560 // introducing new ones.
4561 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4562 for (const SCEV *Reg : ReqRegs) {
4563 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4564 is_contained(F.BaseRegs, Reg)) {
4565 --NumReqRegsToFind;
4566 if (NumReqRegsToFind == 0)
4567 break;
4568 }
4569 }
4570 if (NumReqRegsToFind != 0) {
4571 // If none of the formulae satisfied the required registers, then we could
4572 // clear ReqRegs and try again. Currently, we simply give up in this case.
4573 continue;
4574 }
4575
4576 // Evaluate the cost of the current formula. If it's already worse than
4577 // the current best, prune the search at that point.
4578 NewCost = CurCost;
4579 NewRegs = CurRegs;
4580 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4581 if (NewCost < SolutionCost) {
4582 Workspace.push_back(&F);
4583 if (Workspace.size() != Uses.size()) {
4584 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4585 NewRegs, VisitedRegs);
4586 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4587 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4588 } else {
4589 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\n Regs:"; for (const SCEV *S : NewRegs
) dbgs() << ' ' << *S; dbgs() << '\n'; } } while
(false)
4590 dbgs() << ".\n Regs:";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\n Regs:"; for (const SCEV *S : NewRegs
) dbgs() << ' ' << *S; dbgs() << '\n'; } } while
(false)
4591 for (const SCEV *S : NewRegs)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\n Regs:"; for (const SCEV *S : NewRegs
) dbgs() << ' ' << *S; dbgs() << '\n'; } } while
(false)
4592 dbgs() << ' ' << *S;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\n Regs:"; for (const SCEV *S : NewRegs
) dbgs() << ' ' << *S; dbgs() << '\n'; } } while
(false)
4593 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\n Regs:"; for (const SCEV *S : NewRegs
) dbgs() << ' ' << *S; dbgs() << '\n'; } } while
(false)
;
4594
4595 SolutionCost = NewCost;
4596 Solution = Workspace;
4597 }
4598 Workspace.pop_back();
4599 }
4600 }
4601}
4602
4603/// Choose one formula from each use. Return the results in the given Solution
4604/// vector.
4605void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4606 SmallVector<const Formula *, 8> Workspace;
4607 Cost SolutionCost;
4608 SolutionCost.Lose();
4609 Cost CurCost;
4610 SmallPtrSet<const SCEV *, 16> CurRegs;
4611 DenseSet<const SCEV *> VisitedRegs;
4612 Workspace.reserve(Uses.size());
4613
4614 // SolveRecurse does all the work.
4615 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4616 CurRegs, VisitedRegs);
4617 if (Solution.empty()) {
4618 DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
4619 return;
4620 }
4621
4622 // Ok, we've now made all our decisions.
4623 DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4624 "The chosen solution requires "; SolutionCost.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4625 dbgs() << ":\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4626 for (size_t i = 0, e = Uses.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4627 dbgs() << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4628 Uses[i].print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4629 dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4630 " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4631 Solution[i]->print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4632 dbgs() << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
4633 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
i = 0, e = Uses.size(); i != e; ++i) { dbgs() << " ";
Uses[i].print(dbgs()); dbgs() << "\n" " "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
;
4634
4635 assert(Solution.size() == Uses.size() && "Malformed solution!")((Solution.size() == Uses.size() && "Malformed solution!"
) ? static_cast<void> (0) : __assert_fail ("Solution.size() == Uses.size() && \"Malformed solution!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4635, __PRETTY_FUNCTION__))
;
4636}
4637
4638/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4639/// we can go while still being dominated by the input positions. This helps
4640/// canonicalize the insert position, which encourages sharing.
4641BasicBlock::iterator
4642LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4643 const SmallVectorImpl<Instruction *> &Inputs)
4644 const {
4645 Instruction *Tentative = &*IP;
4646 while (true) {
4647 bool AllDominate = true;
4648 Instruction *BetterPos = nullptr;
4649 // Don't bother attempting to insert before a catchswitch, their basic block
4650 // cannot have other non-PHI instructions.
4651 if (isa<CatchSwitchInst>(Tentative))
4652 return IP;
4653
4654 for (Instruction *Inst : Inputs) {
4655 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4656 AllDominate = false;
4657 break;
4658 }
4659 // Attempt to find an insert position in the middle of the block,
4660 // instead of at the end, so that it can be used for other expansions.
4661 if (Tentative->getParent() == Inst->getParent() &&
4662 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4663 BetterPos = &*std::next(BasicBlock::iterator(Inst));
4664 }
4665 if (!AllDominate)
4666 break;
4667 if (BetterPos)
4668 IP = BetterPos->getIterator();
4669 else
4670 IP = Tentative->getIterator();
4671
4672 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4673 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4674
4675 BasicBlock *IDom;
4676 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4677 if (!Rung) return IP;
4678 Rung = Rung->getIDom();
4679 if (!Rung) return IP;
4680 IDom = Rung->getBlock();
4681
4682 // Don't climb into a loop though.
4683 const Loop *IDomLoop = LI.getLoopFor(IDom);
4684 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4685 if (IDomDepth <= IPLoopDepth &&
4686 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4687 break;
4688 }
4689
4690 Tentative = IDom->getTerminator();
4691 }
4692
4693 return IP;
4694}
4695
4696/// Determine an input position which will be dominated by the operands and
4697/// which will dominate the result.
4698BasicBlock::iterator
4699LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4700 const LSRFixup &LF,
4701 const LSRUse &LU,
4702 SCEVExpander &Rewriter) const {
4703 // Collect some instructions which must be dominated by the
4704 // expanding replacement. These must be dominated by any operands that
4705 // will be required in the expansion.
4706 SmallVector<Instruction *, 4> Inputs;
4707 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4708 Inputs.push_back(I);
4709 if (LU.Kind == LSRUse::ICmpZero)
4710 if (Instruction *I =
4711 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4712 Inputs.push_back(I);
4713 if (LF.PostIncLoops.count(L)) {
4714 if (LF.isUseFullyOutsideLoop(L))
4715 Inputs.push_back(L->getLoopLatch()->getTerminator());
4716 else
4717 Inputs.push_back(IVIncInsertPos);
4718 }
4719 // The expansion must also be dominated by the increment positions of any
4720 // loops it for which it is using post-inc mode.
4721 for (const Loop *PIL : LF.PostIncLoops) {
4722 if (PIL == L) continue;
4723
4724 // Be dominated by the loop exit.
4725 SmallVector<BasicBlock *, 4> ExitingBlocks;
4726 PIL->getExitingBlocks(ExitingBlocks);
4727 if (!ExitingBlocks.empty()) {
4728 BasicBlock *BB = ExitingBlocks[0];
4729 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4730 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4731 Inputs.push_back(BB->getTerminator());
4732 }
4733 }
4734
4735 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()((!isa<PHINode>(LowestIP) && !LowestIP->isEHPad
() && !isa<DbgInfoIntrinsic>(LowestIP) &&
"Insertion point must be a normal instruction") ? static_cast
<void> (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4737, __PRETTY_FUNCTION__))
4736 && !isa<DbgInfoIntrinsic>(LowestIP) &&((!isa<PHINode>(LowestIP) && !LowestIP->isEHPad
() && !isa<DbgInfoIntrinsic>(LowestIP) &&
"Insertion point must be a normal instruction") ? static_cast
<void> (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4737, __PRETTY_FUNCTION__))
4737 "Insertion point must be a normal instruction")((!isa<PHINode>(LowestIP) && !LowestIP->isEHPad
() && !isa<DbgInfoIntrinsic>(LowestIP) &&
"Insertion point must be a normal instruction") ? static_cast
<void> (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4737, __PRETTY_FUNCTION__))
;
4738
4739 // Then, climb up the immediate dominator tree as far as we can go while
4740 // still being dominated by the input positions.
4741 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4742
4743 // Don't insert instructions before PHI nodes.
4744 while (isa<PHINode>(IP)) ++IP;
4745
4746 // Ignore landingpad instructions.
4747 while (IP->isEHPad()) ++IP;
4748
4749 // Ignore debug intrinsics.
4750 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4751
4752 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4753 // IP consistent across expansions and allows the previously inserted
4754 // instructions to be reused by subsequent expansion.
4755 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4756 ++IP;
4757
4758 return IP;
4759}
4760
4761/// Emit instructions for the leading candidate expression for this LSRUse (this
4762/// is called "expanding").
4763Value *LSRInstance::Expand(const LSRUse &LU,
4764 const LSRFixup &LF,
4765 const Formula &F,
4766 BasicBlock::iterator IP,
4767 SCEVExpander &Rewriter,
4768 SmallVectorImpl<WeakVH> &DeadInsts) const {
4769 if (LU.RigidFormula)
4770 return LF.OperandValToReplace;
4771
4772 // Determine an input position which will be dominated by the operands and
4773 // which will dominate the result.
4774 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4775 Rewriter.setInsertPoint(&*IP);
4776
4777 // Inform the Rewriter if we have a post-increment use, so that it can
4778 // perform an advantageous expansion.
4779 Rewriter.setPostInc(LF.PostIncLoops);
4780
4781 // This is the type that the user actually needs.
4782 Type *OpTy = LF.OperandValToReplace->getType();
4783 // This will be the type that we'll initially expand to.
4784 Type *Ty = F.getType();
4785 if (!Ty)
4786 // No type known; just expand directly to the ultimate type.
4787 Ty = OpTy;
4788 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4789 // Expand directly to the ultimate type if it's the right size.
4790 Ty = OpTy;
4791 // This is the type to do integer arithmetic in.
4792 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4793
4794 // Build up a list of operands to add together to form the full base.
4795 SmallVector<const SCEV *, 8> Ops;
4796
4797 // Expand the BaseRegs portion.
4798 for (const SCEV *Reg : F.BaseRegs) {
4799 assert(!Reg->isZero() && "Zero allocated in a base register!")((!Reg->isZero() && "Zero allocated in a base register!"
) ? static_cast<void> (0) : __assert_fail ("!Reg->isZero() && \"Zero allocated in a base register!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4799, __PRETTY_FUNCTION__))
;
4800
4801 // If we're expanding for a post-inc user, make the post-inc adjustment.
4802 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4803 Reg = TransformForPostIncUse(Denormalize, Reg,
4804 LF.UserInst, LF.OperandValToReplace,
4805 Loops, SE, DT);
4806
4807 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4808 }
4809
4810 // Expand the ScaledReg portion.
4811 Value *ICmpScaledV = nullptr;
4812 if (F.Scale != 0) {
4813 const SCEV *ScaledS = F.ScaledReg;
4814
4815 // If we're expanding for a post-inc user, make the post-inc adjustment.
4816 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4817 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4818 LF.UserInst, LF.OperandValToReplace,
4819 Loops, SE, DT);
4820
4821 if (LU.Kind == LSRUse::ICmpZero) {
4822 // Expand ScaleReg as if it was part of the base regs.
4823 if (F.Scale == 1)
4824 Ops.push_back(
4825 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4826 else {
4827 // An interesting way of "folding" with an icmp is to use a negated
4828 // scale, which we'll implement by inserting it into the other operand
4829 // of the icmp.
4830 assert(F.Scale == -1 &&((F.Scale == -1 && "The only scale supported by ICmpZero uses is -1!"
) ? static_cast<void> (0) : __assert_fail ("F.Scale == -1 && \"The only scale supported by ICmpZero uses is -1!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4831, __PRETTY_FUNCTION__))
4831 "The only scale supported by ICmpZero uses is -1!")((F.Scale == -1 && "The only scale supported by ICmpZero uses is -1!"
) ? static_cast<void> (0) : __assert_fail ("F.Scale == -1 && \"The only scale supported by ICmpZero uses is -1!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4831, __PRETTY_FUNCTION__))
;
4832 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4833 }
4834 } else {
4835 // Otherwise just expand the scaled register and an explicit scale,
4836 // which is expected to be matched as part of the address.
4837
4838 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4839 // Unless the addressing mode will not be folded.
4840 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4841 isAMCompletelyFolded(TTI, LU, F)) {
4842 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4843 Ops.clear();
4844 Ops.push_back(SE.getUnknown(FullV));
4845 }
4846 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
4847 if (F.Scale != 1)
4848 ScaledS =
4849 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4850 Ops.push_back(ScaledS);
4851 }
4852 }
4853
4854 // Expand the GV portion.
4855 if (F.BaseGV) {
4856 // Flush the operand list to suppress SCEVExpander hoisting.
4857 if (!Ops.empty()) {
4858 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4859 Ops.clear();
4860 Ops.push_back(SE.getUnknown(FullV));
4861 }
4862 Ops.push_back(SE.getUnknown(F.BaseGV));
4863 }
4864
4865 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4866 // unfolded offsets. LSR assumes they both live next to their uses.
4867 if (!Ops.empty()) {
4868 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4869 Ops.clear();
4870 Ops.push_back(SE.getUnknown(FullV));
4871 }
4872
4873 // Expand the immediate portion.
4874 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4875 if (Offset != 0) {
4876 if (LU.Kind == LSRUse::ICmpZero) {
4877 // The other interesting way of "folding" with an ICmpZero is to use a
4878 // negated immediate.
4879 if (!ICmpScaledV)
4880 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4881 else {
4882 Ops.push_back(SE.getUnknown(ICmpScaledV));
4883 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4884 }
4885 } else {
4886 // Just add the immediate values. These again are expected to be matched
4887 // as part of the address.
4888 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4889 }
4890 }
4891
4892 // Expand the unfolded offset portion.
4893 int64_t UnfoldedOffset = F.UnfoldedOffset;
4894 if (UnfoldedOffset != 0) {
4895 // Just add the immediate values.
4896 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4897 UnfoldedOffset)));
4898 }
4899
4900 // Emit instructions summing all the operands.
4901 const SCEV *FullS = Ops.empty() ?
4902 SE.getConstant(IntTy, 0) :
4903 SE.getAddExpr(Ops);
4904 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
4905
4906 // We're done expanding now, so reset the rewriter.
4907 Rewriter.clearPostInc();
4908
4909 // An ICmpZero Formula represents an ICmp which we're handling as a
4910 // comparison against zero. Now that we've expanded an expression for that
4911 // form, update the ICmp's other operand.
4912 if (LU.Kind == LSRUse::ICmpZero) {
4913 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4914 DeadInsts.emplace_back(CI->getOperand(1));
4915 assert(!F.BaseGV && "ICmp does not support folding a global value and "((!F.BaseGV && "ICmp does not support folding a global value and "
"a scale at the same time!") ? static_cast<void> (0) :
__assert_fail ("!F.BaseGV && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4916, __PRETTY_FUNCTION__))
4916 "a scale at the same time!")((!F.BaseGV && "ICmp does not support folding a global value and "
"a scale at the same time!") ? static_cast<void> (0) :
__assert_fail ("!F.BaseGV && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4916, __PRETTY_FUNCTION__))
;
4917 if (F.Scale == -1) {
4918 if (ICmpScaledV->getType() != OpTy) {
4919 Instruction *Cast =
4920 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4921 OpTy, false),
4922 ICmpScaledV, OpTy, "tmp", CI);
4923 ICmpScaledV = Cast;
4924 }
4925 CI->setOperand(1, ICmpScaledV);
4926 } else {
4927 // A scale of 1 means that the scale has been expanded as part of the
4928 // base regs.
4929 assert((F.Scale == 0 || F.Scale == 1) &&(((F.Scale == 0 || F.Scale == 1) && "ICmp does not support folding a global value and "
"a scale at the same time!") ? static_cast<void> (0) :
__assert_fail ("(F.Scale == 0 || F.Scale == 1) && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4931, __PRETTY_FUNCTION__))
4930 "ICmp does not support folding a global value and "(((F.Scale == 0 || F.Scale == 1) && "ICmp does not support folding a global value and "
"a scale at the same time!") ? static_cast<void> (0) :
__assert_fail ("(F.Scale == 0 || F.Scale == 1) && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4931, __PRETTY_FUNCTION__))
4931 "a scale at the same time!")(((F.Scale == 0 || F.Scale == 1) && "ICmp does not support folding a global value and "
"a scale at the same time!") ? static_cast<void> (0) :
__assert_fail ("(F.Scale == 0 || F.Scale == 1) && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4931, __PRETTY_FUNCTION__))
;
4932 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4933 -(uint64_t)Offset);
4934 if (C->getType() != OpTy)
4935 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4936 OpTy, false),
4937 C, OpTy);
4938
4939 CI->setOperand(1, C);
4940 }
4941 }
4942
4943 return FullV;
4944}
4945
4946/// Helper for Rewrite. PHI nodes are special because the use of their operands
4947/// effectively happens in their predecessor blocks, so the expression may need
4948/// to be expanded in multiple places.
4949void LSRInstance::RewriteForPHI(PHINode *PN,
4950 const LSRUse &LU,
4951 const LSRFixup &LF,
4952 const Formula &F,
4953 SCEVExpander &Rewriter,
4954 SmallVectorImpl<WeakVH> &DeadInsts) const {
4955 DenseMap<BasicBlock *, Value *> Inserted;
4956 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4957 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4958 BasicBlock *BB = PN->getIncomingBlock(i);
4959
4960 // If this is a critical edge, split the edge so that we do not insert
4961 // the code on all predecessor/successor paths. We do this unless this
4962 // is the canonical backedge for this loop, which complicates post-inc
4963 // users.
4964 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4965 !isa<IndirectBrInst>(BB->getTerminator()) &&
4966 !isa<CatchSwitchInst>(BB->getTerminator())) {
4967 BasicBlock *Parent = PN->getParent();
4968 Loop *PNLoop = LI.getLoopFor(Parent);
4969 if (!PNLoop || Parent != PNLoop->getHeader()) {
4970 // Split the critical edge.
4971 BasicBlock *NewBB = nullptr;
4972 if (!Parent->isLandingPad()) {
4973 NewBB = SplitCriticalEdge(BB, Parent,
4974 CriticalEdgeSplittingOptions(&DT, &LI)
4975 .setMergeIdenticalEdges()
4976 .setDontDeleteUselessPHIs());
4977 } else {
4978 SmallVector<BasicBlock*, 2> NewBBs;
4979 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
4980 NewBB = NewBBs[0];
4981 }
4982 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4983 // phi predecessors are identical. The simple thing to do is skip
4984 // splitting in this case rather than complicate the API.
4985 if (NewBB) {
4986 // If PN is outside of the loop and BB is in the loop, we want to
4987 // move the block to be immediately before the PHI block, not
4988 // immediately after BB.
4989 if (L->contains(BB) && !L->contains(PN))
4990 NewBB->moveBefore(PN->getParent());
4991
4992 // Splitting the edge can reduce the number of PHI entries we have.
4993 e = PN->getNumIncomingValues();
4994 BB = NewBB;
4995 i = PN->getBasicBlockIndex(BB);
4996 }
4997 }
4998 }
4999
5000 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5001 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5002 if (!Pair.second)
5003 PN->setIncomingValue(i, Pair.first->second);
5004 else {
5005 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5006 Rewriter, DeadInsts);
5007
5008 // If this is reuse-by-noop-cast, insert the noop cast.
5009 Type *OpTy = LF.OperandValToReplace->getType();
5010 if (FullV->getType() != OpTy)
5011 FullV =
5012 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5013 OpTy, false),
5014 FullV, LF.OperandValToReplace->getType(),
5015 "tmp", BB->getTerminator());
5016
5017 PN->setIncomingValue(i, FullV);
5018 Pair.first->second = FullV;
5019 }
5020 }
5021}
5022
5023/// Emit instructions for the leading candidate expression for this LSRUse (this
5024/// is called "expanding"), and update the UserInst to reference the newly
5025/// expanded value.
5026void LSRInstance::Rewrite(const LSRUse &LU,
5027 const LSRFixup &LF,
5028 const Formula &F,
5029 SCEVExpander &Rewriter,
5030 SmallVectorImpl<WeakVH> &DeadInsts) const {
5031 // First, find an insertion point that dominates UserInst. For PHI nodes,
5032 // find the nearest block which dominates all the relevant uses.
5033 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5034 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5035 } else {
5036 Value *FullV =
5037 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5038
5039 // If this is reuse-by-noop-cast, insert the noop cast.
5040 Type *OpTy = LF.OperandValToReplace->getType();
5041 if (FullV->getType() != OpTy) {
5042 Instruction *Cast =
5043 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5044 FullV, OpTy, "tmp", LF.UserInst);
5045 FullV = Cast;
5046 }
5047
5048 // Update the user. ICmpZero is handled specially here (for now) because
5049 // Expand may have updated one of the operands of the icmp already, and
5050 // its new value may happen to be equal to LF.OperandValToReplace, in
5051 // which case doing replaceUsesOfWith leads to replacing both operands
5052 // with the same value. TODO: Reorganize this.
5053 if (LU.Kind == LSRUse::ICmpZero)
5054 LF.UserInst->setOperand(0, FullV);
5055 else
5056 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5057 }
5058
5059 DeadInsts.emplace_back(LF.OperandValToReplace);
5060}
5061
5062/// Rewrite all the fixup locations with new values, following the chosen
5063/// solution.
5064void LSRInstance::ImplementSolution(
5065 const SmallVectorImpl<const Formula *> &Solution) {
5066 // Keep track of instructions we may have made dead, so that
5067 // we can remove them after we are done working.
5068 SmallVector<WeakVH, 16> DeadInsts;
5069
5070 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5071 "lsr");
5072#ifndef NDEBUG
5073 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5074#endif
5075 Rewriter.disableCanonicalMode();
5076 Rewriter.enableLSRMode();
5077 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5078
5079 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5080 for (const IVChain &Chain : IVChainVec) {
5081 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5082 Rewriter.setChainedPhi(PN);
5083 }
5084
5085 // Expand the new value definitions and update the users.
5086 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5087 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5088 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5089 Changed = true;
5090 }
5091
5092 for (const IVChain &Chain : IVChainVec) {
5093 GenerateIVChain(Chain, Rewriter, DeadInsts);
5094 Changed = true;
5095 }
5096 // Clean up after ourselves. This must be done before deleting any
5097 // instructions.
5098 Rewriter.clear();
5099
5100 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5101}
5102
5103LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5104 DominatorTree &DT, LoopInfo &LI,
5105 const TargetTransformInfo &TTI)
5106 : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L), Changed(false),
5107 IVIncInsertPos(nullptr) {
5108 // If LoopSimplify form is not available, stay out of trouble.
5109 if (!L->isLoopSimplifyForm())
5110 return;
5111
5112 // If there's no interesting work to be done, bail early.
5113 if (IU.empty()) return;
5114
5115 // If there's too much analysis to be done, bail early. We won't be able to
5116 // model the problem anyway.
5117 unsigned NumUsers = 0;
5118 for (const IVStrideUse &U : IU) {
5119 if (++NumUsers > MaxIVUsers) {
5120 (void)U;
5121 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5122 return;
5123 }
5124 // Bail out if we have a PHI on an EHPad that gets a value from a
5125 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5126 // no good place to stick any instructions.
5127 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5128 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5129 if (isa<FuncletPadInst>(FirstNonPHI) ||
5130 isa<CatchSwitchInst>(FirstNonPHI))
5131 for (BasicBlock *PredBB : PN->blocks())
5132 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5133 return;
5134 }
5135 }
5136
5137#ifndef NDEBUG
5138 // All dominating loops must have preheaders, or SCEVExpander may not be able
5139 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5140 //
5141 // IVUsers analysis should only create users that are dominated by simple loop
5142 // headers. Since this loop should dominate all of its users, its user list
5143 // should be empty if this loop itself is not within a simple loop nest.
5144 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5145 Rung; Rung = Rung->getIDom()) {
5146 BasicBlock *BB = Rung->getBlock();
5147 const Loop *DomLoop = LI.getLoopFor(BB);
5148 if (DomLoop && DomLoop->getHeader() == BB) {
5149 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest")((DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"
) ? static_cast<void> (0) : __assert_fail ("DomLoop->getLoopPreheader() && \"LSR needs a simplified loop nest\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5149, __PRETTY_FUNCTION__))
;
5150 }
5151 }
5152#endif // DEBUG
5153
5154 DEBUG(dbgs() << "\nLSR on loop ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
5155 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
5156 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5157
5158 // First, perform some low-level loop optimizations.
5159 OptimizeShadowIV();
5160 OptimizeLoopTermCond();
5161
5162 // If loop preparation eliminates all interesting IV users, bail.
5163 if (IU.empty()) return;
5164
5165 // Skip nested loops until we can model them better with formulae.
5166 if (!L->empty()) {
5167 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping outer loop "
<< *L << "\n"; } } while (false)
;
5168 return;
5169 }
5170
5171 // Start collecting data and preparing for the solver.
5172 CollectChains();
5173 CollectInterestingTypesAndFactors();
5174 CollectFixupsAndInitialFormulae();
5175 CollectLoopInvariantFixupsAndFormulae();
5176
5177 assert(!Uses.empty() && "IVUsers reported at least one use")((!Uses.empty() && "IVUsers reported at least one use"
) ? static_cast<void> (0) : __assert_fail ("!Uses.empty() && \"IVUsers reported at least one use\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5177, __PRETTY_FUNCTION__))
;
5178 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
5179 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
;
5180
5181 // Now use the reuse data to generate a bunch of interesting ways
5182 // to formulate the values needed for the uses.
5183 GenerateAllReuseFormulae();
5184
5185 FilterOutUndesirableDedicatedRegisters();
5186 NarrowSearchSpaceUsingHeuristics();
5187
5188 SmallVector<const Formula *, 8> Solution;
5189 Solve(Solution);
5190
5191 // Release memory that is no longer needed.
5192 Factors.clear();
5193 Types.clear();
5194 RegUses.clear();
5195
5196 if (Solution.empty())
5197 return;
5198
5199#ifndef NDEBUG
5200 // Formulae should be legal.
5201 for (const LSRUse &LU : Uses) {
5202 for (const Formula &F : LU.Formulae)
5203 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,((isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy
, F) && "Illegal formula generated!") ? static_cast<
void> (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Illegal formula generated!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5204, __PRETTY_FUNCTION__))
5204 F) && "Illegal formula generated!")((isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy
, F) && "Illegal formula generated!") ? static_cast<
void> (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Illegal formula generated!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn296300/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5204, __PRETTY_FUNCTION__))
;
5205 };
5206#endif
5207
5208 // Now that we've decided what we want, make it so.
5209 ImplementSolution(Solution);
5210}
5211
5212void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5213 if (Factors.empty() && Types.empty()) return;
5214
5215 OS << "LSR has identified the following interesting factors and types: ";
5216 bool First = true;
5217
5218 for (int64_t Factor : Factors) {
5219 if (!First) OS << ", ";
5220 First = false;
5221 OS << '*' << Factor;
5222 }
5223
5224 for (Type *Ty : Types) {
5225 if (!First) OS << ", ";
5226 First = false;
5227 OS << '(' << *Ty << ')';
5228 }
5229 OS << '\n';
5230}
5231
5232void LSRInstance::print_fixups(raw_ostream &OS) const {
5233 OS << "LSR is examining the following fixup sites:\n";
5234 for (const LSRUse &LU : Uses)
5235 for (const LSRFixup &LF : LU.Fixups) {
5236 dbgs() << " ";
5237 LF.print(OS);
5238 OS << '\n';
5239 }
5240}
5241
5242void LSRInstance::print_uses(raw_ostream &OS) const {
5243 OS << "LSR is examining the following uses:\n";
5244 for (const LSRUse &LU : Uses) {
5245 dbgs() << " ";
5246 LU.print(OS);
5247 OS << '\n';
5248 for (const Formula &F : LU.Formulae) {
5249 OS << " ";
5250 F.print(OS);
5251 OS << '\n';
5252 }
5253 }
5254}
5255
5256void LSRInstance::print(raw_ostream &OS) const {
5257 print_factors_and_types(OS);
5258 print_fixups(OS);
5259 print_uses(OS);
5260}
5261
5262#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5263LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRInstance::dump() const {
5264 print(errs()); errs() << '\n';
5265}
5266#endif
5267
5268namespace {
5269
5270class LoopStrengthReduce : public LoopPass {
5271public:
5272 static char ID; // Pass ID, replacement for typeid
5273
5274 LoopStrengthReduce();
5275
5276private:
5277 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5278 void getAnalysisUsage(AnalysisUsage &AU) const override;
5279};
5280
5281} // end anonymous namespace
5282
5283LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5284 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5285}
5286
5287void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5288 // We split critical edges, so we change the CFG. However, we do update
5289 // many analyses if they are around.
5290 AU.addPreservedID(LoopSimplifyID);
5291
5292 AU.addRequired<LoopInfoWrapperPass>();
5293 AU.addPreserved<LoopInfoWrapperPass>();
5294 AU.addRequiredID(LoopSimplifyID);
5295 AU.addRequired<DominatorTreeWrapperPass>();
5296 AU.addPreserved<DominatorTreeWrapperPass>();
5297 AU.addRequired<ScalarEvolutionWrapperPass>();
5298 AU.addPreserved<ScalarEvolutionWrapperPass>();
5299 // Requiring LoopSimplify a second time here prevents IVUsers from running
5300 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5301 AU.addRequiredID(LoopSimplifyID);
5302 AU.addRequired<IVUsersWrapperPass>();
5303 AU.addPreserved<IVUsersWrapperPass>();
5304 AU.addRequired<TargetTransformInfoWrapperPass>();
5305}
5306
5307static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5308 DominatorTree &DT, LoopInfo &LI,
5309 const TargetTransformInfo &TTI) {
5310 bool Changed = false;
5311
5312 // Run the main LSR transformation.
5313 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5314
5315 // Remove any extra phis created by processing inner loops.
5316 Changed |= DeleteDeadPHIs(L->getHeader());
5317 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5318 SmallVector<WeakVH, 16> DeadInsts;
5319 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5320 SCEVExpander Rewriter(SE, DL, "lsr");
5321#ifndef NDEBUG
5322 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5323#endif
5324 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5325 if (numFolded) {
5326 Changed = true;
5327 DeleteTriviallyDeadInstructions(DeadInsts);
5328 DeleteDeadPHIs(L->getHeader());
5329 }
5330 }
5331 return Changed;
5332}
5333
5334bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5335 if (skipLoop(L))
5336 return false;
5337
5338 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5339 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5340 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5341 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5342 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5343 *L->getHeader()->getParent());
5344 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5345}
5346
5347PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5348 LoopStandardAnalysisResults &AR,
5349 LPMUpdater &) {
5350 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5351 AR.DT, AR.LI, AR.TTI))
5352 return PreservedAnalyses::all();
5353
5354 return getLoopPassPreservedAnalyses();
5355}
5356
5357char LoopStrengthReduce::ID = 0;
5358INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
5359 "Loop Strength Reduction", false, false)static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
5360INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
5361INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
5362INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
5363INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)initializeIVUsersWrapperPassPass(Registry);
5364INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
5365INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
5366INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",PassInfo *PI = new PassInfo( "Loop Strength Reduction", "loop-reduce"
, &LoopStrengthReduce::ID, PassInfo::NormalCtor_t(callDefaultCtor
<LoopStrengthReduce>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeLoopStrengthReducePassFlag
; void llvm::initializeLoopStrengthReducePass(PassRegistry &
Registry) { llvm::call_once(InitializeLoopStrengthReducePassFlag
, initializeLoopStrengthReducePassOnce, std::ref(Registry)); }
5367 "Loop Strength Reduction", false, false)PassInfo *PI = new PassInfo( "Loop Strength Reduction", "loop-reduce"
, &LoopStrengthReduce::ID, PassInfo::NormalCtor_t(callDefaultCtor
<LoopStrengthReduce>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeLoopStrengthReducePassFlag
; void llvm::initializeLoopStrengthReducePass(PassRegistry &
Registry) { llvm::call_once(InitializeLoopStrengthReducePassFlag
, initializeLoopStrengthReducePassOnce, std::ref(Registry)); }
5368
5369Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }