Bug Summary

File:lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 3145, column 16
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name LoopStrengthReduce.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn350071/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-8~svn350071=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-12-27-042839-1215-1 -x c++ /build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp -faddrsig
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/ADT/iterator_range.h"
69#include "llvm/Analysis/IVUsers.h"
70#include "llvm/Analysis/LoopAnalysisManager.h"
71#include "llvm/Analysis/LoopInfo.h"
72#include "llvm/Analysis/LoopPass.h"
73#include "llvm/Analysis/ScalarEvolution.h"
74#include "llvm/Analysis/ScalarEvolutionExpander.h"
75#include "llvm/Analysis/ScalarEvolutionExpressions.h"
76#include "llvm/Analysis/ScalarEvolutionNormalization.h"
77#include "llvm/Analysis/TargetTransformInfo.h"
78#include "llvm/Transforms/Utils/Local.h"
79#include "llvm/Config/llvm-config.h"
80#include "llvm/IR/BasicBlock.h"
81#include "llvm/IR/Constant.h"
82#include "llvm/IR/Constants.h"
83#include "llvm/IR/DerivedTypes.h"
84#include "llvm/IR/Dominators.h"
85#include "llvm/IR/GlobalValue.h"
86#include "llvm/IR/IRBuilder.h"
87#include "llvm/IR/InstrTypes.h"
88#include "llvm/IR/Instruction.h"
89#include "llvm/IR/Instructions.h"
90#include "llvm/IR/IntrinsicInst.h"
91#include "llvm/IR/Intrinsics.h"
92#include "llvm/IR/Module.h"
93#include "llvm/IR/OperandTraits.h"
94#include "llvm/IR/Operator.h"
95#include "llvm/IR/PassManager.h"
96#include "llvm/IR/Type.h"
97#include "llvm/IR/Use.h"
98#include "llvm/IR/User.h"
99#include "llvm/IR/Value.h"
100#include "llvm/IR/ValueHandle.h"
101#include "llvm/Pass.h"
102#include "llvm/Support/Casting.h"
103#include "llvm/Support/CommandLine.h"
104#include "llvm/Support/Compiler.h"
105#include "llvm/Support/Debug.h"
106#include "llvm/Support/ErrorHandling.h"
107#include "llvm/Support/MathExtras.h"
108#include "llvm/Support/raw_ostream.h"
109#include "llvm/Transforms/Scalar.h"
110#include "llvm/Transforms/Utils.h"
111#include "llvm/Transforms/Utils/BasicBlockUtils.h"
112#include <algorithm>
113#include <cassert>
114#include <cstddef>
115#include <cstdint>
116#include <cstdlib>
117#include <iterator>
118#include <limits>
119#include <map>
120#include <utility>
121
122using namespace llvm;
123
124#define DEBUG_TYPE"loop-reduce" "loop-reduce"
125
126/// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
127/// bail out. This threshold is far beyond the number of users that LSR can
128/// conceivably solve, so it should not affect generated code, but catches the
129/// worst cases before LSR burns too much compile time and stack space.
130static const unsigned MaxIVUsers = 200;
131
132// Temporary flag to cleanup congruent phis after LSR phi expansion.
133// It's currently disabled until we can determine whether it's truly useful or
134// not. The flag should be removed after the v3.0 release.
135// This is now needed for ivchains.
136static cl::opt<bool> EnablePhiElim(
137 "enable-lsr-phielim", cl::Hidden, cl::init(true),
138 cl::desc("Enable LSR phi elimination"));
139
140// The flag adds instruction count to solutions cost comparision.
141static cl::opt<bool> InsnsCost(
142 "lsr-insns-cost", cl::Hidden, cl::init(true),
143 cl::desc("Add instruction count to a LSR cost model"));
144
145// Flag to choose how to narrow complex lsr solution
146static cl::opt<bool> LSRExpNarrow(
147 "lsr-exp-narrow", cl::Hidden, cl::init(false),
148 cl::desc("Narrow LSR complex solution using"
149 " expectation of registers number"));
150
151// Flag to narrow search space by filtering non-optimal formulae with
152// the same ScaledReg and Scale.
153static cl::opt<bool> FilterSameScaledReg(
154 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
155 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
156 " with the same ScaledReg and Scale"));
157
158static cl::opt<unsigned> ComplexityLimit(
159 "lsr-complexity-limit", cl::Hidden,
160 cl::init(std::numeric_limits<uint16_t>::max()),
161 cl::desc("LSR search space complexity limit"));
162
163#ifndef NDEBUG
164// Stress test IV chain generation.
165static cl::opt<bool> StressIVChain(
166 "stress-ivchain", cl::Hidden, cl::init(false),
167 cl::desc("Stress test LSR IV chains"));
168#else
169static bool StressIVChain = false;
170#endif
171
172namespace {
173
174struct MemAccessTy {
175 /// Used in situations where the accessed memory type is unknown.
176 static const unsigned UnknownAddressSpace =
177 std::numeric_limits<unsigned>::max();
178
179 Type *MemTy = nullptr;
180 unsigned AddrSpace = UnknownAddressSpace;
181
182 MemAccessTy() = default;
183 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
184
185 bool operator==(MemAccessTy Other) const {
186 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
187 }
188
189 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
190
191 static MemAccessTy getUnknown(LLVMContext &Ctx,
192 unsigned AS = UnknownAddressSpace) {
193 return MemAccessTy(Type::getVoidTy(Ctx), AS);
194 }
195
196 Type *getType() { return MemTy; }
197};
198
199/// This class holds data which is used to order reuse candidates.
200class RegSortData {
201public:
202 /// This represents the set of LSRUse indices which reference
203 /// a particular register.
204 SmallBitVector UsedByIndices;
205
206 void print(raw_ostream &OS) const;
207 void dump() const;
208};
209
210} // end anonymous namespace
211
212#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
213void RegSortData::print(raw_ostream &OS) const {
214 OS << "[NumUses=" << UsedByIndices.count() << ']';
215}
216
217LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
218 print(errs()); errs() << '\n';
219}
220#endif
221
222namespace {
223
224/// Map register candidates to information about how they are used.
225class RegUseTracker {
226 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
227
228 RegUsesTy RegUsesMap;
229 SmallVector<const SCEV *, 16> RegSequence;
230
231public:
232 void countRegister(const SCEV *Reg, size_t LUIdx);
233 void dropRegister(const SCEV *Reg, size_t LUIdx);
234 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
235
236 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
237
238 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
239
240 void clear();
241
242 using iterator = SmallVectorImpl<const SCEV *>::iterator;
243 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
244
245 iterator begin() { return RegSequence.begin(); }
246 iterator end() { return RegSequence.end(); }
247 const_iterator begin() const { return RegSequence.begin(); }
248 const_iterator end() const { return RegSequence.end(); }
249};
250
251} // end anonymous namespace
252
253void
254RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
255 std::pair<RegUsesTy::iterator, bool> Pair =
256 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
257 RegSortData &RSD = Pair.first->second;
258 if (Pair.second)
259 RegSequence.push_back(Reg);
260 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
261 RSD.UsedByIndices.set(LUIdx);
262}
263
264void
265RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
266 RegUsesTy::iterator It = RegUsesMap.find(Reg);
267 assert(It != RegUsesMap.end())((It != RegUsesMap.end()) ? static_cast<void> (0) : __assert_fail
("It != RegUsesMap.end()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 267, __PRETTY_FUNCTION__))
;
268 RegSortData &RSD = It->second;
269 assert(RSD.UsedByIndices.size() > LUIdx)((RSD.UsedByIndices.size() > LUIdx) ? static_cast<void>
(0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 269, __PRETTY_FUNCTION__))
;
270 RSD.UsedByIndices.reset(LUIdx);
271}
272
273void
274RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
275 assert(LUIdx <= LastLUIdx)((LUIdx <= LastLUIdx) ? static_cast<void> (0) : __assert_fail
("LUIdx <= LastLUIdx", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 275, __PRETTY_FUNCTION__))
;
276
277 // Update RegUses. The data structure is not optimized for this purpose;
278 // we must iterate through it and update each of the bit vectors.
279 for (auto &Pair : RegUsesMap) {
280 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
281 if (LUIdx < UsedByIndices.size())
282 UsedByIndices[LUIdx] =
283 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
284 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
285 }
286}
287
288bool
289RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
290 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
291 if (I == RegUsesMap.end())
292 return false;
293 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
294 int i = UsedByIndices.find_first();
295 if (i == -1) return false;
296 if ((size_t)i != LUIdx) return true;
297 return UsedByIndices.find_next(i) != -1;
298}
299
300const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
301 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
302 assert(I != RegUsesMap.end() && "Unknown register!")((I != RegUsesMap.end() && "Unknown register!") ? static_cast
<void> (0) : __assert_fail ("I != RegUsesMap.end() && \"Unknown register!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 302, __PRETTY_FUNCTION__))
;
303 return I->second.UsedByIndices;
304}
305
306void RegUseTracker::clear() {
307 RegUsesMap.clear();
308 RegSequence.clear();
309}
310
311namespace {
312
313/// This class holds information that describes a formula for computing
314/// satisfying a use. It may include broken-out immediates and scaled registers.
315struct Formula {
316 /// Global base address used for complex addressing.
317 GlobalValue *BaseGV = nullptr;
318
319 /// Base offset for complex addressing.
320 int64_t BaseOffset = 0;
321
322 /// Whether any complex addressing has a base register.
323 bool HasBaseReg = false;
324
325 /// The scale of any complex addressing.
326 int64_t Scale = 0;
327
328 /// The list of "base" registers for this use. When this is non-empty. The
329 /// canonical representation of a formula is
330 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
331 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
332 /// 3. The reg containing recurrent expr related with currect loop in the
333 /// formula should be put in the ScaledReg.
334 /// #1 enforces that the scaled register is always used when at least two
335 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
336 /// #2 enforces that 1 * reg is reg.
337 /// #3 ensures invariant regs with respect to current loop can be combined
338 /// together in LSR codegen.
339 /// This invariant can be temporarily broken while building a formula.
340 /// However, every formula inserted into the LSRInstance must be in canonical
341 /// form.
342 SmallVector<const SCEV *, 4> BaseRegs;
343
344 /// The 'scaled' register for this use. This should be non-null when Scale is
345 /// not zero.
346 const SCEV *ScaledReg = nullptr;
347
348 /// An additional constant offset which added near the use. This requires a
349 /// temporary register, but the offset itself can live in an add immediate
350 /// field rather than a register.
351 int64_t UnfoldedOffset = 0;
352
353 Formula() = default;
354
355 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
356
357 bool isCanonical(const Loop &L) const;
358
359 void canonicalize(const Loop &L);
360
361 bool unscale();
362
363 bool hasZeroEnd() const;
364
365 size_t getNumRegs() const;
366 Type *getType() const;
367
368 void deleteBaseReg(const SCEV *&S);
369
370 bool referencesReg(const SCEV *S) const;
371 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
372 const RegUseTracker &RegUses) const;
373
374 void print(raw_ostream &OS) const;
375 void dump() const;
376};
377
378} // end anonymous namespace
379
380/// Recursion helper for initialMatch.
381static void DoInitialMatch(const SCEV *S, Loop *L,
382 SmallVectorImpl<const SCEV *> &Good,
383 SmallVectorImpl<const SCEV *> &Bad,
384 ScalarEvolution &SE) {
385 // Collect expressions which properly dominate the loop header.
386 if (SE.properlyDominates(S, L->getHeader())) {
387 Good.push_back(S);
388 return;
389 }
390
391 // Look at add operands.
392 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
393 for (const SCEV *S : Add->operands())
394 DoInitialMatch(S, L, Good, Bad, SE);
395 return;
396 }
397
398 // Look at addrec operands.
399 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
400 if (!AR->getStart()->isZero() && AR->isAffine()) {
401 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
402 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
403 AR->getStepRecurrence(SE),
404 // FIXME: AR->getNoWrapFlags()
405 AR->getLoop(), SCEV::FlagAnyWrap),
406 L, Good, Bad, SE);
407 return;
408 }
409
410 // Handle a multiplication by -1 (negation) if it didn't fold.
411 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
412 if (Mul->getOperand(0)->isAllOnesValue()) {
413 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
414 const SCEV *NewMul = SE.getMulExpr(Ops);
415
416 SmallVector<const SCEV *, 4> MyGood;
417 SmallVector<const SCEV *, 4> MyBad;
418 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
419 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
420 SE.getEffectiveSCEVType(NewMul->getType())));
421 for (const SCEV *S : MyGood)
422 Good.push_back(SE.getMulExpr(NegOne, S));
423 for (const SCEV *S : MyBad)
424 Bad.push_back(SE.getMulExpr(NegOne, S));
425 return;
426 }
427
428 // Ok, we can't do anything interesting. Just stuff the whole thing into a
429 // register and hope for the best.
430 Bad.push_back(S);
431}
432
433/// Incorporate loop-variant parts of S into this Formula, attempting to keep
434/// all loop-invariant and loop-computable values in a single base register.
435void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
436 SmallVector<const SCEV *, 4> Good;
437 SmallVector<const SCEV *, 4> Bad;
438 DoInitialMatch(S, L, Good, Bad, SE);
439 if (!Good.empty()) {
440 const SCEV *Sum = SE.getAddExpr(Good);
441 if (!Sum->isZero())
442 BaseRegs.push_back(Sum);
443 HasBaseReg = true;
444 }
445 if (!Bad.empty()) {
446 const SCEV *Sum = SE.getAddExpr(Bad);
447 if (!Sum->isZero())
448 BaseRegs.push_back(Sum);
449 HasBaseReg = true;
450 }
451 canonicalize(*L);
452}
453
454/// Check whether or not this formula satisfies the canonical
455/// representation.
456/// \see Formula::BaseRegs.
457bool Formula::isCanonical(const Loop &L) const {
458 if (!ScaledReg)
459 return BaseRegs.size() <= 1;
460
461 if (Scale != 1)
462 return true;
463
464 if (Scale == 1 && BaseRegs.empty())
465 return false;
466
467 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
468 if (SAR && SAR->getLoop() == &L)
469 return true;
470
471 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
472 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
473 // loop, we want to swap the reg in BaseRegs with ScaledReg.
474 auto I =
475 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
476 return isa<const SCEVAddRecExpr>(S) &&
477 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
478 });
479 return I == BaseRegs.end();
480}
481
482/// Helper method to morph a formula into its canonical representation.
483/// \see Formula::BaseRegs.
484/// Every formula having more than one base register, must use the ScaledReg
485/// field. Otherwise, we would have to do special cases everywhere in LSR
486/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
487/// On the other hand, 1*reg should be canonicalized into reg.
488void Formula::canonicalize(const Loop &L) {
489 if (isCanonical(L))
490 return;
491 // So far we did not need this case. This is easy to implement but it is
492 // useless to maintain dead code. Beside it could hurt compile time.
493 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.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 493, __PRETTY_FUNCTION__))
;
494
495 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
496 if (!ScaledReg) {
497 ScaledReg = BaseRegs.back();
498 BaseRegs.pop_back();
499 Scale = 1;
500 }
501
502 // If ScaledReg is an invariant with respect to L, find the reg from
503 // BaseRegs containing the recurrent expr related with Loop L. Swap the
504 // reg with ScaledReg.
505 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
506 if (!SAR || SAR->getLoop() != &L) {
507 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
508 [&](const SCEV *S) {
509 return isa<const SCEVAddRecExpr>(S) &&
510 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
511 });
512 if (I != BaseRegs.end())
513 std::swap(ScaledReg, *I);
514 }
515}
516
517/// Get rid of the scale in the formula.
518/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
519/// \return true if it was possible to get rid of the scale, false otherwise.
520/// \note After this operation the formula may not be in the canonical form.
521bool Formula::unscale() {
522 if (Scale != 1)
523 return false;
524 Scale = 0;
525 BaseRegs.push_back(ScaledReg);
526 ScaledReg = nullptr;
527 return true;
528}
529
530bool Formula::hasZeroEnd() const {
531 if (UnfoldedOffset || BaseOffset)
532 return false;
533 if (BaseRegs.size() != 1 || ScaledReg)
534 return false;
535 return true;
536}
537
538/// Return the total number of register operands used by this formula. This does
539/// not include register uses implied by non-constant addrec strides.
540size_t Formula::getNumRegs() const {
541 return !!ScaledReg + BaseRegs.size();
542}
543
544/// Return the type of this formula, if it has one, or null otherwise. This type
545/// is meaningless except for the bit size.
546Type *Formula::getType() const {
547 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
548 ScaledReg ? ScaledReg->getType() :
549 BaseGV ? BaseGV->getType() :
550 nullptr;
551}
552
553/// Delete the given base reg from the BaseRegs list.
554void Formula::deleteBaseReg(const SCEV *&S) {
555 if (&S != &BaseRegs.back())
556 std::swap(S, BaseRegs.back());
557 BaseRegs.pop_back();
558}
559
560/// Test if this formula references the given register.
561bool Formula::referencesReg(const SCEV *S) const {
562 return S == ScaledReg || is_contained(BaseRegs, S);
563}
564
565/// Test whether this formula uses registers which are used by uses other than
566/// the use with the given index.
567bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
568 const RegUseTracker &RegUses) const {
569 if (ScaledReg)
570 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
571 return true;
572 for (const SCEV *BaseReg : BaseRegs)
573 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
574 return true;
575 return false;
576}
577
578#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
579void Formula::print(raw_ostream &OS) const {
580 bool First = true;
581 if (BaseGV) {
582 if (!First) OS << " + "; else First = false;
583 BaseGV->printAsOperand(OS, /*PrintType=*/false);
584 }
585 if (BaseOffset != 0) {
586 if (!First) OS << " + "; else First = false;
587 OS << BaseOffset;
588 }
589 for (const SCEV *BaseReg : BaseRegs) {
590 if (!First) OS << " + "; else First = false;
591 OS << "reg(" << *BaseReg << ')';
592 }
593 if (HasBaseReg && BaseRegs.empty()) {
594 if (!First) OS << " + "; else First = false;
595 OS << "**error: HasBaseReg**";
596 } else if (!HasBaseReg && !BaseRegs.empty()) {
597 if (!First) OS << " + "; else First = false;
598 OS << "**error: !HasBaseReg**";
599 }
600 if (Scale != 0) {
601 if (!First) OS << " + "; else First = false;
602 OS << Scale << "*reg(";
603 if (ScaledReg)
604 OS << *ScaledReg;
605 else
606 OS << "<unknown>";
607 OS << ')';
608 }
609 if (UnfoldedOffset != 0) {
610 if (!First) OS << " + ";
611 OS << "imm(" << UnfoldedOffset << ')';
612 }
613}
614
615LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
616 print(errs()); errs() << '\n';
617}
618#endif
619
620/// Return true if the given addrec can be sign-extended without changing its
621/// value.
622static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
623 Type *WideTy =
624 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
625 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
626}
627
628/// Return true if the given add can be sign-extended without changing its
629/// value.
630static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
631 Type *WideTy =
632 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
633 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
634}
635
636/// Return true if the given mul can be sign-extended without changing its
637/// value.
638static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
639 Type *WideTy =
640 IntegerType::get(SE.getContext(),
641 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
642 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
643}
644
645/// Return an expression for LHS /s RHS, if it can be determined and if the
646/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
647/// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
648/// the multiplication may overflow, which is useful when the result will be
649/// used in a context where the most significant bits are ignored.
650static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
651 ScalarEvolution &SE,
652 bool IgnoreSignificantBits = false) {
653 // Handle the trivial case, which works for any SCEV type.
654 if (LHS == RHS)
655 return SE.getConstant(LHS->getType(), 1);
656
657 // Handle a few RHS special cases.
658 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
659 if (RC) {
660 const APInt &RA = RC->getAPInt();
661 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
662 // some folding.
663 if (RA.isAllOnesValue())
664 return SE.getMulExpr(LHS, RC);
665 // Handle x /s 1 as x.
666 if (RA == 1)
667 return LHS;
668 }
669
670 // Check for a division of a constant by a constant.
671 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
672 if (!RC)
673 return nullptr;
674 const APInt &LA = C->getAPInt();
675 const APInt &RA = RC->getAPInt();
676 if (LA.srem(RA) != 0)
677 return nullptr;
678 return SE.getConstant(LA.sdiv(RA));
679 }
680
681 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
682 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
683 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
684 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
685 IgnoreSignificantBits);
686 if (!Step) return nullptr;
687 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
688 IgnoreSignificantBits);
689 if (!Start) return nullptr;
690 // FlagNW is independent of the start value, step direction, and is
691 // preserved with smaller magnitude steps.
692 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
693 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
694 }
695 return nullptr;
696 }
697
698 // Distribute the sdiv over add operands, if the add doesn't overflow.
699 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
700 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
701 SmallVector<const SCEV *, 8> Ops;
702 for (const SCEV *S : Add->operands()) {
703 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
704 if (!Op) return nullptr;
705 Ops.push_back(Op);
706 }
707 return SE.getAddExpr(Ops);
708 }
709 return nullptr;
710 }
711
712 // Check for a multiply operand that we can pull RHS out of.
713 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
714 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
715 SmallVector<const SCEV *, 4> Ops;
716 bool Found = false;
717 for (const SCEV *S : Mul->operands()) {
718 if (!Found)
719 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
720 IgnoreSignificantBits)) {
721 S = Q;
722 Found = true;
723 }
724 Ops.push_back(S);
725 }
726 return Found ? SE.getMulExpr(Ops) : nullptr;
727 }
728 return nullptr;
729 }
730
731 // Otherwise we don't know.
732 return nullptr;
733}
734
735/// If S involves the addition of a constant integer value, return that integer
736/// value, and mutate S to point to a new SCEV with that value excluded.
737static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
738 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
739 if (C->getAPInt().getMinSignedBits() <= 64) {
740 S = SE.getConstant(C->getType(), 0);
741 return C->getValue()->getSExtValue();
742 }
743 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
744 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
745 int64_t Result = ExtractImmediate(NewOps.front(), SE);
746 if (Result != 0)
747 S = SE.getAddExpr(NewOps);
748 return Result;
749 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
750 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
751 int64_t Result = ExtractImmediate(NewOps.front(), SE);
752 if (Result != 0)
753 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
754 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
755 SCEV::FlagAnyWrap);
756 return Result;
757 }
758 return 0;
759}
760
761/// If S involves the addition of a GlobalValue address, return that symbol, and
762/// mutate S to point to a new SCEV with that value excluded.
763static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
764 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
765 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
766 S = SE.getConstant(GV->getType(), 0);
767 return GV;
768 }
769 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
770 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
771 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
772 if (Result)
773 S = SE.getAddExpr(NewOps);
774 return Result;
775 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
776 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
777 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
778 if (Result)
779 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
780 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
781 SCEV::FlagAnyWrap);
782 return Result;
783 }
784 return nullptr;
785}
786
787/// Returns true if the specified instruction is using the specified value as an
788/// address.
789static bool isAddressUse(const TargetTransformInfo &TTI,
790 Instruction *Inst, Value *OperandVal) {
791 bool isAddress = isa<LoadInst>(Inst);
792 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
793 if (SI->getPointerOperand() == OperandVal)
794 isAddress = true;
795 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
796 // Addressing modes can also be folded into prefetches and a variety
797 // of intrinsics.
798 switch (II->getIntrinsicID()) {
799 case Intrinsic::memset:
800 case Intrinsic::prefetch:
801 if (II->getArgOperand(0) == OperandVal)
802 isAddress = true;
803 break;
804 case Intrinsic::memmove:
805 case Intrinsic::memcpy:
806 if (II->getArgOperand(0) == OperandVal ||
807 II->getArgOperand(1) == OperandVal)
808 isAddress = true;
809 break;
810 default: {
811 MemIntrinsicInfo IntrInfo;
812 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
813 if (IntrInfo.PtrVal == OperandVal)
814 isAddress = true;
815 }
816 }
817 }
818 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
819 if (RMW->getPointerOperand() == OperandVal)
820 isAddress = true;
821 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
822 if (CmpX->getPointerOperand() == OperandVal)
823 isAddress = true;
824 }
825 return isAddress;
826}
827
828/// Return the type of the memory being accessed.
829static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
830 Instruction *Inst, Value *OperandVal) {
831 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
832 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
833 AccessTy.MemTy = SI->getOperand(0)->getType();
834 AccessTy.AddrSpace = SI->getPointerAddressSpace();
835 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
836 AccessTy.AddrSpace = LI->getPointerAddressSpace();
837 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
838 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
839 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
840 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
841 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
842 switch (II->getIntrinsicID()) {
843 case Intrinsic::prefetch:
844 case Intrinsic::memset:
845 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
846 AccessTy.MemTy = OperandVal->getType();
847 break;
848 case Intrinsic::memmove:
849 case Intrinsic::memcpy:
850 AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
851 AccessTy.MemTy = OperandVal->getType();
852 break;
853 default: {
854 MemIntrinsicInfo IntrInfo;
855 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
856 AccessTy.AddrSpace
857 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
858 }
859
860 break;
861 }
862 }
863 }
864
865 // All pointers have the same requirements, so canonicalize them to an
866 // arbitrary pointer type to minimize variation.
867 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
868 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
869 PTy->getAddressSpace());
870
871 return AccessTy;
872}
873
874/// Return true if this AddRec is already a phi in its loop.
875static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
876 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
877 if (SE.isSCEVable(PN.getType()) &&
878 (SE.getEffectiveSCEVType(PN.getType()) ==
879 SE.getEffectiveSCEVType(AR->getType())) &&
880 SE.getSCEV(&PN) == AR)
881 return true;
882 }
883 return false;
884}
885
886/// Check if expanding this expression is likely to incur significant cost. This
887/// is tricky because SCEV doesn't track which expressions are actually computed
888/// by the current IR.
889///
890/// We currently allow expansion of IV increments that involve adds,
891/// multiplication by constants, and AddRecs from existing phis.
892///
893/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
894/// obvious multiple of the UDivExpr.
895static bool isHighCostExpansion(const SCEV *S,
896 SmallPtrSetImpl<const SCEV*> &Processed,
897 ScalarEvolution &SE) {
898 // Zero/One operand expressions
899 switch (S->getSCEVType()) {
900 case scUnknown:
901 case scConstant:
902 return false;
903 case scTruncate:
904 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
905 Processed, SE);
906 case scZeroExtend:
907 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
908 Processed, SE);
909 case scSignExtend:
910 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
911 Processed, SE);
912 }
913
914 if (!Processed.insert(S).second)
915 return false;
916
917 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
918 for (const SCEV *S : Add->operands()) {
919 if (isHighCostExpansion(S, Processed, SE))
920 return true;
921 }
922 return false;
923 }
924
925 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
926 if (Mul->getNumOperands() == 2) {
927 // Multiplication by a constant is ok
928 if (isa<SCEVConstant>(Mul->getOperand(0)))
929 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
930
931 // If we have the value of one operand, check if an existing
932 // multiplication already generates this expression.
933 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
934 Value *UVal = U->getValue();
935 for (User *UR : UVal->users()) {
936 // If U is a constant, it may be used by a ConstantExpr.
937 Instruction *UI = dyn_cast<Instruction>(UR);
938 if (UI && UI->getOpcode() == Instruction::Mul &&
939 SE.isSCEVable(UI->getType())) {
940 return SE.getSCEV(UI) == Mul;
941 }
942 }
943 }
944 }
945 }
946
947 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
948 if (isExistingPhi(AR, SE))
949 return false;
950 }
951
952 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
953 return true;
954}
955
956/// If any of the instructions in the specified set are trivially dead, delete
957/// them and see if this makes any of their operands subsequently dead.
958static bool
959DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
960 bool Changed = false;
961
962 while (!DeadInsts.empty()) {
963 Value *V = DeadInsts.pop_back_val();
964 Instruction *I = dyn_cast_or_null<Instruction>(V);
965
966 if (!I || !isInstructionTriviallyDead(I))
967 continue;
968
969 for (Use &O : I->operands())
970 if (Instruction *U = dyn_cast<Instruction>(O)) {
971 O = nullptr;
972 if (U->use_empty())
973 DeadInsts.emplace_back(U);
974 }
975
976 I->eraseFromParent();
977 Changed = true;
978 }
979
980 return Changed;
981}
982
983namespace {
984
985class LSRUse;
986
987} // end anonymous namespace
988
989/// Check if the addressing mode defined by \p F is completely
990/// folded in \p LU at isel time.
991/// This includes address-mode folding and special icmp tricks.
992/// This function returns true if \p LU can accommodate what \p F
993/// defines and up to 1 base + 1 scaled + offset.
994/// In other words, if \p F has several base registers, this function may
995/// still return true. Therefore, users still need to account for
996/// additional base registers and/or unfolded offsets to derive an
997/// accurate cost model.
998static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
999 const LSRUse &LU, const Formula &F);
1000
1001// Get the cost of the scaling factor used in F for LU.
1002static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1003 const LSRUse &LU, const Formula &F,
1004 const Loop &L);
1005
1006namespace {
1007
1008/// This class is used to measure and compare candidate formulae.
1009class Cost {
1010 TargetTransformInfo::LSRCost C;
1011
1012public:
1013 Cost() {
1014 C.Insns = 0;
1015 C.NumRegs = 0;
1016 C.AddRecCost = 0;
1017 C.NumIVMuls = 0;
1018 C.NumBaseAdds = 0;
1019 C.ImmCost = 0;
1020 C.SetupCost = 0;
1021 C.ScaleCost = 0;
1022 }
1023
1024 bool isLess(Cost &Other, const TargetTransformInfo &TTI);
1025
1026 void Lose();
1027
1028#ifndef NDEBUG
1029 // Once any of the metrics loses, they must all remain losers.
1030 bool isValid() {
1031 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1032 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1033 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1034 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1035 }
1036#endif
1037
1038 bool isLoser() {
1039 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1039, __PRETTY_FUNCTION__))
;
1040 return C.NumRegs == ~0u;
1041 }
1042
1043 void RateFormula(const TargetTransformInfo &TTI,
1044 const Formula &F,
1045 SmallPtrSetImpl<const SCEV *> &Regs,
1046 const DenseSet<const SCEV *> &VisitedRegs,
1047 const Loop *L,
1048 ScalarEvolution &SE, DominatorTree &DT,
1049 const LSRUse &LU,
1050 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1051
1052 void print(raw_ostream &OS) const;
1053 void dump() const;
1054
1055private:
1056 void RateRegister(const SCEV *Reg,
1057 SmallPtrSetImpl<const SCEV *> &Regs,
1058 const Loop *L,
1059 ScalarEvolution &SE, DominatorTree &DT,
1060 const TargetTransformInfo &TTI);
1061 void RatePrimaryRegister(const SCEV *Reg,
1062 SmallPtrSetImpl<const SCEV *> &Regs,
1063 const Loop *L,
1064 ScalarEvolution &SE, DominatorTree &DT,
1065 SmallPtrSetImpl<const SCEV *> *LoserRegs,
1066 const TargetTransformInfo &TTI);
1067};
1068
1069/// An operand value in an instruction which is to be replaced with some
1070/// equivalent, possibly strength-reduced, replacement.
1071struct LSRFixup {
1072 /// The instruction which will be updated.
1073 Instruction *UserInst = nullptr;
1074
1075 /// The operand of the instruction which will be replaced. The operand may be
1076 /// used more than once; every instance will be replaced.
1077 Value *OperandValToReplace = nullptr;
1078
1079 /// If this user is to use the post-incremented value of an induction
1080 /// variable, this set is non-empty and holds the loops associated with the
1081 /// induction variable.
1082 PostIncLoopSet PostIncLoops;
1083
1084 /// A constant offset to be added to the LSRUse expression. This allows
1085 /// multiple fixups to share the same LSRUse with different offsets, for
1086 /// example in an unrolled loop.
1087 int64_t Offset = 0;
1088
1089 LSRFixup() = default;
1090
1091 bool isUseFullyOutsideLoop(const Loop *L) const;
1092
1093 void print(raw_ostream &OS) const;
1094 void dump() const;
1095};
1096
1097/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1098/// SmallVectors of const SCEV*.
1099struct UniquifierDenseMapInfo {
1100 static SmallVector<const SCEV *, 4> getEmptyKey() {
1101 SmallVector<const SCEV *, 4> V;
1102 V.push_back(reinterpret_cast<const SCEV *>(-1));
1103 return V;
1104 }
1105
1106 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1107 SmallVector<const SCEV *, 4> V;
1108 V.push_back(reinterpret_cast<const SCEV *>(-2));
1109 return V;
1110 }
1111
1112 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1113 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1114 }
1115
1116 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1117 const SmallVector<const SCEV *, 4> &RHS) {
1118 return LHS == RHS;
1119 }
1120};
1121
1122/// This class holds the state that LSR keeps for each use in IVUsers, as well
1123/// as uses invented by LSR itself. It includes information about what kinds of
1124/// things can be folded into the user, information about the user itself, and
1125/// information about how the use may be satisfied. TODO: Represent multiple
1126/// users of the same expression in common?
1127class LSRUse {
1128 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1129
1130public:
1131 /// An enum for a kind of use, indicating what types of scaled and immediate
1132 /// operands it might support.
1133 enum KindType {
1134 Basic, ///< A normal use, with no folding.
1135 Special, ///< A special case of basic, allowing -1 scales.
1136 Address, ///< An address use; folding according to TargetLowering
1137 ICmpZero ///< An equality icmp with both operands folded into one.
1138 // TODO: Add a generic icmp too?
1139 };
1140
1141 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1142
1143 KindType Kind;
1144 MemAccessTy AccessTy;
1145
1146 /// The list of operands which are to be replaced.
1147 SmallVector<LSRFixup, 8> Fixups;
1148
1149 /// Keep track of the min and max offsets of the fixups.
1150 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1151 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1152
1153 /// This records whether all of the fixups using this LSRUse are outside of
1154 /// the loop, in which case some special-case heuristics may be used.
1155 bool AllFixupsOutsideLoop = true;
1156
1157 /// RigidFormula is set to true to guarantee that this use will be associated
1158 /// with a single formula--the one that initially matched. Some SCEV
1159 /// expressions cannot be expanded. This allows LSR to consider the registers
1160 /// used by those expressions without the need to expand them later after
1161 /// changing the formula.
1162 bool RigidFormula = false;
1163
1164 /// This records the widest use type for any fixup using this
1165 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1166 /// fixup widths to be equivalent, because the narrower one may be relying on
1167 /// the implicit truncation to truncate away bogus bits.
1168 Type *WidestFixupType = nullptr;
1169
1170 /// A list of ways to build a value that can satisfy this user. After the
1171 /// list is populated, one of these is selected heuristically and used to
1172 /// formulate a replacement for OperandValToReplace in UserInst.
1173 SmallVector<Formula, 12> Formulae;
1174
1175 /// The set of register candidates used by all formulae in this LSRUse.
1176 SmallPtrSet<const SCEV *, 4> Regs;
1177
1178 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1179
1180 LSRFixup &getNewFixup() {
1181 Fixups.push_back(LSRFixup());
1182 return Fixups.back();
1183 }
1184
1185 void pushFixup(LSRFixup &f) {
1186 Fixups.push_back(f);
1187 if (f.Offset > MaxOffset)
1188 MaxOffset = f.Offset;
1189 if (f.Offset < MinOffset)
1190 MinOffset = f.Offset;
1191 }
1192
1193 bool HasFormulaWithSameRegs(const Formula &F) const;
1194 float getNotSelectedProbability(const SCEV *Reg) const;
1195 bool InsertFormula(const Formula &F, const Loop &L);
1196 void DeleteFormula(Formula &F);
1197 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1198
1199 void print(raw_ostream &OS) const;
1200 void dump() const;
1201};
1202
1203} // end anonymous namespace
1204
1205static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1206 LSRUse::KindType Kind, MemAccessTy AccessTy,
1207 GlobalValue *BaseGV, int64_t BaseOffset,
1208 bool HasBaseReg, int64_t Scale,
1209 Instruction *Fixup = nullptr);
1210
1211/// Tally up interesting quantities from the given register.
1212void Cost::RateRegister(const SCEV *Reg,
1213 SmallPtrSetImpl<const SCEV *> &Regs,
1214 const Loop *L,
1215 ScalarEvolution &SE, DominatorTree &DT,
1216 const TargetTransformInfo &TTI) {
1217 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1218 // If this is an addrec for another loop, it should be an invariant
1219 // with respect to L since L is the innermost loop (at least
1220 // for now LSR only handles innermost loops).
1221 if (AR->getLoop() != L) {
1222 // If the AddRec exists, consider it's register free and leave it alone.
1223 if (isExistingPhi(AR, SE))
1224 return;
1225
1226 // It is bad to allow LSR for current loop to add induction variables
1227 // for its sibling loops.
1228 if (!AR->getLoop()->contains(L)) {
1229 Lose();
1230 return;
1231 }
1232
1233 // Otherwise, it will be an invariant with respect to Loop L.
1234 ++C.NumRegs;
1235 return;
1236 }
1237
1238 unsigned LoopCost = 1;
1239 if (TTI.shouldFavorPostInc()) {
1240 const SCEV *LoopStep = AR->getStepRecurrence(SE);
1241 if (isa<SCEVConstant>(LoopStep)) {
1242 // Check if a post-indexed load/store can be used.
1243 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
1244 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
1245 const SCEV *LoopStart = AR->getStart();
1246 if (!isa<SCEVConstant>(LoopStart) &&
1247 SE.isLoopInvariant(LoopStart, L))
1248 LoopCost = 0;
1249 }
1250 }
1251 }
1252 C.AddRecCost += LoopCost;
1253
1254 // Add the step value register, if it needs one.
1255 // TODO: The non-affine case isn't precisely modeled here.
1256 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1257 if (!Regs.count(AR->getOperand(1))) {
1258 RateRegister(AR->getOperand(1), Regs, L, SE, DT, TTI);
1259 if (isLoser())
1260 return;
1261 }
1262 }
1263 }
1264 ++C.NumRegs;
1265
1266 // Rough heuristic; favor registers which don't require extra setup
1267 // instructions in the preheader.
1268 if (!isa<SCEVUnknown>(Reg) &&
1269 !isa<SCEVConstant>(Reg) &&
1270 !(isa<SCEVAddRecExpr>(Reg) &&
1271 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1272 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1273 ++C.SetupCost;
1274
1275 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1276 SE.hasComputableLoopEvolution(Reg, L);
1277}
1278
1279/// Record this register in the set. If we haven't seen it before, rate
1280/// it. Optional LoserRegs provides a way to declare any formula that refers to
1281/// one of those regs an instant loser.
1282void Cost::RatePrimaryRegister(const SCEV *Reg,
1283 SmallPtrSetImpl<const SCEV *> &Regs,
1284 const Loop *L,
1285 ScalarEvolution &SE, DominatorTree &DT,
1286 SmallPtrSetImpl<const SCEV *> *LoserRegs,
1287 const TargetTransformInfo &TTI) {
1288 if (LoserRegs && LoserRegs->count(Reg)) {
1289 Lose();
1290 return;
1291 }
1292 if (Regs.insert(Reg).second) {
1293 RateRegister(Reg, Regs, L, SE, DT, TTI);
1294 if (LoserRegs && isLoser())
1295 LoserRegs->insert(Reg);
1296 }
1297}
1298
1299void Cost::RateFormula(const TargetTransformInfo &TTI,
1300 const Formula &F,
1301 SmallPtrSetImpl<const SCEV *> &Regs,
1302 const DenseSet<const SCEV *> &VisitedRegs,
1303 const Loop *L,
1304 ScalarEvolution &SE, DominatorTree &DT,
1305 const LSRUse &LU,
1306 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1307 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1307, __PRETTY_FUNCTION__))
;
1308 // Tally up the registers.
1309 unsigned PrevAddRecCost = C.AddRecCost;
1310 unsigned PrevNumRegs = C.NumRegs;
1311 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1312 if (const SCEV *ScaledReg = F.ScaledReg) {
1313 if (VisitedRegs.count(ScaledReg)) {
1314 Lose();
1315 return;
1316 }
1317 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs, TTI);
1318 if (isLoser())
1319 return;
1320 }
1321 for (const SCEV *BaseReg : F.BaseRegs) {
1322 if (VisitedRegs.count(BaseReg)) {
1323 Lose();
1324 return;
1325 }
1326 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs, TTI);
1327 if (isLoser())
1328 return;
1329 }
1330
1331 // Determine how many (unfolded) adds we'll need inside the loop.
1332 size_t NumBaseParts = F.getNumRegs();
1333 if (NumBaseParts > 1)
1334 // Do not count the base and a possible second register if the target
1335 // allows to fold 2 registers.
1336 C.NumBaseAdds +=
1337 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1338 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1339
1340 // Accumulate non-free scaling amounts.
1341 C.ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1342
1343 // Tally up the non-zero immediates.
1344 for (const LSRFixup &Fixup : LU.Fixups) {
1345 int64_t O = Fixup.Offset;
1346 int64_t Offset = (uint64_t)O + F.BaseOffset;
1347 if (F.BaseGV)
1348 C.ImmCost += 64; // Handle symbolic values conservatively.
1349 // TODO: This should probably be the pointer size.
1350 else if (Offset != 0)
1351 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1352
1353 // Check with target if this offset with this instruction is
1354 // specifically not supported.
1355 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1356 !isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1357 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1358 C.NumBaseAdds++;
1359 }
1360
1361 // If we don't count instruction cost exit here.
1362 if (!InsnsCost) {
1363 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1363, __PRETTY_FUNCTION__))
;
1364 return;
1365 }
1366
1367 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1368 // additional instruction (at least fill).
1369 unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1370 if (C.NumRegs > TTIRegNum) {
1371 // Cost already exceeded TTIRegNum, then only newly added register can add
1372 // new instructions.
1373 if (PrevNumRegs > TTIRegNum)
1374 C.Insns += (C.NumRegs - PrevNumRegs);
1375 else
1376 C.Insns += (C.NumRegs - TTIRegNum);
1377 }
1378
1379 // If ICmpZero formula ends with not 0, it could not be replaced by
1380 // just add or sub. We'll need to compare final result of AddRec.
1381 // That means we'll need an additional instruction. But if the target can
1382 // macro-fuse a compare with a branch, don't count this extra instruction.
1383 // For -10 + {0, +, 1}:
1384 // i = i + 1;
1385 // cmp i, 10
1386 //
1387 // For {-10, +, 1}:
1388 // i = i + 1;
1389 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() && !TTI.canMacroFuseCmp())
1390 C.Insns++;
1391 // Each new AddRec adds 1 instruction to calculation.
1392 C.Insns += (C.AddRecCost - PrevAddRecCost);
1393
1394 // BaseAdds adds instructions for unfolded registers.
1395 if (LU.Kind != LSRUse::ICmpZero)
1396 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1397 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1397, __PRETTY_FUNCTION__))
;
1398}
1399
1400/// Set this cost to a losing value.
1401void Cost::Lose() {
1402 C.Insns = std::numeric_limits<unsigned>::max();
1403 C.NumRegs = std::numeric_limits<unsigned>::max();
1404 C.AddRecCost = std::numeric_limits<unsigned>::max();
1405 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1406 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1407 C.ImmCost = std::numeric_limits<unsigned>::max();
1408 C.SetupCost = std::numeric_limits<unsigned>::max();
1409 C.ScaleCost = std::numeric_limits<unsigned>::max();
1410}
1411
1412/// Choose the lower cost.
1413bool Cost::isLess(Cost &Other, const TargetTransformInfo &TTI) {
1414 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1415 C.Insns != Other.C.Insns)
1416 return C.Insns < Other.C.Insns;
1417 return TTI.isLSRCostLess(C, Other.C);
1418}
1419
1420#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1421void Cost::print(raw_ostream &OS) const {
1422 if (InsnsCost)
1423 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1424 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1425 if (C.AddRecCost != 0)
1426 OS << ", with addrec cost " << C.AddRecCost;
1427 if (C.NumIVMuls != 0)
1428 OS << ", plus " << C.NumIVMuls << " IV mul"
1429 << (C.NumIVMuls == 1 ? "" : "s");
1430 if (C.NumBaseAdds != 0)
1431 OS << ", plus " << C.NumBaseAdds << " base add"
1432 << (C.NumBaseAdds == 1 ? "" : "s");
1433 if (C.ScaleCost != 0)
1434 OS << ", plus " << C.ScaleCost << " scale cost";
1435 if (C.ImmCost != 0)
1436 OS << ", plus " << C.ImmCost << " imm cost";
1437 if (C.SetupCost != 0)
1438 OS << ", plus " << C.SetupCost << " setup cost";
1439}
1440
1441LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1442 print(errs()); errs() << '\n';
1443}
1444#endif
1445
1446/// Test whether this fixup always uses its value outside of the given loop.
1447bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1448 // PHI nodes use their value in their incoming blocks.
1449 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1450 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1451 if (PN->getIncomingValue(i) == OperandValToReplace &&
1452 L->contains(PN->getIncomingBlock(i)))
1453 return false;
1454 return true;
1455 }
1456
1457 return !L->contains(UserInst);
1458}
1459
1460#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1461void LSRFixup::print(raw_ostream &OS) const {
1462 OS << "UserInst=";
1463 // Store is common and interesting enough to be worth special-casing.
1464 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1465 OS << "store ";
1466 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1467 } else if (UserInst->getType()->isVoidTy())
1468 OS << UserInst->getOpcodeName();
1469 else
1470 UserInst->printAsOperand(OS, /*PrintType=*/false);
1471
1472 OS << ", OperandValToReplace=";
1473 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1474
1475 for (const Loop *PIL : PostIncLoops) {
1476 OS << ", PostIncLoop=";
1477 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1478 }
1479
1480 if (Offset != 0)
1481 OS << ", Offset=" << Offset;
1482}
1483
1484LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1485 print(errs()); errs() << '\n';
1486}
1487#endif
1488
1489/// Test whether this use as a formula which has the same registers as the given
1490/// formula.
1491bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1492 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1493 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1494 // Unstable sort by host order ok, because this is only used for uniquifying.
1495 llvm::sort(Key);
1496 return Uniquifier.count(Key);
1497}
1498
1499/// The function returns a probability of selecting formula without Reg.
1500float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1501 unsigned FNum = 0;
1502 for (const Formula &F : Formulae)
1503 if (F.referencesReg(Reg))
1504 FNum++;
1505 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1506}
1507
1508/// If the given formula has not yet been inserted, add it to the list, and
1509/// return true. Return false otherwise. The formula must be in canonical form.
1510bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1511 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1511, __PRETTY_FUNCTION__))
;
1512
1513 if (!Formulae.empty() && RigidFormula)
1514 return false;
1515
1516 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1517 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1518 // Unstable sort by host order ok, because this is only used for uniquifying.
1519 llvm::sort(Key);
1520
1521 if (!Uniquifier.insert(Key).second)
1522 return false;
1523
1524 // Using a register to hold the value of 0 is not profitable.
1525 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1526, __PRETTY_FUNCTION__))
1526 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1526, __PRETTY_FUNCTION__))
;
1527#ifndef NDEBUG
1528 for (const SCEV *BaseReg : F.BaseRegs)
1529 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1529, __PRETTY_FUNCTION__))
;
1530#endif
1531
1532 // Add the formula to the list.
1533 Formulae.push_back(F);
1534
1535 // Record registers now being used by this use.
1536 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1537 if (F.ScaledReg)
1538 Regs.insert(F.ScaledReg);
1539
1540 return true;
1541}
1542
1543/// Remove the given formula from this use's list.
1544void LSRUse::DeleteFormula(Formula &F) {
1545 if (&F != &Formulae.back())
1546 std::swap(F, Formulae.back());
1547 Formulae.pop_back();
1548}
1549
1550/// Recompute the Regs field, and update RegUses.
1551void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1552 // Now that we've filtered out some formulae, recompute the Regs set.
1553 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1554 Regs.clear();
1555 for (const Formula &F : Formulae) {
1556 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1557 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1558 }
1559
1560 // Update the RegTracker.
1561 for (const SCEV *S : OldRegs)
1562 if (!Regs.count(S))
1563 RegUses.dropRegister(S, LUIdx);
1564}
1565
1566#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1567void LSRUse::print(raw_ostream &OS) const {
1568 OS << "LSR Use: Kind=";
1569 switch (Kind) {
1570 case Basic: OS << "Basic"; break;
1571 case Special: OS << "Special"; break;
1572 case ICmpZero: OS << "ICmpZero"; break;
1573 case Address:
1574 OS << "Address of ";
1575 if (AccessTy.MemTy->isPointerTy())
1576 OS << "pointer"; // the full pointer type could be really verbose
1577 else {
1578 OS << *AccessTy.MemTy;
1579 }
1580
1581 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1582 }
1583
1584 OS << ", Offsets={";
1585 bool NeedComma = false;
1586 for (const LSRFixup &Fixup : Fixups) {
1587 if (NeedComma) OS << ',';
1588 OS << Fixup.Offset;
1589 NeedComma = true;
1590 }
1591 OS << '}';
1592
1593 if (AllFixupsOutsideLoop)
1594 OS << ", all-fixups-outside-loop";
1595
1596 if (WidestFixupType)
1597 OS << ", widest fixup type: " << *WidestFixupType;
1598}
1599
1600LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1601 print(errs()); errs() << '\n';
1602}
1603#endif
1604
1605static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1606 LSRUse::KindType Kind, MemAccessTy AccessTy,
1607 GlobalValue *BaseGV, int64_t BaseOffset,
1608 bool HasBaseReg, int64_t Scale,
1609 Instruction *Fixup/*= nullptr*/) {
1610 switch (Kind) {
1611 case LSRUse::Address:
1612 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1613 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1614
1615 case LSRUse::ICmpZero:
1616 // There's not even a target hook for querying whether it would be legal to
1617 // fold a GV into an ICmp.
1618 if (BaseGV)
1619 return false;
1620
1621 // ICmp only has two operands; don't allow more than two non-trivial parts.
1622 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1623 return false;
1624
1625 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1626 // putting the scaled register in the other operand of the icmp.
1627 if (Scale != 0 && Scale != -1)
1628 return false;
1629
1630 // If we have low-level target information, ask the target if it can fold an
1631 // integer immediate on an icmp.
1632 if (BaseOffset != 0) {
1633 // We have one of:
1634 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1635 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1636 // Offs is the ICmp immediate.
1637 if (Scale == 0)
1638 // The cast does the right thing with
1639 // std::numeric_limits<int64_t>::min().
1640 BaseOffset = -(uint64_t)BaseOffset;
1641 return TTI.isLegalICmpImmediate(BaseOffset);
1642 }
1643
1644 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1645 return true;
1646
1647 case LSRUse::Basic:
1648 // Only handle single-register values.
1649 return !BaseGV && Scale == 0 && BaseOffset == 0;
1650
1651 case LSRUse::Special:
1652 // Special case Basic to handle -1 scales.
1653 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1654 }
1655
1656 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1656)
;
1657}
1658
1659static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1660 int64_t MinOffset, int64_t MaxOffset,
1661 LSRUse::KindType Kind, MemAccessTy AccessTy,
1662 GlobalValue *BaseGV, int64_t BaseOffset,
1663 bool HasBaseReg, int64_t Scale) {
1664 // Check for overflow.
1665 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1666 (MinOffset > 0))
1667 return false;
1668 MinOffset = (uint64_t)BaseOffset + MinOffset;
1669 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1670 (MaxOffset > 0))
1671 return false;
1672 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1673
1674 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1675 HasBaseReg, Scale) &&
1676 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1677 HasBaseReg, Scale);
1678}
1679
1680static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1681 int64_t MinOffset, int64_t MaxOffset,
1682 LSRUse::KindType Kind, MemAccessTy AccessTy,
1683 const Formula &F, const Loop &L) {
1684 // For the purpose of isAMCompletelyFolded either having a canonical formula
1685 // or a scale not equal to zero is correct.
1686 // Problems may arise from non canonical formulae having a scale == 0.
1687 // Strictly speaking it would best to just rely on canonical formulae.
1688 // However, when we generate the scaled formulae, we first check that the
1689 // scaling factor is profitable before computing the actual ScaledReg for
1690 // compile time sake.
1691 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)", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1691, __PRETTY_FUNCTION__))
;
1692 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1693 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1694}
1695
1696/// Test whether we know how to expand the current formula.
1697static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1698 int64_t MaxOffset, LSRUse::KindType Kind,
1699 MemAccessTy AccessTy, GlobalValue *BaseGV,
1700 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1701 // We know how to expand completely foldable formulae.
1702 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1703 BaseOffset, HasBaseReg, Scale) ||
1704 // Or formulae that use a base register produced by a sum of base
1705 // registers.
1706 (Scale == 1 &&
1707 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1708 BaseGV, BaseOffset, true, 0));
1709}
1710
1711static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1712 int64_t MaxOffset, LSRUse::KindType Kind,
1713 MemAccessTy AccessTy, const Formula &F) {
1714 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1715 F.BaseOffset, F.HasBaseReg, F.Scale);
1716}
1717
1718static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1719 const LSRUse &LU, const Formula &F) {
1720 // Target may want to look at the user instructions.
1721 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1722 for (const LSRFixup &Fixup : LU.Fixups)
1723 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1724 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1725 F.Scale, Fixup.UserInst))
1726 return false;
1727 return true;
1728 }
1729
1730 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1731 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1732 F.Scale);
1733}
1734
1735static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1736 const LSRUse &LU, const Formula &F,
1737 const Loop &L) {
1738 if (!F.Scale)
1739 return 0;
1740
1741 // If the use is not completely folded in that instruction, we will have to
1742 // pay an extra cost only for scale != 1.
1743 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1744 LU.AccessTy, F, L))
1745 return F.Scale != 1;
1746
1747 switch (LU.Kind) {
1748 case LSRUse::Address: {
1749 // Check the scaling factor cost with both the min and max offsets.
1750 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1751 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1752 F.Scale, LU.AccessTy.AddrSpace);
1753 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1754 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1755 F.Scale, LU.AccessTy.AddrSpace);
1756
1757 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1758, __PRETTY_FUNCTION__))
1758 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1758, __PRETTY_FUNCTION__))
;
1759 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1760 }
1761 case LSRUse::ICmpZero:
1762 case LSRUse::Basic:
1763 case LSRUse::Special:
1764 // The use is completely folded, i.e., everything is folded into the
1765 // instruction.
1766 return 0;
1767 }
1768
1769 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1769)
;
1770}
1771
1772static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1773 LSRUse::KindType Kind, MemAccessTy AccessTy,
1774 GlobalValue *BaseGV, int64_t BaseOffset,
1775 bool HasBaseReg) {
1776 // Fast-path: zero is always foldable.
1777 if (BaseOffset == 0 && !BaseGV) return true;
1778
1779 // Conservatively, create an address with an immediate and a
1780 // base and a scale.
1781 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1782
1783 // Canonicalize a scale of 1 to a base register if the formula doesn't
1784 // already have a base register.
1785 if (!HasBaseReg && Scale == 1) {
1786 Scale = 0;
1787 HasBaseReg = true;
1788 }
1789
1790 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1791 HasBaseReg, Scale);
1792}
1793
1794static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1795 ScalarEvolution &SE, int64_t MinOffset,
1796 int64_t MaxOffset, LSRUse::KindType Kind,
1797 MemAccessTy AccessTy, const SCEV *S,
1798 bool HasBaseReg) {
1799 // Fast-path: zero is always foldable.
1800 if (S->isZero()) return true;
1801
1802 // Conservatively, create an address with an immediate and a
1803 // base and a scale.
1804 int64_t BaseOffset = ExtractImmediate(S, SE);
1805 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1806
1807 // If there's anything else involved, it's not foldable.
1808 if (!S->isZero()) return false;
1809
1810 // Fast-path: zero is always foldable.
1811 if (BaseOffset == 0 && !BaseGV) return true;
1812
1813 // Conservatively, create an address with an immediate and a
1814 // base and a scale.
1815 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1816
1817 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1818 BaseOffset, HasBaseReg, Scale);
1819}
1820
1821namespace {
1822
1823/// An individual increment in a Chain of IV increments. Relate an IV user to
1824/// an expression that computes the IV it uses from the IV used by the previous
1825/// link in the Chain.
1826///
1827/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1828/// original IVOperand. The head of the chain's IVOperand is only valid during
1829/// chain collection, before LSR replaces IV users. During chain generation,
1830/// IncExpr can be used to find the new IVOperand that computes the same
1831/// expression.
1832struct IVInc {
1833 Instruction *UserInst;
1834 Value* IVOperand;
1835 const SCEV *IncExpr;
1836
1837 IVInc(Instruction *U, Value *O, const SCEV *E)
1838 : UserInst(U), IVOperand(O), IncExpr(E) {}
1839};
1840
1841// The list of IV increments in program order. We typically add the head of a
1842// chain without finding subsequent links.
1843struct IVChain {
1844 SmallVector<IVInc, 1> Incs;
1845 const SCEV *ExprBase = nullptr;
1846
1847 IVChain() = default;
1848 IVChain(const IVInc &Head, const SCEV *Base)
1849 : Incs(1, Head), ExprBase(Base) {}
1850
1851 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1852
1853 // Return the first increment in the chain.
1854 const_iterator begin() const {
1855 assert(!Incs.empty())((!Incs.empty()) ? static_cast<void> (0) : __assert_fail
("!Incs.empty()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1855, __PRETTY_FUNCTION__))
;
1856 return std::next(Incs.begin());
1857 }
1858 const_iterator end() const {
1859 return Incs.end();
1860 }
1861
1862 // Returns true if this chain contains any increments.
1863 bool hasIncs() const { return Incs.size() >= 2; }
1864
1865 // Add an IVInc to the end of this chain.
1866 void add(const IVInc &X) { Incs.push_back(X); }
1867
1868 // Returns the last UserInst in the chain.
1869 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1870
1871 // Returns true if IncExpr can be profitably added to this chain.
1872 bool isProfitableIncrement(const SCEV *OperExpr,
1873 const SCEV *IncExpr,
1874 ScalarEvolution&);
1875};
1876
1877/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1878/// between FarUsers that definitely cross IV increments and NearUsers that may
1879/// be used between IV increments.
1880struct ChainUsers {
1881 SmallPtrSet<Instruction*, 4> FarUsers;
1882 SmallPtrSet<Instruction*, 4> NearUsers;
1883};
1884
1885/// This class holds state for the main loop strength reduction logic.
1886class LSRInstance {
1887 IVUsers &IU;
1888 ScalarEvolution &SE;
1889 DominatorTree &DT;
1890 LoopInfo &LI;
1891 const TargetTransformInfo &TTI;
1892 Loop *const L;
1893 bool Changed = false;
1894
1895 /// This is the insert position that the current loop's induction variable
1896 /// increment should be placed. In simple loops, this is the latch block's
1897 /// terminator. But in more complicated cases, this is a position which will
1898 /// dominate all the in-loop post-increment users.
1899 Instruction *IVIncInsertPos = nullptr;
1900
1901 /// Interesting factors between use strides.
1902 ///
1903 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1904 /// default, a SmallDenseSet, because we need to use the full range of
1905 /// int64_ts, and there's currently no good way of doing that with
1906 /// SmallDenseSet.
1907 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1908
1909 /// Interesting use types, to facilitate truncation reuse.
1910 SmallSetVector<Type *, 4> Types;
1911
1912 /// The list of interesting uses.
1913 SmallVector<LSRUse, 16> Uses;
1914
1915 /// Track which uses use which register candidates.
1916 RegUseTracker RegUses;
1917
1918 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1919 // have more than a few IV increment chains in a loop. Missing a Chain falls
1920 // back to normal LSR behavior for those uses.
1921 static const unsigned MaxChains = 8;
1922
1923 /// IV users can form a chain of IV increments.
1924 SmallVector<IVChain, MaxChains> IVChainVec;
1925
1926 /// IV users that belong to profitable IVChains.
1927 SmallPtrSet<Use*, MaxChains> IVIncSet;
1928
1929 void OptimizeShadowIV();
1930 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1931 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1932 void OptimizeLoopTermCond();
1933
1934 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1935 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1936 void FinalizeChain(IVChain &Chain);
1937 void CollectChains();
1938 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1939 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1940
1941 void CollectInterestingTypesAndFactors();
1942 void CollectFixupsAndInitialFormulae();
1943
1944 // Support for sharing of LSRUses between LSRFixups.
1945 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
1946 UseMapTy UseMap;
1947
1948 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1949 LSRUse::KindType Kind, MemAccessTy AccessTy);
1950
1951 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1952 MemAccessTy AccessTy);
1953
1954 void DeleteUse(LSRUse &LU, size_t LUIdx);
1955
1956 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1957
1958 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1959 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1960 void CountRegisters(const Formula &F, size_t LUIdx);
1961 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1962
1963 void CollectLoopInvariantFixupsAndFormulae();
1964
1965 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1966 unsigned Depth = 0);
1967
1968 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1969 const Formula &Base, unsigned Depth,
1970 size_t Idx, bool IsScaledReg = false);
1971 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1972 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1973 const Formula &Base, size_t Idx,
1974 bool IsScaledReg = false);
1975 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1976 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1977 const Formula &Base,
1978 const SmallVectorImpl<int64_t> &Worklist,
1979 size_t Idx, bool IsScaledReg = false);
1980 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1981 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1982 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1983 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1984 void GenerateCrossUseConstantOffsets();
1985 void GenerateAllReuseFormulae();
1986
1987 void FilterOutUndesirableDedicatedRegisters();
1988
1989 size_t EstimateSearchSpaceComplexity() const;
1990 void NarrowSearchSpaceByDetectingSupersets();
1991 void NarrowSearchSpaceByCollapsingUnrolledCode();
1992 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1993 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
1994 void NarrowSearchSpaceByDeletingCostlyFormulas();
1995 void NarrowSearchSpaceByPickingWinnerRegs();
1996 void NarrowSearchSpaceUsingHeuristics();
1997
1998 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1999 Cost &SolutionCost,
2000 SmallVectorImpl<const Formula *> &Workspace,
2001 const Cost &CurCost,
2002 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2003 DenseSet<const SCEV *> &VisitedRegs) const;
2004 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2005
2006 BasicBlock::iterator
2007 HoistInsertPosition(BasicBlock::iterator IP,
2008 const SmallVectorImpl<Instruction *> &Inputs) const;
2009 BasicBlock::iterator
2010 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2011 const LSRFixup &LF,
2012 const LSRUse &LU,
2013 SCEVExpander &Rewriter) const;
2014
2015 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2016 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2017 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2018 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2019 const Formula &F, SCEVExpander &Rewriter,
2020 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2021 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2022 SCEVExpander &Rewriter,
2023 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2024 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2025
2026public:
2027 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2028 LoopInfo &LI, const TargetTransformInfo &TTI);
2029
2030 bool getChanged() const { return Changed; }
2031
2032 void print_factors_and_types(raw_ostream &OS) const;
2033 void print_fixups(raw_ostream &OS) const;
2034 void print_uses(raw_ostream &OS) const;
2035 void print(raw_ostream &OS) const;
2036 void dump() const;
2037};
2038
2039} // end anonymous namespace
2040
2041/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2042/// the cast operation.
2043void LSRInstance::OptimizeShadowIV() {
2044 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2045 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2046 return;
2047
2048 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2049 UI != E; /* empty */) {
2050 IVUsers::const_iterator CandidateUI = UI;
2051 ++UI;
2052 Instruction *ShadowUse = CandidateUI->getUser();
2053 Type *DestTy = nullptr;
2054 bool IsSigned = false;
2055
2056 /* If shadow use is a int->float cast then insert a second IV
2057 to eliminate this cast.
2058
2059 for (unsigned i = 0; i < n; ++i)
2060 foo((double)i);
2061
2062 is transformed into
2063
2064 double d = 0.0;
2065 for (unsigned i = 0; i < n; ++i, ++d)
2066 foo(d);
2067 */
2068 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2069 IsSigned = false;
2070 DestTy = UCast->getDestTy();
2071 }
2072 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2073 IsSigned = true;
2074 DestTy = SCast->getDestTy();
2075 }
2076 if (!DestTy) continue;
2077
2078 // If target does not support DestTy natively then do not apply
2079 // this transformation.
2080 if (!TTI.isTypeLegal(DestTy)) continue;
2081
2082 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2083 if (!PH) continue;
2084 if (PH->getNumIncomingValues() != 2) continue;
2085
2086 // If the calculation in integers overflows, the result in FP type will
2087 // differ. So we only can do this transformation if we are guaranteed to not
2088 // deal with overflowing values
2089 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2090 if (!AR) continue;
2091 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2092 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2093
2094 Type *SrcTy = PH->getType();
2095 int Mantissa = DestTy->getFPMantissaWidth();
2096 if (Mantissa == -1) continue;
2097 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2098 continue;
2099
2100 unsigned Entry, Latch;
2101 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2102 Entry = 0;
2103 Latch = 1;
2104 } else {
2105 Entry = 1;
2106 Latch = 0;
2107 }
2108
2109 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2110 if (!Init) continue;
2111 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2112 (double)Init->getSExtValue() :
2113 (double)Init->getZExtValue());
2114
2115 BinaryOperator *Incr =
2116 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2117 if (!Incr) continue;
2118 if (Incr->getOpcode() != Instruction::Add
2119 && Incr->getOpcode() != Instruction::Sub)
2120 continue;
2121
2122 /* Initialize new IV, double d = 0.0 in above example. */
2123 ConstantInt *C = nullptr;
2124 if (Incr->getOperand(0) == PH)
2125 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2126 else if (Incr->getOperand(1) == PH)
2127 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2128 else
2129 continue;
2130
2131 if (!C) continue;
2132
2133 // Ignore negative constants, as the code below doesn't handle them
2134 // correctly. TODO: Remove this restriction.
2135 if (!C->getValue().isStrictlyPositive()) continue;
2136
2137 /* Add new PHINode. */
2138 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2139
2140 /* create new increment. '++d' in above example. */
2141 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2142 BinaryOperator *NewIncr =
2143 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2144 Instruction::FAdd : Instruction::FSub,
2145 NewPH, CFP, "IV.S.next.", Incr);
2146
2147 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2148 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2149
2150 /* Remove cast operation */
2151 ShadowUse->replaceAllUsesWith(NewPH);
2152 ShadowUse->eraseFromParent();
2153 Changed = true;
2154 break;
2155 }
2156}
2157
2158/// If Cond has an operand that is an expression of an IV, set the IV user and
2159/// stride information and return true, otherwise return false.
2160bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2161 for (IVStrideUse &U : IU)
2162 if (U.getUser() == Cond) {
2163 // NOTE: we could handle setcc instructions with multiple uses here, but
2164 // InstCombine does it as well for simple uses, it's not clear that it
2165 // occurs enough in real life to handle.
2166 CondUse = &U;
2167 return true;
2168 }
2169 return false;
2170}
2171
2172/// Rewrite the loop's terminating condition if it uses a max computation.
2173///
2174/// This is a narrow solution to a specific, but acute, problem. For loops
2175/// like this:
2176///
2177/// i = 0;
2178/// do {
2179/// p[i] = 0.0;
2180/// } while (++i < n);
2181///
2182/// the trip count isn't just 'n', because 'n' might not be positive. And
2183/// unfortunately this can come up even for loops where the user didn't use
2184/// a C do-while loop. For example, seemingly well-behaved top-test loops
2185/// will commonly be lowered like this:
2186///
2187/// if (n > 0) {
2188/// i = 0;
2189/// do {
2190/// p[i] = 0.0;
2191/// } while (++i < n);
2192/// }
2193///
2194/// and then it's possible for subsequent optimization to obscure the if
2195/// test in such a way that indvars can't find it.
2196///
2197/// When indvars can't find the if test in loops like this, it creates a
2198/// max expression, which allows it to give the loop a canonical
2199/// induction variable:
2200///
2201/// i = 0;
2202/// max = n < 1 ? 1 : n;
2203/// do {
2204/// p[i] = 0.0;
2205/// } while (++i != max);
2206///
2207/// Canonical induction variables are necessary because the loop passes
2208/// are designed around them. The most obvious example of this is the
2209/// LoopInfo analysis, which doesn't remember trip count values. It
2210/// expects to be able to rediscover the trip count each time it is
2211/// needed, and it does this using a simple analysis that only succeeds if
2212/// the loop has a canonical induction variable.
2213///
2214/// However, when it comes time to generate code, the maximum operation
2215/// can be quite costly, especially if it's inside of an outer loop.
2216///
2217/// This function solves this problem by detecting this type of loop and
2218/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2219/// the instructions for the maximum computation.
2220ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2221 // Check that the loop matches the pattern we're looking for.
2222 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2223 Cond->getPredicate() != CmpInst::ICMP_NE)
2224 return Cond;
2225
2226 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2227 if (!Sel || !Sel->hasOneUse()) return Cond;
2228
2229 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2230 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2231 return Cond;
2232 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2233
2234 // Add one to the backedge-taken count to get the trip count.
2235 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2236 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2237
2238 // Check for a max calculation that matches the pattern. There's no check
2239 // for ICMP_ULE here because the comparison would be with zero, which
2240 // isn't interesting.
2241 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2242 const SCEVNAryExpr *Max = nullptr;
2243 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2244 Pred = ICmpInst::ICMP_SLE;
2245 Max = S;
2246 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2247 Pred = ICmpInst::ICMP_SLT;
2248 Max = S;
2249 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2250 Pred = ICmpInst::ICMP_ULT;
2251 Max = U;
2252 } else {
2253 // No match; bail.
2254 return Cond;
2255 }
2256
2257 // To handle a max with more than two operands, this optimization would
2258 // require additional checking and setup.
2259 if (Max->getNumOperands() != 2)
2260 return Cond;
2261
2262 const SCEV *MaxLHS = Max->getOperand(0);
2263 const SCEV *MaxRHS = Max->getOperand(1);
2264
2265 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2266 // for a comparison with 1. For <= and >=, a comparison with zero.
2267 if (!MaxLHS ||
2268 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2269 return Cond;
2270
2271 // Check the relevant induction variable for conformance to
2272 // the pattern.
2273 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2274 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2275 if (!AR || !AR->isAffine() ||
2276 AR->getStart() != One ||
2277 AR->getStepRecurrence(SE) != One)
2278 return Cond;
2279
2280 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2281, __PRETTY_FUNCTION__))
2281 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2281, __PRETTY_FUNCTION__))
;
2282
2283 // Check the right operand of the select, and remember it, as it will
2284 // be used in the new comparison instruction.
2285 Value *NewRHS = nullptr;
2286 if (ICmpInst::isTrueWhenEqual(Pred)) {
2287 // Look for n+1, and grab n.
2288 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2289 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2290 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2291 NewRHS = BO->getOperand(0);
2292 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2293 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2294 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2295 NewRHS = BO->getOperand(0);
2296 if (!NewRHS)
2297 return Cond;
2298 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2299 NewRHS = Sel->getOperand(1);
2300 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2301 NewRHS = Sel->getOperand(2);
2302 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2303 NewRHS = SU->getValue();
2304 else
2305 // Max doesn't match expected pattern.
2306 return Cond;
2307
2308 // Determine the new comparison opcode. It may be signed or unsigned,
2309 // and the original comparison may be either equality or inequality.
2310 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2311 Pred = CmpInst::getInversePredicate(Pred);
2312
2313 // Ok, everything looks ok to change the condition into an SLT or SGE and
2314 // delete the max calculation.
2315 ICmpInst *NewCond =
2316 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2317
2318 // Delete the max calculation instructions.
2319 Cond->replaceAllUsesWith(NewCond);
2320 CondUse->setUser(NewCond);
2321 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2322 Cond->eraseFromParent();
2323 Sel->eraseFromParent();
2324 if (Cmp->use_empty())
2325 Cmp->eraseFromParent();
2326 return NewCond;
2327}
2328
2329/// Change loop terminating condition to use the postinc iv when possible.
2330void
2331LSRInstance::OptimizeLoopTermCond() {
2332 SmallPtrSet<Instruction *, 4> PostIncs;
2333
2334 // We need a different set of heuristics for rotated and non-rotated loops.
2335 // If a loop is rotated then the latch is also the backedge, so inserting
2336 // post-inc expressions just before the latch is ideal. To reduce live ranges
2337 // it also makes sense to rewrite terminating conditions to use post-inc
2338 // expressions.
2339 //
2340 // If the loop is not rotated then the latch is not a backedge; the latch
2341 // check is done in the loop head. Adding post-inc expressions before the
2342 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2343 // in the loop body. In this case we do *not* want to use post-inc expressions
2344 // in the latch check, and we want to insert post-inc expressions before
2345 // the backedge.
2346 BasicBlock *LatchBlock = L->getLoopLatch();
2347 SmallVector<BasicBlock*, 8> ExitingBlocks;
2348 L->getExitingBlocks(ExitingBlocks);
2349 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2350 return LatchBlock != BB;
2351 })) {
2352 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2353 IVIncInsertPos = LatchBlock->getTerminator();
2354 return;
2355 }
2356
2357 // Otherwise treat this as a rotated loop.
2358 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2359 // Get the terminating condition for the loop if possible. If we
2360 // can, we want to change it to use a post-incremented version of its
2361 // induction variable, to allow coalescing the live ranges for the IV into
2362 // one register value.
2363
2364 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2365 if (!TermBr)
2366 continue;
2367 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2368 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2369 continue;
2370
2371 // Search IVUsesByStride to find Cond's IVUse if there is one.
2372 IVStrideUse *CondUse = nullptr;
2373 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2374 if (!FindIVUserForCond(Cond, CondUse))
2375 continue;
2376
2377 // If the trip count is computed in terms of a max (due to ScalarEvolution
2378 // being unable to find a sufficient guard, for example), change the loop
2379 // comparison to use SLT or ULT instead of NE.
2380 // One consequence of doing this now is that it disrupts the count-down
2381 // optimization. That's not always a bad thing though, because in such
2382 // cases it may still be worthwhile to avoid a max.
2383 Cond = OptimizeMax(Cond, CondUse);
2384
2385 // If this exiting block dominates the latch block, it may also use
2386 // the post-inc value if it won't be shared with other uses.
2387 // Check for dominance.
2388 if (!DT.dominates(ExitingBlock, LatchBlock))
2389 continue;
2390
2391 // Conservatively avoid trying to use the post-inc value in non-latch
2392 // exits if there may be pre-inc users in intervening blocks.
2393 if (LatchBlock != ExitingBlock)
2394 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2395 // Test if the use is reachable from the exiting block. This dominator
2396 // query is a conservative approximation of reachability.
2397 if (&*UI != CondUse &&
2398 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2399 // Conservatively assume there may be reuse if the quotient of their
2400 // strides could be a legal scale.
2401 const SCEV *A = IU.getStride(*CondUse, L);
2402 const SCEV *B = IU.getStride(*UI, L);
2403 if (!A || !B) continue;
2404 if (SE.getTypeSizeInBits(A->getType()) !=
2405 SE.getTypeSizeInBits(B->getType())) {
2406 if (SE.getTypeSizeInBits(A->getType()) >
2407 SE.getTypeSizeInBits(B->getType()))
2408 B = SE.getSignExtendExpr(B, A->getType());
2409 else
2410 A = SE.getSignExtendExpr(A, B->getType());
2411 }
2412 if (const SCEVConstant *D =
2413 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2414 const ConstantInt *C = D->getValue();
2415 // Stride of one or negative one can have reuse with non-addresses.
2416 if (C->isOne() || C->isMinusOne())
2417 goto decline_post_inc;
2418 // Avoid weird situations.
2419 if (C->getValue().getMinSignedBits() >= 64 ||
2420 C->getValue().isMinSignedValue())
2421 goto decline_post_inc;
2422 // Check for possible scaled-address reuse.
2423 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2424 MemAccessTy AccessTy = getAccessType(
2425 TTI, UI->getUser(), UI->getOperandValToReplace());
2426 int64_t Scale = C->getSExtValue();
2427 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2428 /*BaseOffset=*/0,
2429 /*HasBaseReg=*/false, Scale,
2430 AccessTy.AddrSpace))
2431 goto decline_post_inc;
2432 Scale = -Scale;
2433 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2434 /*BaseOffset=*/0,
2435 /*HasBaseReg=*/false, Scale,
2436 AccessTy.AddrSpace))
2437 goto decline_post_inc;
2438 }
2439 }
2440 }
2441
2442 LLVM_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)
2443 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2444
2445 // It's possible for the setcc instruction to be anywhere in the loop, and
2446 // possible for it to have multiple users. If it is not immediately before
2447 // the exiting block branch, move it.
2448 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2449 if (Cond->hasOneUse()) {
2450 Cond->moveBefore(TermBr);
2451 } else {
2452 // Clone the terminating condition and insert into the loopend.
2453 ICmpInst *OldCond = Cond;
2454 Cond = cast<ICmpInst>(Cond->clone());
2455 Cond->setName(L->getHeader()->getName() + ".termcond");
2456 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2457
2458 // Clone the IVUse, as the old use still exists!
2459 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2460 TermBr->replaceUsesOfWith(OldCond, Cond);
2461 }
2462 }
2463
2464 // If we get to here, we know that we can transform the setcc instruction to
2465 // use the post-incremented version of the IV, allowing us to coalesce the
2466 // live ranges for the IV correctly.
2467 CondUse->transformToPostInc(L);
2468 Changed = true;
2469
2470 PostIncs.insert(Cond);
2471 decline_post_inc:;
2472 }
2473
2474 // Determine an insertion point for the loop induction variable increment. It
2475 // must dominate all the post-inc comparisons we just set up, and it must
2476 // dominate the loop latch edge.
2477 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2478 for (Instruction *Inst : PostIncs) {
2479 BasicBlock *BB =
2480 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2481 Inst->getParent());
2482 if (BB == Inst->getParent())
2483 IVIncInsertPos = Inst;
2484 else if (BB != IVIncInsertPos->getParent())
2485 IVIncInsertPos = BB->getTerminator();
2486 }
2487}
2488
2489/// Determine if the given use can accommodate a fixup at the given offset and
2490/// other details. If so, update the use and return true.
2491bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2492 bool HasBaseReg, LSRUse::KindType Kind,
2493 MemAccessTy AccessTy) {
2494 int64_t NewMinOffset = LU.MinOffset;
2495 int64_t NewMaxOffset = LU.MaxOffset;
2496 MemAccessTy NewAccessTy = AccessTy;
2497
2498 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2499 // something conservative, however this can pessimize in the case that one of
2500 // the uses will have all its uses outside the loop, for example.
2501 if (LU.Kind != Kind)
2502 return false;
2503
2504 // Check for a mismatched access type, and fall back conservatively as needed.
2505 // TODO: Be less conservative when the type is similar and can use the same
2506 // addressing modes.
2507 if (Kind == LSRUse::Address) {
2508 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2509 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2510 AccessTy.AddrSpace);
2511 }
2512 }
2513
2514 // Conservatively assume HasBaseReg is true for now.
2515 if (NewOffset < LU.MinOffset) {
2516 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2517 LU.MaxOffset - NewOffset, HasBaseReg))
2518 return false;
2519 NewMinOffset = NewOffset;
2520 } else if (NewOffset > LU.MaxOffset) {
2521 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2522 NewOffset - LU.MinOffset, HasBaseReg))
2523 return false;
2524 NewMaxOffset = NewOffset;
2525 }
2526
2527 // Update the use.
2528 LU.MinOffset = NewMinOffset;
2529 LU.MaxOffset = NewMaxOffset;
2530 LU.AccessTy = NewAccessTy;
2531 return true;
2532}
2533
2534/// Return an LSRUse index and an offset value for a fixup which needs the given
2535/// expression, with the given kind and optional access type. Either reuse an
2536/// existing use or create a new one, as needed.
2537std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2538 LSRUse::KindType Kind,
2539 MemAccessTy AccessTy) {
2540 const SCEV *Copy = Expr;
2541 int64_t Offset = ExtractImmediate(Expr, SE);
2542
2543 // Basic uses can't accept any offset, for example.
2544 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2545 Offset, /*HasBaseReg=*/ true)) {
2546 Expr = Copy;
2547 Offset = 0;
2548 }
2549
2550 std::pair<UseMapTy::iterator, bool> P =
2551 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2552 if (!P.second) {
2553 // A use already existed with this base.
2554 size_t LUIdx = P.first->second;
2555 LSRUse &LU = Uses[LUIdx];
2556 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2557 // Reuse this use.
2558 return std::make_pair(LUIdx, Offset);
2559 }
2560
2561 // Create a new use.
2562 size_t LUIdx = Uses.size();
2563 P.first->second = LUIdx;
2564 Uses.push_back(LSRUse(Kind, AccessTy));
2565 LSRUse &LU = Uses[LUIdx];
2566
2567 LU.MinOffset = Offset;
2568 LU.MaxOffset = Offset;
2569 return std::make_pair(LUIdx, Offset);
2570}
2571
2572/// Delete the given use from the Uses list.
2573void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2574 if (&LU != &Uses.back())
2575 std::swap(LU, Uses.back());
2576 Uses.pop_back();
2577
2578 // Update RegUses.
2579 RegUses.swapAndDropUse(LUIdx, Uses.size());
2580}
2581
2582/// Look for a use distinct from OrigLU which is has a formula that has the same
2583/// registers as the given formula.
2584LSRUse *
2585LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2586 const LSRUse &OrigLU) {
2587 // Search all uses for the formula. This could be more clever.
2588 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2589 LSRUse &LU = Uses[LUIdx];
2590 // Check whether this use is close enough to OrigLU, to see whether it's
2591 // worthwhile looking through its formulae.
2592 // Ignore ICmpZero uses because they may contain formulae generated by
2593 // GenerateICmpZeroScales, in which case adding fixup offsets may
2594 // be invalid.
2595 if (&LU != &OrigLU &&
2596 LU.Kind != LSRUse::ICmpZero &&
2597 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2598 LU.WidestFixupType == OrigLU.WidestFixupType &&
2599 LU.HasFormulaWithSameRegs(OrigF)) {
2600 // Scan through this use's formulae.
2601 for (const Formula &F : LU.Formulae) {
2602 // Check to see if this formula has the same registers and symbols
2603 // as OrigF.
2604 if (F.BaseRegs == OrigF.BaseRegs &&
2605 F.ScaledReg == OrigF.ScaledReg &&
2606 F.BaseGV == OrigF.BaseGV &&
2607 F.Scale == OrigF.Scale &&
2608 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2609 if (F.BaseOffset == 0)
2610 return &LU;
2611 // This is the formula where all the registers and symbols matched;
2612 // there aren't going to be any others. Since we declined it, we
2613 // can skip the rest of the formulae and proceed to the next LSRUse.
2614 break;
2615 }
2616 }
2617 }
2618 }
2619
2620 // Nothing looked good.
2621 return nullptr;
2622}
2623
2624void LSRInstance::CollectInterestingTypesAndFactors() {
2625 SmallSetVector<const SCEV *, 4> Strides;
2626
2627 // Collect interesting types and strides.
2628 SmallVector<const SCEV *, 4> Worklist;
2629 for (const IVStrideUse &U : IU) {
2630 const SCEV *Expr = IU.getExpr(U);
2631
2632 // Collect interesting types.
2633 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2634
2635 // Add strides for mentioned loops.
2636 Worklist.push_back(Expr);
2637 do {
2638 const SCEV *S = Worklist.pop_back_val();
2639 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2640 if (AR->getLoop() == L)
2641 Strides.insert(AR->getStepRecurrence(SE));
2642 Worklist.push_back(AR->getStart());
2643 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2644 Worklist.append(Add->op_begin(), Add->op_end());
2645 }
2646 } while (!Worklist.empty());
2647 }
2648
2649 // Compute interesting factors from the set of interesting strides.
2650 for (SmallSetVector<const SCEV *, 4>::const_iterator
2651 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2652 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2653 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2654 const SCEV *OldStride = *I;
2655 const SCEV *NewStride = *NewStrideIter;
2656
2657 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2658 SE.getTypeSizeInBits(NewStride->getType())) {
2659 if (SE.getTypeSizeInBits(OldStride->getType()) >
2660 SE.getTypeSizeInBits(NewStride->getType()))
2661 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2662 else
2663 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2664 }
2665 if (const SCEVConstant *Factor =
2666 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2667 SE, true))) {
2668 if (Factor->getAPInt().getMinSignedBits() <= 64)
2669 Factors.insert(Factor->getAPInt().getSExtValue());
2670 } else if (const SCEVConstant *Factor =
2671 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2672 NewStride,
2673 SE, true))) {
2674 if (Factor->getAPInt().getMinSignedBits() <= 64)
2675 Factors.insert(Factor->getAPInt().getSExtValue());
2676 }
2677 }
2678
2679 // If all uses use the same type, don't bother looking for truncation-based
2680 // reuse.
2681 if (Types.size() == 1)
2682 Types.clear();
2683
2684 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2685}
2686
2687/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2688/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2689/// IVStrideUses, we could partially skip this.
2690static User::op_iterator
2691findIVOperand(User::op_iterator OI, User::op_iterator OE,
2692 Loop *L, ScalarEvolution &SE) {
2693 for(; OI != OE; ++OI) {
2694 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2695 if (!SE.isSCEVable(Oper->getType()))
2696 continue;
2697
2698 if (const SCEVAddRecExpr *AR =
2699 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2700 if (AR->getLoop() == L)
2701 break;
2702 }
2703 }
2704 }
2705 return OI;
2706}
2707
2708/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2709/// a convenient helper.
2710static Value *getWideOperand(Value *Oper) {
2711 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2712 return Trunc->getOperand(0);
2713 return Oper;
2714}
2715
2716/// Return true if we allow an IV chain to include both types.
2717static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2718 Type *LType = LVal->getType();
2719 Type *RType = RVal->getType();
2720 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2721 // Different address spaces means (possibly)
2722 // different types of the pointer implementation,
2723 // e.g. i16 vs i32 so disallow that.
2724 (LType->getPointerAddressSpace() ==
2725 RType->getPointerAddressSpace()));
2726}
2727
2728/// Return an approximation of this SCEV expression's "base", or NULL for any
2729/// constant. Returning the expression itself is conservative. Returning a
2730/// deeper subexpression is more precise and valid as long as it isn't less
2731/// complex than another subexpression. For expressions involving multiple
2732/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2733/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2734/// IVInc==b-a.
2735///
2736/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2737/// SCEVUnknown, we simply return the rightmost SCEV operand.
2738static const SCEV *getExprBase(const SCEV *S) {
2739 switch (S->getSCEVType()) {
2740 default: // uncluding scUnknown.
2741 return S;
2742 case scConstant:
2743 return nullptr;
2744 case scTruncate:
2745 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2746 case scZeroExtend:
2747 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2748 case scSignExtend:
2749 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2750 case scAddExpr: {
2751 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2752 // there's nothing more complex.
2753 // FIXME: not sure if we want to recognize negation.
2754 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2755 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2756 E(Add->op_begin()); I != E; ++I) {
2757 const SCEV *SubExpr = *I;
2758 if (SubExpr->getSCEVType() == scAddExpr)
2759 return getExprBase(SubExpr);
2760
2761 if (SubExpr->getSCEVType() != scMulExpr)
2762 return SubExpr;
2763 }
2764 return S; // all operands are scaled, be conservative.
2765 }
2766 case scAddRecExpr:
2767 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2768 }
2769}
2770
2771/// Return true if the chain increment is profitable to expand into a loop
2772/// invariant value, which may require its own register. A profitable chain
2773/// increment will be an offset relative to the same base. We allow such offsets
2774/// to potentially be used as chain increment as long as it's not obviously
2775/// expensive to expand using real instructions.
2776bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2777 const SCEV *IncExpr,
2778 ScalarEvolution &SE) {
2779 // Aggressively form chains when -stress-ivchain.
2780 if (StressIVChain)
2781 return true;
2782
2783 // Do not replace a constant offset from IV head with a nonconstant IV
2784 // increment.
2785 if (!isa<SCEVConstant>(IncExpr)) {
2786 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2787 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2788 return false;
2789 }
2790
2791 SmallPtrSet<const SCEV*, 8> Processed;
2792 return !isHighCostExpansion(IncExpr, Processed, SE);
2793}
2794
2795/// Return true if the number of registers needed for the chain is estimated to
2796/// be less than the number required for the individual IV users. First prohibit
2797/// any IV users that keep the IV live across increments (the Users set should
2798/// be empty). Next count the number and type of increments in the chain.
2799///
2800/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2801/// effectively use postinc addressing modes. Only consider it profitable it the
2802/// increments can be computed in fewer registers when chained.
2803///
2804/// TODO: Consider IVInc free if it's already used in another chains.
2805static bool
2806isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2807 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2808 if (StressIVChain)
2809 return true;
2810
2811 if (!Chain.hasIncs())
2812 return false;
2813
2814 if (!Users.empty()) {
2815 LLVM_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)
2816 for (Instruction *Instdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
: Users) { dbgs() << " " << *Inst << "\n"
; }; } } while (false)
2817 : Users) { 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)
;
2818 return false;
2819 }
2820 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2820, __PRETTY_FUNCTION__))
;
2821
2822 // The chain itself may require a register, so intialize cost to 1.
2823 int cost = 1;
2824
2825 // A complete chain likely eliminates the need for keeping the original IV in
2826 // a register. LSR does not currently know how to form a complete chain unless
2827 // the header phi already exists.
2828 if (isa<PHINode>(Chain.tailUserInst())
2829 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2830 --cost;
2831 }
2832 const SCEV *LastIncExpr = nullptr;
2833 unsigned NumConstIncrements = 0;
2834 unsigned NumVarIncrements = 0;
2835 unsigned NumReusedIncrements = 0;
2836 for (const IVInc &Inc : Chain) {
2837 if (Inc.IncExpr->isZero())
2838 continue;
2839
2840 // Incrementing by zero or some constant is neutral. We assume constants can
2841 // be folded into an addressing mode or an add's immediate operand.
2842 if (isa<SCEVConstant>(Inc.IncExpr)) {
2843 ++NumConstIncrements;
2844 continue;
2845 }
2846
2847 if (Inc.IncExpr == LastIncExpr)
2848 ++NumReusedIncrements;
2849 else
2850 ++NumVarIncrements;
2851
2852 LastIncExpr = Inc.IncExpr;
2853 }
2854 // An IV chain with a single increment is handled by LSR's postinc
2855 // uses. However, a chain with multiple increments requires keeping the IV's
2856 // value live longer than it needs to be if chained.
2857 if (NumConstIncrements > 1)
2858 --cost;
2859
2860 // Materializing increment expressions in the preheader that didn't exist in
2861 // the original code may cost a register. For example, sign-extended array
2862 // indices can produce ridiculous increments like this:
2863 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2864 cost += NumVarIncrements;
2865
2866 // Reusing variable increments likely saves a register to hold the multiple of
2867 // the stride.
2868 cost -= NumReusedIncrements;
2869
2870 LLVM_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)
2871 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2872
2873 return cost < 0;
2874}
2875
2876/// Add this IV user to an existing chain or make it the head of a new chain.
2877void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2878 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2879 // When IVs are used as types of varying widths, they are generally converted
2880 // to a wider type with some uses remaining narrow under a (free) trunc.
2881 Value *const NextIV = getWideOperand(IVOper);
2882 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2883 const SCEV *const OperExprBase = getExprBase(OperExpr);
2884
2885 // Visit all existing chains. Check if its IVOper can be computed as a
2886 // profitable loop invariant increment from the last link in the Chain.
2887 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2888 const SCEV *LastIncExpr = nullptr;
2889 for (; ChainIdx < NChains; ++ChainIdx) {
2890 IVChain &Chain = IVChainVec[ChainIdx];
2891
2892 // Prune the solution space aggressively by checking that both IV operands
2893 // are expressions that operate on the same unscaled SCEVUnknown. This
2894 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2895 // first avoids creating extra SCEV expressions.
2896 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2897 continue;
2898
2899 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2900 if (!isCompatibleIVType(PrevIV, NextIV))
2901 continue;
2902
2903 // A phi node terminates a chain.
2904 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2905 continue;
2906
2907 // The increment must be loop-invariant so it can be kept in a register.
2908 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2909 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2910 if (!SE.isLoopInvariant(IncExpr, L))
2911 continue;
2912
2913 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2914 LastIncExpr = IncExpr;
2915 break;
2916 }
2917 }
2918 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2919 // bother for phi nodes, because they must be last in the chain.
2920 if (ChainIdx == NChains) {
2921 if (isa<PHINode>(UserInst))
2922 return;
2923 if (NChains >= MaxChains && !StressIVChain) {
2924 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2925 return;
2926 }
2927 LastIncExpr = OperExpr;
2928 // IVUsers may have skipped over sign/zero extensions. We don't currently
2929 // attempt to form chains involving extensions unless they can be hoisted
2930 // into this loop's AddRec.
2931 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2932 return;
2933 ++NChains;
2934 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2935 OperExprBase));
2936 ChainUsersVec.resize(NChains);
2937 LLVM_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)
2938 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
2939 } else {
2940 LLVM_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)
2941 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
2942 // Add this IV user to the end of the chain.
2943 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2944 }
2945 IVChain &Chain = IVChainVec[ChainIdx];
2946
2947 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2948 // This chain's NearUsers become FarUsers.
2949 if (!LastIncExpr->isZero()) {
2950 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2951 NearUsers.end());
2952 NearUsers.clear();
2953 }
2954
2955 // All other uses of IVOperand become near uses of the chain.
2956 // We currently ignore intermediate values within SCEV expressions, assuming
2957 // they will eventually be used be the current chain, or can be computed
2958 // from one of the chain increments. To be more precise we could
2959 // transitively follow its user and only add leaf IV users to the set.
2960 for (User *U : IVOper->users()) {
2961 Instruction *OtherUse = dyn_cast<Instruction>(U);
2962 if (!OtherUse)
2963 continue;
2964 // Uses in the chain will no longer be uses if the chain is formed.
2965 // Include the head of the chain in this iteration (not Chain.begin()).
2966 IVChain::const_iterator IncIter = Chain.Incs.begin();
2967 IVChain::const_iterator IncEnd = Chain.Incs.end();
2968 for( ; IncIter != IncEnd; ++IncIter) {
2969 if (IncIter->UserInst == OtherUse)
2970 break;
2971 }
2972 if (IncIter != IncEnd)
2973 continue;
2974
2975 if (SE.isSCEVable(OtherUse->getType())
2976 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2977 && IU.isIVUserOrOperand(OtherUse)) {
2978 continue;
2979 }
2980 NearUsers.insert(OtherUse);
2981 }
2982
2983 // Since this user is part of the chain, it's no longer considered a use
2984 // of the chain.
2985 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2986}
2987
2988/// Populate the vector of Chains.
2989///
2990/// This decreases ILP at the architecture level. Targets with ample registers,
2991/// multiple memory ports, and no register renaming probably don't want
2992/// this. However, such targets should probably disable LSR altogether.
2993///
2994/// The job of LSR is to make a reasonable choice of induction variables across
2995/// the loop. Subsequent passes can easily "unchain" computation exposing more
2996/// ILP *within the loop* if the target wants it.
2997///
2998/// Finding the best IV chain is potentially a scheduling problem. Since LSR
2999/// will not reorder memory operations, it will recognize this as a chain, but
3000/// will generate redundant IV increments. Ideally this would be corrected later
3001/// by a smart scheduler:
3002/// = A[i]
3003/// = A[i+x]
3004/// A[i] =
3005/// A[i+x] =
3006///
3007/// TODO: Walk the entire domtree within this loop, not just the path to the
3008/// loop latch. This will discover chains on side paths, but requires
3009/// maintaining multiple copies of the Chains state.
3010void LSRInstance::CollectChains() {
3011 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3012 SmallVector<ChainUsers, 8> ChainUsersVec;
3013
3014 SmallVector<BasicBlock *,8> LatchPath;
3015 BasicBlock *LoopHeader = L->getHeader();
3016 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3017 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3018 LatchPath.push_back(Rung->getBlock());
3019 }
3020 LatchPath.push_back(LoopHeader);
3021
3022 // Walk the instruction stream from the loop header to the loop latch.
3023 for (BasicBlock *BB : reverse(LatchPath)) {
3024 for (Instruction &I : *BB) {
3025 // Skip instructions that weren't seen by IVUsers analysis.
3026 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3027 continue;
3028
3029 // Ignore users that are part of a SCEV expression. This way we only
3030 // consider leaf IV Users. This effectively rediscovers a portion of
3031 // IVUsers analysis but in program order this time.
3032 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3033 continue;
3034
3035 // Remove this instruction from any NearUsers set it may be in.
3036 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3037 ChainIdx < NChains; ++ChainIdx) {
3038 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3039 }
3040 // Search for operands that can be chained.
3041 SmallPtrSet<Instruction*, 4> UniqueOperands;
3042 User::op_iterator IVOpEnd = I.op_end();
3043 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3044 while (IVOpIter != IVOpEnd) {
3045 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3046 if (UniqueOperands.insert(IVOpInst).second)
3047 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3048 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3049 }
3050 } // Continue walking down the instructions.
3051 } // Continue walking down the domtree.
3052 // Visit phi backedges to determine if the chain can generate the IV postinc.
3053 for (PHINode &PN : L->getHeader()->phis()) {
3054 if (!SE.isSCEVable(PN.getType()))
3055 continue;
3056
3057 Instruction *IncV =
3058 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3059 if (IncV)
3060 ChainInstruction(&PN, IncV, ChainUsersVec);
3061 }
3062 // Remove any unprofitable chains.
3063 unsigned ChainIdx = 0;
3064 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3065 UsersIdx < NChains; ++UsersIdx) {
3066 if (!isProfitableChain(IVChainVec[UsersIdx],
3067 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3068 continue;
3069 // Preserve the chain at UsesIdx.
3070 if (ChainIdx != UsersIdx)
3071 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3072 FinalizeChain(IVChainVec[ChainIdx]);
3073 ++ChainIdx;
3074 }
3075 IVChainVec.resize(ChainIdx);
3076}
3077
3078void LSRInstance::FinalizeChain(IVChain &Chain) {
3079 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3079, __PRETTY_FUNCTION__))
;
3080 LLVM_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)
;
3081
3082 for (const IVInc &Inc : Chain) {
3083 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3084 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3085 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3085, __PRETTY_FUNCTION__))
;
3086 IVIncSet.insert(UseI);
3087 }
3088}
3089
3090/// Return true if the IVInc can be folded into an addressing mode.
3091static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3092 Value *Operand, const TargetTransformInfo &TTI) {
3093 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3094 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3095 return false;
3096
3097 if (IncConst->getAPInt().getMinSignedBits() > 64)
3098 return false;
3099
3100 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3101 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3102 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3103 IncOffset, /*HaseBaseReg=*/false))
3104 return false;
3105
3106 return true;
3107}
3108
3109/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3110/// user's operand from the previous IV user's operand.
3111void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3112 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3113 // Find the new IVOperand for the head of the chain. It may have been replaced
3114 // by LSR.
3115 const IVInc &Head = Chain.Incs[0];
3116 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3117 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3118 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3119 IVOpEnd, L, SE);
3120 Value *IVSrc = nullptr;
6
'IVSrc' initialized to a null pointer value
3121 while (IVOpIter != IVOpEnd) {
7
Loop condition is false. Execution continues on line 3138
3122 IVSrc = getWideOperand(*IVOpIter);
3123
3124 // If this operand computes the expression that the chain needs, we may use
3125 // it. (Check this after setting IVSrc which is used below.)
3126 //
3127 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3128 // narrow for the chain, so we can no longer use it. We do allow using a
3129 // wider phi, assuming the LSR checked for free truncation. In that case we
3130 // should already have a truncate on this operand such that
3131 // getSCEV(IVSrc) == IncExpr.
3132 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3133 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3134 break;
3135 }
3136 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3137 }
3138 if (IVOpIter == IVOpEnd) {
8
Assuming 'IVOpIter' is not equal to 'IVOpEnd'
9
Taking false branch
3139 // Gracefully give up on this chain.
3140 LLVM_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)
;
3141 return;
3142 }
3143
3144 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
10
Assuming 'DebugFlag' is 0
11
Loop condition is false. Exiting loop
3145 Type *IVTy = IVSrc->getType();
12
Called C++ object pointer is null
3146 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3147 const SCEV *LeftOverExpr = nullptr;
3148 for (const IVInc &Inc : Chain) {
3149 Instruction *InsertPt = Inc.UserInst;
3150 if (isa<PHINode>(InsertPt))
3151 InsertPt = L->getLoopLatch()->getTerminator();
3152
3153 // IVOper will replace the current IV User's operand. IVSrc is the IV
3154 // value currently held in a register.
3155 Value *IVOper = IVSrc;
3156 if (!Inc.IncExpr->isZero()) {
3157 // IncExpr was the result of subtraction of two narrow values, so must
3158 // be signed.
3159 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3160 LeftOverExpr = LeftOverExpr ?
3161 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3162 }
3163 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3164 // Expand the IV increment.
3165 Rewriter.clearPostInc();
3166 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3167 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3168 SE.getUnknown(IncV));
3169 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3170
3171 // If an IV increment can't be folded, use it as the next IV value.
3172 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3173 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3173, __PRETTY_FUNCTION__))
;
3174 IVSrc = IVOper;
3175 LeftOverExpr = nullptr;
3176 }
3177 }
3178 Type *OperTy = Inc.IVOperand->getType();
3179 if (IVTy != OperTy) {
3180 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3181, __PRETTY_FUNCTION__))
3181 "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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3181, __PRETTY_FUNCTION__))
;
3182 IRBuilder<> Builder(InsertPt);
3183 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3184 }
3185 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3186 DeadInsts.emplace_back(Inc.IVOperand);
3187 }
3188 // If LSR created a new, wider phi, we may also replace its postinc. We only
3189 // do this if we also found a wide value for the head of the chain.
3190 if (isa<PHINode>(Chain.tailUserInst())) {
3191 for (PHINode &Phi : L->getHeader()->phis()) {
3192 if (!isCompatibleIVType(&Phi, IVSrc))
3193 continue;
3194 Instruction *PostIncV = dyn_cast<Instruction>(
3195 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3196 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3197 continue;
3198 Value *IVOper = IVSrc;
3199 Type *PostIncTy = PostIncV->getType();
3200 if (IVTy != PostIncTy) {
3201 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3201, __PRETTY_FUNCTION__))
;
3202 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3203 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3204 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3205 }
3206 Phi.replaceUsesOfWith(PostIncV, IVOper);
3207 DeadInsts.emplace_back(PostIncV);
3208 }
3209 }
3210}
3211
3212void LSRInstance::CollectFixupsAndInitialFormulae() {
3213 for (const IVStrideUse &U : IU) {
3214 Instruction *UserInst = U.getUser();
3215 // Skip IV users that are part of profitable IV Chains.
3216 User::op_iterator UseI =
3217 find(UserInst->operands(), U.getOperandValToReplace());
3218 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3218, __PRETTY_FUNCTION__))
;
3219 if (IVIncSet.count(UseI)) {
3220 LLVM_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)
;
3221 continue;
3222 }
3223
3224 LSRUse::KindType Kind = LSRUse::Basic;
3225 MemAccessTy AccessTy;
3226 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3227 Kind = LSRUse::Address;
3228 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3229 }
3230
3231 const SCEV *S = IU.getExpr(U);
3232 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3233
3234 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3235 // (N - i == 0), and this allows (N - i) to be the expression that we work
3236 // with rather than just N or i, so we can consider the register
3237 // requirements for both N and i at the same time. Limiting this code to
3238 // equality icmps is not a problem because all interesting loops use
3239 // equality icmps, thanks to IndVarSimplify.
3240 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3241 if (CI->isEquality()) {
3242 // Swap the operands if needed to put the OperandValToReplace on the
3243 // left, for consistency.
3244 Value *NV = CI->getOperand(1);
3245 if (NV == U.getOperandValToReplace()) {
3246 CI->setOperand(1, CI->getOperand(0));
3247 CI->setOperand(0, NV);
3248 NV = CI->getOperand(1);
3249 Changed = true;
3250 }
3251
3252 // x == y --> x - y == 0
3253 const SCEV *N = SE.getSCEV(NV);
3254 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3255 // S is normalized, so normalize N before folding it into S
3256 // to keep the result normalized.
3257 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3258 Kind = LSRUse::ICmpZero;
3259 S = SE.getMinusSCEV(N, S);
3260 }
3261
3262 // -1 and the negations of all interesting strides (except the negation
3263 // of -1) are now also interesting.
3264 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3265 if (Factors[i] != -1)
3266 Factors.insert(-(uint64_t)Factors[i]);
3267 Factors.insert(-1);
3268 }
3269
3270 // Get or create an LSRUse.
3271 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3272 size_t LUIdx = P.first;
3273 int64_t Offset = P.second;
3274 LSRUse &LU = Uses[LUIdx];
3275
3276 // Record the fixup.
3277 LSRFixup &LF = LU.getNewFixup();
3278 LF.UserInst = UserInst;
3279 LF.OperandValToReplace = U.getOperandValToReplace();
3280 LF.PostIncLoops = TmpPostIncLoops;
3281 LF.Offset = Offset;
3282 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3283
3284 if (!LU.WidestFixupType ||
3285 SE.getTypeSizeInBits(LU.WidestFixupType) <
3286 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3287 LU.WidestFixupType = LF.OperandValToReplace->getType();
3288
3289 // If this is the first use of this LSRUse, give it a formula.
3290 if (LU.Formulae.empty()) {
3291 InsertInitialFormula(S, LU, LUIdx);
3292 CountRegisters(LU.Formulae.back(), LUIdx);
3293 }
3294 }
3295
3296 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3297}
3298
3299/// Insert a formula for the given expression into the given use, separating out
3300/// loop-variant portions from loop-invariant and loop-computable portions.
3301void
3302LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3303 // Mark uses whose expressions cannot be expanded.
3304 if (!isSafeToExpand(S, SE))
3305 LU.RigidFormula = true;
3306
3307 Formula F;
3308 F.initialMatch(S, L, SE);
3309 bool Inserted = InsertFormula(LU, LUIdx, F);
3310 assert(Inserted && "Initial formula already exists!")((Inserted && "Initial formula already exists!") ? static_cast
<void> (0) : __assert_fail ("Inserted && \"Initial formula already exists!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3310, __PRETTY_FUNCTION__))
; (void)Inserted;
3311}
3312
3313/// Insert a simple single-register formula for the given expression into the
3314/// given use.
3315void
3316LSRInstance::InsertSupplementalFormula(const SCEV *S,
3317 LSRUse &LU, size_t LUIdx) {
3318 Formula F;
3319 F.BaseRegs.push_back(S);
3320 F.HasBaseReg = true;
3321 bool Inserted = InsertFormula(LU, LUIdx, F);
3322 assert(Inserted && "Supplemental formula already exists!")((Inserted && "Supplemental formula already exists!")
? static_cast<void> (0) : __assert_fail ("Inserted && \"Supplemental formula already exists!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3322, __PRETTY_FUNCTION__))
; (void)Inserted;
3323}
3324
3325/// Note which registers are used by the given formula, updating RegUses.
3326void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3327 if (F.ScaledReg)
3328 RegUses.countRegister(F.ScaledReg, LUIdx);
3329 for (const SCEV *BaseReg : F.BaseRegs)
3330 RegUses.countRegister(BaseReg, LUIdx);
3331}
3332
3333/// If the given formula has not yet been inserted, add it to the list, and
3334/// return true. Return false otherwise.
3335bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3336 // Do not insert formula that we will not be able to expand.
3337 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3338, __PRETTY_FUNCTION__))
3338 "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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3338, __PRETTY_FUNCTION__))
;
3339
3340 if (!LU.InsertFormula(F, *L))
3341 return false;
3342
3343 CountRegisters(F, LUIdx);
3344 return true;
3345}
3346
3347/// Check for other uses of loop-invariant values which we're tracking. These
3348/// other uses will pin these values in registers, making them less profitable
3349/// for elimination.
3350/// TODO: This currently misses non-constant addrec step registers.
3351/// TODO: Should this give more weight to users inside the loop?
3352void
3353LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3354 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3355 SmallPtrSet<const SCEV *, 32> Visited;
3356
3357 while (!Worklist.empty()) {
3358 const SCEV *S = Worklist.pop_back_val();
3359
3360 // Don't process the same SCEV twice
3361 if (!Visited.insert(S).second)
3362 continue;
3363
3364 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3365 Worklist.append(N->op_begin(), N->op_end());
3366 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3367 Worklist.push_back(C->getOperand());
3368 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3369 Worklist.push_back(D->getLHS());
3370 Worklist.push_back(D->getRHS());
3371 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3372 const Value *V = US->getValue();
3373 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3374 // Look for instructions defined outside the loop.
3375 if (L->contains(Inst)) continue;
3376 } else if (isa<UndefValue>(V))
3377 // Undef doesn't have a live range, so it doesn't matter.
3378 continue;
3379 for (const Use &U : V->uses()) {
3380 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3381 // Ignore non-instructions.
3382 if (!UserInst)
3383 continue;
3384 // Ignore instructions in other functions (as can happen with
3385 // Constants).
3386 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3387 continue;
3388 // Ignore instructions not dominated by the loop.
3389 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3390 UserInst->getParent() :
3391 cast<PHINode>(UserInst)->getIncomingBlock(
3392 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3393 if (!DT.dominates(L->getHeader(), UseBB))
3394 continue;
3395 // Don't bother if the instruction is in a BB which ends in an EHPad.
3396 if (UseBB->getTerminator()->isEHPad())
3397 continue;
3398 // Don't bother rewriting PHIs in catchswitch blocks.
3399 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3400 continue;
3401 // Ignore uses which are part of other SCEV expressions, to avoid
3402 // analyzing them multiple times.
3403 if (SE.isSCEVable(UserInst->getType())) {
3404 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3405 // If the user is a no-op, look through to its uses.
3406 if (!isa<SCEVUnknown>(UserS))
3407 continue;
3408 if (UserS == US) {
3409 Worklist.push_back(
3410 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3411 continue;
3412 }
3413 }
3414 // Ignore icmp instructions which are already being analyzed.
3415 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3416 unsigned OtherIdx = !U.getOperandNo();
3417 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3418 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3419 continue;
3420 }
3421
3422 std::pair<size_t, int64_t> P = getUse(
3423 S, LSRUse::Basic, MemAccessTy());
3424 size_t LUIdx = P.first;
3425 int64_t Offset = P.second;
3426 LSRUse &LU = Uses[LUIdx];
3427 LSRFixup &LF = LU.getNewFixup();
3428 LF.UserInst = const_cast<Instruction *>(UserInst);
3429 LF.OperandValToReplace = U;
3430 LF.Offset = Offset;
3431 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3432 if (!LU.WidestFixupType ||
3433 SE.getTypeSizeInBits(LU.WidestFixupType) <
3434 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3435 LU.WidestFixupType = LF.OperandValToReplace->getType();
3436 InsertSupplementalFormula(US, LU, LUIdx);
3437 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3438 break;
3439 }
3440 }
3441 }
3442}
3443
3444/// Split S into subexpressions which can be pulled out into separate
3445/// registers. If C is non-null, multiply each subexpression by C.
3446///
3447/// Return remainder expression after factoring the subexpressions captured by
3448/// Ops. If Ops is complete, return NULL.
3449static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3450 SmallVectorImpl<const SCEV *> &Ops,
3451 const Loop *L,
3452 ScalarEvolution &SE,
3453 unsigned Depth = 0) {
3454 // Arbitrarily cap recursion to protect compile time.
3455 if (Depth >= 3)
3456 return S;
3457
3458 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3459 // Break out add operands.
3460 for (const SCEV *S : Add->operands()) {
3461 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3462 if (Remainder)
3463 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3464 }
3465 return nullptr;
3466 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3467 // Split a non-zero base out of an addrec.
3468 if (AR->getStart()->isZero() || !AR->isAffine())
3469 return S;
3470
3471 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3472 C, Ops, L, SE, Depth+1);
3473 // Split the non-zero AddRec unless it is part of a nested recurrence that
3474 // does not pertain to this loop.
3475 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3476 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3477 Remainder = nullptr;
3478 }
3479 if (Remainder != AR->getStart()) {
3480 if (!Remainder)
3481 Remainder = SE.getConstant(AR->getType(), 0);
3482 return SE.getAddRecExpr(Remainder,
3483 AR->getStepRecurrence(SE),
3484 AR->getLoop(),
3485 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3486 SCEV::FlagAnyWrap);
3487 }
3488 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3489 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3490 if (Mul->getNumOperands() != 2)
3491 return S;
3492 if (const SCEVConstant *Op0 =
3493 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3494 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3495 const SCEV *Remainder =
3496 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3497 if (Remainder)
3498 Ops.push_back(SE.getMulExpr(C, Remainder));
3499 return nullptr;
3500 }
3501 }
3502 return S;
3503}
3504
3505/// Return true if the SCEV represents a value that may end up as a
3506/// post-increment operation.
3507static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3508 LSRUse &LU, const SCEV *S, const Loop *L,
3509 ScalarEvolution &SE) {
3510 if (LU.Kind != LSRUse::Address ||
3511 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3512 return false;
3513 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3514 if (!AR)
3515 return false;
3516 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3517 if (!isa<SCEVConstant>(LoopStep))
3518 return false;
3519 if (LU.AccessTy.getType()->getScalarSizeInBits() !=
3520 LoopStep->getType()->getScalarSizeInBits())
3521 return false;
3522 // Check if a post-indexed load/store can be used.
3523 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3524 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3525 const SCEV *LoopStart = AR->getStart();
3526 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3527 return true;
3528 }
3529 return false;
3530}
3531
3532/// Helper function for LSRInstance::GenerateReassociations.
3533void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3534 const Formula &Base,
3535 unsigned Depth, size_t Idx,
3536 bool IsScaledReg) {
3537 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3538 // Don't generate reassociations for the base register of a value that
3539 // may generate a post-increment operator. The reason is that the
3540 // reassociations cause extra base+register formula to be created,
3541 // and possibly chosen, but the post-increment is more efficient.
3542 if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3543 return;
3544 SmallVector<const SCEV *, 8> AddOps;
3545 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3546 if (Remainder)
3547 AddOps.push_back(Remainder);
3548
3549 if (AddOps.size() == 1)
3550 return;
3551
3552 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3553 JE = AddOps.end();
3554 J != JE; ++J) {
3555 // Loop-variant "unknown" values are uninteresting; we won't be able to
3556 // do anything meaningful with them.
3557 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3558 continue;
3559
3560 // Don't pull a constant into a register if the constant could be folded
3561 // into an immediate field.
3562 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3563 LU.AccessTy, *J, Base.getNumRegs() > 1))
3564 continue;
3565
3566 // Collect all operands except *J.
3567 SmallVector<const SCEV *, 8> InnerAddOps(
3568 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3569 InnerAddOps.append(std::next(J),
3570 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3571
3572 // Don't leave just a constant behind in a register if the constant could
3573 // be folded into an immediate field.
3574 if (InnerAddOps.size() == 1 &&
3575 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3576 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3577 continue;
3578
3579 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3580 if (InnerSum->isZero())
3581 continue;
3582 Formula F = Base;
3583
3584 // Add the remaining pieces of the add back into the new formula.
3585 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3586 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3587 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3588 InnerSumSC->getValue()->getZExtValue())) {
3589 F.UnfoldedOffset =
3590 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3591 if (IsScaledReg)
3592 F.ScaledReg = nullptr;
3593 else
3594 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3595 } else if (IsScaledReg)
3596 F.ScaledReg = InnerSum;
3597 else
3598 F.BaseRegs[Idx] = InnerSum;
3599
3600 // Add J as its own register, or an unfolded immediate.
3601 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3602 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3603 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3604 SC->getValue()->getZExtValue()))
3605 F.UnfoldedOffset =
3606 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3607 else
3608 F.BaseRegs.push_back(*J);
3609 // We may have changed the number of register in base regs, adjust the
3610 // formula accordingly.
3611 F.canonicalize(*L);
3612
3613 if (InsertFormula(LU, LUIdx, F))
3614 // If that formula hadn't been seen before, recurse to find more like
3615 // it.
3616 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3617 // Because just Depth is not enough to bound compile time.
3618 // This means that every time AddOps.size() is greater 16^x we will add
3619 // x to Depth.
3620 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3621 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3622 }
3623}
3624
3625/// Split out subexpressions from adds and the bases of addrecs.
3626void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3627 Formula Base, unsigned Depth) {
3628 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3628, __PRETTY_FUNCTION__))
;
3629 // Arbitrarily cap recursion to protect compile time.
3630 if (Depth >= 3)
3631 return;
3632
3633 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3634 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3635
3636 if (Base.Scale == 1)
3637 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3638 /* Idx */ -1, /* IsScaledReg */ true);
3639}
3640
3641/// Generate a formula consisting of all of the loop-dominating registers added
3642/// into a single register.
3643void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3644 Formula Base) {
3645 // This method is only interesting on a plurality of registers.
3646 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3647 (Base.UnfoldedOffset != 0) <= 1)
3648 return;
3649
3650 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3651 // processing the formula.
3652 Base.unscale();
3653 SmallVector<const SCEV *, 4> Ops;
3654 Formula NewBase = Base;
3655 NewBase.BaseRegs.clear();
3656 Type *CombinedIntegerType = nullptr;
3657 for (const SCEV *BaseReg : Base.BaseRegs) {
3658 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3659 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3660 if (!CombinedIntegerType)
3661 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3662 Ops.push_back(BaseReg);
3663 }
3664 else
3665 NewBase.BaseRegs.push_back(BaseReg);
3666 }
3667
3668 // If no register is relevant, we're done.
3669 if (Ops.size() == 0)
3670 return;
3671
3672 // Utility function for generating the required variants of the combined
3673 // registers.
3674 auto GenerateFormula = [&](const SCEV *Sum) {
3675 Formula F = NewBase;
3676
3677 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3678 // opportunity to fold something. For now, just ignore such cases
3679 // rather than proceed with zero in a register.
3680 if (Sum->isZero())
3681 return;
3682
3683 F.BaseRegs.push_back(Sum);
3684 F.canonicalize(*L);
3685 (void)InsertFormula(LU, LUIdx, F);
3686 };
3687
3688 // If we collected at least two registers, generate a formula combining them.
3689 if (Ops.size() > 1) {
3690 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3691 GenerateFormula(SE.getAddExpr(OpsCopy));
3692 }
3693
3694 // If we have an unfolded offset, generate a formula combining it with the
3695 // registers collected.
3696 if (NewBase.UnfoldedOffset) {
3697 assert(CombinedIntegerType && "Missing a type for the unfolded offset")((CombinedIntegerType && "Missing a type for the unfolded offset"
) ? static_cast<void> (0) : __assert_fail ("CombinedIntegerType && \"Missing a type for the unfolded offset\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3697, __PRETTY_FUNCTION__))
;
3698 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3699 true));
3700 NewBase.UnfoldedOffset = 0;
3701 GenerateFormula(SE.getAddExpr(Ops));
3702 }
3703}
3704
3705/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3706void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3707 const Formula &Base, size_t Idx,
3708 bool IsScaledReg) {
3709 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3710 GlobalValue *GV = ExtractSymbol(G, SE);
3711 if (G->isZero() || !GV)
3712 return;
3713 Formula F = Base;
3714 F.BaseGV = GV;
3715 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3716 return;
3717 if (IsScaledReg)
3718 F.ScaledReg = G;
3719 else
3720 F.BaseRegs[Idx] = G;
3721 (void)InsertFormula(LU, LUIdx, F);
3722}
3723
3724/// Generate reuse formulae using symbolic offsets.
3725void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3726 Formula Base) {
3727 // We can't add a symbolic offset if the address already contains one.
3728 if (Base.BaseGV) return;
3729
3730 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3731 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3732 if (Base.Scale == 1)
3733 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3734 /* IsScaledReg */ true);
3735}
3736
3737/// Helper function for LSRInstance::GenerateConstantOffsets.
3738void LSRInstance::GenerateConstantOffsetsImpl(
3739 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3740 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3741 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3742 for (int64_t Offset : Worklist) {
3743 Formula F = Base;
3744 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3745 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3746 LU.AccessTy, F)) {
3747 // Add the offset to the base register.
3748 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3749 // If it cancelled out, drop the base register, otherwise update it.
3750 if (NewG->isZero()) {
3751 if (IsScaledReg) {
3752 F.Scale = 0;
3753 F.ScaledReg = nullptr;
3754 } else
3755 F.deleteBaseReg(F.BaseRegs[Idx]);
3756 F.canonicalize(*L);
3757 } else if (IsScaledReg)
3758 F.ScaledReg = NewG;
3759 else
3760 F.BaseRegs[Idx] = NewG;
3761
3762 (void)InsertFormula(LU, LUIdx, F);
3763 }
3764 }
3765
3766 int64_t Imm = ExtractImmediate(G, SE);
3767 if (G->isZero() || Imm == 0)
3768 return;
3769 Formula F = Base;
3770 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3771 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3772 return;
3773 if (IsScaledReg)
3774 F.ScaledReg = G;
3775 else
3776 F.BaseRegs[Idx] = G;
3777 (void)InsertFormula(LU, LUIdx, F);
3778}
3779
3780/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3781void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3782 Formula Base) {
3783 // TODO: For now, just add the min and max offset, because it usually isn't
3784 // worthwhile looking at everything inbetween.
3785 SmallVector<int64_t, 2> Worklist;
3786 Worklist.push_back(LU.MinOffset);
3787 if (LU.MaxOffset != LU.MinOffset)
3788 Worklist.push_back(LU.MaxOffset);
3789
3790 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3791 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3792 if (Base.Scale == 1)
3793 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3794 /* IsScaledReg */ true);
3795}
3796
3797/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3798/// == y -> x*c == y*c.
3799void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3800 Formula Base) {
3801 if (LU.Kind != LSRUse::ICmpZero) return;
3802
3803 // Determine the integer type for the base formula.
3804 Type *IntTy = Base.getType();
3805 if (!IntTy) return;
3806 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3807
3808 // Don't do this if there is more than one offset.
3809 if (LU.MinOffset != LU.MaxOffset) return;
3810
3811 // Check if transformation is valid. It is illegal to multiply pointer.
3812 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3813 return;
3814 for (const SCEV *BaseReg : Base.BaseRegs)
3815 if (BaseReg->getType()->isPointerTy())
3816 return;
3817 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3817, __PRETTY_FUNCTION__))
;
3818
3819 // Check each interesting stride.
3820 for (int64_t Factor : Factors) {
3821 // Check that the multiplication doesn't overflow.
3822 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3823 continue;
3824 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3825 if (NewBaseOffset / Factor != Base.BaseOffset)
3826 continue;
3827 // If the offset will be truncated at this use, check that it is in bounds.
3828 if (!IntTy->isPointerTy() &&
3829 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3830 continue;
3831
3832 // Check that multiplying with the use offset doesn't overflow.
3833 int64_t Offset = LU.MinOffset;
3834 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3835 continue;
3836 Offset = (uint64_t)Offset * Factor;
3837 if (Offset / Factor != LU.MinOffset)
3838 continue;
3839 // If the offset will be truncated at this use, check that it is in bounds.
3840 if (!IntTy->isPointerTy() &&
3841 !ConstantInt::isValueValidForType(IntTy, Offset))
3842 continue;
3843
3844 Formula F = Base;
3845 F.BaseOffset = NewBaseOffset;
3846
3847 // Check that this scale is legal.
3848 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3849 continue;
3850
3851 // Compensate for the use having MinOffset built into it.
3852 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3853
3854 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3855
3856 // Check that multiplying with each base register doesn't overflow.
3857 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3858 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3859 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3860 goto next;
3861 }
3862
3863 // Check that multiplying with the scaled register doesn't overflow.
3864 if (F.ScaledReg) {
3865 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3866 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3867 continue;
3868 }
3869
3870 // Check that multiplying with the unfolded offset doesn't overflow.
3871 if (F.UnfoldedOffset != 0) {
3872 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3873 Factor == -1)
3874 continue;
3875 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3876 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3877 continue;
3878 // If the offset will be truncated, check that it is in bounds.
3879 if (!IntTy->isPointerTy() &&
3880 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3881 continue;
3882 }
3883
3884 // If we make it here and it's legal, add it.
3885 (void)InsertFormula(LU, LUIdx, F);
3886 next:;
3887 }
3888}
3889
3890/// Generate stride factor reuse formulae by making use of scaled-offset address
3891/// modes, for example.
3892void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3893 // Determine the integer type for the base formula.
3894 Type *IntTy = Base.getType();
3895 if (!IntTy) return;
3896
3897 // If this Formula already has a scaled register, we can't add another one.
3898 // Try to unscale the formula to generate a better scale.
3899 if (Base.Scale != 0 && !Base.unscale())
3900 return;
3901
3902 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3902, __PRETTY_FUNCTION__))
;
3903
3904 // Check each interesting stride.
3905 for (int64_t Factor : Factors) {
3906 Base.Scale = Factor;
3907 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3908 // Check whether this scale is going to be legal.
3909 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3910 Base)) {
3911 // As a special-case, handle special out-of-loop Basic users specially.
3912 // TODO: Reconsider this special case.
3913 if (LU.Kind == LSRUse::Basic &&
3914 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3915 LU.AccessTy, Base) &&
3916 LU.AllFixupsOutsideLoop)
3917 LU.Kind = LSRUse::Special;
3918 else
3919 continue;
3920 }
3921 // For an ICmpZero, negating a solitary base register won't lead to
3922 // new solutions.
3923 if (LU.Kind == LSRUse::ICmpZero &&
3924 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3925 continue;
3926 // For each addrec base reg, if its loop is current loop, apply the scale.
3927 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3928 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3929 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3930 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3931 if (FactorS->isZero())
3932 continue;
3933 // Divide out the factor, ignoring high bits, since we'll be
3934 // scaling the value back up in the end.
3935 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3936 // TODO: This could be optimized to avoid all the copying.
3937 Formula F = Base;
3938 F.ScaledReg = Quotient;
3939 F.deleteBaseReg(F.BaseRegs[i]);
3940 // The canonical representation of 1*reg is reg, which is already in
3941 // Base. In that case, do not try to insert the formula, it will be
3942 // rejected anyway.
3943 if (F.Scale == 1 && (F.BaseRegs.empty() ||
3944 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3945 continue;
3946 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3947 // non canonical Formula with ScaledReg's loop not being L.
3948 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3949 F.canonicalize(*L);
3950 (void)InsertFormula(LU, LUIdx, F);
3951 }
3952 }
3953 }
3954 }
3955}
3956
3957/// Generate reuse formulae from different IV types.
3958void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3959 // Don't bother truncating symbolic values.
3960 if (Base.BaseGV) return;
3961
3962 // Determine the integer type for the base formula.
3963 Type *DstTy = Base.getType();
3964 if (!DstTy) return;
3965 DstTy = SE.getEffectiveSCEVType(DstTy);
3966
3967 for (Type *SrcTy : Types) {
3968 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3969 Formula F = Base;
3970
3971 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3972 for (const SCEV *&BaseReg : F.BaseRegs)
3973 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3974
3975 // TODO: This assumes we've done basic processing on all uses and
3976 // have an idea what the register usage is.
3977 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3978 continue;
3979
3980 F.canonicalize(*L);
3981 (void)InsertFormula(LU, LUIdx, F);
3982 }
3983 }
3984}
3985
3986namespace {
3987
3988/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3989/// modifications so that the search phase doesn't have to worry about the data
3990/// structures moving underneath it.
3991struct WorkItem {
3992 size_t LUIdx;
3993 int64_t Imm;
3994 const SCEV *OrigReg;
3995
3996 WorkItem(size_t LI, int64_t I, const SCEV *R)
3997 : LUIdx(LI), Imm(I), OrigReg(R) {}
3998
3999 void print(raw_ostream &OS) const;
4000 void dump() const;
4001};
4002
4003} // end anonymous namespace
4004
4005#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4006void WorkItem::print(raw_ostream &OS) const {
4007 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4008 << " , add offset " << Imm;
4009}
4010
4011LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4012 print(errs()); errs() << '\n';
4013}
4014#endif
4015
4016/// Look for registers which are a constant distance apart and try to form reuse
4017/// opportunities between them.
4018void LSRInstance::GenerateCrossUseConstantOffsets() {
4019 // Group the registers by their value without any added constant offset.
4020 using ImmMapTy = std::map<int64_t, const SCEV *>;
4021
4022 DenseMap<const SCEV *, ImmMapTy> Map;
4023 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4024 SmallVector<const SCEV *, 8> Sequence;
4025 for (const SCEV *Use : RegUses) {
4026 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4027 int64_t Imm = ExtractImmediate(Reg, SE);
4028 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4029 if (Pair.second)
4030 Sequence.push_back(Reg);
4031 Pair.first->second.insert(std::make_pair(Imm, Use));
4032 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4033 }
4034
4035 // Now examine each set of registers with the same base value. Build up
4036 // a list of work to do and do the work in a separate step so that we're
4037 // not adding formulae and register counts while we're searching.
4038 SmallVector<WorkItem, 32> WorkItems;
4039 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4040 for (const SCEV *Reg : Sequence) {
4041 const ImmMapTy &Imms = Map.find(Reg)->second;
4042
4043 // It's not worthwhile looking for reuse if there's only one offset.
4044 if (Imms.size() == 1)
4045 continue;
4046
4047 LLVM_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)
4048 for (const auto &Entrydo { 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)
4049 : Imms) dbgs()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)
4050 << ' ' << 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)
4051 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)
;
4052
4053 // Examine each offset.
4054 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4055 J != JE; ++J) {
4056 const SCEV *OrigReg = J->second;
4057
4058 int64_t JImm = J->first;
4059 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4060
4061 if (!isa<SCEVConstant>(OrigReg) &&
4062 UsedByIndicesMap[Reg].count() == 1) {
4063 LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigRegdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
4064 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4065 continue;
4066 }
4067
4068 // Conservatively examine offsets between this orig reg a few selected
4069 // other orig regs.
4070 ImmMapTy::const_iterator OtherImms[] = {
4071 Imms.begin(), std::prev(Imms.end()),
4072 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
4073 2)
4074 };
4075 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4076 ImmMapTy::const_iterator M = OtherImms[i];
4077 if (M == J || M == JE) continue;
4078
4079 // Compute the difference between the two.
4080 int64_t Imm = (uint64_t)JImm - M->first;
4081 for (unsigned LUIdx : UsedByIndices.set_bits())
4082 // Make a memo of this use, offset, and register tuple.
4083 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4084 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4085 }
4086 }
4087 }
4088
4089 Map.clear();
4090 Sequence.clear();
4091 UsedByIndicesMap.clear();
4092 UniqueItems.clear();
4093
4094 // Now iterate through the worklist and add new formulae.
4095 for (const WorkItem &WI : WorkItems) {
4096 size_t LUIdx = WI.LUIdx;
4097 LSRUse &LU = Uses[LUIdx];
4098 int64_t Imm = WI.Imm;
4099 const SCEV *OrigReg = WI.OrigReg;
4100
4101 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4102 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4103 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4104
4105 // TODO: Use a more targeted data structure.
4106 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4107 Formula F = LU.Formulae[L];
4108 // FIXME: The code for the scaled and unscaled registers looks
4109 // very similar but slightly different. Investigate if they
4110 // could be merged. That way, we would not have to unscale the
4111 // Formula.
4112 F.unscale();
4113 // Use the immediate in the scaled register.
4114 if (F.ScaledReg == OrigReg) {
4115 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4116 // Don't create 50 + reg(-50).
4117 if (F.referencesReg(SE.getSCEV(
4118 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4119 continue;
4120 Formula NewF = F;
4121 NewF.BaseOffset = Offset;
4122 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4123 NewF))
4124 continue;
4125 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4126
4127 // If the new scale is a constant in a register, and adding the constant
4128 // value to the immediate would produce a value closer to zero than the
4129 // immediate itself, then the formula isn't worthwhile.
4130 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4131 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4132 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4133 .ule(std::abs(NewF.BaseOffset)))
4134 continue;
4135
4136 // OK, looks good.
4137 NewF.canonicalize(*this->L);
4138 (void)InsertFormula(LU, LUIdx, NewF);
4139 } else {
4140 // Use the immediate in a base register.
4141 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4142 const SCEV *BaseReg = F.BaseRegs[N];
4143 if (BaseReg != OrigReg)
4144 continue;
4145 Formula NewF = F;
4146 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4147 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4148 LU.Kind, LU.AccessTy, NewF)) {
4149 if (TTI.shouldFavorPostInc() &&
4150 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4151 continue;
4152 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4153 continue;
4154 NewF = F;
4155 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4156 }
4157 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4158
4159 // If the new formula has a constant in a register, and adding the
4160 // constant value to the immediate would produce a value closer to
4161 // zero than the immediate itself, then the formula isn't worthwhile.
4162 for (const SCEV *NewReg : NewF.BaseRegs)
4163 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4164 if ((C->getAPInt() + NewF.BaseOffset)
4165 .abs()
4166 .slt(std::abs(NewF.BaseOffset)) &&
4167 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4168 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4169 goto skip_formula;
4170
4171 // Ok, looks good.
4172 NewF.canonicalize(*this->L);
4173 (void)InsertFormula(LU, LUIdx, NewF);
4174 break;
4175 skip_formula:;
4176 }
4177 }
4178 }
4179 }
4180}
4181
4182/// Generate formulae for each use.
4183void
4184LSRInstance::GenerateAllReuseFormulae() {
4185 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4186 // queries are more precise.
4187 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4188 LSRUse &LU = Uses[LUIdx];
4189 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4190 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4191 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4192 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4193 }
4194 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4195 LSRUse &LU = Uses[LUIdx];
4196 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4197 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4198 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4199 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4200 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4201 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4202 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4203 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4204 }
4205 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4206 LSRUse &LU = Uses[LUIdx];
4207 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4208 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4209 }
4210
4211 GenerateCrossUseConstantOffsets();
4212
4213 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4214 "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)
4215 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4216}
4217
4218/// If there are multiple formulae with the same set of registers used
4219/// by other uses, pick the best one and delete the others.
4220void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4221 DenseSet<const SCEV *> VisitedRegs;
4222 SmallPtrSet<const SCEV *, 16> Regs;
4223 SmallPtrSet<const SCEV *, 16> LoserRegs;
4224#ifndef NDEBUG
4225 bool ChangedFormulae = false;
4226#endif
4227
4228 // Collect the best formula for each unique set of shared registers. This
4229 // is reset for each use.
4230 using BestFormulaeTy =
4231 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4232
4233 BestFormulaeTy BestFormulae;
4234
4235 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4236 LSRUse &LU = Uses[LUIdx];
4237 LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
4238 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4239
4240 bool Any = false;
4241 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4242 FIdx != NumForms; ++FIdx) {
4243 Formula &F = LU.Formulae[FIdx];
4244
4245 // Some formulas are instant losers. For example, they may depend on
4246 // nonexistent AddRecs from other loops. These need to be filtered
4247 // immediately, otherwise heuristics could choose them over others leading
4248 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4249 // avoids the need to recompute this information across formulae using the
4250 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4251 // the corresponding bad register from the Regs set.
4252 Cost CostF;
4253 Regs.clear();
4254 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4255 if (CostF.isLoser()) {
4256 // During initial formula generation, undesirable formulae are generated
4257 // by uses within other loops that have some non-trivial address mode or
4258 // use the postinc form of the IV. LSR needs to provide these formulae
4259 // as the basis of rediscovering the desired formula that uses an AddRec
4260 // corresponding to the existing phi. Once all formulae have been
4261 // generated, these initial losers may be pruned.
4262 LLVM_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)
4263 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4264 }
4265 else {
4266 SmallVector<const SCEV *, 4> Key;
4267 for (const SCEV *Reg : F.BaseRegs) {
4268 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4269 Key.push_back(Reg);
4270 }
4271 if (F.ScaledReg &&
4272 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4273 Key.push_back(F.ScaledReg);
4274 // Unstable sort by host order ok, because this is only used for
4275 // uniquifying.
4276 llvm::sort(Key);
4277
4278 std::pair<BestFormulaeTy::const_iterator, bool> P =
4279 BestFormulae.insert(std::make_pair(Key, FIdx));
4280 if (P.second)
4281 continue;
4282
4283 Formula &Best = LU.Formulae[P.first->second];
4284
4285 Cost CostBest;
4286 Regs.clear();
4287 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4288 if (CostF.isLess(CostBest, TTI))
4289 std::swap(F, Best);
4290 LLVM_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)
4291 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)
4292 " in favor of formula ";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)
4293 Best.print(dbgs()); 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)
;
4294 }
4295#ifndef NDEBUG
4296 ChangedFormulae = true;
4297#endif
4298 LU.DeleteFormula(F);
4299 --FIdx;
4300 --NumForms;
4301 Any = true;
4302 }
4303
4304 // Now that we've filtered out some formulae, recompute the Regs set.
4305 if (Any)
4306 LU.RecomputeRegs(LUIdx, RegUses);
4307
4308 // Reset this to prepare for the next use.
4309 BestFormulae.clear();
4310 }
4311
4312 LLVM_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)
4313 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)
4314 "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)
4315 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)
4316 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4317}
4318
4319/// Estimate the worst-case number of solutions the solver might have to
4320/// consider. It almost never considers this many solutions because it prune the
4321/// search space, but the pruning isn't always sufficient.
4322size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4323 size_t Power = 1;
4324 for (const LSRUse &LU : Uses) {
4325 size_t FSize = LU.Formulae.size();
4326 if (FSize >= ComplexityLimit) {
4327 Power = ComplexityLimit;
4328 break;
4329 }
4330 Power *= FSize;
4331 if (Power >= ComplexityLimit)
4332 break;
4333 }
4334 return Power;
4335}
4336
4337/// When one formula uses a superset of the registers of another formula, it
4338/// won't help reduce register pressure (though it may not necessarily hurt
4339/// register pressure); remove it to simplify the system.
4340void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4341 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4342 LLVM_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)
;
4343
4344 LLVM_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)
4345 "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)
4346 "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)
;
4347
4348 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4349 LSRUse &LU = Uses[LUIdx];
4350 bool Any = false;
4351 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4352 Formula &F = LU.Formulae[i];
4353 // Look for a formula with a constant or GV in a register. If the use
4354 // also has a formula with that same value in an immediate field,
4355 // delete the one that uses a register.
4356 for (SmallVectorImpl<const SCEV *>::const_iterator
4357 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4358 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4359 Formula NewF = F;
4360 NewF.BaseOffset += C->getValue()->getSExtValue();
4361 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4362 (I - F.BaseRegs.begin()));
4363 if (LU.HasFormulaWithSameRegs(NewF)) {
4364 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4365 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4366 LU.DeleteFormula(F);
4367 --i;
4368 --e;
4369 Any = true;
4370 break;
4371 }
4372 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4373 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4374 if (!F.BaseGV) {
4375 Formula NewF = F;
4376 NewF.BaseGV = GV;
4377 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4378 (I - F.BaseRegs.begin()));
4379 if (LU.HasFormulaWithSameRegs(NewF)) {
4380 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4381 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4382 LU.DeleteFormula(F);
4383 --i;
4384 --e;
4385 Any = true;
4386 break;
4387 }
4388 }
4389 }
4390 }
4391 }
4392 if (Any)
4393 LU.RecomputeRegs(LUIdx, RegUses);
4394 }
4395
4396 LLVM_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)
;
4397 }
4398}
4399
4400/// When there are many registers for expressions like A, A+1, A+2, etc.,
4401/// allocate a single register for them.
4402void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4403 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4404 return;
4405
4406 LLVM_DEBUG(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)
4407 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)
4408 "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)
4409 "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)
;
4410
4411 // This is especially useful for unrolled loops.
4412
4413 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4414 LSRUse &LU = Uses[LUIdx];
4415 for (const Formula &F : LU.Formulae) {
4416 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4417 continue;
4418
4419 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4420 if (!LUThatHas)
4421 continue;
4422
4423 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4424 LU.Kind, LU.AccessTy))
4425 continue;
4426
4427 LLVM_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)
;
4428
4429 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4430
4431 // Transfer the fixups of LU to LUThatHas.
4432 for (LSRFixup &Fixup : LU.Fixups) {
4433 Fixup.Offset += F.BaseOffset;
4434 LUThatHas->pushFixup(Fixup);
4435 LLVM_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)
;
4436 }
4437
4438 // Delete formulae from the new use which are no longer legal.
4439 bool Any = false;
4440 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4441 Formula &F = LUThatHas->Formulae[i];
4442 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4443 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4444 LLVM_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)
;
4445 LUThatHas->DeleteFormula(F);
4446 --i;
4447 --e;
4448 Any = true;
4449 }
4450 }
4451
4452 if (Any)
4453 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4454
4455 // Delete the old use.
4456 DeleteUse(LU, LUIdx);
4457 --LUIdx;
4458 --NumUses;
4459 break;
4460 }
4461 }
4462
4463 LLVM_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)
;
4464}
4465
4466/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4467/// we've done more filtering, as it may be able to find more formulae to
4468/// eliminate.
4469void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4470 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4471 LLVM_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)
;
4472
4473 LLVM_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)
4474 "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)
;
4475
4476 FilterOutUndesirableDedicatedRegisters();
4477
4478 LLVM_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)
;
4479 }
4480}
4481
4482/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4483/// Pick the best one and delete the others.
4484/// This narrowing heuristic is to keep as many formulae with different
4485/// Scale and ScaledReg pair as possible while narrowing the search space.
4486/// The benefit is that it is more likely to find out a better solution
4487/// from a formulae set with more Scale and ScaledReg variations than
4488/// a formulae set with the same Scale and ScaledReg. The picking winner
4489/// reg heuristic will often keep the formulae with the same Scale and
4490/// ScaledReg and filter others, and we want to avoid that if possible.
4491void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4492 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4493 return;
4494
4495 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
4496 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 choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
4497 "Narrowing the search space by choosing the best Formula "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
4498 "from the Formulae with the same Scale and ScaledReg.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
;
4499
4500 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4501 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4502
4503 BestFormulaeTy BestFormulae;
4504#ifndef NDEBUG
4505 bool ChangedFormulae = false;
4506#endif
4507 DenseSet<const SCEV *> VisitedRegs;
4508 SmallPtrSet<const SCEV *, 16> Regs;
4509
4510 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4511 LSRUse &LU = Uses[LUIdx];
4512 LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
4513 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4514
4515 // Return true if Formula FA is better than Formula FB.
4516 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4517 // First we will try to choose the Formula with fewer new registers.
4518 // For a register used by current Formula, the more the register is
4519 // shared among LSRUses, the less we increase the register number
4520 // counter of the formula.
4521 size_t FARegNum = 0;
4522 for (const SCEV *Reg : FA.BaseRegs) {
4523 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4524 FARegNum += (NumUses - UsedByIndices.count() + 1);
4525 }
4526 size_t FBRegNum = 0;
4527 for (const SCEV *Reg : FB.BaseRegs) {
4528 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4529 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4530 }
4531 if (FARegNum != FBRegNum)
4532 return FARegNum < FBRegNum;
4533
4534 // If the new register numbers are the same, choose the Formula with
4535 // less Cost.
4536 Cost CostFA, CostFB;
4537 Regs.clear();
4538 CostFA.RateFormula(TTI, FA, Regs, VisitedRegs, L, SE, DT, LU);
4539 Regs.clear();
4540 CostFB.RateFormula(TTI, FB, Regs, VisitedRegs, L, SE, DT, LU);
4541 return CostFA.isLess(CostFB, TTI);
4542 };
4543
4544 bool Any = false;
4545 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4546 ++FIdx) {
4547 Formula &F = LU.Formulae[FIdx];
4548 if (!F.ScaledReg)
4549 continue;
4550 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4551 if (P.second)
4552 continue;
4553
4554 Formula &Best = LU.Formulae[P.first->second];
4555 if (IsBetterThan(F, Best))
4556 std::swap(F, Best);
4557 LLVM_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)
4558 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)
4559 " in favor of formula ";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)
4560 Best.print(dbgs()); 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)
;
4561#ifndef NDEBUG
4562 ChangedFormulae = true;
4563#endif
4564 LU.DeleteFormula(F);
4565 --FIdx;
4566 --NumForms;
4567 Any = true;
4568 }
4569 if (Any)
4570 LU.RecomputeRegs(LUIdx, RegUses);
4571
4572 // Reset this to prepare for the next use.
4573 BestFormulae.clear();
4574 }
4575
4576 LLVM_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)
4577 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)
4578 "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)
4579 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)
4580 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4581}
4582
4583/// The function delete formulas with high registers number expectation.
4584/// Assuming we don't know the value of each formula (already delete
4585/// all inefficient), generate probability of not selecting for each
4586/// register.
4587/// For example,
4588/// Use1:
4589/// reg(a) + reg({0,+,1})
4590/// reg(a) + reg({-1,+,1}) + 1
4591/// reg({a,+,1})
4592/// Use2:
4593/// reg(b) + reg({0,+,1})
4594/// reg(b) + reg({-1,+,1}) + 1
4595/// reg({b,+,1})
4596/// Use3:
4597/// reg(c) + reg(b) + reg({0,+,1})
4598/// reg(c) + reg({b,+,1})
4599///
4600/// Probability of not selecting
4601/// Use1 Use2 Use3
4602/// reg(a) (1/3) * 1 * 1
4603/// reg(b) 1 * (1/3) * (1/2)
4604/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4605/// reg({-1,+,1}) (2/3) * (2/3) * 1
4606/// reg({a,+,1}) (2/3) * 1 * 1
4607/// reg({b,+,1}) 1 * (2/3) * (2/3)
4608/// reg(c) 1 * 1 * 0
4609///
4610/// Now count registers number mathematical expectation for each formula:
4611/// Note that for each use we exclude probability if not selecting for the use.
4612/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4613/// probabilty 1/3 of not selecting for Use1).
4614/// Use1:
4615/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4616/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4617/// reg({a,+,1}) 1
4618/// Use2:
4619/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4620/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4621/// reg({b,+,1}) 2/3
4622/// Use3:
4623/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4624/// reg(c) + reg({b,+,1}) 1 + 2/3
4625void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4626 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4627 return;
4628 // Ok, we have too many of formulae on our hands to conveniently handle.
4629 // Use a rough heuristic to thin out the list.
4630
4631 // Set of Regs wich will be 100% used in final solution.
4632 // Used in each formula of a solution (in example above this is reg(c)).
4633 // We can skip them in calculations.
4634 SmallPtrSet<const SCEV *, 4> UniqRegs;
4635 LLVM_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)
;
4636
4637 // Map each register to probability of not selecting
4638 DenseMap <const SCEV *, float> RegNumMap;
4639 for (const SCEV *Reg : RegUses) {
4640 if (UniqRegs.count(Reg))
4641 continue;
4642 float PNotSel = 1;
4643 for (const LSRUse &LU : Uses) {
4644 if (!LU.Regs.count(Reg))
4645 continue;
4646 float P = LU.getNotSelectedProbability(Reg);
4647 if (P != 0.0)
4648 PNotSel *= P;
4649 else
4650 UniqRegs.insert(Reg);
4651 }
4652 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4653 }
4654
4655 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4656 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)
;
4657
4658 // Delete formulas where registers number expectation is high.
4659 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4660 LSRUse &LU = Uses[LUIdx];
4661 // If nothing to delete - continue.
4662 if (LU.Formulae.size() < 2)
4663 continue;
4664 // This is temporary solution to test performance. Float should be
4665 // replaced with round independent type (based on integers) to avoid
4666 // different results for different target builds.
4667 float FMinRegNum = LU.Formulae[0].getNumRegs();
4668 float FMinARegNum = LU.Formulae[0].getNumRegs();
4669 size_t MinIdx = 0;
4670 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4671 Formula &F = LU.Formulae[i];
4672 float FRegNum = 0;
4673 float FARegNum = 0;
4674 for (const SCEV *BaseReg : F.BaseRegs) {
4675 if (UniqRegs.count(BaseReg))
4676 continue;
4677 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4678 if (isa<SCEVAddRecExpr>(BaseReg))
4679 FARegNum +=
4680 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4681 }
4682 if (const SCEV *ScaledReg = F.ScaledReg) {
4683 if (!UniqRegs.count(ScaledReg)) {
4684 FRegNum +=
4685 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4686 if (isa<SCEVAddRecExpr>(ScaledReg))
4687 FARegNum +=
4688 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4689 }
4690 }
4691 if (FMinRegNum > FRegNum ||
4692 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4693 FMinRegNum = FRegNum;
4694 FMinARegNum = FARegNum;
4695 MinIdx = i;
4696 }
4697 }
4698 LLVM_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)
4699 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)
;
4700 if (MinIdx != 0)
4701 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4702 while (LU.Formulae.size() != 1) {
4703 LLVM_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
)
4704 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4705 LU.Formulae.pop_back();
4706 }
4707 LU.RecomputeRegs(LUIdx, RegUses);
4708 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4708, __PRETTY_FUNCTION__))
;
4709 Formula &F = LU.Formulae[0];
4710 LLVM_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)
;
4711 // When we choose the formula, the regs become unique.
4712 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4713 if (F.ScaledReg)
4714 UniqRegs.insert(F.ScaledReg);
4715 }
4716 LLVM_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)
;
4717}
4718
4719/// Pick a register which seems likely to be profitable, and then in any use
4720/// which has any reference to that register, delete all formulae which do not
4721/// reference that register.
4722void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4723 // With all other options exhausted, loop until the system is simple
4724 // enough to handle.
4725 SmallPtrSet<const SCEV *, 4> Taken;
4726 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4727 // Ok, we have too many of formulae on our hands to conveniently handle.
4728 // Use a rough heuristic to thin out the list.
4729 LLVM_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)
;
4730
4731 // Pick the register which is used by the most LSRUses, which is likely
4732 // to be a good reuse register candidate.
4733 const SCEV *Best = nullptr;
4734 unsigned BestNum = 0;
4735 for (const SCEV *Reg : RegUses) {
4736 if (Taken.count(Reg))
4737 continue;
4738 if (!Best) {
4739 Best = Reg;
4740 BestNum = RegUses.getUsedByIndices(Reg).count();
4741 } else {
4742 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4743 if (Count > BestNum) {
4744 Best = Reg;
4745 BestNum = Count;
4746 }
4747 }
4748 }
4749
4750 LLVM_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)
4751 << " 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)
;
4752 Taken.insert(Best);
4753
4754 // In any use with formulae which references this register, delete formulae
4755 // which don't reference it.
4756 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4757 LSRUse &LU = Uses[LUIdx];
4758 if (!LU.Regs.count(Best)) continue;
4759
4760 bool Any = false;
4761 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4762 Formula &F = LU.Formulae[i];
4763 if (!F.referencesReg(Best)) {
4764 LLVM_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)
;
4765 LU.DeleteFormula(F);
4766 --e;
4767 --i;
4768 Any = true;
4769 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?\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4769, __PRETTY_FUNCTION__))
;
4770 continue;
4771 }
4772 }
4773
4774 if (Any)
4775 LU.RecomputeRegs(LUIdx, RegUses);
4776 }
4777
4778 LLVM_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)
;
4779 }
4780}
4781
4782/// If there are an extraordinary number of formulae to choose from, use some
4783/// rough heuristics to prune down the number of formulae. This keeps the main
4784/// solver from taking an extraordinary amount of time in some worst-case
4785/// scenarios.
4786void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4787 NarrowSearchSpaceByDetectingSupersets();
4788 NarrowSearchSpaceByCollapsingUnrolledCode();
4789 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4790 if (FilterSameScaledReg)
4791 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4792 if (LSRExpNarrow)
4793 NarrowSearchSpaceByDeletingCostlyFormulas();
4794 else
4795 NarrowSearchSpaceByPickingWinnerRegs();
4796}
4797
4798/// This is the recursive solver.
4799void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4800 Cost &SolutionCost,
4801 SmallVectorImpl<const Formula *> &Workspace,
4802 const Cost &CurCost,
4803 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4804 DenseSet<const SCEV *> &VisitedRegs) const {
4805 // Some ideas:
4806 // - prune more:
4807 // - use more aggressive filtering
4808 // - sort the formula so that the most profitable solutions are found first
4809 // - sort the uses too
4810 // - search faster:
4811 // - don't compute a cost, and then compare. compare while computing a cost
4812 // and bail early.
4813 // - track register sets with SmallBitVector
4814
4815 const LSRUse &LU = Uses[Workspace.size()];
4816
4817 // If this use references any register that's already a part of the
4818 // in-progress solution, consider it a requirement that a formula must
4819 // reference that register in order to be considered. This prunes out
4820 // unprofitable searching.
4821 SmallSetVector<const SCEV *, 4> ReqRegs;
4822 for (const SCEV *S : CurRegs)
4823 if (LU.Regs.count(S))
4824 ReqRegs.insert(S);
4825
4826 SmallPtrSet<const SCEV *, 16> NewRegs;
4827 Cost NewCost;
4828 for (const Formula &F : LU.Formulae) {
4829 // Ignore formulae which may not be ideal in terms of register reuse of
4830 // ReqRegs. The formula should use all required registers before
4831 // introducing new ones.
4832 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4833 for (const SCEV *Reg : ReqRegs) {
4834 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4835 is_contained(F.BaseRegs, Reg)) {
4836 --NumReqRegsToFind;
4837 if (NumReqRegsToFind == 0)
4838 break;
4839 }
4840 }
4841 if (NumReqRegsToFind != 0) {
4842 // If none of the formulae satisfied the required registers, then we could
4843 // clear ReqRegs and try again. Currently, we simply give up in this case.
4844 continue;
4845 }
4846
4847 // Evaluate the cost of the current formula. If it's already worse than
4848 // the current best, prune the search at that point.
4849 NewCost = CurCost;
4850 NewRegs = CurRegs;
4851 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4852 if (NewCost.isLess(SolutionCost, TTI)) {
4853 Workspace.push_back(&F);
4854 if (Workspace.size() != Uses.size()) {
4855 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4856 NewRegs, VisitedRegs);
4857 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4858 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4859 } else {
4860 LLVM_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)
4861 dbgs() << ".\n Regs:"; for (const SCEV *Sdo { 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)
4862 : NewRegs) 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)
4863 << ' ' << *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)
4864 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)
;
4865
4866 SolutionCost = NewCost;
4867 Solution = Workspace;
4868 }
4869 Workspace.pop_back();
4870 }
4871 }
4872}
4873
4874/// Choose one formula from each use. Return the results in the given Solution
4875/// vector.
4876void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4877 SmallVector<const Formula *, 8> Workspace;
4878 Cost SolutionCost;
4879 SolutionCost.Lose();
4880 Cost CurCost;
4881 SmallPtrSet<const SCEV *, 16> CurRegs;
4882 DenseSet<const SCEV *> VisitedRegs;
4883 Workspace.reserve(Uses.size());
4884
4885 // SolveRecurse does all the work.
4886 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4887 CurRegs, VisitedRegs);
4888 if (Solution.empty()) {
4889 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
4890 return;
4891 }
4892
4893 // Ok, we've now made all our decisions.
4894 LLVM_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
)
4895 "The chosen solution requires ";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
)
4896 SolutionCost.print(dbgs()); 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
)
4897 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
)
4898 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
)
4899 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
)
4900 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
)
4901 " ";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
)
4902 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
)
4903 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
)
4904 })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
)
;
4905
4906 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4906, __PRETTY_FUNCTION__))
;
4907}
4908
4909/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4910/// we can go while still being dominated by the input positions. This helps
4911/// canonicalize the insert position, which encourages sharing.
4912BasicBlock::iterator
4913LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4914 const SmallVectorImpl<Instruction *> &Inputs)
4915 const {
4916 Instruction *Tentative = &*IP;
4917 while (true) {
4918 bool AllDominate = true;
4919 Instruction *BetterPos = nullptr;
4920 // Don't bother attempting to insert before a catchswitch, their basic block
4921 // cannot have other non-PHI instructions.
4922 if (isa<CatchSwitchInst>(Tentative))
4923 return IP;
4924
4925 for (Instruction *Inst : Inputs) {
4926 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4927 AllDominate = false;
4928 break;
4929 }
4930 // Attempt to find an insert position in the middle of the block,
4931 // instead of at the end, so that it can be used for other expansions.
4932 if (Tentative->getParent() == Inst->getParent() &&
4933 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4934 BetterPos = &*std::next(BasicBlock::iterator(Inst));
4935 }
4936 if (!AllDominate)
4937 break;
4938 if (BetterPos)
4939 IP = BetterPos->getIterator();
4940 else
4941 IP = Tentative->getIterator();
4942
4943 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4944 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4945
4946 BasicBlock *IDom;
4947 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4948 if (!Rung) return IP;
4949 Rung = Rung->getIDom();
4950 if (!Rung) return IP;
4951 IDom = Rung->getBlock();
4952
4953 // Don't climb into a loop though.
4954 const Loop *IDomLoop = LI.getLoopFor(IDom);
4955 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4956 if (IDomDepth <= IPLoopDepth &&
4957 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4958 break;
4959 }
4960
4961 Tentative = IDom->getTerminator();
4962 }
4963
4964 return IP;
4965}
4966
4967/// Determine an input position which will be dominated by the operands and
4968/// which will dominate the result.
4969BasicBlock::iterator
4970LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4971 const LSRFixup &LF,
4972 const LSRUse &LU,
4973 SCEVExpander &Rewriter) const {
4974 // Collect some instructions which must be dominated by the
4975 // expanding replacement. These must be dominated by any operands that
4976 // will be required in the expansion.
4977 SmallVector<Instruction *, 4> Inputs;
4978 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4979 Inputs.push_back(I);
4980 if (LU.Kind == LSRUse::ICmpZero)
4981 if (Instruction *I =
4982 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4983 Inputs.push_back(I);
4984 if (LF.PostIncLoops.count(L)) {
4985 if (LF.isUseFullyOutsideLoop(L))
4986 Inputs.push_back(L->getLoopLatch()->getTerminator());
4987 else
4988 Inputs.push_back(IVIncInsertPos);
4989 }
4990 // The expansion must also be dominated by the increment positions of any
4991 // loops it for which it is using post-inc mode.
4992 for (const Loop *PIL : LF.PostIncLoops) {
4993 if (PIL == L) continue;
4994
4995 // Be dominated by the loop exit.
4996 SmallVector<BasicBlock *, 4> ExitingBlocks;
4997 PIL->getExitingBlocks(ExitingBlocks);
4998 if (!ExitingBlocks.empty()) {
4999 BasicBlock *BB = ExitingBlocks[0];
5000 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5001 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5002 Inputs.push_back(BB->getTerminator());
5003 }
5004 }
5005
5006 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5008, __PRETTY_FUNCTION__))
5007 && !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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5008, __PRETTY_FUNCTION__))
5008 "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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5008, __PRETTY_FUNCTION__))
;
5009
5010 // Then, climb up the immediate dominator tree as far as we can go while
5011 // still being dominated by the input positions.
5012 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5013
5014 // Don't insert instructions before PHI nodes.
5015 while (isa<PHINode>(IP)) ++IP;
5016
5017 // Ignore landingpad instructions.
5018 while (IP->isEHPad()) ++IP;
5019
5020 // Ignore debug intrinsics.
5021 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5022
5023 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5024 // IP consistent across expansions and allows the previously inserted
5025 // instructions to be reused by subsequent expansion.
5026 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5027 ++IP;
5028
5029 return IP;
5030}
5031
5032/// Emit instructions for the leading candidate expression for this LSRUse (this
5033/// is called "expanding").
5034Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5035 const Formula &F, BasicBlock::iterator IP,
5036 SCEVExpander &Rewriter,
5037 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5038 if (LU.RigidFormula)
5039 return LF.OperandValToReplace;
5040
5041 // Determine an input position which will be dominated by the operands and
5042 // which will dominate the result.
5043 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5044 Rewriter.setInsertPoint(&*IP);
5045
5046 // Inform the Rewriter if we have a post-increment use, so that it can
5047 // perform an advantageous expansion.
5048 Rewriter.setPostInc(LF.PostIncLoops);
5049
5050 // This is the type that the user actually needs.
5051 Type *OpTy = LF.OperandValToReplace->getType();
5052 // This will be the type that we'll initially expand to.
5053 Type *Ty = F.getType();
5054 if (!Ty)
5055 // No type known; just expand directly to the ultimate type.
5056 Ty = OpTy;
5057 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5058 // Expand directly to the ultimate type if it's the right size.
5059 Ty = OpTy;
5060 // This is the type to do integer arithmetic in.
5061 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5062
5063 // Build up a list of operands to add together to form the full base.
5064 SmallVector<const SCEV *, 8> Ops;
5065
5066 // Expand the BaseRegs portion.
5067 for (const SCEV *Reg : F.BaseRegs) {
5068 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5068, __PRETTY_FUNCTION__))
;
5069
5070 // If we're expanding for a post-inc user, make the post-inc adjustment.
5071 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5072 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5073 }
5074
5075 // Expand the ScaledReg portion.
5076 Value *ICmpScaledV = nullptr;
5077 if (F.Scale != 0) {
5078 const SCEV *ScaledS = F.ScaledReg;
5079
5080 // If we're expanding for a post-inc user, make the post-inc adjustment.
5081 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5082 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5083
5084 if (LU.Kind == LSRUse::ICmpZero) {
5085 // Expand ScaleReg as if it was part of the base regs.
5086 if (F.Scale == 1)
5087 Ops.push_back(
5088 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5089 else {
5090 // An interesting way of "folding" with an icmp is to use a negated
5091 // scale, which we'll implement by inserting it into the other operand
5092 // of the icmp.
5093 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5094, __PRETTY_FUNCTION__))
5094 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5094, __PRETTY_FUNCTION__))
;
5095 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5096 }
5097 } else {
5098 // Otherwise just expand the scaled register and an explicit scale,
5099 // which is expected to be matched as part of the address.
5100
5101 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5102 // Unless the addressing mode will not be folded.
5103 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5104 isAMCompletelyFolded(TTI, LU, F)) {
5105 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5106 Ops.clear();
5107 Ops.push_back(SE.getUnknown(FullV));
5108 }
5109 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5110 if (F.Scale != 1)
5111 ScaledS =
5112 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5113 Ops.push_back(ScaledS);
5114 }
5115 }
5116
5117 // Expand the GV portion.
5118 if (F.BaseGV) {
5119 // Flush the operand list to suppress SCEVExpander hoisting.
5120 if (!Ops.empty()) {
5121 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5122 Ops.clear();
5123 Ops.push_back(SE.getUnknown(FullV));
5124 }
5125 Ops.push_back(SE.getUnknown(F.BaseGV));
5126 }
5127
5128 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5129 // unfolded offsets. LSR assumes they both live next to their uses.
5130 if (!Ops.empty()) {
5131 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5132 Ops.clear();
5133 Ops.push_back(SE.getUnknown(FullV));
5134 }
5135
5136 // Expand the immediate portion.
5137 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5138 if (Offset != 0) {
5139 if (LU.Kind == LSRUse::ICmpZero) {
5140 // The other interesting way of "folding" with an ICmpZero is to use a
5141 // negated immediate.
5142 if (!ICmpScaledV)
5143 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5144 else {
5145 Ops.push_back(SE.getUnknown(ICmpScaledV));
5146 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5147 }
5148 } else {
5149 // Just add the immediate values. These again are expected to be matched
5150 // as part of the address.
5151 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5152 }
5153 }
5154
5155 // Expand the unfolded offset portion.
5156 int64_t UnfoldedOffset = F.UnfoldedOffset;
5157 if (UnfoldedOffset != 0) {
5158 // Just add the immediate values.
5159 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5160 UnfoldedOffset)));
5161 }
5162
5163 // Emit instructions summing all the operands.
5164 const SCEV *FullS = Ops.empty() ?
5165 SE.getConstant(IntTy, 0) :
5166 SE.getAddExpr(Ops);
5167 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5168
5169 // We're done expanding now, so reset the rewriter.
5170 Rewriter.clearPostInc();
5171
5172 // An ICmpZero Formula represents an ICmp which we're handling as a
5173 // comparison against zero. Now that we've expanded an expression for that
5174 // form, update the ICmp's other operand.
5175 if (LU.Kind == LSRUse::ICmpZero) {
5176 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5177 DeadInsts.emplace_back(CI->getOperand(1));
5178 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5179, __PRETTY_FUNCTION__))
5179 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5179, __PRETTY_FUNCTION__))
;
5180 if (F.Scale == -1) {
5181 if (ICmpScaledV->getType() != OpTy) {
5182 Instruction *Cast =
5183 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5184 OpTy, false),
5185 ICmpScaledV, OpTy, "tmp", CI);
5186 ICmpScaledV = Cast;
5187 }
5188 CI->setOperand(1, ICmpScaledV);
5189 } else {
5190 // A scale of 1 means that the scale has been expanded as part of the
5191 // base regs.
5192 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5194, __PRETTY_FUNCTION__))
5193 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5194, __PRETTY_FUNCTION__))
5194 "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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5194, __PRETTY_FUNCTION__))
;
5195 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5196 -(uint64_t)Offset);
5197 if (C->getType() != OpTy)
5198 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5199 OpTy, false),
5200 C, OpTy);
5201
5202 CI->setOperand(1, C);
5203 }
5204 }
5205
5206 return FullV;
5207}
5208
5209/// Helper for Rewrite. PHI nodes are special because the use of their operands
5210/// effectively happens in their predecessor blocks, so the expression may need
5211/// to be expanded in multiple places.
5212void LSRInstance::RewriteForPHI(
5213 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5214 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5215 DenseMap<BasicBlock *, Value *> Inserted;
5216 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5217 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5218 BasicBlock *BB = PN->getIncomingBlock(i);
5219
5220 // If this is a critical edge, split the edge so that we do not insert
5221 // the code on all predecessor/successor paths. We do this unless this
5222 // is the canonical backedge for this loop, which complicates post-inc
5223 // users.
5224 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5225 !isa<IndirectBrInst>(BB->getTerminator()) &&
5226 !isa<CatchSwitchInst>(BB->getTerminator())) {
5227 BasicBlock *Parent = PN->getParent();
5228 Loop *PNLoop = LI.getLoopFor(Parent);
5229 if (!PNLoop || Parent != PNLoop->getHeader()) {
5230 // Split the critical edge.
5231 BasicBlock *NewBB = nullptr;
5232 if (!Parent->isLandingPad()) {
5233 NewBB = SplitCriticalEdge(BB, Parent,
5234 CriticalEdgeSplittingOptions(&DT, &LI)
5235 .setMergeIdenticalEdges()
5236 .setDontDeleteUselessPHIs());
5237 } else {
5238 SmallVector<BasicBlock*, 2> NewBBs;
5239 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5240 NewBB = NewBBs[0];
5241 }
5242 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5243 // phi predecessors are identical. The simple thing to do is skip
5244 // splitting in this case rather than complicate the API.
5245 if (NewBB) {
5246 // If PN is outside of the loop and BB is in the loop, we want to
5247 // move the block to be immediately before the PHI block, not
5248 // immediately after BB.
5249 if (L->contains(BB) && !L->contains(PN))
5250 NewBB->moveBefore(PN->getParent());
5251
5252 // Splitting the edge can reduce the number of PHI entries we have.
5253 e = PN->getNumIncomingValues();
5254 BB = NewBB;
5255 i = PN->getBasicBlockIndex(BB);
5256 }
5257 }
5258 }
5259
5260 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5261 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5262 if (!Pair.second)
5263 PN->setIncomingValue(i, Pair.first->second);
5264 else {
5265 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5266 Rewriter, DeadInsts);
5267
5268 // If this is reuse-by-noop-cast, insert the noop cast.
5269 Type *OpTy = LF.OperandValToReplace->getType();
5270 if (FullV->getType() != OpTy)
5271 FullV =
5272 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5273 OpTy, false),
5274 FullV, LF.OperandValToReplace->getType(),
5275 "tmp", BB->getTerminator());
5276
5277 PN->setIncomingValue(i, FullV);
5278 Pair.first->second = FullV;
5279 }
5280 }
5281}
5282
5283/// Emit instructions for the leading candidate expression for this LSRUse (this
5284/// is called "expanding"), and update the UserInst to reference the newly
5285/// expanded value.
5286void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5287 const Formula &F, SCEVExpander &Rewriter,
5288 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5289 // First, find an insertion point that dominates UserInst. For PHI nodes,
5290 // find the nearest block which dominates all the relevant uses.
5291 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5292 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5293 } else {
5294 Value *FullV =
5295 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5296
5297 // If this is reuse-by-noop-cast, insert the noop cast.
5298 Type *OpTy = LF.OperandValToReplace->getType();
5299 if (FullV->getType() != OpTy) {
5300 Instruction *Cast =
5301 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5302 FullV, OpTy, "tmp", LF.UserInst);
5303 FullV = Cast;
5304 }
5305
5306 // Update the user. ICmpZero is handled specially here (for now) because
5307 // Expand may have updated one of the operands of the icmp already, and
5308 // its new value may happen to be equal to LF.OperandValToReplace, in
5309 // which case doing replaceUsesOfWith leads to replacing both operands
5310 // with the same value. TODO: Reorganize this.
5311 if (LU.Kind == LSRUse::ICmpZero)
5312 LF.UserInst->setOperand(0, FullV);
5313 else
5314 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5315 }
5316
5317 DeadInsts.emplace_back(LF.OperandValToReplace);
5318}
5319
5320/// Rewrite all the fixup locations with new values, following the chosen
5321/// solution.
5322void LSRInstance::ImplementSolution(
5323 const SmallVectorImpl<const Formula *> &Solution) {
5324 // Keep track of instructions we may have made dead, so that
5325 // we can remove them after we are done working.
5326 SmallVector<WeakTrackingVH, 16> DeadInsts;
5327
5328 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5329 "lsr");
5330#ifndef NDEBUG
5331 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5332#endif
5333 Rewriter.disableCanonicalMode();
5334 Rewriter.enableLSRMode();
5335 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5336
5337 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5338 for (const IVChain &Chain : IVChainVec) {
1
Assuming '__begin1' is equal to '__end1'
5339 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5340 Rewriter.setChainedPhi(PN);
5341 }
5342
5343 // Expand the new value definitions and update the users.
5344 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
2
Assuming 'LUIdx' is equal to 'NumUses'
3
Loop condition is false. Execution continues on line 5350
5345 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5346 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5347 Changed = true;
5348 }
5349
5350 for (const IVChain &Chain : IVChainVec) {
4
Assuming '__begin1' is not equal to '__end1'
5351 GenerateIVChain(Chain, Rewriter, DeadInsts);
5
Calling 'LSRInstance::GenerateIVChain'
5352 Changed = true;
5353 }
5354 // Clean up after ourselves. This must be done before deleting any
5355 // instructions.
5356 Rewriter.clear();
5357
5358 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5359}
5360
5361LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5362 DominatorTree &DT, LoopInfo &LI,
5363 const TargetTransformInfo &TTI)
5364 : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L) {
5365 // If LoopSimplify form is not available, stay out of trouble.
5366 if (!L->isLoopSimplifyForm())
5367 return;
5368
5369 // If there's no interesting work to be done, bail early.
5370 if (IU.empty()) return;
5371
5372 // If there's too much analysis to be done, bail early. We won't be able to
5373 // model the problem anyway.
5374 unsigned NumUsers = 0;
5375 for (const IVStrideUse &U : IU) {
5376 if (++NumUsers > MaxIVUsers) {
5377 (void)U;
5378 LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
5379 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5380 return;
5381 }
5382 // Bail out if we have a PHI on an EHPad that gets a value from a
5383 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5384 // no good place to stick any instructions.
5385 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5386 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5387 if (isa<FuncletPadInst>(FirstNonPHI) ||
5388 isa<CatchSwitchInst>(FirstNonPHI))
5389 for (BasicBlock *PredBB : PN->blocks())
5390 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5391 return;
5392 }
5393 }
5394
5395#ifndef NDEBUG
5396 // All dominating loops must have preheaders, or SCEVExpander may not be able
5397 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5398 //
5399 // IVUsers analysis should only create users that are dominated by simple loop
5400 // headers. Since this loop should dominate all of its users, its user list
5401 // should be empty if this loop itself is not within a simple loop nest.
5402 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5403 Rung; Rung = Rung->getIDom()) {
5404 BasicBlock *BB = Rung->getBlock();
5405 const Loop *DomLoop = LI.getLoopFor(BB);
5406 if (DomLoop && DomLoop->getHeader() == BB) {
5407 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\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5407, __PRETTY_FUNCTION__))
;
5408 }
5409 }
5410#endif // DEBUG
5411
5412 LLVM_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)
5413 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)
5414 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5415
5416 // First, perform some low-level loop optimizations.
5417 OptimizeShadowIV();
5418 OptimizeLoopTermCond();
5419
5420 // If loop preparation eliminates all interesting IV users, bail.
5421 if (IU.empty()) return;
5422
5423 // Skip nested loops until we can model them better with formulae.
5424 if (!L->empty()) {
5425 LLVM_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)
;
5426 return;
5427 }
5428
5429 // Start collecting data and preparing for the solver.
5430 CollectChains();
5431 CollectInterestingTypesAndFactors();
5432 CollectFixupsAndInitialFormulae();
5433 CollectLoopInvariantFixupsAndFormulae();
5434
5435 if (Uses.empty())
5436 return;
5437
5438 LLVM_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
)
5439 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
;
5440
5441 // Now use the reuse data to generate a bunch of interesting ways
5442 // to formulate the values needed for the uses.
5443 GenerateAllReuseFormulae();
5444
5445 FilterOutUndesirableDedicatedRegisters();
5446 NarrowSearchSpaceUsingHeuristics();
5447
5448 SmallVector<const Formula *, 8> Solution;
5449 Solve(Solution);
5450
5451 // Release memory that is no longer needed.
5452 Factors.clear();
5453 Types.clear();
5454 RegUses.clear();
5455
5456 if (Solution.empty())
5457 return;
5458
5459#ifndef NDEBUG
5460 // Formulae should be legal.
5461 for (const LSRUse &LU : Uses) {
5462 for (const Formula &F : LU.Formulae)
5463 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5464, __PRETTY_FUNCTION__))
5464 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!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5464, __PRETTY_FUNCTION__))
;
5465 };
5466#endif
5467
5468 // Now that we've decided what we want, make it so.
5469 ImplementSolution(Solution);
5470}
5471
5472#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5473void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5474 if (Factors.empty() && Types.empty()) return;
5475
5476 OS << "LSR has identified the following interesting factors and types: ";
5477 bool First = true;
5478
5479 for (int64_t Factor : Factors) {
5480 if (!First) OS << ", ";
5481 First = false;
5482 OS << '*' << Factor;
5483 }
5484
5485 for (Type *Ty : Types) {
5486 if (!First) OS << ", ";
5487 First = false;
5488 OS << '(' << *Ty << ')';
5489 }
5490 OS << '\n';
5491}
5492
5493void LSRInstance::print_fixups(raw_ostream &OS) const {
5494 OS << "LSR is examining the following fixup sites:\n";
5495 for (const LSRUse &LU : Uses)
5496 for (const LSRFixup &LF : LU.Fixups) {
5497 dbgs() << " ";
5498 LF.print(OS);
5499 OS << '\n';
5500 }
5501}
5502
5503void LSRInstance::print_uses(raw_ostream &OS) const {
5504 OS << "LSR is examining the following uses:\n";
5505 for (const LSRUse &LU : Uses) {
5506 dbgs() << " ";
5507 LU.print(OS);
5508 OS << '\n';
5509 for (const Formula &F : LU.Formulae) {
5510 OS << " ";
5511 F.print(OS);
5512 OS << '\n';
5513 }
5514 }
5515}
5516
5517void LSRInstance::print(raw_ostream &OS) const {
5518 print_factors_and_types(OS);
5519 print_fixups(OS);
5520 print_uses(OS);
5521}
5522
5523LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRInstance::dump() const {
5524 print(errs()); errs() << '\n';
5525}
5526#endif
5527
5528namespace {
5529
5530class LoopStrengthReduce : public LoopPass {
5531public:
5532 static char ID; // Pass ID, replacement for typeid
5533
5534 LoopStrengthReduce();
5535
5536private:
5537 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5538 void getAnalysisUsage(AnalysisUsage &AU) const override;
5539};
5540
5541} // end anonymous namespace
5542
5543LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5544 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5545}
5546
5547void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5548 // We split critical edges, so we change the CFG. However, we do update
5549 // many analyses if they are around.
5550 AU.addPreservedID(LoopSimplifyID);
5551
5552 AU.addRequired<LoopInfoWrapperPass>();
5553 AU.addPreserved<LoopInfoWrapperPass>();
5554 AU.addRequiredID(LoopSimplifyID);
5555 AU.addRequired<DominatorTreeWrapperPass>();
5556 AU.addPreserved<DominatorTreeWrapperPass>();
5557 AU.addRequired<ScalarEvolutionWrapperPass>();
5558 AU.addPreserved<ScalarEvolutionWrapperPass>();
5559 // Requiring LoopSimplify a second time here prevents IVUsers from running
5560 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5561 AU.addRequiredID(LoopSimplifyID);
5562 AU.addRequired<IVUsersWrapperPass>();
5563 AU.addPreserved<IVUsersWrapperPass>();
5564 AU.addRequired<TargetTransformInfoWrapperPass>();
5565}
5566
5567static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5568 DominatorTree &DT, LoopInfo &LI,
5569 const TargetTransformInfo &TTI) {
5570 bool Changed = false;
5571
5572 // Run the main LSR transformation.
5573 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5574
5575 // Remove any extra phis created by processing inner loops.
5576 Changed |= DeleteDeadPHIs(L->getHeader());
5577 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5578 SmallVector<WeakTrackingVH, 16> DeadInsts;
5579 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5580 SCEVExpander Rewriter(SE, DL, "lsr");
5581#ifndef NDEBUG
5582 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5583#endif
5584 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5585 if (numFolded) {
5586 Changed = true;
5587 DeleteTriviallyDeadInstructions(DeadInsts);
5588 DeleteDeadPHIs(L->getHeader());
5589 }
5590 }
5591 return Changed;
5592}
5593
5594bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5595 if (skipLoop(L))
5596 return false;
5597
5598 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5599 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5600 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5601 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5602 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5603 *L->getHeader()->getParent());
5604 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5605}
5606
5607PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5608 LoopStandardAnalysisResults &AR,
5609 LPMUpdater &) {
5610 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5611 AR.DT, AR.LI, AR.TTI))
5612 return PreservedAnalyses::all();
5613
5614 return getLoopPassPreservedAnalyses();
5615}
5616
5617char LoopStrengthReduce::ID = 0;
5618
5619INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
5620 "Loop Strength Reduction", false, false)static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
5621INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
5622INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
5623INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
5624INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)initializeIVUsersWrapperPassPass(Registry);
5625INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
5626INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
5627INITIALIZE_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)); }
5628 "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)); }
5629
5630Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }