Bug Summary

File:lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 4839, column 5
Forming reference to null pointer

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 -mframe-pointer=none -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~svn372306/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-10~svn372306/build-llvm/include -I /build/llvm-toolchain-snapshot-10~svn372306/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/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/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++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~svn372306/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~svn372306=. -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 -faddrsig -o /tmp/scan-build-2019-09-19-172240-30738-1 -x c++ /build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp
1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This transformation analyzes and transforms the induction variables (and
10// computations derived from them) into forms suitable for efficient execution
11// on the target.
12//
13// This pass performs a strength reduction on array references inside loops that
14// have as one or more of their components the loop induction variable, it
15// rewrites expressions to take advantage of scaled-index addressing modes
16// available on the target, and it performs a variety of other optimizations
17// related to loop induction variables.
18//
19// Terminology note: this code has a lot of handling for "post-increment" or
20// "post-inc" users. This is not talking about post-increment addressing modes;
21// it is instead talking about code like this:
22//
23// %i = phi [ 0, %entry ], [ %i.next, %latch ]
24// ...
25// %i.next = add %i, 1
26// %c = icmp eq %i.next, %n
27//
28// The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
29// it's useful to think about these as the same register, with some uses using
30// the value of the register before the add and some using it after. In this
31// example, the icmp is a post-increment user, since it uses %i.next, which is
32// the value of the induction variable after the increment. The other common
33// case of post-increment users is users outside the loop.
34//
35// TODO: More sophistication in the way Formulae are generated and filtered.
36//
37// TODO: Handle multiple loops at a time.
38//
39// TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
40// of a GlobalValue?
41//
42// TODO: When truncation is free, truncate ICmp users' operands to make it a
43// smaller encoding (on x86 at least).
44//
45// TODO: When a negated register is used by an add (such as in a list of
46// multiple base registers, or as the increment expression in an addrec),
47// we may not actually need both reg and (-1 * reg) in registers; the
48// negation can be implemented by using a sub instead of an add. The
49// lack of support for taking this into consideration when making
50// register pressure decisions is partly worked around by the "Special"
51// use kind.
52//
53//===----------------------------------------------------------------------===//
54
55#include "llvm/Transforms/Scalar/LoopStrengthReduce.h"
56#include "llvm/ADT/APInt.h"
57#include "llvm/ADT/DenseMap.h"
58#include "llvm/ADT/DenseSet.h"
59#include "llvm/ADT/Hashing.h"
60#include "llvm/ADT/PointerIntPair.h"
61#include "llvm/ADT/STLExtras.h"
62#include "llvm/ADT/SetVector.h"
63#include "llvm/ADT/SmallBitVector.h"
64#include "llvm/ADT/SmallPtrSet.h"
65#include "llvm/ADT/SmallSet.h"
66#include "llvm/ADT/SmallVector.h"
67#include "llvm/ADT/iterator_range.h"
68#include "llvm/Analysis/IVUsers.h"
69#include "llvm/Analysis/LoopAnalysisManager.h"
70#include "llvm/Analysis/LoopInfo.h"
71#include "llvm/Analysis/LoopPass.h"
72#include "llvm/Analysis/ScalarEvolution.h"
73#include "llvm/Analysis/ScalarEvolutionExpander.h"
74#include "llvm/Analysis/ScalarEvolutionExpressions.h"
75#include "llvm/Analysis/ScalarEvolutionNormalization.h"
76#include "llvm/Analysis/TargetTransformInfo.h"
77#include "llvm/Transforms/Utils/Local.h"
78#include "llvm/Config/llvm-config.h"
79#include "llvm/IR/BasicBlock.h"
80#include "llvm/IR/Constant.h"
81#include "llvm/IR/Constants.h"
82#include "llvm/IR/DerivedTypes.h"
83#include "llvm/IR/Dominators.h"
84#include "llvm/IR/GlobalValue.h"
85#include "llvm/IR/IRBuilder.h"
86#include "llvm/IR/InstrTypes.h"
87#include "llvm/IR/Instruction.h"
88#include "llvm/IR/Instructions.h"
89#include "llvm/IR/IntrinsicInst.h"
90#include "llvm/IR/Intrinsics.h"
91#include "llvm/IR/Module.h"
92#include "llvm/IR/OperandTraits.h"
93#include "llvm/IR/Operator.h"
94#include "llvm/IR/PassManager.h"
95#include "llvm/IR/Type.h"
96#include "llvm/IR/Use.h"
97#include "llvm/IR/User.h"
98#include "llvm/IR/Value.h"
99#include "llvm/IR/ValueHandle.h"
100#include "llvm/Pass.h"
101#include "llvm/Support/Casting.h"
102#include "llvm/Support/CommandLine.h"
103#include "llvm/Support/Compiler.h"
104#include "llvm/Support/Debug.h"
105#include "llvm/Support/ErrorHandling.h"
106#include "llvm/Support/MathExtras.h"
107#include "llvm/Support/raw_ostream.h"
108#include "llvm/Transforms/Scalar.h"
109#include "llvm/Transforms/Utils.h"
110#include "llvm/Transforms/Utils/BasicBlockUtils.h"
111#include <algorithm>
112#include <cassert>
113#include <cstddef>
114#include <cstdint>
115#include <cstdlib>
116#include <iterator>
117#include <limits>
118#include <numeric>
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<bool> EnableBackedgeIndexing(
159 "lsr-backedge-indexing", cl::Hidden, cl::init(true),
160 cl::desc("Enable the generation of cross iteration indexed memops"));
161
162static cl::opt<unsigned> ComplexityLimit(
163 "lsr-complexity-limit", cl::Hidden,
164 cl::init(std::numeric_limits<uint16_t>::max()),
165 cl::desc("LSR search space complexity limit"));
166
167static cl::opt<unsigned> SetupCostDepthLimit(
168 "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
169 cl::desc("The limit on recursion depth for LSRs setup cost"));
170
171#ifndef NDEBUG
172// Stress test IV chain generation.
173static cl::opt<bool> StressIVChain(
174 "stress-ivchain", cl::Hidden, cl::init(false),
175 cl::desc("Stress test LSR IV chains"));
176#else
177static bool StressIVChain = false;
178#endif
179
180namespace {
181
182struct MemAccessTy {
183 /// Used in situations where the accessed memory type is unknown.
184 static const unsigned UnknownAddressSpace =
185 std::numeric_limits<unsigned>::max();
186
187 Type *MemTy = nullptr;
188 unsigned AddrSpace = UnknownAddressSpace;
189
190 MemAccessTy() = default;
191 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
192
193 bool operator==(MemAccessTy Other) const {
194 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
195 }
196
197 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
198
199 static MemAccessTy getUnknown(LLVMContext &Ctx,
200 unsigned AS = UnknownAddressSpace) {
201 return MemAccessTy(Type::getVoidTy(Ctx), AS);
202 }
203
204 Type *getType() { return MemTy; }
205};
206
207/// This class holds data which is used to order reuse candidates.
208class RegSortData {
209public:
210 /// This represents the set of LSRUse indices which reference
211 /// a particular register.
212 SmallBitVector UsedByIndices;
213
214 void print(raw_ostream &OS) const;
215 void dump() const;
216};
217
218} // end anonymous namespace
219
220#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
221void RegSortData::print(raw_ostream &OS) const {
222 OS << "[NumUses=" << UsedByIndices.count() << ']';
223}
224
225LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
226 print(errs()); errs() << '\n';
227}
228#endif
229
230namespace {
231
232/// Map register candidates to information about how they are used.
233class RegUseTracker {
234 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
235
236 RegUsesTy RegUsesMap;
237 SmallVector<const SCEV *, 16> RegSequence;
238
239public:
240 void countRegister(const SCEV *Reg, size_t LUIdx);
241 void dropRegister(const SCEV *Reg, size_t LUIdx);
242 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
243
244 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
245
246 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
247
248 void clear();
249
250 using iterator = SmallVectorImpl<const SCEV *>::iterator;
251 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
252
253 iterator begin() { return RegSequence.begin(); }
30
Value assigned to 'DebugFlag', which participates in a condition later
254 iterator end() { return RegSequence.end(); }
255 const_iterator begin() const { return RegSequence.begin(); }
256 const_iterator end() const { return RegSequence.end(); }
257};
258
259} // end anonymous namespace
260
261void
262RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
263 std::pair<RegUsesTy::iterator, bool> Pair =
264 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
265 RegSortData &RSD = Pair.first->second;
266 if (Pair.second)
267 RegSequence.push_back(Reg);
268 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
269 RSD.UsedByIndices.set(LUIdx);
270}
271
272void
273RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
274 RegUsesTy::iterator It = RegUsesMap.find(Reg);
275 assert(It != RegUsesMap.end())((It != RegUsesMap.end()) ? static_cast<void> (0) : __assert_fail
("It != RegUsesMap.end()", "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 275, __PRETTY_FUNCTION__))
;
276 RegSortData &RSD = It->second;
277 assert(RSD.UsedByIndices.size() > LUIdx)((RSD.UsedByIndices.size() > LUIdx) ? static_cast<void>
(0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx", "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 277, __PRETTY_FUNCTION__))
;
278 RSD.UsedByIndices.reset(LUIdx);
279}
280
281void
282RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
283 assert(LUIdx <= LastLUIdx)((LUIdx <= LastLUIdx) ? static_cast<void> (0) : __assert_fail
("LUIdx <= LastLUIdx", "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 283, __PRETTY_FUNCTION__))
;
284
285 // Update RegUses. The data structure is not optimized for this purpose;
286 // we must iterate through it and update each of the bit vectors.
287 for (auto &Pair : RegUsesMap) {
288 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
289 if (LUIdx < UsedByIndices.size())
290 UsedByIndices[LUIdx] =
291 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
292 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
293 }
294}
295
296bool
297RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
298 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
299 if (I == RegUsesMap.end())
300 return false;
301 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
302 int i = UsedByIndices.find_first();
303 if (i == -1) return false;
304 if ((size_t)i != LUIdx) return true;
305 return UsedByIndices.find_next(i) != -1;
306}
307
308const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
309 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
310 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 310, __PRETTY_FUNCTION__))
;
311 return I->second.UsedByIndices;
312}
313
314void RegUseTracker::clear() {
315 RegUsesMap.clear();
316 RegSequence.clear();
317}
318
319namespace {
320
321/// This class holds information that describes a formula for computing
322/// satisfying a use. It may include broken-out immediates and scaled registers.
323struct Formula {
324 /// Global base address used for complex addressing.
325 GlobalValue *BaseGV = nullptr;
326
327 /// Base offset for complex addressing.
328 int64_t BaseOffset = 0;
329
330 /// Whether any complex addressing has a base register.
331 bool HasBaseReg = false;
332
333 /// The scale of any complex addressing.
334 int64_t Scale = 0;
335
336 /// The list of "base" registers for this use. When this is non-empty. The
337 /// canonical representation of a formula is
338 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
339 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
340 /// 3. The reg containing recurrent expr related with currect loop in the
341 /// formula should be put in the ScaledReg.
342 /// #1 enforces that the scaled register is always used when at least two
343 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
344 /// #2 enforces that 1 * reg is reg.
345 /// #3 ensures invariant regs with respect to current loop can be combined
346 /// together in LSR codegen.
347 /// This invariant can be temporarily broken while building a formula.
348 /// However, every formula inserted into the LSRInstance must be in canonical
349 /// form.
350 SmallVector<const SCEV *, 4> BaseRegs;
351
352 /// The 'scaled' register for this use. This should be non-null when Scale is
353 /// not zero.
354 const SCEV *ScaledReg = nullptr;
355
356 /// An additional constant offset which added near the use. This requires a
357 /// temporary register, but the offset itself can live in an add immediate
358 /// field rather than a register.
359 int64_t UnfoldedOffset = 0;
360
361 Formula() = default;
362
363 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
364
365 bool isCanonical(const Loop &L) const;
366
367 void canonicalize(const Loop &L);
368
369 bool unscale();
370
371 bool hasZeroEnd() const;
372
373 size_t getNumRegs() const;
374 Type *getType() const;
375
376 void deleteBaseReg(const SCEV *&S);
377
378 bool referencesReg(const SCEV *S) const;
379 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
380 const RegUseTracker &RegUses) const;
381
382 void print(raw_ostream &OS) const;
383 void dump() const;
384};
385
386} // end anonymous namespace
387
388/// Recursion helper for initialMatch.
389static void DoInitialMatch(const SCEV *S, Loop *L,
390 SmallVectorImpl<const SCEV *> &Good,
391 SmallVectorImpl<const SCEV *> &Bad,
392 ScalarEvolution &SE) {
393 // Collect expressions which properly dominate the loop header.
394 if (SE.properlyDominates(S, L->getHeader())) {
395 Good.push_back(S);
396 return;
397 }
398
399 // Look at add operands.
400 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
401 for (const SCEV *S : Add->operands())
402 DoInitialMatch(S, L, Good, Bad, SE);
403 return;
404 }
405
406 // Look at addrec operands.
407 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
408 if (!AR->getStart()->isZero() && AR->isAffine()) {
409 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
410 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
411 AR->getStepRecurrence(SE),
412 // FIXME: AR->getNoWrapFlags()
413 AR->getLoop(), SCEV::FlagAnyWrap),
414 L, Good, Bad, SE);
415 return;
416 }
417
418 // Handle a multiplication by -1 (negation) if it didn't fold.
419 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
420 if (Mul->getOperand(0)->isAllOnesValue()) {
421 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
422 const SCEV *NewMul = SE.getMulExpr(Ops);
423
424 SmallVector<const SCEV *, 4> MyGood;
425 SmallVector<const SCEV *, 4> MyBad;
426 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
427 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
428 SE.getEffectiveSCEVType(NewMul->getType())));
429 for (const SCEV *S : MyGood)
430 Good.push_back(SE.getMulExpr(NegOne, S));
431 for (const SCEV *S : MyBad)
432 Bad.push_back(SE.getMulExpr(NegOne, S));
433 return;
434 }
435
436 // Ok, we can't do anything interesting. Just stuff the whole thing into a
437 // register and hope for the best.
438 Bad.push_back(S);
439}
440
441/// Incorporate loop-variant parts of S into this Formula, attempting to keep
442/// all loop-invariant and loop-computable values in a single base register.
443void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
444 SmallVector<const SCEV *, 4> Good;
445 SmallVector<const SCEV *, 4> Bad;
446 DoInitialMatch(S, L, Good, Bad, SE);
447 if (!Good.empty()) {
448 const SCEV *Sum = SE.getAddExpr(Good);
449 if (!Sum->isZero())
450 BaseRegs.push_back(Sum);
451 HasBaseReg = true;
452 }
453 if (!Bad.empty()) {
454 const SCEV *Sum = SE.getAddExpr(Bad);
455 if (!Sum->isZero())
456 BaseRegs.push_back(Sum);
457 HasBaseReg = true;
458 }
459 canonicalize(*L);
460}
461
462/// Check whether or not this formula satisfies the canonical
463/// representation.
464/// \see Formula::BaseRegs.
465bool Formula::isCanonical(const Loop &L) const {
466 if (!ScaledReg)
467 return BaseRegs.size() <= 1;
468
469 if (Scale != 1)
470 return true;
471
472 if (Scale == 1 && BaseRegs.empty())
473 return false;
474
475 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
476 if (SAR && SAR->getLoop() == &L)
477 return true;
478
479 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
480 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
481 // loop, we want to swap the reg in BaseRegs with ScaledReg.
482 auto I =
483 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
484 return isa<const SCEVAddRecExpr>(S) &&
485 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
486 });
487 return I == BaseRegs.end();
488}
489
490/// Helper method to morph a formula into its canonical representation.
491/// \see Formula::BaseRegs.
492/// Every formula having more than one base register, must use the ScaledReg
493/// field. Otherwise, we would have to do special cases everywhere in LSR
494/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
495/// On the other hand, 1*reg should be canonicalized into reg.
496void Formula::canonicalize(const Loop &L) {
497 if (isCanonical(L))
498 return;
499 // So far we did not need this case. This is easy to implement but it is
500 // useless to maintain dead code. Beside it could hurt compile time.
501 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 501, __PRETTY_FUNCTION__))
;
502
503 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
504 if (!ScaledReg) {
505 ScaledReg = BaseRegs.back();
506 BaseRegs.pop_back();
507 Scale = 1;
508 }
509
510 // If ScaledReg is an invariant with respect to L, find the reg from
511 // BaseRegs containing the recurrent expr related with Loop L. Swap the
512 // reg with ScaledReg.
513 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
514 if (!SAR || SAR->getLoop() != &L) {
515 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
516 [&](const SCEV *S) {
517 return isa<const SCEVAddRecExpr>(S) &&
518 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
519 });
520 if (I != BaseRegs.end())
521 std::swap(ScaledReg, *I);
522 }
523}
524
525/// Get rid of the scale in the formula.
526/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
527/// \return true if it was possible to get rid of the scale, false otherwise.
528/// \note After this operation the formula may not be in the canonical form.
529bool Formula::unscale() {
530 if (Scale != 1)
531 return false;
532 Scale = 0;
533 BaseRegs.push_back(ScaledReg);
534 ScaledReg = nullptr;
535 return true;
536}
537
538bool Formula::hasZeroEnd() const {
539 if (UnfoldedOffset || BaseOffset)
540 return false;
541 if (BaseRegs.size() != 1 || ScaledReg)
542 return false;
543 return true;
544}
545
546/// Return the total number of register operands used by this formula. This does
547/// not include register uses implied by non-constant addrec strides.
548size_t Formula::getNumRegs() const {
549 return !!ScaledReg + BaseRegs.size();
550}
551
552/// Return the type of this formula, if it has one, or null otherwise. This type
553/// is meaningless except for the bit size.
554Type *Formula::getType() const {
555 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
556 ScaledReg ? ScaledReg->getType() :
557 BaseGV ? BaseGV->getType() :
558 nullptr;
559}
560
561/// Delete the given base reg from the BaseRegs list.
562void Formula::deleteBaseReg(const SCEV *&S) {
563 if (&S != &BaseRegs.back())
564 std::swap(S, BaseRegs.back());
565 BaseRegs.pop_back();
566}
567
568/// Test if this formula references the given register.
569bool Formula::referencesReg(const SCEV *S) const {
570 return S == ScaledReg || is_contained(BaseRegs, S);
571}
572
573/// Test whether this formula uses registers which are used by uses other than
574/// the use with the given index.
575bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
576 const RegUseTracker &RegUses) const {
577 if (ScaledReg)
578 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
579 return true;
580 for (const SCEV *BaseReg : BaseRegs)
581 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
582 return true;
583 return false;
584}
585
586#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
587void Formula::print(raw_ostream &OS) const {
588 bool First = true;
589 if (BaseGV) {
590 if (!First) OS << " + "; else First = false;
591 BaseGV->printAsOperand(OS, /*PrintType=*/false);
592 }
593 if (BaseOffset != 0) {
594 if (!First) OS << " + "; else First = false;
595 OS << BaseOffset;
596 }
597 for (const SCEV *BaseReg : BaseRegs) {
598 if (!First) OS << " + "; else First = false;
599 OS << "reg(" << *BaseReg << ')';
600 }
601 if (HasBaseReg && BaseRegs.empty()) {
602 if (!First) OS << " + "; else First = false;
603 OS << "**error: HasBaseReg**";
604 } else if (!HasBaseReg && !BaseRegs.empty()) {
605 if (!First) OS << " + "; else First = false;
606 OS << "**error: !HasBaseReg**";
607 }
608 if (Scale != 0) {
609 if (!First) OS << " + "; else First = false;
610 OS << Scale << "*reg(";
611 if (ScaledReg)
612 OS << *ScaledReg;
613 else
614 OS << "<unknown>";
615 OS << ')';
616 }
617 if (UnfoldedOffset != 0) {
618 if (!First) OS << " + ";
619 OS << "imm(" << UnfoldedOffset << ')';
620 }
621}
622
623LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
624 print(errs()); errs() << '\n';
625}
626#endif
627
628/// Return true if the given addrec can be sign-extended without changing its
629/// value.
630static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
631 Type *WideTy =
632 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
633 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
634}
635
636/// Return true if the given add can be sign-extended without changing its
637/// value.
638static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
639 Type *WideTy =
640 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
641 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
642}
643
644/// Return true if the given mul can be sign-extended without changing its
645/// value.
646static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
647 Type *WideTy =
648 IntegerType::get(SE.getContext(),
649 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
650 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
651}
652
653/// Return an expression for LHS /s RHS, if it can be determined and if the
654/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
655/// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
656/// the multiplication may overflow, which is useful when the result will be
657/// used in a context where the most significant bits are ignored.
658static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
659 ScalarEvolution &SE,
660 bool IgnoreSignificantBits = false) {
661 // Handle the trivial case, which works for any SCEV type.
662 if (LHS == RHS)
663 return SE.getConstant(LHS->getType(), 1);
664
665 // Handle a few RHS special cases.
666 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
667 if (RC) {
668 const APInt &RA = RC->getAPInt();
669 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
670 // some folding.
671 if (RA.isAllOnesValue())
672 return SE.getMulExpr(LHS, RC);
673 // Handle x /s 1 as x.
674 if (RA == 1)
675 return LHS;
676 }
677
678 // Check for a division of a constant by a constant.
679 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
680 if (!RC)
681 return nullptr;
682 const APInt &LA = C->getAPInt();
683 const APInt &RA = RC->getAPInt();
684 if (LA.srem(RA) != 0)
685 return nullptr;
686 return SE.getConstant(LA.sdiv(RA));
687 }
688
689 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
690 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
691 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
692 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
693 IgnoreSignificantBits);
694 if (!Step) return nullptr;
695 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
696 IgnoreSignificantBits);
697 if (!Start) return nullptr;
698 // FlagNW is independent of the start value, step direction, and is
699 // preserved with smaller magnitude steps.
700 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
701 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
702 }
703 return nullptr;
704 }
705
706 // Distribute the sdiv over add operands, if the add doesn't overflow.
707 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
708 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
709 SmallVector<const SCEV *, 8> Ops;
710 for (const SCEV *S : Add->operands()) {
711 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
712 if (!Op) return nullptr;
713 Ops.push_back(Op);
714 }
715 return SE.getAddExpr(Ops);
716 }
717 return nullptr;
718 }
719
720 // Check for a multiply operand that we can pull RHS out of.
721 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
722 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
723 SmallVector<const SCEV *, 4> Ops;
724 bool Found = false;
725 for (const SCEV *S : Mul->operands()) {
726 if (!Found)
727 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
728 IgnoreSignificantBits)) {
729 S = Q;
730 Found = true;
731 }
732 Ops.push_back(S);
733 }
734 return Found ? SE.getMulExpr(Ops) : nullptr;
735 }
736 return nullptr;
737 }
738
739 // Otherwise we don't know.
740 return nullptr;
741}
742
743/// If S involves the addition of a constant integer value, return that integer
744/// value, and mutate S to point to a new SCEV with that value excluded.
745static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
746 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
747 if (C->getAPInt().getMinSignedBits() <= 64) {
748 S = SE.getConstant(C->getType(), 0);
749 return C->getValue()->getSExtValue();
750 }
751 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
752 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
753 int64_t Result = ExtractImmediate(NewOps.front(), SE);
754 if (Result != 0)
755 S = SE.getAddExpr(NewOps);
756 return Result;
757 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
758 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
759 int64_t Result = ExtractImmediate(NewOps.front(), SE);
760 if (Result != 0)
761 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
762 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
763 SCEV::FlagAnyWrap);
764 return Result;
765 }
766 return 0;
767}
768
769/// If S involves the addition of a GlobalValue address, return that symbol, and
770/// mutate S to point to a new SCEV with that value excluded.
771static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
772 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
773 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
774 S = SE.getConstant(GV->getType(), 0);
775 return GV;
776 }
777 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
778 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
779 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
780 if (Result)
781 S = SE.getAddExpr(NewOps);
782 return Result;
783 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
784 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
785 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
786 if (Result)
787 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
788 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
789 SCEV::FlagAnyWrap);
790 return Result;
791 }
792 return nullptr;
793}
794
795/// Returns true if the specified instruction is using the specified value as an
796/// address.
797static bool isAddressUse(const TargetTransformInfo &TTI,
798 Instruction *Inst, Value *OperandVal) {
799 bool isAddress = isa<LoadInst>(Inst);
800 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
801 if (SI->getPointerOperand() == OperandVal)
802 isAddress = true;
803 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
804 // Addressing modes can also be folded into prefetches and a variety
805 // of intrinsics.
806 switch (II->getIntrinsicID()) {
807 case Intrinsic::memset:
808 case Intrinsic::prefetch:
809 if (II->getArgOperand(0) == OperandVal)
810 isAddress = true;
811 break;
812 case Intrinsic::memmove:
813 case Intrinsic::memcpy:
814 if (II->getArgOperand(0) == OperandVal ||
815 II->getArgOperand(1) == OperandVal)
816 isAddress = true;
817 break;
818 default: {
819 MemIntrinsicInfo IntrInfo;
820 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
821 if (IntrInfo.PtrVal == OperandVal)
822 isAddress = true;
823 }
824 }
825 }
826 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
827 if (RMW->getPointerOperand() == OperandVal)
828 isAddress = true;
829 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
830 if (CmpX->getPointerOperand() == OperandVal)
831 isAddress = true;
832 }
833 return isAddress;
834}
835
836/// Return the type of the memory being accessed.
837static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
838 Instruction *Inst, Value *OperandVal) {
839 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
840 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
841 AccessTy.MemTy = SI->getOperand(0)->getType();
842 AccessTy.AddrSpace = SI->getPointerAddressSpace();
843 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
844 AccessTy.AddrSpace = LI->getPointerAddressSpace();
845 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
846 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
847 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
848 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
849 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
850 switch (II->getIntrinsicID()) {
851 case Intrinsic::prefetch:
852 case Intrinsic::memset:
853 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
854 AccessTy.MemTy = OperandVal->getType();
855 break;
856 case Intrinsic::memmove:
857 case Intrinsic::memcpy:
858 AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
859 AccessTy.MemTy = OperandVal->getType();
860 break;
861 default: {
862 MemIntrinsicInfo IntrInfo;
863 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
864 AccessTy.AddrSpace
865 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
866 }
867
868 break;
869 }
870 }
871 }
872
873 // All pointers have the same requirements, so canonicalize them to an
874 // arbitrary pointer type to minimize variation.
875 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
876 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
877 PTy->getAddressSpace());
878
879 return AccessTy;
880}
881
882/// Return true if this AddRec is already a phi in its loop.
883static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
884 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
885 if (SE.isSCEVable(PN.getType()) &&
886 (SE.getEffectiveSCEVType(PN.getType()) ==
887 SE.getEffectiveSCEVType(AR->getType())) &&
888 SE.getSCEV(&PN) == AR)
889 return true;
890 }
891 return false;
892}
893
894/// Check if expanding this expression is likely to incur significant cost. This
895/// is tricky because SCEV doesn't track which expressions are actually computed
896/// by the current IR.
897///
898/// We currently allow expansion of IV increments that involve adds,
899/// multiplication by constants, and AddRecs from existing phis.
900///
901/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
902/// obvious multiple of the UDivExpr.
903static bool isHighCostExpansion(const SCEV *S,
904 SmallPtrSetImpl<const SCEV*> &Processed,
905 ScalarEvolution &SE) {
906 // Zero/One operand expressions
907 switch (S->getSCEVType()) {
908 case scUnknown:
909 case scConstant:
910 return false;
911 case scTruncate:
912 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
913 Processed, SE);
914 case scZeroExtend:
915 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
916 Processed, SE);
917 case scSignExtend:
918 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
919 Processed, SE);
920 }
921
922 if (!Processed.insert(S).second)
923 return false;
924
925 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
926 for (const SCEV *S : Add->operands()) {
927 if (isHighCostExpansion(S, Processed, SE))
928 return true;
929 }
930 return false;
931 }
932
933 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
934 if (Mul->getNumOperands() == 2) {
935 // Multiplication by a constant is ok
936 if (isa<SCEVConstant>(Mul->getOperand(0)))
937 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
938
939 // If we have the value of one operand, check if an existing
940 // multiplication already generates this expression.
941 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
942 Value *UVal = U->getValue();
943 for (User *UR : UVal->users()) {
944 // If U is a constant, it may be used by a ConstantExpr.
945 Instruction *UI = dyn_cast<Instruction>(UR);
946 if (UI && UI->getOpcode() == Instruction::Mul &&
947 SE.isSCEVable(UI->getType())) {
948 return SE.getSCEV(UI) == Mul;
949 }
950 }
951 }
952 }
953 }
954
955 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
956 if (isExistingPhi(AR, SE))
957 return false;
958 }
959
960 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
961 return true;
962}
963
964/// If any of the instructions in the specified set are trivially dead, delete
965/// them and see if this makes any of their operands subsequently dead.
966static bool
967DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
968 bool Changed = false;
969
970 while (!DeadInsts.empty()) {
971 Value *V = DeadInsts.pop_back_val();
972 Instruction *I = dyn_cast_or_null<Instruction>(V);
973
974 if (!I || !isInstructionTriviallyDead(I))
975 continue;
976
977 for (Use &O : I->operands())
978 if (Instruction *U = dyn_cast<Instruction>(O)) {
979 O = nullptr;
980 if (U->use_empty())
981 DeadInsts.emplace_back(U);
982 }
983
984 I->eraseFromParent();
985 Changed = true;
986 }
987
988 return Changed;
989}
990
991namespace {
992
993class LSRUse;
994
995} // end anonymous namespace
996
997/// Check if the addressing mode defined by \p F is completely
998/// folded in \p LU at isel time.
999/// This includes address-mode folding and special icmp tricks.
1000/// This function returns true if \p LU can accommodate what \p F
1001/// defines and up to 1 base + 1 scaled + offset.
1002/// In other words, if \p F has several base registers, this function may
1003/// still return true. Therefore, users still need to account for
1004/// additional base registers and/or unfolded offsets to derive an
1005/// accurate cost model.
1006static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1007 const LSRUse &LU, const Formula &F);
1008
1009// Get the cost of the scaling factor used in F for LU.
1010static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1011 const LSRUse &LU, const Formula &F,
1012 const Loop &L);
1013
1014namespace {
1015
1016/// This class is used to measure and compare candidate formulae.
1017class Cost {
1018 const Loop *L = nullptr;
1019 ScalarEvolution *SE = nullptr;
1020 const TargetTransformInfo *TTI = nullptr;
1021 TargetTransformInfo::LSRCost C;
1022
1023public:
1024 Cost() = delete;
1025 Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI) :
1026 L(L), SE(&SE), TTI(&TTI) {
1027 C.Insns = 0;
1028 C.NumRegs = 0;
1029 C.AddRecCost = 0;
1030 C.NumIVMuls = 0;
1031 C.NumBaseAdds = 0;
1032 C.ImmCost = 0;
1033 C.SetupCost = 0;
1034 C.ScaleCost = 0;
1035 }
1036
1037 bool isLess(Cost &Other);
1038
1039 void Lose();
1040
1041#ifndef NDEBUG
1042 // Once any of the metrics loses, they must all remain losers.
1043 bool isValid() {
1044 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1045 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1046 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1047 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1048 }
1049#endif
1050
1051 bool isLoser() {
1052 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1052, __PRETTY_FUNCTION__))
;
1053 return C.NumRegs == ~0u;
1054 }
1055
1056 void RateFormula(const Formula &F,
1057 SmallPtrSetImpl<const SCEV *> &Regs,
1058 const DenseSet<const SCEV *> &VisitedRegs,
1059 const LSRUse &LU,
1060 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1061
1062 void print(raw_ostream &OS) const;
1063 void dump() const;
1064
1065private:
1066 void RateRegister(const Formula &F, const SCEV *Reg,
1067 SmallPtrSetImpl<const SCEV *> &Regs);
1068 void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1069 SmallPtrSetImpl<const SCEV *> &Regs,
1070 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1071};
1072
1073/// An operand value in an instruction which is to be replaced with some
1074/// equivalent, possibly strength-reduced, replacement.
1075struct LSRFixup {
1076 /// The instruction which will be updated.
1077 Instruction *UserInst = nullptr;
1078
1079 /// The operand of the instruction which will be replaced. The operand may be
1080 /// used more than once; every instance will be replaced.
1081 Value *OperandValToReplace = nullptr;
1082
1083 /// If this user is to use the post-incremented value of an induction
1084 /// variable, this set is non-empty and holds the loops associated with the
1085 /// induction variable.
1086 PostIncLoopSet PostIncLoops;
1087
1088 /// A constant offset to be added to the LSRUse expression. This allows
1089 /// multiple fixups to share the same LSRUse with different offsets, for
1090 /// example in an unrolled loop.
1091 int64_t Offset = 0;
1092
1093 LSRFixup() = default;
1094
1095 bool isUseFullyOutsideLoop(const Loop *L) const;
1096
1097 void print(raw_ostream &OS) const;
1098 void dump() const;
1099};
1100
1101/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1102/// SmallVectors of const SCEV*.
1103struct UniquifierDenseMapInfo {
1104 static SmallVector<const SCEV *, 4> getEmptyKey() {
1105 SmallVector<const SCEV *, 4> V;
1106 V.push_back(reinterpret_cast<const SCEV *>(-1));
1107 return V;
1108 }
1109
1110 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1111 SmallVector<const SCEV *, 4> V;
1112 V.push_back(reinterpret_cast<const SCEV *>(-2));
1113 return V;
1114 }
1115
1116 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1117 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1118 }
1119
1120 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1121 const SmallVector<const SCEV *, 4> &RHS) {
1122 return LHS == RHS;
1123 }
1124};
1125
1126/// This class holds the state that LSR keeps for each use in IVUsers, as well
1127/// as uses invented by LSR itself. It includes information about what kinds of
1128/// things can be folded into the user, information about the user itself, and
1129/// information about how the use may be satisfied. TODO: Represent multiple
1130/// users of the same expression in common?
1131class LSRUse {
1132 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1133
1134public:
1135 /// An enum for a kind of use, indicating what types of scaled and immediate
1136 /// operands it might support.
1137 enum KindType {
1138 Basic, ///< A normal use, with no folding.
1139 Special, ///< A special case of basic, allowing -1 scales.
1140 Address, ///< An address use; folding according to TargetLowering
1141 ICmpZero ///< An equality icmp with both operands folded into one.
1142 // TODO: Add a generic icmp too?
1143 };
1144
1145 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1146
1147 KindType Kind;
1148 MemAccessTy AccessTy;
1149
1150 /// The list of operands which are to be replaced.
1151 SmallVector<LSRFixup, 8> Fixups;
1152
1153 /// Keep track of the min and max offsets of the fixups.
1154 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1155 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1156
1157 /// This records whether all of the fixups using this LSRUse are outside of
1158 /// the loop, in which case some special-case heuristics may be used.
1159 bool AllFixupsOutsideLoop = true;
1160
1161 /// RigidFormula is set to true to guarantee that this use will be associated
1162 /// with a single formula--the one that initially matched. Some SCEV
1163 /// expressions cannot be expanded. This allows LSR to consider the registers
1164 /// used by those expressions without the need to expand them later after
1165 /// changing the formula.
1166 bool RigidFormula = false;
1167
1168 /// This records the widest use type for any fixup using this
1169 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1170 /// fixup widths to be equivalent, because the narrower one may be relying on
1171 /// the implicit truncation to truncate away bogus bits.
1172 Type *WidestFixupType = nullptr;
1173
1174 /// A list of ways to build a value that can satisfy this user. After the
1175 /// list is populated, one of these is selected heuristically and used to
1176 /// formulate a replacement for OperandValToReplace in UserInst.
1177 SmallVector<Formula, 12> Formulae;
1178
1179 /// The set of register candidates used by all formulae in this LSRUse.
1180 SmallPtrSet<const SCEV *, 4> Regs;
1181
1182 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1183
1184 LSRFixup &getNewFixup() {
1185 Fixups.push_back(LSRFixup());
1186 return Fixups.back();
1187 }
1188
1189 void pushFixup(LSRFixup &f) {
1190 Fixups.push_back(f);
1191 if (f.Offset > MaxOffset)
1192 MaxOffset = f.Offset;
1193 if (f.Offset < MinOffset)
1194 MinOffset = f.Offset;
1195 }
1196
1197 bool HasFormulaWithSameRegs(const Formula &F) const;
1198 float getNotSelectedProbability(const SCEV *Reg) const;
1199 bool InsertFormula(const Formula &F, const Loop &L);
1200 void DeleteFormula(Formula &F);
1201 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1202
1203 void print(raw_ostream &OS) const;
1204 void dump() const;
1205};
1206
1207} // end anonymous namespace
1208
1209static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1210 LSRUse::KindType Kind, MemAccessTy AccessTy,
1211 GlobalValue *BaseGV, int64_t BaseOffset,
1212 bool HasBaseReg, int64_t Scale,
1213 Instruction *Fixup = nullptr);
1214
1215static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1216 if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1217 return 1;
1218 if (Depth == 0)
1219 return 0;
1220 if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1221 return getSetupCost(S->getStart(), Depth - 1);
1222 if (auto S = dyn_cast<SCEVCastExpr>(Reg))
1223 return getSetupCost(S->getOperand(), Depth - 1);
1224 if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1225 return std::accumulate(S->op_begin(), S->op_end(), 0,
1226 [&](unsigned i, const SCEV *Reg) {
1227 return i + getSetupCost(Reg, Depth - 1);
1228 });
1229 if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1230 return getSetupCost(S->getLHS(), Depth - 1) +
1231 getSetupCost(S->getRHS(), Depth - 1);
1232 return 0;
1233}
1234
1235/// Tally up interesting quantities from the given register.
1236void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1237 SmallPtrSetImpl<const SCEV *> &Regs) {
1238 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1239 // If this is an addrec for another loop, it should be an invariant
1240 // with respect to L since L is the innermost loop (at least
1241 // for now LSR only handles innermost loops).
1242 if (AR->getLoop() != L) {
1243 // If the AddRec exists, consider it's register free and leave it alone.
1244 if (isExistingPhi(AR, *SE))
1245 return;
1246
1247 // It is bad to allow LSR for current loop to add induction variables
1248 // for its sibling loops.
1249 if (!AR->getLoop()->contains(L)) {
1250 Lose();
1251 return;
1252 }
1253
1254 // Otherwise, it will be an invariant with respect to Loop L.
1255 ++C.NumRegs;
1256 return;
1257 }
1258
1259 unsigned LoopCost = 1;
1260 if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1261 TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1262
1263 // If the step size matches the base offset, we could use pre-indexed
1264 // addressing.
1265 if (TTI->shouldFavorBackedgeIndex(L)) {
1266 if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1267 if (Step->getAPInt() == F.BaseOffset)
1268 LoopCost = 0;
1269 }
1270
1271 if (TTI->shouldFavorPostInc()) {
1272 const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1273 if (isa<SCEVConstant>(LoopStep)) {
1274 const SCEV *LoopStart = AR->getStart();
1275 if (!isa<SCEVConstant>(LoopStart) &&
1276 SE->isLoopInvariant(LoopStart, L))
1277 LoopCost = 0;
1278 }
1279 }
1280 }
1281 C.AddRecCost += LoopCost;
1282
1283 // Add the step value register, if it needs one.
1284 // TODO: The non-affine case isn't precisely modeled here.
1285 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1286 if (!Regs.count(AR->getOperand(1))) {
1287 RateRegister(F, AR->getOperand(1), Regs);
1288 if (isLoser())
1289 return;
1290 }
1291 }
1292 }
1293 ++C.NumRegs;
1294
1295 // Rough heuristic; favor registers which don't require extra setup
1296 // instructions in the preheader.
1297 C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1298 // Ensure we don't, even with the recusion limit, produce invalid costs.
1299 C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1300
1301 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1302 SE->hasComputableLoopEvolution(Reg, L);
1303}
1304
1305/// Record this register in the set. If we haven't seen it before, rate
1306/// it. Optional LoserRegs provides a way to declare any formula that refers to
1307/// one of those regs an instant loser.
1308void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1309 SmallPtrSetImpl<const SCEV *> &Regs,
1310 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1311 if (LoserRegs && LoserRegs->count(Reg)) {
1312 Lose();
1313 return;
1314 }
1315 if (Regs.insert(Reg).second) {
1316 RateRegister(F, Reg, Regs);
1317 if (LoserRegs && isLoser())
1318 LoserRegs->insert(Reg);
1319 }
1320}
1321
1322void Cost::RateFormula(const Formula &F,
1323 SmallPtrSetImpl<const SCEV *> &Regs,
1324 const DenseSet<const SCEV *> &VisitedRegs,
1325 const LSRUse &LU,
1326 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1327 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1327, __PRETTY_FUNCTION__))
;
1328 // Tally up the registers.
1329 unsigned PrevAddRecCost = C.AddRecCost;
1330 unsigned PrevNumRegs = C.NumRegs;
1331 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1332 if (const SCEV *ScaledReg = F.ScaledReg) {
1333 if (VisitedRegs.count(ScaledReg)) {
1334 Lose();
1335 return;
1336 }
1337 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1338 if (isLoser())
1339 return;
1340 }
1341 for (const SCEV *BaseReg : F.BaseRegs) {
1342 if (VisitedRegs.count(BaseReg)) {
1343 Lose();
1344 return;
1345 }
1346 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1347 if (isLoser())
1348 return;
1349 }
1350
1351 // Determine how many (unfolded) adds we'll need inside the loop.
1352 size_t NumBaseParts = F.getNumRegs();
1353 if (NumBaseParts > 1)
1354 // Do not count the base and a possible second register if the target
1355 // allows to fold 2 registers.
1356 C.NumBaseAdds +=
1357 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1358 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1359
1360 // Accumulate non-free scaling amounts.
1361 C.ScaleCost += getScalingFactorCost(*TTI, LU, F, *L);
1362
1363 // Tally up the non-zero immediates.
1364 for (const LSRFixup &Fixup : LU.Fixups) {
1365 int64_t O = Fixup.Offset;
1366 int64_t Offset = (uint64_t)O + F.BaseOffset;
1367 if (F.BaseGV)
1368 C.ImmCost += 64; // Handle symbolic values conservatively.
1369 // TODO: This should probably be the pointer size.
1370 else if (Offset != 0)
1371 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1372
1373 // Check with target if this offset with this instruction is
1374 // specifically not supported.
1375 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1376 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1377 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1378 C.NumBaseAdds++;
1379 }
1380
1381 // If we don't count instruction cost exit here.
1382 if (!InsnsCost) {
1383 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1383, __PRETTY_FUNCTION__))
;
1384 return;
1385 }
1386
1387 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1388 // additional instruction (at least fill).
1389 unsigned TTIRegNum = TTI->getNumberOfRegisters(false) - 1;
1390 if (C.NumRegs > TTIRegNum) {
1391 // Cost already exceeded TTIRegNum, then only newly added register can add
1392 // new instructions.
1393 if (PrevNumRegs > TTIRegNum)
1394 C.Insns += (C.NumRegs - PrevNumRegs);
1395 else
1396 C.Insns += (C.NumRegs - TTIRegNum);
1397 }
1398
1399 // If ICmpZero formula ends with not 0, it could not be replaced by
1400 // just add or sub. We'll need to compare final result of AddRec.
1401 // That means we'll need an additional instruction. But if the target can
1402 // macro-fuse a compare with a branch, don't count this extra instruction.
1403 // For -10 + {0, +, 1}:
1404 // i = i + 1;
1405 // cmp i, 10
1406 //
1407 // For {-10, +, 1}:
1408 // i = i + 1;
1409 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1410 !TTI->canMacroFuseCmp())
1411 C.Insns++;
1412 // Each new AddRec adds 1 instruction to calculation.
1413 C.Insns += (C.AddRecCost - PrevAddRecCost);
1414
1415 // BaseAdds adds instructions for unfolded registers.
1416 if (LU.Kind != LSRUse::ICmpZero)
1417 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1418 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1418, __PRETTY_FUNCTION__))
;
1419}
1420
1421/// Set this cost to a losing value.
1422void Cost::Lose() {
1423 C.Insns = std::numeric_limits<unsigned>::max();
1424 C.NumRegs = std::numeric_limits<unsigned>::max();
1425 C.AddRecCost = std::numeric_limits<unsigned>::max();
1426 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1427 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1428 C.ImmCost = std::numeric_limits<unsigned>::max();
1429 C.SetupCost = std::numeric_limits<unsigned>::max();
1430 C.ScaleCost = std::numeric_limits<unsigned>::max();
1431}
1432
1433/// Choose the lower cost.
1434bool Cost::isLess(Cost &Other) {
1435 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1436 C.Insns != Other.C.Insns)
1437 return C.Insns < Other.C.Insns;
1438 return TTI->isLSRCostLess(C, Other.C);
1439}
1440
1441#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1442void Cost::print(raw_ostream &OS) const {
1443 if (InsnsCost)
1444 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1445 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1446 if (C.AddRecCost != 0)
1447 OS << ", with addrec cost " << C.AddRecCost;
1448 if (C.NumIVMuls != 0)
1449 OS << ", plus " << C.NumIVMuls << " IV mul"
1450 << (C.NumIVMuls == 1 ? "" : "s");
1451 if (C.NumBaseAdds != 0)
1452 OS << ", plus " << C.NumBaseAdds << " base add"
1453 << (C.NumBaseAdds == 1 ? "" : "s");
1454 if (C.ScaleCost != 0)
1455 OS << ", plus " << C.ScaleCost << " scale cost";
1456 if (C.ImmCost != 0)
1457 OS << ", plus " << C.ImmCost << " imm cost";
1458 if (C.SetupCost != 0)
1459 OS << ", plus " << C.SetupCost << " setup cost";
1460}
1461
1462LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1463 print(errs()); errs() << '\n';
1464}
1465#endif
1466
1467/// Test whether this fixup always uses its value outside of the given loop.
1468bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1469 // PHI nodes use their value in their incoming blocks.
1470 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1471 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1472 if (PN->getIncomingValue(i) == OperandValToReplace &&
1473 L->contains(PN->getIncomingBlock(i)))
1474 return false;
1475 return true;
1476 }
1477
1478 return !L->contains(UserInst);
1479}
1480
1481#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1482void LSRFixup::print(raw_ostream &OS) const {
1483 OS << "UserInst=";
1484 // Store is common and interesting enough to be worth special-casing.
1485 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1486 OS << "store ";
1487 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1488 } else if (UserInst->getType()->isVoidTy())
1489 OS << UserInst->getOpcodeName();
1490 else
1491 UserInst->printAsOperand(OS, /*PrintType=*/false);
1492
1493 OS << ", OperandValToReplace=";
1494 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1495
1496 for (const Loop *PIL : PostIncLoops) {
1497 OS << ", PostIncLoop=";
1498 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1499 }
1500
1501 if (Offset != 0)
1502 OS << ", Offset=" << Offset;
1503}
1504
1505LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1506 print(errs()); errs() << '\n';
1507}
1508#endif
1509
1510/// Test whether this use as a formula which has the same registers as the given
1511/// formula.
1512bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1513 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1514 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1515 // Unstable sort by host order ok, because this is only used for uniquifying.
1516 llvm::sort(Key);
1517 return Uniquifier.count(Key);
1518}
1519
1520/// The function returns a probability of selecting formula without Reg.
1521float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1522 unsigned FNum = 0;
1523 for (const Formula &F : Formulae)
1524 if (F.referencesReg(Reg))
1525 FNum++;
1526 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1527}
1528
1529/// If the given formula has not yet been inserted, add it to the list, and
1530/// return true. Return false otherwise. The formula must be in canonical form.
1531bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1532 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1532, __PRETTY_FUNCTION__))
;
1533
1534 if (!Formulae.empty() && RigidFormula)
1535 return false;
1536
1537 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1538 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1539 // Unstable sort by host order ok, because this is only used for uniquifying.
1540 llvm::sort(Key);
1541
1542 if (!Uniquifier.insert(Key).second)
1543 return false;
1544
1545 // Using a register to hold the value of 0 is not profitable.
1546 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1547, __PRETTY_FUNCTION__))
1547 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1547, __PRETTY_FUNCTION__))
;
1548#ifndef NDEBUG
1549 for (const SCEV *BaseReg : F.BaseRegs)
1550 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1550, __PRETTY_FUNCTION__))
;
1551#endif
1552
1553 // Add the formula to the list.
1554 Formulae.push_back(F);
1555
1556 // Record registers now being used by this use.
1557 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1558 if (F.ScaledReg)
1559 Regs.insert(F.ScaledReg);
1560
1561 return true;
1562}
1563
1564/// Remove the given formula from this use's list.
1565void LSRUse::DeleteFormula(Formula &F) {
1566 if (&F != &Formulae.back())
1567 std::swap(F, Formulae.back());
1568 Formulae.pop_back();
1569}
1570
1571/// Recompute the Regs field, and update RegUses.
1572void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1573 // Now that we've filtered out some formulae, recompute the Regs set.
1574 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1575 Regs.clear();
1576 for (const Formula &F : Formulae) {
1577 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1578 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1579 }
1580
1581 // Update the RegTracker.
1582 for (const SCEV *S : OldRegs)
1583 if (!Regs.count(S))
1584 RegUses.dropRegister(S, LUIdx);
1585}
1586
1587#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1588void LSRUse::print(raw_ostream &OS) const {
1589 OS << "LSR Use: Kind=";
1590 switch (Kind) {
1591 case Basic: OS << "Basic"; break;
1592 case Special: OS << "Special"; break;
1593 case ICmpZero: OS << "ICmpZero"; break;
1594 case Address:
1595 OS << "Address of ";
1596 if (AccessTy.MemTy->isPointerTy())
1597 OS << "pointer"; // the full pointer type could be really verbose
1598 else {
1599 OS << *AccessTy.MemTy;
1600 }
1601
1602 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1603 }
1604
1605 OS << ", Offsets={";
1606 bool NeedComma = false;
1607 for (const LSRFixup &Fixup : Fixups) {
1608 if (NeedComma) OS << ',';
1609 OS << Fixup.Offset;
1610 NeedComma = true;
1611 }
1612 OS << '}';
1613
1614 if (AllFixupsOutsideLoop)
1615 OS << ", all-fixups-outside-loop";
1616
1617 if (WidestFixupType)
1618 OS << ", widest fixup type: " << *WidestFixupType;
1619}
1620
1621LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1622 print(errs()); errs() << '\n';
1623}
1624#endif
1625
1626static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1627 LSRUse::KindType Kind, MemAccessTy AccessTy,
1628 GlobalValue *BaseGV, int64_t BaseOffset,
1629 bool HasBaseReg, int64_t Scale,
1630 Instruction *Fixup/*= nullptr*/) {
1631 switch (Kind) {
1632 case LSRUse::Address:
1633 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1634 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1635
1636 case LSRUse::ICmpZero:
1637 // There's not even a target hook for querying whether it would be legal to
1638 // fold a GV into an ICmp.
1639 if (BaseGV)
1640 return false;
1641
1642 // ICmp only has two operands; don't allow more than two non-trivial parts.
1643 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1644 return false;
1645
1646 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1647 // putting the scaled register in the other operand of the icmp.
1648 if (Scale != 0 && Scale != -1)
1649 return false;
1650
1651 // If we have low-level target information, ask the target if it can fold an
1652 // integer immediate on an icmp.
1653 if (BaseOffset != 0) {
1654 // We have one of:
1655 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1656 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1657 // Offs is the ICmp immediate.
1658 if (Scale == 0)
1659 // The cast does the right thing with
1660 // std::numeric_limits<int64_t>::min().
1661 BaseOffset = -(uint64_t)BaseOffset;
1662 return TTI.isLegalICmpImmediate(BaseOffset);
1663 }
1664
1665 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1666 return true;
1667
1668 case LSRUse::Basic:
1669 // Only handle single-register values.
1670 return !BaseGV && Scale == 0 && BaseOffset == 0;
1671
1672 case LSRUse::Special:
1673 // Special case Basic to handle -1 scales.
1674 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1675 }
1676
1677 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1677)
;
1678}
1679
1680static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1681 int64_t MinOffset, int64_t MaxOffset,
1682 LSRUse::KindType Kind, MemAccessTy AccessTy,
1683 GlobalValue *BaseGV, int64_t BaseOffset,
1684 bool HasBaseReg, int64_t Scale) {
1685 // Check for overflow.
1686 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1687 (MinOffset > 0))
1688 return false;
1689 MinOffset = (uint64_t)BaseOffset + MinOffset;
1690 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1691 (MaxOffset > 0))
1692 return false;
1693 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1694
1695 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1696 HasBaseReg, Scale) &&
1697 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1698 HasBaseReg, Scale);
1699}
1700
1701static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1702 int64_t MinOffset, int64_t MaxOffset,
1703 LSRUse::KindType Kind, MemAccessTy AccessTy,
1704 const Formula &F, const Loop &L) {
1705 // For the purpose of isAMCompletelyFolded either having a canonical formula
1706 // or a scale not equal to zero is correct.
1707 // Problems may arise from non canonical formulae having a scale == 0.
1708 // Strictly speaking it would best to just rely on canonical formulae.
1709 // However, when we generate the scaled formulae, we first check that the
1710 // scaling factor is profitable before computing the actual ScaledReg for
1711 // compile time sake.
1712 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1712, __PRETTY_FUNCTION__))
;
1713 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1714 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1715}
1716
1717/// Test whether we know how to expand the current formula.
1718static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1719 int64_t MaxOffset, LSRUse::KindType Kind,
1720 MemAccessTy AccessTy, GlobalValue *BaseGV,
1721 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1722 // We know how to expand completely foldable formulae.
1723 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1724 BaseOffset, HasBaseReg, Scale) ||
1725 // Or formulae that use a base register produced by a sum of base
1726 // registers.
1727 (Scale == 1 &&
1728 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1729 BaseGV, BaseOffset, true, 0));
1730}
1731
1732static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1733 int64_t MaxOffset, LSRUse::KindType Kind,
1734 MemAccessTy AccessTy, const Formula &F) {
1735 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1736 F.BaseOffset, F.HasBaseReg, F.Scale);
1737}
1738
1739static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1740 const LSRUse &LU, const Formula &F) {
1741 // Target may want to look at the user instructions.
1742 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1743 for (const LSRFixup &Fixup : LU.Fixups)
1744 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1745 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1746 F.Scale, Fixup.UserInst))
1747 return false;
1748 return true;
1749 }
1750
1751 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1752 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1753 F.Scale);
1754}
1755
1756static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1757 const LSRUse &LU, const Formula &F,
1758 const Loop &L) {
1759 if (!F.Scale)
1760 return 0;
1761
1762 // If the use is not completely folded in that instruction, we will have to
1763 // pay an extra cost only for scale != 1.
1764 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1765 LU.AccessTy, F, L))
1766 return F.Scale != 1;
1767
1768 switch (LU.Kind) {
1769 case LSRUse::Address: {
1770 // Check the scaling factor cost with both the min and max offsets.
1771 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1772 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1773 F.Scale, LU.AccessTy.AddrSpace);
1774 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1775 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1776 F.Scale, LU.AccessTy.AddrSpace);
1777
1778 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1779, __PRETTY_FUNCTION__))
1779 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1779, __PRETTY_FUNCTION__))
;
1780 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1781 }
1782 case LSRUse::ICmpZero:
1783 case LSRUse::Basic:
1784 case LSRUse::Special:
1785 // The use is completely folded, i.e., everything is folded into the
1786 // instruction.
1787 return 0;
1788 }
1789
1790 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1790)
;
1791}
1792
1793static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1794 LSRUse::KindType Kind, MemAccessTy AccessTy,
1795 GlobalValue *BaseGV, int64_t BaseOffset,
1796 bool HasBaseReg) {
1797 // Fast-path: zero is always foldable.
1798 if (BaseOffset == 0 && !BaseGV) return true;
1799
1800 // Conservatively, create an address with an immediate and a
1801 // base and a scale.
1802 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1803
1804 // Canonicalize a scale of 1 to a base register if the formula doesn't
1805 // already have a base register.
1806 if (!HasBaseReg && Scale == 1) {
1807 Scale = 0;
1808 HasBaseReg = true;
1809 }
1810
1811 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1812 HasBaseReg, Scale);
1813}
1814
1815static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1816 ScalarEvolution &SE, int64_t MinOffset,
1817 int64_t MaxOffset, LSRUse::KindType Kind,
1818 MemAccessTy AccessTy, const SCEV *S,
1819 bool HasBaseReg) {
1820 // Fast-path: zero is always foldable.
1821 if (S->isZero()) return true;
1822
1823 // Conservatively, create an address with an immediate and a
1824 // base and a scale.
1825 int64_t BaseOffset = ExtractImmediate(S, SE);
1826 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1827
1828 // If there's anything else involved, it's not foldable.
1829 if (!S->isZero()) return false;
1830
1831 // Fast-path: zero is always foldable.
1832 if (BaseOffset == 0 && !BaseGV) return true;
1833
1834 // Conservatively, create an address with an immediate and a
1835 // base and a scale.
1836 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1837
1838 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1839 BaseOffset, HasBaseReg, Scale);
1840}
1841
1842namespace {
1843
1844/// An individual increment in a Chain of IV increments. Relate an IV user to
1845/// an expression that computes the IV it uses from the IV used by the previous
1846/// link in the Chain.
1847///
1848/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1849/// original IVOperand. The head of the chain's IVOperand is only valid during
1850/// chain collection, before LSR replaces IV users. During chain generation,
1851/// IncExpr can be used to find the new IVOperand that computes the same
1852/// expression.
1853struct IVInc {
1854 Instruction *UserInst;
1855 Value* IVOperand;
1856 const SCEV *IncExpr;
1857
1858 IVInc(Instruction *U, Value *O, const SCEV *E)
1859 : UserInst(U), IVOperand(O), IncExpr(E) {}
1860};
1861
1862// The list of IV increments in program order. We typically add the head of a
1863// chain without finding subsequent links.
1864struct IVChain {
1865 SmallVector<IVInc, 1> Incs;
1866 const SCEV *ExprBase = nullptr;
1867
1868 IVChain() = default;
1869 IVChain(const IVInc &Head, const SCEV *Base)
1870 : Incs(1, Head), ExprBase(Base) {}
1871
1872 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1873
1874 // Return the first increment in the chain.
1875 const_iterator begin() const {
1876 assert(!Incs.empty())((!Incs.empty()) ? static_cast<void> (0) : __assert_fail
("!Incs.empty()", "/build/llvm-toolchain-snapshot-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1876, __PRETTY_FUNCTION__))
;
1877 return std::next(Incs.begin());
1878 }
1879 const_iterator end() const {
1880 return Incs.end();
1881 }
1882
1883 // Returns true if this chain contains any increments.
1884 bool hasIncs() const { return Incs.size() >= 2; }
1885
1886 // Add an IVInc to the end of this chain.
1887 void add(const IVInc &X) { Incs.push_back(X); }
1888
1889 // Returns the last UserInst in the chain.
1890 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1891
1892 // Returns true if IncExpr can be profitably added to this chain.
1893 bool isProfitableIncrement(const SCEV *OperExpr,
1894 const SCEV *IncExpr,
1895 ScalarEvolution&);
1896};
1897
1898/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1899/// between FarUsers that definitely cross IV increments and NearUsers that may
1900/// be used between IV increments.
1901struct ChainUsers {
1902 SmallPtrSet<Instruction*, 4> FarUsers;
1903 SmallPtrSet<Instruction*, 4> NearUsers;
1904};
1905
1906/// This class holds state for the main loop strength reduction logic.
1907class LSRInstance {
1908 IVUsers &IU;
1909 ScalarEvolution &SE;
1910 DominatorTree &DT;
1911 LoopInfo &LI;
1912 AssumptionCache &AC;
1913 TargetLibraryInfo &LibInfo;
1914 const TargetTransformInfo &TTI;
1915 Loop *const L;
1916 bool FavorBackedgeIndex = false;
1917 bool Changed = false;
1918
1919 /// This is the insert position that the current loop's induction variable
1920 /// increment should be placed. In simple loops, this is the latch block's
1921 /// terminator. But in more complicated cases, this is a position which will
1922 /// dominate all the in-loop post-increment users.
1923 Instruction *IVIncInsertPos = nullptr;
1924
1925 /// Interesting factors between use strides.
1926 ///
1927 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1928 /// default, a SmallDenseSet, because we need to use the full range of
1929 /// int64_ts, and there's currently no good way of doing that with
1930 /// SmallDenseSet.
1931 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1932
1933 /// Interesting use types, to facilitate truncation reuse.
1934 SmallSetVector<Type *, 4> Types;
1935
1936 /// The list of interesting uses.
1937 mutable SmallVector<LSRUse, 16> Uses;
1938
1939 /// Track which uses use which register candidates.
1940 RegUseTracker RegUses;
1941
1942 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1943 // have more than a few IV increment chains in a loop. Missing a Chain falls
1944 // back to normal LSR behavior for those uses.
1945 static const unsigned MaxChains = 8;
1946
1947 /// IV users can form a chain of IV increments.
1948 SmallVector<IVChain, MaxChains> IVChainVec;
1949
1950 /// IV users that belong to profitable IVChains.
1951 SmallPtrSet<Use*, MaxChains> IVIncSet;
1952
1953 void OptimizeShadowIV();
1954 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1955 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1956 void OptimizeLoopTermCond();
1957
1958 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1959 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1960 void FinalizeChain(IVChain &Chain);
1961 void CollectChains();
1962 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1963 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1964
1965 void CollectInterestingTypesAndFactors();
1966 void CollectFixupsAndInitialFormulae();
1967
1968 // Support for sharing of LSRUses between LSRFixups.
1969 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
1970 UseMapTy UseMap;
1971
1972 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1973 LSRUse::KindType Kind, MemAccessTy AccessTy);
1974
1975 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1976 MemAccessTy AccessTy);
1977
1978 void DeleteUse(LSRUse &LU, size_t LUIdx);
1979
1980 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1981
1982 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1983 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1984 void CountRegisters(const Formula &F, size_t LUIdx);
1985 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1986
1987 void CollectLoopInvariantFixupsAndFormulae();
1988
1989 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1990 unsigned Depth = 0);
1991
1992 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1993 const Formula &Base, unsigned Depth,
1994 size_t Idx, bool IsScaledReg = false);
1995 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1996 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1997 const Formula &Base, size_t Idx,
1998 bool IsScaledReg = false);
1999 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2000 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2001 const Formula &Base,
2002 const SmallVectorImpl<int64_t> &Worklist,
2003 size_t Idx, bool IsScaledReg = false);
2004 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2005 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2006 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2007 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2008 void GenerateCrossUseConstantOffsets();
2009 void GenerateAllReuseFormulae();
2010
2011 void FilterOutUndesirableDedicatedRegisters();
2012
2013 size_t EstimateSearchSpaceComplexity() const;
2014 void NarrowSearchSpaceByDetectingSupersets();
2015 void NarrowSearchSpaceByCollapsingUnrolledCode();
2016 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2017 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2018 void NarrowSearchSpaceByDeletingCostlyFormulas();
2019 void NarrowSearchSpaceByPickingWinnerRegs();
2020 void NarrowSearchSpaceUsingHeuristics();
2021
2022 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2023 Cost &SolutionCost,
2024 SmallVectorImpl<const Formula *> &Workspace,
2025 const Cost &CurCost,
2026 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2027 DenseSet<const SCEV *> &VisitedRegs) const;
2028 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2029
2030 BasicBlock::iterator
2031 HoistInsertPosition(BasicBlock::iterator IP,
2032 const SmallVectorImpl<Instruction *> &Inputs) const;
2033 BasicBlock::iterator
2034 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2035 const LSRFixup &LF,
2036 const LSRUse &LU,
2037 SCEVExpander &Rewriter) const;
2038
2039 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2040 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2041 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2042 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2043 const Formula &F, SCEVExpander &Rewriter,
2044 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2045 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2046 SCEVExpander &Rewriter,
2047 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2048 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2049
2050public:
2051 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2052 LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2053 TargetLibraryInfo &LibInfo);
2054
2055 bool getChanged() const { return Changed; }
2056
2057 void print_factors_and_types(raw_ostream &OS) const;
2058 void print_fixups(raw_ostream &OS) const;
2059 void print_uses(raw_ostream &OS) const;
2060 void print(raw_ostream &OS) const;
2061 void dump() const;
2062};
2063
2064} // end anonymous namespace
2065
2066/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2067/// the cast operation.
2068void LSRInstance::OptimizeShadowIV() {
2069 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2070 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2071 return;
2072
2073 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2074 UI != E; /* empty */) {
2075 IVUsers::const_iterator CandidateUI = UI;
2076 ++UI;
2077 Instruction *ShadowUse = CandidateUI->getUser();
2078 Type *DestTy = nullptr;
2079 bool IsSigned = false;
2080
2081 /* If shadow use is a int->float cast then insert a second IV
2082 to eliminate this cast.
2083
2084 for (unsigned i = 0; i < n; ++i)
2085 foo((double)i);
2086
2087 is transformed into
2088
2089 double d = 0.0;
2090 for (unsigned i = 0; i < n; ++i, ++d)
2091 foo(d);
2092 */
2093 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2094 IsSigned = false;
2095 DestTy = UCast->getDestTy();
2096 }
2097 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2098 IsSigned = true;
2099 DestTy = SCast->getDestTy();
2100 }
2101 if (!DestTy) continue;
2102
2103 // If target does not support DestTy natively then do not apply
2104 // this transformation.
2105 if (!TTI.isTypeLegal(DestTy)) continue;
2106
2107 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2108 if (!PH) continue;
2109 if (PH->getNumIncomingValues() != 2) continue;
2110
2111 // If the calculation in integers overflows, the result in FP type will
2112 // differ. So we only can do this transformation if we are guaranteed to not
2113 // deal with overflowing values
2114 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2115 if (!AR) continue;
2116 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2117 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2118
2119 Type *SrcTy = PH->getType();
2120 int Mantissa = DestTy->getFPMantissaWidth();
2121 if (Mantissa == -1) continue;
2122 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2123 continue;
2124
2125 unsigned Entry, Latch;
2126 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2127 Entry = 0;
2128 Latch = 1;
2129 } else {
2130 Entry = 1;
2131 Latch = 0;
2132 }
2133
2134 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2135 if (!Init) continue;
2136 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2137 (double)Init->getSExtValue() :
2138 (double)Init->getZExtValue());
2139
2140 BinaryOperator *Incr =
2141 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2142 if (!Incr) continue;
2143 if (Incr->getOpcode() != Instruction::Add
2144 && Incr->getOpcode() != Instruction::Sub)
2145 continue;
2146
2147 /* Initialize new IV, double d = 0.0 in above example. */
2148 ConstantInt *C = nullptr;
2149 if (Incr->getOperand(0) == PH)
2150 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2151 else if (Incr->getOperand(1) == PH)
2152 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2153 else
2154 continue;
2155
2156 if (!C) continue;
2157
2158 // Ignore negative constants, as the code below doesn't handle them
2159 // correctly. TODO: Remove this restriction.
2160 if (!C->getValue().isStrictlyPositive()) continue;
2161
2162 /* Add new PHINode. */
2163 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2164
2165 /* create new increment. '++d' in above example. */
2166 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2167 BinaryOperator *NewIncr =
2168 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2169 Instruction::FAdd : Instruction::FSub,
2170 NewPH, CFP, "IV.S.next.", Incr);
2171
2172 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2173 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2174
2175 /* Remove cast operation */
2176 ShadowUse->replaceAllUsesWith(NewPH);
2177 ShadowUse->eraseFromParent();
2178 Changed = true;
2179 break;
2180 }
2181}
2182
2183/// If Cond has an operand that is an expression of an IV, set the IV user and
2184/// stride information and return true, otherwise return false.
2185bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2186 for (IVStrideUse &U : IU)
2187 if (U.getUser() == Cond) {
2188 // NOTE: we could handle setcc instructions with multiple uses here, but
2189 // InstCombine does it as well for simple uses, it's not clear that it
2190 // occurs enough in real life to handle.
2191 CondUse = &U;
2192 return true;
2193 }
2194 return false;
2195}
2196
2197/// Rewrite the loop's terminating condition if it uses a max computation.
2198///
2199/// This is a narrow solution to a specific, but acute, problem. For loops
2200/// like this:
2201///
2202/// i = 0;
2203/// do {
2204/// p[i] = 0.0;
2205/// } while (++i < n);
2206///
2207/// the trip count isn't just 'n', because 'n' might not be positive. And
2208/// unfortunately this can come up even for loops where the user didn't use
2209/// a C do-while loop. For example, seemingly well-behaved top-test loops
2210/// will commonly be lowered like this:
2211///
2212/// if (n > 0) {
2213/// i = 0;
2214/// do {
2215/// p[i] = 0.0;
2216/// } while (++i < n);
2217/// }
2218///
2219/// and then it's possible for subsequent optimization to obscure the if
2220/// test in such a way that indvars can't find it.
2221///
2222/// When indvars can't find the if test in loops like this, it creates a
2223/// max expression, which allows it to give the loop a canonical
2224/// induction variable:
2225///
2226/// i = 0;
2227/// max = n < 1 ? 1 : n;
2228/// do {
2229/// p[i] = 0.0;
2230/// } while (++i != max);
2231///
2232/// Canonical induction variables are necessary because the loop passes
2233/// are designed around them. The most obvious example of this is the
2234/// LoopInfo analysis, which doesn't remember trip count values. It
2235/// expects to be able to rediscover the trip count each time it is
2236/// needed, and it does this using a simple analysis that only succeeds if
2237/// the loop has a canonical induction variable.
2238///
2239/// However, when it comes time to generate code, the maximum operation
2240/// can be quite costly, especially if it's inside of an outer loop.
2241///
2242/// This function solves this problem by detecting this type of loop and
2243/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2244/// the instructions for the maximum computation.
2245ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2246 // Check that the loop matches the pattern we're looking for.
2247 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2248 Cond->getPredicate() != CmpInst::ICMP_NE)
2249 return Cond;
2250
2251 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2252 if (!Sel || !Sel->hasOneUse()) return Cond;
2253
2254 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2255 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2256 return Cond;
2257 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2258
2259 // Add one to the backedge-taken count to get the trip count.
2260 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2261 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2262
2263 // Check for a max calculation that matches the pattern. There's no check
2264 // for ICMP_ULE here because the comparison would be with zero, which
2265 // isn't interesting.
2266 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2267 const SCEVNAryExpr *Max = nullptr;
2268 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2269 Pred = ICmpInst::ICMP_SLE;
2270 Max = S;
2271 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2272 Pred = ICmpInst::ICMP_SLT;
2273 Max = S;
2274 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2275 Pred = ICmpInst::ICMP_ULT;
2276 Max = U;
2277 } else {
2278 // No match; bail.
2279 return Cond;
2280 }
2281
2282 // To handle a max with more than two operands, this optimization would
2283 // require additional checking and setup.
2284 if (Max->getNumOperands() != 2)
2285 return Cond;
2286
2287 const SCEV *MaxLHS = Max->getOperand(0);
2288 const SCEV *MaxRHS = Max->getOperand(1);
2289
2290 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2291 // for a comparison with 1. For <= and >=, a comparison with zero.
2292 if (!MaxLHS ||
2293 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2294 return Cond;
2295
2296 // Check the relevant induction variable for conformance to
2297 // the pattern.
2298 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2299 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2300 if (!AR || !AR->isAffine() ||
2301 AR->getStart() != One ||
2302 AR->getStepRecurrence(SE) != One)
2303 return Cond;
2304
2305 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2306, __PRETTY_FUNCTION__))
2306 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2306, __PRETTY_FUNCTION__))
;
2307
2308 // Check the right operand of the select, and remember it, as it will
2309 // be used in the new comparison instruction.
2310 Value *NewRHS = nullptr;
2311 if (ICmpInst::isTrueWhenEqual(Pred)) {
2312 // Look for n+1, and grab n.
2313 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2314 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2315 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2316 NewRHS = BO->getOperand(0);
2317 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2318 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2319 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2320 NewRHS = BO->getOperand(0);
2321 if (!NewRHS)
2322 return Cond;
2323 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2324 NewRHS = Sel->getOperand(1);
2325 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2326 NewRHS = Sel->getOperand(2);
2327 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2328 NewRHS = SU->getValue();
2329 else
2330 // Max doesn't match expected pattern.
2331 return Cond;
2332
2333 // Determine the new comparison opcode. It may be signed or unsigned,
2334 // and the original comparison may be either equality or inequality.
2335 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2336 Pred = CmpInst::getInversePredicate(Pred);
2337
2338 // Ok, everything looks ok to change the condition into an SLT or SGE and
2339 // delete the max calculation.
2340 ICmpInst *NewCond =
2341 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2342
2343 // Delete the max calculation instructions.
2344 Cond->replaceAllUsesWith(NewCond);
2345 CondUse->setUser(NewCond);
2346 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2347 Cond->eraseFromParent();
2348 Sel->eraseFromParent();
2349 if (Cmp->use_empty())
2350 Cmp->eraseFromParent();
2351 return NewCond;
2352}
2353
2354/// Change loop terminating condition to use the postinc iv when possible.
2355void
2356LSRInstance::OptimizeLoopTermCond() {
2357 SmallPtrSet<Instruction *, 4> PostIncs;
2358
2359 // We need a different set of heuristics for rotated and non-rotated loops.
2360 // If a loop is rotated then the latch is also the backedge, so inserting
2361 // post-inc expressions just before the latch is ideal. To reduce live ranges
2362 // it also makes sense to rewrite terminating conditions to use post-inc
2363 // expressions.
2364 //
2365 // If the loop is not rotated then the latch is not a backedge; the latch
2366 // check is done in the loop head. Adding post-inc expressions before the
2367 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2368 // in the loop body. In this case we do *not* want to use post-inc expressions
2369 // in the latch check, and we want to insert post-inc expressions before
2370 // the backedge.
2371 BasicBlock *LatchBlock = L->getLoopLatch();
2372 SmallVector<BasicBlock*, 8> ExitingBlocks;
2373 L->getExitingBlocks(ExitingBlocks);
2374 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2375 return LatchBlock != BB;
2376 })) {
2377 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2378 IVIncInsertPos = LatchBlock->getTerminator();
2379 return;
2380 }
2381
2382 // Otherwise treat this as a rotated loop.
2383 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2384 // Get the terminating condition for the loop if possible. If we
2385 // can, we want to change it to use a post-incremented version of its
2386 // induction variable, to allow coalescing the live ranges for the IV into
2387 // one register value.
2388
2389 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2390 if (!TermBr)
2391 continue;
2392 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2393 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2394 continue;
2395
2396 // Search IVUsesByStride to find Cond's IVUse if there is one.
2397 IVStrideUse *CondUse = nullptr;
2398 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2399 if (!FindIVUserForCond(Cond, CondUse))
2400 continue;
2401
2402 // If the trip count is computed in terms of a max (due to ScalarEvolution
2403 // being unable to find a sufficient guard, for example), change the loop
2404 // comparison to use SLT or ULT instead of NE.
2405 // One consequence of doing this now is that it disrupts the count-down
2406 // optimization. That's not always a bad thing though, because in such
2407 // cases it may still be worthwhile to avoid a max.
2408 Cond = OptimizeMax(Cond, CondUse);
2409
2410 // If this exiting block dominates the latch block, it may also use
2411 // the post-inc value if it won't be shared with other uses.
2412 // Check for dominance.
2413 if (!DT.dominates(ExitingBlock, LatchBlock))
2414 continue;
2415
2416 // Conservatively avoid trying to use the post-inc value in non-latch
2417 // exits if there may be pre-inc users in intervening blocks.
2418 if (LatchBlock != ExitingBlock)
2419 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2420 // Test if the use is reachable from the exiting block. This dominator
2421 // query is a conservative approximation of reachability.
2422 if (&*UI != CondUse &&
2423 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2424 // Conservatively assume there may be reuse if the quotient of their
2425 // strides could be a legal scale.
2426 const SCEV *A = IU.getStride(*CondUse, L);
2427 const SCEV *B = IU.getStride(*UI, L);
2428 if (!A || !B) continue;
2429 if (SE.getTypeSizeInBits(A->getType()) !=
2430 SE.getTypeSizeInBits(B->getType())) {
2431 if (SE.getTypeSizeInBits(A->getType()) >
2432 SE.getTypeSizeInBits(B->getType()))
2433 B = SE.getSignExtendExpr(B, A->getType());
2434 else
2435 A = SE.getSignExtendExpr(A, B->getType());
2436 }
2437 if (const SCEVConstant *D =
2438 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2439 const ConstantInt *C = D->getValue();
2440 // Stride of one or negative one can have reuse with non-addresses.
2441 if (C->isOne() || C->isMinusOne())
2442 goto decline_post_inc;
2443 // Avoid weird situations.
2444 if (C->getValue().getMinSignedBits() >= 64 ||
2445 C->getValue().isMinSignedValue())
2446 goto decline_post_inc;
2447 // Check for possible scaled-address reuse.
2448 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2449 MemAccessTy AccessTy = getAccessType(
2450 TTI, UI->getUser(), UI->getOperandValToReplace());
2451 int64_t Scale = C->getSExtValue();
2452 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2453 /*BaseOffset=*/0,
2454 /*HasBaseReg=*/false, Scale,
2455 AccessTy.AddrSpace))
2456 goto decline_post_inc;
2457 Scale = -Scale;
2458 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2459 /*BaseOffset=*/0,
2460 /*HasBaseReg=*/false, Scale,
2461 AccessTy.AddrSpace))
2462 goto decline_post_inc;
2463 }
2464 }
2465 }
2466
2467 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)
2468 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2469
2470 // It's possible for the setcc instruction to be anywhere in the loop, and
2471 // possible for it to have multiple users. If it is not immediately before
2472 // the exiting block branch, move it.
2473 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2474 if (Cond->hasOneUse()) {
2475 Cond->moveBefore(TermBr);
2476 } else {
2477 // Clone the terminating condition and insert into the loopend.
2478 ICmpInst *OldCond = Cond;
2479 Cond = cast<ICmpInst>(Cond->clone());
2480 Cond->setName(L->getHeader()->getName() + ".termcond");
2481 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2482
2483 // Clone the IVUse, as the old use still exists!
2484 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2485 TermBr->replaceUsesOfWith(OldCond, Cond);
2486 }
2487 }
2488
2489 // If we get to here, we know that we can transform the setcc instruction to
2490 // use the post-incremented version of the IV, allowing us to coalesce the
2491 // live ranges for the IV correctly.
2492 CondUse->transformToPostInc(L);
2493 Changed = true;
2494
2495 PostIncs.insert(Cond);
2496 decline_post_inc:;
2497 }
2498
2499 // Determine an insertion point for the loop induction variable increment. It
2500 // must dominate all the post-inc comparisons we just set up, and it must
2501 // dominate the loop latch edge.
2502 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2503 for (Instruction *Inst : PostIncs) {
2504 BasicBlock *BB =
2505 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2506 Inst->getParent());
2507 if (BB == Inst->getParent())
2508 IVIncInsertPos = Inst;
2509 else if (BB != IVIncInsertPos->getParent())
2510 IVIncInsertPos = BB->getTerminator();
2511 }
2512}
2513
2514/// Determine if the given use can accommodate a fixup at the given offset and
2515/// other details. If so, update the use and return true.
2516bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2517 bool HasBaseReg, LSRUse::KindType Kind,
2518 MemAccessTy AccessTy) {
2519 int64_t NewMinOffset = LU.MinOffset;
2520 int64_t NewMaxOffset = LU.MaxOffset;
2521 MemAccessTy NewAccessTy = AccessTy;
2522
2523 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2524 // something conservative, however this can pessimize in the case that one of
2525 // the uses will have all its uses outside the loop, for example.
2526 if (LU.Kind != Kind)
2527 return false;
2528
2529 // Check for a mismatched access type, and fall back conservatively as needed.
2530 // TODO: Be less conservative when the type is similar and can use the same
2531 // addressing modes.
2532 if (Kind == LSRUse::Address) {
2533 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2534 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2535 AccessTy.AddrSpace);
2536 }
2537 }
2538
2539 // Conservatively assume HasBaseReg is true for now.
2540 if (NewOffset < LU.MinOffset) {
2541 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2542 LU.MaxOffset - NewOffset, HasBaseReg))
2543 return false;
2544 NewMinOffset = NewOffset;
2545 } else if (NewOffset > LU.MaxOffset) {
2546 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2547 NewOffset - LU.MinOffset, HasBaseReg))
2548 return false;
2549 NewMaxOffset = NewOffset;
2550 }
2551
2552 // Update the use.
2553 LU.MinOffset = NewMinOffset;
2554 LU.MaxOffset = NewMaxOffset;
2555 LU.AccessTy = NewAccessTy;
2556 return true;
2557}
2558
2559/// Return an LSRUse index and an offset value for a fixup which needs the given
2560/// expression, with the given kind and optional access type. Either reuse an
2561/// existing use or create a new one, as needed.
2562std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2563 LSRUse::KindType Kind,
2564 MemAccessTy AccessTy) {
2565 const SCEV *Copy = Expr;
2566 int64_t Offset = ExtractImmediate(Expr, SE);
2567
2568 // Basic uses can't accept any offset, for example.
2569 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2570 Offset, /*HasBaseReg=*/ true)) {
2571 Expr = Copy;
2572 Offset = 0;
2573 }
2574
2575 std::pair<UseMapTy::iterator, bool> P =
2576 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2577 if (!P.second) {
2578 // A use already existed with this base.
2579 size_t LUIdx = P.first->second;
2580 LSRUse &LU = Uses[LUIdx];
2581 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2582 // Reuse this use.
2583 return std::make_pair(LUIdx, Offset);
2584 }
2585
2586 // Create a new use.
2587 size_t LUIdx = Uses.size();
2588 P.first->second = LUIdx;
2589 Uses.push_back(LSRUse(Kind, AccessTy));
2590 LSRUse &LU = Uses[LUIdx];
2591
2592 LU.MinOffset = Offset;
2593 LU.MaxOffset = Offset;
2594 return std::make_pair(LUIdx, Offset);
2595}
2596
2597/// Delete the given use from the Uses list.
2598void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2599 if (&LU != &Uses.back())
2600 std::swap(LU, Uses.back());
2601 Uses.pop_back();
2602
2603 // Update RegUses.
2604 RegUses.swapAndDropUse(LUIdx, Uses.size());
2605}
2606
2607/// Look for a use distinct from OrigLU which is has a formula that has the same
2608/// registers as the given formula.
2609LSRUse *
2610LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2611 const LSRUse &OrigLU) {
2612 // Search all uses for the formula. This could be more clever.
2613 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2614 LSRUse &LU = Uses[LUIdx];
2615 // Check whether this use is close enough to OrigLU, to see whether it's
2616 // worthwhile looking through its formulae.
2617 // Ignore ICmpZero uses because they may contain formulae generated by
2618 // GenerateICmpZeroScales, in which case adding fixup offsets may
2619 // be invalid.
2620 if (&LU != &OrigLU &&
2621 LU.Kind != LSRUse::ICmpZero &&
2622 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2623 LU.WidestFixupType == OrigLU.WidestFixupType &&
2624 LU.HasFormulaWithSameRegs(OrigF)) {
2625 // Scan through this use's formulae.
2626 for (const Formula &F : LU.Formulae) {
2627 // Check to see if this formula has the same registers and symbols
2628 // as OrigF.
2629 if (F.BaseRegs == OrigF.BaseRegs &&
2630 F.ScaledReg == OrigF.ScaledReg &&
2631 F.BaseGV == OrigF.BaseGV &&
2632 F.Scale == OrigF.Scale &&
2633 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2634 if (F.BaseOffset == 0)
2635 return &LU;
2636 // This is the formula where all the registers and symbols matched;
2637 // there aren't going to be any others. Since we declined it, we
2638 // can skip the rest of the formulae and proceed to the next LSRUse.
2639 break;
2640 }
2641 }
2642 }
2643 }
2644
2645 // Nothing looked good.
2646 return nullptr;
2647}
2648
2649void LSRInstance::CollectInterestingTypesAndFactors() {
2650 SmallSetVector<const SCEV *, 4> Strides;
2651
2652 // Collect interesting types and strides.
2653 SmallVector<const SCEV *, 4> Worklist;
2654 for (const IVStrideUse &U : IU) {
2655 const SCEV *Expr = IU.getExpr(U);
2656
2657 // Collect interesting types.
2658 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2659
2660 // Add strides for mentioned loops.
2661 Worklist.push_back(Expr);
2662 do {
2663 const SCEV *S = Worklist.pop_back_val();
2664 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2665 if (AR->getLoop() == L)
2666 Strides.insert(AR->getStepRecurrence(SE));
2667 Worklist.push_back(AR->getStart());
2668 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2669 Worklist.append(Add->op_begin(), Add->op_end());
2670 }
2671 } while (!Worklist.empty());
2672 }
2673
2674 // Compute interesting factors from the set of interesting strides.
2675 for (SmallSetVector<const SCEV *, 4>::const_iterator
2676 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2677 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2678 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2679 const SCEV *OldStride = *I;
2680 const SCEV *NewStride = *NewStrideIter;
2681
2682 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2683 SE.getTypeSizeInBits(NewStride->getType())) {
2684 if (SE.getTypeSizeInBits(OldStride->getType()) >
2685 SE.getTypeSizeInBits(NewStride->getType()))
2686 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2687 else
2688 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2689 }
2690 if (const SCEVConstant *Factor =
2691 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2692 SE, true))) {
2693 if (Factor->getAPInt().getMinSignedBits() <= 64)
2694 Factors.insert(Factor->getAPInt().getSExtValue());
2695 } else if (const SCEVConstant *Factor =
2696 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2697 NewStride,
2698 SE, true))) {
2699 if (Factor->getAPInt().getMinSignedBits() <= 64)
2700 Factors.insert(Factor->getAPInt().getSExtValue());
2701 }
2702 }
2703
2704 // If all uses use the same type, don't bother looking for truncation-based
2705 // reuse.
2706 if (Types.size() == 1)
2707 Types.clear();
2708
2709 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2710}
2711
2712/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2713/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2714/// IVStrideUses, we could partially skip this.
2715static User::op_iterator
2716findIVOperand(User::op_iterator OI, User::op_iterator OE,
2717 Loop *L, ScalarEvolution &SE) {
2718 for(; OI != OE; ++OI) {
2719 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2720 if (!SE.isSCEVable(Oper->getType()))
2721 continue;
2722
2723 if (const SCEVAddRecExpr *AR =
2724 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2725 if (AR->getLoop() == L)
2726 break;
2727 }
2728 }
2729 }
2730 return OI;
2731}
2732
2733/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2734/// a convenient helper.
2735static Value *getWideOperand(Value *Oper) {
2736 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2737 return Trunc->getOperand(0);
2738 return Oper;
2739}
2740
2741/// Return true if we allow an IV chain to include both types.
2742static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2743 Type *LType = LVal->getType();
2744 Type *RType = RVal->getType();
2745 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2746 // Different address spaces means (possibly)
2747 // different types of the pointer implementation,
2748 // e.g. i16 vs i32 so disallow that.
2749 (LType->getPointerAddressSpace() ==
2750 RType->getPointerAddressSpace()));
2751}
2752
2753/// Return an approximation of this SCEV expression's "base", or NULL for any
2754/// constant. Returning the expression itself is conservative. Returning a
2755/// deeper subexpression is more precise and valid as long as it isn't less
2756/// complex than another subexpression. For expressions involving multiple
2757/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2758/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2759/// IVInc==b-a.
2760///
2761/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2762/// SCEVUnknown, we simply return the rightmost SCEV operand.
2763static const SCEV *getExprBase(const SCEV *S) {
2764 switch (S->getSCEVType()) {
2765 default: // uncluding scUnknown.
2766 return S;
2767 case scConstant:
2768 return nullptr;
2769 case scTruncate:
2770 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2771 case scZeroExtend:
2772 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2773 case scSignExtend:
2774 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2775 case scAddExpr: {
2776 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2777 // there's nothing more complex.
2778 // FIXME: not sure if we want to recognize negation.
2779 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2780 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2781 E(Add->op_begin()); I != E; ++I) {
2782 const SCEV *SubExpr = *I;
2783 if (SubExpr->getSCEVType() == scAddExpr)
2784 return getExprBase(SubExpr);
2785
2786 if (SubExpr->getSCEVType() != scMulExpr)
2787 return SubExpr;
2788 }
2789 return S; // all operands are scaled, be conservative.
2790 }
2791 case scAddRecExpr:
2792 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2793 }
2794}
2795
2796/// Return true if the chain increment is profitable to expand into a loop
2797/// invariant value, which may require its own register. A profitable chain
2798/// increment will be an offset relative to the same base. We allow such offsets
2799/// to potentially be used as chain increment as long as it's not obviously
2800/// expensive to expand using real instructions.
2801bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2802 const SCEV *IncExpr,
2803 ScalarEvolution &SE) {
2804 // Aggressively form chains when -stress-ivchain.
2805 if (StressIVChain)
2806 return true;
2807
2808 // Do not replace a constant offset from IV head with a nonconstant IV
2809 // increment.
2810 if (!isa<SCEVConstant>(IncExpr)) {
2811 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2812 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2813 return false;
2814 }
2815
2816 SmallPtrSet<const SCEV*, 8> Processed;
2817 return !isHighCostExpansion(IncExpr, Processed, SE);
2818}
2819
2820/// Return true if the number of registers needed for the chain is estimated to
2821/// be less than the number required for the individual IV users. First prohibit
2822/// any IV users that keep the IV live across increments (the Users set should
2823/// be empty). Next count the number and type of increments in the chain.
2824///
2825/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2826/// effectively use postinc addressing modes. Only consider it profitable it the
2827/// increments can be computed in fewer registers when chained.
2828///
2829/// TODO: Consider IVInc free if it's already used in another chains.
2830static bool
2831isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2832 ScalarEvolution &SE) {
2833 if (StressIVChain)
2834 return true;
2835
2836 if (!Chain.hasIncs())
2837 return false;
2838
2839 if (!Users.empty()) {
2840 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)
2841 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)
2842 : 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)
;
2843 return false;
2844 }
2845 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2845, __PRETTY_FUNCTION__))
;
2846
2847 // The chain itself may require a register, so intialize cost to 1.
2848 int cost = 1;
2849
2850 // A complete chain likely eliminates the need for keeping the original IV in
2851 // a register. LSR does not currently know how to form a complete chain unless
2852 // the header phi already exists.
2853 if (isa<PHINode>(Chain.tailUserInst())
2854 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2855 --cost;
2856 }
2857 const SCEV *LastIncExpr = nullptr;
2858 unsigned NumConstIncrements = 0;
2859 unsigned NumVarIncrements = 0;
2860 unsigned NumReusedIncrements = 0;
2861 for (const IVInc &Inc : Chain) {
2862 if (Inc.IncExpr->isZero())
2863 continue;
2864
2865 // Incrementing by zero or some constant is neutral. We assume constants can
2866 // be folded into an addressing mode or an add's immediate operand.
2867 if (isa<SCEVConstant>(Inc.IncExpr)) {
2868 ++NumConstIncrements;
2869 continue;
2870 }
2871
2872 if (Inc.IncExpr == LastIncExpr)
2873 ++NumReusedIncrements;
2874 else
2875 ++NumVarIncrements;
2876
2877 LastIncExpr = Inc.IncExpr;
2878 }
2879 // An IV chain with a single increment is handled by LSR's postinc
2880 // uses. However, a chain with multiple increments requires keeping the IV's
2881 // value live longer than it needs to be if chained.
2882 if (NumConstIncrements > 1)
2883 --cost;
2884
2885 // Materializing increment expressions in the preheader that didn't exist in
2886 // the original code may cost a register. For example, sign-extended array
2887 // indices can produce ridiculous increments like this:
2888 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2889 cost += NumVarIncrements;
2890
2891 // Reusing variable increments likely saves a register to hold the multiple of
2892 // the stride.
2893 cost -= NumReusedIncrements;
2894
2895 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)
2896 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2897
2898 return cost < 0;
2899}
2900
2901/// Add this IV user to an existing chain or make it the head of a new chain.
2902void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2903 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2904 // When IVs are used as types of varying widths, they are generally converted
2905 // to a wider type with some uses remaining narrow under a (free) trunc.
2906 Value *const NextIV = getWideOperand(IVOper);
2907 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2908 const SCEV *const OperExprBase = getExprBase(OperExpr);
2909
2910 // Visit all existing chains. Check if its IVOper can be computed as a
2911 // profitable loop invariant increment from the last link in the Chain.
2912 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2913 const SCEV *LastIncExpr = nullptr;
2914 for (; ChainIdx < NChains; ++ChainIdx) {
2915 IVChain &Chain = IVChainVec[ChainIdx];
2916
2917 // Prune the solution space aggressively by checking that both IV operands
2918 // are expressions that operate on the same unscaled SCEVUnknown. This
2919 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2920 // first avoids creating extra SCEV expressions.
2921 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2922 continue;
2923
2924 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2925 if (!isCompatibleIVType(PrevIV, NextIV))
2926 continue;
2927
2928 // A phi node terminates a chain.
2929 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2930 continue;
2931
2932 // The increment must be loop-invariant so it can be kept in a register.
2933 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2934 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2935 if (!SE.isLoopInvariant(IncExpr, L))
2936 continue;
2937
2938 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2939 LastIncExpr = IncExpr;
2940 break;
2941 }
2942 }
2943 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2944 // bother for phi nodes, because they must be last in the chain.
2945 if (ChainIdx == NChains) {
2946 if (isa<PHINode>(UserInst))
2947 return;
2948 if (NChains >= MaxChains && !StressIVChain) {
2949 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2950 return;
2951 }
2952 LastIncExpr = OperExpr;
2953 // IVUsers may have skipped over sign/zero extensions. We don't currently
2954 // attempt to form chains involving extensions unless they can be hoisted
2955 // into this loop's AddRec.
2956 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2957 return;
2958 ++NChains;
2959 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2960 OperExprBase));
2961 ChainUsersVec.resize(NChains);
2962 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)
2963 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
2964 } else {
2965 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)
2966 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
2967 // Add this IV user to the end of the chain.
2968 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2969 }
2970 IVChain &Chain = IVChainVec[ChainIdx];
2971
2972 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2973 // This chain's NearUsers become FarUsers.
2974 if (!LastIncExpr->isZero()) {
2975 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2976 NearUsers.end());
2977 NearUsers.clear();
2978 }
2979
2980 // All other uses of IVOperand become near uses of the chain.
2981 // We currently ignore intermediate values within SCEV expressions, assuming
2982 // they will eventually be used be the current chain, or can be computed
2983 // from one of the chain increments. To be more precise we could
2984 // transitively follow its user and only add leaf IV users to the set.
2985 for (User *U : IVOper->users()) {
2986 Instruction *OtherUse = dyn_cast<Instruction>(U);
2987 if (!OtherUse)
2988 continue;
2989 // Uses in the chain will no longer be uses if the chain is formed.
2990 // Include the head of the chain in this iteration (not Chain.begin()).
2991 IVChain::const_iterator IncIter = Chain.Incs.begin();
2992 IVChain::const_iterator IncEnd = Chain.Incs.end();
2993 for( ; IncIter != IncEnd; ++IncIter) {
2994 if (IncIter->UserInst == OtherUse)
2995 break;
2996 }
2997 if (IncIter != IncEnd)
2998 continue;
2999
3000 if (SE.isSCEVable(OtherUse->getType())
3001 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3002 && IU.isIVUserOrOperand(OtherUse)) {
3003 continue;
3004 }
3005 NearUsers.insert(OtherUse);
3006 }
3007
3008 // Since this user is part of the chain, it's no longer considered a use
3009 // of the chain.
3010 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3011}
3012
3013/// Populate the vector of Chains.
3014///
3015/// This decreases ILP at the architecture level. Targets with ample registers,
3016/// multiple memory ports, and no register renaming probably don't want
3017/// this. However, such targets should probably disable LSR altogether.
3018///
3019/// The job of LSR is to make a reasonable choice of induction variables across
3020/// the loop. Subsequent passes can easily "unchain" computation exposing more
3021/// ILP *within the loop* if the target wants it.
3022///
3023/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3024/// will not reorder memory operations, it will recognize this as a chain, but
3025/// will generate redundant IV increments. Ideally this would be corrected later
3026/// by a smart scheduler:
3027/// = A[i]
3028/// = A[i+x]
3029/// A[i] =
3030/// A[i+x] =
3031///
3032/// TODO: Walk the entire domtree within this loop, not just the path to the
3033/// loop latch. This will discover chains on side paths, but requires
3034/// maintaining multiple copies of the Chains state.
3035void LSRInstance::CollectChains() {
3036 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3037 SmallVector<ChainUsers, 8> ChainUsersVec;
3038
3039 SmallVector<BasicBlock *,8> LatchPath;
3040 BasicBlock *LoopHeader = L->getHeader();
3041 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3042 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3043 LatchPath.push_back(Rung->getBlock());
3044 }
3045 LatchPath.push_back(LoopHeader);
3046
3047 // Walk the instruction stream from the loop header to the loop latch.
3048 for (BasicBlock *BB : reverse(LatchPath)) {
3049 for (Instruction &I : *BB) {
3050 // Skip instructions that weren't seen by IVUsers analysis.
3051 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3052 continue;
3053
3054 // Ignore users that are part of a SCEV expression. This way we only
3055 // consider leaf IV Users. This effectively rediscovers a portion of
3056 // IVUsers analysis but in program order this time.
3057 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3058 continue;
3059
3060 // Remove this instruction from any NearUsers set it may be in.
3061 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3062 ChainIdx < NChains; ++ChainIdx) {
3063 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3064 }
3065 // Search for operands that can be chained.
3066 SmallPtrSet<Instruction*, 4> UniqueOperands;
3067 User::op_iterator IVOpEnd = I.op_end();
3068 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3069 while (IVOpIter != IVOpEnd) {
3070 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3071 if (UniqueOperands.insert(IVOpInst).second)
3072 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3073 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3074 }
3075 } // Continue walking down the instructions.
3076 } // Continue walking down the domtree.
3077 // Visit phi backedges to determine if the chain can generate the IV postinc.
3078 for (PHINode &PN : L->getHeader()->phis()) {
3079 if (!SE.isSCEVable(PN.getType()))
3080 continue;
3081
3082 Instruction *IncV =
3083 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3084 if (IncV)
3085 ChainInstruction(&PN, IncV, ChainUsersVec);
3086 }
3087 // Remove any unprofitable chains.
3088 unsigned ChainIdx = 0;
3089 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3090 UsersIdx < NChains; ++UsersIdx) {
3091 if (!isProfitableChain(IVChainVec[UsersIdx],
3092 ChainUsersVec[UsersIdx].FarUsers, SE))
3093 continue;
3094 // Preserve the chain at UsesIdx.
3095 if (ChainIdx != UsersIdx)
3096 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3097 FinalizeChain(IVChainVec[ChainIdx]);
3098 ++ChainIdx;
3099 }
3100 IVChainVec.resize(ChainIdx);
3101}
3102
3103void LSRInstance::FinalizeChain(IVChain &Chain) {
3104 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3104, __PRETTY_FUNCTION__))
;
3105 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)
;
3106
3107 for (const IVInc &Inc : Chain) {
3108 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3109 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3110 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3110, __PRETTY_FUNCTION__))
;
3111 IVIncSet.insert(UseI);
3112 }
3113}
3114
3115/// Return true if the IVInc can be folded into an addressing mode.
3116static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3117 Value *Operand, const TargetTransformInfo &TTI) {
3118 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3119 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3120 return false;
3121
3122 if (IncConst->getAPInt().getMinSignedBits() > 64)
3123 return false;
3124
3125 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3126 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3127 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3128 IncOffset, /*HasBaseReg=*/false))
3129 return false;
3130
3131 return true;
3132}
3133
3134/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3135/// user's operand from the previous IV user's operand.
3136void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3137 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3138 // Find the new IVOperand for the head of the chain. It may have been replaced
3139 // by LSR.
3140 const IVInc &Head = Chain.Incs[0];
3141 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3142 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3143 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3144 IVOpEnd, L, SE);
3145 Value *IVSrc = nullptr;
3146 while (IVOpIter != IVOpEnd) {
3147 IVSrc = getWideOperand(*IVOpIter);
3148
3149 // If this operand computes the expression that the chain needs, we may use
3150 // it. (Check this after setting IVSrc which is used below.)
3151 //
3152 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3153 // narrow for the chain, so we can no longer use it. We do allow using a
3154 // wider phi, assuming the LSR checked for free truncation. In that case we
3155 // should already have a truncate on this operand such that
3156 // getSCEV(IVSrc) == IncExpr.
3157 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3158 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3159 break;
3160 }
3161 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3162 }
3163 if (IVOpIter == IVOpEnd) {
3164 // Gracefully give up on this chain.
3165 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)
;
3166 return;
3167 }
3168
3169 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3170 Type *IVTy = IVSrc->getType();
3171 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3172 const SCEV *LeftOverExpr = nullptr;
3173 for (const IVInc &Inc : Chain) {
3174 Instruction *InsertPt = Inc.UserInst;
3175 if (isa<PHINode>(InsertPt))
3176 InsertPt = L->getLoopLatch()->getTerminator();
3177
3178 // IVOper will replace the current IV User's operand. IVSrc is the IV
3179 // value currently held in a register.
3180 Value *IVOper = IVSrc;
3181 if (!Inc.IncExpr->isZero()) {
3182 // IncExpr was the result of subtraction of two narrow values, so must
3183 // be signed.
3184 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3185 LeftOverExpr = LeftOverExpr ?
3186 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3187 }
3188 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3189 // Expand the IV increment.
3190 Rewriter.clearPostInc();
3191 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3192 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3193 SE.getUnknown(IncV));
3194 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3195
3196 // If an IV increment can't be folded, use it as the next IV value.
3197 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3198 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3198, __PRETTY_FUNCTION__))
;
3199 IVSrc = IVOper;
3200 LeftOverExpr = nullptr;
3201 }
3202 }
3203 Type *OperTy = Inc.IVOperand->getType();
3204 if (IVTy != OperTy) {
3205 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3206, __PRETTY_FUNCTION__))
3206 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3206, __PRETTY_FUNCTION__))
;
3207 IRBuilder<> Builder(InsertPt);
3208 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3209 }
3210 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3211 DeadInsts.emplace_back(Inc.IVOperand);
3212 }
3213 // If LSR created a new, wider phi, we may also replace its postinc. We only
3214 // do this if we also found a wide value for the head of the chain.
3215 if (isa<PHINode>(Chain.tailUserInst())) {
3216 for (PHINode &Phi : L->getHeader()->phis()) {
3217 if (!isCompatibleIVType(&Phi, IVSrc))
3218 continue;
3219 Instruction *PostIncV = dyn_cast<Instruction>(
3220 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3221 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3222 continue;
3223 Value *IVOper = IVSrc;
3224 Type *PostIncTy = PostIncV->getType();
3225 if (IVTy != PostIncTy) {
3226 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3226, __PRETTY_FUNCTION__))
;
3227 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3228 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3229 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3230 }
3231 Phi.replaceUsesOfWith(PostIncV, IVOper);
3232 DeadInsts.emplace_back(PostIncV);
3233 }
3234 }
3235}
3236
3237void LSRInstance::CollectFixupsAndInitialFormulae() {
3238 BranchInst *ExitBranch = nullptr;
3239 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &LibInfo);
3240
3241 for (const IVStrideUse &U : IU) {
3242 Instruction *UserInst = U.getUser();
3243 // Skip IV users that are part of profitable IV Chains.
3244 User::op_iterator UseI =
3245 find(UserInst->operands(), U.getOperandValToReplace());
3246 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3246, __PRETTY_FUNCTION__))
;
3247 if (IVIncSet.count(UseI)) {
3248 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)
;
3249 continue;
3250 }
3251
3252 LSRUse::KindType Kind = LSRUse::Basic;
3253 MemAccessTy AccessTy;
3254 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3255 Kind = LSRUse::Address;
3256 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3257 }
3258
3259 const SCEV *S = IU.getExpr(U);
3260 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3261
3262 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3263 // (N - i == 0), and this allows (N - i) to be the expression that we work
3264 // with rather than just N or i, so we can consider the register
3265 // requirements for both N and i at the same time. Limiting this code to
3266 // equality icmps is not a problem because all interesting loops use
3267 // equality icmps, thanks to IndVarSimplify.
3268 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3269 // If CI can be saved in some target, like replaced inside hardware loop
3270 // in PowerPC, no need to generate initial formulae for it.
3271 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3272 continue;
3273 if (CI->isEquality()) {
3274 // Swap the operands if needed to put the OperandValToReplace on the
3275 // left, for consistency.
3276 Value *NV = CI->getOperand(1);
3277 if (NV == U.getOperandValToReplace()) {
3278 CI->setOperand(1, CI->getOperand(0));
3279 CI->setOperand(0, NV);
3280 NV = CI->getOperand(1);
3281 Changed = true;
3282 }
3283
3284 // x == y --> x - y == 0
3285 const SCEV *N = SE.getSCEV(NV);
3286 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3287 // S is normalized, so normalize N before folding it into S
3288 // to keep the result normalized.
3289 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3290 Kind = LSRUse::ICmpZero;
3291 S = SE.getMinusSCEV(N, S);
3292 }
3293
3294 // -1 and the negations of all interesting strides (except the negation
3295 // of -1) are now also interesting.
3296 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3297 if (Factors[i] != -1)
3298 Factors.insert(-(uint64_t)Factors[i]);
3299 Factors.insert(-1);
3300 }
3301 }
3302
3303 // Get or create an LSRUse.
3304 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3305 size_t LUIdx = P.first;
3306 int64_t Offset = P.second;
3307 LSRUse &LU = Uses[LUIdx];
3308
3309 // Record the fixup.
3310 LSRFixup &LF = LU.getNewFixup();
3311 LF.UserInst = UserInst;
3312 LF.OperandValToReplace = U.getOperandValToReplace();
3313 LF.PostIncLoops = TmpPostIncLoops;
3314 LF.Offset = Offset;
3315 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3316
3317 if (!LU.WidestFixupType ||
3318 SE.getTypeSizeInBits(LU.WidestFixupType) <
3319 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3320 LU.WidestFixupType = LF.OperandValToReplace->getType();
3321
3322 // If this is the first use of this LSRUse, give it a formula.
3323 if (LU.Formulae.empty()) {
3324 InsertInitialFormula(S, LU, LUIdx);
3325 CountRegisters(LU.Formulae.back(), LUIdx);
3326 }
3327 }
3328
3329 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3330}
3331
3332/// Insert a formula for the given expression into the given use, separating out
3333/// loop-variant portions from loop-invariant and loop-computable portions.
3334void
3335LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3336 // Mark uses whose expressions cannot be expanded.
3337 if (!isSafeToExpand(S, SE))
3338 LU.RigidFormula = true;
3339
3340 Formula F;
3341 F.initialMatch(S, L, SE);
3342 bool Inserted = InsertFormula(LU, LUIdx, F);
3343 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3343, __PRETTY_FUNCTION__))
; (void)Inserted;
3344}
3345
3346/// Insert a simple single-register formula for the given expression into the
3347/// given use.
3348void
3349LSRInstance::InsertSupplementalFormula(const SCEV *S,
3350 LSRUse &LU, size_t LUIdx) {
3351 Formula F;
3352 F.BaseRegs.push_back(S);
3353 F.HasBaseReg = true;
3354 bool Inserted = InsertFormula(LU, LUIdx, F);
3355 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3355, __PRETTY_FUNCTION__))
; (void)Inserted;
3356}
3357
3358/// Note which registers are used by the given formula, updating RegUses.
3359void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3360 if (F.ScaledReg)
3361 RegUses.countRegister(F.ScaledReg, LUIdx);
3362 for (const SCEV *BaseReg : F.BaseRegs)
3363 RegUses.countRegister(BaseReg, LUIdx);
3364}
3365
3366/// If the given formula has not yet been inserted, add it to the list, and
3367/// return true. Return false otherwise.
3368bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3369 // Do not insert formula that we will not be able to expand.
3370 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3371, __PRETTY_FUNCTION__))
3371 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3371, __PRETTY_FUNCTION__))
;
3372
3373 if (!LU.InsertFormula(F, *L))
3374 return false;
3375
3376 CountRegisters(F, LUIdx);
3377 return true;
3378}
3379
3380/// Check for other uses of loop-invariant values which we're tracking. These
3381/// other uses will pin these values in registers, making them less profitable
3382/// for elimination.
3383/// TODO: This currently misses non-constant addrec step registers.
3384/// TODO: Should this give more weight to users inside the loop?
3385void
3386LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3387 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3388 SmallPtrSet<const SCEV *, 32> Visited;
3389
3390 while (!Worklist.empty()) {
3391 const SCEV *S = Worklist.pop_back_val();
3392
3393 // Don't process the same SCEV twice
3394 if (!Visited.insert(S).second)
3395 continue;
3396
3397 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3398 Worklist.append(N->op_begin(), N->op_end());
3399 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3400 Worklist.push_back(C->getOperand());
3401 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3402 Worklist.push_back(D->getLHS());
3403 Worklist.push_back(D->getRHS());
3404 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3405 const Value *V = US->getValue();
3406 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3407 // Look for instructions defined outside the loop.
3408 if (L->contains(Inst)) continue;
3409 } else if (isa<UndefValue>(V))
3410 // Undef doesn't have a live range, so it doesn't matter.
3411 continue;
3412 for (const Use &U : V->uses()) {
3413 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3414 // Ignore non-instructions.
3415 if (!UserInst)
3416 continue;
3417 // Ignore instructions in other functions (as can happen with
3418 // Constants).
3419 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3420 continue;
3421 // Ignore instructions not dominated by the loop.
3422 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3423 UserInst->getParent() :
3424 cast<PHINode>(UserInst)->getIncomingBlock(
3425 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3426 if (!DT.dominates(L->getHeader(), UseBB))
3427 continue;
3428 // Don't bother if the instruction is in a BB which ends in an EHPad.
3429 if (UseBB->getTerminator()->isEHPad())
3430 continue;
3431 // Don't bother rewriting PHIs in catchswitch blocks.
3432 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3433 continue;
3434 // Ignore uses which are part of other SCEV expressions, to avoid
3435 // analyzing them multiple times.
3436 if (SE.isSCEVable(UserInst->getType())) {
3437 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3438 // If the user is a no-op, look through to its uses.
3439 if (!isa<SCEVUnknown>(UserS))
3440 continue;
3441 if (UserS == US) {
3442 Worklist.push_back(
3443 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3444 continue;
3445 }
3446 }
3447 // Ignore icmp instructions which are already being analyzed.
3448 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3449 unsigned OtherIdx = !U.getOperandNo();
3450 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3451 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3452 continue;
3453 }
3454
3455 std::pair<size_t, int64_t> P = getUse(
3456 S, LSRUse::Basic, MemAccessTy());
3457 size_t LUIdx = P.first;
3458 int64_t Offset = P.second;
3459 LSRUse &LU = Uses[LUIdx];
3460 LSRFixup &LF = LU.getNewFixup();
3461 LF.UserInst = const_cast<Instruction *>(UserInst);
3462 LF.OperandValToReplace = U;
3463 LF.Offset = Offset;
3464 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3465 if (!LU.WidestFixupType ||
3466 SE.getTypeSizeInBits(LU.WidestFixupType) <
3467 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3468 LU.WidestFixupType = LF.OperandValToReplace->getType();
3469 InsertSupplementalFormula(US, LU, LUIdx);
3470 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3471 break;
3472 }
3473 }
3474 }
3475}
3476
3477/// Split S into subexpressions which can be pulled out into separate
3478/// registers. If C is non-null, multiply each subexpression by C.
3479///
3480/// Return remainder expression after factoring the subexpressions captured by
3481/// Ops. If Ops is complete, return NULL.
3482static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3483 SmallVectorImpl<const SCEV *> &Ops,
3484 const Loop *L,
3485 ScalarEvolution &SE,
3486 unsigned Depth = 0) {
3487 // Arbitrarily cap recursion to protect compile time.
3488 if (Depth >= 3)
3489 return S;
3490
3491 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3492 // Break out add operands.
3493 for (const SCEV *S : Add->operands()) {
3494 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3495 if (Remainder)
3496 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3497 }
3498 return nullptr;
3499 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3500 // Split a non-zero base out of an addrec.
3501 if (AR->getStart()->isZero() || !AR->isAffine())
3502 return S;
3503
3504 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3505 C, Ops, L, SE, Depth+1);
3506 // Split the non-zero AddRec unless it is part of a nested recurrence that
3507 // does not pertain to this loop.
3508 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3509 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3510 Remainder = nullptr;
3511 }
3512 if (Remainder != AR->getStart()) {
3513 if (!Remainder)
3514 Remainder = SE.getConstant(AR->getType(), 0);
3515 return SE.getAddRecExpr(Remainder,
3516 AR->getStepRecurrence(SE),
3517 AR->getLoop(),
3518 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3519 SCEV::FlagAnyWrap);
3520 }
3521 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3522 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3523 if (Mul->getNumOperands() != 2)
3524 return S;
3525 if (const SCEVConstant *Op0 =
3526 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3527 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3528 const SCEV *Remainder =
3529 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3530 if (Remainder)
3531 Ops.push_back(SE.getMulExpr(C, Remainder));
3532 return nullptr;
3533 }
3534 }
3535 return S;
3536}
3537
3538/// Return true if the SCEV represents a value that may end up as a
3539/// post-increment operation.
3540static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3541 LSRUse &LU, const SCEV *S, const Loop *L,
3542 ScalarEvolution &SE) {
3543 if (LU.Kind != LSRUse::Address ||
3544 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3545 return false;
3546 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3547 if (!AR)
3548 return false;
3549 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3550 if (!isa<SCEVConstant>(LoopStep))
3551 return false;
3552 if (LU.AccessTy.getType()->getScalarSizeInBits() !=
3553 LoopStep->getType()->getScalarSizeInBits())
3554 return false;
3555 // Check if a post-indexed load/store can be used.
3556 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3557 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3558 const SCEV *LoopStart = AR->getStart();
3559 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3560 return true;
3561 }
3562 return false;
3563}
3564
3565/// Helper function for LSRInstance::GenerateReassociations.
3566void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3567 const Formula &Base,
3568 unsigned Depth, size_t Idx,
3569 bool IsScaledReg) {
3570 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3571 // Don't generate reassociations for the base register of a value that
3572 // may generate a post-increment operator. The reason is that the
3573 // reassociations cause extra base+register formula to be created,
3574 // and possibly chosen, but the post-increment is more efficient.
3575 if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3576 return;
3577 SmallVector<const SCEV *, 8> AddOps;
3578 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3579 if (Remainder)
3580 AddOps.push_back(Remainder);
3581
3582 if (AddOps.size() == 1)
3583 return;
3584
3585 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3586 JE = AddOps.end();
3587 J != JE; ++J) {
3588 // Loop-variant "unknown" values are uninteresting; we won't be able to
3589 // do anything meaningful with them.
3590 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3591 continue;
3592
3593 // Don't pull a constant into a register if the constant could be folded
3594 // into an immediate field.
3595 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3596 LU.AccessTy, *J, Base.getNumRegs() > 1))
3597 continue;
3598
3599 // Collect all operands except *J.
3600 SmallVector<const SCEV *, 8> InnerAddOps(
3601 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3602 InnerAddOps.append(std::next(J),
3603 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3604
3605 // Don't leave just a constant behind in a register if the constant could
3606 // be folded into an immediate field.
3607 if (InnerAddOps.size() == 1 &&
3608 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3609 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3610 continue;
3611
3612 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3613 if (InnerSum->isZero())
3614 continue;
3615 Formula F = Base;
3616
3617 // Add the remaining pieces of the add back into the new formula.
3618 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3619 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3620 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3621 InnerSumSC->getValue()->getZExtValue())) {
3622 F.UnfoldedOffset =
3623 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3624 if (IsScaledReg)
3625 F.ScaledReg = nullptr;
3626 else
3627 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3628 } else if (IsScaledReg)
3629 F.ScaledReg = InnerSum;
3630 else
3631 F.BaseRegs[Idx] = InnerSum;
3632
3633 // Add J as its own register, or an unfolded immediate.
3634 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3635 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3636 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3637 SC->getValue()->getZExtValue()))
3638 F.UnfoldedOffset =
3639 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3640 else
3641 F.BaseRegs.push_back(*J);
3642 // We may have changed the number of register in base regs, adjust the
3643 // formula accordingly.
3644 F.canonicalize(*L);
3645
3646 if (InsertFormula(LU, LUIdx, F))
3647 // If that formula hadn't been seen before, recurse to find more like
3648 // it.
3649 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3650 // Because just Depth is not enough to bound compile time.
3651 // This means that every time AddOps.size() is greater 16^x we will add
3652 // x to Depth.
3653 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3654 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3655 }
3656}
3657
3658/// Split out subexpressions from adds and the bases of addrecs.
3659void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3660 Formula Base, unsigned Depth) {
3661 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3661, __PRETTY_FUNCTION__))
;
3662 // Arbitrarily cap recursion to protect compile time.
3663 if (Depth >= 3)
3664 return;
3665
3666 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3667 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3668
3669 if (Base.Scale == 1)
3670 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3671 /* Idx */ -1, /* IsScaledReg */ true);
3672}
3673
3674/// Generate a formula consisting of all of the loop-dominating registers added
3675/// into a single register.
3676void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3677 Formula Base) {
3678 // This method is only interesting on a plurality of registers.
3679 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3680 (Base.UnfoldedOffset != 0) <= 1)
3681 return;
3682
3683 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3684 // processing the formula.
3685 Base.unscale();
3686 SmallVector<const SCEV *, 4> Ops;
3687 Formula NewBase = Base;
3688 NewBase.BaseRegs.clear();
3689 Type *CombinedIntegerType = nullptr;
3690 for (const SCEV *BaseReg : Base.BaseRegs) {
3691 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3692 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3693 if (!CombinedIntegerType)
3694 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3695 Ops.push_back(BaseReg);
3696 }
3697 else
3698 NewBase.BaseRegs.push_back(BaseReg);
3699 }
3700
3701 // If no register is relevant, we're done.
3702 if (Ops.size() == 0)
3703 return;
3704
3705 // Utility function for generating the required variants of the combined
3706 // registers.
3707 auto GenerateFormula = [&](const SCEV *Sum) {
3708 Formula F = NewBase;
3709
3710 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3711 // opportunity to fold something. For now, just ignore such cases
3712 // rather than proceed with zero in a register.
3713 if (Sum->isZero())
3714 return;
3715
3716 F.BaseRegs.push_back(Sum);
3717 F.canonicalize(*L);
3718 (void)InsertFormula(LU, LUIdx, F);
3719 };
3720
3721 // If we collected at least two registers, generate a formula combining them.
3722 if (Ops.size() > 1) {
3723 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3724 GenerateFormula(SE.getAddExpr(OpsCopy));
3725 }
3726
3727 // If we have an unfolded offset, generate a formula combining it with the
3728 // registers collected.
3729 if (NewBase.UnfoldedOffset) {
3730 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3730, __PRETTY_FUNCTION__))
;
3731 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3732 true));
3733 NewBase.UnfoldedOffset = 0;
3734 GenerateFormula(SE.getAddExpr(Ops));
3735 }
3736}
3737
3738/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3739void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3740 const Formula &Base, size_t Idx,
3741 bool IsScaledReg) {
3742 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3743 GlobalValue *GV = ExtractSymbol(G, SE);
3744 if (G->isZero() || !GV)
3745 return;
3746 Formula F = Base;
3747 F.BaseGV = GV;
3748 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3749 return;
3750 if (IsScaledReg)
3751 F.ScaledReg = G;
3752 else
3753 F.BaseRegs[Idx] = G;
3754 (void)InsertFormula(LU, LUIdx, F);
3755}
3756
3757/// Generate reuse formulae using symbolic offsets.
3758void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3759 Formula Base) {
3760 // We can't add a symbolic offset if the address already contains one.
3761 if (Base.BaseGV) return;
3762
3763 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3764 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3765 if (Base.Scale == 1)
3766 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3767 /* IsScaledReg */ true);
3768}
3769
3770/// Helper function for LSRInstance::GenerateConstantOffsets.
3771void LSRInstance::GenerateConstantOffsetsImpl(
3772 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3773 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3774
3775 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3776 Formula F = Base;
3777 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3778
3779 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3780 LU.AccessTy, F)) {
3781 // Add the offset to the base register.
3782 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3783 // If it cancelled out, drop the base register, otherwise update it.
3784 if (NewG->isZero()) {
3785 if (IsScaledReg) {
3786 F.Scale = 0;
3787 F.ScaledReg = nullptr;
3788 } else
3789 F.deleteBaseReg(F.BaseRegs[Idx]);
3790 F.canonicalize(*L);
3791 } else if (IsScaledReg)
3792 F.ScaledReg = NewG;
3793 else
3794 F.BaseRegs[Idx] = NewG;
3795
3796 (void)InsertFormula(LU, LUIdx, F);
3797 }
3798 };
3799
3800 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3801
3802 // With constant offsets and constant steps, we can generate pre-inc
3803 // accesses by having the offset equal the step. So, for access #0 with a
3804 // step of 8, we generate a G - 8 base which would require the first access
3805 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3806 // for itself and hopefully becomes the base for other accesses. This means
3807 // means that a single pre-indexed access can be generated to become the new
3808 // base pointer for each iteration of the loop, resulting in no extra add/sub
3809 // instructions for pointer updating.
3810 if (FavorBackedgeIndex && LU.Kind == LSRUse::Address) {
3811 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3812 if (auto *StepRec =
3813 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3814 const APInt &StepInt = StepRec->getAPInt();
3815 int64_t Step = StepInt.isNegative() ?
3816 StepInt.getSExtValue() : StepInt.getZExtValue();
3817
3818 for (int64_t Offset : Worklist) {
3819 Offset -= Step;
3820 GenerateOffset(G, Offset);
3821 }
3822 }
3823 }
3824 }
3825 for (int64_t Offset : Worklist)
3826 GenerateOffset(G, Offset);
3827
3828 int64_t Imm = ExtractImmediate(G, SE);
3829 if (G->isZero() || Imm == 0)
3830 return;
3831 Formula F = Base;
3832 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3833 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3834 return;
3835 if (IsScaledReg)
3836 F.ScaledReg = G;
3837 else
3838 F.BaseRegs[Idx] = G;
3839 (void)InsertFormula(LU, LUIdx, F);
3840}
3841
3842/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3843void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3844 Formula Base) {
3845 // TODO: For now, just add the min and max offset, because it usually isn't
3846 // worthwhile looking at everything inbetween.
3847 SmallVector<int64_t, 2> Worklist;
3848 Worklist.push_back(LU.MinOffset);
3849 if (LU.MaxOffset != LU.MinOffset)
3850 Worklist.push_back(LU.MaxOffset);
3851
3852 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3853 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3854 if (Base.Scale == 1)
3855 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3856 /* IsScaledReg */ true);
3857}
3858
3859/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3860/// == y -> x*c == y*c.
3861void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3862 Formula Base) {
3863 if (LU.Kind != LSRUse::ICmpZero) return;
3864
3865 // Determine the integer type for the base formula.
3866 Type *IntTy = Base.getType();
3867 if (!IntTy) return;
3868 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3869
3870 // Don't do this if there is more than one offset.
3871 if (LU.MinOffset != LU.MaxOffset) return;
3872
3873 // Check if transformation is valid. It is illegal to multiply pointer.
3874 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3875 return;
3876 for (const SCEV *BaseReg : Base.BaseRegs)
3877 if (BaseReg->getType()->isPointerTy())
3878 return;
3879 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3879, __PRETTY_FUNCTION__))
;
3880
3881 // Check each interesting stride.
3882 for (int64_t Factor : Factors) {
3883 // Check that the multiplication doesn't overflow.
3884 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3885 continue;
3886 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3887 if (NewBaseOffset / Factor != Base.BaseOffset)
3888 continue;
3889 // If the offset will be truncated at this use, check that it is in bounds.
3890 if (!IntTy->isPointerTy() &&
3891 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3892 continue;
3893
3894 // Check that multiplying with the use offset doesn't overflow.
3895 int64_t Offset = LU.MinOffset;
3896 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3897 continue;
3898 Offset = (uint64_t)Offset * Factor;
3899 if (Offset / Factor != LU.MinOffset)
3900 continue;
3901 // If the offset will be truncated at this use, check that it is in bounds.
3902 if (!IntTy->isPointerTy() &&
3903 !ConstantInt::isValueValidForType(IntTy, Offset))
3904 continue;
3905
3906 Formula F = Base;
3907 F.BaseOffset = NewBaseOffset;
3908
3909 // Check that this scale is legal.
3910 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3911 continue;
3912
3913 // Compensate for the use having MinOffset built into it.
3914 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3915
3916 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3917
3918 // Check that multiplying with each base register doesn't overflow.
3919 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3920 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3921 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3922 goto next;
3923 }
3924
3925 // Check that multiplying with the scaled register doesn't overflow.
3926 if (F.ScaledReg) {
3927 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3928 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3929 continue;
3930 }
3931
3932 // Check that multiplying with the unfolded offset doesn't overflow.
3933 if (F.UnfoldedOffset != 0) {
3934 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3935 Factor == -1)
3936 continue;
3937 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3938 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3939 continue;
3940 // If the offset will be truncated, check that it is in bounds.
3941 if (!IntTy->isPointerTy() &&
3942 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3943 continue;
3944 }
3945
3946 // If we make it here and it's legal, add it.
3947 (void)InsertFormula(LU, LUIdx, F);
3948 next:;
3949 }
3950}
3951
3952/// Generate stride factor reuse formulae by making use of scaled-offset address
3953/// modes, for example.
3954void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3955 // Determine the integer type for the base formula.
3956 Type *IntTy = Base.getType();
3957 if (!IntTy) return;
3958
3959 // If this Formula already has a scaled register, we can't add another one.
3960 // Try to unscale the formula to generate a better scale.
3961 if (Base.Scale != 0 && !Base.unscale())
3962 return;
3963
3964 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3964, __PRETTY_FUNCTION__))
;
3965
3966 // Check each interesting stride.
3967 for (int64_t Factor : Factors) {
3968 Base.Scale = Factor;
3969 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3970 // Check whether this scale is going to be legal.
3971 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3972 Base)) {
3973 // As a special-case, handle special out-of-loop Basic users specially.
3974 // TODO: Reconsider this special case.
3975 if (LU.Kind == LSRUse::Basic &&
3976 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3977 LU.AccessTy, Base) &&
3978 LU.AllFixupsOutsideLoop)
3979 LU.Kind = LSRUse::Special;
3980 else
3981 continue;
3982 }
3983 // For an ICmpZero, negating a solitary base register won't lead to
3984 // new solutions.
3985 if (LU.Kind == LSRUse::ICmpZero &&
3986 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3987 continue;
3988 // For each addrec base reg, if its loop is current loop, apply the scale.
3989 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3990 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3991 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3992 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3993 if (FactorS->isZero())
3994 continue;
3995 // Divide out the factor, ignoring high bits, since we'll be
3996 // scaling the value back up in the end.
3997 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3998 // TODO: This could be optimized to avoid all the copying.
3999 Formula F = Base;
4000 F.ScaledReg = Quotient;
4001 F.deleteBaseReg(F.BaseRegs[i]);
4002 // The canonical representation of 1*reg is reg, which is already in
4003 // Base. In that case, do not try to insert the formula, it will be
4004 // rejected anyway.
4005 if (F.Scale == 1 && (F.BaseRegs.empty() ||
4006 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4007 continue;
4008 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4009 // non canonical Formula with ScaledReg's loop not being L.
4010 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4011 F.canonicalize(*L);
4012 (void)InsertFormula(LU, LUIdx, F);
4013 }
4014 }
4015 }
4016 }
4017}
4018
4019/// Generate reuse formulae from different IV types.
4020void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4021 // Don't bother truncating symbolic values.
4022 if (Base.BaseGV) return;
4023
4024 // Determine the integer type for the base formula.
4025 Type *DstTy = Base.getType();
4026 if (!DstTy) return;
4027 DstTy = SE.getEffectiveSCEVType(DstTy);
4028
4029 for (Type *SrcTy : Types) {
4030 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4031 Formula F = Base;
4032
4033 // Sometimes SCEV is able to prove zero during ext transform. It may
4034 // happen if SCEV did not do all possible transforms while creating the
4035 // initial node (maybe due to depth limitations), but it can do them while
4036 // taking ext.
4037 if (F.ScaledReg) {
4038 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4039 if (NewScaledReg->isZero())
4040 continue;
4041 F.ScaledReg = NewScaledReg;
4042 }
4043 bool HasZeroBaseReg = false;
4044 for (const SCEV *&BaseReg : F.BaseRegs) {
4045 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4046 if (NewBaseReg->isZero()) {
4047 HasZeroBaseReg = true;
4048 break;
4049 }
4050 BaseReg = NewBaseReg;
4051 }
4052 if (HasZeroBaseReg)
4053 continue;
4054
4055 // TODO: This assumes we've done basic processing on all uses and
4056 // have an idea what the register usage is.
4057 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4058 continue;
4059
4060 F.canonicalize(*L);
4061 (void)InsertFormula(LU, LUIdx, F);
4062 }
4063 }
4064}
4065
4066namespace {
4067
4068/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4069/// modifications so that the search phase doesn't have to worry about the data
4070/// structures moving underneath it.
4071struct WorkItem {
4072 size_t LUIdx;
4073 int64_t Imm;
4074 const SCEV *OrigReg;
4075
4076 WorkItem(size_t LI, int64_t I, const SCEV *R)
4077 : LUIdx(LI), Imm(I), OrigReg(R) {}
4078
4079 void print(raw_ostream &OS) const;
4080 void dump() const;
4081};
4082
4083} // end anonymous namespace
4084
4085#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4086void WorkItem::print(raw_ostream &OS) const {
4087 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4088 << " , add offset " << Imm;
4089}
4090
4091LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4092 print(errs()); errs() << '\n';
4093}
4094#endif
4095
4096/// Look for registers which are a constant distance apart and try to form reuse
4097/// opportunities between them.
4098void LSRInstance::GenerateCrossUseConstantOffsets() {
4099 // Group the registers by their value without any added constant offset.
4100 using ImmMapTy = std::map<int64_t, const SCEV *>;
4101
4102 DenseMap<const SCEV *, ImmMapTy> Map;
4103 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4104 SmallVector<const SCEV *, 8> Sequence;
4105 for (const SCEV *Use : RegUses) {
4106 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4107 int64_t Imm = ExtractImmediate(Reg, SE);
4108 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4109 if (Pair.second)
4110 Sequence.push_back(Reg);
4111 Pair.first->second.insert(std::make_pair(Imm, Use));
4112 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4113 }
4114
4115 // Now examine each set of registers with the same base value. Build up
4116 // a list of work to do and do the work in a separate step so that we're
4117 // not adding formulae and register counts while we're searching.
4118 SmallVector<WorkItem, 32> WorkItems;
4119 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4120 for (const SCEV *Reg : Sequence) {
4121 const ImmMapTy &Imms = Map.find(Reg)->second;
4122
4123 // It's not worthwhile looking for reuse if there's only one offset.
4124 if (Imms.size() == 1)
4125 continue;
4126
4127 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)
4128 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)
4129 : 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)
4130 << ' ' << 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)
4131 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)
;
4132
4133 // Examine each offset.
4134 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4135 J != JE; ++J) {
4136 const SCEV *OrigReg = J->second;
4137
4138 int64_t JImm = J->first;
4139 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4140
4141 if (!isa<SCEVConstant>(OrigReg) &&
4142 UsedByIndicesMap[Reg].count() == 1) {
4143 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)
4144 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4145 continue;
4146 }
4147
4148 // Conservatively examine offsets between this orig reg a few selected
4149 // other orig regs.
4150 int64_t First = Imms.begin()->first;
4151 int64_t Last = std::prev(Imms.end())->first;
4152 // Compute (First + Last) / 2 without overflow using the fact that
4153 // First + Last = 2 * (First + Last) + (First ^ Last).
4154 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4155 // If the result is negative and First is odd and Last even (or vice versa),
4156 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4157 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4158 ImmMapTy::const_iterator OtherImms[] = {
4159 Imms.begin(), std::prev(Imms.end()),
4160 Imms.lower_bound(Avg)};
4161 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4162 ImmMapTy::const_iterator M = OtherImms[i];
4163 if (M == J || M == JE) continue;
4164
4165 // Compute the difference between the two.
4166 int64_t Imm = (uint64_t)JImm - M->first;
4167 for (unsigned LUIdx : UsedByIndices.set_bits())
4168 // Make a memo of this use, offset, and register tuple.
4169 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4170 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4171 }
4172 }
4173 }
4174
4175 Map.clear();
4176 Sequence.clear();
4177 UsedByIndicesMap.clear();
4178 UniqueItems.clear();
4179
4180 // Now iterate through the worklist and add new formulae.
4181 for (const WorkItem &WI : WorkItems) {
4182 size_t LUIdx = WI.LUIdx;
4183 LSRUse &LU = Uses[LUIdx];
4184 int64_t Imm = WI.Imm;
4185 const SCEV *OrigReg = WI.OrigReg;
4186
4187 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4188 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4189 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4190
4191 // TODO: Use a more targeted data structure.
4192 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4193 Formula F = LU.Formulae[L];
4194 // FIXME: The code for the scaled and unscaled registers looks
4195 // very similar but slightly different. Investigate if they
4196 // could be merged. That way, we would not have to unscale the
4197 // Formula.
4198 F.unscale();
4199 // Use the immediate in the scaled register.
4200 if (F.ScaledReg == OrigReg) {
4201 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4202 // Don't create 50 + reg(-50).
4203 if (F.referencesReg(SE.getSCEV(
4204 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4205 continue;
4206 Formula NewF = F;
4207 NewF.BaseOffset = Offset;
4208 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4209 NewF))
4210 continue;
4211 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4212
4213 // If the new scale is a constant in a register, and adding the constant
4214 // value to the immediate would produce a value closer to zero than the
4215 // immediate itself, then the formula isn't worthwhile.
4216 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4217 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4218 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4219 .ule(std::abs(NewF.BaseOffset)))
4220 continue;
4221
4222 // OK, looks good.
4223 NewF.canonicalize(*this->L);
4224 (void)InsertFormula(LU, LUIdx, NewF);
4225 } else {
4226 // Use the immediate in a base register.
4227 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4228 const SCEV *BaseReg = F.BaseRegs[N];
4229 if (BaseReg != OrigReg)
4230 continue;
4231 Formula NewF = F;
4232 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4233 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4234 LU.Kind, LU.AccessTy, NewF)) {
4235 if (TTI.shouldFavorPostInc() &&
4236 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4237 continue;
4238 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4239 continue;
4240 NewF = F;
4241 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4242 }
4243 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4244
4245 // If the new formula has a constant in a register, and adding the
4246 // constant value to the immediate would produce a value closer to
4247 // zero than the immediate itself, then the formula isn't worthwhile.
4248 for (const SCEV *NewReg : NewF.BaseRegs)
4249 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4250 if ((C->getAPInt() + NewF.BaseOffset)
4251 .abs()
4252 .slt(std::abs(NewF.BaseOffset)) &&
4253 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4254 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4255 goto skip_formula;
4256
4257 // Ok, looks good.
4258 NewF.canonicalize(*this->L);
4259 (void)InsertFormula(LU, LUIdx, NewF);
4260 break;
4261 skip_formula:;
4262 }
4263 }
4264 }
4265 }
4266}
4267
4268/// Generate formulae for each use.
4269void
4270LSRInstance::GenerateAllReuseFormulae() {
4271 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4272 // queries are more precise.
4273 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4274 LSRUse &LU = Uses[LUIdx];
4275 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4276 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4277 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4278 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4279 }
4280 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4281 LSRUse &LU = Uses[LUIdx];
4282 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4283 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4284 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4285 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4286 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4287 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4288 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4289 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4290 }
4291 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4292 LSRUse &LU = Uses[LUIdx];
4293 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4294 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4295 }
4296
4297 GenerateCrossUseConstantOffsets();
4298
4299 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4300 "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)
4301 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4302}
4303
4304/// If there are multiple formulae with the same set of registers used
4305/// by other uses, pick the best one and delete the others.
4306void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4307 DenseSet<const SCEV *> VisitedRegs;
4308 SmallPtrSet<const SCEV *, 16> Regs;
4309 SmallPtrSet<const SCEV *, 16> LoserRegs;
4310#ifndef NDEBUG
4311 bool ChangedFormulae = false;
4312#endif
4313
4314 // Collect the best formula for each unique set of shared registers. This
4315 // is reset for each use.
4316 using BestFormulaeTy =
4317 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4318
4319 BestFormulaeTy BestFormulae;
4320
4321 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4322 LSRUse &LU = Uses[LUIdx];
4323 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)
4324 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4325
4326 bool Any = false;
4327 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4328 FIdx != NumForms; ++FIdx) {
4329 Formula &F = LU.Formulae[FIdx];
4330
4331 // Some formulas are instant losers. For example, they may depend on
4332 // nonexistent AddRecs from other loops. These need to be filtered
4333 // immediately, otherwise heuristics could choose them over others leading
4334 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4335 // avoids the need to recompute this information across formulae using the
4336 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4337 // the corresponding bad register from the Regs set.
4338 Cost CostF(L, SE, TTI);
4339 Regs.clear();
4340 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4341 if (CostF.isLoser()) {
4342 // During initial formula generation, undesirable formulae are generated
4343 // by uses within other loops that have some non-trivial address mode or
4344 // use the postinc form of the IV. LSR needs to provide these formulae
4345 // as the basis of rediscovering the desired formula that uses an AddRec
4346 // corresponding to the existing phi. Once all formulae have been
4347 // generated, these initial losers may be pruned.
4348 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)
4349 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4350 }
4351 else {
4352 SmallVector<const SCEV *, 4> Key;
4353 for (const SCEV *Reg : F.BaseRegs) {
4354 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4355 Key.push_back(Reg);
4356 }
4357 if (F.ScaledReg &&
4358 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4359 Key.push_back(F.ScaledReg);
4360 // Unstable sort by host order ok, because this is only used for
4361 // uniquifying.
4362 llvm::sort(Key);
4363
4364 std::pair<BestFormulaeTy::const_iterator, bool> P =
4365 BestFormulae.insert(std::make_pair(Key, FIdx));
4366 if (P.second)
4367 continue;
4368
4369 Formula &Best = LU.Formulae[P.first->second];
4370
4371 Cost CostBest(L, SE, TTI);
4372 Regs.clear();
4373 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4374 if (CostF.isLess(CostBest))
4375 std::swap(F, Best);
4376 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)
4377 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)
4378 " 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)
4379 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)
;
4380 }
4381#ifndef NDEBUG
4382 ChangedFormulae = true;
4383#endif
4384 LU.DeleteFormula(F);
4385 --FIdx;
4386 --NumForms;
4387 Any = true;
4388 }
4389
4390 // Now that we've filtered out some formulae, recompute the Regs set.
4391 if (Any)
4392 LU.RecomputeRegs(LUIdx, RegUses);
4393
4394 // Reset this to prepare for the next use.
4395 BestFormulae.clear();
4396 }
4397
4398 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)
4399 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)
4400 "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)
4401 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)
4402 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4403}
4404
4405/// Estimate the worst-case number of solutions the solver might have to
4406/// consider. It almost never considers this many solutions because it prune the
4407/// search space, but the pruning isn't always sufficient.
4408size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4409 size_t Power = 1;
4410 for (const LSRUse &LU : Uses) {
4411 size_t FSize = LU.Formulae.size();
4412 if (FSize >= ComplexityLimit) {
4413 Power = ComplexityLimit;
4414 break;
4415 }
4416 Power *= FSize;
4417 if (Power >= ComplexityLimit)
4418 break;
4419 }
4420 return Power;
4421}
4422
4423/// When one formula uses a superset of the registers of another formula, it
4424/// won't help reduce register pressure (though it may not necessarily hurt
4425/// register pressure); remove it to simplify the system.
4426void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4427 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4428 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)
;
4429
4430 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)
4431 "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)
4432 "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)
;
4433
4434 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4435 LSRUse &LU = Uses[LUIdx];
4436 bool Any = false;
4437 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4438 Formula &F = LU.Formulae[i];
4439 // Look for a formula with a constant or GV in a register. If the use
4440 // also has a formula with that same value in an immediate field,
4441 // delete the one that uses a register.
4442 for (SmallVectorImpl<const SCEV *>::const_iterator
4443 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4444 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4445 Formula NewF = F;
4446 //FIXME: Formulas should store bitwidth to do wrapping properly.
4447 // See PR41034.
4448 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4449 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4450 (I - F.BaseRegs.begin()));
4451 if (LU.HasFormulaWithSameRegs(NewF)) {
4452 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4453 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4454 LU.DeleteFormula(F);
4455 --i;
4456 --e;
4457 Any = true;
4458 break;
4459 }
4460 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4461 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4462 if (!F.BaseGV) {
4463 Formula NewF = F;
4464 NewF.BaseGV = GV;
4465 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4466 (I - F.BaseRegs.begin()));
4467 if (LU.HasFormulaWithSameRegs(NewF)) {
4468 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4469 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4470 LU.DeleteFormula(F);
4471 --i;
4472 --e;
4473 Any = true;
4474 break;
4475 }
4476 }
4477 }
4478 }
4479 }
4480 if (Any)
4481 LU.RecomputeRegs(LUIdx, RegUses);
4482 }
4483
4484 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)
;
4485 }
4486}
4487
4488/// When there are many registers for expressions like A, A+1, A+2, etc.,
4489/// allocate a single register for them.
4490void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4491 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4492 return;
4493
4494 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)
4495 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)
4496 "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)
4497 "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)
;
4498
4499 // This is especially useful for unrolled loops.
4500
4501 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4502 LSRUse &LU = Uses[LUIdx];
4503 for (const Formula &F : LU.Formulae) {
4504 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4505 continue;
4506
4507 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4508 if (!LUThatHas)
4509 continue;
4510
4511 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4512 LU.Kind, LU.AccessTy))
4513 continue;
4514
4515 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)
;
4516
4517 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4518
4519 // Transfer the fixups of LU to LUThatHas.
4520 for (LSRFixup &Fixup : LU.Fixups) {
4521 Fixup.Offset += F.BaseOffset;
4522 LUThatHas->pushFixup(Fixup);
4523 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)
;
4524 }
4525
4526 // Delete formulae from the new use which are no longer legal.
4527 bool Any = false;
4528 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4529 Formula &F = LUThatHas->Formulae[i];
4530 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4531 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4532 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)
;
4533 LUThatHas->DeleteFormula(F);
4534 --i;
4535 --e;
4536 Any = true;
4537 }
4538 }
4539
4540 if (Any)
4541 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4542
4543 // Delete the old use.
4544 DeleteUse(LU, LUIdx);
4545 --LUIdx;
4546 --NumUses;
4547 break;
4548 }
4549 }
4550
4551 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)
;
4552}
4553
4554/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4555/// we've done more filtering, as it may be able to find more formulae to
4556/// eliminate.
4557void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4558 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4559 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)
;
4560
4561 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)
4562 "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)
;
4563
4564 FilterOutUndesirableDedicatedRegisters();
4565
4566 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)
;
4567 }
4568}
4569
4570/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4571/// Pick the best one and delete the others.
4572/// This narrowing heuristic is to keep as many formulae with different
4573/// Scale and ScaledReg pair as possible while narrowing the search space.
4574/// The benefit is that it is more likely to find out a better solution
4575/// from a formulae set with more Scale and ScaledReg variations than
4576/// a formulae set with the same Scale and ScaledReg. The picking winner
4577/// reg heuristic will often keep the formulae with the same Scale and
4578/// ScaledReg and filter others, and we want to avoid that if possible.
4579void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4580 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4581 return;
4582
4583 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)
4584 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)
4585 "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)
4586 "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)
;
4587
4588 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4589 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4590
4591 BestFormulaeTy BestFormulae;
4592#ifndef NDEBUG
4593 bool ChangedFormulae = false;
4594#endif
4595 DenseSet<const SCEV *> VisitedRegs;
4596 SmallPtrSet<const SCEV *, 16> Regs;
4597
4598 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4599 LSRUse &LU = Uses[LUIdx];
4600 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)
4601 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4602
4603 // Return true if Formula FA is better than Formula FB.
4604 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4605 // First we will try to choose the Formula with fewer new registers.
4606 // For a register used by current Formula, the more the register is
4607 // shared among LSRUses, the less we increase the register number
4608 // counter of the formula.
4609 size_t FARegNum = 0;
4610 for (const SCEV *Reg : FA.BaseRegs) {
4611 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4612 FARegNum += (NumUses - UsedByIndices.count() + 1);
4613 }
4614 size_t FBRegNum = 0;
4615 for (const SCEV *Reg : FB.BaseRegs) {
4616 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4617 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4618 }
4619 if (FARegNum != FBRegNum)
4620 return FARegNum < FBRegNum;
4621
4622 // If the new register numbers are the same, choose the Formula with
4623 // less Cost.
4624 Cost CostFA(L, SE, TTI);
4625 Cost CostFB(L, SE, TTI);
4626 Regs.clear();
4627 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4628 Regs.clear();
4629 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4630 return CostFA.isLess(CostFB);
4631 };
4632
4633 bool Any = false;
4634 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4635 ++FIdx) {
4636 Formula &F = LU.Formulae[FIdx];
4637 if (!F.ScaledReg)
4638 continue;
4639 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4640 if (P.second)
4641 continue;
4642
4643 Formula &Best = LU.Formulae[P.first->second];
4644 if (IsBetterThan(F, Best))
4645 std::swap(F, Best);
4646 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)
4647 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)
4648 " 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)
4649 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)
;
4650#ifndef NDEBUG
4651 ChangedFormulae = true;
4652#endif
4653 LU.DeleteFormula(F);
4654 --FIdx;
4655 --NumForms;
4656 Any = true;
4657 }
4658 if (Any)
4659 LU.RecomputeRegs(LUIdx, RegUses);
4660
4661 // Reset this to prepare for the next use.
4662 BestFormulae.clear();
4663 }
4664
4665 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)
4666 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)
4667 "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)
4668 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)
4669 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4670}
4671
4672/// The function delete formulas with high registers number expectation.
4673/// Assuming we don't know the value of each formula (already delete
4674/// all inefficient), generate probability of not selecting for each
4675/// register.
4676/// For example,
4677/// Use1:
4678/// reg(a) + reg({0,+,1})
4679/// reg(a) + reg({-1,+,1}) + 1
4680/// reg({a,+,1})
4681/// Use2:
4682/// reg(b) + reg({0,+,1})
4683/// reg(b) + reg({-1,+,1}) + 1
4684/// reg({b,+,1})
4685/// Use3:
4686/// reg(c) + reg(b) + reg({0,+,1})
4687/// reg(c) + reg({b,+,1})
4688///
4689/// Probability of not selecting
4690/// Use1 Use2 Use3
4691/// reg(a) (1/3) * 1 * 1
4692/// reg(b) 1 * (1/3) * (1/2)
4693/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4694/// reg({-1,+,1}) (2/3) * (2/3) * 1
4695/// reg({a,+,1}) (2/3) * 1 * 1
4696/// reg({b,+,1}) 1 * (2/3) * (2/3)
4697/// reg(c) 1 * 1 * 0
4698///
4699/// Now count registers number mathematical expectation for each formula:
4700/// Note that for each use we exclude probability if not selecting for the use.
4701/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4702/// probabilty 1/3 of not selecting for Use1).
4703/// Use1:
4704/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4705/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4706/// reg({a,+,1}) 1
4707/// Use2:
4708/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4709/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4710/// reg({b,+,1}) 2/3
4711/// Use3:
4712/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4713/// reg(c) + reg({b,+,1}) 1 + 2/3
4714void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4715 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4716 return;
4717 // Ok, we have too many of formulae on our hands to conveniently handle.
4718 // Use a rough heuristic to thin out the list.
4719
4720 // Set of Regs wich will be 100% used in final solution.
4721 // Used in each formula of a solution (in example above this is reg(c)).
4722 // We can skip them in calculations.
4723 SmallPtrSet<const SCEV *, 4> UniqRegs;
4724 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)
;
4725
4726 // Map each register to probability of not selecting
4727 DenseMap <const SCEV *, float> RegNumMap;
4728 for (const SCEV *Reg : RegUses) {
4729 if (UniqRegs.count(Reg))
4730 continue;
4731 float PNotSel = 1;
4732 for (const LSRUse &LU : Uses) {
4733 if (!LU.Regs.count(Reg))
4734 continue;
4735 float P = LU.getNotSelectedProbability(Reg);
4736 if (P != 0.0)
4737 PNotSel *= P;
4738 else
4739 UniqRegs.insert(Reg);
4740 }
4741 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4742 }
4743
4744 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4745 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)
;
4746
4747 // Delete formulas where registers number expectation is high.
4748 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4749 LSRUse &LU = Uses[LUIdx];
4750 // If nothing to delete - continue.
4751 if (LU.Formulae.size() < 2)
4752 continue;
4753 // This is temporary solution to test performance. Float should be
4754 // replaced with round independent type (based on integers) to avoid
4755 // different results for different target builds.
4756 float FMinRegNum = LU.Formulae[0].getNumRegs();
4757 float FMinARegNum = LU.Formulae[0].getNumRegs();
4758 size_t MinIdx = 0;
4759 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4760 Formula &F = LU.Formulae[i];
4761 float FRegNum = 0;
4762 float FARegNum = 0;
4763 for (const SCEV *BaseReg : F.BaseRegs) {
4764 if (UniqRegs.count(BaseReg))
4765 continue;
4766 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4767 if (isa<SCEVAddRecExpr>(BaseReg))
4768 FARegNum +=
4769 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4770 }
4771 if (const SCEV *ScaledReg = F.ScaledReg) {
4772 if (!UniqRegs.count(ScaledReg)) {
4773 FRegNum +=
4774 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4775 if (isa<SCEVAddRecExpr>(ScaledReg))
4776 FARegNum +=
4777 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4778 }
4779 }
4780 if (FMinRegNum > FRegNum ||
4781 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4782 FMinRegNum = FRegNum;
4783 FMinARegNum = FARegNum;
4784 MinIdx = i;
4785 }
4786 }
4787 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)
4788 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)
;
4789 if (MinIdx != 0)
4790 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4791 while (LU.Formulae.size() != 1) {
4792 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
)
4793 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4794 LU.Formulae.pop_back();
4795 }
4796 LU.RecomputeRegs(LUIdx, RegUses);
4797 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4797, __PRETTY_FUNCTION__))
;
4798 Formula &F = LU.Formulae[0];
4799 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)
;
4800 // When we choose the formula, the regs become unique.
4801 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4802 if (F.ScaledReg)
4803 UniqRegs.insert(F.ScaledReg);
4804 }
4805 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)
;
4806}
4807
4808/// Pick a register which seems likely to be profitable, and then in any use
4809/// which has any reference to that register, delete all formulae which do not
4810/// reference that register.
4811void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4812 // With all other options exhausted, loop until the system is simple
4813 // enough to handle.
4814 SmallPtrSet<const SCEV *, 4> Taken;
4815 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
24
Assuming the condition is true
25
Loop condition is true. Entering loop body
4816 // Ok, we have too many of formulae on our hands to conveniently handle.
4817 // Use a rough heuristic to thin out the list.
4818 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)
;
26
Assuming 'DebugFlag' is false
27
Loop condition is false. Exiting loop
4819
4820 // Pick the register which is used by the most LSRUses, which is likely
4821 // to be a good reuse register candidate.
4822 const SCEV *Best = nullptr;
28
'Best' initialized to a null pointer value
4823 unsigned BestNum = 0;
4824 for (const SCEV *Reg : RegUses) {
29
Calling 'RegUseTracker::begin'
31
Returning from 'RegUseTracker::begin'
32
Assuming '__begin2' is equal to '__end2'
4825 if (Taken.count(Reg))
4826 continue;
4827 if (!Best) {
4828 Best = Reg;
4829 BestNum = RegUses.getUsedByIndices(Reg).count();
4830 } else {
4831 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4832 if (Count > BestNum) {
4833 Best = Reg;
4834 BestNum = Count;
4835 }
4836 }
4837 }
4838
4839 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)
33
Assuming 'DebugFlag' is true
34
Assuming the condition is true
35
Taking true branch
36
Forming reference to null pointer
4840 << " 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)
;
4841 Taken.insert(Best);
4842
4843 // In any use with formulae which references this register, delete formulae
4844 // which don't reference it.
4845 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4846 LSRUse &LU = Uses[LUIdx];
4847 if (!LU.Regs.count(Best)) continue;
4848
4849 bool Any = false;
4850 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4851 Formula &F = LU.Formulae[i];
4852 if (!F.referencesReg(Best)) {
4853 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)
;
4854 LU.DeleteFormula(F);
4855 --e;
4856 --i;
4857 Any = true;
4858 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4858, __PRETTY_FUNCTION__))
;
4859 continue;
4860 }
4861 }
4862
4863 if (Any)
4864 LU.RecomputeRegs(LUIdx, RegUses);
4865 }
4866
4867 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)
;
4868 }
4869}
4870
4871/// If there are an extraordinary number of formulae to choose from, use some
4872/// rough heuristics to prune down the number of formulae. This keeps the main
4873/// solver from taking an extraordinary amount of time in some worst-case
4874/// scenarios.
4875void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4876 NarrowSearchSpaceByDetectingSupersets();
4877 NarrowSearchSpaceByCollapsingUnrolledCode();
4878 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4879 if (FilterSameScaledReg)
19
Assuming the condition is false
20
Taking false branch
4880 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4881 if (LSRExpNarrow)
21
Assuming the condition is false
22
Taking false branch
4882 NarrowSearchSpaceByDeletingCostlyFormulas();
4883 else
4884 NarrowSearchSpaceByPickingWinnerRegs();
23
Calling 'LSRInstance::NarrowSearchSpaceByPickingWinnerRegs'
4885}
4886
4887/// This is the recursive solver.
4888void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4889 Cost &SolutionCost,
4890 SmallVectorImpl<const Formula *> &Workspace,
4891 const Cost &CurCost,
4892 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4893 DenseSet<const SCEV *> &VisitedRegs) const {
4894 // Some ideas:
4895 // - prune more:
4896 // - use more aggressive filtering
4897 // - sort the formula so that the most profitable solutions are found first
4898 // - sort the uses too
4899 // - search faster:
4900 // - don't compute a cost, and then compare. compare while computing a cost
4901 // and bail early.
4902 // - track register sets with SmallBitVector
4903
4904 const LSRUse &LU = Uses[Workspace.size()];
4905
4906 // If this use references any register that's already a part of the
4907 // in-progress solution, consider it a requirement that a formula must
4908 // reference that register in order to be considered. This prunes out
4909 // unprofitable searching.
4910 SmallSetVector<const SCEV *, 4> ReqRegs;
4911 for (const SCEV *S : CurRegs)
4912 if (LU.Regs.count(S))
4913 ReqRegs.insert(S);
4914
4915 SmallPtrSet<const SCEV *, 16> NewRegs;
4916 Cost NewCost(L, SE, TTI);
4917 for (const Formula &F : LU.Formulae) {
4918 // Ignore formulae which may not be ideal in terms of register reuse of
4919 // ReqRegs. The formula should use all required registers before
4920 // introducing new ones.
4921 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4922 for (const SCEV *Reg : ReqRegs) {
4923 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4924 is_contained(F.BaseRegs, Reg)) {
4925 --NumReqRegsToFind;
4926 if (NumReqRegsToFind == 0)
4927 break;
4928 }
4929 }
4930 if (NumReqRegsToFind != 0) {
4931 // If none of the formulae satisfied the required registers, then we could
4932 // clear ReqRegs and try again. Currently, we simply give up in this case.
4933 continue;
4934 }
4935
4936 // Evaluate the cost of the current formula. If it's already worse than
4937 // the current best, prune the search at that point.
4938 NewCost = CurCost;
4939 NewRegs = CurRegs;
4940 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
4941 if (NewCost.isLess(SolutionCost)) {
4942 Workspace.push_back(&F);
4943 if (Workspace.size() != Uses.size()) {
4944 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4945 NewRegs, VisitedRegs);
4946 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4947 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4948 } else {
4949 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() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
'\n'; } } while (false)
4950 dbgs() << ".\nRegs:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
'\n'; } } while (false)
4951 for (const SCEV *S : NewRegs) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
'\n'; } } while (false)
4952 << "- " << *S << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
'\n'; } } while (false)
4953 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
'\n'; } } while (false)
;
4954
4955 SolutionCost = NewCost;
4956 Solution = Workspace;
4957 }
4958 Workspace.pop_back();
4959 }
4960 }
4961}
4962
4963/// Choose one formula from each use. Return the results in the given Solution
4964/// vector.
4965void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4966 SmallVector<const Formula *, 8> Workspace;
4967 Cost SolutionCost(L, SE, TTI);
4968 SolutionCost.Lose();
4969 Cost CurCost(L, SE, TTI);
4970 SmallPtrSet<const SCEV *, 16> CurRegs;
4971 DenseSet<const SCEV *> VisitedRegs;
4972 Workspace.reserve(Uses.size());
4973
4974 // SolveRecurse does all the work.
4975 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4976 CurRegs, VisitedRegs);
4977 if (Solution.empty()) {
4978 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
4979 return;
4980 }
4981
4982 // Ok, we've now made all our decisions.
4983 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
)
4984 "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
)
4985 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
)
4986 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
)
4987 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
)
4988 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
)
4989 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
)
4990 " ";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
)
4991 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
)
4992 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
)
4993 })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
)
;
4994
4995 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4995, __PRETTY_FUNCTION__))
;
4996}
4997
4998/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4999/// we can go while still being dominated by the input positions. This helps
5000/// canonicalize the insert position, which encourages sharing.
5001BasicBlock::iterator
5002LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5003 const SmallVectorImpl<Instruction *> &Inputs)
5004 const {
5005 Instruction *Tentative = &*IP;
5006 while (true) {
5007 bool AllDominate = true;
5008 Instruction *BetterPos = nullptr;
5009 // Don't bother attempting to insert before a catchswitch, their basic block
5010 // cannot have other non-PHI instructions.
5011 if (isa<CatchSwitchInst>(Tentative))
5012 return IP;
5013
5014 for (Instruction *Inst : Inputs) {
5015 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5016 AllDominate = false;
5017 break;
5018 }
5019 // Attempt to find an insert position in the middle of the block,
5020 // instead of at the end, so that it can be used for other expansions.
5021 if (Tentative->getParent() == Inst->getParent() &&
5022 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5023 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5024 }
5025 if (!AllDominate)
5026 break;
5027 if (BetterPos)
5028 IP = BetterPos->getIterator();
5029 else
5030 IP = Tentative->getIterator();
5031
5032 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5033 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5034
5035 BasicBlock *IDom;
5036 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5037 if (!Rung) return IP;
5038 Rung = Rung->getIDom();
5039 if (!Rung) return IP;
5040 IDom = Rung->getBlock();
5041
5042 // Don't climb into a loop though.
5043 const Loop *IDomLoop = LI.getLoopFor(IDom);
5044 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5045 if (IDomDepth <= IPLoopDepth &&
5046 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5047 break;
5048 }
5049
5050 Tentative = IDom->getTerminator();
5051 }
5052
5053 return IP;
5054}
5055
5056/// Determine an input position which will be dominated by the operands and
5057/// which will dominate the result.
5058BasicBlock::iterator
5059LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5060 const LSRFixup &LF,
5061 const LSRUse &LU,
5062 SCEVExpander &Rewriter) const {
5063 // Collect some instructions which must be dominated by the
5064 // expanding replacement. These must be dominated by any operands that
5065 // will be required in the expansion.
5066 SmallVector<Instruction *, 4> Inputs;
5067 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5068 Inputs.push_back(I);
5069 if (LU.Kind == LSRUse::ICmpZero)
5070 if (Instruction *I =
5071 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5072 Inputs.push_back(I);
5073 if (LF.PostIncLoops.count(L)) {
5074 if (LF.isUseFullyOutsideLoop(L))
5075 Inputs.push_back(L->getLoopLatch()->getTerminator());
5076 else
5077 Inputs.push_back(IVIncInsertPos);
5078 }
5079 // The expansion must also be dominated by the increment positions of any
5080 // loops it for which it is using post-inc mode.
5081 for (const Loop *PIL : LF.PostIncLoops) {
5082 if (PIL == L) continue;
5083
5084 // Be dominated by the loop exit.
5085 SmallVector<BasicBlock *, 4> ExitingBlocks;
5086 PIL->getExitingBlocks(ExitingBlocks);
5087 if (!ExitingBlocks.empty()) {
5088 BasicBlock *BB = ExitingBlocks[0];
5089 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5090 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5091 Inputs.push_back(BB->getTerminator());
5092 }
5093 }
5094
5095 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5097, __PRETTY_FUNCTION__))
5096 && !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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5097, __PRETTY_FUNCTION__))
5097 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5097, __PRETTY_FUNCTION__))
;
5098
5099 // Then, climb up the immediate dominator tree as far as we can go while
5100 // still being dominated by the input positions.
5101 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5102
5103 // Don't insert instructions before PHI nodes.
5104 while (isa<PHINode>(IP)) ++IP;
5105
5106 // Ignore landingpad instructions.
5107 while (IP->isEHPad()) ++IP;
5108
5109 // Ignore debug intrinsics.
5110 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5111
5112 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5113 // IP consistent across expansions and allows the previously inserted
5114 // instructions to be reused by subsequent expansion.
5115 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5116 ++IP;
5117
5118 return IP;
5119}
5120
5121/// Emit instructions for the leading candidate expression for this LSRUse (this
5122/// is called "expanding").
5123Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5124 const Formula &F, BasicBlock::iterator IP,
5125 SCEVExpander &Rewriter,
5126 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5127 if (LU.RigidFormula)
5128 return LF.OperandValToReplace;
5129
5130 // Determine an input position which will be dominated by the operands and
5131 // which will dominate the result.
5132 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5133 Rewriter.setInsertPoint(&*IP);
5134
5135 // Inform the Rewriter if we have a post-increment use, so that it can
5136 // perform an advantageous expansion.
5137 Rewriter.setPostInc(LF.PostIncLoops);
5138
5139 // This is the type that the user actually needs.
5140 Type *OpTy = LF.OperandValToReplace->getType();
5141 // This will be the type that we'll initially expand to.
5142 Type *Ty = F.getType();
5143 if (!Ty)
5144 // No type known; just expand directly to the ultimate type.
5145 Ty = OpTy;
5146 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5147 // Expand directly to the ultimate type if it's the right size.
5148 Ty = OpTy;
5149 // This is the type to do integer arithmetic in.
5150 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5151
5152 // Build up a list of operands to add together to form the full base.
5153 SmallVector<const SCEV *, 8> Ops;
5154
5155 // Expand the BaseRegs portion.
5156 for (const SCEV *Reg : F.BaseRegs) {
5157 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5157, __PRETTY_FUNCTION__))
;
5158
5159 // If we're expanding for a post-inc user, make the post-inc adjustment.
5160 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5161 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5162 }
5163
5164 // Expand the ScaledReg portion.
5165 Value *ICmpScaledV = nullptr;
5166 if (F.Scale != 0) {
5167 const SCEV *ScaledS = F.ScaledReg;
5168
5169 // If we're expanding for a post-inc user, make the post-inc adjustment.
5170 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5171 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5172
5173 if (LU.Kind == LSRUse::ICmpZero) {
5174 // Expand ScaleReg as if it was part of the base regs.
5175 if (F.Scale == 1)
5176 Ops.push_back(
5177 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5178 else {
5179 // An interesting way of "folding" with an icmp is to use a negated
5180 // scale, which we'll implement by inserting it into the other operand
5181 // of the icmp.
5182 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5183, __PRETTY_FUNCTION__))
5183 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5183, __PRETTY_FUNCTION__))
;
5184 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5185 }
5186 } else {
5187 // Otherwise just expand the scaled register and an explicit scale,
5188 // which is expected to be matched as part of the address.
5189
5190 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5191 // Unless the addressing mode will not be folded.
5192 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5193 isAMCompletelyFolded(TTI, LU, F)) {
5194 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5195 Ops.clear();
5196 Ops.push_back(SE.getUnknown(FullV));
5197 }
5198 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5199 if (F.Scale != 1)
5200 ScaledS =
5201 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5202 Ops.push_back(ScaledS);
5203 }
5204 }
5205
5206 // Expand the GV portion.
5207 if (F.BaseGV) {
5208 // Flush the operand list to suppress SCEVExpander hoisting.
5209 if (!Ops.empty()) {
5210 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5211 Ops.clear();
5212 Ops.push_back(SE.getUnknown(FullV));
5213 }
5214 Ops.push_back(SE.getUnknown(F.BaseGV));
5215 }
5216
5217 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5218 // unfolded offsets. LSR assumes they both live next to their uses.
5219 if (!Ops.empty()) {
5220 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5221 Ops.clear();
5222 Ops.push_back(SE.getUnknown(FullV));
5223 }
5224
5225 // Expand the immediate portion.
5226 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5227 if (Offset != 0) {
5228 if (LU.Kind == LSRUse::ICmpZero) {
5229 // The other interesting way of "folding" with an ICmpZero is to use a
5230 // negated immediate.
5231 if (!ICmpScaledV)
5232 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5233 else {
5234 Ops.push_back(SE.getUnknown(ICmpScaledV));
5235 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5236 }
5237 } else {
5238 // Just add the immediate values. These again are expected to be matched
5239 // as part of the address.
5240 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5241 }
5242 }
5243
5244 // Expand the unfolded offset portion.
5245 int64_t UnfoldedOffset = F.UnfoldedOffset;
5246 if (UnfoldedOffset != 0) {
5247 // Just add the immediate values.
5248 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5249 UnfoldedOffset)));
5250 }
5251
5252 // Emit instructions summing all the operands.
5253 const SCEV *FullS = Ops.empty() ?
5254 SE.getConstant(IntTy, 0) :
5255 SE.getAddExpr(Ops);
5256 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5257
5258 // We're done expanding now, so reset the rewriter.
5259 Rewriter.clearPostInc();
5260
5261 // An ICmpZero Formula represents an ICmp which we're handling as a
5262 // comparison against zero. Now that we've expanded an expression for that
5263 // form, update the ICmp's other operand.
5264 if (LU.Kind == LSRUse::ICmpZero) {
5265 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5266 DeadInsts.emplace_back(CI->getOperand(1));
5267 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5268, __PRETTY_FUNCTION__))
5268 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5268, __PRETTY_FUNCTION__))
;
5269 if (F.Scale == -1) {
5270 if (ICmpScaledV->getType() != OpTy) {
5271 Instruction *Cast =
5272 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5273 OpTy, false),
5274 ICmpScaledV, OpTy, "tmp", CI);
5275 ICmpScaledV = Cast;
5276 }
5277 CI->setOperand(1, ICmpScaledV);
5278 } else {
5279 // A scale of 1 means that the scale has been expanded as part of the
5280 // base regs.
5281 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5283, __PRETTY_FUNCTION__))
5282 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5283, __PRETTY_FUNCTION__))
5283 "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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5283, __PRETTY_FUNCTION__))
;
5284 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5285 -(uint64_t)Offset);
5286 if (C->getType() != OpTy)
5287 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5288 OpTy, false),
5289 C, OpTy);
5290
5291 CI->setOperand(1, C);
5292 }
5293 }
5294
5295 return FullV;
5296}
5297
5298/// Helper for Rewrite. PHI nodes are special because the use of their operands
5299/// effectively happens in their predecessor blocks, so the expression may need
5300/// to be expanded in multiple places.
5301void LSRInstance::RewriteForPHI(
5302 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5303 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5304 DenseMap<BasicBlock *, Value *> Inserted;
5305 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5306 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5307 bool needUpdateFixups = false;
5308 BasicBlock *BB = PN->getIncomingBlock(i);
5309
5310 // If this is a critical edge, split the edge so that we do not insert
5311 // the code on all predecessor/successor paths. We do this unless this
5312 // is the canonical backedge for this loop, which complicates post-inc
5313 // users.
5314 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5315 !isa<IndirectBrInst>(BB->getTerminator()) &&
5316 !isa<CatchSwitchInst>(BB->getTerminator())) {
5317 BasicBlock *Parent = PN->getParent();
5318 Loop *PNLoop = LI.getLoopFor(Parent);
5319 if (!PNLoop || Parent != PNLoop->getHeader()) {
5320 // Split the critical edge.
5321 BasicBlock *NewBB = nullptr;
5322 if (!Parent->isLandingPad()) {
5323 NewBB = SplitCriticalEdge(BB, Parent,
5324 CriticalEdgeSplittingOptions(&DT, &LI)
5325 .setMergeIdenticalEdges()
5326 .setKeepOneInputPHIs());
5327 } else {
5328 SmallVector<BasicBlock*, 2> NewBBs;
5329 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5330 NewBB = NewBBs[0];
5331 }
5332 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5333 // phi predecessors are identical. The simple thing to do is skip
5334 // splitting in this case rather than complicate the API.
5335 if (NewBB) {
5336 // If PN is outside of the loop and BB is in the loop, we want to
5337 // move the block to be immediately before the PHI block, not
5338 // immediately after BB.
5339 if (L->contains(BB) && !L->contains(PN))
5340 NewBB->moveBefore(PN->getParent());
5341
5342 // Splitting the edge can reduce the number of PHI entries we have.
5343 e = PN->getNumIncomingValues();
5344 BB = NewBB;
5345 i = PN->getBasicBlockIndex(BB);
5346
5347 needUpdateFixups = true;
5348 }
5349 }
5350 }
5351
5352 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5353 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5354 if (!Pair.second)
5355 PN->setIncomingValue(i, Pair.first->second);
5356 else {
5357 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5358 Rewriter, DeadInsts);
5359
5360 // If this is reuse-by-noop-cast, insert the noop cast.
5361 Type *OpTy = LF.OperandValToReplace->getType();
5362 if (FullV->getType() != OpTy)
5363 FullV =
5364 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5365 OpTy, false),
5366 FullV, LF.OperandValToReplace->getType(),
5367 "tmp", BB->getTerminator());
5368
5369 PN->setIncomingValue(i, FullV);
5370 Pair.first->second = FullV;
5371 }
5372
5373 // If LSR splits critical edge and phi node has other pending
5374 // fixup operands, we need to update those pending fixups. Otherwise
5375 // formulae will not be implemented completely and some instructions
5376 // will not be eliminated.
5377 if (needUpdateFixups) {
5378 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5379 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5380 // If fixup is supposed to rewrite some operand in the phi
5381 // that was just updated, it may be already moved to
5382 // another phi node. Such fixup requires update.
5383 if (Fixup.UserInst == PN) {
5384 // Check if the operand we try to replace still exists in the
5385 // original phi.
5386 bool foundInOriginalPHI = false;
5387 for (const auto &val : PN->incoming_values())
5388 if (val == Fixup.OperandValToReplace) {
5389 foundInOriginalPHI = true;
5390 break;
5391 }
5392
5393 // If fixup operand found in original PHI - nothing to do.
5394 if (foundInOriginalPHI)
5395 continue;
5396
5397 // Otherwise it might be moved to another PHI and requires update.
5398 // If fixup operand not found in any of the incoming blocks that
5399 // means we have already rewritten it - nothing to do.
5400 for (const auto &Block : PN->blocks())
5401 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5402 ++I) {
5403 PHINode *NewPN = cast<PHINode>(I);
5404 for (const auto &val : NewPN->incoming_values())
5405 if (val == Fixup.OperandValToReplace)
5406 Fixup.UserInst = NewPN;
5407 }
5408 }
5409 }
5410 }
5411}
5412
5413/// Emit instructions for the leading candidate expression for this LSRUse (this
5414/// is called "expanding"), and update the UserInst to reference the newly
5415/// expanded value.
5416void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5417 const Formula &F, SCEVExpander &Rewriter,
5418 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5419 // First, find an insertion point that dominates UserInst. For PHI nodes,
5420 // find the nearest block which dominates all the relevant uses.
5421 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5422 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5423 } else {
5424 Value *FullV =
5425 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5426
5427 // If this is reuse-by-noop-cast, insert the noop cast.
5428 Type *OpTy = LF.OperandValToReplace->getType();
5429 if (FullV->getType() != OpTy) {
5430 Instruction *Cast =
5431 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5432 FullV, OpTy, "tmp", LF.UserInst);
5433 FullV = Cast;
5434 }
5435
5436 // Update the user. ICmpZero is handled specially here (for now) because
5437 // Expand may have updated one of the operands of the icmp already, and
5438 // its new value may happen to be equal to LF.OperandValToReplace, in
5439 // which case doing replaceUsesOfWith leads to replacing both operands
5440 // with the same value. TODO: Reorganize this.
5441 if (LU.Kind == LSRUse::ICmpZero)
5442 LF.UserInst->setOperand(0, FullV);
5443 else
5444 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5445 }
5446
5447 DeadInsts.emplace_back(LF.OperandValToReplace);
5448}
5449
5450/// Rewrite all the fixup locations with new values, following the chosen
5451/// solution.
5452void LSRInstance::ImplementSolution(
5453 const SmallVectorImpl<const Formula *> &Solution) {
5454 // Keep track of instructions we may have made dead, so that
5455 // we can remove them after we are done working.
5456 SmallVector<WeakTrackingVH, 16> DeadInsts;
5457
5458 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5459 "lsr");
5460#ifndef NDEBUG
5461 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5462#endif
5463 Rewriter.disableCanonicalMode();
5464 Rewriter.enableLSRMode();
5465 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5466
5467 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5468 for (const IVChain &Chain : IVChainVec) {
5469 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5470 Rewriter.setChainedPhi(PN);
5471 }
5472
5473 // Expand the new value definitions and update the users.
5474 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5475 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5476 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5477 Changed = true;
5478 }
5479
5480 for (const IVChain &Chain : IVChainVec) {
5481 GenerateIVChain(Chain, Rewriter, DeadInsts);
5482 Changed = true;
5483 }
5484 // Clean up after ourselves. This must be done before deleting any
5485 // instructions.
5486 Rewriter.clear();
5487
5488 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5489}
5490
5491LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5492 DominatorTree &DT, LoopInfo &LI,
5493 const TargetTransformInfo &TTI, AssumptionCache &AC,
5494 TargetLibraryInfo &LibInfo)
5495 : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), LibInfo(LibInfo), TTI(TTI), L(L),
5496 FavorBackedgeIndex(EnableBackedgeIndexing &&
3
Assuming the condition is false
5497 TTI.shouldFavorBackedgeIndex(L)) {
5498 // If LoopSimplify form is not available, stay out of trouble.
5499 if (!L->isLoopSimplifyForm())
4
Assuming the condition is false
5
Taking false branch
5500 return;
5501
5502 // If there's no interesting work to be done, bail early.
5503 if (IU.empty()) return;
6
Assuming the condition is false
7
Taking false branch
5504
5505 // If there's too much analysis to be done, bail early. We won't be able to
5506 // model the problem anyway.
5507 unsigned NumUsers = 0;
5508 for (const IVStrideUse &U : IU) {
5509 if (++NumUsers > MaxIVUsers) {
5510 (void)U;
5511 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)
5512 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5513 return;
5514 }
5515 // Bail out if we have a PHI on an EHPad that gets a value from a
5516 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5517 // no good place to stick any instructions.
5518 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5519 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5520 if (isa<FuncletPadInst>(FirstNonPHI) ||
5521 isa<CatchSwitchInst>(FirstNonPHI))
5522 for (BasicBlock *PredBB : PN->blocks())
5523 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5524 return;
5525 }
5526 }
5527
5528#ifndef NDEBUG
5529 // All dominating loops must have preheaders, or SCEVExpander may not be able
5530 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5531 //
5532 // IVUsers analysis should only create users that are dominated by simple loop
5533 // headers. Since this loop should dominate all of its users, its user list
5534 // should be empty if this loop itself is not within a simple loop nest.
5535 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
8
Loop condition is false. Execution continues on line 5545
5536 Rung; Rung = Rung->getIDom()) {
5537 BasicBlock *BB = Rung->getBlock();
5538 const Loop *DomLoop = LI.getLoopFor(BB);
5539 if (DomLoop && DomLoop->getHeader() == BB) {
5540 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5540, __PRETTY_FUNCTION__))
;
5541 }
5542 }
5543#endif // DEBUG
5544
5545 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)
9
Assuming 'DebugFlag' is false
10
Loop condition is false. Exiting loop
5546 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)
5547 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5548
5549 // First, perform some low-level loop optimizations.
5550 OptimizeShadowIV();
5551 OptimizeLoopTermCond();
5552
5553 // If loop preparation eliminates all interesting IV users, bail.
5554 if (IU.empty()) return;
11
Assuming the condition is false
12
Taking false branch
5555
5556 // Skip nested loops until we can model them better with formulae.
5557 if (!L->empty()) {
13
Assuming the condition is false
14
Taking false branch
5558 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)
;
5559 return;
5560 }
5561
5562 // Start collecting data and preparing for the solver.
5563 CollectChains();
5564 CollectInterestingTypesAndFactors();
5565 CollectFixupsAndInitialFormulae();
5566 CollectLoopInvariantFixupsAndFormulae();
5567
5568 if (Uses.empty())
15
Taking false branch
5569 return;
5570
5571 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
)
16
Assuming 'DebugFlag' is false
17
Loop condition is false. Exiting loop
5572 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
;
5573
5574 // Now use the reuse data to generate a bunch of interesting ways
5575 // to formulate the values needed for the uses.
5576 GenerateAllReuseFormulae();
5577
5578 FilterOutUndesirableDedicatedRegisters();
5579 NarrowSearchSpaceUsingHeuristics();
18
Calling 'LSRInstance::NarrowSearchSpaceUsingHeuristics'
5580
5581 SmallVector<const Formula *, 8> Solution;
5582 Solve(Solution);
5583
5584 // Release memory that is no longer needed.
5585 Factors.clear();
5586 Types.clear();
5587 RegUses.clear();
5588
5589 if (Solution.empty())
5590 return;
5591
5592#ifndef NDEBUG
5593 // Formulae should be legal.
5594 for (const LSRUse &LU : Uses) {
5595 for (const Formula &F : LU.Formulae)
5596 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5597, __PRETTY_FUNCTION__))
5597 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-10~svn372306/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5597, __PRETTY_FUNCTION__))
;
5598 };
5599#endif
5600
5601 // Now that we've decided what we want, make it so.
5602 ImplementSolution(Solution);
5603}
5604
5605#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5606void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5607 if (Factors.empty() && Types.empty()) return;
5608
5609 OS << "LSR has identified the following interesting factors and types: ";
5610 bool First = true;
5611
5612 for (int64_t Factor : Factors) {
5613 if (!First) OS << ", ";
5614 First = false;
5615 OS << '*' << Factor;
5616 }
5617
5618 for (Type *Ty : Types) {
5619 if (!First) OS << ", ";
5620 First = false;
5621 OS << '(' << *Ty << ')';
5622 }
5623 OS << '\n';
5624}
5625
5626void LSRInstance::print_fixups(raw_ostream &OS) const {
5627 OS << "LSR is examining the following fixup sites:\n";
5628 for (const LSRUse &LU : Uses)
5629 for (const LSRFixup &LF : LU.Fixups) {
5630 dbgs() << " ";
5631 LF.print(OS);
5632 OS << '\n';
5633 }
5634}
5635
5636void LSRInstance::print_uses(raw_ostream &OS) const {
5637 OS << "LSR is examining the following uses:\n";
5638 for (const LSRUse &LU : Uses) {
5639 dbgs() << " ";
5640 LU.print(OS);
5641 OS << '\n';
5642 for (const Formula &F : LU.Formulae) {
5643 OS << " ";
5644 F.print(OS);
5645 OS << '\n';
5646 }
5647 }
5648}
5649
5650void LSRInstance::print(raw_ostream &OS) const {
5651 print_factors_and_types(OS);
5652 print_fixups(OS);
5653 print_uses(OS);
5654}
5655
5656LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRInstance::dump() const {
5657 print(errs()); errs() << '\n';
5658}
5659#endif
5660
5661namespace {
5662
5663class LoopStrengthReduce : public LoopPass {
5664public:
5665 static char ID; // Pass ID, replacement for typeid
5666
5667 LoopStrengthReduce();
5668
5669private:
5670 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5671 void getAnalysisUsage(AnalysisUsage &AU) const override;
5672};
5673
5674} // end anonymous namespace
5675
5676LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5677 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5678}
5679
5680void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5681 // We split critical edges, so we change the CFG. However, we do update
5682 // many analyses if they are around.
5683 AU.addPreservedID(LoopSimplifyID);
5684
5685 AU.addRequired<LoopInfoWrapperPass>();
5686 AU.addPreserved<LoopInfoWrapperPass>();
5687 AU.addRequiredID(LoopSimplifyID);
5688 AU.addRequired<DominatorTreeWrapperPass>();
5689 AU.addPreserved<DominatorTreeWrapperPass>();
5690 AU.addRequired<ScalarEvolutionWrapperPass>();
5691 AU.addPreserved<ScalarEvolutionWrapperPass>();
5692 AU.addRequired<AssumptionCacheTracker>();
5693 AU.addRequired<TargetLibraryInfoWrapperPass>();
5694 // Requiring LoopSimplify a second time here prevents IVUsers from running
5695 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5696 AU.addRequiredID(LoopSimplifyID);
5697 AU.addRequired<IVUsersWrapperPass>();
5698 AU.addPreserved<IVUsersWrapperPass>();
5699 AU.addRequired<TargetTransformInfoWrapperPass>();
5700}
5701
5702static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5703 DominatorTree &DT, LoopInfo &LI,
5704 const TargetTransformInfo &TTI,
5705 AssumptionCache &AC,
5706 TargetLibraryInfo &LibInfo) {
5707
5708 bool Changed = false;
5709
5710 // Run the main LSR transformation.
5711 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI, AC, LibInfo).getChanged();
2
Calling constructor for 'LSRInstance'
5712
5713 // Remove any extra phis created by processing inner loops.
5714 Changed |= DeleteDeadPHIs(L->getHeader());
5715 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5716 SmallVector<WeakTrackingVH, 16> DeadInsts;
5717 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5718 SCEVExpander Rewriter(SE, DL, "lsr");
5719#ifndef NDEBUG
5720 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5721#endif
5722 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5723 if (numFolded) {
5724 Changed = true;
5725 DeleteTriviallyDeadInstructions(DeadInsts);
5726 DeleteDeadPHIs(L->getHeader());
5727 }
5728 }
5729 return Changed;
5730}
5731
5732bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5733 if (skipLoop(L))
5734 return false;
5735
5736 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5737 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5738 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5739 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5740 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5741 *L->getHeader()->getParent());
5742 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
5743 *L->getHeader()->getParent());
5744 auto &LibInfo = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
5745 *L->getHeader()->getParent());
5746 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI, AC, LibInfo);
5747}
5748
5749PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5750 LoopStandardAnalysisResults &AR,
5751 LPMUpdater &) {
5752 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
1
Calling 'ReduceLoopStrength'
5753 AR.DT, AR.LI, AR.TTI, AR.AC, AR.TLI))
5754 return PreservedAnalyses::all();
5755
5756 return getLoopPassPreservedAnalyses();
5757}
5758
5759char LoopStrengthReduce::ID = 0;
5760
5761INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
5762 "Loop Strength Reduction", false, false)static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
5763INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
5764INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
5765INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
5766INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)initializeIVUsersWrapperPassPass(Registry);
5767INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
5768INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
5769INITIALIZE_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)); }
5770 "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)); }
5771
5772Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }