Bug Summary

File:llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 5355, column 22
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name LoopStrengthReduce.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -mframe-pointer=none -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -target-cpu x86-64 -dwarf-column-info -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-11/lib/clang/11.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/build-llvm/include -I /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/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-11/lib/clang/11.0.0/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-11~++20200309111110+2c36c23f347/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-03-09-184146-41876-1 -x c++ /build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/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/MemorySSA.h"
73#include "llvm/Analysis/MemorySSAUpdater.h"
74#include "llvm/Analysis/ScalarEvolution.h"
75#include "llvm/Analysis/ScalarEvolutionExpander.h"
76#include "llvm/Analysis/ScalarEvolutionExpressions.h"
77#include "llvm/Analysis/ScalarEvolutionNormalization.h"
78#include "llvm/Analysis/TargetTransformInfo.h"
79#include "llvm/Config/llvm-config.h"
80#include "llvm/IR/BasicBlock.h"
81#include "llvm/IR/Constant.h"
82#include "llvm/IR/Constants.h"
83#include "llvm/IR/DerivedTypes.h"
84#include "llvm/IR/Dominators.h"
85#include "llvm/IR/GlobalValue.h"
86#include "llvm/IR/IRBuilder.h"
87#include "llvm/IR/InstrTypes.h"
88#include "llvm/IR/Instruction.h"
89#include "llvm/IR/Instructions.h"
90#include "llvm/IR/IntrinsicInst.h"
91#include "llvm/IR/Intrinsics.h"
92#include "llvm/IR/Module.h"
93#include "llvm/IR/OperandTraits.h"
94#include "llvm/IR/Operator.h"
95#include "llvm/IR/PassManager.h"
96#include "llvm/IR/Type.h"
97#include "llvm/IR/Use.h"
98#include "llvm/IR/User.h"
99#include "llvm/IR/Value.h"
100#include "llvm/IR/ValueHandle.h"
101#include "llvm/InitializePasses.h"
102#include "llvm/Pass.h"
103#include "llvm/Support/Casting.h"
104#include "llvm/Support/CommandLine.h"
105#include "llvm/Support/Compiler.h"
106#include "llvm/Support/Debug.h"
107#include "llvm/Support/ErrorHandling.h"
108#include "llvm/Support/MathExtras.h"
109#include "llvm/Support/raw_ostream.h"
110#include "llvm/Transforms/Scalar.h"
111#include "llvm/Transforms/Utils.h"
112#include "llvm/Transforms/Utils/BasicBlockUtils.h"
113#include "llvm/Transforms/Utils/Local.h"
114#include <algorithm>
115#include <cassert>
116#include <cstddef>
117#include <cstdint>
118#include <cstdlib>
119#include <iterator>
120#include <limits>
121#include <map>
122#include <numeric>
123#include <utility>
124
125using namespace llvm;
126
127#define DEBUG_TYPE"loop-reduce" "loop-reduce"
128
129/// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
130/// bail out. This threshold is far beyond the number of users that LSR can
131/// conceivably solve, so it should not affect generated code, but catches the
132/// worst cases before LSR burns too much compile time and stack space.
133static const unsigned MaxIVUsers = 200;
134
135// Temporary flag to cleanup congruent phis after LSR phi expansion.
136// It's currently disabled until we can determine whether it's truly useful or
137// not. The flag should be removed after the v3.0 release.
138// This is now needed for ivchains.
139static cl::opt<bool> EnablePhiElim(
140 "enable-lsr-phielim", cl::Hidden, cl::init(true),
141 cl::desc("Enable LSR phi elimination"));
142
143// The flag adds instruction count to solutions cost comparision.
144static cl::opt<bool> InsnsCost(
145 "lsr-insns-cost", cl::Hidden, cl::init(true),
146 cl::desc("Add instruction count to a LSR cost model"));
147
148// Flag to choose how to narrow complex lsr solution
149static cl::opt<bool> LSRExpNarrow(
150 "lsr-exp-narrow", cl::Hidden, cl::init(false),
151 cl::desc("Narrow LSR complex solution using"
152 " expectation of registers number"));
153
154// Flag to narrow search space by filtering non-optimal formulae with
155// the same ScaledReg and Scale.
156static cl::opt<bool> FilterSameScaledReg(
157 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
158 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
159 " with the same ScaledReg and Scale"));
160
161static cl::opt<bool> EnableBackedgeIndexing(
162 "lsr-backedge-indexing", cl::Hidden, cl::init(true),
163 cl::desc("Enable the generation of cross iteration indexed memops"));
164
165static cl::opt<unsigned> ComplexityLimit(
166 "lsr-complexity-limit", cl::Hidden,
167 cl::init(std::numeric_limits<uint16_t>::max()),
168 cl::desc("LSR search space complexity limit"));
169
170static cl::opt<unsigned> SetupCostDepthLimit(
171 "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
172 cl::desc("The limit on recursion depth for LSRs setup cost"));
173
174#ifndef NDEBUG
175// Stress test IV chain generation.
176static cl::opt<bool> StressIVChain(
177 "stress-ivchain", cl::Hidden, cl::init(false),
178 cl::desc("Stress test LSR IV chains"));
179#else
180static bool StressIVChain = false;
181#endif
182
183namespace {
184
185struct MemAccessTy {
186 /// Used in situations where the accessed memory type is unknown.
187 static const unsigned UnknownAddressSpace =
188 std::numeric_limits<unsigned>::max();
189
190 Type *MemTy = nullptr;
191 unsigned AddrSpace = UnknownAddressSpace;
192
193 MemAccessTy() = default;
194 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
195
196 bool operator==(MemAccessTy Other) const {
197 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
198 }
199
200 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
201
202 static MemAccessTy getUnknown(LLVMContext &Ctx,
203 unsigned AS = UnknownAddressSpace) {
204 return MemAccessTy(Type::getVoidTy(Ctx), AS);
205 }
206
207 Type *getType() { return MemTy; }
208};
209
210/// This class holds data which is used to order reuse candidates.
211class RegSortData {
212public:
213 /// This represents the set of LSRUse indices which reference
214 /// a particular register.
215 SmallBitVector UsedByIndices;
216
217 void print(raw_ostream &OS) const;
218 void dump() const;
219};
220
221} // end anonymous namespace
222
223#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
224void RegSortData::print(raw_ostream &OS) const {
225 OS << "[NumUses=" << UsedByIndices.count() << ']';
226}
227
228LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
229 print(errs()); errs() << '\n';
230}
231#endif
232
233namespace {
234
235/// Map register candidates to information about how they are used.
236class RegUseTracker {
237 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
238
239 RegUsesTy RegUsesMap;
240 SmallVector<const SCEV *, 16> RegSequence;
241
242public:
243 void countRegister(const SCEV *Reg, size_t LUIdx);
244 void dropRegister(const SCEV *Reg, size_t LUIdx);
245 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
246
247 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
248
249 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
250
251 void clear();
252
253 using iterator = SmallVectorImpl<const SCEV *>::iterator;
254 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
255
256 iterator begin() { return RegSequence.begin(); }
257 iterator end() { return RegSequence.end(); }
258 const_iterator begin() const { return RegSequence.begin(); }
259 const_iterator end() const { return RegSequence.end(); }
260};
261
262} // end anonymous namespace
263
264void
265RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
266 std::pair<RegUsesTy::iterator, bool> Pair =
267 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
268 RegSortData &RSD = Pair.first->second;
269 if (Pair.second)
270 RegSequence.push_back(Reg);
271 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
272 RSD.UsedByIndices.set(LUIdx);
273}
274
275void
276RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
277 RegUsesTy::iterator It = RegUsesMap.find(Reg);
278 assert(It != RegUsesMap.end())((It != RegUsesMap.end()) ? static_cast<void> (0) : __assert_fail
("It != RegUsesMap.end()", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 278, __PRETTY_FUNCTION__))
;
279 RegSortData &RSD = It->second;
280 assert(RSD.UsedByIndices.size() > LUIdx)((RSD.UsedByIndices.size() > LUIdx) ? static_cast<void>
(0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 280, __PRETTY_FUNCTION__))
;
281 RSD.UsedByIndices.reset(LUIdx);
282}
283
284void
285RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
286 assert(LUIdx <= LastLUIdx)((LUIdx <= LastLUIdx) ? static_cast<void> (0) : __assert_fail
("LUIdx <= LastLUIdx", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 286, __PRETTY_FUNCTION__))
;
287
288 // Update RegUses. The data structure is not optimized for this purpose;
289 // we must iterate through it and update each of the bit vectors.
290 for (auto &Pair : RegUsesMap) {
291 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
292 if (LUIdx < UsedByIndices.size())
293 UsedByIndices[LUIdx] =
294 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
295 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
296 }
297}
298
299bool
300RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
301 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
302 if (I == RegUsesMap.end())
303 return false;
304 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
305 int i = UsedByIndices.find_first();
306 if (i == -1) return false;
307 if ((size_t)i != LUIdx) return true;
308 return UsedByIndices.find_next(i) != -1;
309}
310
311const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
312 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
313 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 313, __PRETTY_FUNCTION__))
;
314 return I->second.UsedByIndices;
315}
316
317void RegUseTracker::clear() {
318 RegUsesMap.clear();
319 RegSequence.clear();
320}
321
322namespace {
323
324/// This class holds information that describes a formula for computing
325/// satisfying a use. It may include broken-out immediates and scaled registers.
326struct Formula {
327 /// Global base address used for complex addressing.
328 GlobalValue *BaseGV = nullptr;
329
330 /// Base offset for complex addressing.
331 int64_t BaseOffset = 0;
332
333 /// Whether any complex addressing has a base register.
334 bool HasBaseReg = false;
335
336 /// The scale of any complex addressing.
337 int64_t Scale = 0;
338
339 /// The list of "base" registers for this use. When this is non-empty. The
340 /// canonical representation of a formula is
341 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
342 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
343 /// 3. The reg containing recurrent expr related with currect loop in the
344 /// formula should be put in the ScaledReg.
345 /// #1 enforces that the scaled register is always used when at least two
346 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
347 /// #2 enforces that 1 * reg is reg.
348 /// #3 ensures invariant regs with respect to current loop can be combined
349 /// together in LSR codegen.
350 /// This invariant can be temporarily broken while building a formula.
351 /// However, every formula inserted into the LSRInstance must be in canonical
352 /// form.
353 SmallVector<const SCEV *, 4> BaseRegs;
354
355 /// The 'scaled' register for this use. This should be non-null when Scale is
356 /// not zero.
357 const SCEV *ScaledReg = nullptr;
358
359 /// An additional constant offset which added near the use. This requires a
360 /// temporary register, but the offset itself can live in an add immediate
361 /// field rather than a register.
362 int64_t UnfoldedOffset = 0;
363
364 Formula() = default;
365
366 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
367
368 bool isCanonical(const Loop &L) const;
369
370 void canonicalize(const Loop &L);
371
372 bool unscale();
373
374 bool hasZeroEnd() const;
375
376 size_t getNumRegs() const;
377 Type *getType() const;
378
379 void deleteBaseReg(const SCEV *&S);
380
381 bool referencesReg(const SCEV *S) const;
382 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
383 const RegUseTracker &RegUses) const;
384
385 void print(raw_ostream &OS) const;
386 void dump() const;
387};
388
389} // end anonymous namespace
390
391/// Recursion helper for initialMatch.
392static void DoInitialMatch(const SCEV *S, Loop *L,
393 SmallVectorImpl<const SCEV *> &Good,
394 SmallVectorImpl<const SCEV *> &Bad,
395 ScalarEvolution &SE) {
396 // Collect expressions which properly dominate the loop header.
397 if (SE.properlyDominates(S, L->getHeader())) {
398 Good.push_back(S);
399 return;
400 }
401
402 // Look at add operands.
403 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
404 for (const SCEV *S : Add->operands())
405 DoInitialMatch(S, L, Good, Bad, SE);
406 return;
407 }
408
409 // Look at addrec operands.
410 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
411 if (!AR->getStart()->isZero() && AR->isAffine()) {
412 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
413 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
414 AR->getStepRecurrence(SE),
415 // FIXME: AR->getNoWrapFlags()
416 AR->getLoop(), SCEV::FlagAnyWrap),
417 L, Good, Bad, SE);
418 return;
419 }
420
421 // Handle a multiplication by -1 (negation) if it didn't fold.
422 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
423 if (Mul->getOperand(0)->isAllOnesValue()) {
424 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
425 const SCEV *NewMul = SE.getMulExpr(Ops);
426
427 SmallVector<const SCEV *, 4> MyGood;
428 SmallVector<const SCEV *, 4> MyBad;
429 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
430 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
431 SE.getEffectiveSCEVType(NewMul->getType())));
432 for (const SCEV *S : MyGood)
433 Good.push_back(SE.getMulExpr(NegOne, S));
434 for (const SCEV *S : MyBad)
435 Bad.push_back(SE.getMulExpr(NegOne, S));
436 return;
437 }
438
439 // Ok, we can't do anything interesting. Just stuff the whole thing into a
440 // register and hope for the best.
441 Bad.push_back(S);
442}
443
444/// Incorporate loop-variant parts of S into this Formula, attempting to keep
445/// all loop-invariant and loop-computable values in a single base register.
446void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
447 SmallVector<const SCEV *, 4> Good;
448 SmallVector<const SCEV *, 4> Bad;
449 DoInitialMatch(S, L, Good, Bad, SE);
450 if (!Good.empty()) {
451 const SCEV *Sum = SE.getAddExpr(Good);
452 if (!Sum->isZero())
453 BaseRegs.push_back(Sum);
454 HasBaseReg = true;
455 }
456 if (!Bad.empty()) {
457 const SCEV *Sum = SE.getAddExpr(Bad);
458 if (!Sum->isZero())
459 BaseRegs.push_back(Sum);
460 HasBaseReg = true;
461 }
462 canonicalize(*L);
463}
464
465/// Check whether or not this formula satisfies the canonical
466/// representation.
467/// \see Formula::BaseRegs.
468bool Formula::isCanonical(const Loop &L) const {
469 if (!ScaledReg)
470 return BaseRegs.size() <= 1;
471
472 if (Scale != 1)
473 return true;
474
475 if (Scale == 1 && BaseRegs.empty())
476 return false;
477
478 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
479 if (SAR && SAR->getLoop() == &L)
480 return true;
481
482 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
483 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
484 // loop, we want to swap the reg in BaseRegs with ScaledReg.
485 auto I =
486 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
487 return isa<const SCEVAddRecExpr>(S) &&
488 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
489 });
490 return I == BaseRegs.end();
491}
492
493/// Helper method to morph a formula into its canonical representation.
494/// \see Formula::BaseRegs.
495/// Every formula having more than one base register, must use the ScaledReg
496/// field. Otherwise, we would have to do special cases everywhere in LSR
497/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
498/// On the other hand, 1*reg should be canonicalized into reg.
499void Formula::canonicalize(const Loop &L) {
500 if (isCanonical(L))
501 return;
502 // So far we did not need this case. This is easy to implement but it is
503 // useless to maintain dead code. Beside it could hurt compile time.
504 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 504, __PRETTY_FUNCTION__))
;
505
506 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
507 if (!ScaledReg) {
508 ScaledReg = BaseRegs.back();
509 BaseRegs.pop_back();
510 Scale = 1;
511 }
512
513 // If ScaledReg is an invariant with respect to L, find the reg from
514 // BaseRegs containing the recurrent expr related with Loop L. Swap the
515 // reg with ScaledReg.
516 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
517 if (!SAR || SAR->getLoop() != &L) {
518 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
519 [&](const SCEV *S) {
520 return isa<const SCEVAddRecExpr>(S) &&
521 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
522 });
523 if (I != BaseRegs.end())
524 std::swap(ScaledReg, *I);
525 }
526}
527
528/// Get rid of the scale in the formula.
529/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
530/// \return true if it was possible to get rid of the scale, false otherwise.
531/// \note After this operation the formula may not be in the canonical form.
532bool Formula::unscale() {
533 if (Scale != 1)
534 return false;
535 Scale = 0;
536 BaseRegs.push_back(ScaledReg);
537 ScaledReg = nullptr;
538 return true;
539}
540
541bool Formula::hasZeroEnd() const {
542 if (UnfoldedOffset || BaseOffset)
543 return false;
544 if (BaseRegs.size() != 1 || ScaledReg)
545 return false;
546 return true;
547}
548
549/// Return the total number of register operands used by this formula. This does
550/// not include register uses implied by non-constant addrec strides.
551size_t Formula::getNumRegs() const {
552 return !!ScaledReg + BaseRegs.size();
553}
554
555/// Return the type of this formula, if it has one, or null otherwise. This type
556/// is meaningless except for the bit size.
557Type *Formula::getType() const {
558 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
559 ScaledReg ? ScaledReg->getType() :
560 BaseGV ? BaseGV->getType() :
561 nullptr;
562}
563
564/// Delete the given base reg from the BaseRegs list.
565void Formula::deleteBaseReg(const SCEV *&S) {
566 if (&S != &BaseRegs.back())
567 std::swap(S, BaseRegs.back());
568 BaseRegs.pop_back();
569}
570
571/// Test if this formula references the given register.
572bool Formula::referencesReg(const SCEV *S) const {
573 return S == ScaledReg || is_contained(BaseRegs, S);
574}
575
576/// Test whether this formula uses registers which are used by uses other than
577/// the use with the given index.
578bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
579 const RegUseTracker &RegUses) const {
580 if (ScaledReg)
581 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
582 return true;
583 for (const SCEV *BaseReg : BaseRegs)
584 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
585 return true;
586 return false;
587}
588
589#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
590void Formula::print(raw_ostream &OS) const {
591 bool First = true;
592 if (BaseGV) {
593 if (!First) OS << " + "; else First = false;
594 BaseGV->printAsOperand(OS, /*PrintType=*/false);
595 }
596 if (BaseOffset != 0) {
597 if (!First) OS << " + "; else First = false;
598 OS << BaseOffset;
599 }
600 for (const SCEV *BaseReg : BaseRegs) {
601 if (!First) OS << " + "; else First = false;
602 OS << "reg(" << *BaseReg << ')';
603 }
604 if (HasBaseReg && BaseRegs.empty()) {
605 if (!First) OS << " + "; else First = false;
606 OS << "**error: HasBaseReg**";
607 } else if (!HasBaseReg && !BaseRegs.empty()) {
608 if (!First) OS << " + "; else First = false;
609 OS << "**error: !HasBaseReg**";
610 }
611 if (Scale != 0) {
612 if (!First) OS << " + "; else First = false;
613 OS << Scale << "*reg(";
614 if (ScaledReg)
615 OS << *ScaledReg;
616 else
617 OS << "<unknown>";
618 OS << ')';
619 }
620 if (UnfoldedOffset != 0) {
621 if (!First) OS << " + ";
622 OS << "imm(" << UnfoldedOffset << ')';
623 }
624}
625
626LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
627 print(errs()); errs() << '\n';
628}
629#endif
630
631/// Return true if the given addrec can be sign-extended without changing its
632/// value.
633static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
634 Type *WideTy =
635 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
636 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
637}
638
639/// Return true if the given add can be sign-extended without changing its
640/// value.
641static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
642 Type *WideTy =
643 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
644 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
645}
646
647/// Return true if the given mul can be sign-extended without changing its
648/// value.
649static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
650 Type *WideTy =
651 IntegerType::get(SE.getContext(),
652 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
653 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
654}
655
656/// Return an expression for LHS /s RHS, if it can be determined and if the
657/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
658/// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
659/// the multiplication may overflow, which is useful when the result will be
660/// used in a context where the most significant bits are ignored.
661static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
662 ScalarEvolution &SE,
663 bool IgnoreSignificantBits = false) {
664 // Handle the trivial case, which works for any SCEV type.
665 if (LHS == RHS)
666 return SE.getConstant(LHS->getType(), 1);
667
668 // Handle a few RHS special cases.
669 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
670 if (RC) {
671 const APInt &RA = RC->getAPInt();
672 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
673 // some folding.
674 if (RA.isAllOnesValue())
675 return SE.getMulExpr(LHS, RC);
676 // Handle x /s 1 as x.
677 if (RA == 1)
678 return LHS;
679 }
680
681 // Check for a division of a constant by a constant.
682 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
683 if (!RC)
684 return nullptr;
685 const APInt &LA = C->getAPInt();
686 const APInt &RA = RC->getAPInt();
687 if (LA.srem(RA) != 0)
688 return nullptr;
689 return SE.getConstant(LA.sdiv(RA));
690 }
691
692 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
693 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
694 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
695 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
696 IgnoreSignificantBits);
697 if (!Step) return nullptr;
698 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
699 IgnoreSignificantBits);
700 if (!Start) return nullptr;
701 // FlagNW is independent of the start value, step direction, and is
702 // preserved with smaller magnitude steps.
703 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
704 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
705 }
706 return nullptr;
707 }
708
709 // Distribute the sdiv over add operands, if the add doesn't overflow.
710 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
711 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
712 SmallVector<const SCEV *, 8> Ops;
713 for (const SCEV *S : Add->operands()) {
714 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
715 if (!Op) return nullptr;
716 Ops.push_back(Op);
717 }
718 return SE.getAddExpr(Ops);
719 }
720 return nullptr;
721 }
722
723 // Check for a multiply operand that we can pull RHS out of.
724 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
725 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
726 SmallVector<const SCEV *, 4> Ops;
727 bool Found = false;
728 for (const SCEV *S : Mul->operands()) {
729 if (!Found)
730 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
731 IgnoreSignificantBits)) {
732 S = Q;
733 Found = true;
734 }
735 Ops.push_back(S);
736 }
737 return Found ? SE.getMulExpr(Ops) : nullptr;
738 }
739 return nullptr;
740 }
741
742 // Otherwise we don't know.
743 return nullptr;
744}
745
746/// If S involves the addition of a constant integer value, return that integer
747/// value, and mutate S to point to a new SCEV with that value excluded.
748static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
749 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
750 if (C->getAPInt().getMinSignedBits() <= 64) {
751 S = SE.getConstant(C->getType(), 0);
752 return C->getValue()->getSExtValue();
753 }
754 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
755 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
756 int64_t Result = ExtractImmediate(NewOps.front(), SE);
757 if (Result != 0)
758 S = SE.getAddExpr(NewOps);
759 return Result;
760 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
761 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
762 int64_t Result = ExtractImmediate(NewOps.front(), SE);
763 if (Result != 0)
764 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
765 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
766 SCEV::FlagAnyWrap);
767 return Result;
768 }
769 return 0;
770}
771
772/// If S involves the addition of a GlobalValue address, return that symbol, and
773/// mutate S to point to a new SCEV with that value excluded.
774static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
775 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
776 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
777 S = SE.getConstant(GV->getType(), 0);
778 return GV;
779 }
780 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
781 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
782 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
783 if (Result)
784 S = SE.getAddExpr(NewOps);
785 return Result;
786 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
787 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
788 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
789 if (Result)
790 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
791 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
792 SCEV::FlagAnyWrap);
793 return Result;
794 }
795 return nullptr;
796}
797
798/// Returns true if the specified instruction is using the specified value as an
799/// address.
800static bool isAddressUse(const TargetTransformInfo &TTI,
801 Instruction *Inst, Value *OperandVal) {
802 bool isAddress = isa<LoadInst>(Inst);
803 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
804 if (SI->getPointerOperand() == OperandVal)
805 isAddress = true;
806 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
807 // Addressing modes can also be folded into prefetches and a variety
808 // of intrinsics.
809 switch (II->getIntrinsicID()) {
810 case Intrinsic::memset:
811 case Intrinsic::prefetch:
812 case Intrinsic::masked_load:
813 if (II->getArgOperand(0) == OperandVal)
814 isAddress = true;
815 break;
816 case Intrinsic::masked_store:
817 if (II->getArgOperand(1) == OperandVal)
818 isAddress = true;
819 break;
820 case Intrinsic::memmove:
821 case Intrinsic::memcpy:
822 if (II->getArgOperand(0) == OperandVal ||
823 II->getArgOperand(1) == OperandVal)
824 isAddress = true;
825 break;
826 default: {
827 MemIntrinsicInfo IntrInfo;
828 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
829 if (IntrInfo.PtrVal == OperandVal)
830 isAddress = true;
831 }
832 }
833 }
834 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
835 if (RMW->getPointerOperand() == OperandVal)
836 isAddress = true;
837 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
838 if (CmpX->getPointerOperand() == OperandVal)
839 isAddress = true;
840 }
841 return isAddress;
842}
843
844/// Return the type of the memory being accessed.
845static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
846 Instruction *Inst, Value *OperandVal) {
847 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
848 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
849 AccessTy.MemTy = SI->getOperand(0)->getType();
850 AccessTy.AddrSpace = SI->getPointerAddressSpace();
851 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
852 AccessTy.AddrSpace = LI->getPointerAddressSpace();
853 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
854 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
855 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
856 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
857 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
858 switch (II->getIntrinsicID()) {
859 case Intrinsic::prefetch:
860 case Intrinsic::memset:
861 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
862 AccessTy.MemTy = OperandVal->getType();
863 break;
864 case Intrinsic::memmove:
865 case Intrinsic::memcpy:
866 AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
867 AccessTy.MemTy = OperandVal->getType();
868 break;
869 case Intrinsic::masked_load:
870 AccessTy.AddrSpace =
871 II->getArgOperand(0)->getType()->getPointerAddressSpace();
872 break;
873 case Intrinsic::masked_store:
874 AccessTy.MemTy = II->getOperand(0)->getType();
875 AccessTy.AddrSpace =
876 II->getArgOperand(1)->getType()->getPointerAddressSpace();
877 break;
878 default: {
879 MemIntrinsicInfo IntrInfo;
880 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
881 AccessTy.AddrSpace
882 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
883 }
884
885 break;
886 }
887 }
888 }
889
890 // All pointers have the same requirements, so canonicalize them to an
891 // arbitrary pointer type to minimize variation.
892 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
893 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
894 PTy->getAddressSpace());
895
896 return AccessTy;
897}
898
899/// Return true if this AddRec is already a phi in its loop.
900static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
901 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
902 if (SE.isSCEVable(PN.getType()) &&
903 (SE.getEffectiveSCEVType(PN.getType()) ==
904 SE.getEffectiveSCEVType(AR->getType())) &&
905 SE.getSCEV(&PN) == AR)
906 return true;
907 }
908 return false;
909}
910
911/// Check if expanding this expression is likely to incur significant cost. This
912/// is tricky because SCEV doesn't track which expressions are actually computed
913/// by the current IR.
914///
915/// We currently allow expansion of IV increments that involve adds,
916/// multiplication by constants, and AddRecs from existing phis.
917///
918/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
919/// obvious multiple of the UDivExpr.
920static bool isHighCostExpansion(const SCEV *S,
921 SmallPtrSetImpl<const SCEV*> &Processed,
922 ScalarEvolution &SE) {
923 // Zero/One operand expressions
924 switch (S->getSCEVType()) {
925 case scUnknown:
926 case scConstant:
927 return false;
928 case scTruncate:
929 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
930 Processed, SE);
931 case scZeroExtend:
932 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
933 Processed, SE);
934 case scSignExtend:
935 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
936 Processed, SE);
937 }
938
939 if (!Processed.insert(S).second)
940 return false;
941
942 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
943 for (const SCEV *S : Add->operands()) {
944 if (isHighCostExpansion(S, Processed, SE))
945 return true;
946 }
947 return false;
948 }
949
950 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
951 if (Mul->getNumOperands() == 2) {
952 // Multiplication by a constant is ok
953 if (isa<SCEVConstant>(Mul->getOperand(0)))
954 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
955
956 // If we have the value of one operand, check if an existing
957 // multiplication already generates this expression.
958 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
959 Value *UVal = U->getValue();
960 for (User *UR : UVal->users()) {
961 // If U is a constant, it may be used by a ConstantExpr.
962 Instruction *UI = dyn_cast<Instruction>(UR);
963 if (UI && UI->getOpcode() == Instruction::Mul &&
964 SE.isSCEVable(UI->getType())) {
965 return SE.getSCEV(UI) == Mul;
966 }
967 }
968 }
969 }
970 }
971
972 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
973 if (isExistingPhi(AR, SE))
974 return false;
975 }
976
977 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
978 return true;
979}
980
981namespace {
982
983class LSRUse;
984
985} // end anonymous namespace
986
987/// Check if the addressing mode defined by \p F is completely
988/// folded in \p LU at isel time.
989/// This includes address-mode folding and special icmp tricks.
990/// This function returns true if \p LU can accommodate what \p F
991/// defines and up to 1 base + 1 scaled + offset.
992/// In other words, if \p F has several base registers, this function may
993/// still return true. Therefore, users still need to account for
994/// additional base registers and/or unfolded offsets to derive an
995/// accurate cost model.
996static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
997 const LSRUse &LU, const Formula &F);
998
999// Get the cost of the scaling factor used in F for LU.
1000static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1001 const LSRUse &LU, const Formula &F,
1002 const Loop &L);
1003
1004namespace {
1005
1006/// This class is used to measure and compare candidate formulae.
1007class Cost {
1008 const Loop *L = nullptr;
1009 ScalarEvolution *SE = nullptr;
1010 const TargetTransformInfo *TTI = nullptr;
1011 TargetTransformInfo::LSRCost C;
1012
1013public:
1014 Cost() = delete;
1015 Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI) :
1016 L(L), SE(&SE), TTI(&TTI) {
1017 C.Insns = 0;
1018 C.NumRegs = 0;
1019 C.AddRecCost = 0;
1020 C.NumIVMuls = 0;
1021 C.NumBaseAdds = 0;
1022 C.ImmCost = 0;
1023 C.SetupCost = 0;
1024 C.ScaleCost = 0;
1025 }
1026
1027 bool isLess(Cost &Other);
1028
1029 void Lose();
1030
1031#ifndef NDEBUG
1032 // Once any of the metrics loses, they must all remain losers.
1033 bool isValid() {
1034 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1035 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1036 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1037 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1038 }
1039#endif
1040
1041 bool isLoser() {
1042 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1042, __PRETTY_FUNCTION__))
;
1043 return C.NumRegs == ~0u;
1044 }
1045
1046 void RateFormula(const Formula &F,
1047 SmallPtrSetImpl<const SCEV *> &Regs,
1048 const DenseSet<const SCEV *> &VisitedRegs,
1049 const LSRUse &LU,
1050 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1051
1052 void print(raw_ostream &OS) const;
1053 void dump() const;
1054
1055private:
1056 void RateRegister(const Formula &F, const SCEV *Reg,
1057 SmallPtrSetImpl<const SCEV *> &Regs);
1058 void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1059 SmallPtrSetImpl<const SCEV *> &Regs,
1060 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1061};
1062
1063/// An operand value in an instruction which is to be replaced with some
1064/// equivalent, possibly strength-reduced, replacement.
1065struct LSRFixup {
1066 /// The instruction which will be updated.
1067 Instruction *UserInst = nullptr;
1068
1069 /// The operand of the instruction which will be replaced. The operand may be
1070 /// used more than once; every instance will be replaced.
1071 Value *OperandValToReplace = nullptr;
1072
1073 /// If this user is to use the post-incremented value of an induction
1074 /// variable, this set is non-empty and holds the loops associated with the
1075 /// induction variable.
1076 PostIncLoopSet PostIncLoops;
1077
1078 /// A constant offset to be added to the LSRUse expression. This allows
1079 /// multiple fixups to share the same LSRUse with different offsets, for
1080 /// example in an unrolled loop.
1081 int64_t Offset = 0;
1082
1083 LSRFixup() = default;
1084
1085 bool isUseFullyOutsideLoop(const Loop *L) const;
1086
1087 void print(raw_ostream &OS) const;
1088 void dump() const;
1089};
1090
1091/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1092/// SmallVectors of const SCEV*.
1093struct UniquifierDenseMapInfo {
1094 static SmallVector<const SCEV *, 4> getEmptyKey() {
1095 SmallVector<const SCEV *, 4> V;
1096 V.push_back(reinterpret_cast<const SCEV *>(-1));
1097 return V;
1098 }
1099
1100 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1101 SmallVector<const SCEV *, 4> V;
1102 V.push_back(reinterpret_cast<const SCEV *>(-2));
1103 return V;
1104 }
1105
1106 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1107 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1108 }
1109
1110 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1111 const SmallVector<const SCEV *, 4> &RHS) {
1112 return LHS == RHS;
1113 }
1114};
1115
1116/// This class holds the state that LSR keeps for each use in IVUsers, as well
1117/// as uses invented by LSR itself. It includes information about what kinds of
1118/// things can be folded into the user, information about the user itself, and
1119/// information about how the use may be satisfied. TODO: Represent multiple
1120/// users of the same expression in common?
1121class LSRUse {
1122 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1123
1124public:
1125 /// An enum for a kind of use, indicating what types of scaled and immediate
1126 /// operands it might support.
1127 enum KindType {
1128 Basic, ///< A normal use, with no folding.
1129 Special, ///< A special case of basic, allowing -1 scales.
1130 Address, ///< An address use; folding according to TargetLowering
1131 ICmpZero ///< An equality icmp with both operands folded into one.
1132 // TODO: Add a generic icmp too?
1133 };
1134
1135 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1136
1137 KindType Kind;
1138 MemAccessTy AccessTy;
1139
1140 /// The list of operands which are to be replaced.
1141 SmallVector<LSRFixup, 8> Fixups;
1142
1143 /// Keep track of the min and max offsets of the fixups.
1144 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1145 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1146
1147 /// This records whether all of the fixups using this LSRUse are outside of
1148 /// the loop, in which case some special-case heuristics may be used.
1149 bool AllFixupsOutsideLoop = true;
1150
1151 /// RigidFormula is set to true to guarantee that this use will be associated
1152 /// with a single formula--the one that initially matched. Some SCEV
1153 /// expressions cannot be expanded. This allows LSR to consider the registers
1154 /// used by those expressions without the need to expand them later after
1155 /// changing the formula.
1156 bool RigidFormula = false;
1157
1158 /// This records the widest use type for any fixup using this
1159 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1160 /// fixup widths to be equivalent, because the narrower one may be relying on
1161 /// the implicit truncation to truncate away bogus bits.
1162 Type *WidestFixupType = nullptr;
1163
1164 /// A list of ways to build a value that can satisfy this user. After the
1165 /// list is populated, one of these is selected heuristically and used to
1166 /// formulate a replacement for OperandValToReplace in UserInst.
1167 SmallVector<Formula, 12> Formulae;
1168
1169 /// The set of register candidates used by all formulae in this LSRUse.
1170 SmallPtrSet<const SCEV *, 4> Regs;
1171
1172 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1173
1174 LSRFixup &getNewFixup() {
1175 Fixups.push_back(LSRFixup());
1176 return Fixups.back();
1177 }
1178
1179 void pushFixup(LSRFixup &f) {
1180 Fixups.push_back(f);
1181 if (f.Offset > MaxOffset)
1182 MaxOffset = f.Offset;
1183 if (f.Offset < MinOffset)
1184 MinOffset = f.Offset;
1185 }
1186
1187 bool HasFormulaWithSameRegs(const Formula &F) const;
1188 float getNotSelectedProbability(const SCEV *Reg) const;
1189 bool InsertFormula(const Formula &F, const Loop &L);
1190 void DeleteFormula(Formula &F);
1191 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1192
1193 void print(raw_ostream &OS) const;
1194 void dump() const;
1195};
1196
1197} // end anonymous namespace
1198
1199static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1200 LSRUse::KindType Kind, MemAccessTy AccessTy,
1201 GlobalValue *BaseGV, int64_t BaseOffset,
1202 bool HasBaseReg, int64_t Scale,
1203 Instruction *Fixup = nullptr);
1204
1205static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1206 if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1207 return 1;
1208 if (Depth == 0)
1209 return 0;
1210 if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1211 return getSetupCost(S->getStart(), Depth - 1);
1212 if (auto S = dyn_cast<SCEVCastExpr>(Reg))
1213 return getSetupCost(S->getOperand(), Depth - 1);
1214 if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1215 return std::accumulate(S->op_begin(), S->op_end(), 0,
1216 [&](unsigned i, const SCEV *Reg) {
1217 return i + getSetupCost(Reg, Depth - 1);
1218 });
1219 if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1220 return getSetupCost(S->getLHS(), Depth - 1) +
1221 getSetupCost(S->getRHS(), Depth - 1);
1222 return 0;
1223}
1224
1225/// Tally up interesting quantities from the given register.
1226void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1227 SmallPtrSetImpl<const SCEV *> &Regs) {
1228 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1229 // If this is an addrec for another loop, it should be an invariant
1230 // with respect to L since L is the innermost loop (at least
1231 // for now LSR only handles innermost loops).
1232 if (AR->getLoop() != L) {
1233 // If the AddRec exists, consider it's register free and leave it alone.
1234 if (isExistingPhi(AR, *SE))
1235 return;
1236
1237 // It is bad to allow LSR for current loop to add induction variables
1238 // for its sibling loops.
1239 if (!AR->getLoop()->contains(L)) {
1240 Lose();
1241 return;
1242 }
1243
1244 // Otherwise, it will be an invariant with respect to Loop L.
1245 ++C.NumRegs;
1246 return;
1247 }
1248
1249 unsigned LoopCost = 1;
1250 if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1251 TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1252
1253 // If the step size matches the base offset, we could use pre-indexed
1254 // addressing.
1255 if (TTI->shouldFavorBackedgeIndex(L)) {
1256 if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1257 if (Step->getAPInt() == F.BaseOffset)
1258 LoopCost = 0;
1259 }
1260
1261 if (TTI->shouldFavorPostInc()) {
1262 const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1263 if (isa<SCEVConstant>(LoopStep)) {
1264 const SCEV *LoopStart = AR->getStart();
1265 if (!isa<SCEVConstant>(LoopStart) &&
1266 SE->isLoopInvariant(LoopStart, L))
1267 LoopCost = 0;
1268 }
1269 }
1270 }
1271 C.AddRecCost += LoopCost;
1272
1273 // Add the step value register, if it needs one.
1274 // TODO: The non-affine case isn't precisely modeled here.
1275 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1276 if (!Regs.count(AR->getOperand(1))) {
1277 RateRegister(F, AR->getOperand(1), Regs);
1278 if (isLoser())
1279 return;
1280 }
1281 }
1282 }
1283 ++C.NumRegs;
1284
1285 // Rough heuristic; favor registers which don't require extra setup
1286 // instructions in the preheader.
1287 C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1288 // Ensure we don't, even with the recusion limit, produce invalid costs.
1289 C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1290
1291 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1292 SE->hasComputableLoopEvolution(Reg, L);
1293}
1294
1295/// Record this register in the set. If we haven't seen it before, rate
1296/// it. Optional LoserRegs provides a way to declare any formula that refers to
1297/// one of those regs an instant loser.
1298void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1299 SmallPtrSetImpl<const SCEV *> &Regs,
1300 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1301 if (LoserRegs && LoserRegs->count(Reg)) {
1302 Lose();
1303 return;
1304 }
1305 if (Regs.insert(Reg).second) {
1306 RateRegister(F, Reg, Regs);
1307 if (LoserRegs && isLoser())
1308 LoserRegs->insert(Reg);
1309 }
1310}
1311
1312void Cost::RateFormula(const Formula &F,
1313 SmallPtrSetImpl<const SCEV *> &Regs,
1314 const DenseSet<const SCEV *> &VisitedRegs,
1315 const LSRUse &LU,
1316 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1317 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1317, __PRETTY_FUNCTION__))
;
1318 // Tally up the registers.
1319 unsigned PrevAddRecCost = C.AddRecCost;
1320 unsigned PrevNumRegs = C.NumRegs;
1321 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1322 if (const SCEV *ScaledReg = F.ScaledReg) {
1323 if (VisitedRegs.count(ScaledReg)) {
1324 Lose();
1325 return;
1326 }
1327 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1328 if (isLoser())
1329 return;
1330 }
1331 for (const SCEV *BaseReg : F.BaseRegs) {
1332 if (VisitedRegs.count(BaseReg)) {
1333 Lose();
1334 return;
1335 }
1336 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1337 if (isLoser())
1338 return;
1339 }
1340
1341 // Determine how many (unfolded) adds we'll need inside the loop.
1342 size_t NumBaseParts = F.getNumRegs();
1343 if (NumBaseParts > 1)
1344 // Do not count the base and a possible second register if the target
1345 // allows to fold 2 registers.
1346 C.NumBaseAdds +=
1347 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1348 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1349
1350 // Accumulate non-free scaling amounts.
1351 C.ScaleCost += getScalingFactorCost(*TTI, LU, F, *L);
1352
1353 // Tally up the non-zero immediates.
1354 for (const LSRFixup &Fixup : LU.Fixups) {
1355 int64_t O = Fixup.Offset;
1356 int64_t Offset = (uint64_t)O + F.BaseOffset;
1357 if (F.BaseGV)
1358 C.ImmCost += 64; // Handle symbolic values conservatively.
1359 // TODO: This should probably be the pointer size.
1360 else if (Offset != 0)
1361 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1362
1363 // Check with target if this offset with this instruction is
1364 // specifically not supported.
1365 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1366 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1367 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1368 C.NumBaseAdds++;
1369 }
1370
1371 // If we don't count instruction cost exit here.
1372 if (!InsnsCost) {
1373 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1373, __PRETTY_FUNCTION__))
;
1374 return;
1375 }
1376
1377 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1378 // additional instruction (at least fill).
1379 // TODO: Need distinguish register class?
1380 unsigned TTIRegNum = TTI->getNumberOfRegisters(
1381 TTI->getRegisterClassForType(false, F.getType())) - 1;
1382 if (C.NumRegs > TTIRegNum) {
1383 // Cost already exceeded TTIRegNum, then only newly added register can add
1384 // new instructions.
1385 if (PrevNumRegs > TTIRegNum)
1386 C.Insns += (C.NumRegs - PrevNumRegs);
1387 else
1388 C.Insns += (C.NumRegs - TTIRegNum);
1389 }
1390
1391 // If ICmpZero formula ends with not 0, it could not be replaced by
1392 // just add or sub. We'll need to compare final result of AddRec.
1393 // That means we'll need an additional instruction. But if the target can
1394 // macro-fuse a compare with a branch, don't count this extra instruction.
1395 // For -10 + {0, +, 1}:
1396 // i = i + 1;
1397 // cmp i, 10
1398 //
1399 // For {-10, +, 1}:
1400 // i = i + 1;
1401 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1402 !TTI->canMacroFuseCmp())
1403 C.Insns++;
1404 // Each new AddRec adds 1 instruction to calculation.
1405 C.Insns += (C.AddRecCost - PrevAddRecCost);
1406
1407 // BaseAdds adds instructions for unfolded registers.
1408 if (LU.Kind != LSRUse::ICmpZero)
1409 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1410 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1410, __PRETTY_FUNCTION__))
;
1411}
1412
1413/// Set this cost to a losing value.
1414void Cost::Lose() {
1415 C.Insns = std::numeric_limits<unsigned>::max();
1416 C.NumRegs = std::numeric_limits<unsigned>::max();
1417 C.AddRecCost = std::numeric_limits<unsigned>::max();
1418 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1419 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1420 C.ImmCost = std::numeric_limits<unsigned>::max();
1421 C.SetupCost = std::numeric_limits<unsigned>::max();
1422 C.ScaleCost = std::numeric_limits<unsigned>::max();
1423}
1424
1425/// Choose the lower cost.
1426bool Cost::isLess(Cost &Other) {
1427 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1428 C.Insns != Other.C.Insns)
1429 return C.Insns < Other.C.Insns;
1430 return TTI->isLSRCostLess(C, Other.C);
1431}
1432
1433#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1434void Cost::print(raw_ostream &OS) const {
1435 if (InsnsCost)
1436 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1437 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1438 if (C.AddRecCost != 0)
1439 OS << ", with addrec cost " << C.AddRecCost;
1440 if (C.NumIVMuls != 0)
1441 OS << ", plus " << C.NumIVMuls << " IV mul"
1442 << (C.NumIVMuls == 1 ? "" : "s");
1443 if (C.NumBaseAdds != 0)
1444 OS << ", plus " << C.NumBaseAdds << " base add"
1445 << (C.NumBaseAdds == 1 ? "" : "s");
1446 if (C.ScaleCost != 0)
1447 OS << ", plus " << C.ScaleCost << " scale cost";
1448 if (C.ImmCost != 0)
1449 OS << ", plus " << C.ImmCost << " imm cost";
1450 if (C.SetupCost != 0)
1451 OS << ", plus " << C.SetupCost << " setup cost";
1452}
1453
1454LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1455 print(errs()); errs() << '\n';
1456}
1457#endif
1458
1459/// Test whether this fixup always uses its value outside of the given loop.
1460bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1461 // PHI nodes use their value in their incoming blocks.
1462 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1463 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1464 if (PN->getIncomingValue(i) == OperandValToReplace &&
1465 L->contains(PN->getIncomingBlock(i)))
1466 return false;
1467 return true;
1468 }
1469
1470 return !L->contains(UserInst);
1471}
1472
1473#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1474void LSRFixup::print(raw_ostream &OS) const {
1475 OS << "UserInst=";
1476 // Store is common and interesting enough to be worth special-casing.
1477 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1478 OS << "store ";
1479 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1480 } else if (UserInst->getType()->isVoidTy())
1481 OS << UserInst->getOpcodeName();
1482 else
1483 UserInst->printAsOperand(OS, /*PrintType=*/false);
1484
1485 OS << ", OperandValToReplace=";
1486 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1487
1488 for (const Loop *PIL : PostIncLoops) {
1489 OS << ", PostIncLoop=";
1490 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1491 }
1492
1493 if (Offset != 0)
1494 OS << ", Offset=" << Offset;
1495}
1496
1497LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1498 print(errs()); errs() << '\n';
1499}
1500#endif
1501
1502/// Test whether this use as a formula which has the same registers as the given
1503/// formula.
1504bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1505 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1506 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1507 // Unstable sort by host order ok, because this is only used for uniquifying.
1508 llvm::sort(Key);
1509 return Uniquifier.count(Key);
1510}
1511
1512/// The function returns a probability of selecting formula without Reg.
1513float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1514 unsigned FNum = 0;
1515 for (const Formula &F : Formulae)
1516 if (F.referencesReg(Reg))
1517 FNum++;
1518 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1519}
1520
1521/// If the given formula has not yet been inserted, add it to the list, and
1522/// return true. Return false otherwise. The formula must be in canonical form.
1523bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1524 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1524, __PRETTY_FUNCTION__))
;
1525
1526 if (!Formulae.empty() && RigidFormula)
1527 return false;
1528
1529 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1530 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1531 // Unstable sort by host order ok, because this is only used for uniquifying.
1532 llvm::sort(Key);
1533
1534 if (!Uniquifier.insert(Key).second)
1535 return false;
1536
1537 // Using a register to hold the value of 0 is not profitable.
1538 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1539, __PRETTY_FUNCTION__))
1539 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1539, __PRETTY_FUNCTION__))
;
1540#ifndef NDEBUG
1541 for (const SCEV *BaseReg : F.BaseRegs)
1542 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1542, __PRETTY_FUNCTION__))
;
1543#endif
1544
1545 // Add the formula to the list.
1546 Formulae.push_back(F);
1547
1548 // Record registers now being used by this use.
1549 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1550 if (F.ScaledReg)
1551 Regs.insert(F.ScaledReg);
1552
1553 return true;
1554}
1555
1556/// Remove the given formula from this use's list.
1557void LSRUse::DeleteFormula(Formula &F) {
1558 if (&F != &Formulae.back())
1559 std::swap(F, Formulae.back());
1560 Formulae.pop_back();
1561}
1562
1563/// Recompute the Regs field, and update RegUses.
1564void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1565 // Now that we've filtered out some formulae, recompute the Regs set.
1566 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1567 Regs.clear();
1568 for (const Formula &F : Formulae) {
1569 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1570 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1571 }
1572
1573 // Update the RegTracker.
1574 for (const SCEV *S : OldRegs)
1575 if (!Regs.count(S))
1576 RegUses.dropRegister(S, LUIdx);
1577}
1578
1579#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1580void LSRUse::print(raw_ostream &OS) const {
1581 OS << "LSR Use: Kind=";
1582 switch (Kind) {
1583 case Basic: OS << "Basic"; break;
1584 case Special: OS << "Special"; break;
1585 case ICmpZero: OS << "ICmpZero"; break;
1586 case Address:
1587 OS << "Address of ";
1588 if (AccessTy.MemTy->isPointerTy())
1589 OS << "pointer"; // the full pointer type could be really verbose
1590 else {
1591 OS << *AccessTy.MemTy;
1592 }
1593
1594 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1595 }
1596
1597 OS << ", Offsets={";
1598 bool NeedComma = false;
1599 for (const LSRFixup &Fixup : Fixups) {
1600 if (NeedComma) OS << ',';
1601 OS << Fixup.Offset;
1602 NeedComma = true;
1603 }
1604 OS << '}';
1605
1606 if (AllFixupsOutsideLoop)
1607 OS << ", all-fixups-outside-loop";
1608
1609 if (WidestFixupType)
1610 OS << ", widest fixup type: " << *WidestFixupType;
1611}
1612
1613LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1614 print(errs()); errs() << '\n';
1615}
1616#endif
1617
1618static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1619 LSRUse::KindType Kind, MemAccessTy AccessTy,
1620 GlobalValue *BaseGV, int64_t BaseOffset,
1621 bool HasBaseReg, int64_t Scale,
1622 Instruction *Fixup/*= nullptr*/) {
1623 switch (Kind) {
1624 case LSRUse::Address:
1625 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1626 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1627
1628 case LSRUse::ICmpZero:
1629 // There's not even a target hook for querying whether it would be legal to
1630 // fold a GV into an ICmp.
1631 if (BaseGV)
1632 return false;
1633
1634 // ICmp only has two operands; don't allow more than two non-trivial parts.
1635 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1636 return false;
1637
1638 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1639 // putting the scaled register in the other operand of the icmp.
1640 if (Scale != 0 && Scale != -1)
1641 return false;
1642
1643 // If we have low-level target information, ask the target if it can fold an
1644 // integer immediate on an icmp.
1645 if (BaseOffset != 0) {
1646 // We have one of:
1647 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1648 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1649 // Offs is the ICmp immediate.
1650 if (Scale == 0)
1651 // The cast does the right thing with
1652 // std::numeric_limits<int64_t>::min().
1653 BaseOffset = -(uint64_t)BaseOffset;
1654 return TTI.isLegalICmpImmediate(BaseOffset);
1655 }
1656
1657 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1658 return true;
1659
1660 case LSRUse::Basic:
1661 // Only handle single-register values.
1662 return !BaseGV && Scale == 0 && BaseOffset == 0;
1663
1664 case LSRUse::Special:
1665 // Special case Basic to handle -1 scales.
1666 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1667 }
1668
1669 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1669)
;
1670}
1671
1672static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1673 int64_t MinOffset, int64_t MaxOffset,
1674 LSRUse::KindType Kind, MemAccessTy AccessTy,
1675 GlobalValue *BaseGV, int64_t BaseOffset,
1676 bool HasBaseReg, int64_t Scale) {
1677 // Check for overflow.
1678 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1679 (MinOffset > 0))
1680 return false;
1681 MinOffset = (uint64_t)BaseOffset + MinOffset;
1682 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1683 (MaxOffset > 0))
1684 return false;
1685 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1686
1687 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1688 HasBaseReg, Scale) &&
1689 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1690 HasBaseReg, Scale);
1691}
1692
1693static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1694 int64_t MinOffset, int64_t MaxOffset,
1695 LSRUse::KindType Kind, MemAccessTy AccessTy,
1696 const Formula &F, const Loop &L) {
1697 // For the purpose of isAMCompletelyFolded either having a canonical formula
1698 // or a scale not equal to zero is correct.
1699 // Problems may arise from non canonical formulae having a scale == 0.
1700 // Strictly speaking it would best to just rely on canonical formulae.
1701 // However, when we generate the scaled formulae, we first check that the
1702 // scaling factor is profitable before computing the actual ScaledReg for
1703 // compile time sake.
1704 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1704, __PRETTY_FUNCTION__))
;
1705 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1706 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1707}
1708
1709/// Test whether we know how to expand the current formula.
1710static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1711 int64_t MaxOffset, LSRUse::KindType Kind,
1712 MemAccessTy AccessTy, GlobalValue *BaseGV,
1713 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1714 // We know how to expand completely foldable formulae.
1715 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1716 BaseOffset, HasBaseReg, Scale) ||
1717 // Or formulae that use a base register produced by a sum of base
1718 // registers.
1719 (Scale == 1 &&
1720 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1721 BaseGV, BaseOffset, true, 0));
1722}
1723
1724static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1725 int64_t MaxOffset, LSRUse::KindType Kind,
1726 MemAccessTy AccessTy, const Formula &F) {
1727 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1728 F.BaseOffset, F.HasBaseReg, F.Scale);
1729}
1730
1731static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1732 const LSRUse &LU, const Formula &F) {
1733 // Target may want to look at the user instructions.
1734 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1735 for (const LSRFixup &Fixup : LU.Fixups)
1736 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1737 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1738 F.Scale, Fixup.UserInst))
1739 return false;
1740 return true;
1741 }
1742
1743 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1744 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1745 F.Scale);
1746}
1747
1748static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1749 const LSRUse &LU, const Formula &F,
1750 const Loop &L) {
1751 if (!F.Scale)
1752 return 0;
1753
1754 // If the use is not completely folded in that instruction, we will have to
1755 // pay an extra cost only for scale != 1.
1756 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1757 LU.AccessTy, F, L))
1758 return F.Scale != 1;
1759
1760 switch (LU.Kind) {
1761 case LSRUse::Address: {
1762 // Check the scaling factor cost with both the min and max offsets.
1763 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1764 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1765 F.Scale, LU.AccessTy.AddrSpace);
1766 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1767 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1768 F.Scale, LU.AccessTy.AddrSpace);
1769
1770 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1771, __PRETTY_FUNCTION__))
1771 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1771, __PRETTY_FUNCTION__))
;
1772 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1773 }
1774 case LSRUse::ICmpZero:
1775 case LSRUse::Basic:
1776 case LSRUse::Special:
1777 // The use is completely folded, i.e., everything is folded into the
1778 // instruction.
1779 return 0;
1780 }
1781
1782 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1782)
;
1783}
1784
1785static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1786 LSRUse::KindType Kind, MemAccessTy AccessTy,
1787 GlobalValue *BaseGV, int64_t BaseOffset,
1788 bool HasBaseReg) {
1789 // Fast-path: zero is always foldable.
1790 if (BaseOffset == 0 && !BaseGV) return true;
1791
1792 // Conservatively, create an address with an immediate and a
1793 // base and a scale.
1794 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1795
1796 // Canonicalize a scale of 1 to a base register if the formula doesn't
1797 // already have a base register.
1798 if (!HasBaseReg && Scale == 1) {
1799 Scale = 0;
1800 HasBaseReg = true;
1801 }
1802
1803 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1804 HasBaseReg, Scale);
1805}
1806
1807static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1808 ScalarEvolution &SE, int64_t MinOffset,
1809 int64_t MaxOffset, LSRUse::KindType Kind,
1810 MemAccessTy AccessTy, const SCEV *S,
1811 bool HasBaseReg) {
1812 // Fast-path: zero is always foldable.
1813 if (S->isZero()) return true;
1814
1815 // Conservatively, create an address with an immediate and a
1816 // base and a scale.
1817 int64_t BaseOffset = ExtractImmediate(S, SE);
1818 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1819
1820 // If there's anything else involved, it's not foldable.
1821 if (!S->isZero()) return false;
1822
1823 // Fast-path: zero is always foldable.
1824 if (BaseOffset == 0 && !BaseGV) return true;
1825
1826 // Conservatively, create an address with an immediate and a
1827 // base and a scale.
1828 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1829
1830 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1831 BaseOffset, HasBaseReg, Scale);
1832}
1833
1834namespace {
1835
1836/// An individual increment in a Chain of IV increments. Relate an IV user to
1837/// an expression that computes the IV it uses from the IV used by the previous
1838/// link in the Chain.
1839///
1840/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1841/// original IVOperand. The head of the chain's IVOperand is only valid during
1842/// chain collection, before LSR replaces IV users. During chain generation,
1843/// IncExpr can be used to find the new IVOperand that computes the same
1844/// expression.
1845struct IVInc {
1846 Instruction *UserInst;
1847 Value* IVOperand;
1848 const SCEV *IncExpr;
1849
1850 IVInc(Instruction *U, Value *O, const SCEV *E)
1851 : UserInst(U), IVOperand(O), IncExpr(E) {}
1852};
1853
1854// The list of IV increments in program order. We typically add the head of a
1855// chain without finding subsequent links.
1856struct IVChain {
1857 SmallVector<IVInc, 1> Incs;
1858 const SCEV *ExprBase = nullptr;
1859
1860 IVChain() = default;
1861 IVChain(const IVInc &Head, const SCEV *Base)
1862 : Incs(1, Head), ExprBase(Base) {}
1863
1864 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1865
1866 // Return the first increment in the chain.
1867 const_iterator begin() const {
1868 assert(!Incs.empty())((!Incs.empty()) ? static_cast<void> (0) : __assert_fail
("!Incs.empty()", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1868, __PRETTY_FUNCTION__))
;
1869 return std::next(Incs.begin());
1870 }
1871 const_iterator end() const {
1872 return Incs.end();
1873 }
1874
1875 // Returns true if this chain contains any increments.
1876 bool hasIncs() const { return Incs.size() >= 2; }
1877
1878 // Add an IVInc to the end of this chain.
1879 void add(const IVInc &X) { Incs.push_back(X); }
1880
1881 // Returns the last UserInst in the chain.
1882 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1883
1884 // Returns true if IncExpr can be profitably added to this chain.
1885 bool isProfitableIncrement(const SCEV *OperExpr,
1886 const SCEV *IncExpr,
1887 ScalarEvolution&);
1888};
1889
1890/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1891/// between FarUsers that definitely cross IV increments and NearUsers that may
1892/// be used between IV increments.
1893struct ChainUsers {
1894 SmallPtrSet<Instruction*, 4> FarUsers;
1895 SmallPtrSet<Instruction*, 4> NearUsers;
1896};
1897
1898/// This class holds state for the main loop strength reduction logic.
1899class LSRInstance {
1900 IVUsers &IU;
1901 ScalarEvolution &SE;
1902 DominatorTree &DT;
1903 LoopInfo &LI;
1904 AssumptionCache &AC;
1905 TargetLibraryInfo &TLI;
1906 const TargetTransformInfo &TTI;
1907 Loop *const L;
1908 MemorySSAUpdater *MSSAU;
1909 bool FavorBackedgeIndex = false;
1910 bool Changed = false;
1911
1912 /// This is the insert position that the current loop's induction variable
1913 /// increment should be placed. In simple loops, this is the latch block's
1914 /// terminator. But in more complicated cases, this is a position which will
1915 /// dominate all the in-loop post-increment users.
1916 Instruction *IVIncInsertPos = nullptr;
1917
1918 /// Interesting factors between use strides.
1919 ///
1920 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1921 /// default, a SmallDenseSet, because we need to use the full range of
1922 /// int64_ts, and there's currently no good way of doing that with
1923 /// SmallDenseSet.
1924 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1925
1926 /// Interesting use types, to facilitate truncation reuse.
1927 SmallSetVector<Type *, 4> Types;
1928
1929 /// The list of interesting uses.
1930 mutable SmallVector<LSRUse, 16> Uses;
1931
1932 /// Track which uses use which register candidates.
1933 RegUseTracker RegUses;
1934
1935 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1936 // have more than a few IV increment chains in a loop. Missing a Chain falls
1937 // back to normal LSR behavior for those uses.
1938 static const unsigned MaxChains = 8;
1939
1940 /// IV users can form a chain of IV increments.
1941 SmallVector<IVChain, MaxChains> IVChainVec;
1942
1943 /// IV users that belong to profitable IVChains.
1944 SmallPtrSet<Use*, MaxChains> IVIncSet;
1945
1946 void OptimizeShadowIV();
1947 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1948 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1949 void OptimizeLoopTermCond();
1950
1951 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1952 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1953 void FinalizeChain(IVChain &Chain);
1954 void CollectChains();
1955 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1956 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1957
1958 void CollectInterestingTypesAndFactors();
1959 void CollectFixupsAndInitialFormulae();
1960
1961 // Support for sharing of LSRUses between LSRFixups.
1962 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
1963 UseMapTy UseMap;
1964
1965 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1966 LSRUse::KindType Kind, MemAccessTy AccessTy);
1967
1968 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1969 MemAccessTy AccessTy);
1970
1971 void DeleteUse(LSRUse &LU, size_t LUIdx);
1972
1973 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1974
1975 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1976 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1977 void CountRegisters(const Formula &F, size_t LUIdx);
1978 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1979
1980 void CollectLoopInvariantFixupsAndFormulae();
1981
1982 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1983 unsigned Depth = 0);
1984
1985 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1986 const Formula &Base, unsigned Depth,
1987 size_t Idx, bool IsScaledReg = false);
1988 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1989 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1990 const Formula &Base, size_t Idx,
1991 bool IsScaledReg = false);
1992 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1993 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1994 const Formula &Base,
1995 const SmallVectorImpl<int64_t> &Worklist,
1996 size_t Idx, bool IsScaledReg = false);
1997 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1998 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1999 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2000 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2001 void GenerateCrossUseConstantOffsets();
2002 void GenerateAllReuseFormulae();
2003
2004 void FilterOutUndesirableDedicatedRegisters();
2005
2006 size_t EstimateSearchSpaceComplexity() const;
2007 void NarrowSearchSpaceByDetectingSupersets();
2008 void NarrowSearchSpaceByCollapsingUnrolledCode();
2009 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2010 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2011 void NarrowSearchSpaceByDeletingCostlyFormulas();
2012 void NarrowSearchSpaceByPickingWinnerRegs();
2013 void NarrowSearchSpaceUsingHeuristics();
2014
2015 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2016 Cost &SolutionCost,
2017 SmallVectorImpl<const Formula *> &Workspace,
2018 const Cost &CurCost,
2019 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2020 DenseSet<const SCEV *> &VisitedRegs) const;
2021 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2022
2023 BasicBlock::iterator
2024 HoistInsertPosition(BasicBlock::iterator IP,
2025 const SmallVectorImpl<Instruction *> &Inputs) const;
2026 BasicBlock::iterator
2027 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2028 const LSRFixup &LF,
2029 const LSRUse &LU,
2030 SCEVExpander &Rewriter) const;
2031
2032 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2033 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2034 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2035 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2036 const Formula &F, SCEVExpander &Rewriter,
2037 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2038 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2039 SCEVExpander &Rewriter,
2040 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2041 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2042
2043public:
2044 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2045 LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2046 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);
2047
2048 bool getChanged() const { return Changed; }
2049
2050 void print_factors_and_types(raw_ostream &OS) const;
2051 void print_fixups(raw_ostream &OS) const;
2052 void print_uses(raw_ostream &OS) const;
2053 void print(raw_ostream &OS) const;
2054 void dump() const;
2055};
2056
2057} // end anonymous namespace
2058
2059/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2060/// the cast operation.
2061void LSRInstance::OptimizeShadowIV() {
2062 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2063 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2064 return;
2065
2066 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2067 UI != E; /* empty */) {
2068 IVUsers::const_iterator CandidateUI = UI;
2069 ++UI;
2070 Instruction *ShadowUse = CandidateUI->getUser();
2071 Type *DestTy = nullptr;
2072 bool IsSigned = false;
2073
2074 /* If shadow use is a int->float cast then insert a second IV
2075 to eliminate this cast.
2076
2077 for (unsigned i = 0; i < n; ++i)
2078 foo((double)i);
2079
2080 is transformed into
2081
2082 double d = 0.0;
2083 for (unsigned i = 0; i < n; ++i, ++d)
2084 foo(d);
2085 */
2086 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2087 IsSigned = false;
2088 DestTy = UCast->getDestTy();
2089 }
2090 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2091 IsSigned = true;
2092 DestTy = SCast->getDestTy();
2093 }
2094 if (!DestTy) continue;
2095
2096 // If target does not support DestTy natively then do not apply
2097 // this transformation.
2098 if (!TTI.isTypeLegal(DestTy)) continue;
2099
2100 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2101 if (!PH) continue;
2102 if (PH->getNumIncomingValues() != 2) continue;
2103
2104 // If the calculation in integers overflows, the result in FP type will
2105 // differ. So we only can do this transformation if we are guaranteed to not
2106 // deal with overflowing values
2107 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2108 if (!AR) continue;
2109 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2110 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2111
2112 Type *SrcTy = PH->getType();
2113 int Mantissa = DestTy->getFPMantissaWidth();
2114 if (Mantissa == -1) continue;
2115 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2116 continue;
2117
2118 unsigned Entry, Latch;
2119 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2120 Entry = 0;
2121 Latch = 1;
2122 } else {
2123 Entry = 1;
2124 Latch = 0;
2125 }
2126
2127 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2128 if (!Init) continue;
2129 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2130 (double)Init->getSExtValue() :
2131 (double)Init->getZExtValue());
2132
2133 BinaryOperator *Incr =
2134 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2135 if (!Incr) continue;
2136 if (Incr->getOpcode() != Instruction::Add
2137 && Incr->getOpcode() != Instruction::Sub)
2138 continue;
2139
2140 /* Initialize new IV, double d = 0.0 in above example. */
2141 ConstantInt *C = nullptr;
2142 if (Incr->getOperand(0) == PH)
2143 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2144 else if (Incr->getOperand(1) == PH)
2145 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2146 else
2147 continue;
2148
2149 if (!C) continue;
2150
2151 // Ignore negative constants, as the code below doesn't handle them
2152 // correctly. TODO: Remove this restriction.
2153 if (!C->getValue().isStrictlyPositive()) continue;
2154
2155 /* Add new PHINode. */
2156 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2157
2158 /* create new increment. '++d' in above example. */
2159 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2160 BinaryOperator *NewIncr =
2161 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2162 Instruction::FAdd : Instruction::FSub,
2163 NewPH, CFP, "IV.S.next.", Incr);
2164
2165 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2166 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2167
2168 /* Remove cast operation */
2169 ShadowUse->replaceAllUsesWith(NewPH);
2170 ShadowUse->eraseFromParent();
2171 Changed = true;
2172 break;
2173 }
2174}
2175
2176/// If Cond has an operand that is an expression of an IV, set the IV user and
2177/// stride information and return true, otherwise return false.
2178bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2179 for (IVStrideUse &U : IU)
2180 if (U.getUser() == Cond) {
2181 // NOTE: we could handle setcc instructions with multiple uses here, but
2182 // InstCombine does it as well for simple uses, it's not clear that it
2183 // occurs enough in real life to handle.
2184 CondUse = &U;
2185 return true;
2186 }
2187 return false;
2188}
2189
2190/// Rewrite the loop's terminating condition if it uses a max computation.
2191///
2192/// This is a narrow solution to a specific, but acute, problem. For loops
2193/// like this:
2194///
2195/// i = 0;
2196/// do {
2197/// p[i] = 0.0;
2198/// } while (++i < n);
2199///
2200/// the trip count isn't just 'n', because 'n' might not be positive. And
2201/// unfortunately this can come up even for loops where the user didn't use
2202/// a C do-while loop. For example, seemingly well-behaved top-test loops
2203/// will commonly be lowered like this:
2204///
2205/// if (n > 0) {
2206/// i = 0;
2207/// do {
2208/// p[i] = 0.0;
2209/// } while (++i < n);
2210/// }
2211///
2212/// and then it's possible for subsequent optimization to obscure the if
2213/// test in such a way that indvars can't find it.
2214///
2215/// When indvars can't find the if test in loops like this, it creates a
2216/// max expression, which allows it to give the loop a canonical
2217/// induction variable:
2218///
2219/// i = 0;
2220/// max = n < 1 ? 1 : n;
2221/// do {
2222/// p[i] = 0.0;
2223/// } while (++i != max);
2224///
2225/// Canonical induction variables are necessary because the loop passes
2226/// are designed around them. The most obvious example of this is the
2227/// LoopInfo analysis, which doesn't remember trip count values. It
2228/// expects to be able to rediscover the trip count each time it is
2229/// needed, and it does this using a simple analysis that only succeeds if
2230/// the loop has a canonical induction variable.
2231///
2232/// However, when it comes time to generate code, the maximum operation
2233/// can be quite costly, especially if it's inside of an outer loop.
2234///
2235/// This function solves this problem by detecting this type of loop and
2236/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2237/// the instructions for the maximum computation.
2238ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2239 // Check that the loop matches the pattern we're looking for.
2240 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2241 Cond->getPredicate() != CmpInst::ICMP_NE)
2242 return Cond;
2243
2244 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2245 if (!Sel || !Sel->hasOneUse()) return Cond;
2246
2247 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2248 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2249 return Cond;
2250 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2251
2252 // Add one to the backedge-taken count to get the trip count.
2253 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2254 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2255
2256 // Check for a max calculation that matches the pattern. There's no check
2257 // for ICMP_ULE here because the comparison would be with zero, which
2258 // isn't interesting.
2259 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2260 const SCEVNAryExpr *Max = nullptr;
2261 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2262 Pred = ICmpInst::ICMP_SLE;
2263 Max = S;
2264 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2265 Pred = ICmpInst::ICMP_SLT;
2266 Max = S;
2267 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2268 Pred = ICmpInst::ICMP_ULT;
2269 Max = U;
2270 } else {
2271 // No match; bail.
2272 return Cond;
2273 }
2274
2275 // To handle a max with more than two operands, this optimization would
2276 // require additional checking and setup.
2277 if (Max->getNumOperands() != 2)
2278 return Cond;
2279
2280 const SCEV *MaxLHS = Max->getOperand(0);
2281 const SCEV *MaxRHS = Max->getOperand(1);
2282
2283 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2284 // for a comparison with 1. For <= and >=, a comparison with zero.
2285 if (!MaxLHS ||
2286 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2287 return Cond;
2288
2289 // Check the relevant induction variable for conformance to
2290 // the pattern.
2291 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2292 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2293 if (!AR || !AR->isAffine() ||
2294 AR->getStart() != One ||
2295 AR->getStepRecurrence(SE) != One)
2296 return Cond;
2297
2298 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2299, __PRETTY_FUNCTION__))
2299 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2299, __PRETTY_FUNCTION__))
;
2300
2301 // Check the right operand of the select, and remember it, as it will
2302 // be used in the new comparison instruction.
2303 Value *NewRHS = nullptr;
2304 if (ICmpInst::isTrueWhenEqual(Pred)) {
2305 // Look for n+1, and grab n.
2306 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2307 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2308 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2309 NewRHS = BO->getOperand(0);
2310 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2311 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2312 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2313 NewRHS = BO->getOperand(0);
2314 if (!NewRHS)
2315 return Cond;
2316 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2317 NewRHS = Sel->getOperand(1);
2318 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2319 NewRHS = Sel->getOperand(2);
2320 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2321 NewRHS = SU->getValue();
2322 else
2323 // Max doesn't match expected pattern.
2324 return Cond;
2325
2326 // Determine the new comparison opcode. It may be signed or unsigned,
2327 // and the original comparison may be either equality or inequality.
2328 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2329 Pred = CmpInst::getInversePredicate(Pred);
2330
2331 // Ok, everything looks ok to change the condition into an SLT or SGE and
2332 // delete the max calculation.
2333 ICmpInst *NewCond =
2334 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2335
2336 // Delete the max calculation instructions.
2337 Cond->replaceAllUsesWith(NewCond);
2338 CondUse->setUser(NewCond);
2339 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2340 Cond->eraseFromParent();
2341 Sel->eraseFromParent();
2342 if (Cmp->use_empty())
2343 Cmp->eraseFromParent();
2344 return NewCond;
2345}
2346
2347/// Change loop terminating condition to use the postinc iv when possible.
2348void
2349LSRInstance::OptimizeLoopTermCond() {
2350 SmallPtrSet<Instruction *, 4> PostIncs;
2351
2352 // We need a different set of heuristics for rotated and non-rotated loops.
2353 // If a loop is rotated then the latch is also the backedge, so inserting
2354 // post-inc expressions just before the latch is ideal. To reduce live ranges
2355 // it also makes sense to rewrite terminating conditions to use post-inc
2356 // expressions.
2357 //
2358 // If the loop is not rotated then the latch is not a backedge; the latch
2359 // check is done in the loop head. Adding post-inc expressions before the
2360 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2361 // in the loop body. In this case we do *not* want to use post-inc expressions
2362 // in the latch check, and we want to insert post-inc expressions before
2363 // the backedge.
2364 BasicBlock *LatchBlock = L->getLoopLatch();
2365 SmallVector<BasicBlock*, 8> ExitingBlocks;
2366 L->getExitingBlocks(ExitingBlocks);
2367 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2368 return LatchBlock != BB;
2369 })) {
2370 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2371 IVIncInsertPos = LatchBlock->getTerminator();
2372 return;
2373 }
2374
2375 // Otherwise treat this as a rotated loop.
2376 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2377 // Get the terminating condition for the loop if possible. If we
2378 // can, we want to change it to use a post-incremented version of its
2379 // induction variable, to allow coalescing the live ranges for the IV into
2380 // one register value.
2381
2382 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2383 if (!TermBr)
2384 continue;
2385 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2386 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2387 continue;
2388
2389 // Search IVUsesByStride to find Cond's IVUse if there is one.
2390 IVStrideUse *CondUse = nullptr;
2391 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2392 if (!FindIVUserForCond(Cond, CondUse))
2393 continue;
2394
2395 // If the trip count is computed in terms of a max (due to ScalarEvolution
2396 // being unable to find a sufficient guard, for example), change the loop
2397 // comparison to use SLT or ULT instead of NE.
2398 // One consequence of doing this now is that it disrupts the count-down
2399 // optimization. That's not always a bad thing though, because in such
2400 // cases it may still be worthwhile to avoid a max.
2401 Cond = OptimizeMax(Cond, CondUse);
2402
2403 // If this exiting block dominates the latch block, it may also use
2404 // the post-inc value if it won't be shared with other uses.
2405 // Check for dominance.
2406 if (!DT.dominates(ExitingBlock, LatchBlock))
2407 continue;
2408
2409 // Conservatively avoid trying to use the post-inc value in non-latch
2410 // exits if there may be pre-inc users in intervening blocks.
2411 if (LatchBlock != ExitingBlock)
2412 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2413 // Test if the use is reachable from the exiting block. This dominator
2414 // query is a conservative approximation of reachability.
2415 if (&*UI != CondUse &&
2416 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2417 // Conservatively assume there may be reuse if the quotient of their
2418 // strides could be a legal scale.
2419 const SCEV *A = IU.getStride(*CondUse, L);
2420 const SCEV *B = IU.getStride(*UI, L);
2421 if (!A || !B) continue;
2422 if (SE.getTypeSizeInBits(A->getType()) !=
2423 SE.getTypeSizeInBits(B->getType())) {
2424 if (SE.getTypeSizeInBits(A->getType()) >
2425 SE.getTypeSizeInBits(B->getType()))
2426 B = SE.getSignExtendExpr(B, A->getType());
2427 else
2428 A = SE.getSignExtendExpr(A, B->getType());
2429 }
2430 if (const SCEVConstant *D =
2431 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2432 const ConstantInt *C = D->getValue();
2433 // Stride of one or negative one can have reuse with non-addresses.
2434 if (C->isOne() || C->isMinusOne())
2435 goto decline_post_inc;
2436 // Avoid weird situations.
2437 if (C->getValue().getMinSignedBits() >= 64 ||
2438 C->getValue().isMinSignedValue())
2439 goto decline_post_inc;
2440 // Check for possible scaled-address reuse.
2441 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2442 MemAccessTy AccessTy = getAccessType(
2443 TTI, UI->getUser(), UI->getOperandValToReplace());
2444 int64_t Scale = C->getSExtValue();
2445 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2446 /*BaseOffset=*/0,
2447 /*HasBaseReg=*/false, Scale,
2448 AccessTy.AddrSpace))
2449 goto decline_post_inc;
2450 Scale = -Scale;
2451 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2452 /*BaseOffset=*/0,
2453 /*HasBaseReg=*/false, Scale,
2454 AccessTy.AddrSpace))
2455 goto decline_post_inc;
2456 }
2457 }
2458 }
2459
2460 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)
2461 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2462
2463 // It's possible for the setcc instruction to be anywhere in the loop, and
2464 // possible for it to have multiple users. If it is not immediately before
2465 // the exiting block branch, move it.
2466 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2467 if (Cond->hasOneUse()) {
2468 Cond->moveBefore(TermBr);
2469 } else {
2470 // Clone the terminating condition and insert into the loopend.
2471 ICmpInst *OldCond = Cond;
2472 Cond = cast<ICmpInst>(Cond->clone());
2473 Cond->setName(L->getHeader()->getName() + ".termcond");
2474 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2475
2476 // Clone the IVUse, as the old use still exists!
2477 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2478 TermBr->replaceUsesOfWith(OldCond, Cond);
2479 }
2480 }
2481
2482 // If we get to here, we know that we can transform the setcc instruction to
2483 // use the post-incremented version of the IV, allowing us to coalesce the
2484 // live ranges for the IV correctly.
2485 CondUse->transformToPostInc(L);
2486 Changed = true;
2487
2488 PostIncs.insert(Cond);
2489 decline_post_inc:;
2490 }
2491
2492 // Determine an insertion point for the loop induction variable increment. It
2493 // must dominate all the post-inc comparisons we just set up, and it must
2494 // dominate the loop latch edge.
2495 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2496 for (Instruction *Inst : PostIncs) {
2497 BasicBlock *BB =
2498 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2499 Inst->getParent());
2500 if (BB == Inst->getParent())
2501 IVIncInsertPos = Inst;
2502 else if (BB != IVIncInsertPos->getParent())
2503 IVIncInsertPos = BB->getTerminator();
2504 }
2505}
2506
2507/// Determine if the given use can accommodate a fixup at the given offset and
2508/// other details. If so, update the use and return true.
2509bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2510 bool HasBaseReg, LSRUse::KindType Kind,
2511 MemAccessTy AccessTy) {
2512 int64_t NewMinOffset = LU.MinOffset;
2513 int64_t NewMaxOffset = LU.MaxOffset;
2514 MemAccessTy NewAccessTy = AccessTy;
2515
2516 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2517 // something conservative, however this can pessimize in the case that one of
2518 // the uses will have all its uses outside the loop, for example.
2519 if (LU.Kind != Kind)
2520 return false;
2521
2522 // Check for a mismatched access type, and fall back conservatively as needed.
2523 // TODO: Be less conservative when the type is similar and can use the same
2524 // addressing modes.
2525 if (Kind == LSRUse::Address) {
2526 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2527 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2528 AccessTy.AddrSpace);
2529 }
2530 }
2531
2532 // Conservatively assume HasBaseReg is true for now.
2533 if (NewOffset < LU.MinOffset) {
2534 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2535 LU.MaxOffset - NewOffset, HasBaseReg))
2536 return false;
2537 NewMinOffset = NewOffset;
2538 } else if (NewOffset > LU.MaxOffset) {
2539 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2540 NewOffset - LU.MinOffset, HasBaseReg))
2541 return false;
2542 NewMaxOffset = NewOffset;
2543 }
2544
2545 // Update the use.
2546 LU.MinOffset = NewMinOffset;
2547 LU.MaxOffset = NewMaxOffset;
2548 LU.AccessTy = NewAccessTy;
2549 return true;
2550}
2551
2552/// Return an LSRUse index and an offset value for a fixup which needs the given
2553/// expression, with the given kind and optional access type. Either reuse an
2554/// existing use or create a new one, as needed.
2555std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2556 LSRUse::KindType Kind,
2557 MemAccessTy AccessTy) {
2558 const SCEV *Copy = Expr;
2559 int64_t Offset = ExtractImmediate(Expr, SE);
2560
2561 // Basic uses can't accept any offset, for example.
2562 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2563 Offset, /*HasBaseReg=*/ true)) {
2564 Expr = Copy;
2565 Offset = 0;
2566 }
2567
2568 std::pair<UseMapTy::iterator, bool> P =
2569 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2570 if (!P.second) {
2571 // A use already existed with this base.
2572 size_t LUIdx = P.first->second;
2573 LSRUse &LU = Uses[LUIdx];
2574 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2575 // Reuse this use.
2576 return std::make_pair(LUIdx, Offset);
2577 }
2578
2579 // Create a new use.
2580 size_t LUIdx = Uses.size();
2581 P.first->second = LUIdx;
2582 Uses.push_back(LSRUse(Kind, AccessTy));
2583 LSRUse &LU = Uses[LUIdx];
2584
2585 LU.MinOffset = Offset;
2586 LU.MaxOffset = Offset;
2587 return std::make_pair(LUIdx, Offset);
2588}
2589
2590/// Delete the given use from the Uses list.
2591void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2592 if (&LU != &Uses.back())
2593 std::swap(LU, Uses.back());
2594 Uses.pop_back();
2595
2596 // Update RegUses.
2597 RegUses.swapAndDropUse(LUIdx, Uses.size());
2598}
2599
2600/// Look for a use distinct from OrigLU which is has a formula that has the same
2601/// registers as the given formula.
2602LSRUse *
2603LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2604 const LSRUse &OrigLU) {
2605 // Search all uses for the formula. This could be more clever.
2606 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2607 LSRUse &LU = Uses[LUIdx];
2608 // Check whether this use is close enough to OrigLU, to see whether it's
2609 // worthwhile looking through its formulae.
2610 // Ignore ICmpZero uses because they may contain formulae generated by
2611 // GenerateICmpZeroScales, in which case adding fixup offsets may
2612 // be invalid.
2613 if (&LU != &OrigLU &&
2614 LU.Kind != LSRUse::ICmpZero &&
2615 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2616 LU.WidestFixupType == OrigLU.WidestFixupType &&
2617 LU.HasFormulaWithSameRegs(OrigF)) {
2618 // Scan through this use's formulae.
2619 for (const Formula &F : LU.Formulae) {
2620 // Check to see if this formula has the same registers and symbols
2621 // as OrigF.
2622 if (F.BaseRegs == OrigF.BaseRegs &&
2623 F.ScaledReg == OrigF.ScaledReg &&
2624 F.BaseGV == OrigF.BaseGV &&
2625 F.Scale == OrigF.Scale &&
2626 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2627 if (F.BaseOffset == 0)
2628 return &LU;
2629 // This is the formula where all the registers and symbols matched;
2630 // there aren't going to be any others. Since we declined it, we
2631 // can skip the rest of the formulae and proceed to the next LSRUse.
2632 break;
2633 }
2634 }
2635 }
2636 }
2637
2638 // Nothing looked good.
2639 return nullptr;
2640}
2641
2642void LSRInstance::CollectInterestingTypesAndFactors() {
2643 SmallSetVector<const SCEV *, 4> Strides;
2644
2645 // Collect interesting types and strides.
2646 SmallVector<const SCEV *, 4> Worklist;
2647 for (const IVStrideUse &U : IU) {
2648 const SCEV *Expr = IU.getExpr(U);
2649
2650 // Collect interesting types.
2651 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2652
2653 // Add strides for mentioned loops.
2654 Worklist.push_back(Expr);
2655 do {
2656 const SCEV *S = Worklist.pop_back_val();
2657 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2658 if (AR->getLoop() == L)
2659 Strides.insert(AR->getStepRecurrence(SE));
2660 Worklist.push_back(AR->getStart());
2661 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2662 Worklist.append(Add->op_begin(), Add->op_end());
2663 }
2664 } while (!Worklist.empty());
2665 }
2666
2667 // Compute interesting factors from the set of interesting strides.
2668 for (SmallSetVector<const SCEV *, 4>::const_iterator
2669 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2670 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2671 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2672 const SCEV *OldStride = *I;
2673 const SCEV *NewStride = *NewStrideIter;
2674
2675 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2676 SE.getTypeSizeInBits(NewStride->getType())) {
2677 if (SE.getTypeSizeInBits(OldStride->getType()) >
2678 SE.getTypeSizeInBits(NewStride->getType()))
2679 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2680 else
2681 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2682 }
2683 if (const SCEVConstant *Factor =
2684 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2685 SE, true))) {
2686 if (Factor->getAPInt().getMinSignedBits() <= 64)
2687 Factors.insert(Factor->getAPInt().getSExtValue());
2688 } else if (const SCEVConstant *Factor =
2689 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2690 NewStride,
2691 SE, true))) {
2692 if (Factor->getAPInt().getMinSignedBits() <= 64)
2693 Factors.insert(Factor->getAPInt().getSExtValue());
2694 }
2695 }
2696
2697 // If all uses use the same type, don't bother looking for truncation-based
2698 // reuse.
2699 if (Types.size() == 1)
2700 Types.clear();
2701
2702 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2703}
2704
2705/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2706/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2707/// IVStrideUses, we could partially skip this.
2708static User::op_iterator
2709findIVOperand(User::op_iterator OI, User::op_iterator OE,
2710 Loop *L, ScalarEvolution &SE) {
2711 for(; OI != OE; ++OI) {
2712 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2713 if (!SE.isSCEVable(Oper->getType()))
2714 continue;
2715
2716 if (const SCEVAddRecExpr *AR =
2717 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2718 if (AR->getLoop() == L)
2719 break;
2720 }
2721 }
2722 }
2723 return OI;
2724}
2725
2726/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2727/// a convenient helper.
2728static Value *getWideOperand(Value *Oper) {
2729 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2730 return Trunc->getOperand(0);
2731 return Oper;
2732}
2733
2734/// Return true if we allow an IV chain to include both types.
2735static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2736 Type *LType = LVal->getType();
2737 Type *RType = RVal->getType();
2738 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2739 // Different address spaces means (possibly)
2740 // different types of the pointer implementation,
2741 // e.g. i16 vs i32 so disallow that.
2742 (LType->getPointerAddressSpace() ==
2743 RType->getPointerAddressSpace()));
2744}
2745
2746/// Return an approximation of this SCEV expression's "base", or NULL for any
2747/// constant. Returning the expression itself is conservative. Returning a
2748/// deeper subexpression is more precise and valid as long as it isn't less
2749/// complex than another subexpression. For expressions involving multiple
2750/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2751/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2752/// IVInc==b-a.
2753///
2754/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2755/// SCEVUnknown, we simply return the rightmost SCEV operand.
2756static const SCEV *getExprBase(const SCEV *S) {
2757 switch (S->getSCEVType()) {
2758 default: // uncluding scUnknown.
2759 return S;
2760 case scConstant:
2761 return nullptr;
2762 case scTruncate:
2763 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2764 case scZeroExtend:
2765 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2766 case scSignExtend:
2767 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2768 case scAddExpr: {
2769 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2770 // there's nothing more complex.
2771 // FIXME: not sure if we want to recognize negation.
2772 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2773 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2774 E(Add->op_begin()); I != E; ++I) {
2775 const SCEV *SubExpr = *I;
2776 if (SubExpr->getSCEVType() == scAddExpr)
2777 return getExprBase(SubExpr);
2778
2779 if (SubExpr->getSCEVType() != scMulExpr)
2780 return SubExpr;
2781 }
2782 return S; // all operands are scaled, be conservative.
2783 }
2784 case scAddRecExpr:
2785 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2786 }
2787}
2788
2789/// Return true if the chain increment is profitable to expand into a loop
2790/// invariant value, which may require its own register. A profitable chain
2791/// increment will be an offset relative to the same base. We allow such offsets
2792/// to potentially be used as chain increment as long as it's not obviously
2793/// expensive to expand using real instructions.
2794bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2795 const SCEV *IncExpr,
2796 ScalarEvolution &SE) {
2797 // Aggressively form chains when -stress-ivchain.
2798 if (StressIVChain)
2799 return true;
2800
2801 // Do not replace a constant offset from IV head with a nonconstant IV
2802 // increment.
2803 if (!isa<SCEVConstant>(IncExpr)) {
2804 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2805 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2806 return false;
2807 }
2808
2809 SmallPtrSet<const SCEV*, 8> Processed;
2810 return !isHighCostExpansion(IncExpr, Processed, SE);
2811}
2812
2813/// Return true if the number of registers needed for the chain is estimated to
2814/// be less than the number required for the individual IV users. First prohibit
2815/// any IV users that keep the IV live across increments (the Users set should
2816/// be empty). Next count the number and type of increments in the chain.
2817///
2818/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2819/// effectively use postinc addressing modes. Only consider it profitable it the
2820/// increments can be computed in fewer registers when chained.
2821///
2822/// TODO: Consider IVInc free if it's already used in another chains.
2823static bool
2824isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2825 ScalarEvolution &SE) {
2826 if (StressIVChain)
2827 return true;
2828
2829 if (!Chain.hasIncs())
2830 return false;
2831
2832 if (!Users.empty()) {
2833 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)
2834 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)
2835 : 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)
;
2836 return false;
2837 }
2838 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2838, __PRETTY_FUNCTION__))
;
2839
2840 // The chain itself may require a register, so intialize cost to 1.
2841 int cost = 1;
2842
2843 // A complete chain likely eliminates the need for keeping the original IV in
2844 // a register. LSR does not currently know how to form a complete chain unless
2845 // the header phi already exists.
2846 if (isa<PHINode>(Chain.tailUserInst())
2847 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2848 --cost;
2849 }
2850 const SCEV *LastIncExpr = nullptr;
2851 unsigned NumConstIncrements = 0;
2852 unsigned NumVarIncrements = 0;
2853 unsigned NumReusedIncrements = 0;
2854 for (const IVInc &Inc : Chain) {
2855 if (Inc.IncExpr->isZero())
2856 continue;
2857
2858 // Incrementing by zero or some constant is neutral. We assume constants can
2859 // be folded into an addressing mode or an add's immediate operand.
2860 if (isa<SCEVConstant>(Inc.IncExpr)) {
2861 ++NumConstIncrements;
2862 continue;
2863 }
2864
2865 if (Inc.IncExpr == LastIncExpr)
2866 ++NumReusedIncrements;
2867 else
2868 ++NumVarIncrements;
2869
2870 LastIncExpr = Inc.IncExpr;
2871 }
2872 // An IV chain with a single increment is handled by LSR's postinc
2873 // uses. However, a chain with multiple increments requires keeping the IV's
2874 // value live longer than it needs to be if chained.
2875 if (NumConstIncrements > 1)
2876 --cost;
2877
2878 // Materializing increment expressions in the preheader that didn't exist in
2879 // the original code may cost a register. For example, sign-extended array
2880 // indices can produce ridiculous increments like this:
2881 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2882 cost += NumVarIncrements;
2883
2884 // Reusing variable increments likely saves a register to hold the multiple of
2885 // the stride.
2886 cost -= NumReusedIncrements;
2887
2888 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)
2889 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2890
2891 return cost < 0;
2892}
2893
2894/// Add this IV user to an existing chain or make it the head of a new chain.
2895void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2896 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2897 // When IVs are used as types of varying widths, they are generally converted
2898 // to a wider type with some uses remaining narrow under a (free) trunc.
2899 Value *const NextIV = getWideOperand(IVOper);
2900 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2901 const SCEV *const OperExprBase = getExprBase(OperExpr);
2902
2903 // Visit all existing chains. Check if its IVOper can be computed as a
2904 // profitable loop invariant increment from the last link in the Chain.
2905 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2906 const SCEV *LastIncExpr = nullptr;
2907 for (; ChainIdx < NChains; ++ChainIdx) {
2908 IVChain &Chain = IVChainVec[ChainIdx];
2909
2910 // Prune the solution space aggressively by checking that both IV operands
2911 // are expressions that operate on the same unscaled SCEVUnknown. This
2912 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2913 // first avoids creating extra SCEV expressions.
2914 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2915 continue;
2916
2917 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2918 if (!isCompatibleIVType(PrevIV, NextIV))
2919 continue;
2920
2921 // A phi node terminates a chain.
2922 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2923 continue;
2924
2925 // The increment must be loop-invariant so it can be kept in a register.
2926 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2927 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2928 if (!SE.isLoopInvariant(IncExpr, L))
2929 continue;
2930
2931 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2932 LastIncExpr = IncExpr;
2933 break;
2934 }
2935 }
2936 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2937 // bother for phi nodes, because they must be last in the chain.
2938 if (ChainIdx == NChains) {
2939 if (isa<PHINode>(UserInst))
2940 return;
2941 if (NChains >= MaxChains && !StressIVChain) {
2942 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2943 return;
2944 }
2945 LastIncExpr = OperExpr;
2946 // IVUsers may have skipped over sign/zero extensions. We don't currently
2947 // attempt to form chains involving extensions unless they can be hoisted
2948 // into this loop's AddRec.
2949 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2950 return;
2951 ++NChains;
2952 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2953 OperExprBase));
2954 ChainUsersVec.resize(NChains);
2955 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)
2956 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
2957 } else {
2958 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)
2959 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
2960 // Add this IV user to the end of the chain.
2961 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2962 }
2963 IVChain &Chain = IVChainVec[ChainIdx];
2964
2965 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2966 // This chain's NearUsers become FarUsers.
2967 if (!LastIncExpr->isZero()) {
2968 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2969 NearUsers.end());
2970 NearUsers.clear();
2971 }
2972
2973 // All other uses of IVOperand become near uses of the chain.
2974 // We currently ignore intermediate values within SCEV expressions, assuming
2975 // they will eventually be used be the current chain, or can be computed
2976 // from one of the chain increments. To be more precise we could
2977 // transitively follow its user and only add leaf IV users to the set.
2978 for (User *U : IVOper->users()) {
2979 Instruction *OtherUse = dyn_cast<Instruction>(U);
2980 if (!OtherUse)
2981 continue;
2982 // Uses in the chain will no longer be uses if the chain is formed.
2983 // Include the head of the chain in this iteration (not Chain.begin()).
2984 IVChain::const_iterator IncIter = Chain.Incs.begin();
2985 IVChain::const_iterator IncEnd = Chain.Incs.end();
2986 for( ; IncIter != IncEnd; ++IncIter) {
2987 if (IncIter->UserInst == OtherUse)
2988 break;
2989 }
2990 if (IncIter != IncEnd)
2991 continue;
2992
2993 if (SE.isSCEVable(OtherUse->getType())
2994 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2995 && IU.isIVUserOrOperand(OtherUse)) {
2996 continue;
2997 }
2998 NearUsers.insert(OtherUse);
2999 }
3000
3001 // Since this user is part of the chain, it's no longer considered a use
3002 // of the chain.
3003 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3004}
3005
3006/// Populate the vector of Chains.
3007///
3008/// This decreases ILP at the architecture level. Targets with ample registers,
3009/// multiple memory ports, and no register renaming probably don't want
3010/// this. However, such targets should probably disable LSR altogether.
3011///
3012/// The job of LSR is to make a reasonable choice of induction variables across
3013/// the loop. Subsequent passes can easily "unchain" computation exposing more
3014/// ILP *within the loop* if the target wants it.
3015///
3016/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3017/// will not reorder memory operations, it will recognize this as a chain, but
3018/// will generate redundant IV increments. Ideally this would be corrected later
3019/// by a smart scheduler:
3020/// = A[i]
3021/// = A[i+x]
3022/// A[i] =
3023/// A[i+x] =
3024///
3025/// TODO: Walk the entire domtree within this loop, not just the path to the
3026/// loop latch. This will discover chains on side paths, but requires
3027/// maintaining multiple copies of the Chains state.
3028void LSRInstance::CollectChains() {
3029 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3030 SmallVector<ChainUsers, 8> ChainUsersVec;
3031
3032 SmallVector<BasicBlock *,8> LatchPath;
3033 BasicBlock *LoopHeader = L->getHeader();
3034 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3035 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3036 LatchPath.push_back(Rung->getBlock());
3037 }
3038 LatchPath.push_back(LoopHeader);
3039
3040 // Walk the instruction stream from the loop header to the loop latch.
3041 for (BasicBlock *BB : reverse(LatchPath)) {
3042 for (Instruction &I : *BB) {
3043 // Skip instructions that weren't seen by IVUsers analysis.
3044 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3045 continue;
3046
3047 // Ignore users that are part of a SCEV expression. This way we only
3048 // consider leaf IV Users. This effectively rediscovers a portion of
3049 // IVUsers analysis but in program order this time.
3050 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3051 continue;
3052
3053 // Remove this instruction from any NearUsers set it may be in.
3054 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3055 ChainIdx < NChains; ++ChainIdx) {
3056 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3057 }
3058 // Search for operands that can be chained.
3059 SmallPtrSet<Instruction*, 4> UniqueOperands;
3060 User::op_iterator IVOpEnd = I.op_end();
3061 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3062 while (IVOpIter != IVOpEnd) {
3063 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3064 if (UniqueOperands.insert(IVOpInst).second)
3065 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3066 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3067 }
3068 } // Continue walking down the instructions.
3069 } // Continue walking down the domtree.
3070 // Visit phi backedges to determine if the chain can generate the IV postinc.
3071 for (PHINode &PN : L->getHeader()->phis()) {
3072 if (!SE.isSCEVable(PN.getType()))
3073 continue;
3074
3075 Instruction *IncV =
3076 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3077 if (IncV)
3078 ChainInstruction(&PN, IncV, ChainUsersVec);
3079 }
3080 // Remove any unprofitable chains.
3081 unsigned ChainIdx = 0;
3082 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3083 UsersIdx < NChains; ++UsersIdx) {
3084 if (!isProfitableChain(IVChainVec[UsersIdx],
3085 ChainUsersVec[UsersIdx].FarUsers, SE))
3086 continue;
3087 // Preserve the chain at UsesIdx.
3088 if (ChainIdx != UsersIdx)
3089 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3090 FinalizeChain(IVChainVec[ChainIdx]);
3091 ++ChainIdx;
3092 }
3093 IVChainVec.resize(ChainIdx);
3094}
3095
3096void LSRInstance::FinalizeChain(IVChain &Chain) {
3097 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3097, __PRETTY_FUNCTION__))
;
3098 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)
;
3099
3100 for (const IVInc &Inc : Chain) {
3101 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3102 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3103 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3103, __PRETTY_FUNCTION__))
;
3104 IVIncSet.insert(UseI);
3105 }
3106}
3107
3108/// Return true if the IVInc can be folded into an addressing mode.
3109static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3110 Value *Operand, const TargetTransformInfo &TTI) {
3111 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3112 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3113 return false;
3114
3115 if (IncConst->getAPInt().getMinSignedBits() > 64)
3116 return false;
3117
3118 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3119 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3120 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3121 IncOffset, /*HasBaseReg=*/false))
3122 return false;
3123
3124 return true;
3125}
3126
3127/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3128/// user's operand from the previous IV user's operand.
3129void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3130 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3131 // Find the new IVOperand for the head of the chain. It may have been replaced
3132 // by LSR.
3133 const IVInc &Head = Chain.Incs[0];
3134 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3135 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3136 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3137 IVOpEnd, L, SE);
3138 Value *IVSrc = nullptr;
3139 while (IVOpIter != IVOpEnd) {
3140 IVSrc = getWideOperand(*IVOpIter);
3141
3142 // If this operand computes the expression that the chain needs, we may use
3143 // it. (Check this after setting IVSrc which is used below.)
3144 //
3145 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3146 // narrow for the chain, so we can no longer use it. We do allow using a
3147 // wider phi, assuming the LSR checked for free truncation. In that case we
3148 // should already have a truncate on this operand such that
3149 // getSCEV(IVSrc) == IncExpr.
3150 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3151 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3152 break;
3153 }
3154 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3155 }
3156 if (IVOpIter == IVOpEnd) {
3157 // Gracefully give up on this chain.
3158 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)
;
3159 return;
3160 }
3161 assert(IVSrc && "Failed to find IV chain source")((IVSrc && "Failed to find IV chain source") ? static_cast
<void> (0) : __assert_fail ("IVSrc && \"Failed to find IV chain source\""
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3161, __PRETTY_FUNCTION__))
;
3162
3163 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3164 Type *IVTy = IVSrc->getType();
3165 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3166 const SCEV *LeftOverExpr = nullptr;
3167 for (const IVInc &Inc : Chain) {
3168 Instruction *InsertPt = Inc.UserInst;
3169 if (isa<PHINode>(InsertPt))
3170 InsertPt = L->getLoopLatch()->getTerminator();
3171
3172 // IVOper will replace the current IV User's operand. IVSrc is the IV
3173 // value currently held in a register.
3174 Value *IVOper = IVSrc;
3175 if (!Inc.IncExpr->isZero()) {
3176 // IncExpr was the result of subtraction of two narrow values, so must
3177 // be signed.
3178 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3179 LeftOverExpr = LeftOverExpr ?
3180 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3181 }
3182 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3183 // Expand the IV increment.
3184 Rewriter.clearPostInc();
3185 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3186 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3187 SE.getUnknown(IncV));
3188 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3189
3190 // If an IV increment can't be folded, use it as the next IV value.
3191 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3192 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3192, __PRETTY_FUNCTION__))
;
3193 IVSrc = IVOper;
3194 LeftOverExpr = nullptr;
3195 }
3196 }
3197 Type *OperTy = Inc.IVOperand->getType();
3198 if (IVTy != OperTy) {
3199 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3200, __PRETTY_FUNCTION__))
3200 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3200, __PRETTY_FUNCTION__))
;
3201 IRBuilder<> Builder(InsertPt);
3202 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3203 }
3204 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3205 if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3206 DeadInsts.emplace_back(OperandIsInstr);
3207 }
3208 // If LSR created a new, wider phi, we may also replace its postinc. We only
3209 // do this if we also found a wide value for the head of the chain.
3210 if (isa<PHINode>(Chain.tailUserInst())) {
3211 for (PHINode &Phi : L->getHeader()->phis()) {
3212 if (!isCompatibleIVType(&Phi, IVSrc))
3213 continue;
3214 Instruction *PostIncV = dyn_cast<Instruction>(
3215 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3216 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3217 continue;
3218 Value *IVOper = IVSrc;
3219 Type *PostIncTy = PostIncV->getType();
3220 if (IVTy != PostIncTy) {
3221 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3221, __PRETTY_FUNCTION__))
;
3222 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3223 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3224 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3225 }
3226 Phi.replaceUsesOfWith(PostIncV, IVOper);
3227 DeadInsts.emplace_back(PostIncV);
3228 }
3229 }
3230}
3231
3232void LSRInstance::CollectFixupsAndInitialFormulae() {
3233 BranchInst *ExitBranch = nullptr;
3234 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3235
3236 for (const IVStrideUse &U : IU) {
3237 Instruction *UserInst = U.getUser();
3238 // Skip IV users that are part of profitable IV Chains.
3239 User::op_iterator UseI =
3240 find(UserInst->operands(), U.getOperandValToReplace());
3241 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3241, __PRETTY_FUNCTION__))
;
3242 if (IVIncSet.count(UseI)) {
3243 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)
;
3244 continue;
3245 }
3246
3247 LSRUse::KindType Kind = LSRUse::Basic;
3248 MemAccessTy AccessTy;
3249 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3250 Kind = LSRUse::Address;
3251 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3252 }
3253
3254 const SCEV *S = IU.getExpr(U);
3255 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3256
3257 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3258 // (N - i == 0), and this allows (N - i) to be the expression that we work
3259 // with rather than just N or i, so we can consider the register
3260 // requirements for both N and i at the same time. Limiting this code to
3261 // equality icmps is not a problem because all interesting loops use
3262 // equality icmps, thanks to IndVarSimplify.
3263 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3264 // If CI can be saved in some target, like replaced inside hardware loop
3265 // in PowerPC, no need to generate initial formulae for it.
3266 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3267 continue;
3268 if (CI->isEquality()) {
3269 // Swap the operands if needed to put the OperandValToReplace on the
3270 // left, for consistency.
3271 Value *NV = CI->getOperand(1);
3272 if (NV == U.getOperandValToReplace()) {
3273 CI->setOperand(1, CI->getOperand(0));
3274 CI->setOperand(0, NV);
3275 NV = CI->getOperand(1);
3276 Changed = true;
3277 }
3278
3279 // x == y --> x - y == 0
3280 const SCEV *N = SE.getSCEV(NV);
3281 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3282 // S is normalized, so normalize N before folding it into S
3283 // to keep the result normalized.
3284 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3285 Kind = LSRUse::ICmpZero;
3286 S = SE.getMinusSCEV(N, S);
3287 }
3288
3289 // -1 and the negations of all interesting strides (except the negation
3290 // of -1) are now also interesting.
3291 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3292 if (Factors[i] != -1)
3293 Factors.insert(-(uint64_t)Factors[i]);
3294 Factors.insert(-1);
3295 }
3296 }
3297
3298 // Get or create an LSRUse.
3299 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3300 size_t LUIdx = P.first;
3301 int64_t Offset = P.second;
3302 LSRUse &LU = Uses[LUIdx];
3303
3304 // Record the fixup.
3305 LSRFixup &LF = LU.getNewFixup();
3306 LF.UserInst = UserInst;
3307 LF.OperandValToReplace = U.getOperandValToReplace();
3308 LF.PostIncLoops = TmpPostIncLoops;
3309 LF.Offset = Offset;
3310 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3311
3312 if (!LU.WidestFixupType ||
3313 SE.getTypeSizeInBits(LU.WidestFixupType) <
3314 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3315 LU.WidestFixupType = LF.OperandValToReplace->getType();
3316
3317 // If this is the first use of this LSRUse, give it a formula.
3318 if (LU.Formulae.empty()) {
3319 InsertInitialFormula(S, LU, LUIdx);
3320 CountRegisters(LU.Formulae.back(), LUIdx);
3321 }
3322 }
3323
3324 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3325}
3326
3327/// Insert a formula for the given expression into the given use, separating out
3328/// loop-variant portions from loop-invariant and loop-computable portions.
3329void
3330LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3331 // Mark uses whose expressions cannot be expanded.
3332 if (!isSafeToExpand(S, SE))
3333 LU.RigidFormula = true;
3334
3335 Formula F;
3336 F.initialMatch(S, L, SE);
3337 bool Inserted = InsertFormula(LU, LUIdx, F);
3338 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3338, __PRETTY_FUNCTION__))
; (void)Inserted;
3339}
3340
3341/// Insert a simple single-register formula for the given expression into the
3342/// given use.
3343void
3344LSRInstance::InsertSupplementalFormula(const SCEV *S,
3345 LSRUse &LU, size_t LUIdx) {
3346 Formula F;
3347 F.BaseRegs.push_back(S);
3348 F.HasBaseReg = true;
3349 bool Inserted = InsertFormula(LU, LUIdx, F);
3350 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3350, __PRETTY_FUNCTION__))
; (void)Inserted;
3351}
3352
3353/// Note which registers are used by the given formula, updating RegUses.
3354void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3355 if (F.ScaledReg)
3356 RegUses.countRegister(F.ScaledReg, LUIdx);
3357 for (const SCEV *BaseReg : F.BaseRegs)
3358 RegUses.countRegister(BaseReg, LUIdx);
3359}
3360
3361/// If the given formula has not yet been inserted, add it to the list, and
3362/// return true. Return false otherwise.
3363bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3364 // Do not insert formula that we will not be able to expand.
3365 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3366, __PRETTY_FUNCTION__))
3366 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3366, __PRETTY_FUNCTION__))
;
3367
3368 if (!LU.InsertFormula(F, *L))
3369 return false;
3370
3371 CountRegisters(F, LUIdx);
3372 return true;
3373}
3374
3375/// Check for other uses of loop-invariant values which we're tracking. These
3376/// other uses will pin these values in registers, making them less profitable
3377/// for elimination.
3378/// TODO: This currently misses non-constant addrec step registers.
3379/// TODO: Should this give more weight to users inside the loop?
3380void
3381LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3382 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3383 SmallPtrSet<const SCEV *, 32> Visited;
3384
3385 while (!Worklist.empty()) {
3386 const SCEV *S = Worklist.pop_back_val();
3387
3388 // Don't process the same SCEV twice
3389 if (!Visited.insert(S).second)
3390 continue;
3391
3392 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3393 Worklist.append(N->op_begin(), N->op_end());
3394 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3395 Worklist.push_back(C->getOperand());
3396 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3397 Worklist.push_back(D->getLHS());
3398 Worklist.push_back(D->getRHS());
3399 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3400 const Value *V = US->getValue();
3401 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3402 // Look for instructions defined outside the loop.
3403 if (L->contains(Inst)) continue;
3404 } else if (isa<UndefValue>(V))
3405 // Undef doesn't have a live range, so it doesn't matter.
3406 continue;
3407 for (const Use &U : V->uses()) {
3408 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3409 // Ignore non-instructions.
3410 if (!UserInst)
3411 continue;
3412 // Ignore instructions in other functions (as can happen with
3413 // Constants).
3414 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3415 continue;
3416 // Ignore instructions not dominated by the loop.
3417 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3418 UserInst->getParent() :
3419 cast<PHINode>(UserInst)->getIncomingBlock(
3420 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3421 if (!DT.dominates(L->getHeader(), UseBB))
3422 continue;
3423 // Don't bother if the instruction is in a BB which ends in an EHPad.
3424 if (UseBB->getTerminator()->isEHPad())
3425 continue;
3426 // Don't bother rewriting PHIs in catchswitch blocks.
3427 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3428 continue;
3429 // Ignore uses which are part of other SCEV expressions, to avoid
3430 // analyzing them multiple times.
3431 if (SE.isSCEVable(UserInst->getType())) {
3432 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3433 // If the user is a no-op, look through to its uses.
3434 if (!isa<SCEVUnknown>(UserS))
3435 continue;
3436 if (UserS == US) {
3437 Worklist.push_back(
3438 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3439 continue;
3440 }
3441 }
3442 // Ignore icmp instructions which are already being analyzed.
3443 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3444 unsigned OtherIdx = !U.getOperandNo();
3445 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3446 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3447 continue;
3448 }
3449
3450 std::pair<size_t, int64_t> P = getUse(
3451 S, LSRUse::Basic, MemAccessTy());
3452 size_t LUIdx = P.first;
3453 int64_t Offset = P.second;
3454 LSRUse &LU = Uses[LUIdx];
3455 LSRFixup &LF = LU.getNewFixup();
3456 LF.UserInst = const_cast<Instruction *>(UserInst);
3457 LF.OperandValToReplace = U;
3458 LF.Offset = Offset;
3459 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3460 if (!LU.WidestFixupType ||
3461 SE.getTypeSizeInBits(LU.WidestFixupType) <
3462 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3463 LU.WidestFixupType = LF.OperandValToReplace->getType();
3464 InsertSupplementalFormula(US, LU, LUIdx);
3465 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3466 break;
3467 }
3468 }
3469 }
3470}
3471
3472/// Split S into subexpressions which can be pulled out into separate
3473/// registers. If C is non-null, multiply each subexpression by C.
3474///
3475/// Return remainder expression after factoring the subexpressions captured by
3476/// Ops. If Ops is complete, return NULL.
3477static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3478 SmallVectorImpl<const SCEV *> &Ops,
3479 const Loop *L,
3480 ScalarEvolution &SE,
3481 unsigned Depth = 0) {
3482 // Arbitrarily cap recursion to protect compile time.
3483 if (Depth >= 3)
3484 return S;
3485
3486 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3487 // Break out add operands.
3488 for (const SCEV *S : Add->operands()) {
3489 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3490 if (Remainder)
3491 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3492 }
3493 return nullptr;
3494 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3495 // Split a non-zero base out of an addrec.
3496 if (AR->getStart()->isZero() || !AR->isAffine())
3497 return S;
3498
3499 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3500 C, Ops, L, SE, Depth+1);
3501 // Split the non-zero AddRec unless it is part of a nested recurrence that
3502 // does not pertain to this loop.
3503 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3504 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3505 Remainder = nullptr;
3506 }
3507 if (Remainder != AR->getStart()) {
3508 if (!Remainder)
3509 Remainder = SE.getConstant(AR->getType(), 0);
3510 return SE.getAddRecExpr(Remainder,
3511 AR->getStepRecurrence(SE),
3512 AR->getLoop(),
3513 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3514 SCEV::FlagAnyWrap);
3515 }
3516 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3517 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3518 if (Mul->getNumOperands() != 2)
3519 return S;
3520 if (const SCEVConstant *Op0 =
3521 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3522 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3523 const SCEV *Remainder =
3524 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3525 if (Remainder)
3526 Ops.push_back(SE.getMulExpr(C, Remainder));
3527 return nullptr;
3528 }
3529 }
3530 return S;
3531}
3532
3533/// Return true if the SCEV represents a value that may end up as a
3534/// post-increment operation.
3535static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3536 LSRUse &LU, const SCEV *S, const Loop *L,
3537 ScalarEvolution &SE) {
3538 if (LU.Kind != LSRUse::Address ||
3539 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3540 return false;
3541 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3542 if (!AR)
3543 return false;
3544 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3545 if (!isa<SCEVConstant>(LoopStep))
3546 return false;
3547 // Check if a post-indexed load/store can be used.
3548 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3549 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3550 const SCEV *LoopStart = AR->getStart();
3551 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3552 return true;
3553 }
3554 return false;
3555}
3556
3557/// Helper function for LSRInstance::GenerateReassociations.
3558void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3559 const Formula &Base,
3560 unsigned Depth, size_t Idx,
3561 bool IsScaledReg) {
3562 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3563 // Don't generate reassociations for the base register of a value that
3564 // may generate a post-increment operator. The reason is that the
3565 // reassociations cause extra base+register formula to be created,
3566 // and possibly chosen, but the post-increment is more efficient.
3567 if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3568 return;
3569 SmallVector<const SCEV *, 8> AddOps;
3570 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3571 if (Remainder)
3572 AddOps.push_back(Remainder);
3573
3574 if (AddOps.size() == 1)
3575 return;
3576
3577 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3578 JE = AddOps.end();
3579 J != JE; ++J) {
3580 // Loop-variant "unknown" values are uninteresting; we won't be able to
3581 // do anything meaningful with them.
3582 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3583 continue;
3584
3585 // Don't pull a constant into a register if the constant could be folded
3586 // into an immediate field.
3587 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3588 LU.AccessTy, *J, Base.getNumRegs() > 1))
3589 continue;
3590
3591 // Collect all operands except *J.
3592 SmallVector<const SCEV *, 8> InnerAddOps(
3593 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3594 InnerAddOps.append(std::next(J),
3595 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3596
3597 // Don't leave just a constant behind in a register if the constant could
3598 // be folded into an immediate field.
3599 if (InnerAddOps.size() == 1 &&
3600 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3601 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3602 continue;
3603
3604 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3605 if (InnerSum->isZero())
3606 continue;
3607 Formula F = Base;
3608
3609 // Add the remaining pieces of the add back into the new formula.
3610 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3611 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3612 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3613 InnerSumSC->getValue()->getZExtValue())) {
3614 F.UnfoldedOffset =
3615 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3616 if (IsScaledReg)
3617 F.ScaledReg = nullptr;
3618 else
3619 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3620 } else if (IsScaledReg)
3621 F.ScaledReg = InnerSum;
3622 else
3623 F.BaseRegs[Idx] = InnerSum;
3624
3625 // Add J as its own register, or an unfolded immediate.
3626 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3627 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3628 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3629 SC->getValue()->getZExtValue()))
3630 F.UnfoldedOffset =
3631 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3632 else
3633 F.BaseRegs.push_back(*J);
3634 // We may have changed the number of register in base regs, adjust the
3635 // formula accordingly.
3636 F.canonicalize(*L);
3637
3638 if (InsertFormula(LU, LUIdx, F))
3639 // If that formula hadn't been seen before, recurse to find more like
3640 // it.
3641 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3642 // Because just Depth is not enough to bound compile time.
3643 // This means that every time AddOps.size() is greater 16^x we will add
3644 // x to Depth.
3645 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3646 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3647 }
3648}
3649
3650/// Split out subexpressions from adds and the bases of addrecs.
3651void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3652 Formula Base, unsigned Depth) {
3653 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3653, __PRETTY_FUNCTION__))
;
3654 // Arbitrarily cap recursion to protect compile time.
3655 if (Depth >= 3)
3656 return;
3657
3658 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3659 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3660
3661 if (Base.Scale == 1)
3662 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3663 /* Idx */ -1, /* IsScaledReg */ true);
3664}
3665
3666/// Generate a formula consisting of all of the loop-dominating registers added
3667/// into a single register.
3668void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3669 Formula Base) {
3670 // This method is only interesting on a plurality of registers.
3671 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3672 (Base.UnfoldedOffset != 0) <= 1)
3673 return;
3674
3675 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3676 // processing the formula.
3677 Base.unscale();
3678 SmallVector<const SCEV *, 4> Ops;
3679 Formula NewBase = Base;
3680 NewBase.BaseRegs.clear();
3681 Type *CombinedIntegerType = nullptr;
3682 for (const SCEV *BaseReg : Base.BaseRegs) {
3683 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3684 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3685 if (!CombinedIntegerType)
3686 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3687 Ops.push_back(BaseReg);
3688 }
3689 else
3690 NewBase.BaseRegs.push_back(BaseReg);
3691 }
3692
3693 // If no register is relevant, we're done.
3694 if (Ops.size() == 0)
3695 return;
3696
3697 // Utility function for generating the required variants of the combined
3698 // registers.
3699 auto GenerateFormula = [&](const SCEV *Sum) {
3700 Formula F = NewBase;
3701
3702 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3703 // opportunity to fold something. For now, just ignore such cases
3704 // rather than proceed with zero in a register.
3705 if (Sum->isZero())
3706 return;
3707
3708 F.BaseRegs.push_back(Sum);
3709 F.canonicalize(*L);
3710 (void)InsertFormula(LU, LUIdx, F);
3711 };
3712
3713 // If we collected at least two registers, generate a formula combining them.
3714 if (Ops.size() > 1) {
3715 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3716 GenerateFormula(SE.getAddExpr(OpsCopy));
3717 }
3718
3719 // If we have an unfolded offset, generate a formula combining it with the
3720 // registers collected.
3721 if (NewBase.UnfoldedOffset) {
3722 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3722, __PRETTY_FUNCTION__))
;
3723 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3724 true));
3725 NewBase.UnfoldedOffset = 0;
3726 GenerateFormula(SE.getAddExpr(Ops));
3727 }
3728}
3729
3730/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3731void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3732 const Formula &Base, size_t Idx,
3733 bool IsScaledReg) {
3734 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3735 GlobalValue *GV = ExtractSymbol(G, SE);
3736 if (G->isZero() || !GV)
3737 return;
3738 Formula F = Base;
3739 F.BaseGV = GV;
3740 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3741 return;
3742 if (IsScaledReg)
3743 F.ScaledReg = G;
3744 else
3745 F.BaseRegs[Idx] = G;
3746 (void)InsertFormula(LU, LUIdx, F);
3747}
3748
3749/// Generate reuse formulae using symbolic offsets.
3750void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3751 Formula Base) {
3752 // We can't add a symbolic offset if the address already contains one.
3753 if (Base.BaseGV) return;
3754
3755 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3756 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3757 if (Base.Scale == 1)
3758 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3759 /* IsScaledReg */ true);
3760}
3761
3762/// Helper function for LSRInstance::GenerateConstantOffsets.
3763void LSRInstance::GenerateConstantOffsetsImpl(
3764 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3765 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3766
3767 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3768 Formula F = Base;
3769 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3770
3771 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3772 LU.AccessTy, F)) {
3773 // Add the offset to the base register.
3774 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3775 // If it cancelled out, drop the base register, otherwise update it.
3776 if (NewG->isZero()) {
3777 if (IsScaledReg) {
3778 F.Scale = 0;
3779 F.ScaledReg = nullptr;
3780 } else
3781 F.deleteBaseReg(F.BaseRegs[Idx]);
3782 F.canonicalize(*L);
3783 } else if (IsScaledReg)
3784 F.ScaledReg = NewG;
3785 else
3786 F.BaseRegs[Idx] = NewG;
3787
3788 (void)InsertFormula(LU, LUIdx, F);
3789 }
3790 };
3791
3792 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3793
3794 // With constant offsets and constant steps, we can generate pre-inc
3795 // accesses by having the offset equal the step. So, for access #0 with a
3796 // step of 8, we generate a G - 8 base which would require the first access
3797 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3798 // for itself and hopefully becomes the base for other accesses. This means
3799 // means that a single pre-indexed access can be generated to become the new
3800 // base pointer for each iteration of the loop, resulting in no extra add/sub
3801 // instructions for pointer updating.
3802 if (FavorBackedgeIndex && LU.Kind == LSRUse::Address) {
3803 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3804 if (auto *StepRec =
3805 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3806 const APInt &StepInt = StepRec->getAPInt();
3807 int64_t Step = StepInt.isNegative() ?
3808 StepInt.getSExtValue() : StepInt.getZExtValue();
3809
3810 for (int64_t Offset : Worklist) {
3811 Offset -= Step;
3812 GenerateOffset(G, Offset);
3813 }
3814 }
3815 }
3816 }
3817 for (int64_t Offset : Worklist)
3818 GenerateOffset(G, Offset);
3819
3820 int64_t Imm = ExtractImmediate(G, SE);
3821 if (G->isZero() || Imm == 0)
3822 return;
3823 Formula F = Base;
3824 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3825 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3826 return;
3827 if (IsScaledReg)
3828 F.ScaledReg = G;
3829 else
3830 F.BaseRegs[Idx] = G;
3831 (void)InsertFormula(LU, LUIdx, F);
3832}
3833
3834/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3835void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3836 Formula Base) {
3837 // TODO: For now, just add the min and max offset, because it usually isn't
3838 // worthwhile looking at everything inbetween.
3839 SmallVector<int64_t, 2> Worklist;
3840 Worklist.push_back(LU.MinOffset);
3841 if (LU.MaxOffset != LU.MinOffset)
3842 Worklist.push_back(LU.MaxOffset);
3843
3844 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3845 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3846 if (Base.Scale == 1)
3847 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3848 /* IsScaledReg */ true);
3849}
3850
3851/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3852/// == y -> x*c == y*c.
3853void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3854 Formula Base) {
3855 if (LU.Kind != LSRUse::ICmpZero) return;
3856
3857 // Determine the integer type for the base formula.
3858 Type *IntTy = Base.getType();
3859 if (!IntTy) return;
3860 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3861
3862 // Don't do this if there is more than one offset.
3863 if (LU.MinOffset != LU.MaxOffset) return;
3864
3865 // Check if transformation is valid. It is illegal to multiply pointer.
3866 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3867 return;
3868 for (const SCEV *BaseReg : Base.BaseRegs)
3869 if (BaseReg->getType()->isPointerTy())
3870 return;
3871 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3871, __PRETTY_FUNCTION__))
;
3872
3873 // Check each interesting stride.
3874 for (int64_t Factor : Factors) {
3875 // Check that the multiplication doesn't overflow.
3876 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3877 continue;
3878 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3879 if (NewBaseOffset / Factor != Base.BaseOffset)
3880 continue;
3881 // If the offset will be truncated at this use, check that it is in bounds.
3882 if (!IntTy->isPointerTy() &&
3883 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3884 continue;
3885
3886 // Check that multiplying with the use offset doesn't overflow.
3887 int64_t Offset = LU.MinOffset;
3888 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3889 continue;
3890 Offset = (uint64_t)Offset * Factor;
3891 if (Offset / Factor != LU.MinOffset)
3892 continue;
3893 // If the offset will be truncated at this use, check that it is in bounds.
3894 if (!IntTy->isPointerTy() &&
3895 !ConstantInt::isValueValidForType(IntTy, Offset))
3896 continue;
3897
3898 Formula F = Base;
3899 F.BaseOffset = NewBaseOffset;
3900
3901 // Check that this scale is legal.
3902 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3903 continue;
3904
3905 // Compensate for the use having MinOffset built into it.
3906 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3907
3908 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3909
3910 // Check that multiplying with each base register doesn't overflow.
3911 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3912 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3913 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3914 goto next;
3915 }
3916
3917 // Check that multiplying with the scaled register doesn't overflow.
3918 if (F.ScaledReg) {
3919 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3920 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3921 continue;
3922 }
3923
3924 // Check that multiplying with the unfolded offset doesn't overflow.
3925 if (F.UnfoldedOffset != 0) {
3926 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3927 Factor == -1)
3928 continue;
3929 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3930 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3931 continue;
3932 // If the offset will be truncated, check that it is in bounds.
3933 if (!IntTy->isPointerTy() &&
3934 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3935 continue;
3936 }
3937
3938 // If we make it here and it's legal, add it.
3939 (void)InsertFormula(LU, LUIdx, F);
3940 next:;
3941 }
3942}
3943
3944/// Generate stride factor reuse formulae by making use of scaled-offset address
3945/// modes, for example.
3946void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3947 // Determine the integer type for the base formula.
3948 Type *IntTy = Base.getType();
3949 if (!IntTy) return;
3950
3951 // If this Formula already has a scaled register, we can't add another one.
3952 // Try to unscale the formula to generate a better scale.
3953 if (Base.Scale != 0 && !Base.unscale())
3954 return;
3955
3956 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3956, __PRETTY_FUNCTION__))
;
3957
3958 // Check each interesting stride.
3959 for (int64_t Factor : Factors) {
3960 Base.Scale = Factor;
3961 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3962 // Check whether this scale is going to be legal.
3963 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3964 Base)) {
3965 // As a special-case, handle special out-of-loop Basic users specially.
3966 // TODO: Reconsider this special case.
3967 if (LU.Kind == LSRUse::Basic &&
3968 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3969 LU.AccessTy, Base) &&
3970 LU.AllFixupsOutsideLoop)
3971 LU.Kind = LSRUse::Special;
3972 else
3973 continue;
3974 }
3975 // For an ICmpZero, negating a solitary base register won't lead to
3976 // new solutions.
3977 if (LU.Kind == LSRUse::ICmpZero &&
3978 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3979 continue;
3980 // For each addrec base reg, if its loop is current loop, apply the scale.
3981 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3982 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3983 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3984 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3985 if (FactorS->isZero())
3986 continue;
3987 // Divide out the factor, ignoring high bits, since we'll be
3988 // scaling the value back up in the end.
3989 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3990 // TODO: This could be optimized to avoid all the copying.
3991 Formula F = Base;
3992 F.ScaledReg = Quotient;
3993 F.deleteBaseReg(F.BaseRegs[i]);
3994 // The canonical representation of 1*reg is reg, which is already in
3995 // Base. In that case, do not try to insert the formula, it will be
3996 // rejected anyway.
3997 if (F.Scale == 1 && (F.BaseRegs.empty() ||
3998 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3999 continue;
4000 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4001 // non canonical Formula with ScaledReg's loop not being L.
4002 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4003 F.canonicalize(*L);
4004 (void)InsertFormula(LU, LUIdx, F);
4005 }
4006 }
4007 }
4008 }
4009}
4010
4011/// Generate reuse formulae from different IV types.
4012void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4013 // Don't bother truncating symbolic values.
4014 if (Base.BaseGV) return;
4015
4016 // Determine the integer type for the base formula.
4017 Type *DstTy = Base.getType();
4018 if (!DstTy) return;
4019 DstTy = SE.getEffectiveSCEVType(DstTy);
4020
4021 for (Type *SrcTy : Types) {
4022 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4023 Formula F = Base;
4024
4025 // Sometimes SCEV is able to prove zero during ext transform. It may
4026 // happen if SCEV did not do all possible transforms while creating the
4027 // initial node (maybe due to depth limitations), but it can do them while
4028 // taking ext.
4029 if (F.ScaledReg) {
4030 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4031 if (NewScaledReg->isZero())
4032 continue;
4033 F.ScaledReg = NewScaledReg;
4034 }
4035 bool HasZeroBaseReg = false;
4036 for (const SCEV *&BaseReg : F.BaseRegs) {
4037 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4038 if (NewBaseReg->isZero()) {
4039 HasZeroBaseReg = true;
4040 break;
4041 }
4042 BaseReg = NewBaseReg;
4043 }
4044 if (HasZeroBaseReg)
4045 continue;
4046
4047 // TODO: This assumes we've done basic processing on all uses and
4048 // have an idea what the register usage is.
4049 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4050 continue;
4051
4052 F.canonicalize(*L);
4053 (void)InsertFormula(LU, LUIdx, F);
4054 }
4055 }
4056}
4057
4058namespace {
4059
4060/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4061/// modifications so that the search phase doesn't have to worry about the data
4062/// structures moving underneath it.
4063struct WorkItem {
4064 size_t LUIdx;
4065 int64_t Imm;
4066 const SCEV *OrigReg;
4067
4068 WorkItem(size_t LI, int64_t I, const SCEV *R)
4069 : LUIdx(LI), Imm(I), OrigReg(R) {}
4070
4071 void print(raw_ostream &OS) const;
4072 void dump() const;
4073};
4074
4075} // end anonymous namespace
4076
4077#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4078void WorkItem::print(raw_ostream &OS) const {
4079 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4080 << " , add offset " << Imm;
4081}
4082
4083LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4084 print(errs()); errs() << '\n';
4085}
4086#endif
4087
4088/// Look for registers which are a constant distance apart and try to form reuse
4089/// opportunities between them.
4090void LSRInstance::GenerateCrossUseConstantOffsets() {
4091 // Group the registers by their value without any added constant offset.
4092 using ImmMapTy = std::map<int64_t, const SCEV *>;
4093
4094 DenseMap<const SCEV *, ImmMapTy> Map;
4095 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4096 SmallVector<const SCEV *, 8> Sequence;
4097 for (const SCEV *Use : RegUses) {
4098 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4099 int64_t Imm = ExtractImmediate(Reg, SE);
4100 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4101 if (Pair.second)
4102 Sequence.push_back(Reg);
4103 Pair.first->second.insert(std::make_pair(Imm, Use));
4104 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4105 }
4106
4107 // Now examine each set of registers with the same base value. Build up
4108 // a list of work to do and do the work in a separate step so that we're
4109 // not adding formulae and register counts while we're searching.
4110 SmallVector<WorkItem, 32> WorkItems;
4111 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4112 for (const SCEV *Reg : Sequence) {
4113 const ImmMapTy &Imms = Map.find(Reg)->second;
4114
4115 // It's not worthwhile looking for reuse if there's only one offset.
4116 if (Imms.size() == 1)
4117 continue;
4118
4119 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)
4120 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)
4121 : 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)
4122 << ' ' << 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)
4123 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)
;
4124
4125 // Examine each offset.
4126 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4127 J != JE; ++J) {
4128 const SCEV *OrigReg = J->second;
4129
4130 int64_t JImm = J->first;
4131 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4132
4133 if (!isa<SCEVConstant>(OrigReg) &&
4134 UsedByIndicesMap[Reg].count() == 1) {
4135 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)
4136 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4137 continue;
4138 }
4139
4140 // Conservatively examine offsets between this orig reg a few selected
4141 // other orig regs.
4142 int64_t First = Imms.begin()->first;
4143 int64_t Last = std::prev(Imms.end())->first;
4144 // Compute (First + Last) / 2 without overflow using the fact that
4145 // First + Last = 2 * (First + Last) + (First ^ Last).
4146 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4147 // If the result is negative and First is odd and Last even (or vice versa),
4148 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4149 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4150 ImmMapTy::const_iterator OtherImms[] = {
4151 Imms.begin(), std::prev(Imms.end()),
4152 Imms.lower_bound(Avg)};
4153 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4154 ImmMapTy::const_iterator M = OtherImms[i];
4155 if (M == J || M == JE) continue;
4156
4157 // Compute the difference between the two.
4158 int64_t Imm = (uint64_t)JImm - M->first;
4159 for (unsigned LUIdx : UsedByIndices.set_bits())
4160 // Make a memo of this use, offset, and register tuple.
4161 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4162 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4163 }
4164 }
4165 }
4166
4167 Map.clear();
4168 Sequence.clear();
4169 UsedByIndicesMap.clear();
4170 UniqueItems.clear();
4171
4172 // Now iterate through the worklist and add new formulae.
4173 for (const WorkItem &WI : WorkItems) {
4174 size_t LUIdx = WI.LUIdx;
4175 LSRUse &LU = Uses[LUIdx];
4176 int64_t Imm = WI.Imm;
4177 const SCEV *OrigReg = WI.OrigReg;
4178
4179 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4180 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4181 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4182
4183 // TODO: Use a more targeted data structure.
4184 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4185 Formula F = LU.Formulae[L];
4186 // FIXME: The code for the scaled and unscaled registers looks
4187 // very similar but slightly different. Investigate if they
4188 // could be merged. That way, we would not have to unscale the
4189 // Formula.
4190 F.unscale();
4191 // Use the immediate in the scaled register.
4192 if (F.ScaledReg == OrigReg) {
4193 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4194 // Don't create 50 + reg(-50).
4195 if (F.referencesReg(SE.getSCEV(
4196 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4197 continue;
4198 Formula NewF = F;
4199 NewF.BaseOffset = Offset;
4200 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4201 NewF))
4202 continue;
4203 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4204
4205 // If the new scale is a constant in a register, and adding the constant
4206 // value to the immediate would produce a value closer to zero than the
4207 // immediate itself, then the formula isn't worthwhile.
4208 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4209 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4210 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4211 .ule(std::abs(NewF.BaseOffset)))
4212 continue;
4213
4214 // OK, looks good.
4215 NewF.canonicalize(*this->L);
4216 (void)InsertFormula(LU, LUIdx, NewF);
4217 } else {
4218 // Use the immediate in a base register.
4219 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4220 const SCEV *BaseReg = F.BaseRegs[N];
4221 if (BaseReg != OrigReg)
4222 continue;
4223 Formula NewF = F;
4224 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4225 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4226 LU.Kind, LU.AccessTy, NewF)) {
4227 if (TTI.shouldFavorPostInc() &&
4228 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4229 continue;
4230 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4231 continue;
4232 NewF = F;
4233 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4234 }
4235 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4236
4237 // If the new formula has a constant in a register, and adding the
4238 // constant value to the immediate would produce a value closer to
4239 // zero than the immediate itself, then the formula isn't worthwhile.
4240 for (const SCEV *NewReg : NewF.BaseRegs)
4241 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4242 if ((C->getAPInt() + NewF.BaseOffset)
4243 .abs()
4244 .slt(std::abs(NewF.BaseOffset)) &&
4245 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4246 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4247 goto skip_formula;
4248
4249 // Ok, looks good.
4250 NewF.canonicalize(*this->L);
4251 (void)InsertFormula(LU, LUIdx, NewF);
4252 break;
4253 skip_formula:;
4254 }
4255 }
4256 }
4257 }
4258}
4259
4260/// Generate formulae for each use.
4261void
4262LSRInstance::GenerateAllReuseFormulae() {
4263 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4264 // queries are more precise.
4265 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4266 LSRUse &LU = Uses[LUIdx];
4267 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4268 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4269 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4270 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4271 }
4272 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4273 LSRUse &LU = Uses[LUIdx];
4274 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4275 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4276 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4277 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4278 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4279 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4280 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4281 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4282 }
4283 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4284 LSRUse &LU = Uses[LUIdx];
4285 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4286 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4287 }
4288
4289 GenerateCrossUseConstantOffsets();
4290
4291 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4292 "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)
4293 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4294}
4295
4296/// If there are multiple formulae with the same set of registers used
4297/// by other uses, pick the best one and delete the others.
4298void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4299 DenseSet<const SCEV *> VisitedRegs;
4300 SmallPtrSet<const SCEV *, 16> Regs;
4301 SmallPtrSet<const SCEV *, 16> LoserRegs;
4302#ifndef NDEBUG
4303 bool ChangedFormulae = false;
4304#endif
4305
4306 // Collect the best formula for each unique set of shared registers. This
4307 // is reset for each use.
4308 using BestFormulaeTy =
4309 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4310
4311 BestFormulaeTy BestFormulae;
4312
4313 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4314 LSRUse &LU = Uses[LUIdx];
4315 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)
4316 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4317
4318 bool Any = false;
4319 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4320 FIdx != NumForms; ++FIdx) {
4321 Formula &F = LU.Formulae[FIdx];
4322
4323 // Some formulas are instant losers. For example, they may depend on
4324 // nonexistent AddRecs from other loops. These need to be filtered
4325 // immediately, otherwise heuristics could choose them over others leading
4326 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4327 // avoids the need to recompute this information across formulae using the
4328 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4329 // the corresponding bad register from the Regs set.
4330 Cost CostF(L, SE, TTI);
4331 Regs.clear();
4332 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4333 if (CostF.isLoser()) {
4334 // During initial formula generation, undesirable formulae are generated
4335 // by uses within other loops that have some non-trivial address mode or
4336 // use the postinc form of the IV. LSR needs to provide these formulae
4337 // as the basis of rediscovering the desired formula that uses an AddRec
4338 // corresponding to the existing phi. Once all formulae have been
4339 // generated, these initial losers may be pruned.
4340 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)
4341 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4342 }
4343 else {
4344 SmallVector<const SCEV *, 4> Key;
4345 for (const SCEV *Reg : F.BaseRegs) {
4346 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4347 Key.push_back(Reg);
4348 }
4349 if (F.ScaledReg &&
4350 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4351 Key.push_back(F.ScaledReg);
4352 // Unstable sort by host order ok, because this is only used for
4353 // uniquifying.
4354 llvm::sort(Key);
4355
4356 std::pair<BestFormulaeTy::const_iterator, bool> P =
4357 BestFormulae.insert(std::make_pair(Key, FIdx));
4358 if (P.second)
4359 continue;
4360
4361 Formula &Best = LU.Formulae[P.first->second];
4362
4363 Cost CostBest(L, SE, TTI);
4364 Regs.clear();
4365 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4366 if (CostF.isLess(CostBest))
4367 std::swap(F, Best);
4368 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)
4369 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)
4370 " 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)
4371 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)
;
4372 }
4373#ifndef NDEBUG
4374 ChangedFormulae = true;
4375#endif
4376 LU.DeleteFormula(F);
4377 --FIdx;
4378 --NumForms;
4379 Any = true;
4380 }
4381
4382 // Now that we've filtered out some formulae, recompute the Regs set.
4383 if (Any)
4384 LU.RecomputeRegs(LUIdx, RegUses);
4385
4386 // Reset this to prepare for the next use.
4387 BestFormulae.clear();
4388 }
4389
4390 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)
4391 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)
4392 "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)
4393 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)
4394 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4395}
4396
4397/// Estimate the worst-case number of solutions the solver might have to
4398/// consider. It almost never considers this many solutions because it prune the
4399/// search space, but the pruning isn't always sufficient.
4400size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4401 size_t Power = 1;
4402 for (const LSRUse &LU : Uses) {
4403 size_t FSize = LU.Formulae.size();
4404 if (FSize >= ComplexityLimit) {
4405 Power = ComplexityLimit;
4406 break;
4407 }
4408 Power *= FSize;
4409 if (Power >= ComplexityLimit)
4410 break;
4411 }
4412 return Power;
4413}
4414
4415/// When one formula uses a superset of the registers of another formula, it
4416/// won't help reduce register pressure (though it may not necessarily hurt
4417/// register pressure); remove it to simplify the system.
4418void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4419 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4420 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)
;
4421
4422 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)
4423 "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)
4424 "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)
;
4425
4426 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4427 LSRUse &LU = Uses[LUIdx];
4428 bool Any = false;
4429 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4430 Formula &F = LU.Formulae[i];
4431 // Look for a formula with a constant or GV in a register. If the use
4432 // also has a formula with that same value in an immediate field,
4433 // delete the one that uses a register.
4434 for (SmallVectorImpl<const SCEV *>::const_iterator
4435 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4436 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4437 Formula NewF = F;
4438 //FIXME: Formulas should store bitwidth to do wrapping properly.
4439 // See PR41034.
4440 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4441 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4442 (I - F.BaseRegs.begin()));
4443 if (LU.HasFormulaWithSameRegs(NewF)) {
4444 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4445 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4446 LU.DeleteFormula(F);
4447 --i;
4448 --e;
4449 Any = true;
4450 break;
4451 }
4452 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4453 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4454 if (!F.BaseGV) {
4455 Formula NewF = F;
4456 NewF.BaseGV = GV;
4457 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4458 (I - F.BaseRegs.begin()));
4459 if (LU.HasFormulaWithSameRegs(NewF)) {
4460 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4461 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4462 LU.DeleteFormula(F);
4463 --i;
4464 --e;
4465 Any = true;
4466 break;
4467 }
4468 }
4469 }
4470 }
4471 }
4472 if (Any)
4473 LU.RecomputeRegs(LUIdx, RegUses);
4474 }
4475
4476 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)
;
4477 }
4478}
4479
4480/// When there are many registers for expressions like A, A+1, A+2, etc.,
4481/// allocate a single register for them.
4482void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4483 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4484 return;
4485
4486 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)
4487 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)
4488 "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)
4489 "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)
;
4490
4491 // This is especially useful for unrolled loops.
4492
4493 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4494 LSRUse &LU = Uses[LUIdx];
4495 for (const Formula &F : LU.Formulae) {
4496 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4497 continue;
4498
4499 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4500 if (!LUThatHas)
4501 continue;
4502
4503 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4504 LU.Kind, LU.AccessTy))
4505 continue;
4506
4507 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)
;
4508
4509 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4510
4511 // Transfer the fixups of LU to LUThatHas.
4512 for (LSRFixup &Fixup : LU.Fixups) {
4513 Fixup.Offset += F.BaseOffset;
4514 LUThatHas->pushFixup(Fixup);
4515 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)
;
4516 }
4517
4518 // Delete formulae from the new use which are no longer legal.
4519 bool Any = false;
4520 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4521 Formula &F = LUThatHas->Formulae[i];
4522 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4523 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4524 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)
;
4525 LUThatHas->DeleteFormula(F);
4526 --i;
4527 --e;
4528 Any = true;
4529 }
4530 }
4531
4532 if (Any)
4533 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4534
4535 // Delete the old use.
4536 DeleteUse(LU, LUIdx);
4537 --LUIdx;
4538 --NumUses;
4539 break;
4540 }
4541 }
4542
4543 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)
;
4544}
4545
4546/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4547/// we've done more filtering, as it may be able to find more formulae to
4548/// eliminate.
4549void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4550 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4551 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)
;
4552
4553 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)
4554 "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)
;
4555
4556 FilterOutUndesirableDedicatedRegisters();
4557
4558 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)
;
4559 }
4560}
4561
4562/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4563/// Pick the best one and delete the others.
4564/// This narrowing heuristic is to keep as many formulae with different
4565/// Scale and ScaledReg pair as possible while narrowing the search space.
4566/// The benefit is that it is more likely to find out a better solution
4567/// from a formulae set with more Scale and ScaledReg variations than
4568/// a formulae set with the same Scale and ScaledReg. The picking winner
4569/// reg heuristic will often keep the formulae with the same Scale and
4570/// ScaledReg and filter others, and we want to avoid that if possible.
4571void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4572 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4573 return;
4574
4575 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)
4576 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)
4577 "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)
4578 "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)
;
4579
4580 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4581 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4582
4583 BestFormulaeTy BestFormulae;
4584#ifndef NDEBUG
4585 bool ChangedFormulae = false;
4586#endif
4587 DenseSet<const SCEV *> VisitedRegs;
4588 SmallPtrSet<const SCEV *, 16> Regs;
4589
4590 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4591 LSRUse &LU = Uses[LUIdx];
4592 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)
4593 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4594
4595 // Return true if Formula FA is better than Formula FB.
4596 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4597 // First we will try to choose the Formula with fewer new registers.
4598 // For a register used by current Formula, the more the register is
4599 // shared among LSRUses, the less we increase the register number
4600 // counter of the formula.
4601 size_t FARegNum = 0;
4602 for (const SCEV *Reg : FA.BaseRegs) {
4603 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4604 FARegNum += (NumUses - UsedByIndices.count() + 1);
4605 }
4606 size_t FBRegNum = 0;
4607 for (const SCEV *Reg : FB.BaseRegs) {
4608 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4609 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4610 }
4611 if (FARegNum != FBRegNum)
4612 return FARegNum < FBRegNum;
4613
4614 // If the new register numbers are the same, choose the Formula with
4615 // less Cost.
4616 Cost CostFA(L, SE, TTI);
4617 Cost CostFB(L, SE, TTI);
4618 Regs.clear();
4619 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4620 Regs.clear();
4621 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4622 return CostFA.isLess(CostFB);
4623 };
4624
4625 bool Any = false;
4626 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4627 ++FIdx) {
4628 Formula &F = LU.Formulae[FIdx];
4629 if (!F.ScaledReg)
4630 continue;
4631 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4632 if (P.second)
4633 continue;
4634
4635 Formula &Best = LU.Formulae[P.first->second];
4636 if (IsBetterThan(F, Best))
4637 std::swap(F, Best);
4638 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)
4639 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)
4640 " 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)
4641 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)
;
4642#ifndef NDEBUG
4643 ChangedFormulae = true;
4644#endif
4645 LU.DeleteFormula(F);
4646 --FIdx;
4647 --NumForms;
4648 Any = true;
4649 }
4650 if (Any)
4651 LU.RecomputeRegs(LUIdx, RegUses);
4652
4653 // Reset this to prepare for the next use.
4654 BestFormulae.clear();
4655 }
4656
4657 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)
4658 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)
4659 "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)
4660 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)
4661 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4662}
4663
4664/// The function delete formulas with high registers number expectation.
4665/// Assuming we don't know the value of each formula (already delete
4666/// all inefficient), generate probability of not selecting for each
4667/// register.
4668/// For example,
4669/// Use1:
4670/// reg(a) + reg({0,+,1})
4671/// reg(a) + reg({-1,+,1}) + 1
4672/// reg({a,+,1})
4673/// Use2:
4674/// reg(b) + reg({0,+,1})
4675/// reg(b) + reg({-1,+,1}) + 1
4676/// reg({b,+,1})
4677/// Use3:
4678/// reg(c) + reg(b) + reg({0,+,1})
4679/// reg(c) + reg({b,+,1})
4680///
4681/// Probability of not selecting
4682/// Use1 Use2 Use3
4683/// reg(a) (1/3) * 1 * 1
4684/// reg(b) 1 * (1/3) * (1/2)
4685/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4686/// reg({-1,+,1}) (2/3) * (2/3) * 1
4687/// reg({a,+,1}) (2/3) * 1 * 1
4688/// reg({b,+,1}) 1 * (2/3) * (2/3)
4689/// reg(c) 1 * 1 * 0
4690///
4691/// Now count registers number mathematical expectation for each formula:
4692/// Note that for each use we exclude probability if not selecting for the use.
4693/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4694/// probabilty 1/3 of not selecting for Use1).
4695/// Use1:
4696/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4697/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4698/// reg({a,+,1}) 1
4699/// Use2:
4700/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4701/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4702/// reg({b,+,1}) 2/3
4703/// Use3:
4704/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4705/// reg(c) + reg({b,+,1}) 1 + 2/3
4706void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4707 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4708 return;
4709 // Ok, we have too many of formulae on our hands to conveniently handle.
4710 // Use a rough heuristic to thin out the list.
4711
4712 // Set of Regs wich will be 100% used in final solution.
4713 // Used in each formula of a solution (in example above this is reg(c)).
4714 // We can skip them in calculations.
4715 SmallPtrSet<const SCEV *, 4> UniqRegs;
4716 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)
;
4717
4718 // Map each register to probability of not selecting
4719 DenseMap <const SCEV *, float> RegNumMap;
4720 for (const SCEV *Reg : RegUses) {
4721 if (UniqRegs.count(Reg))
4722 continue;
4723 float PNotSel = 1;
4724 for (const LSRUse &LU : Uses) {
4725 if (!LU.Regs.count(Reg))
4726 continue;
4727 float P = LU.getNotSelectedProbability(Reg);
4728 if (P != 0.0)
4729 PNotSel *= P;
4730 else
4731 UniqRegs.insert(Reg);
4732 }
4733 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4734 }
4735
4736 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4737 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)
;
4738
4739 // Delete formulas where registers number expectation is high.
4740 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4741 LSRUse &LU = Uses[LUIdx];
4742 // If nothing to delete - continue.
4743 if (LU.Formulae.size() < 2)
4744 continue;
4745 // This is temporary solution to test performance. Float should be
4746 // replaced with round independent type (based on integers) to avoid
4747 // different results for different target builds.
4748 float FMinRegNum = LU.Formulae[0].getNumRegs();
4749 float FMinARegNum = LU.Formulae[0].getNumRegs();
4750 size_t MinIdx = 0;
4751 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4752 Formula &F = LU.Formulae[i];
4753 float FRegNum = 0;
4754 float FARegNum = 0;
4755 for (const SCEV *BaseReg : F.BaseRegs) {
4756 if (UniqRegs.count(BaseReg))
4757 continue;
4758 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4759 if (isa<SCEVAddRecExpr>(BaseReg))
4760 FARegNum +=
4761 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4762 }
4763 if (const SCEV *ScaledReg = F.ScaledReg) {
4764 if (!UniqRegs.count(ScaledReg)) {
4765 FRegNum +=
4766 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4767 if (isa<SCEVAddRecExpr>(ScaledReg))
4768 FARegNum +=
4769 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4770 }
4771 }
4772 if (FMinRegNum > FRegNum ||
4773 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4774 FMinRegNum = FRegNum;
4775 FMinARegNum = FARegNum;
4776 MinIdx = i;
4777 }
4778 }
4779 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)
4780 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)
;
4781 if (MinIdx != 0)
4782 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4783 while (LU.Formulae.size() != 1) {
4784 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
)
4785 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4786 LU.Formulae.pop_back();
4787 }
4788 LU.RecomputeRegs(LUIdx, RegUses);
4789 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4789, __PRETTY_FUNCTION__))
;
4790 Formula &F = LU.Formulae[0];
4791 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)
;
4792 // When we choose the formula, the regs become unique.
4793 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4794 if (F.ScaledReg)
4795 UniqRegs.insert(F.ScaledReg);
4796 }
4797 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)
;
4798}
4799
4800/// Pick a register which seems likely to be profitable, and then in any use
4801/// which has any reference to that register, delete all formulae which do not
4802/// reference that register.
4803void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4804 // With all other options exhausted, loop until the system is simple
4805 // enough to handle.
4806 SmallPtrSet<const SCEV *, 4> Taken;
4807 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4808 // Ok, we have too many of formulae on our hands to conveniently handle.
4809 // Use a rough heuristic to thin out the list.
4810 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)
;
4811
4812 // Pick the register which is used by the most LSRUses, which is likely
4813 // to be a good reuse register candidate.
4814 const SCEV *Best = nullptr;
4815 unsigned BestNum = 0;
4816 for (const SCEV *Reg : RegUses) {
4817 if (Taken.count(Reg))
4818 continue;
4819 if (!Best) {
4820 Best = Reg;
4821 BestNum = RegUses.getUsedByIndices(Reg).count();
4822 } else {
4823 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4824 if (Count > BestNum) {
4825 Best = Reg;
4826 BestNum = Count;
4827 }
4828 }
4829 }
4830 assert(Best && "Failed to find best LSRUse candidate")((Best && "Failed to find best LSRUse candidate") ? static_cast
<void> (0) : __assert_fail ("Best && \"Failed to find best LSRUse candidate\""
, "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4830, __PRETTY_FUNCTION__))
;
4831
4832 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)
4833 << " 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)
;
4834 Taken.insert(Best);
4835
4836 // In any use with formulae which references this register, delete formulae
4837 // which don't reference it.
4838 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4839 LSRUse &LU = Uses[LUIdx];
4840 if (!LU.Regs.count(Best)) continue;
4841
4842 bool Any = false;
4843 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4844 Formula &F = LU.Formulae[i];
4845 if (!F.referencesReg(Best)) {
4846 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)
;
4847 LU.DeleteFormula(F);
4848 --e;
4849 --i;
4850 Any = true;
4851 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4851, __PRETTY_FUNCTION__))
;
4852 continue;
4853 }
4854 }
4855
4856 if (Any)
4857 LU.RecomputeRegs(LUIdx, RegUses);
4858 }
4859
4860 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)
;
4861 }
4862}
4863
4864/// If there are an extraordinary number of formulae to choose from, use some
4865/// rough heuristics to prune down the number of formulae. This keeps the main
4866/// solver from taking an extraordinary amount of time in some worst-case
4867/// scenarios.
4868void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4869 NarrowSearchSpaceByDetectingSupersets();
4870 NarrowSearchSpaceByCollapsingUnrolledCode();
4871 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4872 if (FilterSameScaledReg)
4873 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4874 if (LSRExpNarrow)
4875 NarrowSearchSpaceByDeletingCostlyFormulas();
4876 else
4877 NarrowSearchSpaceByPickingWinnerRegs();
4878}
4879
4880/// This is the recursive solver.
4881void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4882 Cost &SolutionCost,
4883 SmallVectorImpl<const Formula *> &Workspace,
4884 const Cost &CurCost,
4885 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4886 DenseSet<const SCEV *> &VisitedRegs) const {
4887 // Some ideas:
4888 // - prune more:
4889 // - use more aggressive filtering
4890 // - sort the formula so that the most profitable solutions are found first
4891 // - sort the uses too
4892 // - search faster:
4893 // - don't compute a cost, and then compare. compare while computing a cost
4894 // and bail early.
4895 // - track register sets with SmallBitVector
4896
4897 const LSRUse &LU = Uses[Workspace.size()];
4898
4899 // If this use references any register that's already a part of the
4900 // in-progress solution, consider it a requirement that a formula must
4901 // reference that register in order to be considered. This prunes out
4902 // unprofitable searching.
4903 SmallSetVector<const SCEV *, 4> ReqRegs;
4904 for (const SCEV *S : CurRegs)
4905 if (LU.Regs.count(S))
4906 ReqRegs.insert(S);
4907
4908 SmallPtrSet<const SCEV *, 16> NewRegs;
4909 Cost NewCost(L, SE, TTI);
4910 for (const Formula &F : LU.Formulae) {
4911 // Ignore formulae which may not be ideal in terms of register reuse of
4912 // ReqRegs. The formula should use all required registers before
4913 // introducing new ones.
4914 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4915 for (const SCEV *Reg : ReqRegs) {
4916 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4917 is_contained(F.BaseRegs, Reg)) {
4918 --NumReqRegsToFind;
4919 if (NumReqRegsToFind == 0)
4920 break;
4921 }
4922 }
4923 if (NumReqRegsToFind != 0) {
4924 // If none of the formulae satisfied the required registers, then we could
4925 // clear ReqRegs and try again. Currently, we simply give up in this case.
4926 continue;
4927 }
4928
4929 // Evaluate the cost of the current formula. If it's already worse than
4930 // the current best, prune the search at that point.
4931 NewCost = CurCost;
4932 NewRegs = CurRegs;
4933 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
4934 if (NewCost.isLess(SolutionCost)) {
4935 Workspace.push_back(&F);
4936 if (Workspace.size() != Uses.size()) {
4937 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4938 NewRegs, VisitedRegs);
4939 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4940 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4941 } else {
4942 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)
4943 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)
4944 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)
4945 << "- " << *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)
4946 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)
;
4947
4948 SolutionCost = NewCost;
4949 Solution = Workspace;
4950 }
4951 Workspace.pop_back();
4952 }
4953 }
4954}
4955
4956/// Choose one formula from each use. Return the results in the given Solution
4957/// vector.
4958void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4959 SmallVector<const Formula *, 8> Workspace;
4960 Cost SolutionCost(L, SE, TTI);
4961 SolutionCost.Lose();
4962 Cost CurCost(L, SE, TTI);
4963 SmallPtrSet<const SCEV *, 16> CurRegs;
4964 DenseSet<const SCEV *> VisitedRegs;
4965 Workspace.reserve(Uses.size());
4966
4967 // SolveRecurse does all the work.
4968 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4969 CurRegs, VisitedRegs);
4970 if (Solution.empty()) {
4971 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
4972 return;
4973 }
4974
4975 // Ok, we've now made all our decisions.
4976 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
)
4977 "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
)
4978 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
)
4979 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
)
4980 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
)
4981 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
)
4982 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
)
4983 " ";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 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
)
4985 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 })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
4988 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4988, __PRETTY_FUNCTION__))
;
4989}
4990
4991/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4992/// we can go while still being dominated by the input positions. This helps
4993/// canonicalize the insert position, which encourages sharing.
4994BasicBlock::iterator
4995LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4996 const SmallVectorImpl<Instruction *> &Inputs)
4997 const {
4998 Instruction *Tentative = &*IP;
4999 while (true) {
5000 bool AllDominate = true;
5001 Instruction *BetterPos = nullptr;
5002 // Don't bother attempting to insert before a catchswitch, their basic block
5003 // cannot have other non-PHI instructions.
5004 if (isa<CatchSwitchInst>(Tentative))
5005 return IP;
5006
5007 for (Instruction *Inst : Inputs) {
5008 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5009 AllDominate = false;
5010 break;
5011 }
5012 // Attempt to find an insert position in the middle of the block,
5013 // instead of at the end, so that it can be used for other expansions.
5014 if (Tentative->getParent() == Inst->getParent() &&
5015 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5016 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5017 }
5018 if (!AllDominate)
5019 break;
5020 if (BetterPos)
5021 IP = BetterPos->getIterator();
5022 else
5023 IP = Tentative->getIterator();
5024
5025 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5026 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5027
5028 BasicBlock *IDom;
5029 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5030 if (!Rung) return IP;
5031 Rung = Rung->getIDom();
5032 if (!Rung) return IP;
5033 IDom = Rung->getBlock();
5034
5035 // Don't climb into a loop though.
5036 const Loop *IDomLoop = LI.getLoopFor(IDom);
5037 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5038 if (IDomDepth <= IPLoopDepth &&
5039 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5040 break;
5041 }
5042
5043 Tentative = IDom->getTerminator();
5044 }
5045
5046 return IP;
5047}
5048
5049/// Determine an input position which will be dominated by the operands and
5050/// which will dominate the result.
5051BasicBlock::iterator
5052LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5053 const LSRFixup &LF,
5054 const LSRUse &LU,
5055 SCEVExpander &Rewriter) const {
5056 // Collect some instructions which must be dominated by the
5057 // expanding replacement. These must be dominated by any operands that
5058 // will be required in the expansion.
5059 SmallVector<Instruction *, 4> Inputs;
5060 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5061 Inputs.push_back(I);
5062 if (LU.Kind == LSRUse::ICmpZero)
5063 if (Instruction *I =
5064 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5065 Inputs.push_back(I);
5066 if (LF.PostIncLoops.count(L)) {
5067 if (LF.isUseFullyOutsideLoop(L))
5068 Inputs.push_back(L->getLoopLatch()->getTerminator());
5069 else
5070 Inputs.push_back(IVIncInsertPos);
5071 }
5072 // The expansion must also be dominated by the increment positions of any
5073 // loops it for which it is using post-inc mode.
5074 for (const Loop *PIL : LF.PostIncLoops) {
5075 if (PIL == L) continue;
5076
5077 // Be dominated by the loop exit.
5078 SmallVector<BasicBlock *, 4> ExitingBlocks;
5079 PIL->getExitingBlocks(ExitingBlocks);
5080 if (!ExitingBlocks.empty()) {
5081 BasicBlock *BB = ExitingBlocks[0];
5082 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5083 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5084 Inputs.push_back(BB->getTerminator());
5085 }
5086 }
5087
5088 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5090, __PRETTY_FUNCTION__))
5089 && !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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5090, __PRETTY_FUNCTION__))
5090 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5090, __PRETTY_FUNCTION__))
;
5091
5092 // Then, climb up the immediate dominator tree as far as we can go while
5093 // still being dominated by the input positions.
5094 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5095
5096 // Don't insert instructions before PHI nodes.
5097 while (isa<PHINode>(IP)) ++IP;
5098
5099 // Ignore landingpad instructions.
5100 while (IP->isEHPad()) ++IP;
5101
5102 // Ignore debug intrinsics.
5103 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5104
5105 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5106 // IP consistent across expansions and allows the previously inserted
5107 // instructions to be reused by subsequent expansion.
5108 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5109 ++IP;
5110
5111 return IP;
5112}
5113
5114/// Emit instructions for the leading candidate expression for this LSRUse (this
5115/// is called "expanding").
5116Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5117 const Formula &F, BasicBlock::iterator IP,
5118 SCEVExpander &Rewriter,
5119 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5120 if (LU.RigidFormula)
5121 return LF.OperandValToReplace;
5122
5123 // Determine an input position which will be dominated by the operands and
5124 // which will dominate the result.
5125 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5126 Rewriter.setInsertPoint(&*IP);
5127
5128 // Inform the Rewriter if we have a post-increment use, so that it can
5129 // perform an advantageous expansion.
5130 Rewriter.setPostInc(LF.PostIncLoops);
5131
5132 // This is the type that the user actually needs.
5133 Type *OpTy = LF.OperandValToReplace->getType();
5134 // This will be the type that we'll initially expand to.
5135 Type *Ty = F.getType();
5136 if (!Ty)
5137 // No type known; just expand directly to the ultimate type.
5138 Ty = OpTy;
5139 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5140 // Expand directly to the ultimate type if it's the right size.
5141 Ty = OpTy;
5142 // This is the type to do integer arithmetic in.
5143 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5144
5145 // Build up a list of operands to add together to form the full base.
5146 SmallVector<const SCEV *, 8> Ops;
5147
5148 // Expand the BaseRegs portion.
5149 for (const SCEV *Reg : F.BaseRegs) {
5150 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5150, __PRETTY_FUNCTION__))
;
5151
5152 // If we're expanding for a post-inc user, make the post-inc adjustment.
5153 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5154 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5155 }
5156
5157 // Expand the ScaledReg portion.
5158 Value *ICmpScaledV = nullptr;
5159 if (F.Scale != 0) {
5160 const SCEV *ScaledS = F.ScaledReg;
5161
5162 // If we're expanding for a post-inc user, make the post-inc adjustment.
5163 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5164 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5165
5166 if (LU.Kind == LSRUse::ICmpZero) {
5167 // Expand ScaleReg as if it was part of the base regs.
5168 if (F.Scale == 1)
5169 Ops.push_back(
5170 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5171 else {
5172 // An interesting way of "folding" with an icmp is to use a negated
5173 // scale, which we'll implement by inserting it into the other operand
5174 // of the icmp.
5175 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5176, __PRETTY_FUNCTION__))
5176 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5176, __PRETTY_FUNCTION__))
;
5177 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5178 }
5179 } else {
5180 // Otherwise just expand the scaled register and an explicit scale,
5181 // which is expected to be matched as part of the address.
5182
5183 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5184 // Unless the addressing mode will not be folded.
5185 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5186 isAMCompletelyFolded(TTI, LU, F)) {
5187 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5188 Ops.clear();
5189 Ops.push_back(SE.getUnknown(FullV));
5190 }
5191 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5192 if (F.Scale != 1)
5193 ScaledS =
5194 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5195 Ops.push_back(ScaledS);
5196 }
5197 }
5198
5199 // Expand the GV portion.
5200 if (F.BaseGV) {
5201 // Flush the operand list to suppress SCEVExpander hoisting.
5202 if (!Ops.empty()) {
5203 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5204 Ops.clear();
5205 Ops.push_back(SE.getUnknown(FullV));
5206 }
5207 Ops.push_back(SE.getUnknown(F.BaseGV));
5208 }
5209
5210 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5211 // unfolded offsets. LSR assumes they both live next to their uses.
5212 if (!Ops.empty()) {
5213 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5214 Ops.clear();
5215 Ops.push_back(SE.getUnknown(FullV));
5216 }
5217
5218 // Expand the immediate portion.
5219 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5220 if (Offset != 0) {
5221 if (LU.Kind == LSRUse::ICmpZero) {
5222 // The other interesting way of "folding" with an ICmpZero is to use a
5223 // negated immediate.
5224 if (!ICmpScaledV)
5225 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5226 else {
5227 Ops.push_back(SE.getUnknown(ICmpScaledV));
5228 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5229 }
5230 } else {
5231 // Just add the immediate values. These again are expected to be matched
5232 // as part of the address.
5233 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5234 }
5235 }
5236
5237 // Expand the unfolded offset portion.
5238 int64_t UnfoldedOffset = F.UnfoldedOffset;
5239 if (UnfoldedOffset != 0) {
5240 // Just add the immediate values.
5241 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5242 UnfoldedOffset)));
5243 }
5244
5245 // Emit instructions summing all the operands.
5246 const SCEV *FullS = Ops.empty() ?
5247 SE.getConstant(IntTy, 0) :
5248 SE.getAddExpr(Ops);
5249 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5250
5251 // We're done expanding now, so reset the rewriter.
5252 Rewriter.clearPostInc();
5253
5254 // An ICmpZero Formula represents an ICmp which we're handling as a
5255 // comparison against zero. Now that we've expanded an expression for that
5256 // form, update the ICmp's other operand.
5257 if (LU.Kind == LSRUse::ICmpZero) {
5258 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5259 if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
5260 DeadInsts.emplace_back(OperandIsInstr);
5261 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5262, __PRETTY_FUNCTION__))
5262 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5262, __PRETTY_FUNCTION__))
;
5263 if (F.Scale == -1) {
5264 if (ICmpScaledV->getType() != OpTy) {
5265 Instruction *Cast =
5266 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5267 OpTy, false),
5268 ICmpScaledV, OpTy, "tmp", CI);
5269 ICmpScaledV = Cast;
5270 }
5271 CI->setOperand(1, ICmpScaledV);
5272 } else {
5273 // A scale of 1 means that the scale has been expanded as part of the
5274 // base regs.
5275 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5277, __PRETTY_FUNCTION__))
5276 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5277, __PRETTY_FUNCTION__))
5277 "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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5277, __PRETTY_FUNCTION__))
;
5278 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5279 -(uint64_t)Offset);
5280 if (C->getType() != OpTy)
5281 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5282 OpTy, false),
5283 C, OpTy);
5284
5285 CI->setOperand(1, C);
5286 }
5287 }
5288
5289 return FullV;
5290}
5291
5292/// Helper for Rewrite. PHI nodes are special because the use of their operands
5293/// effectively happens in their predecessor blocks, so the expression may need
5294/// to be expanded in multiple places.
5295void LSRInstance::RewriteForPHI(
5296 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5297 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5298 DenseMap<BasicBlock *, Value *> Inserted;
5299 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1
Assuming 'i' is not equal to 'e'
2
Loop condition is true. Entering loop body
5300 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3
Assuming pointer value is null
4
Taking true branch
5301 bool needUpdateFixups = false;
5302 BasicBlock *BB = PN->getIncomingBlock(i);
5303
5304 // If this is a critical edge, split the edge so that we do not insert
5305 // the code on all predecessor/successor paths. We do this unless this
5306 // is the canonical backedge for this loop, which complicates post-inc
5307 // users.
5308 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5
Assuming 'e' is equal to 1
6
Taking false branch
5309 !isa<IndirectBrInst>(BB->getTerminator()) &&
5310 !isa<CatchSwitchInst>(BB->getTerminator())) {
5311 BasicBlock *Parent = PN->getParent();
5312 Loop *PNLoop = LI.getLoopFor(Parent);
5313 if (!PNLoop || Parent != PNLoop->getHeader()) {
5314 // Split the critical edge.
5315 BasicBlock *NewBB = nullptr;
5316 if (!Parent->isLandingPad()) {
5317 NewBB = SplitCriticalEdge(BB, Parent,
5318 CriticalEdgeSplittingOptions(&DT, &LI)
5319 .setMergeIdenticalEdges()
5320 .setKeepOneInputPHIs());
5321 } else {
5322 SmallVector<BasicBlock*, 2> NewBBs;
5323 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5324 NewBB = NewBBs[0];
5325 }
5326 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5327 // phi predecessors are identical. The simple thing to do is skip
5328 // splitting in this case rather than complicate the API.
5329 if (NewBB) {
5330 // If PN is outside of the loop and BB is in the loop, we want to
5331 // move the block to be immediately before the PHI block, not
5332 // immediately after BB.
5333 if (L->contains(BB) && !L->contains(PN))
5334 NewBB->moveBefore(PN->getParent());
5335
5336 // Splitting the edge can reduce the number of PHI entries we have.
5337 e = PN->getNumIncomingValues();
5338 BB = NewBB;
5339 i = PN->getBasicBlockIndex(BB);
5340
5341 needUpdateFixups = true;
5342 }
5343 }
5344 }
5345
5346 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5347 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5348 if (!Pair.second)
7
Assuming field 'second' is true
8
Taking false branch
5349 PN->setIncomingValue(i, Pair.first->second);
5350 else {
5351 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5352 Rewriter, DeadInsts);
5353
5354 // If this is reuse-by-noop-cast, insert the noop cast.
5355 Type *OpTy = LF.OperandValToReplace->getType();
9
Called C++ object pointer is null
5356 if (FullV->getType() != OpTy)
5357 FullV =
5358 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5359 OpTy, false),
5360 FullV, LF.OperandValToReplace->getType(),
5361 "tmp", BB->getTerminator());
5362
5363 PN->setIncomingValue(i, FullV);
5364 Pair.first->second = FullV;
5365 }
5366
5367 // If LSR splits critical edge and phi node has other pending
5368 // fixup operands, we need to update those pending fixups. Otherwise
5369 // formulae will not be implemented completely and some instructions
5370 // will not be eliminated.
5371 if (needUpdateFixups) {
5372 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5373 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5374 // If fixup is supposed to rewrite some operand in the phi
5375 // that was just updated, it may be already moved to
5376 // another phi node. Such fixup requires update.
5377 if (Fixup.UserInst == PN) {
5378 // Check if the operand we try to replace still exists in the
5379 // original phi.
5380 bool foundInOriginalPHI = false;
5381 for (const auto &val : PN->incoming_values())
5382 if (val == Fixup.OperandValToReplace) {
5383 foundInOriginalPHI = true;
5384 break;
5385 }
5386
5387 // If fixup operand found in original PHI - nothing to do.
5388 if (foundInOriginalPHI)
5389 continue;
5390
5391 // Otherwise it might be moved to another PHI and requires update.
5392 // If fixup operand not found in any of the incoming blocks that
5393 // means we have already rewritten it - nothing to do.
5394 for (const auto &Block : PN->blocks())
5395 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5396 ++I) {
5397 PHINode *NewPN = cast<PHINode>(I);
5398 for (const auto &val : NewPN->incoming_values())
5399 if (val == Fixup.OperandValToReplace)
5400 Fixup.UserInst = NewPN;
5401 }
5402 }
5403 }
5404 }
5405}
5406
5407/// Emit instructions for the leading candidate expression for this LSRUse (this
5408/// is called "expanding"), and update the UserInst to reference the newly
5409/// expanded value.
5410void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5411 const Formula &F, SCEVExpander &Rewriter,
5412 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5413 // First, find an insertion point that dominates UserInst. For PHI nodes,
5414 // find the nearest block which dominates all the relevant uses.
5415 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5416 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5417 } else {
5418 Value *FullV =
5419 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5420
5421 // If this is reuse-by-noop-cast, insert the noop cast.
5422 Type *OpTy = LF.OperandValToReplace->getType();
5423 if (FullV->getType() != OpTy) {
5424 Instruction *Cast =
5425 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5426 FullV, OpTy, "tmp", LF.UserInst);
5427 FullV = Cast;
5428 }
5429
5430 // Update the user. ICmpZero is handled specially here (for now) because
5431 // Expand may have updated one of the operands of the icmp already, and
5432 // its new value may happen to be equal to LF.OperandValToReplace, in
5433 // which case doing replaceUsesOfWith leads to replacing both operands
5434 // with the same value. TODO: Reorganize this.
5435 if (LU.Kind == LSRUse::ICmpZero)
5436 LF.UserInst->setOperand(0, FullV);
5437 else
5438 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5439 }
5440
5441 if (auto *OperandIsInstr = dyn_cast<Instruction>(LF.OperandValToReplace))
5442 DeadInsts.emplace_back(OperandIsInstr);
5443}
5444
5445/// Rewrite all the fixup locations with new values, following the chosen
5446/// solution.
5447void LSRInstance::ImplementSolution(
5448 const SmallVectorImpl<const Formula *> &Solution) {
5449 // Keep track of instructions we may have made dead, so that
5450 // we can remove them after we are done working.
5451 SmallVector<WeakTrackingVH, 16> DeadInsts;
5452
5453 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5454 "lsr");
5455#ifndef NDEBUG
5456 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5457#endif
5458 Rewriter.disableCanonicalMode();
5459 Rewriter.enableLSRMode();
5460 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5461
5462 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5463 for (const IVChain &Chain : IVChainVec) {
5464 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5465 Rewriter.setChainedPhi(PN);
5466 }
5467
5468 // Expand the new value definitions and update the users.
5469 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5470 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5471 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5472 Changed = true;
5473 }
5474
5475 for (const IVChain &Chain : IVChainVec) {
5476 GenerateIVChain(Chain, Rewriter, DeadInsts);
5477 Changed = true;
5478 }
5479 // Clean up after ourselves. This must be done before deleting any
5480 // instructions.
5481 Rewriter.clear();
5482
5483 Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts,
5484 &TLI, MSSAU);
5485}
5486
5487LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5488 DominatorTree &DT, LoopInfo &LI,
5489 const TargetTransformInfo &TTI, AssumptionCache &AC,
5490 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU)
5491 : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), TLI(TLI), TTI(TTI), L(L),
5492 MSSAU(MSSAU), FavorBackedgeIndex(EnableBackedgeIndexing &&
5493 TTI.shouldFavorBackedgeIndex(L)) {
5494 // If LoopSimplify form is not available, stay out of trouble.
5495 if (!L->isLoopSimplifyForm())
5496 return;
5497
5498 // If there's no interesting work to be done, bail early.
5499 if (IU.empty()) return;
5500
5501 // If there's too much analysis to be done, bail early. We won't be able to
5502 // model the problem anyway.
5503 unsigned NumUsers = 0;
5504 for (const IVStrideUse &U : IU) {
5505 if (++NumUsers > MaxIVUsers) {
5506 (void)U;
5507 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)
5508 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5509 return;
5510 }
5511 // Bail out if we have a PHI on an EHPad that gets a value from a
5512 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5513 // no good place to stick any instructions.
5514 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5515 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5516 if (isa<FuncletPadInst>(FirstNonPHI) ||
5517 isa<CatchSwitchInst>(FirstNonPHI))
5518 for (BasicBlock *PredBB : PN->blocks())
5519 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5520 return;
5521 }
5522 }
5523
5524#ifndef NDEBUG
5525 // All dominating loops must have preheaders, or SCEVExpander may not be able
5526 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5527 //
5528 // IVUsers analysis should only create users that are dominated by simple loop
5529 // headers. Since this loop should dominate all of its users, its user list
5530 // should be empty if this loop itself is not within a simple loop nest.
5531 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5532 Rung; Rung = Rung->getIDom()) {
5533 BasicBlock *BB = Rung->getBlock();
5534 const Loop *DomLoop = LI.getLoopFor(BB);
5535 if (DomLoop && DomLoop->getHeader() == BB) {
5536 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5536, __PRETTY_FUNCTION__))
;
5537 }
5538 }
5539#endif // DEBUG
5540
5541 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)
5542 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)
5543 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5544
5545 // First, perform some low-level loop optimizations.
5546 OptimizeShadowIV();
5547 OptimizeLoopTermCond();
5548
5549 // If loop preparation eliminates all interesting IV users, bail.
5550 if (IU.empty()) return;
5551
5552 // Skip nested loops until we can model them better with formulae.
5553 if (!L->empty()) {
5554 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)
;
5555 return;
5556 }
5557
5558 // Start collecting data and preparing for the solver.
5559 CollectChains();
5560 CollectInterestingTypesAndFactors();
5561 CollectFixupsAndInitialFormulae();
5562 CollectLoopInvariantFixupsAndFormulae();
5563
5564 if (Uses.empty())
5565 return;
5566
5567 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
)
5568 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
;
5569
5570 // Now use the reuse data to generate a bunch of interesting ways
5571 // to formulate the values needed for the uses.
5572 GenerateAllReuseFormulae();
5573
5574 FilterOutUndesirableDedicatedRegisters();
5575 NarrowSearchSpaceUsingHeuristics();
5576
5577 SmallVector<const Formula *, 8> Solution;
5578 Solve(Solution);
5579
5580 // Release memory that is no longer needed.
5581 Factors.clear();
5582 Types.clear();
5583 RegUses.clear();
5584
5585 if (Solution.empty())
5586 return;
5587
5588#ifndef NDEBUG
5589 // Formulae should be legal.
5590 for (const LSRUse &LU : Uses) {
5591 for (const Formula &F : LU.Formulae)
5592 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5593, __PRETTY_FUNCTION__))
5593 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-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5593, __PRETTY_FUNCTION__))
;
5594 };
5595#endif
5596
5597 // Now that we've decided what we want, make it so.
5598 ImplementSolution(Solution);
5599}
5600
5601#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5602void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5603 if (Factors.empty() && Types.empty()) return;
5604
5605 OS << "LSR has identified the following interesting factors and types: ";
5606 bool First = true;
5607
5608 for (int64_t Factor : Factors) {
5609 if (!First) OS << ", ";
5610 First = false;
5611 OS << '*' << Factor;
5612 }
5613
5614 for (Type *Ty : Types) {
5615 if (!First) OS << ", ";
5616 First = false;
5617 OS << '(' << *Ty << ')';
5618 }
5619 OS << '\n';
5620}
5621
5622void LSRInstance::print_fixups(raw_ostream &OS) const {
5623 OS << "LSR is examining the following fixup sites:\n";
5624 for (const LSRUse &LU : Uses)
5625 for (const LSRFixup &LF : LU.Fixups) {
5626 dbgs() << " ";
5627 LF.print(OS);
5628 OS << '\n';
5629 }
5630}
5631
5632void LSRInstance::print_uses(raw_ostream &OS) const {
5633 OS << "LSR is examining the following uses:\n";
5634 for (const LSRUse &LU : Uses) {
5635 dbgs() << " ";
5636 LU.print(OS);
5637 OS << '\n';
5638 for (const Formula &F : LU.Formulae) {
5639 OS << " ";
5640 F.print(OS);
5641 OS << '\n';
5642 }
5643 }
5644}
5645
5646void LSRInstance::print(raw_ostream &OS) const {
5647 print_factors_and_types(OS);
5648 print_fixups(OS);
5649 print_uses(OS);
5650}
5651
5652LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRInstance::dump() const {
5653 print(errs()); errs() << '\n';
5654}
5655#endif
5656
5657namespace {
5658
5659class LoopStrengthReduce : public LoopPass {
5660public:
5661 static char ID; // Pass ID, replacement for typeid
5662
5663 LoopStrengthReduce();
5664
5665private:
5666 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5667 void getAnalysisUsage(AnalysisUsage &AU) const override;
5668};
5669
5670} // end anonymous namespace
5671
5672LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5673 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5674}
5675
5676void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5677 // We split critical edges, so we change the CFG. However, we do update
5678 // many analyses if they are around.
5679 AU.addPreservedID(LoopSimplifyID);
5680
5681 AU.addRequired<LoopInfoWrapperPass>();
5682 AU.addPreserved<LoopInfoWrapperPass>();
5683 AU.addRequiredID(LoopSimplifyID);
5684 AU.addRequired<DominatorTreeWrapperPass>();
5685 AU.addPreserved<DominatorTreeWrapperPass>();
5686 AU.addRequired<ScalarEvolutionWrapperPass>();
5687 AU.addPreserved<ScalarEvolutionWrapperPass>();
5688 AU.addRequired<AssumptionCacheTracker>();
5689 AU.addRequired<TargetLibraryInfoWrapperPass>();
5690 // Requiring LoopSimplify a second time here prevents IVUsers from running
5691 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5692 AU.addRequiredID(LoopSimplifyID);
5693 AU.addRequired<IVUsersWrapperPass>();
5694 AU.addPreserved<IVUsersWrapperPass>();
5695 AU.addRequired<TargetTransformInfoWrapperPass>();
5696 AU.addPreserved<MemorySSAWrapperPass>();
5697}
5698
5699static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5700 DominatorTree &DT, LoopInfo &LI,
5701 const TargetTransformInfo &TTI,
5702 AssumptionCache &AC, TargetLibraryInfo &TLI,
5703 MemorySSA *MSSA) {
5704
5705 bool Changed = false;
5706 std::unique_ptr<MemorySSAUpdater> MSSAU;
5707 if (MSSA)
5708 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
5709
5710 // Run the main LSR transformation.
5711 Changed |=
5712 LSRInstance(L, IU, SE, DT, LI, TTI, AC, TLI, MSSAU.get()).getChanged();
5713
5714 // Remove any extra phis created by processing inner loops.
5715 Changed |= DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
5716 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5717 SmallVector<WeakTrackingVH, 16> DeadInsts;
5718 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5719 SCEVExpander Rewriter(SE, DL, "lsr");
5720#ifndef NDEBUG
5721 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5722#endif
5723 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5724 if (numFolded) {
5725 Changed = true;
5726 RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, &TLI,
5727 MSSAU.get());
5728 DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
5729 }
5730 }
5731 return Changed;
5732}
5733
5734bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5735 if (skipLoop(L))
5736 return false;
5737
5738 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5739 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5740 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5741 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5742 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5743 *L->getHeader()->getParent());
5744 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
5745 *L->getHeader()->getParent());
5746 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
5747 *L->getHeader()->getParent());
5748 auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>();
5749 MemorySSA *MSSA = nullptr;
5750 if (MSSAAnalysis)
5751 MSSA = &MSSAAnalysis->getMSSA();
5752 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI, AC, TLI, MSSA);
5753}
5754
5755PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5756 LoopStandardAnalysisResults &AR,
5757 LPMUpdater &) {
5758 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5759 AR.DT, AR.LI, AR.TTI, AR.AC, AR.TLI, AR.MSSA))
5760 return PreservedAnalyses::all();
5761
5762 auto PA = getLoopPassPreservedAnalyses();
5763 if (AR.MSSA)
5764 PA.preserve<MemorySSAAnalysis>();
5765 return PA;
5766}