Bug Summary

File:llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 5425, column 22
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -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 -fhalf-no-semantic-interposition -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/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-12/lib/clang/12.0.0/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-12~++20210124100612+2afaf072f5c1/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2021-01-24-223304-31662-1 -x c++ /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/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/AssumptionCache.h"
69#include "llvm/Analysis/IVUsers.h"
70#include "llvm/Analysis/LoopAnalysisManager.h"
71#include "llvm/Analysis/LoopInfo.h"
72#include "llvm/Analysis/LoopPass.h"
73#include "llvm/Analysis/MemorySSA.h"
74#include "llvm/Analysis/MemorySSAUpdater.h"
75#include "llvm/Analysis/ScalarEvolution.h"
76#include "llvm/Analysis/ScalarEvolutionExpressions.h"
77#include "llvm/Analysis/ScalarEvolutionNormalization.h"
78#include "llvm/Analysis/TargetLibraryInfo.h"
79#include "llvm/Analysis/TargetTransformInfo.h"
80#include "llvm/Config/llvm-config.h"
81#include "llvm/IR/BasicBlock.h"
82#include "llvm/IR/Constant.h"
83#include "llvm/IR/Constants.h"
84#include "llvm/IR/DebugInfoMetadata.h"
85#include "llvm/IR/DerivedTypes.h"
86#include "llvm/IR/Dominators.h"
87#include "llvm/IR/GlobalValue.h"
88#include "llvm/IR/IRBuilder.h"
89#include "llvm/IR/InstrTypes.h"
90#include "llvm/IR/Instruction.h"
91#include "llvm/IR/Instructions.h"
92#include "llvm/IR/IntrinsicInst.h"
93#include "llvm/IR/Intrinsics.h"
94#include "llvm/IR/Module.h"
95#include "llvm/IR/OperandTraits.h"
96#include "llvm/IR/Operator.h"
97#include "llvm/IR/PassManager.h"
98#include "llvm/IR/Type.h"
99#include "llvm/IR/Use.h"
100#include "llvm/IR/User.h"
101#include "llvm/IR/Value.h"
102#include "llvm/IR/ValueHandle.h"
103#include "llvm/InitializePasses.h"
104#include "llvm/Pass.h"
105#include "llvm/Support/Casting.h"
106#include "llvm/Support/CommandLine.h"
107#include "llvm/Support/Compiler.h"
108#include "llvm/Support/Debug.h"
109#include "llvm/Support/ErrorHandling.h"
110#include "llvm/Support/MathExtras.h"
111#include "llvm/Support/raw_ostream.h"
112#include "llvm/Transforms/Scalar.h"
113#include "llvm/Transforms/Utils.h"
114#include "llvm/Transforms/Utils/BasicBlockUtils.h"
115#include "llvm/Transforms/Utils/Local.h"
116#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
117#include <algorithm>
118#include <cassert>
119#include <cstddef>
120#include <cstdint>
121#include <cstdlib>
122#include <iterator>
123#include <limits>
124#include <map>
125#include <numeric>
126#include <utility>
127
128using namespace llvm;
129
130#define DEBUG_TYPE"loop-reduce" "loop-reduce"
131
132/// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
133/// bail out. This threshold is far beyond the number of users that LSR can
134/// conceivably solve, so it should not affect generated code, but catches the
135/// worst cases before LSR burns too much compile time and stack space.
136static const unsigned MaxIVUsers = 200;
137
138// Temporary flag to cleanup congruent phis after LSR phi expansion.
139// It's currently disabled until we can determine whether it's truly useful or
140// not. The flag should be removed after the v3.0 release.
141// This is now needed for ivchains.
142static cl::opt<bool> EnablePhiElim(
143 "enable-lsr-phielim", cl::Hidden, cl::init(true),
144 cl::desc("Enable LSR phi elimination"));
145
146// The flag adds instruction count to solutions cost comparision.
147static cl::opt<bool> InsnsCost(
148 "lsr-insns-cost", cl::Hidden, cl::init(true),
149 cl::desc("Add instruction count to a LSR cost model"));
150
151// Flag to choose how to narrow complex lsr solution
152static cl::opt<bool> LSRExpNarrow(
153 "lsr-exp-narrow", cl::Hidden, cl::init(false),
154 cl::desc("Narrow LSR complex solution using"
155 " expectation of registers number"));
156
157// Flag to narrow search space by filtering non-optimal formulae with
158// the same ScaledReg and Scale.
159static cl::opt<bool> FilterSameScaledReg(
160 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
161 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
162 " with the same ScaledReg and Scale"));
163
164static cl::opt<bool> EnableBackedgeIndexing(
165 "lsr-backedge-indexing", cl::Hidden, cl::init(true),
166 cl::desc("Enable the generation of cross iteration indexed memops"));
167
168static cl::opt<unsigned> ComplexityLimit(
169 "lsr-complexity-limit", cl::Hidden,
170 cl::init(std::numeric_limits<uint16_t>::max()),
171 cl::desc("LSR search space complexity limit"));
172
173static cl::opt<unsigned> SetupCostDepthLimit(
174 "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
175 cl::desc("The limit on recursion depth for LSRs setup cost"));
176
177#ifndef NDEBUG
178// Stress test IV chain generation.
179static cl::opt<bool> StressIVChain(
180 "stress-ivchain", cl::Hidden, cl::init(false),
181 cl::desc("Stress test LSR IV chains"));
182#else
183static bool StressIVChain = false;
184#endif
185
186namespace {
187
188struct MemAccessTy {
189 /// Used in situations where the accessed memory type is unknown.
190 static const unsigned UnknownAddressSpace =
191 std::numeric_limits<unsigned>::max();
192
193 Type *MemTy = nullptr;
194 unsigned AddrSpace = UnknownAddressSpace;
195
196 MemAccessTy() = default;
197 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
198
199 bool operator==(MemAccessTy Other) const {
200 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
201 }
202
203 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
204
205 static MemAccessTy getUnknown(LLVMContext &Ctx,
206 unsigned AS = UnknownAddressSpace) {
207 return MemAccessTy(Type::getVoidTy(Ctx), AS);
208 }
209
210 Type *getType() { return MemTy; }
211};
212
213/// This class holds data which is used to order reuse candidates.
214class RegSortData {
215public:
216 /// This represents the set of LSRUse indices which reference
217 /// a particular register.
218 SmallBitVector UsedByIndices;
219
220 void print(raw_ostream &OS) const;
221 void dump() const;
222};
223
224} // end anonymous namespace
225
226#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
227void RegSortData::print(raw_ostream &OS) const {
228 OS << "[NumUses=" << UsedByIndices.count() << ']';
229}
230
231LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
232 print(errs()); errs() << '\n';
233}
234#endif
235
236namespace {
237
238/// Map register candidates to information about how they are used.
239class RegUseTracker {
240 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
241
242 RegUsesTy RegUsesMap;
243 SmallVector<const SCEV *, 16> RegSequence;
244
245public:
246 void countRegister(const SCEV *Reg, size_t LUIdx);
247 void dropRegister(const SCEV *Reg, size_t LUIdx);
248 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
249
250 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
251
252 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
253
254 void clear();
255
256 using iterator = SmallVectorImpl<const SCEV *>::iterator;
257 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
258
259 iterator begin() { return RegSequence.begin(); }
260 iterator end() { return RegSequence.end(); }
261 const_iterator begin() const { return RegSequence.begin(); }
262 const_iterator end() const { return RegSequence.end(); }
263};
264
265} // end anonymous namespace
266
267void
268RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
269 std::pair<RegUsesTy::iterator, bool> Pair =
270 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
271 RegSortData &RSD = Pair.first->second;
272 if (Pair.second)
273 RegSequence.push_back(Reg);
274 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
275 RSD.UsedByIndices.set(LUIdx);
276}
277
278void
279RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
280 RegUsesTy::iterator It = RegUsesMap.find(Reg);
281 assert(It != RegUsesMap.end())((It != RegUsesMap.end()) ? static_cast<void> (0) : __assert_fail
("It != RegUsesMap.end()", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 281, __PRETTY_FUNCTION__))
;
282 RegSortData &RSD = It->second;
283 assert(RSD.UsedByIndices.size() > LUIdx)((RSD.UsedByIndices.size() > LUIdx) ? static_cast<void>
(0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 283, __PRETTY_FUNCTION__))
;
284 RSD.UsedByIndices.reset(LUIdx);
285}
286
287void
288RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
289 assert(LUIdx <= LastLUIdx)((LUIdx <= LastLUIdx) ? static_cast<void> (0) : __assert_fail
("LUIdx <= LastLUIdx", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 289, __PRETTY_FUNCTION__))
;
290
291 // Update RegUses. The data structure is not optimized for this purpose;
292 // we must iterate through it and update each of the bit vectors.
293 for (auto &Pair : RegUsesMap) {
294 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
295 if (LUIdx < UsedByIndices.size())
296 UsedByIndices[LUIdx] =
297 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
298 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
299 }
300}
301
302bool
303RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
304 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
305 if (I == RegUsesMap.end())
306 return false;
307 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
308 int i = UsedByIndices.find_first();
309 if (i == -1) return false;
310 if ((size_t)i != LUIdx) return true;
311 return UsedByIndices.find_next(i) != -1;
312}
313
314const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
315 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
316 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 316, __PRETTY_FUNCTION__))
;
317 return I->second.UsedByIndices;
318}
319
320void RegUseTracker::clear() {
321 RegUsesMap.clear();
322 RegSequence.clear();
323}
324
325namespace {
326
327/// This class holds information that describes a formula for computing
328/// satisfying a use. It may include broken-out immediates and scaled registers.
329struct Formula {
330 /// Global base address used for complex addressing.
331 GlobalValue *BaseGV = nullptr;
332
333 /// Base offset for complex addressing.
334 int64_t BaseOffset = 0;
335
336 /// Whether any complex addressing has a base register.
337 bool HasBaseReg = false;
338
339 /// The scale of any complex addressing.
340 int64_t Scale = 0;
341
342 /// The list of "base" registers for this use. When this is non-empty. The
343 /// canonical representation of a formula is
344 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
345 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
346 /// 3. The reg containing recurrent expr related with currect loop in the
347 /// formula should be put in the ScaledReg.
348 /// #1 enforces that the scaled register is always used when at least two
349 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
350 /// #2 enforces that 1 * reg is reg.
351 /// #3 ensures invariant regs with respect to current loop can be combined
352 /// together in LSR codegen.
353 /// This invariant can be temporarily broken while building a formula.
354 /// However, every formula inserted into the LSRInstance must be in canonical
355 /// form.
356 SmallVector<const SCEV *, 4> BaseRegs;
357
358 /// The 'scaled' register for this use. This should be non-null when Scale is
359 /// not zero.
360 const SCEV *ScaledReg = nullptr;
361
362 /// An additional constant offset which added near the use. This requires a
363 /// temporary register, but the offset itself can live in an add immediate
364 /// field rather than a register.
365 int64_t UnfoldedOffset = 0;
366
367 Formula() = default;
368
369 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
370
371 bool isCanonical(const Loop &L) const;
372
373 void canonicalize(const Loop &L);
374
375 bool unscale();
376
377 bool hasZeroEnd() const;
378
379 size_t getNumRegs() const;
380 Type *getType() const;
381
382 void deleteBaseReg(const SCEV *&S);
383
384 bool referencesReg(const SCEV *S) const;
385 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
386 const RegUseTracker &RegUses) const;
387
388 void print(raw_ostream &OS) const;
389 void dump() const;
390};
391
392} // end anonymous namespace
393
394/// Recursion helper for initialMatch.
395static void DoInitialMatch(const SCEV *S, Loop *L,
396 SmallVectorImpl<const SCEV *> &Good,
397 SmallVectorImpl<const SCEV *> &Bad,
398 ScalarEvolution &SE) {
399 // Collect expressions which properly dominate the loop header.
400 if (SE.properlyDominates(S, L->getHeader())) {
401 Good.push_back(S);
402 return;
403 }
404
405 // Look at add operands.
406 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
407 for (const SCEV *S : Add->operands())
408 DoInitialMatch(S, L, Good, Bad, SE);
409 return;
410 }
411
412 // Look at addrec operands.
413 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
414 if (!AR->getStart()->isZero() && AR->isAffine()) {
415 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
416 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
417 AR->getStepRecurrence(SE),
418 // FIXME: AR->getNoWrapFlags()
419 AR->getLoop(), SCEV::FlagAnyWrap),
420 L, Good, Bad, SE);
421 return;
422 }
423
424 // Handle a multiplication by -1 (negation) if it didn't fold.
425 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
426 if (Mul->getOperand(0)->isAllOnesValue()) {
427 SmallVector<const SCEV *, 4> Ops(drop_begin(Mul->operands()));
428 const SCEV *NewMul = SE.getMulExpr(Ops);
429
430 SmallVector<const SCEV *, 4> MyGood;
431 SmallVector<const SCEV *, 4> MyBad;
432 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
433 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
434 SE.getEffectiveSCEVType(NewMul->getType())));
435 for (const SCEV *S : MyGood)
436 Good.push_back(SE.getMulExpr(NegOne, S));
437 for (const SCEV *S : MyBad)
438 Bad.push_back(SE.getMulExpr(NegOne, S));
439 return;
440 }
441
442 // Ok, we can't do anything interesting. Just stuff the whole thing into a
443 // register and hope for the best.
444 Bad.push_back(S);
445}
446
447/// Incorporate loop-variant parts of S into this Formula, attempting to keep
448/// all loop-invariant and loop-computable values in a single base register.
449void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
450 SmallVector<const SCEV *, 4> Good;
451 SmallVector<const SCEV *, 4> Bad;
452 DoInitialMatch(S, L, Good, Bad, SE);
453 if (!Good.empty()) {
454 const SCEV *Sum = SE.getAddExpr(Good);
455 if (!Sum->isZero())
456 BaseRegs.push_back(Sum);
457 HasBaseReg = true;
458 }
459 if (!Bad.empty()) {
460 const SCEV *Sum = SE.getAddExpr(Bad);
461 if (!Sum->isZero())
462 BaseRegs.push_back(Sum);
463 HasBaseReg = true;
464 }
465 canonicalize(*L);
466}
467
468/// Check whether or not this formula satisfies the canonical
469/// representation.
470/// \see Formula::BaseRegs.
471bool Formula::isCanonical(const Loop &L) const {
472 if (!ScaledReg)
473 return BaseRegs.size() <= 1;
474
475 if (Scale != 1)
476 return true;
477
478 if (Scale == 1 && BaseRegs.empty())
479 return false;
480
481 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
482 if (SAR && SAR->getLoop() == &L)
483 return true;
484
485 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
486 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
487 // loop, we want to swap the reg in BaseRegs with ScaledReg.
488 auto I = find_if(BaseRegs, [&](const SCEV *S) {
489 return isa<const SCEVAddRecExpr>(S) &&
490 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
491 });
492 return I == BaseRegs.end();
493}
494
495/// Helper method to morph a formula into its canonical representation.
496/// \see Formula::BaseRegs.
497/// Every formula having more than one base register, must use the ScaledReg
498/// field. Otherwise, we would have to do special cases everywhere in LSR
499/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
500/// On the other hand, 1*reg should be canonicalized into reg.
501void Formula::canonicalize(const Loop &L) {
502 if (isCanonical(L))
503 return;
504 // So far we did not need this case. This is easy to implement but it is
505 // useless to maintain dead code. Beside it could hurt compile time.
506 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 506, __PRETTY_FUNCTION__))
;
507
508 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
509 if (!ScaledReg) {
510 ScaledReg = BaseRegs.pop_back_val();
511 Scale = 1;
512 }
513
514 // If ScaledReg is an invariant with respect to L, find the reg from
515 // BaseRegs containing the recurrent expr related with Loop L. Swap the
516 // reg with ScaledReg.
517 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
518 if (!SAR || SAR->getLoop() != &L) {
519 auto I = find_if(BaseRegs, [&](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->operands());
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->operands());
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->operands());
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->operands());
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 default:
938 break;
939 }
940
941 if (!Processed.insert(S).second)
942 return false;
943
944 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
945 for (const SCEV *S : Add->operands()) {
946 if (isHighCostExpansion(S, Processed, SE))
947 return true;
948 }
949 return false;
950 }
951
952 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
953 if (Mul->getNumOperands() == 2) {
954 // Multiplication by a constant is ok
955 if (isa<SCEVConstant>(Mul->getOperand(0)))
956 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
957
958 // If we have the value of one operand, check if an existing
959 // multiplication already generates this expression.
960 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
961 Value *UVal = U->getValue();
962 for (User *UR : UVal->users()) {
963 // If U is a constant, it may be used by a ConstantExpr.
964 Instruction *UI = dyn_cast<Instruction>(UR);
965 if (UI && UI->getOpcode() == Instruction::Mul &&
966 SE.isSCEVable(UI->getType())) {
967 return SE.getSCEV(UI) == Mul;
968 }
969 }
970 }
971 }
972 }
973
974 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
975 if (isExistingPhi(AR, SE))
976 return false;
977 }
978
979 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
980 return true;
981}
982
983namespace {
984
985class LSRUse;
986
987} // end anonymous namespace
988
989/// Check if the addressing mode defined by \p F is completely
990/// folded in \p LU at isel time.
991/// This includes address-mode folding and special icmp tricks.
992/// This function returns true if \p LU can accommodate what \p F
993/// defines and up to 1 base + 1 scaled + offset.
994/// In other words, if \p F has several base registers, this function may
995/// still return true. Therefore, users still need to account for
996/// additional base registers and/or unfolded offsets to derive an
997/// accurate cost model.
998static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
999 const LSRUse &LU, const Formula &F);
1000
1001// Get the cost of the scaling factor used in F for LU.
1002static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1003 const LSRUse &LU, const Formula &F,
1004 const Loop &L);
1005
1006namespace {
1007
1008/// This class is used to measure and compare candidate formulae.
1009class Cost {
1010 const Loop *L = nullptr;
1011 ScalarEvolution *SE = nullptr;
1012 const TargetTransformInfo *TTI = nullptr;
1013 TargetTransformInfo::LSRCost C;
1014
1015public:
1016 Cost() = delete;
1017 Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI) :
1018 L(L), SE(&SE), TTI(&TTI) {
1019 C.Insns = 0;
1020 C.NumRegs = 0;
1021 C.AddRecCost = 0;
1022 C.NumIVMuls = 0;
1023 C.NumBaseAdds = 0;
1024 C.ImmCost = 0;
1025 C.SetupCost = 0;
1026 C.ScaleCost = 0;
1027 }
1028
1029 bool isLess(Cost &Other);
1030
1031 void Lose();
1032
1033#ifndef NDEBUG
1034 // Once any of the metrics loses, they must all remain losers.
1035 bool isValid() {
1036 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1037 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1038 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1039 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1040 }
1041#endif
1042
1043 bool isLoser() {
1044 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1044, __PRETTY_FUNCTION__))
;
1045 return C.NumRegs == ~0u;
1046 }
1047
1048 void RateFormula(const Formula &F,
1049 SmallPtrSetImpl<const SCEV *> &Regs,
1050 const DenseSet<const SCEV *> &VisitedRegs,
1051 const LSRUse &LU,
1052 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1053
1054 void print(raw_ostream &OS) const;
1055 void dump() const;
1056
1057private:
1058 void RateRegister(const Formula &F, const SCEV *Reg,
1059 SmallPtrSetImpl<const SCEV *> &Regs);
1060 void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1061 SmallPtrSetImpl<const SCEV *> &Regs,
1062 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1063};
1064
1065/// An operand value in an instruction which is to be replaced with some
1066/// equivalent, possibly strength-reduced, replacement.
1067struct LSRFixup {
1068 /// The instruction which will be updated.
1069 Instruction *UserInst = nullptr;
1070
1071 /// The operand of the instruction which will be replaced. The operand may be
1072 /// used more than once; every instance will be replaced.
1073 Value *OperandValToReplace = nullptr;
1074
1075 /// If this user is to use the post-incremented value of an induction
1076 /// variable, this set is non-empty and holds the loops associated with the
1077 /// induction variable.
1078 PostIncLoopSet PostIncLoops;
1079
1080 /// A constant offset to be added to the LSRUse expression. This allows
1081 /// multiple fixups to share the same LSRUse with different offsets, for
1082 /// example in an unrolled loop.
1083 int64_t Offset = 0;
1084
1085 LSRFixup() = default;
1086
1087 bool isUseFullyOutsideLoop(const Loop *L) const;
1088
1089 void print(raw_ostream &OS) const;
1090 void dump() const;
1091};
1092
1093/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1094/// SmallVectors of const SCEV*.
1095struct UniquifierDenseMapInfo {
1096 static SmallVector<const SCEV *, 4> getEmptyKey() {
1097 SmallVector<const SCEV *, 4> V;
1098 V.push_back(reinterpret_cast<const SCEV *>(-1));
1099 return V;
1100 }
1101
1102 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1103 SmallVector<const SCEV *, 4> V;
1104 V.push_back(reinterpret_cast<const SCEV *>(-2));
1105 return V;
1106 }
1107
1108 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1109 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1110 }
1111
1112 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1113 const SmallVector<const SCEV *, 4> &RHS) {
1114 return LHS == RHS;
1115 }
1116};
1117
1118/// This class holds the state that LSR keeps for each use in IVUsers, as well
1119/// as uses invented by LSR itself. It includes information about what kinds of
1120/// things can be folded into the user, information about the user itself, and
1121/// information about how the use may be satisfied. TODO: Represent multiple
1122/// users of the same expression in common?
1123class LSRUse {
1124 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1125
1126public:
1127 /// An enum for a kind of use, indicating what types of scaled and immediate
1128 /// operands it might support.
1129 enum KindType {
1130 Basic, ///< A normal use, with no folding.
1131 Special, ///< A special case of basic, allowing -1 scales.
1132 Address, ///< An address use; folding according to TargetLowering
1133 ICmpZero ///< An equality icmp with both operands folded into one.
1134 // TODO: Add a generic icmp too?
1135 };
1136
1137 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1138
1139 KindType Kind;
1140 MemAccessTy AccessTy;
1141
1142 /// The list of operands which are to be replaced.
1143 SmallVector<LSRFixup, 8> Fixups;
1144
1145 /// Keep track of the min and max offsets of the fixups.
1146 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1147 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1148
1149 /// This records whether all of the fixups using this LSRUse are outside of
1150 /// the loop, in which case some special-case heuristics may be used.
1151 bool AllFixupsOutsideLoop = true;
1152
1153 /// RigidFormula is set to true to guarantee that this use will be associated
1154 /// with a single formula--the one that initially matched. Some SCEV
1155 /// expressions cannot be expanded. This allows LSR to consider the registers
1156 /// used by those expressions without the need to expand them later after
1157 /// changing the formula.
1158 bool RigidFormula = false;
1159
1160 /// This records the widest use type for any fixup using this
1161 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1162 /// fixup widths to be equivalent, because the narrower one may be relying on
1163 /// the implicit truncation to truncate away bogus bits.
1164 Type *WidestFixupType = nullptr;
1165
1166 /// A list of ways to build a value that can satisfy this user. After the
1167 /// list is populated, one of these is selected heuristically and used to
1168 /// formulate a replacement for OperandValToReplace in UserInst.
1169 SmallVector<Formula, 12> Formulae;
1170
1171 /// The set of register candidates used by all formulae in this LSRUse.
1172 SmallPtrSet<const SCEV *, 4> Regs;
1173
1174 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1175
1176 LSRFixup &getNewFixup() {
1177 Fixups.push_back(LSRFixup());
1178 return Fixups.back();
1179 }
1180
1181 void pushFixup(LSRFixup &f) {
1182 Fixups.push_back(f);
1183 if (f.Offset > MaxOffset)
1184 MaxOffset = f.Offset;
1185 if (f.Offset < MinOffset)
1186 MinOffset = f.Offset;
1187 }
1188
1189 bool HasFormulaWithSameRegs(const Formula &F) const;
1190 float getNotSelectedProbability(const SCEV *Reg) const;
1191 bool InsertFormula(const Formula &F, const Loop &L);
1192 void DeleteFormula(Formula &F);
1193 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1194
1195 void print(raw_ostream &OS) const;
1196 void dump() const;
1197};
1198
1199} // end anonymous namespace
1200
1201static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1202 LSRUse::KindType Kind, MemAccessTy AccessTy,
1203 GlobalValue *BaseGV, int64_t BaseOffset,
1204 bool HasBaseReg, int64_t Scale,
1205 Instruction *Fixup = nullptr);
1206
1207static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1208 if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1209 return 1;
1210 if (Depth == 0)
1211 return 0;
1212 if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1213 return getSetupCost(S->getStart(), Depth - 1);
1214 if (auto S = dyn_cast<SCEVIntegralCastExpr>(Reg))
1215 return getSetupCost(S->getOperand(), Depth - 1);
1216 if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1217 return std::accumulate(S->op_begin(), S->op_end(), 0,
1218 [&](unsigned i, const SCEV *Reg) {
1219 return i + getSetupCost(Reg, Depth - 1);
1220 });
1221 if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1222 return getSetupCost(S->getLHS(), Depth - 1) +
1223 getSetupCost(S->getRHS(), Depth - 1);
1224 return 0;
1225}
1226
1227/// Tally up interesting quantities from the given register.
1228void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1229 SmallPtrSetImpl<const SCEV *> &Regs) {
1230 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1231 // If this is an addrec for another loop, it should be an invariant
1232 // with respect to L since L is the innermost loop (at least
1233 // for now LSR only handles innermost loops).
1234 if (AR->getLoop() != L) {
1235 // If the AddRec exists, consider it's register free and leave it alone.
1236 if (isExistingPhi(AR, *SE) && !TTI->shouldFavorPostInc())
1237 return;
1238
1239 // It is bad to allow LSR for current loop to add induction variables
1240 // for its sibling loops.
1241 if (!AR->getLoop()->contains(L)) {
1242 Lose();
1243 return;
1244 }
1245
1246 // Otherwise, it will be an invariant with respect to Loop L.
1247 ++C.NumRegs;
1248 return;
1249 }
1250
1251 unsigned LoopCost = 1;
1252 if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1253 TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1254
1255 // If the step size matches the base offset, we could use pre-indexed
1256 // addressing.
1257 if (TTI->shouldFavorBackedgeIndex(L)) {
1258 if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1259 if (Step->getAPInt() == F.BaseOffset)
1260 LoopCost = 0;
1261 }
1262
1263 if (TTI->shouldFavorPostInc()) {
1264 const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1265 if (isa<SCEVConstant>(LoopStep)) {
1266 const SCEV *LoopStart = AR->getStart();
1267 if (!isa<SCEVConstant>(LoopStart) &&
1268 SE->isLoopInvariant(LoopStart, L))
1269 LoopCost = 0;
1270 }
1271 }
1272 }
1273 C.AddRecCost += LoopCost;
1274
1275 // Add the step value register, if it needs one.
1276 // TODO: The non-affine case isn't precisely modeled here.
1277 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1278 if (!Regs.count(AR->getOperand(1))) {
1279 RateRegister(F, AR->getOperand(1), Regs);
1280 if (isLoser())
1281 return;
1282 }
1283 }
1284 }
1285 ++C.NumRegs;
1286
1287 // Rough heuristic; favor registers which don't require extra setup
1288 // instructions in the preheader.
1289 C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1290 // Ensure we don't, even with the recusion limit, produce invalid costs.
1291 C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1292
1293 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1294 SE->hasComputableLoopEvolution(Reg, L);
1295}
1296
1297/// Record this register in the set. If we haven't seen it before, rate
1298/// it. Optional LoserRegs provides a way to declare any formula that refers to
1299/// one of those regs an instant loser.
1300void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1301 SmallPtrSetImpl<const SCEV *> &Regs,
1302 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1303 if (LoserRegs && LoserRegs->count(Reg)) {
1304 Lose();
1305 return;
1306 }
1307 if (Regs.insert(Reg).second) {
1308 RateRegister(F, Reg, Regs);
1309 if (LoserRegs && isLoser())
1310 LoserRegs->insert(Reg);
1311 }
1312}
1313
1314void Cost::RateFormula(const Formula &F,
1315 SmallPtrSetImpl<const SCEV *> &Regs,
1316 const DenseSet<const SCEV *> &VisitedRegs,
1317 const LSRUse &LU,
1318 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1319 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1319, __PRETTY_FUNCTION__))
;
1320 // Tally up the registers.
1321 unsigned PrevAddRecCost = C.AddRecCost;
1322 unsigned PrevNumRegs = C.NumRegs;
1323 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1324 if (const SCEV *ScaledReg = F.ScaledReg) {
1325 if (VisitedRegs.count(ScaledReg)) {
1326 Lose();
1327 return;
1328 }
1329 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1330 if (isLoser())
1331 return;
1332 }
1333 for (const SCEV *BaseReg : F.BaseRegs) {
1334 if (VisitedRegs.count(BaseReg)) {
1335 Lose();
1336 return;
1337 }
1338 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1339 if (isLoser())
1340 return;
1341 }
1342
1343 // Determine how many (unfolded) adds we'll need inside the loop.
1344 size_t NumBaseParts = F.getNumRegs();
1345 if (NumBaseParts > 1)
1346 // Do not count the base and a possible second register if the target
1347 // allows to fold 2 registers.
1348 C.NumBaseAdds +=
1349 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1350 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1351
1352 // Accumulate non-free scaling amounts.
1353 C.ScaleCost += getScalingFactorCost(*TTI, LU, F, *L);
1354
1355 // Tally up the non-zero immediates.
1356 for (const LSRFixup &Fixup : LU.Fixups) {
1357 int64_t O = Fixup.Offset;
1358 int64_t Offset = (uint64_t)O + F.BaseOffset;
1359 if (F.BaseGV)
1360 C.ImmCost += 64; // Handle symbolic values conservatively.
1361 // TODO: This should probably be the pointer size.
1362 else if (Offset != 0)
1363 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1364
1365 // Check with target if this offset with this instruction is
1366 // specifically not supported.
1367 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1368 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1369 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1370 C.NumBaseAdds++;
1371 }
1372
1373 // If we don't count instruction cost exit here.
1374 if (!InsnsCost) {
1375 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1375, __PRETTY_FUNCTION__))
;
1376 return;
1377 }
1378
1379 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1380 // additional instruction (at least fill).
1381 // TODO: Need distinguish register class?
1382 unsigned TTIRegNum = TTI->getNumberOfRegisters(
1383 TTI->getRegisterClassForType(false, F.getType())) - 1;
1384 if (C.NumRegs > TTIRegNum) {
1385 // Cost already exceeded TTIRegNum, then only newly added register can add
1386 // new instructions.
1387 if (PrevNumRegs > TTIRegNum)
1388 C.Insns += (C.NumRegs - PrevNumRegs);
1389 else
1390 C.Insns += (C.NumRegs - TTIRegNum);
1391 }
1392
1393 // If ICmpZero formula ends with not 0, it could not be replaced by
1394 // just add or sub. We'll need to compare final result of AddRec.
1395 // That means we'll need an additional instruction. But if the target can
1396 // macro-fuse a compare with a branch, don't count this extra instruction.
1397 // For -10 + {0, +, 1}:
1398 // i = i + 1;
1399 // cmp i, 10
1400 //
1401 // For {-10, +, 1}:
1402 // i = i + 1;
1403 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1404 !TTI->canMacroFuseCmp())
1405 C.Insns++;
1406 // Each new AddRec adds 1 instruction to calculation.
1407 C.Insns += (C.AddRecCost - PrevAddRecCost);
1408
1409 // BaseAdds adds instructions for unfolded registers.
1410 if (LU.Kind != LSRUse::ICmpZero)
1411 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1412 assert(isValid() && "invalid cost")((isValid() && "invalid cost") ? static_cast<void>
(0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1412, __PRETTY_FUNCTION__))
;
1413}
1414
1415/// Set this cost to a losing value.
1416void Cost::Lose() {
1417 C.Insns = std::numeric_limits<unsigned>::max();
1418 C.NumRegs = std::numeric_limits<unsigned>::max();
1419 C.AddRecCost = std::numeric_limits<unsigned>::max();
1420 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1421 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1422 C.ImmCost = std::numeric_limits<unsigned>::max();
1423 C.SetupCost = std::numeric_limits<unsigned>::max();
1424 C.ScaleCost = std::numeric_limits<unsigned>::max();
1425}
1426
1427/// Choose the lower cost.
1428bool Cost::isLess(Cost &Other) {
1429 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1430 C.Insns != Other.C.Insns)
1431 return C.Insns < Other.C.Insns;
1432 return TTI->isLSRCostLess(C, Other.C);
1433}
1434
1435#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1436void Cost::print(raw_ostream &OS) const {
1437 if (InsnsCost)
1438 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1439 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1440 if (C.AddRecCost != 0)
1441 OS << ", with addrec cost " << C.AddRecCost;
1442 if (C.NumIVMuls != 0)
1443 OS << ", plus " << C.NumIVMuls << " IV mul"
1444 << (C.NumIVMuls == 1 ? "" : "s");
1445 if (C.NumBaseAdds != 0)
1446 OS << ", plus " << C.NumBaseAdds << " base add"
1447 << (C.NumBaseAdds == 1 ? "" : "s");
1448 if (C.ScaleCost != 0)
1449 OS << ", plus " << C.ScaleCost << " scale cost";
1450 if (C.ImmCost != 0)
1451 OS << ", plus " << C.ImmCost << " imm cost";
1452 if (C.SetupCost != 0)
1453 OS << ", plus " << C.SetupCost << " setup cost";
1454}
1455
1456LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1457 print(errs()); errs() << '\n';
1458}
1459#endif
1460
1461/// Test whether this fixup always uses its value outside of the given loop.
1462bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1463 // PHI nodes use their value in their incoming blocks.
1464 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1465 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1466 if (PN->getIncomingValue(i) == OperandValToReplace &&
1467 L->contains(PN->getIncomingBlock(i)))
1468 return false;
1469 return true;
1470 }
1471
1472 return !L->contains(UserInst);
1473}
1474
1475#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1476void LSRFixup::print(raw_ostream &OS) const {
1477 OS << "UserInst=";
1478 // Store is common and interesting enough to be worth special-casing.
1479 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1480 OS << "store ";
1481 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1482 } else if (UserInst->getType()->isVoidTy())
1483 OS << UserInst->getOpcodeName();
1484 else
1485 UserInst->printAsOperand(OS, /*PrintType=*/false);
1486
1487 OS << ", OperandValToReplace=";
1488 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1489
1490 for (const Loop *PIL : PostIncLoops) {
1491 OS << ", PostIncLoop=";
1492 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1493 }
1494
1495 if (Offset != 0)
1496 OS << ", Offset=" << Offset;
1497}
1498
1499LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1500 print(errs()); errs() << '\n';
1501}
1502#endif
1503
1504/// Test whether this use as a formula which has the same registers as the given
1505/// formula.
1506bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1507 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1508 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1509 // Unstable sort by host order ok, because this is only used for uniquifying.
1510 llvm::sort(Key);
1511 return Uniquifier.count(Key);
1512}
1513
1514/// The function returns a probability of selecting formula without Reg.
1515float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1516 unsigned FNum = 0;
1517 for (const Formula &F : Formulae)
1518 if (F.referencesReg(Reg))
1519 FNum++;
1520 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1521}
1522
1523/// If the given formula has not yet been inserted, add it to the list, and
1524/// return true. Return false otherwise. The formula must be in canonical form.
1525bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1526 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1526, __PRETTY_FUNCTION__))
;
1527
1528 if (!Formulae.empty() && RigidFormula)
1529 return false;
1530
1531 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1532 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1533 // Unstable sort by host order ok, because this is only used for uniquifying.
1534 llvm::sort(Key);
1535
1536 if (!Uniquifier.insert(Key).second)
1537 return false;
1538
1539 // Using a register to hold the value of 0 is not profitable.
1540 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1541, __PRETTY_FUNCTION__))
1541 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1541, __PRETTY_FUNCTION__))
;
1542#ifndef NDEBUG
1543 for (const SCEV *BaseReg : F.BaseRegs)
1544 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1544, __PRETTY_FUNCTION__))
;
1545#endif
1546
1547 // Add the formula to the list.
1548 Formulae.push_back(F);
1549
1550 // Record registers now being used by this use.
1551 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1552 if (F.ScaledReg)
1553 Regs.insert(F.ScaledReg);
1554
1555 return true;
1556}
1557
1558/// Remove the given formula from this use's list.
1559void LSRUse::DeleteFormula(Formula &F) {
1560 if (&F != &Formulae.back())
1561 std::swap(F, Formulae.back());
1562 Formulae.pop_back();
1563}
1564
1565/// Recompute the Regs field, and update RegUses.
1566void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1567 // Now that we've filtered out some formulae, recompute the Regs set.
1568 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1569 Regs.clear();
1570 for (const Formula &F : Formulae) {
1571 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1572 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1573 }
1574
1575 // Update the RegTracker.
1576 for (const SCEV *S : OldRegs)
1577 if (!Regs.count(S))
1578 RegUses.dropRegister(S, LUIdx);
1579}
1580
1581#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1582void LSRUse::print(raw_ostream &OS) const {
1583 OS << "LSR Use: Kind=";
1584 switch (Kind) {
1585 case Basic: OS << "Basic"; break;
1586 case Special: OS << "Special"; break;
1587 case ICmpZero: OS << "ICmpZero"; break;
1588 case Address:
1589 OS << "Address of ";
1590 if (AccessTy.MemTy->isPointerTy())
1591 OS << "pointer"; // the full pointer type could be really verbose
1592 else {
1593 OS << *AccessTy.MemTy;
1594 }
1595
1596 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1597 }
1598
1599 OS << ", Offsets={";
1600 bool NeedComma = false;
1601 for (const LSRFixup &Fixup : Fixups) {
1602 if (NeedComma) OS << ',';
1603 OS << Fixup.Offset;
1604 NeedComma = true;
1605 }
1606 OS << '}';
1607
1608 if (AllFixupsOutsideLoop)
1609 OS << ", all-fixups-outside-loop";
1610
1611 if (WidestFixupType)
1612 OS << ", widest fixup type: " << *WidestFixupType;
1613}
1614
1615LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1616 print(errs()); errs() << '\n';
1617}
1618#endif
1619
1620static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1621 LSRUse::KindType Kind, MemAccessTy AccessTy,
1622 GlobalValue *BaseGV, int64_t BaseOffset,
1623 bool HasBaseReg, int64_t Scale,
1624 Instruction *Fixup/*= nullptr*/) {
1625 switch (Kind) {
1626 case LSRUse::Address:
1627 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1628 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1629
1630 case LSRUse::ICmpZero:
1631 // There's not even a target hook for querying whether it would be legal to
1632 // fold a GV into an ICmp.
1633 if (BaseGV)
1634 return false;
1635
1636 // ICmp only has two operands; don't allow more than two non-trivial parts.
1637 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1638 return false;
1639
1640 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1641 // putting the scaled register in the other operand of the icmp.
1642 if (Scale != 0 && Scale != -1)
1643 return false;
1644
1645 // If we have low-level target information, ask the target if it can fold an
1646 // integer immediate on an icmp.
1647 if (BaseOffset != 0) {
1648 // We have one of:
1649 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1650 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1651 // Offs is the ICmp immediate.
1652 if (Scale == 0)
1653 // The cast does the right thing with
1654 // std::numeric_limits<int64_t>::min().
1655 BaseOffset = -(uint64_t)BaseOffset;
1656 return TTI.isLegalICmpImmediate(BaseOffset);
1657 }
1658
1659 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1660 return true;
1661
1662 case LSRUse::Basic:
1663 // Only handle single-register values.
1664 return !BaseGV && Scale == 0 && BaseOffset == 0;
1665
1666 case LSRUse::Special:
1667 // Special case Basic to handle -1 scales.
1668 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1669 }
1670
1671 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1671)
;
1672}
1673
1674static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1675 int64_t MinOffset, int64_t MaxOffset,
1676 LSRUse::KindType Kind, MemAccessTy AccessTy,
1677 GlobalValue *BaseGV, int64_t BaseOffset,
1678 bool HasBaseReg, int64_t Scale) {
1679 // Check for overflow.
1680 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1681 (MinOffset > 0))
1682 return false;
1683 MinOffset = (uint64_t)BaseOffset + MinOffset;
1684 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1685 (MaxOffset > 0))
1686 return false;
1687 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1688
1689 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1690 HasBaseReg, Scale) &&
1691 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1692 HasBaseReg, Scale);
1693}
1694
1695static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1696 int64_t MinOffset, int64_t MaxOffset,
1697 LSRUse::KindType Kind, MemAccessTy AccessTy,
1698 const Formula &F, const Loop &L) {
1699 // For the purpose of isAMCompletelyFolded either having a canonical formula
1700 // or a scale not equal to zero is correct.
1701 // Problems may arise from non canonical formulae having a scale == 0.
1702 // Strictly speaking it would best to just rely on canonical formulae.
1703 // However, when we generate the scaled formulae, we first check that the
1704 // scaling factor is profitable before computing the actual ScaledReg for
1705 // compile time sake.
1706 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1706, __PRETTY_FUNCTION__))
;
1707 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1708 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1709}
1710
1711/// Test whether we know how to expand the current formula.
1712static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1713 int64_t MaxOffset, LSRUse::KindType Kind,
1714 MemAccessTy AccessTy, GlobalValue *BaseGV,
1715 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1716 // We know how to expand completely foldable formulae.
1717 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1718 BaseOffset, HasBaseReg, Scale) ||
1719 // Or formulae that use a base register produced by a sum of base
1720 // registers.
1721 (Scale == 1 &&
1722 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1723 BaseGV, BaseOffset, true, 0));
1724}
1725
1726static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1727 int64_t MaxOffset, LSRUse::KindType Kind,
1728 MemAccessTy AccessTy, const Formula &F) {
1729 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1730 F.BaseOffset, F.HasBaseReg, F.Scale);
1731}
1732
1733static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1734 const LSRUse &LU, const Formula &F) {
1735 // Target may want to look at the user instructions.
1736 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1737 for (const LSRFixup &Fixup : LU.Fixups)
1738 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1739 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1740 F.Scale, Fixup.UserInst))
1741 return false;
1742 return true;
1743 }
1744
1745 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1746 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1747 F.Scale);
1748}
1749
1750static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1751 const LSRUse &LU, const Formula &F,
1752 const Loop &L) {
1753 if (!F.Scale)
1754 return 0;
1755
1756 // If the use is not completely folded in that instruction, we will have to
1757 // pay an extra cost only for scale != 1.
1758 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1759 LU.AccessTy, F, L))
1760 return F.Scale != 1;
1761
1762 switch (LU.Kind) {
1763 case LSRUse::Address: {
1764 // Check the scaling factor cost with both the min and max offsets.
1765 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1766 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1767 F.Scale, LU.AccessTy.AddrSpace);
1768 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1769 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1770 F.Scale, LU.AccessTy.AddrSpace);
1771
1772 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1773, __PRETTY_FUNCTION__))
1773 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1773, __PRETTY_FUNCTION__))
;
1774 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1775 }
1776 case LSRUse::ICmpZero:
1777 case LSRUse::Basic:
1778 case LSRUse::Special:
1779 // The use is completely folded, i.e., everything is folded into the
1780 // instruction.
1781 return 0;
1782 }
1783
1784 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1784)
;
1785}
1786
1787static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1788 LSRUse::KindType Kind, MemAccessTy AccessTy,
1789 GlobalValue *BaseGV, int64_t BaseOffset,
1790 bool HasBaseReg) {
1791 // Fast-path: zero is always foldable.
1792 if (BaseOffset == 0 && !BaseGV) return true;
1793
1794 // Conservatively, create an address with an immediate and a
1795 // base and a scale.
1796 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1797
1798 // Canonicalize a scale of 1 to a base register if the formula doesn't
1799 // already have a base register.
1800 if (!HasBaseReg && Scale == 1) {
1801 Scale = 0;
1802 HasBaseReg = true;
1803 }
1804
1805 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1806 HasBaseReg, Scale);
1807}
1808
1809static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1810 ScalarEvolution &SE, int64_t MinOffset,
1811 int64_t MaxOffset, LSRUse::KindType Kind,
1812 MemAccessTy AccessTy, const SCEV *S,
1813 bool HasBaseReg) {
1814 // Fast-path: zero is always foldable.
1815 if (S->isZero()) return true;
1816
1817 // Conservatively, create an address with an immediate and a
1818 // base and a scale.
1819 int64_t BaseOffset = ExtractImmediate(S, SE);
1820 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1821
1822 // If there's anything else involved, it's not foldable.
1823 if (!S->isZero()) return false;
1824
1825 // Fast-path: zero is always foldable.
1826 if (BaseOffset == 0 && !BaseGV) return true;
1827
1828 // Conservatively, create an address with an immediate and a
1829 // base and a scale.
1830 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1831
1832 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1833 BaseOffset, HasBaseReg, Scale);
1834}
1835
1836namespace {
1837
1838/// An individual increment in a Chain of IV increments. Relate an IV user to
1839/// an expression that computes the IV it uses from the IV used by the previous
1840/// link in the Chain.
1841///
1842/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1843/// original IVOperand. The head of the chain's IVOperand is only valid during
1844/// chain collection, before LSR replaces IV users. During chain generation,
1845/// IncExpr can be used to find the new IVOperand that computes the same
1846/// expression.
1847struct IVInc {
1848 Instruction *UserInst;
1849 Value* IVOperand;
1850 const SCEV *IncExpr;
1851
1852 IVInc(Instruction *U, Value *O, const SCEV *E)
1853 : UserInst(U), IVOperand(O), IncExpr(E) {}
1854};
1855
1856// The list of IV increments in program order. We typically add the head of a
1857// chain without finding subsequent links.
1858struct IVChain {
1859 SmallVector<IVInc, 1> Incs;
1860 const SCEV *ExprBase = nullptr;
1861
1862 IVChain() = default;
1863 IVChain(const IVInc &Head, const SCEV *Base)
1864 : Incs(1, Head), ExprBase(Base) {}
1865
1866 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1867
1868 // Return the first increment in the chain.
1869 const_iterator begin() const {
1870 assert(!Incs.empty())((!Incs.empty()) ? static_cast<void> (0) : __assert_fail
("!Incs.empty()", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1870, __PRETTY_FUNCTION__))
;
1871 return std::next(Incs.begin());
1872 }
1873 const_iterator end() const {
1874 return Incs.end();
1875 }
1876
1877 // Returns true if this chain contains any increments.
1878 bool hasIncs() const { return Incs.size() >= 2; }
1879
1880 // Add an IVInc to the end of this chain.
1881 void add(const IVInc &X) { Incs.push_back(X); }
1882
1883 // Returns the last UserInst in the chain.
1884 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1885
1886 // Returns true if IncExpr can be profitably added to this chain.
1887 bool isProfitableIncrement(const SCEV *OperExpr,
1888 const SCEV *IncExpr,
1889 ScalarEvolution&);
1890};
1891
1892/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1893/// between FarUsers that definitely cross IV increments and NearUsers that may
1894/// be used between IV increments.
1895struct ChainUsers {
1896 SmallPtrSet<Instruction*, 4> FarUsers;
1897 SmallPtrSet<Instruction*, 4> NearUsers;
1898};
1899
1900/// This class holds state for the main loop strength reduction logic.
1901class LSRInstance {
1902 IVUsers &IU;
1903 ScalarEvolution &SE;
1904 DominatorTree &DT;
1905 LoopInfo &LI;
1906 AssumptionCache &AC;
1907 TargetLibraryInfo &TLI;
1908 const TargetTransformInfo &TTI;
1909 Loop *const L;
1910 MemorySSAUpdater *MSSAU;
1911 bool FavorBackedgeIndex = false;
1912 bool Changed = false;
1913
1914 /// This is the insert position that the current loop's induction variable
1915 /// increment should be placed. In simple loops, this is the latch block's
1916 /// terminator. But in more complicated cases, this is a position which will
1917 /// dominate all the in-loop post-increment users.
1918 Instruction *IVIncInsertPos = nullptr;
1919
1920 /// Interesting factors between use strides.
1921 ///
1922 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1923 /// default, a SmallDenseSet, because we need to use the full range of
1924 /// int64_ts, and there's currently no good way of doing that with
1925 /// SmallDenseSet.
1926 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1927
1928 /// Interesting use types, to facilitate truncation reuse.
1929 SmallSetVector<Type *, 4> Types;
1930
1931 /// The list of interesting uses.
1932 mutable SmallVector<LSRUse, 16> Uses;
1933
1934 /// Track which uses use which register candidates.
1935 RegUseTracker RegUses;
1936
1937 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1938 // have more than a few IV increment chains in a loop. Missing a Chain falls
1939 // back to normal LSR behavior for those uses.
1940 static const unsigned MaxChains = 8;
1941
1942 /// IV users can form a chain of IV increments.
1943 SmallVector<IVChain, MaxChains> IVChainVec;
1944
1945 /// IV users that belong to profitable IVChains.
1946 SmallPtrSet<Use*, MaxChains> IVIncSet;
1947
1948 void OptimizeShadowIV();
1949 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1950 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1951 void OptimizeLoopTermCond();
1952
1953 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1954 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1955 void FinalizeChain(IVChain &Chain);
1956 void CollectChains();
1957 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1958 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1959
1960 void CollectInterestingTypesAndFactors();
1961 void CollectFixupsAndInitialFormulae();
1962
1963 // Support for sharing of LSRUses between LSRFixups.
1964 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
1965 UseMapTy UseMap;
1966
1967 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1968 LSRUse::KindType Kind, MemAccessTy AccessTy);
1969
1970 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1971 MemAccessTy AccessTy);
1972
1973 void DeleteUse(LSRUse &LU, size_t LUIdx);
1974
1975 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1976
1977 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1978 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1979 void CountRegisters(const Formula &F, size_t LUIdx);
1980 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1981
1982 void CollectLoopInvariantFixupsAndFormulae();
1983
1984 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1985 unsigned Depth = 0);
1986
1987 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1988 const Formula &Base, unsigned Depth,
1989 size_t Idx, bool IsScaledReg = false);
1990 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1991 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1992 const Formula &Base, size_t Idx,
1993 bool IsScaledReg = false);
1994 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1995 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1996 const Formula &Base,
1997 const SmallVectorImpl<int64_t> &Worklist,
1998 size_t Idx, bool IsScaledReg = false);
1999 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2000 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2001 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2002 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2003 void GenerateCrossUseConstantOffsets();
2004 void GenerateAllReuseFormulae();
2005
2006 void FilterOutUndesirableDedicatedRegisters();
2007
2008 size_t EstimateSearchSpaceComplexity() const;
2009 void NarrowSearchSpaceByDetectingSupersets();
2010 void NarrowSearchSpaceByCollapsingUnrolledCode();
2011 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2012 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2013 void NarrowSearchSpaceByFilterPostInc();
2014 void NarrowSearchSpaceByDeletingCostlyFormulas();
2015 void NarrowSearchSpaceByPickingWinnerRegs();
2016 void NarrowSearchSpaceUsingHeuristics();
2017
2018 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2019 Cost &SolutionCost,
2020 SmallVectorImpl<const Formula *> &Workspace,
2021 const Cost &CurCost,
2022 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2023 DenseSet<const SCEV *> &VisitedRegs) const;
2024 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2025
2026 BasicBlock::iterator
2027 HoistInsertPosition(BasicBlock::iterator IP,
2028 const SmallVectorImpl<Instruction *> &Inputs) const;
2029 BasicBlock::iterator
2030 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2031 const LSRFixup &LF,
2032 const LSRUse &LU,
2033 SCEVExpander &Rewriter) const;
2034
2035 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2036 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2037 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2038 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2039 const Formula &F, SCEVExpander &Rewriter,
2040 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2041 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2042 SCEVExpander &Rewriter,
2043 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2044 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2045
2046public:
2047 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2048 LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2049 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);
2050
2051 bool getChanged() const { return Changed; }
2052
2053 void print_factors_and_types(raw_ostream &OS) const;
2054 void print_fixups(raw_ostream &OS) const;
2055 void print_uses(raw_ostream &OS) const;
2056 void print(raw_ostream &OS) const;
2057 void dump() const;
2058};
2059
2060} // end anonymous namespace
2061
2062/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2063/// the cast operation.
2064void LSRInstance::OptimizeShadowIV() {
2065 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2066 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2067 return;
2068
2069 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2070 UI != E; /* empty */) {
2071 IVUsers::const_iterator CandidateUI = UI;
2072 ++UI;
2073 Instruction *ShadowUse = CandidateUI->getUser();
2074 Type *DestTy = nullptr;
2075 bool IsSigned = false;
2076
2077 /* If shadow use is a int->float cast then insert a second IV
2078 to eliminate this cast.
2079
2080 for (unsigned i = 0; i < n; ++i)
2081 foo((double)i);
2082
2083 is transformed into
2084
2085 double d = 0.0;
2086 for (unsigned i = 0; i < n; ++i, ++d)
2087 foo(d);
2088 */
2089 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2090 IsSigned = false;
2091 DestTy = UCast->getDestTy();
2092 }
2093 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2094 IsSigned = true;
2095 DestTy = SCast->getDestTy();
2096 }
2097 if (!DestTy) continue;
2098
2099 // If target does not support DestTy natively then do not apply
2100 // this transformation.
2101 if (!TTI.isTypeLegal(DestTy)) continue;
2102
2103 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2104 if (!PH) continue;
2105 if (PH->getNumIncomingValues() != 2) continue;
2106
2107 // If the calculation in integers overflows, the result in FP type will
2108 // differ. So we only can do this transformation if we are guaranteed to not
2109 // deal with overflowing values
2110 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2111 if (!AR) continue;
2112 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2113 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2114
2115 Type *SrcTy = PH->getType();
2116 int Mantissa = DestTy->getFPMantissaWidth();
2117 if (Mantissa == -1) continue;
2118 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2119 continue;
2120
2121 unsigned Entry, Latch;
2122 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2123 Entry = 0;
2124 Latch = 1;
2125 } else {
2126 Entry = 1;
2127 Latch = 0;
2128 }
2129
2130 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2131 if (!Init) continue;
2132 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2133 (double)Init->getSExtValue() :
2134 (double)Init->getZExtValue());
2135
2136 BinaryOperator *Incr =
2137 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2138 if (!Incr) continue;
2139 if (Incr->getOpcode() != Instruction::Add
2140 && Incr->getOpcode() != Instruction::Sub)
2141 continue;
2142
2143 /* Initialize new IV, double d = 0.0 in above example. */
2144 ConstantInt *C = nullptr;
2145 if (Incr->getOperand(0) == PH)
2146 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2147 else if (Incr->getOperand(1) == PH)
2148 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2149 else
2150 continue;
2151
2152 if (!C) continue;
2153
2154 // Ignore negative constants, as the code below doesn't handle them
2155 // correctly. TODO: Remove this restriction.
2156 if (!C->getValue().isStrictlyPositive()) continue;
2157
2158 /* Add new PHINode. */
2159 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2160
2161 /* create new increment. '++d' in above example. */
2162 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2163 BinaryOperator *NewIncr =
2164 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2165 Instruction::FAdd : Instruction::FSub,
2166 NewPH, CFP, "IV.S.next.", Incr);
2167
2168 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2169 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2170
2171 /* Remove cast operation */
2172 ShadowUse->replaceAllUsesWith(NewPH);
2173 ShadowUse->eraseFromParent();
2174 Changed = true;
2175 break;
2176 }
2177}
2178
2179/// If Cond has an operand that is an expression of an IV, set the IV user and
2180/// stride information and return true, otherwise return false.
2181bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2182 for (IVStrideUse &U : IU)
2183 if (U.getUser() == Cond) {
2184 // NOTE: we could handle setcc instructions with multiple uses here, but
2185 // InstCombine does it as well for simple uses, it's not clear that it
2186 // occurs enough in real life to handle.
2187 CondUse = &U;
2188 return true;
2189 }
2190 return false;
2191}
2192
2193/// Rewrite the loop's terminating condition if it uses a max computation.
2194///
2195/// This is a narrow solution to a specific, but acute, problem. For loops
2196/// like this:
2197///
2198/// i = 0;
2199/// do {
2200/// p[i] = 0.0;
2201/// } while (++i < n);
2202///
2203/// the trip count isn't just 'n', because 'n' might not be positive. And
2204/// unfortunately this can come up even for loops where the user didn't use
2205/// a C do-while loop. For example, seemingly well-behaved top-test loops
2206/// will commonly be lowered like this:
2207///
2208/// if (n > 0) {
2209/// i = 0;
2210/// do {
2211/// p[i] = 0.0;
2212/// } while (++i < n);
2213/// }
2214///
2215/// and then it's possible for subsequent optimization to obscure the if
2216/// test in such a way that indvars can't find it.
2217///
2218/// When indvars can't find the if test in loops like this, it creates a
2219/// max expression, which allows it to give the loop a canonical
2220/// induction variable:
2221///
2222/// i = 0;
2223/// max = n < 1 ? 1 : n;
2224/// do {
2225/// p[i] = 0.0;
2226/// } while (++i != max);
2227///
2228/// Canonical induction variables are necessary because the loop passes
2229/// are designed around them. The most obvious example of this is the
2230/// LoopInfo analysis, which doesn't remember trip count values. It
2231/// expects to be able to rediscover the trip count each time it is
2232/// needed, and it does this using a simple analysis that only succeeds if
2233/// the loop has a canonical induction variable.
2234///
2235/// However, when it comes time to generate code, the maximum operation
2236/// can be quite costly, especially if it's inside of an outer loop.
2237///
2238/// This function solves this problem by detecting this type of loop and
2239/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2240/// the instructions for the maximum computation.
2241ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2242 // Check that the loop matches the pattern we're looking for.
2243 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2244 Cond->getPredicate() != CmpInst::ICMP_NE)
2245 return Cond;
2246
2247 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2248 if (!Sel || !Sel->hasOneUse()) return Cond;
2249
2250 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2251 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2252 return Cond;
2253 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2254
2255 // Add one to the backedge-taken count to get the trip count.
2256 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2257 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2258
2259 // Check for a max calculation that matches the pattern. There's no check
2260 // for ICMP_ULE here because the comparison would be with zero, which
2261 // isn't interesting.
2262 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2263 const SCEVNAryExpr *Max = nullptr;
2264 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2265 Pred = ICmpInst::ICMP_SLE;
2266 Max = S;
2267 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2268 Pred = ICmpInst::ICMP_SLT;
2269 Max = S;
2270 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2271 Pred = ICmpInst::ICMP_ULT;
2272 Max = U;
2273 } else {
2274 // No match; bail.
2275 return Cond;
2276 }
2277
2278 // To handle a max with more than two operands, this optimization would
2279 // require additional checking and setup.
2280 if (Max->getNumOperands() != 2)
2281 return Cond;
2282
2283 const SCEV *MaxLHS = Max->getOperand(0);
2284 const SCEV *MaxRHS = Max->getOperand(1);
2285
2286 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2287 // for a comparison with 1. For <= and >=, a comparison with zero.
2288 if (!MaxLHS ||
2289 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2290 return Cond;
2291
2292 // Check the relevant induction variable for conformance to
2293 // the pattern.
2294 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2295 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2296 if (!AR || !AR->isAffine() ||
2297 AR->getStart() != One ||
2298 AR->getStepRecurrence(SE) != One)
2299 return Cond;
2300
2301 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2302, __PRETTY_FUNCTION__))
2302 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2302, __PRETTY_FUNCTION__))
;
2303
2304 // Check the right operand of the select, and remember it, as it will
2305 // be used in the new comparison instruction.
2306 Value *NewRHS = nullptr;
2307 if (ICmpInst::isTrueWhenEqual(Pred)) {
2308 // Look for n+1, and grab n.
2309 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2310 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2311 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2312 NewRHS = BO->getOperand(0);
2313 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2314 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2315 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2316 NewRHS = BO->getOperand(0);
2317 if (!NewRHS)
2318 return Cond;
2319 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2320 NewRHS = Sel->getOperand(1);
2321 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2322 NewRHS = Sel->getOperand(2);
2323 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2324 NewRHS = SU->getValue();
2325 else
2326 // Max doesn't match expected pattern.
2327 return Cond;
2328
2329 // Determine the new comparison opcode. It may be signed or unsigned,
2330 // and the original comparison may be either equality or inequality.
2331 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2332 Pred = CmpInst::getInversePredicate(Pred);
2333
2334 // Ok, everything looks ok to change the condition into an SLT or SGE and
2335 // delete the max calculation.
2336 ICmpInst *NewCond =
2337 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2338
2339 // Delete the max calculation instructions.
2340 Cond->replaceAllUsesWith(NewCond);
2341 CondUse->setUser(NewCond);
2342 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2343 Cond->eraseFromParent();
2344 Sel->eraseFromParent();
2345 if (Cmp->use_empty())
2346 Cmp->eraseFromParent();
2347 return NewCond;
2348}
2349
2350/// Change loop terminating condition to use the postinc iv when possible.
2351void
2352LSRInstance::OptimizeLoopTermCond() {
2353 SmallPtrSet<Instruction *, 4> PostIncs;
2354
2355 // We need a different set of heuristics for rotated and non-rotated loops.
2356 // If a loop is rotated then the latch is also the backedge, so inserting
2357 // post-inc expressions just before the latch is ideal. To reduce live ranges
2358 // it also makes sense to rewrite terminating conditions to use post-inc
2359 // expressions.
2360 //
2361 // If the loop is not rotated then the latch is not a backedge; the latch
2362 // check is done in the loop head. Adding post-inc expressions before the
2363 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2364 // in the loop body. In this case we do *not* want to use post-inc expressions
2365 // in the latch check, and we want to insert post-inc expressions before
2366 // the backedge.
2367 BasicBlock *LatchBlock = L->getLoopLatch();
2368 SmallVector<BasicBlock*, 8> ExitingBlocks;
2369 L->getExitingBlocks(ExitingBlocks);
2370 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2371 return LatchBlock != BB;
2372 })) {
2373 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2374 IVIncInsertPos = LatchBlock->getTerminator();
2375 return;
2376 }
2377
2378 // Otherwise treat this as a rotated loop.
2379 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2380 // Get the terminating condition for the loop if possible. If we
2381 // can, we want to change it to use a post-incremented version of its
2382 // induction variable, to allow coalescing the live ranges for the IV into
2383 // one register value.
2384
2385 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2386 if (!TermBr)
2387 continue;
2388 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2389 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2390 continue;
2391
2392 // Search IVUsesByStride to find Cond's IVUse if there is one.
2393 IVStrideUse *CondUse = nullptr;
2394 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2395 if (!FindIVUserForCond(Cond, CondUse))
2396 continue;
2397
2398 // If the trip count is computed in terms of a max (due to ScalarEvolution
2399 // being unable to find a sufficient guard, for example), change the loop
2400 // comparison to use SLT or ULT instead of NE.
2401 // One consequence of doing this now is that it disrupts the count-down
2402 // optimization. That's not always a bad thing though, because in such
2403 // cases it may still be worthwhile to avoid a max.
2404 Cond = OptimizeMax(Cond, CondUse);
2405
2406 // If this exiting block dominates the latch block, it may also use
2407 // the post-inc value if it won't be shared with other uses.
2408 // Check for dominance.
2409 if (!DT.dominates(ExitingBlock, LatchBlock))
2410 continue;
2411
2412 // Conservatively avoid trying to use the post-inc value in non-latch
2413 // exits if there may be pre-inc users in intervening blocks.
2414 if (LatchBlock != ExitingBlock)
2415 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2416 // Test if the use is reachable from the exiting block. This dominator
2417 // query is a conservative approximation of reachability.
2418 if (&*UI != CondUse &&
2419 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2420 // Conservatively assume there may be reuse if the quotient of their
2421 // strides could be a legal scale.
2422 const SCEV *A = IU.getStride(*CondUse, L);
2423 const SCEV *B = IU.getStride(*UI, L);
2424 if (!A || !B) continue;
2425 if (SE.getTypeSizeInBits(A->getType()) !=
2426 SE.getTypeSizeInBits(B->getType())) {
2427 if (SE.getTypeSizeInBits(A->getType()) >
2428 SE.getTypeSizeInBits(B->getType()))
2429 B = SE.getSignExtendExpr(B, A->getType());
2430 else
2431 A = SE.getSignExtendExpr(A, B->getType());
2432 }
2433 if (const SCEVConstant *D =
2434 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2435 const ConstantInt *C = D->getValue();
2436 // Stride of one or negative one can have reuse with non-addresses.
2437 if (C->isOne() || C->isMinusOne())
2438 goto decline_post_inc;
2439 // Avoid weird situations.
2440 if (C->getValue().getMinSignedBits() >= 64 ||
2441 C->getValue().isMinSignedValue())
2442 goto decline_post_inc;
2443 // Check for possible scaled-address reuse.
2444 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2445 MemAccessTy AccessTy = getAccessType(
2446 TTI, UI->getUser(), UI->getOperandValToReplace());
2447 int64_t Scale = C->getSExtValue();
2448 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2449 /*BaseOffset=*/0,
2450 /*HasBaseReg=*/false, Scale,
2451 AccessTy.AddrSpace))
2452 goto decline_post_inc;
2453 Scale = -Scale;
2454 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2455 /*BaseOffset=*/0,
2456 /*HasBaseReg=*/false, Scale,
2457 AccessTy.AddrSpace))
2458 goto decline_post_inc;
2459 }
2460 }
2461 }
2462
2463 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)
2464 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2465
2466 // It's possible for the setcc instruction to be anywhere in the loop, and
2467 // possible for it to have multiple users. If it is not immediately before
2468 // the exiting block branch, move it.
2469 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2470 if (Cond->hasOneUse()) {
2471 Cond->moveBefore(TermBr);
2472 } else {
2473 // Clone the terminating condition and insert into the loopend.
2474 ICmpInst *OldCond = Cond;
2475 Cond = cast<ICmpInst>(Cond->clone());
2476 Cond->setName(L->getHeader()->getName() + ".termcond");
2477 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2478
2479 // Clone the IVUse, as the old use still exists!
2480 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2481 TermBr->replaceUsesOfWith(OldCond, Cond);
2482 }
2483 }
2484
2485 // If we get to here, we know that we can transform the setcc instruction to
2486 // use the post-incremented version of the IV, allowing us to coalesce the
2487 // live ranges for the IV correctly.
2488 CondUse->transformToPostInc(L);
2489 Changed = true;
2490
2491 PostIncs.insert(Cond);
2492 decline_post_inc:;
2493 }
2494
2495 // Determine an insertion point for the loop induction variable increment. It
2496 // must dominate all the post-inc comparisons we just set up, and it must
2497 // dominate the loop latch edge.
2498 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2499 for (Instruction *Inst : PostIncs) {
2500 BasicBlock *BB =
2501 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2502 Inst->getParent());
2503 if (BB == Inst->getParent())
2504 IVIncInsertPos = Inst;
2505 else if (BB != IVIncInsertPos->getParent())
2506 IVIncInsertPos = BB->getTerminator();
2507 }
2508}
2509
2510/// Determine if the given use can accommodate a fixup at the given offset and
2511/// other details. If so, update the use and return true.
2512bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2513 bool HasBaseReg, LSRUse::KindType Kind,
2514 MemAccessTy AccessTy) {
2515 int64_t NewMinOffset = LU.MinOffset;
2516 int64_t NewMaxOffset = LU.MaxOffset;
2517 MemAccessTy NewAccessTy = AccessTy;
2518
2519 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2520 // something conservative, however this can pessimize in the case that one of
2521 // the uses will have all its uses outside the loop, for example.
2522 if (LU.Kind != Kind)
2523 return false;
2524
2525 // Check for a mismatched access type, and fall back conservatively as needed.
2526 // TODO: Be less conservative when the type is similar and can use the same
2527 // addressing modes.
2528 if (Kind == LSRUse::Address) {
2529 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2530 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2531 AccessTy.AddrSpace);
2532 }
2533 }
2534
2535 // Conservatively assume HasBaseReg is true for now.
2536 if (NewOffset < LU.MinOffset) {
2537 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2538 LU.MaxOffset - NewOffset, HasBaseReg))
2539 return false;
2540 NewMinOffset = NewOffset;
2541 } else if (NewOffset > LU.MaxOffset) {
2542 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2543 NewOffset - LU.MinOffset, HasBaseReg))
2544 return false;
2545 NewMaxOffset = NewOffset;
2546 }
2547
2548 // Update the use.
2549 LU.MinOffset = NewMinOffset;
2550 LU.MaxOffset = NewMaxOffset;
2551 LU.AccessTy = NewAccessTy;
2552 return true;
2553}
2554
2555/// Return an LSRUse index and an offset value for a fixup which needs the given
2556/// expression, with the given kind and optional access type. Either reuse an
2557/// existing use or create a new one, as needed.
2558std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2559 LSRUse::KindType Kind,
2560 MemAccessTy AccessTy) {
2561 const SCEV *Copy = Expr;
2562 int64_t Offset = ExtractImmediate(Expr, SE);
2563
2564 // Basic uses can't accept any offset, for example.
2565 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2566 Offset, /*HasBaseReg=*/ true)) {
2567 Expr = Copy;
2568 Offset = 0;
2569 }
2570
2571 std::pair<UseMapTy::iterator, bool> P =
2572 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2573 if (!P.second) {
2574 // A use already existed with this base.
2575 size_t LUIdx = P.first->second;
2576 LSRUse &LU = Uses[LUIdx];
2577 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2578 // Reuse this use.
2579 return std::make_pair(LUIdx, Offset);
2580 }
2581
2582 // Create a new use.
2583 size_t LUIdx = Uses.size();
2584 P.first->second = LUIdx;
2585 Uses.push_back(LSRUse(Kind, AccessTy));
2586 LSRUse &LU = Uses[LUIdx];
2587
2588 LU.MinOffset = Offset;
2589 LU.MaxOffset = Offset;
2590 return std::make_pair(LUIdx, Offset);
2591}
2592
2593/// Delete the given use from the Uses list.
2594void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2595 if (&LU != &Uses.back())
2596 std::swap(LU, Uses.back());
2597 Uses.pop_back();
2598
2599 // Update RegUses.
2600 RegUses.swapAndDropUse(LUIdx, Uses.size());
2601}
2602
2603/// Look for a use distinct from OrigLU which is has a formula that has the same
2604/// registers as the given formula.
2605LSRUse *
2606LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2607 const LSRUse &OrigLU) {
2608 // Search all uses for the formula. This could be more clever.
2609 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2610 LSRUse &LU = Uses[LUIdx];
2611 // Check whether this use is close enough to OrigLU, to see whether it's
2612 // worthwhile looking through its formulae.
2613 // Ignore ICmpZero uses because they may contain formulae generated by
2614 // GenerateICmpZeroScales, in which case adding fixup offsets may
2615 // be invalid.
2616 if (&LU != &OrigLU &&
2617 LU.Kind != LSRUse::ICmpZero &&
2618 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2619 LU.WidestFixupType == OrigLU.WidestFixupType &&
2620 LU.HasFormulaWithSameRegs(OrigF)) {
2621 // Scan through this use's formulae.
2622 for (const Formula &F : LU.Formulae) {
2623 // Check to see if this formula has the same registers and symbols
2624 // as OrigF.
2625 if (F.BaseRegs == OrigF.BaseRegs &&
2626 F.ScaledReg == OrigF.ScaledReg &&
2627 F.BaseGV == OrigF.BaseGV &&
2628 F.Scale == OrigF.Scale &&
2629 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2630 if (F.BaseOffset == 0)
2631 return &LU;
2632 // This is the formula where all the registers and symbols matched;
2633 // there aren't going to be any others. Since we declined it, we
2634 // can skip the rest of the formulae and proceed to the next LSRUse.
2635 break;
2636 }
2637 }
2638 }
2639 }
2640
2641 // Nothing looked good.
2642 return nullptr;
2643}
2644
2645void LSRInstance::CollectInterestingTypesAndFactors() {
2646 SmallSetVector<const SCEV *, 4> Strides;
2647
2648 // Collect interesting types and strides.
2649 SmallVector<const SCEV *, 4> Worklist;
2650 for (const IVStrideUse &U : IU) {
2651 const SCEV *Expr = IU.getExpr(U);
2652
2653 // Collect interesting types.
2654 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2655
2656 // Add strides for mentioned loops.
2657 Worklist.push_back(Expr);
2658 do {
2659 const SCEV *S = Worklist.pop_back_val();
2660 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2661 if (AR->getLoop() == L)
2662 Strides.insert(AR->getStepRecurrence(SE));
2663 Worklist.push_back(AR->getStart());
2664 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2665 Worklist.append(Add->op_begin(), Add->op_end());
2666 }
2667 } while (!Worklist.empty());
2668 }
2669
2670 // Compute interesting factors from the set of interesting strides.
2671 for (SmallSetVector<const SCEV *, 4>::const_iterator
2672 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2673 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2674 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2675 const SCEV *OldStride = *I;
2676 const SCEV *NewStride = *NewStrideIter;
2677
2678 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2679 SE.getTypeSizeInBits(NewStride->getType())) {
2680 if (SE.getTypeSizeInBits(OldStride->getType()) >
2681 SE.getTypeSizeInBits(NewStride->getType()))
2682 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2683 else
2684 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2685 }
2686 if (const SCEVConstant *Factor =
2687 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2688 SE, true))) {
2689 if (Factor->getAPInt().getMinSignedBits() <= 64)
2690 Factors.insert(Factor->getAPInt().getSExtValue());
2691 } else if (const SCEVConstant *Factor =
2692 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2693 NewStride,
2694 SE, true))) {
2695 if (Factor->getAPInt().getMinSignedBits() <= 64)
2696 Factors.insert(Factor->getAPInt().getSExtValue());
2697 }
2698 }
2699
2700 // If all uses use the same type, don't bother looking for truncation-based
2701 // reuse.
2702 if (Types.size() == 1)
2703 Types.clear();
2704
2705 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2706}
2707
2708/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2709/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2710/// IVStrideUses, we could partially skip this.
2711static User::op_iterator
2712findIVOperand(User::op_iterator OI, User::op_iterator OE,
2713 Loop *L, ScalarEvolution &SE) {
2714 for(; OI != OE; ++OI) {
2715 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2716 if (!SE.isSCEVable(Oper->getType()))
2717 continue;
2718
2719 if (const SCEVAddRecExpr *AR =
2720 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2721 if (AR->getLoop() == L)
2722 break;
2723 }
2724 }
2725 }
2726 return OI;
2727}
2728
2729/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2730/// a convenient helper.
2731static Value *getWideOperand(Value *Oper) {
2732 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2733 return Trunc->getOperand(0);
2734 return Oper;
2735}
2736
2737/// Return true if we allow an IV chain to include both types.
2738static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2739 Type *LType = LVal->getType();
2740 Type *RType = RVal->getType();
2741 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2742 // Different address spaces means (possibly)
2743 // different types of the pointer implementation,
2744 // e.g. i16 vs i32 so disallow that.
2745 (LType->getPointerAddressSpace() ==
2746 RType->getPointerAddressSpace()));
2747}
2748
2749/// Return an approximation of this SCEV expression's "base", or NULL for any
2750/// constant. Returning the expression itself is conservative. Returning a
2751/// deeper subexpression is more precise and valid as long as it isn't less
2752/// complex than another subexpression. For expressions involving multiple
2753/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2754/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2755/// IVInc==b-a.
2756///
2757/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2758/// SCEVUnknown, we simply return the rightmost SCEV operand.
2759static const SCEV *getExprBase(const SCEV *S) {
2760 switch (S->getSCEVType()) {
2761 default: // uncluding scUnknown.
2762 return S;
2763 case scConstant:
2764 return nullptr;
2765 case scTruncate:
2766 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2767 case scZeroExtend:
2768 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2769 case scSignExtend:
2770 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2771 case scAddExpr: {
2772 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2773 // there's nothing more complex.
2774 // FIXME: not sure if we want to recognize negation.
2775 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2776 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2777 E(Add->op_begin()); I != E; ++I) {
2778 const SCEV *SubExpr = *I;
2779 if (SubExpr->getSCEVType() == scAddExpr)
2780 return getExprBase(SubExpr);
2781
2782 if (SubExpr->getSCEVType() != scMulExpr)
2783 return SubExpr;
2784 }
2785 return S; // all operands are scaled, be conservative.
2786 }
2787 case scAddRecExpr:
2788 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2789 }
2790 llvm_unreachable("Unknown SCEV kind!")::llvm::llvm_unreachable_internal("Unknown SCEV kind!", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2790)
;
2791}
2792
2793/// Return true if the chain increment is profitable to expand into a loop
2794/// invariant value, which may require its own register. A profitable chain
2795/// increment will be an offset relative to the same base. We allow such offsets
2796/// to potentially be used as chain increment as long as it's not obviously
2797/// expensive to expand using real instructions.
2798bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2799 const SCEV *IncExpr,
2800 ScalarEvolution &SE) {
2801 // Aggressively form chains when -stress-ivchain.
2802 if (StressIVChain)
2803 return true;
2804
2805 // Do not replace a constant offset from IV head with a nonconstant IV
2806 // increment.
2807 if (!isa<SCEVConstant>(IncExpr)) {
2808 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2809 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2810 return false;
2811 }
2812
2813 SmallPtrSet<const SCEV*, 8> Processed;
2814 return !isHighCostExpansion(IncExpr, Processed, SE);
2815}
2816
2817/// Return true if the number of registers needed for the chain is estimated to
2818/// be less than the number required for the individual IV users. First prohibit
2819/// any IV users that keep the IV live across increments (the Users set should
2820/// be empty). Next count the number and type of increments in the chain.
2821///
2822/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2823/// effectively use postinc addressing modes. Only consider it profitable it the
2824/// increments can be computed in fewer registers when chained.
2825///
2826/// TODO: Consider IVInc free if it's already used in another chains.
2827static bool isProfitableChain(IVChain &Chain,
2828 SmallPtrSetImpl<Instruction *> &Users,
2829 ScalarEvolution &SE,
2830 const TargetTransformInfo &TTI) {
2831 if (StressIVChain)
2832 return true;
2833
2834 if (!Chain.hasIncs())
2835 return false;
2836
2837 if (!Users.empty()) {
2838 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)
2839 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)
2840 : 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)
;
2841 return false;
2842 }
2843 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2843, __PRETTY_FUNCTION__))
;
2844
2845 // The chain itself may require a register, so intialize cost to 1.
2846 int cost = 1;
2847
2848 // A complete chain likely eliminates the need for keeping the original IV in
2849 // a register. LSR does not currently know how to form a complete chain unless
2850 // the header phi already exists.
2851 if (isa<PHINode>(Chain.tailUserInst())
2852 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2853 --cost;
2854 }
2855 const SCEV *LastIncExpr = nullptr;
2856 unsigned NumConstIncrements = 0;
2857 unsigned NumVarIncrements = 0;
2858 unsigned NumReusedIncrements = 0;
2859
2860 if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
2861 return true;
2862
2863 for (const IVInc &Inc : Chain) {
2864 if (TTI.isProfitableLSRChainElement(Inc.UserInst))
2865 return true;
2866 if (Inc.IncExpr->isZero())
2867 continue;
2868
2869 // Incrementing by zero or some constant is neutral. We assume constants can
2870 // be folded into an addressing mode or an add's immediate operand.
2871 if (isa<SCEVConstant>(Inc.IncExpr)) {
2872 ++NumConstIncrements;
2873 continue;
2874 }
2875
2876 if (Inc.IncExpr == LastIncExpr)
2877 ++NumReusedIncrements;
2878 else
2879 ++NumVarIncrements;
2880
2881 LastIncExpr = Inc.IncExpr;
2882 }
2883 // An IV chain with a single increment is handled by LSR's postinc
2884 // uses. However, a chain with multiple increments requires keeping the IV's
2885 // value live longer than it needs to be if chained.
2886 if (NumConstIncrements > 1)
2887 --cost;
2888
2889 // Materializing increment expressions in the preheader that didn't exist in
2890 // the original code may cost a register. For example, sign-extended array
2891 // indices can produce ridiculous increments like this:
2892 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2893 cost += NumVarIncrements;
2894
2895 // Reusing variable increments likely saves a register to hold the multiple of
2896 // the stride.
2897 cost -= NumReusedIncrements;
2898
2899 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)
2900 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2901
2902 return cost < 0;
2903}
2904
2905/// Add this IV user to an existing chain or make it the head of a new chain.
2906void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2907 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2908 // When IVs are used as types of varying widths, they are generally converted
2909 // to a wider type with some uses remaining narrow under a (free) trunc.
2910 Value *const NextIV = getWideOperand(IVOper);
2911 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2912 const SCEV *const OperExprBase = getExprBase(OperExpr);
2913
2914 // Visit all existing chains. Check if its IVOper can be computed as a
2915 // profitable loop invariant increment from the last link in the Chain.
2916 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2917 const SCEV *LastIncExpr = nullptr;
2918 for (; ChainIdx < NChains; ++ChainIdx) {
2919 IVChain &Chain = IVChainVec[ChainIdx];
2920
2921 // Prune the solution space aggressively by checking that both IV operands
2922 // are expressions that operate on the same unscaled SCEVUnknown. This
2923 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2924 // first avoids creating extra SCEV expressions.
2925 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2926 continue;
2927
2928 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2929 if (!isCompatibleIVType(PrevIV, NextIV))
2930 continue;
2931
2932 // A phi node terminates a chain.
2933 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2934 continue;
2935
2936 // The increment must be loop-invariant so it can be kept in a register.
2937 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2938 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2939 if (!SE.isLoopInvariant(IncExpr, L))
2940 continue;
2941
2942 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2943 LastIncExpr = IncExpr;
2944 break;
2945 }
2946 }
2947 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2948 // bother for phi nodes, because they must be last in the chain.
2949 if (ChainIdx == NChains) {
2950 if (isa<PHINode>(UserInst))
2951 return;
2952 if (NChains >= MaxChains && !StressIVChain) {
2953 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2954 return;
2955 }
2956 LastIncExpr = OperExpr;
2957 // IVUsers may have skipped over sign/zero extensions. We don't currently
2958 // attempt to form chains involving extensions unless they can be hoisted
2959 // into this loop's AddRec.
2960 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2961 return;
2962 ++NChains;
2963 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2964 OperExprBase));
2965 ChainUsersVec.resize(NChains);
2966 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)
2967 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
2968 } else {
2969 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)
2970 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
2971 // Add this IV user to the end of the chain.
2972 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2973 }
2974 IVChain &Chain = IVChainVec[ChainIdx];
2975
2976 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2977 // This chain's NearUsers become FarUsers.
2978 if (!LastIncExpr->isZero()) {
2979 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2980 NearUsers.end());
2981 NearUsers.clear();
2982 }
2983
2984 // All other uses of IVOperand become near uses of the chain.
2985 // We currently ignore intermediate values within SCEV expressions, assuming
2986 // they will eventually be used be the current chain, or can be computed
2987 // from one of the chain increments. To be more precise we could
2988 // transitively follow its user and only add leaf IV users to the set.
2989 for (User *U : IVOper->users()) {
2990 Instruction *OtherUse = dyn_cast<Instruction>(U);
2991 if (!OtherUse)
2992 continue;
2993 // Uses in the chain will no longer be uses if the chain is formed.
2994 // Include the head of the chain in this iteration (not Chain.begin()).
2995 IVChain::const_iterator IncIter = Chain.Incs.begin();
2996 IVChain::const_iterator IncEnd = Chain.Incs.end();
2997 for( ; IncIter != IncEnd; ++IncIter) {
2998 if (IncIter->UserInst == OtherUse)
2999 break;
3000 }
3001 if (IncIter != IncEnd)
3002 continue;
3003
3004 if (SE.isSCEVable(OtherUse->getType())
3005 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3006 && IU.isIVUserOrOperand(OtherUse)) {
3007 continue;
3008 }
3009 NearUsers.insert(OtherUse);
3010 }
3011
3012 // Since this user is part of the chain, it's no longer considered a use
3013 // of the chain.
3014 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3015}
3016
3017/// Populate the vector of Chains.
3018///
3019/// This decreases ILP at the architecture level. Targets with ample registers,
3020/// multiple memory ports, and no register renaming probably don't want
3021/// this. However, such targets should probably disable LSR altogether.
3022///
3023/// The job of LSR is to make a reasonable choice of induction variables across
3024/// the loop. Subsequent passes can easily "unchain" computation exposing more
3025/// ILP *within the loop* if the target wants it.
3026///
3027/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3028/// will not reorder memory operations, it will recognize this as a chain, but
3029/// will generate redundant IV increments. Ideally this would be corrected later
3030/// by a smart scheduler:
3031/// = A[i]
3032/// = A[i+x]
3033/// A[i] =
3034/// A[i+x] =
3035///
3036/// TODO: Walk the entire domtree within this loop, not just the path to the
3037/// loop latch. This will discover chains on side paths, but requires
3038/// maintaining multiple copies of the Chains state.
3039void LSRInstance::CollectChains() {
3040 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3041 SmallVector<ChainUsers, 8> ChainUsersVec;
3042
3043 SmallVector<BasicBlock *,8> LatchPath;
3044 BasicBlock *LoopHeader = L->getHeader();
3045 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3046 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3047 LatchPath.push_back(Rung->getBlock());
3048 }
3049 LatchPath.push_back(LoopHeader);
3050
3051 // Walk the instruction stream from the loop header to the loop latch.
3052 for (BasicBlock *BB : reverse(LatchPath)) {
3053 for (Instruction &I : *BB) {
3054 // Skip instructions that weren't seen by IVUsers analysis.
3055 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3056 continue;
3057
3058 // Ignore users that are part of a SCEV expression. This way we only
3059 // consider leaf IV Users. This effectively rediscovers a portion of
3060 // IVUsers analysis but in program order this time.
3061 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3062 continue;
3063
3064 // Remove this instruction from any NearUsers set it may be in.
3065 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3066 ChainIdx < NChains; ++ChainIdx) {
3067 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3068 }
3069 // Search for operands that can be chained.
3070 SmallPtrSet<Instruction*, 4> UniqueOperands;
3071 User::op_iterator IVOpEnd = I.op_end();
3072 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3073 while (IVOpIter != IVOpEnd) {
3074 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3075 if (UniqueOperands.insert(IVOpInst).second)
3076 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3077 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3078 }
3079 } // Continue walking down the instructions.
3080 } // Continue walking down the domtree.
3081 // Visit phi backedges to determine if the chain can generate the IV postinc.
3082 for (PHINode &PN : L->getHeader()->phis()) {
3083 if (!SE.isSCEVable(PN.getType()))
3084 continue;
3085
3086 Instruction *IncV =
3087 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3088 if (IncV)
3089 ChainInstruction(&PN, IncV, ChainUsersVec);
3090 }
3091 // Remove any unprofitable chains.
3092 unsigned ChainIdx = 0;
3093 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3094 UsersIdx < NChains; ++UsersIdx) {
3095 if (!isProfitableChain(IVChainVec[UsersIdx],
3096 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3097 continue;
3098 // Preserve the chain at UsesIdx.
3099 if (ChainIdx != UsersIdx)
3100 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3101 FinalizeChain(IVChainVec[ChainIdx]);
3102 ++ChainIdx;
3103 }
3104 IVChainVec.resize(ChainIdx);
3105}
3106
3107void LSRInstance::FinalizeChain(IVChain &Chain) {
3108 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3108, __PRETTY_FUNCTION__))
;
3109 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)
;
3110
3111 for (const IVInc &Inc : Chain) {
3112 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3113 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3114 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3114, __PRETTY_FUNCTION__))
;
3115 IVIncSet.insert(UseI);
3116 }
3117}
3118
3119/// Return true if the IVInc can be folded into an addressing mode.
3120static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3121 Value *Operand, const TargetTransformInfo &TTI) {
3122 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3123 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3124 return false;
3125
3126 if (IncConst->getAPInt().getMinSignedBits() > 64)
3127 return false;
3128
3129 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3130 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3131 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3132 IncOffset, /*HasBaseReg=*/false))
3133 return false;
3134
3135 return true;
3136}
3137
3138/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3139/// user's operand from the previous IV user's operand.
3140void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3141 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3142 // Find the new IVOperand for the head of the chain. It may have been replaced
3143 // by LSR.
3144 const IVInc &Head = Chain.Incs[0];
3145 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3146 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3147 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3148 IVOpEnd, L, SE);
3149 Value *IVSrc = nullptr;
3150 while (IVOpIter != IVOpEnd) {
3151 IVSrc = getWideOperand(*IVOpIter);
3152
3153 // If this operand computes the expression that the chain needs, we may use
3154 // it. (Check this after setting IVSrc which is used below.)
3155 //
3156 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3157 // narrow for the chain, so we can no longer use it. We do allow using a
3158 // wider phi, assuming the LSR checked for free truncation. In that case we
3159 // should already have a truncate on this operand such that
3160 // getSCEV(IVSrc) == IncExpr.
3161 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3162 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3163 break;
3164 }
3165 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3166 }
3167 if (IVOpIter == IVOpEnd) {
3168 // Gracefully give up on this chain.
3169 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)
;
3170 return;
3171 }
3172 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3172, __PRETTY_FUNCTION__))
;
3173
3174 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3175 Type *IVTy = IVSrc->getType();
3176 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3177 const SCEV *LeftOverExpr = nullptr;
3178 for (const IVInc &Inc : Chain) {
3179 Instruction *InsertPt = Inc.UserInst;
3180 if (isa<PHINode>(InsertPt))
3181 InsertPt = L->getLoopLatch()->getTerminator();
3182
3183 // IVOper will replace the current IV User's operand. IVSrc is the IV
3184 // value currently held in a register.
3185 Value *IVOper = IVSrc;
3186 if (!Inc.IncExpr->isZero()) {
3187 // IncExpr was the result of subtraction of two narrow values, so must
3188 // be signed.
3189 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3190 LeftOverExpr = LeftOverExpr ?
3191 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3192 }
3193 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3194 // Expand the IV increment.
3195 Rewriter.clearPostInc();
3196 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3197 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3198 SE.getUnknown(IncV));
3199 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3200
3201 // If an IV increment can't be folded, use it as the next IV value.
3202 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3203 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3203, __PRETTY_FUNCTION__))
;
3204 IVSrc = IVOper;
3205 LeftOverExpr = nullptr;
3206 }
3207 }
3208 Type *OperTy = Inc.IVOperand->getType();
3209 if (IVTy != OperTy) {
3210 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3211, __PRETTY_FUNCTION__))
3211 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3211, __PRETTY_FUNCTION__))
;
3212 IRBuilder<> Builder(InsertPt);
3213 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3214 }
3215 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3216 if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3217 DeadInsts.emplace_back(OperandIsInstr);
3218 }
3219 // If LSR created a new, wider phi, we may also replace its postinc. We only
3220 // do this if we also found a wide value for the head of the chain.
3221 if (isa<PHINode>(Chain.tailUserInst())) {
3222 for (PHINode &Phi : L->getHeader()->phis()) {
3223 if (!isCompatibleIVType(&Phi, IVSrc))
3224 continue;
3225 Instruction *PostIncV = dyn_cast<Instruction>(
3226 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3227 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3228 continue;
3229 Value *IVOper = IVSrc;
3230 Type *PostIncTy = PostIncV->getType();
3231 if (IVTy != PostIncTy) {
3232 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3232, __PRETTY_FUNCTION__))
;
3233 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3234 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3235 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3236 }
3237 Phi.replaceUsesOfWith(PostIncV, IVOper);
3238 DeadInsts.emplace_back(PostIncV);
3239 }
3240 }
3241}
3242
3243void LSRInstance::CollectFixupsAndInitialFormulae() {
3244 BranchInst *ExitBranch = nullptr;
3245 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3246
3247 for (const IVStrideUse &U : IU) {
3248 Instruction *UserInst = U.getUser();
3249 // Skip IV users that are part of profitable IV Chains.
3250 User::op_iterator UseI =
3251 find(UserInst->operands(), U.getOperandValToReplace());
3252 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3252, __PRETTY_FUNCTION__))
;
3253 if (IVIncSet.count(UseI)) {
3254 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)
;
3255 continue;
3256 }
3257
3258 LSRUse::KindType Kind = LSRUse::Basic;
3259 MemAccessTy AccessTy;
3260 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3261 Kind = LSRUse::Address;
3262 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3263 }
3264
3265 const SCEV *S = IU.getExpr(U);
3266 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3267
3268 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3269 // (N - i == 0), and this allows (N - i) to be the expression that we work
3270 // with rather than just N or i, so we can consider the register
3271 // requirements for both N and i at the same time. Limiting this code to
3272 // equality icmps is not a problem because all interesting loops use
3273 // equality icmps, thanks to IndVarSimplify.
3274 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3275 // If CI can be saved in some target, like replaced inside hardware loop
3276 // in PowerPC, no need to generate initial formulae for it.
3277 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3278 continue;
3279 if (CI->isEquality()) {
3280 // Swap the operands if needed to put the OperandValToReplace on the
3281 // left, for consistency.
3282 Value *NV = CI->getOperand(1);
3283 if (NV == U.getOperandValToReplace()) {
3284 CI->setOperand(1, CI->getOperand(0));
3285 CI->setOperand(0, NV);
3286 NV = CI->getOperand(1);
3287 Changed = true;
3288 }
3289
3290 // x == y --> x - y == 0
3291 const SCEV *N = SE.getSCEV(NV);
3292 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3293 // S is normalized, so normalize N before folding it into S
3294 // to keep the result normalized.
3295 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3296 Kind = LSRUse::ICmpZero;
3297 S = SE.getMinusSCEV(N, S);
3298 }
3299
3300 // -1 and the negations of all interesting strides (except the negation
3301 // of -1) are now also interesting.
3302 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3303 if (Factors[i] != -1)
3304 Factors.insert(-(uint64_t)Factors[i]);
3305 Factors.insert(-1);
3306 }
3307 }
3308
3309 // Get or create an LSRUse.
3310 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3311 size_t LUIdx = P.first;
3312 int64_t Offset = P.second;
3313 LSRUse &LU = Uses[LUIdx];
3314
3315 // Record the fixup.
3316 LSRFixup &LF = LU.getNewFixup();
3317 LF.UserInst = UserInst;
3318 LF.OperandValToReplace = U.getOperandValToReplace();
3319 LF.PostIncLoops = TmpPostIncLoops;
3320 LF.Offset = Offset;
3321 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3322
3323 if (!LU.WidestFixupType ||
3324 SE.getTypeSizeInBits(LU.WidestFixupType) <
3325 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3326 LU.WidestFixupType = LF.OperandValToReplace->getType();
3327
3328 // If this is the first use of this LSRUse, give it a formula.
3329 if (LU.Formulae.empty()) {
3330 InsertInitialFormula(S, LU, LUIdx);
3331 CountRegisters(LU.Formulae.back(), LUIdx);
3332 }
3333 }
3334
3335 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3336}
3337
3338/// Insert a formula for the given expression into the given use, separating out
3339/// loop-variant portions from loop-invariant and loop-computable portions.
3340void
3341LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3342 // Mark uses whose expressions cannot be expanded.
3343 if (!isSafeToExpand(S, SE))
3344 LU.RigidFormula = true;
3345
3346 Formula F;
3347 F.initialMatch(S, L, SE);
3348 bool Inserted = InsertFormula(LU, LUIdx, F);
3349 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3349, __PRETTY_FUNCTION__))
; (void)Inserted;
3350}
3351
3352/// Insert a simple single-register formula for the given expression into the
3353/// given use.
3354void
3355LSRInstance::InsertSupplementalFormula(const SCEV *S,
3356 LSRUse &LU, size_t LUIdx) {
3357 Formula F;
3358 F.BaseRegs.push_back(S);
3359 F.HasBaseReg = true;
3360 bool Inserted = InsertFormula(LU, LUIdx, F);
3361 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3361, __PRETTY_FUNCTION__))
; (void)Inserted;
3362}
3363
3364/// Note which registers are used by the given formula, updating RegUses.
3365void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3366 if (F.ScaledReg)
3367 RegUses.countRegister(F.ScaledReg, LUIdx);
3368 for (const SCEV *BaseReg : F.BaseRegs)
3369 RegUses.countRegister(BaseReg, LUIdx);
3370}
3371
3372/// If the given formula has not yet been inserted, add it to the list, and
3373/// return true. Return false otherwise.
3374bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3375 // Do not insert formula that we will not be able to expand.
3376 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3377, __PRETTY_FUNCTION__))
3377 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3377, __PRETTY_FUNCTION__))
;
3378
3379 if (!LU.InsertFormula(F, *L))
3380 return false;
3381
3382 CountRegisters(F, LUIdx);
3383 return true;
3384}
3385
3386/// Check for other uses of loop-invariant values which we're tracking. These
3387/// other uses will pin these values in registers, making them less profitable
3388/// for elimination.
3389/// TODO: This currently misses non-constant addrec step registers.
3390/// TODO: Should this give more weight to users inside the loop?
3391void
3392LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3393 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3394 SmallPtrSet<const SCEV *, 32> Visited;
3395
3396 while (!Worklist.empty()) {
3397 const SCEV *S = Worklist.pop_back_val();
3398
3399 // Don't process the same SCEV twice
3400 if (!Visited.insert(S).second)
3401 continue;
3402
3403 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3404 Worklist.append(N->op_begin(), N->op_end());
3405 else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
3406 Worklist.push_back(C->getOperand());
3407 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3408 Worklist.push_back(D->getLHS());
3409 Worklist.push_back(D->getRHS());
3410 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3411 const Value *V = US->getValue();
3412 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3413 // Look for instructions defined outside the loop.
3414 if (L->contains(Inst)) continue;
3415 } else if (isa<UndefValue>(V))
3416 // Undef doesn't have a live range, so it doesn't matter.
3417 continue;
3418 for (const Use &U : V->uses()) {
3419 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3420 // Ignore non-instructions.
3421 if (!UserInst)
3422 continue;
3423 // Ignore instructions in other functions (as can happen with
3424 // Constants).
3425 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3426 continue;
3427 // Ignore instructions not dominated by the loop.
3428 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3429 UserInst->getParent() :
3430 cast<PHINode>(UserInst)->getIncomingBlock(
3431 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3432 if (!DT.dominates(L->getHeader(), UseBB))
3433 continue;
3434 // Don't bother if the instruction is in a BB which ends in an EHPad.
3435 if (UseBB->getTerminator()->isEHPad())
3436 continue;
3437 // Don't bother rewriting PHIs in catchswitch blocks.
3438 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3439 continue;
3440 // Ignore uses which are part of other SCEV expressions, to avoid
3441 // analyzing them multiple times.
3442 if (SE.isSCEVable(UserInst->getType())) {
3443 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3444 // If the user is a no-op, look through to its uses.
3445 if (!isa<SCEVUnknown>(UserS))
3446 continue;
3447 if (UserS == US) {
3448 Worklist.push_back(
3449 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3450 continue;
3451 }
3452 }
3453 // Ignore icmp instructions which are already being analyzed.
3454 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3455 unsigned OtherIdx = !U.getOperandNo();
3456 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3457 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3458 continue;
3459 }
3460
3461 std::pair<size_t, int64_t> P = getUse(
3462 S, LSRUse::Basic, MemAccessTy());
3463 size_t LUIdx = P.first;
3464 int64_t Offset = P.second;
3465 LSRUse &LU = Uses[LUIdx];
3466 LSRFixup &LF = LU.getNewFixup();
3467 LF.UserInst = const_cast<Instruction *>(UserInst);
3468 LF.OperandValToReplace = U;
3469 LF.Offset = Offset;
3470 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3471 if (!LU.WidestFixupType ||
3472 SE.getTypeSizeInBits(LU.WidestFixupType) <
3473 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3474 LU.WidestFixupType = LF.OperandValToReplace->getType();
3475 InsertSupplementalFormula(US, LU, LUIdx);
3476 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3477 break;
3478 }
3479 }
3480 }
3481}
3482
3483/// Split S into subexpressions which can be pulled out into separate
3484/// registers. If C is non-null, multiply each subexpression by C.
3485///
3486/// Return remainder expression after factoring the subexpressions captured by
3487/// Ops. If Ops is complete, return NULL.
3488static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3489 SmallVectorImpl<const SCEV *> &Ops,
3490 const Loop *L,
3491 ScalarEvolution &SE,
3492 unsigned Depth = 0) {
3493 // Arbitrarily cap recursion to protect compile time.
3494 if (Depth >= 3)
3495 return S;
3496
3497 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3498 // Break out add operands.
3499 for (const SCEV *S : Add->operands()) {
3500 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3501 if (Remainder)
3502 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3503 }
3504 return nullptr;
3505 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3506 // Split a non-zero base out of an addrec.
3507 if (AR->getStart()->isZero() || !AR->isAffine())
3508 return S;
3509
3510 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3511 C, Ops, L, SE, Depth+1);
3512 // Split the non-zero AddRec unless it is part of a nested recurrence that
3513 // does not pertain to this loop.
3514 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3515 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3516 Remainder = nullptr;
3517 }
3518 if (Remainder != AR->getStart()) {
3519 if (!Remainder)
3520 Remainder = SE.getConstant(AR->getType(), 0);
3521 return SE.getAddRecExpr(Remainder,
3522 AR->getStepRecurrence(SE),
3523 AR->getLoop(),
3524 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3525 SCEV::FlagAnyWrap);
3526 }
3527 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3528 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3529 if (Mul->getNumOperands() != 2)
3530 return S;
3531 if (const SCEVConstant *Op0 =
3532 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3533 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3534 const SCEV *Remainder =
3535 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3536 if (Remainder)
3537 Ops.push_back(SE.getMulExpr(C, Remainder));
3538 return nullptr;
3539 }
3540 }
3541 return S;
3542}
3543
3544/// Return true if the SCEV represents a value that may end up as a
3545/// post-increment operation.
3546static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3547 LSRUse &LU, const SCEV *S, const Loop *L,
3548 ScalarEvolution &SE) {
3549 if (LU.Kind != LSRUse::Address ||
3550 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3551 return false;
3552 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3553 if (!AR)
3554 return false;
3555 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3556 if (!isa<SCEVConstant>(LoopStep))
3557 return false;
3558 // Check if a post-indexed load/store can be used.
3559 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3560 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3561 const SCEV *LoopStart = AR->getStart();
3562 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3563 return true;
3564 }
3565 return false;
3566}
3567
3568/// Helper function for LSRInstance::GenerateReassociations.
3569void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3570 const Formula &Base,
3571 unsigned Depth, size_t Idx,
3572 bool IsScaledReg) {
3573 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3574 // Don't generate reassociations for the base register of a value that
3575 // may generate a post-increment operator. The reason is that the
3576 // reassociations cause extra base+register formula to be created,
3577 // and possibly chosen, but the post-increment is more efficient.
3578 if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3579 return;
3580 SmallVector<const SCEV *, 8> AddOps;
3581 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3582 if (Remainder)
3583 AddOps.push_back(Remainder);
3584
3585 if (AddOps.size() == 1)
3586 return;
3587
3588 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3589 JE = AddOps.end();
3590 J != JE; ++J) {
3591 // Loop-variant "unknown" values are uninteresting; we won't be able to
3592 // do anything meaningful with them.
3593 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3594 continue;
3595
3596 // Don't pull a constant into a register if the constant could be folded
3597 // into an immediate field.
3598 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3599 LU.AccessTy, *J, Base.getNumRegs() > 1))
3600 continue;
3601
3602 // Collect all operands except *J.
3603 SmallVector<const SCEV *, 8> InnerAddOps(
3604 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3605 InnerAddOps.append(std::next(J),
3606 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3607
3608 // Don't leave just a constant behind in a register if the constant could
3609 // be folded into an immediate field.
3610 if (InnerAddOps.size() == 1 &&
3611 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3612 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3613 continue;
3614
3615 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3616 if (InnerSum->isZero())
3617 continue;
3618 Formula F = Base;
3619
3620 // Add the remaining pieces of the add back into the new formula.
3621 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3622 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3623 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3624 InnerSumSC->getValue()->getZExtValue())) {
3625 F.UnfoldedOffset =
3626 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3627 if (IsScaledReg)
3628 F.ScaledReg = nullptr;
3629 else
3630 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3631 } else if (IsScaledReg)
3632 F.ScaledReg = InnerSum;
3633 else
3634 F.BaseRegs[Idx] = InnerSum;
3635
3636 // Add J as its own register, or an unfolded immediate.
3637 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3638 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3639 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3640 SC->getValue()->getZExtValue()))
3641 F.UnfoldedOffset =
3642 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3643 else
3644 F.BaseRegs.push_back(*J);
3645 // We may have changed the number of register in base regs, adjust the
3646 // formula accordingly.
3647 F.canonicalize(*L);
3648
3649 if (InsertFormula(LU, LUIdx, F))
3650 // If that formula hadn't been seen before, recurse to find more like
3651 // it.
3652 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3653 // Because just Depth is not enough to bound compile time.
3654 // This means that every time AddOps.size() is greater 16^x we will add
3655 // x to Depth.
3656 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3657 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3658 }
3659}
3660
3661/// Split out subexpressions from adds and the bases of addrecs.
3662void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3663 Formula Base, unsigned Depth) {
3664 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3664, __PRETTY_FUNCTION__))
;
3665 // Arbitrarily cap recursion to protect compile time.
3666 if (Depth >= 3)
3667 return;
3668
3669 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3670 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3671
3672 if (Base.Scale == 1)
3673 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3674 /* Idx */ -1, /* IsScaledReg */ true);
3675}
3676
3677/// Generate a formula consisting of all of the loop-dominating registers added
3678/// into a single register.
3679void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3680 Formula Base) {
3681 // This method is only interesting on a plurality of registers.
3682 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3683 (Base.UnfoldedOffset != 0) <= 1)
3684 return;
3685
3686 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3687 // processing the formula.
3688 Base.unscale();
3689 SmallVector<const SCEV *, 4> Ops;
3690 Formula NewBase = Base;
3691 NewBase.BaseRegs.clear();
3692 Type *CombinedIntegerType = nullptr;
3693 for (const SCEV *BaseReg : Base.BaseRegs) {
3694 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3695 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3696 if (!CombinedIntegerType)
3697 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3698 Ops.push_back(BaseReg);
3699 }
3700 else
3701 NewBase.BaseRegs.push_back(BaseReg);
3702 }
3703
3704 // If no register is relevant, we're done.
3705 if (Ops.size() == 0)
3706 return;
3707
3708 // Utility function for generating the required variants of the combined
3709 // registers.
3710 auto GenerateFormula = [&](const SCEV *Sum) {
3711 Formula F = NewBase;
3712
3713 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3714 // opportunity to fold something. For now, just ignore such cases
3715 // rather than proceed with zero in a register.
3716 if (Sum->isZero())
3717 return;
3718
3719 F.BaseRegs.push_back(Sum);
3720 F.canonicalize(*L);
3721 (void)InsertFormula(LU, LUIdx, F);
3722 };
3723
3724 // If we collected at least two registers, generate a formula combining them.
3725 if (Ops.size() > 1) {
3726 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3727 GenerateFormula(SE.getAddExpr(OpsCopy));
3728 }
3729
3730 // If we have an unfolded offset, generate a formula combining it with the
3731 // registers collected.
3732 if (NewBase.UnfoldedOffset) {
3733 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3733, __PRETTY_FUNCTION__))
;
3734 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3735 true));
3736 NewBase.UnfoldedOffset = 0;
3737 GenerateFormula(SE.getAddExpr(Ops));
3738 }
3739}
3740
3741/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3742void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3743 const Formula &Base, size_t Idx,
3744 bool IsScaledReg) {
3745 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3746 GlobalValue *GV = ExtractSymbol(G, SE);
3747 if (G->isZero() || !GV)
3748 return;
3749 Formula F = Base;
3750 F.BaseGV = GV;
3751 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3752 return;
3753 if (IsScaledReg)
3754 F.ScaledReg = G;
3755 else
3756 F.BaseRegs[Idx] = G;
3757 (void)InsertFormula(LU, LUIdx, F);
3758}
3759
3760/// Generate reuse formulae using symbolic offsets.
3761void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3762 Formula Base) {
3763 // We can't add a symbolic offset if the address already contains one.
3764 if (Base.BaseGV) return;
3765
3766 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3767 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3768 if (Base.Scale == 1)
3769 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3770 /* IsScaledReg */ true);
3771}
3772
3773/// Helper function for LSRInstance::GenerateConstantOffsets.
3774void LSRInstance::GenerateConstantOffsetsImpl(
3775 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3776 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3777
3778 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3779 Formula F = Base;
3780 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3781
3782 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3783 LU.AccessTy, F)) {
3784 // Add the offset to the base register.
3785 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3786 // If it cancelled out, drop the base register, otherwise update it.
3787 if (NewG->isZero()) {
3788 if (IsScaledReg) {
3789 F.Scale = 0;
3790 F.ScaledReg = nullptr;
3791 } else
3792 F.deleteBaseReg(F.BaseRegs[Idx]);
3793 F.canonicalize(*L);
3794 } else if (IsScaledReg)
3795 F.ScaledReg = NewG;
3796 else
3797 F.BaseRegs[Idx] = NewG;
3798
3799 (void)InsertFormula(LU, LUIdx, F);
3800 }
3801 };
3802
3803 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3804
3805 // With constant offsets and constant steps, we can generate pre-inc
3806 // accesses by having the offset equal the step. So, for access #0 with a
3807 // step of 8, we generate a G - 8 base which would require the first access
3808 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3809 // for itself and hopefully becomes the base for other accesses. This means
3810 // means that a single pre-indexed access can be generated to become the new
3811 // base pointer for each iteration of the loop, resulting in no extra add/sub
3812 // instructions for pointer updating.
3813 if (FavorBackedgeIndex && LU.Kind == LSRUse::Address) {
3814 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3815 if (auto *StepRec =
3816 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3817 const APInt &StepInt = StepRec->getAPInt();
3818 int64_t Step = StepInt.isNegative() ?
3819 StepInt.getSExtValue() : StepInt.getZExtValue();
3820
3821 for (int64_t Offset : Worklist) {
3822 Offset -= Step;
3823 GenerateOffset(G, Offset);
3824 }
3825 }
3826 }
3827 }
3828 for (int64_t Offset : Worklist)
3829 GenerateOffset(G, Offset);
3830
3831 int64_t Imm = ExtractImmediate(G, SE);
3832 if (G->isZero() || Imm == 0)
3833 return;
3834 Formula F = Base;
3835 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3836 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3837 return;
3838 if (IsScaledReg) {
3839 F.ScaledReg = G;
3840 } else {
3841 F.BaseRegs[Idx] = G;
3842 // We may generate non canonical Formula if G is a recurrent expr reg
3843 // related with current loop while F.ScaledReg is not.
3844 F.canonicalize(*L);
3845 }
3846 (void)InsertFormula(LU, LUIdx, F);
3847}
3848
3849/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3850void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3851 Formula Base) {
3852 // TODO: For now, just add the min and max offset, because it usually isn't
3853 // worthwhile looking at everything inbetween.
3854 SmallVector<int64_t, 2> Worklist;
3855 Worklist.push_back(LU.MinOffset);
3856 if (LU.MaxOffset != LU.MinOffset)
3857 Worklist.push_back(LU.MaxOffset);
3858
3859 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3860 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3861 if (Base.Scale == 1)
3862 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3863 /* IsScaledReg */ true);
3864}
3865
3866/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3867/// == y -> x*c == y*c.
3868void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3869 Formula Base) {
3870 if (LU.Kind != LSRUse::ICmpZero) return;
3871
3872 // Determine the integer type for the base formula.
3873 Type *IntTy = Base.getType();
3874 if (!IntTy) return;
3875 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3876
3877 // Don't do this if there is more than one offset.
3878 if (LU.MinOffset != LU.MaxOffset) return;
3879
3880 // Check if transformation is valid. It is illegal to multiply pointer.
3881 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3882 return;
3883 for (const SCEV *BaseReg : Base.BaseRegs)
3884 if (BaseReg->getType()->isPointerTy())
3885 return;
3886 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3886, __PRETTY_FUNCTION__))
;
3887
3888 // Check each interesting stride.
3889 for (int64_t Factor : Factors) {
3890 // Check that the multiplication doesn't overflow.
3891 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3892 continue;
3893 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3894 if (NewBaseOffset / Factor != Base.BaseOffset)
3895 continue;
3896 // If the offset will be truncated at this use, check that it is in bounds.
3897 if (!IntTy->isPointerTy() &&
3898 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3899 continue;
3900
3901 // Check that multiplying with the use offset doesn't overflow.
3902 int64_t Offset = LU.MinOffset;
3903 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3904 continue;
3905 Offset = (uint64_t)Offset * Factor;
3906 if (Offset / Factor != LU.MinOffset)
3907 continue;
3908 // If the offset will be truncated at this use, check that it is in bounds.
3909 if (!IntTy->isPointerTy() &&
3910 !ConstantInt::isValueValidForType(IntTy, Offset))
3911 continue;
3912
3913 Formula F = Base;
3914 F.BaseOffset = NewBaseOffset;
3915
3916 // Check that this scale is legal.
3917 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3918 continue;
3919
3920 // Compensate for the use having MinOffset built into it.
3921 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3922
3923 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3924
3925 // Check that multiplying with each base register doesn't overflow.
3926 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3927 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3928 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3929 goto next;
3930 }
3931
3932 // Check that multiplying with the scaled register doesn't overflow.
3933 if (F.ScaledReg) {
3934 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3935 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3936 continue;
3937 }
3938
3939 // Check that multiplying with the unfolded offset doesn't overflow.
3940 if (F.UnfoldedOffset != 0) {
3941 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3942 Factor == -1)
3943 continue;
3944 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3945 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3946 continue;
3947 // If the offset will be truncated, check that it is in bounds.
3948 if (!IntTy->isPointerTy() &&
3949 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3950 continue;
3951 }
3952
3953 // If we make it here and it's legal, add it.
3954 (void)InsertFormula(LU, LUIdx, F);
3955 next:;
3956 }
3957}
3958
3959/// Generate stride factor reuse formulae by making use of scaled-offset address
3960/// modes, for example.
3961void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3962 // Determine the integer type for the base formula.
3963 Type *IntTy = Base.getType();
3964 if (!IntTy) return;
3965
3966 // If this Formula already has a scaled register, we can't add another one.
3967 // Try to unscale the formula to generate a better scale.
3968 if (Base.Scale != 0 && !Base.unscale())
3969 return;
3970
3971 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3971, __PRETTY_FUNCTION__))
;
3972
3973 // Check each interesting stride.
3974 for (int64_t Factor : Factors) {
3975 Base.Scale = Factor;
3976 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3977 // Check whether this scale is going to be legal.
3978 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3979 Base)) {
3980 // As a special-case, handle special out-of-loop Basic users specially.
3981 // TODO: Reconsider this special case.
3982 if (LU.Kind == LSRUse::Basic &&
3983 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3984 LU.AccessTy, Base) &&
3985 LU.AllFixupsOutsideLoop)
3986 LU.Kind = LSRUse::Special;
3987 else
3988 continue;
3989 }
3990 // For an ICmpZero, negating a solitary base register won't lead to
3991 // new solutions.
3992 if (LU.Kind == LSRUse::ICmpZero &&
3993 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3994 continue;
3995 // For each addrec base reg, if its loop is current loop, apply the scale.
3996 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3997 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3998 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3999 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4000 if (FactorS->isZero())
4001 continue;
4002 // Divide out the factor, ignoring high bits, since we'll be
4003 // scaling the value back up in the end.
4004 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
4005 // TODO: This could be optimized to avoid all the copying.
4006 Formula F = Base;
4007 F.ScaledReg = Quotient;
4008 F.deleteBaseReg(F.BaseRegs[i]);
4009 // The canonical representation of 1*reg is reg, which is already in
4010 // Base. In that case, do not try to insert the formula, it will be
4011 // rejected anyway.
4012 if (F.Scale == 1 && (F.BaseRegs.empty() ||
4013 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4014 continue;
4015 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4016 // non canonical Formula with ScaledReg's loop not being L.
4017 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4018 F.canonicalize(*L);
4019 (void)InsertFormula(LU, LUIdx, F);
4020 }
4021 }
4022 }
4023 }
4024}
4025
4026/// Generate reuse formulae from different IV types.
4027void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4028 // Don't bother truncating symbolic values.
4029 if (Base.BaseGV) return;
4030
4031 // Determine the integer type for the base formula.
4032 Type *DstTy = Base.getType();
4033 if (!DstTy) return;
4034 DstTy = SE.getEffectiveSCEVType(DstTy);
4035
4036 for (Type *SrcTy : Types) {
4037 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4038 Formula F = Base;
4039
4040 // Sometimes SCEV is able to prove zero during ext transform. It may
4041 // happen if SCEV did not do all possible transforms while creating the
4042 // initial node (maybe due to depth limitations), but it can do them while
4043 // taking ext.
4044 if (F.ScaledReg) {
4045 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4046 if (NewScaledReg->isZero())
4047 continue;
4048 F.ScaledReg = NewScaledReg;
4049 }
4050 bool HasZeroBaseReg = false;
4051 for (const SCEV *&BaseReg : F.BaseRegs) {
4052 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4053 if (NewBaseReg->isZero()) {
4054 HasZeroBaseReg = true;
4055 break;
4056 }
4057 BaseReg = NewBaseReg;
4058 }
4059 if (HasZeroBaseReg)
4060 continue;
4061
4062 // TODO: This assumes we've done basic processing on all uses and
4063 // have an idea what the register usage is.
4064 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4065 continue;
4066
4067 F.canonicalize(*L);
4068 (void)InsertFormula(LU, LUIdx, F);
4069 }
4070 }
4071}
4072
4073namespace {
4074
4075/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4076/// modifications so that the search phase doesn't have to worry about the data
4077/// structures moving underneath it.
4078struct WorkItem {
4079 size_t LUIdx;
4080 int64_t Imm;
4081 const SCEV *OrigReg;
4082
4083 WorkItem(size_t LI, int64_t I, const SCEV *R)
4084 : LUIdx(LI), Imm(I), OrigReg(R) {}
4085
4086 void print(raw_ostream &OS) const;
4087 void dump() const;
4088};
4089
4090} // end anonymous namespace
4091
4092#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4093void WorkItem::print(raw_ostream &OS) const {
4094 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4095 << " , add offset " << Imm;
4096}
4097
4098LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4099 print(errs()); errs() << '\n';
4100}
4101#endif
4102
4103/// Look for registers which are a constant distance apart and try to form reuse
4104/// opportunities between them.
4105void LSRInstance::GenerateCrossUseConstantOffsets() {
4106 // Group the registers by their value without any added constant offset.
4107 using ImmMapTy = std::map<int64_t, const SCEV *>;
4108
4109 DenseMap<const SCEV *, ImmMapTy> Map;
4110 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4111 SmallVector<const SCEV *, 8> Sequence;
4112 for (const SCEV *Use : RegUses) {
4113 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4114 int64_t Imm = ExtractImmediate(Reg, SE);
4115 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4116 if (Pair.second)
4117 Sequence.push_back(Reg);
4118 Pair.first->second.insert(std::make_pair(Imm, Use));
4119 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4120 }
4121
4122 // Now examine each set of registers with the same base value. Build up
4123 // a list of work to do and do the work in a separate step so that we're
4124 // not adding formulae and register counts while we're searching.
4125 SmallVector<WorkItem, 32> WorkItems;
4126 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4127 for (const SCEV *Reg : Sequence) {
4128 const ImmMapTy &Imms = Map.find(Reg)->second;
4129
4130 // It's not worthwhile looking for reuse if there's only one offset.
4131 if (Imms.size() == 1)
4132 continue;
4133
4134 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)
4135 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)
4136 : 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)
4137 << ' ' << 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)
4138 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)
;
4139
4140 // Examine each offset.
4141 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4142 J != JE; ++J) {
4143 const SCEV *OrigReg = J->second;
4144
4145 int64_t JImm = J->first;
4146 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4147
4148 if (!isa<SCEVConstant>(OrigReg) &&
4149 UsedByIndicesMap[Reg].count() == 1) {
4150 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)
4151 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4152 continue;
4153 }
4154
4155 // Conservatively examine offsets between this orig reg a few selected
4156 // other orig regs.
4157 int64_t First = Imms.begin()->first;
4158 int64_t Last = std::prev(Imms.end())->first;
4159 // Compute (First + Last) / 2 without overflow using the fact that
4160 // First + Last = 2 * (First + Last) + (First ^ Last).
4161 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4162 // If the result is negative and First is odd and Last even (or vice versa),
4163 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4164 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4165 ImmMapTy::const_iterator OtherImms[] = {
4166 Imms.begin(), std::prev(Imms.end()),
4167 Imms.lower_bound(Avg)};
4168 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4169 ImmMapTy::const_iterator M = OtherImms[i];
4170 if (M == J || M == JE) continue;
4171
4172 // Compute the difference between the two.
4173 int64_t Imm = (uint64_t)JImm - M->first;
4174 for (unsigned LUIdx : UsedByIndices.set_bits())
4175 // Make a memo of this use, offset, and register tuple.
4176 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4177 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4178 }
4179 }
4180 }
4181
4182 Map.clear();
4183 Sequence.clear();
4184 UsedByIndicesMap.clear();
4185 UniqueItems.clear();
4186
4187 // Now iterate through the worklist and add new formulae.
4188 for (const WorkItem &WI : WorkItems) {
4189 size_t LUIdx = WI.LUIdx;
4190 LSRUse &LU = Uses[LUIdx];
4191 int64_t Imm = WI.Imm;
4192 const SCEV *OrigReg = WI.OrigReg;
4193
4194 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4195 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4196 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4197
4198 // TODO: Use a more targeted data structure.
4199 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4200 Formula F = LU.Formulae[L];
4201 // FIXME: The code for the scaled and unscaled registers looks
4202 // very similar but slightly different. Investigate if they
4203 // could be merged. That way, we would not have to unscale the
4204 // Formula.
4205 F.unscale();
4206 // Use the immediate in the scaled register.
4207 if (F.ScaledReg == OrigReg) {
4208 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4209 // Don't create 50 + reg(-50).
4210 if (F.referencesReg(SE.getSCEV(
4211 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4212 continue;
4213 Formula NewF = F;
4214 NewF.BaseOffset = Offset;
4215 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4216 NewF))
4217 continue;
4218 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4219
4220 // If the new scale is a constant in a register, and adding the constant
4221 // value to the immediate would produce a value closer to zero than the
4222 // immediate itself, then the formula isn't worthwhile.
4223 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4224 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4225 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4226 .ule(std::abs(NewF.BaseOffset)))
4227 continue;
4228
4229 // OK, looks good.
4230 NewF.canonicalize(*this->L);
4231 (void)InsertFormula(LU, LUIdx, NewF);
4232 } else {
4233 // Use the immediate in a base register.
4234 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4235 const SCEV *BaseReg = F.BaseRegs[N];
4236 if (BaseReg != OrigReg)
4237 continue;
4238 Formula NewF = F;
4239 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4240 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4241 LU.Kind, LU.AccessTy, NewF)) {
4242 if (TTI.shouldFavorPostInc() &&
4243 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4244 continue;
4245 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4246 continue;
4247 NewF = F;
4248 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4249 }
4250 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4251
4252 // If the new formula has a constant in a register, and adding the
4253 // constant value to the immediate would produce a value closer to
4254 // zero than the immediate itself, then the formula isn't worthwhile.
4255 for (const SCEV *NewReg : NewF.BaseRegs)
4256 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4257 if ((C->getAPInt() + NewF.BaseOffset)
4258 .abs()
4259 .slt(std::abs(NewF.BaseOffset)) &&
4260 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4261 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4262 goto skip_formula;
4263
4264 // Ok, looks good.
4265 NewF.canonicalize(*this->L);
4266 (void)InsertFormula(LU, LUIdx, NewF);
4267 break;
4268 skip_formula:;
4269 }
4270 }
4271 }
4272 }
4273}
4274
4275/// Generate formulae for each use.
4276void
4277LSRInstance::GenerateAllReuseFormulae() {
4278 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4279 // queries are more precise.
4280 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4281 LSRUse &LU = Uses[LUIdx];
4282 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4283 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4284 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4285 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4286 }
4287 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4288 LSRUse &LU = Uses[LUIdx];
4289 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4290 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4291 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4292 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4293 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4294 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4295 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4296 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4297 }
4298 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4299 LSRUse &LU = Uses[LUIdx];
4300 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4301 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4302 }
4303
4304 GenerateCrossUseConstantOffsets();
4305
4306 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4307 "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)
4308 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4309}
4310
4311/// If there are multiple formulae with the same set of registers used
4312/// by other uses, pick the best one and delete the others.
4313void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4314 DenseSet<const SCEV *> VisitedRegs;
4315 SmallPtrSet<const SCEV *, 16> Regs;
4316 SmallPtrSet<const SCEV *, 16> LoserRegs;
4317#ifndef NDEBUG
4318 bool ChangedFormulae = false;
4319#endif
4320
4321 // Collect the best formula for each unique set of shared registers. This
4322 // is reset for each use.
4323 using BestFormulaeTy =
4324 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4325
4326 BestFormulaeTy BestFormulae;
4327
4328 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4329 LSRUse &LU = Uses[LUIdx];
4330 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)
4331 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4332
4333 bool Any = false;
4334 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4335 FIdx != NumForms; ++FIdx) {
4336 Formula &F = LU.Formulae[FIdx];
4337
4338 // Some formulas are instant losers. For example, they may depend on
4339 // nonexistent AddRecs from other loops. These need to be filtered
4340 // immediately, otherwise heuristics could choose them over others leading
4341 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4342 // avoids the need to recompute this information across formulae using the
4343 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4344 // the corresponding bad register from the Regs set.
4345 Cost CostF(L, SE, TTI);
4346 Regs.clear();
4347 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4348 if (CostF.isLoser()) {
4349 // During initial formula generation, undesirable formulae are generated
4350 // by uses within other loops that have some non-trivial address mode or
4351 // use the postinc form of the IV. LSR needs to provide these formulae
4352 // as the basis of rediscovering the desired formula that uses an AddRec
4353 // corresponding to the existing phi. Once all formulae have been
4354 // generated, these initial losers may be pruned.
4355 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)
4356 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4357 }
4358 else {
4359 SmallVector<const SCEV *, 4> Key;
4360 for (const SCEV *Reg : F.BaseRegs) {
4361 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4362 Key.push_back(Reg);
4363 }
4364 if (F.ScaledReg &&
4365 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4366 Key.push_back(F.ScaledReg);
4367 // Unstable sort by host order ok, because this is only used for
4368 // uniquifying.
4369 llvm::sort(Key);
4370
4371 std::pair<BestFormulaeTy::const_iterator, bool> P =
4372 BestFormulae.insert(std::make_pair(Key, FIdx));
4373 if (P.second)
4374 continue;
4375
4376 Formula &Best = LU.Formulae[P.first->second];
4377
4378 Cost CostBest(L, SE, TTI);
4379 Regs.clear();
4380 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4381 if (CostF.isLess(CostBest))
4382 std::swap(F, Best);
4383 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)
4384 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)
4385 " 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)
4386 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)
;
4387 }
4388#ifndef NDEBUG
4389 ChangedFormulae = true;
4390#endif
4391 LU.DeleteFormula(F);
4392 --FIdx;
4393 --NumForms;
4394 Any = true;
4395 }
4396
4397 // Now that we've filtered out some formulae, recompute the Regs set.
4398 if (Any)
4399 LU.RecomputeRegs(LUIdx, RegUses);
4400
4401 // Reset this to prepare for the next use.
4402 BestFormulae.clear();
4403 }
4404
4405 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)
4406 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)
4407 "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)
4408 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)
4409 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4410}
4411
4412/// Estimate the worst-case number of solutions the solver might have to
4413/// consider. It almost never considers this many solutions because it prune the
4414/// search space, but the pruning isn't always sufficient.
4415size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4416 size_t Power = 1;
4417 for (const LSRUse &LU : Uses) {
4418 size_t FSize = LU.Formulae.size();
4419 if (FSize >= ComplexityLimit) {
4420 Power = ComplexityLimit;
4421 break;
4422 }
4423 Power *= FSize;
4424 if (Power >= ComplexityLimit)
4425 break;
4426 }
4427 return Power;
4428}
4429
4430/// When one formula uses a superset of the registers of another formula, it
4431/// won't help reduce register pressure (though it may not necessarily hurt
4432/// register pressure); remove it to simplify the system.
4433void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4434 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4435 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)
;
4436
4437 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)
4438 "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)
4439 "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)
;
4440
4441 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4442 LSRUse &LU = Uses[LUIdx];
4443 bool Any = false;
4444 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4445 Formula &F = LU.Formulae[i];
4446 // Look for a formula with a constant or GV in a register. If the use
4447 // also has a formula with that same value in an immediate field,
4448 // delete the one that uses a register.
4449 for (SmallVectorImpl<const SCEV *>::const_iterator
4450 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4451 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4452 Formula NewF = F;
4453 //FIXME: Formulas should store bitwidth to do wrapping properly.
4454 // See PR41034.
4455 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4456 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4457 (I - F.BaseRegs.begin()));
4458 if (LU.HasFormulaWithSameRegs(NewF)) {
4459 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4460 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4461 LU.DeleteFormula(F);
4462 --i;
4463 --e;
4464 Any = true;
4465 break;
4466 }
4467 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4468 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4469 if (!F.BaseGV) {
4470 Formula NewF = F;
4471 NewF.BaseGV = GV;
4472 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4473 (I - F.BaseRegs.begin()));
4474 if (LU.HasFormulaWithSameRegs(NewF)) {
4475 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4476 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4477 LU.DeleteFormula(F);
4478 --i;
4479 --e;
4480 Any = true;
4481 break;
4482 }
4483 }
4484 }
4485 }
4486 }
4487 if (Any)
4488 LU.RecomputeRegs(LUIdx, RegUses);
4489 }
4490
4491 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)
;
4492 }
4493}
4494
4495/// When there are many registers for expressions like A, A+1, A+2, etc.,
4496/// allocate a single register for them.
4497void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4498 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4499 return;
4500
4501 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)
4502 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)
4503 "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)
4504 "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)
;
4505
4506 // This is especially useful for unrolled loops.
4507
4508 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4509 LSRUse &LU = Uses[LUIdx];
4510 for (const Formula &F : LU.Formulae) {
4511 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4512 continue;
4513
4514 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4515 if (!LUThatHas)
4516 continue;
4517
4518 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4519 LU.Kind, LU.AccessTy))
4520 continue;
4521
4522 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)
;
4523
4524 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4525
4526 // Transfer the fixups of LU to LUThatHas.
4527 for (LSRFixup &Fixup : LU.Fixups) {
4528 Fixup.Offset += F.BaseOffset;
4529 LUThatHas->pushFixup(Fixup);
4530 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)
;
4531 }
4532
4533 // Delete formulae from the new use which are no longer legal.
4534 bool Any = false;
4535 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4536 Formula &F = LUThatHas->Formulae[i];
4537 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4538 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4539 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)
;
4540 LUThatHas->DeleteFormula(F);
4541 --i;
4542 --e;
4543 Any = true;
4544 }
4545 }
4546
4547 if (Any)
4548 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4549
4550 // Delete the old use.
4551 DeleteUse(LU, LUIdx);
4552 --LUIdx;
4553 --NumUses;
4554 break;
4555 }
4556 }
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/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4562/// we've done more filtering, as it may be able to find more formulae to
4563/// eliminate.
4564void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4565 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4566 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)
;
4567
4568 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)
4569 "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)
;
4570
4571 FilterOutUndesirableDedicatedRegisters();
4572
4573 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)
;
4574 }
4575}
4576
4577/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4578/// Pick the best one and delete the others.
4579/// This narrowing heuristic is to keep as many formulae with different
4580/// Scale and ScaledReg pair as possible while narrowing the search space.
4581/// The benefit is that it is more likely to find out a better solution
4582/// from a formulae set with more Scale and ScaledReg variations than
4583/// a formulae set with the same Scale and ScaledReg. The picking winner
4584/// reg heuristic will often keep the formulae with the same Scale and
4585/// ScaledReg and filter others, and we want to avoid that if possible.
4586void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4587 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4588 return;
4589
4590 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)
4591 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)
4592 "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)
4593 "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)
;
4594
4595 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4596 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4597
4598 BestFormulaeTy BestFormulae;
4599#ifndef NDEBUG
4600 bool ChangedFormulae = false;
4601#endif
4602 DenseSet<const SCEV *> VisitedRegs;
4603 SmallPtrSet<const SCEV *, 16> Regs;
4604
4605 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4606 LSRUse &LU = Uses[LUIdx];
4607 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)
4608 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4609
4610 // Return true if Formula FA is better than Formula FB.
4611 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4612 // First we will try to choose the Formula with fewer new registers.
4613 // For a register used by current Formula, the more the register is
4614 // shared among LSRUses, the less we increase the register number
4615 // counter of the formula.
4616 size_t FARegNum = 0;
4617 for (const SCEV *Reg : FA.BaseRegs) {
4618 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4619 FARegNum += (NumUses - UsedByIndices.count() + 1);
4620 }
4621 size_t FBRegNum = 0;
4622 for (const SCEV *Reg : FB.BaseRegs) {
4623 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4624 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4625 }
4626 if (FARegNum != FBRegNum)
4627 return FARegNum < FBRegNum;
4628
4629 // If the new register numbers are the same, choose the Formula with
4630 // less Cost.
4631 Cost CostFA(L, SE, TTI);
4632 Cost CostFB(L, SE, TTI);
4633 Regs.clear();
4634 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4635 Regs.clear();
4636 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4637 return CostFA.isLess(CostFB);
4638 };
4639
4640 bool Any = false;
4641 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4642 ++FIdx) {
4643 Formula &F = LU.Formulae[FIdx];
4644 if (!F.ScaledReg)
4645 continue;
4646 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4647 if (P.second)
4648 continue;
4649
4650 Formula &Best = LU.Formulae[P.first->second];
4651 if (IsBetterThan(F, Best))
4652 std::swap(F, Best);
4653 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)
4654 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)
4655 " 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)
4656 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)
;
4657#ifndef NDEBUG
4658 ChangedFormulae = true;
4659#endif
4660 LU.DeleteFormula(F);
4661 --FIdx;
4662 --NumForms;
4663 Any = true;
4664 }
4665 if (Any)
4666 LU.RecomputeRegs(LUIdx, RegUses);
4667
4668 // Reset this to prepare for the next use.
4669 BestFormulae.clear();
4670 }
4671
4672 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)
4673 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)
4674 "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)
4675 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)
4676 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4677}
4678
4679/// If we are over the complexity limit, filter out any post-inc prefering
4680/// variables to only post-inc values.
4681void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
4682 if (!TTI.shouldFavorPostInc())
4683 return;
4684 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4685 return;
4686
4687 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"
"Narrowing the search space by choosing the lowest " "register Formula for PostInc Uses.\n"
; } } while (false)
4688 "Narrowing the search space by choosing the lowest "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by choosing the lowest " "register Formula for PostInc Uses.\n"
; } } while (false)
4689 "register Formula for PostInc Uses.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
"Narrowing the search space by choosing the lowest " "register Formula for PostInc Uses.\n"
; } } while (false)
;
4690
4691 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4692 LSRUse &LU = Uses[LUIdx];
4693
4694 if (LU.Kind != LSRUse::Address)
4695 continue;
4696 if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
4697 !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
4698 continue;
4699
4700 size_t MinRegs = std::numeric_limits<size_t>::max();
4701 for (const Formula &F : LU.Formulae)
4702 MinRegs = std::min(F.getNumRegs(), MinRegs);
4703
4704 bool Any = false;
4705 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4706 ++FIdx) {
4707 Formula &F = LU.Formulae[FIdx];
4708 if (F.getNumRegs() > MinRegs) {
4709 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"; } } while (false)
4710 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n"; } } while (false)
;
4711 LU.DeleteFormula(F);
4712 --FIdx;
4713 --NumForms;
4714 Any = true;
4715 }
4716 }
4717 if (Any)
4718 LU.RecomputeRegs(LUIdx, RegUses);
4719
4720 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4721 break;
4722 }
4723
4724 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)
;
4725}
4726
4727/// The function delete formulas with high registers number expectation.
4728/// Assuming we don't know the value of each formula (already delete
4729/// all inefficient), generate probability of not selecting for each
4730/// register.
4731/// For example,
4732/// Use1:
4733/// reg(a) + reg({0,+,1})
4734/// reg(a) + reg({-1,+,1}) + 1
4735/// reg({a,+,1})
4736/// Use2:
4737/// reg(b) + reg({0,+,1})
4738/// reg(b) + reg({-1,+,1}) + 1
4739/// reg({b,+,1})
4740/// Use3:
4741/// reg(c) + reg(b) + reg({0,+,1})
4742/// reg(c) + reg({b,+,1})
4743///
4744/// Probability of not selecting
4745/// Use1 Use2 Use3
4746/// reg(a) (1/3) * 1 * 1
4747/// reg(b) 1 * (1/3) * (1/2)
4748/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4749/// reg({-1,+,1}) (2/3) * (2/3) * 1
4750/// reg({a,+,1}) (2/3) * 1 * 1
4751/// reg({b,+,1}) 1 * (2/3) * (2/3)
4752/// reg(c) 1 * 1 * 0
4753///
4754/// Now count registers number mathematical expectation for each formula:
4755/// Note that for each use we exclude probability if not selecting for the use.
4756/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4757/// probabilty 1/3 of not selecting for Use1).
4758/// Use1:
4759/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4760/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4761/// reg({a,+,1}) 1
4762/// Use2:
4763/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4764/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4765/// reg({b,+,1}) 2/3
4766/// Use3:
4767/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4768/// reg(c) + reg({b,+,1}) 1 + 2/3
4769void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4770 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4771 return;
4772 // Ok, we have too many of formulae on our hands to conveniently handle.
4773 // Use a rough heuristic to thin out the list.
4774
4775 // Set of Regs wich will be 100% used in final solution.
4776 // Used in each formula of a solution (in example above this is reg(c)).
4777 // We can skip them in calculations.
4778 SmallPtrSet<const SCEV *, 4> UniqRegs;
4779 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)
;
4780
4781 // Map each register to probability of not selecting
4782 DenseMap <const SCEV *, float> RegNumMap;
4783 for (const SCEV *Reg : RegUses) {
4784 if (UniqRegs.count(Reg))
4785 continue;
4786 float PNotSel = 1;
4787 for (const LSRUse &LU : Uses) {
4788 if (!LU.Regs.count(Reg))
4789 continue;
4790 float P = LU.getNotSelectedProbability(Reg);
4791 if (P != 0.0)
4792 PNotSel *= P;
4793 else
4794 UniqRegs.insert(Reg);
4795 }
4796 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4797 }
4798
4799 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4800 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)
;
4801
4802 // Delete formulas where registers number expectation is high.
4803 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4804 LSRUse &LU = Uses[LUIdx];
4805 // If nothing to delete - continue.
4806 if (LU.Formulae.size() < 2)
4807 continue;
4808 // This is temporary solution to test performance. Float should be
4809 // replaced with round independent type (based on integers) to avoid
4810 // different results for different target builds.
4811 float FMinRegNum = LU.Formulae[0].getNumRegs();
4812 float FMinARegNum = LU.Formulae[0].getNumRegs();
4813 size_t MinIdx = 0;
4814 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4815 Formula &F = LU.Formulae[i];
4816 float FRegNum = 0;
4817 float FARegNum = 0;
4818 for (const SCEV *BaseReg : F.BaseRegs) {
4819 if (UniqRegs.count(BaseReg))
4820 continue;
4821 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4822 if (isa<SCEVAddRecExpr>(BaseReg))
4823 FARegNum +=
4824 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4825 }
4826 if (const SCEV *ScaledReg = F.ScaledReg) {
4827 if (!UniqRegs.count(ScaledReg)) {
4828 FRegNum +=
4829 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4830 if (isa<SCEVAddRecExpr>(ScaledReg))
4831 FARegNum +=
4832 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4833 }
4834 }
4835 if (FMinRegNum > FRegNum ||
4836 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4837 FMinRegNum = FRegNum;
4838 FMinARegNum = FARegNum;
4839 MinIdx = i;
4840 }
4841 }
4842 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)
4843 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)
;
4844 if (MinIdx != 0)
4845 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4846 while (LU.Formulae.size() != 1) {
4847 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
)
4848 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4849 LU.Formulae.pop_back();
4850 }
4851 LU.RecomputeRegs(LUIdx, RegUses);
4852 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4852, __PRETTY_FUNCTION__))
;
4853 Formula &F = LU.Formulae[0];
4854 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)
;
4855 // When we choose the formula, the regs become unique.
4856 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4857 if (F.ScaledReg)
4858 UniqRegs.insert(F.ScaledReg);
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/// Pick a register which seems likely to be profitable, and then in any use
4864/// which has any reference to that register, delete all formulae which do not
4865/// reference that register.
4866void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4867 // With all other options exhausted, loop until the system is simple
4868 // enough to handle.
4869 SmallPtrSet<const SCEV *, 4> Taken;
4870 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4871 // Ok, we have too many of formulae on our hands to conveniently handle.
4872 // Use a rough heuristic to thin out the list.
4873 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)
;
4874
4875 // Pick the register which is used by the most LSRUses, which is likely
4876 // to be a good reuse register candidate.
4877 const SCEV *Best = nullptr;
4878 unsigned BestNum = 0;
4879 for (const SCEV *Reg : RegUses) {
4880 if (Taken.count(Reg))
4881 continue;
4882 if (!Best) {
4883 Best = Reg;
4884 BestNum = RegUses.getUsedByIndices(Reg).count();
4885 } else {
4886 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4887 if (Count > BestNum) {
4888 Best = Reg;
4889 BestNum = Count;
4890 }
4891 }
4892 }
4893 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4893, __PRETTY_FUNCTION__))
;
4894
4895 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)
4896 << " 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)
;
4897 Taken.insert(Best);
4898
4899 // In any use with formulae which references this register, delete formulae
4900 // which don't reference it.
4901 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4902 LSRUse &LU = Uses[LUIdx];
4903 if (!LU.Regs.count(Best)) continue;
4904
4905 bool Any = false;
4906 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4907 Formula &F = LU.Formulae[i];
4908 if (!F.referencesReg(Best)) {
4909 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)
;
4910 LU.DeleteFormula(F);
4911 --e;
4912 --i;
4913 Any = true;
4914 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4914, __PRETTY_FUNCTION__))
;
4915 continue;
4916 }
4917 }
4918
4919 if (Any)
4920 LU.RecomputeRegs(LUIdx, RegUses);
4921 }
4922
4923 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)
;
4924 }
4925}
4926
4927/// If there are an extraordinary number of formulae to choose from, use some
4928/// rough heuristics to prune down the number of formulae. This keeps the main
4929/// solver from taking an extraordinary amount of time in some worst-case
4930/// scenarios.
4931void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4932 NarrowSearchSpaceByDetectingSupersets();
4933 NarrowSearchSpaceByCollapsingUnrolledCode();
4934 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4935 if (FilterSameScaledReg)
4936 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4937 NarrowSearchSpaceByFilterPostInc();
4938 if (LSRExpNarrow)
4939 NarrowSearchSpaceByDeletingCostlyFormulas();
4940 else
4941 NarrowSearchSpaceByPickingWinnerRegs();
4942}
4943
4944/// This is the recursive solver.
4945void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4946 Cost &SolutionCost,
4947 SmallVectorImpl<const Formula *> &Workspace,
4948 const Cost &CurCost,
4949 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4950 DenseSet<const SCEV *> &VisitedRegs) const {
4951 // Some ideas:
4952 // - prune more:
4953 // - use more aggressive filtering
4954 // - sort the formula so that the most profitable solutions are found first
4955 // - sort the uses too
4956 // - search faster:
4957 // - don't compute a cost, and then compare. compare while computing a cost
4958 // and bail early.
4959 // - track register sets with SmallBitVector
4960
4961 const LSRUse &LU = Uses[Workspace.size()];
4962
4963 // If this use references any register that's already a part of the
4964 // in-progress solution, consider it a requirement that a formula must
4965 // reference that register in order to be considered. This prunes out
4966 // unprofitable searching.
4967 SmallSetVector<const SCEV *, 4> ReqRegs;
4968 for (const SCEV *S : CurRegs)
4969 if (LU.Regs.count(S))
4970 ReqRegs.insert(S);
4971
4972 SmallPtrSet<const SCEV *, 16> NewRegs;
4973 Cost NewCost(L, SE, TTI);
4974 for (const Formula &F : LU.Formulae) {
4975 // Ignore formulae which may not be ideal in terms of register reuse of
4976 // ReqRegs. The formula should use all required registers before
4977 // introducing new ones.
4978 // This can sometimes (notably when trying to favour postinc) lead to
4979 // sub-optimial decisions. There it is best left to the cost modelling to
4980 // get correct.
4981 if (!TTI.shouldFavorPostInc() || LU.Kind != LSRUse::Address) {
4982 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4983 for (const SCEV *Reg : ReqRegs) {
4984 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4985 is_contained(F.BaseRegs, Reg)) {
4986 --NumReqRegsToFind;
4987 if (NumReqRegsToFind == 0)
4988 break;
4989 }
4990 }
4991 if (NumReqRegsToFind != 0) {
4992 // If none of the formulae satisfied the required registers, then we could
4993 // clear ReqRegs and try again. Currently, we simply give up in this case.
4994 continue;
4995 }
4996 }
4997
4998 // Evaluate the cost of the current formula. If it's already worse than
4999 // the current best, prune the search at that point.
5000 NewCost = CurCost;
5001 NewRegs = CurRegs;
5002 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
5003 if (NewCost.isLess(SolutionCost)) {
5004 Workspace.push_back(&F);
5005 if (Workspace.size() != Uses.size()) {
5006 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
5007 NewRegs, VisitedRegs);
5008 if (F.getNumRegs() == 1 && Workspace.size() == 1)
5009 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
5010 } else {
5011 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)
5012 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)
5013 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)
5014 << "- " << *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)
5015 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)
;
5016
5017 SolutionCost = NewCost;
5018 Solution = Workspace;
5019 }
5020 Workspace.pop_back();
5021 }
5022 }
5023}
5024
5025/// Choose one formula from each use. Return the results in the given Solution
5026/// vector.
5027void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
5028 SmallVector<const Formula *, 8> Workspace;
5029 Cost SolutionCost(L, SE, TTI);
5030 SolutionCost.Lose();
5031 Cost CurCost(L, SE, TTI);
5032 SmallPtrSet<const SCEV *, 16> CurRegs;
5033 DenseSet<const SCEV *> VisitedRegs;
5034 Workspace.reserve(Uses.size());
5035
5036 // SolveRecurse does all the work.
5037 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
5038 CurRegs, VisitedRegs);
5039 if (Solution.empty()) {
5040 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
5041 return;
5042 }
5043
5044 // Ok, we've now made all our decisions.
5045 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
)
5046 "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
)
5047 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
)
5048 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
)
5049 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
)
5050 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
)
5051 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
)
5052 " ";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
)
5053 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
)
5054 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
)
5055 })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
)
;
5056
5057 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5057, __PRETTY_FUNCTION__))
;
5058}
5059
5060/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5061/// we can go while still being dominated by the input positions. This helps
5062/// canonicalize the insert position, which encourages sharing.
5063BasicBlock::iterator
5064LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5065 const SmallVectorImpl<Instruction *> &Inputs)
5066 const {
5067 Instruction *Tentative = &*IP;
5068 while (true) {
5069 bool AllDominate = true;
5070 Instruction *BetterPos = nullptr;
5071 // Don't bother attempting to insert before a catchswitch, their basic block
5072 // cannot have other non-PHI instructions.
5073 if (isa<CatchSwitchInst>(Tentative))
5074 return IP;
5075
5076 for (Instruction *Inst : Inputs) {
5077 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5078 AllDominate = false;
5079 break;
5080 }
5081 // Attempt to find an insert position in the middle of the block,
5082 // instead of at the end, so that it can be used for other expansions.
5083 if (Tentative->getParent() == Inst->getParent() &&
5084 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5085 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5086 }
5087 if (!AllDominate)
5088 break;
5089 if (BetterPos)
5090 IP = BetterPos->getIterator();
5091 else
5092 IP = Tentative->getIterator();
5093
5094 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5095 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5096
5097 BasicBlock *IDom;
5098 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5099 if (!Rung) return IP;
5100 Rung = Rung->getIDom();
5101 if (!Rung) return IP;
5102 IDom = Rung->getBlock();
5103
5104 // Don't climb into a loop though.
5105 const Loop *IDomLoop = LI.getLoopFor(IDom);
5106 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5107 if (IDomDepth <= IPLoopDepth &&
5108 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5109 break;
5110 }
5111
5112 Tentative = IDom->getTerminator();
5113 }
5114
5115 return IP;
5116}
5117
5118/// Determine an input position which will be dominated by the operands and
5119/// which will dominate the result.
5120BasicBlock::iterator
5121LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5122 const LSRFixup &LF,
5123 const LSRUse &LU,
5124 SCEVExpander &Rewriter) const {
5125 // Collect some instructions which must be dominated by the
5126 // expanding replacement. These must be dominated by any operands that
5127 // will be required in the expansion.
5128 SmallVector<Instruction *, 4> Inputs;
5129 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5130 Inputs.push_back(I);
5131 if (LU.Kind == LSRUse::ICmpZero)
5132 if (Instruction *I =
5133 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5134 Inputs.push_back(I);
5135 if (LF.PostIncLoops.count(L)) {
5136 if (LF.isUseFullyOutsideLoop(L))
5137 Inputs.push_back(L->getLoopLatch()->getTerminator());
5138 else
5139 Inputs.push_back(IVIncInsertPos);
5140 }
5141 // The expansion must also be dominated by the increment positions of any
5142 // loops it for which it is using post-inc mode.
5143 for (const Loop *PIL : LF.PostIncLoops) {
5144 if (PIL == L) continue;
5145
5146 // Be dominated by the loop exit.
5147 SmallVector<BasicBlock *, 4> ExitingBlocks;
5148 PIL->getExitingBlocks(ExitingBlocks);
5149 if (!ExitingBlocks.empty()) {
5150 BasicBlock *BB = ExitingBlocks[0];
5151 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5152 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5153 Inputs.push_back(BB->getTerminator());
5154 }
5155 }
5156
5157 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5159, __PRETTY_FUNCTION__))
5158 && !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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5159, __PRETTY_FUNCTION__))
5159 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5159, __PRETTY_FUNCTION__))
;
5160
5161 // Then, climb up the immediate dominator tree as far as we can go while
5162 // still being dominated by the input positions.
5163 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5164
5165 // Don't insert instructions before PHI nodes.
5166 while (isa<PHINode>(IP)) ++IP;
5167
5168 // Ignore landingpad instructions.
5169 while (IP->isEHPad()) ++IP;
5170
5171 // Ignore debug intrinsics.
5172 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5173
5174 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5175 // IP consistent across expansions and allows the previously inserted
5176 // instructions to be reused by subsequent expansion.
5177 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5178 ++IP;
5179
5180 return IP;
5181}
5182
5183/// Emit instructions for the leading candidate expression for this LSRUse (this
5184/// is called "expanding").
5185Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5186 const Formula &F, BasicBlock::iterator IP,
5187 SCEVExpander &Rewriter,
5188 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5189 if (LU.RigidFormula)
5190 return LF.OperandValToReplace;
5191
5192 // Determine an input position which will be dominated by the operands and
5193 // which will dominate the result.
5194 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5195 Rewriter.setInsertPoint(&*IP);
5196
5197 // Inform the Rewriter if we have a post-increment use, so that it can
5198 // perform an advantageous expansion.
5199 Rewriter.setPostInc(LF.PostIncLoops);
5200
5201 // This is the type that the user actually needs.
5202 Type *OpTy = LF.OperandValToReplace->getType();
5203 // This will be the type that we'll initially expand to.
5204 Type *Ty = F.getType();
5205 if (!Ty)
5206 // No type known; just expand directly to the ultimate type.
5207 Ty = OpTy;
5208 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5209 // Expand directly to the ultimate type if it's the right size.
5210 Ty = OpTy;
5211 // This is the type to do integer arithmetic in.
5212 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5213
5214 // Build up a list of operands to add together to form the full base.
5215 SmallVector<const SCEV *, 8> Ops;
5216
5217 // Expand the BaseRegs portion.
5218 for (const SCEV *Reg : F.BaseRegs) {
5219 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5219, __PRETTY_FUNCTION__))
;
5220
5221 // If we're expanding for a post-inc user, make the post-inc adjustment.
5222 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5223 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5224 }
5225
5226 // Expand the ScaledReg portion.
5227 Value *ICmpScaledV = nullptr;
5228 if (F.Scale != 0) {
5229 const SCEV *ScaledS = F.ScaledReg;
5230
5231 // If we're expanding for a post-inc user, make the post-inc adjustment.
5232 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5233 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5234
5235 if (LU.Kind == LSRUse::ICmpZero) {
5236 // Expand ScaleReg as if it was part of the base regs.
5237 if (F.Scale == 1)
5238 Ops.push_back(
5239 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5240 else {
5241 // An interesting way of "folding" with an icmp is to use a negated
5242 // scale, which we'll implement by inserting it into the other operand
5243 // of the icmp.
5244 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5245, __PRETTY_FUNCTION__))
5245 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5245, __PRETTY_FUNCTION__))
;
5246 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5247 }
5248 } else {
5249 // Otherwise just expand the scaled register and an explicit scale,
5250 // which is expected to be matched as part of the address.
5251
5252 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5253 // Unless the addressing mode will not be folded.
5254 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5255 isAMCompletelyFolded(TTI, LU, F)) {
5256 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5257 Ops.clear();
5258 Ops.push_back(SE.getUnknown(FullV));
5259 }
5260 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5261 if (F.Scale != 1)
5262 ScaledS =
5263 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5264 Ops.push_back(ScaledS);
5265 }
5266 }
5267
5268 // Expand the GV portion.
5269 if (F.BaseGV) {
5270 // Flush the operand list to suppress SCEVExpander hoisting.
5271 if (!Ops.empty()) {
5272 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5273 Ops.clear();
5274 Ops.push_back(SE.getUnknown(FullV));
5275 }
5276 Ops.push_back(SE.getUnknown(F.BaseGV));
5277 }
5278
5279 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5280 // unfolded offsets. LSR assumes they both live next to their uses.
5281 if (!Ops.empty()) {
5282 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5283 Ops.clear();
5284 Ops.push_back(SE.getUnknown(FullV));
5285 }
5286
5287 // Expand the immediate portion.
5288 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5289 if (Offset != 0) {
5290 if (LU.Kind == LSRUse::ICmpZero) {
5291 // The other interesting way of "folding" with an ICmpZero is to use a
5292 // negated immediate.
5293 if (!ICmpScaledV)
5294 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5295 else {
5296 Ops.push_back(SE.getUnknown(ICmpScaledV));
5297 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5298 }
5299 } else {
5300 // Just add the immediate values. These again are expected to be matched
5301 // as part of the address.
5302 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5303 }
5304 }
5305
5306 // Expand the unfolded offset portion.
5307 int64_t UnfoldedOffset = F.UnfoldedOffset;
5308 if (UnfoldedOffset != 0) {
5309 // Just add the immediate values.
5310 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5311 UnfoldedOffset)));
5312 }
5313
5314 // Emit instructions summing all the operands.
5315 const SCEV *FullS = Ops.empty() ?
5316 SE.getConstant(IntTy, 0) :
5317 SE.getAddExpr(Ops);
5318 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5319
5320 // We're done expanding now, so reset the rewriter.
5321 Rewriter.clearPostInc();
5322
5323 // An ICmpZero Formula represents an ICmp which we're handling as a
5324 // comparison against zero. Now that we've expanded an expression for that
5325 // form, update the ICmp's other operand.
5326 if (LU.Kind == LSRUse::ICmpZero) {
5327 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5328 if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
5329 DeadInsts.emplace_back(OperandIsInstr);
5330 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5331, __PRETTY_FUNCTION__))
5331 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5331, __PRETTY_FUNCTION__))
;
5332 if (F.Scale == -1) {
5333 if (ICmpScaledV->getType() != OpTy) {
5334 Instruction *Cast =
5335 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5336 OpTy, false),
5337 ICmpScaledV, OpTy, "tmp", CI);
5338 ICmpScaledV = Cast;
5339 }
5340 CI->setOperand(1, ICmpScaledV);
5341 } else {
5342 // A scale of 1 means that the scale has been expanded as part of the
5343 // base regs.
5344 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5346, __PRETTY_FUNCTION__))
5345 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5346, __PRETTY_FUNCTION__))
5346 "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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5346, __PRETTY_FUNCTION__))
;
5347 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5348 -(uint64_t)Offset);
5349 if (C->getType() != OpTy)
5350 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5351 OpTy, false),
5352 C, OpTy);
5353
5354 CI->setOperand(1, C);
5355 }
5356 }
5357
5358 return FullV;
5359}
5360
5361/// Helper for Rewrite. PHI nodes are special because the use of their operands
5362/// effectively happens in their predecessor blocks, so the expression may need
5363/// to be expanded in multiple places.
5364void LSRInstance::RewriteForPHI(
5365 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5366 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5367 DenseMap<BasicBlock *, Value *> Inserted;
5368 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
5369 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3
Assuming pointer value is null
4
Taking true branch
5370 bool needUpdateFixups = false;
5371 BasicBlock *BB = PN->getIncomingBlock(i);
5372
5373 // If this is a critical edge, split the edge so that we do not insert
5374 // the code on all predecessor/successor paths. We do this unless this
5375 // is the canonical backedge for this loop, which complicates post-inc
5376 // users.
5377 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5
Assuming 'e' is equal to 1
6
Taking false branch
5378 !isa<IndirectBrInst>(BB->getTerminator()) &&
5379 !isa<CatchSwitchInst>(BB->getTerminator())) {
5380 BasicBlock *Parent = PN->getParent();
5381 Loop *PNLoop = LI.getLoopFor(Parent);
5382 if (!PNLoop || Parent != PNLoop->getHeader()) {
5383 // Split the critical edge.
5384 BasicBlock *NewBB = nullptr;
5385 if (!Parent->isLandingPad()) {
5386 NewBB =
5387 SplitCriticalEdge(BB, Parent,
5388 CriticalEdgeSplittingOptions(&DT, &LI, MSSAU)
5389 .setMergeIdenticalEdges()
5390 .setKeepOneInputPHIs());
5391 } else {
5392 SmallVector<BasicBlock*, 2> NewBBs;
5393 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5394 NewBB = NewBBs[0];
5395 }
5396 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5397 // phi predecessors are identical. The simple thing to do is skip
5398 // splitting in this case rather than complicate the API.
5399 if (NewBB) {
5400 // If PN is outside of the loop and BB is in the loop, we want to
5401 // move the block to be immediately before the PHI block, not
5402 // immediately after BB.
5403 if (L->contains(BB) && !L->contains(PN))
5404 NewBB->moveBefore(PN->getParent());
5405
5406 // Splitting the edge can reduce the number of PHI entries we have.
5407 e = PN->getNumIncomingValues();
5408 BB = NewBB;
5409 i = PN->getBasicBlockIndex(BB);
5410
5411 needUpdateFixups = true;
5412 }
5413 }
5414 }
5415
5416 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5417 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5418 if (!Pair.second)
7
Assuming field 'second' is true
8
Taking false branch
5419 PN->setIncomingValue(i, Pair.first->second);
5420 else {
5421 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5422 Rewriter, DeadInsts);
5423
5424 // If this is reuse-by-noop-cast, insert the noop cast.
5425 Type *OpTy = LF.OperandValToReplace->getType();
9
Called C++ object pointer is null
5426 if (FullV->getType() != OpTy)
5427 FullV =
5428 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5429 OpTy, false),
5430 FullV, LF.OperandValToReplace->getType(),
5431 "tmp", BB->getTerminator());
5432
5433 PN->setIncomingValue(i, FullV);
5434 Pair.first->second = FullV;
5435 }
5436
5437 // If LSR splits critical edge and phi node has other pending
5438 // fixup operands, we need to update those pending fixups. Otherwise
5439 // formulae will not be implemented completely and some instructions
5440 // will not be eliminated.
5441 if (needUpdateFixups) {
5442 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5443 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5444 // If fixup is supposed to rewrite some operand in the phi
5445 // that was just updated, it may be already moved to
5446 // another phi node. Such fixup requires update.
5447 if (Fixup.UserInst == PN) {
5448 // Check if the operand we try to replace still exists in the
5449 // original phi.
5450 bool foundInOriginalPHI = false;
5451 for (const auto &val : PN->incoming_values())
5452 if (val == Fixup.OperandValToReplace) {
5453 foundInOriginalPHI = true;
5454 break;
5455 }
5456
5457 // If fixup operand found in original PHI - nothing to do.
5458 if (foundInOriginalPHI)
5459 continue;
5460
5461 // Otherwise it might be moved to another PHI and requires update.
5462 // If fixup operand not found in any of the incoming blocks that
5463 // means we have already rewritten it - nothing to do.
5464 for (const auto &Block : PN->blocks())
5465 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5466 ++I) {
5467 PHINode *NewPN = cast<PHINode>(I);
5468 for (const auto &val : NewPN->incoming_values())
5469 if (val == Fixup.OperandValToReplace)
5470 Fixup.UserInst = NewPN;
5471 }
5472 }
5473 }
5474 }
5475}
5476
5477/// Emit instructions for the leading candidate expression for this LSRUse (this
5478/// is called "expanding"), and update the UserInst to reference the newly
5479/// expanded value.
5480void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5481 const Formula &F, SCEVExpander &Rewriter,
5482 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5483 // First, find an insertion point that dominates UserInst. For PHI nodes,
5484 // find the nearest block which dominates all the relevant uses.
5485 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5486 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5487 } else {
5488 Value *FullV =
5489 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5490
5491 // If this is reuse-by-noop-cast, insert the noop cast.
5492 Type *OpTy = LF.OperandValToReplace->getType();
5493 if (FullV->getType() != OpTy) {
5494 Instruction *Cast =
5495 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5496 FullV, OpTy, "tmp", LF.UserInst);
5497 FullV = Cast;
5498 }
5499
5500 // Update the user. ICmpZero is handled specially here (for now) because
5501 // Expand may have updated one of the operands of the icmp already, and
5502 // its new value may happen to be equal to LF.OperandValToReplace, in
5503 // which case doing replaceUsesOfWith leads to replacing both operands
5504 // with the same value. TODO: Reorganize this.
5505 if (LU.Kind == LSRUse::ICmpZero)
5506 LF.UserInst->setOperand(0, FullV);
5507 else
5508 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5509 }
5510
5511 if (auto *OperandIsInstr = dyn_cast<Instruction>(LF.OperandValToReplace))
5512 DeadInsts.emplace_back(OperandIsInstr);
5513}
5514
5515/// Rewrite all the fixup locations with new values, following the chosen
5516/// solution.
5517void LSRInstance::ImplementSolution(
5518 const SmallVectorImpl<const Formula *> &Solution) {
5519 // Keep track of instructions we may have made dead, so that
5520 // we can remove them after we are done working.
5521 SmallVector<WeakTrackingVH, 16> DeadInsts;
5522
5523 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(), "lsr",
5524 false);
5525#ifndef NDEBUG
5526 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5527#endif
5528 Rewriter.disableCanonicalMode();
5529 Rewriter.enableLSRMode();
5530 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5531
5532 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5533 for (const IVChain &Chain : IVChainVec) {
5534 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5535 Rewriter.setChainedPhi(PN);
5536 }
5537
5538 // Expand the new value definitions and update the users.
5539 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5540 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5541 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5542 Changed = true;
5543 }
5544
5545 for (const IVChain &Chain : IVChainVec) {
5546 GenerateIVChain(Chain, Rewriter, DeadInsts);
5547 Changed = true;
5548 }
5549 // Clean up after ourselves. This must be done before deleting any
5550 // instructions.
5551 Rewriter.clear();
5552
5553 Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts,
5554 &TLI, MSSAU);
5555}
5556
5557LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5558 DominatorTree &DT, LoopInfo &LI,
5559 const TargetTransformInfo &TTI, AssumptionCache &AC,
5560 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU)
5561 : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), TLI(TLI), TTI(TTI), L(L),
5562 MSSAU(MSSAU), FavorBackedgeIndex(EnableBackedgeIndexing &&
5563 TTI.shouldFavorBackedgeIndex(L)) {
5564 // If LoopSimplify form is not available, stay out of trouble.
5565 if (!L->isLoopSimplifyForm())
5566 return;
5567
5568 // If there's no interesting work to be done, bail early.
5569 if (IU.empty()) return;
5570
5571 // If there's too much analysis to be done, bail early. We won't be able to
5572 // model the problem anyway.
5573 unsigned NumUsers = 0;
5574 for (const IVStrideUse &U : IU) {
5575 if (++NumUsers > MaxIVUsers) {
5576 (void)U;
5577 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)
5578 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5579 return;
5580 }
5581 // Bail out if we have a PHI on an EHPad that gets a value from a
5582 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5583 // no good place to stick any instructions.
5584 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5585 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5586 if (isa<FuncletPadInst>(FirstNonPHI) ||
5587 isa<CatchSwitchInst>(FirstNonPHI))
5588 for (BasicBlock *PredBB : PN->blocks())
5589 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5590 return;
5591 }
5592 }
5593
5594#ifndef NDEBUG
5595 // All dominating loops must have preheaders, or SCEVExpander may not be able
5596 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5597 //
5598 // IVUsers analysis should only create users that are dominated by simple loop
5599 // headers. Since this loop should dominate all of its users, its user list
5600 // should be empty if this loop itself is not within a simple loop nest.
5601 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5602 Rung; Rung = Rung->getIDom()) {
5603 BasicBlock *BB = Rung->getBlock();
5604 const Loop *DomLoop = LI.getLoopFor(BB);
5605 if (DomLoop && DomLoop->getHeader() == BB) {
5606 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5606, __PRETTY_FUNCTION__))
;
5607 }
5608 }
5609#endif // DEBUG
5610
5611 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)
5612 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)
5613 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5614
5615 // First, perform some low-level loop optimizations.
5616 OptimizeShadowIV();
5617 OptimizeLoopTermCond();
5618
5619 // If loop preparation eliminates all interesting IV users, bail.
5620 if (IU.empty()) return;
5621
5622 // Skip nested loops until we can model them better with formulae.
5623 if (!L->isInnermost()) {
5624 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)
;
5625 return;
5626 }
5627
5628 // Start collecting data and preparing for the solver.
5629 // If number of registers is not the major cost, we cannot benefit from the
5630 // current profitable chain optimization which is based on number of
5631 // registers.
5632 // FIXME: add profitable chain optimization for other kinds major cost, for
5633 // example number of instructions.
5634 if (TTI.isNumRegsMajorCostOfLSR() || StressIVChain)
5635 CollectChains();
5636 CollectInterestingTypesAndFactors();
5637 CollectFixupsAndInitialFormulae();
5638 CollectLoopInvariantFixupsAndFormulae();
5639
5640 if (Uses.empty())
5641 return;
5642
5643 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
)
5644 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
;
5645
5646 // Now use the reuse data to generate a bunch of interesting ways
5647 // to formulate the values needed for the uses.
5648 GenerateAllReuseFormulae();
5649
5650 FilterOutUndesirableDedicatedRegisters();
5651 NarrowSearchSpaceUsingHeuristics();
5652
5653 SmallVector<const Formula *, 8> Solution;
5654 Solve(Solution);
5655
5656 // Release memory that is no longer needed.
5657 Factors.clear();
5658 Types.clear();
5659 RegUses.clear();
5660
5661 if (Solution.empty())
5662 return;
5663
5664#ifndef NDEBUG
5665 // Formulae should be legal.
5666 for (const LSRUse &LU : Uses) {
5667 for (const Formula &F : LU.Formulae)
5668 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5669, __PRETTY_FUNCTION__))
5669 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-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5669, __PRETTY_FUNCTION__))
;
5670 };
5671#endif
5672
5673 // Now that we've decided what we want, make it so.
5674 ImplementSolution(Solution);
5675}
5676
5677#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5678void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5679 if (Factors.empty() && Types.empty()) return;
5680
5681 OS << "LSR has identified the following interesting factors and types: ";
5682 bool First = true;
5683
5684 for (int64_t Factor : Factors) {
5685 if (!First) OS << ", ";
5686 First = false;
5687 OS << '*' << Factor;
5688 }
5689
5690 for (Type *Ty : Types) {
5691 if (!First) OS << ", ";
5692 First = false;
5693 OS << '(' << *Ty << ')';
5694 }
5695 OS << '\n';
5696}
5697
5698void LSRInstance::print_fixups(raw_ostream &OS) const {
5699 OS << "LSR is examining the following fixup sites:\n";
5700 for (const LSRUse &LU : Uses)
5701 for (const LSRFixup &LF : LU.Fixups) {
5702 dbgs() << " ";
5703 LF.print(OS);
5704 OS << '\n';
5705 }
5706}
5707
5708void LSRInstance::print_uses(raw_ostream &OS) const {
5709 OS << "LSR is examining the following uses:\n";
5710 for (const LSRUse &LU : Uses) {
5711 dbgs() << " ";
5712 LU.print(OS);
5713 OS << '\n';
5714 for (const Formula &F : LU.Formulae) {
5715 OS << " ";
5716 F.print(OS);
5717 OS << '\n';
5718 }
5719 }
5720}
5721
5722void LSRInstance::print(raw_ostream &OS) const {
5723 print_factors_and_types(OS);
5724 print_fixups(OS);
5725 print_uses(OS);
5726}
5727
5728LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRInstance::dump() const {
5729 print(errs()); errs() << '\n';
5730}
5731#endif
5732
5733namespace {
5734
5735class LoopStrengthReduce : public LoopPass {
5736public:
5737 static char ID; // Pass ID, replacement for typeid
5738
5739 LoopStrengthReduce();
5740
5741private:
5742 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5743 void getAnalysisUsage(AnalysisUsage &AU) const override;
5744};
5745
5746} // end anonymous namespace
5747
5748LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5749 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5750}
5751
5752void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5753 // We split critical edges, so we change the CFG. However, we do update
5754 // many analyses if they are around.