Bug Summary

File:llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 5468, 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 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/lib/Transforms/Scalar -resource-dir /usr/lib/llvm-13/lib/clang/13.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/include -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/include -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-13/lib/clang/13.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/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-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c=. -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 -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-07-26-235520-9401-1 -x c++ /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/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/Analysis/ValueTracking.h"
81#include "llvm/Config/llvm-config.h"
82#include "llvm/IR/BasicBlock.h"
83#include "llvm/IR/Constant.h"
84#include "llvm/IR/Constants.h"
85#include "llvm/IR/DebugInfoMetadata.h"
86#include "llvm/IR/DerivedTypes.h"
87#include "llvm/IR/Dominators.h"
88#include "llvm/IR/GlobalValue.h"
89#include "llvm/IR/IRBuilder.h"
90#include "llvm/IR/InstrTypes.h"
91#include "llvm/IR/Instruction.h"
92#include "llvm/IR/Instructions.h"
93#include "llvm/IR/IntrinsicInst.h"
94#include "llvm/IR/Intrinsics.h"
95#include "llvm/IR/Module.h"
96#include "llvm/IR/OperandTraits.h"
97#include "llvm/IR/Operator.h"
98#include "llvm/IR/PassManager.h"
99#include "llvm/IR/Type.h"
100#include "llvm/IR/Use.h"
101#include "llvm/IR/User.h"
102#include "llvm/IR/Value.h"
103#include "llvm/IR/ValueHandle.h"
104#include "llvm/InitializePasses.h"
105#include "llvm/Pass.h"
106#include "llvm/Support/Casting.h"
107#include "llvm/Support/CommandLine.h"
108#include "llvm/Support/Compiler.h"
109#include "llvm/Support/Debug.h"
110#include "llvm/Support/ErrorHandling.h"
111#include "llvm/Support/MathExtras.h"
112#include "llvm/Support/raw_ostream.h"
113#include "llvm/Transforms/Scalar.h"
114#include "llvm/Transforms/Utils.h"
115#include "llvm/Transforms/Utils/BasicBlockUtils.h"
116#include "llvm/Transforms/Utils/Local.h"
117#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
118#include <algorithm>
119#include <cassert>
120#include <cstddef>
121#include <cstdint>
122#include <cstdlib>
123#include <iterator>
124#include <limits>
125#include <map>
126#include <numeric>
127#include <utility>
128
129using namespace llvm;
130
131#define DEBUG_TYPE"loop-reduce" "loop-reduce"
132
133/// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
134/// bail out. This threshold is far beyond the number of users that LSR can
135/// conceivably solve, so it should not affect generated code, but catches the
136/// worst cases before LSR burns too much compile time and stack space.
137static const unsigned MaxIVUsers = 200;
138
139// Temporary flag to cleanup congruent phis after LSR phi expansion.
140// It's currently disabled until we can determine whether it's truly useful or
141// not. The flag should be removed after the v3.0 release.
142// This is now needed for ivchains.
143static cl::opt<bool> EnablePhiElim(
144 "enable-lsr-phielim", cl::Hidden, cl::init(true),
145 cl::desc("Enable LSR phi elimination"));
146
147// The flag adds instruction count to solutions cost comparision.
148static cl::opt<bool> InsnsCost(
149 "lsr-insns-cost", cl::Hidden, cl::init(true),
150 cl::desc("Add instruction count to a LSR cost model"));
151
152// Flag to choose how to narrow complex lsr solution
153static cl::opt<bool> LSRExpNarrow(
154 "lsr-exp-narrow", cl::Hidden, cl::init(false),
155 cl::desc("Narrow LSR complex solution using"
156 " expectation of registers number"));
157
158// Flag to narrow search space by filtering non-optimal formulae with
159// the same ScaledReg and Scale.
160static cl::opt<bool> FilterSameScaledReg(
161 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
162 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
163 " with the same ScaledReg and Scale"));
164
165static cl::opt<TTI::AddressingModeKind> PreferredAddresingMode(
166 "lsr-preferred-addressing-mode", cl::Hidden, cl::init(TTI::AMK_None),
167 cl::desc("A flag that overrides the target's preferred addressing mode."),
168 cl::values(clEnumValN(TTI::AMK_None,llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
}
169 "none",llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
}
170 "Don't prefer any addressing mode")llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
}
,
171 clEnumValN(TTI::AMK_PreIndexed,llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
172 "preindexed",llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
173 "Prefer pre-indexed addressing mode")llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
,
174 clEnumValN(TTI::AMK_PostIndexed,llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
175 "postindexed",llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
176 "Prefer post-indexed addressing mode")llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
));
177
178static cl::opt<unsigned> ComplexityLimit(
179 "lsr-complexity-limit", cl::Hidden,
180 cl::init(std::numeric_limits<uint16_t>::max()),
181 cl::desc("LSR search space complexity limit"));
182
183static cl::opt<unsigned> SetupCostDepthLimit(
184 "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
185 cl::desc("The limit on recursion depth for LSRs setup cost"));
186
187#ifndef NDEBUG
188// Stress test IV chain generation.
189static cl::opt<bool> StressIVChain(
190 "stress-ivchain", cl::Hidden, cl::init(false),
191 cl::desc("Stress test LSR IV chains"));
192#else
193static bool StressIVChain = false;
194#endif
195
196namespace {
197
198struct MemAccessTy {
199 /// Used in situations where the accessed memory type is unknown.
200 static const unsigned UnknownAddressSpace =
201 std::numeric_limits<unsigned>::max();
202
203 Type *MemTy = nullptr;
204 unsigned AddrSpace = UnknownAddressSpace;
205
206 MemAccessTy() = default;
207 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
208
209 bool operator==(MemAccessTy Other) const {
210 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
211 }
212
213 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
214
215 static MemAccessTy getUnknown(LLVMContext &Ctx,
216 unsigned AS = UnknownAddressSpace) {
217 return MemAccessTy(Type::getVoidTy(Ctx), AS);
218 }
219
220 Type *getType() { return MemTy; }
221};
222
223/// This class holds data which is used to order reuse candidates.
224class RegSortData {
225public:
226 /// This represents the set of LSRUse indices which reference
227 /// a particular register.
228 SmallBitVector UsedByIndices;
229
230 void print(raw_ostream &OS) const;
231 void dump() const;
232};
233
234} // end anonymous namespace
235
236#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
237void RegSortData::print(raw_ostream &OS) const {
238 OS << "[NumUses=" << UsedByIndices.count() << ']';
239}
240
241LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
242 print(errs()); errs() << '\n';
243}
244#endif
245
246namespace {
247
248/// Map register candidates to information about how they are used.
249class RegUseTracker {
250 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
251
252 RegUsesTy RegUsesMap;
253 SmallVector<const SCEV *, 16> RegSequence;
254
255public:
256 void countRegister(const SCEV *Reg, size_t LUIdx);
257 void dropRegister(const SCEV *Reg, size_t LUIdx);
258 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
259
260 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
261
262 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
263
264 void clear();
265
266 using iterator = SmallVectorImpl<const SCEV *>::iterator;
267 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
268
269 iterator begin() { return RegSequence.begin(); }
270 iterator end() { return RegSequence.end(); }
271 const_iterator begin() const { return RegSequence.begin(); }
272 const_iterator end() const { return RegSequence.end(); }
273};
274
275} // end anonymous namespace
276
277void
278RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
279 std::pair<RegUsesTy::iterator, bool> Pair =
280 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
281 RegSortData &RSD = Pair.first->second;
282 if (Pair.second)
283 RegSequence.push_back(Reg);
284 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
285 RSD.UsedByIndices.set(LUIdx);
286}
287
288void
289RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
290 RegUsesTy::iterator It = RegUsesMap.find(Reg);
291 assert(It != RegUsesMap.end())(static_cast <bool> (It != RegUsesMap.end()) ? void (0)
: __assert_fail ("It != RegUsesMap.end()", "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 291, __extension__ __PRETTY_FUNCTION__))
;
292 RegSortData &RSD = It->second;
293 assert(RSD.UsedByIndices.size() > LUIdx)(static_cast <bool> (RSD.UsedByIndices.size() > LUIdx
) ? void (0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx"
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 293, __extension__ __PRETTY_FUNCTION__))
;
294 RSD.UsedByIndices.reset(LUIdx);
295}
296
297void
298RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
299 assert(LUIdx <= LastLUIdx)(static_cast <bool> (LUIdx <= LastLUIdx) ? void (0) :
__assert_fail ("LUIdx <= LastLUIdx", "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 299, __extension__ __PRETTY_FUNCTION__))
;
300
301 // Update RegUses. The data structure is not optimized for this purpose;
302 // we must iterate through it and update each of the bit vectors.
303 for (auto &Pair : RegUsesMap) {
304 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
305 if (LUIdx < UsedByIndices.size())
306 UsedByIndices[LUIdx] =
307 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
308 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
309 }
310}
311
312bool
313RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
314 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
315 if (I == RegUsesMap.end())
316 return false;
317 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
318 int i = UsedByIndices.find_first();
319 if (i == -1) return false;
320 if ((size_t)i != LUIdx) return true;
321 return UsedByIndices.find_next(i) != -1;
322}
323
324const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
325 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
326 assert(I != RegUsesMap.end() && "Unknown register!")(static_cast <bool> (I != RegUsesMap.end() && "Unknown register!"
) ? void (0) : __assert_fail ("I != RegUsesMap.end() && \"Unknown register!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 326, __extension__ __PRETTY_FUNCTION__))
;
327 return I->second.UsedByIndices;
328}
329
330void RegUseTracker::clear() {
331 RegUsesMap.clear();
332 RegSequence.clear();
333}
334
335namespace {
336
337/// This class holds information that describes a formula for computing
338/// satisfying a use. It may include broken-out immediates and scaled registers.
339struct Formula {
340 /// Global base address used for complex addressing.
341 GlobalValue *BaseGV = nullptr;
342
343 /// Base offset for complex addressing.
344 int64_t BaseOffset = 0;
345
346 /// Whether any complex addressing has a base register.
347 bool HasBaseReg = false;
348
349 /// The scale of any complex addressing.
350 int64_t Scale = 0;
351
352 /// The list of "base" registers for this use. When this is non-empty. The
353 /// canonical representation of a formula is
354 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
355 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
356 /// 3. The reg containing recurrent expr related with currect loop in the
357 /// formula should be put in the ScaledReg.
358 /// #1 enforces that the scaled register is always used when at least two
359 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
360 /// #2 enforces that 1 * reg is reg.
361 /// #3 ensures invariant regs with respect to current loop can be combined
362 /// together in LSR codegen.
363 /// This invariant can be temporarily broken while building a formula.
364 /// However, every formula inserted into the LSRInstance must be in canonical
365 /// form.
366 SmallVector<const SCEV *, 4> BaseRegs;
367
368 /// The 'scaled' register for this use. This should be non-null when Scale is
369 /// not zero.
370 const SCEV *ScaledReg = nullptr;
371
372 /// An additional constant offset which added near the use. This requires a
373 /// temporary register, but the offset itself can live in an add immediate
374 /// field rather than a register.
375 int64_t UnfoldedOffset = 0;
376
377 Formula() = default;
378
379 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
380
381 bool isCanonical(const Loop &L) const;
382
383 void canonicalize(const Loop &L);
384
385 bool unscale();
386
387 bool hasZeroEnd() const;
388
389 size_t getNumRegs() const;
390 Type *getType() const;
391
392 void deleteBaseReg(const SCEV *&S);
393
394 bool referencesReg(const SCEV *S) const;
395 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
396 const RegUseTracker &RegUses) const;
397
398 void print(raw_ostream &OS) const;
399 void dump() const;
400};
401
402} // end anonymous namespace
403
404/// Recursion helper for initialMatch.
405static void DoInitialMatch(const SCEV *S, Loop *L,
406 SmallVectorImpl<const SCEV *> &Good,
407 SmallVectorImpl<const SCEV *> &Bad,
408 ScalarEvolution &SE) {
409 // Collect expressions which properly dominate the loop header.
410 if (SE.properlyDominates(S, L->getHeader())) {
411 Good.push_back(S);
412 return;
413 }
414
415 // Look at add operands.
416 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
417 for (const SCEV *S : Add->operands())
418 DoInitialMatch(S, L, Good, Bad, SE);
419 return;
420 }
421
422 // Look at addrec operands.
423 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
424 if (!AR->getStart()->isZero() && AR->isAffine()) {
425 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
426 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
427 AR->getStepRecurrence(SE),
428 // FIXME: AR->getNoWrapFlags()
429 AR->getLoop(), SCEV::FlagAnyWrap),
430 L, Good, Bad, SE);
431 return;
432 }
433
434 // Handle a multiplication by -1 (negation) if it didn't fold.
435 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
436 if (Mul->getOperand(0)->isAllOnesValue()) {
437 SmallVector<const SCEV *, 4> Ops(drop_begin(Mul->operands()));
438 const SCEV *NewMul = SE.getMulExpr(Ops);
439
440 SmallVector<const SCEV *, 4> MyGood;
441 SmallVector<const SCEV *, 4> MyBad;
442 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
443 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
444 SE.getEffectiveSCEVType(NewMul->getType())));
445 for (const SCEV *S : MyGood)
446 Good.push_back(SE.getMulExpr(NegOne, S));
447 for (const SCEV *S : MyBad)
448 Bad.push_back(SE.getMulExpr(NegOne, S));
449 return;
450 }
451
452 // Ok, we can't do anything interesting. Just stuff the whole thing into a
453 // register and hope for the best.
454 Bad.push_back(S);
455}
456
457/// Incorporate loop-variant parts of S into this Formula, attempting to keep
458/// all loop-invariant and loop-computable values in a single base register.
459void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
460 SmallVector<const SCEV *, 4> Good;
461 SmallVector<const SCEV *, 4> Bad;
462 DoInitialMatch(S, L, Good, Bad, SE);
463 if (!Good.empty()) {
464 const SCEV *Sum = SE.getAddExpr(Good);
465 if (!Sum->isZero())
466 BaseRegs.push_back(Sum);
467 HasBaseReg = true;
468 }
469 if (!Bad.empty()) {
470 const SCEV *Sum = SE.getAddExpr(Bad);
471 if (!Sum->isZero())
472 BaseRegs.push_back(Sum);
473 HasBaseReg = true;
474 }
475 canonicalize(*L);
476}
477
478/// Check whether or not this formula satisfies the canonical
479/// representation.
480/// \see Formula::BaseRegs.
481bool Formula::isCanonical(const Loop &L) const {
482 if (!ScaledReg)
483 return BaseRegs.size() <= 1;
484
485 if (Scale != 1)
486 return true;
487
488 if (Scale == 1 && BaseRegs.empty())
489 return false;
490
491 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
492 if (SAR && SAR->getLoop() == &L)
493 return true;
494
495 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
496 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
497 // loop, we want to swap the reg in BaseRegs with ScaledReg.
498 auto I = find_if(BaseRegs, [&](const SCEV *S) {
499 return isa<const SCEVAddRecExpr>(S) &&
500 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
501 });
502 return I == BaseRegs.end();
503}
504
505/// Helper method to morph a formula into its canonical representation.
506/// \see Formula::BaseRegs.
507/// Every formula having more than one base register, must use the ScaledReg
508/// field. Otherwise, we would have to do special cases everywhere in LSR
509/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
510/// On the other hand, 1*reg should be canonicalized into reg.
511void Formula::canonicalize(const Loop &L) {
512 if (isCanonical(L))
513 return;
514
515 if (BaseRegs.empty()) {
516 // No base reg? Use scale reg with scale = 1 as such.
517 assert(ScaledReg && "Expected 1*reg => reg")(static_cast <bool> (ScaledReg && "Expected 1*reg => reg"
) ? void (0) : __assert_fail ("ScaledReg && \"Expected 1*reg => reg\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 517, __extension__ __PRETTY_FUNCTION__))
;
518 assert(Scale == 1 && "Expected 1*reg => reg")(static_cast <bool> (Scale == 1 && "Expected 1*reg => reg"
) ? void (0) : __assert_fail ("Scale == 1 && \"Expected 1*reg => reg\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 518, __extension__ __PRETTY_FUNCTION__))
;
519 BaseRegs.push_back(ScaledReg);
520 Scale = 0;
521 ScaledReg = nullptr;
522 return;
523 }
524
525 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
526 if (!ScaledReg) {
527 ScaledReg = BaseRegs.pop_back_val();
528 Scale = 1;
529 }
530
531 // If ScaledReg is an invariant with respect to L, find the reg from
532 // BaseRegs containing the recurrent expr related with Loop L. Swap the
533 // reg with ScaledReg.
534 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
535 if (!SAR || SAR->getLoop() != &L) {
536 auto I = find_if(BaseRegs, [&](const SCEV *S) {
537 return isa<const SCEVAddRecExpr>(S) &&
538 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
539 });
540 if (I != BaseRegs.end())
541 std::swap(ScaledReg, *I);
542 }
543 assert(isCanonical(L) && "Failed to canonicalize?")(static_cast <bool> (isCanonical(L) && "Failed to canonicalize?"
) ? void (0) : __assert_fail ("isCanonical(L) && \"Failed to canonicalize?\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 543, __extension__ __PRETTY_FUNCTION__))
;
544}
545
546/// Get rid of the scale in the formula.
547/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
548/// \return true if it was possible to get rid of the scale, false otherwise.
549/// \note After this operation the formula may not be in the canonical form.
550bool Formula::unscale() {
551 if (Scale != 1)
552 return false;
553 Scale = 0;
554 BaseRegs.push_back(ScaledReg);
555 ScaledReg = nullptr;
556 return true;
557}
558
559bool Formula::hasZeroEnd() const {
560 if (UnfoldedOffset || BaseOffset)
561 return false;
562 if (BaseRegs.size() != 1 || ScaledReg)
563 return false;
564 return true;
565}
566
567/// Return the total number of register operands used by this formula. This does
568/// not include register uses implied by non-constant addrec strides.
569size_t Formula::getNumRegs() const {
570 return !!ScaledReg + BaseRegs.size();
571}
572
573/// Return the type of this formula, if it has one, or null otherwise. This type
574/// is meaningless except for the bit size.
575Type *Formula::getType() const {
576 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
577 ScaledReg ? ScaledReg->getType() :
578 BaseGV ? BaseGV->getType() :
579 nullptr;
580}
581
582/// Delete the given base reg from the BaseRegs list.
583void Formula::deleteBaseReg(const SCEV *&S) {
584 if (&S != &BaseRegs.back())
585 std::swap(S, BaseRegs.back());
586 BaseRegs.pop_back();
587}
588
589/// Test if this formula references the given register.
590bool Formula::referencesReg(const SCEV *S) const {
591 return S == ScaledReg || is_contained(BaseRegs, S);
592}
593
594/// Test whether this formula uses registers which are used by uses other than
595/// the use with the given index.
596bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
597 const RegUseTracker &RegUses) const {
598 if (ScaledReg)
599 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
600 return true;
601 for (const SCEV *BaseReg : BaseRegs)
602 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
603 return true;
604 return false;
605}
606
607#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
608void Formula::print(raw_ostream &OS) const {
609 bool First = true;
610 if (BaseGV) {
611 if (!First) OS << " + "; else First = false;
612 BaseGV->printAsOperand(OS, /*PrintType=*/false);
613 }
614 if (BaseOffset != 0) {
615 if (!First) OS << " + "; else First = false;
616 OS << BaseOffset;
617 }
618 for (const SCEV *BaseReg : BaseRegs) {
619 if (!First) OS << " + "; else First = false;
620 OS << "reg(" << *BaseReg << ')';
621 }
622 if (HasBaseReg && BaseRegs.empty()) {
623 if (!First) OS << " + "; else First = false;
624 OS << "**error: HasBaseReg**";
625 } else if (!HasBaseReg && !BaseRegs.empty()) {
626 if (!First) OS << " + "; else First = false;
627 OS << "**error: !HasBaseReg**";
628 }
629 if (Scale != 0) {
630 if (!First) OS << " + "; else First = false;
631 OS << Scale << "*reg(";
632 if (ScaledReg)
633 OS << *ScaledReg;
634 else
635 OS << "<unknown>";
636 OS << ')';
637 }
638 if (UnfoldedOffset != 0) {
639 if (!First) OS << " + ";
640 OS << "imm(" << UnfoldedOffset << ')';
641 }
642}
643
644LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
645 print(errs()); errs() << '\n';
646}
647#endif
648
649/// Return true if the given addrec can be sign-extended without changing its
650/// value.
651static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
652 Type *WideTy =
653 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
654 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
655}
656
657/// Return true if the given add can be sign-extended without changing its
658/// value.
659static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
660 Type *WideTy =
661 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
662 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
663}
664
665/// Return true if the given mul can be sign-extended without changing its
666/// value.
667static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
668 Type *WideTy =
669 IntegerType::get(SE.getContext(),
670 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
671 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
672}
673
674/// Return an expression for LHS /s RHS, if it can be determined and if the
675/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
676/// is true, expressions like (X * Y) /s Y are simplified to X, ignoring that
677/// the multiplication may overflow, which is useful when the result will be
678/// used in a context where the most significant bits are ignored.
679static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
680 ScalarEvolution &SE,
681 bool IgnoreSignificantBits = false) {
682 // Handle the trivial case, which works for any SCEV type.
683 if (LHS == RHS)
684 return SE.getConstant(LHS->getType(), 1);
685
686 // Handle a few RHS special cases.
687 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
688 if (RC) {
689 const APInt &RA = RC->getAPInt();
690 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
691 // some folding.
692 if (RA.isAllOnesValue()) {
693 if (LHS->getType()->isPointerTy())
694 return nullptr;
695 return SE.getMulExpr(LHS, RC);
696 }
697 // Handle x /s 1 as x.
698 if (RA == 1)
699 return LHS;
700 }
701
702 // Check for a division of a constant by a constant.
703 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
704 if (!RC)
705 return nullptr;
706 const APInt &LA = C->getAPInt();
707 const APInt &RA = RC->getAPInt();
708 if (LA.srem(RA) != 0)
709 return nullptr;
710 return SE.getConstant(LA.sdiv(RA));
711 }
712
713 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
714 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
715 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
716 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
717 IgnoreSignificantBits);
718 if (!Step) return nullptr;
719 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
720 IgnoreSignificantBits);
721 if (!Start) return nullptr;
722 // FlagNW is independent of the start value, step direction, and is
723 // preserved with smaller magnitude steps.
724 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
725 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
726 }
727 return nullptr;
728 }
729
730 // Distribute the sdiv over add operands, if the add doesn't overflow.
731 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
732 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
733 SmallVector<const SCEV *, 8> Ops;
734 for (const SCEV *S : Add->operands()) {
735 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
736 if (!Op) return nullptr;
737 Ops.push_back(Op);
738 }
739 return SE.getAddExpr(Ops);
740 }
741 return nullptr;
742 }
743
744 // Check for a multiply operand that we can pull RHS out of.
745 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
746 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
747 // Handle special case C1*X*Y /s C2*X*Y.
748 if (const SCEVMulExpr *MulRHS = dyn_cast<SCEVMulExpr>(RHS)) {
749 if (IgnoreSignificantBits || isMulSExtable(MulRHS, SE)) {
750 const SCEVConstant *LC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
751 const SCEVConstant *RC =
752 dyn_cast<SCEVConstant>(MulRHS->getOperand(0));
753 if (LC && RC) {
754 SmallVector<const SCEV *, 4> LOps(drop_begin(Mul->operands()));
755 SmallVector<const SCEV *, 4> ROps(drop_begin(MulRHS->operands()));
756 if (LOps == ROps)
757 return getExactSDiv(LC, RC, SE, IgnoreSignificantBits);
758 }
759 }
760 }
761
762 SmallVector<const SCEV *, 4> Ops;
763 bool Found = false;
764 for (const SCEV *S : Mul->operands()) {
765 if (!Found)
766 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
767 IgnoreSignificantBits)) {
768 S = Q;
769 Found = true;
770 }
771 Ops.push_back(S);
772 }
773 return Found ? SE.getMulExpr(Ops) : nullptr;
774 }
775 return nullptr;
776 }
777
778 // Otherwise we don't know.
779 return nullptr;
780}
781
782/// If S involves the addition of a constant integer value, return that integer
783/// value, and mutate S to point to a new SCEV with that value excluded.
784static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
785 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
786 if (C->getAPInt().getMinSignedBits() <= 64) {
787 S = SE.getConstant(C->getType(), 0);
788 return C->getValue()->getSExtValue();
789 }
790 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
791 SmallVector<const SCEV *, 8> NewOps(Add->operands());
792 int64_t Result = ExtractImmediate(NewOps.front(), SE);
793 if (Result != 0)
794 S = SE.getAddExpr(NewOps);
795 return Result;
796 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
797 SmallVector<const SCEV *, 8> NewOps(AR->operands());
798 int64_t Result = ExtractImmediate(NewOps.front(), SE);
799 if (Result != 0)
800 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
801 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
802 SCEV::FlagAnyWrap);
803 return Result;
804 }
805 return 0;
806}
807
808/// If S involves the addition of a GlobalValue address, return that symbol, and
809/// mutate S to point to a new SCEV with that value excluded.
810static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
811 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
812 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
813 S = SE.getConstant(GV->getType(), 0);
814 return GV;
815 }
816 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
817 SmallVector<const SCEV *, 8> NewOps(Add->operands());
818 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
819 if (Result)
820 S = SE.getAddExpr(NewOps);
821 return Result;
822 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
823 SmallVector<const SCEV *, 8> NewOps(AR->operands());
824 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
825 if (Result)
826 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
827 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
828 SCEV::FlagAnyWrap);
829 return Result;
830 }
831 return nullptr;
832}
833
834/// Returns true if the specified instruction is using the specified value as an
835/// address.
836static bool isAddressUse(const TargetTransformInfo &TTI,
837 Instruction *Inst, Value *OperandVal) {
838 bool isAddress = isa<LoadInst>(Inst);
839 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
840 if (SI->getPointerOperand() == OperandVal)
841 isAddress = true;
842 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
843 // Addressing modes can also be folded into prefetches and a variety
844 // of intrinsics.
845 switch (II->getIntrinsicID()) {
846 case Intrinsic::memset:
847 case Intrinsic::prefetch:
848 case Intrinsic::masked_load:
849 if (II->getArgOperand(0) == OperandVal)
850 isAddress = true;
851 break;
852 case Intrinsic::masked_store:
853 if (II->getArgOperand(1) == OperandVal)
854 isAddress = true;
855 break;
856 case Intrinsic::memmove:
857 case Intrinsic::memcpy:
858 if (II->getArgOperand(0) == OperandVal ||
859 II->getArgOperand(1) == OperandVal)
860 isAddress = true;
861 break;
862 default: {
863 MemIntrinsicInfo IntrInfo;
864 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
865 if (IntrInfo.PtrVal == OperandVal)
866 isAddress = true;
867 }
868 }
869 }
870 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
871 if (RMW->getPointerOperand() == OperandVal)
872 isAddress = true;
873 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
874 if (CmpX->getPointerOperand() == OperandVal)
875 isAddress = true;
876 }
877 return isAddress;
878}
879
880/// Return the type of the memory being accessed.
881static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
882 Instruction *Inst, Value *OperandVal) {
883 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
884 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
885 AccessTy.MemTy = SI->getOperand(0)->getType();
886 AccessTy.AddrSpace = SI->getPointerAddressSpace();
887 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
888 AccessTy.AddrSpace = LI->getPointerAddressSpace();
889 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
890 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
891 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
892 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
893 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
894 switch (II->getIntrinsicID()) {
895 case Intrinsic::prefetch:
896 case Intrinsic::memset:
897 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
898 AccessTy.MemTy = OperandVal->getType();
899 break;
900 case Intrinsic::memmove:
901 case Intrinsic::memcpy:
902 AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
903 AccessTy.MemTy = OperandVal->getType();
904 break;
905 case Intrinsic::masked_load:
906 AccessTy.AddrSpace =
907 II->getArgOperand(0)->getType()->getPointerAddressSpace();
908 break;
909 case Intrinsic::masked_store:
910 AccessTy.MemTy = II->getOperand(0)->getType();
911 AccessTy.AddrSpace =
912 II->getArgOperand(1)->getType()->getPointerAddressSpace();
913 break;
914 default: {
915 MemIntrinsicInfo IntrInfo;
916 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
917 AccessTy.AddrSpace
918 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
919 }
920
921 break;
922 }
923 }
924 }
925
926 // All pointers have the same requirements, so canonicalize them to an
927 // arbitrary pointer type to minimize variation.
928 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
929 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
930 PTy->getAddressSpace());
931
932 return AccessTy;
933}
934
935/// Return true if this AddRec is already a phi in its loop.
936static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
937 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
938 if (SE.isSCEVable(PN.getType()) &&
939 (SE.getEffectiveSCEVType(PN.getType()) ==
940 SE.getEffectiveSCEVType(AR->getType())) &&
941 SE.getSCEV(&PN) == AR)
942 return true;
943 }
944 return false;
945}
946
947/// Check if expanding this expression is likely to incur significant cost. This
948/// is tricky because SCEV doesn't track which expressions are actually computed
949/// by the current IR.
950///
951/// We currently allow expansion of IV increments that involve adds,
952/// multiplication by constants, and AddRecs from existing phis.
953///
954/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
955/// obvious multiple of the UDivExpr.
956static bool isHighCostExpansion(const SCEV *S,
957 SmallPtrSetImpl<const SCEV*> &Processed,
958 ScalarEvolution &SE) {
959 // Zero/One operand expressions
960 switch (S->getSCEVType()) {
961 case scUnknown:
962 case scConstant:
963 return false;
964 case scTruncate:
965 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
966 Processed, SE);
967 case scZeroExtend:
968 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
969 Processed, SE);
970 case scSignExtend:
971 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
972 Processed, SE);
973 default:
974 break;
975 }
976
977 if (!Processed.insert(S).second)
978 return false;
979
980 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
981 for (const SCEV *S : Add->operands()) {
982 if (isHighCostExpansion(S, Processed, SE))
983 return true;
984 }
985 return false;
986 }
987
988 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
989 if (Mul->getNumOperands() == 2) {
990 // Multiplication by a constant is ok
991 if (isa<SCEVConstant>(Mul->getOperand(0)))
992 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
993
994 // If we have the value of one operand, check if an existing
995 // multiplication already generates this expression.
996 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
997 Value *UVal = U->getValue();
998 for (User *UR : UVal->users()) {
999 // If U is a constant, it may be used by a ConstantExpr.
1000 Instruction *UI = dyn_cast<Instruction>(UR);
1001 if (UI && UI->getOpcode() == Instruction::Mul &&
1002 SE.isSCEVable(UI->getType())) {
1003 return SE.getSCEV(UI) == Mul;
1004 }
1005 }
1006 }
1007 }
1008 }
1009
1010 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1011 if (isExistingPhi(AR, SE))
1012 return false;
1013 }
1014
1015 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
1016 return true;
1017}
1018
1019namespace {
1020
1021class LSRUse;
1022
1023} // end anonymous namespace
1024
1025/// Check if the addressing mode defined by \p F is completely
1026/// folded in \p LU at isel time.
1027/// This includes address-mode folding and special icmp tricks.
1028/// This function returns true if \p LU can accommodate what \p F
1029/// defines and up to 1 base + 1 scaled + offset.
1030/// In other words, if \p F has several base registers, this function may
1031/// still return true. Therefore, users still need to account for
1032/// additional base registers and/or unfolded offsets to derive an
1033/// accurate cost model.
1034static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1035 const LSRUse &LU, const Formula &F);
1036
1037// Get the cost of the scaling factor used in F for LU.
1038static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1039 const LSRUse &LU, const Formula &F,
1040 const Loop &L);
1041
1042namespace {
1043
1044/// This class is used to measure and compare candidate formulae.
1045class Cost {
1046 const Loop *L = nullptr;
1047 ScalarEvolution *SE = nullptr;
1048 const TargetTransformInfo *TTI = nullptr;
1049 TargetTransformInfo::LSRCost C;
1050 TTI::AddressingModeKind AMK = TTI::AMK_None;
1051
1052public:
1053 Cost() = delete;
1054 Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
1055 TTI::AddressingModeKind AMK) :
1056 L(L), SE(&SE), TTI(&TTI), AMK(AMK) {
1057 C.Insns = 0;
1058 C.NumRegs = 0;
1059 C.AddRecCost = 0;
1060 C.NumIVMuls = 0;
1061 C.NumBaseAdds = 0;
1062 C.ImmCost = 0;
1063 C.SetupCost = 0;
1064 C.ScaleCost = 0;
1065 }
1066
1067 bool isLess(Cost &Other);
1068
1069 void Lose();
1070
1071#ifndef NDEBUG
1072 // Once any of the metrics loses, they must all remain losers.
1073 bool isValid() {
1074 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1075 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1076 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1077 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1078 }
1079#endif
1080
1081 bool isLoser() {
1082 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1082, __extension__ __PRETTY_FUNCTION__))
;
1083 return C.NumRegs == ~0u;
1084 }
1085
1086 void RateFormula(const Formula &F,
1087 SmallPtrSetImpl<const SCEV *> &Regs,
1088 const DenseSet<const SCEV *> &VisitedRegs,
1089 const LSRUse &LU,
1090 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1091
1092 void print(raw_ostream &OS) const;
1093 void dump() const;
1094
1095private:
1096 void RateRegister(const Formula &F, const SCEV *Reg,
1097 SmallPtrSetImpl<const SCEV *> &Regs);
1098 void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1099 SmallPtrSetImpl<const SCEV *> &Regs,
1100 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1101};
1102
1103/// An operand value in an instruction which is to be replaced with some
1104/// equivalent, possibly strength-reduced, replacement.
1105struct LSRFixup {
1106 /// The instruction which will be updated.
1107 Instruction *UserInst = nullptr;
1108
1109 /// The operand of the instruction which will be replaced. The operand may be
1110 /// used more than once; every instance will be replaced.
1111 Value *OperandValToReplace = nullptr;
1112
1113 /// If this user is to use the post-incremented value of an induction
1114 /// variable, this set is non-empty and holds the loops associated with the
1115 /// induction variable.
1116 PostIncLoopSet PostIncLoops;
1117
1118 /// A constant offset to be added to the LSRUse expression. This allows
1119 /// multiple fixups to share the same LSRUse with different offsets, for
1120 /// example in an unrolled loop.
1121 int64_t Offset = 0;
1122
1123 LSRFixup() = default;
1124
1125 bool isUseFullyOutsideLoop(const Loop *L) const;
1126
1127 void print(raw_ostream &OS) const;
1128 void dump() const;
1129};
1130
1131/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1132/// SmallVectors of const SCEV*.
1133struct UniquifierDenseMapInfo {
1134 static SmallVector<const SCEV *, 4> getEmptyKey() {
1135 SmallVector<const SCEV *, 4> V;
1136 V.push_back(reinterpret_cast<const SCEV *>(-1));
1137 return V;
1138 }
1139
1140 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1141 SmallVector<const SCEV *, 4> V;
1142 V.push_back(reinterpret_cast<const SCEV *>(-2));
1143 return V;
1144 }
1145
1146 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1147 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1148 }
1149
1150 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1151 const SmallVector<const SCEV *, 4> &RHS) {
1152 return LHS == RHS;
1153 }
1154};
1155
1156/// This class holds the state that LSR keeps for each use in IVUsers, as well
1157/// as uses invented by LSR itself. It includes information about what kinds of
1158/// things can be folded into the user, information about the user itself, and
1159/// information about how the use may be satisfied. TODO: Represent multiple
1160/// users of the same expression in common?
1161class LSRUse {
1162 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1163
1164public:
1165 /// An enum for a kind of use, indicating what types of scaled and immediate
1166 /// operands it might support.
1167 enum KindType {
1168 Basic, ///< A normal use, with no folding.
1169 Special, ///< A special case of basic, allowing -1 scales.
1170 Address, ///< An address use; folding according to TargetLowering
1171 ICmpZero ///< An equality icmp with both operands folded into one.
1172 // TODO: Add a generic icmp too?
1173 };
1174
1175 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1176
1177 KindType Kind;
1178 MemAccessTy AccessTy;
1179
1180 /// The list of operands which are to be replaced.
1181 SmallVector<LSRFixup, 8> Fixups;
1182
1183 /// Keep track of the min and max offsets of the fixups.
1184 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1185 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1186
1187 /// This records whether all of the fixups using this LSRUse are outside of
1188 /// the loop, in which case some special-case heuristics may be used.
1189 bool AllFixupsOutsideLoop = true;
1190
1191 /// RigidFormula is set to true to guarantee that this use will be associated
1192 /// with a single formula--the one that initially matched. Some SCEV
1193 /// expressions cannot be expanded. This allows LSR to consider the registers
1194 /// used by those expressions without the need to expand them later after
1195 /// changing the formula.
1196 bool RigidFormula = false;
1197
1198 /// This records the widest use type for any fixup using this
1199 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1200 /// fixup widths to be equivalent, because the narrower one may be relying on
1201 /// the implicit truncation to truncate away bogus bits.
1202 Type *WidestFixupType = nullptr;
1203
1204 /// A list of ways to build a value that can satisfy this user. After the
1205 /// list is populated, one of these is selected heuristically and used to
1206 /// formulate a replacement for OperandValToReplace in UserInst.
1207 SmallVector<Formula, 12> Formulae;
1208
1209 /// The set of register candidates used by all formulae in this LSRUse.
1210 SmallPtrSet<const SCEV *, 4> Regs;
1211
1212 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1213
1214 LSRFixup &getNewFixup() {
1215 Fixups.push_back(LSRFixup());
1216 return Fixups.back();
1217 }
1218
1219 void pushFixup(LSRFixup &f) {
1220 Fixups.push_back(f);
1221 if (f.Offset > MaxOffset)
1222 MaxOffset = f.Offset;
1223 if (f.Offset < MinOffset)
1224 MinOffset = f.Offset;
1225 }
1226
1227 bool HasFormulaWithSameRegs(const Formula &F) const;
1228 float getNotSelectedProbability(const SCEV *Reg) const;
1229 bool InsertFormula(const Formula &F, const Loop &L);
1230 void DeleteFormula(Formula &F);
1231 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1232
1233 void print(raw_ostream &OS) const;
1234 void dump() const;
1235};
1236
1237} // end anonymous namespace
1238
1239static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1240 LSRUse::KindType Kind, MemAccessTy AccessTy,
1241 GlobalValue *BaseGV, int64_t BaseOffset,
1242 bool HasBaseReg, int64_t Scale,
1243 Instruction *Fixup = nullptr);
1244
1245static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1246 if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1247 return 1;
1248 if (Depth == 0)
1249 return 0;
1250 if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1251 return getSetupCost(S->getStart(), Depth - 1);
1252 if (auto S = dyn_cast<SCEVIntegralCastExpr>(Reg))
1253 return getSetupCost(S->getOperand(), Depth - 1);
1254 if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1255 return std::accumulate(S->op_begin(), S->op_end(), 0,
1256 [&](unsigned i, const SCEV *Reg) {
1257 return i + getSetupCost(Reg, Depth - 1);
1258 });
1259 if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1260 return getSetupCost(S->getLHS(), Depth - 1) +
1261 getSetupCost(S->getRHS(), Depth - 1);
1262 return 0;
1263}
1264
1265/// Tally up interesting quantities from the given register.
1266void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1267 SmallPtrSetImpl<const SCEV *> &Regs) {
1268 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1269 // If this is an addrec for another loop, it should be an invariant
1270 // with respect to L since L is the innermost loop (at least
1271 // for now LSR only handles innermost loops).
1272 if (AR->getLoop() != L) {
1273 // If the AddRec exists, consider it's register free and leave it alone.
1274 if (isExistingPhi(AR, *SE) && AMK != TTI::AMK_PostIndexed)
1275 return;
1276
1277 // It is bad to allow LSR for current loop to add induction variables
1278 // for its sibling loops.
1279 if (!AR->getLoop()->contains(L)) {
1280 Lose();
1281 return;
1282 }
1283
1284 // Otherwise, it will be an invariant with respect to Loop L.
1285 ++C.NumRegs;
1286 return;
1287 }
1288
1289 unsigned LoopCost = 1;
1290 if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1291 TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1292
1293 // If the step size matches the base offset, we could use pre-indexed
1294 // addressing.
1295 if (AMK == TTI::AMK_PreIndexed) {
1296 if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1297 if (Step->getAPInt() == F.BaseOffset)
1298 LoopCost = 0;
1299 } else if (AMK == TTI::AMK_PostIndexed) {
1300 const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1301 if (isa<SCEVConstant>(LoopStep)) {
1302 const SCEV *LoopStart = AR->getStart();
1303 if (!isa<SCEVConstant>(LoopStart) &&
1304 SE->isLoopInvariant(LoopStart, L))
1305 LoopCost = 0;
1306 }
1307 }
1308 }
1309 C.AddRecCost += LoopCost;
1310
1311 // Add the step value register, if it needs one.
1312 // TODO: The non-affine case isn't precisely modeled here.
1313 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1314 if (!Regs.count(AR->getOperand(1))) {
1315 RateRegister(F, AR->getOperand(1), Regs);
1316 if (isLoser())
1317 return;
1318 }
1319 }
1320 }
1321 ++C.NumRegs;
1322
1323 // Rough heuristic; favor registers which don't require extra setup
1324 // instructions in the preheader.
1325 C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1326 // Ensure we don't, even with the recusion limit, produce invalid costs.
1327 C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1328
1329 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1330 SE->hasComputableLoopEvolution(Reg, L);
1331}
1332
1333/// Record this register in the set. If we haven't seen it before, rate
1334/// it. Optional LoserRegs provides a way to declare any formula that refers to
1335/// one of those regs an instant loser.
1336void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1337 SmallPtrSetImpl<const SCEV *> &Regs,
1338 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1339 if (LoserRegs && LoserRegs->count(Reg)) {
1340 Lose();
1341 return;
1342 }
1343 if (Regs.insert(Reg).second) {
1344 RateRegister(F, Reg, Regs);
1345 if (LoserRegs && isLoser())
1346 LoserRegs->insert(Reg);
1347 }
1348}
1349
1350void Cost::RateFormula(const Formula &F,
1351 SmallPtrSetImpl<const SCEV *> &Regs,
1352 const DenseSet<const SCEV *> &VisitedRegs,
1353 const LSRUse &LU,
1354 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1355 assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula")(static_cast <bool> (F.isCanonical(*L) && "Cost is accurate only for canonical formula"
) ? void (0) : __assert_fail ("F.isCanonical(*L) && \"Cost is accurate only for canonical formula\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1355, __extension__ __PRETTY_FUNCTION__))
;
1356 // Tally up the registers.
1357 unsigned PrevAddRecCost = C.AddRecCost;
1358 unsigned PrevNumRegs = C.NumRegs;
1359 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1360 if (const SCEV *ScaledReg = F.ScaledReg) {
1361 if (VisitedRegs.count(ScaledReg)) {
1362 Lose();
1363 return;
1364 }
1365 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1366 if (isLoser())
1367 return;
1368 }
1369 for (const SCEV *BaseReg : F.BaseRegs) {
1370 if (VisitedRegs.count(BaseReg)) {
1371 Lose();
1372 return;
1373 }
1374 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1375 if (isLoser())
1376 return;
1377 }
1378
1379 // Determine how many (unfolded) adds we'll need inside the loop.
1380 size_t NumBaseParts = F.getNumRegs();
1381 if (NumBaseParts > 1)
1382 // Do not count the base and a possible second register if the target
1383 // allows to fold 2 registers.
1384 C.NumBaseAdds +=
1385 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1386 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1387
1388 // Accumulate non-free scaling amounts.
1389 C.ScaleCost += *getScalingFactorCost(*TTI, LU, F, *L).getValue();
1390
1391 // Tally up the non-zero immediates.
1392 for (const LSRFixup &Fixup : LU.Fixups) {
1393 int64_t O = Fixup.Offset;
1394 int64_t Offset = (uint64_t)O + F.BaseOffset;
1395 if (F.BaseGV)
1396 C.ImmCost += 64; // Handle symbolic values conservatively.
1397 // TODO: This should probably be the pointer size.
1398 else if (Offset != 0)
1399 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1400
1401 // Check with target if this offset with this instruction is
1402 // specifically not supported.
1403 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1404 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1405 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1406 C.NumBaseAdds++;
1407 }
1408
1409 // If we don't count instruction cost exit here.
1410 if (!InsnsCost) {
1411 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1411, __extension__ __PRETTY_FUNCTION__))
;
1412 return;
1413 }
1414
1415 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1416 // additional instruction (at least fill).
1417 // TODO: Need distinguish register class?
1418 unsigned TTIRegNum = TTI->getNumberOfRegisters(
1419 TTI->getRegisterClassForType(false, F.getType())) - 1;
1420 if (C.NumRegs > TTIRegNum) {
1421 // Cost already exceeded TTIRegNum, then only newly added register can add
1422 // new instructions.
1423 if (PrevNumRegs > TTIRegNum)
1424 C.Insns += (C.NumRegs - PrevNumRegs);
1425 else
1426 C.Insns += (C.NumRegs - TTIRegNum);
1427 }
1428
1429 // If ICmpZero formula ends with not 0, it could not be replaced by
1430 // just add or sub. We'll need to compare final result of AddRec.
1431 // That means we'll need an additional instruction. But if the target can
1432 // macro-fuse a compare with a branch, don't count this extra instruction.
1433 // For -10 + {0, +, 1}:
1434 // i = i + 1;
1435 // cmp i, 10
1436 //
1437 // For {-10, +, 1}:
1438 // i = i + 1;
1439 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1440 !TTI->canMacroFuseCmp())
1441 C.Insns++;
1442 // Each new AddRec adds 1 instruction to calculation.
1443 C.Insns += (C.AddRecCost - PrevAddRecCost);
1444
1445 // BaseAdds adds instructions for unfolded registers.
1446 if (LU.Kind != LSRUse::ICmpZero)
1447 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1448 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1448, __extension__ __PRETTY_FUNCTION__))
;
1449}
1450
1451/// Set this cost to a losing value.
1452void Cost::Lose() {
1453 C.Insns = std::numeric_limits<unsigned>::max();
1454 C.NumRegs = std::numeric_limits<unsigned>::max();
1455 C.AddRecCost = std::numeric_limits<unsigned>::max();
1456 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1457 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1458 C.ImmCost = std::numeric_limits<unsigned>::max();
1459 C.SetupCost = std::numeric_limits<unsigned>::max();
1460 C.ScaleCost = std::numeric_limits<unsigned>::max();
1461}
1462
1463/// Choose the lower cost.
1464bool Cost::isLess(Cost &Other) {
1465 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1466 C.Insns != Other.C.Insns)
1467 return C.Insns < Other.C.Insns;
1468 return TTI->isLSRCostLess(C, Other.C);
1469}
1470
1471#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1472void Cost::print(raw_ostream &OS) const {
1473 if (InsnsCost)
1474 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1475 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1476 if (C.AddRecCost != 0)
1477 OS << ", with addrec cost " << C.AddRecCost;
1478 if (C.NumIVMuls != 0)
1479 OS << ", plus " << C.NumIVMuls << " IV mul"
1480 << (C.NumIVMuls == 1 ? "" : "s");
1481 if (C.NumBaseAdds != 0)
1482 OS << ", plus " << C.NumBaseAdds << " base add"
1483 << (C.NumBaseAdds == 1 ? "" : "s");
1484 if (C.ScaleCost != 0)
1485 OS << ", plus " << C.ScaleCost << " scale cost";
1486 if (C.ImmCost != 0)
1487 OS << ", plus " << C.ImmCost << " imm cost";
1488 if (C.SetupCost != 0)
1489 OS << ", plus " << C.SetupCost << " setup cost";
1490}
1491
1492LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1493 print(errs()); errs() << '\n';
1494}
1495#endif
1496
1497/// Test whether this fixup always uses its value outside of the given loop.
1498bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1499 // PHI nodes use their value in their incoming blocks.
1500 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1501 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1502 if (PN->getIncomingValue(i) == OperandValToReplace &&
1503 L->contains(PN->getIncomingBlock(i)))
1504 return false;
1505 return true;
1506 }
1507
1508 return !L->contains(UserInst);
1509}
1510
1511#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1512void LSRFixup::print(raw_ostream &OS) const {
1513 OS << "UserInst=";
1514 // Store is common and interesting enough to be worth special-casing.
1515 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1516 OS << "store ";
1517 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1518 } else if (UserInst->getType()->isVoidTy())
1519 OS << UserInst->getOpcodeName();
1520 else
1521 UserInst->printAsOperand(OS, /*PrintType=*/false);
1522
1523 OS << ", OperandValToReplace=";
1524 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1525
1526 for (const Loop *PIL : PostIncLoops) {
1527 OS << ", PostIncLoop=";
1528 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1529 }
1530
1531 if (Offset != 0)
1532 OS << ", Offset=" << Offset;
1533}
1534
1535LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1536 print(errs()); errs() << '\n';
1537}
1538#endif
1539
1540/// Test whether this use as a formula which has the same registers as the given
1541/// formula.
1542bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1543 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1544 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1545 // Unstable sort by host order ok, because this is only used for uniquifying.
1546 llvm::sort(Key);
1547 return Uniquifier.count(Key);
1548}
1549
1550/// The function returns a probability of selecting formula without Reg.
1551float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1552 unsigned FNum = 0;
1553 for (const Formula &F : Formulae)
1554 if (F.referencesReg(Reg))
1555 FNum++;
1556 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1557}
1558
1559/// If the given formula has not yet been inserted, add it to the list, and
1560/// return true. Return false otherwise. The formula must be in canonical form.
1561bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1562 assert(F.isCanonical(L) && "Invalid canonical representation")(static_cast <bool> (F.isCanonical(L) && "Invalid canonical representation"
) ? void (0) : __assert_fail ("F.isCanonical(L) && \"Invalid canonical representation\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1562, __extension__ __PRETTY_FUNCTION__))
;
1563
1564 if (!Formulae.empty() && RigidFormula)
1565 return false;
1566
1567 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1568 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1569 // Unstable sort by host order ok, because this is only used for uniquifying.
1570 llvm::sort(Key);
1571
1572 if (!Uniquifier.insert(Key).second)
1573 return false;
1574
1575 // Using a register to hold the value of 0 is not profitable.
1576 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&(static_cast <bool> ((!F.ScaledReg || !F.ScaledReg->
isZero()) && "Zero allocated in a scaled register!") ?
void (0) : __assert_fail ("(!F.ScaledReg || !F.ScaledReg->isZero()) && \"Zero allocated in a scaled register!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1577, __extension__ __PRETTY_FUNCTION__))
1577 "Zero allocated in a scaled register!")(static_cast <bool> ((!F.ScaledReg || !F.ScaledReg->
isZero()) && "Zero allocated in a scaled register!") ?
void (0) : __assert_fail ("(!F.ScaledReg || !F.ScaledReg->isZero()) && \"Zero allocated in a scaled register!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1577, __extension__ __PRETTY_FUNCTION__))
;
1578#ifndef NDEBUG
1579 for (const SCEV *BaseReg : F.BaseRegs)
1580 assert(!BaseReg->isZero() && "Zero allocated in a base register!")(static_cast <bool> (!BaseReg->isZero() && "Zero allocated in a base register!"
) ? void (0) : __assert_fail ("!BaseReg->isZero() && \"Zero allocated in a base register!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1580, __extension__ __PRETTY_FUNCTION__))
;
1581#endif
1582
1583 // Add the formula to the list.
1584 Formulae.push_back(F);
1585
1586 // Record registers now being used by this use.
1587 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1588 if (F.ScaledReg)
1589 Regs.insert(F.ScaledReg);
1590
1591 return true;
1592}
1593
1594/// Remove the given formula from this use's list.
1595void LSRUse::DeleteFormula(Formula &F) {
1596 if (&F != &Formulae.back())
1597 std::swap(F, Formulae.back());
1598 Formulae.pop_back();
1599}
1600
1601/// Recompute the Regs field, and update RegUses.
1602void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1603 // Now that we've filtered out some formulae, recompute the Regs set.
1604 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1605 Regs.clear();
1606 for (const Formula &F : Formulae) {
1607 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1608 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1609 }
1610
1611 // Update the RegTracker.
1612 for (const SCEV *S : OldRegs)
1613 if (!Regs.count(S))
1614 RegUses.dropRegister(S, LUIdx);
1615}
1616
1617#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1618void LSRUse::print(raw_ostream &OS) const {
1619 OS << "LSR Use: Kind=";
1620 switch (Kind) {
1621 case Basic: OS << "Basic"; break;
1622 case Special: OS << "Special"; break;
1623 case ICmpZero: OS << "ICmpZero"; break;
1624 case Address:
1625 OS << "Address of ";
1626 if (AccessTy.MemTy->isPointerTy())
1627 OS << "pointer"; // the full pointer type could be really verbose
1628 else {
1629 OS << *AccessTy.MemTy;
1630 }
1631
1632 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1633 }
1634
1635 OS << ", Offsets={";
1636 bool NeedComma = false;
1637 for (const LSRFixup &Fixup : Fixups) {
1638 if (NeedComma) OS << ',';
1639 OS << Fixup.Offset;
1640 NeedComma = true;
1641 }
1642 OS << '}';
1643
1644 if (AllFixupsOutsideLoop)
1645 OS << ", all-fixups-outside-loop";
1646
1647 if (WidestFixupType)
1648 OS << ", widest fixup type: " << *WidestFixupType;
1649}
1650
1651LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1652 print(errs()); errs() << '\n';
1653}
1654#endif
1655
1656static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1657 LSRUse::KindType Kind, MemAccessTy AccessTy,
1658 GlobalValue *BaseGV, int64_t BaseOffset,
1659 bool HasBaseReg, int64_t Scale,
1660 Instruction *Fixup/*= nullptr*/) {
1661 switch (Kind) {
1662 case LSRUse::Address:
1663 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1664 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1665
1666 case LSRUse::ICmpZero:
1667 // There's not even a target hook for querying whether it would be legal to
1668 // fold a GV into an ICmp.
1669 if (BaseGV)
1670 return false;
1671
1672 // ICmp only has two operands; don't allow more than two non-trivial parts.
1673 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1674 return false;
1675
1676 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1677 // putting the scaled register in the other operand of the icmp.
1678 if (Scale != 0 && Scale != -1)
1679 return false;
1680
1681 // If we have low-level target information, ask the target if it can fold an
1682 // integer immediate on an icmp.
1683 if (BaseOffset != 0) {
1684 // We have one of:
1685 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1686 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1687 // Offs is the ICmp immediate.
1688 if (Scale == 0)
1689 // The cast does the right thing with
1690 // std::numeric_limits<int64_t>::min().
1691 BaseOffset = -(uint64_t)BaseOffset;
1692 return TTI.isLegalICmpImmediate(BaseOffset);
1693 }
1694
1695 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1696 return true;
1697
1698 case LSRUse::Basic:
1699 // Only handle single-register values.
1700 return !BaseGV && Scale == 0 && BaseOffset == 0;
1701
1702 case LSRUse::Special:
1703 // Special case Basic to handle -1 scales.
1704 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1705 }
1706
1707 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1707)
;
1708}
1709
1710static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1711 int64_t MinOffset, int64_t MaxOffset,
1712 LSRUse::KindType Kind, MemAccessTy AccessTy,
1713 GlobalValue *BaseGV, int64_t BaseOffset,
1714 bool HasBaseReg, int64_t Scale) {
1715 // Check for overflow.
1716 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1717 (MinOffset > 0))
1718 return false;
1719 MinOffset = (uint64_t)BaseOffset + MinOffset;
1720 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1721 (MaxOffset > 0))
1722 return false;
1723 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1724
1725 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1726 HasBaseReg, Scale) &&
1727 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1728 HasBaseReg, Scale);
1729}
1730
1731static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1732 int64_t MinOffset, int64_t MaxOffset,
1733 LSRUse::KindType Kind, MemAccessTy AccessTy,
1734 const Formula &F, const Loop &L) {
1735 // For the purpose of isAMCompletelyFolded either having a canonical formula
1736 // or a scale not equal to zero is correct.
1737 // Problems may arise from non canonical formulae having a scale == 0.
1738 // Strictly speaking it would best to just rely on canonical formulae.
1739 // However, when we generate the scaled formulae, we first check that the
1740 // scaling factor is profitable before computing the actual ScaledReg for
1741 // compile time sake.
1742 assert((F.isCanonical(L) || F.Scale != 0))(static_cast <bool> ((F.isCanonical(L) || F.Scale != 0)
) ? void (0) : __assert_fail ("(F.isCanonical(L) || F.Scale != 0)"
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1742, __extension__ __PRETTY_FUNCTION__))
;
1743 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1744 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1745}
1746
1747/// Test whether we know how to expand the current formula.
1748static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1749 int64_t MaxOffset, LSRUse::KindType Kind,
1750 MemAccessTy AccessTy, GlobalValue *BaseGV,
1751 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1752 // We know how to expand completely foldable formulae.
1753 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1754 BaseOffset, HasBaseReg, Scale) ||
1755 // Or formulae that use a base register produced by a sum of base
1756 // registers.
1757 (Scale == 1 &&
1758 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1759 BaseGV, BaseOffset, true, 0));
1760}
1761
1762static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1763 int64_t MaxOffset, LSRUse::KindType Kind,
1764 MemAccessTy AccessTy, const Formula &F) {
1765 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1766 F.BaseOffset, F.HasBaseReg, F.Scale);
1767}
1768
1769static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1770 const LSRUse &LU, const Formula &F) {
1771 // Target may want to look at the user instructions.
1772 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1773 for (const LSRFixup &Fixup : LU.Fixups)
1774 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1775 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1776 F.Scale, Fixup.UserInst))
1777 return false;
1778 return true;
1779 }
1780
1781 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1782 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1783 F.Scale);
1784}
1785
1786static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1787 const LSRUse &LU, const Formula &F,
1788 const Loop &L) {
1789 if (!F.Scale)
1790 return 0;
1791
1792 // If the use is not completely folded in that instruction, we will have to
1793 // pay an extra cost only for scale != 1.
1794 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1795 LU.AccessTy, F, L))
1796 return F.Scale != 1;
1797
1798 switch (LU.Kind) {
1799 case LSRUse::Address: {
1800 // Check the scaling factor cost with both the min and max offsets.
1801 InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost(
1802 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1803 F.Scale, LU.AccessTy.AddrSpace);
1804 InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost(
1805 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1806 F.Scale, LU.AccessTy.AddrSpace);
1807
1808 assert(ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() &&(static_cast <bool> (ScaleCostMinOffset.isValid() &&
ScaleCostMaxOffset.isValid() && "Legal addressing mode has an illegal cost!"
) ? void (0) : __assert_fail ("ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() && \"Legal addressing mode has an illegal cost!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1809, __extension__ __PRETTY_FUNCTION__))
1809 "Legal addressing mode has an illegal cost!")(static_cast <bool> (ScaleCostMinOffset.isValid() &&
ScaleCostMaxOffset.isValid() && "Legal addressing mode has an illegal cost!"
) ? void (0) : __assert_fail ("ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() && \"Legal addressing mode has an illegal cost!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1809, __extension__ __PRETTY_FUNCTION__))
;
1810 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1811 }
1812 case LSRUse::ICmpZero:
1813 case LSRUse::Basic:
1814 case LSRUse::Special:
1815 // The use is completely folded, i.e., everything is folded into the
1816 // instruction.
1817 return 0;
1818 }
1819
1820 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1820)
;
1821}
1822
1823static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1824 LSRUse::KindType Kind, MemAccessTy AccessTy,
1825 GlobalValue *BaseGV, int64_t BaseOffset,
1826 bool HasBaseReg) {
1827 // Fast-path: zero is always foldable.
1828 if (BaseOffset == 0 && !BaseGV) return true;
1829
1830 // Conservatively, create an address with an immediate and a
1831 // base and a scale.
1832 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1833
1834 // Canonicalize a scale of 1 to a base register if the formula doesn't
1835 // already have a base register.
1836 if (!HasBaseReg && Scale == 1) {
1837 Scale = 0;
1838 HasBaseReg = true;
1839 }
1840
1841 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1842 HasBaseReg, Scale);
1843}
1844
1845static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1846 ScalarEvolution &SE, int64_t MinOffset,
1847 int64_t MaxOffset, LSRUse::KindType Kind,
1848 MemAccessTy AccessTy, const SCEV *S,
1849 bool HasBaseReg) {
1850 // Fast-path: zero is always foldable.
1851 if (S->isZero()) return true;
1852
1853 // Conservatively, create an address with an immediate and a
1854 // base and a scale.
1855 int64_t BaseOffset = ExtractImmediate(S, SE);
1856 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1857
1858 // If there's anything else involved, it's not foldable.
1859 if (!S->isZero()) return false;
1860
1861 // Fast-path: zero is always foldable.
1862 if (BaseOffset == 0 && !BaseGV) return true;
1863
1864 // Conservatively, create an address with an immediate and a
1865 // base and a scale.
1866 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1867
1868 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1869 BaseOffset, HasBaseReg, Scale);
1870}
1871
1872namespace {
1873
1874/// An individual increment in a Chain of IV increments. Relate an IV user to
1875/// an expression that computes the IV it uses from the IV used by the previous
1876/// link in the Chain.
1877///
1878/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1879/// original IVOperand. The head of the chain's IVOperand is only valid during
1880/// chain collection, before LSR replaces IV users. During chain generation,
1881/// IncExpr can be used to find the new IVOperand that computes the same
1882/// expression.
1883struct IVInc {
1884 Instruction *UserInst;
1885 Value* IVOperand;
1886 const SCEV *IncExpr;
1887
1888 IVInc(Instruction *U, Value *O, const SCEV *E)
1889 : UserInst(U), IVOperand(O), IncExpr(E) {}
1890};
1891
1892// The list of IV increments in program order. We typically add the head of a
1893// chain without finding subsequent links.
1894struct IVChain {
1895 SmallVector<IVInc, 1> Incs;
1896 const SCEV *ExprBase = nullptr;
1897
1898 IVChain() = default;
1899 IVChain(const IVInc &Head, const SCEV *Base)
1900 : Incs(1, Head), ExprBase(Base) {}
1901
1902 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1903
1904 // Return the first increment in the chain.
1905 const_iterator begin() const {
1906 assert(!Incs.empty())(static_cast <bool> (!Incs.empty()) ? void (0) : __assert_fail
("!Incs.empty()", "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1906, __extension__ __PRETTY_FUNCTION__))
;
1907 return std::next(Incs.begin());
1908 }
1909 const_iterator end() const {
1910 return Incs.end();
1911 }
1912
1913 // Returns true if this chain contains any increments.
1914 bool hasIncs() const { return Incs.size() >= 2; }
1915
1916 // Add an IVInc to the end of this chain.
1917 void add(const IVInc &X) { Incs.push_back(X); }
1918
1919 // Returns the last UserInst in the chain.
1920 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1921
1922 // Returns true if IncExpr can be profitably added to this chain.
1923 bool isProfitableIncrement(const SCEV *OperExpr,
1924 const SCEV *IncExpr,
1925 ScalarEvolution&);
1926};
1927
1928/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1929/// between FarUsers that definitely cross IV increments and NearUsers that may
1930/// be used between IV increments.
1931struct ChainUsers {
1932 SmallPtrSet<Instruction*, 4> FarUsers;
1933 SmallPtrSet<Instruction*, 4> NearUsers;
1934};
1935
1936/// This class holds state for the main loop strength reduction logic.
1937class LSRInstance {
1938 IVUsers &IU;
1939 ScalarEvolution &SE;
1940 DominatorTree &DT;
1941 LoopInfo &LI;
1942 AssumptionCache &AC;
1943 TargetLibraryInfo &TLI;
1944 const TargetTransformInfo &TTI;
1945 Loop *const L;
1946 MemorySSAUpdater *MSSAU;
1947 TTI::AddressingModeKind AMK;
1948 bool Changed = false;
1949
1950 /// This is the insert position that the current loop's induction variable
1951 /// increment should be placed. In simple loops, this is the latch block's
1952 /// terminator. But in more complicated cases, this is a position which will
1953 /// dominate all the in-loop post-increment users.
1954 Instruction *IVIncInsertPos = nullptr;
1955
1956 /// Interesting factors between use strides.
1957 ///
1958 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1959 /// default, a SmallDenseSet, because we need to use the full range of
1960 /// int64_ts, and there's currently no good way of doing that with
1961 /// SmallDenseSet.
1962 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1963
1964 /// Interesting use types, to facilitate truncation reuse.
1965 SmallSetVector<Type *, 4> Types;
1966
1967 /// The list of interesting uses.
1968 mutable SmallVector<LSRUse, 16> Uses;
1969
1970 /// Track which uses use which register candidates.
1971 RegUseTracker RegUses;
1972
1973 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1974 // have more than a few IV increment chains in a loop. Missing a Chain falls
1975 // back to normal LSR behavior for those uses.
1976 static const unsigned MaxChains = 8;
1977
1978 /// IV users can form a chain of IV increments.
1979 SmallVector<IVChain, MaxChains> IVChainVec;
1980
1981 /// IV users that belong to profitable IVChains.
1982 SmallPtrSet<Use*, MaxChains> IVIncSet;
1983
1984 void OptimizeShadowIV();
1985 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1986 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1987 void OptimizeLoopTermCond();
1988
1989 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1990 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1991 void FinalizeChain(IVChain &Chain);
1992 void CollectChains();
1993 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1994 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1995
1996 void CollectInterestingTypesAndFactors();
1997 void CollectFixupsAndInitialFormulae();
1998
1999 // Support for sharing of LSRUses between LSRFixups.
2000 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
2001 UseMapTy UseMap;
2002
2003 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2004 LSRUse::KindType Kind, MemAccessTy AccessTy);
2005
2006 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
2007 MemAccessTy AccessTy);
2008
2009 void DeleteUse(LSRUse &LU, size_t LUIdx);
2010
2011 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
2012
2013 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2014 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2015 void CountRegisters(const Formula &F, size_t LUIdx);
2016 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
2017
2018 void CollectLoopInvariantFixupsAndFormulae();
2019
2020 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
2021 unsigned Depth = 0);
2022
2023 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
2024 const Formula &Base, unsigned Depth,
2025 size_t Idx, bool IsScaledReg = false);
2026 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
2027 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2028 const Formula &Base, size_t Idx,
2029 bool IsScaledReg = false);
2030 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2031 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2032 const Formula &Base,
2033 const SmallVectorImpl<int64_t> &Worklist,
2034 size_t Idx, bool IsScaledReg = false);
2035 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2036 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2037 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2038 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2039 void GenerateCrossUseConstantOffsets();
2040 void GenerateAllReuseFormulae();
2041
2042 void FilterOutUndesirableDedicatedRegisters();
2043
2044 size_t EstimateSearchSpaceComplexity() const;
2045 void NarrowSearchSpaceByDetectingSupersets();
2046 void NarrowSearchSpaceByCollapsingUnrolledCode();
2047 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2048 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2049 void NarrowSearchSpaceByFilterPostInc();
2050 void NarrowSearchSpaceByDeletingCostlyFormulas();
2051 void NarrowSearchSpaceByPickingWinnerRegs();
2052 void NarrowSearchSpaceUsingHeuristics();
2053
2054 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2055 Cost &SolutionCost,
2056 SmallVectorImpl<const Formula *> &Workspace,
2057 const Cost &CurCost,
2058 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2059 DenseSet<const SCEV *> &VisitedRegs) const;
2060 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2061
2062 BasicBlock::iterator
2063 HoistInsertPosition(BasicBlock::iterator IP,
2064 const SmallVectorImpl<Instruction *> &Inputs) const;
2065 BasicBlock::iterator
2066 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2067 const LSRFixup &LF,
2068 const LSRUse &LU,
2069 SCEVExpander &Rewriter) const;
2070
2071 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2072 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2073 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2074 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2075 const Formula &F, SCEVExpander &Rewriter,
2076 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2077 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2078 SCEVExpander &Rewriter,
2079 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2080 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2081
2082public:
2083 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2084 LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2085 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);
2086
2087 bool getChanged() const { return Changed; }
2088
2089 void print_factors_and_types(raw_ostream &OS) const;
2090 void print_fixups(raw_ostream &OS) const;
2091 void print_uses(raw_ostream &OS) const;
2092 void print(raw_ostream &OS) const;
2093 void dump() const;
2094};
2095
2096} // end anonymous namespace
2097
2098/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2099/// the cast operation.
2100void LSRInstance::OptimizeShadowIV() {
2101 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2102 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2103 return;
2104
2105 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2106 UI != E; /* empty */) {
2107 IVUsers::const_iterator CandidateUI = UI;
2108 ++UI;
2109 Instruction *ShadowUse = CandidateUI->getUser();
2110 Type *DestTy = nullptr;
2111 bool IsSigned = false;
2112
2113 /* If shadow use is a int->float cast then insert a second IV
2114 to eliminate this cast.
2115
2116 for (unsigned i = 0; i < n; ++i)
2117 foo((double)i);
2118
2119 is transformed into
2120
2121 double d = 0.0;
2122 for (unsigned i = 0; i < n; ++i, ++d)
2123 foo(d);
2124 */
2125 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2126 IsSigned = false;
2127 DestTy = UCast->getDestTy();
2128 }
2129 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2130 IsSigned = true;
2131 DestTy = SCast->getDestTy();
2132 }
2133 if (!DestTy) continue;
2134
2135 // If target does not support DestTy natively then do not apply
2136 // this transformation.
2137 if (!TTI.isTypeLegal(DestTy)) continue;
2138
2139 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2140 if (!PH) continue;
2141 if (PH->getNumIncomingValues() != 2) continue;
2142
2143 // If the calculation in integers overflows, the result in FP type will
2144 // differ. So we only can do this transformation if we are guaranteed to not
2145 // deal with overflowing values
2146 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2147 if (!AR) continue;
2148 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2149 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2150
2151 Type *SrcTy = PH->getType();
2152 int Mantissa = DestTy->getFPMantissaWidth();
2153 if (Mantissa == -1) continue;
2154 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2155 continue;
2156
2157 unsigned Entry, Latch;
2158 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2159 Entry = 0;
2160 Latch = 1;
2161 } else {
2162 Entry = 1;
2163 Latch = 0;
2164 }
2165
2166 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2167 if (!Init) continue;
2168 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2169 (double)Init->getSExtValue() :
2170 (double)Init->getZExtValue());
2171
2172 BinaryOperator *Incr =
2173 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2174 if (!Incr) continue;
2175 if (Incr->getOpcode() != Instruction::Add
2176 && Incr->getOpcode() != Instruction::Sub)
2177 continue;
2178
2179 /* Initialize new IV, double d = 0.0 in above example. */
2180 ConstantInt *C = nullptr;
2181 if (Incr->getOperand(0) == PH)
2182 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2183 else if (Incr->getOperand(1) == PH)
2184 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2185 else
2186 continue;
2187
2188 if (!C) continue;
2189
2190 // Ignore negative constants, as the code below doesn't handle them
2191 // correctly. TODO: Remove this restriction.
2192 if (!C->getValue().isStrictlyPositive()) continue;
2193
2194 /* Add new PHINode. */
2195 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2196
2197 /* create new increment. '++d' in above example. */
2198 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2199 BinaryOperator *NewIncr =
2200 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2201 Instruction::FAdd : Instruction::FSub,
2202 NewPH, CFP, "IV.S.next.", Incr);
2203
2204 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2205 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2206
2207 /* Remove cast operation */
2208 ShadowUse->replaceAllUsesWith(NewPH);
2209 ShadowUse->eraseFromParent();
2210 Changed = true;
2211 break;
2212 }
2213}
2214
2215/// If Cond has an operand that is an expression of an IV, set the IV user and
2216/// stride information and return true, otherwise return false.
2217bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2218 for (IVStrideUse &U : IU)
2219 if (U.getUser() == Cond) {
2220 // NOTE: we could handle setcc instructions with multiple uses here, but
2221 // InstCombine does it as well for simple uses, it's not clear that it
2222 // occurs enough in real life to handle.
2223 CondUse = &U;
2224 return true;
2225 }
2226 return false;
2227}
2228
2229/// Rewrite the loop's terminating condition if it uses a max computation.
2230///
2231/// This is a narrow solution to a specific, but acute, problem. For loops
2232/// like this:
2233///
2234/// i = 0;
2235/// do {
2236/// p[i] = 0.0;
2237/// } while (++i < n);
2238///
2239/// the trip count isn't just 'n', because 'n' might not be positive. And
2240/// unfortunately this can come up even for loops where the user didn't use
2241/// a C do-while loop. For example, seemingly well-behaved top-test loops
2242/// will commonly be lowered like this:
2243///
2244/// if (n > 0) {
2245/// i = 0;
2246/// do {
2247/// p[i] = 0.0;
2248/// } while (++i < n);
2249/// }
2250///
2251/// and then it's possible for subsequent optimization to obscure the if
2252/// test in such a way that indvars can't find it.
2253///
2254/// When indvars can't find the if test in loops like this, it creates a
2255/// max expression, which allows it to give the loop a canonical
2256/// induction variable:
2257///
2258/// i = 0;
2259/// max = n < 1 ? 1 : n;
2260/// do {
2261/// p[i] = 0.0;
2262/// } while (++i != max);
2263///
2264/// Canonical induction variables are necessary because the loop passes
2265/// are designed around them. The most obvious example of this is the
2266/// LoopInfo analysis, which doesn't remember trip count values. It
2267/// expects to be able to rediscover the trip count each time it is
2268/// needed, and it does this using a simple analysis that only succeeds if
2269/// the loop has a canonical induction variable.
2270///
2271/// However, when it comes time to generate code, the maximum operation
2272/// can be quite costly, especially if it's inside of an outer loop.
2273///
2274/// This function solves this problem by detecting this type of loop and
2275/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2276/// the instructions for the maximum computation.
2277ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2278 // Check that the loop matches the pattern we're looking for.
2279 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2280 Cond->getPredicate() != CmpInst::ICMP_NE)
2281 return Cond;
2282
2283 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2284 if (!Sel || !Sel->hasOneUse()) return Cond;
2285
2286 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2287 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2288 return Cond;
2289 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2290
2291 // Add one to the backedge-taken count to get the trip count.
2292 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2293 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2294
2295 // Check for a max calculation that matches the pattern. There's no check
2296 // for ICMP_ULE here because the comparison would be with zero, which
2297 // isn't interesting.
2298 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2299 const SCEVNAryExpr *Max = nullptr;
2300 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2301 Pred = ICmpInst::ICMP_SLE;
2302 Max = S;
2303 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2304 Pred = ICmpInst::ICMP_SLT;
2305 Max = S;
2306 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2307 Pred = ICmpInst::ICMP_ULT;
2308 Max = U;
2309 } else {
2310 // No match; bail.
2311 return Cond;
2312 }
2313
2314 // To handle a max with more than two operands, this optimization would
2315 // require additional checking and setup.
2316 if (Max->getNumOperands() != 2)
2317 return Cond;
2318
2319 const SCEV *MaxLHS = Max->getOperand(0);
2320 const SCEV *MaxRHS = Max->getOperand(1);
2321
2322 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2323 // for a comparison with 1. For <= and >=, a comparison with zero.
2324 if (!MaxLHS ||
2325 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2326 return Cond;
2327
2328 // Check the relevant induction variable for conformance to
2329 // the pattern.
2330 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2331 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2332 if (!AR || !AR->isAffine() ||
2333 AR->getStart() != One ||
2334 AR->getStepRecurrence(SE) != One)
2335 return Cond;
2336
2337 assert(AR->getLoop() == L &&(static_cast <bool> (AR->getLoop() == L && "Loop condition operand is an addrec in a different loop!"
) ? void (0) : __assert_fail ("AR->getLoop() == L && \"Loop condition operand is an addrec in a different loop!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2338, __extension__ __PRETTY_FUNCTION__))
2338 "Loop condition operand is an addrec in a different loop!")(static_cast <bool> (AR->getLoop() == L && "Loop condition operand is an addrec in a different loop!"
) ? void (0) : __assert_fail ("AR->getLoop() == L && \"Loop condition operand is an addrec in a different loop!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2338, __extension__ __PRETTY_FUNCTION__))
;
2339
2340 // Check the right operand of the select, and remember it, as it will
2341 // be used in the new comparison instruction.
2342 Value *NewRHS = nullptr;
2343 if (ICmpInst::isTrueWhenEqual(Pred)) {
2344 // Look for n+1, and grab n.
2345 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2346 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2347 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2348 NewRHS = BO->getOperand(0);
2349 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2350 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2351 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2352 NewRHS = BO->getOperand(0);
2353 if (!NewRHS)
2354 return Cond;
2355 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2356 NewRHS = Sel->getOperand(1);
2357 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2358 NewRHS = Sel->getOperand(2);
2359 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2360 NewRHS = SU->getValue();
2361 else
2362 // Max doesn't match expected pattern.
2363 return Cond;
2364
2365 // Determine the new comparison opcode. It may be signed or unsigned,
2366 // and the original comparison may be either equality or inequality.
2367 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2368 Pred = CmpInst::getInversePredicate(Pred);
2369
2370 // Ok, everything looks ok to change the condition into an SLT or SGE and
2371 // delete the max calculation.
2372 ICmpInst *NewCond =
2373 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2374
2375 // Delete the max calculation instructions.
2376 NewCond->setDebugLoc(Cond->getDebugLoc());
2377 Cond->replaceAllUsesWith(NewCond);
2378 CondUse->setUser(NewCond);
2379 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2380 Cond->eraseFromParent();
2381 Sel->eraseFromParent();
2382 if (Cmp->use_empty())
2383 Cmp->eraseFromParent();
2384 return NewCond;
2385}
2386
2387/// Change loop terminating condition to use the postinc iv when possible.
2388void
2389LSRInstance::OptimizeLoopTermCond() {
2390 SmallPtrSet<Instruction *, 4> PostIncs;
2391
2392 // We need a different set of heuristics for rotated and non-rotated loops.
2393 // If a loop is rotated then the latch is also the backedge, so inserting
2394 // post-inc expressions just before the latch is ideal. To reduce live ranges
2395 // it also makes sense to rewrite terminating conditions to use post-inc
2396 // expressions.
2397 //
2398 // If the loop is not rotated then the latch is not a backedge; the latch
2399 // check is done in the loop head. Adding post-inc expressions before the
2400 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2401 // in the loop body. In this case we do *not* want to use post-inc expressions
2402 // in the latch check, and we want to insert post-inc expressions before
2403 // the backedge.
2404 BasicBlock *LatchBlock = L->getLoopLatch();
2405 SmallVector<BasicBlock*, 8> ExitingBlocks;
2406 L->getExitingBlocks(ExitingBlocks);
2407 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2408 return LatchBlock != BB;
2409 })) {
2410 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2411 IVIncInsertPos = LatchBlock->getTerminator();
2412 return;
2413 }
2414
2415 // Otherwise treat this as a rotated loop.
2416 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2417 // Get the terminating condition for the loop if possible. If we
2418 // can, we want to change it to use a post-incremented version of its
2419 // induction variable, to allow coalescing the live ranges for the IV into
2420 // one register value.
2421
2422 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2423 if (!TermBr)
2424 continue;
2425 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2426 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2427 continue;
2428
2429 // Search IVUsesByStride to find Cond's IVUse if there is one.
2430 IVStrideUse *CondUse = nullptr;
2431 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2432 if (!FindIVUserForCond(Cond, CondUse))
2433 continue;
2434
2435 // If the trip count is computed in terms of a max (due to ScalarEvolution
2436 // being unable to find a sufficient guard, for example), change the loop
2437 // comparison to use SLT or ULT instead of NE.
2438 // One consequence of doing this now is that it disrupts the count-down
2439 // optimization. That's not always a bad thing though, because in such
2440 // cases it may still be worthwhile to avoid a max.
2441 Cond = OptimizeMax(Cond, CondUse);
2442
2443 // If this exiting block dominates the latch block, it may also use
2444 // the post-inc value if it won't be shared with other uses.
2445 // Check for dominance.
2446 if (!DT.dominates(ExitingBlock, LatchBlock))
2447 continue;
2448
2449 // Conservatively avoid trying to use the post-inc value in non-latch
2450 // exits if there may be pre-inc users in intervening blocks.
2451 if (LatchBlock != ExitingBlock)
2452 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2453 // Test if the use is reachable from the exiting block. This dominator
2454 // query is a conservative approximation of reachability.
2455 if (&*UI != CondUse &&
2456 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2457 // Conservatively assume there may be reuse if the quotient of their
2458 // strides could be a legal scale.
2459 const SCEV *A = IU.getStride(*CondUse, L);
2460 const SCEV *B = IU.getStride(*UI, L);
2461 if (!A || !B) continue;
2462 if (SE.getTypeSizeInBits(A->getType()) !=
2463 SE.getTypeSizeInBits(B->getType())) {
2464 if (SE.getTypeSizeInBits(A->getType()) >
2465 SE.getTypeSizeInBits(B->getType()))
2466 B = SE.getSignExtendExpr(B, A->getType());
2467 else
2468 A = SE.getSignExtendExpr(A, B->getType());
2469 }
2470 if (const SCEVConstant *D =
2471 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2472 const ConstantInt *C = D->getValue();
2473 // Stride of one or negative one can have reuse with non-addresses.
2474 if (C->isOne() || C->isMinusOne())
2475 goto decline_post_inc;
2476 // Avoid weird situations.
2477 if (C->getValue().getMinSignedBits() >= 64 ||
2478 C->getValue().isMinSignedValue())
2479 goto decline_post_inc;
2480 // Check for possible scaled-address reuse.
2481 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2482 MemAccessTy AccessTy = getAccessType(
2483 TTI, UI->getUser(), UI->getOperandValToReplace());
2484 int64_t Scale = C->getSExtValue();
2485 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2486 /*BaseOffset=*/0,
2487 /*HasBaseReg=*/false, Scale,
2488 AccessTy.AddrSpace))
2489 goto decline_post_inc;
2490 Scale = -Scale;
2491 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2492 /*BaseOffset=*/0,
2493 /*HasBaseReg=*/false, Scale,
2494 AccessTy.AddrSpace))
2495 goto decline_post_inc;
2496 }
2497 }
2498 }
2499
2500 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)
2501 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2502
2503 // It's possible for the setcc instruction to be anywhere in the loop, and
2504 // possible for it to have multiple users. If it is not immediately before
2505 // the exiting block branch, move it.
2506 if (Cond->getNextNonDebugInstruction() != TermBr) {
2507 if (Cond->hasOneUse()) {
2508 Cond->moveBefore(TermBr);
2509 } else {
2510 // Clone the terminating condition and insert into the loopend.
2511 ICmpInst *OldCond = Cond;
2512 Cond = cast<ICmpInst>(Cond->clone());
2513 Cond->setName(L->getHeader()->getName() + ".termcond");
2514 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2515
2516 // Clone the IVUse, as the old use still exists!
2517 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2518 TermBr->replaceUsesOfWith(OldCond, Cond);
2519 }
2520 }
2521
2522 // If we get to here, we know that we can transform the setcc instruction to
2523 // use the post-incremented version of the IV, allowing us to coalesce the
2524 // live ranges for the IV correctly.
2525 CondUse->transformToPostInc(L);
2526 Changed = true;
2527
2528 PostIncs.insert(Cond);
2529 decline_post_inc:;
2530 }
2531
2532 // Determine an insertion point for the loop induction variable increment. It
2533 // must dominate all the post-inc comparisons we just set up, and it must
2534 // dominate the loop latch edge.
2535 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2536 for (Instruction *Inst : PostIncs) {
2537 BasicBlock *BB =
2538 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2539 Inst->getParent());
2540 if (BB == Inst->getParent())
2541 IVIncInsertPos = Inst;
2542 else if (BB != IVIncInsertPos->getParent())
2543 IVIncInsertPos = BB->getTerminator();
2544 }
2545}
2546
2547/// Determine if the given use can accommodate a fixup at the given offset and
2548/// other details. If so, update the use and return true.
2549bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2550 bool HasBaseReg, LSRUse::KindType Kind,
2551 MemAccessTy AccessTy) {
2552 int64_t NewMinOffset = LU.MinOffset;
2553 int64_t NewMaxOffset = LU.MaxOffset;
2554 MemAccessTy NewAccessTy = AccessTy;
2555
2556 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2557 // something conservative, however this can pessimize in the case that one of
2558 // the uses will have all its uses outside the loop, for example.
2559 if (LU.Kind != Kind)
2560 return false;
2561
2562 // Check for a mismatched access type, and fall back conservatively as needed.
2563 // TODO: Be less conservative when the type is similar and can use the same
2564 // addressing modes.
2565 if (Kind == LSRUse::Address) {
2566 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2567 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2568 AccessTy.AddrSpace);
2569 }
2570 }
2571
2572 // Conservatively assume HasBaseReg is true for now.
2573 if (NewOffset < LU.MinOffset) {
2574 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2575 LU.MaxOffset - NewOffset, HasBaseReg))
2576 return false;
2577 NewMinOffset = NewOffset;
2578 } else if (NewOffset > LU.MaxOffset) {
2579 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2580 NewOffset - LU.MinOffset, HasBaseReg))
2581 return false;
2582 NewMaxOffset = NewOffset;
2583 }
2584
2585 // Update the use.
2586 LU.MinOffset = NewMinOffset;
2587 LU.MaxOffset = NewMaxOffset;
2588 LU.AccessTy = NewAccessTy;
2589 return true;
2590}
2591
2592/// Return an LSRUse index and an offset value for a fixup which needs the given
2593/// expression, with the given kind and optional access type. Either reuse an
2594/// existing use or create a new one, as needed.
2595std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2596 LSRUse::KindType Kind,
2597 MemAccessTy AccessTy) {
2598 const SCEV *Copy = Expr;
2599 int64_t Offset = ExtractImmediate(Expr, SE);
2600
2601 // Basic uses can't accept any offset, for example.
2602 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2603 Offset, /*HasBaseReg=*/ true)) {
2604 Expr = Copy;
2605 Offset = 0;
2606 }
2607
2608 std::pair<UseMapTy::iterator, bool> P =
2609 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2610 if (!P.second) {
2611 // A use already existed with this base.
2612 size_t LUIdx = P.first->second;
2613 LSRUse &LU = Uses[LUIdx];
2614 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2615 // Reuse this use.
2616 return std::make_pair(LUIdx, Offset);
2617 }
2618
2619 // Create a new use.
2620 size_t LUIdx = Uses.size();
2621 P.first->second = LUIdx;
2622 Uses.push_back(LSRUse(Kind, AccessTy));
2623 LSRUse &LU = Uses[LUIdx];
2624
2625 LU.MinOffset = Offset;
2626 LU.MaxOffset = Offset;
2627 return std::make_pair(LUIdx, Offset);
2628}
2629
2630/// Delete the given use from the Uses list.
2631void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2632 if (&LU != &Uses.back())
2633 std::swap(LU, Uses.back());
2634 Uses.pop_back();
2635
2636 // Update RegUses.
2637 RegUses.swapAndDropUse(LUIdx, Uses.size());
2638}
2639
2640/// Look for a use distinct from OrigLU which is has a formula that has the same
2641/// registers as the given formula.
2642LSRUse *
2643LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2644 const LSRUse &OrigLU) {
2645 // Search all uses for the formula. This could be more clever.
2646 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2647 LSRUse &LU = Uses[LUIdx];
2648 // Check whether this use is close enough to OrigLU, to see whether it's
2649 // worthwhile looking through its formulae.
2650 // Ignore ICmpZero uses because they may contain formulae generated by
2651 // GenerateICmpZeroScales, in which case adding fixup offsets may
2652 // be invalid.
2653 if (&LU != &OrigLU &&
2654 LU.Kind != LSRUse::ICmpZero &&
2655 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2656 LU.WidestFixupType == OrigLU.WidestFixupType &&
2657 LU.HasFormulaWithSameRegs(OrigF)) {
2658 // Scan through this use's formulae.
2659 for (const Formula &F : LU.Formulae) {
2660 // Check to see if this formula has the same registers and symbols
2661 // as OrigF.
2662 if (F.BaseRegs == OrigF.BaseRegs &&
2663 F.ScaledReg == OrigF.ScaledReg &&
2664 F.BaseGV == OrigF.BaseGV &&
2665 F.Scale == OrigF.Scale &&
2666 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2667 if (F.BaseOffset == 0)
2668 return &LU;
2669 // This is the formula where all the registers and symbols matched;
2670 // there aren't going to be any others. Since we declined it, we
2671 // can skip the rest of the formulae and proceed to the next LSRUse.
2672 break;
2673 }
2674 }
2675 }
2676 }
2677
2678 // Nothing looked good.
2679 return nullptr;
2680}
2681
2682void LSRInstance::CollectInterestingTypesAndFactors() {
2683 SmallSetVector<const SCEV *, 4> Strides;
2684
2685 // Collect interesting types and strides.
2686 SmallVector<const SCEV *, 4> Worklist;
2687 for (const IVStrideUse &U : IU) {
2688 const SCEV *Expr = IU.getExpr(U);
2689
2690 // Collect interesting types.
2691 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2692
2693 // Add strides for mentioned loops.
2694 Worklist.push_back(Expr);
2695 do {
2696 const SCEV *S = Worklist.pop_back_val();
2697 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2698 if (AR->getLoop() == L)
2699 Strides.insert(AR->getStepRecurrence(SE));
2700 Worklist.push_back(AR->getStart());
2701 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2702 Worklist.append(Add->op_begin(), Add->op_end());
2703 }
2704 } while (!Worklist.empty());
2705 }
2706
2707 // Compute interesting factors from the set of interesting strides.
2708 for (SmallSetVector<const SCEV *, 4>::const_iterator
2709 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2710 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2711 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2712 const SCEV *OldStride = *I;
2713 const SCEV *NewStride = *NewStrideIter;
2714
2715 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2716 SE.getTypeSizeInBits(NewStride->getType())) {
2717 if (SE.getTypeSizeInBits(OldStride->getType()) >
2718 SE.getTypeSizeInBits(NewStride->getType()))
2719 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2720 else
2721 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2722 }
2723 if (const SCEVConstant *Factor =
2724 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2725 SE, true))) {
2726 if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2727 Factors.insert(Factor->getAPInt().getSExtValue());
2728 } else if (const SCEVConstant *Factor =
2729 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2730 NewStride,
2731 SE, true))) {
2732 if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2733 Factors.insert(Factor->getAPInt().getSExtValue());
2734 }
2735 }
2736
2737 // If all uses use the same type, don't bother looking for truncation-based
2738 // reuse.
2739 if (Types.size() == 1)
2740 Types.clear();
2741
2742 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2743}
2744
2745/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2746/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2747/// IVStrideUses, we could partially skip this.
2748static User::op_iterator
2749findIVOperand(User::op_iterator OI, User::op_iterator OE,
2750 Loop *L, ScalarEvolution &SE) {
2751 for(; OI != OE; ++OI) {
2752 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2753 if (!SE.isSCEVable(Oper->getType()))
2754 continue;
2755
2756 if (const SCEVAddRecExpr *AR =
2757 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2758 if (AR->getLoop() == L)
2759 break;
2760 }
2761 }
2762 }
2763 return OI;
2764}
2765
2766/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2767/// a convenient helper.
2768static Value *getWideOperand(Value *Oper) {
2769 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2770 return Trunc->getOperand(0);
2771 return Oper;
2772}
2773
2774/// Return true if we allow an IV chain to include both types.
2775static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2776 Type *LType = LVal->getType();
2777 Type *RType = RVal->getType();
2778 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2779 // Different address spaces means (possibly)
2780 // different types of the pointer implementation,
2781 // e.g. i16 vs i32 so disallow that.
2782 (LType->getPointerAddressSpace() ==
2783 RType->getPointerAddressSpace()));
2784}
2785
2786/// Return an approximation of this SCEV expression's "base", or NULL for any
2787/// constant. Returning the expression itself is conservative. Returning a
2788/// deeper subexpression is more precise and valid as long as it isn't less
2789/// complex than another subexpression. For expressions involving multiple
2790/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2791/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2792/// IVInc==b-a.
2793///
2794/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2795/// SCEVUnknown, we simply return the rightmost SCEV operand.
2796static const SCEV *getExprBase(const SCEV *S) {
2797 switch (S->getSCEVType()) {
2798 default: // uncluding scUnknown.
2799 return S;
2800 case scConstant:
2801 return nullptr;
2802 case scTruncate:
2803 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2804 case scZeroExtend:
2805 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2806 case scSignExtend:
2807 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2808 case scAddExpr: {
2809 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2810 // there's nothing more complex.
2811 // FIXME: not sure if we want to recognize negation.
2812 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2813 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2814 E(Add->op_begin()); I != E; ++I) {
2815 const SCEV *SubExpr = *I;
2816 if (SubExpr->getSCEVType() == scAddExpr)
2817 return getExprBase(SubExpr);
2818
2819 if (SubExpr->getSCEVType() != scMulExpr)
2820 return SubExpr;
2821 }
2822 return S; // all operands are scaled, be conservative.
2823 }
2824 case scAddRecExpr:
2825 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2826 }
2827 llvm_unreachable("Unknown SCEV kind!")::llvm::llvm_unreachable_internal("Unknown SCEV kind!", "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2827)
;
2828}
2829
2830/// Return true if the chain increment is profitable to expand into a loop
2831/// invariant value, which may require its own register. A profitable chain
2832/// increment will be an offset relative to the same base. We allow such offsets
2833/// to potentially be used as chain increment as long as it's not obviously
2834/// expensive to expand using real instructions.
2835bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2836 const SCEV *IncExpr,
2837 ScalarEvolution &SE) {
2838 // Aggressively form chains when -stress-ivchain.
2839 if (StressIVChain)
2840 return true;
2841
2842 // Do not replace a constant offset from IV head with a nonconstant IV
2843 // increment.
2844 if (!isa<SCEVConstant>(IncExpr)) {
2845 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2846 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2847 return false;
2848 }
2849
2850 SmallPtrSet<const SCEV*, 8> Processed;
2851 return !isHighCostExpansion(IncExpr, Processed, SE);
2852}
2853
2854/// Return true if the number of registers needed for the chain is estimated to
2855/// be less than the number required for the individual IV users. First prohibit
2856/// any IV users that keep the IV live across increments (the Users set should
2857/// be empty). Next count the number and type of increments in the chain.
2858///
2859/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2860/// effectively use postinc addressing modes. Only consider it profitable it the
2861/// increments can be computed in fewer registers when chained.
2862///
2863/// TODO: Consider IVInc free if it's already used in another chains.
2864static bool isProfitableChain(IVChain &Chain,
2865 SmallPtrSetImpl<Instruction *> &Users,
2866 ScalarEvolution &SE,
2867 const TargetTransformInfo &TTI) {
2868 if (StressIVChain)
2869 return true;
2870
2871 if (!Chain.hasIncs())
2872 return false;
2873
2874 if (!Users.empty()) {
2875 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)
2876 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)
2877 : 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)
;
2878 return false;
2879 }
2880 assert(!Chain.Incs.empty() && "empty IV chains are not allowed")(static_cast <bool> (!Chain.Incs.empty() && "empty IV chains are not allowed"
) ? void (0) : __assert_fail ("!Chain.Incs.empty() && \"empty IV chains are not allowed\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2880, __extension__ __PRETTY_FUNCTION__))
;
2881
2882 // The chain itself may require a register, so intialize cost to 1.
2883 int cost = 1;
2884
2885 // A complete chain likely eliminates the need for keeping the original IV in
2886 // a register. LSR does not currently know how to form a complete chain unless
2887 // the header phi already exists.
2888 if (isa<PHINode>(Chain.tailUserInst())
2889 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2890 --cost;
2891 }
2892 const SCEV *LastIncExpr = nullptr;
2893 unsigned NumConstIncrements = 0;
2894 unsigned NumVarIncrements = 0;
2895 unsigned NumReusedIncrements = 0;
2896
2897 if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
2898 return true;
2899
2900 for (const IVInc &Inc : Chain) {
2901 if (TTI.isProfitableLSRChainElement(Inc.UserInst))
2902 return true;
2903 if (Inc.IncExpr->isZero())
2904 continue;
2905
2906 // Incrementing by zero or some constant is neutral. We assume constants can
2907 // be folded into an addressing mode or an add's immediate operand.
2908 if (isa<SCEVConstant>(Inc.IncExpr)) {
2909 ++NumConstIncrements;
2910 continue;
2911 }
2912
2913 if (Inc.IncExpr == LastIncExpr)
2914 ++NumReusedIncrements;
2915 else
2916 ++NumVarIncrements;
2917
2918 LastIncExpr = Inc.IncExpr;
2919 }
2920 // An IV chain with a single increment is handled by LSR's postinc
2921 // uses. However, a chain with multiple increments requires keeping the IV's
2922 // value live longer than it needs to be if chained.
2923 if (NumConstIncrements > 1)
2924 --cost;
2925
2926 // Materializing increment expressions in the preheader that didn't exist in
2927 // the original code may cost a register. For example, sign-extended array
2928 // indices can produce ridiculous increments like this:
2929 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2930 cost += NumVarIncrements;
2931
2932 // Reusing variable increments likely saves a register to hold the multiple of
2933 // the stride.
2934 cost -= NumReusedIncrements;
2935
2936 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)
2937 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2938
2939 return cost < 0;
2940}
2941
2942/// Add this IV user to an existing chain or make it the head of a new chain.
2943void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2944 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2945 // When IVs are used as types of varying widths, they are generally converted
2946 // to a wider type with some uses remaining narrow under a (free) trunc.
2947 Value *const NextIV = getWideOperand(IVOper);
2948 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2949 const SCEV *const OperExprBase = getExprBase(OperExpr);
2950
2951 // Visit all existing chains. Check if its IVOper can be computed as a
2952 // profitable loop invariant increment from the last link in the Chain.
2953 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2954 const SCEV *LastIncExpr = nullptr;
2955 for (; ChainIdx < NChains; ++ChainIdx) {
2956 IVChain &Chain = IVChainVec[ChainIdx];
2957
2958 // Prune the solution space aggressively by checking that both IV operands
2959 // are expressions that operate on the same unscaled SCEVUnknown. This
2960 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2961 // first avoids creating extra SCEV expressions.
2962 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2963 continue;
2964
2965 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2966 if (!isCompatibleIVType(PrevIV, NextIV))
2967 continue;
2968
2969 // A phi node terminates a chain.
2970 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2971 continue;
2972
2973 // The increment must be loop-invariant so it can be kept in a register.
2974 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2975 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2976 if (isa<SCEVCouldNotCompute>(IncExpr) || !SE.isLoopInvariant(IncExpr, L))
2977 continue;
2978
2979 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2980 LastIncExpr = IncExpr;
2981 break;
2982 }
2983 }
2984 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2985 // bother for phi nodes, because they must be last in the chain.
2986 if (ChainIdx == NChains) {
2987 if (isa<PHINode>(UserInst))
2988 return;
2989 if (NChains >= MaxChains && !StressIVChain) {
2990 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2991 return;
2992 }
2993 LastIncExpr = OperExpr;
2994 // IVUsers may have skipped over sign/zero extensions. We don't currently
2995 // attempt to form chains involving extensions unless they can be hoisted
2996 // into this loop's AddRec.
2997 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2998 return;
2999 ++NChains;
3000 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
3001 OperExprBase));
3002 ChainUsersVec.resize(NChains);
3003 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)
3004 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
3005 } else {
3006 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)
3007 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
3008 // Add this IV user to the end of the chain.
3009 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
3010 }
3011 IVChain &Chain = IVChainVec[ChainIdx];
3012
3013 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
3014 // This chain's NearUsers become FarUsers.
3015 if (!LastIncExpr->isZero()) {
3016 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
3017 NearUsers.end());
3018 NearUsers.clear();
3019 }
3020
3021 // All other uses of IVOperand become near uses of the chain.
3022 // We currently ignore intermediate values within SCEV expressions, assuming
3023 // they will eventually be used be the current chain, or can be computed
3024 // from one of the chain increments. To be more precise we could
3025 // transitively follow its user and only add leaf IV users to the set.
3026 for (User *U : IVOper->users()) {
3027 Instruction *OtherUse = dyn_cast<Instruction>(U);
3028 if (!OtherUse)
3029 continue;
3030 // Uses in the chain will no longer be uses if the chain is formed.
3031 // Include the head of the chain in this iteration (not Chain.begin()).
3032 IVChain::const_iterator IncIter = Chain.Incs.begin();
3033 IVChain::const_iterator IncEnd = Chain.Incs.end();
3034 for( ; IncIter != IncEnd; ++IncIter) {
3035 if (IncIter->UserInst == OtherUse)
3036 break;
3037 }
3038 if (IncIter != IncEnd)
3039 continue;
3040
3041 if (SE.isSCEVable(OtherUse->getType())
3042 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3043 && IU.isIVUserOrOperand(OtherUse)) {
3044 continue;
3045 }
3046 NearUsers.insert(OtherUse);
3047 }
3048
3049 // Since this user is part of the chain, it's no longer considered a use
3050 // of the chain.
3051 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3052}
3053
3054/// Populate the vector of Chains.
3055///
3056/// This decreases ILP at the architecture level. Targets with ample registers,
3057/// multiple memory ports, and no register renaming probably don't want
3058/// this. However, such targets should probably disable LSR altogether.
3059///
3060/// The job of LSR is to make a reasonable choice of induction variables across
3061/// the loop. Subsequent passes can easily "unchain" computation exposing more
3062/// ILP *within the loop* if the target wants it.
3063///
3064/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3065/// will not reorder memory operations, it will recognize this as a chain, but
3066/// will generate redundant IV increments. Ideally this would be corrected later
3067/// by a smart scheduler:
3068/// = A[i]
3069/// = A[i+x]
3070/// A[i] =
3071/// A[i+x] =
3072///
3073/// TODO: Walk the entire domtree within this loop, not just the path to the
3074/// loop latch. This will discover chains on side paths, but requires
3075/// maintaining multiple copies of the Chains state.
3076void LSRInstance::CollectChains() {
3077 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3078 SmallVector<ChainUsers, 8> ChainUsersVec;
3079
3080 SmallVector<BasicBlock *,8> LatchPath;
3081 BasicBlock *LoopHeader = L->getHeader();
3082 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3083 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3084 LatchPath.push_back(Rung->getBlock());
3085 }
3086 LatchPath.push_back(LoopHeader);
3087
3088 // Walk the instruction stream from the loop header to the loop latch.
3089 for (BasicBlock *BB : reverse(LatchPath)) {
3090 for (Instruction &I : *BB) {
3091 // Skip instructions that weren't seen by IVUsers analysis.
3092 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3093 continue;
3094
3095 // Ignore users that are part of a SCEV expression. This way we only
3096 // consider leaf IV Users. This effectively rediscovers a portion of
3097 // IVUsers analysis but in program order this time.
3098 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3099 continue;
3100
3101 // Remove this instruction from any NearUsers set it may be in.
3102 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3103 ChainIdx < NChains; ++ChainIdx) {
3104 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3105 }
3106 // Search for operands that can be chained.
3107 SmallPtrSet<Instruction*, 4> UniqueOperands;
3108 User::op_iterator IVOpEnd = I.op_end();
3109 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3110 while (IVOpIter != IVOpEnd) {
3111 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3112 if (UniqueOperands.insert(IVOpInst).second)
3113 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3114 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3115 }
3116 } // Continue walking down the instructions.
3117 } // Continue walking down the domtree.
3118 // Visit phi backedges to determine if the chain can generate the IV postinc.
3119 for (PHINode &PN : L->getHeader()->phis()) {
3120 if (!SE.isSCEVable(PN.getType()))
3121 continue;
3122
3123 Instruction *IncV =
3124 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3125 if (IncV)
3126 ChainInstruction(&PN, IncV, ChainUsersVec);
3127 }
3128 // Remove any unprofitable chains.
3129 unsigned ChainIdx = 0;
3130 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3131 UsersIdx < NChains; ++UsersIdx) {
3132 if (!isProfitableChain(IVChainVec[UsersIdx],
3133 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3134 continue;
3135 // Preserve the chain at UsesIdx.
3136 if (ChainIdx != UsersIdx)
3137 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3138 FinalizeChain(IVChainVec[ChainIdx]);
3139 ++ChainIdx;
3140 }
3141 IVChainVec.resize(ChainIdx);
3142}
3143
3144void LSRInstance::FinalizeChain(IVChain &Chain) {
3145 assert(!Chain.Incs.empty() && "empty IV chains are not allowed")(static_cast <bool> (!Chain.Incs.empty() && "empty IV chains are not allowed"
) ? void (0) : __assert_fail ("!Chain.Incs.empty() && \"empty IV chains are not allowed\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3145, __extension__ __PRETTY_FUNCTION__))
;
3146 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)
;
3147
3148 for (const IVInc &Inc : Chain) {
3149 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3150 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3151 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand")(static_cast <bool> (UseI != Inc.UserInst->op_end() &&
"cannot find IV operand") ? void (0) : __assert_fail ("UseI != Inc.UserInst->op_end() && \"cannot find IV operand\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3151, __extension__ __PRETTY_FUNCTION__))
;
3152 IVIncSet.insert(UseI);
3153 }
3154}
3155
3156/// Return true if the IVInc can be folded into an addressing mode.
3157static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3158 Value *Operand, const TargetTransformInfo &TTI) {
3159 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3160 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3161 return false;
3162
3163 if (IncConst->getAPInt().getMinSignedBits() > 64)
3164 return false;
3165
3166 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3167 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3168 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3169 IncOffset, /*HasBaseReg=*/false))
3170 return false;
3171
3172 return true;
3173}
3174
3175/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3176/// user's operand from the previous IV user's operand.
3177void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3178 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3179 // Find the new IVOperand for the head of the chain. It may have been replaced
3180 // by LSR.
3181 const IVInc &Head = Chain.Incs[0];
3182 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3183 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3184 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3185 IVOpEnd, L, SE);
3186 Value *IVSrc = nullptr;
3187 while (IVOpIter != IVOpEnd) {
3188 IVSrc = getWideOperand(*IVOpIter);
3189
3190 // If this operand computes the expression that the chain needs, we may use
3191 // it. (Check this after setting IVSrc which is used below.)
3192 //
3193 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3194 // narrow for the chain, so we can no longer use it. We do allow using a
3195 // wider phi, assuming the LSR checked for free truncation. In that case we
3196 // should already have a truncate on this operand such that
3197 // getSCEV(IVSrc) == IncExpr.
3198 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3199 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3200 break;
3201 }
3202 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3203 }
3204 if (IVOpIter == IVOpEnd) {
3205 // Gracefully give up on this chain.
3206 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)
;
3207 return;
3208 }
3209 assert(IVSrc && "Failed to find IV chain source")(static_cast <bool> (IVSrc && "Failed to find IV chain source"
) ? void (0) : __assert_fail ("IVSrc && \"Failed to find IV chain source\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3209, __extension__ __PRETTY_FUNCTION__))
;
3210
3211 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3212 Type *IVTy = IVSrc->getType();
3213 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3214 const SCEV *LeftOverExpr = nullptr;
3215 for (const IVInc &Inc : Chain) {
3216 Instruction *InsertPt = Inc.UserInst;
3217 if (isa<PHINode>(InsertPt))
3218 InsertPt = L->getLoopLatch()->getTerminator();
3219
3220 // IVOper will replace the current IV User's operand. IVSrc is the IV
3221 // value currently held in a register.
3222 Value *IVOper = IVSrc;
3223 if (!Inc.IncExpr->isZero()) {
3224 // IncExpr was the result of subtraction of two narrow values, so must
3225 // be signed.
3226 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3227 LeftOverExpr = LeftOverExpr ?
3228 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3229 }
3230 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3231 // Expand the IV increment.
3232 Rewriter.clearPostInc();
3233 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3234 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3235 SE.getUnknown(IncV));
3236 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3237
3238 // If an IV increment can't be folded, use it as the next IV value.
3239 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3240 assert(IVTy == IVOper->getType() && "inconsistent IV increment type")(static_cast <bool> (IVTy == IVOper->getType() &&
"inconsistent IV increment type") ? void (0) : __assert_fail
("IVTy == IVOper->getType() && \"inconsistent IV increment type\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3240, __extension__ __PRETTY_FUNCTION__))
;
3241 IVSrc = IVOper;
3242 LeftOverExpr = nullptr;
3243 }
3244 }
3245 Type *OperTy = Inc.IVOperand->getType();
3246 if (IVTy != OperTy) {
3247 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&(static_cast <bool> (SE.getTypeSizeInBits(IVTy) >= SE
.getTypeSizeInBits(OperTy) && "cannot extend a chained IV"
) ? void (0) : __assert_fail ("SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && \"cannot extend a chained IV\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3248, __extension__ __PRETTY_FUNCTION__))
3248 "cannot extend a chained IV")(static_cast <bool> (SE.getTypeSizeInBits(IVTy) >= SE
.getTypeSizeInBits(OperTy) && "cannot extend a chained IV"
) ? void (0) : __assert_fail ("SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && \"cannot extend a chained IV\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3248, __extension__ __PRETTY_FUNCTION__))
;
3249 IRBuilder<> Builder(InsertPt);
3250 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3251 }
3252 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3253 if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3254 DeadInsts.emplace_back(OperandIsInstr);
3255 }
3256 // If LSR created a new, wider phi, we may also replace its postinc. We only
3257 // do this if we also found a wide value for the head of the chain.
3258 if (isa<PHINode>(Chain.tailUserInst())) {
3259 for (PHINode &Phi : L->getHeader()->phis()) {
3260 if (!isCompatibleIVType(&Phi, IVSrc))
3261 continue;
3262 Instruction *PostIncV = dyn_cast<Instruction>(
3263 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3264 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3265 continue;
3266 Value *IVOper = IVSrc;
3267 Type *PostIncTy = PostIncV->getType();
3268 if (IVTy != PostIncTy) {
3269 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types")(static_cast <bool> (PostIncTy->isPointerTy() &&
"mixing int/ptr IV types") ? void (0) : __assert_fail ("PostIncTy->isPointerTy() && \"mixing int/ptr IV types\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3269, __extension__ __PRETTY_FUNCTION__))
;
3270 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3271 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3272 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3273 }
3274 Phi.replaceUsesOfWith(PostIncV, IVOper);
3275 DeadInsts.emplace_back(PostIncV);
3276 }
3277 }
3278}
3279
3280void LSRInstance::CollectFixupsAndInitialFormulae() {
3281 BranchInst *ExitBranch = nullptr;
3282 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3283
3284 for (const IVStrideUse &U : IU) {
3285 Instruction *UserInst = U.getUser();
3286 // Skip IV users that are part of profitable IV Chains.
3287 User::op_iterator UseI =
3288 find(UserInst->operands(), U.getOperandValToReplace());
3289 assert(UseI != UserInst->op_end() && "cannot find IV operand")(static_cast <bool> (UseI != UserInst->op_end() &&
"cannot find IV operand") ? void (0) : __assert_fail ("UseI != UserInst->op_end() && \"cannot find IV operand\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3289, __extension__ __PRETTY_FUNCTION__))
;
3290 if (IVIncSet.count(UseI)) {
3291 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)
;
3292 continue;
3293 }
3294
3295 LSRUse::KindType Kind = LSRUse::Basic;
3296 MemAccessTy AccessTy;
3297 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3298 Kind = LSRUse::Address;
3299 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3300 }
3301
3302 const SCEV *S = IU.getExpr(U);
3303 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3304
3305 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3306 // (N - i == 0), and this allows (N - i) to be the expression that we work
3307 // with rather than just N or i, so we can consider the register
3308 // requirements for both N and i at the same time. Limiting this code to
3309 // equality icmps is not a problem because all interesting loops use
3310 // equality icmps, thanks to IndVarSimplify.
3311 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3312 // If CI can be saved in some target, like replaced inside hardware loop
3313 // in PowerPC, no need to generate initial formulae for it.
3314 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3315 continue;
3316 if (CI->isEquality()) {
3317 // Swap the operands if needed to put the OperandValToReplace on the
3318 // left, for consistency.
3319 Value *NV = CI->getOperand(1);
3320 if (NV == U.getOperandValToReplace()) {
3321 CI->setOperand(1, CI->getOperand(0));
3322 CI->setOperand(0, NV);
3323 NV = CI->getOperand(1);
3324 Changed = true;
3325 }
3326
3327 // x == y --> x - y == 0
3328 const SCEV *N = SE.getSCEV(NV);
3329 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE) &&
3330 (!NV->getType()->isPointerTy() ||
3331 SE.getPointerBase(N) == SE.getPointerBase(S))) {
3332 // S is normalized, so normalize N before folding it into S
3333 // to keep the result normalized.
3334 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3335 Kind = LSRUse::ICmpZero;
3336 S = SE.getMinusSCEV(N, S);
3337 }
3338
3339 // -1 and the negations of all interesting strides (except the negation
3340 // of -1) are now also interesting.
3341 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3342 if (Factors[i] != -1)
3343 Factors.insert(-(uint64_t)Factors[i]);
3344 Factors.insert(-1);
3345 }
3346 }
3347
3348 // Get or create an LSRUse.
3349 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3350 size_t LUIdx = P.first;
3351 int64_t Offset = P.second;
3352 LSRUse &LU = Uses[LUIdx];
3353
3354 // Record the fixup.
3355 LSRFixup &LF = LU.getNewFixup();
3356 LF.UserInst = UserInst;
3357 LF.OperandValToReplace = U.getOperandValToReplace();
3358 LF.PostIncLoops = TmpPostIncLoops;
3359 LF.Offset = Offset;
3360 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3361
3362 if (!LU.WidestFixupType ||
3363 SE.getTypeSizeInBits(LU.WidestFixupType) <
3364 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3365 LU.WidestFixupType = LF.OperandValToReplace->getType();
3366
3367 // If this is the first use of this LSRUse, give it a formula.
3368 if (LU.Formulae.empty()) {
3369 InsertInitialFormula(S, LU, LUIdx);
3370 CountRegisters(LU.Formulae.back(), LUIdx);
3371 }
3372 }
3373
3374 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3375}
3376
3377/// Insert a formula for the given expression into the given use, separating out
3378/// loop-variant portions from loop-invariant and loop-computable portions.
3379void
3380LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3381 // Mark uses whose expressions cannot be expanded.
3382 if (!isSafeToExpand(S, SE))
3383 LU.RigidFormula = true;
3384
3385 Formula F;
3386 F.initialMatch(S, L, SE);
3387 bool Inserted = InsertFormula(LU, LUIdx, F);
3388 assert(Inserted && "Initial formula already exists!")(static_cast <bool> (Inserted && "Initial formula already exists!"
) ? void (0) : __assert_fail ("Inserted && \"Initial formula already exists!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3388, __extension__ __PRETTY_FUNCTION__))
; (void)Inserted;
3389}
3390
3391/// Insert a simple single-register formula for the given expression into the
3392/// given use.
3393void
3394LSRInstance::InsertSupplementalFormula(const SCEV *S,
3395 LSRUse &LU, size_t LUIdx) {
3396 Formula F;
3397 F.BaseRegs.push_back(S);
3398 F.HasBaseReg = true;
3399 bool Inserted = InsertFormula(LU, LUIdx, F);
3400 assert(Inserted && "Supplemental formula already exists!")(static_cast <bool> (Inserted && "Supplemental formula already exists!"
) ? void (0) : __assert_fail ("Inserted && \"Supplemental formula already exists!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3400, __extension__ __PRETTY_FUNCTION__))
; (void)Inserted;
3401}
3402
3403/// Note which registers are used by the given formula, updating RegUses.
3404void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3405 if (F.ScaledReg)
3406 RegUses.countRegister(F.ScaledReg, LUIdx);
3407 for (const SCEV *BaseReg : F.BaseRegs)
3408 RegUses.countRegister(BaseReg, LUIdx);
3409}
3410
3411/// If the given formula has not yet been inserted, add it to the list, and
3412/// return true. Return false otherwise.
3413bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3414 // Do not insert formula that we will not be able to expand.
3415 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&(static_cast <bool> (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset
, LU.Kind, LU.AccessTy, F) && "Formula is illegal") ?
void (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Formula is illegal\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3416, __extension__ __PRETTY_FUNCTION__))
3416 "Formula is illegal")(static_cast <bool> (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset
, LU.Kind, LU.AccessTy, F) && "Formula is illegal") ?
void (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Formula is illegal\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3416, __extension__ __PRETTY_FUNCTION__))
;
3417
3418 if (!LU.InsertFormula(F, *L))
3419 return false;
3420
3421 CountRegisters(F, LUIdx);
3422 return true;
3423}
3424
3425/// Check for other uses of loop-invariant values which we're tracking. These
3426/// other uses will pin these values in registers, making them less profitable
3427/// for elimination.
3428/// TODO: This currently misses non-constant addrec step registers.
3429/// TODO: Should this give more weight to users inside the loop?
3430void
3431LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3432 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3433 SmallPtrSet<const SCEV *, 32> Visited;
3434
3435 while (!Worklist.empty()) {
3436 const SCEV *S = Worklist.pop_back_val();
3437
3438 // Don't process the same SCEV twice
3439 if (!Visited.insert(S).second)
3440 continue;
3441
3442 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3443 Worklist.append(N->op_begin(), N->op_end());
3444 else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
3445 Worklist.push_back(C->getOperand());
3446 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3447 Worklist.push_back(D->getLHS());
3448 Worklist.push_back(D->getRHS());
3449 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3450 const Value *V = US->getValue();
3451 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3452 // Look for instructions defined outside the loop.
3453 if (L->contains(Inst)) continue;
3454 } else if (isa<UndefValue>(V))
3455 // Undef doesn't have a live range, so it doesn't matter.
3456 continue;
3457 for (const Use &U : V->uses()) {
3458 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3459 // Ignore non-instructions.
3460 if (!UserInst)
3461 continue;
3462 // Don't bother if the instruction is an EHPad.
3463 if (UserInst->isEHPad())
3464 continue;
3465 // Ignore instructions in other functions (as can happen with
3466 // Constants).
3467 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3468 continue;
3469 // Ignore instructions not dominated by the loop.
3470 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3471 UserInst->getParent() :
3472 cast<PHINode>(UserInst)->getIncomingBlock(
3473 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3474 if (!DT.dominates(L->getHeader(), UseBB))
3475 continue;
3476 // Don't bother if the instruction is in a BB which ends in an EHPad.
3477 if (UseBB->getTerminator()->isEHPad())
3478 continue;
3479 // Don't bother rewriting PHIs in catchswitch blocks.
3480 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3481 continue;
3482 // Ignore uses which are part of other SCEV expressions, to avoid
3483 // analyzing them multiple times.
3484 if (SE.isSCEVable(UserInst->getType())) {
3485 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3486 // If the user is a no-op, look through to its uses.
3487 if (!isa<SCEVUnknown>(UserS))
3488 continue;
3489 if (UserS == US) {
3490 Worklist.push_back(
3491 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3492 continue;
3493 }
3494 }
3495 // Ignore icmp instructions which are already being analyzed.
3496 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3497 unsigned OtherIdx = !U.getOperandNo();
3498 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3499 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3500 continue;
3501 }
3502
3503 std::pair<size_t, int64_t> P = getUse(
3504 S, LSRUse::Basic, MemAccessTy());
3505 size_t LUIdx = P.first;
3506 int64_t Offset = P.second;
3507 LSRUse &LU = Uses[LUIdx];
3508 LSRFixup &LF = LU.getNewFixup();
3509 LF.UserInst = const_cast<Instruction *>(UserInst);
3510 LF.OperandValToReplace = U;
3511 LF.Offset = Offset;
3512 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3513 if (!LU.WidestFixupType ||
3514 SE.getTypeSizeInBits(LU.WidestFixupType) <
3515 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3516 LU.WidestFixupType = LF.OperandValToReplace->getType();
3517 InsertSupplementalFormula(US, LU, LUIdx);
3518 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3519 break;
3520 }
3521 }
3522 }
3523}
3524
3525/// Split S into subexpressions which can be pulled out into separate
3526/// registers. If C is non-null, multiply each subexpression by C.
3527///
3528/// Return remainder expression after factoring the subexpressions captured by
3529/// Ops. If Ops is complete, return NULL.
3530static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3531 SmallVectorImpl<const SCEV *> &Ops,
3532 const Loop *L,
3533 ScalarEvolution &SE,
3534 unsigned Depth = 0) {
3535 // Arbitrarily cap recursion to protect compile time.
3536 if (Depth >= 3)
3537 return S;
3538
3539 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3540 // Break out add operands.
3541 for (const SCEV *S : Add->operands()) {
3542 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3543 if (Remainder)
3544 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3545 }
3546 return nullptr;
3547 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3548 // Split a non-zero base out of an addrec.
3549 if (AR->getStart()->isZero() || !AR->isAffine())
3550 return S;
3551
3552 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3553 C, Ops, L, SE, Depth+1);
3554 // Split the non-zero AddRec unless it is part of a nested recurrence that
3555 // does not pertain to this loop.
3556 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3557 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3558 Remainder = nullptr;
3559 }
3560 if (Remainder != AR->getStart()) {
3561 if (!Remainder)
3562 Remainder = SE.getConstant(AR->getType(), 0);
3563 return SE.getAddRecExpr(Remainder,
3564 AR->getStepRecurrence(SE),
3565 AR->getLoop(),
3566 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3567 SCEV::FlagAnyWrap);
3568 }
3569 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3570 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3571 if (Mul->getNumOperands() != 2)
3572 return S;
3573 if (const SCEVConstant *Op0 =
3574 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3575 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3576 const SCEV *Remainder =
3577 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3578 if (Remainder)
3579 Ops.push_back(SE.getMulExpr(C, Remainder));
3580 return nullptr;
3581 }
3582 }
3583 return S;
3584}
3585
3586/// Return true if the SCEV represents a value that may end up as a
3587/// post-increment operation.
3588static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3589 LSRUse &LU, const SCEV *S, const Loop *L,
3590 ScalarEvolution &SE) {
3591 if (LU.Kind != LSRUse::Address ||
3592 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3593 return false;
3594 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3595 if (!AR)
3596 return false;
3597 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3598 if (!isa<SCEVConstant>(LoopStep))
3599 return false;
3600 // Check if a post-indexed load/store can be used.
3601 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3602 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3603 const SCEV *LoopStart = AR->getStart();
3604 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3605 return true;
3606 }
3607 return false;
3608}
3609
3610/// Helper function for LSRInstance::GenerateReassociations.
3611void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3612 const Formula &Base,
3613 unsigned Depth, size_t Idx,
3614 bool IsScaledReg) {
3615 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3616 // Don't generate reassociations for the base register of a value that
3617 // may generate a post-increment operator. The reason is that the
3618 // reassociations cause extra base+register formula to be created,
3619 // and possibly chosen, but the post-increment is more efficient.
3620 if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3621 return;
3622 SmallVector<const SCEV *, 8> AddOps;
3623 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3624 if (Remainder)
3625 AddOps.push_back(Remainder);
3626
3627 if (AddOps.size() == 1)
3628 return;
3629
3630 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3631 JE = AddOps.end();
3632 J != JE; ++J) {
3633 // Loop-variant "unknown" values are uninteresting; we won't be able to
3634 // do anything meaningful with them.
3635 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3636 continue;
3637
3638 // Don't pull a constant into a register if the constant could be folded
3639 // into an immediate field.
3640 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3641 LU.AccessTy, *J, Base.getNumRegs() > 1))
3642 continue;
3643
3644 // Collect all operands except *J.
3645 SmallVector<const SCEV *, 8> InnerAddOps(
3646 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3647 InnerAddOps.append(std::next(J),
3648 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3649
3650 // Don't leave just a constant behind in a register if the constant could
3651 // be folded into an immediate field.
3652 if (InnerAddOps.size() == 1 &&
3653 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3654 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3655 continue;
3656
3657 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3658 if (InnerSum->isZero())
3659 continue;
3660 Formula F = Base;
3661
3662 // Add the remaining pieces of the add back into the new formula.
3663 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3664 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3665 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3666 InnerSumSC->getValue()->getZExtValue())) {
3667 F.UnfoldedOffset =
3668 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3669 if (IsScaledReg)
3670 F.ScaledReg = nullptr;
3671 else
3672 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3673 } else if (IsScaledReg)
3674 F.ScaledReg = InnerSum;
3675 else
3676 F.BaseRegs[Idx] = InnerSum;
3677
3678 // Add J as its own register, or an unfolded immediate.
3679 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3680 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3681 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3682 SC->getValue()->getZExtValue()))
3683 F.UnfoldedOffset =
3684 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3685 else
3686 F.BaseRegs.push_back(*J);
3687 // We may have changed the number of register in base regs, adjust the
3688 // formula accordingly.
3689 F.canonicalize(*L);
3690
3691 if (InsertFormula(LU, LUIdx, F))
3692 // If that formula hadn't been seen before, recurse to find more like
3693 // it.
3694 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3695 // Because just Depth is not enough to bound compile time.
3696 // This means that every time AddOps.size() is greater 16^x we will add
3697 // x to Depth.
3698 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3699 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3700 }
3701}
3702
3703/// Split out subexpressions from adds and the bases of addrecs.
3704void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3705 Formula Base, unsigned Depth) {
3706 assert(Base.isCanonical(*L) && "Input must be in the canonical form")(static_cast <bool> (Base.isCanonical(*L) && "Input must be in the canonical form"
) ? void (0) : __assert_fail ("Base.isCanonical(*L) && \"Input must be in the canonical form\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3706, __extension__ __PRETTY_FUNCTION__))
;
3707 // Arbitrarily cap recursion to protect compile time.
3708 if (Depth >= 3)
3709 return;
3710
3711 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3712 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3713
3714 if (Base.Scale == 1)
3715 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3716 /* Idx */ -1, /* IsScaledReg */ true);
3717}
3718
3719/// Generate a formula consisting of all of the loop-dominating registers added
3720/// into a single register.
3721void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3722 Formula Base) {
3723 // This method is only interesting on a plurality of registers.
3724 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3725 (Base.UnfoldedOffset != 0) <= 1)
3726 return;
3727
3728 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3729 // processing the formula.
3730 Base.unscale();
3731 SmallVector<const SCEV *, 4> Ops;
3732 Formula NewBase = Base;
3733 NewBase.BaseRegs.clear();
3734 Type *CombinedIntegerType = nullptr;
3735 for (const SCEV *BaseReg : Base.BaseRegs) {
3736 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3737 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3738 if (!CombinedIntegerType)
3739 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3740 Ops.push_back(BaseReg);
3741 }
3742 else
3743 NewBase.BaseRegs.push_back(BaseReg);
3744 }
3745
3746 // If no register is relevant, we're done.
3747 if (Ops.size() == 0)
3748 return;
3749
3750 // Utility function for generating the required variants of the combined
3751 // registers.
3752 auto GenerateFormula = [&](const SCEV *Sum) {
3753 Formula F = NewBase;
3754
3755 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3756 // opportunity to fold something. For now, just ignore such cases
3757 // rather than proceed with zero in a register.
3758 if (Sum->isZero())
3759 return;
3760
3761 F.BaseRegs.push_back(Sum);
3762 F.canonicalize(*L);
3763 (void)InsertFormula(LU, LUIdx, F);
3764 };
3765
3766 // If we collected at least two registers, generate a formula combining them.
3767 if (Ops.size() > 1) {
3768 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3769 GenerateFormula(SE.getAddExpr(OpsCopy));
3770 }
3771
3772 // If we have an unfolded offset, generate a formula combining it with the
3773 // registers collected.
3774 if (NewBase.UnfoldedOffset) {
3775 assert(CombinedIntegerType && "Missing a type for the unfolded offset")(static_cast <bool> (CombinedIntegerType && "Missing a type for the unfolded offset"
) ? void (0) : __assert_fail ("CombinedIntegerType && \"Missing a type for the unfolded offset\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3775, __extension__ __PRETTY_FUNCTION__))
;
3776 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3777 true));
3778 NewBase.UnfoldedOffset = 0;
3779 GenerateFormula(SE.getAddExpr(Ops));
3780 }
3781}
3782
3783/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3784void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3785 const Formula &Base, size_t Idx,
3786 bool IsScaledReg) {
3787 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3788 GlobalValue *GV = ExtractSymbol(G, SE);
3789 if (G->isZero() || !GV)
3790 return;
3791 Formula F = Base;
3792 F.BaseGV = GV;
3793 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3794 return;
3795 if (IsScaledReg)
3796 F.ScaledReg = G;
3797 else
3798 F.BaseRegs[Idx] = G;
3799 (void)InsertFormula(LU, LUIdx, F);
3800}
3801
3802/// Generate reuse formulae using symbolic offsets.
3803void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3804 Formula Base) {
3805 // We can't add a symbolic offset if the address already contains one.
3806 if (Base.BaseGV) return;
3807
3808 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3809 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3810 if (Base.Scale == 1)
3811 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3812 /* IsScaledReg */ true);
3813}
3814
3815/// Helper function for LSRInstance::GenerateConstantOffsets.
3816void LSRInstance::GenerateConstantOffsetsImpl(
3817 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3818 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3819
3820 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3821 Formula F = Base;
3822 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3823
3824 if (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) {
3825 // Add the offset to the base register.
3826 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3827 // If it cancelled out, drop the base register, otherwise update it.
3828 if (NewG->isZero()) {
3829 if (IsScaledReg) {
3830 F.Scale = 0;
3831 F.ScaledReg = nullptr;
3832 } else
3833 F.deleteBaseReg(F.BaseRegs[Idx]);
3834 F.canonicalize(*L);
3835 } else if (IsScaledReg)
3836 F.ScaledReg = NewG;
3837 else
3838 F.BaseRegs[Idx] = NewG;
3839
3840 (void)InsertFormula(LU, LUIdx, F);
3841 }
3842 };
3843
3844 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3845
3846 // With constant offsets and constant steps, we can generate pre-inc
3847 // accesses by having the offset equal the step. So, for access #0 with a
3848 // step of 8, we generate a G - 8 base which would require the first access
3849 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3850 // for itself and hopefully becomes the base for other accesses. This means
3851 // means that a single pre-indexed access can be generated to become the new
3852 // base pointer for each iteration of the loop, resulting in no extra add/sub
3853 // instructions for pointer updating.
3854 if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) {
3855 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3856 if (auto *StepRec =
3857 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3858 const APInt &StepInt = StepRec->getAPInt();
3859 int64_t Step = StepInt.isNegative() ?
3860 StepInt.getSExtValue() : StepInt.getZExtValue();
3861
3862 for (int64_t Offset : Worklist) {
3863 Offset -= Step;
3864 GenerateOffset(G, Offset);
3865 }
3866 }
3867 }
3868 }
3869 for (int64_t Offset : Worklist)
3870 GenerateOffset(G, Offset);
3871
3872 int64_t Imm = ExtractImmediate(G, SE);
3873 if (G->isZero() || Imm == 0)
3874 return;
3875 Formula F = Base;
3876 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3877 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3878 return;
3879 if (IsScaledReg) {
3880 F.ScaledReg = G;
3881 } else {
3882 F.BaseRegs[Idx] = G;
3883 // We may generate non canonical Formula if G is a recurrent expr reg
3884 // related with current loop while F.ScaledReg is not.
3885 F.canonicalize(*L);
3886 }
3887 (void)InsertFormula(LU, LUIdx, F);
3888}
3889
3890/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3891void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3892 Formula Base) {
3893 // TODO: For now, just add the min and max offset, because it usually isn't
3894 // worthwhile looking at everything inbetween.
3895 SmallVector<int64_t, 2> Worklist;
3896 Worklist.push_back(LU.MinOffset);
3897 if (LU.MaxOffset != LU.MinOffset)
3898 Worklist.push_back(LU.MaxOffset);
3899
3900 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3901 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3902 if (Base.Scale == 1)
3903 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3904 /* IsScaledReg */ true);
3905}
3906
3907/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3908/// == y -> x*c == y*c.
3909void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3910 Formula Base) {
3911 if (LU.Kind != LSRUse::ICmpZero) return;
3912
3913 // Determine the integer type for the base formula.
3914 Type *IntTy = Base.getType();
3915 if (!IntTy) return;
3916 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3917
3918 // Don't do this if there is more than one offset.
3919 if (LU.MinOffset != LU.MaxOffset) return;
3920
3921 // Check if transformation is valid. It is illegal to multiply pointer.
3922 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3923 return;
3924 for (const SCEV *BaseReg : Base.BaseRegs)
3925 if (BaseReg->getType()->isPointerTy())
3926 return;
3927 assert(!Base.BaseGV && "ICmpZero use is not legal!")(static_cast <bool> (!Base.BaseGV && "ICmpZero use is not legal!"
) ? void (0) : __assert_fail ("!Base.BaseGV && \"ICmpZero use is not legal!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3927, __extension__ __PRETTY_FUNCTION__))
;
3928
3929 // Check each interesting stride.
3930 for (int64_t Factor : Factors) {
3931 // Check that the multiplication doesn't overflow.
3932 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3933 continue;
3934 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3935 assert(Factor != 0 && "Zero factor not expected!")(static_cast <bool> (Factor != 0 && "Zero factor not expected!"
) ? void (0) : __assert_fail ("Factor != 0 && \"Zero factor not expected!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 3935, __extension__ __PRETTY_FUNCTION__))
;
3936 if (NewBaseOffset / Factor != Base.BaseOffset)
3937 continue;
3938 // If the offset will be truncated at this use, check that it is in bounds.
3939 if (!IntTy->isPointerTy() &&
3940 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3941 continue;
3942
3943 // Check that multiplying with the use offset doesn't overflow.
3944 int64_t Offset = LU.MinOffset;
3945 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3946 continue;
3947 Offset = (uint64_t)Offset * Factor;
3948 if (Offset / Factor != LU.MinOffset)
3949 continue;
3950 // If the offset will be truncated at this use, check that it is in bounds.
3951 if (!IntTy->isPointerTy() &&
3952 !ConstantInt::isValueValidForType(IntTy, Offset))
3953 continue;
3954
3955 Formula F = Base;
3956 F.BaseOffset = NewBaseOffset;
3957
3958 // Check that this scale is legal.
3959 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3960 continue;
3961
3962 // Compensate for the use having MinOffset built into it.
3963 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3964
3965 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3966
3967 // Check that multiplying with each base register doesn't overflow.
3968 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3969 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3970 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3971 goto next;
3972 }
3973
3974 // Check that multiplying with the scaled register doesn't overflow.
3975 if (F.ScaledReg) {
3976 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3977 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3978 continue;
3979 }
3980
3981 // Check that multiplying with the unfolded offset doesn't overflow.
3982 if (F.UnfoldedOffset != 0) {
3983 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3984 Factor == -1)
3985 continue;
3986 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3987 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3988 continue;
3989 // If the offset will be truncated, check that it is in bounds.
3990 if (!IntTy->isPointerTy() &&
3991 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3992 continue;
3993 }
3994
3995 // If we make it here and it's legal, add it.
3996 (void)InsertFormula(LU, LUIdx, F);
3997 next:;
3998 }
3999}
4000
4001/// Generate stride factor reuse formulae by making use of scaled-offset address
4002/// modes, for example.
4003void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
4004 // Determine the integer type for the base formula.
4005 Type *IntTy = Base.getType();
4006 if (!IntTy) return;
4007
4008 // If this Formula already has a scaled register, we can't add another one.
4009 // Try to unscale the formula to generate a better scale.
4010 if (Base.Scale != 0 && !Base.unscale())
4011 return;
4012
4013 assert(Base.Scale == 0 && "unscale did not did its job!")(static_cast <bool> (Base.Scale == 0 && "unscale did not did its job!"
) ? void (0) : __assert_fail ("Base.Scale == 0 && \"unscale did not did its job!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4013, __extension__ __PRETTY_FUNCTION__))
;
4014
4015 // Check each interesting stride.
4016 for (int64_t Factor : Factors) {
4017 Base.Scale = Factor;
4018 Base.HasBaseReg = Base.BaseRegs.size() > 1;
4019 // Check whether this scale is going to be legal.
4020 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4021 Base)) {
4022 // As a special-case, handle special out-of-loop Basic users specially.
4023 // TODO: Reconsider this special case.
4024 if (LU.Kind == LSRUse::Basic &&
4025 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
4026 LU.AccessTy, Base) &&
4027 LU.AllFixupsOutsideLoop)
4028 LU.Kind = LSRUse::Special;
4029 else
4030 continue;
4031 }
4032 // For an ICmpZero, negating a solitary base register won't lead to
4033 // new solutions.
4034 if (LU.Kind == LSRUse::ICmpZero &&
4035 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
4036 continue;
4037 // For each addrec base reg, if its loop is current loop, apply the scale.
4038 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
4039 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
4040 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
4041 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4042 if (FactorS->isZero())
4043 continue;
4044 // Divide out the factor, ignoring high bits, since we'll be
4045 // scaling the value back up in the end.
4046 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
4047 // TODO: This could be optimized to avoid all the copying.
4048 Formula F = Base;
4049 F.ScaledReg = Quotient;
4050 F.deleteBaseReg(F.BaseRegs[i]);
4051 // The canonical representation of 1*reg is reg, which is already in
4052 // Base. In that case, do not try to insert the formula, it will be
4053 // rejected anyway.
4054 if (F.Scale == 1 && (F.BaseRegs.empty() ||
4055 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4056 continue;
4057 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4058 // non canonical Formula with ScaledReg's loop not being L.
4059 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4060 F.canonicalize(*L);
4061 (void)InsertFormula(LU, LUIdx, F);
4062 }
4063 }
4064 }
4065 }
4066}
4067
4068/// Generate reuse formulae from different IV types.
4069void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4070 // Don't bother truncating symbolic values.
4071 if (Base.BaseGV) return;
4072
4073 // Determine the integer type for the base formula.
4074 Type *DstTy = Base.getType();
4075 if (!DstTy) return;
4076 if (DstTy->isPointerTy())
4077 return;
4078
4079 for (Type *SrcTy : Types) {
4080 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4081 Formula F = Base;
4082
4083 // Sometimes SCEV is able to prove zero during ext transform. It may
4084 // happen if SCEV did not do all possible transforms while creating the
4085 // initial node (maybe due to depth limitations), but it can do them while
4086 // taking ext.
4087 if (F.ScaledReg) {
4088 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4089 if (NewScaledReg->isZero())
4090 continue;
4091 F.ScaledReg = NewScaledReg;
4092 }
4093 bool HasZeroBaseReg = false;
4094 for (const SCEV *&BaseReg : F.BaseRegs) {
4095 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4096 if (NewBaseReg->isZero()) {
4097 HasZeroBaseReg = true;
4098 break;
4099 }
4100 BaseReg = NewBaseReg;
4101 }
4102 if (HasZeroBaseReg)
4103 continue;
4104
4105 // TODO: This assumes we've done basic processing on all uses and
4106 // have an idea what the register usage is.
4107 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4108 continue;
4109
4110 F.canonicalize(*L);
4111 (void)InsertFormula(LU, LUIdx, F);
4112 }
4113 }
4114}
4115
4116namespace {
4117
4118/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4119/// modifications so that the search phase doesn't have to worry about the data
4120/// structures moving underneath it.
4121struct WorkItem {
4122 size_t LUIdx;
4123 int64_t Imm;
4124 const SCEV *OrigReg;
4125
4126 WorkItem(size_t LI, int64_t I, const SCEV *R)
4127 : LUIdx(LI), Imm(I), OrigReg(R) {}
4128
4129 void print(raw_ostream &OS) const;
4130 void dump() const;
4131};
4132
4133} // end anonymous namespace
4134
4135#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4136void WorkItem::print(raw_ostream &OS) const {
4137 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4138 << " , add offset " << Imm;
4139}
4140
4141LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4142 print(errs()); errs() << '\n';
4143}
4144#endif
4145
4146/// Look for registers which are a constant distance apart and try to form reuse
4147/// opportunities between them.
4148void LSRInstance::GenerateCrossUseConstantOffsets() {
4149 // Group the registers by their value without any added constant offset.
4150 using ImmMapTy = std::map<int64_t, const SCEV *>;
4151
4152 DenseMap<const SCEV *, ImmMapTy> Map;
4153 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4154 SmallVector<const SCEV *, 8> Sequence;
4155 for (const SCEV *Use : RegUses) {
4156 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4157 int64_t Imm = ExtractImmediate(Reg, SE);
4158 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4159 if (Pair.second)
4160 Sequence.push_back(Reg);
4161 Pair.first->second.insert(std::make_pair(Imm, Use));
4162 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4163 }
4164
4165 // Now examine each set of registers with the same base value. Build up
4166 // a list of work to do and do the work in a separate step so that we're
4167 // not adding formulae and register counts while we're searching.
4168 SmallVector<WorkItem, 32> WorkItems;
4169 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4170 for (const SCEV *Reg : Sequence) {
4171 const ImmMapTy &Imms = Map.find(Reg)->second;
4172
4173 // It's not worthwhile looking for reuse if there's only one offset.
4174 if (Imms.size() == 1)
4175 continue;
4176
4177 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)
4178 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)
4179 : 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)
4180 << ' ' << 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)
4181 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)
;
4182
4183 // Examine each offset.
4184 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4185 J != JE; ++J) {
4186 const SCEV *OrigReg = J->second;
4187
4188 int64_t JImm = J->first;
4189 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4190
4191 if (!isa<SCEVConstant>(OrigReg) &&
4192 UsedByIndicesMap[Reg].count() == 1) {
4193 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)
4194 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4195 continue;
4196 }
4197
4198 // Conservatively examine offsets between this orig reg a few selected
4199 // other orig regs.
4200 int64_t First = Imms.begin()->first;
4201 int64_t Last = std::prev(Imms.end())->first;
4202 // Compute (First + Last) / 2 without overflow using the fact that
4203 // First + Last = 2 * (First + Last) + (First ^ Last).
4204 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4205 // If the result is negative and First is odd and Last even (or vice versa),
4206 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4207 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4208 ImmMapTy::const_iterator OtherImms[] = {
4209 Imms.begin(), std::prev(Imms.end()),
4210 Imms.lower_bound(Avg)};
4211 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4212 ImmMapTy::const_iterator M = OtherImms[i];
4213 if (M == J || M == JE) continue;
4214
4215 // Compute the difference between the two.
4216 int64_t Imm = (uint64_t)JImm - M->first;
4217 for (unsigned LUIdx : UsedByIndices.set_bits())
4218 // Make a memo of this use, offset, and register tuple.
4219 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4220 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4221 }
4222 }
4223 }
4224
4225 Map.clear();
4226 Sequence.clear();
4227 UsedByIndicesMap.clear();
4228 UniqueItems.clear();
4229
4230 // Now iterate through the worklist and add new formulae.
4231 for (const WorkItem &WI : WorkItems) {
4232 size_t LUIdx = WI.LUIdx;
4233 LSRUse &LU = Uses[LUIdx];
4234 int64_t Imm = WI.Imm;
4235 const SCEV *OrigReg = WI.OrigReg;
4236
4237 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4238 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4239 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4240
4241 // TODO: Use a more targeted data structure.
4242 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4243 Formula F = LU.Formulae[L];
4244 // FIXME: The code for the scaled and unscaled registers looks
4245 // very similar but slightly different. Investigate if they
4246 // could be merged. That way, we would not have to unscale the
4247 // Formula.
4248 F.unscale();
4249 // Use the immediate in the scaled register.
4250 if (F.ScaledReg == OrigReg) {
4251 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4252 // Don't create 50 + reg(-50).
4253 if (F.referencesReg(SE.getSCEV(
4254 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4255 continue;
4256 Formula NewF = F;
4257 NewF.BaseOffset = Offset;
4258 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4259 NewF))
4260 continue;
4261 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4262
4263 // If the new scale is a constant in a register, and adding the constant
4264 // value to the immediate would produce a value closer to zero than the
4265 // immediate itself, then the formula isn't worthwhile.
4266 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4267 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4268 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4269 .ule(std::abs(NewF.BaseOffset)))
4270 continue;
4271
4272 // OK, looks good.
4273 NewF.canonicalize(*this->L);
4274 (void)InsertFormula(LU, LUIdx, NewF);
4275 } else {
4276 // Use the immediate in a base register.
4277 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4278 const SCEV *BaseReg = F.BaseRegs[N];
4279 if (BaseReg != OrigReg)
4280 continue;
4281 Formula NewF = F;
4282 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4283 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4284 LU.Kind, LU.AccessTy, NewF)) {
4285 if (AMK == TTI::AMK_PostIndexed &&
4286 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4287 continue;
4288 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4289 continue;
4290 NewF = F;
4291 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4292 }
4293 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4294
4295 // If the new formula has a constant in a register, and adding the
4296 // constant value to the immediate would produce a value closer to
4297 // zero than the immediate itself, then the formula isn't worthwhile.
4298 for (const SCEV *NewReg : NewF.BaseRegs)
4299 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4300 if ((C->getAPInt() + NewF.BaseOffset)
4301 .abs()
4302 .slt(std::abs(NewF.BaseOffset)) &&
4303 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4304 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4305 goto skip_formula;
4306
4307 // Ok, looks good.
4308 NewF.canonicalize(*this->L);
4309 (void)InsertFormula(LU, LUIdx, NewF);
4310 break;
4311 skip_formula:;
4312 }
4313 }
4314 }
4315 }
4316}
4317
4318/// Generate formulae for each use.
4319void
4320LSRInstance::GenerateAllReuseFormulae() {
4321 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4322 // queries are more precise.
4323 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4324 LSRUse &LU = Uses[LUIdx];
4325 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4326 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4327 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4328 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4329 }
4330 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4331 LSRUse &LU = Uses[LUIdx];
4332 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4333 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4334 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4335 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4336 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4337 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4338 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4339 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4340 }
4341 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4342 LSRUse &LU = Uses[LUIdx];
4343 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4344 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4345 }
4346
4347 GenerateCrossUseConstantOffsets();
4348
4349 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4350 "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)
4351 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4352}
4353
4354/// If there are multiple formulae with the same set of registers used
4355/// by other uses, pick the best one and delete the others.
4356void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4357 DenseSet<const SCEV *> VisitedRegs;
4358 SmallPtrSet<const SCEV *, 16> Regs;
4359 SmallPtrSet<const SCEV *, 16> LoserRegs;
4360#ifndef NDEBUG
4361 bool ChangedFormulae = false;
4362#endif
4363
4364 // Collect the best formula for each unique set of shared registers. This
4365 // is reset for each use.
4366 using BestFormulaeTy =
4367 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4368
4369 BestFormulaeTy BestFormulae;
4370
4371 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4372 LSRUse &LU = Uses[LUIdx];
4373 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)
4374 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4375
4376 bool Any = false;
4377 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4378 FIdx != NumForms; ++FIdx) {
4379 Formula &F = LU.Formulae[FIdx];
4380
4381 // Some formulas are instant losers. For example, they may depend on
4382 // nonexistent AddRecs from other loops. These need to be filtered
4383 // immediately, otherwise heuristics could choose them over others leading
4384 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4385 // avoids the need to recompute this information across formulae using the
4386 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4387 // the corresponding bad register from the Regs set.
4388 Cost CostF(L, SE, TTI, AMK);
4389 Regs.clear();
4390 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4391 if (CostF.isLoser()) {
4392 // During initial formula generation, undesirable formulae are generated
4393 // by uses within other loops that have some non-trivial address mode or
4394 // use the postinc form of the IV. LSR needs to provide these formulae
4395 // as the basis of rediscovering the desired formula that uses an AddRec
4396 // corresponding to the existing phi. Once all formulae have been
4397 // generated, these initial losers may be pruned.
4398 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)
4399 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4400 }
4401 else {
4402 SmallVector<const SCEV *, 4> Key;
4403 for (const SCEV *Reg : F.BaseRegs) {
4404 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4405 Key.push_back(Reg);
4406 }
4407 if (F.ScaledReg &&
4408 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4409 Key.push_back(F.ScaledReg);
4410 // Unstable sort by host order ok, because this is only used for
4411 // uniquifying.
4412 llvm::sort(Key);
4413
4414 std::pair<BestFormulaeTy::const_iterator, bool> P =
4415 BestFormulae.insert(std::make_pair(Key, FIdx));
4416 if (P.second)
4417 continue;
4418
4419 Formula &Best = LU.Formulae[P.first->second];
4420
4421 Cost CostBest(L, SE, TTI, AMK);
4422 Regs.clear();
4423 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4424 if (CostF.isLess(CostBest))
4425 std::swap(F, Best);
4426 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)
4427 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)
4428 " 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)
4429 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)
;
4430 }
4431#ifndef NDEBUG
4432 ChangedFormulae = true;
4433#endif
4434 LU.DeleteFormula(F);
4435 --FIdx;
4436 --NumForms;
4437 Any = true;
4438 }
4439
4440 // Now that we've filtered out some formulae, recompute the Regs set.
4441 if (Any)
4442 LU.RecomputeRegs(LUIdx, RegUses);
4443
4444 // Reset this to prepare for the next use.
4445 BestFormulae.clear();
4446 }
4447
4448 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)
4449 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)
4450 "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)
4451 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)
4452 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4453}
4454
4455/// Estimate the worst-case number of solutions the solver might have to
4456/// consider. It almost never considers this many solutions because it prune the
4457/// search space, but the pruning isn't always sufficient.
4458size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4459 size_t Power = 1;
4460 for (const LSRUse &LU : Uses) {
4461 size_t FSize = LU.Formulae.size();
4462 if (FSize >= ComplexityLimit) {
4463 Power = ComplexityLimit;
4464 break;
4465 }
4466 Power *= FSize;
4467 if (Power >= ComplexityLimit)
4468 break;
4469 }
4470 return Power;
4471}
4472
4473/// When one formula uses a superset of the registers of another formula, it
4474/// won't help reduce register pressure (though it may not necessarily hurt
4475/// register pressure); remove it to simplify the system.
4476void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4477 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4478 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)
;
4479
4480 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)
4481 "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)
4482 "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)
;
4483
4484 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4485 LSRUse &LU = Uses[LUIdx];
4486 bool Any = false;
4487 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4488 Formula &F = LU.Formulae[i];
4489 // Look for a formula with a constant or GV in a register. If the use
4490 // also has a formula with that same value in an immediate field,
4491 // delete the one that uses a register.
4492 for (SmallVectorImpl<const SCEV *>::const_iterator
4493 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4494 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4495 Formula NewF = F;
4496 //FIXME: Formulas should store bitwidth to do wrapping properly.
4497 // See PR41034.
4498 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4499 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4500 (I - F.BaseRegs.begin()));
4501 if (LU.HasFormulaWithSameRegs(NewF)) {
4502 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4503 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4504 LU.DeleteFormula(F);
4505 --i;
4506 --e;
4507 Any = true;
4508 break;
4509 }
4510 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4511 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4512 if (!F.BaseGV) {
4513 Formula NewF = F;
4514 NewF.BaseGV = GV;
4515 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4516 (I - F.BaseRegs.begin()));
4517 if (LU.HasFormulaWithSameRegs(NewF)) {
4518 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4519 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4520 LU.DeleteFormula(F);
4521 --i;
4522 --e;
4523 Any = true;
4524 break;
4525 }
4526 }
4527 }
4528 }
4529 }
4530 if (Any)
4531 LU.RecomputeRegs(LUIdx, RegUses);
4532 }
4533
4534 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)
;
4535 }
4536}
4537
4538/// When there are many registers for expressions like A, A+1, A+2, etc.,
4539/// allocate a single register for them.
4540void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4541 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4542 return;
4543
4544 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)
4545 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)
4546 "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)
4547 "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)
;
4548
4549 // This is especially useful for unrolled loops.
4550
4551 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4552 LSRUse &LU = Uses[LUIdx];
4553 for (const Formula &F : LU.Formulae) {
4554 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4555 continue;
4556
4557 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4558 if (!LUThatHas)
4559 continue;
4560
4561 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4562 LU.Kind, LU.AccessTy))
4563 continue;
4564
4565 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)
;
4566
4567 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4568
4569 // Transfer the fixups of LU to LUThatHas.
4570 for (LSRFixup &Fixup : LU.Fixups) {
4571 Fixup.Offset += F.BaseOffset;
4572 LUThatHas->pushFixup(Fixup);
4573 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)
;
4574 }
4575
4576 // Delete formulae from the new use which are no longer legal.
4577 bool Any = false;
4578 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4579 Formula &F = LUThatHas->Formulae[i];
4580 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4581 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4582 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)
;
4583 LUThatHas->DeleteFormula(F);
4584 --i;
4585 --e;
4586 Any = true;
4587 }
4588 }
4589
4590 if (Any)
4591 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4592
4593 // Delete the old use.
4594 DeleteUse(LU, LUIdx);
4595 --LUIdx;
4596 --NumUses;
4597 break;
4598 }
4599 }
4600
4601 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)
;
4602}
4603
4604/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4605/// we've done more filtering, as it may be able to find more formulae to
4606/// eliminate.
4607void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4608 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4609 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)
;
4610
4611 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)
4612 "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)
;
4613
4614 FilterOutUndesirableDedicatedRegisters();
4615
4616 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)
;
4617 }
4618}
4619
4620/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4621/// Pick the best one and delete the others.
4622/// This narrowing heuristic is to keep as many formulae with different
4623/// Scale and ScaledReg pair as possible while narrowing the search space.
4624/// The benefit is that it is more likely to find out a better solution
4625/// from a formulae set with more Scale and ScaledReg variations than
4626/// a formulae set with the same Scale and ScaledReg. The picking winner
4627/// reg heuristic will often keep the formulae with the same Scale and
4628/// ScaledReg and filter others, and we want to avoid that if possible.
4629void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4630 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4631 return;
4632
4633 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)
4634 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)
4635 "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)
4636 "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)
;
4637
4638 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4639 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4640
4641 BestFormulaeTy BestFormulae;
4642#ifndef NDEBUG
4643 bool ChangedFormulae = false;
4644#endif
4645 DenseSet<const SCEV *> VisitedRegs;
4646 SmallPtrSet<const SCEV *, 16> Regs;
4647
4648 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4649 LSRUse &LU = Uses[LUIdx];
4650 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)
4651 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4652
4653 // Return true if Formula FA is better than Formula FB.
4654 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4655 // First we will try to choose the Formula with fewer new registers.
4656 // For a register used by current Formula, the more the register is
4657 // shared among LSRUses, the less we increase the register number
4658 // counter of the formula.
4659 size_t FARegNum = 0;
4660 for (const SCEV *Reg : FA.BaseRegs) {
4661 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4662 FARegNum += (NumUses - UsedByIndices.count() + 1);
4663 }
4664 size_t FBRegNum = 0;
4665 for (const SCEV *Reg : FB.BaseRegs) {
4666 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4667 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4668 }
4669 if (FARegNum != FBRegNum)
4670 return FARegNum < FBRegNum;
4671
4672 // If the new register numbers are the same, choose the Formula with
4673 // less Cost.
4674 Cost CostFA(L, SE, TTI, AMK);
4675 Cost CostFB(L, SE, TTI, AMK);
4676 Regs.clear();
4677 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4678 Regs.clear();
4679 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4680 return CostFA.isLess(CostFB);
4681 };
4682
4683 bool Any = false;
4684 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4685 ++FIdx) {
4686 Formula &F = LU.Formulae[FIdx];
4687 if (!F.ScaledReg)
4688 continue;
4689 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4690 if (P.second)
4691 continue;
4692
4693 Formula &Best = LU.Formulae[P.first->second];
4694 if (IsBetterThan(F, Best))
4695 std::swap(F, Best);
4696 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)
4697 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)
4698 " 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)
4699 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)
;
4700#ifndef NDEBUG
4701 ChangedFormulae = true;
4702#endif
4703 LU.DeleteFormula(F);
4704 --FIdx;
4705 --NumForms;
4706 Any = true;
4707 }
4708 if (Any)
4709 LU.RecomputeRegs(LUIdx, RegUses);
4710
4711 // Reset this to prepare for the next use.
4712 BestFormulae.clear();
4713 }
4714
4715 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)
4716 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)
4717 "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)
4718 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)
4719 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4720}
4721
4722/// If we are over the complexity limit, filter out any post-inc prefering
4723/// variables to only post-inc values.
4724void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
4725 if (AMK != TTI::AMK_PostIndexed)
4726 return;
4727 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4728 return;
4729
4730 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)
4731 "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)
4732 "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)
;
4733
4734 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4735 LSRUse &LU = Uses[LUIdx];
4736
4737 if (LU.Kind != LSRUse::Address)
4738 continue;
4739 if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
4740 !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
4741 continue;
4742
4743 size_t MinRegs = std::numeric_limits<size_t>::max();
4744 for (const Formula &F : LU.Formulae)
4745 MinRegs = std::min(F.getNumRegs(), MinRegs);
4746
4747 bool Any = false;
4748 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4749 ++FIdx) {
4750 Formula &F = LU.Formulae[FIdx];
4751 if (F.getNumRegs() > MinRegs) {
4752 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)
4753 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n"; } } while (false)
;
4754 LU.DeleteFormula(F);
4755 --FIdx;
4756 --NumForms;
4757 Any = true;
4758 }
4759 }
4760 if (Any)
4761 LU.RecomputeRegs(LUIdx, RegUses);
4762
4763 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4764 break;
4765 }
4766
4767 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)
;
4768}
4769
4770/// The function delete formulas with high registers number expectation.
4771/// Assuming we don't know the value of each formula (already delete
4772/// all inefficient), generate probability of not selecting for each
4773/// register.
4774/// For example,
4775/// Use1:
4776/// reg(a) + reg({0,+,1})
4777/// reg(a) + reg({-1,+,1}) + 1
4778/// reg({a,+,1})
4779/// Use2:
4780/// reg(b) + reg({0,+,1})
4781/// reg(b) + reg({-1,+,1}) + 1
4782/// reg({b,+,1})
4783/// Use3:
4784/// reg(c) + reg(b) + reg({0,+,1})
4785/// reg(c) + reg({b,+,1})
4786///
4787/// Probability of not selecting
4788/// Use1 Use2 Use3
4789/// reg(a) (1/3) * 1 * 1
4790/// reg(b) 1 * (1/3) * (1/2)
4791/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4792/// reg({-1,+,1}) (2/3) * (2/3) * 1
4793/// reg({a,+,1}) (2/3) * 1 * 1
4794/// reg({b,+,1}) 1 * (2/3) * (2/3)
4795/// reg(c) 1 * 1 * 0
4796///
4797/// Now count registers number mathematical expectation for each formula:
4798/// Note that for each use we exclude probability if not selecting for the use.
4799/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4800/// probabilty 1/3 of not selecting for Use1).
4801/// Use1:
4802/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4803/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4804/// reg({a,+,1}) 1
4805/// Use2:
4806/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4807/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4808/// reg({b,+,1}) 2/3
4809/// Use3:
4810/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4811/// reg(c) + reg({b,+,1}) 1 + 2/3
4812void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4813 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4814 return;
4815 // Ok, we have too many of formulae on our hands to conveniently handle.
4816 // Use a rough heuristic to thin out the list.
4817
4818 // Set of Regs wich will be 100% used in final solution.
4819 // Used in each formula of a solution (in example above this is reg(c)).
4820 // We can skip them in calculations.
4821 SmallPtrSet<const SCEV *, 4> UniqRegs;
4822 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)
;
4823
4824 // Map each register to probability of not selecting
4825 DenseMap <const SCEV *, float> RegNumMap;
4826 for (const SCEV *Reg : RegUses) {
4827 if (UniqRegs.count(Reg))
4828 continue;
4829 float PNotSel = 1;
4830 for (const LSRUse &LU : Uses) {
4831 if (!LU.Regs.count(Reg))
4832 continue;
4833 float P = LU.getNotSelectedProbability(Reg);
4834 if (P != 0.0)
4835 PNotSel *= P;
4836 else
4837 UniqRegs.insert(Reg);
4838 }
4839 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4840 }
4841
4842 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4843 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)
;
4844
4845 // Delete formulas where registers number expectation is high.
4846 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4847 LSRUse &LU = Uses[LUIdx];
4848 // If nothing to delete - continue.
4849 if (LU.Formulae.size() < 2)
4850 continue;
4851 // This is temporary solution to test performance. Float should be
4852 // replaced with round independent type (based on integers) to avoid
4853 // different results for different target builds.
4854 float FMinRegNum = LU.Formulae[0].getNumRegs();
4855 float FMinARegNum = LU.Formulae[0].getNumRegs();
4856 size_t MinIdx = 0;
4857 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4858 Formula &F = LU.Formulae[i];
4859 float FRegNum = 0;
4860 float FARegNum = 0;
4861 for (const SCEV *BaseReg : F.BaseRegs) {
4862 if (UniqRegs.count(BaseReg))
4863 continue;
4864 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4865 if (isa<SCEVAddRecExpr>(BaseReg))
4866 FARegNum +=
4867 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4868 }
4869 if (const SCEV *ScaledReg = F.ScaledReg) {
4870 if (!UniqRegs.count(ScaledReg)) {
4871 FRegNum +=
4872 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4873 if (isa<SCEVAddRecExpr>(ScaledReg))
4874 FARegNum +=
4875 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4876 }
4877 }
4878 if (FMinRegNum > FRegNum ||
4879 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4880 FMinRegNum = FRegNum;
4881 FMinARegNum = FARegNum;
4882 MinIdx = i;
4883 }
4884 }
4885 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)
4886 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)
;
4887 if (MinIdx != 0)
4888 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4889 while (LU.Formulae.size() != 1) {
4890 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
)
4891 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4892 LU.Formulae.pop_back();
4893 }
4894 LU.RecomputeRegs(LUIdx, RegUses);
4895 assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula")(static_cast <bool> (LU.Formulae.size() == 1 &&
"Should be exactly 1 min regs formula") ? void (0) : __assert_fail
("LU.Formulae.size() == 1 && \"Should be exactly 1 min regs formula\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4895, __extension__ __PRETTY_FUNCTION__))
;
4896 Formula &F = LU.Formulae[0];
4897 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)
;
4898 // When we choose the formula, the regs become unique.
4899 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4900 if (F.ScaledReg)
4901 UniqRegs.insert(F.ScaledReg);
4902 }
4903 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)
;
4904}
4905
4906/// Pick a register which seems likely to be profitable, and then in any use
4907/// which has any reference to that register, delete all formulae which do not
4908/// reference that register.
4909void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4910 // With all other options exhausted, loop until the system is simple
4911 // enough to handle.
4912 SmallPtrSet<const SCEV *, 4> Taken;
4913 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4914 // Ok, we have too many of formulae on our hands to conveniently handle.
4915 // Use a rough heuristic to thin out the list.
4916 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)
;
4917
4918 // Pick the register which is used by the most LSRUses, which is likely
4919 // to be a good reuse register candidate.
4920 const SCEV *Best = nullptr;
4921 unsigned BestNum = 0;
4922 for (const SCEV *Reg : RegUses) {
4923 if (Taken.count(Reg))
4924 continue;
4925 if (!Best) {
4926 Best = Reg;
4927 BestNum = RegUses.getUsedByIndices(Reg).count();
4928 } else {
4929 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4930 if (Count > BestNum) {
4931 Best = Reg;
4932 BestNum = Count;
4933 }
4934 }
4935 }
4936 assert(Best && "Failed to find best LSRUse candidate")(static_cast <bool> (Best && "Failed to find best LSRUse candidate"
) ? void (0) : __assert_fail ("Best && \"Failed to find best LSRUse candidate\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4936, __extension__ __PRETTY_FUNCTION__))
;
4937
4938 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)
4939 << " 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)
;
4940 Taken.insert(Best);
4941
4942 // In any use with formulae which references this register, delete formulae
4943 // which don't reference it.
4944 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4945 LSRUse &LU = Uses[LUIdx];
4946 if (!LU.Regs.count(Best)) continue;
4947
4948 bool Any = false;
4949 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4950 Formula &F = LU.Formulae[i];
4951 if (!F.referencesReg(Best)) {
4952 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)
;
4953 LU.DeleteFormula(F);
4954 --e;
4955 --i;
4956 Any = true;
4957 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?")(static_cast <bool> (e != 0 && "Use has no formulae left! Is Regs inconsistent?"
) ? void (0) : __assert_fail ("e != 0 && \"Use has no formulae left! Is Regs inconsistent?\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 4957, __extension__ __PRETTY_FUNCTION__))
;
4958 continue;
4959 }
4960 }
4961
4962 if (Any)
4963 LU.RecomputeRegs(LUIdx, RegUses);
4964 }
4965
4966 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)
;
4967 }
4968}
4969
4970/// If there are an extraordinary number of formulae to choose from, use some
4971/// rough heuristics to prune down the number of formulae. This keeps the main
4972/// solver from taking an extraordinary amount of time in some worst-case
4973/// scenarios.
4974void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4975 NarrowSearchSpaceByDetectingSupersets();
4976 NarrowSearchSpaceByCollapsingUnrolledCode();
4977 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4978 if (FilterSameScaledReg)
4979 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4980 NarrowSearchSpaceByFilterPostInc();
4981 if (LSRExpNarrow)
4982 NarrowSearchSpaceByDeletingCostlyFormulas();
4983 else
4984 NarrowSearchSpaceByPickingWinnerRegs();
4985}
4986
4987/// This is the recursive solver.
4988void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4989 Cost &SolutionCost,
4990 SmallVectorImpl<const Formula *> &Workspace,
4991 const Cost &CurCost,
4992 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4993 DenseSet<const SCEV *> &VisitedRegs) const {
4994 // Some ideas:
4995 // - prune more:
4996 // - use more aggressive filtering
4997 // - sort the formula so that the most profitable solutions are found first
4998 // - sort the uses too
4999 // - search faster:
5000 // - don't compute a cost, and then compare. compare while computing a cost
5001 // and bail early.
5002 // - track register sets with SmallBitVector
5003
5004 const LSRUse &LU = Uses[Workspace.size()];
5005
5006 // If this use references any register that's already a part of the
5007 // in-progress solution, consider it a requirement that a formula must
5008 // reference that register in order to be considered. This prunes out
5009 // unprofitable searching.
5010 SmallSetVector<const SCEV *, 4> ReqRegs;
5011 for (const SCEV *S : CurRegs)
5012 if (LU.Regs.count(S))
5013 ReqRegs.insert(S);
5014
5015 SmallPtrSet<const SCEV *, 16> NewRegs;
5016 Cost NewCost(L, SE, TTI, AMK);
5017 for (const Formula &F : LU.Formulae) {
5018 // Ignore formulae which may not be ideal in terms of register reuse of
5019 // ReqRegs. The formula should use all required registers before
5020 // introducing new ones.
5021 // This can sometimes (notably when trying to favour postinc) lead to
5022 // sub-optimial decisions. There it is best left to the cost modelling to
5023 // get correct.
5024 if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) {
5025 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
5026 for (const SCEV *Reg : ReqRegs) {
5027 if ((F.ScaledReg && F.ScaledReg == Reg) ||
5028 is_contained(F.BaseRegs, Reg)) {
5029 --NumReqRegsToFind;
5030 if (NumReqRegsToFind == 0)
5031 break;
5032 }
5033 }
5034 if (NumReqRegsToFind != 0) {
5035 // If none of the formulae satisfied the required registers, then we could
5036 // clear ReqRegs and try again. Currently, we simply give up in this case.
5037 continue;
5038 }
5039 }
5040
5041 // Evaluate the cost of the current formula. If it's already worse than
5042 // the current best, prune the search at that point.
5043 NewCost = CurCost;
5044 NewRegs = CurRegs;
5045 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
5046 if (NewCost.isLess(SolutionCost)) {
5047 Workspace.push_back(&F);
5048 if (Workspace.size() != Uses.size()) {
5049 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
5050 NewRegs, VisitedRegs);
5051 if (F.getNumRegs() == 1 && Workspace.size() == 1)
5052 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
5053 } else {
5054 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)
5055 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)
5056 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)
5057 << "- " << *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)
5058 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)
;
5059
5060 SolutionCost = NewCost;
5061 Solution = Workspace;
5062 }
5063 Workspace.pop_back();
5064 }
5065 }
5066}
5067
5068/// Choose one formula from each use. Return the results in the given Solution
5069/// vector.
5070void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
5071 SmallVector<const Formula *, 8> Workspace;
5072 Cost SolutionCost(L, SE, TTI, AMK);
5073 SolutionCost.Lose();
5074 Cost CurCost(L, SE, TTI, AMK);
5075 SmallPtrSet<const SCEV *, 16> CurRegs;
5076 DenseSet<const SCEV *> VisitedRegs;
5077 Workspace.reserve(Uses.size());
5078
5079 // SolveRecurse does all the work.
5080 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
5081 CurRegs, VisitedRegs);
5082 if (Solution.empty()) {
5083 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
5084 return;
5085 }
5086
5087 // Ok, we've now made all our decisions.
5088 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
)
5089 "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
)
5090 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
)
5091 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
)
5092 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
)
5093 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
)
5094 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
)
5095 " ";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
)
5096 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
)
5097 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
)
5098 })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
)
;
5099
5100 assert(Solution.size() == Uses.size() && "Malformed solution!")(static_cast <bool> (Solution.size() == Uses.size() &&
"Malformed solution!") ? void (0) : __assert_fail ("Solution.size() == Uses.size() && \"Malformed solution!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5100, __extension__ __PRETTY_FUNCTION__))
;
5101}
5102
5103/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5104/// we can go while still being dominated by the input positions. This helps
5105/// canonicalize the insert position, which encourages sharing.
5106BasicBlock::iterator
5107LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5108 const SmallVectorImpl<Instruction *> &Inputs)
5109 const {
5110 Instruction *Tentative = &*IP;
5111 while (true) {
5112 bool AllDominate = true;
5113 Instruction *BetterPos = nullptr;
5114 // Don't bother attempting to insert before a catchswitch, their basic block
5115 // cannot have other non-PHI instructions.
5116 if (isa<CatchSwitchInst>(Tentative))
5117 return IP;
5118
5119 for (Instruction *Inst : Inputs) {
5120 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5121 AllDominate = false;
5122 break;
5123 }
5124 // Attempt to find an insert position in the middle of the block,
5125 // instead of at the end, so that it can be used for other expansions.
5126 if (Tentative->getParent() == Inst->getParent() &&
5127 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5128 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5129 }
5130 if (!AllDominate)
5131 break;
5132 if (BetterPos)
5133 IP = BetterPos->getIterator();
5134 else
5135 IP = Tentative->getIterator();
5136
5137 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5138 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5139
5140 BasicBlock *IDom;
5141 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5142 if (!Rung) return IP;
5143 Rung = Rung->getIDom();
5144 if (!Rung) return IP;
5145 IDom = Rung->getBlock();
5146
5147 // Don't climb into a loop though.
5148 const Loop *IDomLoop = LI.getLoopFor(IDom);
5149 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5150 if (IDomDepth <= IPLoopDepth &&
5151 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5152 break;
5153 }
5154
5155 Tentative = IDom->getTerminator();
5156 }
5157
5158 return IP;
5159}
5160
5161/// Determine an input position which will be dominated by the operands and
5162/// which will dominate the result.
5163BasicBlock::iterator
5164LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5165 const LSRFixup &LF,
5166 const LSRUse &LU,
5167 SCEVExpander &Rewriter) const {
5168 // Collect some instructions which must be dominated by the
5169 // expanding replacement. These must be dominated by any operands that
5170 // will be required in the expansion.
5171 SmallVector<Instruction *, 4> Inputs;
5172 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5173 Inputs.push_back(I);
5174 if (LU.Kind == LSRUse::ICmpZero)
5175 if (Instruction *I =
5176 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5177 Inputs.push_back(I);
5178 if (LF.PostIncLoops.count(L)) {
5179 if (LF.isUseFullyOutsideLoop(L))
5180 Inputs.push_back(L->getLoopLatch()->getTerminator());
5181 else
5182 Inputs.push_back(IVIncInsertPos);
5183 }
5184 // The expansion must also be dominated by the increment positions of any
5185 // loops it for which it is using post-inc mode.
5186 for (const Loop *PIL : LF.PostIncLoops) {
5187 if (PIL == L) continue;
5188
5189 // Be dominated by the loop exit.
5190 SmallVector<BasicBlock *, 4> ExitingBlocks;
5191 PIL->getExitingBlocks(ExitingBlocks);
5192 if (!ExitingBlocks.empty()) {
5193 BasicBlock *BB = ExitingBlocks[0];
5194 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5195 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5196 Inputs.push_back(BB->getTerminator());
5197 }
5198 }
5199
5200 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()(static_cast <bool> (!isa<PHINode>(LowestIP) &&
!LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>
(LowestIP) && "Insertion point must be a normal instruction"
) ? void (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5202, __extension__ __PRETTY_FUNCTION__))
5201 && !isa<DbgInfoIntrinsic>(LowestIP) &&(static_cast <bool> (!isa<PHINode>(LowestIP) &&
!LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>
(LowestIP) && "Insertion point must be a normal instruction"
) ? void (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5202, __extension__ __PRETTY_FUNCTION__))
5202 "Insertion point must be a normal instruction")(static_cast <bool> (!isa<PHINode>(LowestIP) &&
!LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>
(LowestIP) && "Insertion point must be a normal instruction"
) ? void (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5202, __extension__ __PRETTY_FUNCTION__))
;
5203
5204 // Then, climb up the immediate dominator tree as far as we can go while
5205 // still being dominated by the input positions.
5206 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5207
5208 // Don't insert instructions before PHI nodes.
5209 while (isa<PHINode>(IP)) ++IP;
5210
5211 // Ignore landingpad instructions.
5212 while (IP->isEHPad()) ++IP;
5213
5214 // Ignore debug intrinsics.
5215 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5216
5217 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5218 // IP consistent across expansions and allows the previously inserted
5219 // instructions to be reused by subsequent expansion.
5220 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5221 ++IP;
5222
5223 return IP;
5224}
5225
5226/// Emit instructions for the leading candidate expression for this LSRUse (this
5227/// is called "expanding").
5228Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5229 const Formula &F, BasicBlock::iterator IP,
5230 SCEVExpander &Rewriter,
5231 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5232 if (LU.RigidFormula)
5233 return LF.OperandValToReplace;
5234
5235 // Determine an input position which will be dominated by the operands and
5236 // which will dominate the result.
5237 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5238 Rewriter.setInsertPoint(&*IP);
5239
5240 // Inform the Rewriter if we have a post-increment use, so that it can
5241 // perform an advantageous expansion.
5242 Rewriter.setPostInc(LF.PostIncLoops);
5243
5244 // This is the type that the user actually needs.
5245 Type *OpTy = LF.OperandValToReplace->getType();
5246 // This will be the type that we'll initially expand to.
5247 Type *Ty = F.getType();
5248 if (!Ty)
5249 // No type known; just expand directly to the ultimate type.
5250 Ty = OpTy;
5251 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5252 // Expand directly to the ultimate type if it's the right size.
5253 Ty = OpTy;
5254 // This is the type to do integer arithmetic in.
5255 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5256
5257 // Build up a list of operands to add together to form the full base.
5258 SmallVector<const SCEV *, 8> Ops;
5259
5260 // Expand the BaseRegs portion.
5261 for (const SCEV *Reg : F.BaseRegs) {
5262 assert(!Reg->isZero() && "Zero allocated in a base register!")(static_cast <bool> (!Reg->isZero() && "Zero allocated in a base register!"
) ? void (0) : __assert_fail ("!Reg->isZero() && \"Zero allocated in a base register!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5262, __extension__ __PRETTY_FUNCTION__))
;
5263
5264 // If we're expanding for a post-inc user, make the post-inc adjustment.
5265 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5266 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5267 }
5268
5269 // Expand the ScaledReg portion.
5270 Value *ICmpScaledV = nullptr;
5271 if (F.Scale != 0) {
5272 const SCEV *ScaledS = F.ScaledReg;
5273
5274 // If we're expanding for a post-inc user, make the post-inc adjustment.
5275 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5276 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5277
5278 if (LU.Kind == LSRUse::ICmpZero) {
5279 // Expand ScaleReg as if it was part of the base regs.
5280 if (F.Scale == 1)
5281 Ops.push_back(
5282 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5283 else {
5284 // An interesting way of "folding" with an icmp is to use a negated
5285 // scale, which we'll implement by inserting it into the other operand
5286 // of the icmp.
5287 assert(F.Scale == -1 &&(static_cast <bool> (F.Scale == -1 && "The only scale supported by ICmpZero uses is -1!"
) ? void (0) : __assert_fail ("F.Scale == -1 && \"The only scale supported by ICmpZero uses is -1!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5288, __extension__ __PRETTY_FUNCTION__))
5288 "The only scale supported by ICmpZero uses is -1!")(static_cast <bool> (F.Scale == -1 && "The only scale supported by ICmpZero uses is -1!"
) ? void (0) : __assert_fail ("F.Scale == -1 && \"The only scale supported by ICmpZero uses is -1!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5288, __extension__ __PRETTY_FUNCTION__))
;
5289 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5290 }
5291 } else {
5292 // Otherwise just expand the scaled register and an explicit scale,
5293 // which is expected to be matched as part of the address.
5294
5295 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5296 // Unless the addressing mode will not be folded.
5297 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5298 isAMCompletelyFolded(TTI, LU, F)) {
5299 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5300 Ops.clear();
5301 Ops.push_back(SE.getUnknown(FullV));
5302 }
5303 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5304 if (F.Scale != 1)
5305 ScaledS =
5306 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5307 Ops.push_back(ScaledS);
5308 }
5309 }
5310
5311 // Expand the GV portion.
5312 if (F.BaseGV) {
5313 // Flush the operand list to suppress SCEVExpander hoisting.
5314 if (!Ops.empty()) {
5315 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), IntTy);
5316 Ops.clear();
5317 Ops.push_back(SE.getUnknown(FullV));
5318 }
5319 Ops.push_back(SE.getUnknown(F.BaseGV));
5320 }
5321
5322 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5323 // unfolded offsets. LSR assumes they both live next to their uses.
5324 if (!Ops.empty()) {
5325 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5326 Ops.clear();
5327 Ops.push_back(SE.getUnknown(FullV));
5328 }
5329
5330 // Expand the immediate portion.
5331 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5332 if (Offset != 0) {
5333 if (LU.Kind == LSRUse::ICmpZero) {
5334 // The other interesting way of "folding" with an ICmpZero is to use a
5335 // negated immediate.
5336 if (!ICmpScaledV)
5337 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5338 else {
5339 Ops.push_back(SE.getUnknown(ICmpScaledV));
5340 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5341 }
5342 } else {
5343 // Just add the immediate values. These again are expected to be matched
5344 // as part of the address.
5345 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5346 }
5347 }
5348
5349 // Expand the unfolded offset portion.
5350 int64_t UnfoldedOffset = F.UnfoldedOffset;
5351 if (UnfoldedOffset != 0) {
5352 // Just add the immediate values.
5353 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5354 UnfoldedOffset)));
5355 }
5356
5357 // Emit instructions summing all the operands.
5358 const SCEV *FullS = Ops.empty() ?
5359 SE.getConstant(IntTy, 0) :
5360 SE.getAddExpr(Ops);
5361 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5362
5363 // We're done expanding now, so reset the rewriter.
5364 Rewriter.clearPostInc();
5365
5366 // An ICmpZero Formula represents an ICmp which we're handling as a
5367 // comparison against zero. Now that we've expanded an expression for that
5368 // form, update the ICmp's other operand.
5369 if (LU.Kind == LSRUse::ICmpZero) {
5370 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5371 if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
5372 DeadInsts.emplace_back(OperandIsInstr);
5373 assert(!F.BaseGV && "ICmp does not support folding a global value and "(static_cast <bool> (!F.BaseGV && "ICmp does not support folding a global value and "
"a scale at the same time!") ? 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-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5374, __extension__ __PRETTY_FUNCTION__))
5374 "a scale at the same time!")(static_cast <bool> (!F.BaseGV && "ICmp does not support folding a global value and "
"a scale at the same time!") ? 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-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5374, __extension__ __PRETTY_FUNCTION__))
;
5375 if (F.Scale == -1) {
5376 if (ICmpScaledV->getType() != OpTy) {
5377 Instruction *Cast =
5378 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5379 OpTy, false),
5380 ICmpScaledV, OpTy, "tmp", CI);
5381 ICmpScaledV = Cast;
5382 }
5383 CI->setOperand(1, ICmpScaledV);
5384 } else {
5385 // A scale of 1 means that the scale has been expanded as part of the
5386 // base regs.
5387 assert((F.Scale == 0 || F.Scale == 1) &&(static_cast <bool> ((F.Scale == 0 || F.Scale == 1) &&
"ICmp does not support folding a global value and " "a scale at the same time!"
) ? 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-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5389, __extension__ __PRETTY_FUNCTION__))
5388 "ICmp does not support folding a global value and "(static_cast <bool> ((F.Scale == 0 || F.Scale == 1) &&
"ICmp does not support folding a global value and " "a scale at the same time!"
) ? 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-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5389, __extension__ __PRETTY_FUNCTION__))
5389 "a scale at the same time!")(static_cast <bool> ((F.Scale == 0 || F.Scale == 1) &&
"ICmp does not support folding a global value and " "a scale at the same time!"
) ? 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-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5389, __extension__ __PRETTY_FUNCTION__))
;
5390 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5391 -(uint64_t)Offset);
5392 if (C->getType() != OpTy)
5393 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5394 OpTy, false),
5395 C, OpTy);
5396
5397 CI->setOperand(1, C);
5398 }
5399 }
5400
5401 return FullV;
5402}
5403
5404/// Helper for Rewrite. PHI nodes are special because the use of their operands
5405/// effectively happens in their predecessor blocks, so the expression may need
5406/// to be expanded in multiple places.
5407void LSRInstance::RewriteForPHI(
5408 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5409 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5410 DenseMap<BasicBlock *, Value *> Inserted;
5411 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
5412 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3
Assuming pointer value is null
4
Taking true branch
5413 bool needUpdateFixups = false;
5414 BasicBlock *BB = PN->getIncomingBlock(i);
5415
5416 // If this is a critical edge, split the edge so that we do not insert
5417 // the code on all predecessor/successor paths. We do this unless this
5418 // is the canonical backedge for this loop, which complicates post-inc
5419 // users.
5420 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5
Assuming 'e' is equal to 1
6
Taking false branch
5421 !isa<IndirectBrInst>(BB->getTerminator()) &&
5422 !isa<CatchSwitchInst>(BB->getTerminator())) {
5423 BasicBlock *Parent = PN->getParent();
5424 Loop *PNLoop = LI.getLoopFor(Parent);
5425 if (!PNLoop || Parent != PNLoop->getHeader()) {
5426 // Split the critical edge.
5427 BasicBlock *NewBB = nullptr;
5428 if (!Parent->isLandingPad()) {
5429 NewBB =
5430 SplitCriticalEdge(BB, Parent,
5431 CriticalEdgeSplittingOptions(&DT, &LI, MSSAU)
5432 .setMergeIdenticalEdges()
5433 .setKeepOneInputPHIs());
5434 } else {
5435 SmallVector<BasicBlock*, 2> NewBBs;
5436 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5437 NewBB = NewBBs[0];
5438 }
5439 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5440 // phi predecessors are identical. The simple thing to do is skip
5441 // splitting in this case rather than complicate the API.
5442 if (NewBB) {
5443 // If PN is outside of the loop and BB is in the loop, we want to
5444 // move the block to be immediately before the PHI block, not
5445 // immediately after BB.
5446 if (L->contains(BB) && !L->contains(PN))
5447 NewBB->moveBefore(PN->getParent());
5448
5449 // Splitting the edge can reduce the number of PHI entries we have.
5450 e = PN->getNumIncomingValues();
5451 BB = NewBB;
5452 i = PN->getBasicBlockIndex(BB);
5453
5454 needUpdateFixups = true;
5455 }
5456 }
5457 }
5458
5459 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5460 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5461 if (!Pair.second)
7
Assuming field 'second' is true
8
Taking false branch
5462 PN->setIncomingValue(i, Pair.first->second);
5463 else {
5464 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5465 Rewriter, DeadInsts);
5466
5467 // If this is reuse-by-noop-cast, insert the noop cast.
5468 Type *OpTy = LF.OperandValToReplace->getType();
9
Called C++ object pointer is null
5469 if (FullV->getType() != OpTy)
5470 FullV =
5471 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5472 OpTy, false),
5473 FullV, LF.OperandValToReplace->getType(),
5474 "tmp", BB->getTerminator());
5475
5476 PN->setIncomingValue(i, FullV);
5477 Pair.first->second = FullV;
5478 }
5479
5480 // If LSR splits critical edge and phi node has other pending
5481 // fixup operands, we need to update those pending fixups. Otherwise
5482 // formulae will not be implemented completely and some instructions
5483 // will not be eliminated.
5484 if (needUpdateFixups) {
5485 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5486 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5487 // If fixup is supposed to rewrite some operand in the phi
5488 // that was just updated, it may be already moved to
5489 // another phi node. Such fixup requires update.
5490 if (Fixup.UserInst == PN) {
5491 // Check if the operand we try to replace still exists in the
5492 // original phi.
5493 bool foundInOriginalPHI = false;
5494 for (const auto &val : PN->incoming_values())
5495 if (val == Fixup.OperandValToReplace) {
5496 foundInOriginalPHI = true;
5497 break;
5498 }
5499
5500 // If fixup operand found in original PHI - nothing to do.
5501 if (foundInOriginalPHI)
5502 continue;
5503
5504 // Otherwise it might be moved to another PHI and requires update.
5505 // If fixup operand not found in any of the incoming blocks that
5506 // means we have already rewritten it - nothing to do.
5507 for (const auto &Block : PN->blocks())
5508 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5509 ++I) {
5510 PHINode *NewPN = cast<PHINode>(I);
5511 for (const auto &val : NewPN->incoming_values())
5512 if (val == Fixup.OperandValToReplace)
5513 Fixup.UserInst = NewPN;
5514 }
5515 }
5516 }
5517 }
5518}
5519
5520/// Emit instructions for the leading candidate expression for this LSRUse (this
5521/// is called "expanding"), and update the UserInst to reference the newly
5522/// expanded value.
5523void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5524 const Formula &F, SCEVExpander &Rewriter,
5525 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5526 // First, find an insertion point that dominates UserInst. For PHI nodes,
5527 // find the nearest block which dominates all the relevant uses.
5528 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5529 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5530 } else {
5531 Value *FullV =
5532 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5533
5534 // If this is reuse-by-noop-cast, insert the noop cast.
5535 Type *OpTy = LF.OperandValToReplace->getType();
5536 if (FullV->getType() != OpTy) {
5537 Instruction *Cast =
5538 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5539 FullV, OpTy, "tmp", LF.UserInst);
5540 FullV = Cast;
5541 }
5542
5543 // Update the user. ICmpZero is handled specially here (for now) because
5544 // Expand may have updated one of the operands of the icmp already, and
5545 // its new value may happen to be equal to LF.OperandValToReplace, in
5546 // which case doing replaceUsesOfWith leads to replacing both operands
5547 // with the same value. TODO: Reorganize this.
5548 if (LU.Kind == LSRUse::ICmpZero)
5549 LF.UserInst->setOperand(0, FullV);
5550 else
5551 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5552 }
5553
5554 if (auto *OperandIsInstr = dyn_cast<Instruction>(LF.OperandValToReplace))
5555 DeadInsts.emplace_back(OperandIsInstr);
5556}
5557
5558/// Rewrite all the fixup locations with new values, following the chosen
5559/// solution.
5560void LSRInstance::ImplementSolution(
5561 const SmallVectorImpl<const Formula *> &Solution) {
5562 // Keep track of instructions we may have made dead, so that
5563 // we can remove them after we are done working.
5564 SmallVector<WeakTrackingVH, 16> DeadInsts;
5565
5566 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(), "lsr",
5567 false);
5568#ifndef NDEBUG
5569 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5570#endif
5571 Rewriter.disableCanonicalMode();
5572 Rewriter.enableLSRMode();
5573 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5574
5575 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5576 for (const IVChain &Chain : IVChainVec) {
5577 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5578 Rewriter.setChainedPhi(PN);
5579 }
5580
5581 // Expand the new value definitions and update the users.
5582 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5583 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5584 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5585 Changed = true;
5586 }
5587
5588 for (const IVChain &Chain : IVChainVec) {
5589 GenerateIVChain(Chain, Rewriter, DeadInsts);
5590 Changed = true;
5591 }
5592 // Clean up after ourselves. This must be done before deleting any
5593 // instructions.
5594 Rewriter.clear();
5595
5596 Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts,
5597 &TLI, MSSAU);
5598
5599 // In our cost analysis above, we assume that each addrec consumes exactly
5600 // one register, and arrange to have increments inserted just before the
5601 // latch to maximimize the chance this is true. However, if we reused
5602 // existing IVs, we now need to move the increments to match our
5603 // expectations. Otherwise, our cost modeling results in us having a
5604 // chosen a non-optimal result for the actual schedule. (And yes, this
5605 // scheduling decision does impact later codegen.)
5606 for (PHINode &PN : L->getHeader()->phis()) {
5607 BinaryOperator *BO = nullptr;
5608 Value *Start = nullptr, *Step = nullptr;
5609 if (!matchSimpleRecurrence(&PN, BO, Start, Step))
5610 continue;
5611
5612 switch (BO->getOpcode()) {
5613 case Instruction::Sub:
5614 if (BO->getOperand(0) != &PN)
5615 // sub is non-commutative - match handling elsewhere in LSR
5616 continue;
5617 break;
5618 case Instruction::Add:
5619 break;
5620 default:
5621 continue;
5622 };
5623
5624 if (!isa<Constant>(Step))
5625 // If not a constant step, might increase register pressure
5626 // (We assume constants have been canonicalized to RHS)
5627 continue;
5628
5629 if (BO->getParent() == IVIncInsertPos->getParent())
5630 // Only bother moving across blocks. Isel can handle block local case.
5631 continue;
5632
5633 // Can we legally schedule inc at the desired point?
5634 if (!llvm::all_of(BO->uses(),
5635 [&](Use &U) {return DT.dominates(IVIncInsertPos, U);}))
5636 continue;
5637 BO->moveBefore(IVIncInsertPos);
5638 Changed = true;
5639 }
5640
5641
5642}
5643
5644LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5645 DominatorTree &DT, LoopInfo &LI,
5646 const TargetTransformInfo &TTI, AssumptionCache &AC,
5647 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU)
5648 : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), TLI(TLI), TTI(TTI), L(L),
5649 MSSAU(MSSAU), AMK(PreferredAddresingMode.getNumOccurrences() > 0 ?
5650 PreferredAddresingMode : TTI.getPreferredAddressingMode(L, &SE)) {
5651 // If LoopSimplify form is not available, stay out of trouble.
5652 if (!L->isLoopSimplifyForm())
5653 return;
5654
5655 // If there's no interesting work to be done, bail early.
5656 if (IU.empty()) return;
5657
5658 // If there's too much analysis to be done, bail early. We won't be able to
5659 // model the problem anyway.
5660 unsigned NumUsers = 0;
5661 for (const IVStrideUse &U : IU) {
5662 if (++NumUsers > MaxIVUsers) {
5663 (void)U;
5664 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)
5665 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5666 return;
5667 }
5668 // Bail out if we have a PHI on an EHPad that gets a value from a
5669 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5670 // no good place to stick any instructions.
5671 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5672 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5673 if (isa<FuncletPadInst>(FirstNonPHI) ||
5674 isa<CatchSwitchInst>(FirstNonPHI))
5675 for (BasicBlock *PredBB : PN->blocks())
5676 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5677 return;
5678 }
5679 }
5680
5681#ifndef NDEBUG
5682 // All dominating loops must have preheaders, or SCEVExpander may not be able
5683 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5684 //
5685 // IVUsers analysis should only create users that are dominated by simple loop
5686 // headers. Since this loop should dominate all of its users, its user list
5687 // should be empty if this loop itself is not within a simple loop nest.
5688 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5689 Rung; Rung = Rung->getIDom()) {
5690 BasicBlock *BB = Rung->getBlock();
5691 const Loop *DomLoop = LI.getLoopFor(BB);
5692 if (DomLoop && DomLoop->getHeader() == BB) {
5693 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest")(static_cast <bool> (DomLoop->getLoopPreheader() &&
"LSR needs a simplified loop nest") ? void (0) : __assert_fail
("DomLoop->getLoopPreheader() && \"LSR needs a simplified loop nest\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 5693, __extension__ __PRETTY_FUNCTION__))
;
5694 }
5695 }
5696#endif // DEBUG
5697
5698 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)
5699 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)
5700 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5701
5702 // First, perform some low-level loop optimizations.
5703 OptimizeShadowIV();
5704 OptimizeLoopTermCond();
5705
5706 // If loop preparation eliminates all interesting IV users, bail.
5707 if (IU.empty()) return;
5708
5709 // Skip nested loops until we can model them better with formulae.
5710 if (!L->isInnermost()) {
5711 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)
;
5712 return;
5713 }
5714
5715 // Start collecting data and preparing for the solver.
5716 // If number of registers is not the major cost, we cannot benefit from the
5717 // current profitable chain optimization which is based on number of
5718 // registers.
5719 // FIXME: add profitable chain optimization for other kinds major cost, for
5720 // example number of instructions.
5721 if (TTI.isNumRegsMajorCostOfLSR() || StressIVChain)
5722 CollectChains();
5723 CollectInterestingTypesAndFactors();
5724 CollectFixupsAndInitialFormulae();
5725 CollectLoopInvariantFixupsAndFormulae();
5726
5727 if (Uses.empty())
5728 return;
5729
5730 LLVM_DEBUG(dbgs() <<