Bug Summary

File:build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 5638, column 15
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 -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name LoopStrengthReduce.cpp -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 -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm -resource-dir /usr/lib/llvm-16/lib/clang/16.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/lib/Transforms/Scalar -I include -I /build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/include -D _FORTIFY_SOURCE=2 -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-16/lib/clang/16.0.0/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 -fmacro-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/= -O3 -Wno-unused-command-line-argument -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 -Wno-misleading-indentation -std=c++17 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -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-2022-09-04-125545-48738-1 -x c++ /build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/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/BinaryFormat/Dwarf.h"
82#include "llvm/Config/llvm-config.h"
83#include "llvm/IR/BasicBlock.h"
84#include "llvm/IR/Constant.h"
85#include "llvm/IR/Constants.h"
86#include "llvm/IR/DebugInfoMetadata.h"
87#include "llvm/IR/DerivedTypes.h"
88#include "llvm/IR/Dominators.h"
89#include "llvm/IR/GlobalValue.h"
90#include "llvm/IR/IRBuilder.h"
91#include "llvm/IR/InstrTypes.h"
92#include "llvm/IR/Instruction.h"
93#include "llvm/IR/Instructions.h"
94#include "llvm/IR/IntrinsicInst.h"
95#include "llvm/IR/Module.h"
96#include "llvm/IR/Operator.h"
97#include "llvm/IR/PassManager.h"
98#include "llvm/IR/Type.h"
99#include "llvm/IR/Use.h"
100#include "llvm/IR/User.h"
101#include "llvm/IR/Value.h"
102#include "llvm/IR/ValueHandle.h"
103#include "llvm/InitializePasses.h"
104#include "llvm/Pass.h"
105#include "llvm/Support/Casting.h"
106#include "llvm/Support/CommandLine.h"
107#include "llvm/Support/Compiler.h"
108#include "llvm/Support/Debug.h"
109#include "llvm/Support/ErrorHandling.h"
110#include "llvm/Support/MathExtras.h"
111#include "llvm/Support/raw_ostream.h"
112#include "llvm/Transforms/Scalar.h"
113#include "llvm/Transforms/Utils.h"
114#include "llvm/Transforms/Utils/BasicBlockUtils.h"
115#include "llvm/Transforms/Utils/Local.h"
116#include "llvm/Transforms/Utils/LoopUtils.h"
117#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
118#include <algorithm>
119#include <cassert>
120#include <cstddef>
121#include <cstdint>
122#include <iterator>
123#include <limits>
124#include <map>
125#include <numeric>
126#include <utility>
127
128using namespace llvm;
129
130#define DEBUG_TYPE"loop-reduce" "loop-reduce"
131
132/// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
133/// bail out. This threshold is far beyond the number of users that LSR can
134/// conceivably solve, so it should not affect generated code, but catches the
135/// worst cases before LSR burns too much compile time and stack space.
136static const unsigned MaxIVUsers = 200;
137
138/// Limit the size of expression that SCEV-based salvaging will attempt to
139/// translate into a DIExpression.
140/// Choose a maximum size such that debuginfo is not excessively increased and
141/// the salvaging is not too expensive for the compiler.
142static const unsigned MaxSCEVSalvageExpressionSize = 64;
143
144// Cleanup congruent phis after LSR phi expansion.
145static cl::opt<bool> EnablePhiElim(
146 "enable-lsr-phielim", cl::Hidden, cl::init(true),
147 cl::desc("Enable LSR phi elimination"));
148
149// The flag adds instruction count to solutions cost comparison.
150static cl::opt<bool> InsnsCost(
151 "lsr-insns-cost", cl::Hidden, cl::init(true),
152 cl::desc("Add instruction count to a LSR cost model"));
153
154// Flag to choose how to narrow complex lsr solution
155static cl::opt<bool> LSRExpNarrow(
156 "lsr-exp-narrow", cl::Hidden, cl::init(false),
157 cl::desc("Narrow LSR complex solution using"
158 " expectation of registers number"));
159
160// Flag to narrow search space by filtering non-optimal formulae with
161// the same ScaledReg and Scale.
162static cl::opt<bool> FilterSameScaledReg(
163 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
164 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
165 " with the same ScaledReg and Scale"));
166
167static cl::opt<TTI::AddressingModeKind> PreferredAddresingMode(
168 "lsr-preferred-addressing-mode", cl::Hidden, cl::init(TTI::AMK_None),
169 cl::desc("A flag that overrides the target's preferred addressing mode."),
170 cl::values(clEnumValN(TTI::AMK_None,llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
}
171 "none",llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
}
172 "Don't prefer any addressing mode")llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
}
,
173 clEnumValN(TTI::AMK_PreIndexed,llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
174 "preindexed",llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
175 "Prefer pre-indexed addressing mode")llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
,
176 clEnumValN(TTI::AMK_PostIndexed,llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
177 "postindexed",llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
178 "Prefer post-indexed addressing mode")llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
));
179
180static cl::opt<unsigned> ComplexityLimit(
181 "lsr-complexity-limit", cl::Hidden,
182 cl::init(std::numeric_limits<uint16_t>::max()),
183 cl::desc("LSR search space complexity limit"));
184
185static cl::opt<unsigned> SetupCostDepthLimit(
186 "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
187 cl::desc("The limit on recursion depth for LSRs setup cost"));
188
189#ifndef NDEBUG
190// Stress test IV chain generation.
191static cl::opt<bool> StressIVChain(
192 "stress-ivchain", cl::Hidden, cl::init(false),
193 cl::desc("Stress test LSR IV chains"));
194#else
195static bool StressIVChain = false;
196#endif
197
198namespace {
199
200struct MemAccessTy {
201 /// Used in situations where the accessed memory type is unknown.
202 static const unsigned UnknownAddressSpace =
203 std::numeric_limits<unsigned>::max();
204
205 Type *MemTy = nullptr;
206 unsigned AddrSpace = UnknownAddressSpace;
207
208 MemAccessTy() = default;
209 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
210
211 bool operator==(MemAccessTy Other) const {
212 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
213 }
214
215 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
216
217 static MemAccessTy getUnknown(LLVMContext &Ctx,
218 unsigned AS = UnknownAddressSpace) {
219 return MemAccessTy(Type::getVoidTy(Ctx), AS);
220 }
221
222 Type *getType() { return MemTy; }
223};
224
225/// This class holds data which is used to order reuse candidates.
226class RegSortData {
227public:
228 /// This represents the set of LSRUse indices which reference
229 /// a particular register.
230 SmallBitVector UsedByIndices;
231
232 void print(raw_ostream &OS) const;
233 void dump() const;
234};
235
236} // end anonymous namespace
237
238#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
239void RegSortData::print(raw_ostream &OS) const {
240 OS << "[NumUses=" << UsedByIndices.count() << ']';
241}
242
243LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
244 print(errs()); errs() << '\n';
245}
246#endif
247
248namespace {
249
250/// Map register candidates to information about how they are used.
251class RegUseTracker {
252 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
253
254 RegUsesTy RegUsesMap;
255 SmallVector<const SCEV *, 16> RegSequence;
256
257public:
258 void countRegister(const SCEV *Reg, size_t LUIdx);
259 void dropRegister(const SCEV *Reg, size_t LUIdx);
260 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
261
262 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
263
264 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
265
266 void clear();
267
268 using iterator = SmallVectorImpl<const SCEV *>::iterator;
269 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
270
271 iterator begin() { return RegSequence.begin(); }
272 iterator end() { return RegSequence.end(); }
273 const_iterator begin() const { return RegSequence.begin(); }
274 const_iterator end() const { return RegSequence.end(); }
275};
276
277} // end anonymous namespace
278
279void
280RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
281 std::pair<RegUsesTy::iterator, bool> Pair =
282 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
283 RegSortData &RSD = Pair.first->second;
284 if (Pair.second)
285 RegSequence.push_back(Reg);
286 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
287 RSD.UsedByIndices.set(LUIdx);
288}
289
290void
291RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
292 RegUsesTy::iterator It = RegUsesMap.find(Reg);
293 assert(It != RegUsesMap.end())(static_cast <bool> (It != RegUsesMap.end()) ? void (0)
: __assert_fail ("It != RegUsesMap.end()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 293, __extension__ __PRETTY_FUNCTION__))
;
294 RegSortData &RSD = It->second;
295 assert(RSD.UsedByIndices.size() > LUIdx)(static_cast <bool> (RSD.UsedByIndices.size() > LUIdx
) ? void (0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx"
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 295, __extension__
__PRETTY_FUNCTION__))
;
296 RSD.UsedByIndices.reset(LUIdx);
297}
298
299void
300RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
301 assert(LUIdx <= LastLUIdx)(static_cast <bool> (LUIdx <= LastLUIdx) ? void (0) :
__assert_fail ("LUIdx <= LastLUIdx", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 301, __extension__ __PRETTY_FUNCTION__))
;
302
303 // Update RegUses. The data structure is not optimized for this purpose;
304 // we must iterate through it and update each of the bit vectors.
305 for (auto &Pair : RegUsesMap) {
306 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
307 if (LUIdx < UsedByIndices.size())
308 UsedByIndices[LUIdx] =
309 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
310 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
311 }
312}
313
314bool
315RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
316 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
317 if (I == RegUsesMap.end())
318 return false;
319 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
320 int i = UsedByIndices.find_first();
321 if (i == -1) return false;
322 if ((size_t)i != LUIdx) return true;
323 return UsedByIndices.find_next(i) != -1;
324}
325
326const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
327 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
328 assert(I != RegUsesMap.end() && "Unknown register!")(static_cast <bool> (I != RegUsesMap.end() && "Unknown register!"
) ? void (0) : __assert_fail ("I != RegUsesMap.end() && \"Unknown register!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 328, __extension__
__PRETTY_FUNCTION__))
;
329 return I->second.UsedByIndices;
330}
331
332void RegUseTracker::clear() {
333 RegUsesMap.clear();
334 RegSequence.clear();
335}
336
337namespace {
338
339/// This class holds information that describes a formula for computing
340/// satisfying a use. It may include broken-out immediates and scaled registers.
341struct Formula {
342 /// Global base address used for complex addressing.
343 GlobalValue *BaseGV = nullptr;
344
345 /// Base offset for complex addressing.
346 int64_t BaseOffset = 0;
347
348 /// Whether any complex addressing has a base register.
349 bool HasBaseReg = false;
350
351 /// The scale of any complex addressing.
352 int64_t Scale = 0;
353
354 /// The list of "base" registers for this use. When this is non-empty. The
355 /// canonical representation of a formula is
356 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
357 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
358 /// 3. The reg containing recurrent expr related with currect loop in the
359 /// formula should be put in the ScaledReg.
360 /// #1 enforces that the scaled register is always used when at least two
361 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
362 /// #2 enforces that 1 * reg is reg.
363 /// #3 ensures invariant regs with respect to current loop can be combined
364 /// together in LSR codegen.
365 /// This invariant can be temporarily broken while building a formula.
366 /// However, every formula inserted into the LSRInstance must be in canonical
367 /// form.
368 SmallVector<const SCEV *, 4> BaseRegs;
369
370 /// The 'scaled' register for this use. This should be non-null when Scale is
371 /// not zero.
372 const SCEV *ScaledReg = nullptr;
373
374 /// An additional constant offset which added near the use. This requires a
375 /// temporary register, but the offset itself can live in an add immediate
376 /// field rather than a register.
377 int64_t UnfoldedOffset = 0;
378
379 Formula() = default;
380
381 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
382
383 bool isCanonical(const Loop &L) const;
384
385 void canonicalize(const Loop &L);
386
387 bool unscale();
388
389 bool hasZeroEnd() const;
390
391 size_t getNumRegs() const;
392 Type *getType() const;
393
394 void deleteBaseReg(const SCEV *&S);
395
396 bool referencesReg(const SCEV *S) const;
397 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
398 const RegUseTracker &RegUses) const;
399
400 void print(raw_ostream &OS) const;
401 void dump() const;
402};
403
404} // end anonymous namespace
405
406/// Recursion helper for initialMatch.
407static void DoInitialMatch(const SCEV *S, Loop *L,
408 SmallVectorImpl<const SCEV *> &Good,
409 SmallVectorImpl<const SCEV *> &Bad,
410 ScalarEvolution &SE) {
411 // Collect expressions which properly dominate the loop header.
412 if (SE.properlyDominates(S, L->getHeader())) {
413 Good.push_back(S);
414 return;
415 }
416
417 // Look at add operands.
418 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
419 for (const SCEV *S : Add->operands())
420 DoInitialMatch(S, L, Good, Bad, SE);
421 return;
422 }
423
424 // Look at addrec operands.
425 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
426 if (!AR->getStart()->isZero() && AR->isAffine()) {
427 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
428 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
429 AR->getStepRecurrence(SE),
430 // FIXME: AR->getNoWrapFlags()
431 AR->getLoop(), SCEV::FlagAnyWrap),
432 L, Good, Bad, SE);
433 return;
434 }
435
436 // Handle a multiplication by -1 (negation) if it didn't fold.
437 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
438 if (Mul->getOperand(0)->isAllOnesValue()) {
439 SmallVector<const SCEV *, 4> Ops(drop_begin(Mul->operands()));
440 const SCEV *NewMul = SE.getMulExpr(Ops);
441
442 SmallVector<const SCEV *, 4> MyGood;
443 SmallVector<const SCEV *, 4> MyBad;
444 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
445 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
446 SE.getEffectiveSCEVType(NewMul->getType())));
447 for (const SCEV *S : MyGood)
448 Good.push_back(SE.getMulExpr(NegOne, S));
449 for (const SCEV *S : MyBad)
450 Bad.push_back(SE.getMulExpr(NegOne, S));
451 return;
452 }
453
454 // Ok, we can't do anything interesting. Just stuff the whole thing into a
455 // register and hope for the best.
456 Bad.push_back(S);
457}
458
459/// Incorporate loop-variant parts of S into this Formula, attempting to keep
460/// all loop-invariant and loop-computable values in a single base register.
461void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
462 SmallVector<const SCEV *, 4> Good;
463 SmallVector<const SCEV *, 4> Bad;
464 DoInitialMatch(S, L, Good, Bad, SE);
465 if (!Good.empty()) {
466 const SCEV *Sum = SE.getAddExpr(Good);
467 if (!Sum->isZero())
468 BaseRegs.push_back(Sum);
469 HasBaseReg = true;
470 }
471 if (!Bad.empty()) {
472 const SCEV *Sum = SE.getAddExpr(Bad);
473 if (!Sum->isZero())
474 BaseRegs.push_back(Sum);
475 HasBaseReg = true;
476 }
477 canonicalize(*L);
478}
479
480static bool containsAddRecDependentOnLoop(const SCEV *S, const Loop &L) {
481 return SCEVExprContains(S, [&L](const SCEV *S) {
482 return isa<SCEVAddRecExpr>(S) && (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
483 });
484}
485
486/// Check whether or not this formula satisfies the canonical
487/// representation.
488/// \see Formula::BaseRegs.
489bool Formula::isCanonical(const Loop &L) const {
490 if (!ScaledReg)
491 return BaseRegs.size() <= 1;
492
493 if (Scale != 1)
494 return true;
495
496 if (Scale == 1 && BaseRegs.empty())
497 return false;
498
499 if (containsAddRecDependentOnLoop(ScaledReg, L))
500 return true;
501
502 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
503 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
504 // loop, we want to swap the reg in BaseRegs with ScaledReg.
505 return none_of(BaseRegs, [&L](const SCEV *S) {
506 return containsAddRecDependentOnLoop(S, L);
507 });
508}
509
510/// Helper method to morph a formula into its canonical representation.
511/// \see Formula::BaseRegs.
512/// Every formula having more than one base register, must use the ScaledReg
513/// field. Otherwise, we would have to do special cases everywhere in LSR
514/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
515/// On the other hand, 1*reg should be canonicalized into reg.
516void Formula::canonicalize(const Loop &L) {
517 if (isCanonical(L))
518 return;
519
520 if (BaseRegs.empty()) {
521 // No base reg? Use scale reg with scale = 1 as such.
522 assert(ScaledReg && "Expected 1*reg => reg")(static_cast <bool> (ScaledReg && "Expected 1*reg => reg"
) ? void (0) : __assert_fail ("ScaledReg && \"Expected 1*reg => reg\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 522, __extension__
__PRETTY_FUNCTION__))
;
523 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 523, __extension__
__PRETTY_FUNCTION__))
;
524 BaseRegs.push_back(ScaledReg);
525 Scale = 0;
526 ScaledReg = nullptr;
527 return;
528 }
529
530 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
531 if (!ScaledReg) {
532 ScaledReg = BaseRegs.pop_back_val();
533 Scale = 1;
534 }
535
536 // If ScaledReg is an invariant with respect to L, find the reg from
537 // BaseRegs containing the recurrent expr related with Loop L. Swap the
538 // reg with ScaledReg.
539 if (!containsAddRecDependentOnLoop(ScaledReg, L)) {
540 auto I = find_if(BaseRegs, [&L](const SCEV *S) {
541 return containsAddRecDependentOnLoop(S, L);
542 });
543 if (I != BaseRegs.end())
544 std::swap(ScaledReg, *I);
545 }
546 assert(isCanonical(L) && "Failed to canonicalize?")(static_cast <bool> (isCanonical(L) && "Failed to canonicalize?"
) ? void (0) : __assert_fail ("isCanonical(L) && \"Failed to canonicalize?\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 546, __extension__
__PRETTY_FUNCTION__))
;
547}
548
549/// Get rid of the scale in the formula.
550/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
551/// \return true if it was possible to get rid of the scale, false otherwise.
552/// \note After this operation the formula may not be in the canonical form.
553bool Formula::unscale() {
554 if (Scale != 1)
555 return false;
556 Scale = 0;
557 BaseRegs.push_back(ScaledReg);
558 ScaledReg = nullptr;
559 return true;
560}
561
562bool Formula::hasZeroEnd() const {
563 if (UnfoldedOffset || BaseOffset)
564 return false;
565 if (BaseRegs.size() != 1 || ScaledReg)
566 return false;
567 return true;
568}
569
570/// Return the total number of register operands used by this formula. This does
571/// not include register uses implied by non-constant addrec strides.
572size_t Formula::getNumRegs() const {
573 return !!ScaledReg + BaseRegs.size();
574}
575
576/// Return the type of this formula, if it has one, or null otherwise. This type
577/// is meaningless except for the bit size.
578Type *Formula::getType() const {
579 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
580 ScaledReg ? ScaledReg->getType() :
581 BaseGV ? BaseGV->getType() :
582 nullptr;
583}
584
585/// Delete the given base reg from the BaseRegs list.
586void Formula::deleteBaseReg(const SCEV *&S) {
587 if (&S != &BaseRegs.back())
588 std::swap(S, BaseRegs.back());
589 BaseRegs.pop_back();
590}
591
592/// Test if this formula references the given register.
593bool Formula::referencesReg(const SCEV *S) const {
594 return S == ScaledReg || is_contained(BaseRegs, S);
595}
596
597/// Test whether this formula uses registers which are used by uses other than
598/// the use with the given index.
599bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
600 const RegUseTracker &RegUses) const {
601 if (ScaledReg)
602 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
603 return true;
604 for (const SCEV *BaseReg : BaseRegs)
605 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
606 return true;
607 return false;
608}
609
610#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
611void Formula::print(raw_ostream &OS) const {
612 bool First = true;
613 if (BaseGV) {
614 if (!First) OS << " + "; else First = false;
615 BaseGV->printAsOperand(OS, /*PrintType=*/false);
616 }
617 if (BaseOffset != 0) {
618 if (!First) OS << " + "; else First = false;
619 OS << BaseOffset;
620 }
621 for (const SCEV *BaseReg : BaseRegs) {
622 if (!First) OS << " + "; else First = false;
623 OS << "reg(" << *BaseReg << ')';
624 }
625 if (HasBaseReg && BaseRegs.empty()) {
626 if (!First) OS << " + "; else First = false;
627 OS << "**error: HasBaseReg**";
628 } else if (!HasBaseReg && !BaseRegs.empty()) {
629 if (!First) OS << " + "; else First = false;
630 OS << "**error: !HasBaseReg**";
631 }
632 if (Scale != 0) {
633 if (!First) OS << " + "; else First = false;
634 OS << Scale << "*reg(";
635 if (ScaledReg)
636 OS << *ScaledReg;
637 else
638 OS << "<unknown>";
639 OS << ')';
640 }
641 if (UnfoldedOffset != 0) {
642 if (!First) OS << " + ";
643 OS << "imm(" << UnfoldedOffset << ')';
644 }
645}
646
647LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
648 print(errs()); errs() << '\n';
649}
650#endif
651
652/// Return true if the given addrec can be sign-extended without changing its
653/// value.
654static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
655 Type *WideTy =
656 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
657 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
658}
659
660/// Return true if the given add can be sign-extended without changing its
661/// value.
662static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
663 Type *WideTy =
664 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
665 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
666}
667
668/// Return true if the given mul can be sign-extended without changing its
669/// value.
670static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
671 Type *WideTy =
672 IntegerType::get(SE.getContext(),
673 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
674 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
675}
676
677/// Return an expression for LHS /s RHS, if it can be determined and if the
678/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
679/// is true, expressions like (X * Y) /s Y are simplified to X, ignoring that
680/// the multiplication may overflow, which is useful when the result will be
681/// used in a context where the most significant bits are ignored.
682static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
683 ScalarEvolution &SE,
684 bool IgnoreSignificantBits = false) {
685 // Handle the trivial case, which works for any SCEV type.
686 if (LHS == RHS)
687 return SE.getConstant(LHS->getType(), 1);
688
689 // Handle a few RHS special cases.
690 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
691 if (RC) {
692 const APInt &RA = RC->getAPInt();
693 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
694 // some folding.
695 if (RA.isAllOnes()) {
696 if (LHS->getType()->isPointerTy())
697 return nullptr;
698 return SE.getMulExpr(LHS, RC);
699 }
700 // Handle x /s 1 as x.
701 if (RA == 1)
702 return LHS;
703 }
704
705 // Check for a division of a constant by a constant.
706 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
707 if (!RC)
708 return nullptr;
709 const APInt &LA = C->getAPInt();
710 const APInt &RA = RC->getAPInt();
711 if (LA.srem(RA) != 0)
712 return nullptr;
713 return SE.getConstant(LA.sdiv(RA));
714 }
715
716 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
717 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
718 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
719 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
720 IgnoreSignificantBits);
721 if (!Step) return nullptr;
722 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
723 IgnoreSignificantBits);
724 if (!Start) return nullptr;
725 // FlagNW is independent of the start value, step direction, and is
726 // preserved with smaller magnitude steps.
727 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
728 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
729 }
730 return nullptr;
731 }
732
733 // Distribute the sdiv over add operands, if the add doesn't overflow.
734 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
735 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
736 SmallVector<const SCEV *, 8> Ops;
737 for (const SCEV *S : Add->operands()) {
738 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
739 if (!Op) return nullptr;
740 Ops.push_back(Op);
741 }
742 return SE.getAddExpr(Ops);
743 }
744 return nullptr;
745 }
746
747 // Check for a multiply operand that we can pull RHS out of.
748 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
749 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
750 // Handle special case C1*X*Y /s C2*X*Y.
751 if (const SCEVMulExpr *MulRHS = dyn_cast<SCEVMulExpr>(RHS)) {
752 if (IgnoreSignificantBits || isMulSExtable(MulRHS, SE)) {
753 const SCEVConstant *LC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
754 const SCEVConstant *RC =
755 dyn_cast<SCEVConstant>(MulRHS->getOperand(0));
756 if (LC && RC) {
757 SmallVector<const SCEV *, 4> LOps(drop_begin(Mul->operands()));
758 SmallVector<const SCEV *, 4> ROps(drop_begin(MulRHS->operands()));
759 if (LOps == ROps)
760 return getExactSDiv(LC, RC, SE, IgnoreSignificantBits);
761 }
762 }
763 }
764
765 SmallVector<const SCEV *, 4> Ops;
766 bool Found = false;
767 for (const SCEV *S : Mul->operands()) {
768 if (!Found)
769 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
770 IgnoreSignificantBits)) {
771 S = Q;
772 Found = true;
773 }
774 Ops.push_back(S);
775 }
776 return Found ? SE.getMulExpr(Ops) : nullptr;
777 }
778 return nullptr;
779 }
780
781 // Otherwise we don't know.
782 return nullptr;
783}
784
785/// If S involves the addition of a constant integer value, return that integer
786/// value, and mutate S to point to a new SCEV with that value excluded.
787static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
788 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
789 if (C->getAPInt().getMinSignedBits() <= 64) {
790 S = SE.getConstant(C->getType(), 0);
791 return C->getValue()->getSExtValue();
792 }
793 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
794 SmallVector<const SCEV *, 8> NewOps(Add->operands());
795 int64_t Result = ExtractImmediate(NewOps.front(), SE);
796 if (Result != 0)
797 S = SE.getAddExpr(NewOps);
798 return Result;
799 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
800 SmallVector<const SCEV *, 8> NewOps(AR->operands());
801 int64_t Result = ExtractImmediate(NewOps.front(), SE);
802 if (Result != 0)
803 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
804 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
805 SCEV::FlagAnyWrap);
806 return Result;
807 }
808 return 0;
809}
810
811/// If S involves the addition of a GlobalValue address, return that symbol, and
812/// mutate S to point to a new SCEV with that value excluded.
813static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
814 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
815 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
816 S = SE.getConstant(GV->getType(), 0);
817 return GV;
818 }
819 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
820 SmallVector<const SCEV *, 8> NewOps(Add->operands());
821 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
822 if (Result)
823 S = SE.getAddExpr(NewOps);
824 return Result;
825 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
826 SmallVector<const SCEV *, 8> NewOps(AR->operands());
827 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
828 if (Result)
829 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
830 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
831 SCEV::FlagAnyWrap);
832 return Result;
833 }
834 return nullptr;
835}
836
837/// Returns true if the specified instruction is using the specified value as an
838/// address.
839static bool isAddressUse(const TargetTransformInfo &TTI,
840 Instruction *Inst, Value *OperandVal) {
841 bool isAddress = isa<LoadInst>(Inst);
842 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
843 if (SI->getPointerOperand() == OperandVal)
844 isAddress = true;
845 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
846 // Addressing modes can also be folded into prefetches and a variety
847 // of intrinsics.
848 switch (II->getIntrinsicID()) {
849 case Intrinsic::memset:
850 case Intrinsic::prefetch:
851 case Intrinsic::masked_load:
852 if (II->getArgOperand(0) == OperandVal)
853 isAddress = true;
854 break;
855 case Intrinsic::masked_store:
856 if (II->getArgOperand(1) == OperandVal)
857 isAddress = true;
858 break;
859 case Intrinsic::memmove:
860 case Intrinsic::memcpy:
861 if (II->getArgOperand(0) == OperandVal ||
862 II->getArgOperand(1) == OperandVal)
863 isAddress = true;
864 break;
865 default: {
866 MemIntrinsicInfo IntrInfo;
867 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
868 if (IntrInfo.PtrVal == OperandVal)
869 isAddress = true;
870 }
871 }
872 }
873 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
874 if (RMW->getPointerOperand() == OperandVal)
875 isAddress = true;
876 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
877 if (CmpX->getPointerOperand() == OperandVal)
878 isAddress = true;
879 }
880 return isAddress;
881}
882
883/// Return the type of the memory being accessed.
884static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
885 Instruction *Inst, Value *OperandVal) {
886 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
887 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
888 AccessTy.MemTy = SI->getOperand(0)->getType();
889 AccessTy.AddrSpace = SI->getPointerAddressSpace();
890 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
891 AccessTy.AddrSpace = LI->getPointerAddressSpace();
892 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
893 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
894 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
895 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
896 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
897 switch (II->getIntrinsicID()) {
898 case Intrinsic::prefetch:
899 case Intrinsic::memset:
900 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
901 AccessTy.MemTy = OperandVal->getType();
902 break;
903 case Intrinsic::memmove:
904 case Intrinsic::memcpy:
905 AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
906 AccessTy.MemTy = OperandVal->getType();
907 break;
908 case Intrinsic::masked_load:
909 AccessTy.AddrSpace =
910 II->getArgOperand(0)->getType()->getPointerAddressSpace();
911 break;
912 case Intrinsic::masked_store:
913 AccessTy.MemTy = II->getOperand(0)->getType();
914 AccessTy.AddrSpace =
915 II->getArgOperand(1)->getType()->getPointerAddressSpace();
916 break;
917 default: {
918 MemIntrinsicInfo IntrInfo;
919 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
920 AccessTy.AddrSpace
921 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
922 }
923
924 break;
925 }
926 }
927 }
928
929 // All pointers have the same requirements, so canonicalize them to an
930 // arbitrary pointer type to minimize variation.
931 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
932 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
933 PTy->getAddressSpace());
934
935 return AccessTy;
936}
937
938/// Return true if this AddRec is already a phi in its loop.
939static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
940 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
941 if (SE.isSCEVable(PN.getType()) &&
942 (SE.getEffectiveSCEVType(PN.getType()) ==
943 SE.getEffectiveSCEVType(AR->getType())) &&
944 SE.getSCEV(&PN) == AR)
945 return true;
946 }
947 return false;
948}
949
950/// Check if expanding this expression is likely to incur significant cost. This
951/// is tricky because SCEV doesn't track which expressions are actually computed
952/// by the current IR.
953///
954/// We currently allow expansion of IV increments that involve adds,
955/// multiplication by constants, and AddRecs from existing phis.
956///
957/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
958/// obvious multiple of the UDivExpr.
959static bool isHighCostExpansion(const SCEV *S,
960 SmallPtrSetImpl<const SCEV*> &Processed,
961 ScalarEvolution &SE) {
962 // Zero/One operand expressions
963 switch (S->getSCEVType()) {
964 case scUnknown:
965 case scConstant:
966 return false;
967 case scTruncate:
968 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
969 Processed, SE);
970 case scZeroExtend:
971 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
972 Processed, SE);
973 case scSignExtend:
974 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
975 Processed, SE);
976 default:
977 break;
978 }
979
980 if (!Processed.insert(S).second)
981 return false;
982
983 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
984 for (const SCEV *S : Add->operands()) {
985 if (isHighCostExpansion(S, Processed, SE))
986 return true;
987 }
988 return false;
989 }
990
991 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
992 if (Mul->getNumOperands() == 2) {
993 // Multiplication by a constant is ok
994 if (isa<SCEVConstant>(Mul->getOperand(0)))
995 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
996
997 // If we have the value of one operand, check if an existing
998 // multiplication already generates this expression.
999 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
1000 Value *UVal = U->getValue();
1001 for (User *UR : UVal->users()) {
1002 // If U is a constant, it may be used by a ConstantExpr.
1003 Instruction *UI = dyn_cast<Instruction>(UR);
1004 if (UI && UI->getOpcode() == Instruction::Mul &&
1005 SE.isSCEVable(UI->getType())) {
1006 return SE.getSCEV(UI) == Mul;
1007 }
1008 }
1009 }
1010 }
1011 }
1012
1013 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1014 if (isExistingPhi(AR, SE))
1015 return false;
1016 }
1017
1018 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
1019 return true;
1020}
1021
1022namespace {
1023
1024class LSRUse;
1025
1026} // end anonymous namespace
1027
1028/// Check if the addressing mode defined by \p F is completely
1029/// folded in \p LU at isel time.
1030/// This includes address-mode folding and special icmp tricks.
1031/// This function returns true if \p LU can accommodate what \p F
1032/// defines and up to 1 base + 1 scaled + offset.
1033/// In other words, if \p F has several base registers, this function may
1034/// still return true. Therefore, users still need to account for
1035/// additional base registers and/or unfolded offsets to derive an
1036/// accurate cost model.
1037static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1038 const LSRUse &LU, const Formula &F);
1039
1040// Get the cost of the scaling factor used in F for LU.
1041static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1042 const LSRUse &LU, const Formula &F,
1043 const Loop &L);
1044
1045namespace {
1046
1047/// This class is used to measure and compare candidate formulae.
1048class Cost {
1049 const Loop *L = nullptr;
1050 ScalarEvolution *SE = nullptr;
1051 const TargetTransformInfo *TTI = nullptr;
1052 TargetTransformInfo::LSRCost C;
1053 TTI::AddressingModeKind AMK = TTI::AMK_None;
1054
1055public:
1056 Cost() = delete;
1057 Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
1058 TTI::AddressingModeKind AMK) :
1059 L(L), SE(&SE), TTI(&TTI), AMK(AMK) {
1060 C.Insns = 0;
1061 C.NumRegs = 0;
1062 C.AddRecCost = 0;
1063 C.NumIVMuls = 0;
1064 C.NumBaseAdds = 0;
1065 C.ImmCost = 0;
1066 C.SetupCost = 0;
1067 C.ScaleCost = 0;
1068 }
1069
1070 bool isLess(const Cost &Other);
1071
1072 void Lose();
1073
1074#ifndef NDEBUG
1075 // Once any of the metrics loses, they must all remain losers.
1076 bool isValid() {
1077 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1078 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1079 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1080 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1081 }
1082#endif
1083
1084 bool isLoser() {
1085 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1085, __extension__
__PRETTY_FUNCTION__))
;
1086 return C.NumRegs == ~0u;
1087 }
1088
1089 void RateFormula(const Formula &F,
1090 SmallPtrSetImpl<const SCEV *> &Regs,
1091 const DenseSet<const SCEV *> &VisitedRegs,
1092 const LSRUse &LU,
1093 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1094
1095 void print(raw_ostream &OS) const;
1096 void dump() const;
1097
1098private:
1099 void RateRegister(const Formula &F, const SCEV *Reg,
1100 SmallPtrSetImpl<const SCEV *> &Regs);
1101 void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1102 SmallPtrSetImpl<const SCEV *> &Regs,
1103 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1104};
1105
1106/// An operand value in an instruction which is to be replaced with some
1107/// equivalent, possibly strength-reduced, replacement.
1108struct LSRFixup {
1109 /// The instruction which will be updated.
1110 Instruction *UserInst = nullptr;
1111
1112 /// The operand of the instruction which will be replaced. The operand may be
1113 /// used more than once; every instance will be replaced.
1114 Value *OperandValToReplace = nullptr;
1115
1116 /// If this user is to use the post-incremented value of an induction
1117 /// variable, this set is non-empty and holds the loops associated with the
1118 /// induction variable.
1119 PostIncLoopSet PostIncLoops;
1120
1121 /// A constant offset to be added to the LSRUse expression. This allows
1122 /// multiple fixups to share the same LSRUse with different offsets, for
1123 /// example in an unrolled loop.
1124 int64_t Offset = 0;
1125
1126 LSRFixup() = default;
1127
1128 bool isUseFullyOutsideLoop(const Loop *L) const;
1129
1130 void print(raw_ostream &OS) const;
1131 void dump() const;
1132};
1133
1134/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1135/// SmallVectors of const SCEV*.
1136struct UniquifierDenseMapInfo {
1137 static SmallVector<const SCEV *, 4> getEmptyKey() {
1138 SmallVector<const SCEV *, 4> V;
1139 V.push_back(reinterpret_cast<const SCEV *>(-1));
1140 return V;
1141 }
1142
1143 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1144 SmallVector<const SCEV *, 4> V;
1145 V.push_back(reinterpret_cast<const SCEV *>(-2));
1146 return V;
1147 }
1148
1149 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1150 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1151 }
1152
1153 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1154 const SmallVector<const SCEV *, 4> &RHS) {
1155 return LHS == RHS;
1156 }
1157};
1158
1159/// This class holds the state that LSR keeps for each use in IVUsers, as well
1160/// as uses invented by LSR itself. It includes information about what kinds of
1161/// things can be folded into the user, information about the user itself, and
1162/// information about how the use may be satisfied. TODO: Represent multiple
1163/// users of the same expression in common?
1164class LSRUse {
1165 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1166
1167public:
1168 /// An enum for a kind of use, indicating what types of scaled and immediate
1169 /// operands it might support.
1170 enum KindType {
1171 Basic, ///< A normal use, with no folding.
1172 Special, ///< A special case of basic, allowing -1 scales.
1173 Address, ///< An address use; folding according to TargetLowering
1174 ICmpZero ///< An equality icmp with both operands folded into one.
1175 // TODO: Add a generic icmp too?
1176 };
1177
1178 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1179
1180 KindType Kind;
1181 MemAccessTy AccessTy;
1182
1183 /// The list of operands which are to be replaced.
1184 SmallVector<LSRFixup, 8> Fixups;
1185
1186 /// Keep track of the min and max offsets of the fixups.
1187 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1188 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1189
1190 /// This records whether all of the fixups using this LSRUse are outside of
1191 /// the loop, in which case some special-case heuristics may be used.
1192 bool AllFixupsOutsideLoop = true;
1193
1194 /// RigidFormula is set to true to guarantee that this use will be associated
1195 /// with a single formula--the one that initially matched. Some SCEV
1196 /// expressions cannot be expanded. This allows LSR to consider the registers
1197 /// used by those expressions without the need to expand them later after
1198 /// changing the formula.
1199 bool RigidFormula = false;
1200
1201 /// This records the widest use type for any fixup using this
1202 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1203 /// fixup widths to be equivalent, because the narrower one may be relying on
1204 /// the implicit truncation to truncate away bogus bits.
1205 Type *WidestFixupType = nullptr;
1206
1207 /// A list of ways to build a value that can satisfy this user. After the
1208 /// list is populated, one of these is selected heuristically and used to
1209 /// formulate a replacement for OperandValToReplace in UserInst.
1210 SmallVector<Formula, 12> Formulae;
1211
1212 /// The set of register candidates used by all formulae in this LSRUse.
1213 SmallPtrSet<const SCEV *, 4> Regs;
1214
1215 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1216
1217 LSRFixup &getNewFixup() {
1218 Fixups.push_back(LSRFixup());
1219 return Fixups.back();
1220 }
1221
1222 void pushFixup(LSRFixup &f) {
1223 Fixups.push_back(f);
1224 if (f.Offset > MaxOffset)
1225 MaxOffset = f.Offset;
1226 if (f.Offset < MinOffset)
1227 MinOffset = f.Offset;
1228 }
1229
1230 bool HasFormulaWithSameRegs(const Formula &F) const;
1231 float getNotSelectedProbability(const SCEV *Reg) const;
1232 bool InsertFormula(const Formula &F, const Loop &L);
1233 void DeleteFormula(Formula &F);
1234 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1235
1236 void print(raw_ostream &OS) const;
1237 void dump() const;
1238};
1239
1240} // end anonymous namespace
1241
1242static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1243 LSRUse::KindType Kind, MemAccessTy AccessTy,
1244 GlobalValue *BaseGV, int64_t BaseOffset,
1245 bool HasBaseReg, int64_t Scale,
1246 Instruction *Fixup = nullptr);
1247
1248static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1249 if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1250 return 1;
1251 if (Depth == 0)
1252 return 0;
1253 if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1254 return getSetupCost(S->getStart(), Depth - 1);
1255 if (auto S = dyn_cast<SCEVIntegralCastExpr>(Reg))
1256 return getSetupCost(S->getOperand(), Depth - 1);
1257 if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1258 return std::accumulate(S->op_begin(), S->op_end(), 0,
1259 [&](unsigned i, const SCEV *Reg) {
1260 return i + getSetupCost(Reg, Depth - 1);
1261 });
1262 if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1263 return getSetupCost(S->getLHS(), Depth - 1) +
1264 getSetupCost(S->getRHS(), Depth - 1);
1265 return 0;
1266}
1267
1268/// Tally up interesting quantities from the given register.
1269void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1270 SmallPtrSetImpl<const SCEV *> &Regs) {
1271 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1272 // If this is an addrec for another loop, it should be an invariant
1273 // with respect to L since L is the innermost loop (at least
1274 // for now LSR only handles innermost loops).
1275 if (AR->getLoop() != L) {
1276 // If the AddRec exists, consider it's register free and leave it alone.
1277 if (isExistingPhi(AR, *SE) && AMK != TTI::AMK_PostIndexed)
1278 return;
1279
1280 // It is bad to allow LSR for current loop to add induction variables
1281 // for its sibling loops.
1282 if (!AR->getLoop()->contains(L)) {
1283 Lose();
1284 return;
1285 }
1286
1287 // Otherwise, it will be an invariant with respect to Loop L.
1288 ++C.NumRegs;
1289 return;
1290 }
1291
1292 unsigned LoopCost = 1;
1293 if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1294 TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1295
1296 // If the step size matches the base offset, we could use pre-indexed
1297 // addressing.
1298 if (AMK == TTI::AMK_PreIndexed) {
1299 if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1300 if (Step->getAPInt() == F.BaseOffset)
1301 LoopCost = 0;
1302 } else if (AMK == TTI::AMK_PostIndexed) {
1303 const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1304 if (isa<SCEVConstant>(LoopStep)) {
1305 const SCEV *LoopStart = AR->getStart();
1306 if (!isa<SCEVConstant>(LoopStart) &&
1307 SE->isLoopInvariant(LoopStart, L))
1308 LoopCost = 0;
1309 }
1310 }
1311 }
1312 C.AddRecCost += LoopCost;
1313
1314 // Add the step value register, if it needs one.
1315 // TODO: The non-affine case isn't precisely modeled here.
1316 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1317 if (!Regs.count(AR->getOperand(1))) {
1318 RateRegister(F, AR->getOperand(1), Regs);
1319 if (isLoser())
1320 return;
1321 }
1322 }
1323 }
1324 ++C.NumRegs;
1325
1326 // Rough heuristic; favor registers which don't require extra setup
1327 // instructions in the preheader.
1328 C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1329 // Ensure we don't, even with the recusion limit, produce invalid costs.
1330 C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1331
1332 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1333 SE->hasComputableLoopEvolution(Reg, L);
1334}
1335
1336/// Record this register in the set. If we haven't seen it before, rate
1337/// it. Optional LoserRegs provides a way to declare any formula that refers to
1338/// one of those regs an instant loser.
1339void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1340 SmallPtrSetImpl<const SCEV *> &Regs,
1341 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1342 if (LoserRegs && LoserRegs->count(Reg)) {
1343 Lose();
1344 return;
1345 }
1346 if (Regs.insert(Reg).second) {
1347 RateRegister(F, Reg, Regs);
1348 if (LoserRegs && isLoser())
1349 LoserRegs->insert(Reg);
1350 }
1351}
1352
1353void Cost::RateFormula(const Formula &F,
1354 SmallPtrSetImpl<const SCEV *> &Regs,
1355 const DenseSet<const SCEV *> &VisitedRegs,
1356 const LSRUse &LU,
1357 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1358 if (isLoser())
1359 return;
1360 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1360, __extension__
__PRETTY_FUNCTION__))
;
1361 // Tally up the registers.
1362 unsigned PrevAddRecCost = C.AddRecCost;
1363 unsigned PrevNumRegs = C.NumRegs;
1364 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1365 if (const SCEV *ScaledReg = F.ScaledReg) {
1366 if (VisitedRegs.count(ScaledReg)) {
1367 Lose();
1368 return;
1369 }
1370 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1371 if (isLoser())
1372 return;
1373 }
1374 for (const SCEV *BaseReg : F.BaseRegs) {
1375 if (VisitedRegs.count(BaseReg)) {
1376 Lose();
1377 return;
1378 }
1379 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1380 if (isLoser())
1381 return;
1382 }
1383
1384 // Determine how many (unfolded) adds we'll need inside the loop.
1385 size_t NumBaseParts = F.getNumRegs();
1386 if (NumBaseParts > 1)
1387 // Do not count the base and a possible second register if the target
1388 // allows to fold 2 registers.
1389 C.NumBaseAdds +=
1390 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1391 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1392
1393 // Accumulate non-free scaling amounts.
1394 C.ScaleCost += *getScalingFactorCost(*TTI, LU, F, *L).getValue();
1395
1396 // Tally up the non-zero immediates.
1397 for (const LSRFixup &Fixup : LU.Fixups) {
1398 int64_t O = Fixup.Offset;
1399 int64_t Offset = (uint64_t)O + F.BaseOffset;
1400 if (F.BaseGV)
1401 C.ImmCost += 64; // Handle symbolic values conservatively.
1402 // TODO: This should probably be the pointer size.
1403 else if (Offset != 0)
1404 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1405
1406 // Check with target if this offset with this instruction is
1407 // specifically not supported.
1408 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1409 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1410 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1411 C.NumBaseAdds++;
1412 }
1413
1414 // If we don't count instruction cost exit here.
1415 if (!InsnsCost) {
1416 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1416, __extension__
__PRETTY_FUNCTION__))
;
1417 return;
1418 }
1419
1420 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1421 // additional instruction (at least fill).
1422 // TODO: Need distinguish register class?
1423 unsigned TTIRegNum = TTI->getNumberOfRegisters(
1424 TTI->getRegisterClassForType(false, F.getType())) - 1;
1425 if (C.NumRegs > TTIRegNum) {
1426 // Cost already exceeded TTIRegNum, then only newly added register can add
1427 // new instructions.
1428 if (PrevNumRegs > TTIRegNum)
1429 C.Insns += (C.NumRegs - PrevNumRegs);
1430 else
1431 C.Insns += (C.NumRegs - TTIRegNum);
1432 }
1433
1434 // If ICmpZero formula ends with not 0, it could not be replaced by
1435 // just add or sub. We'll need to compare final result of AddRec.
1436 // That means we'll need an additional instruction. But if the target can
1437 // macro-fuse a compare with a branch, don't count this extra instruction.
1438 // For -10 + {0, +, 1}:
1439 // i = i + 1;
1440 // cmp i, 10
1441 //
1442 // For {-10, +, 1}:
1443 // i = i + 1;
1444 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1445 !TTI->canMacroFuseCmp())
1446 C.Insns++;
1447 // Each new AddRec adds 1 instruction to calculation.
1448 C.Insns += (C.AddRecCost - PrevAddRecCost);
1449
1450 // BaseAdds adds instructions for unfolded registers.
1451 if (LU.Kind != LSRUse::ICmpZero)
1452 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1453 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1453, __extension__
__PRETTY_FUNCTION__))
;
1454}
1455
1456/// Set this cost to a losing value.
1457void Cost::Lose() {
1458 C.Insns = std::numeric_limits<unsigned>::max();
1459 C.NumRegs = std::numeric_limits<unsigned>::max();
1460 C.AddRecCost = std::numeric_limits<unsigned>::max();
1461 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1462 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1463 C.ImmCost = std::numeric_limits<unsigned>::max();
1464 C.SetupCost = std::numeric_limits<unsigned>::max();
1465 C.ScaleCost = std::numeric_limits<unsigned>::max();
1466}
1467
1468/// Choose the lower cost.
1469bool Cost::isLess(const Cost &Other) {
1470 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1471 C.Insns != Other.C.Insns)
1472 return C.Insns < Other.C.Insns;
1473 return TTI->isLSRCostLess(C, Other.C);
1474}
1475
1476#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1477void Cost::print(raw_ostream &OS) const {
1478 if (InsnsCost)
1479 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1480 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1481 if (C.AddRecCost != 0)
1482 OS << ", with addrec cost " << C.AddRecCost;
1483 if (C.NumIVMuls != 0)
1484 OS << ", plus " << C.NumIVMuls << " IV mul"
1485 << (C.NumIVMuls == 1 ? "" : "s");
1486 if (C.NumBaseAdds != 0)
1487 OS << ", plus " << C.NumBaseAdds << " base add"
1488 << (C.NumBaseAdds == 1 ? "" : "s");
1489 if (C.ScaleCost != 0)
1490 OS << ", plus " << C.ScaleCost << " scale cost";
1491 if (C.ImmCost != 0)
1492 OS << ", plus " << C.ImmCost << " imm cost";
1493 if (C.SetupCost != 0)
1494 OS << ", plus " << C.SetupCost << " setup cost";
1495}
1496
1497LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1498 print(errs()); errs() << '\n';
1499}
1500#endif
1501
1502/// Test whether this fixup always uses its value outside of the given loop.
1503bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1504 // PHI nodes use their value in their incoming blocks.
1505 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1506 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1507 if (PN->getIncomingValue(i) == OperandValToReplace &&
1508 L->contains(PN->getIncomingBlock(i)))
1509 return false;
1510 return true;
1511 }
1512
1513 return !L->contains(UserInst);
1514}
1515
1516#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1517void LSRFixup::print(raw_ostream &OS) const {
1518 OS << "UserInst=";
1519 // Store is common and interesting enough to be worth special-casing.
1520 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1521 OS << "store ";
1522 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1523 } else if (UserInst->getType()->isVoidTy())
1524 OS << UserInst->getOpcodeName();
1525 else
1526 UserInst->printAsOperand(OS, /*PrintType=*/false);
1527
1528 OS << ", OperandValToReplace=";
1529 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1530
1531 for (const Loop *PIL : PostIncLoops) {
1532 OS << ", PostIncLoop=";
1533 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1534 }
1535
1536 if (Offset != 0)
1537 OS << ", Offset=" << Offset;
1538}
1539
1540LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1541 print(errs()); errs() << '\n';
1542}
1543#endif
1544
1545/// Test whether this use as a formula which has the same registers as the given
1546/// formula.
1547bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1548 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1549 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1550 // Unstable sort by host order ok, because this is only used for uniquifying.
1551 llvm::sort(Key);
1552 return Uniquifier.count(Key);
1553}
1554
1555/// The function returns a probability of selecting formula without Reg.
1556float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1557 unsigned FNum = 0;
1558 for (const Formula &F : Formulae)
1559 if (F.referencesReg(Reg))
1560 FNum++;
1561 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1562}
1563
1564/// If the given formula has not yet been inserted, add it to the list, and
1565/// return true. Return false otherwise. The formula must be in canonical form.
1566bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1567 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1567, __extension__
__PRETTY_FUNCTION__))
;
1568
1569 if (!Formulae.empty() && RigidFormula)
1570 return false;
1571
1572 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1573 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1574 // Unstable sort by host order ok, because this is only used for uniquifying.
1575 llvm::sort(Key);
1576
1577 if (!Uniquifier.insert(Key).second)
1578 return false;
1579
1580 // Using a register to hold the value of 0 is not profitable.
1581 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1582, __extension__
__PRETTY_FUNCTION__))
1582 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1582, __extension__
__PRETTY_FUNCTION__))
;
1583#ifndef NDEBUG
1584 for (const SCEV *BaseReg : F.BaseRegs)
1585 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1585, __extension__
__PRETTY_FUNCTION__))
;
1586#endif
1587
1588 // Add the formula to the list.
1589 Formulae.push_back(F);
1590
1591 // Record registers now being used by this use.
1592 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1593 if (F.ScaledReg)
1594 Regs.insert(F.ScaledReg);
1595
1596 return true;
1597}
1598
1599/// Remove the given formula from this use's list.
1600void LSRUse::DeleteFormula(Formula &F) {
1601 if (&F != &Formulae.back())
1602 std::swap(F, Formulae.back());
1603 Formulae.pop_back();
1604}
1605
1606/// Recompute the Regs field, and update RegUses.
1607void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1608 // Now that we've filtered out some formulae, recompute the Regs set.
1609 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1610 Regs.clear();
1611 for (const Formula &F : Formulae) {
1612 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1613 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1614 }
1615
1616 // Update the RegTracker.
1617 for (const SCEV *S : OldRegs)
1618 if (!Regs.count(S))
1619 RegUses.dropRegister(S, LUIdx);
1620}
1621
1622#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1623void LSRUse::print(raw_ostream &OS) const {
1624 OS << "LSR Use: Kind=";
1625 switch (Kind) {
1626 case Basic: OS << "Basic"; break;
1627 case Special: OS << "Special"; break;
1628 case ICmpZero: OS << "ICmpZero"; break;
1629 case Address:
1630 OS << "Address of ";
1631 if (AccessTy.MemTy->isPointerTy())
1632 OS << "pointer"; // the full pointer type could be really verbose
1633 else {
1634 OS << *AccessTy.MemTy;
1635 }
1636
1637 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1638 }
1639
1640 OS << ", Offsets={";
1641 bool NeedComma = false;
1642 for (const LSRFixup &Fixup : Fixups) {
1643 if (NeedComma) OS << ',';
1644 OS << Fixup.Offset;
1645 NeedComma = true;
1646 }
1647 OS << '}';
1648
1649 if (AllFixupsOutsideLoop)
1650 OS << ", all-fixups-outside-loop";
1651
1652 if (WidestFixupType)
1653 OS << ", widest fixup type: " << *WidestFixupType;
1654}
1655
1656LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1657 print(errs()); errs() << '\n';
1658}
1659#endif
1660
1661static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1662 LSRUse::KindType Kind, MemAccessTy AccessTy,
1663 GlobalValue *BaseGV, int64_t BaseOffset,
1664 bool HasBaseReg, int64_t Scale,
1665 Instruction *Fixup/*= nullptr*/) {
1666 switch (Kind) {
1667 case LSRUse::Address:
1668 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1669 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1670
1671 case LSRUse::ICmpZero:
1672 // There's not even a target hook for querying whether it would be legal to
1673 // fold a GV into an ICmp.
1674 if (BaseGV)
1675 return false;
1676
1677 // ICmp only has two operands; don't allow more than two non-trivial parts.
1678 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1679 return false;
1680
1681 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1682 // putting the scaled register in the other operand of the icmp.
1683 if (Scale != 0 && Scale != -1)
1684 return false;
1685
1686 // If we have low-level target information, ask the target if it can fold an
1687 // integer immediate on an icmp.
1688 if (BaseOffset != 0) {
1689 // We have one of:
1690 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1691 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1692 // Offs is the ICmp immediate.
1693 if (Scale == 0)
1694 // The cast does the right thing with
1695 // std::numeric_limits<int64_t>::min().
1696 BaseOffset = -(uint64_t)BaseOffset;
1697 return TTI.isLegalICmpImmediate(BaseOffset);
1698 }
1699
1700 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1701 return true;
1702
1703 case LSRUse::Basic:
1704 // Only handle single-register values.
1705 return !BaseGV && Scale == 0 && BaseOffset == 0;
1706
1707 case LSRUse::Special:
1708 // Special case Basic to handle -1 scales.
1709 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1710 }
1711
1712 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1712)
;
1713}
1714
1715static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1716 int64_t MinOffset, int64_t MaxOffset,
1717 LSRUse::KindType Kind, MemAccessTy AccessTy,
1718 GlobalValue *BaseGV, int64_t BaseOffset,
1719 bool HasBaseReg, int64_t Scale) {
1720 // Check for overflow.
1721 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1722 (MinOffset > 0))
1723 return false;
1724 MinOffset = (uint64_t)BaseOffset + MinOffset;
1725 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1726 (MaxOffset > 0))
1727 return false;
1728 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1729
1730 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1731 HasBaseReg, Scale) &&
1732 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1733 HasBaseReg, Scale);
1734}
1735
1736static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1737 int64_t MinOffset, int64_t MaxOffset,
1738 LSRUse::KindType Kind, MemAccessTy AccessTy,
1739 const Formula &F, const Loop &L) {
1740 // For the purpose of isAMCompletelyFolded either having a canonical formula
1741 // or a scale not equal to zero is correct.
1742 // Problems may arise from non canonical formulae having a scale == 0.
1743 // Strictly speaking it would best to just rely on canonical formulae.
1744 // However, when we generate the scaled formulae, we first check that the
1745 // scaling factor is profitable before computing the actual ScaledReg for
1746 // compile time sake.
1747 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)"
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1747, __extension__
__PRETTY_FUNCTION__))
;
1748 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1749 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1750}
1751
1752/// Test whether we know how to expand the current formula.
1753static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1754 int64_t MaxOffset, LSRUse::KindType Kind,
1755 MemAccessTy AccessTy, GlobalValue *BaseGV,
1756 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1757 // We know how to expand completely foldable formulae.
1758 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1759 BaseOffset, HasBaseReg, Scale) ||
1760 // Or formulae that use a base register produced by a sum of base
1761 // registers.
1762 (Scale == 1 &&
1763 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1764 BaseGV, BaseOffset, true, 0));
1765}
1766
1767static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1768 int64_t MaxOffset, LSRUse::KindType Kind,
1769 MemAccessTy AccessTy, const Formula &F) {
1770 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1771 F.BaseOffset, F.HasBaseReg, F.Scale);
1772}
1773
1774static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1775 const LSRUse &LU, const Formula &F) {
1776 // Target may want to look at the user instructions.
1777 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1778 for (const LSRFixup &Fixup : LU.Fixups)
1779 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1780 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1781 F.Scale, Fixup.UserInst))
1782 return false;
1783 return true;
1784 }
1785
1786 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1787 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1788 F.Scale);
1789}
1790
1791static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1792 const LSRUse &LU, const Formula &F,
1793 const Loop &L) {
1794 if (!F.Scale)
1795 return 0;
1796
1797 // If the use is not completely folded in that instruction, we will have to
1798 // pay an extra cost only for scale != 1.
1799 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1800 LU.AccessTy, F, L))
1801 return F.Scale != 1;
1802
1803 switch (LU.Kind) {
1804 case LSRUse::Address: {
1805 // Check the scaling factor cost with both the min and max offsets.
1806 InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost(
1807 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1808 F.Scale, LU.AccessTy.AddrSpace);
1809 InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost(
1810 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1811 F.Scale, LU.AccessTy.AddrSpace);
1812
1813 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1814, __extension__
__PRETTY_FUNCTION__))
1814 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1814, __extension__
__PRETTY_FUNCTION__))
;
1815 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1816 }
1817 case LSRUse::ICmpZero:
1818 case LSRUse::Basic:
1819 case LSRUse::Special:
1820 // The use is completely folded, i.e., everything is folded into the
1821 // instruction.
1822 return 0;
1823 }
1824
1825 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1825)
;
1826}
1827
1828static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1829 LSRUse::KindType Kind, MemAccessTy AccessTy,
1830 GlobalValue *BaseGV, int64_t BaseOffset,
1831 bool HasBaseReg) {
1832 // Fast-path: zero is always foldable.
1833 if (BaseOffset == 0 && !BaseGV) return true;
1834
1835 // Conservatively, create an address with an immediate and a
1836 // base and a scale.
1837 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1838
1839 // Canonicalize a scale of 1 to a base register if the formula doesn't
1840 // already have a base register.
1841 if (!HasBaseReg && Scale == 1) {
1842 Scale = 0;
1843 HasBaseReg = true;
1844 }
1845
1846 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1847 HasBaseReg, Scale);
1848}
1849
1850static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1851 ScalarEvolution &SE, int64_t MinOffset,
1852 int64_t MaxOffset, LSRUse::KindType Kind,
1853 MemAccessTy AccessTy, const SCEV *S,
1854 bool HasBaseReg) {
1855 // Fast-path: zero is always foldable.
1856 if (S->isZero()) return true;
1857
1858 // Conservatively, create an address with an immediate and a
1859 // base and a scale.
1860 int64_t BaseOffset = ExtractImmediate(S, SE);
1861 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1862
1863 // If there's anything else involved, it's not foldable.
1864 if (!S->isZero()) return false;
1865
1866 // Fast-path: zero is always foldable.
1867 if (BaseOffset == 0 && !BaseGV) return true;
1868
1869 // Conservatively, create an address with an immediate and a
1870 // base and a scale.
1871 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1872
1873 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1874 BaseOffset, HasBaseReg, Scale);
1875}
1876
1877namespace {
1878
1879/// An individual increment in a Chain of IV increments. Relate an IV user to
1880/// an expression that computes the IV it uses from the IV used by the previous
1881/// link in the Chain.
1882///
1883/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1884/// original IVOperand. The head of the chain's IVOperand is only valid during
1885/// chain collection, before LSR replaces IV users. During chain generation,
1886/// IncExpr can be used to find the new IVOperand that computes the same
1887/// expression.
1888struct IVInc {
1889 Instruction *UserInst;
1890 Value* IVOperand;
1891 const SCEV *IncExpr;
1892
1893 IVInc(Instruction *U, Value *O, const SCEV *E)
1894 : UserInst(U), IVOperand(O), IncExpr(E) {}
1895};
1896
1897// The list of IV increments in program order. We typically add the head of a
1898// chain without finding subsequent links.
1899struct IVChain {
1900 SmallVector<IVInc, 1> Incs;
1901 const SCEV *ExprBase = nullptr;
1902
1903 IVChain() = default;
1904 IVChain(const IVInc &Head, const SCEV *Base)
1905 : Incs(1, Head), ExprBase(Base) {}
1906
1907 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1908
1909 // Return the first increment in the chain.
1910 const_iterator begin() const {
1911 assert(!Incs.empty())(static_cast <bool> (!Incs.empty()) ? void (0) : __assert_fail
("!Incs.empty()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1911, __extension__ __PRETTY_FUNCTION__))
;
1912 return std::next(Incs.begin());
1913 }
1914 const_iterator end() const {
1915 return Incs.end();
1916 }
1917
1918 // Returns true if this chain contains any increments.
1919 bool hasIncs() const { return Incs.size() >= 2; }
1920
1921 // Add an IVInc to the end of this chain.
1922 void add(const IVInc &X) { Incs.push_back(X); }
1923
1924 // Returns the last UserInst in the chain.
1925 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1926
1927 // Returns true if IncExpr can be profitably added to this chain.
1928 bool isProfitableIncrement(const SCEV *OperExpr,
1929 const SCEV *IncExpr,
1930 ScalarEvolution&);
1931};
1932
1933/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1934/// between FarUsers that definitely cross IV increments and NearUsers that may
1935/// be used between IV increments.
1936struct ChainUsers {
1937 SmallPtrSet<Instruction*, 4> FarUsers;
1938 SmallPtrSet<Instruction*, 4> NearUsers;
1939};
1940
1941/// This class holds state for the main loop strength reduction logic.
1942class LSRInstance {
1943 IVUsers &IU;
1944 ScalarEvolution &SE;
1945 DominatorTree &DT;
1946 LoopInfo &LI;
1947 AssumptionCache &AC;
1948 TargetLibraryInfo &TLI;
1949 const TargetTransformInfo &TTI;
1950 Loop *const L;
1951 MemorySSAUpdater *MSSAU;
1952 TTI::AddressingModeKind AMK;
1953 mutable SCEVExpander Rewriter;
1954 bool Changed = false;
1955
1956 /// This is the insert position that the current loop's induction variable
1957 /// increment should be placed. In simple loops, this is the latch block's
1958 /// terminator. But in more complicated cases, this is a position which will
1959 /// dominate all the in-loop post-increment users.
1960 Instruction *IVIncInsertPos = nullptr;
1961
1962 /// Interesting factors between use strides.
1963 ///
1964 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1965 /// default, a SmallDenseSet, because we need to use the full range of
1966 /// int64_ts, and there's currently no good way of doing that with
1967 /// SmallDenseSet.
1968 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1969
1970 /// Interesting use types, to facilitate truncation reuse.
1971 SmallSetVector<Type *, 4> Types;
1972
1973 /// The list of interesting uses.
1974 mutable SmallVector<LSRUse, 16> Uses;
1975
1976 /// Track which uses use which register candidates.
1977 RegUseTracker RegUses;
1978
1979 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1980 // have more than a few IV increment chains in a loop. Missing a Chain falls
1981 // back to normal LSR behavior for those uses.
1982 static const unsigned MaxChains = 8;
1983
1984 /// IV users can form a chain of IV increments.
1985 SmallVector<IVChain, MaxChains> IVChainVec;
1986
1987 /// IV users that belong to profitable IVChains.
1988 SmallPtrSet<Use*, MaxChains> IVIncSet;
1989
1990 /// Induction variables that were generated and inserted by the SCEV Expander.
1991 SmallVector<llvm::WeakVH, 2> ScalarEvolutionIVs;
1992
1993 void OptimizeShadowIV();
1994 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1995 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1996 void OptimizeLoopTermCond();
1997
1998 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1999 SmallVectorImpl<ChainUsers> &ChainUsersVec);
2000 void FinalizeChain(IVChain &Chain);
2001 void CollectChains();
2002 void GenerateIVChain(const IVChain &Chain,
2003 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
2004
2005 void CollectInterestingTypesAndFactors();
2006 void CollectFixupsAndInitialFormulae();
2007
2008 // Support for sharing of LSRUses between LSRFixups.
2009 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
2010 UseMapTy UseMap;
2011
2012 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2013 LSRUse::KindType Kind, MemAccessTy AccessTy);
2014
2015 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
2016 MemAccessTy AccessTy);
2017
2018 void DeleteUse(LSRUse &LU, size_t LUIdx);
2019
2020 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
2021
2022 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2023 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2024 void CountRegisters(const Formula &F, size_t LUIdx);
2025 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
2026
2027 void CollectLoopInvariantFixupsAndFormulae();
2028
2029 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
2030 unsigned Depth = 0);
2031
2032 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
2033 const Formula &Base, unsigned Depth,
2034 size_t Idx, bool IsScaledReg = false);
2035 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
2036 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2037 const Formula &Base, size_t Idx,
2038 bool IsScaledReg = false);
2039 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2040 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2041 const Formula &Base,
2042 const SmallVectorImpl<int64_t> &Worklist,
2043 size_t Idx, bool IsScaledReg = false);
2044 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2045 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2046 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2047 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2048 void GenerateCrossUseConstantOffsets();
2049 void GenerateAllReuseFormulae();
2050
2051 void FilterOutUndesirableDedicatedRegisters();
2052
2053 size_t EstimateSearchSpaceComplexity() const;
2054 void NarrowSearchSpaceByDetectingSupersets();
2055 void NarrowSearchSpaceByCollapsingUnrolledCode();
2056 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2057 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2058 void NarrowSearchSpaceByFilterPostInc();
2059 void NarrowSearchSpaceByDeletingCostlyFormulas();
2060 void NarrowSearchSpaceByPickingWinnerRegs();
2061 void NarrowSearchSpaceUsingHeuristics();
2062
2063 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2064 Cost &SolutionCost,
2065 SmallVectorImpl<const Formula *> &Workspace,
2066 const Cost &CurCost,
2067 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2068 DenseSet<const SCEV *> &VisitedRegs) const;
2069 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2070
2071 BasicBlock::iterator
2072 HoistInsertPosition(BasicBlock::iterator IP,
2073 const SmallVectorImpl<Instruction *> &Inputs) const;
2074 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2075 const LSRFixup &LF,
2076 const LSRUse &LU) const;
2077
2078 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2079 BasicBlock::iterator IP,
2080 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2081 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2082 const Formula &F,
2083 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2084 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2085 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2086 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2087
2088public:
2089 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2090 LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2091 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);
2092
2093 bool getChanged() const { return Changed; }
2094 const SmallVectorImpl<WeakVH> &getScalarEvolutionIVs() const {
2095 return ScalarEvolutionIVs;
2096 }
2097
2098 void print_factors_and_types(raw_ostream &OS) const;
2099 void print_fixups(raw_ostream &OS) const;
2100 void print_uses(raw_ostream &OS) const;
2101 void print(raw_ostream &OS) const;
2102 void dump() const;
2103};
2104
2105} // end anonymous namespace
2106
2107/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2108/// the cast operation.
2109void LSRInstance::OptimizeShadowIV() {
2110 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2111 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2112 return;
2113
2114 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2115 UI != E; /* empty */) {
2116 IVUsers::const_iterator CandidateUI = UI;
2117 ++UI;
2118 Instruction *ShadowUse = CandidateUI->getUser();
2119 Type *DestTy = nullptr;
2120 bool IsSigned = false;
2121
2122 /* If shadow use is a int->float cast then insert a second IV
2123 to eliminate this cast.
2124
2125 for (unsigned i = 0; i < n; ++i)
2126 foo((double)i);
2127
2128 is transformed into
2129
2130 double d = 0.0;
2131 for (unsigned i = 0; i < n; ++i, ++d)
2132 foo(d);
2133 */
2134 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2135 IsSigned = false;
2136 DestTy = UCast->getDestTy();
2137 }
2138 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2139 IsSigned = true;
2140 DestTy = SCast->getDestTy();
2141 }
2142 if (!DestTy) continue;
2143
2144 // If target does not support DestTy natively then do not apply
2145 // this transformation.
2146 if (!TTI.isTypeLegal(DestTy)) continue;
2147
2148 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2149 if (!PH) continue;
2150 if (PH->getNumIncomingValues() != 2) continue;
2151
2152 // If the calculation in integers overflows, the result in FP type will
2153 // differ. So we only can do this transformation if we are guaranteed to not
2154 // deal with overflowing values
2155 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2156 if (!AR) continue;
2157 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2158 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2159
2160 Type *SrcTy = PH->getType();
2161 int Mantissa = DestTy->getFPMantissaWidth();
2162 if (Mantissa == -1) continue;
2163 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2164 continue;
2165
2166 unsigned Entry, Latch;
2167 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2168 Entry = 0;
2169 Latch = 1;
2170 } else {
2171 Entry = 1;
2172 Latch = 0;
2173 }
2174
2175 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2176 if (!Init) continue;
2177 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2178 (double)Init->getSExtValue() :
2179 (double)Init->getZExtValue());
2180
2181 BinaryOperator *Incr =
2182 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2183 if (!Incr) continue;
2184 if (Incr->getOpcode() != Instruction::Add
2185 && Incr->getOpcode() != Instruction::Sub)
2186 continue;
2187
2188 /* Initialize new IV, double d = 0.0 in above example. */
2189 ConstantInt *C = nullptr;
2190 if (Incr->getOperand(0) == PH)
2191 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2192 else if (Incr->getOperand(1) == PH)
2193 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2194 else
2195 continue;
2196
2197 if (!C) continue;
2198
2199 // Ignore negative constants, as the code below doesn't handle them
2200 // correctly. TODO: Remove this restriction.
2201 if (!C->getValue().isStrictlyPositive()) continue;
2202
2203 /* Add new PHINode. */
2204 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2205
2206 /* create new increment. '++d' in above example. */
2207 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2208 BinaryOperator *NewIncr =
2209 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2210 Instruction::FAdd : Instruction::FSub,
2211 NewPH, CFP, "IV.S.next.", Incr);
2212
2213 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2214 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2215
2216 /* Remove cast operation */
2217 ShadowUse->replaceAllUsesWith(NewPH);
2218 ShadowUse->eraseFromParent();
2219 Changed = true;
2220 break;
2221 }
2222}
2223
2224/// If Cond has an operand that is an expression of an IV, set the IV user and
2225/// stride information and return true, otherwise return false.
2226bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2227 for (IVStrideUse &U : IU)
2228 if (U.getUser() == Cond) {
2229 // NOTE: we could handle setcc instructions with multiple uses here, but
2230 // InstCombine does it as well for simple uses, it's not clear that it
2231 // occurs enough in real life to handle.
2232 CondUse = &U;
2233 return true;
2234 }
2235 return false;
2236}
2237
2238/// Rewrite the loop's terminating condition if it uses a max computation.
2239///
2240/// This is a narrow solution to a specific, but acute, problem. For loops
2241/// like this:
2242///
2243/// i = 0;
2244/// do {
2245/// p[i] = 0.0;
2246/// } while (++i < n);
2247///
2248/// the trip count isn't just 'n', because 'n' might not be positive. And
2249/// unfortunately this can come up even for loops where the user didn't use
2250/// a C do-while loop. For example, seemingly well-behaved top-test loops
2251/// will commonly be lowered like this:
2252///
2253/// if (n > 0) {
2254/// i = 0;
2255/// do {
2256/// p[i] = 0.0;
2257/// } while (++i < n);
2258/// }
2259///
2260/// and then it's possible for subsequent optimization to obscure the if
2261/// test in such a way that indvars can't find it.
2262///
2263/// When indvars can't find the if test in loops like this, it creates a
2264/// max expression, which allows it to give the loop a canonical
2265/// induction variable:
2266///
2267/// i = 0;
2268/// max = n < 1 ? 1 : n;
2269/// do {
2270/// p[i] = 0.0;
2271/// } while (++i != max);
2272///
2273/// Canonical induction variables are necessary because the loop passes
2274/// are designed around them. The most obvious example of this is the
2275/// LoopInfo analysis, which doesn't remember trip count values. It
2276/// expects to be able to rediscover the trip count each time it is
2277/// needed, and it does this using a simple analysis that only succeeds if
2278/// the loop has a canonical induction variable.
2279///
2280/// However, when it comes time to generate code, the maximum operation
2281/// can be quite costly, especially if it's inside of an outer loop.
2282///
2283/// This function solves this problem by detecting this type of loop and
2284/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2285/// the instructions for the maximum computation.
2286ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2287 // Check that the loop matches the pattern we're looking for.
2288 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2289 Cond->getPredicate() != CmpInst::ICMP_NE)
2290 return Cond;
2291
2292 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2293 if (!Sel || !Sel->hasOneUse()) return Cond;
2294
2295 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2296 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2297 return Cond;
2298 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2299
2300 // Add one to the backedge-taken count to get the trip count.
2301 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2302 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2303
2304 // Check for a max calculation that matches the pattern. There's no check
2305 // for ICMP_ULE here because the comparison would be with zero, which
2306 // isn't interesting.
2307 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2308 const SCEVNAryExpr *Max = nullptr;
2309 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2310 Pred = ICmpInst::ICMP_SLE;
2311 Max = S;
2312 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2313 Pred = ICmpInst::ICMP_SLT;
2314 Max = S;
2315 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2316 Pred = ICmpInst::ICMP_ULT;
2317 Max = U;
2318 } else {
2319 // No match; bail.
2320 return Cond;
2321 }
2322
2323 // To handle a max with more than two operands, this optimization would
2324 // require additional checking and setup.
2325 if (Max->getNumOperands() != 2)
2326 return Cond;
2327
2328 const SCEV *MaxLHS = Max->getOperand(0);
2329 const SCEV *MaxRHS = Max->getOperand(1);
2330
2331 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2332 // for a comparison with 1. For <= and >=, a comparison with zero.
2333 if (!MaxLHS ||
2334 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2335 return Cond;
2336
2337 // Check the relevant induction variable for conformance to
2338 // the pattern.
2339 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2340 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2341 if (!AR || !AR->isAffine() ||
2342 AR->getStart() != One ||
2343 AR->getStepRecurrence(SE) != One)
2344 return Cond;
2345
2346 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 2347, __extension__
__PRETTY_FUNCTION__))
2347 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 2347, __extension__
__PRETTY_FUNCTION__))
;
2348
2349 // Check the right operand of the select, and remember it, as it will
2350 // be used in the new comparison instruction.
2351 Value *NewRHS = nullptr;
2352 if (ICmpInst::isTrueWhenEqual(Pred)) {
2353 // Look for n+1, and grab n.
2354 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2355 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2356 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2357 NewRHS = BO->getOperand(0);
2358 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2359 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2360 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2361 NewRHS = BO->getOperand(0);
2362 if (!NewRHS)
2363 return Cond;
2364 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2365 NewRHS = Sel->getOperand(1);
2366 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2367 NewRHS = Sel->getOperand(2);
2368 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2369 NewRHS = SU->getValue();
2370 else
2371 // Max doesn't match expected pattern.
2372 return Cond;
2373
2374 // Determine the new comparison opcode. It may be signed or unsigned,
2375 // and the original comparison may be either equality or inequality.
2376 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2377 Pred = CmpInst::getInversePredicate(Pred);
2378
2379 // Ok, everything looks ok to change the condition into an SLT or SGE and
2380 // delete the max calculation.
2381 ICmpInst *NewCond =
2382 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2383
2384 // Delete the max calculation instructions.
2385 NewCond->setDebugLoc(Cond->getDebugLoc());
2386 Cond->replaceAllUsesWith(NewCond);
2387 CondUse->setUser(NewCond);
2388 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2389 Cond->eraseFromParent();
2390 Sel->eraseFromParent();
2391 if (Cmp->use_empty())
2392 Cmp->eraseFromParent();
2393 return NewCond;
2394}
2395
2396/// Change loop terminating condition to use the postinc iv when possible.
2397void
2398LSRInstance::OptimizeLoopTermCond() {
2399 SmallPtrSet<Instruction *, 4> PostIncs;
2400
2401 // We need a different set of heuristics for rotated and non-rotated loops.
2402 // If a loop is rotated then the latch is also the backedge, so inserting
2403 // post-inc expressions just before the latch is ideal. To reduce live ranges
2404 // it also makes sense to rewrite terminating conditions to use post-inc
2405 // expressions.
2406 //
2407 // If the loop is not rotated then the latch is not a backedge; the latch
2408 // check is done in the loop head. Adding post-inc expressions before the
2409 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2410 // in the loop body. In this case we do *not* want to use post-inc expressions
2411 // in the latch check, and we want to insert post-inc expressions before
2412 // the backedge.
2413 BasicBlock *LatchBlock = L->getLoopLatch();
2414 SmallVector<BasicBlock*, 8> ExitingBlocks;
2415 L->getExitingBlocks(ExitingBlocks);
2416 if (!llvm::is_contained(ExitingBlocks, LatchBlock)) {
2417 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2418 IVIncInsertPos = LatchBlock->getTerminator();
2419 return;
2420 }
2421
2422 // Otherwise treat this as a rotated loop.
2423 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2424 // Get the terminating condition for the loop if possible. If we
2425 // can, we want to change it to use a post-incremented version of its
2426 // induction variable, to allow coalescing the live ranges for the IV into
2427 // one register value.
2428
2429 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2430 if (!TermBr)
2431 continue;
2432 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2433 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2434 continue;
2435
2436 // Search IVUsesByStride to find Cond's IVUse if there is one.
2437 IVStrideUse *CondUse = nullptr;
2438 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2439 if (!FindIVUserForCond(Cond, CondUse))
2440 continue;
2441
2442 // If the trip count is computed in terms of a max (due to ScalarEvolution
2443 // being unable to find a sufficient guard, for example), change the loop
2444 // comparison to use SLT or ULT instead of NE.
2445 // One consequence of doing this now is that it disrupts the count-down
2446 // optimization. That's not always a bad thing though, because in such
2447 // cases it may still be worthwhile to avoid a max.
2448 Cond = OptimizeMax(Cond, CondUse);
2449
2450 // If this exiting block dominates the latch block, it may also use
2451 // the post-inc value if it won't be shared with other uses.
2452 // Check for dominance.
2453 if (!DT.dominates(ExitingBlock, LatchBlock))
2454 continue;
2455
2456 // Conservatively avoid trying to use the post-inc value in non-latch
2457 // exits if there may be pre-inc users in intervening blocks.
2458 if (LatchBlock != ExitingBlock)
2459 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2460 // Test if the use is reachable from the exiting block. This dominator
2461 // query is a conservative approximation of reachability.
2462 if (&*UI != CondUse &&
2463 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2464 // Conservatively assume there may be reuse if the quotient of their
2465 // strides could be a legal scale.
2466 const SCEV *A = IU.getStride(*CondUse, L);
2467 const SCEV *B = IU.getStride(*UI, L);
2468 if (!A || !B) continue;
2469 if (SE.getTypeSizeInBits(A->getType()) !=
2470 SE.getTypeSizeInBits(B->getType())) {
2471 if (SE.getTypeSizeInBits(A->getType()) >
2472 SE.getTypeSizeInBits(B->getType()))
2473 B = SE.getSignExtendExpr(B, A->getType());
2474 else
2475 A = SE.getSignExtendExpr(A, B->getType());
2476 }
2477 if (const SCEVConstant *D =
2478 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2479 const ConstantInt *C = D->getValue();
2480 // Stride of one or negative one can have reuse with non-addresses.
2481 if (C->isOne() || C->isMinusOne())
2482 goto decline_post_inc;
2483 // Avoid weird situations.
2484 if (C->getValue().getMinSignedBits() >= 64 ||
2485 C->getValue().isMinSignedValue())
2486 goto decline_post_inc;
2487 // Check for possible scaled-address reuse.
2488 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2489 MemAccessTy AccessTy = getAccessType(
2490 TTI, UI->getUser(), UI->getOperandValToReplace());
2491 int64_t Scale = C->getSExtValue();
2492 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2493 /*BaseOffset=*/0,
2494 /*HasBaseReg=*/false, Scale,
2495 AccessTy.AddrSpace))
2496 goto decline_post_inc;
2497 Scale = -Scale;
2498 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2499 /*BaseOffset=*/0,
2500 /*HasBaseReg=*/false, Scale,
2501 AccessTy.AddrSpace))
2502 goto decline_post_inc;
2503 }
2504 }
2505 }
2506
2507 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)
2508 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2509
2510 // It's possible for the setcc instruction to be anywhere in the loop, and
2511 // possible for it to have multiple users. If it is not immediately before
2512 // the exiting block branch, move it.
2513 if (Cond->getNextNonDebugInstruction() != TermBr) {
2514 if (Cond->hasOneUse()) {
2515 Cond->moveBefore(TermBr);
2516 } else {
2517 // Clone the terminating condition and insert into the loopend.
2518 ICmpInst *OldCond = Cond;
2519 Cond = cast<ICmpInst>(Cond->clone());
2520 Cond->setName(L->getHeader()->getName() + ".termcond");
2521 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2522
2523 // Clone the IVUse, as the old use still exists!
2524 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2525 TermBr->replaceUsesOfWith(OldCond, Cond);
2526 }
2527 }
2528
2529 // If we get to here, we know that we can transform the setcc instruction to
2530 // use the post-incremented version of the IV, allowing us to coalesce the
2531 // live ranges for the IV correctly.
2532 CondUse->transformToPostInc(L);
2533 Changed = true;
2534
2535 PostIncs.insert(Cond);
2536 decline_post_inc:;
2537 }
2538
2539 // Determine an insertion point for the loop induction variable increment. It
2540 // must dominate all the post-inc comparisons we just set up, and it must
2541 // dominate the loop latch edge.
2542 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2543 for (Instruction *Inst : PostIncs) {
2544 BasicBlock *BB =
2545 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2546 Inst->getParent());
2547 if (BB == Inst->getParent())
2548 IVIncInsertPos = Inst;
2549 else if (BB != IVIncInsertPos->getParent())
2550 IVIncInsertPos = BB->getTerminator();
2551 }
2552}
2553
2554/// Determine if the given use can accommodate a fixup at the given offset and
2555/// other details. If so, update the use and return true.
2556bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2557 bool HasBaseReg, LSRUse::KindType Kind,
2558 MemAccessTy AccessTy) {
2559 int64_t NewMinOffset = LU.MinOffset;
2560 int64_t NewMaxOffset = LU.MaxOffset;
2561 MemAccessTy NewAccessTy = AccessTy;
2562
2563 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2564 // something conservative, however this can pessimize in the case that one of
2565 // the uses will have all its uses outside the loop, for example.
2566 if (LU.Kind != Kind)
2567 return false;
2568
2569 // Check for a mismatched access type, and fall back conservatively as needed.
2570 // TODO: Be less conservative when the type is similar and can use the same
2571 // addressing modes.
2572 if (Kind == LSRUse::Address) {
2573 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2574 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2575 AccessTy.AddrSpace);
2576 }
2577 }
2578
2579 // Conservatively assume HasBaseReg is true for now.
2580 if (NewOffset < LU.MinOffset) {
2581 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2582 LU.MaxOffset - NewOffset, HasBaseReg))
2583 return false;
2584 NewMinOffset = NewOffset;
2585 } else if (NewOffset > LU.MaxOffset) {
2586 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2587 NewOffset - LU.MinOffset, HasBaseReg))
2588 return false;
2589 NewMaxOffset = NewOffset;
2590 }
2591
2592 // Update the use.
2593 LU.MinOffset = NewMinOffset;
2594 LU.MaxOffset = NewMaxOffset;
2595 LU.AccessTy = NewAccessTy;
2596 return true;
2597}
2598
2599/// Return an LSRUse index and an offset value for a fixup which needs the given
2600/// expression, with the given kind and optional access type. Either reuse an
2601/// existing use or create a new one, as needed.
2602std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2603 LSRUse::KindType Kind,
2604 MemAccessTy AccessTy) {
2605 const SCEV *Copy = Expr;
2606 int64_t Offset = ExtractImmediate(Expr, SE);
2607
2608 // Basic uses can't accept any offset, for example.
2609 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2610 Offset, /*HasBaseReg=*/ true)) {
2611 Expr = Copy;
2612 Offset = 0;
2613 }
2614
2615 std::pair<UseMapTy::iterator, bool> P =
2616 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2617 if (!P.second) {
2618 // A use already existed with this base.
2619 size_t LUIdx = P.first->second;
2620 LSRUse &LU = Uses[LUIdx];
2621 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2622 // Reuse this use.
2623 return std::make_pair(LUIdx, Offset);
2624 }
2625
2626 // Create a new use.
2627 size_t LUIdx = Uses.size();
2628 P.first->second = LUIdx;
2629 Uses.push_back(LSRUse(Kind, AccessTy));
2630 LSRUse &LU = Uses[LUIdx];
2631
2632 LU.MinOffset = Offset;
2633 LU.MaxOffset = Offset;
2634 return std::make_pair(LUIdx, Offset);
2635}
2636
2637/// Delete the given use from the Uses list.
2638void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2639 if (&LU != &Uses.back())
2640 std::swap(LU, Uses.back());
2641 Uses.pop_back();
2642
2643 // Update RegUses.
2644 RegUses.swapAndDropUse(LUIdx, Uses.size());
2645}
2646
2647/// Look for a use distinct from OrigLU which is has a formula that has the same
2648/// registers as the given formula.
2649LSRUse *
2650LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2651 const LSRUse &OrigLU) {
2652 // Search all uses for the formula. This could be more clever.
2653 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2654 LSRUse &LU = Uses[LUIdx];
2655 // Check whether this use is close enough to OrigLU, to see whether it's
2656 // worthwhile looking through its formulae.
2657 // Ignore ICmpZero uses because they may contain formulae generated by
2658 // GenerateICmpZeroScales, in which case adding fixup offsets may
2659 // be invalid.
2660 if (&LU != &OrigLU &&
2661 LU.Kind != LSRUse::ICmpZero &&
2662 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2663 LU.WidestFixupType == OrigLU.WidestFixupType &&
2664 LU.HasFormulaWithSameRegs(OrigF)) {
2665 // Scan through this use's formulae.
2666 for (const Formula &F : LU.Formulae) {
2667 // Check to see if this formula has the same registers and symbols
2668 // as OrigF.
2669 if (F.BaseRegs == OrigF.BaseRegs &&
2670 F.ScaledReg == OrigF.ScaledReg &&
2671 F.BaseGV == OrigF.BaseGV &&
2672 F.Scale == OrigF.Scale &&
2673 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2674 if (F.BaseOffset == 0)
2675 return &LU;
2676 // This is the formula where all the registers and symbols matched;
2677 // there aren't going to be any others. Since we declined it, we
2678 // can skip the rest of the formulae and proceed to the next LSRUse.
2679 break;
2680 }
2681 }
2682 }
2683 }
2684
2685 // Nothing looked good.
2686 return nullptr;
2687}
2688
2689void LSRInstance::CollectInterestingTypesAndFactors() {
2690 SmallSetVector<const SCEV *, 4> Strides;
2691
2692 // Collect interesting types and strides.
2693 SmallVector<const SCEV *, 4> Worklist;
2694 for (const IVStrideUse &U : IU) {
2695 const SCEV *Expr = IU.getExpr(U);
2696
2697 // Collect interesting types.
2698 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2699
2700 // Add strides for mentioned loops.
2701 Worklist.push_back(Expr);
2702 do {
2703 const SCEV *S = Worklist.pop_back_val();
2704 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2705 if (AR->getLoop() == L)
2706 Strides.insert(AR->getStepRecurrence(SE));
2707 Worklist.push_back(AR->getStart());
2708 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2709 Worklist.append(Add->op_begin(), Add->op_end());
2710 }
2711 } while (!Worklist.empty());
2712 }
2713
2714 // Compute interesting factors from the set of interesting strides.
2715 for (SmallSetVector<const SCEV *, 4>::const_iterator
2716 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2717 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2718 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2719 const SCEV *OldStride = *I;
2720 const SCEV *NewStride = *NewStrideIter;
2721
2722 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2723 SE.getTypeSizeInBits(NewStride->getType())) {
2724 if (SE.getTypeSizeInBits(OldStride->getType()) >
2725 SE.getTypeSizeInBits(NewStride->getType()))
2726 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2727 else
2728 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2729 }
2730 if (const SCEVConstant *Factor =
2731 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2732 SE, true))) {
2733 if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2734 Factors.insert(Factor->getAPInt().getSExtValue());
2735 } else if (const SCEVConstant *Factor =
2736 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2737 NewStride,
2738 SE, true))) {
2739 if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2740 Factors.insert(Factor->getAPInt().getSExtValue());
2741 }
2742 }
2743
2744 // If all uses use the same type, don't bother looking for truncation-based
2745 // reuse.
2746 if (Types.size() == 1)
2747 Types.clear();
2748
2749 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2750}
2751
2752/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2753/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2754/// IVStrideUses, we could partially skip this.
2755static User::op_iterator
2756findIVOperand(User::op_iterator OI, User::op_iterator OE,
2757 Loop *L, ScalarEvolution &SE) {
2758 for(; OI != OE; ++OI) {
2759 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2760 if (!SE.isSCEVable(Oper->getType()))
2761 continue;
2762
2763 if (const SCEVAddRecExpr *AR =
2764 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2765 if (AR->getLoop() == L)
2766 break;
2767 }
2768 }
2769 }
2770 return OI;
2771}
2772
2773/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2774/// a convenient helper.
2775static Value *getWideOperand(Value *Oper) {
2776 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2777 return Trunc->getOperand(0);
2778 return Oper;
2779}
2780
2781/// Return true if we allow an IV chain to include both types.
2782static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2783 Type *LType = LVal->getType();
2784 Type *RType = RVal->getType();
2785 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2786 // Different address spaces means (possibly)
2787 // different types of the pointer implementation,
2788 // e.g. i16 vs i32 so disallow that.
2789 (LType->getPointerAddressSpace() ==
2790 RType->getPointerAddressSpace()));
2791}
2792
2793/// Return an approximation of this SCEV expression's "base", or NULL for any
2794/// constant. Returning the expression itself is conservative. Returning a
2795/// deeper subexpression is more precise and valid as long as it isn't less
2796/// complex than another subexpression. For expressions involving multiple
2797/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2798/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2799/// IVInc==b-a.
2800///
2801/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2802/// SCEVUnknown, we simply return the rightmost SCEV operand.
2803static const SCEV *getExprBase(const SCEV *S) {
2804 switch (S->getSCEVType()) {
2805 default: // uncluding scUnknown.
2806 return S;
2807 case scConstant:
2808 return nullptr;
2809 case scTruncate:
2810 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2811 case scZeroExtend:
2812 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2813 case scSignExtend:
2814 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2815 case scAddExpr: {
2816 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2817 // there's nothing more complex.
2818 // FIXME: not sure if we want to recognize negation.
2819 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2820 for (const SCEV *SubExpr : reverse(Add->operands())) {
2821 if (SubExpr->getSCEVType() == scAddExpr)
2822 return getExprBase(SubExpr);
2823
2824 if (SubExpr->getSCEVType() != scMulExpr)
2825 return SubExpr;
2826 }
2827 return S; // all operands are scaled, be conservative.
2828 }
2829 case scAddRecExpr:
2830 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2831 }
2832 llvm_unreachable("Unknown SCEV kind!")::llvm::llvm_unreachable_internal("Unknown SCEV kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2832)
;
2833}
2834
2835/// Return true if the chain increment is profitable to expand into a loop
2836/// invariant value, which may require its own register. A profitable chain
2837/// increment will be an offset relative to the same base. We allow such offsets
2838/// to potentially be used as chain increment as long as it's not obviously
2839/// expensive to expand using real instructions.
2840bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2841 const SCEV *IncExpr,
2842 ScalarEvolution &SE) {
2843 // Aggressively form chains when -stress-ivchain.
2844 if (StressIVChain)
2845 return true;
2846
2847 // Do not replace a constant offset from IV head with a nonconstant IV
2848 // increment.
2849 if (!isa<SCEVConstant>(IncExpr)) {
2850 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2851 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2852 return false;
2853 }
2854
2855 SmallPtrSet<const SCEV*, 8> Processed;
2856 return !isHighCostExpansion(IncExpr, Processed, SE);
2857}
2858
2859/// Return true if the number of registers needed for the chain is estimated to
2860/// be less than the number required for the individual IV users. First prohibit
2861/// any IV users that keep the IV live across increments (the Users set should
2862/// be empty). Next count the number and type of increments in the chain.
2863///
2864/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2865/// effectively use postinc addressing modes. Only consider it profitable it the
2866/// increments can be computed in fewer registers when chained.
2867///
2868/// TODO: Consider IVInc free if it's already used in another chains.
2869static bool isProfitableChain(IVChain &Chain,
2870 SmallPtrSetImpl<Instruction *> &Users,
2871 ScalarEvolution &SE,
2872 const TargetTransformInfo &TTI) {
2873 if (StressIVChain)
2874 return true;
2875
2876 if (!Chain.hasIncs())
2877 return false;
2878
2879 if (!Users.empty()) {
2880 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)
2881 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)
2882 : 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)
;
2883 return false;
2884 }
2885 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 2885, __extension__
__PRETTY_FUNCTION__))
;
2886
2887 // The chain itself may require a register, so intialize cost to 1.
2888 int cost = 1;
2889
2890 // A complete chain likely eliminates the need for keeping the original IV in
2891 // a register. LSR does not currently know how to form a complete chain unless
2892 // the header phi already exists.
2893 if (isa<PHINode>(Chain.tailUserInst())
2894 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2895 --cost;
2896 }
2897 const SCEV *LastIncExpr = nullptr;
2898 unsigned NumConstIncrements = 0;
2899 unsigned NumVarIncrements = 0;
2900 unsigned NumReusedIncrements = 0;
2901
2902 if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
2903 return true;
2904
2905 for (const IVInc &Inc : Chain) {
2906 if (TTI.isProfitableLSRChainElement(Inc.UserInst))
2907 return true;
2908 if (Inc.IncExpr->isZero())
2909 continue;
2910
2911 // Incrementing by zero or some constant is neutral. We assume constants can
2912 // be folded into an addressing mode or an add's immediate operand.
2913 if (isa<SCEVConstant>(Inc.IncExpr)) {
2914 ++NumConstIncrements;
2915 continue;
2916 }
2917
2918 if (Inc.IncExpr == LastIncExpr)
2919 ++NumReusedIncrements;
2920 else
2921 ++NumVarIncrements;
2922
2923 LastIncExpr = Inc.IncExpr;
2924 }
2925 // An IV chain with a single increment is handled by LSR's postinc
2926 // uses. However, a chain with multiple increments requires keeping the IV's
2927 // value live longer than it needs to be if chained.
2928 if (NumConstIncrements > 1)
2929 --cost;
2930
2931 // Materializing increment expressions in the preheader that didn't exist in
2932 // the original code may cost a register. For example, sign-extended array
2933 // indices can produce ridiculous increments like this:
2934 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2935 cost += NumVarIncrements;
2936
2937 // Reusing variable increments likely saves a register to hold the multiple of
2938 // the stride.
2939 cost -= NumReusedIncrements;
2940
2941 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)
2942 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2943
2944 return cost < 0;
2945}
2946
2947/// Add this IV user to an existing chain or make it the head of a new chain.
2948void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2949 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2950 // When IVs are used as types of varying widths, they are generally converted
2951 // to a wider type with some uses remaining narrow under a (free) trunc.
2952 Value *const NextIV = getWideOperand(IVOper);
2953 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2954 const SCEV *const OperExprBase = getExprBase(OperExpr);
2955
2956 // Visit all existing chains. Check if its IVOper can be computed as a
2957 // profitable loop invariant increment from the last link in the Chain.
2958 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2959 const SCEV *LastIncExpr = nullptr;
2960 for (; ChainIdx < NChains; ++ChainIdx) {
2961 IVChain &Chain = IVChainVec[ChainIdx];
2962
2963 // Prune the solution space aggressively by checking that both IV operands
2964 // are expressions that operate on the same unscaled SCEVUnknown. This
2965 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2966 // first avoids creating extra SCEV expressions.
2967 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2968 continue;
2969
2970 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2971 if (!isCompatibleIVType(PrevIV, NextIV))
2972 continue;
2973
2974 // A phi node terminates a chain.
2975 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2976 continue;
2977
2978 // The increment must be loop-invariant so it can be kept in a register.
2979 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2980 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2981 if (isa<SCEVCouldNotCompute>(IncExpr) || !SE.isLoopInvariant(IncExpr, L))
2982 continue;
2983
2984 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2985 LastIncExpr = IncExpr;
2986 break;
2987 }
2988 }
2989 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2990 // bother for phi nodes, because they must be last in the chain.
2991 if (ChainIdx == NChains) {
2992 if (isa<PHINode>(UserInst))
2993 return;
2994 if (NChains >= MaxChains && !StressIVChain) {
2995 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2996 return;
2997 }
2998 LastIncExpr = OperExpr;
2999 // IVUsers may have skipped over sign/zero extensions. We don't currently
3000 // attempt to form chains involving extensions unless they can be hoisted
3001 // into this loop's AddRec.
3002 if (!isa<SCEVAddRecExpr>(LastIncExpr))
3003 return;
3004 ++NChains;
3005 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
3006 OperExprBase));
3007 ChainUsersVec.resize(NChains);
3008 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)
3009 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
3010 } else {
3011 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)
3012 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
3013 // Add this IV user to the end of the chain.
3014 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
3015 }
3016 IVChain &Chain = IVChainVec[ChainIdx];
3017
3018 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
3019 // This chain's NearUsers become FarUsers.
3020 if (!LastIncExpr->isZero()) {
3021 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
3022 NearUsers.end());
3023 NearUsers.clear();
3024 }
3025
3026 // All other uses of IVOperand become near uses of the chain.
3027 // We currently ignore intermediate values within SCEV expressions, assuming
3028 // they will eventually be used be the current chain, or can be computed
3029 // from one of the chain increments. To be more precise we could
3030 // transitively follow its user and only add leaf IV users to the set.
3031 for (User *U : IVOper->users()) {
3032 Instruction *OtherUse = dyn_cast<Instruction>(U);
3033 if (!OtherUse)
3034 continue;
3035 // Uses in the chain will no longer be uses if the chain is formed.
3036 // Include the head of the chain in this iteration (not Chain.begin()).
3037 IVChain::const_iterator IncIter = Chain.Incs.begin();
3038 IVChain::const_iterator IncEnd = Chain.Incs.end();
3039 for( ; IncIter != IncEnd; ++IncIter) {
3040 if (IncIter->UserInst == OtherUse)
3041 break;
3042 }
3043 if (IncIter != IncEnd)
3044 continue;
3045
3046 if (SE.isSCEVable(OtherUse->getType())
3047 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3048 && IU.isIVUserOrOperand(OtherUse)) {
3049 continue;
3050 }
3051 NearUsers.insert(OtherUse);
3052 }
3053
3054 // Since this user is part of the chain, it's no longer considered a use
3055 // of the chain.
3056 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3057}
3058
3059/// Populate the vector of Chains.
3060///
3061/// This decreases ILP at the architecture level. Targets with ample registers,
3062/// multiple memory ports, and no register renaming probably don't want
3063/// this. However, such targets should probably disable LSR altogether.
3064///
3065/// The job of LSR is to make a reasonable choice of induction variables across
3066/// the loop. Subsequent passes can easily "unchain" computation exposing more
3067/// ILP *within the loop* if the target wants it.
3068///
3069/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3070/// will not reorder memory operations, it will recognize this as a chain, but
3071/// will generate redundant IV increments. Ideally this would be corrected later
3072/// by a smart scheduler:
3073/// = A[i]
3074/// = A[i+x]
3075/// A[i] =
3076/// A[i+x] =
3077///
3078/// TODO: Walk the entire domtree within this loop, not just the path to the
3079/// loop latch. This will discover chains on side paths, but requires
3080/// maintaining multiple copies of the Chains state.
3081void LSRInstance::CollectChains() {
3082 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3083 SmallVector<ChainUsers, 8> ChainUsersVec;
3084
3085 SmallVector<BasicBlock *,8> LatchPath;
3086 BasicBlock *LoopHeader = L->getHeader();
3087 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3088 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3089 LatchPath.push_back(Rung->getBlock());
3090 }
3091 LatchPath.push_back(LoopHeader);
3092
3093 // Walk the instruction stream from the loop header to the loop latch.
3094 for (BasicBlock *BB : reverse(LatchPath)) {
3095 for (Instruction &I : *BB) {
3096 // Skip instructions that weren't seen by IVUsers analysis.
3097 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3098 continue;
3099
3100 // Ignore users that are part of a SCEV expression. This way we only
3101 // consider leaf IV Users. This effectively rediscovers a portion of
3102 // IVUsers analysis but in program order this time.
3103 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3104 continue;
3105
3106 // Remove this instruction from any NearUsers set it may be in.
3107 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3108 ChainIdx < NChains; ++ChainIdx) {
3109 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3110 }
3111 // Search for operands that can be chained.
3112 SmallPtrSet<Instruction*, 4> UniqueOperands;
3113 User::op_iterator IVOpEnd = I.op_end();
3114 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3115 while (IVOpIter != IVOpEnd) {
3116 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3117 if (UniqueOperands.insert(IVOpInst).second)
3118 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3119 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3120 }
3121 } // Continue walking down the instructions.
3122 } // Continue walking down the domtree.
3123 // Visit phi backedges to determine if the chain can generate the IV postinc.
3124 for (PHINode &PN : L->getHeader()->phis()) {
3125 if (!SE.isSCEVable(PN.getType()))
3126 continue;
3127
3128 Instruction *IncV =
3129 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3130 if (IncV)
3131 ChainInstruction(&PN, IncV, ChainUsersVec);
3132 }
3133 // Remove any unprofitable chains.
3134 unsigned ChainIdx = 0;
3135 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3136 UsersIdx < NChains; ++UsersIdx) {
3137 if (!isProfitableChain(IVChainVec[UsersIdx],
3138 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3139 continue;
3140 // Preserve the chain at UsesIdx.
3141 if (ChainIdx != UsersIdx)
3142 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3143 FinalizeChain(IVChainVec[ChainIdx]);
3144 ++ChainIdx;
3145 }
3146 IVChainVec.resize(ChainIdx);
3147}
3148
3149void LSRInstance::FinalizeChain(IVChain &Chain) {
3150 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3150, __extension__
__PRETTY_FUNCTION__))
;
3151 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)
;
3152
3153 for (const IVInc &Inc : Chain) {
3154 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3155 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3156 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3156, __extension__
__PRETTY_FUNCTION__))
;
3157 IVIncSet.insert(UseI);
3158 }
3159}
3160
3161/// Return true if the IVInc can be folded into an addressing mode.
3162static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3163 Value *Operand, const TargetTransformInfo &TTI) {
3164 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3165 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3166 return false;
3167
3168 if (IncConst->getAPInt().getMinSignedBits() > 64)
3169 return false;
3170
3171 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3172 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3173 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3174 IncOffset, /*HasBaseReg=*/false))
3175 return false;
3176
3177 return true;
3178}
3179
3180/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3181/// user's operand from the previous IV user's operand.
3182void LSRInstance::GenerateIVChain(const IVChain &Chain,
3183 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3184 // Find the new IVOperand for the head of the chain. It may have been replaced
3185 // by LSR.
3186 const IVInc &Head = Chain.Incs[0];
3187 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3188 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3189 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3190 IVOpEnd, L, SE);
3191 Value *IVSrc = nullptr;
3192 while (IVOpIter != IVOpEnd) {
3193 IVSrc = getWideOperand(*IVOpIter);
3194
3195 // If this operand computes the expression that the chain needs, we may use
3196 // it. (Check this after setting IVSrc which is used below.)
3197 //
3198 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3199 // narrow for the chain, so we can no longer use it. We do allow using a
3200 // wider phi, assuming the LSR checked for free truncation. In that case we
3201 // should already have a truncate on this operand such that
3202 // getSCEV(IVSrc) == IncExpr.
3203 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3204 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3205 break;
3206 }
3207 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3208 }
3209 if (IVOpIter == IVOpEnd) {
3210 // Gracefully give up on this chain.
3211 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)
;
3212 return;
3213 }
3214 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3214, __extension__
__PRETTY_FUNCTION__))
;
3215
3216 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3217 Type *IVTy = IVSrc->getType();
3218 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3219 const SCEV *LeftOverExpr = nullptr;
3220 for (const IVInc &Inc : Chain) {
3221 Instruction *InsertPt = Inc.UserInst;
3222 if (isa<PHINode>(InsertPt))
3223 InsertPt = L->getLoopLatch()->getTerminator();
3224
3225 // IVOper will replace the current IV User's operand. IVSrc is the IV
3226 // value currently held in a register.
3227 Value *IVOper = IVSrc;
3228 if (!Inc.IncExpr->isZero()) {
3229 // IncExpr was the result of subtraction of two narrow values, so must
3230 // be signed.
3231 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3232 LeftOverExpr = LeftOverExpr ?
3233 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3234 }
3235 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3236 // Expand the IV increment.
3237 Rewriter.clearPostInc();
3238 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3239 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3240 SE.getUnknown(IncV));
3241 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3242
3243 // If an IV increment can't be folded, use it as the next IV value.
3244 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3245 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3245, __extension__
__PRETTY_FUNCTION__))
;
3246 IVSrc = IVOper;
3247 LeftOverExpr = nullptr;
3248 }
3249 }
3250 Type *OperTy = Inc.IVOperand->getType();
3251 if (IVTy != OperTy) {
3252 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3253, __extension__
__PRETTY_FUNCTION__))
3253 "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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3253, __extension__
__PRETTY_FUNCTION__))
;
3254 IRBuilder<> Builder(InsertPt);
3255 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3256 }
3257 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3258 if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3259 DeadInsts.emplace_back(OperandIsInstr);
3260 }
3261 // If LSR created a new, wider phi, we may also replace its postinc. We only
3262 // do this if we also found a wide value for the head of the chain.
3263 if (isa<PHINode>(Chain.tailUserInst())) {
3264 for (PHINode &Phi : L->getHeader()->phis()) {
3265 if (!isCompatibleIVType(&Phi, IVSrc))
3266 continue;
3267 Instruction *PostIncV = dyn_cast<Instruction>(
3268 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3269 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3270 continue;
3271 Value *IVOper = IVSrc;
3272 Type *PostIncTy = PostIncV->getType();
3273 if (IVTy != PostIncTy) {
3274 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3274, __extension__
__PRETTY_FUNCTION__))
;
3275 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3276 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3277 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3278 }
3279 Phi.replaceUsesOfWith(PostIncV, IVOper);
3280 DeadInsts.emplace_back(PostIncV);
3281 }
3282 }
3283}
3284
3285void LSRInstance::CollectFixupsAndInitialFormulae() {
3286 BranchInst *ExitBranch = nullptr;
3287 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3288
3289 for (const IVStrideUse &U : IU) {
3290 Instruction *UserInst = U.getUser();
3291 // Skip IV users that are part of profitable IV Chains.
3292 User::op_iterator UseI =
3293 find(UserInst->operands(), U.getOperandValToReplace());
3294 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3294, __extension__
__PRETTY_FUNCTION__))
;
3295 if (IVIncSet.count(UseI)) {
3296 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)
;
3297 continue;
3298 }
3299
3300 LSRUse::KindType Kind = LSRUse::Basic;
3301 MemAccessTy AccessTy;
3302 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3303 Kind = LSRUse::Address;
3304 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3305 }
3306
3307 const SCEV *S = IU.getExpr(U);
3308 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3309
3310 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3311 // (N - i == 0), and this allows (N - i) to be the expression that we work
3312 // with rather than just N or i, so we can consider the register
3313 // requirements for both N and i at the same time. Limiting this code to
3314 // equality icmps is not a problem because all interesting loops use
3315 // equality icmps, thanks to IndVarSimplify.
3316 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3317 // If CI can be saved in some target, like replaced inside hardware loop
3318 // in PowerPC, no need to generate initial formulae for it.
3319 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3320 continue;
3321 if (CI->isEquality()) {
3322 // Swap the operands if needed to put the OperandValToReplace on the
3323 // left, for consistency.
3324 Value *NV = CI->getOperand(1);
3325 if (NV == U.getOperandValToReplace()) {
3326 CI->setOperand(1, CI->getOperand(0));
3327 CI->setOperand(0, NV);
3328 NV = CI->getOperand(1);
3329 Changed = true;
3330 }
3331
3332 // x == y --> x - y == 0
3333 const SCEV *N = SE.getSCEV(NV);
3334 if (SE.isLoopInvariant(N, L) && Rewriter.isSafeToExpand(N) &&
3335 (!NV->getType()->isPointerTy() ||
3336 SE.getPointerBase(N) == SE.getPointerBase(S))) {
3337 // S is normalized, so normalize N before folding it into S
3338 // to keep the result normalized.
3339 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3340 Kind = LSRUse::ICmpZero;
3341 S = SE.getMinusSCEV(N, S);
3342 } else if (L->isLoopInvariant(NV) &&
3343 (!isa<Instruction>(NV) ||
3344 DT.dominates(cast<Instruction>(NV), L->getHeader())) &&
3345 !NV->getType()->isPointerTy()) {
3346 // If we can't generally expand the expression (e.g. it contains
3347 // a divide), but it is already at a loop invariant point before the
3348 // loop, wrap it in an unknown (to prevent the expander from trying
3349 // to re-expand in a potentially unsafe way.) The restriction to
3350 // integer types is required because the unknown hides the base, and
3351 // SCEV can't compute the difference of two unknown pointers.
3352 N = SE.getUnknown(NV);
3353 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3354 Kind = LSRUse::ICmpZero;
3355 S = SE.getMinusSCEV(N, S);
3356 assert(!isa<SCEVCouldNotCompute>(S))(static_cast <bool> (!isa<SCEVCouldNotCompute>(S)
) ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(S)"
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3356, __extension__
__PRETTY_FUNCTION__))
;
3357 }
3358
3359 // -1 and the negations of all interesting strides (except the negation
3360 // of -1) are now also interesting.
3361 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3362 if (Factors[i] != -1)
3363 Factors.insert(-(uint64_t)Factors[i]);
3364 Factors.insert(-1);
3365 }
3366 }
3367
3368 // Get or create an LSRUse.
3369 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3370 size_t LUIdx = P.first;
3371 int64_t Offset = P.second;
3372 LSRUse &LU = Uses[LUIdx];
3373
3374 // Record the fixup.
3375 LSRFixup &LF = LU.getNewFixup();
3376 LF.UserInst = UserInst;
3377 LF.OperandValToReplace = U.getOperandValToReplace();
3378 LF.PostIncLoops = TmpPostIncLoops;
3379 LF.Offset = Offset;
3380 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3381
3382 if (!LU.WidestFixupType ||
3383 SE.getTypeSizeInBits(LU.WidestFixupType) <
3384 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3385 LU.WidestFixupType = LF.OperandValToReplace->getType();
3386
3387 // If this is the first use of this LSRUse, give it a formula.
3388 if (LU.Formulae.empty()) {
3389 InsertInitialFormula(S, LU, LUIdx);
3390 CountRegisters(LU.Formulae.back(), LUIdx);
3391 }
3392 }
3393
3394 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3395}
3396
3397/// Insert a formula for the given expression into the given use, separating out
3398/// loop-variant portions from loop-invariant and loop-computable portions.
3399void LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU,
3400 size_t LUIdx) {
3401 // Mark uses whose expressions cannot be expanded.
3402 if (!Rewriter.isSafeToExpand(S))
3403 LU.RigidFormula = true;
3404
3405 Formula F;
3406 F.initialMatch(S, L, SE);
3407 bool Inserted = InsertFormula(LU, LUIdx, F);
3408 assert(Inserted && "Initial formula already exists!")(static_cast <bool> (Inserted && "Initial formula already exists!"
) ? void (0) : __assert_fail ("Inserted && \"Initial formula already exists!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3408, __extension__
__PRETTY_FUNCTION__))
; (void)Inserted;
3409}
3410
3411/// Insert a simple single-register formula for the given expression into the
3412/// given use.
3413void
3414LSRInstance::InsertSupplementalFormula(const SCEV *S,
3415 LSRUse &LU, size_t LUIdx) {
3416 Formula F;
3417 F.BaseRegs.push_back(S);
3418 F.HasBaseReg = true;
3419 bool Inserted = InsertFormula(LU, LUIdx, F);
3420 assert(Inserted && "Supplemental formula already exists!")(static_cast <bool> (Inserted && "Supplemental formula already exists!"
) ? void (0) : __assert_fail ("Inserted && \"Supplemental formula already exists!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3420, __extension__
__PRETTY_FUNCTION__))
; (void)Inserted;
3421}
3422
3423/// Note which registers are used by the given formula, updating RegUses.
3424void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3425 if (F.ScaledReg)
3426 RegUses.countRegister(F.ScaledReg, LUIdx);
3427 for (const SCEV *BaseReg : F.BaseRegs)
3428 RegUses.countRegister(BaseReg, LUIdx);
3429}
3430
3431/// If the given formula has not yet been inserted, add it to the list, and
3432/// return true. Return false otherwise.
3433bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3434 // Do not insert formula that we will not be able to expand.
3435 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3436, __extension__
__PRETTY_FUNCTION__))
3436 "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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3436, __extension__
__PRETTY_FUNCTION__))
;
3437
3438 if (!LU.InsertFormula(F, *L))
3439 return false;
3440
3441 CountRegisters(F, LUIdx);
3442 return true;
3443}
3444
3445/// Check for other uses of loop-invariant values which we're tracking. These
3446/// other uses will pin these values in registers, making them less profitable
3447/// for elimination.
3448/// TODO: This currently misses non-constant addrec step registers.
3449/// TODO: Should this give more weight to users inside the loop?
3450void
3451LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3452 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3453 SmallPtrSet<const SCEV *, 32> Visited;
3454
3455 while (!Worklist.empty()) {
3456 const SCEV *S = Worklist.pop_back_val();
3457
3458 // Don't process the same SCEV twice
3459 if (!Visited.insert(S).second)
3460 continue;
3461
3462 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3463 Worklist.append(N->op_begin(), N->op_end());
3464 else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
3465 Worklist.push_back(C->getOperand());
3466 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3467 Worklist.push_back(D->getLHS());
3468 Worklist.push_back(D->getRHS());
3469 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3470 const Value *V = US->getValue();
3471 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3472 // Look for instructions defined outside the loop.
3473 if (L->contains(Inst)) continue;
3474 } else if (isa<UndefValue>(V))
3475 // Undef doesn't have a live range, so it doesn't matter.
3476 continue;
3477 for (const Use &U : V->uses()) {
3478 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3479 // Ignore non-instructions.
3480 if (!UserInst)
3481 continue;
3482 // Don't bother if the instruction is an EHPad.
3483 if (UserInst->isEHPad())
3484 continue;
3485 // Ignore instructions in other functions (as can happen with
3486 // Constants).
3487 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3488 continue;
3489 // Ignore instructions not dominated by the loop.
3490 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3491 UserInst->getParent() :
3492 cast<PHINode>(UserInst)->getIncomingBlock(
3493 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3494 if (!DT.dominates(L->getHeader(), UseBB))
3495 continue;
3496 // Don't bother if the instruction is in a BB which ends in an EHPad.
3497 if (UseBB->getTerminator()->isEHPad())
3498 continue;
3499
3500 // Ignore cases in which the currently-examined value could come from
3501 // a basic block terminated with an EHPad. This checks all incoming
3502 // blocks of the phi node since it is possible that the same incoming
3503 // value comes from multiple basic blocks, only some of which may end
3504 // in an EHPad. If any of them do, a subsequent rewrite attempt by this
3505 // pass would try to insert instructions into an EHPad, hitting an
3506 // assertion.
3507 if (isa<PHINode>(UserInst)) {
3508 const auto *PhiNode = cast<PHINode>(UserInst);
3509 bool HasIncompatibleEHPTerminatedBlock = false;
3510 llvm::Value *ExpectedValue = U;
3511 for (unsigned int I = 0; I < PhiNode->getNumIncomingValues(); I++) {
3512 if (PhiNode->getIncomingValue(I) == ExpectedValue) {
3513 if (PhiNode->getIncomingBlock(I)->getTerminator()->isEHPad()) {
3514 HasIncompatibleEHPTerminatedBlock = true;
3515 break;
3516 }
3517 }
3518 }
3519 if (HasIncompatibleEHPTerminatedBlock) {
3520 continue;
3521 }
3522 }
3523
3524 // Don't bother rewriting PHIs in catchswitch blocks.
3525 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3526 continue;
3527 // Ignore uses which are part of other SCEV expressions, to avoid
3528 // analyzing them multiple times.
3529 if (SE.isSCEVable(UserInst->getType())) {
3530 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3531 // If the user is a no-op, look through to its uses.
3532 if (!isa<SCEVUnknown>(UserS))
3533 continue;
3534 if (UserS == US) {
3535 Worklist.push_back(
3536 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3537 continue;
3538 }
3539 }
3540 // Ignore icmp instructions which are already being analyzed.
3541 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3542 unsigned OtherIdx = !U.getOperandNo();
3543 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3544 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3545 continue;
3546 }
3547
3548 std::pair<size_t, int64_t> P = getUse(
3549 S, LSRUse::Basic, MemAccessTy());
3550 size_t LUIdx = P.first;
3551 int64_t Offset = P.second;
3552 LSRUse &LU = Uses[LUIdx];
3553 LSRFixup &LF = LU.getNewFixup();
3554 LF.UserInst = const_cast<Instruction *>(UserInst);
3555 LF.OperandValToReplace = U;
3556 LF.Offset = Offset;
3557 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3558 if (!LU.WidestFixupType ||
3559 SE.getTypeSizeInBits(LU.WidestFixupType) <
3560 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3561 LU.WidestFixupType = LF.OperandValToReplace->getType();
3562 InsertSupplementalFormula(US, LU, LUIdx);
3563 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3564 break;
3565 }
3566 }
3567 }
3568}
3569
3570/// Split S into subexpressions which can be pulled out into separate
3571/// registers. If C is non-null, multiply each subexpression by C.
3572///
3573/// Return remainder expression after factoring the subexpressions captured by
3574/// Ops. If Ops is complete, return NULL.
3575static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3576 SmallVectorImpl<const SCEV *> &Ops,
3577 const Loop *L,
3578 ScalarEvolution &SE,
3579 unsigned Depth = 0) {
3580 // Arbitrarily cap recursion to protect compile time.
3581 if (Depth >= 3)
3582 return S;
3583
3584 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3585 // Break out add operands.
3586 for (const SCEV *S : Add->operands()) {
3587 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3588 if (Remainder)
3589 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3590 }
3591 return nullptr;
3592 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3593 // Split a non-zero base out of an addrec.
3594 if (AR->getStart()->isZero() || !AR->isAffine())
3595 return S;
3596
3597 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3598 C, Ops, L, SE, Depth+1);
3599 // Split the non-zero AddRec unless it is part of a nested recurrence that
3600 // does not pertain to this loop.
3601 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3602 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3603 Remainder = nullptr;
3604 }
3605 if (Remainder != AR->getStart()) {
3606 if (!Remainder)
3607 Remainder = SE.getConstant(AR->getType(), 0);
3608 return SE.getAddRecExpr(Remainder,
3609 AR->getStepRecurrence(SE),
3610 AR->getLoop(),
3611 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3612 SCEV::FlagAnyWrap);
3613 }
3614 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3615 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3616 if (Mul->getNumOperands() != 2)
3617 return S;
3618 if (const SCEVConstant *Op0 =
3619 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3620 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3621 const SCEV *Remainder =
3622 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3623 if (Remainder)
3624 Ops.push_back(SE.getMulExpr(C, Remainder));
3625 return nullptr;
3626 }
3627 }
3628 return S;
3629}
3630
3631/// Return true if the SCEV represents a value that may end up as a
3632/// post-increment operation.
3633static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3634 LSRUse &LU, const SCEV *S, const Loop *L,
3635 ScalarEvolution &SE) {
3636 if (LU.Kind != LSRUse::Address ||
3637 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3638 return false;
3639 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3640 if (!AR)
3641 return false;
3642 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3643 if (!isa<SCEVConstant>(LoopStep))
3644 return false;
3645 // Check if a post-indexed load/store can be used.
3646 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3647 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3648 const SCEV *LoopStart = AR->getStart();
3649 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3650 return true;
3651 }
3652 return false;
3653}
3654
3655/// Helper function for LSRInstance::GenerateReassociations.
3656void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3657 const Formula &Base,
3658 unsigned Depth, size_t Idx,
3659 bool IsScaledReg) {
3660 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3661 // Don't generate reassociations for the base register of a value that
3662 // may generate a post-increment operator. The reason is that the
3663 // reassociations cause extra base+register formula to be created,
3664 // and possibly chosen, but the post-increment is more efficient.
3665 if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3666 return;
3667 SmallVector<const SCEV *, 8> AddOps;
3668 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3669 if (Remainder)
3670 AddOps.push_back(Remainder);
3671
3672 if (AddOps.size() == 1)
3673 return;
3674
3675 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3676 JE = AddOps.end();
3677 J != JE; ++J) {
3678 // Loop-variant "unknown" values are uninteresting; we won't be able to
3679 // do anything meaningful with them.
3680 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3681 continue;
3682
3683 // Don't pull a constant into a register if the constant could be folded
3684 // into an immediate field.
3685 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3686 LU.AccessTy, *J, Base.getNumRegs() > 1))
3687 continue;
3688
3689 // Collect all operands except *J.
3690 SmallVector<const SCEV *, 8> InnerAddOps(
3691 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3692 InnerAddOps.append(std::next(J),
3693 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3694
3695 // Don't leave just a constant behind in a register if the constant could
3696 // be folded into an immediate field.
3697 if (InnerAddOps.size() == 1 &&
3698 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3699 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3700 continue;
3701
3702 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3703 if (InnerSum->isZero())
3704 continue;
3705 Formula F = Base;
3706
3707 // Add the remaining pieces of the add back into the new formula.
3708 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3709 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3710 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3711 InnerSumSC->getValue()->getZExtValue())) {
3712 F.UnfoldedOffset =
3713 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3714 if (IsScaledReg)
3715 F.ScaledReg = nullptr;
3716 else
3717 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3718 } else if (IsScaledReg)
3719 F.ScaledReg = InnerSum;
3720 else
3721 F.BaseRegs[Idx] = InnerSum;
3722
3723 // Add J as its own register, or an unfolded immediate.
3724 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3725 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3726 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3727 SC->getValue()->getZExtValue()))
3728 F.UnfoldedOffset =
3729 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3730 else
3731 F.BaseRegs.push_back(*J);
3732 // We may have changed the number of register in base regs, adjust the
3733 // formula accordingly.
3734 F.canonicalize(*L);
3735
3736 if (InsertFormula(LU, LUIdx, F))
3737 // If that formula hadn't been seen before, recurse to find more like
3738 // it.
3739 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3740 // Because just Depth is not enough to bound compile time.
3741 // This means that every time AddOps.size() is greater 16^x we will add
3742 // x to Depth.
3743 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3744 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3745 }
3746}
3747
3748/// Split out subexpressions from adds and the bases of addrecs.
3749void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3750 Formula Base, unsigned Depth) {
3751 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3751, __extension__
__PRETTY_FUNCTION__))
;
3752 // Arbitrarily cap recursion to protect compile time.
3753 if (Depth >= 3)
3754 return;
3755
3756 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3757 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3758
3759 if (Base.Scale == 1)
3760 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3761 /* Idx */ -1, /* IsScaledReg */ true);
3762}
3763
3764/// Generate a formula consisting of all of the loop-dominating registers added
3765/// into a single register.
3766void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3767 Formula Base) {
3768 // This method is only interesting on a plurality of registers.
3769 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3770 (Base.UnfoldedOffset != 0) <= 1)
3771 return;
3772
3773 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3774 // processing the formula.
3775 Base.unscale();
3776 SmallVector<const SCEV *, 4> Ops;
3777 Formula NewBase = Base;
3778 NewBase.BaseRegs.clear();
3779 Type *CombinedIntegerType = nullptr;
3780 for (const SCEV *BaseReg : Base.BaseRegs) {
3781 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3782 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3783 if (!CombinedIntegerType)
3784 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3785 Ops.push_back(BaseReg);
3786 }
3787 else
3788 NewBase.BaseRegs.push_back(BaseReg);
3789 }
3790
3791 // If no register is relevant, we're done.
3792 if (Ops.size() == 0)
3793 return;
3794
3795 // Utility function for generating the required variants of the combined
3796 // registers.
3797 auto GenerateFormula = [&](const SCEV *Sum) {
3798 Formula F = NewBase;
3799
3800 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3801 // opportunity to fold something. For now, just ignore such cases
3802 // rather than proceed with zero in a register.
3803 if (Sum->isZero())
3804 return;
3805
3806 F.BaseRegs.push_back(Sum);
3807 F.canonicalize(*L);
3808 (void)InsertFormula(LU, LUIdx, F);
3809 };
3810
3811 // If we collected at least two registers, generate a formula combining them.
3812 if (Ops.size() > 1) {
3813 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3814 GenerateFormula(SE.getAddExpr(OpsCopy));
3815 }
3816
3817 // If we have an unfolded offset, generate a formula combining it with the
3818 // registers collected.
3819 if (NewBase.UnfoldedOffset) {
3820 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3820, __extension__
__PRETTY_FUNCTION__))
;
3821 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3822 true));
3823 NewBase.UnfoldedOffset = 0;
3824 GenerateFormula(SE.getAddExpr(Ops));
3825 }
3826}
3827
3828/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3829void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3830 const Formula &Base, size_t Idx,
3831 bool IsScaledReg) {
3832 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3833 GlobalValue *GV = ExtractSymbol(G, SE);
3834 if (G->isZero() || !GV)
3835 return;
3836 Formula F = Base;
3837 F.BaseGV = GV;
3838 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3839 return;
3840 if (IsScaledReg)
3841 F.ScaledReg = G;
3842 else
3843 F.BaseRegs[Idx] = G;
3844 (void)InsertFormula(LU, LUIdx, F);
3845}
3846
3847/// Generate reuse formulae using symbolic offsets.
3848void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3849 Formula Base) {
3850 // We can't add a symbolic offset if the address already contains one.
3851 if (Base.BaseGV) return;
3852
3853 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3854 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3855 if (Base.Scale == 1)
3856 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3857 /* IsScaledReg */ true);
3858}
3859
3860/// Helper function for LSRInstance::GenerateConstantOffsets.
3861void LSRInstance::GenerateConstantOffsetsImpl(
3862 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3863 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3864
3865 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3866 Formula F = Base;
3867 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3868
3869 if (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) {
3870 // Add the offset to the base register.
3871 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3872 // If it cancelled out, drop the base register, otherwise update it.
3873 if (NewG->isZero()) {
3874 if (IsScaledReg) {
3875 F.Scale = 0;
3876 F.ScaledReg = nullptr;
3877 } else
3878 F.deleteBaseReg(F.BaseRegs[Idx]);
3879 F.canonicalize(*L);
3880 } else if (IsScaledReg)
3881 F.ScaledReg = NewG;
3882 else
3883 F.BaseRegs[Idx] = NewG;
3884
3885 (void)InsertFormula(LU, LUIdx, F);
3886 }
3887 };
3888
3889 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3890
3891 // With constant offsets and constant steps, we can generate pre-inc
3892 // accesses by having the offset equal the step. So, for access #0 with a
3893 // step of 8, we generate a G - 8 base which would require the first access
3894 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3895 // for itself and hopefully becomes the base for other accesses. This means
3896 // means that a single pre-indexed access can be generated to become the new
3897 // base pointer for each iteration of the loop, resulting in no extra add/sub
3898 // instructions for pointer updating.
3899 if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) {
3900 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3901 if (auto *StepRec =
3902 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3903 const APInt &StepInt = StepRec->getAPInt();
3904 int64_t Step = StepInt.isNegative() ?
3905 StepInt.getSExtValue() : StepInt.getZExtValue();
3906
3907 for (int64_t Offset : Worklist) {
3908 Offset -= Step;
3909 GenerateOffset(G, Offset);
3910 }
3911 }
3912 }
3913 }
3914 for (int64_t Offset : Worklist)
3915 GenerateOffset(G, Offset);
3916
3917 int64_t Imm = ExtractImmediate(G, SE);
3918 if (G->isZero() || Imm == 0)
3919 return;
3920 Formula F = Base;
3921 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3922 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3923 return;
3924 if (IsScaledReg) {
3925 F.ScaledReg = G;
3926 } else {
3927 F.BaseRegs[Idx] = G;
3928 // We may generate non canonical Formula if G is a recurrent expr reg
3929 // related with current loop while F.ScaledReg is not.
3930 F.canonicalize(*L);
3931 }
3932 (void)InsertFormula(LU, LUIdx, F);
3933}
3934
3935/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3936void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3937 Formula Base) {
3938 // TODO: For now, just add the min and max offset, because it usually isn't
3939 // worthwhile looking at everything inbetween.
3940 SmallVector<int64_t, 2> Worklist;
3941 Worklist.push_back(LU.MinOffset);
3942 if (LU.MaxOffset != LU.MinOffset)
3943 Worklist.push_back(LU.MaxOffset);
3944
3945 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3946 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3947 if (Base.Scale == 1)
3948 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3949 /* IsScaledReg */ true);
3950}
3951
3952/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3953/// == y -> x*c == y*c.
3954void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3955 Formula Base) {
3956 if (LU.Kind != LSRUse::ICmpZero) return;
3957
3958 // Determine the integer type for the base formula.
3959 Type *IntTy = Base.getType();
3960 if (!IntTy) return;
3961 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3962
3963 // Don't do this if there is more than one offset.
3964 if (LU.MinOffset != LU.MaxOffset) return;
3965
3966 // Check if transformation is valid. It is illegal to multiply pointer.
3967 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3968 return;
3969 for (const SCEV *BaseReg : Base.BaseRegs)
3970 if (BaseReg->getType()->isPointerTy())
3971 return;
3972 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3972, __extension__
__PRETTY_FUNCTION__))
;
3973
3974 // Check each interesting stride.
3975 for (int64_t Factor : Factors) {
3976 // Check that Factor can be represented by IntTy
3977 if (!ConstantInt::isValueValidForType(IntTy, Factor))
3978 continue;
3979 // Check that the multiplication doesn't overflow.
3980 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3981 continue;
3982 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3983 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3983, __extension__
__PRETTY_FUNCTION__))
;
3984 if (NewBaseOffset / Factor != Base.BaseOffset)
3985 continue;
3986 // If the offset will be truncated at this use, check that it is in bounds.
3987 if (!IntTy->isPointerTy() &&
3988 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3989 continue;
3990
3991 // Check that multiplying with the use offset doesn't overflow.
3992 int64_t Offset = LU.MinOffset;
3993 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3994 continue;
3995 Offset = (uint64_t)Offset * Factor;
3996 if (Offset / Factor != LU.MinOffset)
3997 continue;
3998 // If the offset will be truncated at this use, check that it is in bounds.
3999 if (!IntTy->isPointerTy() &&
4000 !ConstantInt::isValueValidForType(IntTy, Offset))
4001 continue;
4002
4003 Formula F = Base;
4004 F.BaseOffset = NewBaseOffset;
4005
4006 // Check that this scale is legal.
4007 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
4008 continue;
4009
4010 // Compensate for the use having MinOffset built into it.
4011 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
4012
4013 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4014
4015 // Check that multiplying with each base register doesn't overflow.
4016 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
4017 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
4018 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
4019 goto next;
4020 }
4021
4022 // Check that multiplying with the scaled register doesn't overflow.
4023 if (F.ScaledReg) {
4024 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
4025 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
4026 continue;
4027 }
4028
4029 // Check that multiplying with the unfolded offset doesn't overflow.
4030 if (F.UnfoldedOffset != 0) {
4031 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
4032 Factor == -1)
4033 continue;
4034 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
4035 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
4036 continue;
4037 // If the offset will be truncated, check that it is in bounds.
4038 if (!IntTy->isPointerTy() &&
4039 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
4040 continue;
4041 }
4042
4043 // If we make it here and it's legal, add it.
4044 (void)InsertFormula(LU, LUIdx, F);
4045 next:;
4046 }
4047}
4048
4049/// Generate stride factor reuse formulae by making use of scaled-offset address
4050/// modes, for example.
4051void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
4052 // Determine the integer type for the base formula.
4053 Type *IntTy = Base.getType();
4054 if (!IntTy) return;
4055
4056 // If this Formula already has a scaled register, we can't add another one.
4057 // Try to unscale the formula to generate a better scale.
4058 if (Base.Scale != 0 && !Base.unscale())
4059 return;
4060
4061 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 4061, __extension__
__PRETTY_FUNCTION__))
;
4062
4063 // Check each interesting stride.
4064 for (int64_t Factor : Factors) {
4065 Base.Scale = Factor;
4066 Base.HasBaseReg = Base.BaseRegs.size() > 1;
4067 // Check whether this scale is going to be legal.
4068 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4069 Base)) {
4070 // As a special-case, handle special out-of-loop Basic users specially.
4071 // TODO: Reconsider this special case.
4072 if (LU.Kind == LSRUse::Basic &&
4073 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
4074 LU.AccessTy, Base) &&
4075 LU.AllFixupsOutsideLoop)
4076 LU.Kind = LSRUse::Special;
4077 else
4078 continue;
4079 }
4080 // For an ICmpZero, negating a solitary base register won't lead to
4081 // new solutions.
4082 if (LU.Kind == LSRUse::ICmpZero &&
4083 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
4084 continue;
4085 // For each addrec base reg, if its loop is current loop, apply the scale.
4086 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
4087 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
4088 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
4089 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4090 if (FactorS->isZero())
4091 continue;
4092 // Divide out the factor, ignoring high bits, since we'll be
4093 // scaling the value back up in the end.
4094 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true))
4095 if (!Quotient->isZero()) {
4096 // TODO: This could be optimized to avoid all the copying.
4097 Formula F = Base;
4098 F.ScaledReg = Quotient;
4099 F.deleteBaseReg(F.BaseRegs[i]);
4100 // The canonical representation of 1*reg is reg, which is already in
4101 // Base. In that case, do not try to insert the formula, it will be
4102 // rejected anyway.
4103 if (F.Scale == 1 && (F.BaseRegs.empty() ||
4104 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4105 continue;
4106 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4107 // non canonical Formula with ScaledReg's loop not being L.
4108 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4109 F.canonicalize(*L);
4110 (void)InsertFormula(LU, LUIdx, F);
4111 }
4112 }
4113 }
4114 }
4115}
4116
4117/// Generate reuse formulae from different IV types.
4118void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4119 // Don't bother truncating symbolic values.
4120 if (Base.BaseGV) return;
4121
4122 // Determine the integer type for the base formula.
4123 Type *DstTy = Base.getType();
4124 if (!DstTy) return;
4125 if (DstTy->isPointerTy())
4126 return;
4127
4128 // It is invalid to extend a pointer type so exit early if ScaledReg or
4129 // any of the BaseRegs are pointers.
4130 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
4131 return;
4132 if (any_of(Base.BaseRegs,
4133 [](const SCEV *S) { return S->getType()->isPointerTy(); }))
4134 return;
4135
4136 for (Type *SrcTy : Types) {
4137 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4138 Formula F = Base;
4139
4140 // Sometimes SCEV is able to prove zero during ext transform. It may
4141 // happen if SCEV did not do all possible transforms while creating the
4142 // initial node (maybe due to depth limitations), but it can do them while
4143 // taking ext.
4144 if (F.ScaledReg) {
4145 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4146 if (NewScaledReg->isZero())
4147 continue;
4148 F.ScaledReg = NewScaledReg;
4149 }
4150 bool HasZeroBaseReg = false;
4151 for (const SCEV *&BaseReg : F.BaseRegs) {
4152 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4153 if (NewBaseReg->isZero()) {
4154 HasZeroBaseReg = true;
4155 break;
4156 }
4157 BaseReg = NewBaseReg;
4158 }
4159 if (HasZeroBaseReg)
4160 continue;
4161
4162 // TODO: This assumes we've done basic processing on all uses and
4163 // have an idea what the register usage is.
4164 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4165 continue;
4166
4167 F.canonicalize(*L);
4168 (void)InsertFormula(LU, LUIdx, F);
4169 }
4170 }
4171}
4172
4173namespace {
4174
4175/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4176/// modifications so that the search phase doesn't have to worry about the data
4177/// structures moving underneath it.
4178struct WorkItem {
4179 size_t LUIdx;
4180 int64_t Imm;
4181 const SCEV *OrigReg;
4182
4183 WorkItem(size_t LI, int64_t I, const SCEV *R)
4184 : LUIdx(LI), Imm(I), OrigReg(R) {}
4185
4186 void print(raw_ostream &OS) const;
4187 void dump() const;
4188};
4189
4190} // end anonymous namespace
4191
4192#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4193void WorkItem::print(raw_ostream &OS) const {
4194 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4195 << " , add offset " << Imm;
4196}
4197
4198LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4199 print(errs()); errs() << '\n';
4200}
4201#endif
4202
4203/// Look for registers which are a constant distance apart and try to form reuse
4204/// opportunities between them.
4205void LSRInstance::GenerateCrossUseConstantOffsets() {
4206 // Group the registers by their value without any added constant offset.
4207 using ImmMapTy = std::map<int64_t, const SCEV *>;
4208
4209 DenseMap<const SCEV *, ImmMapTy> Map;
4210 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4211 SmallVector<const SCEV *, 8> Sequence;
4212 for (const SCEV *Use : RegUses) {
4213 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4214 int64_t Imm = ExtractImmediate(Reg, SE);
4215 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4216 if (Pair.second)
4217 Sequence.push_back(Reg);
4218 Pair.first->second.insert(std::make_pair(Imm, Use));
4219 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4220 }
4221
4222 // Now examine each set of registers with the same base value. Build up
4223 // a list of work to do and do the work in a separate step so that we're
4224 // not adding formulae and register counts while we're searching.
4225 SmallVector<WorkItem, 32> WorkItems;
4226 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4227 for (const SCEV *Reg : Sequence) {
4228 const ImmMapTy &Imms = Map.find(Reg)->second;
4229
4230 // It's not worthwhile looking for reuse if there's only one offset.
4231 if (Imms.size() == 1)
4232 continue;
4233
4234 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)
4235 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)
4236 : 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)
4237 << ' ' << 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)
4238 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)
;
4239
4240 // Examine each offset.
4241 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4242 J != JE; ++J) {
4243 const SCEV *OrigReg = J->second;
4244
4245 int64_t JImm = J->first;
4246 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4247
4248 if (!isa<SCEVConstant>(OrigReg) &&
4249 UsedByIndicesMap[Reg].count() == 1) {
4250 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)
4251 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4252 continue;
4253 }
4254
4255 // Conservatively examine offsets between this orig reg a few selected
4256 // other orig regs.
4257 int64_t First = Imms.begin()->first;
4258 int64_t Last = std::prev(Imms.end())->first;
4259 // Compute (First + Last) / 2 without overflow using the fact that
4260 // First + Last = 2 * (First + Last) + (First ^ Last).
4261 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4262 // If the result is negative and First is odd and Last even (or vice versa),
4263 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4264 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4265 ImmMapTy::const_iterator OtherImms[] = {
4266 Imms.begin(), std::prev(Imms.end()),
4267 Imms.lower_bound(Avg)};
4268 for (const auto &M : OtherImms) {
4269 if (M == J || M == JE) continue;
4270
4271 // Compute the difference between the two.
4272 int64_t Imm = (uint64_t)JImm - M->first;
4273 for (unsigned LUIdx : UsedByIndices.set_bits())
4274 // Make a memo of this use, offset, and register tuple.
4275 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4276 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4277 }
4278 }
4279 }
4280
4281 Map.clear();
4282 Sequence.clear();
4283 UsedByIndicesMap.clear();
4284 UniqueItems.clear();
4285
4286 // Now iterate through the worklist and add new formulae.
4287 for (const WorkItem &WI : WorkItems) {
4288 size_t LUIdx = WI.LUIdx;
4289 LSRUse &LU = Uses[LUIdx];
4290 int64_t Imm = WI.Imm;
4291 const SCEV *OrigReg = WI.OrigReg;
4292
4293 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4294 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4295 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4296
4297 // TODO: Use a more targeted data structure.
4298 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4299 Formula F = LU.Formulae[L];
4300 // FIXME: The code for the scaled and unscaled registers looks
4301 // very similar but slightly different. Investigate if they
4302 // could be merged. That way, we would not have to unscale the
4303 // Formula.
4304 F.unscale();
4305 // Use the immediate in the scaled register.
4306 if (F.ScaledReg == OrigReg) {
4307 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4308 // Don't create 50 + reg(-50).
4309 if (F.referencesReg(SE.getSCEV(
4310 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4311 continue;
4312 Formula NewF = F;
4313 NewF.BaseOffset = Offset;
4314 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4315 NewF))
4316 continue;
4317 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4318
4319 // If the new scale is a constant in a register, and adding the constant
4320 // value to the immediate would produce a value closer to zero than the
4321 // immediate itself, then the formula isn't worthwhile.
4322 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4323 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4324 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4325 .ule(std::abs(NewF.BaseOffset)))
4326 continue;
4327
4328 // OK, looks good.
4329 NewF.canonicalize(*this->L);
4330 (void)InsertFormula(LU, LUIdx, NewF);
4331 } else {
4332 // Use the immediate in a base register.
4333 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4334 const SCEV *BaseReg = F.BaseRegs[N];
4335 if (BaseReg != OrigReg)
4336 continue;
4337 Formula NewF = F;
4338 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4339 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4340 LU.Kind, LU.AccessTy, NewF)) {
4341 if (AMK == TTI::AMK_PostIndexed &&
4342 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4343 continue;
4344 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4345 continue;
4346 NewF = F;
4347 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4348 }
4349 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4350
4351 // If the new formula has a constant in a register, and adding the
4352 // constant value to the immediate would produce a value closer to
4353 // zero than the immediate itself, then the formula isn't worthwhile.
4354 for (const SCEV *NewReg : NewF.BaseRegs)
4355 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4356 if ((C->getAPInt() + NewF.BaseOffset)
4357 .abs()
4358 .slt(std::abs(NewF.BaseOffset)) &&
4359 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4360 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4361 goto skip_formula;
4362
4363 // Ok, looks good.
4364 NewF.canonicalize(*this->L);
4365 (void)InsertFormula(LU, LUIdx, NewF);
4366 break;
4367 skip_formula:;
4368 }
4369 }
4370 }
4371 }
4372}
4373
4374/// Generate formulae for each use.
4375void
4376LSRInstance::GenerateAllReuseFormulae() {
4377 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4378 // queries are more precise.
4379 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4380 LSRUse &LU = Uses[LUIdx];
4381 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4382 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4383 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4384 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4385 }
4386 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4387 LSRUse &LU = Uses[LUIdx];
4388 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4389 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4390 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4391 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4392 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4393 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4394 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4395 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4396 }
4397 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4398 LSRUse &LU = Uses[LUIdx];
4399 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4400 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4401 }
4402
4403 GenerateCrossUseConstantOffsets();
4404
4405 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4406 "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)
4407 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4408}
4409
4410/// If there are multiple formulae with the same set of registers used
4411/// by other uses, pick the best one and delete the others.
4412void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4413 DenseSet<const SCEV *> VisitedRegs;
4414 SmallPtrSet<const SCEV *, 16> Regs;
4415 SmallPtrSet<const SCEV *, 16> LoserRegs;
4416#ifndef NDEBUG
4417 bool ChangedFormulae = false;
4418#endif
4419
4420 // Collect the best formula for each unique set of shared registers. This
4421 // is reset for each use.
4422 using BestFormulaeTy =
4423 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4424
4425 BestFormulaeTy BestFormulae;
4426
4427 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4428 LSRUse &LU = Uses[LUIdx];
4429 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)
4430 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4431
4432 bool Any = false;
4433 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4434 FIdx != NumForms; ++FIdx) {
4435 Formula &F = LU.Formulae[FIdx];
4436
4437 // Some formulas are instant losers. For example, they may depend on
4438 // nonexistent AddRecs from other loops. These need to be filtered
4439 // immediately, otherwise heuristics could choose them over others leading
4440 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4441 // avoids the need to recompute this information across formulae using the
4442 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4443 // the corresponding bad register from the Regs set.
4444 Cost CostF(L, SE, TTI, AMK);
4445 Regs.clear();
4446 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4447 if (CostF.isLoser()) {
4448 // During initial formula generation, undesirable formulae are generated
4449 // by uses within other loops that have some non-trivial address mode or
4450 // use the postinc form of the IV. LSR needs to provide these formulae
4451 // as the basis of rediscovering the desired formula that uses an AddRec
4452 // corresponding to the existing phi. Once all formulae have been
4453 // generated, these initial losers may be pruned.
4454 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)
4455 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4456 }
4457 else {
4458 SmallVector<const SCEV *, 4> Key;
4459 for (const SCEV *Reg : F.BaseRegs) {
4460 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4461 Key.push_back(Reg);
4462 }
4463 if (F.ScaledReg &&
4464 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4465 Key.push_back(F.ScaledReg);
4466 // Unstable sort by host order ok, because this is only used for
4467 // uniquifying.
4468 llvm::sort(Key);
4469
4470 std::pair<BestFormulaeTy::const_iterator, bool> P =
4471 BestFormulae.insert(std::make_pair(Key, FIdx));
4472 if (P.second)
4473 continue;
4474
4475 Formula &Best = LU.Formulae[P.first->second];
4476
4477 Cost CostBest(L, SE, TTI, AMK);
4478 Regs.clear();
4479 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4480 if (CostF.isLess(CostBest))
4481 std::swap(F, Best);
4482 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)
4483 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)
4484 " 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)
4485 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)
;
4486 }
4487#ifndef NDEBUG
4488 ChangedFormulae = true;
4489#endif
4490 LU.DeleteFormula(F);
4491 --FIdx;
4492 --NumForms;
4493 Any = true;
4494 }
4495
4496 // Now that we've filtered out some formulae, recompute the Regs set.
4497 if (Any)
4498 LU.RecomputeRegs(LUIdx, RegUses);
4499
4500 // Reset this to prepare for the next use.
4501 BestFormulae.clear();
4502 }
4503
4504 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)
4505 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)
4506 "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)
4507 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)
4508 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4509}
4510
4511/// Estimate the worst-case number of solutions the solver might have to
4512/// consider. It almost never considers this many solutions because it prune the
4513/// search space, but the pruning isn't always sufficient.
4514size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4515 size_t Power = 1;
4516 for (const LSRUse &LU : Uses) {
4517 size_t FSize = LU.Formulae.size();
4518 if (FSize >= ComplexityLimit) {
4519 Power = ComplexityLimit;
4520 break;
4521 }
4522 Power *= FSize;
4523 if (Power >= ComplexityLimit)
4524 break;
4525 }
4526 return Power;
4527}
4528
4529/// When one formula uses a superset of the registers of another formula, it
4530/// won't help reduce register pressure (though it may not necessarily hurt
4531/// register pressure); remove it to simplify the system.
4532void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4533 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4534 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)
;
4535
4536 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)
4537 "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)
4538 "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)
;
4539
4540 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4541 LSRUse &LU = Uses[LUIdx];
4542 bool Any = false;
4543 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4544 Formula &F = LU.Formulae[i];
4545 // Look for a formula with a constant or GV in a register. If the use
4546 // also has a formula with that same value in an immediate field,
4547 // delete the one that uses a register.
4548 for (SmallVectorImpl<const SCEV *>::const_iterator
4549 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4550 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4551 Formula NewF = F;
4552 //FIXME: Formulas should store bitwidth to do wrapping properly.
4553 // See PR41034.
4554 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4555 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4556 (I - F.BaseRegs.begin()));
4557 if (LU.HasFormulaWithSameRegs(NewF)) {
4558 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4559 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4560 LU.DeleteFormula(F);
4561 --i;
4562 --e;
4563 Any = true;
4564 break;
4565 }
4566 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4567 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4568 if (!F.BaseGV) {
4569 Formula NewF = F;
4570 NewF.BaseGV = GV;
4571 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4572 (I - F.BaseRegs.begin()));
4573 if (LU.HasFormulaWithSameRegs(NewF)) {
4574 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4575 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4576 LU.DeleteFormula(F);
4577 --i;
4578 --e;
4579 Any = true;
4580 break;
4581 }
4582 }
4583 }
4584 }
4585 }
4586 if (Any)
4587 LU.RecomputeRegs(LUIdx, RegUses);
4588 }
4589
4590 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)
;
4591 }
4592}
4593
4594/// When there are many registers for expressions like A, A+1, A+2, etc.,
4595/// allocate a single register for them.
4596void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4597 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4598 return;
4599
4600 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)
4601 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)
4602 "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)
4603 "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)
;
4604
4605 // This is especially useful for unrolled loops.
4606
4607 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4608 LSRUse &LU = Uses[LUIdx];
4609 for (const Formula &F : LU.Formulae) {
4610 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4611 continue;
4612
4613 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4614 if (!LUThatHas)
4615 continue;
4616
4617 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4618 LU.Kind, LU.AccessTy))
4619 continue;
4620
4621 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)
;
4622
4623 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4624
4625 // Transfer the fixups of LU to LUThatHas.
4626 for (LSRFixup &Fixup : LU.Fixups) {
4627 Fixup.Offset += F.BaseOffset;
4628 LUThatHas->pushFixup(Fixup);
4629 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)
;
4630 }
4631
4632 // Delete formulae from the new use which are no longer legal.
4633 bool Any = false;
4634 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4635 Formula &F = LUThatHas->Formulae[i];
4636 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4637 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4638 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)
;
4639 LUThatHas->DeleteFormula(F);
4640 --i;
4641 --e;
4642 Any = true;
4643 }
4644 }
4645
4646 if (Any)
4647 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4648
4649 // Delete the old use.
4650 DeleteUse(LU, LUIdx);
4651 --LUIdx;
4652 --NumUses;
4653 break;
4654 }
4655 }
4656
4657 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)
;
4658}
4659
4660/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4661/// we've done more filtering, as it may be able to find more formulae to
4662/// eliminate.
4663void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4664 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4665 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)
;
4666
4667 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)
4668 "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)
;
4669
4670 FilterOutUndesirableDedicatedRegisters();
4671
4672 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)
;
4673 }
4674}
4675
4676/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4677/// Pick the best one and delete the others.
4678/// This narrowing heuristic is to keep as many formulae with different
4679/// Scale and ScaledReg pair as possible while narrowing the search space.
4680/// The benefit is that it is more likely to find out a better solution
4681/// from a formulae set with more Scale and ScaledReg variations than
4682/// a formulae set with the same Scale and ScaledReg. The picking winner
4683/// reg heuristic will often keep the formulae with the same Scale and
4684/// ScaledReg and filter others, and we want to avoid that if possible.
4685void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4686 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4687 return;
4688
4689 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)
4690 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)
4691 "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)
4692 "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)
;
4693
4694 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4695 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4696
4697 BestFormulaeTy BestFormulae;
4698#ifndef NDEBUG
4699 bool ChangedFormulae = false;
4700#endif
4701 DenseSet<const SCEV *> VisitedRegs;
4702 SmallPtrSet<const SCEV *, 16> Regs;
4703
4704 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4705 LSRUse &LU = Uses[LUIdx];
4706 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)
4707 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4708
4709 // Return true if Formula FA is better than Formula FB.
4710 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4711 // First we will try to choose the Formula with fewer new registers.
4712 // For a register used by current Formula, the more the register is
4713 // shared among LSRUses, the less we increase the register number
4714 // counter of the formula.
4715 size_t FARegNum = 0;
4716 for (const SCEV *Reg : FA.BaseRegs) {
4717 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4718 FARegNum += (NumUses - UsedByIndices.count() + 1);
4719 }
4720 size_t FBRegNum = 0;
4721 for (const SCEV *Reg : FB.BaseRegs) {
4722 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4723 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4724 }
4725 if (FARegNum != FBRegNum)
4726 return FARegNum < FBRegNum;
4727
4728 // If the new register numbers are the same, choose the Formula with
4729 // less Cost.
4730 Cost CostFA(L, SE, TTI, AMK);
4731 Cost CostFB(L, SE, TTI, AMK);
4732 Regs.clear();
4733 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4734 Regs.clear();
4735 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4736 return CostFA.isLess(CostFB);
4737 };
4738
4739 bool Any = false;
4740 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4741 ++FIdx) {
4742 Formula &F = LU.Formulae[FIdx];
4743 if (!F.ScaledReg)
4744 continue;
4745 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4746 if (P.second)
4747 continue;
4748
4749 Formula &Best = LU.Formulae[P.first->second];
4750 if (IsBetterThan(F, Best))
4751 std::swap(F, Best);
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" " in favor of formula "
; Best.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" " in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
4754 " 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)
4755 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)
;
4756#ifndef NDEBUG
4757 ChangedFormulae = true;
4758#endif
4759 LU.DeleteFormula(F);
4760 --FIdx;
4761 --NumForms;
4762 Any = true;
4763 }
4764 if (Any)
4765 LU.RecomputeRegs(LUIdx, RegUses);
4766
4767 // Reset this to prepare for the next use.
4768 BestFormulae.clear();
4769 }
4770
4771 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)
4772 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)
4773 "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)
4774 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)
4775 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4776}
4777
4778/// If we are over the complexity limit, filter out any post-inc prefering
4779/// variables to only post-inc values.
4780void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
4781 if (AMK != TTI::AMK_PostIndexed)
4782 return;
4783 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4784 return;
4785
4786 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)
4787 "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)
4788 "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)
;
4789
4790 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4791 LSRUse &LU = Uses[LUIdx];
4792
4793 if (LU.Kind != LSRUse::Address)
4794 continue;
4795 if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
4796 !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
4797 continue;
4798
4799 size_t MinRegs = std::numeric_limits<size_t>::max();
4800 for (const Formula &F : LU.Formulae)
4801 MinRegs = std::min(F.getNumRegs(), MinRegs);
4802
4803 bool Any = false;
4804 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4805 ++FIdx) {
4806 Formula &F = LU.Formulae[FIdx];
4807 if (F.getNumRegs() > MinRegs) {
4808 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)
4809 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n"; } } while (false)
;
4810 LU.DeleteFormula(F);
4811 --FIdx;
4812 --NumForms;
4813 Any = true;
4814 }
4815 }
4816 if (Any)
4817 LU.RecomputeRegs(LUIdx, RegUses);
4818
4819 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4820 break;
4821 }
4822
4823 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)
;
4824}
4825
4826/// The function delete formulas with high registers number expectation.
4827/// Assuming we don't know the value of each formula (already delete
4828/// all inefficient), generate probability of not selecting for each
4829/// register.
4830/// For example,
4831/// Use1:
4832/// reg(a) + reg({0,+,1})
4833/// reg(a) + reg({-1,+,1}) + 1
4834/// reg({a,+,1})
4835/// Use2:
4836/// reg(b) + reg({0,+,1})
4837/// reg(b) + reg({-1,+,1}) + 1
4838/// reg({b,+,1})
4839/// Use3:
4840/// reg(c) + reg(b) + reg({0,+,1})
4841/// reg(c) + reg({b,+,1})
4842///
4843/// Probability of not selecting
4844/// Use1 Use2 Use3
4845/// reg(a) (1/3) * 1 * 1
4846/// reg(b) 1 * (1/3) * (1/2)
4847/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4848/// reg({-1,+,1}) (2/3) * (2/3) * 1
4849/// reg({a,+,1}) (2/3) * 1 * 1
4850/// reg({b,+,1}) 1 * (2/3) * (2/3)
4851/// reg(c) 1 * 1 * 0
4852///
4853/// Now count registers number mathematical expectation for each formula:
4854/// Note that for each use we exclude probability if not selecting for the use.
4855/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4856/// probabilty 1/3 of not selecting for Use1).
4857/// Use1:
4858/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4859/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4860/// reg({a,+,1}) 1
4861/// Use2:
4862/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4863/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4864/// reg({b,+,1}) 2/3
4865/// Use3:
4866/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4867/// reg(c) + reg({b,+,1}) 1 + 2/3
4868void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4869 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4870 return;
4871 // Ok, we have too many of formulae on our hands to conveniently handle.
4872 // Use a rough heuristic to thin out the list.
4873
4874 // Set of Regs wich will be 100% used in final solution.
4875 // Used in each formula of a solution (in example above this is reg(c)).
4876 // We can skip them in calculations.
4877 SmallPtrSet<const SCEV *, 4> UniqRegs;
4878 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)
;
4879
4880 // Map each register to probability of not selecting
4881 DenseMap <const SCEV *, float> RegNumMap;
4882 for (const SCEV *Reg : RegUses) {
4883 if (UniqRegs.count(Reg))
4884 continue;
4885 float PNotSel = 1;
4886 for (const LSRUse &LU : Uses) {
4887 if (!LU.Regs.count(Reg))
4888 continue;
4889 float P = LU.getNotSelectedProbability(Reg);
4890 if (P != 0.0)
4891 PNotSel *= P;
4892 else
4893 UniqRegs.insert(Reg);
4894 }
4895 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4896 }
4897
4898 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4899 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)
;
4900
4901 // Delete formulas where registers number expectation is high.
4902 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4903 LSRUse &LU = Uses[LUIdx];
4904 // If nothing to delete - continue.
4905 if (LU.Formulae.size() < 2)
4906 continue;
4907 // This is temporary solution to test performance. Float should be
4908 // replaced with round independent type (based on integers) to avoid
4909 // different results for different target builds.
4910 float FMinRegNum = LU.Formulae[0].getNumRegs();
4911 float FMinARegNum = LU.Formulae[0].getNumRegs();
4912 size_t MinIdx = 0;
4913 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4914 Formula &F = LU.Formulae[i];
4915 float FRegNum = 0;
4916 float FARegNum = 0;
4917 for (const SCEV *BaseReg : F.BaseRegs) {
4918 if (UniqRegs.count(BaseReg))
4919 continue;
4920 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4921 if (isa<SCEVAddRecExpr>(BaseReg))
4922 FARegNum +=
4923 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4924 }
4925 if (const SCEV *ScaledReg = F.ScaledReg) {
4926 if (!UniqRegs.count(ScaledReg)) {
4927 FRegNum +=
4928 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4929 if (isa<SCEVAddRecExpr>(ScaledReg))
4930 FARegNum +=
4931 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4932 }
4933 }
4934 if (FMinRegNum > FRegNum ||
4935 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4936 FMinRegNum = FRegNum;
4937 FMinARegNum = FARegNum;
4938 MinIdx = i;
4939 }
4940 }
4941 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)
4942 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)
;
4943 if (MinIdx != 0)
4944 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4945 while (LU.Formulae.size() != 1) {
4946 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
)
4947 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4948 LU.Formulae.pop_back();
4949 }
4950 LU.RecomputeRegs(LUIdx, RegUses);
4951 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 4951, __extension__
__PRETTY_FUNCTION__))
;
4952 Formula &F = LU.Formulae[0];
4953 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)
;
4954 // When we choose the formula, the regs become unique.
4955 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4956 if (F.ScaledReg)
4957 UniqRegs.insert(F.ScaledReg);
4958 }
4959 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)
;
4960}
4961
4962/// Pick a register which seems likely to be profitable, and then in any use
4963/// which has any reference to that register, delete all formulae which do not
4964/// reference that register.
4965void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4966 // With all other options exhausted, loop until the system is simple
4967 // enough to handle.
4968 SmallPtrSet<const SCEV *, 4> Taken;
4969 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4970 // Ok, we have too many of formulae on our hands to conveniently handle.
4971 // Use a rough heuristic to thin out the list.
4972 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)
;
4973
4974 // Pick the register which is used by the most LSRUses, which is likely
4975 // to be a good reuse register candidate.
4976 const SCEV *Best = nullptr;
4977 unsigned BestNum = 0;
4978 for (const SCEV *Reg : RegUses) {
4979 if (Taken.count(Reg))
4980 continue;
4981 if (!Best) {
4982 Best = Reg;
4983 BestNum = RegUses.getUsedByIndices(Reg).count();
4984 } else {
4985 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4986 if (Count > BestNum) {
4987 Best = Reg;
4988 BestNum = Count;
4989 }
4990 }
4991 }
4992 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 4992, __extension__
__PRETTY_FUNCTION__))
;
4993
4994 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)
4995 << " 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)
;
4996 Taken.insert(Best);
4997
4998 // In any use with formulae which references this register, delete formulae
4999 // which don't reference it.
5000 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
5001 LSRUse &LU = Uses[LUIdx];
5002 if (!LU.Regs.count(Best)) continue;
5003
5004 bool Any = false;
5005 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
5006 Formula &F = LU.Formulae[i];
5007 if (!F.referencesReg(Best)) {
5008 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)
;
5009 LU.DeleteFormula(F);
5010 --e;
5011 --i;
5012 Any = true;
5013 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?\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5013, __extension__
__PRETTY_FUNCTION__))
;
5014 continue;
5015 }
5016 }
5017
5018 if (Any)
5019 LU.RecomputeRegs(LUIdx, RegUses);
5020 }
5021
5022 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)
;
5023 }
5024}
5025
5026/// If there are an extraordinary number of formulae to choose from, use some
5027/// rough heuristics to prune down the number of formulae. This keeps the main
5028/// solver from taking an extraordinary amount of time in some worst-case
5029/// scenarios.
5030void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
5031 NarrowSearchSpaceByDetectingSupersets();
5032 NarrowSearchSpaceByCollapsingUnrolledCode();
5033 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
5034 if (FilterSameScaledReg)
5035 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
5036 NarrowSearchSpaceByFilterPostInc();
5037 if (LSRExpNarrow)
5038 NarrowSearchSpaceByDeletingCostlyFormulas();
5039 else
5040 NarrowSearchSpaceByPickingWinnerRegs();
5041}
5042
5043/// This is the recursive solver.
5044void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
5045 Cost &SolutionCost,
5046 SmallVectorImpl<const Formula *> &Workspace,
5047 const Cost &CurCost,
5048 const SmallPtrSet<const SCEV *, 16> &CurRegs,
5049 DenseSet<const SCEV *> &VisitedRegs) const {
5050 // Some ideas:
5051 // - prune more:
5052 // - use more aggressive filtering
5053 // - sort the formula so that the most profitable solutions are found first
5054 // - sort the uses too
5055 // - search faster:
5056 // - don't compute a cost, and then compare. compare while computing a cost
5057 // and bail early.
5058 // - track register sets with SmallBitVector
5059
5060 const LSRUse &LU = Uses[Workspace.size()];
5061
5062 // If this use references any register that's already a part of the
5063 // in-progress solution, consider it a requirement that a formula must
5064 // reference that register in order to be considered. This prunes out
5065 // unprofitable searching.
5066 SmallSetVector<const SCEV *, 4> ReqRegs;
5067 for (const SCEV *S : CurRegs)
5068 if (LU.Regs.count(S))
5069 ReqRegs.insert(S);
5070
5071 SmallPtrSet<const SCEV *, 16> NewRegs;
5072 Cost NewCost(L, SE, TTI, AMK);
5073 for (const Formula &F : LU.Formulae) {
5074 // Ignore formulae which may not be ideal in terms of register reuse of
5075 // ReqRegs. The formula should use all required registers before
5076 // introducing new ones.
5077 // This can sometimes (notably when trying to favour postinc) lead to
5078 // sub-optimial decisions. There it is best left to the cost modelling to
5079 // get correct.
5080 if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) {
5081 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
5082 for (const SCEV *Reg : ReqRegs) {
5083 if ((F.ScaledReg && F.ScaledReg == Reg) ||
5084 is_contained(F.BaseRegs, Reg)) {
5085 --NumReqRegsToFind;
5086 if (NumReqRegsToFind == 0)
5087 break;
5088 }
5089 }
5090 if (NumReqRegsToFind != 0) {
5091 // If none of the formulae satisfied the required registers, then we could
5092 // clear ReqRegs and try again. Currently, we simply give up in this case.
5093 continue;
5094 }
5095 }
5096
5097 // Evaluate the cost of the current formula. If it's already worse than
5098 // the current best, prune the search at that point.
5099 NewCost = CurCost;
5100 NewRegs = CurRegs;
5101 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
5102 if (NewCost.isLess(SolutionCost)) {
5103 Workspace.push_back(&F);
5104 if (Workspace.size() != Uses.size()) {
5105 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
5106 NewRegs, VisitedRegs);
5107 if (F.getNumRegs() == 1 && Workspace.size() == 1)
5108 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
5109 } else {
5110 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)
5111 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)
5112 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)
5113 << "- " << *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)
5114 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)
;
5115
5116 SolutionCost = NewCost;
5117 Solution = Workspace;
5118 }
5119 Workspace.pop_back();
5120 }
5121 }
5122}
5123
5124/// Choose one formula from each use. Return the results in the given Solution
5125/// vector.
5126void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
5127 SmallVector<const Formula *, 8> Workspace;
5128 Cost SolutionCost(L, SE, TTI, AMK);
5129 SolutionCost.Lose();
5130 Cost CurCost(L, SE, TTI, AMK);
5131 SmallPtrSet<const SCEV *, 16> CurRegs;
5132 DenseSet<const SCEV *> VisitedRegs;
5133 Workspace.reserve(Uses.size());
5134
5135 // SolveRecurse does all the work.
5136 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
5137 CurRegs, VisitedRegs);
5138 if (Solution.empty()) {
5139 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
5140 return;
5141 }
5142
5143 // Ok, we've now made all our decisions.
5144 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
)
5145 "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
)
5146 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
)
5147 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
)
5148 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
)
5149 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
)
5150 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
)
5151 " ";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
)
5152 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
)
5153 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
)
5154 })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
)
;
5155
5156 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5156, __extension__
__PRETTY_FUNCTION__))
;
5157}
5158
5159/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5160/// we can go while still being dominated by the input positions. This helps
5161/// canonicalize the insert position, which encourages sharing.
5162BasicBlock::iterator
5163LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5164 const SmallVectorImpl<Instruction *> &Inputs)
5165 const {
5166 Instruction *Tentative = &*IP;
5167 while (true) {
5168 bool AllDominate = true;
5169 Instruction *BetterPos = nullptr;
5170 // Don't bother attempting to insert before a catchswitch, their basic block
5171 // cannot have other non-PHI instructions.
5172 if (isa<CatchSwitchInst>(Tentative))
5173 return IP;
5174
5175 for (Instruction *Inst : Inputs) {
5176 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5177 AllDominate = false;
5178 break;
5179 }
5180 // Attempt to find an insert position in the middle of the block,
5181 // instead of at the end, so that it can be used for other expansions.
5182 if (Tentative->getParent() == Inst->getParent() &&
5183 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5184 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5185 }
5186 if (!AllDominate)
5187 break;
5188 if (BetterPos)
5189 IP = BetterPos->getIterator();
5190 else
5191 IP = Tentative->getIterator();
5192
5193 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5194 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5195
5196 BasicBlock *IDom;
5197 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5198 if (!Rung) return IP;
5199 Rung = Rung->getIDom();
5200 if (!Rung) return IP;
5201 IDom = Rung->getBlock();
5202
5203 // Don't climb into a loop though.
5204 const Loop *IDomLoop = LI.getLoopFor(IDom);
5205 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5206 if (IDomDepth <= IPLoopDepth &&
5207 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5208 break;
5209 }
5210
5211 Tentative = IDom->getTerminator();
5212 }
5213
5214 return IP;
5215}
5216
5217/// Determine an input position which will be dominated by the operands and
5218/// which will dominate the result.
5219BasicBlock::iterator LSRInstance::AdjustInsertPositionForExpand(
5220 BasicBlock::iterator LowestIP, const LSRFixup &LF, const LSRUse &LU) const {
5221 // Collect some instructions which must be dominated by the
5222 // expanding replacement. These must be dominated by any operands that
5223 // will be required in the expansion.
5224 SmallVector<Instruction *, 4> Inputs;
5225 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5226 Inputs.push_back(I);
5227 if (LU.Kind == LSRUse::ICmpZero)
5228 if (Instruction *I =
5229 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5230 Inputs.push_back(I);
5231 if (LF.PostIncLoops.count(L)) {
5232 if (LF.isUseFullyOutsideLoop(L))
5233 Inputs.push_back(L->getLoopLatch()->getTerminator());
5234 else
5235 Inputs.push_back(IVIncInsertPos);
5236 }
5237 // The expansion must also be dominated by the increment positions of any
5238 // loops it for which it is using post-inc mode.
5239 for (const Loop *PIL : LF.PostIncLoops) {
5240 if (PIL == L) continue;
5241
5242 // Be dominated by the loop exit.
5243 SmallVector<BasicBlock *, 4> ExitingBlocks;
5244 PIL->getExitingBlocks(ExitingBlocks);
5245 if (!ExitingBlocks.empty()) {
5246 BasicBlock *BB = ExitingBlocks[0];
5247 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5248 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5249 Inputs.push_back(BB->getTerminator());
5250 }
5251 }
5252
5253 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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5255, __extension__
__PRETTY_FUNCTION__))
5254 && !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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5255, __extension__
__PRETTY_FUNCTION__))
5255 "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\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5255, __extension__
__PRETTY_FUNCTION__))
;
5256
5257 // Then, climb up the immediate dominator tree as far as we can go while
5258 // still being dominated by the input positions.
5259 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5260
5261 // Don't insert instructions before PHI nodes.
5262 while (isa<PHINode>(IP)) ++IP;
5263
5264 // Ignore landingpad instructions.
5265 while (IP->isEHPad()) ++IP;
5266
5267 // Ignore debug intrinsics.
5268 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5269
5270 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5271 // IP consistent across expansions and allows the previously inserted
5272 // instructions to be reused by subsequent expansion.
5273 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5274 ++IP;
5275
5276 return IP;
5277}
5278
5279/// Emit instructions for the leading candidate expression for this LSRUse (this
5280/// is called "expanding").
5281Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5282 const Formula &F, BasicBlock::iterator IP,
5283 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5284 if (LU.RigidFormula)
5285 return LF.OperandValToReplace;
5286
5287 // Determine an input position which will be dominated by the operands and
5288 // which will dominate the result.
5289 IP = AdjustInsertPositionForExpand(IP, LF, LU);
5290 Rewriter.setInsertPoint(&*IP);
5291
5292 // Inform the Rewriter if we have a post-increment use, so that it can
5293 // perform an advantageous expansion.
5294 Rewriter.setPostInc(LF.PostIncLoops);
5295
5296 // This is the type that the user actually needs.
5297 Type *OpTy = LF.OperandValToReplace->getType();
5298 // This will be the type that we'll initially expand to.
5299 Type *Ty = F.getType();
5300 if (!Ty)
5301 // No type known; just expand directly to the ultimate type.
5302 Ty = OpTy;
5303 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5304 // Expand directly to the ultimate type if it's the right size.
5305 Ty = OpTy;
5306 // This is the type to do integer arithmetic in.
5307 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5308
5309 // Build up a list of operands to add together to form the full base.
5310 SmallVector<const SCEV *, 8> Ops;
5311
5312 // Expand the BaseRegs portion.
5313 for (const SCEV *Reg : F.BaseRegs) {
5314 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5314, __extension__
__PRETTY_FUNCTION__))
;
5315
5316 // If we're expanding for a post-inc user, make the post-inc adjustment.
5317 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5318 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5319 }
5320
5321 // Expand the ScaledReg portion.
5322 Value *ICmpScaledV = nullptr;
5323 if (F.Scale != 0) {
5324 const SCEV *ScaledS = F.ScaledReg;
5325
5326 // If we're expanding for a post-inc user, make the post-inc adjustment.
5327 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5328 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5329
5330 if (LU.Kind == LSRUse::ICmpZero) {
5331 // Expand ScaleReg as if it was part of the base regs.
5332 if (F.Scale == 1)
5333 Ops.push_back(
5334 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5335 else {
5336 // An interesting way of "folding" with an icmp is to use a negated
5337 // scale, which we'll implement by inserting it into the other operand
5338 // of the icmp.
5339 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5340, __extension__
__PRETTY_FUNCTION__))
5340 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5340, __extension__
__PRETTY_FUNCTION__))
;
5341 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5342 }
5343 } else {
5344 // Otherwise just expand the scaled register and an explicit scale,
5345 // which is expected to be matched as part of the address.
5346
5347 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5348 // Unless the addressing mode will not be folded.
5349 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5350 isAMCompletelyFolded(TTI, LU, F)) {
5351 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5352 Ops.clear();
5353 Ops.push_back(SE.getUnknown(FullV));
5354 }
5355 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5356 if (F.Scale != 1)
5357 ScaledS =
5358 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5359 Ops.push_back(ScaledS);
5360 }
5361 }
5362
5363 // Expand the GV portion.
5364 if (F.BaseGV) {
5365 // Flush the operand list to suppress SCEVExpander hoisting.
5366 if (!Ops.empty()) {
5367 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), IntTy);
5368 Ops.clear();
5369 Ops.push_back(SE.getUnknown(FullV));
5370 }
5371 Ops.push_back(SE.getUnknown(F.BaseGV));
5372 }
5373
5374 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5375 // unfolded offsets. LSR assumes they both live next to their uses.
5376 if (!Ops.empty()) {
5377 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5378 Ops.clear();
5379 Ops.push_back(SE.getUnknown(FullV));
5380 }
5381
5382 // Expand the immediate portion.
5383 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5384 if (Offset != 0) {
5385 if (LU.Kind == LSRUse::ICmpZero) {
5386 // The other interesting way of "folding" with an ICmpZero is to use a
5387 // negated immediate.
5388 if (!ICmpScaledV)
5389 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5390 else {
5391 Ops.push_back(SE.getUnknown(ICmpScaledV));
5392 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5393 }
5394 } else {
5395 // Just add the immediate values. These again are expected to be matched
5396 // as part of the address.
5397 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5398 }
5399 }
5400
5401 // Expand the unfolded offset portion.
5402 int64_t UnfoldedOffset = F.UnfoldedOffset;
5403 if (UnfoldedOffset != 0) {
5404 // Just add the immediate values.
5405 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5406 UnfoldedOffset)));
5407 }
5408
5409 // Emit instructions summing all the operands.
5410 const SCEV *FullS = Ops.empty() ?
5411 SE.getConstant(IntTy, 0) :
5412 SE.getAddExpr(Ops);
5413 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5414
5415 // We're done expanding now, so reset the rewriter.
5416 Rewriter.clearPostInc();
5417
5418 // An ICmpZero Formula represents an ICmp which we're handling as a
5419 // comparison against zero. Now that we've expanded an expression for that
5420 // form, update the ICmp's other operand.
5421 if (LU.Kind == LSRUse::ICmpZero) {
5422 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5423 if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
5424 DeadInsts.emplace_back(OperandIsInstr);
5425 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5426, __extension__
__PRETTY_FUNCTION__))
5426 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5426, __extension__
__PRETTY_FUNCTION__))
;
5427 if (F.Scale == -1) {
5428 if (ICmpScaledV->getType() != OpTy) {
5429 Instruction *Cast =
5430 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5431 OpTy, false),
5432 ICmpScaledV, OpTy, "tmp", CI);
5433 ICmpScaledV = Cast;
5434 }
5435 CI->setOperand(1, ICmpScaledV);
5436 } else {
5437 // A scale of 1 means that the scale has been expanded as part of the
5438 // base regs.
5439 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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5441, __extension__
__PRETTY_FUNCTION__))
5440 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5441, __extension__
__PRETTY_FUNCTION__))
5441 "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!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5441, __extension__
__PRETTY_FUNCTION__))
;
5442 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5443 -(uint64_t)Offset);
5444 if (C->getType() != OpTy)
5445 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5446 OpTy, false),
5447 C, OpTy);
5448
5449 CI->setOperand(1, C);
5450 }
5451 }
5452
5453 return FullV;
5454}
5455
5456/// Helper for Rewrite. PHI nodes are special because the use of their operands
5457/// effectively happens in their predecessor blocks, so the expression may need
5458/// to be expanded in multiple places.
5459void LSRInstance::RewriteForPHI(
5460 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5461 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5462 DenseMap<BasicBlock *, Value *> Inserted;
5463 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5464 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5465 bool needUpdateFixups = false;
5466 BasicBlock *BB = PN->getIncomingBlock(i);
5467
5468 // If this is a critical edge, split the edge so that we do not insert
5469 // the code on all predecessor/successor paths. We do this unless this
5470 // is the canonical backedge for this loop, which complicates post-inc
5471 // users.
5472 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5473 !isa<IndirectBrInst>(BB->getTerminator()) &&
5474 !isa<CatchSwitchInst>(BB->getTerminator())) {
5475 BasicBlock *Parent = PN->getParent();
5476 Loop *PNLoop = LI.getLoopFor(Parent);
5477 if (!PNLoop || Parent != PNLoop->getHeader()) {
5478 // Split the critical edge.
5479 BasicBlock *NewBB = nullptr;
5480 if (!Parent->isLandingPad()) {
5481 NewBB =
5482 SplitCriticalEdge(BB, Parent,
5483 CriticalEdgeSplittingOptions(&DT, &LI, MSSAU)
5484 .setMergeIdenticalEdges()
5485 .setKeepOneInputPHIs());
5486 } else {
5487 SmallVector<BasicBlock*, 2> NewBBs;
5488 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5489 NewBB = NewBBs[0];
5490 }
5491 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5492 // phi predecessors are identical. The simple thing to do is skip
5493 // splitting in this case rather than complicate the API.
5494 if (NewBB) {
5495 // If PN is outside of the loop and BB is in the loop, we want to
5496 // move the block to be immediately before the PHI block, not
5497 // immediately after BB.
5498 if (L->contains(BB) && !L->contains(PN))
5499 NewBB->moveBefore(PN->getParent());
5500
5501 // Splitting the edge can reduce the number of PHI entries we have.
5502 e = PN->getNumIncomingValues();
5503 BB = NewBB;
5504 i = PN->getBasicBlockIndex(BB);
5505
5506 needUpdateFixups = true;
5507 }
5508 }
5509 }
5510
5511 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5512 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5513 if (!Pair.second)
5514 PN->setIncomingValue(i, Pair.first->second);
5515 else {
5516 Value *FullV =
5517 Expand(LU, LF, F, BB->getTerminator()->getIterator(), DeadInsts);
5518
5519 // If this is reuse-by-noop-cast, insert the noop cast.
5520 Type *OpTy = LF.OperandValToReplace->getType();
5521 if (FullV->getType() != OpTy)
5522 FullV =
5523 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5524 OpTy, false),
5525 FullV, LF.OperandValToReplace->getType(),
5526 "tmp", BB->getTerminator());
5527
5528 PN->setIncomingValue(i, FullV);
5529 Pair.first->second = FullV;
5530 }
5531
5532 // If LSR splits critical edge and phi node has other pending
5533 // fixup operands, we need to update those pending fixups. Otherwise
5534 // formulae will not be implemented completely and some instructions
5535 // will not be eliminated.
5536 if (needUpdateFixups) {
5537 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5538 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5539 // If fixup is supposed to rewrite some operand in the phi
5540 // that was just updated, it may be already moved to
5541 // another phi node. Such fixup requires update.
5542 if (Fixup.UserInst == PN) {
5543 // Check if the operand we try to replace still exists in the
5544 // original phi.
5545 bool foundInOriginalPHI = false;
5546 for (const auto &val : PN->incoming_values())
5547 if (val == Fixup.OperandValToReplace) {
5548 foundInOriginalPHI = true;
5549 break;
5550 }
5551
5552 // If fixup operand found in original PHI - nothing to do.
5553 if (foundInOriginalPHI)
5554 continue;
5555
5556 // Otherwise it might be moved to another PHI and requires update.
5557 // If fixup operand not found in any of the incoming blocks that
5558 // means we have already rewritten it - nothing to do.
5559 for (const auto &Block : PN->blocks())
5560 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5561 ++I) {
5562 PHINode *NewPN = cast<PHINode>(I);
5563 for (const auto &val : NewPN->incoming_values())
5564 if (val == Fixup.OperandValToReplace)
5565 Fixup.UserInst = NewPN;
5566 }
5567 }
5568 }
5569 }
5570}
5571
5572/// Emit instructions for the leading candidate expression for this LSRUse (this
5573/// is called "expanding"), and update the UserInst to reference the newly
5574/// expanded value.
5575void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5576 const Formula &F,
5577 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5578 // First, find an insertion point that dominates UserInst. For PHI nodes,
5579 // find the nearest block which dominates all the relevant uses.
5580 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5581 RewriteForPHI(PN, LU, LF, F, DeadInsts);
5582 } else {
5583 Value *FullV = Expand(LU, LF, F, LF.UserInst->getIterator(), DeadInsts);
5584
5585 // If this is reuse-by-noop-cast, insert the noop cast.
5586 Type *OpTy = LF.OperandValToReplace->getType();
5587 if (FullV->getType() != OpTy) {
5588 Instruction *Cast =
5589 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5590 FullV, OpTy, "tmp", LF.UserInst);
5591 FullV = Cast;
5592 }
5593
5594 // Update the user. ICmpZero is handled specially here (for now) because
5595 // Expand may have updated one of the operands of the icmp already, and
5596 // its new value may happen to be equal to LF.OperandValToReplace, in
5597 // which case doing replaceUsesOfWith leads to replacing both operands
5598 // with the same value. TODO: Reorganize this.
5599 if (LU.Kind == LSRUse::ICmpZero)
5600 LF.UserInst->setOperand(0, FullV);
5601 else
5602 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5603 }
5604
5605 if (auto *OperandIsInstr = dyn_cast<Instruction>(LF.OperandValToReplace))
5606 DeadInsts.emplace_back(OperandIsInstr);
5607}
5608
5609/// Rewrite all the fixup locations with new values, following the chosen
5610/// solution.
5611void LSRInstance::ImplementSolution(
5612 const SmallVectorImpl<const Formula *> &Solution) {
5613 // Keep track of instructions we may have made dead, so that
5614 // we can remove them after we are done working.
5615 SmallVector<WeakTrackingVH, 16> DeadInsts;
5616
5617 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5618
5619 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5620 for (const IVChain &Chain : IVChainVec) {
25
Assuming '__begin1' is equal to '__end1'
5621 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5622 Rewriter.setChainedPhi(PN);
5623 }
5624
5625 // Expand the new value definitions and update the users.
5626 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
26
Assuming 'LUIdx' is equal to 'NumUses'
27
Loop condition is false. Execution continues on line 5632
5627 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5628 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], DeadInsts);
5629 Changed = true;
5630 }
5631
5632 for (const IVChain &Chain : IVChainVec) {
28
Assuming '__begin1' is equal to '__end1'
5633 GenerateIVChain(Chain, DeadInsts);
5634 Changed = true;
5635 }
5636
5637 for (const WeakVH &IV : Rewriter.getInsertedIVs())
29
Assuming '__begin1' is not equal to '__end1'
5638 if (IV && dyn_cast<Instruction>(&*IV)->getParent())
30
Assuming the condition is true
31
Assuming the object is not a 'CastReturnType'
32
Called C++ object pointer is null
5639 ScalarEvolutionIVs.push_back(IV);
5640
5641 // Clean up after ourselves. This must be done before deleting any
5642 // instructions.
5643 Rewriter.clear();
5644
5645 Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts,
5646 &TLI, MSSAU);
5647
5648 // In our cost analysis above, we assume that each addrec consumes exactly
5649 // one register, and arrange to have increments inserted just before the
5650 // latch to maximimize the chance this is true. However, if we reused
5651 // existing IVs, we now need to move the increments to match our