Bug Summary

File:build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:line 5288, column 16
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-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -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-15~++20220420111733+e13d2efed663/build-llvm -resource-dir /usr/lib/llvm-15/lib/clang/15.0.0 -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-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/Scalar -I include -I /build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/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-15/lib/clang/15.0.0/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-15~++20220420111733+e13d2efed663/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/= -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 -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/= -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-2022-04-20-140412-16051-1 -x c++ /build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/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 comparision.
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(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 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", 1358, __extension__
__PRETTY_FUNCTION__))
;
1359 // Tally up the registers.
1360 unsigned PrevAddRecCost = C.AddRecCost;
1361 unsigned PrevNumRegs = C.NumRegs;
1362 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1363 if (const SCEV *ScaledReg = F.ScaledReg) {
1364 if (VisitedRegs.count(ScaledReg)) {
1365 Lose();
1366 return;
1367 }
1368 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1369 if (isLoser())
1370 return;
1371 }
1372 for (const SCEV *BaseReg : F.BaseRegs) {
1373 if (VisitedRegs.count(BaseReg)) {
1374 Lose();
1375 return;
1376 }
1377 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1378 if (isLoser())
1379 return;
1380 }
1381
1382 // Determine how many (unfolded) adds we'll need inside the loop.
1383 size_t NumBaseParts = F.getNumRegs();
1384 if (NumBaseParts > 1)
1385 // Do not count the base and a possible second register if the target
1386 // allows to fold 2 registers.
1387 C.NumBaseAdds +=
1388 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1389 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1390
1391 // Accumulate non-free scaling amounts.
1392 C.ScaleCost += *getScalingFactorCost(*TTI, LU, F, *L).getValue();
1393
1394 // Tally up the non-zero immediates.
1395 for (const LSRFixup &Fixup : LU.Fixups) {
1396 int64_t O = Fixup.Offset;
1397 int64_t Offset = (uint64_t)O + F.BaseOffset;
1398 if (F.BaseGV)
1399 C.ImmCost += 64; // Handle symbolic values conservatively.
1400 // TODO: This should probably be the pointer size.
1401 else if (Offset != 0)
1402 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1403
1404 // Check with target if this offset with this instruction is
1405 // specifically not supported.
1406 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1407 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1408 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1409 C.NumBaseAdds++;
1410 }
1411
1412 // If we don't count instruction cost exit here.
1413 if (!InsnsCost) {
1414 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1414, __extension__
__PRETTY_FUNCTION__))
;
1415 return;
1416 }
1417
1418 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1419 // additional instruction (at least fill).
1420 // TODO: Need distinguish register class?
1421 unsigned TTIRegNum = TTI->getNumberOfRegisters(
1422 TTI->getRegisterClassForType(false, F.getType())) - 1;
1423 if (C.NumRegs > TTIRegNum) {
1424 // Cost already exceeded TTIRegNum, then only newly added register can add
1425 // new instructions.
1426 if (PrevNumRegs > TTIRegNum)
1427 C.Insns += (C.NumRegs - PrevNumRegs);
1428 else
1429 C.Insns += (C.NumRegs - TTIRegNum);
1430 }
1431
1432 // If ICmpZero formula ends with not 0, it could not be replaced by
1433 // just add or sub. We'll need to compare final result of AddRec.
1434 // That means we'll need an additional instruction. But if the target can
1435 // macro-fuse a compare with a branch, don't count this extra instruction.
1436 // For -10 + {0, +, 1}:
1437 // i = i + 1;
1438 // cmp i, 10
1439 //
1440 // For {-10, +, 1}:
1441 // i = i + 1;
1442 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1443 !TTI->canMacroFuseCmp())
1444 C.Insns++;
1445 // Each new AddRec adds 1 instruction to calculation.
1446 C.Insns += (C.AddRecCost - PrevAddRecCost);
1447
1448 // BaseAdds adds instructions for unfolded registers.
1449 if (LU.Kind != LSRUse::ICmpZero)
1450 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1451 assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1451, __extension__
__PRETTY_FUNCTION__))
;
1452}
1453
1454/// Set this cost to a losing value.
1455void Cost::Lose() {
1456 C.Insns = std::numeric_limits<unsigned>::max();
1457 C.NumRegs = std::numeric_limits<unsigned>::max();
1458 C.AddRecCost = std::numeric_limits<unsigned>::max();
1459 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1460 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1461 C.ImmCost = std::numeric_limits<unsigned>::max();
1462 C.SetupCost = std::numeric_limits<unsigned>::max();
1463 C.ScaleCost = std::numeric_limits<unsigned>::max();
1464}
1465
1466/// Choose the lower cost.
1467bool Cost::isLess(Cost &Other) {
1468 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1469 C.Insns != Other.C.Insns)
1470 return C.Insns < Other.C.Insns;
1471 return TTI->isLSRCostLess(C, Other.C);
1472}
1473
1474#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1475void Cost::print(raw_ostream &OS) const {
1476 if (InsnsCost)
1477 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1478 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1479 if (C.AddRecCost != 0)
1480 OS << ", with addrec cost " << C.AddRecCost;
1481 if (C.NumIVMuls != 0)
1482 OS << ", plus " << C.NumIVMuls << " IV mul"
1483 << (C.NumIVMuls == 1 ? "" : "s");
1484 if (C.NumBaseAdds != 0)
1485 OS << ", plus " << C.NumBaseAdds << " base add"
1486 << (C.NumBaseAdds == 1 ? "" : "s");
1487 if (C.ScaleCost != 0)
1488 OS << ", plus " << C.ScaleCost << " scale cost";
1489 if (C.ImmCost != 0)
1490 OS << ", plus " << C.ImmCost << " imm cost";
1491 if (C.SetupCost != 0)
1492 OS << ", plus " << C.SetupCost << " setup cost";
1493}
1494
1495LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
1496 print(errs()); errs() << '\n';
1497}
1498#endif
1499
1500/// Test whether this fixup always uses its value outside of the given loop.
1501bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1502 // PHI nodes use their value in their incoming blocks.
1503 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1504 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1505 if (PN->getIncomingValue(i) == OperandValToReplace &&
1506 L->contains(PN->getIncomingBlock(i)))
1507 return false;
1508 return true;
1509 }
1510
1511 return !L->contains(UserInst);
1512}
1513
1514#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1515void LSRFixup::print(raw_ostream &OS) const {
1516 OS << "UserInst=";
1517 // Store is common and interesting enough to be worth special-casing.
1518 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1519 OS << "store ";
1520 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1521 } else if (UserInst->getType()->isVoidTy())
1522 OS << UserInst->getOpcodeName();
1523 else
1524 UserInst->printAsOperand(OS, /*PrintType=*/false);
1525
1526 OS << ", OperandValToReplace=";
1527 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1528
1529 for (const Loop *PIL : PostIncLoops) {
1530 OS << ", PostIncLoop=";
1531 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1532 }
1533
1534 if (Offset != 0)
1535 OS << ", Offset=" << Offset;
1536}
1537
1538LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
1539 print(errs()); errs() << '\n';
1540}
1541#endif
1542
1543/// Test whether this use as a formula which has the same registers as the given
1544/// formula.
1545bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1546 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1547 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1548 // Unstable sort by host order ok, because this is only used for uniquifying.
1549 llvm::sort(Key);
1550 return Uniquifier.count(Key);
1551}
1552
1553/// The function returns a probability of selecting formula without Reg.
1554float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1555 unsigned FNum = 0;
1556 for (const Formula &F : Formulae)
1557 if (F.referencesReg(Reg))
1558 FNum++;
1559 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1560}
1561
1562/// If the given formula has not yet been inserted, add it to the list, and
1563/// return true. Return false otherwise. The formula must be in canonical form.
1564bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1565 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", 1565, __extension__
__PRETTY_FUNCTION__))
;
1566
1567 if (!Formulae.empty() && RigidFormula)
1568 return false;
1569
1570 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1571 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1572 // Unstable sort by host order ok, because this is only used for uniquifying.
1573 llvm::sort(Key);
1574
1575 if (!Uniquifier.insert(Key).second)
1576 return false;
1577
1578 // Using a register to hold the value of 0 is not profitable.
1579 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", 1580, __extension__
__PRETTY_FUNCTION__))
1580 "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", 1580, __extension__
__PRETTY_FUNCTION__))
;
1581#ifndef NDEBUG
1582 for (const SCEV *BaseReg : F.BaseRegs)
1583 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", 1583, __extension__
__PRETTY_FUNCTION__))
;
1584#endif
1585
1586 // Add the formula to the list.
1587 Formulae.push_back(F);
1588
1589 // Record registers now being used by this use.
1590 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1591 if (F.ScaledReg)
1592 Regs.insert(F.ScaledReg);
1593
1594 return true;
1595}
1596
1597/// Remove the given formula from this use's list.
1598void LSRUse::DeleteFormula(Formula &F) {
1599 if (&F != &Formulae.back())
1600 std::swap(F, Formulae.back());
1601 Formulae.pop_back();
1602}
1603
1604/// Recompute the Regs field, and update RegUses.
1605void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1606 // Now that we've filtered out some formulae, recompute the Regs set.
1607 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1608 Regs.clear();
1609 for (const Formula &F : Formulae) {
1610 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1611 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1612 }
1613
1614 // Update the RegTracker.
1615 for (const SCEV *S : OldRegs)
1616 if (!Regs.count(S))
1617 RegUses.dropRegister(S, LUIdx);
1618}
1619
1620#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1621void LSRUse::print(raw_ostream &OS) const {
1622 OS << "LSR Use: Kind=";
1623 switch (Kind) {
1624 case Basic: OS << "Basic"; break;
1625 case Special: OS << "Special"; break;
1626 case ICmpZero: OS << "ICmpZero"; break;
1627 case Address:
1628 OS << "Address of ";
1629 if (AccessTy.MemTy->isPointerTy())
1630 OS << "pointer"; // the full pointer type could be really verbose
1631 else {
1632 OS << *AccessTy.MemTy;
1633 }
1634
1635 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1636 }
1637
1638 OS << ", Offsets={";
1639 bool NeedComma = false;
1640 for (const LSRFixup &Fixup : Fixups) {
1641 if (NeedComma) OS << ',';
1642 OS << Fixup.Offset;
1643 NeedComma = true;
1644 }
1645 OS << '}';
1646
1647 if (AllFixupsOutsideLoop)
1648 OS << ", all-fixups-outside-loop";
1649
1650 if (WidestFixupType)
1651 OS << ", widest fixup type: " << *WidestFixupType;
1652}
1653
1654LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
1655 print(errs()); errs() << '\n';
1656}
1657#endif
1658
1659static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1660 LSRUse::KindType Kind, MemAccessTy AccessTy,
1661 GlobalValue *BaseGV, int64_t BaseOffset,
1662 bool HasBaseReg, int64_t Scale,
1663 Instruction *Fixup/*= nullptr*/) {
1664 switch (Kind) {
1665 case LSRUse::Address:
1666 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1667 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1668
1669 case LSRUse::ICmpZero:
1670 // There's not even a target hook for querying whether it would be legal to
1671 // fold a GV into an ICmp.
1672 if (BaseGV)
1673 return false;
1674
1675 // ICmp only has two operands; don't allow more than two non-trivial parts.
1676 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1677 return false;
1678
1679 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1680 // putting the scaled register in the other operand of the icmp.
1681 if (Scale != 0 && Scale != -1)
1682 return false;
1683
1684 // If we have low-level target information, ask the target if it can fold an
1685 // integer immediate on an icmp.
1686 if (BaseOffset != 0) {
1687 // We have one of:
1688 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1689 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1690 // Offs is the ICmp immediate.
1691 if (Scale == 0)
1692 // The cast does the right thing with
1693 // std::numeric_limits<int64_t>::min().
1694 BaseOffset = -(uint64_t)BaseOffset;
1695 return TTI.isLegalICmpImmediate(BaseOffset);
1696 }
1697
1698 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1699 return true;
1700
1701 case LSRUse::Basic:
1702 // Only handle single-register values.
1703 return !BaseGV && Scale == 0 && BaseOffset == 0;
1704
1705 case LSRUse::Special:
1706 // Special case Basic to handle -1 scales.
1707 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1708 }
1709
1710 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1710)
;
1711}
1712
1713static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1714 int64_t MinOffset, int64_t MaxOffset,
1715 LSRUse::KindType Kind, MemAccessTy AccessTy,
1716 GlobalValue *BaseGV, int64_t BaseOffset,
1717 bool HasBaseReg, int64_t Scale) {
1718 // Check for overflow.
1719 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1720 (MinOffset > 0))
1721 return false;
1722 MinOffset = (uint64_t)BaseOffset + MinOffset;
1723 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1724 (MaxOffset > 0))
1725 return false;
1726 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1727
1728 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1729 HasBaseReg, Scale) &&
1730 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1731 HasBaseReg, Scale);
1732}
1733
1734static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1735 int64_t MinOffset, int64_t MaxOffset,
1736 LSRUse::KindType Kind, MemAccessTy AccessTy,
1737 const Formula &F, const Loop &L) {
1738 // For the purpose of isAMCompletelyFolded either having a canonical formula
1739 // or a scale not equal to zero is correct.
1740 // Problems may arise from non canonical formulae having a scale == 0.
1741 // Strictly speaking it would best to just rely on canonical formulae.
1742 // However, when we generate the scaled formulae, we first check that the
1743 // scaling factor is profitable before computing the actual ScaledReg for
1744 // compile time sake.
1745 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", 1745, __extension__
__PRETTY_FUNCTION__))
;
1746 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1747 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1748}
1749
1750/// Test whether we know how to expand the current formula.
1751static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1752 int64_t MaxOffset, LSRUse::KindType Kind,
1753 MemAccessTy AccessTy, GlobalValue *BaseGV,
1754 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1755 // We know how to expand completely foldable formulae.
1756 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1757 BaseOffset, HasBaseReg, Scale) ||
1758 // Or formulae that use a base register produced by a sum of base
1759 // registers.
1760 (Scale == 1 &&
1761 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1762 BaseGV, BaseOffset, true, 0));
1763}
1764
1765static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1766 int64_t MaxOffset, LSRUse::KindType Kind,
1767 MemAccessTy AccessTy, const Formula &F) {
1768 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1769 F.BaseOffset, F.HasBaseReg, F.Scale);
1770}
1771
1772static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1773 const LSRUse &LU, const Formula &F) {
1774 // Target may want to look at the user instructions.
1775 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1776 for (const LSRFixup &Fixup : LU.Fixups)
1777 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1778 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1779 F.Scale, Fixup.UserInst))
1780 return false;
1781 return true;
1782 }
1783
1784 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1785 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1786 F.Scale);
1787}
1788
1789static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1790 const LSRUse &LU, const Formula &F,
1791 const Loop &L) {
1792 if (!F.Scale)
1793 return 0;
1794
1795 // If the use is not completely folded in that instruction, we will have to
1796 // pay an extra cost only for scale != 1.
1797 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1798 LU.AccessTy, F, L))
1799 return F.Scale != 1;
1800
1801 switch (LU.Kind) {
1802 case LSRUse::Address: {
1803 // Check the scaling factor cost with both the min and max offsets.
1804 InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost(
1805 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1806 F.Scale, LU.AccessTy.AddrSpace);
1807 InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost(
1808 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1809 F.Scale, LU.AccessTy.AddrSpace);
1810
1811 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", 1812, __extension__
__PRETTY_FUNCTION__))
1812 "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", 1812, __extension__
__PRETTY_FUNCTION__))
;
1813 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1814 }
1815 case LSRUse::ICmpZero:
1816 case LSRUse::Basic:
1817 case LSRUse::Special:
1818 // The use is completely folded, i.e., everything is folded into the
1819 // instruction.
1820 return 0;
1821 }
1822
1823 llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1823)
;
1824}
1825
1826static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1827 LSRUse::KindType Kind, MemAccessTy AccessTy,
1828 GlobalValue *BaseGV, int64_t BaseOffset,
1829 bool HasBaseReg) {
1830 // Fast-path: zero is always foldable.
1831 if (BaseOffset == 0 && !BaseGV) return true;
1832
1833 // Conservatively, create an address with an immediate and a
1834 // base and a scale.
1835 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1836
1837 // Canonicalize a scale of 1 to a base register if the formula doesn't
1838 // already have a base register.
1839 if (!HasBaseReg && Scale == 1) {
1840 Scale = 0;
1841 HasBaseReg = true;
1842 }
1843
1844 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1845 HasBaseReg, Scale);
1846}
1847
1848static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1849 ScalarEvolution &SE, int64_t MinOffset,
1850 int64_t MaxOffset, LSRUse::KindType Kind,
1851 MemAccessTy AccessTy, const SCEV *S,
1852 bool HasBaseReg) {
1853 // Fast-path: zero is always foldable.
1854 if (S->isZero()) return true;
1855
1856 // Conservatively, create an address with an immediate and a
1857 // base and a scale.
1858 int64_t BaseOffset = ExtractImmediate(S, SE);
1859 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1860
1861 // If there's anything else involved, it's not foldable.
1862 if (!S->isZero()) return false;
1863
1864 // Fast-path: zero is always foldable.
1865 if (BaseOffset == 0 && !BaseGV) return true;
1866
1867 // Conservatively, create an address with an immediate and a
1868 // base and a scale.
1869 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1870
1871 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1872 BaseOffset, HasBaseReg, Scale);
1873}
1874
1875namespace {
1876
1877/// An individual increment in a Chain of IV increments. Relate an IV user to
1878/// an expression that computes the IV it uses from the IV used by the previous
1879/// link in the Chain.
1880///
1881/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1882/// original IVOperand. The head of the chain's IVOperand is only valid during
1883/// chain collection, before LSR replaces IV users. During chain generation,
1884/// IncExpr can be used to find the new IVOperand that computes the same
1885/// expression.
1886struct IVInc {
1887 Instruction *UserInst;
1888 Value* IVOperand;
1889 const SCEV *IncExpr;
1890
1891 IVInc(Instruction *U, Value *O, const SCEV *E)
1892 : UserInst(U), IVOperand(O), IncExpr(E) {}
1893};
1894
1895// The list of IV increments in program order. We typically add the head of a
1896// chain without finding subsequent links.
1897struct IVChain {
1898 SmallVector<IVInc, 1> Incs;
1899 const SCEV *ExprBase = nullptr;
1900
1901 IVChain() = default;
1902 IVChain(const IVInc &Head, const SCEV *Base)
1903 : Incs(1, Head), ExprBase(Base) {}
1904
1905 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1906
1907 // Return the first increment in the chain.
1908 const_iterator begin() const {
1909 assert(!Incs.empty())(static_cast <bool> (!Incs.empty()) ? void (0) : __assert_fail
("!Incs.empty()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1909, __extension__ __PRETTY_FUNCTION__))
;
1910 return std::next(Incs.begin());
1911 }
1912 const_iterator end() const {
1913 return Incs.end();
1914 }
1915
1916 // Returns true if this chain contains any increments.
1917 bool hasIncs() const { return Incs.size() >= 2; }
1918
1919 // Add an IVInc to the end of this chain.
1920 void add(const IVInc &X) { Incs.push_back(X); }
1921
1922 // Returns the last UserInst in the chain.
1923 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1924
1925 // Returns true if IncExpr can be profitably added to this chain.
1926 bool isProfitableIncrement(const SCEV *OperExpr,
1927 const SCEV *IncExpr,
1928 ScalarEvolution&);
1929};
1930
1931/// Helper for CollectChains to track multiple IV increment uses. Distinguish
1932/// between FarUsers that definitely cross IV increments and NearUsers that may
1933/// be used between IV increments.
1934struct ChainUsers {
1935 SmallPtrSet<Instruction*, 4> FarUsers;
1936 SmallPtrSet<Instruction*, 4> NearUsers;
1937};
1938
1939/// This class holds state for the main loop strength reduction logic.
1940class LSRInstance {
1941 IVUsers &IU;
1942 ScalarEvolution &SE;
1943 DominatorTree &DT;
1944 LoopInfo &LI;
1945 AssumptionCache &AC;
1946 TargetLibraryInfo &TLI;
1947 const TargetTransformInfo &TTI;
1948 Loop *const L;
1949 MemorySSAUpdater *MSSAU;
1950 TTI::AddressingModeKind AMK;
1951 bool Changed = false;
1952
1953 /// This is the insert position that the current loop's induction variable
1954 /// increment should be placed. In simple loops, this is the latch block's
1955 /// terminator. But in more complicated cases, this is a position which will
1956 /// dominate all the in-loop post-increment users.
1957 Instruction *IVIncInsertPos = nullptr;
1958
1959 /// Interesting factors between use strides.
1960 ///
1961 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1962 /// default, a SmallDenseSet, because we need to use the full range of
1963 /// int64_ts, and there's currently no good way of doing that with
1964 /// SmallDenseSet.
1965 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1966
1967 /// Interesting use types, to facilitate truncation reuse.
1968 SmallSetVector<Type *, 4> Types;
1969
1970 /// The list of interesting uses.
1971 mutable SmallVector<LSRUse, 16> Uses;
1972
1973 /// Track which uses use which register candidates.
1974 RegUseTracker RegUses;
1975
1976 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1977 // have more than a few IV increment chains in a loop. Missing a Chain falls
1978 // back to normal LSR behavior for those uses.
1979 static const unsigned MaxChains = 8;
1980
1981 /// IV users can form a chain of IV increments.
1982 SmallVector<IVChain, MaxChains> IVChainVec;
1983
1984 /// IV users that belong to profitable IVChains.
1985 SmallPtrSet<Use*, MaxChains> IVIncSet;
1986
1987 /// Induction variables that were generated and inserted by the SCEV Expander.
1988 SmallVector<llvm::WeakVH, 2> ScalarEvolutionIVs;
1989
1990 void OptimizeShadowIV();
1991 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1992 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1993 void OptimizeLoopTermCond();
1994
1995 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1996 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1997 void FinalizeChain(IVChain &Chain);
1998 void CollectChains();
1999 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2000 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
2001
2002 void CollectInterestingTypesAndFactors();
2003 void CollectFixupsAndInitialFormulae();
2004
2005 // Support for sharing of LSRUses between LSRFixups.
2006 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
2007 UseMapTy UseMap;
2008
2009 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2010 LSRUse::KindType Kind, MemAccessTy AccessTy);
2011
2012 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
2013 MemAccessTy AccessTy);
2014
2015 void DeleteUse(LSRUse &LU, size_t LUIdx);
2016
2017 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
2018
2019 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2020 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2021 void CountRegisters(const Formula &F, size_t LUIdx);
2022 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
2023
2024 void CollectLoopInvariantFixupsAndFormulae();
2025
2026 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
2027 unsigned Depth = 0);
2028
2029 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
2030 const Formula &Base, unsigned Depth,
2031 size_t Idx, bool IsScaledReg = false);
2032 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
2033 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2034 const Formula &Base, size_t Idx,
2035 bool IsScaledReg = false);
2036 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2037 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2038 const Formula &Base,
2039 const SmallVectorImpl<int64_t> &Worklist,
2040 size_t Idx, bool IsScaledReg = false);
2041 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2042 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2043 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2044 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2045 void GenerateCrossUseConstantOffsets();
2046 void GenerateAllReuseFormulae();
2047
2048 void FilterOutUndesirableDedicatedRegisters();
2049
2050 size_t EstimateSearchSpaceComplexity() const;
2051 void NarrowSearchSpaceByDetectingSupersets();
2052 void NarrowSearchSpaceByCollapsingUnrolledCode();
2053 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2054 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2055 void NarrowSearchSpaceByFilterPostInc();
2056 void NarrowSearchSpaceByDeletingCostlyFormulas();
2057 void NarrowSearchSpaceByPickingWinnerRegs();
2058 void NarrowSearchSpaceUsingHeuristics();
2059
2060 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2061 Cost &SolutionCost,
2062 SmallVectorImpl<const Formula *> &Workspace,
2063 const Cost &CurCost,
2064 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2065 DenseSet<const SCEV *> &VisitedRegs) const;
2066 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2067
2068 BasicBlock::iterator
2069 HoistInsertPosition(BasicBlock::iterator IP,
2070 const SmallVectorImpl<Instruction *> &Inputs) const;
2071 BasicBlock::iterator
2072 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2073 const LSRFixup &LF,
2074 const LSRUse &LU,
2075 SCEVExpander &Rewriter) const;
2076
2077 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2078 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2079 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2080 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2081 const Formula &F, SCEVExpander &Rewriter,
2082 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2083 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2084 SCEVExpander &Rewriter,
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::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2417 return LatchBlock != BB;
2418 })) {
2419 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2420 IVIncInsertPos = LatchBlock->getTerminator();
2421 return;
2422 }
2423
2424 // Otherwise treat this as a rotated loop.
2425 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2426 // Get the terminating condition for the loop if possible. If we
2427 // can, we want to change it to use a post-incremented version of its
2428 // induction variable, to allow coalescing the live ranges for the IV into
2429 // one register value.
2430
2431 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2432 if (!TermBr)
2433 continue;
2434 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2435 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2436 continue;
2437
2438 // Search IVUsesByStride to find Cond's IVUse if there is one.
2439 IVStrideUse *CondUse = nullptr;
2440 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2441 if (!FindIVUserForCond(Cond, CondUse))
2442 continue;
2443
2444 // If the trip count is computed in terms of a max (due to ScalarEvolution
2445 // being unable to find a sufficient guard, for example), change the loop
2446 // comparison to use SLT or ULT instead of NE.
2447 // One consequence of doing this now is that it disrupts the count-down
2448 // optimization. That's not always a bad thing though, because in such
2449 // cases it may still be worthwhile to avoid a max.
2450 Cond = OptimizeMax(Cond, CondUse);
2451
2452 // If this exiting block dominates the latch block, it may also use
2453 // the post-inc value if it won't be shared with other uses.
2454 // Check for dominance.
2455 if (!DT.dominates(ExitingBlock, LatchBlock))
2456 continue;
2457
2458 // Conservatively avoid trying to use the post-inc value in non-latch
2459 // exits if there may be pre-inc users in intervening blocks.
2460 if (LatchBlock != ExitingBlock)
2461 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2462 // Test if the use is reachable from the exiting block. This dominator
2463 // query is a conservative approximation of reachability.
2464 if (&*UI != CondUse &&
2465 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2466 // Conservatively assume there may be reuse if the quotient of their
2467 // strides could be a legal scale.
2468 const SCEV *A = IU.getStride(*CondUse, L);
2469 const SCEV *B = IU.getStride(*UI, L);
2470 if (!A || !B) continue;
2471 if (SE.getTypeSizeInBits(A->getType()) !=
2472 SE.getTypeSizeInBits(B->getType())) {
2473 if (SE.getTypeSizeInBits(A->getType()) >
2474 SE.getTypeSizeInBits(B->getType()))
2475 B = SE.getSignExtendExpr(B, A->getType());
2476 else
2477 A = SE.getSignExtendExpr(A, B->getType());
2478 }
2479 if (const SCEVConstant *D =
2480 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2481 const ConstantInt *C = D->getValue();
2482 // Stride of one or negative one can have reuse with non-addresses.
2483 if (C->isOne() || C->isMinusOne())
2484 goto decline_post_inc;
2485 // Avoid weird situations.
2486 if (C->getValue().getMinSignedBits() >= 64 ||
2487 C->getValue().isMinSignedValue())
2488 goto decline_post_inc;
2489 // Check for possible scaled-address reuse.
2490 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2491 MemAccessTy AccessTy = getAccessType(
2492 TTI, UI->getUser(), UI->getOperandValToReplace());
2493 int64_t Scale = C->getSExtValue();
2494 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2495 /*BaseOffset=*/0,
2496 /*HasBaseReg=*/false, Scale,
2497 AccessTy.AddrSpace))
2498 goto decline_post_inc;
2499 Scale = -Scale;
2500 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2501 /*BaseOffset=*/0,
2502 /*HasBaseReg=*/false, Scale,
2503 AccessTy.AddrSpace))
2504 goto decline_post_inc;
2505 }
2506 }
2507 }
2508
2509 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)
2510 << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Change loop exiting icmp to use postinc iv: "
<< *Cond << '\n'; } } while (false)
;
2511
2512 // It's possible for the setcc instruction to be anywhere in the loop, and
2513 // possible for it to have multiple users. If it is not immediately before
2514 // the exiting block branch, move it.
2515 if (Cond->getNextNonDebugInstruction() != TermBr) {
2516 if (Cond->hasOneUse()) {
2517 Cond->moveBefore(TermBr);
2518 } else {
2519 // Clone the terminating condition and insert into the loopend.
2520 ICmpInst *OldCond = Cond;
2521 Cond = cast<ICmpInst>(Cond->clone());
2522 Cond->setName(L->getHeader()->getName() + ".termcond");
2523 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2524
2525 // Clone the IVUse, as the old use still exists!
2526 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2527 TermBr->replaceUsesOfWith(OldCond, Cond);
2528 }
2529 }
2530
2531 // If we get to here, we know that we can transform the setcc instruction to
2532 // use the post-incremented version of the IV, allowing us to coalesce the
2533 // live ranges for the IV correctly.
2534 CondUse->transformToPostInc(L);
2535 Changed = true;
2536
2537 PostIncs.insert(Cond);
2538 decline_post_inc:;
2539 }
2540
2541 // Determine an insertion point for the loop induction variable increment. It
2542 // must dominate all the post-inc comparisons we just set up, and it must
2543 // dominate the loop latch edge.
2544 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2545 for (Instruction *Inst : PostIncs) {
2546 BasicBlock *BB =
2547 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2548 Inst->getParent());
2549 if (BB == Inst->getParent())
2550 IVIncInsertPos = Inst;
2551 else if (BB != IVIncInsertPos->getParent())
2552 IVIncInsertPos = BB->getTerminator();
2553 }
2554}
2555
2556/// Determine if the given use can accommodate a fixup at the given offset and
2557/// other details. If so, update the use and return true.
2558bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2559 bool HasBaseReg, LSRUse::KindType Kind,
2560 MemAccessTy AccessTy) {
2561 int64_t NewMinOffset = LU.MinOffset;
2562 int64_t NewMaxOffset = LU.MaxOffset;
2563 MemAccessTy NewAccessTy = AccessTy;
2564
2565 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2566 // something conservative, however this can pessimize in the case that one of
2567 // the uses will have all its uses outside the loop, for example.
2568 if (LU.Kind != Kind)
2569 return false;
2570
2571 // Check for a mismatched access type, and fall back conservatively as needed.
2572 // TODO: Be less conservative when the type is similar and can use the same
2573 // addressing modes.
2574 if (Kind == LSRUse::Address) {
2575 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2576 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2577 AccessTy.AddrSpace);
2578 }
2579 }
2580
2581 // Conservatively assume HasBaseReg is true for now.
2582 if (NewOffset < LU.MinOffset) {
2583 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2584 LU.MaxOffset - NewOffset, HasBaseReg))
2585 return false;
2586 NewMinOffset = NewOffset;
2587 } else if (NewOffset > LU.MaxOffset) {
2588 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2589 NewOffset - LU.MinOffset, HasBaseReg))
2590 return false;
2591 NewMaxOffset = NewOffset;
2592 }
2593
2594 // Update the use.
2595 LU.MinOffset = NewMinOffset;
2596 LU.MaxOffset = NewMaxOffset;
2597 LU.AccessTy = NewAccessTy;
2598 return true;
2599}
2600
2601/// Return an LSRUse index and an offset value for a fixup which needs the given
2602/// expression, with the given kind and optional access type. Either reuse an
2603/// existing use or create a new one, as needed.
2604std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2605 LSRUse::KindType Kind,
2606 MemAccessTy AccessTy) {
2607 const SCEV *Copy = Expr;
2608 int64_t Offset = ExtractImmediate(Expr, SE);
2609
2610 // Basic uses can't accept any offset, for example.
2611 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2612 Offset, /*HasBaseReg=*/ true)) {
2613 Expr = Copy;
2614 Offset = 0;
2615 }
2616
2617 std::pair<UseMapTy::iterator, bool> P =
2618 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2619 if (!P.second) {
2620 // A use already existed with this base.
2621 size_t LUIdx = P.first->second;
2622 LSRUse &LU = Uses[LUIdx];
2623 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2624 // Reuse this use.
2625 return std::make_pair(LUIdx, Offset);
2626 }
2627
2628 // Create a new use.
2629 size_t LUIdx = Uses.size();
2630 P.first->second = LUIdx;
2631 Uses.push_back(LSRUse(Kind, AccessTy));
2632 LSRUse &LU = Uses[LUIdx];
2633
2634 LU.MinOffset = Offset;
2635 LU.MaxOffset = Offset;
2636 return std::make_pair(LUIdx, Offset);
2637}
2638
2639/// Delete the given use from the Uses list.
2640void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2641 if (&LU != &Uses.back())
2642 std::swap(LU, Uses.back());
2643 Uses.pop_back();
2644
2645 // Update RegUses.
2646 RegUses.swapAndDropUse(LUIdx, Uses.size());
2647}
2648
2649/// Look for a use distinct from OrigLU which is has a formula that has the same
2650/// registers as the given formula.
2651LSRUse *
2652LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2653 const LSRUse &OrigLU) {
2654 // Search all uses for the formula. This could be more clever.
2655 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2656 LSRUse &LU = Uses[LUIdx];
2657 // Check whether this use is close enough to OrigLU, to see whether it's
2658 // worthwhile looking through its formulae.
2659 // Ignore ICmpZero uses because they may contain formulae generated by
2660 // GenerateICmpZeroScales, in which case adding fixup offsets may
2661 // be invalid.
2662 if (&LU != &OrigLU &&
2663 LU.Kind != LSRUse::ICmpZero &&
2664 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2665 LU.WidestFixupType == OrigLU.WidestFixupType &&
2666 LU.HasFormulaWithSameRegs(OrigF)) {
2667 // Scan through this use's formulae.
2668 for (const Formula &F : LU.Formulae) {
2669 // Check to see if this formula has the same registers and symbols
2670 // as OrigF.
2671 if (F.BaseRegs == OrigF.BaseRegs &&
2672 F.ScaledReg == OrigF.ScaledReg &&
2673 F.BaseGV == OrigF.BaseGV &&
2674 F.Scale == OrigF.Scale &&
2675 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2676 if (F.BaseOffset == 0)
2677 return &LU;
2678 // This is the formula where all the registers and symbols matched;
2679 // there aren't going to be any others. Since we declined it, we
2680 // can skip the rest of the formulae and proceed to the next LSRUse.
2681 break;
2682 }
2683 }
2684 }
2685 }
2686
2687 // Nothing looked good.
2688 return nullptr;
2689}
2690
2691void LSRInstance::CollectInterestingTypesAndFactors() {
2692 SmallSetVector<const SCEV *, 4> Strides;
2693
2694 // Collect interesting types and strides.
2695 SmallVector<const SCEV *, 4> Worklist;
2696 for (const IVStrideUse &U : IU) {
2697 const SCEV *Expr = IU.getExpr(U);
2698
2699 // Collect interesting types.
2700 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2701
2702 // Add strides for mentioned loops.
2703 Worklist.push_back(Expr);
2704 do {
2705 const SCEV *S = Worklist.pop_back_val();
2706 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2707 if (AR->getLoop() == L)
2708 Strides.insert(AR->getStepRecurrence(SE));
2709 Worklist.push_back(AR->getStart());
2710 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2711 Worklist.append(Add->op_begin(), Add->op_end());
2712 }
2713 } while (!Worklist.empty());
2714 }
2715
2716 // Compute interesting factors from the set of interesting strides.
2717 for (SmallSetVector<const SCEV *, 4>::const_iterator
2718 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2719 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2720 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2721 const SCEV *OldStride = *I;
2722 const SCEV *NewStride = *NewStrideIter;
2723
2724 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2725 SE.getTypeSizeInBits(NewStride->getType())) {
2726 if (SE.getTypeSizeInBits(OldStride->getType()) >
2727 SE.getTypeSizeInBits(NewStride->getType()))
2728 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2729 else
2730 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2731 }
2732 if (const SCEVConstant *Factor =
2733 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2734 SE, true))) {
2735 if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2736 Factors.insert(Factor->getAPInt().getSExtValue());
2737 } else if (const SCEVConstant *Factor =
2738 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2739 NewStride,
2740 SE, true))) {
2741 if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2742 Factors.insert(Factor->getAPInt().getSExtValue());
2743 }
2744 }
2745
2746 // If all uses use the same type, don't bother looking for truncation-based
2747 // reuse.
2748 if (Types.size() == 1)
2749 Types.clear();
2750
2751 LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
(false)
;
2752}
2753
2754/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2755/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2756/// IVStrideUses, we could partially skip this.
2757static User::op_iterator
2758findIVOperand(User::op_iterator OI, User::op_iterator OE,
2759 Loop *L, ScalarEvolution &SE) {
2760 for(; OI != OE; ++OI) {
2761 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2762 if (!SE.isSCEVable(Oper->getType()))
2763 continue;
2764
2765 if (const SCEVAddRecExpr *AR =
2766 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2767 if (AR->getLoop() == L)
2768 break;
2769 }
2770 }
2771 }
2772 return OI;
2773}
2774
2775/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2776/// a convenient helper.
2777static Value *getWideOperand(Value *Oper) {
2778 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2779 return Trunc->getOperand(0);
2780 return Oper;
2781}
2782
2783/// Return true if we allow an IV chain to include both types.
2784static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2785 Type *LType = LVal->getType();
2786 Type *RType = RVal->getType();
2787 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2788 // Different address spaces means (possibly)
2789 // different types of the pointer implementation,
2790 // e.g. i16 vs i32 so disallow that.
2791 (LType->getPointerAddressSpace() ==
2792 RType->getPointerAddressSpace()));
2793}
2794
2795/// Return an approximation of this SCEV expression's "base", or NULL for any
2796/// constant. Returning the expression itself is conservative. Returning a
2797/// deeper subexpression is more precise and valid as long as it isn't less
2798/// complex than another subexpression. For expressions involving multiple
2799/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2800/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2801/// IVInc==b-a.
2802///
2803/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2804/// SCEVUnknown, we simply return the rightmost SCEV operand.
2805static const SCEV *getExprBase(const SCEV *S) {
2806 switch (S->getSCEVType()) {
2807 default: // uncluding scUnknown.
2808 return S;
2809 case scConstant:
2810 return nullptr;
2811 case scTruncate:
2812 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2813 case scZeroExtend:
2814 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2815 case scSignExtend:
2816 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2817 case scAddExpr: {
2818 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2819 // there's nothing more complex.
2820 // FIXME: not sure if we want to recognize negation.
2821 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2822 for (const SCEV *SubExpr : reverse(Add->operands())) {
2823 if (SubExpr->getSCEVType() == scAddExpr)
2824 return getExprBase(SubExpr);
2825
2826 if (SubExpr->getSCEVType() != scMulExpr)
2827 return SubExpr;
2828 }
2829 return S; // all operands are scaled, be conservative.
2830 }
2831 case scAddRecExpr:
2832 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2833 }
2834 llvm_unreachable("Unknown SCEV kind!")::llvm::llvm_unreachable_internal("Unknown SCEV kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2834)
;
2835}
2836
2837/// Return true if the chain increment is profitable to expand into a loop
2838/// invariant value, which may require its own register. A profitable chain
2839/// increment will be an offset relative to the same base. We allow such offsets
2840/// to potentially be used as chain increment as long as it's not obviously
2841/// expensive to expand using real instructions.
2842bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2843 const SCEV *IncExpr,
2844 ScalarEvolution &SE) {
2845 // Aggressively form chains when -stress-ivchain.
2846 if (StressIVChain)
2847 return true;
2848
2849 // Do not replace a constant offset from IV head with a nonconstant IV
2850 // increment.
2851 if (!isa<SCEVConstant>(IncExpr)) {
2852 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2853 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2854 return false;
2855 }
2856
2857 SmallPtrSet<const SCEV*, 8> Processed;
2858 return !isHighCostExpansion(IncExpr, Processed, SE);
2859}
2860
2861/// Return true if the number of registers needed for the chain is estimated to
2862/// be less than the number required for the individual IV users. First prohibit
2863/// any IV users that keep the IV live across increments (the Users set should
2864/// be empty). Next count the number and type of increments in the chain.
2865///
2866/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2867/// effectively use postinc addressing modes. Only consider it profitable it the
2868/// increments can be computed in fewer registers when chained.
2869///
2870/// TODO: Consider IVInc free if it's already used in another chains.
2871static bool isProfitableChain(IVChain &Chain,
2872 SmallPtrSetImpl<Instruction *> &Users,
2873 ScalarEvolution &SE,
2874 const TargetTransformInfo &TTI) {
2875 if (StressIVChain)
2876 return true;
2877
2878 if (!Chain.hasIncs())
2879 return false;
2880
2881 if (!Users.empty()) {
2882 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)
2883 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)
2884 : 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)
;
2885 return false;
2886 }
2887 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", 2887, __extension__
__PRETTY_FUNCTION__))
;
2888
2889 // The chain itself may require a register, so intialize cost to 1.
2890 int cost = 1;
2891
2892 // A complete chain likely eliminates the need for keeping the original IV in
2893 // a register. LSR does not currently know how to form a complete chain unless
2894 // the header phi already exists.
2895 if (isa<PHINode>(Chain.tailUserInst())
2896 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2897 --cost;
2898 }
2899 const SCEV *LastIncExpr = nullptr;
2900 unsigned NumConstIncrements = 0;
2901 unsigned NumVarIncrements = 0;
2902 unsigned NumReusedIncrements = 0;
2903
2904 if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
2905 return true;
2906
2907 for (const IVInc &Inc : Chain) {
2908 if (TTI.isProfitableLSRChainElement(Inc.UserInst))
2909 return true;
2910 if (Inc.IncExpr->isZero())
2911 continue;
2912
2913 // Incrementing by zero or some constant is neutral. We assume constants can
2914 // be folded into an addressing mode or an add's immediate operand.
2915 if (isa<SCEVConstant>(Inc.IncExpr)) {
2916 ++NumConstIncrements;
2917 continue;
2918 }
2919
2920 if (Inc.IncExpr == LastIncExpr)
2921 ++NumReusedIncrements;
2922 else
2923 ++NumVarIncrements;
2924
2925 LastIncExpr = Inc.IncExpr;
2926 }
2927 // An IV chain with a single increment is handled by LSR's postinc
2928 // uses. However, a chain with multiple increments requires keeping the IV's
2929 // value live longer than it needs to be if chained.
2930 if (NumConstIncrements > 1)
2931 --cost;
2932
2933 // Materializing increment expressions in the preheader that didn't exist in
2934 // the original code may cost a register. For example, sign-extended array
2935 // indices can produce ridiculous increments like this:
2936 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2937 cost += NumVarIncrements;
2938
2939 // Reusing variable increments likely saves a register to hold the multiple of
2940 // the stride.
2941 cost -= NumReusedIncrements;
2942
2943 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)
2944 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
;
2945
2946 return cost < 0;
2947}
2948
2949/// Add this IV user to an existing chain or make it the head of a new chain.
2950void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2951 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2952 // When IVs are used as types of varying widths, they are generally converted
2953 // to a wider type with some uses remaining narrow under a (free) trunc.
2954 Value *const NextIV = getWideOperand(IVOper);
2955 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2956 const SCEV *const OperExprBase = getExprBase(OperExpr);
2957
2958 // Visit all existing chains. Check if its IVOper can be computed as a
2959 // profitable loop invariant increment from the last link in the Chain.
2960 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2961 const SCEV *LastIncExpr = nullptr;
2962 for (; ChainIdx < NChains; ++ChainIdx) {
2963 IVChain &Chain = IVChainVec[ChainIdx];
2964
2965 // Prune the solution space aggressively by checking that both IV operands
2966 // are expressions that operate on the same unscaled SCEVUnknown. This
2967 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2968 // first avoids creating extra SCEV expressions.
2969 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2970 continue;
2971
2972 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2973 if (!isCompatibleIVType(PrevIV, NextIV))
2974 continue;
2975
2976 // A phi node terminates a chain.
2977 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2978 continue;
2979
2980 // The increment must be loop-invariant so it can be kept in a register.
2981 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2982 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2983 if (isa<SCEVCouldNotCompute>(IncExpr) || !SE.isLoopInvariant(IncExpr, L))
2984 continue;
2985
2986 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2987 LastIncExpr = IncExpr;
2988 break;
2989 }
2990 }
2991 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2992 // bother for phi nodes, because they must be last in the chain.
2993 if (ChainIdx == NChains) {
2994 if (isa<PHINode>(UserInst))
2995 return;
2996 if (NChains >= MaxChains && !StressIVChain) {
2997 LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
(false)
;
2998 return;
2999 }
3000 LastIncExpr = OperExpr;
3001 // IVUsers may have skipped over sign/zero extensions. We don't currently
3002 // attempt to form chains involving extensions unless they can be hoisted
3003 // into this loop's AddRec.
3004 if (!isa<SCEVAddRecExpr>(LastIncExpr))
3005 return;
3006 ++NChains;
3007 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
3008 OperExprBase));
3009 ChainUsersVec.resize(NChains);
3010 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)
3011 << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Head: (" << *UserInst << ") IV=" <<
*LastIncExpr << "\n"; } } while (false)
;
3012 } else {
3013 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)
3014 << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
<< " Inc: (" << *UserInst << ") IV+" <<
*LastIncExpr << "\n"; } } while (false)
;
3015 // Add this IV user to the end of the chain.
3016 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
3017 }
3018 IVChain &Chain = IVChainVec[ChainIdx];
3019
3020 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
3021 // This chain's NearUsers become FarUsers.
3022 if (!LastIncExpr->isZero()) {
3023 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
3024 NearUsers.end());
3025 NearUsers.clear();
3026 }
3027
3028 // All other uses of IVOperand become near uses of the chain.
3029 // We currently ignore intermediate values within SCEV expressions, assuming
3030 // they will eventually be used be the current chain, or can be computed
3031 // from one of the chain increments. To be more precise we could
3032 // transitively follow its user and only add leaf IV users to the set.
3033 for (User *U : IVOper->users()) {
3034 Instruction *OtherUse = dyn_cast<Instruction>(U);
3035 if (!OtherUse)
3036 continue;
3037 // Uses in the chain will no longer be uses if the chain is formed.
3038 // Include the head of the chain in this iteration (not Chain.begin()).
3039 IVChain::const_iterator IncIter = Chain.Incs.begin();
3040 IVChain::const_iterator IncEnd = Chain.Incs.end();
3041 for( ; IncIter != IncEnd; ++IncIter) {
3042 if (IncIter->UserInst == OtherUse)
3043 break;
3044 }
3045 if (IncIter != IncEnd)
3046 continue;
3047
3048 if (SE.isSCEVable(OtherUse->getType())
3049 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3050 && IU.isIVUserOrOperand(OtherUse)) {
3051 continue;
3052 }
3053 NearUsers.insert(OtherUse);
3054 }
3055
3056 // Since this user is part of the chain, it's no longer considered a use
3057 // of the chain.
3058 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3059}
3060
3061/// Populate the vector of Chains.
3062///
3063/// This decreases ILP at the architecture level. Targets with ample registers,
3064/// multiple memory ports, and no register renaming probably don't want
3065/// this. However, such targets should probably disable LSR altogether.
3066///
3067/// The job of LSR is to make a reasonable choice of induction variables across
3068/// the loop. Subsequent passes can easily "unchain" computation exposing more
3069/// ILP *within the loop* if the target wants it.
3070///
3071/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3072/// will not reorder memory operations, it will recognize this as a chain, but
3073/// will generate redundant IV increments. Ideally this would be corrected later
3074/// by a smart scheduler:
3075/// = A[i]
3076/// = A[i+x]
3077/// A[i] =
3078/// A[i+x] =
3079///
3080/// TODO: Walk the entire domtree within this loop, not just the path to the
3081/// loop latch. This will discover chains on side paths, but requires
3082/// maintaining multiple copies of the Chains state.
3083void LSRInstance::CollectChains() {
3084 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
} } while (false)
;
3085 SmallVector<ChainUsers, 8> ChainUsersVec;
3086
3087 SmallVector<BasicBlock *,8> LatchPath;
3088 BasicBlock *LoopHeader = L->getHeader();
3089 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3090 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3091 LatchPath.push_back(Rung->getBlock());
3092 }
3093 LatchPath.push_back(LoopHeader);
3094
3095 // Walk the instruction stream from the loop header to the loop latch.
3096 for (BasicBlock *BB : reverse(LatchPath)) {
3097 for (Instruction &I : *BB) {
3098 // Skip instructions that weren't seen by IVUsers analysis.
3099 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3100 continue;
3101
3102 // Ignore users that are part of a SCEV expression. This way we only
3103 // consider leaf IV Users. This effectively rediscovers a portion of
3104 // IVUsers analysis but in program order this time.
3105 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3106 continue;
3107
3108 // Remove this instruction from any NearUsers set it may be in.
3109 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3110 ChainIdx < NChains; ++ChainIdx) {
3111 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3112 }
3113 // Search for operands that can be chained.
3114 SmallPtrSet<Instruction*, 4> UniqueOperands;
3115 User::op_iterator IVOpEnd = I.op_end();
3116 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3117 while (IVOpIter != IVOpEnd) {
3118 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3119 if (UniqueOperands.insert(IVOpInst).second)
3120 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3121 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3122 }
3123 } // Continue walking down the instructions.
3124 } // Continue walking down the domtree.
3125 // Visit phi backedges to determine if the chain can generate the IV postinc.
3126 for (PHINode &PN : L->getHeader()->phis()) {
3127 if (!SE.isSCEVable(PN.getType()))
3128 continue;
3129
3130 Instruction *IncV =
3131 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3132 if (IncV)
3133 ChainInstruction(&PN, IncV, ChainUsersVec);
3134 }
3135 // Remove any unprofitable chains.
3136 unsigned ChainIdx = 0;
3137 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3138 UsersIdx < NChains; ++UsersIdx) {
3139 if (!isProfitableChain(IVChainVec[UsersIdx],
3140 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3141 continue;
3142 // Preserve the chain at UsesIdx.
3143 if (ChainIdx != UsersIdx)
3144 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3145 FinalizeChain(IVChainVec[ChainIdx]);
3146 ++ChainIdx;
3147 }
3148 IVChainVec.resize(ChainIdx);
3149}
3150
3151void LSRInstance::FinalizeChain(IVChain &Chain) {
3152 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", 3152, __extension__
__PRETTY_FUNCTION__))
;
3153 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)
;
3154
3155 for (const IVInc &Inc : Chain) {
3156 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Inc: " << *
Inc.UserInst << "\n"; } } while (false)
;
3157 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3158 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", 3158, __extension__
__PRETTY_FUNCTION__))
;
3159 IVIncSet.insert(UseI);
3160 }
3161}
3162
3163/// Return true if the IVInc can be folded into an addressing mode.
3164static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3165 Value *Operand, const TargetTransformInfo &TTI) {
3166 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3167 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3168 return false;
3169
3170 if (IncConst->getAPInt().getMinSignedBits() > 64)
3171 return false;
3172
3173 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3174 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3175 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3176 IncOffset, /*HasBaseReg=*/false))
3177 return false;
3178
3179 return true;
3180}
3181
3182/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3183/// user's operand from the previous IV user's operand.
3184void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3185 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3186 // Find the new IVOperand for the head of the chain. It may have been replaced
3187 // by LSR.
3188 const IVInc &Head = Chain.Incs[0];
3189 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3190 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3191 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3192 IVOpEnd, L, SE);
3193 Value *IVSrc = nullptr;
3194 while (IVOpIter != IVOpEnd) {
3195 IVSrc = getWideOperand(*IVOpIter);
3196
3197 // If this operand computes the expression that the chain needs, we may use
3198 // it. (Check this after setting IVSrc which is used below.)
3199 //
3200 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3201 // narrow for the chain, so we can no longer use it. We do allow using a
3202 // wider phi, assuming the LSR checked for free truncation. In that case we
3203 // should already have a truncate on this operand such that
3204 // getSCEV(IVSrc) == IncExpr.
3205 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3206 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3207 break;
3208 }
3209 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3210 }
3211 if (IVOpIter == IVOpEnd) {
3212 // Gracefully give up on this chain.
3213 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)
;
3214 return;
3215 }
3216 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", 3216, __extension__
__PRETTY_FUNCTION__))
;
3217
3218 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
*IVSrc << "\n"; } } while (false)
;
3219 Type *IVTy = IVSrc->getType();
3220 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3221 const SCEV *LeftOverExpr = nullptr;
3222 for (const IVInc &Inc : Chain) {
3223 Instruction *InsertPt = Inc.UserInst;
3224 if (isa<PHINode>(InsertPt))
3225 InsertPt = L->getLoopLatch()->getTerminator();
3226
3227 // IVOper will replace the current IV User's operand. IVSrc is the IV
3228 // value currently held in a register.
3229 Value *IVOper = IVSrc;
3230 if (!Inc.IncExpr->isZero()) {
3231 // IncExpr was the result of subtraction of two narrow values, so must
3232 // be signed.
3233 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3234 LeftOverExpr = LeftOverExpr ?
3235 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3236 }
3237 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3238 // Expand the IV increment.
3239 Rewriter.clearPostInc();
3240 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3241 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3242 SE.getUnknown(IncV));
3243 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3244
3245 // If an IV increment can't be folded, use it as the next IV value.
3246 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3247 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", 3247, __extension__
__PRETTY_FUNCTION__))
;
3248 IVSrc = IVOper;
3249 LeftOverExpr = nullptr;
3250 }
3251 }
3252 Type *OperTy = Inc.IVOperand->getType();
3253 if (IVTy != OperTy) {
3254 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", 3255, __extension__
__PRETTY_FUNCTION__))
3255 "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", 3255, __extension__
__PRETTY_FUNCTION__))
;
3256 IRBuilder<> Builder(InsertPt);
3257 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3258 }
3259 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3260 if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3261 DeadInsts.emplace_back(OperandIsInstr);
3262 }
3263 // If LSR created a new, wider phi, we may also replace its postinc. We only
3264 // do this if we also found a wide value for the head of the chain.
3265 if (isa<PHINode>(Chain.tailUserInst())) {
3266 for (PHINode &Phi : L->getHeader()->phis()) {
3267 if (!isCompatibleIVType(&Phi, IVSrc))
3268 continue;
3269 Instruction *PostIncV = dyn_cast<Instruction>(
3270 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3271 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3272 continue;
3273 Value *IVOper = IVSrc;
3274 Type *PostIncTy = PostIncV->getType();
3275 if (IVTy != PostIncTy) {
3276 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", 3276, __extension__
__PRETTY_FUNCTION__))
;
3277 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3278 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3279 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3280 }
3281 Phi.replaceUsesOfWith(PostIncV, IVOper);
3282 DeadInsts.emplace_back(PostIncV);
3283 }
3284 }
3285}
3286
3287void LSRInstance::CollectFixupsAndInitialFormulae() {
3288 BranchInst *ExitBranch = nullptr;
3289 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3290
3291 for (const IVStrideUse &U : IU) {
3292 Instruction *UserInst = U.getUser();
3293 // Skip IV users that are part of profitable IV Chains.
3294 User::op_iterator UseI =
3295 find(UserInst->operands(), U.getOperandValToReplace());
3296 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", 3296, __extension__
__PRETTY_FUNCTION__))
;
3297 if (IVIncSet.count(UseI)) {
3298 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)
;
3299 continue;
3300 }
3301
3302 LSRUse::KindType Kind = LSRUse::Basic;
3303 MemAccessTy AccessTy;
3304 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3305 Kind = LSRUse::Address;
3306 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3307 }
3308
3309 const SCEV *S = IU.getExpr(U);
3310 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3311
3312 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3313 // (N - i == 0), and this allows (N - i) to be the expression that we work
3314 // with rather than just N or i, so we can consider the register
3315 // requirements for both N and i at the same time. Limiting this code to
3316 // equality icmps is not a problem because all interesting loops use
3317 // equality icmps, thanks to IndVarSimplify.
3318 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3319 // If CI can be saved in some target, like replaced inside hardware loop
3320 // in PowerPC, no need to generate initial formulae for it.
3321 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3322 continue;
3323 if (CI->isEquality()) {
3324 // Swap the operands if needed to put the OperandValToReplace on the
3325 // left, for consistency.
3326 Value *NV = CI->getOperand(1);
3327 if (NV == U.getOperandValToReplace()) {
3328 CI->setOperand(1, CI->getOperand(0));
3329 CI->setOperand(0, NV);
3330 NV = CI->getOperand(1);
3331 Changed = true;
3332 }
3333
3334 // x == y --> x - y == 0
3335 const SCEV *N = SE.getSCEV(NV);
3336 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE) &&
3337 (!NV->getType()->isPointerTy() ||
3338 SE.getPointerBase(N) == SE.getPointerBase(S))) {
3339 // S is normalized, so normalize N before folding it into S
3340 // to keep the result normalized.
3341 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3342 Kind = LSRUse::ICmpZero;
3343 S = SE.getMinusSCEV(N, S);
3344 }
3345
3346 // -1 and the negations of all interesting strides (except the negation
3347 // of -1) are now also interesting.
3348 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3349 if (Factors[i] != -1)
3350 Factors.insert(-(uint64_t)Factors[i]);
3351 Factors.insert(-1);
3352 }
3353 }
3354
3355 // Get or create an LSRUse.
3356 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3357 size_t LUIdx = P.first;
3358 int64_t Offset = P.second;
3359 LSRUse &LU = Uses[LUIdx];
3360
3361 // Record the fixup.
3362 LSRFixup &LF = LU.getNewFixup();
3363 LF.UserInst = UserInst;
3364 LF.OperandValToReplace = U.getOperandValToReplace();
3365 LF.PostIncLoops = TmpPostIncLoops;
3366 LF.Offset = Offset;
3367 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3368
3369 if (!LU.WidestFixupType ||
3370 SE.getTypeSizeInBits(LU.WidestFixupType) <
3371 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3372 LU.WidestFixupType = LF.OperandValToReplace->getType();
3373
3374 // If this is the first use of this LSRUse, give it a formula.
3375 if (LU.Formulae.empty()) {
3376 InsertInitialFormula(S, LU, LUIdx);
3377 CountRegisters(LU.Formulae.back(), LUIdx);
3378 }
3379 }
3380
3381 LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false)
;
3382}
3383
3384/// Insert a formula for the given expression into the given use, separating out
3385/// loop-variant portions from loop-invariant and loop-computable portions.
3386void
3387LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3388 // Mark uses whose expressions cannot be expanded.
3389 if (!isSafeToExpand(S, SE, /*CanonicalMode*/ false))
3390 LU.RigidFormula = true;
3391
3392 Formula F;
3393 F.initialMatch(S, L, SE);
3394 bool Inserted = InsertFormula(LU, LUIdx, F);
3395 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", 3395, __extension__
__PRETTY_FUNCTION__))
; (void)Inserted;
3396}
3397
3398/// Insert a simple single-register formula for the given expression into the
3399/// given use.
3400void
3401LSRInstance::InsertSupplementalFormula(const SCEV *S,
3402 LSRUse &LU, size_t LUIdx) {
3403 Formula F;
3404 F.BaseRegs.push_back(S);
3405 F.HasBaseReg = true;
3406 bool Inserted = InsertFormula(LU, LUIdx, F);
3407 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", 3407, __extension__
__PRETTY_FUNCTION__))
; (void)Inserted;
3408}
3409
3410/// Note which registers are used by the given formula, updating RegUses.
3411void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3412 if (F.ScaledReg)
3413 RegUses.countRegister(F.ScaledReg, LUIdx);
3414 for (const SCEV *BaseReg : F.BaseRegs)
3415 RegUses.countRegister(BaseReg, LUIdx);
3416}
3417
3418/// If the given formula has not yet been inserted, add it to the list, and
3419/// return true. Return false otherwise.
3420bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3421 // Do not insert formula that we will not be able to expand.
3422 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", 3423, __extension__
__PRETTY_FUNCTION__))
3423 "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", 3423, __extension__
__PRETTY_FUNCTION__))
;
3424
3425 if (!LU.InsertFormula(F, *L))
3426 return false;
3427
3428 CountRegisters(F, LUIdx);
3429 return true;
3430}
3431
3432/// Check for other uses of loop-invariant values which we're tracking. These
3433/// other uses will pin these values in registers, making them less profitable
3434/// for elimination.
3435/// TODO: This currently misses non-constant addrec step registers.
3436/// TODO: Should this give more weight to users inside the loop?
3437void
3438LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3439 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3440 SmallPtrSet<const SCEV *, 32> Visited;
3441
3442 while (!Worklist.empty()) {
3443 const SCEV *S = Worklist.pop_back_val();
3444
3445 // Don't process the same SCEV twice
3446 if (!Visited.insert(S).second)
3447 continue;
3448
3449 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3450 Worklist.append(N->op_begin(), N->op_end());
3451 else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
3452 Worklist.push_back(C->getOperand());
3453 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3454 Worklist.push_back(D->getLHS());
3455 Worklist.push_back(D->getRHS());
3456 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3457 const Value *V = US->getValue();
3458 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3459 // Look for instructions defined outside the loop.
3460 if (L->contains(Inst)) continue;
3461 } else if (isa<UndefValue>(V))
3462 // Undef doesn't have a live range, so it doesn't matter.
3463 continue;
3464 for (const Use &U : V->uses()) {
3465 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3466 // Ignore non-instructions.
3467 if (!UserInst)
3468 continue;
3469 // Don't bother if the instruction is an EHPad.
3470 if (UserInst->isEHPad())
3471 continue;
3472 // Ignore instructions in other functions (as can happen with
3473 // Constants).
3474 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3475 continue;
3476 // Ignore instructions not dominated by the loop.
3477 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3478 UserInst->getParent() :
3479 cast<PHINode>(UserInst)->getIncomingBlock(
3480 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3481 if (!DT.dominates(L->getHeader(), UseBB))
3482 continue;
3483 // Don't bother if the instruction is in a BB which ends in an EHPad.
3484 if (UseBB->getTerminator()->isEHPad())
3485 continue;
3486
3487 // Ignore cases in which the currently-examined value could come from
3488 // a basic block terminated with an EHPad. This checks all incoming
3489 // blocks of the phi node since it is possible that the same incoming
3490 // value comes from multiple basic blocks, only some of which may end
3491 // in an EHPad. If any of them do, a subsequent rewrite attempt by this
3492 // pass would try to insert instructions into an EHPad, hitting an
3493 // assertion.
3494 if (isa<PHINode>(UserInst)) {
3495 const auto *PhiNode = cast<PHINode>(UserInst);
3496 bool HasIncompatibleEHPTerminatedBlock = false;
3497 llvm::Value *ExpectedValue = U;
3498 for (unsigned int I = 0; I < PhiNode->getNumIncomingValues(); I++) {
3499 if (PhiNode->getIncomingValue(I) == ExpectedValue) {
3500 if (PhiNode->getIncomingBlock(I)->getTerminator()->isEHPad()) {
3501 HasIncompatibleEHPTerminatedBlock = true;
3502 break;
3503 }
3504 }
3505 }
3506 if (HasIncompatibleEHPTerminatedBlock) {
3507 continue;
3508 }
3509 }
3510
3511 // Don't bother rewriting PHIs in catchswitch blocks.
3512 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3513 continue;
3514 // Ignore uses which are part of other SCEV expressions, to avoid
3515 // analyzing them multiple times.
3516 if (SE.isSCEVable(UserInst->getType())) {
3517 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3518 // If the user is a no-op, look through to its uses.
3519 if (!isa<SCEVUnknown>(UserS))
3520 continue;
3521 if (UserS == US) {
3522 Worklist.push_back(
3523 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3524 continue;
3525 }
3526 }
3527 // Ignore icmp instructions which are already being analyzed.
3528 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3529 unsigned OtherIdx = !U.getOperandNo();
3530 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3531 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3532 continue;
3533 }
3534
3535 std::pair<size_t, int64_t> P = getUse(
3536 S, LSRUse::Basic, MemAccessTy());
3537 size_t LUIdx = P.first;
3538 int64_t Offset = P.second;
3539 LSRUse &LU = Uses[LUIdx];
3540 LSRFixup &LF = LU.getNewFixup();
3541 LF.UserInst = const_cast<Instruction *>(UserInst);
3542 LF.OperandValToReplace = U;
3543 LF.Offset = Offset;
3544 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3545 if (!LU.WidestFixupType ||
3546 SE.getTypeSizeInBits(LU.WidestFixupType) <
3547 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3548 LU.WidestFixupType = LF.OperandValToReplace->getType();
3549 InsertSupplementalFormula(US, LU, LUIdx);
3550 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3551 break;
3552 }
3553 }
3554 }
3555}
3556
3557/// Split S into subexpressions which can be pulled out into separate
3558/// registers. If C is non-null, multiply each subexpression by C.
3559///
3560/// Return remainder expression after factoring the subexpressions captured by
3561/// Ops. If Ops is complete, return NULL.
3562static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3563 SmallVectorImpl<const SCEV *> &Ops,
3564 const Loop *L,
3565 ScalarEvolution &SE,
3566 unsigned Depth = 0) {
3567 // Arbitrarily cap recursion to protect compile time.
3568 if (Depth >= 3)
3569 return S;
3570
3571 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3572 // Break out add operands.
3573 for (const SCEV *S : Add->operands()) {
3574 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3575 if (Remainder)
3576 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3577 }
3578 return nullptr;
3579 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3580 // Split a non-zero base out of an addrec.
3581 if (AR->getStart()->isZero() || !AR->isAffine())
3582 return S;
3583
3584 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3585 C, Ops, L, SE, Depth+1);
3586 // Split the non-zero AddRec unless it is part of a nested recurrence that
3587 // does not pertain to this loop.
3588 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3589 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3590 Remainder = nullptr;
3591 }
3592 if (Remainder != AR->getStart()) {
3593 if (!Remainder)
3594 Remainder = SE.getConstant(AR->getType(), 0);
3595 return SE.getAddRecExpr(Remainder,
3596 AR->getStepRecurrence(SE),
3597 AR->getLoop(),
3598 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3599 SCEV::FlagAnyWrap);
3600 }
3601 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3602 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3603 if (Mul->getNumOperands() != 2)
3604 return S;
3605 if (const SCEVConstant *Op0 =
3606 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3607 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3608 const SCEV *Remainder =
3609 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3610 if (Remainder)
3611 Ops.push_back(SE.getMulExpr(C, Remainder));
3612 return nullptr;
3613 }
3614 }
3615 return S;
3616}
3617
3618/// Return true if the SCEV represents a value that may end up as a
3619/// post-increment operation.
3620static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3621 LSRUse &LU, const SCEV *S, const Loop *L,
3622 ScalarEvolution &SE) {
3623 if (LU.Kind != LSRUse::Address ||
3624 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3625 return false;
3626 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3627 if (!AR)
3628 return false;
3629 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3630 if (!isa<SCEVConstant>(LoopStep))
3631 return false;
3632 // Check if a post-indexed load/store can be used.
3633 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3634 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3635 const SCEV *LoopStart = AR->getStart();
3636 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3637 return true;
3638 }
3639 return false;
3640}
3641
3642/// Helper function for LSRInstance::GenerateReassociations.
3643void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3644 const Formula &Base,
3645 unsigned Depth, size_t Idx,
3646 bool IsScaledReg) {
3647 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3648 // Don't generate reassociations for the base register of a value that
3649 // may generate a post-increment operator. The reason is that the
3650 // reassociations cause extra base+register formula to be created,
3651 // and possibly chosen, but the post-increment is more efficient.
3652 if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3653 return;
3654 SmallVector<const SCEV *, 8> AddOps;
3655 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3656 if (Remainder)
3657 AddOps.push_back(Remainder);
3658
3659 if (AddOps.size() == 1)
3660 return;
3661
3662 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3663 JE = AddOps.end();
3664 J != JE; ++J) {
3665 // Loop-variant "unknown" values are uninteresting; we won't be able to
3666 // do anything meaningful with them.
3667 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3668 continue;
3669
3670 // Don't pull a constant into a register if the constant could be folded
3671 // into an immediate field.
3672 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3673 LU.AccessTy, *J, Base.getNumRegs() > 1))
3674 continue;
3675
3676 // Collect all operands except *J.
3677 SmallVector<const SCEV *, 8> InnerAddOps(
3678 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3679 InnerAddOps.append(std::next(J),
3680 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3681
3682 // Don't leave just a constant behind in a register if the constant could
3683 // be folded into an immediate field.
3684 if (InnerAddOps.size() == 1 &&
3685 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3686 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3687 continue;
3688
3689 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3690 if (InnerSum->isZero())
3691 continue;
3692 Formula F = Base;
3693
3694 // Add the remaining pieces of the add back into the new formula.
3695 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3696 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3697 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3698 InnerSumSC->getValue()->getZExtValue())) {
3699 F.UnfoldedOffset =
3700 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3701 if (IsScaledReg)
3702 F.ScaledReg = nullptr;
3703 else
3704 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3705 } else if (IsScaledReg)
3706 F.ScaledReg = InnerSum;
3707 else
3708 F.BaseRegs[Idx] = InnerSum;
3709
3710 // Add J as its own register, or an unfolded immediate.
3711 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3712 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3713 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3714 SC->getValue()->getZExtValue()))
3715 F.UnfoldedOffset =
3716 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3717 else
3718 F.BaseRegs.push_back(*J);
3719 // We may have changed the number of register in base regs, adjust the
3720 // formula accordingly.
3721 F.canonicalize(*L);
3722
3723 if (InsertFormula(LU, LUIdx, F))
3724 // If that formula hadn't been seen before, recurse to find more like
3725 // it.
3726 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3727 // Because just Depth is not enough to bound compile time.
3728 // This means that every time AddOps.size() is greater 16^x we will add
3729 // x to Depth.
3730 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3731 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3732 }
3733}
3734
3735/// Split out subexpressions from adds and the bases of addrecs.
3736void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3737 Formula Base, unsigned Depth) {
3738 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", 3738, __extension__
__PRETTY_FUNCTION__))
;
3739 // Arbitrarily cap recursion to protect compile time.
3740 if (Depth >= 3)
3741 return;
3742
3743 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3744 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3745
3746 if (Base.Scale == 1)
3747 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3748 /* Idx */ -1, /* IsScaledReg */ true);
3749}
3750
3751/// Generate a formula consisting of all of the loop-dominating registers added
3752/// into a single register.
3753void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3754 Formula Base) {
3755 // This method is only interesting on a plurality of registers.
3756 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3757 (Base.UnfoldedOffset != 0) <= 1)
3758 return;
3759
3760 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3761 // processing the formula.
3762 Base.unscale();
3763 SmallVector<const SCEV *, 4> Ops;
3764 Formula NewBase = Base;
3765 NewBase.BaseRegs.clear();
3766 Type *CombinedIntegerType = nullptr;
3767 for (const SCEV *BaseReg : Base.BaseRegs) {
3768 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3769 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3770 if (!CombinedIntegerType)
3771 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3772 Ops.push_back(BaseReg);
3773 }
3774 else
3775 NewBase.BaseRegs.push_back(BaseReg);
3776 }
3777
3778 // If no register is relevant, we're done.
3779 if (Ops.size() == 0)
3780 return;
3781
3782 // Utility function for generating the required variants of the combined
3783 // registers.
3784 auto GenerateFormula = [&](const SCEV *Sum) {
3785 Formula F = NewBase;
3786
3787 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3788 // opportunity to fold something. For now, just ignore such cases
3789 // rather than proceed with zero in a register.
3790 if (Sum->isZero())
3791 return;
3792
3793 F.BaseRegs.push_back(Sum);
3794 F.canonicalize(*L);
3795 (void)InsertFormula(LU, LUIdx, F);
3796 };
3797
3798 // If we collected at least two registers, generate a formula combining them.
3799 if (Ops.size() > 1) {
3800 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3801 GenerateFormula(SE.getAddExpr(OpsCopy));
3802 }
3803
3804 // If we have an unfolded offset, generate a formula combining it with the
3805 // registers collected.
3806 if (NewBase.UnfoldedOffset) {
3807 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", 3807, __extension__
__PRETTY_FUNCTION__))
;
3808 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3809 true));
3810 NewBase.UnfoldedOffset = 0;
3811 GenerateFormula(SE.getAddExpr(Ops));
3812 }
3813}
3814
3815/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3816void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3817 const Formula &Base, size_t Idx,
3818 bool IsScaledReg) {
3819 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3820 GlobalValue *GV = ExtractSymbol(G, SE);
3821 if (G->isZero() || !GV)
3822 return;
3823 Formula F = Base;
3824 F.BaseGV = GV;
3825 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3826 return;
3827 if (IsScaledReg)
3828 F.ScaledReg = G;
3829 else
3830 F.BaseRegs[Idx] = G;
3831 (void)InsertFormula(LU, LUIdx, F);
3832}
3833
3834/// Generate reuse formulae using symbolic offsets.
3835void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3836 Formula Base) {
3837 // We can't add a symbolic offset if the address already contains one.
3838 if (Base.BaseGV) return;
3839
3840 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3841 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3842 if (Base.Scale == 1)
3843 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3844 /* IsScaledReg */ true);
3845}
3846
3847/// Helper function for LSRInstance::GenerateConstantOffsets.
3848void LSRInstance::GenerateConstantOffsetsImpl(
3849 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3850 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3851
3852 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3853 Formula F = Base;
3854 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3855
3856 if (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) {
3857 // Add the offset to the base register.
3858 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3859 // If it cancelled out, drop the base register, otherwise update it.
3860 if (NewG->isZero()) {
3861 if (IsScaledReg) {
3862 F.Scale = 0;
3863 F.ScaledReg = nullptr;
3864 } else
3865 F.deleteBaseReg(F.BaseRegs[Idx]);
3866 F.canonicalize(*L);
3867 } else if (IsScaledReg)
3868 F.ScaledReg = NewG;
3869 else
3870 F.BaseRegs[Idx] = NewG;
3871
3872 (void)InsertFormula(LU, LUIdx, F);
3873 }
3874 };
3875
3876 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3877
3878 // With constant offsets and constant steps, we can generate pre-inc
3879 // accesses by having the offset equal the step. So, for access #0 with a
3880 // step of 8, we generate a G - 8 base which would require the first access
3881 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3882 // for itself and hopefully becomes the base for other accesses. This means
3883 // means that a single pre-indexed access can be generated to become the new
3884 // base pointer for each iteration of the loop, resulting in no extra add/sub
3885 // instructions for pointer updating.
3886 if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) {
3887 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3888 if (auto *StepRec =
3889 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3890 const APInt &StepInt = StepRec->getAPInt();
3891 int64_t Step = StepInt.isNegative() ?
3892 StepInt.getSExtValue() : StepInt.getZExtValue();
3893
3894 for (int64_t Offset : Worklist) {
3895 Offset -= Step;
3896 GenerateOffset(G, Offset);
3897 }
3898 }
3899 }
3900 }
3901 for (int64_t Offset : Worklist)
3902 GenerateOffset(G, Offset);
3903
3904 int64_t Imm = ExtractImmediate(G, SE);
3905 if (G->isZero() || Imm == 0)
3906 return;
3907 Formula F = Base;
3908 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3909 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3910 return;
3911 if (IsScaledReg) {
3912 F.ScaledReg = G;
3913 } else {
3914 F.BaseRegs[Idx] = G;
3915 // We may generate non canonical Formula if G is a recurrent expr reg
3916 // related with current loop while F.ScaledReg is not.
3917 F.canonicalize(*L);
3918 }
3919 (void)InsertFormula(LU, LUIdx, F);
3920}
3921
3922/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3923void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3924 Formula Base) {
3925 // TODO: For now, just add the min and max offset, because it usually isn't
3926 // worthwhile looking at everything inbetween.
3927 SmallVector<int64_t, 2> Worklist;
3928 Worklist.push_back(LU.MinOffset);
3929 if (LU.MaxOffset != LU.MinOffset)
3930 Worklist.push_back(LU.MaxOffset);
3931
3932 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3933 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3934 if (Base.Scale == 1)
3935 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3936 /* IsScaledReg */ true);
3937}
3938
3939/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3940/// == y -> x*c == y*c.
3941void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3942 Formula Base) {
3943 if (LU.Kind != LSRUse::ICmpZero) return;
3944
3945 // Determine the integer type for the base formula.
3946 Type *IntTy = Base.getType();
3947 if (!IntTy) return;
3948 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3949
3950 // Don't do this if there is more than one offset.
3951 if (LU.MinOffset != LU.MaxOffset) return;
3952
3953 // Check if transformation is valid. It is illegal to multiply pointer.
3954 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3955 return;
3956 for (const SCEV *BaseReg : Base.BaseRegs)
3957 if (BaseReg->getType()->isPointerTy())
3958 return;
3959 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", 3959, __extension__
__PRETTY_FUNCTION__))
;
3960
3961 // Check each interesting stride.
3962 for (int64_t Factor : Factors) {
3963 // Check that Factor can be represented by IntTy
3964 if (!ConstantInt::isValueValidForType(IntTy, Factor))
3965 continue;
3966 // Check that the multiplication doesn't overflow.
3967 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3968 continue;
3969 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3970 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", 3970, __extension__
__PRETTY_FUNCTION__))
;
3971 if (NewBaseOffset / Factor != Base.BaseOffset)
3972 continue;
3973 // If the offset will be truncated at this use, check that it is in bounds.
3974 if (!IntTy->isPointerTy() &&
3975 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3976 continue;
3977
3978 // Check that multiplying with the use offset doesn't overflow.
3979 int64_t Offset = LU.MinOffset;
3980 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3981 continue;
3982 Offset = (uint64_t)Offset * Factor;
3983 if (Offset / Factor != LU.MinOffset)
3984 continue;
3985 // If the offset will be truncated at this use, check that it is in bounds.
3986 if (!IntTy->isPointerTy() &&
3987 !ConstantInt::isValueValidForType(IntTy, Offset))
3988 continue;
3989
3990 Formula F = Base;
3991 F.BaseOffset = NewBaseOffset;
3992
3993 // Check that this scale is legal.
3994 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3995 continue;
3996
3997 // Compensate for the use having MinOffset built into it.
3998 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3999
4000 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4001
4002 // Check that multiplying with each base register doesn't overflow.
4003 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
4004 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
4005 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
4006 goto next;
4007 }
4008
4009 // Check that multiplying with the scaled register doesn't overflow.
4010 if (F.ScaledReg) {
4011 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
4012 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
4013 continue;
4014 }
4015
4016 // Check that multiplying with the unfolded offset doesn't overflow.
4017 if (F.UnfoldedOffset != 0) {
4018 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
4019 Factor == -1)
4020 continue;
4021 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
4022 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
4023 continue;
4024 // If the offset will be truncated, check that it is in bounds.
4025 if (!IntTy->isPointerTy() &&
4026 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
4027 continue;
4028 }
4029
4030 // If we make it here and it's legal, add it.
4031 (void)InsertFormula(LU, LUIdx, F);
4032 next:;
4033 }
4034}
4035
4036/// Generate stride factor reuse formulae by making use of scaled-offset address
4037/// modes, for example.
4038void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
4039 // Determine the integer type for the base formula.
4040 Type *IntTy = Base.getType();
4041 if (!IntTy) return;
4042
4043 // If this Formula already has a scaled register, we can't add another one.
4044 // Try to unscale the formula to generate a better scale.
4045 if (Base.Scale != 0 && !Base.unscale())
4046 return;
4047
4048 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", 4048, __extension__
__PRETTY_FUNCTION__))
;
4049
4050 // Check each interesting stride.
4051 for (int64_t Factor : Factors) {
4052 Base.Scale = Factor;
4053 Base.HasBaseReg = Base.BaseRegs.size() > 1;
4054 // Check whether this scale is going to be legal.
4055 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4056 Base)) {
4057 // As a special-case, handle special out-of-loop Basic users specially.
4058 // TODO: Reconsider this special case.
4059 if (LU.Kind == LSRUse::Basic &&
4060 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
4061 LU.AccessTy, Base) &&
4062 LU.AllFixupsOutsideLoop)
4063 LU.Kind = LSRUse::Special;
4064 else
4065 continue;
4066 }
4067 // For an ICmpZero, negating a solitary base register won't lead to
4068 // new solutions.
4069 if (LU.Kind == LSRUse::ICmpZero &&
4070 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
4071 continue;
4072 // For each addrec base reg, if its loop is current loop, apply the scale.
4073 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
4074 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
4075 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
4076 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4077 if (FactorS->isZero())
4078 continue;
4079 // Divide out the factor, ignoring high bits, since we'll be
4080 // scaling the value back up in the end.
4081 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
4082 // TODO: This could be optimized to avoid all the copying.
4083 Formula F = Base;
4084 F.ScaledReg = Quotient;
4085 F.deleteBaseReg(F.BaseRegs[i]);
4086 // The canonical representation of 1*reg is reg, which is already in
4087 // Base. In that case, do not try to insert the formula, it will be
4088 // rejected anyway.
4089 if (F.Scale == 1 && (F.BaseRegs.empty() ||
4090 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4091 continue;
4092 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4093 // non canonical Formula with ScaledReg's loop not being L.
4094 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4095 F.canonicalize(*L);
4096 (void)InsertFormula(LU, LUIdx, F);
4097 }
4098 }
4099 }
4100 }
4101}
4102
4103/// Generate reuse formulae from different IV types.
4104void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4105 // Don't bother truncating symbolic values.
4106 if (Base.BaseGV) return;
4107
4108 // Determine the integer type for the base formula.
4109 Type *DstTy = Base.getType();
4110 if (!DstTy) return;
4111 if (DstTy->isPointerTy())
4112 return;
4113
4114 // It is invalid to extend a pointer type so exit early if ScaledReg or
4115 // any of the BaseRegs are pointers.
4116 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
4117 return;
4118 if (any_of(Base.BaseRegs,
4119 [](const SCEV *S) { return S->getType()->isPointerTy(); }))
4120 return;
4121
4122 for (Type *SrcTy : Types) {
4123 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4124 Formula F = Base;
4125
4126 // Sometimes SCEV is able to prove zero during ext transform. It may
4127 // happen if SCEV did not do all possible transforms while creating the
4128 // initial node (maybe due to depth limitations), but it can do them while
4129 // taking ext.
4130 if (F.ScaledReg) {
4131 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4132 if (NewScaledReg->isZero())
4133 continue;
4134 F.ScaledReg = NewScaledReg;
4135 }
4136 bool HasZeroBaseReg = false;
4137 for (const SCEV *&BaseReg : F.BaseRegs) {
4138 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4139 if (NewBaseReg->isZero()) {
4140 HasZeroBaseReg = true;
4141 break;
4142 }
4143 BaseReg = NewBaseReg;
4144 }
4145 if (HasZeroBaseReg)
4146 continue;
4147
4148 // TODO: This assumes we've done basic processing on all uses and
4149 // have an idea what the register usage is.
4150 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4151 continue;
4152
4153 F.canonicalize(*L);
4154 (void)InsertFormula(LU, LUIdx, F);
4155 }
4156 }
4157}
4158
4159namespace {
4160
4161/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4162/// modifications so that the search phase doesn't have to worry about the data
4163/// structures moving underneath it.
4164struct WorkItem {
4165 size_t LUIdx;
4166 int64_t Imm;
4167 const SCEV *OrigReg;
4168
4169 WorkItem(size_t LI, int64_t I, const SCEV *R)
4170 : LUIdx(LI), Imm(I), OrigReg(R) {}
4171
4172 void print(raw_ostream &OS) const;
4173 void dump() const;
4174};
4175
4176} // end anonymous namespace
4177
4178#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4179void WorkItem::print(raw_ostream &OS) const {
4180 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4181 << " , add offset " << Imm;
4182}
4183
4184LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
4185 print(errs()); errs() << '\n';
4186}
4187#endif
4188
4189/// Look for registers which are a constant distance apart and try to form reuse
4190/// opportunities between them.
4191void LSRInstance::GenerateCrossUseConstantOffsets() {
4192 // Group the registers by their value without any added constant offset.
4193 using ImmMapTy = std::map<int64_t, const SCEV *>;
4194
4195 DenseMap<const SCEV *, ImmMapTy> Map;
4196 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4197 SmallVector<const SCEV *, 8> Sequence;
4198 for (const SCEV *Use : RegUses) {
4199 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4200 int64_t Imm = ExtractImmediate(Reg, SE);
4201 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4202 if (Pair.second)
4203 Sequence.push_back(Reg);
4204 Pair.first->second.insert(std::make_pair(Imm, Use));
4205 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4206 }
4207
4208 // Now examine each set of registers with the same base value. Build up
4209 // a list of work to do and do the work in a separate step so that we're
4210 // not adding formulae and register counts while we're searching.
4211 SmallVector<WorkItem, 32> WorkItems;
4212 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4213 for (const SCEV *Reg : Sequence) {
4214 const ImmMapTy &Imms = Map.find(Reg)->second;
4215
4216 // It's not worthwhile looking for reuse if there's only one offset.
4217 if (Imms.size() == 1)
4218 continue;
4219
4220 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)
4221 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)
4222 : 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)
4223 << ' ' << 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)
4224 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)
;
4225
4226 // Examine each offset.
4227 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4228 J != JE; ++J) {
4229 const SCEV *OrigReg = J->second;
4230
4231 int64_t JImm = J->first;
4232 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4233
4234 if (!isa<SCEVConstant>(OrigReg) &&
4235 UsedByIndicesMap[Reg].count() == 1) {
4236 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)
4237 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
<< *OrigReg << '\n'; } } while (false)
;
4238 continue;
4239 }
4240
4241 // Conservatively examine offsets between this orig reg a few selected
4242 // other orig regs.
4243 int64_t First = Imms.begin()->first;
4244 int64_t Last = std::prev(Imms.end())->first;
4245 // Compute (First + Last) / 2 without overflow using the fact that
4246 // First + Last = 2 * (First + Last) + (First ^ Last).
4247 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4248 // If the result is negative and First is odd and Last even (or vice versa),
4249 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4250 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4251 ImmMapTy::const_iterator OtherImms[] = {
4252 Imms.begin(), std::prev(Imms.end()),
4253 Imms.lower_bound(Avg)};
4254 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4255 ImmMapTy::const_iterator M = OtherImms[i];
4256 if (M == J || M == JE) continue;
4257
4258 // Compute the difference between the two.
4259 int64_t Imm = (uint64_t)JImm - M->first;
4260 for (unsigned LUIdx : UsedByIndices.set_bits())
4261 // Make a memo of this use, offset, and register tuple.
4262 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4263 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4264 }
4265 }
4266 }
4267
4268 Map.clear();
4269 Sequence.clear();
4270 UsedByIndicesMap.clear();
4271 UniqueItems.clear();
4272
4273 // Now iterate through the worklist and add new formulae.
4274 for (const WorkItem &WI : WorkItems) {
4275 size_t LUIdx = WI.LUIdx;
4276 LSRUse &LU = Uses[LUIdx];
4277 int64_t Imm = WI.Imm;
4278 const SCEV *OrigReg = WI.OrigReg;
4279
4280 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4281 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4282 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4283
4284 // TODO: Use a more targeted data structure.
4285 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4286 Formula F = LU.Formulae[L];
4287 // FIXME: The code for the scaled and unscaled registers looks
4288 // very similar but slightly different. Investigate if they
4289 // could be merged. That way, we would not have to unscale the
4290 // Formula.
4291 F.unscale();
4292 // Use the immediate in the scaled register.
4293 if (F.ScaledReg == OrigReg) {
4294 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4295 // Don't create 50 + reg(-50).
4296 if (F.referencesReg(SE.getSCEV(
4297 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4298 continue;
4299 Formula NewF = F;
4300 NewF.BaseOffset = Offset;
4301 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4302 NewF))
4303 continue;
4304 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4305
4306 // If the new scale is a constant in a register, and adding the constant
4307 // value to the immediate would produce a value closer to zero than the
4308 // immediate itself, then the formula isn't worthwhile.
4309 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4310 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4311 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4312 .ule(std::abs(NewF.BaseOffset)))
4313 continue;
4314
4315 // OK, looks good.
4316 NewF.canonicalize(*this->L);
4317 (void)InsertFormula(LU, LUIdx, NewF);
4318 } else {
4319 // Use the immediate in a base register.
4320 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4321 const SCEV *BaseReg = F.BaseRegs[N];
4322 if (BaseReg != OrigReg)
4323 continue;
4324 Formula NewF = F;
4325 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4326 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4327 LU.Kind, LU.AccessTy, NewF)) {
4328 if (AMK == TTI::AMK_PostIndexed &&
4329 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4330 continue;
4331 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4332 continue;
4333 NewF = F;
4334 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4335 }
4336 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4337
4338 // If the new formula has a constant in a register, and adding the
4339 // constant value to the immediate would produce a value closer to
4340 // zero than the immediate itself, then the formula isn't worthwhile.
4341 for (const SCEV *NewReg : NewF.BaseRegs)
4342 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4343 if ((C->getAPInt() + NewF.BaseOffset)
4344 .abs()
4345 .slt(std::abs(NewF.BaseOffset)) &&
4346 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4347 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4348 goto skip_formula;
4349
4350 // Ok, looks good.
4351 NewF.canonicalize(*this->L);
4352 (void)InsertFormula(LU, LUIdx, NewF);
4353 break;
4354 skip_formula:;
4355 }
4356 }
4357 }
4358 }
4359}
4360
4361/// Generate formulae for each use.
4362void
4363LSRInstance::GenerateAllReuseFormulae() {
4364 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4365 // queries are more precise.
4366 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4367 LSRUse &LU = Uses[LUIdx];
4368 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4369 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4370 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4371 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4372 }
4373 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4374 LSRUse &LU = Uses[LUIdx];
4375 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4376 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4377 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4378 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4379 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4380 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4381 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4382 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4383 }
4384 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4385 LSRUse &LU = Uses[LUIdx];
4386 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4387 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4388 }
4389
4390 GenerateCrossUseConstantOffsets();
4391
4392 LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
4393 "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)
4394 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
;
4395}
4396
4397/// If there are multiple formulae with the same set of registers used
4398/// by other uses, pick the best one and delete the others.
4399void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4400 DenseSet<const SCEV *> VisitedRegs;
4401 SmallPtrSet<const SCEV *, 16> Regs;
4402 SmallPtrSet<const SCEV *, 16> LoserRegs;
4403#ifndef NDEBUG
4404 bool ChangedFormulae = false;
4405#endif
4406
4407 // Collect the best formula for each unique set of shared registers. This
4408 // is reset for each use.
4409 using BestFormulaeTy =
4410 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4411
4412 BestFormulaeTy BestFormulae;
4413
4414 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4415 LSRUse &LU = Uses[LUIdx];
4416 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)
4417 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4418
4419 bool Any = false;
4420 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4421 FIdx != NumForms; ++FIdx) {
4422 Formula &F = LU.Formulae[FIdx];
4423
4424 // Some formulas are instant losers. For example, they may depend on
4425 // nonexistent AddRecs from other loops. These need to be filtered
4426 // immediately, otherwise heuristics could choose them over others leading
4427 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4428 // avoids the need to recompute this information across formulae using the
4429 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4430 // the corresponding bad register from the Regs set.
4431 Cost CostF(L, SE, TTI, AMK);
4432 Regs.clear();
4433 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4434 if (CostF.isLoser()) {
4435 // During initial formula generation, undesirable formulae are generated
4436 // by uses within other loops that have some non-trivial address mode or
4437 // use the postinc form of the IV. LSR needs to provide these formulae
4438 // as the basis of rediscovering the desired formula that uses an AddRec
4439 // corresponding to the existing phi. Once all formulae have been
4440 // generated, these initial losers may be pruned.
4441 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)
4442 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
;
4443 }
4444 else {
4445 SmallVector<const SCEV *, 4> Key;
4446 for (const SCEV *Reg : F.BaseRegs) {
4447 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4448 Key.push_back(Reg);
4449 }
4450 if (F.ScaledReg &&
4451 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4452 Key.push_back(F.ScaledReg);
4453 // Unstable sort by host order ok, because this is only used for
4454 // uniquifying.
4455 llvm::sort(Key);
4456
4457 std::pair<BestFormulaeTy::const_iterator, bool> P =
4458 BestFormulae.insert(std::make_pair(Key, FIdx));
4459 if (P.second)
4460 continue;
4461
4462 Formula &Best = LU.Formulae[P.first->second];
4463
4464 Cost CostBest(L, SE, TTI, AMK);
4465 Regs.clear();
4466 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4467 if (CostF.isLess(CostBest))
4468 std::swap(F, Best);
4469 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)
4470 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)
4471 " 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)
4472 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)
;
4473 }
4474#ifndef NDEBUG
4475 ChangedFormulae = true;
4476#endif
4477 LU.DeleteFormula(F);
4478 --FIdx;
4479 --NumForms;
4480 Any = true;
4481 }
4482
4483 // Now that we've filtered out some formulae, recompute the Regs set.
4484 if (Any)
4485 LU.RecomputeRegs(LUIdx, RegUses);
4486
4487 // Reset this to prepare for the next use.
4488 BestFormulae.clear();
4489 }
4490
4491 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)
4492 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)
4493 "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)
4494 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)
4495 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4496}
4497
4498/// Estimate the worst-case number of solutions the solver might have to
4499/// consider. It almost never considers this many solutions because it prune the
4500/// search space, but the pruning isn't always sufficient.
4501size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4502 size_t Power = 1;
4503 for (const LSRUse &LU : Uses) {
4504 size_t FSize = LU.Formulae.size();
4505 if (FSize >= ComplexityLimit) {
4506 Power = ComplexityLimit;
4507 break;
4508 }
4509 Power *= FSize;
4510 if (Power >= ComplexityLimit)
4511 break;
4512 }
4513 return Power;
4514}
4515
4516/// When one formula uses a superset of the registers of another formula, it
4517/// won't help reduce register pressure (though it may not necessarily hurt
4518/// register pressure); remove it to simplify the system.
4519void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4520 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4521 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)
;
4522
4523 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)
4524 "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)
4525 "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)
;
4526
4527 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4528 LSRUse &LU = Uses[LUIdx];
4529 bool Any = false;
4530 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4531 Formula &F = LU.Formulae[i];
4532 // Look for a formula with a constant or GV in a register. If the use
4533 // also has a formula with that same value in an immediate field,
4534 // delete the one that uses a register.
4535 for (SmallVectorImpl<const SCEV *>::const_iterator
4536 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4537 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4538 Formula NewF = F;
4539 //FIXME: Formulas should store bitwidth to do wrapping properly.
4540 // See PR41034.
4541 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4542 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4543 (I - F.BaseRegs.begin()));
4544 if (LU.HasFormulaWithSameRegs(NewF)) {
4545 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4546 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4547 LU.DeleteFormula(F);
4548 --i;
4549 --e;
4550 Any = true;
4551 break;
4552 }
4553 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4554 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4555 if (!F.BaseGV) {
4556 Formula NewF = F;
4557 NewF.BaseGV = GV;
4558 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4559 (I - F.BaseRegs.begin()));
4560 if (LU.HasFormulaWithSameRegs(NewF)) {
4561 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
4562 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
;
4563 LU.DeleteFormula(F);
4564 --i;
4565 --e;
4566 Any = true;
4567 break;
4568 }
4569 }
4570 }
4571 }
4572 }
4573 if (Any)
4574 LU.RecomputeRegs(LUIdx, RegUses);
4575 }
4576
4577 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)
;
4578 }
4579}
4580
4581/// When there are many registers for expressions like A, A+1, A+2, etc.,
4582/// allocate a single register for them.
4583void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4584 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4585 return;
4586
4587 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)
4588 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)
4589 "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)
4590 "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)
;
4591
4592 // This is especially useful for unrolled loops.
4593
4594 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4595 LSRUse &LU = Uses[LUIdx];
4596 for (const Formula &F : LU.Formulae) {
4597 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4598 continue;
4599
4600 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4601 if (!LUThatHas)
4602 continue;
4603
4604 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4605 LU.Kind, LU.AccessTy))
4606 continue;
4607
4608 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)
;
4609
4610 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4611
4612 // Transfer the fixups of LU to LUThatHas.
4613 for (LSRFixup &Fixup : LU.Fixups) {
4614 Fixup.Offset += F.BaseOffset;
4615 LUThatHas->pushFixup(Fixup);
4616 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)
;
4617 }
4618
4619 // Delete formulae from the new use which are no longer legal.
4620 bool Any = false;
4621 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4622 Formula &F = LUThatHas->Formulae[i];
4623 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4624 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4625 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)
;
4626 LUThatHas->DeleteFormula(F);
4627 --i;
4628 --e;
4629 Any = true;
4630 }
4631 }
4632
4633 if (Any)
4634 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4635
4636 // Delete the old use.
4637 DeleteUse(LU, LUIdx);
4638 --LUIdx;
4639 --NumUses;
4640 break;
4641 }
4642 }
4643
4644 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)
;
4645}
4646
4647/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4648/// we've done more filtering, as it may be able to find more formulae to
4649/// eliminate.
4650void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4651 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4652 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)
;
4653
4654 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)
4655 "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)
;
4656
4657 FilterOutUndesirableDedicatedRegisters();
4658
4659 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)
;
4660 }
4661}
4662
4663/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4664/// Pick the best one and delete the others.
4665/// This narrowing heuristic is to keep as many formulae with different
4666/// Scale and ScaledReg pair as possible while narrowing the search space.
4667/// The benefit is that it is more likely to find out a better solution
4668/// from a formulae set with more Scale and ScaledReg variations than
4669/// a formulae set with the same Scale and ScaledReg. The picking winner
4670/// reg heuristic will often keep the formulae with the same Scale and
4671/// ScaledReg and filter others, and we want to avoid that if possible.
4672void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4673 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4674 return;
4675
4676 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)
4677 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)
4678 "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)
4679 "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)
;
4680
4681 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4682 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4683
4684 BestFormulaeTy BestFormulae;
4685#ifndef NDEBUG
4686 bool ChangedFormulae = false;
4687#endif
4688 DenseSet<const SCEV *> VisitedRegs;
4689 SmallPtrSet<const SCEV *, 16> Regs;
4690
4691 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4692 LSRUse &LU = Uses[LUIdx];
4693 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)
4694 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
;
4695
4696 // Return true if Formula FA is better than Formula FB.
4697 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4698 // First we will try to choose the Formula with fewer new registers.
4699 // For a register used by current Formula, the more the register is
4700 // shared among LSRUses, the less we increase the register number
4701 // counter of the formula.
4702 size_t FARegNum = 0;
4703 for (const SCEV *Reg : FA.BaseRegs) {
4704 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4705 FARegNum += (NumUses - UsedByIndices.count() + 1);
4706 }
4707 size_t FBRegNum = 0;
4708 for (const SCEV *Reg : FB.BaseRegs) {
4709 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4710 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4711 }
4712 if (FARegNum != FBRegNum)
4713 return FARegNum < FBRegNum;
4714
4715 // If the new register numbers are the same, choose the Formula with
4716 // less Cost.
4717 Cost CostFA(L, SE, TTI, AMK);
4718 Cost CostFB(L, SE, TTI, AMK);
4719 Regs.clear();
4720 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4721 Regs.clear();
4722 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4723 return CostFA.isLess(CostFB);
4724 };
4725
4726 bool Any = false;
4727 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4728 ++FIdx) {
4729 Formula &F = LU.Formulae[FIdx];
4730 if (!F.ScaledReg)
4731 continue;
4732 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4733 if (P.second)
4734 continue;
4735
4736 Formula &Best = LU.Formulae[P.first->second];
4737 if (IsBetterThan(F, Best))
4738 std::swap(F, Best);
4739 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)
4740 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)
4741 " 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)
4742 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)
;
4743#ifndef NDEBUG
4744 ChangedFormulae = true;
4745#endif
4746 LU.DeleteFormula(F);
4747 --FIdx;
4748 --NumForms;
4749 Any = true;
4750 }
4751 if (Any)
4752 LU.RecomputeRegs(LUIdx, RegUses);
4753
4754 // Reset this to prepare for the next use.
4755 BestFormulae.clear();
4756 }
4757
4758 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)
4759 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)
4760 "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)
4761 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)
4762 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
"After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
;
4763}
4764
4765/// If we are over the complexity limit, filter out any post-inc prefering
4766/// variables to only post-inc values.
4767void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
4768 if (AMK != TTI::AMK_PostIndexed)
4769 return;
4770 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4771 return;
4772
4773 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)
4774 "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)
4775 "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)
;
4776
4777 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4778 LSRUse &LU = Uses[LUIdx];
4779
4780 if (LU.Kind != LSRUse::Address)
4781 continue;
4782 if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
4783 !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
4784 continue;
4785
4786 size_t MinRegs = std::numeric_limits<size_t>::max();
4787 for (const Formula &F : LU.Formulae)
4788 MinRegs = std::min(F.getNumRegs(), MinRegs);
4789
4790 bool Any = false;
4791 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4792 ++FIdx) {
4793 Formula &F = LU.Formulae[FIdx];
4794 if (F.getNumRegs() > MinRegs) {
4795 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)
4796 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Filtering out formula "
; F.print(dbgs()); dbgs() << "\n"; } } while (false)
;
4797 LU.DeleteFormula(F);
4798 --FIdx;
4799 --NumForms;
4800 Any = true;
4801 }
4802 }
4803 if (Any)
4804 LU.RecomputeRegs(LUIdx, RegUses);
4805
4806 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4807 break;
4808 }
4809
4810 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)
;
4811}
4812
4813/// The function delete formulas with high registers number expectation.
4814/// Assuming we don't know the value of each formula (already delete
4815/// all inefficient), generate probability of not selecting for each
4816/// register.
4817/// For example,
4818/// Use1:
4819/// reg(a) + reg({0,+,1})
4820/// reg(a) + reg({-1,+,1}) + 1
4821/// reg({a,+,1})
4822/// Use2:
4823/// reg(b) + reg({0,+,1})
4824/// reg(b) + reg({-1,+,1}) + 1
4825/// reg({b,+,1})
4826/// Use3:
4827/// reg(c) + reg(b) + reg({0,+,1})
4828/// reg(c) + reg({b,+,1})
4829///
4830/// Probability of not selecting
4831/// Use1 Use2 Use3
4832/// reg(a) (1/3) * 1 * 1
4833/// reg(b) 1 * (1/3) * (1/2)
4834/// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4835/// reg({-1,+,1}) (2/3) * (2/3) * 1
4836/// reg({a,+,1}) (2/3) * 1 * 1
4837/// reg({b,+,1}) 1 * (2/3) * (2/3)
4838/// reg(c) 1 * 1 * 0
4839///
4840/// Now count registers number mathematical expectation for each formula:
4841/// Note that for each use we exclude probability if not selecting for the use.
4842/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4843/// probabilty 1/3 of not selecting for Use1).
4844/// Use1:
4845/// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4846/// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4847/// reg({a,+,1}) 1
4848/// Use2:
4849/// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4850/// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4851/// reg({b,+,1}) 2/3
4852/// Use3:
4853/// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4854/// reg(c) + reg({b,+,1}) 1 + 2/3
4855void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4856 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4857 return;
4858 // Ok, we have too many of formulae on our hands to conveniently handle.
4859 // Use a rough heuristic to thin out the list.
4860
4861 // Set of Regs wich will be 100% used in final solution.
4862 // Used in each formula of a solution (in example above this is reg(c)).
4863 // We can skip them in calculations.
4864 SmallPtrSet<const SCEV *, 4> UniqRegs;
4865 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)
;
4866
4867 // Map each register to probability of not selecting
4868 DenseMap <const SCEV *, float> RegNumMap;
4869 for (const SCEV *Reg : RegUses) {
4870 if (UniqRegs.count(Reg))
4871 continue;
4872 float PNotSel = 1;
4873 for (const LSRUse &LU : Uses) {
4874 if (!LU.Regs.count(Reg))
4875 continue;
4876 float P = LU.getNotSelectedProbability(Reg);
4877 if (P != 0.0)
4878 PNotSel *= P;
4879 else
4880 UniqRegs.insert(Reg);
4881 }
4882 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4883 }
4884
4885 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
4886 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)
;
4887
4888 // Delete formulas where registers number expectation is high.
4889 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4890 LSRUse &LU = Uses[LUIdx];
4891 // If nothing to delete - continue.
4892 if (LU.Formulae.size() < 2)
4893 continue;
4894 // This is temporary solution to test performance. Float should be
4895 // replaced with round independent type (based on integers) to avoid
4896 // different results for different target builds.
4897 float FMinRegNum = LU.Formulae[0].getNumRegs();
4898 float FMinARegNum = LU.Formulae[0].getNumRegs();
4899 size_t MinIdx = 0;
4900 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4901 Formula &F = LU.Formulae[i];
4902 float FRegNum = 0;
4903 float FARegNum = 0;
4904 for (const SCEV *BaseReg : F.BaseRegs) {
4905 if (UniqRegs.count(BaseReg))
4906 continue;
4907 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4908 if (isa<SCEVAddRecExpr>(BaseReg))
4909 FARegNum +=
4910 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4911 }
4912 if (const SCEV *ScaledReg = F.ScaledReg) {
4913 if (!UniqRegs.count(ScaledReg)) {
4914 FRegNum +=
4915 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4916 if (isa<SCEVAddRecExpr>(ScaledReg))
4917 FARegNum +=
4918 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4919 }
4920 }
4921 if (FMinRegNum > FRegNum ||
4922 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4923 FMinRegNum = FRegNum;
4924 FMinARegNum = FARegNum;
4925 MinIdx = i;
4926 }
4927 }
4928 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)
4929 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)
;
4930 if (MinIdx != 0)
4931 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4932 while (LU.Formulae.size() != 1) {
4933 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
)
4934 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << " Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
;
4935 LU.Formulae.pop_back();
4936 }
4937 LU.RecomputeRegs(LUIdx, RegUses);
4938 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", 4938, __extension__
__PRETTY_FUNCTION__))
;
4939 Formula &F = LU.Formulae[0];
4940 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)
;
4941 // When we choose the formula, the regs become unique.
4942 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4943 if (F.ScaledReg)
4944 UniqRegs.insert(F.ScaledReg);
4945 }
4946 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)
;
4947}
4948
4949/// Pick a register which seems likely to be profitable, and then in any use
4950/// which has any reference to that register, delete all formulae which do not
4951/// reference that register.
4952void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4953 // With all other options exhausted, loop until the system is simple
4954 // enough to handle.
4955 SmallPtrSet<const SCEV *, 4> Taken;
4956 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4957 // Ok, we have too many of formulae on our hands to conveniently handle.
4958 // Use a rough heuristic to thin out the list.
4959 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)
;
4960
4961 // Pick the register which is used by the most LSRUses, which is likely
4962 // to be a good reuse register candidate.
4963 const SCEV *Best = nullptr;
4964 unsigned BestNum = 0;
4965 for (const SCEV *Reg : RegUses) {
4966 if (Taken.count(Reg))
4967 continue;
4968 if (!Best) {
4969 Best = Reg;
4970 BestNum = RegUses.getUsedByIndices(Reg).count();
4971 } else {
4972 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4973 if (Count > BestNum) {
4974 Best = Reg;
4975 BestNum = Count;
4976 }
4977 }
4978 }
4979 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", 4979, __extension__
__PRETTY_FUNCTION__))
;
4980
4981 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)
4982 << " 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)
;
4983 Taken.insert(Best);
4984
4985 // In any use with formulae which references this register, delete formulae
4986 // which don't reference it.
4987 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4988 LSRUse &LU = Uses[LUIdx];
4989 if (!LU.Regs.count(Best)) continue;
4990
4991 bool Any = false;
4992 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4993 Formula &F = LU.Formulae[i];
4994 if (!F.referencesReg(Best)) {
4995 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)
;
4996 LU.DeleteFormula(F);
4997 --e;
4998 --i;
4999 Any = true;
5000 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", 5000, __extension__
__PRETTY_FUNCTION__))
;
5001 continue;
5002 }
5003 }
5004
5005 if (Any)
5006 LU.RecomputeRegs(LUIdx, RegUses);
5007 }
5008
5009 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)
;
5010 }
5011}
5012
5013/// If there are an extraordinary number of formulae to choose from, use some
5014/// rough heuristics to prune down the number of formulae. This keeps the main
5015/// solver from taking an extraordinary amount of time in some worst-case
5016/// scenarios.
5017void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
5018 NarrowSearchSpaceByDetectingSupersets();
5019 NarrowSearchSpaceByCollapsingUnrolledCode();
5020 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
5021 if (FilterSameScaledReg)
5022 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
5023 NarrowSearchSpaceByFilterPostInc();
5024 if (LSRExpNarrow)
5025 NarrowSearchSpaceByDeletingCostlyFormulas();
5026 else
5027 NarrowSearchSpaceByPickingWinnerRegs();
5028}
5029
5030/// This is the recursive solver.
5031void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
5032 Cost &SolutionCost,
5033 SmallVectorImpl<const Formula *> &Workspace,
5034 const Cost &CurCost,
5035 const SmallPtrSet<const SCEV *, 16> &CurRegs,
5036 DenseSet<const SCEV *> &VisitedRegs) const {
5037 // Some ideas:
5038 // - prune more:
5039 // - use more aggressive filtering
5040 // - sort the formula so that the most profitable solutions are found first
5041 // - sort the uses too
5042 // - search faster:
5043 // - don't compute a cost, and then compare. compare while computing a cost
5044 // and bail early.
5045 // - track register sets with SmallBitVector
5046
5047 const LSRUse &LU = Uses[Workspace.size()];
5048
5049 // If this use references any register that's already a part of the
5050 // in-progress solution, consider it a requirement that a formula must
5051 // reference that register in order to be considered. This prunes out
5052 // unprofitable searching.
5053 SmallSetVector<const SCEV *, 4> ReqRegs;
5054 for (const SCEV *S : CurRegs)
5055 if (LU.Regs.count(S))
5056 ReqRegs.insert(S);
5057
5058 SmallPtrSet<const SCEV *, 16> NewRegs;
5059 Cost NewCost(L, SE, TTI, AMK);
5060 for (const Formula &F : LU.Formulae) {
5061 // Ignore formulae which may not be ideal in terms of register reuse of
5062 // ReqRegs. The formula should use all required registers before
5063 // introducing new ones.
5064 // This can sometimes (notably when trying to favour postinc) lead to
5065 // sub-optimial decisions. There it is best left to the cost modelling to
5066 // get correct.
5067 if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) {
5068 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
5069 for (const SCEV *Reg : ReqRegs) {
5070 if ((F.ScaledReg && F.ScaledReg == Reg) ||
5071 is_contained(F.BaseRegs, Reg)) {
5072 --NumReqRegsToFind;
5073 if (NumReqRegsToFind == 0)
5074 break;
5075 }
5076 }
5077 if (NumReqRegsToFind != 0) {
5078 // If none of the formulae satisfied the required registers, then we could
5079 // clear ReqRegs and try again. Currently, we simply give up in this case.
5080 continue;
5081 }
5082 }
5083
5084 // Evaluate the cost of the current formula. If it's already worse than
5085 // the current best, prune the search at that point.
5086 NewCost = CurCost;
5087 NewRegs = CurRegs;
5088 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
5089 if (NewCost.isLess(SolutionCost)) {
5090 Workspace.push_back(&F);
5091 if (Workspace.size() != Uses.size()) {
5092 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
5093 NewRegs, VisitedRegs);
5094 if (F.getNumRegs() == 1 && Workspace.size() == 1)
5095 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
5096 } else {
5097 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)
5098 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)
5099 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)
5100 << "- " << *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)
5101 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)
;
5102
5103 SolutionCost = NewCost;
5104 Solution = Workspace;
5105 }
5106 Workspace.pop_back();
5107 }
5108 }
5109}
5110
5111/// Choose one formula from each use. Return the results in the given Solution
5112/// vector.
5113void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
5114 SmallVector<const Formula *, 8> Workspace;
5115 Cost SolutionCost(L, SE, TTI, AMK);
5116 SolutionCost.Lose();
5117 Cost CurCost(L, SE, TTI, AMK);
5118 SmallPtrSet<const SCEV *, 16> CurRegs;
5119 DenseSet<const SCEV *> VisitedRegs;
5120 Workspace.reserve(Uses.size());
5121
5122 // SolveRecurse does all the work.
5123 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
5124 CurRegs, VisitedRegs);
5125 if (Solution.empty()) {
5126 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false)
;
5127 return;
5128 }
5129
5130 // Ok, we've now made all our decisions.
5131 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
)
5132 "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
)
5133 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
)
5134 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
)
5135 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
)
5136 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
)
5137 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
)
5138 " ";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
)
5139 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
)
5140 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
)
5141 })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
)
;
5142
5143 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", 5143, __extension__
__PRETTY_FUNCTION__))
;
5144}
5145
5146/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5147/// we can go while still being dominated by the input positions. This helps
5148/// canonicalize the insert position, which encourages sharing.
5149BasicBlock::iterator
5150LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5151 const SmallVectorImpl<Instruction *> &Inputs)
5152 const {
5153 Instruction *Tentative = &*IP;
5154 while (true) {
5155 bool AllDominate = true;
5156 Instruction *BetterPos = nullptr;
5157 // Don't bother attempting to insert before a catchswitch, their basic block
5158 // cannot have other non-PHI instructions.
5159 if (isa<CatchSwitchInst>(Tentative))
5160 return IP;
5161
5162 for (Instruction *Inst : Inputs) {
5163 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5164 AllDominate = false;
5165 break;
5166 }
5167 // Attempt to find an insert position in the middle of the block,
5168 // instead of at the end, so that it can be used for other expansions.
5169 if (Tentative->getParent() == Inst->getParent() &&
5170 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5171 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5172 }
5173 if (!AllDominate)
5174 break;
5175 if (BetterPos)
5176 IP = BetterPos->getIterator();
5177 else
5178 IP = Tentative->getIterator();
5179
5180 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5181 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5182
5183 BasicBlock *IDom;
5184 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5185 if (!Rung) return IP;
5186 Rung = Rung->getIDom();
5187 if (!Rung) return IP;
5188 IDom = Rung->getBlock();
5189
5190 // Don't climb into a loop though.
5191 const Loop *IDomLoop = LI.getLoopFor(IDom);
5192 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5193 if (IDomDepth <= IPLoopDepth &&
5194 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5195 break;
5196 }
5197
5198 Tentative = IDom->getTerminator();
5199 }
5200
5201 return IP;
5202}
5203
5204/// Determine an input position which will be dominated by the operands and
5205/// which will dominate the result.
5206BasicBlock::iterator
5207LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5208 const LSRFixup &LF,
5209 const LSRUse &LU,
5210 SCEVExpander &Rewriter) const {
5211 // Collect some instructions which must be dominated by the
5212 // expanding replacement. These must be dominated by any operands that
5213 // will be required in the expansion.
5214 SmallVector<Instruction *, 4> Inputs;
5215 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5216 Inputs.push_back(I);
5217 if (LU.Kind == LSRUse::ICmpZero)
5218 if (Instruction *I =
5219 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5220 Inputs.push_back(I);
5221 if (LF.PostIncLoops.count(L)) {
5222 if (LF.isUseFullyOutsideLoop(L))
5223 Inputs.push_back(L->getLoopLatch()->getTerminator());
5224 else
5225 Inputs.push_back(IVIncInsertPos);
5226 }
5227 // The expansion must also be dominated by the increment positions of any
5228 // loops it for which it is using post-inc mode.
5229 for (const Loop *PIL : LF.PostIncLoops) {
5230 if (PIL == L) continue;
5231
5232 // Be dominated by the loop exit.
5233 SmallVector<BasicBlock *, 4> ExitingBlocks;
5234 PIL->getExitingBlocks(ExitingBlocks);
5235 if (!ExitingBlocks.empty()) {
5236 BasicBlock *BB = ExitingBlocks[0];
5237 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5238 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5239 Inputs.push_back(BB->getTerminator());
5240 }
5241 }
5242
5243 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", 5245, __extension__
__PRETTY_FUNCTION__))
5244 && !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", 5245, __extension__
__PRETTY_FUNCTION__))
5245 "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", 5245, __extension__
__PRETTY_FUNCTION__))
;
5246
5247 // Then, climb up the immediate dominator tree as far as we can go while
5248 // still being dominated by the input positions.
5249 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5250
5251 // Don't insert instructions before PHI nodes.
5252 while (isa<PHINode>(IP)) ++IP;
5253
5254 // Ignore landingpad instructions.
5255 while (IP->isEHPad()) ++IP;
5256
5257 // Ignore debug intrinsics.
5258 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5259
5260 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5261 // IP consistent across expansions and allows the previously inserted
5262 // instructions to be reused by subsequent expansion.
5263 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5264 ++IP;
5265
5266 return IP;
5267}
5268
5269/// Emit instructions for the leading candidate expression for this LSRUse (this
5270/// is called "expanding").
5271Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5272 const Formula &F, BasicBlock::iterator IP,
5273 SCEVExpander &Rewriter,
5274 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5275 if (LU.RigidFormula)
10
Assuming field 'RigidFormula' is false
11
Taking false branch
5276 return LF.OperandValToReplace;
5277
5278 // Determine an input position which will be dominated by the operands and
5279 // which will dominate the result.
5280 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5281 Rewriter.setInsertPoint(&*IP);
5282
5283 // Inform the Rewriter if we have a post-increment use, so that it can
5284 // perform an advantageous expansion.
5285 Rewriter.setPostInc(LF.PostIncLoops);
5286
5287 // This is the type that the user actually needs.
5288 Type *OpTy = LF.OperandValToReplace->getType();
12
Called C++ object pointer is null
5289 // This will be the type that we'll initially expand to.
5290 Type *Ty = F.getType();
5291 if (!Ty)
5292 // No type known; just expand directly to the ultimate type.
5293 Ty = OpTy;
5294 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5295 // Expand directly to the ultimate type if it's the right size.
5296 Ty = OpTy;
5297 // This is the type to do integer arithmetic in.
5298 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5299
5300 // Build up a list of operands to add together to form the full base.
5301 SmallVector<const SCEV *, 8> Ops;
5302
5303 // Expand the BaseRegs portion.
5304 for (const SCEV *Reg : F.BaseRegs) {
5305 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", 5305, __extension__
__PRETTY_FUNCTION__))
;
5306
5307 // If we're expanding for a post-inc user, make the post-inc adjustment.
5308 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5309 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5310 }
5311
5312 // Expand the ScaledReg portion.
5313 Value *ICmpScaledV = nullptr;
5314 if (F.Scale != 0) {
5315 const SCEV *ScaledS = F.ScaledReg;
5316
5317 // If we're expanding for a post-inc user, make the post-inc adjustment.
5318 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5319 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5320
5321 if (LU.Kind == LSRUse::ICmpZero) {
5322 // Expand ScaleReg as if it was part of the base regs.
5323 if (F.Scale == 1)
5324 Ops.push_back(
5325 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5326 else {
5327 // An interesting way of "folding" with an icmp is to use a negated
5328 // scale, which we'll implement by inserting it into the other operand
5329 // of the icmp.
5330 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", 5331, __extension__
__PRETTY_FUNCTION__))
5331 "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", 5331, __extension__
__PRETTY_FUNCTION__))
;
5332 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5333 }
5334 } else {
5335 // Otherwise just expand the scaled register and an explicit scale,
5336 // which is expected to be matched as part of the address.
5337
5338 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5339 // Unless the addressing mode will not be folded.
5340 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5341 isAMCompletelyFolded(TTI, LU, F)) {
5342 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5343 Ops.clear();
5344 Ops.push_back(SE.getUnknown(FullV));
5345 }
5346 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5347 if (F.Scale != 1)
5348 ScaledS =
5349 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5350 Ops.push_back(ScaledS);
5351 }
5352 }
5353
5354 // Expand the GV portion.
5355 if (F.BaseGV) {
5356 // Flush the operand list to suppress SCEVExpander hoisting.
5357 if (!Ops.empty()) {
5358 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), IntTy);
5359 Ops.clear();
5360 Ops.push_back(SE.getUnknown(FullV));
5361 }
5362 Ops.push_back(SE.getUnknown(F.BaseGV));
5363 }
5364
5365 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5366 // unfolded offsets. LSR assumes they both live next to their uses.
5367 if (!Ops.empty()) {
5368 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5369 Ops.clear();
5370 Ops.push_back(SE.getUnknown(FullV));
5371 }
5372
5373 // Expand the immediate portion.
5374 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5375 if (Offset != 0) {
5376 if (LU.Kind == LSRUse::ICmpZero) {
5377 // The other interesting way of "folding" with an ICmpZero is to use a
5378 // negated immediate.
5379 if (!ICmpScaledV)
5380 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5381 else {
5382 Ops.push_back(SE.getUnknown(ICmpScaledV));
5383 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5384 }
5385 } else {
5386 // Just add the immediate values. These again are expected to be matched
5387 // as part of the address.
5388 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5389 }
5390 }
5391
5392 // Expand the unfolded offset portion.
5393 int64_t UnfoldedOffset = F.UnfoldedOffset;
5394 if (UnfoldedOffset != 0) {
5395 // Just add the immediate values.
5396 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5397 UnfoldedOffset)));
5398 }
5399
5400 // Emit instructions summing all the operands.
5401 const SCEV *FullS = Ops.empty() ?
5402 SE.getConstant(IntTy, 0) :
5403 SE.getAddExpr(Ops);
5404 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5405
5406 // We're done expanding now, so reset the rewriter.
5407 Rewriter.clearPostInc();
5408
5409 // An ICmpZero Formula represents an ICmp which we're handling as a
5410 // comparison against zero. Now that we've expanded an expression for that
5411 // form, update the ICmp's other operand.
5412 if (LU.Kind == LSRUse::ICmpZero) {
5413 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5414 if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
5415 DeadInsts.emplace_back(OperandIsInstr);
5416 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", 5417, __extension__
__PRETTY_FUNCTION__))
5417 "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", 5417, __extension__
__PRETTY_FUNCTION__))
;
5418 if (F.Scale == -1) {
5419 if (ICmpScaledV->getType() != OpTy) {
5420 Instruction *Cast =
5421 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5422 OpTy, false),
5423 ICmpScaledV, OpTy, "tmp", CI);
5424 ICmpScaledV = Cast;
5425 }
5426 CI->setOperand(1, ICmpScaledV);
5427 } else {
5428 // A scale of 1 means that the scale has been expanded as part of the
5429 // base regs.
5430 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", 5432, __extension__
__PRETTY_FUNCTION__))
5431 "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", 5432, __extension__
__PRETTY_FUNCTION__))
5432 "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", 5432, __extension__
__PRETTY_FUNCTION__))
;
5433 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5434 -(uint64_t)Offset);
5435 if (C->getType() != OpTy)
5436 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5437 OpTy, false),
5438 C, OpTy);
5439
5440 CI->setOperand(1, C);
5441 }
5442 }
5443
5444 return FullV;
5445}
5446
5447/// Helper for Rewrite. PHI nodes are special because the use of their operands
5448/// effectively happens in their predecessor blocks, so the expression may need
5449/// to be expanded in multiple places.
5450void LSRInstance::RewriteForPHI(
5451 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5452 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5453 DenseMap<BasicBlock *, Value *> Inserted;
5454 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1
Assuming 'i' is not equal to 'e'
2
Loop condition is true. Entering loop body
5455 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3
Assuming pointer value is null
4
Taking true branch
5456 bool needUpdateFixups = false;
5457 BasicBlock *BB = PN->getIncomingBlock(i);
5458
5459 // If this is a critical edge, split the edge so that we do not insert
5460 // the code on all predecessor/successor paths. We do this unless this
5461 // is the canonical backedge for this loop, which complicates post-inc
5462 // users.
5463 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5
Assuming 'e' is equal to 1
6
Taking false branch
5464 !isa<IndirectBrInst>(BB->getTerminator()) &&
5465 !isa<CatchSwitchInst>(BB->getTerminator())) {
5466 BasicBlock *Parent = PN->getParent();
5467 Loop *PNLoop = LI.getLoopFor(Parent);
5468 if (!PNLoop || Parent != PNLoop->getHeader()) {
5469 // Split the critical edge.
5470 BasicBlock *NewBB = nullptr;
5471 if (!Parent->isLandingPad()) {
5472 NewBB =
5473 SplitCriticalEdge(BB, Parent,
5474 CriticalEdgeSplittingOptions(&DT, &LI, MSSAU)
5475 .setMergeIdenticalEdges()
5476 .setKeepOneInputPHIs());
5477 } else {
5478 SmallVector<BasicBlock*, 2> NewBBs;
5479 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5480 NewBB = NewBBs[0];
5481 }
5482 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5483 // phi predecessors are identical. The simple thing to do is skip
5484 // splitting in this case rather than complicate the API.
5485 if (NewBB) {
5486 // If PN is outside of the loop and BB is in the loop, we want to
5487 // move the block to be immediately before the PHI block, not
5488 // immediately after BB.
5489 if (L->contains(BB) && !L->contains(PN))
5490 NewBB->moveBefore(PN->getParent());
5491
5492 // Splitting the edge can reduce the number of PHI entries we have.
5493 e = PN->getNumIncomingValues();
5494 BB = NewBB;
5495 i = PN->getBasicBlockIndex(BB);
5496
5497 needUpdateFixups = true;
5498 }
5499 }
5500 }
5501
5502 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5503 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5504 if (!Pair.second)
7
Assuming field 'second' is true
8
Taking false branch
5505 PN->setIncomingValue(i, Pair.first->second);
5506 else {
5507 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
9
Calling 'LSRInstance::Expand'
5508 Rewriter, DeadInsts);
5509
5510 // If this is reuse-by-noop-cast, insert the noop cast.
5511 Type *OpTy = LF.OperandValToReplace->getType();
5512 if (FullV->getType() != OpTy)
5513 FullV =
5514 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5515 OpTy, false),
5516 FullV, LF.OperandValToReplace->getType(),
5517 "tmp", BB->getTerminator());
5518
5519 PN->setIncomingValue(i, FullV);
5520 Pair.first->second = FullV;
5521 }
5522
5523 // If LSR splits critical edge and phi node has other pending
5524 // fixup operands, we need to update those pending fixups. Otherwise
5525 // formulae will not be implemented completely and some instructions
5526 // will not be eliminated.
5527 if (needUpdateFixups) {
5528 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5529 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5530 // If fixup is supposed to rewrite some operand in the phi
5531 // that was just updated, it may be already moved to
5532 // another phi node. Such fixup requires update.
5533 if (Fixup.UserInst == PN) {
5534 // Check if the operand we try to replace still exists in the
5535 // original phi.
5536 bool foundInOriginalPHI = false;
5537 for (const auto &val : PN->incoming_values())
5538 if (val == Fixup.OperandValToReplace) {
5539 foundInOriginalPHI = true;
5540 break;
5541 }
5542
5543 // If fixup operand found in original PHI - nothing to do.
5544 if (foundInOriginalPHI)
5545 continue;
5546
5547 // Otherwise it might be moved to another PHI and requires update.
5548 // If fixup operand not found in any of the incoming blocks that
5549 // means we have already rewritten it - nothing to do.
5550 for (const auto &Block : PN->blocks())
5551 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5552 ++I) {
5553 PHINode *NewPN = cast<PHINode>(I);
5554 for (const auto &val : NewPN->incoming_values())
5555 if (val == Fixup.OperandValToReplace)
5556 Fixup.UserInst = NewPN;
5557 }
5558 }
5559 }
5560 }
5561}
5562
5563/// Emit instructions for the leading candidate expression for this LSRUse (this
5564/// is called "expanding"), and update the UserInst to reference the newly
5565/// expanded value.
5566void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5567 const Formula &F, SCEVExpander &Rewriter,
5568 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5569 // First, find an insertion point that dominates UserInst. For PHI nodes,
5570 // find the nearest block which dominates all the relevant uses.
5571 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5572 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5573 } else {
5574 Value *FullV =
5575 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5576
5577 // If this is reuse-by-noop-cast, insert the noop cast.
5578 Type *OpTy = LF.OperandValToReplace->getType();
5579 if (FullV->getType() != OpTy) {
5580 Instruction *Cast =
5581 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5582 FullV, OpTy, "tmp", LF.UserInst);
5583 FullV = Cast;
5584 }
5585
5586 // Update the user. ICmpZero is handled specially here (for now) because
5587 // Expand may have updated one of the operands of the icmp already, and
5588 // its new value may happen to be equal to LF.OperandValToReplace, in
5589 // which case doing replaceUsesOfWith leads to replacing both operands
5590 // with the same value. TODO: Reorganize this.
5591 if (LU.Kind == LSRUse::ICmpZero)
5592 LF.UserInst->setOperand(0, FullV);
5593 else
5594 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5595 }
5596
5597 if (auto *OperandIsInstr = dyn_cast<Instruction>(LF.OperandValToReplace))
5598 DeadInsts.emplace_back(OperandIsInstr);
5599}
5600
5601// Check if there are any loop exit values which are only used once within the
5602// loop which may potentially be optimized with a call to rewriteLoopExitValue.
5603static bool LoopExitValHasSingleUse(Loop *L) {
5604 BasicBlock *ExitBB = L->getExitBlock();
5605 if (!ExitBB)
5606 return false;
5607
5608 for (PHINode &ExitPhi : ExitBB->phis()) {
5609 if (ExitPhi.getNumIncomingValues() != 1)
5610 break;
5611
5612 BasicBlock *Pred = ExitPhi.getIncomingBlock(0);
5613 Value *IVNext = ExitPhi.getIncomingValueForBlock(Pred);
5614 // One use would be the exit phi node, and there should be only one other
5615 // use for this to be considered.
5616 if (IVNext->getNumUses() == 2)
5617 return true;
5618 }
5619 return false;
5620}
5621
5622/// Rewrite all the fixup locations with new values, following the chosen
5623/// solution.
5624void LSRInstance::ImplementSolution(
5625 const SmallVectorImpl<const Formula *> &Solution) {
5626 // Keep track of instructions we may have made dead, so that
5627 // we can remove them after we are done working.
5628 SmallVector<WeakTrackingVH, 16> DeadInsts;
5629
5630 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(), "lsr",
5631 false);
5632#ifndef NDEBUG
5633 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
5634#endif
5635 Rewriter.disableCanonicalMode();
5636 Rewriter.enableLSRMode();
5637 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5638
5639 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5640 for (const IVChain &Chain : IVChainVec) {
5641 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5642 Rewriter.setChainedPhi(PN);
5643 }
5644
5645 // Expand the new value definitions and update the users.
5646 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5647 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5648 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5649 Changed = true;
5650 }
5651
5652 for (const IVChain &Chain : IVChainVec) {
5653 GenerateIVChain(Chain, Rewriter, DeadInsts);
5654 Changed = true;
5655 }
5656
5657 for (const WeakVH &IV : Rewriter.getInsertedIVs())
5658 if (IV && dyn_cast<Instruction>(&*IV)->getParent())
5659 ScalarEvolutionIVs.push_back(IV);
5660
5661 // Clean up after ourselves. This must be done before deleting any
5662 // instructions.
5663 Rewriter.clear();
5664
5665 Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts,
5666 &TLI, MSSAU);
5667
5668 // In our cost analysis above, we assume that each addrec consumes exactly
5669 // one register, and arrange to have increments inserted just before the
5670 // latch to maximimize the chance this is true. However, if we reused
5671 // existing IVs, we now need to move the increments to match our
5672 // expectations. Otherwise, our cost modeling results in us having a
5673 // chosen a non-optimal result for the actual schedule. (And yes, this
5674 // scheduling decision does impact later codegen.)
5675 for (PHINode &PN : L->getHeader()->phis()) {
5676 BinaryOperator *BO = nullptr;
5677 Value *Start = nullptr, *Step = nullptr;
5678 if (!matchSimpleRecurrence(&PN, BO, Start, Step))
5679 continue;
5680
5681 switch (BO->getOpcode()) {
5682 case Instruction::Sub:
5683 if (BO->getOperand(0) != &PN)
5684 // sub is non-commutative - match handling elsewhere in LSR
5685 continue;
5686 break;
5687 case Instruction::Add:
5688 break;
5689 default:
5690 continue;
5691 };
5692
5693 if (!isa<Constant>(Step))
5694 // If not a constant step, might increase register pressure
5695 // (We assume constants have been canonicalized to RHS)
5696 continue;
5697
5698 if (BO->getParent() == IVIncInsertPos->getParent())
5699 // Only bother moving across blocks. Isel can handle block local case.
5700 continue;
5701
5702 // Can we legally schedule inc at the desired point?
5703 if (!llvm::all_of(BO->uses(),
5704 [&](Use &U) {return DT.dominates(IVIncInsertPos, U);}))
5705 continue;
5706 BO->moveBefore(IVIncInsertPos);
5707 Changed = true;
5708 }
5709
5710
5711}
5712
5713LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5714 DominatorTree &DT, LoopInfo &LI,
5715 const TargetTransformInfo &TTI, AssumptionCache &AC,
5716 TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU)
5717 : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), TLI(TLI), TTI(TTI), L(L),
5718 MSSAU(MSSAU), AMK(PreferredAddresingMode.getNumOccurrences() > 0 ?
5719 PreferredAddresingMode : TTI.getPreferredAddressingMode(L, &SE)) {
5720 // If LoopSimplify form is not available, stay out of trouble.
5721 if (!L->isLoopSimplifyForm())
5722 return;
5723
5724 // If there's no interesting work to be done, bail early.
5725 if (IU.empty()) return;
5726
5727 // If there's too much analysis to be done, bail early. We won't be able to
5728 // model the problem anyway.
5729 unsigned NumUsers = 0;
5730 for (const IVStrideUse &U : IU) {
5731 if (++NumUsers > MaxIVUsers) {
5732 (void)U;
5733 LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
5734 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping loop, too many IV Users in "
<< U << "\n"; } } while (false)
;
5735 return;
5736 }
5737 // Bail out if we have a PHI on an EHPad that gets a value from a
5738 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5739 // no good place to stick any instructions.
5740 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5741 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5742 if (isa<FuncletPadInst>(FirstNonPHI) ||
5743 isa<CatchSwitchInst>(FirstNonPHI))
5744 for (BasicBlock *PredBB : PN->blocks())
5745 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5746 return;
5747 }
5748 }
5749
5750 LLVM_DEBUG(dbgs() << "\nLSR on loop ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
5751 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
5752 dbgs() << ":\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nLSR on loop "; L->getHeader
()->printAsOperand(dbgs(), false); dbgs() << ":\n"; }
} while (false)
;
5753
5754 // First, perform some low-level loop optimizations.
5755 OptimizeShadowIV();
5756 OptimizeLoopTermCond();
5757
5758 // If loop preparation eliminates all interesting IV users, bail.
5759 if (IU.empty()) return;
5760
5761 // Skip nested loops until we can model them better with formulae.
5762 if (!L->isInnermost()) {
5763 LLVM_DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR skipping outer loop "
<< *L << "\n"; } } while (false)
;
5764 return;
5765 }
5766
5767 // Start collecting data and preparing for the solver.
5768 // If number of registers is not the major cost, we cannot benefit from the
5769 // current profitable chain optimization which is based on number of
5770 // registers.
5771 // FIXME: add profitable chain optimization for other kinds major cost, for
5772 // example number of instructions.
5773 if (TTI.isNumRegsMajorCostOfLSR() || StressIVChain)
5774 CollectChains();
5775 CollectInterestingTypesAndFactors();
5776 CollectFixupsAndInitialFormulae();
5777 CollectLoopInvariantFixupsAndFormulae();
5778
5779 if (Uses.empty())
5780 return;
5781
5782 LLVM_DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
5783 print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "LSR found " << Uses
.size() << " uses:\n"; print_uses(dbgs()); } } while (false
)
;
5784
5785 // Now use the reuse data to generate a bunch of interesting ways
5786 // to formulate the values needed for the uses.
5787 GenerateAllReuseFormulae();
5788
5789 FilterOutUndesirableDedicatedRegisters();
5790 NarrowSearchSpaceUsingHeuristics();
5791
5792 SmallVector<const Formula *, 8> Solution;
5793 Solve(Solution);
5794
5795 // Release memory that is no longer needed.
5796 Factors.clear();
5797 Types.clear();
5798 RegUses.clear();
5799
5800 if (Solution.empty())
5801 return;
5802
5803#ifndef NDEBUG
5804 // Formulae should be legal.
5805 for (const LSRUse &LU : Uses) {
5806 for (const Formula &F : LU.Formulae)
5807 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,(static_cast <bool> (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset
, LU.Kind, LU.AccessTy, F) && "Illegal formula generated!"
) ? void (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Illegal formula generated!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5808, __extension__
__PRETTY_FUNCTION__))
5808 F) && "Illegal formula generated!")(static_cast <bool> (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset
, LU.Kind, LU.AccessTy, F) && "Illegal formula generated!"
) ? void (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Illegal formula generated!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5808, __extension__
__PRETTY_FUNCTION__))
;
5809 };
5810#endif
5811
5812 // Now that we've decided what we want, make it so.
5813 ImplementSolution(Solution);
5814}
5815
5816#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5817void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5818 if (Factors.empty() && Types.empty()) return;
5819
5820 OS << "LSR has identified the following interesting factors and types: ";
5821 bool First = true;
5822
5823 for (int64_t Factor : Factors) {
5824 if (!First) OS << ", ";
5825 First = false;
5826 OS << '*' << Factor;
5827 }
5828
5829 for (Type *Ty : Types) {
5830 if (!First) OS << ", ";
5831 First = false;
5832 OS << '(' << *Ty << ')';
5833 }
5834 OS << '\n';
5835}
5836
5837void LSRInstance::print_fixups(raw_ostream &OS) const {
5838 OS << "LSR is examining the following fixup sites:\n";
5839 for (const LSRUse &LU : Uses)
5840 for (const LSRFixup &LF : LU.Fixups) {
5841 dbgs() << " ";
5842 LF.print(OS);
5843 OS << '\n';
5844 }
5845}
5846
5847void LSRInstance::print_uses(raw_ostream &OS) const {
5848 OS << "LSR is examining the following uses:\n";
5849 for (const LSRUse &LU : Uses) {
5850 dbgs() << " ";
5851 LU.print(OS);
5852 OS << '\n';
5853 for (const Formula &F : LU.Formulae) {
5854 OS << " ";
5855 F.print(OS);
5856 OS << '\n';
5857 }
5858 }
5859}
5860
5861void LSRInstance::print(raw_ostream &OS) const {
5862 print_factors_and_types(OS);
5863 print_fixups(OS);
5864 print_uses(OS);
5865}
5866
5867LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRInstance::dump() const {
5868 print(errs()); errs() << '\n';
5869}
5870#endif
5871
5872namespace {
5873
5874class LoopStrengthReduce : public LoopPass {
5875public:
5876 static char ID; // Pass ID, replacement for typeid
5877
5878 LoopStrengthReduce();
5879
5880private:
5881 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5882 void getAnalysisUsage(AnalysisUsage &AU) const override;
5883};
5884
5885} // end anonymous namespace
5886
5887LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5888 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5889}
5890
5891void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5892 // We split critical edges, so we change the CFG. However, we do update
5893 // many analyses if they are around.
5894 AU.addPreservedID(LoopSimplifyID);
5895
5896 AU.addRequired<LoopInfoWrapperPass>();
5897 AU.addPreserved<LoopInfoWrapperPass>();
5898 AU.addRequiredID(LoopSimplifyID);
5899 AU.addRequired<DominatorTreeWrapperPass>();
5900 AU.addPreserved<DominatorTreeWrapperPass>();
5901 AU.addRequired<ScalarEvolutionWrapperPass>();
5902 AU.addPreserved<ScalarEvolutionWrapperPass>();
5903 AU.addRequired<AssumptionCacheTracker>();
5904 AU.addRequired<TargetLibraryInfoWrapperPass>();
5905 // Requiring LoopSimplify a second time here prevents IVUsers from running
5906 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5907 AU.addRequiredID(LoopSimplifyID);
5908 AU.addRequired<IVUsersWrapperPass>();
5909 AU.addPreserved<IVUsersWrapperPass>();
5910 AU.addRequired<TargetTransformInfoWrapperPass>();
5911 AU.addPreserved<MemorySSAWrapperPass>();
5912}
5913
5914namespace {
5915struct SCEVDbgValueBuilder {
5916 SCEVDbgValueBuilder() = default;
5917 SCEVDbgValueBuilder(const SCEVDbgValueBuilder &Base) {
5918 Values = Base.Values;
5919 Expr = Base.Expr;
5920 }
5921
5922 /// The DIExpression as we translate the SCEV.
5923 SmallVector<uint64_t, 6> Expr;
5924 /// The location ops of the DIExpression.
5925 SmallVector<llvm::ValueAsMetadata *, 2> Values;
5926
5927 void pushOperator(uint64_t Op) { Expr.push_back(Op); }
5928 void pushUInt(uint64_t Operand) { Expr.push_back(Operand); }
5929
5930 /// Add a DW_OP_LLVM_arg to the expression, followed by the index of the value
5931 /// in the set of values referenced by the expression.
5932 void pushValue(llvm::Value *V) {
5933 Expr.push_back(llvm::dwarf::DW_OP_LLVM_arg);
5934 auto *It =
5935 std::find(Values.begin(), Values.end(), llvm::ValueAsMetadata::get(V));
5936 unsigned ArgIndex = 0;
5937 if (It != Values.end()) {
5938 ArgIndex = std::distance(Values.begin(), It);
5939 } else {
5940 ArgIndex = Values.size();
5941 Values.push_back(llvm::ValueAsMetadata::get(V));
5942 }
5943 Expr.push_back(ArgIndex);
5944 }
5945
5946 void pushValue(const SCEVUnknown *U) {
5947 llvm::Value *V = cast<SCEVUnknown>(U)->getValue();
5948 pushValue(V);
5949 }
5950
5951 bool pushConst(const SCEVConstant *C) {
5952 if (C->getAPInt().getMinSignedBits() > 64)
5953 return false;
5954 Expr.push_back(llvm::dwarf::DW_OP_consts);
5955 Expr.push_back(C->getAPInt().getSExtValue());
5956 return true;
5957 }
5958
5959 /// Several SCEV types are sequences of the same arithmetic operator applied
5960 /// to constants and values that may be extended or truncated.
5961 bool pushArithmeticExpr(const llvm::SCEVCommutativeExpr *CommExpr,
5962 uint64_t DwarfOp) {
5963 assert((isa<llvm::SCEVAddExpr>(CommExpr) || isa<SCEVMulExpr>(CommExpr)) &&(static_cast <bool> ((isa<llvm::SCEVAddExpr>(CommExpr
) || isa<SCEVMulExpr>(CommExpr)) && "Expected arithmetic SCEV type"
) ? void (0) : __assert_fail ("(isa<llvm::SCEVAddExpr>(CommExpr) || isa<SCEVMulExpr>(CommExpr)) && \"Expected arithmetic SCEV type\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5964, __extension__
__PRETTY_FUNCTION__))
5964 "Expected arithmetic SCEV type")(static_cast <bool> ((isa<llvm::SCEVAddExpr>(CommExpr
) || isa<SCEVMulExpr>(CommExpr)) && "Expected arithmetic SCEV type"
) ? void (0) : __assert_fail ("(isa<llvm::SCEVAddExpr>(CommExpr) || isa<SCEVMulExpr>(CommExpr)) && \"Expected arithmetic SCEV type\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5964, __extension__
__PRETTY_FUNCTION__))
;
5965 bool Success = true;
5966 unsigned EmitOperator = 0;
5967 for (auto &Op : CommExpr->operands()) {
5968 Success &= pushSCEV(Op);
5969
5970 if (EmitOperator >= 1)
5971 pushOperator(DwarfOp);
5972 ++EmitOperator;
5973 }
5974 return Success;
5975 }
5976
5977 // TODO: Identify and omit noop casts.
5978 bool pushCast(const llvm::SCEVCastExpr *C, bool IsSigned) {
5979 const llvm::SCEV *Inner = C->getOperand(0);
5980 const llvm::Type *Type = C->getType();
5981 uint64_t ToWidth = Type->getIntegerBitWidth();
5982 bool Success = pushSCEV(Inner);
5983 uint64_t CastOps[] = {dwarf::DW_OP_LLVM_convert, ToWidth,
5984 IsSigned ? llvm::dwarf::DW_ATE_signed
5985 : llvm::dwarf::DW_ATE_unsigned};
5986 for (const auto &Op : CastOps)
5987 pushOperator(Op);
5988 return Success;
5989 }
5990
5991 // TODO: MinMax - although these haven't been encountered in the test suite.
5992 bool pushSCEV(const llvm::SCEV *S) {
5993 bool Success = true;
5994 if (const SCEVConstant *StartInt = dyn_cast<SCEVConstant>(S)) {
5995 Success &= pushConst(StartInt);
5996
5997 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
5998 if (!U->getValue())
5999 return false;
6000 pushValue(U->getValue());
6001
6002 } else if (const SCEVMulExpr *MulRec = dyn_cast<SCEVMulExpr>(S)) {
6003 Success &= pushArithmeticExpr(MulRec, llvm::dwarf::DW_OP_mul);
6004
6005 } else if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
6006 Success &= pushSCEV(UDiv->getLHS());
6007 Success &= pushSCEV(UDiv->getRHS());
6008 pushOperator(llvm::dwarf::DW_OP_div);
6009
6010 } else if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(S)) {
6011 // Assert if a new and unknown SCEVCastEXpr type is encountered.
6012 assert((isa<SCEVZeroExtendExpr>(Cast) || isa<SCEVTruncateExpr>(Cast) ||(static_cast <bool> ((isa<SCEVZeroExtendExpr>(Cast
) || isa<SCEVTruncateExpr>(Cast) || isa<SCEVPtrToIntExpr
>(Cast) || isa<SCEVSignExtendExpr>(Cast)) &&
"Unexpected cast type in SCEV.") ? void (0) : __assert_fail (
"(isa<SCEVZeroExtendExpr>(Cast) || isa<SCEVTruncateExpr>(Cast) || isa<SCEVPtrToIntExpr>(Cast) || isa<SCEVSignExtendExpr>(Cast)) && \"Unexpected cast type in SCEV.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6014, __extension__
__PRETTY_FUNCTION__))
6013 isa<SCEVPtrToIntExpr>(Cast) || isa<SCEVSignExtendExpr>(Cast)) &&(static_cast <bool> ((isa<SCEVZeroExtendExpr>(Cast
) || isa<SCEVTruncateExpr>(Cast) || isa<SCEVPtrToIntExpr
>(Cast) || isa<SCEVSignExtendExpr>(Cast)) &&
"Unexpected cast type in SCEV.") ? void (0) : __assert_fail (
"(isa<SCEVZeroExtendExpr>(Cast) || isa<SCEVTruncateExpr>(Cast) || isa<SCEVPtrToIntExpr>(Cast) || isa<SCEVSignExtendExpr>(Cast)) && \"Unexpected cast type in SCEV.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6014, __extension__
__PRETTY_FUNCTION__))
6014 "Unexpected cast type in SCEV.")(static_cast <bool> ((isa<SCEVZeroExtendExpr>(Cast
) || isa<SCEVTruncateExpr>(Cast) || isa<SCEVPtrToIntExpr
>(Cast) || isa<SCEVSignExtendExpr>(Cast)) &&
"Unexpected cast type in SCEV.") ? void (0) : __assert_fail (
"(isa<SCEVZeroExtendExpr>(Cast) || isa<SCEVTruncateExpr>(Cast) || isa<SCEVPtrToIntExpr>(Cast) || isa<SCEVSignExtendExpr>(Cast)) && \"Unexpected cast type in SCEV.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6014, __extension__
__PRETTY_FUNCTION__))
;
6015 Success &= pushCast(Cast, (isa<SCEVSignExtendExpr>(Cast)));
6016
6017 } else if (const SCEVAddExpr *AddExpr = dyn_cast<SCEVAddExpr>(S)) {
6018 Success &= pushArithmeticExpr(AddExpr, llvm::dwarf::DW_OP_plus);
6019
6020 } else if (isa<SCEVAddRecExpr>(S)) {
6021 // Nested SCEVAddRecExpr are generated by nested loops and are currently
6022 // unsupported.
6023 return false;
6024
6025 } else {
6026 return false;
6027 }
6028 return Success;
6029 }
6030
6031 void setFinalExpression(llvm::DbgValueInst &DI, const DIExpression *OldExpr) {
6032 // Re-state assumption that this dbg.value is not variadic. Any remaining
6033 // opcodes in its expression operate on a single value already on the
6034 // expression stack. Prepend our operations, which will re-compute and
6035 // place that value on the expression stack.
6036 assert(!DI.hasArgList())(static_cast <bool> (!DI.hasArgList()) ? void (0) : __assert_fail
("!DI.hasArgList()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 6036, __extension__ __PRETTY_FUNCTION__))
;
6037 auto *NewExpr =
6038 DIExpression::prependOpcodes(OldExpr, Expr, /*StackValue*/ true);
6039 DI.setExpression(NewExpr);
6040
6041 auto ValArrayRef = llvm::ArrayRef<llvm::ValueAsMetadata *>(Values);
6042 DI.setRawLocation(llvm::DIArgList::get(DI.getContext(), ValArrayRef));
6043 }
6044
6045 /// If a DVI can be emitted without a DIArgList, omit DW_OP_llvm_arg and the
6046 /// location op index 0.
6047 void setShortFinalExpression(llvm::DbgValueInst &DI,
6048 const DIExpression *OldExpr) {
6049 assert((Expr[0] == llvm::dwarf::DW_OP_LLVM_arg && Expr[1] == 0) &&(static_cast <bool> ((Expr[0] == llvm::dwarf::DW_OP_LLVM_arg
&& Expr[1] == 0) && "Expected DW_OP_llvm_arg and 0."
) ? void (0) : __assert_fail ("(Expr[0] == llvm::dwarf::DW_OP_LLVM_arg && Expr[1] == 0) && \"Expected DW_OP_llvm_arg and 0.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6050, __extension__
__PRETTY_FUNCTION__))
6050 "Expected DW_OP_llvm_arg and 0.")(static_cast <bool> ((Expr[0] == llvm::dwarf::DW_OP_LLVM_arg
&& Expr[1] == 0) && "Expected DW_OP_llvm_arg and 0."
) ? void (0) : __assert_fail ("(Expr[0] == llvm::dwarf::DW_OP_LLVM_arg && Expr[1] == 0) && \"Expected DW_OP_llvm_arg and 0.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6050, __extension__
__PRETTY_FUNCTION__))
;
6051 DI.replaceVariableLocationOp(
6052 0u, llvm::MetadataAsValue::get(DI.getContext(), Values[0]));
6053
6054 // See setFinalExpression: prepend our opcodes on the start of any old
6055 // expression opcodes.
6056 assert(!DI.hasArgList())(static_cast <bool> (!DI.hasArgList()) ? void (0) : __assert_fail
("!DI.hasArgList()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 6056, __extension__ __PRETTY_FUNCTION__))
;
6057 llvm::SmallVector<uint64_t, 6> FinalExpr(llvm::drop_begin(Expr, 2));
6058 auto *NewExpr =
6059 DIExpression::prependOpcodes(OldExpr, FinalExpr, /*StackValue*/ true);
6060 DI.setExpression(NewExpr);
6061 }
6062
6063 /// Once the IV and variable SCEV translation is complete, write it to the
6064 /// source DVI.
6065 void applyExprToDbgValue(llvm::DbgValueInst &DI,
6066 const DIExpression *OldExpr) {
6067 assert(!Expr.empty() && "Unexpected empty expression.")(static_cast <bool> (!Expr.empty() && "Unexpected empty expression."
) ? void (0) : __assert_fail ("!Expr.empty() && \"Unexpected empty expression.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6067, __extension__
__PRETTY_FUNCTION__))
;
6068 // Emit a simpler form if only a single location is referenced.
6069 if (Values.size() == 1 && Expr[0] == llvm::dwarf::DW_OP_LLVM_arg &&
6070 Expr[1] == 0) {
6071 setShortFinalExpression(DI, OldExpr);
6072 } else {
6073 setFinalExpression(DI, OldExpr);
6074 }
6075 }
6076
6077 /// Return true if the combination of arithmetic operator and underlying
6078 /// SCEV constant value is an identity function.
6079 bool isIdentityFunction(uint64_t Op, const SCEV *S) {
6080 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
6081 if (C->getAPInt().getMinSignedBits() > 64)
6082 return false;
6083 int64_t I = C->getAPInt().getSExtValue();
6084 switch (Op) {
6085 case llvm::dwarf::DW_OP_plus:
6086 case llvm::dwarf::DW_OP_minus:
6087 return I == 0;
6088 case llvm::dwarf::DW_OP_mul:
6089 case llvm::dwarf::DW_OP_div:
6090 return I == 1;
6091 }
6092 }
6093 return false;
6094 }
6095
6096 /// Convert a SCEV of a value to a DIExpression that is pushed onto the
6097 /// builder's expression stack. The stack should already contain an
6098 /// expression for the iteration count, so that it can be multiplied by
6099 /// the stride and added to the start.
6100 /// Components of the expression are omitted if they are an identity function.
6101 /// Chain (non-affine) SCEVs are not supported.
6102 bool SCEVToValueExpr(const llvm::SCEVAddRecExpr &SAR, ScalarEvolution &SE) {
6103 assert(SAR.isAffine() && "Expected affine SCEV")(static_cast <bool> (SAR.isAffine() && "Expected affine SCEV"
) ? void (0) : __assert_fail ("SAR.isAffine() && \"Expected affine SCEV\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6103, __extension__
__PRETTY_FUNCTION__))
;
6104 // TODO: Is this check needed?
6105 if (isa<SCEVAddRecExpr>(SAR.getStart()))
6106 return false;
6107
6108 const SCEV *Start = SAR.getStart();
6109 const SCEV *Stride = SAR.getStepRecurrence(SE);
6110
6111 // Skip pushing arithmetic noops.
6112 if (!isIdentityFunction(llvm::dwarf::DW_OP_mul, Stride)) {
6113 if (!pushSCEV(Stride))
6114 return false;
6115 pushOperator(llvm::dwarf::DW_OP_mul);
6116 }
6117 if (!isIdentityFunction(llvm::dwarf::DW_OP_plus, Start)) {
6118 if (!pushSCEV(Start))
6119 return false;
6120 pushOperator(llvm::dwarf::DW_OP_plus);
6121 }
6122 return true;
6123 }
6124
6125 /// Convert a SCEV of a value to a DIExpression that is pushed onto the
6126 /// builder's expression stack. The stack should already contain an
6127 /// expression for the iteration count, so that it can be multiplied by
6128 /// the stride and added to the start.
6129 /// Components of the expression are omitted if they are an identity function.
6130 bool SCEVToIterCountExpr(const llvm::SCEVAddRecExpr &SAR,
6131 ScalarEvolution &SE) {
6132 assert(SAR.isAffine() && "Expected affine SCEV")(static_cast <bool> (SAR.isAffine() && "Expected affine SCEV"
) ? void (0) : __assert_fail ("SAR.isAffine() && \"Expected affine SCEV\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6132, __extension__
__PRETTY_FUNCTION__))
;
6133 if (isa<SCEVAddRecExpr>(SAR.getStart())) {
6134 LLVM_DEBUG(dbgs() << "scev-salvage: IV SCEV. Unsupported nested AddRec: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: IV SCEV. Unsupported nested AddRec: "
<< SAR << '\n'; } } while (false)
6135 << SAR << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: IV SCEV. Unsupported nested AddRec: "
<< SAR << '\n'; } } while (false)
;
6136 return false;
6137 }
6138 const SCEV *Start = SAR.getStart();
6139 const SCEV *Stride = SAR.getStepRecurrence(SE);
6140
6141 // Skip pushing arithmetic noops.
6142 if (!isIdentityFunction(llvm::dwarf::DW_OP_minus, Start)) {
6143 if (!pushSCEV(Start))
6144 return false;
6145 pushOperator(llvm::dwarf::DW_OP_minus);
6146 }
6147 if (!isIdentityFunction(llvm::dwarf::DW_OP_div, Stride)) {
6148 if (!pushSCEV(Stride))
6149 return false;
6150 pushOperator(llvm::dwarf::DW_OP_div);
6151 }
6152 return true;
6153 }
6154};
6155
6156struct DVIRecoveryRec {
6157 DbgValueInst *DVI;
6158 DIExpression *Expr;
6159 Metadata *LocationOp;
6160 const llvm::SCEV *SCEV;
6161};
6162} // namespace
6163
6164static void RewriteDVIUsingIterCount(DVIRecoveryRec CachedDVI,
6165 const SCEVDbgValueBuilder &IterationCount,
6166 ScalarEvolution &SE) {
6167 // LSR may add locations to previously single location-op DVIs which
6168 // are currently not supported.
6169 if (CachedDVI.DVI->getNumVariableLocationOps() != 1)
6170 return;
6171
6172 // SCEVs for SSA values are most frquently of the form
6173 // {start,+,stride}, but sometimes they are ({start,+,stride} + %a + ..).
6174 // This is because %a is a PHI node that is not the IV. However, these
6175 // SCEVs have not been observed to result in debuginfo-lossy optimisations,
6176 // so its not expected this point will be reached.
6177 if (!isa<SCEVAddRecExpr>(CachedDVI.SCEV))
6178 return;
6179
6180 LLVM_DEBUG(dbgs() << "scev-salvage: Value to salvage SCEV: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: Value to salvage SCEV: "
<< *CachedDVI.SCEV << '\n'; } } while (false)
6181 << *CachedDVI.SCEV << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: Value to salvage SCEV: "
<< *CachedDVI.SCEV << '\n'; } } while (false)
;
6182
6183 const auto *Rec = cast<SCEVAddRecExpr>(CachedDVI.SCEV);
6184 if (!Rec->isAffine())
6185 return;
6186
6187 if (CachedDVI.SCEV->getExpressionSize() > MaxSCEVSalvageExpressionSize)
6188 return;
6189
6190 // Initialise a new builder with the iteration count expression. In
6191 // combination with the value's SCEV this enables recovery.
6192 SCEVDbgValueBuilder RecoverValue(IterationCount);
6193 if (!RecoverValue.SCEVToValueExpr(*Rec, SE))
6194 return;
6195
6196 LLVM_DEBUG(dbgs() << "scev-salvage: Updating: " << *CachedDVI.DVI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: Updating: "
<< *CachedDVI.DVI << '\n'; } } while (false)
;
6197 RecoverValue.applyExprToDbgValue(*CachedDVI.DVI, CachedDVI.Expr);
6198 LLVM_DEBUG(dbgs() << "scev-salvage: to: " << *CachedDVI.DVI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: to: " <<
*CachedDVI.DVI << '\n'; } } while (false)
;
6199}
6200
6201static void RewriteDVIUsingOffset(DVIRecoveryRec &DVIRec, llvm::PHINode &IV,
6202 int64_t Offset) {
6203 assert(!DVIRec.DVI->hasArgList() && "Expected single location-op dbg.value.")(static_cast <bool> (!DVIRec.DVI->hasArgList() &&
"Expected single location-op dbg.value.") ? void (0) : __assert_fail
("!DVIRec.DVI->hasArgList() && \"Expected single location-op dbg.value.\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6203, __extension__
__PRETTY_FUNCTION__))
;
6204 DbgValueInst *DVI = DVIRec.DVI;
6205 SmallVector<uint64_t, 8> Ops;
6206 DIExpression::appendOffset(Ops, Offset);
6207 DIExpression *Expr = DIExpression::prependOpcodes(DVIRec.Expr, Ops, true);
6208 LLVM_DEBUG(dbgs() << "scev-salvage: Updating: " << *DVIRec.DVI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: Updating: "
<< *DVIRec.DVI << '\n'; } } while (false)
;
6209 DVI->setExpression(Expr);
6210 llvm::Value *ValIV = dyn_cast<llvm::Value>(&IV);
6211 DVI->replaceVariableLocationOp(
6212 0u, llvm::MetadataAsValue::get(DVI->getContext(),
6213 llvm::ValueAsMetadata::get(ValIV)));
6214 LLVM_DEBUG(dbgs() << "scev-salvage: updated with offset to IV: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: updated with offset to IV: "
<< *DVIRec.DVI << '\n'; } } while (false)
6215 << *DVIRec.DVI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: updated with offset to IV: "
<< *DVIRec.DVI << '\n'; } } while (false)
;
6216}
6217
6218static void
6219DbgRewriteSalvageableDVIs(llvm::Loop *L, ScalarEvolution &SE,
6220 llvm::PHINode *LSRInductionVar,
6221 SmallVector<DVIRecoveryRec, 2> &DVIToUpdate) {
6222 if (DVIToUpdate.empty())
6223 return;
6224
6225 const llvm::SCEV *SCEVInductionVar = SE.getSCEV(LSRInductionVar);
6226 assert(SCEVInductionVar &&(static_cast <bool> (SCEVInductionVar && "Anticipated a SCEV for the post-LSR induction variable"
) ? void (0) : __assert_fail ("SCEVInductionVar && \"Anticipated a SCEV for the post-LSR induction variable\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6227, __extension__
__PRETTY_FUNCTION__))
6227 "Anticipated a SCEV for the post-LSR induction variable")(static_cast <bool> (SCEVInductionVar && "Anticipated a SCEV for the post-LSR induction variable"
) ? void (0) : __assert_fail ("SCEVInductionVar && \"Anticipated a SCEV for the post-LSR induction variable\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 6227, __extension__
__PRETTY_FUNCTION__))
;
6228
6229 if (const SCEVAddRecExpr *IVAddRec =
6230 dyn_cast<SCEVAddRecExpr>(SCEVInductionVar)) {
6231 if (!IVAddRec->isAffine())
6232 return;
6233
6234 if (IVAddRec->getExpressionSize() > MaxSCEVSalvageExpressionSize)
6235 return;
6236
6237 // The iteration count is required to recover location values.
6238 SCEVDbgValueBuilder IterCountExpr;
6239 IterCountExpr.pushValue(LSRInductionVar);
6240 if (!IterCountExpr.SCEVToIterCountExpr(*IVAddRec, SE))
6241 return;
6242
6243 LLVM_DEBUG(dbgs() << "scev-salvage: IV SCEV: " << *SCEVInductionVardo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: IV SCEV: " <<
*SCEVInductionVar << '\n'; } } while (false)
6244 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: IV SCEV: " <<
*SCEVInductionVar << '\n'; } } while (false)
;
6245
6246 // Needn't salvage if the location op hasn't been undef'd by LSR.
6247 for (auto &DVIRec : DVIToUpdate) {
6248 if (!DVIRec.DVI->isUndef())
6249 continue;
6250
6251 // Some DVIs that were single location-op when cached are now multi-op,
6252 // due to LSR optimisations. However, multi-op salvaging is not yet
6253 // supported by SCEV salvaging. But, we can attempt a salvage by restoring
6254 // the pre-LSR single-op expression.
6255 if (DVIRec.DVI->hasArgList()) {
6256 if (!DVIRec.DVI->getVariableLocationOp(0))
6257 continue;
6258 llvm::Type *Ty = DVIRec.DVI->getVariableLocationOp(0)->getType();
6259 DVIRec.DVI->setRawLocation(
6260 llvm::ValueAsMetadata::get(UndefValue::get(Ty)));
6261 DVIRec.DVI->setExpression(DVIRec.Expr);
6262 }
6263
6264 LLVM_DEBUG(dbgs() << "scev-salvage: value to recover SCEV: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: value to recover SCEV: "
<< *DVIRec.SCEV << '\n'; } } while (false)
6265 << *DVIRec.SCEV << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: value to recover SCEV: "
<< *DVIRec.SCEV << '\n'; } } while (false)
;
6266
6267 // Create a simple expression if the IV and value to salvage SCEVs
6268 // start values differ by only a constant value.
6269 if (Optional<APInt> Offset =
6270 SE.computeConstantDifference(DVIRec.SCEV, SCEVInductionVar)) {
6271 if (Offset.getValue().getMinSignedBits() <= 64)
6272 RewriteDVIUsingOffset(DVIRec, *LSRInductionVar,
6273 Offset.getValue().getSExtValue());
6274 } else {
6275 RewriteDVIUsingIterCount(DVIRec, IterCountExpr, SE);
6276 }
6277 }
6278 }
6279}
6280
6281/// Identify and cache salvageable DVI locations and expressions along with the
6282/// corresponding SCEV(s). Also ensure that the DVI is not deleted between
6283/// cacheing and salvaging.
6284static void
6285DbgGatherSalvagableDVI(Loop *L, ScalarEvolution &SE,
6286 SmallVector<DVIRecoveryRec, 2> &SalvageableDVISCEVs,
6287 SmallSet<AssertingVH<DbgValueInst>, 2> &DVIHandles) {
6288 for (auto &B : L->getBlocks()) {
6289 for (auto &I : *B) {
6290 auto DVI = dyn_cast<DbgValueInst>(&I);
6291 if (!DVI)
6292 continue;
6293
6294 if (DVI->isUndef())
6295 continue;
6296
6297 if (DVI->hasArgList())
6298 continue;
6299
6300 if (!DVI->getVariableLocationOp(0) ||
6301 !SE.isSCEVable(DVI->getVariableLocationOp(0)->getType()))
6302 continue;
6303
6304 // SCEVUnknown wraps an llvm::Value, it does not have a start and stride.
6305 // Therefore no translation to DIExpression is performed.
6306 const SCEV *S = SE.getSCEV(DVI->getVariableLocationOp(0));
6307 if (isa<SCEVUnknown>(S))
6308 continue;
6309
6310 // Avoid wasting resources generating an expression containing undef.
6311 if (SE.containsUndefs(S))
6312 continue;
6313
6314 SalvageableDVISCEVs.push_back(
6315 {DVI, DVI->getExpression(), DVI->getRawLocation(),
6316 SE.getSCEV(DVI->getVariableLocationOp(0))});
6317 DVIHandles.insert(DVI);
6318 }
6319 }
6320}
6321
6322/// Ideally pick the PHI IV inserted by ScalarEvolutionExpander. As a fallback
6323/// any PHi from the loop header is usable, but may have less chance of
6324/// surviving subsequent transforms.
6325static llvm::PHINode *GetInductionVariable(const Loop &L, ScalarEvolution &SE,
6326 const LSRInstance &LSR) {
6327
6328 auto IsSuitableIV = [&](PHINode *P) {
6329 if (!SE.isSCEVable(P->getType()))
6330 return false;
6331 if (const SCEVAddRecExpr *Rec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(P)))
6332 return Rec->isAffine() && !SE.containsUndefs(SE.getSCEV(P));
6333 return false;
6334 };
6335
6336 // For now, just pick the first IV that was generated and inserted by
6337 // ScalarEvolution. Ideally pick an IV that is unlikely to be optimised away
6338 // by subsequent transforms.
6339 for (const WeakVH &IV : LSR.getScalarEvolutionIVs()) {
6340 if (!IV)
6341 continue;
6342
6343 // There should only be PHI node IVs.
6344 PHINode *P = cast<PHINode>(&*IV);
6345
6346 if (IsSuitableIV(P))
6347 return P;
6348 }
6349
6350 for (PHINode &P : L.getHeader()->phis()) {
6351 if (IsSuitableIV(&P))
6352 return &P;
6353 }
6354 return nullptr;
6355}
6356
6357static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
6358 DominatorTree &DT, LoopInfo &LI,
6359 const TargetTransformInfo &TTI,
6360 AssumptionCache &AC, TargetLibraryInfo &TLI,
6361 MemorySSA *MSSA) {
6362
6363 // Debug preservation - before we start removing anything identify which DVI
6364 // meet the salvageable criteria and store their DIExpression and SCEVs.
6365 SmallVector<DVIRecoveryRec, 2> SalvageableDVI;
6366 SmallSet<AssertingVH<DbgValueInst>, 2> DVIHandles;
6367 DbgGatherSalvagableDVI(L, SE, SalvageableDVI, DVIHandles);
6368
6369 bool Changed = false;
6370 std::unique_ptr<MemorySSAUpdater> MSSAU;
6371 if (MSSA)
6372 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
6373
6374 // Run the main LSR transformation.
6375 const LSRInstance &Reducer =
6376 LSRInstance(L, IU, SE, DT, LI, TTI, AC, TLI, MSSAU.get());
6377 Changed |= Reducer.getChanged();
6378
6379 // Remove any extra phis created by processing inner loops.
6380 Changed |= DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
6381 if (EnablePhiElim && L->isLoopSimplifyForm()) {
6382 SmallVector<WeakTrackingVH, 16> DeadInsts;
6383 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
6384 SCEVExpander Rewriter(SE, DL, "lsr", false);
6385#ifndef NDEBUG
6386 Rewriter.setDebugType(DEBUG_TYPE"loop-reduce");
6387#endif
6388 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
6389 if (numFolded) {
6390 Changed = true;
6391 RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, &TLI,
6392 MSSAU.get());
6393 DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
6394 }
6395 }
6396 // LSR may at times remove all uses of an induction variable from a loop.
6397 // The only remaining use is the PHI in the exit block.
6398 // When this is the case, if the exit value of the IV can be calculated using
6399 // SCEV, we can replace the exit block PHI with the final value of the IV and
6400 // skip the updates in each loop iteration.
6401 if (L->isRecursivelyLCSSAForm(DT, LI) && LoopExitValHasSingleUse(L)) {
6402 SmallVector<WeakTrackingVH, 16> DeadInsts;
6403 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
6404 SCEVExpander Rewriter(SE, DL, "lsr", false);
6405 int Rewrites = rewriteLoopExitValues(L, &LI, &TLI, &SE, &TTI, Rewriter, &DT,
6406 OnlyCheapRepl, DeadInsts);
6407 if (Rewrites) {
6408 Changed = true;
6409 RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, &TLI,
6410 MSSAU.get());
6411 DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
6412 }
6413 }
6414
6415 if (SalvageableDVI.empty())
6416 return Changed;
6417
6418 // Obtain relevant IVs and attempt to rewrite the salvageable DVIs with
6419 // expressions composed using the derived iteration count.
6420 // TODO: Allow for multiple IV references for nested AddRecSCEVs
6421 for (auto &L : LI) {
6422 if (llvm::PHINode *IV = GetInductionVariable(*L, SE, Reducer))
6423 DbgRewriteSalvageableDVIs(L, SE, IV, SalvageableDVI);
6424 else {
6425 LLVM_DEBUG(dbgs() << "scev-salvage: SCEV salvaging not possible. An IV "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: SCEV salvaging not possible. An IV "
"could not be identified.\n"; } } while (false)
6426 "could not be identified.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "scev-salvage: SCEV salvaging not possible. An IV "
"could not be identified.\n"; } } while (false)
;
6427 }
6428 }
6429
6430 DVIHandles.clear();
6431 return Changed;
6432}
6433
6434bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
6435 if (skipLoop(L))
6436 return false;
6437
6438 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
6439 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
6440 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
6441 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
6442 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
6443 *L->getHeader()->getParent());
6444 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
6445 *L->getHeader()->getParent());
6446 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
6447 *L->getHeader()->getParent());
6448 auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>();
6449 MemorySSA *MSSA = nullptr;
6450 if (MSSAAnalysis)
6451 MSSA = &MSSAAnalysis->getMSSA();
6452 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI, AC, TLI, MSSA);
6453}
6454
6455PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
6456 LoopStandardAnalysisResults &AR,
6457 LPMUpdater &) {
6458 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
6459 AR.DT, AR.LI, AR.TTI, AR.AC, AR.TLI, AR.MSSA))
6460 return PreservedAnalyses::all();
6461
6462 auto PA = getLoopPassPreservedAnalyses();
6463 if (AR.MSSA)
6464 PA.preserve<MemorySSAAnalysis>();
6465 return PA;
6466}
6467
6468char LoopStrengthReduce::ID = 0;
6469
6470INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
6471 "Loop Strength Reduction", false, false)static void *initializeLoopStrengthReducePassOnce(PassRegistry
&Registry) {
6472INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
6473INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
6474INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
6475INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)initializeIVUsersWrapperPassPass(Registry);
6476INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
6477INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
6478INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",PassInfo *PI = new PassInfo( "Loop Strength Reduction", "loop-reduce"
, &LoopStrengthReduce::ID, PassInfo::NormalCtor_t(callDefaultCtor
<LoopStrengthReduce>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeLoopStrengthReducePassFlag
; void llvm::initializeLoopStrengthReducePass(PassRegistry &
Registry) { llvm::call_once(InitializeLoopStrengthReducePassFlag
, initializeLoopStrengthReducePassOnce, std::ref(Registry)); }
6479 "Loop Strength Reduction", false, false)PassInfo *PI = new PassInfo( "Loop Strength Reduction", "loop-reduce"
, &LoopStrengthReduce::ID, PassInfo::NormalCtor_t(callDefaultCtor
<LoopStrengthReduce>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeLoopStrengthReducePassFlag
; void llvm::initializeLoopStrengthReducePass(PassRegistry &
Registry) { llvm::call_once(InitializeLoopStrengthReducePassFlag
, initializeLoopStrengthReducePassOnce, std::ref(Registry)); }
6480
6481Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }