1 //===---- MachineOutliner.cpp - Outline instructions -----------*- C++ -*-===//
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 /// \file
10 /// Replaces repeated sequences of instructions with function calls.
11 ///
12 /// This works by placing every instruction from every basic block in a
13 /// suffix tree, and repeatedly querying that tree for repeated sequences of
14 /// instructions. If a sequence of instructions appears often, then it ought
15 /// to be beneficial to pull out into a function.
16 ///
17 /// The MachineOutliner communicates with a given target using hooks defined in
18 /// TargetInstrInfo.h. The target supplies the outliner with information on how
19 /// a specific sequence of instructions should be outlined. This information
20 /// is used to deduce the number of instructions necessary to
21 ///
22 /// * Create an outlined function
23 /// * Call that outlined function
24 ///
25 /// Targets must implement
26 /// * getOutliningCandidateInfo
27 /// * buildOutlinedFrame
28 /// * insertOutlinedCall
29 /// * isFunctionSafeToOutlineFrom
30 ///
31 /// in order to make use of the MachineOutliner.
32 ///
33 /// This was originally presented at the 2016 LLVM Developers' Meeting in the
34 /// talk "Reducing Code Size Using Outlining". For a high-level overview of
35 /// how this pass works, the talk is available on YouTube at
36 ///
37 /// https://www.youtube.com/watch?v=yorld-WSOeU
38 ///
39 /// The slides for the talk are available at
40 ///
41 /// http://www.llvm.org/devmtg/2016-11/Slides/Paquette-Outliner.pdf
42 ///
43 /// The talk provides an overview of how the outliner finds candidates and
44 /// ultimately outlines them. It describes how the main data structure for this
45 /// pass, the suffix tree, is queried and purged for candidates. It also gives
46 /// a simplified suffix tree construction algorithm for suffix trees based off
47 /// of the algorithm actually used here, Ukkonen's algorithm.
48 ///
49 /// For the original RFC for this pass, please see
50 ///
51 /// http://lists.llvm.org/pipermail/llvm-dev/2016-August/104170.html
52 ///
53 /// For more information on the suffix tree data structure, please see
54 /// https://www.cs.helsinki.fi/u/ukkonen/SuffixT1withFigs.pdf
55 ///
56 //===----------------------------------------------------------------------===//
57 #include "llvm/CodeGen/MachineOutliner.h"
58 #include "llvm/ADT/DenseMap.h"
59 #include "llvm/ADT/SmallSet.h"
60 #include "llvm/ADT/Statistic.h"
61 #include "llvm/ADT/Twine.h"
62 #include "llvm/CodeGen/MachineModuleInfo.h"
63 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
64 #include "llvm/CodeGen/Passes.h"
65 #include "llvm/CodeGen/TargetInstrInfo.h"
66 #include "llvm/CodeGen/TargetSubtargetInfo.h"
67 #include "llvm/IR/DIBuilder.h"
68 #include "llvm/IR/IRBuilder.h"
69 #include "llvm/IR/Mangler.h"
70 #include "llvm/InitializePasses.h"
71 #include "llvm/Support/CommandLine.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/SuffixTree.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include <functional>
76 #include <tuple>
77 #include <vector>
78
79 #define DEBUG_TYPE "machine-outliner"
80
81 using namespace llvm;
82 using namespace ore;
83 using namespace outliner;
84
85 STATISTIC(NumOutlined, "Number of candidates outlined");
86 STATISTIC(FunctionsCreated, "Number of functions created");
87
88 // Set to true if the user wants the outliner to run on linkonceodr linkage
89 // functions. This is false by default because the linker can dedupe linkonceodr
90 // functions. Since the outliner is confined to a single module (modulo LTO),
91 // this is off by default. It should, however, be the default behaviour in
92 // LTO.
93 static cl::opt<bool> EnableLinkOnceODROutlining(
94 "enable-linkonceodr-outlining", cl::Hidden,
95 cl::desc("Enable the machine outliner on linkonceodr functions"),
96 cl::init(false));
97
98 /// Number of times to re-run the outliner. This is not the total number of runs
99 /// as the outliner will run at least one time. The default value is set to 0,
100 /// meaning the outliner will run one time and rerun zero times after that.
101 static cl::opt<unsigned> OutlinerReruns(
102 "machine-outliner-reruns", cl::init(0), cl::Hidden,
103 cl::desc(
104 "Number of times to rerun the outliner after the initial outline"));
105
106 namespace {
107
108 /// Maps \p MachineInstrs to unsigned integers and stores the mappings.
109 struct InstructionMapper {
110
111 /// The next available integer to assign to a \p MachineInstr that
112 /// cannot be outlined.
113 ///
114 /// Set to -3 for compatability with \p DenseMapInfo<unsigned>.
115 unsigned IllegalInstrNumber = -3;
116
117 /// The next available integer to assign to a \p MachineInstr that can
118 /// be outlined.
119 unsigned LegalInstrNumber = 0;
120
121 /// Correspondence from \p MachineInstrs to unsigned integers.
122 DenseMap<MachineInstr *, unsigned, MachineInstrExpressionTrait>
123 InstructionIntegerMap;
124
125 /// Correspondence between \p MachineBasicBlocks and target-defined flags.
126 DenseMap<MachineBasicBlock *, unsigned> MBBFlagsMap;
127
128 /// The vector of unsigned integers that the module is mapped to.
129 std::vector<unsigned> UnsignedVec;
130
131 /// Stores the location of the instruction associated with the integer
132 /// at index i in \p UnsignedVec for each index i.
133 std::vector<MachineBasicBlock::iterator> InstrList;
134
135 // Set if we added an illegal number in the previous step.
136 // Since each illegal number is unique, we only need one of them between
137 // each range of legal numbers. This lets us make sure we don't add more
138 // than one illegal number per range.
139 bool AddedIllegalLastTime = false;
140
141 /// Maps \p *It to a legal integer.
142 ///
143 /// Updates \p CanOutlineWithPrevInstr, \p HaveLegalRange, \p InstrListForMBB,
144 /// \p UnsignedVecForMBB, \p InstructionIntegerMap, and \p LegalInstrNumber.
145 ///
146 /// \returns The integer that \p *It was mapped to.
mapToLegalUnsigned__anon3e97bd940111::InstructionMapper147 unsigned mapToLegalUnsigned(
148 MachineBasicBlock::iterator &It, bool &CanOutlineWithPrevInstr,
149 bool &HaveLegalRange, unsigned &NumLegalInBlock,
150 std::vector<unsigned> &UnsignedVecForMBB,
151 std::vector<MachineBasicBlock::iterator> &InstrListForMBB) {
152 // We added something legal, so we should unset the AddedLegalLastTime
153 // flag.
154 AddedIllegalLastTime = false;
155
156 // If we have at least two adjacent legal instructions (which may have
157 // invisible instructions in between), remember that.
158 if (CanOutlineWithPrevInstr)
159 HaveLegalRange = true;
160 CanOutlineWithPrevInstr = true;
161
162 // Keep track of the number of legal instructions we insert.
163 NumLegalInBlock++;
164
165 // Get the integer for this instruction or give it the current
166 // LegalInstrNumber.
167 InstrListForMBB.push_back(It);
168 MachineInstr &MI = *It;
169 bool WasInserted;
170 DenseMap<MachineInstr *, unsigned, MachineInstrExpressionTrait>::iterator
171 ResultIt;
172 std::tie(ResultIt, WasInserted) =
173 InstructionIntegerMap.insert(std::make_pair(&MI, LegalInstrNumber));
174 unsigned MINumber = ResultIt->second;
175
176 // There was an insertion.
177 if (WasInserted)
178 LegalInstrNumber++;
179
180 UnsignedVecForMBB.push_back(MINumber);
181
182 // Make sure we don't overflow or use any integers reserved by the DenseMap.
183 if (LegalInstrNumber >= IllegalInstrNumber)
184 report_fatal_error("Instruction mapping overflow!");
185
186 assert(LegalInstrNumber != DenseMapInfo<unsigned>::getEmptyKey() &&
187 "Tried to assign DenseMap tombstone or empty key to instruction.");
188 assert(LegalInstrNumber != DenseMapInfo<unsigned>::getTombstoneKey() &&
189 "Tried to assign DenseMap tombstone or empty key to instruction.");
190
191 return MINumber;
192 }
193
194 /// Maps \p *It to an illegal integer.
195 ///
196 /// Updates \p InstrListForMBB, \p UnsignedVecForMBB, and \p
197 /// IllegalInstrNumber.
198 ///
199 /// \returns The integer that \p *It was mapped to.
mapToIllegalUnsigned__anon3e97bd940111::InstructionMapper200 unsigned mapToIllegalUnsigned(
201 MachineBasicBlock::iterator &It, bool &CanOutlineWithPrevInstr,
202 std::vector<unsigned> &UnsignedVecForMBB,
203 std::vector<MachineBasicBlock::iterator> &InstrListForMBB) {
204 // Can't outline an illegal instruction. Set the flag.
205 CanOutlineWithPrevInstr = false;
206
207 // Only add one illegal number per range of legal numbers.
208 if (AddedIllegalLastTime)
209 return IllegalInstrNumber;
210
211 // Remember that we added an illegal number last time.
212 AddedIllegalLastTime = true;
213 unsigned MINumber = IllegalInstrNumber;
214
215 InstrListForMBB.push_back(It);
216 UnsignedVecForMBB.push_back(IllegalInstrNumber);
217 IllegalInstrNumber--;
218
219 assert(LegalInstrNumber < IllegalInstrNumber &&
220 "Instruction mapping overflow!");
221
222 assert(IllegalInstrNumber != DenseMapInfo<unsigned>::getEmptyKey() &&
223 "IllegalInstrNumber cannot be DenseMap tombstone or empty key!");
224
225 assert(IllegalInstrNumber != DenseMapInfo<unsigned>::getTombstoneKey() &&
226 "IllegalInstrNumber cannot be DenseMap tombstone or empty key!");
227
228 return MINumber;
229 }
230
231 /// Transforms a \p MachineBasicBlock into a \p vector of \p unsigneds
232 /// and appends it to \p UnsignedVec and \p InstrList.
233 ///
234 /// Two instructions are assigned the same integer if they are identical.
235 /// If an instruction is deemed unsafe to outline, then it will be assigned an
236 /// unique integer. The resulting mapping is placed into a suffix tree and
237 /// queried for candidates.
238 ///
239 /// \param MBB The \p MachineBasicBlock to be translated into integers.
240 /// \param TII \p TargetInstrInfo for the function.
convertToUnsignedVec__anon3e97bd940111::InstructionMapper241 void convertToUnsignedVec(MachineBasicBlock &MBB,
242 const TargetInstrInfo &TII) {
243 unsigned Flags = 0;
244
245 // Don't even map in this case.
246 if (!TII.isMBBSafeToOutlineFrom(MBB, Flags))
247 return;
248
249 // Store info for the MBB for later outlining.
250 MBBFlagsMap[&MBB] = Flags;
251
252 MachineBasicBlock::iterator It = MBB.begin();
253
254 // The number of instructions in this block that will be considered for
255 // outlining.
256 unsigned NumLegalInBlock = 0;
257
258 // True if we have at least two legal instructions which aren't separated
259 // by an illegal instruction.
260 bool HaveLegalRange = false;
261
262 // True if we can perform outlining given the last mapped (non-invisible)
263 // instruction. This lets us know if we have a legal range.
264 bool CanOutlineWithPrevInstr = false;
265
266 // FIXME: Should this all just be handled in the target, rather than using
267 // repeated calls to getOutliningType?
268 std::vector<unsigned> UnsignedVecForMBB;
269 std::vector<MachineBasicBlock::iterator> InstrListForMBB;
270
271 for (MachineBasicBlock::iterator Et = MBB.end(); It != Et; ++It) {
272 // Keep track of where this instruction is in the module.
273 switch (TII.getOutliningType(It, Flags)) {
274 case InstrType::Illegal:
275 mapToIllegalUnsigned(It, CanOutlineWithPrevInstr, UnsignedVecForMBB,
276 InstrListForMBB);
277 break;
278
279 case InstrType::Legal:
280 mapToLegalUnsigned(It, CanOutlineWithPrevInstr, HaveLegalRange,
281 NumLegalInBlock, UnsignedVecForMBB, InstrListForMBB);
282 break;
283
284 case InstrType::LegalTerminator:
285 mapToLegalUnsigned(It, CanOutlineWithPrevInstr, HaveLegalRange,
286 NumLegalInBlock, UnsignedVecForMBB, InstrListForMBB);
287 // The instruction also acts as a terminator, so we have to record that
288 // in the string.
289 mapToIllegalUnsigned(It, CanOutlineWithPrevInstr, UnsignedVecForMBB,
290 InstrListForMBB);
291 break;
292
293 case InstrType::Invisible:
294 // Normally this is set by mapTo(Blah)Unsigned, but we just want to
295 // skip this instruction. So, unset the flag here.
296 AddedIllegalLastTime = false;
297 break;
298 }
299 }
300
301 // Are there enough legal instructions in the block for outlining to be
302 // possible?
303 if (HaveLegalRange) {
304 // After we're done every insertion, uniquely terminate this part of the
305 // "string". This makes sure we won't match across basic block or function
306 // boundaries since the "end" is encoded uniquely and thus appears in no
307 // repeated substring.
308 mapToIllegalUnsigned(It, CanOutlineWithPrevInstr, UnsignedVecForMBB,
309 InstrListForMBB);
310 InstrList.insert(InstrList.end(), InstrListForMBB.begin(),
311 InstrListForMBB.end());
312 UnsignedVec.insert(UnsignedVec.end(), UnsignedVecForMBB.begin(),
313 UnsignedVecForMBB.end());
314 }
315 }
316
InstructionMapper__anon3e97bd940111::InstructionMapper317 InstructionMapper() {
318 // Make sure that the implementation of DenseMapInfo<unsigned> hasn't
319 // changed.
320 assert(DenseMapInfo<unsigned>::getEmptyKey() == (unsigned)-1 &&
321 "DenseMapInfo<unsigned>'s empty key isn't -1!");
322 assert(DenseMapInfo<unsigned>::getTombstoneKey() == (unsigned)-2 &&
323 "DenseMapInfo<unsigned>'s tombstone key isn't -2!");
324 }
325 };
326
327 /// An interprocedural pass which finds repeated sequences of
328 /// instructions and replaces them with calls to functions.
329 ///
330 /// Each instruction is mapped to an unsigned integer and placed in a string.
331 /// The resulting mapping is then placed in a \p SuffixTree. The \p SuffixTree
332 /// is then repeatedly queried for repeated sequences of instructions. Each
333 /// non-overlapping repeated sequence is then placed in its own
334 /// \p MachineFunction and each instance is then replaced with a call to that
335 /// function.
336 struct MachineOutliner : public ModulePass {
337
338 static char ID;
339
340 /// Set to true if the outliner should consider functions with
341 /// linkonceodr linkage.
342 bool OutlineFromLinkOnceODRs = false;
343
344 /// The current repeat number of machine outlining.
345 unsigned OutlineRepeatedNum = 0;
346
347 /// Set to true if the outliner should run on all functions in the module
348 /// considered safe for outlining.
349 /// Set to true by default for compatibility with llc's -run-pass option.
350 /// Set when the pass is constructed in TargetPassConfig.
351 bool RunOnAllFunctions = true;
352
getPassName__anon3e97bd940111::MachineOutliner353 StringRef getPassName() const override { return "Machine Outliner"; }
354
getAnalysisUsage__anon3e97bd940111::MachineOutliner355 void getAnalysisUsage(AnalysisUsage &AU) const override {
356 AU.addRequired<MachineModuleInfoWrapperPass>();
357 AU.addPreserved<MachineModuleInfoWrapperPass>();
358 AU.setPreservesAll();
359 ModulePass::getAnalysisUsage(AU);
360 }
361
MachineOutliner__anon3e97bd940111::MachineOutliner362 MachineOutliner() : ModulePass(ID) {
363 initializeMachineOutlinerPass(*PassRegistry::getPassRegistry());
364 }
365
366 /// Remark output explaining that not outlining a set of candidates would be
367 /// better than outlining that set.
368 void emitNotOutliningCheaperRemark(
369 unsigned StringLen, std::vector<Candidate> &CandidatesForRepeatedSeq,
370 OutlinedFunction &OF);
371
372 /// Remark output explaining that a function was outlined.
373 void emitOutlinedFunctionRemark(OutlinedFunction &OF);
374
375 /// Find all repeated substrings that satisfy the outlining cost model by
376 /// constructing a suffix tree.
377 ///
378 /// If a substring appears at least twice, then it must be represented by
379 /// an internal node which appears in at least two suffixes. Each suffix
380 /// is represented by a leaf node. To do this, we visit each internal node
381 /// in the tree, using the leaf children of each internal node. If an
382 /// internal node represents a beneficial substring, then we use each of
383 /// its leaf children to find the locations of its substring.
384 ///
385 /// \param Mapper Contains outlining mapping information.
386 /// \param[out] FunctionList Filled with a list of \p OutlinedFunctions
387 /// each type of candidate.
388 void findCandidates(InstructionMapper &Mapper,
389 std::vector<OutlinedFunction> &FunctionList);
390
391 /// Replace the sequences of instructions represented by \p OutlinedFunctions
392 /// with calls to functions.
393 ///
394 /// \param M The module we are outlining from.
395 /// \param FunctionList A list of functions to be inserted into the module.
396 /// \param Mapper Contains the instruction mappings for the module.
397 bool outline(Module &M, std::vector<OutlinedFunction> &FunctionList,
398 InstructionMapper &Mapper, unsigned &OutlinedFunctionNum);
399
400 /// Creates a function for \p OF and inserts it into the module.
401 MachineFunction *createOutlinedFunction(Module &M, OutlinedFunction &OF,
402 InstructionMapper &Mapper,
403 unsigned Name);
404
405 /// Calls 'doOutline()' 1 + OutlinerReruns times.
406 bool runOnModule(Module &M) override;
407
408 /// Construct a suffix tree on the instructions in \p M and outline repeated
409 /// strings from that tree.
410 bool doOutline(Module &M, unsigned &OutlinedFunctionNum);
411
412 /// Return a DISubprogram for OF if one exists, and null otherwise. Helper
413 /// function for remark emission.
getSubprogramOrNull__anon3e97bd940111::MachineOutliner414 DISubprogram *getSubprogramOrNull(const OutlinedFunction &OF) {
415 for (const Candidate &C : OF.Candidates)
416 if (MachineFunction *MF = C.getMF())
417 if (DISubprogram *SP = MF->getFunction().getSubprogram())
418 return SP;
419 return nullptr;
420 }
421
422 /// Populate and \p InstructionMapper with instruction-to-integer mappings.
423 /// These are used to construct a suffix tree.
424 void populateMapper(InstructionMapper &Mapper, Module &M,
425 MachineModuleInfo &MMI);
426
427 /// Initialize information necessary to output a size remark.
428 /// FIXME: This should be handled by the pass manager, not the outliner.
429 /// FIXME: This is nearly identical to the initSizeRemarkInfo in the legacy
430 /// pass manager.
431 void initSizeRemarkInfo(const Module &M, const MachineModuleInfo &MMI,
432 StringMap<unsigned> &FunctionToInstrCount);
433
434 /// Emit the remark.
435 // FIXME: This should be handled by the pass manager, not the outliner.
436 void
437 emitInstrCountChangedRemark(const Module &M, const MachineModuleInfo &MMI,
438 const StringMap<unsigned> &FunctionToInstrCount);
439 };
440 } // Anonymous namespace.
441
442 char MachineOutliner::ID = 0;
443
444 namespace llvm {
createMachineOutlinerPass(bool RunOnAllFunctions)445 ModulePass *createMachineOutlinerPass(bool RunOnAllFunctions) {
446 MachineOutliner *OL = new MachineOutliner();
447 OL->RunOnAllFunctions = RunOnAllFunctions;
448 return OL;
449 }
450
451 } // namespace llvm
452
453 INITIALIZE_PASS(MachineOutliner, DEBUG_TYPE, "Machine Function Outliner", false,
454 false)
455
emitNotOutliningCheaperRemark(unsigned StringLen,std::vector<Candidate> & CandidatesForRepeatedSeq,OutlinedFunction & OF)456 void MachineOutliner::emitNotOutliningCheaperRemark(
457 unsigned StringLen, std::vector<Candidate> &CandidatesForRepeatedSeq,
458 OutlinedFunction &OF) {
459 // FIXME: Right now, we arbitrarily choose some Candidate from the
460 // OutlinedFunction. This isn't necessarily fixed, nor does it have to be.
461 // We should probably sort these by function name or something to make sure
462 // the remarks are stable.
463 Candidate &C = CandidatesForRepeatedSeq.front();
464 MachineOptimizationRemarkEmitter MORE(*(C.getMF()), nullptr);
465 MORE.emit([&]() {
466 MachineOptimizationRemarkMissed R(DEBUG_TYPE, "NotOutliningCheaper",
467 C.front()->getDebugLoc(), C.getMBB());
468 R << "Did not outline " << NV("Length", StringLen) << " instructions"
469 << " from " << NV("NumOccurrences", CandidatesForRepeatedSeq.size())
470 << " locations."
471 << " Bytes from outlining all occurrences ("
472 << NV("OutliningCost", OF.getOutliningCost()) << ")"
473 << " >= Unoutlined instruction bytes ("
474 << NV("NotOutliningCost", OF.getNotOutlinedCost()) << ")"
475 << " (Also found at: ";
476
477 // Tell the user the other places the candidate was found.
478 for (unsigned i = 1, e = CandidatesForRepeatedSeq.size(); i < e; i++) {
479 R << NV((Twine("OtherStartLoc") + Twine(i)).str(),
480 CandidatesForRepeatedSeq[i].front()->getDebugLoc());
481 if (i != e - 1)
482 R << ", ";
483 }
484
485 R << ")";
486 return R;
487 });
488 }
489
emitOutlinedFunctionRemark(OutlinedFunction & OF)490 void MachineOutliner::emitOutlinedFunctionRemark(OutlinedFunction &OF) {
491 MachineBasicBlock *MBB = &*OF.MF->begin();
492 MachineOptimizationRemarkEmitter MORE(*OF.MF, nullptr);
493 MachineOptimizationRemark R(DEBUG_TYPE, "OutlinedFunction",
494 MBB->findDebugLoc(MBB->begin()), MBB);
495 R << "Saved " << NV("OutliningBenefit", OF.getBenefit()) << " bytes by "
496 << "outlining " << NV("Length", OF.getNumInstrs()) << " instructions "
497 << "from " << NV("NumOccurrences", OF.getOccurrenceCount())
498 << " locations. "
499 << "(Found at: ";
500
501 // Tell the user the other places the candidate was found.
502 for (size_t i = 0, e = OF.Candidates.size(); i < e; i++) {
503
504 R << NV((Twine("StartLoc") + Twine(i)).str(),
505 OF.Candidates[i].front()->getDebugLoc());
506 if (i != e - 1)
507 R << ", ";
508 }
509
510 R << ")";
511
512 MORE.emit(R);
513 }
514
findCandidates(InstructionMapper & Mapper,std::vector<OutlinedFunction> & FunctionList)515 void MachineOutliner::findCandidates(
516 InstructionMapper &Mapper, std::vector<OutlinedFunction> &FunctionList) {
517 FunctionList.clear();
518 SuffixTree ST(Mapper.UnsignedVec);
519
520 // First, find all of the repeated substrings in the tree of minimum length
521 // 2.
522 std::vector<Candidate> CandidatesForRepeatedSeq;
523 for (auto It = ST.begin(), Et = ST.end(); It != Et; ++It) {
524 CandidatesForRepeatedSeq.clear();
525 SuffixTree::RepeatedSubstring RS = *It;
526 unsigned StringLen = RS.Length;
527 for (const unsigned &StartIdx : RS.StartIndices) {
528 unsigned EndIdx = StartIdx + StringLen - 1;
529 // Trick: Discard some candidates that would be incompatible with the
530 // ones we've already found for this sequence. This will save us some
531 // work in candidate selection.
532 //
533 // If two candidates overlap, then we can't outline them both. This
534 // happens when we have candidates that look like, say
535 //
536 // AA (where each "A" is an instruction).
537 //
538 // We might have some portion of the module that looks like this:
539 // AAAAAA (6 A's)
540 //
541 // In this case, there are 5 different copies of "AA" in this range, but
542 // at most 3 can be outlined. If only outlining 3 of these is going to
543 // be unbeneficial, then we ought to not bother.
544 //
545 // Note that two things DON'T overlap when they look like this:
546 // start1...end1 .... start2...end2
547 // That is, one must either
548 // * End before the other starts
549 // * Start after the other ends
550 if (std::all_of(
551 CandidatesForRepeatedSeq.begin(), CandidatesForRepeatedSeq.end(),
552 [&StartIdx, &EndIdx](const Candidate &C) {
553 return (EndIdx < C.getStartIdx() || StartIdx > C.getEndIdx());
554 })) {
555 // It doesn't overlap with anything, so we can outline it.
556 // Each sequence is over [StartIt, EndIt].
557 // Save the candidate and its location.
558
559 MachineBasicBlock::iterator StartIt = Mapper.InstrList[StartIdx];
560 MachineBasicBlock::iterator EndIt = Mapper.InstrList[EndIdx];
561 MachineBasicBlock *MBB = StartIt->getParent();
562
563 CandidatesForRepeatedSeq.emplace_back(StartIdx, StringLen, StartIt,
564 EndIt, MBB, FunctionList.size(),
565 Mapper.MBBFlagsMap[MBB]);
566 }
567 }
568
569 // We've found something we might want to outline.
570 // Create an OutlinedFunction to store it and check if it'd be beneficial
571 // to outline.
572 if (CandidatesForRepeatedSeq.size() < 2)
573 continue;
574
575 // Arbitrarily choose a TII from the first candidate.
576 // FIXME: Should getOutliningCandidateInfo move to TargetMachine?
577 const TargetInstrInfo *TII =
578 CandidatesForRepeatedSeq[0].getMF()->getSubtarget().getInstrInfo();
579
580 OutlinedFunction OF =
581 TII->getOutliningCandidateInfo(CandidatesForRepeatedSeq);
582
583 // If we deleted too many candidates, then there's nothing worth outlining.
584 // FIXME: This should take target-specified instruction sizes into account.
585 if (OF.Candidates.size() < 2)
586 continue;
587
588 // Is it better to outline this candidate than not?
589 if (OF.getBenefit() < 1) {
590 emitNotOutliningCheaperRemark(StringLen, CandidatesForRepeatedSeq, OF);
591 continue;
592 }
593
594 FunctionList.push_back(OF);
595 }
596 }
597
createOutlinedFunction(Module & M,OutlinedFunction & OF,InstructionMapper & Mapper,unsigned Name)598 MachineFunction *MachineOutliner::createOutlinedFunction(
599 Module &M, OutlinedFunction &OF, InstructionMapper &Mapper, unsigned Name) {
600
601 // Create the function name. This should be unique.
602 // FIXME: We should have a better naming scheme. This should be stable,
603 // regardless of changes to the outliner's cost model/traversal order.
604 std::string FunctionName = "OUTLINED_FUNCTION_";
605 if (OutlineRepeatedNum > 0)
606 FunctionName += std::to_string(OutlineRepeatedNum + 1) + "_";
607 FunctionName += std::to_string(Name);
608
609 // Create the function using an IR-level function.
610 LLVMContext &C = M.getContext();
611 Function *F = Function::Create(FunctionType::get(Type::getVoidTy(C), false),
612 Function::ExternalLinkage, FunctionName, M);
613
614 // NOTE: If this is linkonceodr, then we can take advantage of linker deduping
615 // which gives us better results when we outline from linkonceodr functions.
616 F->setLinkage(GlobalValue::InternalLinkage);
617 F->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
618
619 // Set optsize/minsize, so we don't insert padding between outlined
620 // functions.
621 F->addFnAttr(Attribute::OptimizeForSize);
622 F->addFnAttr(Attribute::MinSize);
623
624 // Include target features from an arbitrary candidate for the outlined
625 // function. This makes sure the outlined function knows what kinds of
626 // instructions are going into it. This is fine, since all parent functions
627 // must necessarily support the instructions that are in the outlined region.
628 Candidate &FirstCand = OF.Candidates.front();
629 const Function &ParentFn = FirstCand.getMF()->getFunction();
630 if (ParentFn.hasFnAttribute("target-features"))
631 F->addFnAttr(ParentFn.getFnAttribute("target-features"));
632
633 // Set nounwind, so we don't generate eh_frame.
634 if (llvm::all_of(OF.Candidates, [](const outliner::Candidate &C) {
635 return C.getMF()->getFunction().hasFnAttribute(Attribute::NoUnwind);
636 }))
637 F->addFnAttr(Attribute::NoUnwind);
638
639 BasicBlock *EntryBB = BasicBlock::Create(C, "entry", F);
640 IRBuilder<> Builder(EntryBB);
641 Builder.CreateRetVoid();
642
643 MachineModuleInfo &MMI = getAnalysis<MachineModuleInfoWrapperPass>().getMMI();
644 MachineFunction &MF = MMI.getOrCreateMachineFunction(*F);
645 MachineBasicBlock &MBB = *MF.CreateMachineBasicBlock();
646 const TargetSubtargetInfo &STI = MF.getSubtarget();
647 const TargetInstrInfo &TII = *STI.getInstrInfo();
648
649 // Insert the new function into the module.
650 MF.insert(MF.begin(), &MBB);
651
652 MachineFunction *OriginalMF = FirstCand.front()->getMF();
653 const std::vector<MCCFIInstruction> &Instrs =
654 OriginalMF->getFrameInstructions();
655 for (auto I = FirstCand.front(), E = std::next(FirstCand.back()); I != E;
656 ++I) {
657 if (I->isDebugInstr())
658 continue;
659 MachineInstr *NewMI = MF.CloneMachineInstr(&*I);
660 if (I->isCFIInstruction()) {
661 unsigned CFIIndex = NewMI->getOperand(0).getCFIIndex();
662 MCCFIInstruction CFI = Instrs[CFIIndex];
663 (void)MF.addFrameInst(CFI);
664 }
665 NewMI->dropMemRefs(MF);
666
667 // Don't keep debug information for outlined instructions.
668 NewMI->setDebugLoc(DebugLoc());
669 MBB.insert(MBB.end(), NewMI);
670 }
671
672 // Set normal properties for a late MachineFunction.
673 MF.getProperties().reset(MachineFunctionProperties::Property::IsSSA);
674 MF.getProperties().set(MachineFunctionProperties::Property::NoPHIs);
675 MF.getProperties().set(MachineFunctionProperties::Property::NoVRegs);
676 MF.getProperties().set(MachineFunctionProperties::Property::TracksLiveness);
677 MF.getRegInfo().freezeReservedRegs(MF);
678
679 // Compute live-in set for outlined fn
680 const MachineRegisterInfo &MRI = MF.getRegInfo();
681 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
682 LivePhysRegs LiveIns(TRI);
683 for (auto &Cand : OF.Candidates) {
684 // Figure out live-ins at the first instruction.
685 MachineBasicBlock &OutlineBB = *Cand.front()->getParent();
686 LivePhysRegs CandLiveIns(TRI);
687 CandLiveIns.addLiveOuts(OutlineBB);
688 for (const MachineInstr &MI :
689 reverse(make_range(Cand.front(), OutlineBB.end())))
690 CandLiveIns.stepBackward(MI);
691
692 // The live-in set for the outlined function is the union of the live-ins
693 // from all the outlining points.
694 for (MCPhysReg Reg : make_range(CandLiveIns.begin(), CandLiveIns.end()))
695 LiveIns.addReg(Reg);
696 }
697 addLiveIns(MBB, LiveIns);
698
699 TII.buildOutlinedFrame(MBB, MF, OF);
700
701 // If there's a DISubprogram associated with this outlined function, then
702 // emit debug info for the outlined function.
703 if (DISubprogram *SP = getSubprogramOrNull(OF)) {
704 // We have a DISubprogram. Get its DICompileUnit.
705 DICompileUnit *CU = SP->getUnit();
706 DIBuilder DB(M, true, CU);
707 DIFile *Unit = SP->getFile();
708 Mangler Mg;
709 // Get the mangled name of the function for the linkage name.
710 std::string Dummy;
711 llvm::raw_string_ostream MangledNameStream(Dummy);
712 Mg.getNameWithPrefix(MangledNameStream, F, false);
713
714 DISubprogram *OutlinedSP = DB.createFunction(
715 Unit /* Context */, F->getName(), StringRef(MangledNameStream.str()),
716 Unit /* File */,
717 0 /* Line 0 is reserved for compiler-generated code. */,
718 DB.createSubroutineType(DB.getOrCreateTypeArray(None)), /* void type */
719 0, /* Line 0 is reserved for compiler-generated code. */
720 DINode::DIFlags::FlagArtificial /* Compiler-generated code. */,
721 /* Outlined code is optimized code by definition. */
722 DISubprogram::SPFlagDefinition | DISubprogram::SPFlagOptimized);
723
724 // Don't add any new variables to the subprogram.
725 DB.finalizeSubprogram(OutlinedSP);
726
727 // Attach subprogram to the function.
728 F->setSubprogram(OutlinedSP);
729 // We're done with the DIBuilder.
730 DB.finalize();
731 }
732
733 return &MF;
734 }
735
outline(Module & M,std::vector<OutlinedFunction> & FunctionList,InstructionMapper & Mapper,unsigned & OutlinedFunctionNum)736 bool MachineOutliner::outline(Module &M,
737 std::vector<OutlinedFunction> &FunctionList,
738 InstructionMapper &Mapper,
739 unsigned &OutlinedFunctionNum) {
740
741 bool OutlinedSomething = false;
742
743 // Sort by benefit. The most beneficial functions should be outlined first.
744 llvm::stable_sort(FunctionList, [](const OutlinedFunction &LHS,
745 const OutlinedFunction &RHS) {
746 return LHS.getBenefit() > RHS.getBenefit();
747 });
748
749 // Walk over each function, outlining them as we go along. Functions are
750 // outlined greedily, based off the sort above.
751 for (OutlinedFunction &OF : FunctionList) {
752 // If we outlined something that overlapped with a candidate in a previous
753 // step, then we can't outline from it.
754 erase_if(OF.Candidates, [&Mapper](Candidate &C) {
755 return std::any_of(
756 Mapper.UnsignedVec.begin() + C.getStartIdx(),
757 Mapper.UnsignedVec.begin() + C.getEndIdx() + 1,
758 [](unsigned I) { return (I == static_cast<unsigned>(-1)); });
759 });
760
761 // If we made it unbeneficial to outline this function, skip it.
762 if (OF.getBenefit() < 1)
763 continue;
764
765 // It's beneficial. Create the function and outline its sequence's
766 // occurrences.
767 OF.MF = createOutlinedFunction(M, OF, Mapper, OutlinedFunctionNum);
768 emitOutlinedFunctionRemark(OF);
769 FunctionsCreated++;
770 OutlinedFunctionNum++; // Created a function, move to the next name.
771 MachineFunction *MF = OF.MF;
772 const TargetSubtargetInfo &STI = MF->getSubtarget();
773 const TargetInstrInfo &TII = *STI.getInstrInfo();
774
775 // Replace occurrences of the sequence with calls to the new function.
776 for (Candidate &C : OF.Candidates) {
777 MachineBasicBlock &MBB = *C.getMBB();
778 MachineBasicBlock::iterator StartIt = C.front();
779 MachineBasicBlock::iterator EndIt = C.back();
780
781 // Insert the call.
782 auto CallInst = TII.insertOutlinedCall(M, MBB, StartIt, *MF, C);
783
784 // If the caller tracks liveness, then we need to make sure that
785 // anything we outline doesn't break liveness assumptions. The outlined
786 // functions themselves currently don't track liveness, but we should
787 // make sure that the ranges we yank things out of aren't wrong.
788 if (MBB.getParent()->getProperties().hasProperty(
789 MachineFunctionProperties::Property::TracksLiveness)) {
790 // The following code is to add implicit def operands to the call
791 // instruction. It also updates call site information for moved
792 // code.
793 SmallSet<Register, 2> UseRegs, DefRegs;
794 // Copy over the defs in the outlined range.
795 // First inst in outlined range <-- Anything that's defined in this
796 // ... .. range has to be added as an
797 // implicit Last inst in outlined range <-- def to the call
798 // instruction. Also remove call site information for outlined block
799 // of code. The exposed uses need to be copied in the outlined range.
800 for (MachineBasicBlock::reverse_iterator
801 Iter = EndIt.getReverse(),
802 Last = std::next(CallInst.getReverse());
803 Iter != Last; Iter++) {
804 MachineInstr *MI = &*Iter;
805 for (MachineOperand &MOP : MI->operands()) {
806 // Skip over anything that isn't a register.
807 if (!MOP.isReg())
808 continue;
809
810 if (MOP.isDef()) {
811 // Introduce DefRegs set to skip the redundant register.
812 DefRegs.insert(MOP.getReg());
813 if (UseRegs.count(MOP.getReg()))
814 // Since the regiester is modeled as defined,
815 // it is not necessary to be put in use register set.
816 UseRegs.erase(MOP.getReg());
817 } else if (!MOP.isUndef()) {
818 // Any register which is not undefined should
819 // be put in the use register set.
820 UseRegs.insert(MOP.getReg());
821 }
822 }
823 if (MI->isCandidateForCallSiteEntry())
824 MI->getMF()->eraseCallSiteInfo(MI);
825 }
826
827 for (const Register &I : DefRegs)
828 // If it's a def, add it to the call instruction.
829 CallInst->addOperand(
830 MachineOperand::CreateReg(I, true, /* isDef = true */
831 true /* isImp = true */));
832
833 for (const Register &I : UseRegs)
834 // If it's a exposed use, add it to the call instruction.
835 CallInst->addOperand(
836 MachineOperand::CreateReg(I, false, /* isDef = false */
837 true /* isImp = true */));
838 }
839
840 // Erase from the point after where the call was inserted up to, and
841 // including, the final instruction in the sequence.
842 // Erase needs one past the end, so we need std::next there too.
843 MBB.erase(std::next(StartIt), std::next(EndIt));
844
845 // Keep track of what we removed by marking them all as -1.
846 std::for_each(Mapper.UnsignedVec.begin() + C.getStartIdx(),
847 Mapper.UnsignedVec.begin() + C.getEndIdx() + 1,
848 [](unsigned &I) { I = static_cast<unsigned>(-1); });
849 OutlinedSomething = true;
850
851 // Statistics.
852 NumOutlined++;
853 }
854 }
855
856 LLVM_DEBUG(dbgs() << "OutlinedSomething = " << OutlinedSomething << "\n";);
857 return OutlinedSomething;
858 }
859
populateMapper(InstructionMapper & Mapper,Module & M,MachineModuleInfo & MMI)860 void MachineOutliner::populateMapper(InstructionMapper &Mapper, Module &M,
861 MachineModuleInfo &MMI) {
862 // Build instruction mappings for each function in the module. Start by
863 // iterating over each Function in M.
864 for (Function &F : M) {
865
866 // If there's nothing in F, then there's no reason to try and outline from
867 // it.
868 if (F.empty())
869 continue;
870
871 // There's something in F. Check if it has a MachineFunction associated with
872 // it.
873 MachineFunction *MF = MMI.getMachineFunction(F);
874
875 // If it doesn't, then there's nothing to outline from. Move to the next
876 // Function.
877 if (!MF)
878 continue;
879
880 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
881
882 if (!RunOnAllFunctions && !TII->shouldOutlineFromFunctionByDefault(*MF))
883 continue;
884
885 // We have a MachineFunction. Ask the target if it's suitable for outlining.
886 // If it isn't, then move on to the next Function in the module.
887 if (!TII->isFunctionSafeToOutlineFrom(*MF, OutlineFromLinkOnceODRs))
888 continue;
889
890 // We have a function suitable for outlining. Iterate over every
891 // MachineBasicBlock in MF and try to map its instructions to a list of
892 // unsigned integers.
893 for (MachineBasicBlock &MBB : *MF) {
894 // If there isn't anything in MBB, then there's no point in outlining from
895 // it.
896 // If there are fewer than 2 instructions in the MBB, then it can't ever
897 // contain something worth outlining.
898 // FIXME: This should be based off of the maximum size in B of an outlined
899 // call versus the size in B of the MBB.
900 if (MBB.empty() || MBB.size() < 2)
901 continue;
902
903 // Check if MBB could be the target of an indirect branch. If it is, then
904 // we don't want to outline from it.
905 if (MBB.hasAddressTaken())
906 continue;
907
908 // MBB is suitable for outlining. Map it to a list of unsigneds.
909 Mapper.convertToUnsignedVec(MBB, *TII);
910 }
911 }
912 }
913
initSizeRemarkInfo(const Module & M,const MachineModuleInfo & MMI,StringMap<unsigned> & FunctionToInstrCount)914 void MachineOutliner::initSizeRemarkInfo(
915 const Module &M, const MachineModuleInfo &MMI,
916 StringMap<unsigned> &FunctionToInstrCount) {
917 // Collect instruction counts for every function. We'll use this to emit
918 // per-function size remarks later.
919 for (const Function &F : M) {
920 MachineFunction *MF = MMI.getMachineFunction(F);
921
922 // We only care about MI counts here. If there's no MachineFunction at this
923 // point, then there won't be after the outliner runs, so let's move on.
924 if (!MF)
925 continue;
926 FunctionToInstrCount[F.getName().str()] = MF->getInstructionCount();
927 }
928 }
929
emitInstrCountChangedRemark(const Module & M,const MachineModuleInfo & MMI,const StringMap<unsigned> & FunctionToInstrCount)930 void MachineOutliner::emitInstrCountChangedRemark(
931 const Module &M, const MachineModuleInfo &MMI,
932 const StringMap<unsigned> &FunctionToInstrCount) {
933 // Iterate over each function in the module and emit remarks.
934 // Note that we won't miss anything by doing this, because the outliner never
935 // deletes functions.
936 for (const Function &F : M) {
937 MachineFunction *MF = MMI.getMachineFunction(F);
938
939 // The outliner never deletes functions. If we don't have a MF here, then we
940 // didn't have one prior to outlining either.
941 if (!MF)
942 continue;
943
944 std::string Fname = std::string(F.getName());
945 unsigned FnCountAfter = MF->getInstructionCount();
946 unsigned FnCountBefore = 0;
947
948 // Check if the function was recorded before.
949 auto It = FunctionToInstrCount.find(Fname);
950
951 // Did we have a previously-recorded size? If yes, then set FnCountBefore
952 // to that.
953 if (It != FunctionToInstrCount.end())
954 FnCountBefore = It->second;
955
956 // Compute the delta and emit a remark if there was a change.
957 int64_t FnDelta = static_cast<int64_t>(FnCountAfter) -
958 static_cast<int64_t>(FnCountBefore);
959 if (FnDelta == 0)
960 continue;
961
962 MachineOptimizationRemarkEmitter MORE(*MF, nullptr);
963 MORE.emit([&]() {
964 MachineOptimizationRemarkAnalysis R("size-info", "FunctionMISizeChange",
965 DiagnosticLocation(), &MF->front());
966 R << DiagnosticInfoOptimizationBase::Argument("Pass", "Machine Outliner")
967 << ": Function: "
968 << DiagnosticInfoOptimizationBase::Argument("Function", F.getName())
969 << ": MI instruction count changed from "
970 << DiagnosticInfoOptimizationBase::Argument("MIInstrsBefore",
971 FnCountBefore)
972 << " to "
973 << DiagnosticInfoOptimizationBase::Argument("MIInstrsAfter",
974 FnCountAfter)
975 << "; Delta: "
976 << DiagnosticInfoOptimizationBase::Argument("Delta", FnDelta);
977 return R;
978 });
979 }
980 }
981
runOnModule(Module & M)982 bool MachineOutliner::runOnModule(Module &M) {
983 // Check if there's anything in the module. If it's empty, then there's
984 // nothing to outline.
985 if (M.empty())
986 return false;
987
988 // Number to append to the current outlined function.
989 unsigned OutlinedFunctionNum = 0;
990
991 OutlineRepeatedNum = 0;
992 if (!doOutline(M, OutlinedFunctionNum))
993 return false;
994
995 for (unsigned I = 0; I < OutlinerReruns; ++I) {
996 OutlinedFunctionNum = 0;
997 OutlineRepeatedNum++;
998 if (!doOutline(M, OutlinedFunctionNum)) {
999 LLVM_DEBUG({
1000 dbgs() << "Did not outline on iteration " << I + 2 << " out of "
1001 << OutlinerReruns + 1 << "\n";
1002 });
1003 break;
1004 }
1005 }
1006
1007 return true;
1008 }
1009
doOutline(Module & M,unsigned & OutlinedFunctionNum)1010 bool MachineOutliner::doOutline(Module &M, unsigned &OutlinedFunctionNum) {
1011 MachineModuleInfo &MMI = getAnalysis<MachineModuleInfoWrapperPass>().getMMI();
1012
1013 // If the user passed -enable-machine-outliner=always or
1014 // -enable-machine-outliner, the pass will run on all functions in the module.
1015 // Otherwise, if the target supports default outlining, it will run on all
1016 // functions deemed by the target to be worth outlining from by default. Tell
1017 // the user how the outliner is running.
1018 LLVM_DEBUG({
1019 dbgs() << "Machine Outliner: Running on ";
1020 if (RunOnAllFunctions)
1021 dbgs() << "all functions";
1022 else
1023 dbgs() << "target-default functions";
1024 dbgs() << "\n";
1025 });
1026
1027 // If the user specifies that they want to outline from linkonceodrs, set
1028 // it here.
1029 OutlineFromLinkOnceODRs = EnableLinkOnceODROutlining;
1030 InstructionMapper Mapper;
1031
1032 // Prepare instruction mappings for the suffix tree.
1033 populateMapper(Mapper, M, MMI);
1034 std::vector<OutlinedFunction> FunctionList;
1035
1036 // Find all of the outlining candidates.
1037 findCandidates(Mapper, FunctionList);
1038
1039 // If we've requested size remarks, then collect the MI counts of every
1040 // function before outlining, and the MI counts after outlining.
1041 // FIXME: This shouldn't be in the outliner at all; it should ultimately be
1042 // the pass manager's responsibility.
1043 // This could pretty easily be placed in outline instead, but because we
1044 // really ultimately *don't* want this here, it's done like this for now
1045 // instead.
1046
1047 // Check if we want size remarks.
1048 bool ShouldEmitSizeRemarks = M.shouldEmitInstrCountChangedRemark();
1049 StringMap<unsigned> FunctionToInstrCount;
1050 if (ShouldEmitSizeRemarks)
1051 initSizeRemarkInfo(M, MMI, FunctionToInstrCount);
1052
1053 // Outline each of the candidates and return true if something was outlined.
1054 bool OutlinedSomething =
1055 outline(M, FunctionList, Mapper, OutlinedFunctionNum);
1056
1057 // If we outlined something, we definitely changed the MI count of the
1058 // module. If we've asked for size remarks, then output them.
1059 // FIXME: This should be in the pass manager.
1060 if (ShouldEmitSizeRemarks && OutlinedSomething)
1061 emitInstrCountChangedRemark(M, MMI, FunctionToInstrCount);
1062
1063 LLVM_DEBUG({
1064 if (!OutlinedSomething)
1065 dbgs() << "Stopped outlining at iteration " << OutlineRepeatedNum
1066 << " because no changes were found.\n";
1067 });
1068
1069 return OutlinedSomething;
1070 }
1071