1//===- ShrinkWrap.cpp - Compute safe point for prolog/epilog insertion ----===//
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 pass looks for safe point where the prologue and epilogue can be
10// inserted.
11// The safe point for the prologue (resp. epilogue) is called Save
12// (resp. Restore).
13// A point is safe for prologue (resp. epilogue) if and only if
14// it 1) dominates (resp. post-dominates) all the frame related operations and
15// between 2) two executions of the Save (resp. Restore) point there is an
16// execution of the Restore (resp. Save) point.
17//
18// For instance, the following points are safe:
19// for (int i = 0; i < 10; ++i) {
20// Save
21// ...
22// Restore
23// }
24// Indeed, the execution looks like Save -> Restore -> Save -> Restore ...
25// And the following points are not:
26// for (int i = 0; i < 10; ++i) {
27// Save
28// ...
29// }
30// for (int i = 0; i < 10; ++i) {
31// ...
32// Restore
33// }
34// Indeed, the execution looks like Save -> Save -> ... -> Restore -> Restore.
35//
36// This pass also ensures that the safe points are 3) cheaper than the regular
37// entry and exits blocks.
38//
39// Property #1 is ensured via the use of MachineDominatorTree and
40// MachinePostDominatorTree.
41// Property #2 is ensured via property #1 and MachineLoopInfo, i.e., both
42// points must be in the same loop.
43// Property #3 is ensured via the MachineBlockFrequencyInfo.
44//
45// If this pass found points matching all these properties, then
46// MachineFrameInfo is updated with this information.
47//
48//===----------------------------------------------------------------------===//
49
50#include "llvm/ADT/BitVector.h"
51#include "llvm/ADT/PostOrderIterator.h"
52#include "llvm/ADT/SetVector.h"
53#include "llvm/ADT/SmallVector.h"
54#include "llvm/ADT/Statistic.h"
55#include "llvm/Analysis/CFG.h"
56#include "llvm/Analysis/ValueTracking.h"
57#include "llvm/CodeGen/MachineBasicBlock.h"
58#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
59#include "llvm/CodeGen/MachineDominators.h"
60#include "llvm/CodeGen/MachineFrameInfo.h"
61#include "llvm/CodeGen/MachineFunction.h"
62#include "llvm/CodeGen/MachineFunctionPass.h"
63#include "llvm/CodeGen/MachineInstr.h"
64#include "llvm/CodeGen/MachineLoopInfo.h"
65#include "llvm/CodeGen/MachineOperand.h"
66#include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
67#include "llvm/CodeGen/MachinePostDominators.h"
68#include "llvm/CodeGen/RegisterClassInfo.h"
69#include "llvm/CodeGen/RegisterScavenging.h"
70#include "llvm/CodeGen/TargetFrameLowering.h"
71#include "llvm/CodeGen/TargetInstrInfo.h"
72#include "llvm/CodeGen/TargetLowering.h"
73#include "llvm/CodeGen/TargetRegisterInfo.h"
74#include "llvm/CodeGen/TargetSubtargetInfo.h"
75#include "llvm/IR/Attributes.h"
76#include "llvm/IR/Function.h"
77#include "llvm/InitializePasses.h"
78#include "llvm/MC/MCAsmInfo.h"
79#include "llvm/Pass.h"
80#include "llvm/Support/CommandLine.h"
81#include "llvm/Support/Debug.h"
82#include "llvm/Support/ErrorHandling.h"
83#include "llvm/Support/raw_ostream.h"
84#include "llvm/Target/TargetMachine.h"
85#include <cassert>
86#include <cstdint>
87#include <memory>
88
89using namespace llvm;
90
91#define DEBUG_TYPE "shrink-wrap"
92
93STATISTIC(NumFunc, "Number of functions");
94STATISTIC(NumCandidates, "Number of shrink-wrapping candidates");
95STATISTIC(NumCandidatesDropped,
96 "Number of shrink-wrapping candidates dropped because of frequency");
97
98static cl::opt<cl::boolOrDefault>
99EnableShrinkWrapOpt("enable-shrink-wrap", cl::Hidden,
100 cl::desc("enable the shrink-wrapping pass"));
101static cl::opt<bool> EnablePostShrinkWrapOpt(
102 "enable-shrink-wrap-region-split", cl::init(Val: true), cl::Hidden,
103 cl::desc("enable splitting of the restore block if possible"));
104
105namespace {
106
107/// Class to determine where the safe point to insert the
108/// prologue and epilogue are.
109/// Unlike the paper from Fred C. Chow, PLDI'88, that introduces the
110/// shrink-wrapping term for prologue/epilogue placement, this pass
111/// does not rely on expensive data-flow analysis. Instead we use the
112/// dominance properties and loop information to decide which point
113/// are safe for such insertion.
114class ShrinkWrap : public MachineFunctionPass {
115 /// Hold callee-saved information.
116 RegisterClassInfo RCI;
117 MachineDominatorTree *MDT = nullptr;
118 MachinePostDominatorTree *MPDT = nullptr;
119
120 /// Current safe point found for the prologue.
121 /// The prologue will be inserted before the first instruction
122 /// in this basic block.
123 MachineBasicBlock *Save = nullptr;
124
125 /// Current safe point found for the epilogue.
126 /// The epilogue will be inserted before the first terminator instruction
127 /// in this basic block.
128 MachineBasicBlock *Restore = nullptr;
129
130 /// Hold the information of the basic block frequency.
131 /// Use to check the profitability of the new points.
132 MachineBlockFrequencyInfo *MBFI = nullptr;
133
134 /// Hold the loop information. Used to determine if Save and Restore
135 /// are in the same loop.
136 MachineLoopInfo *MLI = nullptr;
137
138 // Emit remarks.
139 MachineOptimizationRemarkEmitter *ORE = nullptr;
140
141 /// Frequency of the Entry block.
142 BlockFrequency EntryFreq;
143
144 /// Current opcode for frame setup.
145 unsigned FrameSetupOpcode = ~0u;
146
147 /// Current opcode for frame destroy.
148 unsigned FrameDestroyOpcode = ~0u;
149
150 /// Stack pointer register, used by llvm.{savestack,restorestack}
151 Register SP;
152
153 /// Entry block.
154 const MachineBasicBlock *Entry = nullptr;
155
156 using SetOfRegs = SmallSetVector<unsigned, 16>;
157
158 /// Registers that need to be saved for the current function.
159 mutable SetOfRegs CurrentCSRs;
160
161 /// Current MachineFunction.
162 MachineFunction *MachineFunc = nullptr;
163
164 /// Is `true` for the block numbers where we assume possible stack accesses
165 /// or computation of stack-relative addresses on any CFG path including the
166 /// block itself. Is `false` for basic blocks where we can guarantee the
167 /// opposite. False positives won't lead to incorrect analysis results,
168 /// therefore this approach is fair.
169 BitVector StackAddressUsedBlockInfo;
170
171 /// Check if \p MI uses or defines a callee-saved register or
172 /// a frame index. If this is the case, this means \p MI must happen
173 /// after Save and before Restore.
174 bool useOrDefCSROrFI(const MachineInstr &MI, RegScavenger *RS,
175 bool StackAddressUsed) const;
176
177 const SetOfRegs &getCurrentCSRs(RegScavenger *RS) const {
178 if (CurrentCSRs.empty()) {
179 BitVector SavedRegs;
180 const TargetFrameLowering *TFI =
181 MachineFunc->getSubtarget().getFrameLowering();
182
183 TFI->determineCalleeSaves(MF&: *MachineFunc, SavedRegs, RS);
184
185 for (int Reg = SavedRegs.find_first(); Reg != -1;
186 Reg = SavedRegs.find_next(Prev: Reg))
187 CurrentCSRs.insert(X: (unsigned)Reg);
188 }
189 return CurrentCSRs;
190 }
191
192 /// Update the Save and Restore points such that \p MBB is in
193 /// the region that is dominated by Save and post-dominated by Restore
194 /// and Save and Restore still match the safe point definition.
195 /// Such point may not exist and Save and/or Restore may be null after
196 /// this call.
197 void updateSaveRestorePoints(MachineBasicBlock &MBB, RegScavenger *RS);
198
199 // Try to find safe point based on dominance and block frequency without
200 // any change in IR.
201 bool performShrinkWrapping(
202 const ReversePostOrderTraversal<MachineBasicBlock *> &RPOT,
203 RegScavenger *RS);
204
205 /// This function tries to split the restore point if doing so can shrink the
206 /// save point further. \return True if restore point is split.
207 bool postShrinkWrapping(bool HasCandidate, MachineFunction &MF,
208 RegScavenger *RS);
209
210 /// This function analyzes if the restore point can split to create a new
211 /// restore point. This function collects
212 /// 1. Any preds of current restore that are reachable by callee save/FI
213 /// blocks
214 /// - indicated by DirtyPreds
215 /// 2. Any preds of current restore that are not DirtyPreds - indicated by
216 /// CleanPreds
217 /// Both sets should be non-empty for considering restore point split.
218 bool checkIfRestoreSplittable(
219 const MachineBasicBlock *CurRestore,
220 const DenseSet<const MachineBasicBlock *> &ReachableByDirty,
221 SmallVectorImpl<MachineBasicBlock *> &DirtyPreds,
222 SmallVectorImpl<MachineBasicBlock *> &CleanPreds,
223 const TargetInstrInfo *TII, RegScavenger *RS);
224
225 /// Initialize the pass for \p MF.
226 void init(MachineFunction &MF) {
227 RCI.runOnMachineFunction(MF);
228 MDT = &getAnalysis<MachineDominatorTree>();
229 MPDT = &getAnalysis<MachinePostDominatorTree>();
230 Save = nullptr;
231 Restore = nullptr;
232 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
233 MLI = &getAnalysis<MachineLoopInfo>();
234 ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE();
235 EntryFreq = MBFI->getEntryFreq();
236 const TargetSubtargetInfo &Subtarget = MF.getSubtarget();
237 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
238 FrameSetupOpcode = TII.getCallFrameSetupOpcode();
239 FrameDestroyOpcode = TII.getCallFrameDestroyOpcode();
240 SP = Subtarget.getTargetLowering()->getStackPointerRegisterToSaveRestore();
241 Entry = &MF.front();
242 CurrentCSRs.clear();
243 MachineFunc = &MF;
244
245 ++NumFunc;
246 }
247
248 /// Check whether or not Save and Restore points are still interesting for
249 /// shrink-wrapping.
250 bool ArePointsInteresting() const { return Save != Entry && Save && Restore; }
251
252 /// Check if shrink wrapping is enabled for this target and function.
253 static bool isShrinkWrapEnabled(const MachineFunction &MF);
254
255public:
256 static char ID;
257
258 ShrinkWrap() : MachineFunctionPass(ID) {
259 initializeShrinkWrapPass(*PassRegistry::getPassRegistry());
260 }
261
262 void getAnalysisUsage(AnalysisUsage &AU) const override {
263 AU.setPreservesAll();
264 AU.addRequired<MachineBlockFrequencyInfo>();
265 AU.addRequired<MachineDominatorTree>();
266 AU.addRequired<MachinePostDominatorTree>();
267 AU.addRequired<MachineLoopInfo>();
268 AU.addRequired<MachineOptimizationRemarkEmitterPass>();
269 MachineFunctionPass::getAnalysisUsage(AU);
270 }
271
272 MachineFunctionProperties getRequiredProperties() const override {
273 return MachineFunctionProperties().set(
274 MachineFunctionProperties::Property::NoVRegs);
275 }
276
277 StringRef getPassName() const override { return "Shrink Wrapping analysis"; }
278
279 /// Perform the shrink-wrapping analysis and update
280 /// the MachineFrameInfo attached to \p MF with the results.
281 bool runOnMachineFunction(MachineFunction &MF) override;
282};
283
284} // end anonymous namespace
285
286char ShrinkWrap::ID = 0;
287
288char &llvm::ShrinkWrapID = ShrinkWrap::ID;
289
290INITIALIZE_PASS_BEGIN(ShrinkWrap, DEBUG_TYPE, "Shrink Wrap Pass", false, false)
291INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
292INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
293INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
294INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
295INITIALIZE_PASS_DEPENDENCY(MachineOptimizationRemarkEmitterPass)
296INITIALIZE_PASS_END(ShrinkWrap, DEBUG_TYPE, "Shrink Wrap Pass", false, false)
297
298bool ShrinkWrap::useOrDefCSROrFI(const MachineInstr &MI, RegScavenger *RS,
299 bool StackAddressUsed) const {
300 /// Check if \p Op is known to access an address not on the function's stack .
301 /// At the moment, accesses where the underlying object is a global, function
302 /// argument, or jump table are considered non-stack accesses. Note that the
303 /// caller's stack may get accessed when passing an argument via the stack,
304 /// but not the stack of the current function.
305 ///
306 auto IsKnownNonStackPtr = [](MachineMemOperand *Op) {
307 if (Op->getValue()) {
308 const Value *UO = getUnderlyingObject(V: Op->getValue());
309 if (!UO)
310 return false;
311 if (auto *Arg = dyn_cast<Argument>(Val: UO))
312 return !Arg->hasPassPointeeByValueCopyAttr();
313 return isa<GlobalValue>(Val: UO);
314 }
315 if (const PseudoSourceValue *PSV = Op->getPseudoValue())
316 return PSV->isJumpTable();
317 return false;
318 };
319 // Load/store operations may access the stack indirectly when we previously
320 // computed an address to a stack location.
321 if (StackAddressUsed && MI.mayLoadOrStore() &&
322 (MI.isCall() || MI.hasUnmodeledSideEffects() || MI.memoperands_empty() ||
323 !all_of(Range: MI.memoperands(), P: IsKnownNonStackPtr)))
324 return true;
325
326 if (MI.getOpcode() == FrameSetupOpcode ||
327 MI.getOpcode() == FrameDestroyOpcode) {
328 LLVM_DEBUG(dbgs() << "Frame instruction: " << MI << '\n');
329 return true;
330 }
331 const MachineFunction *MF = MI.getParent()->getParent();
332 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
333 for (const MachineOperand &MO : MI.operands()) {
334 bool UseOrDefCSR = false;
335 if (MO.isReg()) {
336 // Ignore instructions like DBG_VALUE which don't read/def the register.
337 if (!MO.isDef() && !MO.readsReg())
338 continue;
339 Register PhysReg = MO.getReg();
340 if (!PhysReg)
341 continue;
342 assert(PhysReg.isPhysical() && "Unallocated register?!");
343 // The stack pointer is not normally described as a callee-saved register
344 // in calling convention definitions, so we need to watch for it
345 // separately. An SP mentioned by a call instruction, we can ignore,
346 // though, as it's harmless and we do not want to effectively disable tail
347 // calls by forcing the restore point to post-dominate them.
348 // PPC's LR is also not normally described as a callee-saved register in
349 // calling convention definitions, so we need to watch for it, too. An LR
350 // mentioned implicitly by a return (or "branch to link register")
351 // instruction we can ignore, otherwise we may pessimize shrinkwrapping.
352 UseOrDefCSR =
353 (!MI.isCall() && PhysReg == SP) ||
354 RCI.getLastCalleeSavedAlias(PhysReg) ||
355 (!MI.isReturn() && TRI->isNonallocatableRegisterCalleeSave(Reg: PhysReg));
356 } else if (MO.isRegMask()) {
357 // Check if this regmask clobbers any of the CSRs.
358 for (unsigned Reg : getCurrentCSRs(RS)) {
359 if (MO.clobbersPhysReg(PhysReg: Reg)) {
360 UseOrDefCSR = true;
361 break;
362 }
363 }
364 }
365 // Skip FrameIndex operands in DBG_VALUE instructions.
366 if (UseOrDefCSR || (MO.isFI() && !MI.isDebugValue())) {
367 LLVM_DEBUG(dbgs() << "Use or define CSR(" << UseOrDefCSR << ") or FI("
368 << MO.isFI() << "): " << MI << '\n');
369 return true;
370 }
371 }
372 return false;
373}
374
375/// Helper function to find the immediate (post) dominator.
376template <typename ListOfBBs, typename DominanceAnalysis>
377static MachineBasicBlock *FindIDom(MachineBasicBlock &Block, ListOfBBs BBs,
378 DominanceAnalysis &Dom, bool Strict = true) {
379 MachineBasicBlock *IDom = &Block;
380 for (MachineBasicBlock *BB : BBs) {
381 IDom = Dom.findNearestCommonDominator(IDom, BB);
382 if (!IDom)
383 break;
384 }
385 if (Strict && IDom == &Block)
386 return nullptr;
387 return IDom;
388}
389
390static bool isAnalyzableBB(const TargetInstrInfo &TII,
391 MachineBasicBlock &Entry) {
392 // Check if the block is analyzable.
393 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
394 SmallVector<MachineOperand, 4> Cond;
395 return !TII.analyzeBranch(MBB&: Entry, TBB, FBB, Cond);
396}
397
398/// Determines if any predecessor of MBB is on the path from block that has use
399/// or def of CSRs/FI to MBB.
400/// ReachableByDirty: All blocks reachable from block that has use or def of
401/// CSR/FI.
402static bool
403hasDirtyPred(const DenseSet<const MachineBasicBlock *> &ReachableByDirty,
404 const MachineBasicBlock &MBB) {
405 for (const MachineBasicBlock *PredBB : MBB.predecessors())
406 if (ReachableByDirty.count(V: PredBB))
407 return true;
408 return false;
409}
410
411/// Derives the list of all the basic blocks reachable from MBB.
412static void markAllReachable(DenseSet<const MachineBasicBlock *> &Visited,
413 const MachineBasicBlock &MBB) {
414 SmallVector<MachineBasicBlock *, 4> Worklist(MBB.succ_begin(),
415 MBB.succ_end());
416 Visited.insert(V: &MBB);
417 while (!Worklist.empty()) {
418 MachineBasicBlock *SuccMBB = Worklist.pop_back_val();
419 if (!Visited.insert(V: SuccMBB).second)
420 continue;
421 Worklist.append(in_start: SuccMBB->succ_begin(), in_end: SuccMBB->succ_end());
422 }
423}
424
425/// Collect blocks reachable by use or def of CSRs/FI.
426static void collectBlocksReachableByDirty(
427 const DenseSet<const MachineBasicBlock *> &DirtyBBs,
428 DenseSet<const MachineBasicBlock *> &ReachableByDirty) {
429 for (const MachineBasicBlock *MBB : DirtyBBs) {
430 if (ReachableByDirty.count(V: MBB))
431 continue;
432 // Mark all offsprings as reachable.
433 markAllReachable(Visited&: ReachableByDirty, MBB: *MBB);
434 }
435}
436
437/// \return true if there is a clean path from SavePoint to the original
438/// Restore.
439static bool
440isSaveReachableThroughClean(const MachineBasicBlock *SavePoint,
441 ArrayRef<MachineBasicBlock *> CleanPreds) {
442 DenseSet<const MachineBasicBlock *> Visited;
443 SmallVector<MachineBasicBlock *, 4> Worklist(CleanPreds.begin(),
444 CleanPreds.end());
445 while (!Worklist.empty()) {
446 MachineBasicBlock *CleanBB = Worklist.pop_back_val();
447 if (CleanBB == SavePoint)
448 return true;
449 if (!Visited.insert(V: CleanBB).second || !CleanBB->pred_size())
450 continue;
451 Worklist.append(in_start: CleanBB->pred_begin(), in_end: CleanBB->pred_end());
452 }
453 return false;
454}
455
456/// This function updates the branches post restore point split.
457///
458/// Restore point has been split.
459/// Old restore point: MBB
460/// New restore point: NMBB
461/// Any basic block(say BBToUpdate) which had a fallthrough to MBB
462/// previously should
463/// 1. Fallthrough to NMBB iff NMBB is inserted immediately above MBB in the
464/// block layout OR
465/// 2. Branch unconditionally to NMBB iff NMBB is inserted at any other place.
466static void updateTerminator(MachineBasicBlock *BBToUpdate,
467 MachineBasicBlock *NMBB,
468 const TargetInstrInfo *TII) {
469 DebugLoc DL = BBToUpdate->findBranchDebugLoc();
470 // if NMBB isn't the new layout successor for BBToUpdate, insert unconditional
471 // branch to it
472 if (!BBToUpdate->isLayoutSuccessor(MBB: NMBB))
473 TII->insertUnconditionalBranch(MBB&: *BBToUpdate, DestBB: NMBB, DL);
474}
475
476/// This function splits the restore point and returns new restore point/BB.
477///
478/// DirtyPreds: Predessors of \p MBB that are ReachableByDirty
479///
480/// Decision has been made to split the restore point.
481/// old restore point: \p MBB
482/// new restore point: \p NMBB
483/// This function makes the necessary block layout changes so that
484/// 1. \p NMBB points to \p MBB unconditionally
485/// 2. All dirtyPreds that previously pointed to \p MBB point to \p NMBB
486static MachineBasicBlock *
487tryToSplitRestore(MachineBasicBlock *MBB,
488 ArrayRef<MachineBasicBlock *> DirtyPreds,
489 const TargetInstrInfo *TII) {
490 MachineFunction *MF = MBB->getParent();
491
492 // get the list of DirtyPreds who have a fallthrough to MBB
493 // before the block layout change. This is just to ensure that if the NMBB is
494 // inserted after MBB, then we create unconditional branch from
495 // DirtyPred/CleanPred to NMBB
496 SmallPtrSet<MachineBasicBlock *, 8> MBBFallthrough;
497 for (MachineBasicBlock *BB : DirtyPreds)
498 if (BB->getFallThrough(JumpToFallThrough: false) == MBB)
499 MBBFallthrough.insert(Ptr: BB);
500
501 MachineBasicBlock *NMBB = MF->CreateMachineBasicBlock();
502 // Insert this block at the end of the function. Inserting in between may
503 // interfere with control flow optimizer decisions.
504 MF->insert(MBBI: MF->end(), MBB: NMBB);
505
506 for (const MachineBasicBlock::RegisterMaskPair &LI : MBB->liveins())
507 NMBB->addLiveIn(PhysReg: LI.PhysReg);
508
509 TII->insertUnconditionalBranch(MBB&: *NMBB, DestBB: MBB, DL: DebugLoc());
510
511 // After splitting, all predecessors of the restore point should be dirty
512 // blocks.
513 for (MachineBasicBlock *SuccBB : DirtyPreds)
514 SuccBB->ReplaceUsesOfBlockWith(Old: MBB, New: NMBB);
515
516 NMBB->addSuccessor(Succ: MBB);
517
518 for (MachineBasicBlock *BBToUpdate : MBBFallthrough)
519 updateTerminator(BBToUpdate, NMBB, TII);
520
521 return NMBB;
522}
523
524/// This function undoes the restore point split done earlier.
525///
526/// DirtyPreds: All predecessors of \p NMBB that are ReachableByDirty.
527///
528/// Restore point was split and the change needs to be unrolled. Make necessary
529/// changes to reset restore point from \p NMBB to \p MBB.
530static void rollbackRestoreSplit(MachineFunction &MF, MachineBasicBlock *NMBB,
531 MachineBasicBlock *MBB,
532 ArrayRef<MachineBasicBlock *> DirtyPreds,
533 const TargetInstrInfo *TII) {
534 // For a BB, if NMBB is fallthrough in the current layout, then in the new
535 // layout a. BB should fallthrough to MBB OR b. BB should undconditionally
536 // branch to MBB
537 SmallPtrSet<MachineBasicBlock *, 8> NMBBFallthrough;
538 for (MachineBasicBlock *BB : DirtyPreds)
539 if (BB->getFallThrough(JumpToFallThrough: false) == NMBB)
540 NMBBFallthrough.insert(Ptr: BB);
541
542 NMBB->removeSuccessor(Succ: MBB);
543 for (MachineBasicBlock *SuccBB : DirtyPreds)
544 SuccBB->ReplaceUsesOfBlockWith(Old: NMBB, New: MBB);
545
546 NMBB->erase(I: NMBB->begin(), E: NMBB->end());
547 NMBB->eraseFromParent();
548
549 for (MachineBasicBlock *BBToUpdate : NMBBFallthrough)
550 updateTerminator(BBToUpdate, NMBB: MBB, TII);
551}
552
553// A block is deemed fit for restore point split iff there exist
554// 1. DirtyPreds - preds of CurRestore reachable from use or def of CSR/FI
555// 2. CleanPreds - preds of CurRestore that arent DirtyPreds
556bool ShrinkWrap::checkIfRestoreSplittable(
557 const MachineBasicBlock *CurRestore,
558 const DenseSet<const MachineBasicBlock *> &ReachableByDirty,
559 SmallVectorImpl<MachineBasicBlock *> &DirtyPreds,
560 SmallVectorImpl<MachineBasicBlock *> &CleanPreds,
561 const TargetInstrInfo *TII, RegScavenger *RS) {
562 for (const MachineInstr &MI : *CurRestore)
563 if (useOrDefCSROrFI(MI, RS, /*StackAddressUsed=*/true))
564 return false;
565
566 for (MachineBasicBlock *PredBB : CurRestore->predecessors()) {
567 if (!isAnalyzableBB(TII: *TII, Entry&: *PredBB))
568 return false;
569
570 if (ReachableByDirty.count(V: PredBB))
571 DirtyPreds.push_back(Elt: PredBB);
572 else
573 CleanPreds.push_back(Elt: PredBB);
574 }
575
576 return !(CleanPreds.empty() || DirtyPreds.empty());
577}
578
579bool ShrinkWrap::postShrinkWrapping(bool HasCandidate, MachineFunction &MF,
580 RegScavenger *RS) {
581 if (!EnablePostShrinkWrapOpt)
582 return false;
583
584 MachineBasicBlock *InitSave = nullptr;
585 MachineBasicBlock *InitRestore = nullptr;
586
587 if (HasCandidate) {
588 InitSave = Save;
589 InitRestore = Restore;
590 } else {
591 InitRestore = nullptr;
592 InitSave = &MF.front();
593 for (MachineBasicBlock &MBB : MF) {
594 if (MBB.isEHFuncletEntry())
595 return false;
596 if (MBB.isReturnBlock()) {
597 // Do not support multiple restore points.
598 if (InitRestore)
599 return false;
600 InitRestore = &MBB;
601 }
602 }
603 }
604
605 if (!InitSave || !InitRestore || InitRestore == InitSave ||
606 !MDT->dominates(A: InitSave, B: InitRestore) ||
607 !MPDT->dominates(A: InitRestore, B: InitSave))
608 return false;
609
610 // Bail out of the optimization if any of the basic block is target of
611 // INLINEASM_BR instruction
612 for (MachineBasicBlock &MBB : MF)
613 if (MBB.isInlineAsmBrIndirectTarget())
614 return false;
615
616 DenseSet<const MachineBasicBlock *> DirtyBBs;
617 for (MachineBasicBlock &MBB : MF) {
618 if (MBB.isEHPad()) {
619 DirtyBBs.insert(V: &MBB);
620 continue;
621 }
622 for (const MachineInstr &MI : MBB)
623 if (useOrDefCSROrFI(MI, RS, /*StackAddressUsed=*/true)) {
624 DirtyBBs.insert(V: &MBB);
625 break;
626 }
627 }
628
629 // Find blocks reachable from the use or def of CSRs/FI.
630 DenseSet<const MachineBasicBlock *> ReachableByDirty;
631 collectBlocksReachableByDirty(DirtyBBs, ReachableByDirty);
632
633 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
634 SmallVector<MachineBasicBlock *, 2> DirtyPreds;
635 SmallVector<MachineBasicBlock *, 2> CleanPreds;
636 if (!checkIfRestoreSplittable(CurRestore: InitRestore, ReachableByDirty, DirtyPreds,
637 CleanPreds, TII, RS))
638 return false;
639
640 // Trying to reach out to the new save point which dominates all dirty blocks.
641 MachineBasicBlock *NewSave =
642 FindIDom<>(Block&: **DirtyPreds.begin(), BBs: DirtyPreds, Dom&: *MDT, Strict: false);
643
644 while (NewSave && (hasDirtyPred(ReachableByDirty, MBB: *NewSave) ||
645 EntryFreq < MBFI->getBlockFreq(MBB: NewSave) ||
646 /*Entry freq has been observed more than a loop block in
647 some cases*/
648 MLI->getLoopFor(BB: NewSave)))
649 NewSave = FindIDom<>(Block&: **NewSave->pred_begin(), BBs: NewSave->predecessors(), Dom&: *MDT,
650 Strict: false);
651
652 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
653 if (!NewSave || NewSave == InitSave ||
654 isSaveReachableThroughClean(SavePoint: NewSave, CleanPreds) ||
655 !TFI->canUseAsPrologue(MBB: *NewSave))
656 return false;
657
658 // Now we know that splitting a restore point can isolate the restore point
659 // from clean blocks and doing so can shrink the save point.
660 MachineBasicBlock *NewRestore =
661 tryToSplitRestore(MBB: InitRestore, DirtyPreds, TII);
662
663 // Make sure if the new restore point is valid as an epilogue, depending on
664 // targets.
665 if (!TFI->canUseAsEpilogue(MBB: *NewRestore)) {
666 rollbackRestoreSplit(MF, NMBB: NewRestore, MBB: InitRestore, DirtyPreds, TII);
667 return false;
668 }
669
670 Save = NewSave;
671 Restore = NewRestore;
672
673 MDT->runOnMachineFunction(F&: MF);
674 MPDT->runOnMachineFunction(MF);
675
676 assert((MDT->dominates(Save, Restore) && MPDT->dominates(Restore, Save)) &&
677 "Incorrect save or restore point due to dominance relations");
678 assert((!MLI->getLoopFor(Save) && !MLI->getLoopFor(Restore)) &&
679 "Unexpected save or restore point in a loop");
680 assert((EntryFreq >= MBFI->getBlockFreq(Save) &&
681 EntryFreq >= MBFI->getBlockFreq(Restore)) &&
682 "Incorrect save or restore point based on block frequency");
683 return true;
684}
685
686void ShrinkWrap::updateSaveRestorePoints(MachineBasicBlock &MBB,
687 RegScavenger *RS) {
688 // Get rid of the easy cases first.
689 if (!Save)
690 Save = &MBB;
691 else
692 Save = MDT->findNearestCommonDominator(A: Save, B: &MBB);
693 assert(Save);
694
695 if (!Restore)
696 Restore = &MBB;
697 else if (MPDT->getNode(BB: &MBB)) // If the block is not in the post dom tree, it
698 // means the block never returns. If that's the
699 // case, we don't want to call
700 // `findNearestCommonDominator`, which will
701 // return `Restore`.
702 Restore = MPDT->findNearestCommonDominator(A: Restore, B: &MBB);
703 else
704 Restore = nullptr; // Abort, we can't find a restore point in this case.
705
706 // Make sure we would be able to insert the restore code before the
707 // terminator.
708 if (Restore == &MBB) {
709 for (const MachineInstr &Terminator : MBB.terminators()) {
710 if (!useOrDefCSROrFI(MI: Terminator, RS, /*StackAddressUsed=*/true))
711 continue;
712 // One of the terminator needs to happen before the restore point.
713 if (MBB.succ_empty()) {
714 Restore = nullptr; // Abort, we can't find a restore point in this case.
715 break;
716 }
717 // Look for a restore point that post-dominates all the successors.
718 // The immediate post-dominator is what we are looking for.
719 Restore = FindIDom<>(Block&: *Restore, BBs: Restore->successors(), Dom&: *MPDT);
720 break;
721 }
722 }
723
724 if (!Restore) {
725 LLVM_DEBUG(
726 dbgs() << "Restore point needs to be spanned on several blocks\n");
727 return;
728 }
729
730 // Make sure Save and Restore are suitable for shrink-wrapping:
731 // 1. all path from Save needs to lead to Restore before exiting.
732 // 2. all path to Restore needs to go through Save from Entry.
733 // We achieve that by making sure that:
734 // A. Save dominates Restore.
735 // B. Restore post-dominates Save.
736 // C. Save and Restore are in the same loop.
737 bool SaveDominatesRestore = false;
738 bool RestorePostDominatesSave = false;
739 while (Restore &&
740 (!(SaveDominatesRestore = MDT->dominates(A: Save, B: Restore)) ||
741 !(RestorePostDominatesSave = MPDT->dominates(A: Restore, B: Save)) ||
742 // Post-dominance is not enough in loops to ensure that all uses/defs
743 // are after the prologue and before the epilogue at runtime.
744 // E.g.,
745 // while(1) {
746 // Save
747 // Restore
748 // if (...)
749 // break;
750 // use/def CSRs
751 // }
752 // All the uses/defs of CSRs are dominated by Save and post-dominated
753 // by Restore. However, the CSRs uses are still reachable after
754 // Restore and before Save are executed.
755 //
756 // For now, just push the restore/save points outside of loops.
757 // FIXME: Refine the criteria to still find interesting cases
758 // for loops.
759 MLI->getLoopFor(BB: Save) || MLI->getLoopFor(BB: Restore))) {
760 // Fix (A).
761 if (!SaveDominatesRestore) {
762 Save = MDT->findNearestCommonDominator(A: Save, B: Restore);
763 continue;
764 }
765 // Fix (B).
766 if (!RestorePostDominatesSave)
767 Restore = MPDT->findNearestCommonDominator(A: Restore, B: Save);
768
769 // Fix (C).
770 if (Restore && (MLI->getLoopFor(BB: Save) || MLI->getLoopFor(BB: Restore))) {
771 if (MLI->getLoopDepth(BB: Save) > MLI->getLoopDepth(BB: Restore)) {
772 // Push Save outside of this loop if immediate dominator is different
773 // from save block. If immediate dominator is not different, bail out.
774 Save = FindIDom<>(Block&: *Save, BBs: Save->predecessors(), Dom&: *MDT);
775 if (!Save)
776 break;
777 } else {
778 // If the loop does not exit, there is no point in looking
779 // for a post-dominator outside the loop.
780 SmallVector<MachineBasicBlock*, 4> ExitBlocks;
781 MLI->getLoopFor(BB: Restore)->getExitingBlocks(ExitingBlocks&: ExitBlocks);
782 // Push Restore outside of this loop.
783 // Look for the immediate post-dominator of the loop exits.
784 MachineBasicBlock *IPdom = Restore;
785 for (MachineBasicBlock *LoopExitBB: ExitBlocks) {
786 IPdom = FindIDom<>(Block&: *IPdom, BBs: LoopExitBB->successors(), Dom&: *MPDT);
787 if (!IPdom)
788 break;
789 }
790 // If the immediate post-dominator is not in a less nested loop,
791 // then we are stuck in a program with an infinite loop.
792 // In that case, we will not find a safe point, hence, bail out.
793 if (IPdom && MLI->getLoopDepth(BB: IPdom) < MLI->getLoopDepth(BB: Restore))
794 Restore = IPdom;
795 else {
796 Restore = nullptr;
797 break;
798 }
799 }
800 }
801 }
802}
803
804static bool giveUpWithRemarks(MachineOptimizationRemarkEmitter *ORE,
805 StringRef RemarkName, StringRef RemarkMessage,
806 const DiagnosticLocation &Loc,
807 const MachineBasicBlock *MBB) {
808 ORE->emit(RemarkBuilder: [&]() {
809 return MachineOptimizationRemarkMissed(DEBUG_TYPE, RemarkName, Loc, MBB)
810 << RemarkMessage;
811 });
812
813 LLVM_DEBUG(dbgs() << RemarkMessage << '\n');
814 return false;
815}
816
817bool ShrinkWrap::performShrinkWrapping(
818 const ReversePostOrderTraversal<MachineBasicBlock *> &RPOT,
819 RegScavenger *RS) {
820 for (MachineBasicBlock *MBB : RPOT) {
821 LLVM_DEBUG(dbgs() << "Look into: " << printMBBReference(*MBB) << '\n');
822
823 if (MBB->isEHFuncletEntry())
824 return giveUpWithRemarks(ORE, RemarkName: "UnsupportedEHFunclets",
825 RemarkMessage: "EH Funclets are not supported yet.",
826 Loc: MBB->front().getDebugLoc(), MBB);
827
828 if (MBB->isEHPad() || MBB->isInlineAsmBrIndirectTarget()) {
829 // Push the prologue and epilogue outside of the region that may throw (or
830 // jump out via inlineasm_br), by making sure that all the landing pads
831 // are at least at the boundary of the save and restore points. The
832 // problem is that a basic block can jump out from the middle in these
833 // cases, which we do not handle.
834 updateSaveRestorePoints(MBB&: *MBB, RS);
835 if (!ArePointsInteresting()) {
836 LLVM_DEBUG(dbgs() << "EHPad/inlineasm_br prevents shrink-wrapping\n");
837 return false;
838 }
839 continue;
840 }
841
842 bool StackAddressUsed = false;
843 // Check if we found any stack accesses in the predecessors. We are not
844 // doing a full dataflow analysis here to keep things simple but just
845 // rely on a reverse portorder traversal (RPOT) to guarantee predecessors
846 // are already processed except for loops (and accept the conservative
847 // result for loops).
848 for (const MachineBasicBlock *Pred : MBB->predecessors()) {
849 if (StackAddressUsedBlockInfo.test(Idx: Pred->getNumber())) {
850 StackAddressUsed = true;
851 break;
852 }
853 }
854
855 for (const MachineInstr &MI : *MBB) {
856 if (useOrDefCSROrFI(MI, RS, StackAddressUsed)) {
857 // Save (resp. restore) point must dominate (resp. post dominate)
858 // MI. Look for the proper basic block for those.
859 updateSaveRestorePoints(MBB&: *MBB, RS);
860 // If we are at a point where we cannot improve the placement of
861 // save/restore instructions, just give up.
862 if (!ArePointsInteresting()) {
863 LLVM_DEBUG(dbgs() << "No Shrink wrap candidate found\n");
864 return false;
865 }
866 // No need to look for other instructions, this basic block
867 // will already be part of the handled region.
868 StackAddressUsed = true;
869 break;
870 }
871 }
872 StackAddressUsedBlockInfo[MBB->getNumber()] = StackAddressUsed;
873 }
874 if (!ArePointsInteresting()) {
875 // If the points are not interesting at this point, then they must be null
876 // because it means we did not encounter any frame/CSR related code.
877 // Otherwise, we would have returned from the previous loop.
878 assert(!Save && !Restore && "We miss a shrink-wrap opportunity?!");
879 LLVM_DEBUG(dbgs() << "Nothing to shrink-wrap\n");
880 return false;
881 }
882
883 LLVM_DEBUG(dbgs() << "\n ** Results **\nFrequency of the Entry: "
884 << EntryFreq.getFrequency() << '\n');
885
886 const TargetFrameLowering *TFI =
887 MachineFunc->getSubtarget().getFrameLowering();
888 do {
889 LLVM_DEBUG(dbgs() << "Shrink wrap candidates (#, Name, Freq):\nSave: "
890 << printMBBReference(*Save) << ' '
891 << printBlockFreq(*MBFI, *Save)
892 << "\nRestore: " << printMBBReference(*Restore) << ' '
893 << printBlockFreq(*MBFI, *Restore) << '\n');
894
895 bool IsSaveCheap, TargetCanUseSaveAsPrologue = false;
896 if (((IsSaveCheap = EntryFreq >= MBFI->getBlockFreq(MBB: Save)) &&
897 EntryFreq >= MBFI->getBlockFreq(MBB: Restore)) &&
898 ((TargetCanUseSaveAsPrologue = TFI->canUseAsPrologue(MBB: *Save)) &&
899 TFI->canUseAsEpilogue(MBB: *Restore)))
900 break;
901 LLVM_DEBUG(
902 dbgs() << "New points are too expensive or invalid for the target\n");
903 MachineBasicBlock *NewBB;
904 if (!IsSaveCheap || !TargetCanUseSaveAsPrologue) {
905 Save = FindIDom<>(Block&: *Save, BBs: Save->predecessors(), Dom&: *MDT);
906 if (!Save)
907 break;
908 NewBB = Save;
909 } else {
910 // Restore is expensive.
911 Restore = FindIDom<>(Block&: *Restore, BBs: Restore->successors(), Dom&: *MPDT);
912 if (!Restore)
913 break;
914 NewBB = Restore;
915 }
916 updateSaveRestorePoints(MBB&: *NewBB, RS);
917 } while (Save && Restore);
918
919 if (!ArePointsInteresting()) {
920 ++NumCandidatesDropped;
921 return false;
922 }
923 return true;
924}
925
926bool ShrinkWrap::runOnMachineFunction(MachineFunction &MF) {
927 if (skipFunction(F: MF.getFunction()) || MF.empty() || !isShrinkWrapEnabled(MF))
928 return false;
929
930 LLVM_DEBUG(dbgs() << "**** Analysing " << MF.getName() << '\n');
931
932 init(MF);
933
934 ReversePostOrderTraversal<MachineBasicBlock *> RPOT(&*MF.begin());
935 if (containsIrreducibleCFG<MachineBasicBlock *>(RPOTraversal&: RPOT, LI: *MLI)) {
936 // If MF is irreducible, a block may be in a loop without
937 // MachineLoopInfo reporting it. I.e., we may use the
938 // post-dominance property in loops, which lead to incorrect
939 // results. Moreover, we may miss that the prologue and
940 // epilogue are not in the same loop, leading to unbalanced
941 // construction/deconstruction of the stack frame.
942 return giveUpWithRemarks(ORE, RemarkName: "UnsupportedIrreducibleCFG",
943 RemarkMessage: "Irreducible CFGs are not supported yet.",
944 Loc: MF.getFunction().getSubprogram(), MBB: &MF.front());
945 }
946
947 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
948 std::unique_ptr<RegScavenger> RS(
949 TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : nullptr);
950
951 bool Changed = false;
952
953 // Initially, conservatively assume that stack addresses can be used in each
954 // basic block and change the state only for those basic blocks for which we
955 // were able to prove the opposite.
956 StackAddressUsedBlockInfo.resize(N: MF.getNumBlockIDs(), t: true);
957 bool HasCandidate = performShrinkWrapping(RPOT, RS: RS.get());
958 StackAddressUsedBlockInfo.clear();
959 Changed = postShrinkWrapping(HasCandidate, MF, RS: RS.get());
960 if (!HasCandidate && !Changed)
961 return false;
962 if (!ArePointsInteresting())
963 return Changed;
964
965 LLVM_DEBUG(dbgs() << "Final shrink wrap candidates:\nSave: "
966 << printMBBReference(*Save) << ' '
967 << "\nRestore: " << printMBBReference(*Restore) << '\n');
968
969 MachineFrameInfo &MFI = MF.getFrameInfo();
970 MFI.setSavePoint(Save);
971 MFI.setRestorePoint(Restore);
972 ++NumCandidates;
973 return Changed;
974}
975
976bool ShrinkWrap::isShrinkWrapEnabled(const MachineFunction &MF) {
977 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
978
979 switch (EnableShrinkWrapOpt) {
980 case cl::BOU_UNSET:
981 return TFI->enableShrinkWrapping(MF) &&
982 // Windows with CFI has some limitations that make it impossible
983 // to use shrink-wrapping.
984 !MF.getTarget().getMCAsmInfo()->usesWindowsCFI() &&
985 // Sanitizers look at the value of the stack at the location
986 // of the crash. Since a crash can happen anywhere, the
987 // frame must be lowered before anything else happen for the
988 // sanitizers to be able to get a correct stack frame.
989 !(MF.getFunction().hasFnAttribute(Attribute::SanitizeAddress) ||
990 MF.getFunction().hasFnAttribute(Attribute::SanitizeThread) ||
991 MF.getFunction().hasFnAttribute(Attribute::SanitizeMemory) ||
992 MF.getFunction().hasFnAttribute(Attribute::SanitizeHWAddress));
993 // If EnableShrinkWrap is set, it takes precedence on whatever the
994 // target sets. The rational is that we assume we want to test
995 // something related to shrink-wrapping.
996 case cl::BOU_TRUE:
997 return true;
998 case cl::BOU_FALSE:
999 return false;
1000 }
1001 llvm_unreachable("Invalid shrink-wrapping state");
1002}
1003

source code of llvm/lib/CodeGen/ShrinkWrap.cpp