Line data Source code
1 : //===- BranchFolding.cpp - Fold machine code branch instructions ----------===//
2 : //
3 : // The LLVM Compiler Infrastructure
4 : //
5 : // This file is distributed under the University of Illinois Open Source
6 : // License. See LICENSE.TXT for details.
7 : //
8 : //===----------------------------------------------------------------------===//
9 : //
10 : // This pass forwards branches to unconditional branches to make them branch
11 : // directly to the target block. This pass often results in dead MBB's, which
12 : // it then removes.
13 : //
14 : // Note that this pass must be run after register allocation, it cannot handle
15 : // SSA form. It also must handle virtual registers for targets that emit virtual
16 : // ISA (e.g. NVPTX).
17 : //
18 : //===----------------------------------------------------------------------===//
19 :
20 : #include "BranchFolding.h"
21 : #include "llvm/ADT/BitVector.h"
22 : #include "llvm/ADT/DenseMap.h"
23 : #include "llvm/ADT/STLExtras.h"
24 : #include "llvm/ADT/SmallPtrSet.h"
25 : #include "llvm/ADT/SmallSet.h"
26 : #include "llvm/ADT/SmallVector.h"
27 : #include "llvm/ADT/Statistic.h"
28 : #include "llvm/CodeGen/Analysis.h"
29 : #include "llvm/CodeGen/LivePhysRegs.h"
30 : #include "llvm/CodeGen/MachineBasicBlock.h"
31 : #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
32 : #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
33 : #include "llvm/CodeGen/MachineFunction.h"
34 : #include "llvm/CodeGen/MachineFunctionPass.h"
35 : #include "llvm/CodeGen/MachineInstr.h"
36 : #include "llvm/CodeGen/MachineInstrBuilder.h"
37 : #include "llvm/CodeGen/MachineJumpTableInfo.h"
38 : #include "llvm/CodeGen/MachineLoopInfo.h"
39 : #include "llvm/CodeGen/MachineModuleInfo.h"
40 : #include "llvm/CodeGen/MachineOperand.h"
41 : #include "llvm/CodeGen/MachineRegisterInfo.h"
42 : #include "llvm/CodeGen/TargetInstrInfo.h"
43 : #include "llvm/CodeGen/TargetOpcodes.h"
44 : #include "llvm/CodeGen/TargetPassConfig.h"
45 : #include "llvm/CodeGen/TargetRegisterInfo.h"
46 : #include "llvm/CodeGen/TargetSubtargetInfo.h"
47 : #include "llvm/IR/DebugInfoMetadata.h"
48 : #include "llvm/IR/DebugLoc.h"
49 : #include "llvm/IR/Function.h"
50 : #include "llvm/MC/LaneBitmask.h"
51 : #include "llvm/MC/MCRegisterInfo.h"
52 : #include "llvm/Pass.h"
53 : #include "llvm/Support/BlockFrequency.h"
54 : #include "llvm/Support/BranchProbability.h"
55 : #include "llvm/Support/CommandLine.h"
56 : #include "llvm/Support/Debug.h"
57 : #include "llvm/Support/ErrorHandling.h"
58 : #include "llvm/Support/raw_ostream.h"
59 : #include "llvm/Target/TargetMachine.h"
60 : #include <cassert>
61 : #include <cstddef>
62 : #include <iterator>
63 : #include <numeric>
64 : #include <vector>
65 :
66 : using namespace llvm;
67 :
68 : #define DEBUG_TYPE "branch-folder"
69 :
70 : STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
71 : STATISTIC(NumBranchOpts, "Number of branches optimized");
72 : STATISTIC(NumTailMerge , "Number of block tails merged");
73 : STATISTIC(NumHoist , "Number of times common instructions are hoisted");
74 : STATISTIC(NumTailCalls, "Number of tail calls optimized");
75 :
76 : static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
77 : cl::init(cl::BOU_UNSET), cl::Hidden);
78 :
79 : // Throttle for huge numbers of predecessors (compile speed problems)
80 : static cl::opt<unsigned>
81 : TailMergeThreshold("tail-merge-threshold",
82 : cl::desc("Max number of predecessors to consider tail merging"),
83 : cl::init(150), cl::Hidden);
84 :
85 : // Heuristic for tail merging (and, inversely, tail duplication).
86 : // TODO: This should be replaced with a target query.
87 : static cl::opt<unsigned>
88 : TailMergeSize("tail-merge-size",
89 : cl::desc("Min number of instructions to consider tail merging"),
90 : cl::init(3), cl::Hidden);
91 :
92 : namespace {
93 :
94 : /// BranchFolderPass - Wrap branch folder in a machine function pass.
95 : class BranchFolderPass : public MachineFunctionPass {
96 : public:
97 : static char ID;
98 :
99 20212 : explicit BranchFolderPass(): MachineFunctionPass(ID) {}
100 :
101 : bool runOnMachineFunction(MachineFunction &MF) override;
102 :
103 20053 : void getAnalysisUsage(AnalysisUsage &AU) const override {
104 : AU.addRequired<MachineBlockFrequencyInfo>();
105 : AU.addRequired<MachineBranchProbabilityInfo>();
106 : AU.addRequired<TargetPassConfig>();
107 20053 : MachineFunctionPass::getAnalysisUsage(AU);
108 20053 : }
109 : };
110 :
111 : } // end anonymous namespace
112 :
113 : char BranchFolderPass::ID = 0;
114 :
115 : char &llvm::BranchFolderPassID = BranchFolderPass::ID;
116 :
117 85147 : INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,
118 : "Control Flow Optimizer", false, false)
119 :
120 197862 : bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
121 197862 : if (skipFunction(MF.getFunction()))
122 : return false;
123 :
124 197676 : TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
125 : // TailMerge can create jump into if branches that make CFG irreducible for
126 : // HW that requires structurized CFG.
127 395352 : bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
128 194343 : PassConfig->getEnableTailMerge();
129 : BranchFolder::MBFIWrapper MBBFreqInfo(
130 197676 : getAnalysis<MachineBlockFrequencyInfo>());
131 : BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo,
132 395352 : getAnalysis<MachineBranchProbabilityInfo>());
133 395352 : return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(),
134 197676 : MF.getSubtarget().getRegisterInfo(),
135 : getAnalysisIfAvailable<MachineModuleInfo>());
136 : }
137 :
138 241482 : BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist,
139 : MBFIWrapper &FreqInfo,
140 : const MachineBranchProbabilityInfo &ProbInfo,
141 241482 : unsigned MinTailLength)
142 : : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength),
143 241482 : MBBFreqInfo(FreqInfo), MBPI(ProbInfo) {
144 241482 : if (MinCommonTailLength == 0)
145 233021 : MinCommonTailLength = TailMergeSize;
146 241482 : switch (FlagEnableTailMerge) {
147 241430 : case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
148 3 : case cl::BOU_TRUE: EnableTailMerge = true; break;
149 49 : case cl::BOU_FALSE: EnableTailMerge = false; break;
150 : }
151 241482 : }
152 :
153 15269 : void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
154 : assert(MBB->pred_empty() && "MBB must be dead!");
155 : LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
156 :
157 15269 : MachineFunction *MF = MBB->getParent();
158 : // drop all successors.
159 24202 : while (!MBB->succ_empty())
160 8933 : MBB->removeSuccessor(MBB->succ_end()-1);
161 :
162 : // Avoid matching if this pointer gets reused.
163 : TriedMerging.erase(MBB);
164 :
165 : // Remove the block.
166 : MF->erase(MBB);
167 15269 : EHScopeMembership.erase(MBB);
168 15269 : if (MLI)
169 : MLI->removeBlock(MBB);
170 15269 : }
171 :
172 241482 : bool BranchFolder::OptimizeFunction(MachineFunction &MF,
173 : const TargetInstrInfo *tii,
174 : const TargetRegisterInfo *tri,
175 : MachineModuleInfo *mmi,
176 : MachineLoopInfo *mli, bool AfterPlacement) {
177 241482 : if (!tii) return false;
178 :
179 241482 : TriedMerging.clear();
180 :
181 241482 : MachineRegisterInfo &MRI = MF.getRegInfo();
182 241482 : AfterBlockPlacement = AfterPlacement;
183 241482 : TII = tii;
184 241482 : TRI = tri;
185 241482 : MMI = mmi;
186 241482 : MLI = mli;
187 241482 : this->MRI = &MRI;
188 :
189 482964 : UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF);
190 241482 : if (!UpdateLiveIns)
191 9606 : MRI.invalidateLiveness();
192 :
193 : // Fix CFG. The later algorithms expect it to be right.
194 : bool MadeChange = false;
195 1006930 : for (MachineBasicBlock &MBB : MF) {
196 765449 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
197 : SmallVector<MachineOperand, 4> Cond;
198 765449 : if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true))
199 520186 : MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
200 : }
201 :
202 : // Recalculate EH scope membership.
203 482963 : EHScopeMembership = getEHScopeMembership(MF);
204 :
205 : bool MadeChangeThisIteration = true;
206 495729 : while (MadeChangeThisIteration) {
207 254247 : MadeChangeThisIteration = TailMergeBlocks(MF);
208 : // No need to clean up if tail merging does not change anything after the
209 : // block placement.
210 254246 : if (!AfterBlockPlacement || MadeChangeThisIteration)
211 245785 : MadeChangeThisIteration |= OptimizeBranches(MF);
212 254247 : if (EnableHoistCommonCode)
213 208111 : MadeChangeThisIteration |= HoistCommonCode(MF);
214 254247 : MadeChange |= MadeChangeThisIteration;
215 : }
216 :
217 : // See if any jump tables have become dead as the code generator
218 : // did its thing.
219 241482 : MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
220 241482 : if (!JTI)
221 : return MadeChange;
222 :
223 : // Walk the function to find jump tables that are live.
224 1246 : BitVector JTIsLive(JTI->getJumpTables().size());
225 21156 : for (const MachineBasicBlock &BB : MF) {
226 169735 : for (const MachineInstr &I : BB)
227 770307 : for (const MachineOperand &Op : I.operands()) {
228 621105 : if (!Op.isJTI()) continue;
229 :
230 : // Remember that this JT is live.
231 1007 : JTIsLive.set(Op.getIndex());
232 : }
233 : }
234 :
235 : // Finally, remove dead jump tables. This happens when the
236 : // indirect jump was unreachable (and thus deleted).
237 1316 : for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
238 693 : if (!JTIsLive.test(i)) {
239 : JTI->RemoveJumpTable(i);
240 : MadeChange = true;
241 : }
242 :
243 : return MadeChange;
244 : }
245 :
246 : //===----------------------------------------------------------------------===//
247 : // Tail Merging of Blocks
248 : //===----------------------------------------------------------------------===//
249 :
250 : /// HashMachineInstr - Compute a hash value for MI and its operands.
251 851713 : static unsigned HashMachineInstr(const MachineInstr &MI) {
252 851713 : unsigned Hash = MI.getOpcode();
253 4145373 : for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
254 3293660 : const MachineOperand &Op = MI.getOperand(i);
255 :
256 : // Merge in bits from the operand if easy. We can't use MachineOperand's
257 : // hash_code here because it's not deterministic and we sort by hash value
258 : // later.
259 : unsigned OperandHash = 0;
260 3293660 : switch (Op.getType()) {
261 2146776 : case MachineOperand::MO_Register:
262 2146776 : OperandHash = Op.getReg();
263 2146776 : break;
264 600526 : case MachineOperand::MO_Immediate:
265 600526 : OperandHash = Op.getImm();
266 600526 : break;
267 95075 : case MachineOperand::MO_MachineBasicBlock:
268 95075 : OperandHash = Op.getMBB()->getNumber();
269 95075 : break;
270 381 : case MachineOperand::MO_FrameIndex:
271 : case MachineOperand::MO_ConstantPoolIndex:
272 : case MachineOperand::MO_JumpTableIndex:
273 381 : OperandHash = Op.getIndex();
274 381 : break;
275 334778 : case MachineOperand::MO_GlobalAddress:
276 : case MachineOperand::MO_ExternalSymbol:
277 : // Global address / external symbol are too hard, don't bother, but do
278 : // pull in the offset.
279 334778 : OperandHash = Op.getOffset();
280 334778 : break;
281 : default:
282 : break;
283 : }
284 :
285 3293660 : Hash += ((OperandHash << 3) | Op.getType()) << (i & 31);
286 : }
287 851713 : return Hash;
288 : }
289 :
290 : /// HashEndOfMBB - Hash the last instruction in the MBB.
291 862005 : static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) {
292 : MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr();
293 862005 : if (I == MBB.end())
294 : return 0;
295 :
296 851713 : return HashMachineInstr(*I);
297 : }
298 :
299 : /// Whether MI should be counted as an instruction when calculating common tail.
300 : static bool countsAsInstruction(const MachineInstr &MI) {
301 6471992 : return !(MI.isDebugValue() || MI.isCFIInstruction());
302 : }
303 :
304 : /// ComputeCommonTailLength - Given two machine basic blocks, compute the number
305 : /// of instructions they actually have in common together at their end. Return
306 : /// iterators for the first shared instruction in each block.
307 2131948 : static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
308 : MachineBasicBlock *MBB2,
309 : MachineBasicBlock::iterator &I1,
310 : MachineBasicBlock::iterator &I2) {
311 2131948 : I1 = MBB1->end();
312 2131948 : I2 = MBB2->end();
313 :
314 : unsigned TailLen = 0;
315 5869912 : while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
316 : --I1; --I2;
317 : // Skip debugging pseudos; necessary to avoid changing the code.
318 : while (!countsAsInstruction(*I1)) {
319 234322 : if (I1==MBB1->begin()) {
320 : while (!countsAsInstruction(*I2)) {
321 2740 : if (I2==MBB2->begin()) {
322 : // I1==DBG at begin; I2==DBG at begin
323 : goto SkipTopCFIAndReturn;
324 : }
325 : --I2;
326 : }
327 : ++I2;
328 : // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
329 889 : goto SkipTopCFIAndReturn;
330 : }
331 : --I1;
332 : }
333 : // I1==first (untested) non-DBG preceding known match
334 : while (!countsAsInstruction(*I2)) {
335 231589 : if (I2==MBB2->begin()) {
336 : ++I1;
337 : // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
338 183 : goto SkipTopCFIAndReturn;
339 : }
340 : --I2;
341 : }
342 : // I1, I2==first (untested) non-DBGs preceding known match
343 5612370 : if (!I1->isIdenticalTo(*I2) ||
344 : // FIXME: This check is dubious. It's used to get around a problem where
345 : // people incorrectly expect inline asm directives to remain in the same
346 : // relative order. This is untenable because normal compiler
347 : // optimizations (like this one) may reorder and/or merge these
348 : // directives.
349 : I1->isInlineAsm()) {
350 : ++I1; ++I2;
351 : break;
352 : }
353 3737964 : ++TailLen;
354 : }
355 : // Back past possible debugging pseudos at beginning of block. This matters
356 : // when one block differs from the other only by whether debugging pseudos
357 : // are present at the beginning. (This way, the various checks later for
358 : // I1==MBB1->begin() work as expected.)
359 2130731 : if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
360 : --I2;
361 : while (I2->isDebugInstr()) {
362 56 : if (I2 == MBB2->begin())
363 : return TailLen;
364 : --I2;
365 : }
366 : ++I2;
367 : }
368 2130714 : if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
369 : --I1;
370 : while (I1->isDebugInstr()) {
371 74 : if (I1 == MBB1->begin())
372 : return TailLen;
373 : --I1;
374 : }
375 : ++I1;
376 : }
377 :
378 2131907 : SkipTopCFIAndReturn:
379 : // Ensure that I1 and I2 do not point to a CFI_INSTRUCTION. This can happen if
380 : // I1 and I2 are non-identical when compared and then one or both of them ends
381 : // up pointing to a CFI instruction after being incremented. For example:
382 : /*
383 : BB1:
384 : ...
385 : INSTRUCTION_A
386 : ADD32ri8 <- last common instruction
387 : ...
388 : BB2:
389 : ...
390 : INSTRUCTION_B
391 : CFI_INSTRUCTION
392 : ADD32ri8 <- last common instruction
393 : ...
394 : */
395 : // When INSTRUCTION_A and INSTRUCTION_B are compared as not equal, after
396 : // incrementing the iterators, I1 will point to ADD, however I2 will point to
397 : // the CFI instruction. Later on, this leads to BB2 being 'hacked off' at the
398 : // wrong place (in ReplaceTailWithBranchTo()) which results in losing this CFI
399 : // instruction.
400 2184360 : while (I1 != MBB1->end() && I1->isCFIInstruction()) {
401 : ++I1;
402 : }
403 :
404 2184359 : while (I2 != MBB2->end() && I2->isCFIInstruction()) {
405 : ++I2;
406 : }
407 :
408 : return TailLen;
409 : }
410 :
411 26347 : void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
412 : MachineBasicBlock &NewDest) {
413 26347 : if (UpdateLiveIns) {
414 : // OldInst should always point to an instruction.
415 26324 : MachineBasicBlock &OldMBB = *OldInst->getParent();
416 : LiveRegs.clear();
417 26324 : LiveRegs.addLiveOuts(OldMBB);
418 : // Move backward to the place where will insert the jump.
419 26324 : MachineBasicBlock::iterator I = OldMBB.end();
420 : do {
421 : --I;
422 105176 : LiveRegs.stepBackward(*I);
423 105176 : } while (I != OldInst);
424 :
425 : // Merging the tails may have switched some undef operand to non-undef ones.
426 : // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the
427 : // register.
428 77611 : for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) {
429 : // We computed the liveins with computeLiveIn earlier and should only see
430 : // full registers:
431 : assert(P.LaneMask == LaneBitmask::getAll() &&
432 : "Can only handle full register.");
433 : MCPhysReg Reg = P.PhysReg;
434 51287 : if (!LiveRegs.available(*MRI, Reg))
435 51286 : continue;
436 1 : DebugLoc DL;
437 2 : BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
438 : }
439 : }
440 :
441 26347 : TII->ReplaceTailWithBranchTo(OldInst, &NewDest);
442 : ++NumTailMerge;
443 26347 : }
444 :
445 7802 : MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
446 : MachineBasicBlock::iterator BBI1,
447 : const BasicBlock *BB) {
448 7802 : if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
449 : return nullptr;
450 :
451 7802 : MachineFunction &MF = *CurMBB.getParent();
452 :
453 : // Create the fall-through block.
454 7802 : MachineFunction::iterator MBBI = CurMBB.getIterator();
455 7802 : MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB);
456 7802 : CurMBB.getParent()->insert(++MBBI, NewMBB);
457 :
458 : // Move all the successors of this block to the specified block.
459 7802 : NewMBB->transferSuccessors(&CurMBB);
460 :
461 : // Add an edge from CurMBB to NewMBB for the fall-through.
462 7802 : CurMBB.addSuccessor(NewMBB);
463 :
464 : // Splice the code over.
465 : NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
466 :
467 : // NewMBB belongs to the same loop as CurMBB.
468 7802 : if (MLI)
469 26 : if (MachineLoop *ML = MLI->getLoopFor(&CurMBB))
470 52 : ML->addBasicBlockToLoop(NewMBB, MLI->getBase());
471 :
472 : // NewMBB inherits CurMBB's block frequency.
473 7802 : MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
474 :
475 7802 : if (UpdateLiveIns)
476 7793 : computeAndAddLiveIns(LiveRegs, *NewMBB);
477 :
478 : // Add the new block to the EH scope.
479 7802 : const auto &EHScopeI = EHScopeMembership.find(&CurMBB);
480 7802 : if (EHScopeI != EHScopeMembership.end()) {
481 0 : auto n = EHScopeI->second;
482 0 : EHScopeMembership[NewMBB] = n;
483 : }
484 :
485 : return NewMBB;
486 : }
487 :
488 : /// EstimateRuntime - Make a rough estimate for how long it will take to run
489 : /// the specified code.
490 22646 : static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
491 : MachineBasicBlock::iterator E) {
492 : unsigned Time = 0;
493 66049 : for (; I != E; ++I) {
494 : if (!countsAsInstruction(*I))
495 : continue;
496 38873 : if (I->isCall())
497 212 : Time += 10;
498 38661 : else if (I->mayLoad() || I->mayStore())
499 11646 : Time += 2;
500 : else
501 27015 : ++Time;
502 : }
503 22646 : return Time;
504 : }
505 :
506 : // CurMBB needs to add an unconditional branch to SuccMBB (we removed these
507 : // branches temporarily for tail merging). In the case where CurMBB ends
508 : // with a conditional branch to the next block, optimize by reversing the
509 : // test and conditionally branching to SuccMBB instead.
510 366665 : static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
511 : const TargetInstrInfo *TII) {
512 366665 : MachineFunction *MF = CurMBB->getParent();
513 : MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB));
514 366665 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
515 : SmallVector<MachineOperand, 4> Cond;
516 366665 : DebugLoc dl = CurMBB->findBranchDebugLoc();
517 366665 : if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
518 : MachineBasicBlock *NextBB = &*I;
519 362425 : if (TBB == NextBB && !Cond.empty() && !FBB) {
520 72612 : if (!TII->reverseBranchCondition(Cond)) {
521 72610 : TII->removeBranch(*CurMBB);
522 145220 : TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl);
523 : return;
524 : }
525 : }
526 : }
527 294055 : TII->insertBranch(*CurMBB, SuccBB, nullptr,
528 588110 : SmallVector<MachineOperand, 0>(), dl);
529 : }
530 :
531 : bool
532 0 : BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
533 937534 : if (getHash() < o.getHash())
534 : return true;
535 437450 : if (getHash() > o.getHash())
536 : return false;
537 387639 : if (getBlock()->getNumber() < o.getBlock()->getNumber())
538 : return true;
539 134752 : if (getBlock()->getNumber() > o.getBlock()->getNumber())
540 : return false;
541 : // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
542 : // an object with itself.
543 : #ifndef _GLIBCXX_DEBUG
544 0 : llvm_unreachable("Predecessor appears twice");
545 : #else
546 : return false;
547 : #endif
548 : }
549 :
550 : BlockFrequency
551 2557158 : BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const {
552 2557158 : auto I = MergedBBFreq.find(MBB);
553 :
554 2557158 : if (I != MergedBBFreq.end())
555 11725 : return I->second;
556 :
557 2545433 : return MBFI.getBlockFreq(MBB);
558 : }
559 :
560 0 : void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB,
561 : BlockFrequency F) {
562 7802 : MergedBBFreq[MBB] = F;
563 0 : }
564 :
565 : raw_ostream &
566 0 : BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
567 : const MachineBasicBlock *MBB) const {
568 0 : return MBFI.printBlockFreq(OS, getBlockFreq(MBB));
569 : }
570 :
571 : raw_ostream &
572 0 : BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS,
573 : const BlockFrequency Freq) const {
574 0 : return MBFI.printBlockFreq(OS, Freq);
575 : }
576 :
577 0 : void BranchFolder::MBFIWrapper::view(const Twine &Name, bool isSimple) {
578 0 : MBFI.view(Name, isSimple);
579 0 : }
580 :
581 : uint64_t
582 3119 : BranchFolder::MBFIWrapper::getEntryFreq() const {
583 3119 : return MBFI.getEntryFreq();
584 : }
585 :
586 : /// CountTerminators - Count the number of terminators in the given
587 : /// block and set I to the position of the first non-terminator, if there
588 : /// is one, or MBB->end() otherwise.
589 8074 : static unsigned CountTerminators(MachineBasicBlock *MBB,
590 : MachineBasicBlock::iterator &I) {
591 8074 : I = MBB->end();
592 : unsigned NumTerms = 0;
593 : while (true) {
594 9045 : if (I == MBB->begin()) {
595 1 : I = MBB->end();
596 1 : break;
597 : }
598 : --I;
599 9044 : if (!I->isTerminator()) break;
600 971 : ++NumTerms;
601 : }
602 8074 : return NumTerms;
603 : }
604 :
605 : /// A no successor, non-return block probably ends in unreachable and is cold.
606 : /// Also consider a block that ends in an indirect branch to be a return block,
607 : /// since many targets use plain indirect branches to return.
608 147000 : static bool blockEndsInUnreachable(const MachineBasicBlock *MBB) {
609 147000 : if (!MBB->succ_empty())
610 : return false;
611 119263 : if (MBB->empty())
612 : return true;
613 132165 : return !(MBB->back().isReturn() || MBB->back().isIndirectBranch());
614 : }
615 :
616 : /// ProfitableToMerge - Check if two machine basic blocks have a common tail
617 : /// and decide if it would be profitable to merge those tails. Return the
618 : /// length of the common tail and iterators to the first common instruction
619 : /// in each block.
620 : /// MBB1, MBB2 The blocks to check
621 : /// MinCommonTailLength Minimum size of tail block to be merged.
622 : /// CommonTailLen Out parameter to record the size of the shared tail between
623 : /// MBB1 and MBB2
624 : /// I1, I2 Iterator references that will be changed to point to the first
625 : /// instruction in the common tail shared by MBB1,MBB2
626 : /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form
627 : /// relative to SuccBB
628 : /// PredBB The layout predecessor of SuccBB, if any.
629 : /// EHScopeMembership map from block to EH scope #.
630 : /// AfterPlacement True if we are merging blocks after layout. Stricter
631 : /// thresholds apply to prevent undoing tail-duplication.
632 : static bool
633 2132016 : ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2,
634 : unsigned MinCommonTailLength, unsigned &CommonTailLen,
635 : MachineBasicBlock::iterator &I1,
636 : MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB,
637 : MachineBasicBlock *PredBB,
638 : DenseMap<const MachineBasicBlock *, int> &EHScopeMembership,
639 : bool AfterPlacement) {
640 : // It is never profitable to tail-merge blocks from two different EH scopes.
641 2132016 : if (!EHScopeMembership.empty()) {
642 91 : auto EHScope1 = EHScopeMembership.find(MBB1);
643 : assert(EHScope1 != EHScopeMembership.end());
644 91 : auto EHScope2 = EHScopeMembership.find(MBB2);
645 : assert(EHScope2 != EHScopeMembership.end());
646 91 : if (EHScope1->second != EHScope2->second)
647 68 : return false;
648 : }
649 :
650 2131948 : CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
651 2131948 : if (CommonTailLen == 0)
652 : return false;
653 : LLVM_DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1)
654 : << " and " << printMBBReference(*MBB2) << " is "
655 : << CommonTailLen << '\n');
656 :
657 : // It's almost always profitable to merge any number of non-terminator
658 : // instructions with the block that falls through into the common successor.
659 : // This is true only for a single successor. For multiple successors, we are
660 : // trading a conditional branch for an unconditional one.
661 : // TODO: Re-visit successor size for non-layout tail merging.
662 1340404 : if ((MBB1 == PredBB || MBB2 == PredBB) &&
663 5063 : (!AfterPlacement || MBB1->succ_size() == 1)) {
664 : MachineBasicBlock::iterator I;
665 11494 : unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
666 8074 : if (CommonTailLen > NumTerms)
667 7472 : return true;
668 : }
669 :
670 : // If these are identical non-return blocks with no successors, merge them.
671 : // Such blocks are typically cold calls to noreturn functions like abort, and
672 : // are unlikely to become a fallthrough target after machine block placement.
673 : // Tail merging these blocks is unlikely to create additional unconditional
674 : // branches, and will reduce the size of this cold code.
675 346787 : if (I1 == MBB1->begin() && I2 == MBB2->begin() &&
676 1479932 : blockEndsInUnreachable(MBB1) && blockEndsInUnreachable(MBB2))
677 : return true;
678 :
679 : // If one of the blocks can be completely merged and happens to be in
680 : // a position where the other could fall through into it, merge any number
681 : // of instructions, because it can be done without a branch.
682 : // TODO: If the blocks are not adjacent, move one of them so that they are?
683 1326481 : if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
684 : return true;
685 1326365 : if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
686 : return true;
687 :
688 : // If both blocks are identical and end in a branch, merge them unless they
689 : // both have a fallthrough predecessor and successor.
690 : // We can only do this after block placement because it depends on whether
691 : // there are fallthroughs, and we don't know until after layout.
692 1323279 : if (AfterPlacement && I1 == MBB1->begin() && I2 == MBB2->begin()) {
693 : auto BothFallThrough = [](MachineBasicBlock *MBB) {
694 : if (MBB->succ_size() != 0 && !MBB->canFallThrough())
695 : return false;
696 : MachineFunction::iterator I(MBB);
697 : MachineFunction *MF = MBB->getParent();
698 : return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough();
699 : };
700 5619 : if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2))
701 : return true;
702 : }
703 :
704 : // If both blocks have an unconditional branch temporarily stripped out,
705 : // count that as an additional common instruction for the following
706 : // heuristics. This heuristic is only accurate for single-succ blocks, so to
707 : // make sure that during layout merging and duplicating don't crash, we check
708 : // for that when merging during layout.
709 1317692 : unsigned EffectiveTailLen = CommonTailLen;
710 52124 : if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
711 50904 : (MBB1->succ_size() == 1 || !AfterPlacement) &&
712 1366001 : !MBB1->back().isBarrier() &&
713 : !MBB2->back().isBarrier())
714 48309 : ++EffectiveTailLen;
715 :
716 : // Check if the common tail is long enough to be worthwhile.
717 1317692 : if (EffectiveTailLen >= MinCommonTailLength)
718 : return true;
719 :
720 : // If we are optimizing for code size, 2 instructions in common is enough if
721 : // we don't have to split a block. At worst we will be introducing 1 new
722 : // branch instruction, which is likely to be smaller than the 2
723 : // instructions that would be deleted in the merge.
724 568749 : MachineFunction *MF = MBB1->getParent();
725 568749 : return EffectiveTailLen >= 2 && MF->getFunction().optForSize() &&
726 50 : (I1 == MBB1->begin() || I2 == MBB2->begin());
727 : }
728 :
729 313096 : unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
730 : unsigned MinCommonTailLength,
731 : MachineBasicBlock *SuccBB,
732 : MachineBasicBlock *PredBB) {
733 : unsigned maxCommonTailLength = 0U;
734 313096 : SameTails.clear();
735 : MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
736 : MPIterator HighestMPIter = std::prev(MergePotentials.end());
737 : for (MPIterator CurMPIter = std::prev(MergePotentials.end()),
738 : B = MergePotentials.begin();
739 753081 : CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) {
740 2477753 : for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) {
741 : unsigned CommonTailLen;
742 2132016 : if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
743 : MinCommonTailLength,
744 : CommonTailLen, TrialBBI1, TrialBBI2,
745 : SuccBB, PredBB,
746 2132016 : EHScopeMembership,
747 2132016 : AfterBlockPlacement)) {
748 771681 : if (CommonTailLen > maxCommonTailLength) {
749 : SameTails.clear();
750 : maxCommonTailLength = CommonTailLen;
751 : HighestMPIter = CurMPIter;
752 11933 : SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
753 : }
754 771681 : if (HighestMPIter == CurMPIter &&
755 49008 : CommonTailLen == maxCommonTailLength)
756 35678 : SameTails.push_back(SameTailElt(I, TrialBBI2));
757 : }
758 2132016 : if (I == B)
759 : break;
760 : }
761 : }
762 313096 : return maxCommonTailLength;
763 : }
764 :
765 302751 : void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
766 : MachineBasicBlock *SuccBB,
767 : MachineBasicBlock *PredBB) {
768 : MPIterator CurMPIter, B;
769 : for (CurMPIter = std::prev(MergePotentials.end()),
770 : B = MergePotentials.begin();
771 672342 : CurMPIter->getHash() == CurHash; --CurMPIter) {
772 : // Put the unconditional branch back, if we need one.
773 392095 : MachineBasicBlock *CurMBB = CurMPIter->getBlock();
774 392095 : if (SuccBB && CurMBB != PredBB)
775 196888 : FixTail(CurMBB, SuccBB, TII);
776 392095 : if (CurMPIter == B)
777 : break;
778 : }
779 302751 : if (CurMPIter->getHash() != CurHash)
780 : CurMPIter++;
781 : MergePotentials.erase(CurMPIter, MergePotentials.end());
782 302751 : }
783 :
784 7802 : bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
785 : MachineBasicBlock *SuccBB,
786 : unsigned maxCommonTailLength,
787 : unsigned &commonTailIndex) {
788 7802 : commonTailIndex = 0;
789 : unsigned TimeEstimate = ~0U;
790 38250 : for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
791 : // Use PredBB if possible; that doesn't require a new branch.
792 52904 : if (SameTails[i].getBlock() == PredBB) {
793 3806 : commonTailIndex = i;
794 3806 : break;
795 : }
796 : // Otherwise, make a (fairly bogus) choice based on estimate of
797 : // how long it will take the various blocks to execute.
798 22646 : unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
799 : SameTails[i].getTailStartPos());
800 22646 : if (t <= TimeEstimate) {
801 : TimeEstimate = t;
802 22388 : commonTailIndex = i;
803 : }
804 : }
805 :
806 : MachineBasicBlock::iterator BBI =
807 15604 : SameTails[commonTailIndex].getTailStartPos();
808 : MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
809 :
810 : LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "
811 : << maxCommonTailLength);
812 :
813 : // If the split block unconditionally falls-thru to SuccBB, it will be
814 : // merged. In control flow terms it should then take SuccBB's name. e.g. If
815 : // SuccBB is an inner loop, the common tail is still part of the inner loop.
816 7802 : const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ?
817 7802 : SuccBB->getBasicBlock() : MBB->getBasicBlock();
818 7802 : MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB);
819 7802 : if (!newMBB) {
820 : LLVM_DEBUG(dbgs() << "... failed!");
821 : return false;
822 : }
823 :
824 7802 : SameTails[commonTailIndex].setBlock(newMBB);
825 : SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
826 :
827 : // If we split PredBB, newMBB is the new predecessor.
828 7802 : if (PredBB == MBB)
829 3806 : PredBB = newMBB;
830 :
831 : return true;
832 : }
833 :
834 : static void
835 26347 : mergeOperations(MachineBasicBlock::iterator MBBIStartPos,
836 : MachineBasicBlock &MBBCommon) {
837 26347 : MachineBasicBlock *MBB = MBBIStartPos->getParent();
838 : // Note CommonTailLen does not necessarily matches the size of
839 : // the common BB nor all its instructions because of debug
840 : // instructions differences.
841 : unsigned CommonTailLen = 0;
842 131561 : for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos)
843 105214 : ++CommonTailLen;
844 :
845 : MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin();
846 : MachineBasicBlock::reverse_iterator MBBIE = MBB->rend();
847 : MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin();
848 : MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend();
849 :
850 131561 : while (CommonTailLen--) {
851 : assert(MBBI != MBBIE && "Reached BB end within common tail length!");
852 : (void)MBBIE;
853 :
854 : if (!countsAsInstruction(*MBBI)) {
855 : ++MBBI;
856 6509 : continue;
857 : }
858 :
859 104552 : while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon))
860 : ++MBBICommon;
861 :
862 : assert(MBBICommon != MBBIECommon &&
863 : "Reached BB end within common tail length!");
864 : assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!");
865 :
866 : // Merge MMOs from memory operations in the common block.
867 98705 : if (MBBICommon->mayLoad() || MBBICommon->mayStore())
868 79626 : MBBICommon->cloneMergedMemRefs(*MBB->getParent(), {&*MBBICommon, &*MBBI});
869 : // Drop undef flags if they aren't present in all merged instructions.
870 638285 : for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) {
871 539580 : MachineOperand &MO = MBBICommon->getOperand(I);
872 539580 : if (MO.isReg() && MO.isUndef()) {
873 987 : const MachineOperand &OtherMO = MBBI->getOperand(I);
874 987 : if (!OtherMO.isUndef())
875 : MO.setIsUndef(false);
876 : }
877 : }
878 :
879 : ++MBBI;
880 : ++MBBICommon;
881 : }
882 26347 : }
883 :
884 10345 : void BranchFolder::mergeCommonTails(unsigned commonTailIndex) {
885 10345 : MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
886 :
887 20690 : std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size());
888 57382 : for (unsigned int i = 0 ; i != SameTails.size() ; ++i) {
889 36692 : if (i != commonTailIndex) {
890 26347 : NextCommonInsts[i] = SameTails[i].getTailStartPos();
891 26347 : mergeOperations(SameTails[i].getTailStartPos(), *MBB);
892 : } else {
893 : assert(SameTails[i].getTailStartPos() == MBB->begin() &&
894 : "MBB is not a common tail only block");
895 : }
896 : }
897 :
898 42817 : for (auto &MI : *MBB) {
899 : if (!countsAsInstruction(MI))
900 2466 : continue;
901 : DebugLoc DL = MI.getDebugLoc();
902 188723 : for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) {
903 128711 : if (i == commonTailIndex)
904 : continue;
905 :
906 : auto &Pos = NextCommonInsts[i];
907 : assert(Pos != SameTails[i].getBlock()->end() &&
908 : "Reached BB end within common tail");
909 : while (!countsAsInstruction(*Pos)) {
910 : ++Pos;
911 : assert(Pos != SameTails[i].getBlock()->end() &&
912 : "Reached BB end within common tail");
913 : }
914 : assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!");
915 197410 : DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc());
916 197410 : NextCommonInsts[i] = ++Pos;
917 : }
918 30006 : MI.setDebugLoc(DL);
919 : }
920 :
921 10345 : if (UpdateLiveIns) {
922 10322 : LivePhysRegs NewLiveIns(*TRI);
923 10322 : computeLiveIns(NewLiveIns, *MBB);
924 10322 : LiveRegs.init(*TRI);
925 :
926 : // The flag merging may lead to some register uses no longer using the
927 : // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary.
928 23432 : for (MachineBasicBlock *Pred : MBB->predecessors()) {
929 : LiveRegs.clear();
930 13110 : LiveRegs.addLiveOuts(*Pred);
931 13110 : MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator();
932 215117 : for (unsigned Reg : NewLiveIns) {
933 202007 : if (!LiveRegs.available(*MRI, Reg))
934 201971 : continue;
935 36 : DebugLoc DL;
936 36 : BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF),
937 72 : Reg);
938 : }
939 : }
940 :
941 10322 : MBB->clearLiveIns();
942 10322 : addLiveIns(*MBB, NewLiveIns);
943 : }
944 10345 : }
945 :
946 : // See if any of the blocks in MergePotentials (which all have SuccBB as a
947 : // successor, or all have no successor if it is null) can be tail-merged.
948 : // If there is a successor, any blocks in MergePotentials that are not
949 : // tail-merged and are not immediately before Succ must have an unconditional
950 : // branch to Succ added (but the predecessor/successor lists need no
951 : // adjustment). The lone predecessor of Succ that falls through into Succ,
952 : // if any, is given in PredBB.
953 : // MinCommonTailLength - Except for the special cases below, tail-merge if
954 : // there are at least this many instructions in common.
955 239633 : bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
956 : MachineBasicBlock *PredBB,
957 : unsigned MinCommonTailLength) {
958 : bool MadeChange = false;
959 :
960 : LLVM_DEBUG(
961 : dbgs() << "\nTryTailMergeBlocks: ";
962 : for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs()
963 : << printMBBReference(*MergePotentials[i].getBlock())
964 : << (i == e - 1 ? "" : ", ");
965 : dbgs() << "\n"; if (SuccBB) {
966 : dbgs() << " with successor " << printMBBReference(*SuccBB) << '\n';
967 : if (PredBB)
968 : dbgs() << " which has fall-through from "
969 : << printMBBReference(*PredBB) << "\n";
970 : } dbgs() << "Looking for common tails of at least "
971 : << MinCommonTailLength << " instruction"
972 : << (MinCommonTailLength == 1 ? "" : "s") << '\n';);
973 :
974 : // Sort by hash value so that blocks with identical end sequences sort
975 : // together.
976 : array_pod_sort(MergePotentials.begin(), MergePotentials.end());
977 :
978 : // Walk through equivalence sets looking for actual exact matches.
979 1105458 : while (MergePotentials.size() > 1) {
980 313096 : unsigned CurHash = MergePotentials.back().getHash();
981 :
982 : // Build SameTails, identifying the set of blocks with this hash code
983 : // and with the maximum number of instructions in common.
984 313096 : unsigned maxCommonTailLength = ComputeSameTails(CurHash,
985 : MinCommonTailLength,
986 : SuccBB, PredBB);
987 :
988 : // If we didn't find any pair that has at least MinCommonTailLength
989 : // instructions in common, remove all blocks with this hash code and retry.
990 313096 : if (SameTails.empty()) {
991 302751 : RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
992 302751 : continue;
993 : }
994 :
995 : // If one of the blocks is the entire common tail (and not the entry
996 : // block, which we can't jump to), we can treat all blocks with this same
997 : // tail at once. Use PredBB if that is one of the possibilities, as that
998 : // will not introduce any extra branches.
999 : MachineBasicBlock *EntryBB =
1000 10345 : &MergePotentials.front().getBlock()->getParent()->front();
1001 10345 : unsigned commonTailIndex = SameTails.size();
1002 : // If there are two blocks, check to see if one can be made to fall through
1003 : // into the other.
1004 8216 : if (SameTails.size() == 2 &&
1005 10463 : SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
1006 118 : SameTails[1].tailIsWholeBlock())
1007 35 : commonTailIndex = 1;
1008 18491 : else if (SameTails.size() == 2 &&
1009 8181 : SameTails[1].getBlock()->isLayoutSuccessor(
1010 11732 : SameTails[0].getBlock()) &&
1011 1422 : SameTails[0].tailIsWholeBlock())
1012 759 : commonTailIndex = 0;
1013 : else {
1014 : // Otherwise just pick one, favoring the fall-through predecessor if
1015 : // there is one.
1016 45431 : for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
1017 31195 : MachineBasicBlock *MBB = SameTails[i].getBlock();
1018 31195 : if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
1019 : continue;
1020 31193 : if (MBB == PredBB) {
1021 4866 : commonTailIndex = i;
1022 4866 : break;
1023 : }
1024 26327 : if (SameTails[i].tailIsWholeBlock())
1025 2850 : commonTailIndex = i;
1026 : }
1027 : }
1028 :
1029 20690 : if (commonTailIndex == SameTails.size() ||
1030 6348 : (SameTails[commonTailIndex].getBlock() == PredBB &&
1031 : !SameTails[commonTailIndex].tailIsWholeBlock())) {
1032 : // None of the blocks consist entirely of the common tail.
1033 : // Split a block so that one does.
1034 7802 : if (!CreateCommonTailOnlyBlock(PredBB, SuccBB,
1035 : maxCommonTailLength, commonTailIndex)) {
1036 0 : RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
1037 0 : continue;
1038 : }
1039 : }
1040 :
1041 10345 : MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
1042 :
1043 : // Recompute common tail MBB's edge weights and block frequency.
1044 10345 : setCommonTailEdgeWeights(*MBB);
1045 :
1046 : // Merge debug locations, MMOs and undef flags across identical instructions
1047 : // for common tail.
1048 10345 : mergeCommonTails(commonTailIndex);
1049 :
1050 : // MBB is common tail. Adjust all other BB's to jump to this one.
1051 : // Traversal must be forwards so erases work.
1052 : LLVM_DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB)
1053 : << " for ");
1054 57382 : for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
1055 36692 : if (commonTailIndex == i)
1056 : continue;
1057 : LLVM_DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock())
1058 : << (i == e - 1 ? "" : ", "));
1059 : // Hack the end off BB i, making it jump to BB commonTailIndex instead.
1060 52694 : replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB);
1061 : // BB i is no longer a predecessor of SuccBB; remove it from the worklist.
1062 26347 : MergePotentials.erase(SameTails[i].getMPIter());
1063 : }
1064 : LLVM_DEBUG(dbgs() << "\n");
1065 : // We leave commonTailIndex in the worklist in case there are other blocks
1066 : // that match it with a smaller number of instructions.
1067 : MadeChange = true;
1068 : }
1069 239633 : return MadeChange;
1070 : }
1071 :
1072 254247 : bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
1073 : bool MadeChange = false;
1074 254247 : if (!EnableTailMerge) return MadeChange;
1075 :
1076 : // First find blocks with no successors.
1077 : // Block placement does not create new tail merging opportunities for these
1078 : // blocks.
1079 234926 : if (!AfterBlockPlacement) {
1080 225696 : MergePotentials.clear();
1081 1113315 : for (MachineBasicBlock &MBB : MF) {
1082 1775258 : if (MergePotentials.size() == TailMergeThreshold)
1083 : break;
1084 887619 : if (!TriedMerging.count(&MBB) && MBB.succ_empty())
1085 538307 : MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB));
1086 : }
1087 :
1088 : // If this is a large problem, avoid visiting the same basic blocks
1089 : // multiple times.
1090 451392 : if (MergePotentials.size() == TailMergeThreshold)
1091 1510 : for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1092 3000 : TriedMerging.insert(MergePotentials[i].getBlock());
1093 :
1094 : // See if we can do any tail merging on those.
1095 451392 : if (MergePotentials.size() >= 2)
1096 13503 : MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength);
1097 : }
1098 :
1099 : // Look at blocks (IBB) with multiple predecessors (PBB).
1100 : // We change each predecessor to a canonical form, by
1101 : // (1) temporarily removing any unconditional branch from the predecessor
1102 : // to IBB, and
1103 : // (2) alter conditional branches so they branch to the other block
1104 : // not IBB; this may require adding back an unconditional branch to IBB
1105 : // later, where there wasn't one coming in. E.g.
1106 : // Bcc IBB
1107 : // fallthrough to QBB
1108 : // here becomes
1109 : // Bncc QBB
1110 : // with a conceptual B to IBB after that, which never actually exists.
1111 : // With those changes, we see whether the predecessors' tails match,
1112 : // and merge them if so. We change things out of canonical form and
1113 : // back to the way they were later in the process. (OptimizeBranches
1114 : // would undo some of this, but we can't use it, because we'd get into
1115 : // a compile-time infinite loop repeatedly doing and undoing the same
1116 : // transformations.)
1117 :
1118 : for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1119 1225586 : I != E; ++I) {
1120 997769 : if (I->pred_size() < 2) continue;
1121 : SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
1122 : MachineBasicBlock *IBB = &*I;
1123 : MachineBasicBlock *PredBB = &*std::prev(I);
1124 260524 : MergePotentials.clear();
1125 : MachineLoop *ML;
1126 :
1127 : // Bail if merging after placement and IBB is the loop header because
1128 : // -- If merging predecessors that belong to the same loop as IBB, the
1129 : // common tail of merged predecessors may become the loop top if block
1130 : // placement is called again and the predecessors may branch to this common
1131 : // tail and require more branches. This can be relaxed if
1132 : // MachineBlockPlacement::findBestLoopTop is more flexible.
1133 : // --If merging predecessors that do not belong to the same loop as IBB, the
1134 : // loop info of IBB's loop and the other loops may be affected. Calling the
1135 : // block placement again may make big change to the layout and eliminate the
1136 : // reason to do tail merging here.
1137 260524 : if (AfterBlockPlacement && MLI) {
1138 : ML = MLI->getLoopFor(IBB);
1139 17170 : if (ML && IBB == ML->getHeader())
1140 : continue;
1141 : }
1142 :
1143 908335 : for (MachineBasicBlock *PBB : I->predecessors()) {
1144 1309848 : if (MergePotentials.size() == TailMergeThreshold)
1145 : break;
1146 :
1147 654920 : if (TriedMerging.count(PBB))
1148 58217 : continue;
1149 :
1150 : // Skip blocks that loop to themselves, can't tail merge these.
1151 650449 : if (PBB == IBB)
1152 : continue;
1153 :
1154 : // Visit each predecessor only once.
1155 640807 : if (!UniquePreds.insert(PBB).second)
1156 : continue;
1157 :
1158 : // Skip blocks which may jump to a landing pad. Can't tail merge these.
1159 640807 : if (PBB->hasEHPadSuccessor())
1160 : continue;
1161 :
1162 : // After block placement, only consider predecessors that belong to the
1163 : // same loop as IBB. The reason is the same as above when skipping loop
1164 : // header.
1165 601975 : if (AfterBlockPlacement && MLI)
1166 192773 : if (ML != MLI->getLoopFor(PBB))
1167 : continue;
1168 :
1169 596771 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1170 : SmallVector<MachineOperand, 4> Cond;
1171 596771 : if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) {
1172 : // Failing case: IBB is the target of a cbr, and we cannot reverse the
1173 : // branch.
1174 : SmallVector<MachineOperand, 4> NewCond(Cond);
1175 592919 : if (!Cond.empty() && TBB == IBB) {
1176 78505 : if (TII->reverseBranchCondition(NewCond))
1177 : continue;
1178 : // This is the QBB case described above
1179 78437 : if (!FBB) {
1180 71701 : auto Next = ++PBB->getIterator();
1181 71701 : if (Next != MF.end())
1182 71701 : FBB = &*Next;
1183 : }
1184 : }
1185 :
1186 : // Failing case: the only way IBB can be reached from PBB is via
1187 : // exception handling. Happens for landing pads. Would be nice to have
1188 : // a bit in the edge so we didn't have to do all this.
1189 592851 : if (IBB->isEHPad()) {
1190 0 : MachineFunction::iterator IP = ++PBB->getIterator();
1191 : MachineBasicBlock *PredNextBB = nullptr;
1192 0 : if (IP != MF.end())
1193 : PredNextBB = &*IP;
1194 0 : if (!TBB) {
1195 0 : if (IBB != PredNextBB) // fallthrough
1196 : continue;
1197 0 : } else if (FBB) {
1198 0 : if (TBB != IBB && FBB != IBB) // cbr then ubr
1199 : continue;
1200 0 : } else if (Cond.empty()) {
1201 0 : if (TBB != IBB) // ubr
1202 : continue;
1203 : } else {
1204 0 : if (TBB != IBB && IBB != PredNextBB) // cbr
1205 : continue;
1206 : }
1207 : }
1208 :
1209 : // Remove the unconditional branch at the end, if any.
1210 592851 : if (TBB && (Cond.empty() || FBB)) {
1211 374335 : DebugLoc dl = PBB->findBranchDebugLoc();
1212 374335 : TII->removeBranch(*PBB);
1213 374335 : if (!Cond.empty())
1214 : // reinsert conditional branch only, for now
1215 81482 : TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr,
1216 81482 : NewCond, dl);
1217 : }
1218 :
1219 1185702 : MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB));
1220 : }
1221 : }
1222 :
1223 : // If this is a large problem, avoid visiting the same basic blocks multiple
1224 : // times.
1225 506830 : if (MergePotentials.size() == TailMergeThreshold)
1226 308 : for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
1227 608 : TriedMerging.insert(MergePotentials[i].getBlock());
1228 :
1229 506830 : if (MergePotentials.size() >= 2)
1230 226130 : MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength);
1231 :
1232 : // Reinsert an unconditional branch if needed. The 1 below can occur as a
1233 : // result of removing blocks in TryTailMergeBlocks.
1234 : PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks
1235 506830 : if (MergePotentials.size() == 1 &&
1236 220596 : MergePotentials.begin()->getBlock() != PredBB)
1237 169777 : FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
1238 : }
1239 :
1240 : return MadeChange;
1241 : }
1242 :
1243 10345 : void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) {
1244 10345 : SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size());
1245 : BlockFrequency AccumulatedMBBFreq;
1246 :
1247 : // Aggregate edge frequency of successor edge j:
1248 : // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)),
1249 : // where bb is a basic block that is in SameTails.
1250 47037 : for (const auto &Src : SameTails) {
1251 : const MachineBasicBlock *SrcMBB = Src.getBlock();
1252 36692 : BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB);
1253 36692 : AccumulatedMBBFreq += BlockFreq;
1254 :
1255 : // It is not necessary to recompute edge weights if TailBB has less than two
1256 : // successors.
1257 36692 : if (TailMBB.succ_size() <= 1)
1258 35770 : continue;
1259 :
1260 : auto EdgeFreq = EdgeFreqLs.begin();
1261 :
1262 1844 : for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1263 2766 : SuccI != SuccE; ++SuccI, ++EdgeFreq)
1264 1844 : *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI);
1265 : }
1266 :
1267 10345 : MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq);
1268 :
1269 10345 : if (TailMBB.succ_size() <= 1)
1270 : return;
1271 :
1272 : auto SumEdgeFreq =
1273 429 : std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0))
1274 : .getFrequency();
1275 : auto EdgeFreq = EdgeFreqLs.begin();
1276 :
1277 429 : if (SumEdgeFreq > 0) {
1278 854 : for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end();
1279 1281 : SuccI != SuccE; ++SuccI, ++EdgeFreq) {
1280 : auto Prob = BranchProbability::getBranchProbability(
1281 854 : EdgeFreq->getFrequency(), SumEdgeFreq);
1282 854 : TailMBB.setSuccProbability(SuccI, Prob);
1283 : }
1284 : }
1285 : }
1286 :
1287 : //===----------------------------------------------------------------------===//
1288 : // Branch Optimization
1289 : //===----------------------------------------------------------------------===//
1290 :
1291 245786 : bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
1292 : bool MadeChange = false;
1293 :
1294 : // Make sure blocks are numbered in order
1295 245786 : MF.RenumberBlocks();
1296 : // Renumbering blocks alters EH scope membership, recalculate it.
1297 491572 : EHScopeMembership = getEHScopeMembership(MF);
1298 :
1299 : for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end();
1300 1031589 : I != E; ) {
1301 : MachineBasicBlock *MBB = &*I++;
1302 785803 : MadeChange |= OptimizeBlock(MBB);
1303 :
1304 : // If it is dead, remove it.
1305 785803 : if (MBB->pred_empty()) {
1306 15269 : RemoveDeadBlock(MBB);
1307 : MadeChange = true;
1308 : ++NumDeadBlocks;
1309 : }
1310 : }
1311 :
1312 245786 : return MadeChange;
1313 : }
1314 :
1315 : // Blocks should be considered empty if they contain only debug info;
1316 : // else the debug info would affect codegen.
1317 : static bool IsEmptyBlock(MachineBasicBlock *MBB) {
1318 1625345 : return MBB->getFirstNonDebugInstr() == MBB->end();
1319 : }
1320 :
1321 : // Blocks with only debug info and branches should be considered the same
1322 : // as blocks with only branches.
1323 276121 : static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
1324 276121 : MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr();
1325 : assert(I != MBB->end() && "empty block!");
1326 276121 : return I->isBranch();
1327 : }
1328 :
1329 : /// IsBetterFallthrough - Return true if it would be clearly better to
1330 : /// fall-through to MBB1 than to fall through into MBB2. This has to return
1331 : /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
1332 : /// result in infinite loops.
1333 4868 : static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
1334 : MachineBasicBlock *MBB2) {
1335 : // Right now, we use a simple heuristic. If MBB2 ends with a call, and
1336 : // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
1337 : // optimize branches that branch to either a return block or an assert block
1338 : // into a fallthrough to the return.
1339 4868 : MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr();
1340 4868 : MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr();
1341 4868 : if (MBB1I == MBB1->end() || MBB2I == MBB2->end())
1342 : return false;
1343 :
1344 : // If there is a clear successor ordering we make sure that one block
1345 : // will fall through to the next
1346 4861 : if (MBB1->isSuccessor(MBB2)) return true;
1347 4830 : if (MBB2->isSuccessor(MBB1)) return false;
1348 :
1349 5690 : return MBB2I->isCall() && !MBB1I->isCall();
1350 : }
1351 :
1352 : /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
1353 : /// instructions on the block.
1354 17723 : static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
1355 17723 : MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
1356 35446 : if (I != MBB.end() && I->isBranch())
1357 : return I->getDebugLoc();
1358 911 : return DebugLoc();
1359 : }
1360 :
1361 543 : static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII,
1362 : MachineBasicBlock &MBB,
1363 : MachineBasicBlock &PredMBB) {
1364 543 : auto InsertBefore = PredMBB.getFirstTerminator();
1365 583 : for (MachineInstr &MI : MBB.instrs())
1366 40 : if (MI.isDebugValue()) {
1367 40 : TII->duplicate(PredMBB, InsertBefore, MI);
1368 : LLVM_DEBUG(dbgs() << "Copied debug value from empty block to pred: "
1369 : << MI);
1370 : }
1371 543 : }
1372 :
1373 105 : static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII,
1374 : MachineBasicBlock &MBB,
1375 : MachineBasicBlock &SuccMBB) {
1376 105 : auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin());
1377 106 : for (MachineInstr &MI : MBB.instrs())
1378 1 : if (MI.isDebugValue()) {
1379 1 : TII->duplicate(SuccMBB, InsertBefore, MI);
1380 : LLVM_DEBUG(dbgs() << "Copied debug value from empty block to succ: "
1381 : << MI);
1382 : }
1383 105 : }
1384 :
1385 : // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such
1386 : // a basic block is removed we would lose the debug information unless we have
1387 : // copied the information to a predecessor/successor.
1388 : //
1389 : // TODO: This function only handles some simple cases. An alternative would be
1390 : // to run a heavier analysis, such as the LiveDebugValues pass, before we do
1391 : // branch folding.
1392 6617 : static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII,
1393 : MachineBasicBlock &MBB) {
1394 : assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info).");
1395 : // If this MBB is the only predecessor of a successor it is legal to copy
1396 : // DBG_VALUE instructions to the beginning of the successor.
1397 9185 : for (MachineBasicBlock *SuccBB : MBB.successors())
1398 2568 : if (SuccBB->pred_size() == 1)
1399 105 : copyDebugInfoToSuccessor(TII, MBB, *SuccBB);
1400 : // If this MBB is the only successor of a predecessor it is legal to copy the
1401 : // DBG_VALUE instructions to the end of the predecessor (just before the
1402 : // terminators, assuming that the terminator isn't affecting the DBG_VALUE).
1403 13207 : for (MachineBasicBlock *PredBB : MBB.predecessors())
1404 6590 : if (PredBB->succ_size() == 1)
1405 543 : copyDebugInfoToPredecessor(TII, MBB, *PredBB);
1406 6617 : }
1407 :
1408 785803 : bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
1409 : bool MadeChange = false;
1410 785803 : MachineFunction &MF = *MBB->getParent();
1411 818883 : ReoptimizeBlock:
1412 :
1413 818883 : MachineFunction::iterator FallThrough = MBB->getIterator();
1414 : ++FallThrough;
1415 :
1416 : // Make sure MBB and FallThrough belong to the same EH scope.
1417 : bool SameEHScope = true;
1418 818883 : if (!EHScopeMembership.empty() && FallThrough != MF.end()) {
1419 628 : auto MBBEHScope = EHScopeMembership.find(MBB);
1420 : assert(MBBEHScope != EHScopeMembership.end());
1421 628 : auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough);
1422 : assert(FallThroughEHScope != EHScopeMembership.end());
1423 628 : SameEHScope = MBBEHScope->second == FallThroughEHScope->second;
1424 : }
1425 :
1426 : // If this block is empty, make everyone use its fall-through, not the block
1427 : // explicitly. Landing pads should not do this since the landing-pad table
1428 : // points to this block. Blocks with their addresses taken shouldn't be
1429 : // optimized away.
1430 818883 : if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() &&
1431 : SameEHScope) {
1432 6617 : salvageDebugInfoFromEmptyBlock(TII, *MBB);
1433 : // Dead block? Leave for cleanup later.
1434 6617 : if (MBB->pred_empty()) return MadeChange;
1435 :
1436 5826 : if (FallThrough == MF.end()) {
1437 : // TODO: Simplify preds to not branch here if possible!
1438 5590 : } else if (FallThrough->isEHPad()) {
1439 : // Don't rewrite to a landing pad fallthough. That could lead to the case
1440 : // where a BB jumps to more than one landing pad.
1441 : // TODO: Is it ever worth rewriting predecessors which don't already
1442 : // jump to a landing pad, and so can safely jump to the fallthrough?
1443 4566 : } else if (MBB->isSuccessor(&*FallThrough)) {
1444 : // Rewrite all predecessors of the old block to go to the fallthrough
1445 : // instead.
1446 5723 : while (!MBB->pred_empty()) {
1447 3155 : MachineBasicBlock *Pred = *(MBB->pred_end()-1);
1448 3155 : Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough);
1449 : }
1450 : // If MBB was the target of a jump table, update jump tables to go to the
1451 : // fallthrough instead.
1452 2568 : if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1453 333 : MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough);
1454 : MadeChange = true;
1455 : }
1456 5826 : return MadeChange;
1457 : }
1458 :
1459 : // Check to see if we can simplify the terminator of the block before this
1460 : // one.
1461 : MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB));
1462 :
1463 812266 : MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
1464 : SmallVector<MachineOperand, 4> PriorCond;
1465 : bool PriorUnAnalyzable =
1466 812266 : TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
1467 812266 : if (!PriorUnAnalyzable) {
1468 : // If the CFG for the prior block has extra edges, remove them.
1469 1583802 : MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
1470 791901 : !PriorCond.empty());
1471 :
1472 : // If the previous branch is conditional and both conditions go to the same
1473 : // destination, remove the branch, replacing it with an unconditional one or
1474 : // a fall-through.
1475 791901 : if (PriorTBB && PriorTBB == PriorFBB) {
1476 15 : DebugLoc dl = getBranchDebugLoc(PrevBB);
1477 15 : TII->removeBranch(PrevBB);
1478 : PriorCond.clear();
1479 15 : if (PriorTBB != MBB)
1480 6 : TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1481 : MadeChange = true;
1482 : ++NumBranchOpts;
1483 : goto ReoptimizeBlock;
1484 : }
1485 :
1486 : // If the previous block unconditionally falls through to this block and
1487 : // this block has no other predecessors, move the contents of this block
1488 : // into the prior block. This doesn't usually happen when SimplifyCFG
1489 : // has been used, but it can happen if tail merging splits a fall-through
1490 : // predecessor of a block.
1491 : // This has to check PrevBB->succ_size() because EH edges are ignored by
1492 : // AnalyzeBranch.
1493 577378 : if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
1494 5598 : PrevBB.succ_size() == 1 &&
1495 797484 : !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1496 : LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBB
1497 : << "From MBB: " << *MBB);
1498 : // Remove redundant DBG_VALUEs first.
1499 5519 : if (PrevBB.begin() != PrevBB.end()) {
1500 5362 : MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
1501 : --PrevBBIter;
1502 : MachineBasicBlock::iterator MBBIter = MBB->begin();
1503 : // Check if DBG_VALUE at the end of PrevBB is identical to the
1504 : // DBG_VALUE at the beginning of MBB.
1505 1753 : while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
1506 5364 : && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) {
1507 4 : if (!MBBIter->isIdenticalTo(*PrevBBIter))
1508 : break;
1509 : MachineInstr &DuplicateDbg = *MBBIter;
1510 : ++MBBIter; -- PrevBBIter;
1511 2 : DuplicateDbg.eraseFromParent();
1512 : }
1513 : }
1514 : PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
1515 5519 : PrevBB.removeSuccessor(PrevBB.succ_begin());
1516 : assert(PrevBB.succ_empty());
1517 5519 : PrevBB.transferSuccessors(MBB);
1518 : MadeChange = true;
1519 5519 : return MadeChange;
1520 : }
1521 :
1522 : // If the previous branch *only* branches to *this* block (conditional or
1523 : // not) remove the branch.
1524 786367 : if (PriorTBB == MBB && !PriorFBB) {
1525 694 : TII->removeBranch(PrevBB);
1526 : MadeChange = true;
1527 : ++NumBranchOpts;
1528 694 : goto ReoptimizeBlock;
1529 : }
1530 :
1531 : // If the prior block branches somewhere else on the condition and here if
1532 : // the condition is false, remove the uncond second branch.
1533 785673 : if (PriorFBB == MBB) {
1534 2546 : DebugLoc dl = getBranchDebugLoc(PrevBB);
1535 2546 : TII->removeBranch(PrevBB);
1536 5092 : TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl);
1537 : MadeChange = true;
1538 : ++NumBranchOpts;
1539 : goto ReoptimizeBlock;
1540 : }
1541 :
1542 : // If the prior block branches here on true and somewhere else on false, and
1543 : // if the branch condition is reversible, reverse the branch to create a
1544 : // fall-through.
1545 783127 : if (PriorTBB == MBB) {
1546 : SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1547 704 : if (!TII->reverseBranchCondition(NewPriorCond)) {
1548 697 : DebugLoc dl = getBranchDebugLoc(PrevBB);
1549 697 : TII->removeBranch(PrevBB);
1550 1394 : TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl);
1551 : MadeChange = true;
1552 : ++NumBranchOpts;
1553 : goto ReoptimizeBlock;
1554 : }
1555 : }
1556 :
1557 : // If this block has no successors (e.g. it is a return block or ends with
1558 : // a call to a no-return function like abort or __cxa_throw) and if the pred
1559 : // falls through into this block, and if it would otherwise fall through
1560 : // into the block after this, move this block to the end of the function.
1561 : //
1562 : // We consider it more likely that execution will stay in the function (e.g.
1563 : // due to loops) than it is to exit it. This asserts in loops etc, moving
1564 : // the assert condition out of the loop body.
1565 53926 : if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB &&
1566 806386 : MachineFunction::iterator(PriorTBB) == FallThrough &&
1567 6875 : !MBB->canFallThrough()) {
1568 : bool DoTransform = true;
1569 :
1570 : // We have to be careful that the succs of PredBB aren't both no-successor
1571 : // blocks. If neither have successors and if PredBB is the second from
1572 : // last block in the function, we'd just keep swapping the two blocks for
1573 : // last. Only do the swap if one is clearly better to fall through than
1574 : // the other.
1575 11743 : if (FallThrough == --MF.end() &&
1576 4868 : !IsBetterFallthrough(PriorTBB, MBB))
1577 : DoTransform = false;
1578 :
1579 : if (DoTransform) {
1580 : // Reverse the branch so we will fall through on the previous true cond.
1581 : SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
1582 2700 : if (!TII->reverseBranchCondition(NewPriorCond)) {
1583 : LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBB
1584 : << "To make fallthrough to: " << *PriorTBB << "\n");
1585 :
1586 2700 : DebugLoc dl = getBranchDebugLoc(PrevBB);
1587 2700 : TII->removeBranch(PrevBB);
1588 5400 : TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl);
1589 :
1590 : // Move this block to the end of the function.
1591 2700 : MBB->moveAfter(&MF.back());
1592 : MadeChange = true;
1593 : ++NumBranchOpts;
1594 : return MadeChange;
1595 : }
1596 : }
1597 : }
1598 : }
1599 :
1600 1389814 : if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 &&
1601 589719 : MF.getFunction().optForSize()) {
1602 : // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
1603 : // direction, thereby defeating careful block placement and regressing
1604 : // performance. Therefore, only consider this for optsize functions.
1605 1446 : MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
1606 1446 : if (TII->isUnconditionalTailCall(TailCall)) {
1607 41 : MachineBasicBlock *Pred = *MBB->pred_begin();
1608 41 : MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1609 : SmallVector<MachineOperand, 4> PredCond;
1610 : bool PredAnalyzable =
1611 41 : !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
1612 :
1613 41 : if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB &&
1614 27 : PredTBB != PredFBB) {
1615 : // The predecessor has a conditional branch to this block which consists
1616 : // of only a tail call. Try to fold the tail call into the conditional
1617 : // branch.
1618 26 : if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
1619 : // TODO: It would be nice if analyzeBranch() could provide a pointer
1620 : // to the branch instruction so replaceBranchWithTailCall() doesn't
1621 : // have to search for it.
1622 26 : TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
1623 : ++NumTailCalls;
1624 26 : Pred->removeSuccessor(MBB);
1625 : MadeChange = true;
1626 : return MadeChange;
1627 : }
1628 : }
1629 : // If the predecessor is falling through to this block, we could reverse
1630 : // the branch condition and fold the tail call into that. However, after
1631 : // that we might have to re-arrange the CFG to fall through to the other
1632 : // block and there is a high risk of regressing code size rather than
1633 : // improving it.
1634 : }
1635 : }
1636 :
1637 : // Analyze the branch in the current block.
1638 800069 : MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
1639 : SmallVector<MachineOperand, 4> CurCond;
1640 : bool CurUnAnalyzable =
1641 800069 : TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
1642 800069 : if (!CurUnAnalyzable) {
1643 : // If the CFG for the prior block has extra edges, remove them.
1644 757086 : MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
1645 :
1646 : // If this is a two-way branch, and the FBB branches to this block, reverse
1647 : // the condition so the single-basic-block loop is faster. Instead of:
1648 : // Loop: xxx; jcc Out; jmp Loop
1649 : // we want:
1650 : // Loop: xxx; jncc Loop; jmp Out
1651 757086 : if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
1652 : SmallVector<MachineOperand, 4> NewCond(CurCond);
1653 34 : if (!TII->reverseBranchCondition(NewCond)) {
1654 34 : DebugLoc dl = getBranchDebugLoc(*MBB);
1655 34 : TII->removeBranch(*MBB);
1656 68 : TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
1657 : MadeChange = true;
1658 : ++NumBranchOpts;
1659 : goto ReoptimizeBlock;
1660 : }
1661 : }
1662 :
1663 : // If this branch is the only thing in its block, see if we can forward
1664 : // other blocks across it.
1665 739319 : if (CurTBB && CurCond.empty() && !CurFBB &&
1666 282859 : IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
1667 763432 : !MBB->hasAddressTaken() && !MBB->isEHPad()) {
1668 6367 : DebugLoc dl = getBranchDebugLoc(*MBB);
1669 : // This block may contain just an unconditional branch. Because there can
1670 : // be 'non-branch terminators' in the block, try removing the branch and
1671 : // then seeing if the block is empty.
1672 6367 : TII->removeBranch(*MBB);
1673 : // If the only things remaining in the block are debug info, remove these
1674 : // as well, so this will behave the same as an empty block in non-debug
1675 : // mode.
1676 6367 : if (IsEmptyBlock(MBB)) {
1677 : // Make the block empty, losing the debug info (we could probably
1678 : // improve this in some cases.)
1679 : MBB->erase(MBB->begin(), MBB->end());
1680 : }
1681 : // If this block is just an unconditional branch to CurTBB, we can
1682 : // usually completely eliminate the block. The only case we cannot
1683 : // completely eliminate the block is when the block before this one
1684 : // falls through into MBB and we can't understand the prior block's branch
1685 : // condition.
1686 6367 : if (MBB->empty()) {
1687 6367 : bool PredHasNoFallThrough = !PrevBB.canFallThrough();
1688 6369 : if (PredHasNoFallThrough || !PriorUnAnalyzable ||
1689 2 : !PrevBB.isSuccessor(MBB)) {
1690 : // If the prior block falls through into us, turn it into an
1691 : // explicit branch to us to make updates simpler.
1692 5327 : if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
1693 11692 : PriorTBB != MBB && PriorFBB != MBB) {
1694 5327 : if (!PriorTBB) {
1695 : assert(PriorCond.empty() && !PriorFBB &&
1696 : "Bad branch analysis");
1697 911 : PriorTBB = MBB;
1698 : } else {
1699 : assert(!PriorFBB && "Machine CFG out of date!");
1700 4416 : PriorFBB = MBB;
1701 : }
1702 5327 : DebugLoc pdl = getBranchDebugLoc(PrevBB);
1703 5327 : TII->removeBranch(PrevBB);
1704 10654 : TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
1705 : }
1706 :
1707 : // Iterate through all the predecessors, revectoring each in-turn.
1708 : size_t PI = 0;
1709 : bool DidChange = false;
1710 : bool HasBranchToSelf = false;
1711 13887 : while(PI != MBB->pred_size()) {
1712 7522 : MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
1713 7522 : if (PMBB == MBB) {
1714 : // If this block has an uncond branch to itself, leave it.
1715 0 : ++PI;
1716 : HasBranchToSelf = true;
1717 : } else {
1718 : DidChange = true;
1719 7522 : PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
1720 : // If this change resulted in PMBB ending in a conditional
1721 : // branch where both conditions go to the same destination,
1722 : // change this to an unconditional branch (and fix the CFG).
1723 7522 : MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr;
1724 : SmallVector<MachineOperand, 4> NewCurCond;
1725 15044 : bool NewCurUnAnalyzable = TII->analyzeBranch(
1726 7522 : *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true);
1727 7522 : if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
1728 37 : DebugLoc pdl = getBranchDebugLoc(*PMBB);
1729 37 : TII->removeBranch(*PMBB);
1730 : NewCurCond.clear();
1731 74 : TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl);
1732 : MadeChange = true;
1733 : ++NumBranchOpts;
1734 37 : PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false);
1735 : }
1736 : }
1737 : }
1738 :
1739 : // Change any jumptables to go to the new MBB.
1740 6365 : if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
1741 1397 : MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
1742 6365 : if (DidChange) {
1743 : ++NumBranchOpts;
1744 : MadeChange = true;
1745 6365 : if (!HasBranchToSelf) return MadeChange;
1746 : }
1747 : }
1748 : }
1749 :
1750 : // Add the branch back if the block is more than just an uncond branch.
1751 4 : TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl);
1752 : }
1753 : }
1754 :
1755 : // If the prior block doesn't fall through into this block, and if this
1756 : // block doesn't fall through into some other block, see if we can find a
1757 : // place to move this block where a fall-through will happen.
1758 793670 : if (!PrevBB.canFallThrough()) {
1759 : // Now we know that there was no fall-through into this block, check to
1760 : // see if it has a fall-through into its successor.
1761 292380 : bool CurFallsThru = MBB->canFallThrough();
1762 :
1763 292380 : if (!MBB->isEHPad()) {
1764 : // Check all the predecessors of this block. If one of them has no fall
1765 : // throughs, move this block right after it.
1766 309968 : for (MachineBasicBlock *PredBB : MBB->predecessors()) {
1767 : // Analyze the branch at the end of the pred.
1768 172764 : MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
1769 : SmallVector<MachineOperand, 4> PredCond;
1770 207107 : if (PredBB != MBB && !PredBB->canFallThrough() &&
1771 61577 : !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) &&
1772 199728 : (!CurFallsThru || !CurTBB || !CurFBB) &&
1773 14141 : (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
1774 : // If the current block doesn't fall through, just move it.
1775 : // If the current block can fall through and does not end with a
1776 : // conditional branch, we need to append an unconditional jump to
1777 : // the (current) next block. To avoid a possible compile-time
1778 : // infinite loop, move blocks only backward in this case.
1779 : // Also, if there are already 2 branches here, we cannot add a third;
1780 : // this means we have the case
1781 : // Bcc next
1782 : // B elsewhere
1783 : // next:
1784 25581 : if (CurFallsThru) {
1785 : MachineBasicBlock *NextBB = &*std::next(MBB->getIterator());
1786 : CurCond.clear();
1787 25544 : TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc());
1788 : }
1789 25581 : MBB->moveAfter(PredBB);
1790 : MadeChange = true;
1791 : goto ReoptimizeBlock;
1792 : }
1793 : }
1794 : }
1795 :
1796 266799 : if (!CurFallsThru) {
1797 : // Check all successors to see if we can move this block before it.
1798 221826 : for (MachineBasicBlock *SuccBB : MBB->successors()) {
1799 : // Analyze the branch at the end of the block before the succ.
1800 105573 : MachineFunction::iterator SuccPrev = --SuccBB->getIterator();
1801 :
1802 : // If this block doesn't already fall-through to that successor, and if
1803 : // the succ doesn't already have a block that can fall through into it,
1804 : // and if the successor isn't an EH destination, we can arrange for the
1805 : // fallthrough to happen.
1806 210538 : if (SuccBB != MBB && &*SuccPrev != MBB &&
1807 215638 : !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
1808 3737 : !SuccBB->isEHPad()) {
1809 3513 : MBB->moveBefore(SuccBB);
1810 : MadeChange = true;
1811 : goto ReoptimizeBlock;
1812 : }
1813 : }
1814 :
1815 : // Okay, there is no really great place to put this block. If, however,
1816 : // the block before this one would be a fall-through if this block were
1817 : // removed, move this block to the end of the function. There is no real
1818 : // advantage in "falling through" to an EH block, so we don't want to
1819 : // perform this transformation for that case.
1820 : //
1821 : // Also, Windows EH introduced the possibility of an arbitrary number of
1822 : // successors to a given block. The analyzeBranch call does not consider
1823 : // exception handling and so we can get in a state where a block
1824 : // containing a call is followed by multiple EH blocks that would be
1825 : // rotated infinitely at the end of the function if the transformation
1826 : // below were performed for EH "FallThrough" blocks. Therefore, even if
1827 : // that appears not to be happening anymore, we should assume that it is
1828 : // possible and not remove the "!FallThrough()->isEHPad" condition below.
1829 116253 : MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr;
1830 : SmallVector<MachineOperand, 4> PrevCond;
1831 102491 : if (FallThrough != MF.end() &&
1832 136156 : !FallThrough->isEHPad() &&
1833 180726 : !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
1834 30808 : PrevBB.isSuccessor(&*FallThrough)) {
1835 11415 : MBB->moveAfter(&MF.back());
1836 : MadeChange = true;
1837 : return MadeChange;
1838 : }
1839 : }
1840 : }
1841 :
1842 : return MadeChange;
1843 : }
1844 :
1845 : //===----------------------------------------------------------------------===//
1846 : // Hoist Common Code
1847 : //===----------------------------------------------------------------------===//
1848 :
1849 0 : bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
1850 : bool MadeChange = false;
1851 1067036 : for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
1852 : MachineBasicBlock *MBB = &*I++;
1853 858925 : MadeChange |= HoistCommonCodeInSuccs(MBB);
1854 : }
1855 :
1856 0 : return MadeChange;
1857 : }
1858 :
1859 : /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
1860 : /// its 'true' successor.
1861 : static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
1862 : MachineBasicBlock *TrueBB) {
1863 244680 : for (MachineBasicBlock *SuccBB : BB->successors())
1864 244660 : if (SuccBB != TrueBB)
1865 : return SuccBB;
1866 : return nullptr;
1867 : }
1868 :
1869 : template <class Container>
1870 315744 : static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
1871 : Container &Set) {
1872 631488 : if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1873 1662465 : for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1874 1347031 : Set.insert(*AI);
1875 : } else {
1876 310 : Set.insert(Reg);
1877 : }
1878 315744 : }
1879 :
1880 : /// findHoistingInsertPosAndDeps - Find the location to move common instructions
1881 : /// in successors to. The location is usually just before the terminator,
1882 : /// however if the terminator is a conditional branch and its previous
1883 : /// instruction is the flag setting instruction, the previous instruction is
1884 : /// the preferred location. This function also gathers uses and defs of the
1885 : /// instructions from the insertion point to the end of the block. The data is
1886 : /// used by HoistCommonCodeInSuccs to ensure safety.
1887 : static
1888 108068 : MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
1889 : const TargetInstrInfo *TII,
1890 : const TargetRegisterInfo *TRI,
1891 : SmallSet<unsigned,4> &Uses,
1892 : SmallSet<unsigned,4> &Defs) {
1893 108068 : MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
1894 108068 : if (!TII->isUnpredicatedTerminator(*Loc))
1895 : return MBB->end();
1896 :
1897 327885 : for (const MachineOperand &MO : Loc->operands()) {
1898 220371 : if (!MO.isReg())
1899 : continue;
1900 109015 : unsigned Reg = MO.getReg();
1901 109015 : if (!Reg)
1902 : continue;
1903 109015 : if (MO.isUse()) {
1904 108270 : addRegAndItsAliases(Reg, TRI, Uses);
1905 : } else {
1906 745 : if (!MO.isDead())
1907 : // Don't try to hoist code in the rare case the terminator defines a
1908 : // register that is later used.
1909 : return MBB->end();
1910 :
1911 : // If the terminator defines a register, make sure we don't hoist
1912 : // the instruction whose def might be clobbered by the terminator.
1913 191 : addRegAndItsAliases(Reg, TRI, Defs);
1914 : }
1915 : }
1916 :
1917 : if (Uses.empty())
1918 0 : return Loc;
1919 : // If the terminator is the only instruction in the block and Uses is not
1920 : // empty (or we would have returned above), we can still safely hoist
1921 : // instructions just before the terminator as long as the Defs/Uses are not
1922 : // violated (which is checked in HoistCommonCodeInSuccs).
1923 107514 : if (Loc == MBB->begin())
1924 450 : return Loc;
1925 :
1926 : // The terminator is probably a conditional branch, try not to separate the
1927 : // branch from condition setting instruction.
1928 : MachineBasicBlock::iterator PI =
1929 107064 : skipDebugInstructionsBackward(std::prev(Loc), MBB->begin());
1930 :
1931 : bool IsDef = false;
1932 457814 : for (const MachineOperand &MO : PI->operands()) {
1933 : // If PI has a regmask operand, it is probably a call. Separate away.
1934 453428 : if (MO.isRegMask())
1935 2 : return Loc;
1936 453426 : if (!MO.isReg() || MO.isUse())
1937 346448 : continue;
1938 106978 : unsigned Reg = MO.getReg();
1939 106978 : if (!Reg)
1940 : continue;
1941 106978 : if (Uses.count(Reg)) {
1942 : IsDef = true;
1943 102676 : break;
1944 : }
1945 : }
1946 : if (!IsDef)
1947 : // The condition setting instruction is not just before the conditional
1948 : // branch.
1949 4386 : return Loc;
1950 :
1951 : // Be conservative, don't insert instruction above something that may have
1952 : // side-effects. And since it's potentially bad to separate flag setting
1953 : // instruction from the conditional branch, just abort the optimization
1954 : // completely.
1955 : // Also avoid moving code above predicated instruction since it's hard to
1956 : // reason about register liveness with predicated instruction.
1957 102676 : bool DontMoveAcrossStore = true;
1958 102676 : if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI))
1959 : return MBB->end();
1960 :
1961 : // Find out what registers are live. Note this routine is ignoring other live
1962 : // registers which are only used by instructions in successor blocks.
1963 297973 : for (const MachineOperand &MO : PI->operands()) {
1964 225212 : if (!MO.isReg())
1965 18399 : continue;
1966 208593 : unsigned Reg = MO.getReg();
1967 208593 : if (!Reg)
1968 : continue;
1969 206813 : if (MO.isUse()) {
1970 132316 : addRegAndItsAliases(Reg, TRI, Uses);
1971 : } else {
1972 74497 : if (Uses.erase(Reg)) {
1973 145536 : if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
1974 147504 : for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
1975 2068 : Uses.erase(*SubRegs); // Use sub-registers to be conservative
1976 : }
1977 : }
1978 74497 : addRegAndItsAliases(Reg, TRI, Defs);
1979 : }
1980 : }
1981 :
1982 72761 : return PI;
1983 : }
1984 :
1985 858925 : bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
1986 858925 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1987 : SmallVector<MachineOperand, 4> Cond;
1988 858925 : if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
1989 : return false;
1990 :
1991 341246 : if (!FBB) FBB = findFalseBlock(MBB, TBB);
1992 171697 : if (!FBB)
1993 : // Malformed bcc? True and false blocks are the same?
1994 : return false;
1995 :
1996 : // Restrict the optimization to cases where MBB is the only predecessor,
1997 : // it is an obvious win.
1998 171677 : if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
1999 : return false;
2000 :
2001 : // Find a suitable position to hoist the common instructions to. Also figure
2002 : // out which registers are used or defined by instructions from the insertion
2003 : // point to the end of the block.
2004 108068 : SmallSet<unsigned, 4> Uses, Defs;
2005 : MachineBasicBlock::iterator Loc =
2006 108068 : findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
2007 108068 : if (Loc == MBB->end())
2008 : return false;
2009 :
2010 : bool HasDups = false;
2011 77599 : SmallSet<unsigned, 4> ActiveDefsSet, AllDefsSet;
2012 77599 : MachineBasicBlock::iterator TIB = TBB->begin();
2013 77599 : MachineBasicBlock::iterator FIB = FBB->begin();
2014 : MachineBasicBlock::iterator TIE = TBB->end();
2015 : MachineBasicBlock::iterator FIE = FBB->end();
2016 77814 : while (TIB != TIE && FIB != FIE) {
2017 : // Skip dbg_value instructions. These do not count.
2018 77430 : TIB = skipDebugInstructionsForward(TIB, TIE);
2019 77430 : FIB = skipDebugInstructionsForward(FIB, FIE);
2020 77430 : if (TIB == TIE || FIB == FIE)
2021 : break;
2022 :
2023 77430 : if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead))
2024 : break;
2025 :
2026 780 : if (TII->isPredicated(*TIB))
2027 : // Hard to reason about register liveness with predicated instruction.
2028 : break;
2029 :
2030 : bool IsSafe = true;
2031 2550 : for (MachineOperand &MO : TIB->operands()) {
2032 : // Don't attempt to hoist instructions with register masks.
2033 2197 : if (MO.isRegMask()) {
2034 : IsSafe = false;
2035 426 : break;
2036 : }
2037 2137 : if (!MO.isReg())
2038 642 : continue;
2039 1688 : unsigned Reg = MO.getReg();
2040 1688 : if (!Reg)
2041 : continue;
2042 1495 : if (MO.isDef()) {
2043 792 : if (Uses.count(Reg)) {
2044 : // Avoid clobbering a register that's used by the instruction at
2045 : // the point of insertion.
2046 : IsSafe = false;
2047 : break;
2048 : }
2049 :
2050 606 : if (Defs.count(Reg) && !MO.isDead()) {
2051 : // Don't hoist the instruction if the def would be clobber by the
2052 : // instruction at the point insertion. FIXME: This is overly
2053 : // conservative. It should be possible to hoist the instructions
2054 : // in BB2 in the following example:
2055 : // BB1:
2056 : // r1, eflag = op1 r2, r3
2057 : // brcc eflag
2058 : //
2059 : // BB2:
2060 : // r1 = op2, ...
2061 : // = op3, killed r1
2062 : IsSafe = false;
2063 : break;
2064 : }
2065 703 : } else if (!ActiveDefsSet.count(Reg)) {
2066 691 : if (Defs.count(Reg)) {
2067 : // Use is defined by the instruction at the point of insertion.
2068 : IsSafe = false;
2069 : break;
2070 : }
2071 :
2072 691 : if (MO.isKill() && Uses.count(Reg))
2073 : // Kills a register that's read by the instruction at the point of
2074 : // insertion. Remove the kill marker.
2075 : MO.setIsKill(false);
2076 : }
2077 : }
2078 : if (!IsSafe)
2079 : break;
2080 :
2081 353 : bool DontMoveAcrossStore = true;
2082 353 : if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore))
2083 : break;
2084 :
2085 : // Remove kills from ActiveDefsSet, these registers had short live ranges.
2086 746 : for (const MachineOperand &MO : TIB->operands()) {
2087 531 : if (!MO.isReg() || !MO.isUse() || !MO.isKill())
2088 527 : continue;
2089 46 : unsigned Reg = MO.getReg();
2090 46 : if (!Reg)
2091 : continue;
2092 46 : if (!AllDefsSet.count(Reg)) {
2093 : continue;
2094 : }
2095 4 : if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
2096 38 : for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
2097 34 : ActiveDefsSet.erase(*AI);
2098 : } else {
2099 0 : ActiveDefsSet.erase(Reg);
2100 : }
2101 : }
2102 :
2103 : // Track local defs so we can update liveins.
2104 746 : for (const MachineOperand &MO : TIB->operands()) {
2105 531 : if (!MO.isReg() || !MO.isDef() || MO.isDead())
2106 : continue;
2107 235 : unsigned Reg = MO.getReg();
2108 235 : if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg))
2109 : continue;
2110 235 : addRegAndItsAliases(Reg, TRI, ActiveDefsSet);
2111 235 : addRegAndItsAliases(Reg, TRI, AllDefsSet);
2112 : }
2113 :
2114 : HasDups = true;
2115 : ++TIB;
2116 : ++FIB;
2117 : }
2118 :
2119 77599 : if (!HasDups)
2120 : return false;
2121 :
2122 178 : MBB->splice(Loc, TBB, TBB->begin(), TIB);
2123 178 : FBB->erase(FBB->begin(), FIB);
2124 :
2125 178 : if (UpdateLiveIns) {
2126 172 : recomputeLiveIns(*TBB);
2127 172 : recomputeLiveIns(*FBB);
2128 : }
2129 :
2130 : ++NumHoist;
2131 : return true;
2132 : }
|
