1 //===-- AMDILCFGStructurizer.cpp - CFG Structurizer -----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
9 //==-----------------------------------------------------------------------===//
11 #define DEBUG_TYPE "structcfg"
14 #include "AMDGPUInstrInfo.h"
15 #include "R600InstrInfo.h"
16 #include "llvm/ADT/DepthFirstIterator.h"
17 #include "llvm/ADT/SCCIterator.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/CodeGen/MachineDominators.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/MachineLoopInfo.h"
27 #include "llvm/CodeGen/MachinePostDominators.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetInstrInfo.h"
33 #include "llvm/Target/TargetMachine.h"
37 #define DEFAULT_VEC_SLOTS 8
41 //===----------------------------------------------------------------------===//
43 // Statistics for CFGStructurizer.
45 //===----------------------------------------------------------------------===//
47 STATISTIC(numSerialPatternMatch, "CFGStructurizer number of serial pattern "
49 STATISTIC(numIfPatternMatch, "CFGStructurizer number of if pattern "
51 STATISTIC(numLoopcontPatternMatch, "CFGStructurizer number of loop-continue "
53 STATISTIC(numClonedBlock, "CFGStructurizer cloned blocks");
54 STATISTIC(numClonedInstr, "CFGStructurizer cloned instructions");
57 void initializeAMDGPUCFGStructurizerPass(PassRegistry&);
60 //===----------------------------------------------------------------------===//
62 // Miscellaneous utility for CFGStructurizer.
64 //===----------------------------------------------------------------------===//
66 #define SHOWNEWINSTR(i) \
67 DEBUG(dbgs() << "New instr: " << *i << "\n");
69 #define SHOWNEWBLK(b, msg) \
71 dbgs() << msg << "BB" << b->getNumber() << "size " << b->size(); \
75 #define SHOWBLK_DETAIL(b, msg) \
78 dbgs() << msg << "BB" << b->getNumber() << "size " << b->size(); \
84 #define INVALIDSCCNUM -1
87 void ReverseVector(SmallVectorImpl<NodeT *> &Src) {
88 size_t sz = Src.size();
89 for (size_t i = 0; i < sz/2; ++i) {
91 Src[i] = Src[sz - i - 1];
96 } // end anonymous namespace
98 //===----------------------------------------------------------------------===//
100 // supporting data structure for CFGStructurizer
102 //===----------------------------------------------------------------------===//
107 class BlockInformation {
111 BlockInformation() : IsRetired(false), SccNum(INVALIDSCCNUM) {}
114 } // end anonymous namespace
116 //===----------------------------------------------------------------------===//
120 //===----------------------------------------------------------------------===//
123 class AMDGPUCFGStructurizer : public MachineFunctionPass {
125 typedef SmallVector<MachineBasicBlock *, 32> MBBVector;
126 typedef std::map<MachineBasicBlock *, BlockInformation *> MBBInfoMap;
127 typedef std::map<MachineLoop *, MachineBasicBlock *> LoopLandInfoMap;
131 SinglePath_InPath = 1,
132 SinglePath_NotInPath = 2
137 AMDGPUCFGStructurizer() :
138 MachineFunctionPass(ID), TII(NULL), TRI(NULL) {
139 initializeAMDGPUCFGStructurizerPass(*PassRegistry::getPassRegistry());
142 const char *getPassName() const {
143 return "AMDGPU Control Flow Graph structurizer Pass";
146 void getAnalysisUsage(AnalysisUsage &AU) const {
147 AU.addPreserved<MachineFunctionAnalysis>();
148 AU.addRequired<MachineFunctionAnalysis>();
149 AU.addRequired<MachineDominatorTree>();
150 AU.addRequired<MachinePostDominatorTree>();
151 AU.addRequired<MachineLoopInfo>();
154 /// Perform the CFG structurization
157 /// Perform the CFG preparation
158 /// This step will remove every unconditionnal/dead jump instructions and make
159 /// sure all loops have an exit block
162 bool runOnMachineFunction(MachineFunction &MF) {
163 TII = static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo());
164 TRI = &TII->getRegisterInfo();
168 MLI = &getAnalysis<MachineLoopInfo>();
169 DEBUG(dbgs() << "LoopInfo:\n"; PrintLoopinfo(*MLI););
170 MDT = &getAnalysis<MachineDominatorTree>();
171 DEBUG(MDT->print(dbgs(), (const llvm::Module*)0););
172 PDT = &getAnalysis<MachinePostDominatorTree>();
173 DEBUG(PDT->print(dbgs()););
181 MachineDominatorTree *MDT;
182 MachinePostDominatorTree *PDT;
183 MachineLoopInfo *MLI;
184 const R600InstrInfo *TII;
185 const AMDGPURegisterInfo *TRI;
188 /// Print the ordered Blocks.
189 void printOrderedBlocks() const {
191 for (MBBVector::const_iterator iterBlk = OrderedBlks.begin(),
192 iterBlkEnd = OrderedBlks.end(); iterBlk != iterBlkEnd; ++iterBlk, ++i) {
193 dbgs() << "BB" << (*iterBlk)->getNumber();
194 dbgs() << "(" << getSCCNum(*iterBlk) << "," << (*iterBlk)->size() << ")";
195 if (i != 0 && i % 10 == 0) {
202 static void PrintLoopinfo(const MachineLoopInfo &LoopInfo) {
203 for (MachineLoop::iterator iter = LoopInfo.begin(),
204 iterEnd = LoopInfo.end(); iter != iterEnd; ++iter) {
205 (*iter)->print(dbgs(), 0);
210 int getSCCNum(MachineBasicBlock *MBB) const;
211 MachineBasicBlock *getLoopLandInfo(MachineLoop *LoopRep) const;
212 bool hasBackEdge(MachineBasicBlock *MBB) const;
213 static unsigned getLoopDepth(MachineLoop *LoopRep);
214 bool isRetiredBlock(MachineBasicBlock *MBB) const;
215 bool isActiveLoophead(MachineBasicBlock *MBB) const;
216 PathToKind singlePathTo(MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB,
217 bool AllowSideEntry = true) const;
218 int countActiveBlock(MBBVector::const_iterator It,
219 MBBVector::const_iterator E) const;
220 bool needMigrateBlock(MachineBasicBlock *MBB) const;
223 void reversePredicateSetter(MachineBasicBlock::iterator I);
224 /// Compute the reversed DFS post order of Blocks
225 void orderBlocks(MachineFunction *MF);
227 // Function originally from CFGStructTraits
228 void insertInstrEnd(MachineBasicBlock *MBB, int NewOpcode,
229 DebugLoc DL = DebugLoc());
230 MachineInstr *insertInstrBefore(MachineBasicBlock *MBB, int NewOpcode,
231 DebugLoc DL = DebugLoc());
232 MachineInstr *insertInstrBefore(MachineBasicBlock::iterator I, int NewOpcode);
233 void insertCondBranchBefore(MachineBasicBlock::iterator I, int NewOpcode,
235 void insertCondBranchBefore(MachineBasicBlock *MBB,
236 MachineBasicBlock::iterator I, int NewOpcode, int RegNum,
238 void insertCondBranchEnd(MachineBasicBlock *MBB, int NewOpcode, int RegNum);
239 static int getBranchNzeroOpcode(int OldOpcode);
240 static int getBranchZeroOpcode(int OldOpcode);
241 static int getContinueNzeroOpcode(int OldOpcode);
242 static int getContinueZeroOpcode(int OldOpcode);
243 static MachineBasicBlock *getTrueBranch(MachineInstr *MI);
244 static void setTrueBranch(MachineInstr *MI, MachineBasicBlock *MBB);
245 static MachineBasicBlock *getFalseBranch(MachineBasicBlock *MBB,
247 static bool isCondBranch(MachineInstr *MI);
248 static bool isUncondBranch(MachineInstr *MI);
249 static DebugLoc getLastDebugLocInBB(MachineBasicBlock *MBB);
250 static MachineInstr *getNormalBlockBranchInstr(MachineBasicBlock *MBB);
251 /// The correct naming for this is getPossibleLoopendBlockBranchInstr.
253 /// BB with backward-edge could have move instructions after the branch
254 /// instruction. Such move instruction "belong to" the loop backward-edge.
255 MachineInstr *getLoopendBlockBranchInstr(MachineBasicBlock *MBB);
256 static MachineInstr *getReturnInstr(MachineBasicBlock *MBB);
257 static MachineInstr *getContinueInstr(MachineBasicBlock *MBB);
258 static bool isReturnBlock(MachineBasicBlock *MBB);
259 static void cloneSuccessorList(MachineBasicBlock *DstMBB,
260 MachineBasicBlock *SrcMBB) ;
261 static MachineBasicBlock *clone(MachineBasicBlock *MBB);
262 /// MachineBasicBlock::ReplaceUsesOfBlockWith doesn't serve the purpose
263 /// because the AMDGPU instruction is not recognized as terminator fix this
264 /// and retire this routine
265 void replaceInstrUseOfBlockWith(MachineBasicBlock *SrcMBB,
266 MachineBasicBlock *OldMBB, MachineBasicBlock *NewBlk);
267 static void wrapup(MachineBasicBlock *MBB);
270 int patternMatch(MachineBasicBlock *MBB);
271 int patternMatchGroup(MachineBasicBlock *MBB);
272 int serialPatternMatch(MachineBasicBlock *MBB);
273 int ifPatternMatch(MachineBasicBlock *MBB);
274 int loopendPatternMatch();
275 int mergeLoop(MachineLoop *LoopRep);
276 int loopcontPatternMatch(MachineLoop *LoopRep, MachineBasicBlock *LoopHeader);
278 void handleLoopcontBlock(MachineBasicBlock *ContingMBB,
279 MachineLoop *ContingLoop, MachineBasicBlock *ContMBB,
280 MachineLoop *ContLoop);
281 /// return true iff src1Blk->succ_size() == 0 && src1Blk and src2Blk are in
282 /// the same loop with LoopLandInfo without explicitly keeping track of
283 /// loopContBlks and loopBreakBlks, this is a method to get the information.
284 bool isSameloopDetachedContbreak(MachineBasicBlock *Src1MBB,
285 MachineBasicBlock *Src2MBB);
286 int handleJumpintoIf(MachineBasicBlock *HeadMBB,
287 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB);
288 int handleJumpintoIfImp(MachineBasicBlock *HeadMBB,
289 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB);
290 int improveSimpleJumpintoIf(MachineBasicBlock *HeadMBB,
291 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB,
292 MachineBasicBlock **LandMBBPtr);
293 void showImproveSimpleJumpintoIf(MachineBasicBlock *HeadMBB,
294 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB,
295 MachineBasicBlock *LandMBB, bool Detail = false);
296 int cloneOnSideEntryTo(MachineBasicBlock *PreMBB,
297 MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB);
298 void mergeSerialBlock(MachineBasicBlock *DstMBB,
299 MachineBasicBlock *SrcMBB);
301 void mergeIfthenelseBlock(MachineInstr *BranchMI,
302 MachineBasicBlock *MBB, MachineBasicBlock *TrueMBB,
303 MachineBasicBlock *FalseMBB, MachineBasicBlock *LandMBB);
304 void mergeLooplandBlock(MachineBasicBlock *DstMBB,
305 MachineBasicBlock *LandMBB);
306 void mergeLoopbreakBlock(MachineBasicBlock *ExitingMBB,
307 MachineBasicBlock *LandMBB);
308 void settleLoopcontBlock(MachineBasicBlock *ContingMBB,
309 MachineBasicBlock *ContMBB);
310 /// normalizeInfiniteLoopExit change
312 /// uncond_br LoopHeader
316 /// cond_br 1 LoopHeader dummyExit
317 /// and return the newly added dummy exit block
318 MachineBasicBlock *normalizeInfiniteLoopExit(MachineLoop *LoopRep);
319 void removeUnconditionalBranch(MachineBasicBlock *MBB);
320 /// Remove duplicate branches instructions in a block.
325 /// is transformed to
328 void removeRedundantConditionalBranch(MachineBasicBlock *MBB);
329 void addDummyExitBlock(SmallVectorImpl<MachineBasicBlock *> &RetMBB);
330 void removeSuccessor(MachineBasicBlock *MBB);
331 MachineBasicBlock *cloneBlockForPredecessor(MachineBasicBlock *MBB,
332 MachineBasicBlock *PredMBB);
333 void migrateInstruction(MachineBasicBlock *SrcMBB,
334 MachineBasicBlock *DstMBB, MachineBasicBlock::iterator I);
335 void recordSccnum(MachineBasicBlock *MBB, int SCCNum);
336 void retireBlock(MachineBasicBlock *MBB);
337 void setLoopLandBlock(MachineLoop *LoopRep, MachineBasicBlock *MBB = NULL);
339 MachineBasicBlock *findNearestCommonPostDom(std::set<MachineBasicBlock *>&);
340 /// This is work around solution for findNearestCommonDominator not avaiable
341 /// to post dom a proper fix should go to Dominators.h.
342 MachineBasicBlock *findNearestCommonPostDom(MachineBasicBlock *MBB1,
343 MachineBasicBlock *MBB2);
346 MBBInfoMap BlockInfoMap;
347 LoopLandInfoMap LLInfoMap;
348 std::map<MachineLoop *, bool> Visited;
349 MachineFunction *FuncRep;
350 SmallVector<MachineBasicBlock *, DEFAULT_VEC_SLOTS> OrderedBlks;
353 int AMDGPUCFGStructurizer::getSCCNum(MachineBasicBlock *MBB) const {
354 MBBInfoMap::const_iterator It = BlockInfoMap.find(MBB);
355 if (It == BlockInfoMap.end())
356 return INVALIDSCCNUM;
357 return (*It).second->SccNum;
360 MachineBasicBlock *AMDGPUCFGStructurizer::getLoopLandInfo(MachineLoop *LoopRep)
362 LoopLandInfoMap::const_iterator It = LLInfoMap.find(LoopRep);
363 if (It == LLInfoMap.end())
368 bool AMDGPUCFGStructurizer::hasBackEdge(MachineBasicBlock *MBB) const {
369 MachineLoop *LoopRep = MLI->getLoopFor(MBB);
372 MachineBasicBlock *LoopHeader = LoopRep->getHeader();
373 return MBB->isSuccessor(LoopHeader);
376 unsigned AMDGPUCFGStructurizer::getLoopDepth(MachineLoop *LoopRep) {
377 return LoopRep ? LoopRep->getLoopDepth() : 0;
380 bool AMDGPUCFGStructurizer::isRetiredBlock(MachineBasicBlock *MBB) const {
381 MBBInfoMap::const_iterator It = BlockInfoMap.find(MBB);
382 if (It == BlockInfoMap.end())
384 return (*It).second->IsRetired;
387 bool AMDGPUCFGStructurizer::isActiveLoophead(MachineBasicBlock *MBB) const {
388 MachineLoop *LoopRep = MLI->getLoopFor(MBB);
389 while (LoopRep && LoopRep->getHeader() == MBB) {
390 MachineBasicBlock *LoopLand = getLoopLandInfo(LoopRep);
393 if (!isRetiredBlock(LoopLand))
395 LoopRep = LoopRep->getParentLoop();
399 AMDGPUCFGStructurizer::PathToKind AMDGPUCFGStructurizer::singlePathTo(
400 MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB,
401 bool AllowSideEntry) const {
403 if (SrcMBB == DstMBB)
404 return SinglePath_InPath;
405 while (SrcMBB && SrcMBB->succ_size() == 1) {
406 SrcMBB = *SrcMBB->succ_begin();
407 if (SrcMBB == DstMBB)
408 return SinglePath_InPath;
409 if (!AllowSideEntry && SrcMBB->pred_size() > 1)
410 return Not_SinglePath;
412 if (SrcMBB && SrcMBB->succ_size()==0)
413 return SinglePath_NotInPath;
414 return Not_SinglePath;
417 int AMDGPUCFGStructurizer::countActiveBlock(MBBVector::const_iterator It,
418 MBBVector::const_iterator E) const {
421 if (!isRetiredBlock(*It))
428 bool AMDGPUCFGStructurizer::needMigrateBlock(MachineBasicBlock *MBB) const {
429 unsigned BlockSizeThreshold = 30;
430 unsigned CloneInstrThreshold = 100;
431 bool MultiplePreds = MBB && (MBB->pred_size() > 1);
435 unsigned BlkSize = MBB->size();
436 return ((BlkSize > BlockSizeThreshold) &&
437 (BlkSize * (MBB->pred_size() - 1) > CloneInstrThreshold));
440 void AMDGPUCFGStructurizer::reversePredicateSetter(
441 MachineBasicBlock::iterator I) {
443 if (I->getOpcode() == AMDGPU::PRED_X) {
444 switch (static_cast<MachineInstr *>(I)->getOperand(2).getImm()) {
445 case OPCODE_IS_ZERO_INT:
446 static_cast<MachineInstr *>(I)->getOperand(2)
447 .setImm(OPCODE_IS_NOT_ZERO_INT);
449 case OPCODE_IS_NOT_ZERO_INT:
450 static_cast<MachineInstr *>(I)->getOperand(2)
451 .setImm(OPCODE_IS_ZERO_INT);
454 static_cast<MachineInstr *>(I)->getOperand(2)
455 .setImm(OPCODE_IS_NOT_ZERO);
457 case OPCODE_IS_NOT_ZERO:
458 static_cast<MachineInstr *>(I)->getOperand(2)
459 .setImm(OPCODE_IS_ZERO);
462 llvm_unreachable("PRED_X Opcode invalid!");
468 void AMDGPUCFGStructurizer::insertInstrEnd(MachineBasicBlock *MBB,
469 int NewOpcode, DebugLoc DL) {
470 MachineInstr *MI = MBB->getParent()
471 ->CreateMachineInstr(TII->get(NewOpcode), DL);
473 //assume the instruction doesn't take any reg operand ...
477 MachineInstr *AMDGPUCFGStructurizer::insertInstrBefore(MachineBasicBlock *MBB,
478 int NewOpcode, DebugLoc DL) {
480 MBB->getParent()->CreateMachineInstr(TII->get(NewOpcode), DL);
481 if (MBB->begin() != MBB->end())
482 MBB->insert(MBB->begin(), MI);
489 MachineInstr *AMDGPUCFGStructurizer::insertInstrBefore(
490 MachineBasicBlock::iterator I, int NewOpcode) {
491 MachineInstr *OldMI = &(*I);
492 MachineBasicBlock *MBB = OldMI->getParent();
493 MachineInstr *NewMBB =
494 MBB->getParent()->CreateMachineInstr(TII->get(NewOpcode), DebugLoc());
495 MBB->insert(I, NewMBB);
496 //assume the instruction doesn't take any reg operand ...
497 SHOWNEWINSTR(NewMBB);
501 void AMDGPUCFGStructurizer::insertCondBranchBefore(
502 MachineBasicBlock::iterator I, int NewOpcode, DebugLoc DL) {
503 MachineInstr *OldMI = &(*I);
504 MachineBasicBlock *MBB = OldMI->getParent();
505 MachineFunction *MF = MBB->getParent();
506 MachineInstr *NewMI = MF->CreateMachineInstr(TII->get(NewOpcode), DL);
507 MBB->insert(I, NewMI);
508 MachineInstrBuilder MIB(*MF, NewMI);
509 MIB.addReg(OldMI->getOperand(1).getReg(), false);
511 //erase later oldInstr->eraseFromParent();
514 void AMDGPUCFGStructurizer::insertCondBranchBefore(MachineBasicBlock *blk,
515 MachineBasicBlock::iterator I, int NewOpcode, int RegNum,
517 MachineFunction *MF = blk->getParent();
518 MachineInstr *NewInstr = MF->CreateMachineInstr(TII->get(NewOpcode), DL);
520 blk->insert(I, NewInstr);
521 MachineInstrBuilder(*MF, NewInstr).addReg(RegNum, false);
522 SHOWNEWINSTR(NewInstr);
525 void AMDGPUCFGStructurizer::insertCondBranchEnd(MachineBasicBlock *MBB,
526 int NewOpcode, int RegNum) {
527 MachineFunction *MF = MBB->getParent();
528 MachineInstr *NewInstr =
529 MF->CreateMachineInstr(TII->get(NewOpcode), DebugLoc());
530 MBB->push_back(NewInstr);
531 MachineInstrBuilder(*MF, NewInstr).addReg(RegNum, false);
532 SHOWNEWINSTR(NewInstr);
535 int AMDGPUCFGStructurizer::getBranchNzeroOpcode(int OldOpcode) {
537 case AMDGPU::JUMP_COND:
538 case AMDGPU::JUMP: return AMDGPU::IF_PREDICATE_SET;
539 case AMDGPU::BRANCH_COND_i32:
540 case AMDGPU::BRANCH_COND_f32: return AMDGPU::IF_LOGICALNZ_f32;
541 default: llvm_unreachable("internal error");
546 int AMDGPUCFGStructurizer::getBranchZeroOpcode(int OldOpcode) {
548 case AMDGPU::JUMP_COND:
549 case AMDGPU::JUMP: return AMDGPU::IF_PREDICATE_SET;
550 case AMDGPU::BRANCH_COND_i32:
551 case AMDGPU::BRANCH_COND_f32: return AMDGPU::IF_LOGICALZ_f32;
552 default: llvm_unreachable("internal error");
557 int AMDGPUCFGStructurizer::getContinueNzeroOpcode(int OldOpcode) {
559 case AMDGPU::JUMP_COND:
560 case AMDGPU::JUMP: return AMDGPU::CONTINUE_LOGICALNZ_i32;
561 default: llvm_unreachable("internal error");
566 int AMDGPUCFGStructurizer::getContinueZeroOpcode(int OldOpcode) {
568 case AMDGPU::JUMP_COND:
569 case AMDGPU::JUMP: return AMDGPU::CONTINUE_LOGICALZ_i32;
570 default: llvm_unreachable("internal error");
575 MachineBasicBlock *AMDGPUCFGStructurizer::getTrueBranch(MachineInstr *MI) {
576 return MI->getOperand(0).getMBB();
579 void AMDGPUCFGStructurizer::setTrueBranch(MachineInstr *MI,
580 MachineBasicBlock *MBB) {
581 MI->getOperand(0).setMBB(MBB);
585 AMDGPUCFGStructurizer::getFalseBranch(MachineBasicBlock *MBB,
587 assert(MBB->succ_size() == 2);
588 MachineBasicBlock *TrueBranch = getTrueBranch(MI);
589 MachineBasicBlock::succ_iterator It = MBB->succ_begin();
590 MachineBasicBlock::succ_iterator Next = It;
592 return (*It == TrueBranch) ? *Next : *It;
595 bool AMDGPUCFGStructurizer::isCondBranch(MachineInstr *MI) {
596 switch (MI->getOpcode()) {
597 case AMDGPU::JUMP_COND:
598 case AMDGPU::BRANCH_COND_i32:
599 case AMDGPU::BRANCH_COND_f32: return true;
606 bool AMDGPUCFGStructurizer::isUncondBranch(MachineInstr *MI) {
607 switch (MI->getOpcode()) {
617 DebugLoc AMDGPUCFGStructurizer::getLastDebugLocInBB(MachineBasicBlock *MBB) {
618 //get DebugLoc from the first MachineBasicBlock instruction with debug info
620 for (MachineBasicBlock::iterator It = MBB->begin(); It != MBB->end();
622 MachineInstr *instr = &(*It);
623 if (instr->getDebugLoc().isUnknown() == false)
624 DL = instr->getDebugLoc();
629 MachineInstr *AMDGPUCFGStructurizer::getNormalBlockBranchInstr(
630 MachineBasicBlock *MBB) {
631 MachineBasicBlock::reverse_iterator It = MBB->rbegin();
632 MachineInstr *MI = &*It;
633 if (MI && (isCondBranch(MI) || isUncondBranch(MI)))
638 MachineInstr *AMDGPUCFGStructurizer::getLoopendBlockBranchInstr(
639 MachineBasicBlock *MBB) {
640 for (MachineBasicBlock::reverse_iterator It = MBB->rbegin(), E = MBB->rend();
643 MachineInstr *MI = &*It;
645 if (isCondBranch(MI) || isUncondBranch(MI))
647 else if (!TII->isMov(MI->getOpcode()))
654 MachineInstr *AMDGPUCFGStructurizer::getReturnInstr(MachineBasicBlock *MBB) {
655 MachineBasicBlock::reverse_iterator It = MBB->rbegin();
656 if (It != MBB->rend()) {
657 MachineInstr *instr = &(*It);
658 if (instr->getOpcode() == AMDGPU::RETURN)
664 MachineInstr *AMDGPUCFGStructurizer::getContinueInstr(MachineBasicBlock *MBB) {
665 MachineBasicBlock::reverse_iterator It = MBB->rbegin();
666 if (It != MBB->rend()) {
667 MachineInstr *MI = &(*It);
668 if (MI->getOpcode() == AMDGPU::CONTINUE)
674 bool AMDGPUCFGStructurizer::isReturnBlock(MachineBasicBlock *MBB) {
675 MachineInstr *MI = getReturnInstr(MBB);
676 bool IsReturn = (MBB->succ_size() == 0);
681 dbgs() << "BB" << MBB->getNumber()
682 <<" is return block without RETURN instr\n";);
686 void AMDGPUCFGStructurizer::cloneSuccessorList(MachineBasicBlock *DstMBB,
687 MachineBasicBlock *SrcMBB) {
688 for (MachineBasicBlock::succ_iterator It = SrcMBB->succ_begin(),
689 iterEnd = SrcMBB->succ_end(); It != iterEnd; ++It)
690 DstMBB->addSuccessor(*It); // *iter's predecessor is also taken care of
693 MachineBasicBlock *AMDGPUCFGStructurizer::clone(MachineBasicBlock *MBB) {
694 MachineFunction *Func = MBB->getParent();
695 MachineBasicBlock *NewMBB = Func->CreateMachineBasicBlock();
696 Func->push_back(NewMBB); //insert to function
697 for (MachineBasicBlock::iterator It = MBB->begin(), E = MBB->end();
699 MachineInstr *MI = Func->CloneMachineInstr(It);
700 NewMBB->push_back(MI);
705 void AMDGPUCFGStructurizer::replaceInstrUseOfBlockWith(
706 MachineBasicBlock *SrcMBB, MachineBasicBlock *OldMBB,
707 MachineBasicBlock *NewBlk) {
708 MachineInstr *BranchMI = getLoopendBlockBranchInstr(SrcMBB);
709 if (BranchMI && isCondBranch(BranchMI) &&
710 getTrueBranch(BranchMI) == OldMBB)
711 setTrueBranch(BranchMI, NewBlk);
714 void AMDGPUCFGStructurizer::wrapup(MachineBasicBlock *MBB) {
715 assert((!MBB->getParent()->getJumpTableInfo()
716 || MBB->getParent()->getJumpTableInfo()->isEmpty())
717 && "found a jump table");
719 //collect continue right before endloop
720 SmallVector<MachineInstr *, DEFAULT_VEC_SLOTS> ContInstr;
721 MachineBasicBlock::iterator Pre = MBB->begin();
722 MachineBasicBlock::iterator E = MBB->end();
723 MachineBasicBlock::iterator It = Pre;
725 if (Pre->getOpcode() == AMDGPU::CONTINUE
726 && It->getOpcode() == AMDGPU::ENDLOOP)
727 ContInstr.push_back(Pre);
732 //delete continue right before endloop
733 for (unsigned i = 0; i < ContInstr.size(); ++i)
734 ContInstr[i]->eraseFromParent();
736 // TODO to fix up jump table so later phase won't be confused. if
737 // (jumpTableInfo->isEmpty() == false) { need to clean the jump table, but
738 // there isn't such an interface yet. alternatively, replace all the other
739 // blocks in the jump table with the entryBlk //}
744 bool AMDGPUCFGStructurizer::prepare() {
745 bool Changed = false;
747 //FIXME: if not reducible flow graph, make it so ???
749 DEBUG(dbgs() << "AMDGPUCFGStructurizer::prepare\n";);
751 orderBlocks(FuncRep);
753 SmallVector<MachineBasicBlock *, DEFAULT_VEC_SLOTS> RetBlks;
755 // Add an ExitBlk to loop that don't have one
756 for (MachineLoopInfo::iterator It = MLI->begin(),
757 E = MLI->end(); It != E; ++It) {
758 MachineLoop *LoopRep = (*It);
759 MBBVector ExitingMBBs;
760 LoopRep->getExitingBlocks(ExitingMBBs);
762 if (ExitingMBBs.size() == 0) {
763 MachineBasicBlock* DummyExitBlk = normalizeInfiniteLoopExit(LoopRep);
765 RetBlks.push_back(DummyExitBlk);
769 // Remove unconditional branch instr.
770 // Add dummy exit block iff there are multiple returns.
771 for (SmallVectorImpl<MachineBasicBlock *>::const_iterator
772 It = OrderedBlks.begin(), E = OrderedBlks.end(); It != E; ++It) {
773 MachineBasicBlock *MBB = *It;
774 removeUnconditionalBranch(MBB);
775 removeRedundantConditionalBranch(MBB);
776 if (isReturnBlock(MBB)) {
777 RetBlks.push_back(MBB);
779 assert(MBB->succ_size() <= 2);
782 if (RetBlks.size() >= 2) {
783 addDummyExitBlock(RetBlks);
790 bool AMDGPUCFGStructurizer::run() {
792 //Assume reducible CFG...
793 DEBUG(dbgs() << "AMDGPUCFGStructurizer::run\n");
796 //Use the worse block ordering to test the algorithm.
797 ReverseVector(orderedBlks);
800 DEBUG(dbgs() << "Ordered blocks:\n"; printOrderedBlocks(););
803 MachineBasicBlock *MBB;
804 bool MakeProgress = false;
805 int NumRemainedBlk = countActiveBlock(OrderedBlks.begin(),
811 dbgs() << "numIter = " << NumIter
812 << ", numRemaintedBlk = " << NumRemainedBlk << "\n";
815 SmallVectorImpl<MachineBasicBlock *>::const_iterator It =
817 SmallVectorImpl<MachineBasicBlock *>::const_iterator E =
820 SmallVectorImpl<MachineBasicBlock *>::const_iterator SccBeginIter =
822 MachineBasicBlock *SccBeginMBB = NULL;
823 int SccNumBlk = 0; // The number of active blocks, init to a
824 // maximum possible number.
825 int SccNumIter; // Number of iteration in this SCC.
834 SccNumBlk = NumRemainedBlk; // Init to maximum possible number.
836 dbgs() << "start processing SCC" << getSCCNum(SccBeginMBB);
841 if (!isRetiredBlock(MBB))
846 bool ContNextScc = true;
848 || getSCCNum(SccBeginMBB) != getSCCNum(*It)) {
849 // Just finish one scc.
851 int sccRemainedNumBlk = countActiveBlock(SccBeginIter, It);
852 if (sccRemainedNumBlk != 1 && sccRemainedNumBlk >= SccNumBlk) {
854 dbgs() << "Can't reduce SCC " << getSCCNum(MBB)
855 << ", sccNumIter = " << SccNumIter;
856 dbgs() << "doesn't make any progress\n";
859 } else if (sccRemainedNumBlk != 1 && sccRemainedNumBlk < SccNumBlk) {
860 SccNumBlk = sccRemainedNumBlk;
864 dbgs() << "repeat processing SCC" << getSCCNum(MBB)
865 << "sccNumIter = " << SccNumIter << '\n';
868 // Finish the current scc.
872 // Continue on next component in the current scc.
878 } //while, "one iteration" over the function.
880 MachineBasicBlock *EntryMBB =
881 GraphTraits<MachineFunction *>::nodes_begin(FuncRep);
882 if (EntryMBB->succ_size() == 0) {
885 dbgs() << "Reduce to one block\n";
888 int NewnumRemainedBlk
889 = countActiveBlock(OrderedBlks.begin(), OrderedBlks.end());
890 // consider cloned blocks ??
891 if (NewnumRemainedBlk == 1 || NewnumRemainedBlk < NumRemainedBlk) {
893 NumRemainedBlk = NewnumRemainedBlk;
895 MakeProgress = false;
897 dbgs() << "No progress\n";
901 } while (!Finish && MakeProgress);
903 // Misc wrap up to maintain the consistency of the Function representation.
904 wrapup(GraphTraits<MachineFunction *>::nodes_begin(FuncRep));
906 // Detach retired Block, release memory.
907 for (MBBInfoMap::iterator It = BlockInfoMap.begin(), E = BlockInfoMap.end();
909 if ((*It).second && (*It).second->IsRetired) {
910 assert(((*It).first)->getNumber() != -1);
912 dbgs() << "Erase BB" << ((*It).first)->getNumber() << "\n";
914 (*It).first->eraseFromParent(); //Remove from the parent Function.
918 BlockInfoMap.clear();
922 DEBUG(FuncRep->viewCFG());
923 llvm_unreachable("IRREDUCIBLE_CFG");
931 void AMDGPUCFGStructurizer::orderBlocks(MachineFunction *MF) {
933 MachineBasicBlock *MBB;
934 for (scc_iterator<MachineFunction *> It = scc_begin(MF); !It.isAtEnd();
936 std::vector<MachineBasicBlock *> &SccNext = *It;
937 for (std::vector<MachineBasicBlock *>::const_iterator
938 blockIter = SccNext.begin(), blockEnd = SccNext.end();
939 blockIter != blockEnd; ++blockIter) {
941 OrderedBlks.push_back(MBB);
942 recordSccnum(MBB, SccNum);
946 //walk through all the block in func to check for unreachable
947 typedef GraphTraits<MachineFunction *> GTM;
948 MachineFunction::iterator It = GTM::nodes_begin(MF), E = GTM::nodes_end(MF);
949 for (; It != E; ++It) {
950 MachineBasicBlock *MBB = &(*It);
951 SccNum = getSCCNum(MBB);
952 if (SccNum == INVALIDSCCNUM)
953 dbgs() << "unreachable block BB" << MBB->getNumber() << "\n";
957 int AMDGPUCFGStructurizer::patternMatch(MachineBasicBlock *MBB) {
962 dbgs() << "Begin patternMatch BB" << MBB->getNumber() << "\n";
965 while ((CurMatch = patternMatchGroup(MBB)) > 0)
966 NumMatch += CurMatch;
969 dbgs() << "End patternMatch BB" << MBB->getNumber()
970 << ", numMatch = " << NumMatch << "\n";
976 int AMDGPUCFGStructurizer::patternMatchGroup(MachineBasicBlock *MBB) {
978 NumMatch += loopendPatternMatch();
979 NumMatch += serialPatternMatch(MBB);
980 NumMatch += ifPatternMatch(MBB);
985 int AMDGPUCFGStructurizer::serialPatternMatch(MachineBasicBlock *MBB) {
986 if (MBB->succ_size() != 1)
989 MachineBasicBlock *childBlk = *MBB->succ_begin();
990 if (childBlk->pred_size() != 1 || isActiveLoophead(childBlk))
993 mergeSerialBlock(MBB, childBlk);
994 ++numSerialPatternMatch;
998 int AMDGPUCFGStructurizer::ifPatternMatch(MachineBasicBlock *MBB) {
1000 if (MBB->succ_size() != 2)
1002 if (hasBackEdge(MBB))
1004 MachineInstr *BranchMI = getNormalBlockBranchInstr(MBB);
1008 assert(isCondBranch(BranchMI));
1011 MachineBasicBlock *TrueMBB = getTrueBranch(BranchMI);
1012 NumMatch += serialPatternMatch(TrueMBB);
1013 NumMatch += ifPatternMatch(TrueMBB);
1014 MachineBasicBlock *FalseMBB = getFalseBranch(MBB, BranchMI);
1015 NumMatch += serialPatternMatch(FalseMBB);
1016 NumMatch += ifPatternMatch(FalseMBB);
1017 MachineBasicBlock *LandBlk;
1020 assert (!TrueMBB->succ_empty() || !FalseMBB->succ_empty());
1022 if (TrueMBB->succ_size() == 1 && FalseMBB->succ_size() == 1
1023 && *TrueMBB->succ_begin() == *FalseMBB->succ_begin()) {
1025 LandBlk = *TrueMBB->succ_begin();
1026 } else if (TrueMBB->succ_size() == 1 && *TrueMBB->succ_begin() == FalseMBB) {
1027 // Triangle pattern, false is empty
1030 } else if (FalseMBB->succ_size() == 1
1031 && *FalseMBB->succ_begin() == TrueMBB) {
1032 // Triangle pattern, true is empty
1033 // We reverse the predicate to make a triangle, empty false pattern;
1034 std::swap(TrueMBB, FalseMBB);
1035 reversePredicateSetter(MBB->end());
1038 } else if (FalseMBB->succ_size() == 1
1039 && isSameloopDetachedContbreak(TrueMBB, FalseMBB)) {
1040 LandBlk = *FalseMBB->succ_begin();
1041 } else if (TrueMBB->succ_size() == 1
1042 && isSameloopDetachedContbreak(FalseMBB, TrueMBB)) {
1043 LandBlk = *TrueMBB->succ_begin();
1045 return NumMatch + handleJumpintoIf(MBB, TrueMBB, FalseMBB);
1048 // improveSimpleJumpinfoIf can handle the case where landBlk == NULL but the
1049 // new BB created for landBlk==NULL may introduce new challenge to the
1050 // reduction process.
1052 ((TrueMBB && TrueMBB->pred_size() > 1)
1053 || (FalseMBB && FalseMBB->pred_size() > 1))) {
1054 Cloned += improveSimpleJumpintoIf(MBB, TrueMBB, FalseMBB, &LandBlk);
1057 if (TrueMBB && TrueMBB->pred_size() > 1) {
1058 TrueMBB = cloneBlockForPredecessor(TrueMBB, MBB);
1062 if (FalseMBB && FalseMBB->pred_size() > 1) {
1063 FalseMBB = cloneBlockForPredecessor(FalseMBB, MBB);
1067 mergeIfthenelseBlock(BranchMI, MBB, TrueMBB, FalseMBB, LandBlk);
1069 ++numIfPatternMatch;
1071 numClonedBlock += Cloned;
1073 return 1 + Cloned + NumMatch;
1076 int AMDGPUCFGStructurizer::loopendPatternMatch() {
1077 std::vector<MachineLoop *> NestedLoops;
1078 for (MachineLoopInfo::iterator It = MLI->begin(), E = MLI->end();
1080 df_iterator<MachineLoop *> LpIt = df_begin(*It),
1082 for (; LpIt != LpE; ++LpIt)
1083 NestedLoops.push_back(*LpIt);
1085 if (NestedLoops.size() == 0)
1088 // Process nested loop outside->inside, so "continue" to a outside loop won't
1089 // be mistaken as "break" of the current loop.
1091 for (std::vector<MachineLoop *>::reverse_iterator It = NestedLoops.rbegin(),
1092 E = NestedLoops.rend(); It != E; ++It) {
1093 MachineLoop *ExaminedLoop = *It;
1094 if (ExaminedLoop->getNumBlocks() == 0 || Visited[ExaminedLoop])
1096 DEBUG(dbgs() << "Processing:\n"; ExaminedLoop->dump(););
1097 int NumBreak = mergeLoop(ExaminedLoop);
1105 int AMDGPUCFGStructurizer::mergeLoop(MachineLoop *LoopRep) {
1106 MachineBasicBlock *LoopHeader = LoopRep->getHeader();
1107 MBBVector ExitingMBBs;
1108 LoopRep->getExitingBlocks(ExitingMBBs);
1109 assert(!ExitingMBBs.empty() && "Infinite Loop not supported");
1110 DEBUG(dbgs() << "Loop has " << ExitingMBBs.size() << " exiting blocks\n";);
1111 // We assume a single ExitBlk
1113 LoopRep->getExitBlocks(ExitBlks);
1114 SmallPtrSet<MachineBasicBlock *, 2> ExitBlkSet;
1115 for (unsigned i = 0, e = ExitBlks.size(); i < e; ++i)
1116 ExitBlkSet.insert(ExitBlks[i]);
1117 assert(ExitBlkSet.size() == 1);
1118 MachineBasicBlock *ExitBlk = *ExitBlks.begin();
1119 assert(ExitBlk && "Loop has several exit block");
1120 MBBVector LatchBlks;
1121 typedef GraphTraits<Inverse<MachineBasicBlock*> > InvMBBTraits;
1122 InvMBBTraits::ChildIteratorType PI = InvMBBTraits::child_begin(LoopHeader),
1123 PE = InvMBBTraits::child_end(LoopHeader);
1124 for (; PI != PE; PI++) {
1125 if (LoopRep->contains(*PI))
1126 LatchBlks.push_back(*PI);
1129 for (unsigned i = 0, e = ExitingMBBs.size(); i < e; ++i)
1130 mergeLoopbreakBlock(ExitingMBBs[i], ExitBlk);
1131 for (unsigned i = 0, e = LatchBlks.size(); i < e; ++i)
1132 settleLoopcontBlock(LatchBlks[i], LoopHeader);
1136 Match += serialPatternMatch(LoopHeader);
1137 Match += ifPatternMatch(LoopHeader);
1138 } while (Match > 0);
1139 mergeLooplandBlock(LoopHeader, ExitBlk);
1140 MachineLoop *ParentLoop = LoopRep->getParentLoop();
1142 MLI->changeLoopFor(LoopHeader, ParentLoop);
1144 MLI->removeBlock(LoopHeader);
1145 Visited[LoopRep] = true;
1149 int AMDGPUCFGStructurizer::loopcontPatternMatch(MachineLoop *LoopRep,
1150 MachineBasicBlock *LoopHeader) {
1152 SmallVector<MachineBasicBlock *, DEFAULT_VEC_SLOTS> ContMBB;
1153 typedef GraphTraits<Inverse<MachineBasicBlock *> > GTIM;
1154 GTIM::ChildIteratorType It = GTIM::child_begin(LoopHeader),
1155 E = GTIM::child_end(LoopHeader);
1156 for (; It != E; ++It) {
1157 MachineBasicBlock *MBB = *It;
1158 if (LoopRep->contains(MBB)) {
1159 handleLoopcontBlock(MBB, MLI->getLoopFor(MBB),
1160 LoopHeader, LoopRep);
1161 ContMBB.push_back(MBB);
1166 for (SmallVectorImpl<MachineBasicBlock *>::iterator It = ContMBB.begin(),
1167 E = ContMBB.end(); It != E; ++It) {
1168 (*It)->removeSuccessor(LoopHeader);
1171 numLoopcontPatternMatch += NumCont;
1177 bool AMDGPUCFGStructurizer::isSameloopDetachedContbreak(
1178 MachineBasicBlock *Src1MBB, MachineBasicBlock *Src2MBB) {
1179 if (Src1MBB->succ_size() == 0) {
1180 MachineLoop *LoopRep = MLI->getLoopFor(Src1MBB);
1181 if (LoopRep&& LoopRep == MLI->getLoopFor(Src2MBB)) {
1182 MachineBasicBlock *&TheEntry = LLInfoMap[LoopRep];
1185 dbgs() << "isLoopContBreakBlock yes src1 = BB"
1186 << Src1MBB->getNumber()
1187 << " src2 = BB" << Src2MBB->getNumber() << "\n";
1196 int AMDGPUCFGStructurizer::handleJumpintoIf(MachineBasicBlock *HeadMBB,
1197 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB) {
1198 int Num = handleJumpintoIfImp(HeadMBB, TrueMBB, FalseMBB);
1201 dbgs() << "handleJumpintoIf swap trueBlk and FalseBlk" << "\n";
1203 Num = handleJumpintoIfImp(HeadMBB, FalseMBB, TrueMBB);
1208 int AMDGPUCFGStructurizer::handleJumpintoIfImp(MachineBasicBlock *HeadMBB,
1209 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB) {
1211 MachineBasicBlock *DownBlk;
1213 //trueBlk could be the common post dominator
1217 dbgs() << "handleJumpintoIfImp head = BB" << HeadMBB->getNumber()
1218 << " true = BB" << TrueMBB->getNumber()
1219 << ", numSucc=" << TrueMBB->succ_size()
1220 << " false = BB" << FalseMBB->getNumber() << "\n";
1225 dbgs() << "check down = BB" << DownBlk->getNumber();
1228 if (singlePathTo(FalseMBB, DownBlk) == SinglePath_InPath) {
1230 dbgs() << " working\n";
1233 Num += cloneOnSideEntryTo(HeadMBB, TrueMBB, DownBlk);
1234 Num += cloneOnSideEntryTo(HeadMBB, FalseMBB, DownBlk);
1236 numClonedBlock += Num;
1237 Num += serialPatternMatch(*HeadMBB->succ_begin());
1238 Num += serialPatternMatch(*std::next(HeadMBB->succ_begin()));
1239 Num += ifPatternMatch(HeadMBB);
1245 dbgs() << " not working\n";
1247 DownBlk = (DownBlk->succ_size() == 1) ? (*DownBlk->succ_begin()) : NULL;
1248 } // walk down the postDomTree
1253 void AMDGPUCFGStructurizer::showImproveSimpleJumpintoIf(
1254 MachineBasicBlock *HeadMBB, MachineBasicBlock *TrueMBB,
1255 MachineBasicBlock *FalseMBB, MachineBasicBlock *LandMBB, bool Detail) {
1256 dbgs() << "head = BB" << HeadMBB->getNumber()
1257 << " size = " << HeadMBB->size();
1260 HeadMBB->print(dbgs());
1265 dbgs() << ", true = BB" << TrueMBB->getNumber() << " size = "
1266 << TrueMBB->size() << " numPred = " << TrueMBB->pred_size();
1269 TrueMBB->print(dbgs());
1274 dbgs() << ", false = BB" << FalseMBB->getNumber() << " size = "
1275 << FalseMBB->size() << " numPred = " << FalseMBB->pred_size();
1278 FalseMBB->print(dbgs());
1283 dbgs() << ", land = BB" << LandMBB->getNumber() << " size = "
1284 << LandMBB->size() << " numPred = " << LandMBB->pred_size();
1287 LandMBB->print(dbgs());
1295 int AMDGPUCFGStructurizer::improveSimpleJumpintoIf(MachineBasicBlock *HeadMBB,
1296 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB,
1297 MachineBasicBlock **LandMBBPtr) {
1298 bool MigrateTrue = false;
1299 bool MigrateFalse = false;
1301 MachineBasicBlock *LandBlk = *LandMBBPtr;
1303 assert((!TrueMBB || TrueMBB->succ_size() <= 1)
1304 && (!FalseMBB || FalseMBB->succ_size() <= 1));
1306 if (TrueMBB == FalseMBB)
1309 MigrateTrue = needMigrateBlock(TrueMBB);
1310 MigrateFalse = needMigrateBlock(FalseMBB);
1312 if (!MigrateTrue && !MigrateFalse)
1315 // If we need to migrate either trueBlk and falseBlk, migrate the rest that
1316 // have more than one predecessors. without doing this, its predecessor
1317 // rather than headBlk will have undefined value in initReg.
1318 if (!MigrateTrue && TrueMBB && TrueMBB->pred_size() > 1)
1320 if (!MigrateFalse && FalseMBB && FalseMBB->pred_size() > 1)
1321 MigrateFalse = true;
1324 dbgs() << "before improveSimpleJumpintoIf: ";
1325 showImproveSimpleJumpintoIf(HeadMBB, TrueMBB, FalseMBB, LandBlk, 0);
1328 // org: headBlk => if () {trueBlk} else {falseBlk} => landBlk
1330 // new: headBlk => if () {initReg = 1; org trueBlk branch} else
1331 // {initReg = 0; org falseBlk branch }
1332 // => landBlk => if (initReg) {org trueBlk} else {org falseBlk}
1334 // if landBlk->pred_size() > 2, put the about if-else inside
1335 // if (initReg !=2) {...}
1337 // add initReg = initVal to headBlk
1339 const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
1340 if (!MigrateTrue || !MigrateFalse) {
1341 // XXX: We have an opportunity here to optimize the "branch into if" case
1342 // here. Branch into if looks like this:
1345 // diamond_head branch_from
1347 // diamond_false diamond_true
1351 // The diamond_head block begins the "if" and the diamond_true block
1352 // is the block being "branched into".
1354 // If MigrateTrue is true, then TrueBB is the block being "branched into"
1355 // and if MigrateFalse is true, then FalseBB is the block being
1358 // Here is the pseudo code for how I think the optimization should work:
1359 // 1. Insert MOV GPR0, 0 before the branch instruction in diamond_head.
1360 // 2. Insert MOV GPR0, 1 before the branch instruction in branch_from.
1361 // 3. Move the branch instruction from diamond_head into its own basic
1362 // block (new_block).
1363 // 4. Add an unconditional branch from diamond_head to new_block
1364 // 5. Replace the branch instruction in branch_from with an unconditional
1365 // branch to new_block. If branch_from has multiple predecessors, then
1366 // we need to replace the True/False block in the branch
1367 // instruction instead of replacing it.
1368 // 6. Change the condition of the branch instruction in new_block from
1369 // COND to (COND || GPR0)
1371 // In order insert these MOV instruction, we will need to use the
1372 // RegisterScavenger. Usually liveness stops being tracked during
1373 // the late machine optimization passes, however if we implement
1374 // bool TargetRegisterInfo::requiresRegisterScavenging(
1375 // const MachineFunction &MF)
1376 // and have it return true, liveness will be tracked correctly
1377 // by generic optimization passes. We will also need to make sure that
1378 // all of our target-specific passes that run after regalloc and before
1379 // the CFGStructurizer track liveness and we will need to modify this pass
1380 // to correctly track liveness.
1382 // After the above changes, the new CFG should look like this:
1385 // diamond_head branch_from
1389 // diamond_false diamond_true
1393 // Without this optimization, we are forced to duplicate the diamond_true
1394 // block and we will end up with a CFG like this:
1398 // diamond_head branch_from
1400 // diamond_false diamond_true diamond_true (duplicate)
1402 // done --------------------|
1404 // Duplicating diamond_true can be very costly especially if it has a
1405 // lot of instructions.
1411 bool LandBlkHasOtherPred = (LandBlk->pred_size() > 2);
1413 //insert AMDGPU::ENDIF to avoid special case "input landBlk == NULL"
1414 MachineBasicBlock::iterator I = insertInstrBefore(LandBlk, AMDGPU::ENDIF);
1416 if (LandBlkHasOtherPred) {
1417 llvm_unreachable("Extra register needed to handle CFG");
1418 unsigned CmpResReg =
1419 HeadMBB->getParent()->getRegInfo().createVirtualRegister(I32RC);
1420 llvm_unreachable("Extra compare instruction needed to handle CFG");
1421 insertCondBranchBefore(LandBlk, I, AMDGPU::IF_PREDICATE_SET,
1422 CmpResReg, DebugLoc());
1425 // XXX: We are running this after RA, so creating virtual registers will
1426 // cause an assertion failure in the PostRA scheduling pass.
1428 HeadMBB->getParent()->getRegInfo().createVirtualRegister(I32RC);
1429 insertCondBranchBefore(LandBlk, I, AMDGPU::IF_PREDICATE_SET, InitReg,
1433 migrateInstruction(TrueMBB, LandBlk, I);
1434 // need to uncondionally insert the assignment to ensure a path from its
1435 // predecessor rather than headBlk has valid value in initReg if
1437 llvm_unreachable("Extra register needed to handle CFG");
1439 insertInstrBefore(I, AMDGPU::ELSE);
1442 migrateInstruction(FalseMBB, LandBlk, I);
1443 // need to uncondionally insert the assignment to ensure a path from its
1444 // predecessor rather than headBlk has valid value in initReg if
1446 llvm_unreachable("Extra register needed to handle CFG");
1449 if (LandBlkHasOtherPred) {
1451 insertInstrBefore(I, AMDGPU::ENDIF);
1453 // put initReg = 2 to other predecessors of landBlk
1454 for (MachineBasicBlock::pred_iterator PI = LandBlk->pred_begin(),
1455 PE = LandBlk->pred_end(); PI != PE; ++PI) {
1456 MachineBasicBlock *MBB = *PI;
1457 if (MBB != TrueMBB && MBB != FalseMBB)
1458 llvm_unreachable("Extra register needed to handle CFG");
1462 dbgs() << "result from improveSimpleJumpintoIf: ";
1463 showImproveSimpleJumpintoIf(HeadMBB, TrueMBB, FalseMBB, LandBlk, 0);
1467 *LandMBBPtr = LandBlk;
1472 void AMDGPUCFGStructurizer::handleLoopcontBlock(MachineBasicBlock *ContingMBB,
1473 MachineLoop *ContingLoop, MachineBasicBlock *ContMBB,
1474 MachineLoop *ContLoop) {
1475 DEBUG(dbgs() << "loopcontPattern cont = BB" << ContingMBB->getNumber()
1476 << " header = BB" << ContMBB->getNumber() << "\n";
1477 dbgs() << "Trying to continue loop-depth = "
1478 << getLoopDepth(ContLoop)
1479 << " from loop-depth = " << getLoopDepth(ContingLoop) << "\n";);
1480 settleLoopcontBlock(ContingMBB, ContMBB);
1483 void AMDGPUCFGStructurizer::mergeSerialBlock(MachineBasicBlock *DstMBB,
1484 MachineBasicBlock *SrcMBB) {
1486 dbgs() << "serialPattern BB" << DstMBB->getNumber()
1487 << " <= BB" << SrcMBB->getNumber() << "\n";
1489 DstMBB->splice(DstMBB->end(), SrcMBB, SrcMBB->begin(), SrcMBB->end());
1491 DstMBB->removeSuccessor(SrcMBB);
1492 cloneSuccessorList(DstMBB, SrcMBB);
1494 removeSuccessor(SrcMBB);
1495 MLI->removeBlock(SrcMBB);
1496 retireBlock(SrcMBB);
1499 void AMDGPUCFGStructurizer::mergeIfthenelseBlock(MachineInstr *BranchMI,
1500 MachineBasicBlock *MBB, MachineBasicBlock *TrueMBB,
1501 MachineBasicBlock *FalseMBB, MachineBasicBlock *LandMBB) {
1504 dbgs() << "ifPattern BB" << MBB->getNumber();
1507 dbgs() << "BB" << TrueMBB->getNumber();
1509 dbgs() << " } else ";
1512 dbgs() << "BB" << FalseMBB->getNumber();
1515 dbgs() << "landBlock: ";
1519 dbgs() << "BB" << LandMBB->getNumber();
1524 int OldOpcode = BranchMI->getOpcode();
1525 DebugLoc BranchDL = BranchMI->getDebugLoc();
1535 MachineBasicBlock::iterator I = BranchMI;
1536 insertCondBranchBefore(I, getBranchNzeroOpcode(OldOpcode),
1540 MBB->splice(I, TrueMBB, TrueMBB->begin(), TrueMBB->end());
1541 MBB->removeSuccessor(TrueMBB);
1542 if (LandMBB && TrueMBB->succ_size()!=0)
1543 TrueMBB->removeSuccessor(LandMBB);
1544 retireBlock(TrueMBB);
1545 MLI->removeBlock(TrueMBB);
1549 insertInstrBefore(I, AMDGPU::ELSE);
1550 MBB->splice(I, FalseMBB, FalseMBB->begin(),
1552 MBB->removeSuccessor(FalseMBB);
1553 if (LandMBB && FalseMBB->succ_size() != 0)
1554 FalseMBB->removeSuccessor(LandMBB);
1555 retireBlock(FalseMBB);
1556 MLI->removeBlock(FalseMBB);
1558 insertInstrBefore(I, AMDGPU::ENDIF);
1560 BranchMI->eraseFromParent();
1562 if (LandMBB && TrueMBB && FalseMBB)
1563 MBB->addSuccessor(LandMBB);
1567 void AMDGPUCFGStructurizer::mergeLooplandBlock(MachineBasicBlock *DstBlk,
1568 MachineBasicBlock *LandMBB) {
1569 DEBUG(dbgs() << "loopPattern header = BB" << DstBlk->getNumber()
1570 << " land = BB" << LandMBB->getNumber() << "\n";);
1572 insertInstrBefore(DstBlk, AMDGPU::WHILELOOP, DebugLoc());
1573 insertInstrEnd(DstBlk, AMDGPU::ENDLOOP, DebugLoc());
1574 DstBlk->addSuccessor(LandMBB);
1575 DstBlk->removeSuccessor(DstBlk);
1579 void AMDGPUCFGStructurizer::mergeLoopbreakBlock(MachineBasicBlock *ExitingMBB,
1580 MachineBasicBlock *LandMBB) {
1581 DEBUG(dbgs() << "loopbreakPattern exiting = BB" << ExitingMBB->getNumber()
1582 << " land = BB" << LandMBB->getNumber() << "\n";);
1583 MachineInstr *BranchMI = getLoopendBlockBranchInstr(ExitingMBB);
1584 assert(BranchMI && isCondBranch(BranchMI));
1585 DebugLoc DL = BranchMI->getDebugLoc();
1586 MachineBasicBlock *TrueBranch = getTrueBranch(BranchMI);
1587 MachineBasicBlock::iterator I = BranchMI;
1588 if (TrueBranch != LandMBB)
1589 reversePredicateSetter(I);
1590 insertCondBranchBefore(ExitingMBB, I, AMDGPU::IF_PREDICATE_SET, AMDGPU::PREDICATE_BIT, DL);
1591 insertInstrBefore(I, AMDGPU::BREAK);
1592 insertInstrBefore(I, AMDGPU::ENDIF);
1593 //now branchInst can be erase safely
1594 BranchMI->eraseFromParent();
1595 //now take care of successors, retire blocks
1596 ExitingMBB->removeSuccessor(LandMBB);
1599 void AMDGPUCFGStructurizer::settleLoopcontBlock(MachineBasicBlock *ContingMBB,
1600 MachineBasicBlock *ContMBB) {
1601 DEBUG(dbgs() << "settleLoopcontBlock conting = BB"
1602 << ContingMBB->getNumber()
1603 << ", cont = BB" << ContMBB->getNumber() << "\n";);
1605 MachineInstr *MI = getLoopendBlockBranchInstr(ContingMBB);
1607 assert(isCondBranch(MI));
1608 MachineBasicBlock::iterator I = MI;
1609 MachineBasicBlock *TrueBranch = getTrueBranch(MI);
1610 int OldOpcode = MI->getOpcode();
1611 DebugLoc DL = MI->getDebugLoc();
1613 bool UseContinueLogical = ((&*ContingMBB->rbegin()) == MI);
1615 if (UseContinueLogical == false) {
1617 TrueBranch == ContMBB ? getBranchNzeroOpcode(OldOpcode) :
1618 getBranchZeroOpcode(OldOpcode);
1619 insertCondBranchBefore(I, BranchOpcode, DL);
1620 // insertEnd to ensure phi-moves, if exist, go before the continue-instr.
1621 insertInstrEnd(ContingMBB, AMDGPU::CONTINUE, DL);
1622 insertInstrEnd(ContingMBB, AMDGPU::ENDIF, DL);
1625 TrueBranch == ContMBB ? getContinueNzeroOpcode(OldOpcode) :
1626 getContinueZeroOpcode(OldOpcode);
1627 insertCondBranchBefore(I, BranchOpcode, DL);
1630 MI->eraseFromParent();
1632 // if we've arrived here then we've already erased the branch instruction
1633 // travel back up the basic block to see the last reference of our debug
1634 // location we've just inserted that reference here so it should be
1635 // representative insertEnd to ensure phi-moves, if exist, go before the
1637 insertInstrEnd(ContingMBB, AMDGPU::CONTINUE,
1638 getLastDebugLocInBB(ContingMBB));
1642 int AMDGPUCFGStructurizer::cloneOnSideEntryTo(MachineBasicBlock *PreMBB,
1643 MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB) {
1645 assert(PreMBB->isSuccessor(SrcMBB));
1646 while (SrcMBB && SrcMBB != DstMBB) {
1647 assert(SrcMBB->succ_size() == 1);
1648 if (SrcMBB->pred_size() > 1) {
1649 SrcMBB = cloneBlockForPredecessor(SrcMBB, PreMBB);
1654 SrcMBB = *SrcMBB->succ_begin();
1661 AMDGPUCFGStructurizer::cloneBlockForPredecessor(MachineBasicBlock *MBB,
1662 MachineBasicBlock *PredMBB) {
1663 assert(PredMBB->isSuccessor(MBB) &&
1664 "succBlk is not a prececessor of curBlk");
1666 MachineBasicBlock *CloneMBB = clone(MBB); //clone instructions
1667 replaceInstrUseOfBlockWith(PredMBB, MBB, CloneMBB);
1668 //srcBlk, oldBlk, newBlk
1670 PredMBB->removeSuccessor(MBB);
1671 PredMBB->addSuccessor(CloneMBB);
1673 // add all successor to cloneBlk
1674 cloneSuccessorList(CloneMBB, MBB);
1676 numClonedInstr += MBB->size();
1679 dbgs() << "Cloned block: " << "BB"
1680 << MBB->getNumber() << "size " << MBB->size() << "\n";
1683 SHOWNEWBLK(CloneMBB, "result of Cloned block: ");
1688 void AMDGPUCFGStructurizer::migrateInstruction(MachineBasicBlock *SrcMBB,
1689 MachineBasicBlock *DstMBB, MachineBasicBlock::iterator I) {
1690 MachineBasicBlock::iterator SpliceEnd;
1691 //look for the input branchinstr, not the AMDGPU branchinstr
1692 MachineInstr *BranchMI = getNormalBlockBranchInstr(SrcMBB);
1695 dbgs() << "migrateInstruction don't see branch instr\n" ;
1697 SpliceEnd = SrcMBB->end();
1700 dbgs() << "migrateInstruction see branch instr\n" ;
1703 SpliceEnd = BranchMI;
1706 dbgs() << "migrateInstruction before splice dstSize = " << DstMBB->size()
1707 << "srcSize = " << SrcMBB->size() << "\n";
1710 //splice insert before insertPos
1711 DstMBB->splice(I, SrcMBB, SrcMBB->begin(), SpliceEnd);
1714 dbgs() << "migrateInstruction after splice dstSize = " << DstMBB->size()
1715 << "srcSize = " << SrcMBB->size() << "\n";
1720 AMDGPUCFGStructurizer::normalizeInfiniteLoopExit(MachineLoop* LoopRep) {
1721 MachineBasicBlock *LoopHeader = LoopRep->getHeader();
1722 MachineBasicBlock *LoopLatch = LoopRep->getLoopLatch();
1723 const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
1725 if (!LoopHeader || !LoopLatch)
1727 MachineInstr *BranchMI = getLoopendBlockBranchInstr(LoopLatch);
1728 // Is LoopRep an infinite loop ?
1729 if (!BranchMI || !isUncondBranch(BranchMI))
1732 MachineBasicBlock *DummyExitBlk = FuncRep->CreateMachineBasicBlock();
1733 FuncRep->push_back(DummyExitBlk); //insert to function
1734 SHOWNEWBLK(DummyExitBlk, "DummyExitBlock to normalize infiniteLoop: ");
1735 DEBUG(dbgs() << "Old branch instr: " << *BranchMI << "\n";);
1736 MachineBasicBlock::iterator I = BranchMI;
1737 unsigned ImmReg = FuncRep->getRegInfo().createVirtualRegister(I32RC);
1738 llvm_unreachable("Extra register needed to handle CFG");
1739 MachineInstr *NewMI = insertInstrBefore(I, AMDGPU::BRANCH_COND_i32);
1740 MachineInstrBuilder MIB(*FuncRep, NewMI);
1741 MIB.addMBB(LoopHeader);
1742 MIB.addReg(ImmReg, false);
1743 SHOWNEWINSTR(NewMI);
1744 BranchMI->eraseFromParent();
1745 LoopLatch->addSuccessor(DummyExitBlk);
1747 return DummyExitBlk;
1750 void AMDGPUCFGStructurizer::removeUnconditionalBranch(MachineBasicBlock *MBB) {
1751 MachineInstr *BranchMI;
1753 // I saw two unconditional branch in one basic block in example
1754 // test_fc_do_while_or.c need to fix the upstream on this to remove the loop.
1755 while ((BranchMI = getLoopendBlockBranchInstr(MBB))
1756 && isUncondBranch(BranchMI)) {
1757 DEBUG(dbgs() << "Removing uncond branch instr"; BranchMI->dump(););
1758 BranchMI->eraseFromParent();
1762 void AMDGPUCFGStructurizer::removeRedundantConditionalBranch(
1763 MachineBasicBlock *MBB) {
1764 if (MBB->succ_size() != 2)
1766 MachineBasicBlock *MBB1 = *MBB->succ_begin();
1767 MachineBasicBlock *MBB2 = *std::next(MBB->succ_begin());
1771 MachineInstr *BranchMI = getNormalBlockBranchInstr(MBB);
1772 assert(BranchMI && isCondBranch(BranchMI));
1773 DEBUG(dbgs() << "Removing unneeded cond branch instr"; BranchMI->dump(););
1774 BranchMI->eraseFromParent();
1775 SHOWNEWBLK(MBB1, "Removing redundant successor");
1776 MBB->removeSuccessor(MBB1);
1779 void AMDGPUCFGStructurizer::addDummyExitBlock(
1780 SmallVectorImpl<MachineBasicBlock*> &RetMBB) {
1781 MachineBasicBlock *DummyExitBlk = FuncRep->CreateMachineBasicBlock();
1782 FuncRep->push_back(DummyExitBlk); //insert to function
1783 insertInstrEnd(DummyExitBlk, AMDGPU::RETURN);
1785 for (SmallVectorImpl<MachineBasicBlock *>::iterator It = RetMBB.begin(),
1786 E = RetMBB.end(); It != E; ++It) {
1787 MachineBasicBlock *MBB = *It;
1788 MachineInstr *MI = getReturnInstr(MBB);
1790 MI->eraseFromParent();
1791 MBB->addSuccessor(DummyExitBlk);
1793 dbgs() << "Add dummyExitBlock to BB" << MBB->getNumber()
1797 SHOWNEWBLK(DummyExitBlk, "DummyExitBlock: ");
1800 void AMDGPUCFGStructurizer::removeSuccessor(MachineBasicBlock *MBB) {
1801 while (MBB->succ_size())
1802 MBB->removeSuccessor(*MBB->succ_begin());
1805 void AMDGPUCFGStructurizer::recordSccnum(MachineBasicBlock *MBB,
1807 BlockInformation *&srcBlkInfo = BlockInfoMap[MBB];
1809 srcBlkInfo = new BlockInformation();
1810 srcBlkInfo->SccNum = SccNum;
1813 void AMDGPUCFGStructurizer::retireBlock(MachineBasicBlock *MBB) {
1815 dbgs() << "Retiring BB" << MBB->getNumber() << "\n";
1818 BlockInformation *&SrcBlkInfo = BlockInfoMap[MBB];
1821 SrcBlkInfo = new BlockInformation();
1823 SrcBlkInfo->IsRetired = true;
1824 assert(MBB->succ_size() == 0 && MBB->pred_size() == 0
1825 && "can't retire block yet");
1828 void AMDGPUCFGStructurizer::setLoopLandBlock(MachineLoop *loopRep,
1829 MachineBasicBlock *MBB) {
1830 MachineBasicBlock *&TheEntry = LLInfoMap[loopRep];
1832 MBB = FuncRep->CreateMachineBasicBlock();
1833 FuncRep->push_back(MBB); //insert to function
1834 SHOWNEWBLK(MBB, "DummyLandingBlock for loop without break: ");
1838 dbgs() << "setLoopLandBlock loop-header = BB"
1839 << loopRep->getHeader()->getNumber()
1840 << " landing-block = BB" << MBB->getNumber() << "\n";
1845 AMDGPUCFGStructurizer::findNearestCommonPostDom(MachineBasicBlock *MBB1,
1846 MachineBasicBlock *MBB2) {
1848 if (PDT->dominates(MBB1, MBB2))
1850 if (PDT->dominates(MBB2, MBB1))
1853 MachineDomTreeNode *Node1 = PDT->getNode(MBB1);
1854 MachineDomTreeNode *Node2 = PDT->getNode(MBB2);
1856 // Handle newly cloned node.
1857 if (!Node1 && MBB1->succ_size() == 1)
1858 return findNearestCommonPostDom(*MBB1->succ_begin(), MBB2);
1859 if (!Node2 && MBB2->succ_size() == 1)
1860 return findNearestCommonPostDom(MBB1, *MBB2->succ_begin());
1862 if (!Node1 || !Node2)
1865 Node1 = Node1->getIDom();
1867 if (PDT->dominates(Node1, Node2))
1868 return Node1->getBlock();
1869 Node1 = Node1->getIDom();
1876 AMDGPUCFGStructurizer::findNearestCommonPostDom(
1877 std::set<MachineBasicBlock *> &MBBs) {
1878 MachineBasicBlock *CommonDom;
1879 std::set<MachineBasicBlock *>::const_iterator It = MBBs.begin();
1880 std::set<MachineBasicBlock *>::const_iterator E = MBBs.end();
1881 for (CommonDom = *It; It != E && CommonDom; ++It) {
1882 MachineBasicBlock *MBB = *It;
1883 if (MBB != CommonDom)
1884 CommonDom = findNearestCommonPostDom(MBB, CommonDom);
1888 dbgs() << "Common post dominator for exit blocks is ";
1890 dbgs() << "BB" << CommonDom->getNumber() << "\n";
1898 char AMDGPUCFGStructurizer::ID = 0;
1900 } // end anonymous namespace
1903 INITIALIZE_PASS_BEGIN(AMDGPUCFGStructurizer, "amdgpustructurizer",
1904 "AMDGPU CFG Structurizer", false, false)
1905 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
1906 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
1907 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
1908 INITIALIZE_PASS_END(AMDGPUCFGStructurizer, "amdgpustructurizer",
1909 "AMDGPU CFG Structurizer", false, false)
1911 FunctionPass *llvm::createAMDGPUCFGStructurizerPass() {
1912 return new AMDGPUCFGStructurizer();