[InstCombine] Signed saturation patterns
[llvm-complete.git] / lib / Target / AMDGPU / AMDILCFGStructurizer.cpp
blob101ecfc0c87c817128859bf48a97a670bb3170bb
1 //===- AMDILCFGStructurizer.cpp - CFG Structurizer ------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //==-----------------------------------------------------------------------===//
9 #include "AMDGPU.h"
10 #include "AMDGPUSubtarget.h"
11 #include "R600InstrInfo.h"
12 #include "R600RegisterInfo.h"
13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
14 #include "llvm/ADT/DepthFirstIterator.h"
15 #include "llvm/ADT/SCCIterator.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/CodeGen/MachineBasicBlock.h"
21 #include "llvm/CodeGen/MachineDominators.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstr.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineJumpTableInfo.h"
27 #include "llvm/CodeGen/MachineLoopInfo.h"
28 #include "llvm/CodeGen/MachineOperand.h"
29 #include "llvm/CodeGen/MachinePostDominators.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/IR/DebugLoc.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MachineValueType.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include <cassert>
39 #include <cstddef>
40 #include <deque>
41 #include <iterator>
42 #include <map>
43 #include <utility>
44 #include <vector>
46 using namespace llvm;
48 #define DEBUG_TYPE "structcfg"
50 #define DEFAULT_VEC_SLOTS 8
52 // TODO: move-begin.
54 //===----------------------------------------------------------------------===//
56 // Statistics for CFGStructurizer.
58 //===----------------------------------------------------------------------===//
60 STATISTIC(numSerialPatternMatch, "CFGStructurizer number of serial pattern "
61 "matched");
62 STATISTIC(numIfPatternMatch, "CFGStructurizer number of if pattern "
63 "matched");
64 STATISTIC(numClonedBlock, "CFGStructurizer cloned blocks");
65 STATISTIC(numClonedInstr, "CFGStructurizer cloned instructions");
67 namespace llvm {
69 void initializeAMDGPUCFGStructurizerPass(PassRegistry &);
71 } // end namespace llvm
73 namespace {
75 //===----------------------------------------------------------------------===//
77 // Miscellaneous utility for CFGStructurizer.
79 //===----------------------------------------------------------------------===//
81 #define SHOWNEWINSTR(i) LLVM_DEBUG(dbgs() << "New instr: " << *i << "\n");
83 #define SHOWNEWBLK(b, msg) \
84 LLVM_DEBUG(dbgs() << msg << "BB" << b->getNumber() << "size " << b->size(); \
85 dbgs() << "\n";);
87 #define SHOWBLK_DETAIL(b, msg) \
88 LLVM_DEBUG(if (b) { \
89 dbgs() << msg << "BB" << b->getNumber() << "size " << b->size(); \
90 b->print(dbgs()); \
91 dbgs() << "\n"; \
92 });
94 #define INVALIDSCCNUM -1
96 //===----------------------------------------------------------------------===//
98 // supporting data structure for CFGStructurizer
100 //===----------------------------------------------------------------------===//
102 class BlockInformation {
103 public:
104 bool IsRetired = false;
105 int SccNum = INVALIDSCCNUM;
107 BlockInformation() = default;
110 //===----------------------------------------------------------------------===//
112 // CFGStructurizer
114 //===----------------------------------------------------------------------===//
116 class AMDGPUCFGStructurizer : public MachineFunctionPass {
117 public:
118 using MBBVector = SmallVector<MachineBasicBlock *, 32>;
119 using MBBInfoMap = std::map<MachineBasicBlock *, BlockInformation *>;
120 using LoopLandInfoMap = std::map<MachineLoop *, MachineBasicBlock *>;
122 enum PathToKind {
123 Not_SinglePath = 0,
124 SinglePath_InPath = 1,
125 SinglePath_NotInPath = 2
128 static char ID;
130 AMDGPUCFGStructurizer() : MachineFunctionPass(ID) {
131 initializeAMDGPUCFGStructurizerPass(*PassRegistry::getPassRegistry());
134 StringRef getPassName() const override {
135 return "AMDGPU Control Flow Graph structurizer Pass";
138 void getAnalysisUsage(AnalysisUsage &AU) const override {
139 AU.addRequired<MachineDominatorTree>();
140 AU.addRequired<MachinePostDominatorTree>();
141 AU.addRequired<MachineLoopInfo>();
142 MachineFunctionPass::getAnalysisUsage(AU);
145 /// Perform the CFG structurization
146 bool run();
148 /// Perform the CFG preparation
149 /// This step will remove every unconditionnal/dead jump instructions and make
150 /// sure all loops have an exit block
151 bool prepare();
153 bool runOnMachineFunction(MachineFunction &MF) override {
154 TII = MF.getSubtarget<R600Subtarget>().getInstrInfo();
155 TRI = &TII->getRegisterInfo();
156 LLVM_DEBUG(MF.dump(););
157 OrderedBlks.clear();
158 Visited.clear();
159 FuncRep = &MF;
160 MLI = &getAnalysis<MachineLoopInfo>();
161 LLVM_DEBUG(dbgs() << "LoopInfo:\n"; PrintLoopinfo(*MLI););
162 MDT = &getAnalysis<MachineDominatorTree>();
163 LLVM_DEBUG(MDT->print(dbgs(), (const Module *)nullptr););
164 PDT = &getAnalysis<MachinePostDominatorTree>();
165 LLVM_DEBUG(PDT->print(dbgs()););
166 prepare();
167 run();
168 LLVM_DEBUG(MF.dump(););
169 return true;
172 protected:
173 MachineDominatorTree *MDT;
174 MachinePostDominatorTree *PDT;
175 MachineLoopInfo *MLI;
176 const R600InstrInfo *TII = nullptr;
177 const R600RegisterInfo *TRI = nullptr;
179 // PRINT FUNCTIONS
180 /// Print the ordered Blocks.
181 void printOrderedBlocks() const {
182 size_t i = 0;
183 for (MBBVector::const_iterator iterBlk = OrderedBlks.begin(),
184 iterBlkEnd = OrderedBlks.end(); iterBlk != iterBlkEnd; ++iterBlk, ++i) {
185 dbgs() << "BB" << (*iterBlk)->getNumber();
186 dbgs() << "(" << getSCCNum(*iterBlk) << "," << (*iterBlk)->size() << ")";
187 if (i != 0 && i % 10 == 0) {
188 dbgs() << "\n";
189 } else {
190 dbgs() << " ";
195 static void PrintLoopinfo(const MachineLoopInfo &LoopInfo) {
196 for (MachineLoop::iterator iter = LoopInfo.begin(),
197 iterEnd = LoopInfo.end(); iter != iterEnd; ++iter) {
198 (*iter)->print(dbgs(), 0);
202 // UTILITY FUNCTIONS
203 int getSCCNum(MachineBasicBlock *MBB) const;
204 MachineBasicBlock *getLoopLandInfo(MachineLoop *LoopRep) const;
205 bool hasBackEdge(MachineBasicBlock *MBB) const;
206 bool isRetiredBlock(MachineBasicBlock *MBB) const;
207 bool isActiveLoophead(MachineBasicBlock *MBB) const;
208 PathToKind singlePathTo(MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB,
209 bool AllowSideEntry = true) const;
210 int countActiveBlock(MBBVector::const_iterator It,
211 MBBVector::const_iterator E) const;
212 bool needMigrateBlock(MachineBasicBlock *MBB) const;
214 // Utility Functions
215 void reversePredicateSetter(MachineBasicBlock::iterator I,
216 MachineBasicBlock &MBB);
217 /// Compute the reversed DFS post order of Blocks
218 void orderBlocks(MachineFunction *MF);
220 // Function originally from CFGStructTraits
221 void insertInstrEnd(MachineBasicBlock *MBB, int NewOpcode,
222 const DebugLoc &DL = DebugLoc());
223 MachineInstr *insertInstrBefore(MachineBasicBlock *MBB, int NewOpcode,
224 const DebugLoc &DL = DebugLoc());
225 MachineInstr *insertInstrBefore(MachineBasicBlock::iterator I, int NewOpcode);
226 void insertCondBranchBefore(MachineBasicBlock::iterator I, int NewOpcode,
227 const DebugLoc &DL);
228 void insertCondBranchBefore(MachineBasicBlock *MBB,
229 MachineBasicBlock::iterator I, int NewOpcode,
230 int RegNum, const DebugLoc &DL);
232 static int getBranchNzeroOpcode(int OldOpcode);
233 static int getBranchZeroOpcode(int OldOpcode);
234 static int getContinueNzeroOpcode(int OldOpcode);
235 static int getContinueZeroOpcode(int OldOpcode);
236 static MachineBasicBlock *getTrueBranch(MachineInstr *MI);
237 static void setTrueBranch(MachineInstr *MI, MachineBasicBlock *MBB);
238 static MachineBasicBlock *getFalseBranch(MachineBasicBlock *MBB,
239 MachineInstr *MI);
240 static bool isCondBranch(MachineInstr *MI);
241 static bool isUncondBranch(MachineInstr *MI);
242 static DebugLoc getLastDebugLocInBB(MachineBasicBlock *MBB);
243 static MachineInstr *getNormalBlockBranchInstr(MachineBasicBlock *MBB);
245 /// The correct naming for this is getPossibleLoopendBlockBranchInstr.
247 /// BB with backward-edge could have move instructions after the branch
248 /// instruction. Such move instruction "belong to" the loop backward-edge.
249 MachineInstr *getLoopendBlockBranchInstr(MachineBasicBlock *MBB);
251 static MachineInstr *getReturnInstr(MachineBasicBlock *MBB);
252 static bool isReturnBlock(MachineBasicBlock *MBB);
253 static void cloneSuccessorList(MachineBasicBlock *DstMBB,
254 MachineBasicBlock *SrcMBB);
255 static MachineBasicBlock *clone(MachineBasicBlock *MBB);
257 /// MachineBasicBlock::ReplaceUsesOfBlockWith doesn't serve the purpose
258 /// because the AMDGPU instruction is not recognized as terminator fix this
259 /// and retire this routine
260 void replaceInstrUseOfBlockWith(MachineBasicBlock *SrcMBB,
261 MachineBasicBlock *OldMBB, MachineBasicBlock *NewBlk);
263 static void wrapup(MachineBasicBlock *MBB);
265 int patternMatch(MachineBasicBlock *MBB);
266 int patternMatchGroup(MachineBasicBlock *MBB);
267 int serialPatternMatch(MachineBasicBlock *MBB);
268 int ifPatternMatch(MachineBasicBlock *MBB);
269 int loopendPatternMatch();
270 int mergeLoop(MachineLoop *LoopRep);
272 /// return true iff src1Blk->succ_size() == 0 && src1Blk and src2Blk are in
273 /// the same loop with LoopLandInfo without explicitly keeping track of
274 /// loopContBlks and loopBreakBlks, this is a method to get the information.
275 bool isSameloopDetachedContbreak(MachineBasicBlock *Src1MBB,
276 MachineBasicBlock *Src2MBB);
277 int handleJumpintoIf(MachineBasicBlock *HeadMBB,
278 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB);
279 int handleJumpintoIfImp(MachineBasicBlock *HeadMBB,
280 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB);
281 int improveSimpleJumpintoIf(MachineBasicBlock *HeadMBB,
282 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB,
283 MachineBasicBlock **LandMBBPtr);
284 void showImproveSimpleJumpintoIf(MachineBasicBlock *HeadMBB,
285 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB,
286 MachineBasicBlock *LandMBB, bool Detail = false);
287 int cloneOnSideEntryTo(MachineBasicBlock *PreMBB,
288 MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB);
289 void mergeSerialBlock(MachineBasicBlock *DstMBB,
290 MachineBasicBlock *SrcMBB);
292 void mergeIfthenelseBlock(MachineInstr *BranchMI,
293 MachineBasicBlock *MBB, MachineBasicBlock *TrueMBB,
294 MachineBasicBlock *FalseMBB, MachineBasicBlock *LandMBB);
295 void mergeLooplandBlock(MachineBasicBlock *DstMBB,
296 MachineBasicBlock *LandMBB);
297 void mergeLoopbreakBlock(MachineBasicBlock *ExitingMBB,
298 MachineBasicBlock *LandMBB);
299 void settleLoopcontBlock(MachineBasicBlock *ContingMBB,
300 MachineBasicBlock *ContMBB);
302 /// normalizeInfiniteLoopExit change
303 /// B1:
304 /// uncond_br LoopHeader
306 /// to
307 /// B1:
308 /// cond_br 1 LoopHeader dummyExit
309 /// and return the newly added dummy exit block
310 MachineBasicBlock *normalizeInfiniteLoopExit(MachineLoop *LoopRep);
311 void removeUnconditionalBranch(MachineBasicBlock *MBB);
313 /// Remove duplicate branches instructions in a block.
314 /// For instance
315 /// B0:
316 /// cond_br X B1 B2
317 /// cond_br X B1 B2
318 /// is transformed to
319 /// B0:
320 /// cond_br X B1 B2
321 void removeRedundantConditionalBranch(MachineBasicBlock *MBB);
323 void addDummyExitBlock(SmallVectorImpl<MachineBasicBlock *> &RetMBB);
324 void removeSuccessor(MachineBasicBlock *MBB);
325 MachineBasicBlock *cloneBlockForPredecessor(MachineBasicBlock *MBB,
326 MachineBasicBlock *PredMBB);
327 void migrateInstruction(MachineBasicBlock *SrcMBB,
328 MachineBasicBlock *DstMBB, MachineBasicBlock::iterator I);
329 void recordSccnum(MachineBasicBlock *MBB, int SCCNum);
330 void retireBlock(MachineBasicBlock *MBB);
332 private:
333 MBBInfoMap BlockInfoMap;
334 LoopLandInfoMap LLInfoMap;
335 std::map<MachineLoop *, bool> Visited;
336 MachineFunction *FuncRep;
337 SmallVector<MachineBasicBlock *, DEFAULT_VEC_SLOTS> OrderedBlks;
340 } // end anonymous namespace
342 char AMDGPUCFGStructurizer::ID = 0;
344 int AMDGPUCFGStructurizer::getSCCNum(MachineBasicBlock *MBB) const {
345 MBBInfoMap::const_iterator It = BlockInfoMap.find(MBB);
346 if (It == BlockInfoMap.end())
347 return INVALIDSCCNUM;
348 return (*It).second->SccNum;
351 MachineBasicBlock *AMDGPUCFGStructurizer::getLoopLandInfo(MachineLoop *LoopRep)
352 const {
353 LoopLandInfoMap::const_iterator It = LLInfoMap.find(LoopRep);
354 if (It == LLInfoMap.end())
355 return nullptr;
356 return (*It).second;
359 bool AMDGPUCFGStructurizer::hasBackEdge(MachineBasicBlock *MBB) const {
360 MachineLoop *LoopRep = MLI->getLoopFor(MBB);
361 if (!LoopRep)
362 return false;
363 MachineBasicBlock *LoopHeader = LoopRep->getHeader();
364 return MBB->isSuccessor(LoopHeader);
367 bool AMDGPUCFGStructurizer::isRetiredBlock(MachineBasicBlock *MBB) const {
368 MBBInfoMap::const_iterator It = BlockInfoMap.find(MBB);
369 if (It == BlockInfoMap.end())
370 return false;
371 return (*It).second->IsRetired;
374 bool AMDGPUCFGStructurizer::isActiveLoophead(MachineBasicBlock *MBB) const {
375 MachineLoop *LoopRep = MLI->getLoopFor(MBB);
376 while (LoopRep && LoopRep->getHeader() == MBB) {
377 MachineBasicBlock *LoopLand = getLoopLandInfo(LoopRep);
378 if(!LoopLand)
379 return true;
380 if (!isRetiredBlock(LoopLand))
381 return true;
382 LoopRep = LoopRep->getParentLoop();
384 return false;
387 AMDGPUCFGStructurizer::PathToKind AMDGPUCFGStructurizer::singlePathTo(
388 MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB,
389 bool AllowSideEntry) const {
390 assert(DstMBB);
391 if (SrcMBB == DstMBB)
392 return SinglePath_InPath;
393 while (SrcMBB && SrcMBB->succ_size() == 1) {
394 SrcMBB = *SrcMBB->succ_begin();
395 if (SrcMBB == DstMBB)
396 return SinglePath_InPath;
397 if (!AllowSideEntry && SrcMBB->pred_size() > 1)
398 return Not_SinglePath;
400 if (SrcMBB && SrcMBB->succ_size()==0)
401 return SinglePath_NotInPath;
402 return Not_SinglePath;
405 int AMDGPUCFGStructurizer::countActiveBlock(MBBVector::const_iterator It,
406 MBBVector::const_iterator E) const {
407 int Count = 0;
408 while (It != E) {
409 if (!isRetiredBlock(*It))
410 ++Count;
411 ++It;
413 return Count;
416 bool AMDGPUCFGStructurizer::needMigrateBlock(MachineBasicBlock *MBB) const {
417 unsigned BlockSizeThreshold = 30;
418 unsigned CloneInstrThreshold = 100;
419 bool MultiplePreds = MBB && (MBB->pred_size() > 1);
421 if(!MultiplePreds)
422 return false;
423 unsigned BlkSize = MBB->size();
424 return ((BlkSize > BlockSizeThreshold) &&
425 (BlkSize * (MBB->pred_size() - 1) > CloneInstrThreshold));
428 void AMDGPUCFGStructurizer::reversePredicateSetter(
429 MachineBasicBlock::iterator I, MachineBasicBlock &MBB) {
430 assert(I.isValid() && "Expected valid iterator");
431 for (;; --I) {
432 if (I == MBB.end())
433 continue;
434 if (I->getOpcode() == R600::PRED_X) {
435 switch (I->getOperand(2).getImm()) {
436 case R600::PRED_SETE_INT:
437 I->getOperand(2).setImm(R600::PRED_SETNE_INT);
438 return;
439 case R600::PRED_SETNE_INT:
440 I->getOperand(2).setImm(R600::PRED_SETE_INT);
441 return;
442 case R600::PRED_SETE:
443 I->getOperand(2).setImm(R600::PRED_SETNE);
444 return;
445 case R600::PRED_SETNE:
446 I->getOperand(2).setImm(R600::PRED_SETE);
447 return;
448 default:
449 llvm_unreachable("PRED_X Opcode invalid!");
455 void AMDGPUCFGStructurizer::insertInstrEnd(MachineBasicBlock *MBB,
456 int NewOpcode, const DebugLoc &DL) {
457 MachineInstr *MI =
458 MBB->getParent()->CreateMachineInstr(TII->get(NewOpcode), DL);
459 MBB->push_back(MI);
460 //assume the instruction doesn't take any reg operand ...
461 SHOWNEWINSTR(MI);
464 MachineInstr *AMDGPUCFGStructurizer::insertInstrBefore(MachineBasicBlock *MBB,
465 int NewOpcode,
466 const DebugLoc &DL) {
467 MachineInstr *MI =
468 MBB->getParent()->CreateMachineInstr(TII->get(NewOpcode), DL);
469 if (MBB->begin() != MBB->end())
470 MBB->insert(MBB->begin(), MI);
471 else
472 MBB->push_back(MI);
473 SHOWNEWINSTR(MI);
474 return MI;
477 MachineInstr *AMDGPUCFGStructurizer::insertInstrBefore(
478 MachineBasicBlock::iterator I, int NewOpcode) {
479 MachineInstr *OldMI = &(*I);
480 MachineBasicBlock *MBB = OldMI->getParent();
481 MachineInstr *NewMBB =
482 MBB->getParent()->CreateMachineInstr(TII->get(NewOpcode), DebugLoc());
483 MBB->insert(I, NewMBB);
484 //assume the instruction doesn't take any reg operand ...
485 SHOWNEWINSTR(NewMBB);
486 return NewMBB;
489 void AMDGPUCFGStructurizer::insertCondBranchBefore(
490 MachineBasicBlock::iterator I, int NewOpcode, const DebugLoc &DL) {
491 MachineInstr *OldMI = &(*I);
492 MachineBasicBlock *MBB = OldMI->getParent();
493 MachineFunction *MF = MBB->getParent();
494 MachineInstr *NewMI = MF->CreateMachineInstr(TII->get(NewOpcode), DL);
495 MBB->insert(I, NewMI);
496 MachineInstrBuilder MIB(*MF, NewMI);
497 MIB.addReg(OldMI->getOperand(1).getReg(), false);
498 SHOWNEWINSTR(NewMI);
499 //erase later oldInstr->eraseFromParent();
502 void AMDGPUCFGStructurizer::insertCondBranchBefore(
503 MachineBasicBlock *blk, MachineBasicBlock::iterator I, int NewOpcode,
504 int RegNum, const DebugLoc &DL) {
505 MachineFunction *MF = blk->getParent();
506 MachineInstr *NewInstr = MF->CreateMachineInstr(TII->get(NewOpcode), DL);
507 //insert before
508 blk->insert(I, NewInstr);
509 MachineInstrBuilder(*MF, NewInstr).addReg(RegNum, false);
510 SHOWNEWINSTR(NewInstr);
513 int AMDGPUCFGStructurizer::getBranchNzeroOpcode(int OldOpcode) {
514 switch(OldOpcode) {
515 case R600::JUMP_COND:
516 case R600::JUMP: return R600::IF_PREDICATE_SET;
517 case R600::BRANCH_COND_i32:
518 case R600::BRANCH_COND_f32: return R600::IF_LOGICALNZ_f32;
519 default: llvm_unreachable("internal error");
521 return -1;
524 int AMDGPUCFGStructurizer::getBranchZeroOpcode(int OldOpcode) {
525 switch(OldOpcode) {
526 case R600::JUMP_COND:
527 case R600::JUMP: return R600::IF_PREDICATE_SET;
528 case R600::BRANCH_COND_i32:
529 case R600::BRANCH_COND_f32: return R600::IF_LOGICALZ_f32;
530 default: llvm_unreachable("internal error");
532 return -1;
535 int AMDGPUCFGStructurizer::getContinueNzeroOpcode(int OldOpcode) {
536 switch(OldOpcode) {
537 case R600::JUMP_COND:
538 case R600::JUMP: return R600::CONTINUE_LOGICALNZ_i32;
539 default: llvm_unreachable("internal error");
541 return -1;
544 int AMDGPUCFGStructurizer::getContinueZeroOpcode(int OldOpcode) {
545 switch(OldOpcode) {
546 case R600::JUMP_COND:
547 case R600::JUMP: return R600::CONTINUE_LOGICALZ_i32;
548 default: llvm_unreachable("internal error");
550 return -1;
553 MachineBasicBlock *AMDGPUCFGStructurizer::getTrueBranch(MachineInstr *MI) {
554 return MI->getOperand(0).getMBB();
557 void AMDGPUCFGStructurizer::setTrueBranch(MachineInstr *MI,
558 MachineBasicBlock *MBB) {
559 MI->getOperand(0).setMBB(MBB);
562 MachineBasicBlock *
563 AMDGPUCFGStructurizer::getFalseBranch(MachineBasicBlock *MBB,
564 MachineInstr *MI) {
565 assert(MBB->succ_size() == 2);
566 MachineBasicBlock *TrueBranch = getTrueBranch(MI);
567 MachineBasicBlock::succ_iterator It = MBB->succ_begin();
568 MachineBasicBlock::succ_iterator Next = It;
569 ++Next;
570 return (*It == TrueBranch) ? *Next : *It;
573 bool AMDGPUCFGStructurizer::isCondBranch(MachineInstr *MI) {
574 switch (MI->getOpcode()) {
575 case R600::JUMP_COND:
576 case R600::BRANCH_COND_i32:
577 case R600::BRANCH_COND_f32: return true;
578 default:
579 return false;
581 return false;
584 bool AMDGPUCFGStructurizer::isUncondBranch(MachineInstr *MI) {
585 switch (MI->getOpcode()) {
586 case R600::JUMP:
587 case R600::BRANCH:
588 return true;
589 default:
590 return false;
592 return false;
595 DebugLoc AMDGPUCFGStructurizer::getLastDebugLocInBB(MachineBasicBlock *MBB) {
596 //get DebugLoc from the first MachineBasicBlock instruction with debug info
597 DebugLoc DL;
598 for (MachineBasicBlock::iterator It = MBB->begin(); It != MBB->end();
599 ++It) {
600 MachineInstr *instr = &(*It);
601 if (instr->getDebugLoc())
602 DL = instr->getDebugLoc();
604 return DL;
607 MachineInstr *AMDGPUCFGStructurizer::getNormalBlockBranchInstr(
608 MachineBasicBlock *MBB) {
609 MachineBasicBlock::reverse_iterator It = MBB->rbegin();
610 MachineInstr *MI = &*It;
611 if (MI && (isCondBranch(MI) || isUncondBranch(MI)))
612 return MI;
613 return nullptr;
616 MachineInstr *AMDGPUCFGStructurizer::getLoopendBlockBranchInstr(
617 MachineBasicBlock *MBB) {
618 for (MachineBasicBlock::reverse_iterator It = MBB->rbegin(), E = MBB->rend();
619 It != E; ++It) {
620 // FIXME: Simplify
621 MachineInstr *MI = &*It;
622 if (MI) {
623 if (isCondBranch(MI) || isUncondBranch(MI))
624 return MI;
625 else if (!TII->isMov(MI->getOpcode()))
626 break;
629 return nullptr;
632 MachineInstr *AMDGPUCFGStructurizer::getReturnInstr(MachineBasicBlock *MBB) {
633 MachineBasicBlock::reverse_iterator It = MBB->rbegin();
634 if (It != MBB->rend()) {
635 MachineInstr *instr = &(*It);
636 if (instr->getOpcode() == R600::RETURN)
637 return instr;
639 return nullptr;
642 bool AMDGPUCFGStructurizer::isReturnBlock(MachineBasicBlock *MBB) {
643 MachineInstr *MI = getReturnInstr(MBB);
644 bool IsReturn = (MBB->succ_size() == 0);
645 if (MI)
646 assert(IsReturn);
647 else if (IsReturn)
648 LLVM_DEBUG(dbgs() << "BB" << MBB->getNumber()
649 << " is return block without RETURN instr\n";);
650 return IsReturn;
653 void AMDGPUCFGStructurizer::cloneSuccessorList(MachineBasicBlock *DstMBB,
654 MachineBasicBlock *SrcMBB) {
655 for (MachineBasicBlock::succ_iterator It = SrcMBB->succ_begin(),
656 iterEnd = SrcMBB->succ_end(); It != iterEnd; ++It)
657 DstMBB->addSuccessor(*It); // *iter's predecessor is also taken care of
660 MachineBasicBlock *AMDGPUCFGStructurizer::clone(MachineBasicBlock *MBB) {
661 MachineFunction *Func = MBB->getParent();
662 MachineBasicBlock *NewMBB = Func->CreateMachineBasicBlock();
663 Func->push_back(NewMBB); //insert to function
664 for (const MachineInstr &It : *MBB)
665 NewMBB->push_back(Func->CloneMachineInstr(&It));
666 return NewMBB;
669 void AMDGPUCFGStructurizer::replaceInstrUseOfBlockWith(
670 MachineBasicBlock *SrcMBB, MachineBasicBlock *OldMBB,
671 MachineBasicBlock *NewBlk) {
672 MachineInstr *BranchMI = getLoopendBlockBranchInstr(SrcMBB);
673 if (BranchMI && isCondBranch(BranchMI) &&
674 getTrueBranch(BranchMI) == OldMBB)
675 setTrueBranch(BranchMI, NewBlk);
678 void AMDGPUCFGStructurizer::wrapup(MachineBasicBlock *MBB) {
679 assert((!MBB->getParent()->getJumpTableInfo()
680 || MBB->getParent()->getJumpTableInfo()->isEmpty())
681 && "found a jump table");
683 //collect continue right before endloop
684 SmallVector<MachineInstr *, DEFAULT_VEC_SLOTS> ContInstr;
685 MachineBasicBlock::iterator Pre = MBB->begin();
686 MachineBasicBlock::iterator E = MBB->end();
687 MachineBasicBlock::iterator It = Pre;
688 while (It != E) {
689 if (Pre->getOpcode() == R600::CONTINUE
690 && It->getOpcode() == R600::ENDLOOP)
691 ContInstr.push_back(&*Pre);
692 Pre = It;
693 ++It;
696 //delete continue right before endloop
697 for (unsigned i = 0; i < ContInstr.size(); ++i)
698 ContInstr[i]->eraseFromParent();
700 // TODO to fix up jump table so later phase won't be confused. if
701 // (jumpTableInfo->isEmpty() == false) { need to clean the jump table, but
702 // there isn't such an interface yet. alternatively, replace all the other
703 // blocks in the jump table with the entryBlk //}
706 bool AMDGPUCFGStructurizer::prepare() {
707 bool Changed = false;
709 //FIXME: if not reducible flow graph, make it so ???
711 LLVM_DEBUG(dbgs() << "AMDGPUCFGStructurizer::prepare\n";);
713 orderBlocks(FuncRep);
715 SmallVector<MachineBasicBlock *, DEFAULT_VEC_SLOTS> RetBlks;
717 // Add an ExitBlk to loop that don't have one
718 for (MachineLoopInfo::iterator It = MLI->begin(),
719 E = MLI->end(); It != E; ++It) {
720 MachineLoop *LoopRep = (*It);
721 MBBVector ExitingMBBs;
722 LoopRep->getExitingBlocks(ExitingMBBs);
724 if (ExitingMBBs.size() == 0) {
725 MachineBasicBlock* DummyExitBlk = normalizeInfiniteLoopExit(LoopRep);
726 if (DummyExitBlk)
727 RetBlks.push_back(DummyExitBlk);
731 // Remove unconditional branch instr.
732 // Add dummy exit block iff there are multiple returns.
733 for (SmallVectorImpl<MachineBasicBlock *>::const_iterator
734 It = OrderedBlks.begin(), E = OrderedBlks.end(); It != E; ++It) {
735 MachineBasicBlock *MBB = *It;
736 removeUnconditionalBranch(MBB);
737 removeRedundantConditionalBranch(MBB);
738 if (isReturnBlock(MBB)) {
739 RetBlks.push_back(MBB);
741 assert(MBB->succ_size() <= 2);
744 if (RetBlks.size() >= 2) {
745 addDummyExitBlock(RetBlks);
746 Changed = true;
749 return Changed;
752 bool AMDGPUCFGStructurizer::run() {
753 //Assume reducible CFG...
754 LLVM_DEBUG(dbgs() << "AMDGPUCFGStructurizer::run\n");
756 #ifdef STRESSTEST
757 //Use the worse block ordering to test the algorithm.
758 ReverseVector(orderedBlks);
759 #endif
761 LLVM_DEBUG(dbgs() << "Ordered blocks:\n"; printOrderedBlocks(););
762 int NumIter = 0;
763 bool Finish = false;
764 MachineBasicBlock *MBB;
765 bool MakeProgress = false;
766 int NumRemainedBlk = countActiveBlock(OrderedBlks.begin(),
767 OrderedBlks.end());
769 do {
770 ++NumIter;
771 LLVM_DEBUG(dbgs() << "numIter = " << NumIter
772 << ", numRemaintedBlk = " << NumRemainedBlk << "\n";);
774 SmallVectorImpl<MachineBasicBlock *>::const_iterator It =
775 OrderedBlks.begin();
776 SmallVectorImpl<MachineBasicBlock *>::const_iterator E =
777 OrderedBlks.end();
779 SmallVectorImpl<MachineBasicBlock *>::const_iterator SccBeginIter =
781 MachineBasicBlock *SccBeginMBB = nullptr;
782 int SccNumBlk = 0; // The number of active blocks, init to a
783 // maximum possible number.
784 int SccNumIter; // Number of iteration in this SCC.
786 while (It != E) {
787 MBB = *It;
789 if (!SccBeginMBB) {
790 SccBeginIter = It;
791 SccBeginMBB = MBB;
792 SccNumIter = 0;
793 SccNumBlk = NumRemainedBlk; // Init to maximum possible number.
794 LLVM_DEBUG(dbgs() << "start processing SCC" << getSCCNum(SccBeginMBB);
795 dbgs() << "\n";);
798 if (!isRetiredBlock(MBB))
799 patternMatch(MBB);
801 ++It;
803 bool ContNextScc = true;
804 if (It == E
805 || getSCCNum(SccBeginMBB) != getSCCNum(*It)) {
806 // Just finish one scc.
807 ++SccNumIter;
808 int sccRemainedNumBlk = countActiveBlock(SccBeginIter, It);
809 if (sccRemainedNumBlk != 1 && sccRemainedNumBlk >= SccNumBlk) {
810 LLVM_DEBUG(dbgs() << "Can't reduce SCC " << getSCCNum(MBB)
811 << ", sccNumIter = " << SccNumIter;
812 dbgs() << "doesn't make any progress\n";);
813 ContNextScc = true;
814 } else if (sccRemainedNumBlk != 1 && sccRemainedNumBlk < SccNumBlk) {
815 SccNumBlk = sccRemainedNumBlk;
816 It = SccBeginIter;
817 ContNextScc = false;
818 LLVM_DEBUG(dbgs() << "repeat processing SCC" << getSCCNum(MBB)
819 << "sccNumIter = " << SccNumIter << '\n';);
820 } else {
821 // Finish the current scc.
822 ContNextScc = true;
824 } else {
825 // Continue on next component in the current scc.
826 ContNextScc = false;
829 if (ContNextScc)
830 SccBeginMBB = nullptr;
831 } //while, "one iteration" over the function.
833 MachineBasicBlock *EntryMBB =
834 *GraphTraits<MachineFunction *>::nodes_begin(FuncRep);
835 if (EntryMBB->succ_size() == 0) {
836 Finish = true;
837 LLVM_DEBUG(dbgs() << "Reduce to one block\n";);
838 } else {
839 int NewnumRemainedBlk
840 = countActiveBlock(OrderedBlks.begin(), OrderedBlks.end());
841 // consider cloned blocks ??
842 if (NewnumRemainedBlk == 1 || NewnumRemainedBlk < NumRemainedBlk) {
843 MakeProgress = true;
844 NumRemainedBlk = NewnumRemainedBlk;
845 } else {
846 MakeProgress = false;
847 LLVM_DEBUG(dbgs() << "No progress\n";);
850 } while (!Finish && MakeProgress);
852 // Misc wrap up to maintain the consistency of the Function representation.
853 wrapup(*GraphTraits<MachineFunction *>::nodes_begin(FuncRep));
855 // Detach retired Block, release memory.
856 for (MBBInfoMap::iterator It = BlockInfoMap.begin(), E = BlockInfoMap.end();
857 It != E; ++It) {
858 if ((*It).second && (*It).second->IsRetired) {
859 assert(((*It).first)->getNumber() != -1);
860 LLVM_DEBUG(dbgs() << "Erase BB" << ((*It).first)->getNumber() << "\n";);
861 (*It).first->eraseFromParent(); //Remove from the parent Function.
863 delete (*It).second;
865 BlockInfoMap.clear();
866 LLInfoMap.clear();
868 if (!Finish) {
869 LLVM_DEBUG(FuncRep->viewCFG());
870 report_fatal_error("IRREDUCIBLE_CFG");
873 return true;
876 void AMDGPUCFGStructurizer::orderBlocks(MachineFunction *MF) {
877 int SccNum = 0;
878 MachineBasicBlock *MBB;
879 for (scc_iterator<MachineFunction *> It = scc_begin(MF); !It.isAtEnd();
880 ++It, ++SccNum) {
881 const std::vector<MachineBasicBlock *> &SccNext = *It;
882 for (std::vector<MachineBasicBlock *>::const_iterator
883 blockIter = SccNext.begin(), blockEnd = SccNext.end();
884 blockIter != blockEnd; ++blockIter) {
885 MBB = *blockIter;
886 OrderedBlks.push_back(MBB);
887 recordSccnum(MBB, SccNum);
891 // walk through all the block in func to check for unreachable
892 for (auto *MBB : nodes(MF)) {
893 SccNum = getSCCNum(MBB);
894 if (SccNum == INVALIDSCCNUM)
895 dbgs() << "unreachable block BB" << MBB->getNumber() << "\n";
899 int AMDGPUCFGStructurizer::patternMatch(MachineBasicBlock *MBB) {
900 int NumMatch = 0;
901 int CurMatch;
903 LLVM_DEBUG(dbgs() << "Begin patternMatch BB" << MBB->getNumber() << "\n";);
905 while ((CurMatch = patternMatchGroup(MBB)) > 0)
906 NumMatch += CurMatch;
908 LLVM_DEBUG(dbgs() << "End patternMatch BB" << MBB->getNumber()
909 << ", numMatch = " << NumMatch << "\n";);
911 return NumMatch;
914 int AMDGPUCFGStructurizer::patternMatchGroup(MachineBasicBlock *MBB) {
915 int NumMatch = 0;
916 NumMatch += loopendPatternMatch();
917 NumMatch += serialPatternMatch(MBB);
918 NumMatch += ifPatternMatch(MBB);
919 return NumMatch;
922 int AMDGPUCFGStructurizer::serialPatternMatch(MachineBasicBlock *MBB) {
923 if (MBB->succ_size() != 1)
924 return 0;
926 MachineBasicBlock *childBlk = *MBB->succ_begin();
927 if (childBlk->pred_size() != 1 || isActiveLoophead(childBlk))
928 return 0;
930 mergeSerialBlock(MBB, childBlk);
931 ++numSerialPatternMatch;
932 return 1;
935 int AMDGPUCFGStructurizer::ifPatternMatch(MachineBasicBlock *MBB) {
936 //two edges
937 if (MBB->succ_size() != 2)
938 return 0;
939 if (hasBackEdge(MBB))
940 return 0;
941 MachineInstr *BranchMI = getNormalBlockBranchInstr(MBB);
942 if (!BranchMI)
943 return 0;
945 assert(isCondBranch(BranchMI));
946 int NumMatch = 0;
948 MachineBasicBlock *TrueMBB = getTrueBranch(BranchMI);
949 NumMatch += serialPatternMatch(TrueMBB);
950 NumMatch += ifPatternMatch(TrueMBB);
951 MachineBasicBlock *FalseMBB = getFalseBranch(MBB, BranchMI);
952 NumMatch += serialPatternMatch(FalseMBB);
953 NumMatch += ifPatternMatch(FalseMBB);
954 MachineBasicBlock *LandBlk;
955 int Cloned = 0;
957 assert (!TrueMBB->succ_empty() || !FalseMBB->succ_empty());
958 // TODO: Simplify
959 if (TrueMBB->succ_size() == 1 && FalseMBB->succ_size() == 1
960 && *TrueMBB->succ_begin() == *FalseMBB->succ_begin()) {
961 // Diamond pattern
962 LandBlk = *TrueMBB->succ_begin();
963 } else if (TrueMBB->succ_size() == 1 && *TrueMBB->succ_begin() == FalseMBB) {
964 // Triangle pattern, false is empty
965 LandBlk = FalseMBB;
966 FalseMBB = nullptr;
967 } else if (FalseMBB->succ_size() == 1
968 && *FalseMBB->succ_begin() == TrueMBB) {
969 // Triangle pattern, true is empty
970 // We reverse the predicate to make a triangle, empty false pattern;
971 std::swap(TrueMBB, FalseMBB);
972 reversePredicateSetter(MBB->end(), *MBB);
973 LandBlk = FalseMBB;
974 FalseMBB = nullptr;
975 } else if (FalseMBB->succ_size() == 1
976 && isSameloopDetachedContbreak(TrueMBB, FalseMBB)) {
977 LandBlk = *FalseMBB->succ_begin();
978 } else if (TrueMBB->succ_size() == 1
979 && isSameloopDetachedContbreak(FalseMBB, TrueMBB)) {
980 LandBlk = *TrueMBB->succ_begin();
981 } else {
982 return NumMatch + handleJumpintoIf(MBB, TrueMBB, FalseMBB);
985 // improveSimpleJumpinfoIf can handle the case where landBlk == NULL but the
986 // new BB created for landBlk==NULL may introduce new challenge to the
987 // reduction process.
988 if (LandBlk &&
989 ((TrueMBB && TrueMBB->pred_size() > 1)
990 || (FalseMBB && FalseMBB->pred_size() > 1))) {
991 Cloned += improveSimpleJumpintoIf(MBB, TrueMBB, FalseMBB, &LandBlk);
994 if (TrueMBB && TrueMBB->pred_size() > 1) {
995 TrueMBB = cloneBlockForPredecessor(TrueMBB, MBB);
996 ++Cloned;
999 if (FalseMBB && FalseMBB->pred_size() > 1) {
1000 FalseMBB = cloneBlockForPredecessor(FalseMBB, MBB);
1001 ++Cloned;
1004 mergeIfthenelseBlock(BranchMI, MBB, TrueMBB, FalseMBB, LandBlk);
1006 ++numIfPatternMatch;
1008 numClonedBlock += Cloned;
1010 return 1 + Cloned + NumMatch;
1013 int AMDGPUCFGStructurizer::loopendPatternMatch() {
1014 std::deque<MachineLoop *> NestedLoops;
1015 for (auto &It: *MLI)
1016 for (MachineLoop *ML : depth_first(It))
1017 NestedLoops.push_front(ML);
1019 if (NestedLoops.empty())
1020 return 0;
1022 // Process nested loop outside->inside (we did push_front),
1023 // so "continue" to a outside loop won't be mistaken as "break"
1024 // of the current loop.
1025 int Num = 0;
1026 for (MachineLoop *ExaminedLoop : NestedLoops) {
1027 if (ExaminedLoop->getNumBlocks() == 0 || Visited[ExaminedLoop])
1028 continue;
1029 LLVM_DEBUG(dbgs() << "Processing:\n"; ExaminedLoop->dump(););
1030 int NumBreak = mergeLoop(ExaminedLoop);
1031 if (NumBreak == -1)
1032 break;
1033 Num += NumBreak;
1035 return Num;
1038 int AMDGPUCFGStructurizer::mergeLoop(MachineLoop *LoopRep) {
1039 MachineBasicBlock *LoopHeader = LoopRep->getHeader();
1040 MBBVector ExitingMBBs;
1041 LoopRep->getExitingBlocks(ExitingMBBs);
1042 assert(!ExitingMBBs.empty() && "Infinite Loop not supported");
1043 LLVM_DEBUG(dbgs() << "Loop has " << ExitingMBBs.size()
1044 << " exiting blocks\n";);
1045 // We assume a single ExitBlk
1046 MBBVector ExitBlks;
1047 LoopRep->getExitBlocks(ExitBlks);
1048 SmallPtrSet<MachineBasicBlock *, 2> ExitBlkSet;
1049 for (unsigned i = 0, e = ExitBlks.size(); i < e; ++i)
1050 ExitBlkSet.insert(ExitBlks[i]);
1051 assert(ExitBlkSet.size() == 1);
1052 MachineBasicBlock *ExitBlk = *ExitBlks.begin();
1053 assert(ExitBlk && "Loop has several exit block");
1054 MBBVector LatchBlks;
1055 for (auto *LB : inverse_children<MachineBasicBlock*>(LoopHeader))
1056 if (LoopRep->contains(LB))
1057 LatchBlks.push_back(LB);
1059 for (unsigned i = 0, e = ExitingMBBs.size(); i < e; ++i)
1060 mergeLoopbreakBlock(ExitingMBBs[i], ExitBlk);
1061 for (unsigned i = 0, e = LatchBlks.size(); i < e; ++i)
1062 settleLoopcontBlock(LatchBlks[i], LoopHeader);
1063 int Match = 0;
1064 do {
1065 Match = 0;
1066 Match += serialPatternMatch(LoopHeader);
1067 Match += ifPatternMatch(LoopHeader);
1068 } while (Match > 0);
1069 mergeLooplandBlock(LoopHeader, ExitBlk);
1070 MachineLoop *ParentLoop = LoopRep->getParentLoop();
1071 if (ParentLoop)
1072 MLI->changeLoopFor(LoopHeader, ParentLoop);
1073 else
1074 MLI->removeBlock(LoopHeader);
1075 Visited[LoopRep] = true;
1076 return 1;
1079 bool AMDGPUCFGStructurizer::isSameloopDetachedContbreak(
1080 MachineBasicBlock *Src1MBB, MachineBasicBlock *Src2MBB) {
1081 if (Src1MBB->succ_size() == 0) {
1082 MachineLoop *LoopRep = MLI->getLoopFor(Src1MBB);
1083 if (LoopRep&& LoopRep == MLI->getLoopFor(Src2MBB)) {
1084 MachineBasicBlock *&TheEntry = LLInfoMap[LoopRep];
1085 if (TheEntry) {
1086 LLVM_DEBUG(dbgs() << "isLoopContBreakBlock yes src1 = BB"
1087 << Src1MBB->getNumber() << " src2 = BB"
1088 << Src2MBB->getNumber() << "\n";);
1089 return true;
1093 return false;
1096 int AMDGPUCFGStructurizer::handleJumpintoIf(MachineBasicBlock *HeadMBB,
1097 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB) {
1098 int Num = handleJumpintoIfImp(HeadMBB, TrueMBB, FalseMBB);
1099 if (Num == 0) {
1100 LLVM_DEBUG(dbgs() << "handleJumpintoIf swap trueBlk and FalseBlk"
1101 << "\n";);
1102 Num = handleJumpintoIfImp(HeadMBB, FalseMBB, TrueMBB);
1104 return Num;
1107 int AMDGPUCFGStructurizer::handleJumpintoIfImp(MachineBasicBlock *HeadMBB,
1108 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB) {
1109 int Num = 0;
1110 MachineBasicBlock *DownBlk;
1112 //trueBlk could be the common post dominator
1113 DownBlk = TrueMBB;
1115 LLVM_DEBUG(dbgs() << "handleJumpintoIfImp head = BB" << HeadMBB->getNumber()
1116 << " true = BB" << TrueMBB->getNumber()
1117 << ", numSucc=" << TrueMBB->succ_size() << " false = BB"
1118 << FalseMBB->getNumber() << "\n";);
1120 while (DownBlk) {
1121 LLVM_DEBUG(dbgs() << "check down = BB" << DownBlk->getNumber(););
1123 if (singlePathTo(FalseMBB, DownBlk) == SinglePath_InPath) {
1124 LLVM_DEBUG(dbgs() << " working\n";);
1126 Num += cloneOnSideEntryTo(HeadMBB, TrueMBB, DownBlk);
1127 Num += cloneOnSideEntryTo(HeadMBB, FalseMBB, DownBlk);
1129 numClonedBlock += Num;
1130 Num += serialPatternMatch(*HeadMBB->succ_begin());
1131 Num += serialPatternMatch(*std::next(HeadMBB->succ_begin()));
1132 Num += ifPatternMatch(HeadMBB);
1133 assert(Num > 0);
1135 break;
1137 LLVM_DEBUG(dbgs() << " not working\n";);
1138 DownBlk = (DownBlk->succ_size() == 1) ? (*DownBlk->succ_begin()) : nullptr;
1139 } // walk down the postDomTree
1141 return Num;
1144 #ifndef NDEBUG
1145 void AMDGPUCFGStructurizer::showImproveSimpleJumpintoIf(
1146 MachineBasicBlock *HeadMBB, MachineBasicBlock *TrueMBB,
1147 MachineBasicBlock *FalseMBB, MachineBasicBlock *LandMBB, bool Detail) {
1148 dbgs() << "head = BB" << HeadMBB->getNumber()
1149 << " size = " << HeadMBB->size();
1150 if (Detail) {
1151 dbgs() << "\n";
1152 HeadMBB->print(dbgs());
1153 dbgs() << "\n";
1156 if (TrueMBB) {
1157 dbgs() << ", true = BB" << TrueMBB->getNumber() << " size = "
1158 << TrueMBB->size() << " numPred = " << TrueMBB->pred_size();
1159 if (Detail) {
1160 dbgs() << "\n";
1161 TrueMBB->print(dbgs());
1162 dbgs() << "\n";
1165 if (FalseMBB) {
1166 dbgs() << ", false = BB" << FalseMBB->getNumber() << " size = "
1167 << FalseMBB->size() << " numPred = " << FalseMBB->pred_size();
1168 if (Detail) {
1169 dbgs() << "\n";
1170 FalseMBB->print(dbgs());
1171 dbgs() << "\n";
1174 if (LandMBB) {
1175 dbgs() << ", land = BB" << LandMBB->getNumber() << " size = "
1176 << LandMBB->size() << " numPred = " << LandMBB->pred_size();
1177 if (Detail) {
1178 dbgs() << "\n";
1179 LandMBB->print(dbgs());
1180 dbgs() << "\n";
1184 dbgs() << "\n";
1186 #endif
1188 int AMDGPUCFGStructurizer::improveSimpleJumpintoIf(MachineBasicBlock *HeadMBB,
1189 MachineBasicBlock *TrueMBB, MachineBasicBlock *FalseMBB,
1190 MachineBasicBlock **LandMBBPtr) {
1191 bool MigrateTrue = false;
1192 bool MigrateFalse = false;
1194 MachineBasicBlock *LandBlk = *LandMBBPtr;
1196 assert((!TrueMBB || TrueMBB->succ_size() <= 1)
1197 && (!FalseMBB || FalseMBB->succ_size() <= 1));
1199 if (TrueMBB == FalseMBB)
1200 return 0;
1202 MigrateTrue = needMigrateBlock(TrueMBB);
1203 MigrateFalse = needMigrateBlock(FalseMBB);
1205 if (!MigrateTrue && !MigrateFalse)
1206 return 0;
1208 // If we need to migrate either trueBlk and falseBlk, migrate the rest that
1209 // have more than one predecessors. without doing this, its predecessor
1210 // rather than headBlk will have undefined value in initReg.
1211 if (!MigrateTrue && TrueMBB && TrueMBB->pred_size() > 1)
1212 MigrateTrue = true;
1213 if (!MigrateFalse && FalseMBB && FalseMBB->pred_size() > 1)
1214 MigrateFalse = true;
1216 LLVM_DEBUG(
1217 dbgs() << "before improveSimpleJumpintoIf: ";
1218 showImproveSimpleJumpintoIf(HeadMBB, TrueMBB, FalseMBB, LandBlk, 0););
1220 // org: headBlk => if () {trueBlk} else {falseBlk} => landBlk
1222 // new: headBlk => if () {initReg = 1; org trueBlk branch} else
1223 // {initReg = 0; org falseBlk branch }
1224 // => landBlk => if (initReg) {org trueBlk} else {org falseBlk}
1225 // => org landBlk
1226 // if landBlk->pred_size() > 2, put the about if-else inside
1227 // if (initReg !=2) {...}
1229 // add initReg = initVal to headBlk
1231 const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
1232 if (!MigrateTrue || !MigrateFalse) {
1233 // XXX: We have an opportunity here to optimize the "branch into if" case
1234 // here. Branch into if looks like this:
1235 // entry
1236 // / |
1237 // diamond_head branch_from
1238 // / \ |
1239 // diamond_false diamond_true
1240 // \ /
1241 // done
1243 // The diamond_head block begins the "if" and the diamond_true block
1244 // is the block being "branched into".
1246 // If MigrateTrue is true, then TrueBB is the block being "branched into"
1247 // and if MigrateFalse is true, then FalseBB is the block being
1248 // "branched into"
1250 // Here is the pseudo code for how I think the optimization should work:
1251 // 1. Insert MOV GPR0, 0 before the branch instruction in diamond_head.
1252 // 2. Insert MOV GPR0, 1 before the branch instruction in branch_from.
1253 // 3. Move the branch instruction from diamond_head into its own basic
1254 // block (new_block).
1255 // 4. Add an unconditional branch from diamond_head to new_block
1256 // 5. Replace the branch instruction in branch_from with an unconditional
1257 // branch to new_block. If branch_from has multiple predecessors, then
1258 // we need to replace the True/False block in the branch
1259 // instruction instead of replacing it.
1260 // 6. Change the condition of the branch instruction in new_block from
1261 // COND to (COND || GPR0)
1263 // In order insert these MOV instruction, we will need to use the
1264 // RegisterScavenger. Usually liveness stops being tracked during
1265 // the late machine optimization passes, however if we implement
1266 // bool TargetRegisterInfo::requiresRegisterScavenging(
1267 // const MachineFunction &MF)
1268 // and have it return true, liveness will be tracked correctly
1269 // by generic optimization passes. We will also need to make sure that
1270 // all of our target-specific passes that run after regalloc and before
1271 // the CFGStructurizer track liveness and we will need to modify this pass
1272 // to correctly track liveness.
1274 // After the above changes, the new CFG should look like this:
1275 // entry
1276 // / |
1277 // diamond_head branch_from
1278 // \ /
1279 // new_block
1280 // / |
1281 // diamond_false diamond_true
1282 // \ /
1283 // done
1285 // Without this optimization, we are forced to duplicate the diamond_true
1286 // block and we will end up with a CFG like this:
1288 // entry
1289 // / |
1290 // diamond_head branch_from
1291 // / \ |
1292 // diamond_false diamond_true diamond_true (duplicate)
1293 // \ / |
1294 // done --------------------|
1296 // Duplicating diamond_true can be very costly especially if it has a
1297 // lot of instructions.
1298 return 0;
1301 int NumNewBlk = 0;
1303 bool LandBlkHasOtherPred = (LandBlk->pred_size() > 2);
1305 //insert R600::ENDIF to avoid special case "input landBlk == NULL"
1306 MachineBasicBlock::iterator I = insertInstrBefore(LandBlk, R600::ENDIF);
1308 if (LandBlkHasOtherPred) {
1309 report_fatal_error("Extra register needed to handle CFG");
1310 Register CmpResReg =
1311 HeadMBB->getParent()->getRegInfo().createVirtualRegister(I32RC);
1312 report_fatal_error("Extra compare instruction needed to handle CFG");
1313 insertCondBranchBefore(LandBlk, I, R600::IF_PREDICATE_SET,
1314 CmpResReg, DebugLoc());
1317 // XXX: We are running this after RA, so creating virtual registers will
1318 // cause an assertion failure in the PostRA scheduling pass.
1319 Register InitReg =
1320 HeadMBB->getParent()->getRegInfo().createVirtualRegister(I32RC);
1321 insertCondBranchBefore(LandBlk, I, R600::IF_PREDICATE_SET, InitReg,
1322 DebugLoc());
1324 if (MigrateTrue) {
1325 migrateInstruction(TrueMBB, LandBlk, I);
1326 // need to uncondionally insert the assignment to ensure a path from its
1327 // predecessor rather than headBlk has valid value in initReg if
1328 // (initVal != 1).
1329 report_fatal_error("Extra register needed to handle CFG");
1331 insertInstrBefore(I, R600::ELSE);
1333 if (MigrateFalse) {
1334 migrateInstruction(FalseMBB, LandBlk, I);
1335 // need to uncondionally insert the assignment to ensure a path from its
1336 // predecessor rather than headBlk has valid value in initReg if
1337 // (initVal != 0)
1338 report_fatal_error("Extra register needed to handle CFG");
1341 if (LandBlkHasOtherPred) {
1342 // add endif
1343 insertInstrBefore(I, R600::ENDIF);
1345 // put initReg = 2 to other predecessors of landBlk
1346 for (MachineBasicBlock::pred_iterator PI = LandBlk->pred_begin(),
1347 PE = LandBlk->pred_end(); PI != PE; ++PI) {
1348 MachineBasicBlock *MBB = *PI;
1349 if (MBB != TrueMBB && MBB != FalseMBB)
1350 report_fatal_error("Extra register needed to handle CFG");
1353 LLVM_DEBUG(
1354 dbgs() << "result from improveSimpleJumpintoIf: ";
1355 showImproveSimpleJumpintoIf(HeadMBB, TrueMBB, FalseMBB, LandBlk, 0););
1357 // update landBlk
1358 *LandMBBPtr = LandBlk;
1360 return NumNewBlk;
1363 void AMDGPUCFGStructurizer::mergeSerialBlock(MachineBasicBlock *DstMBB,
1364 MachineBasicBlock *SrcMBB) {
1365 LLVM_DEBUG(dbgs() << "serialPattern BB" << DstMBB->getNumber() << " <= BB"
1366 << SrcMBB->getNumber() << "\n";);
1367 DstMBB->splice(DstMBB->end(), SrcMBB, SrcMBB->begin(), SrcMBB->end());
1369 DstMBB->removeSuccessor(SrcMBB, true);
1370 cloneSuccessorList(DstMBB, SrcMBB);
1372 removeSuccessor(SrcMBB);
1373 MLI->removeBlock(SrcMBB);
1374 retireBlock(SrcMBB);
1377 void AMDGPUCFGStructurizer::mergeIfthenelseBlock(MachineInstr *BranchMI,
1378 MachineBasicBlock *MBB, MachineBasicBlock *TrueMBB,
1379 MachineBasicBlock *FalseMBB, MachineBasicBlock *LandMBB) {
1380 assert (TrueMBB);
1381 LLVM_DEBUG(dbgs() << "ifPattern BB" << MBB->getNumber(); dbgs() << "{ ";
1382 if (TrueMBB) { dbgs() << "BB" << TrueMBB->getNumber(); } dbgs()
1383 << " } else ";
1384 dbgs() << "{ "; if (FalseMBB) {
1385 dbgs() << "BB" << FalseMBB->getNumber();
1386 } dbgs() << " }\n ";
1387 dbgs() << "landBlock: "; if (!LandMBB) { dbgs() << "NULL"; } else {
1388 dbgs() << "BB" << LandMBB->getNumber();
1389 } dbgs() << "\n";);
1391 int OldOpcode = BranchMI->getOpcode();
1392 DebugLoc BranchDL = BranchMI->getDebugLoc();
1394 // transform to
1395 // if cond
1396 // trueBlk
1397 // else
1398 // falseBlk
1399 // endif
1400 // landBlk
1402 MachineBasicBlock::iterator I = BranchMI;
1403 insertCondBranchBefore(I, getBranchNzeroOpcode(OldOpcode),
1404 BranchDL);
1406 if (TrueMBB) {
1407 MBB->splice(I, TrueMBB, TrueMBB->begin(), TrueMBB->end());
1408 MBB->removeSuccessor(TrueMBB, true);
1409 if (LandMBB && TrueMBB->succ_size()!=0)
1410 TrueMBB->removeSuccessor(LandMBB, true);
1411 retireBlock(TrueMBB);
1412 MLI->removeBlock(TrueMBB);
1415 if (FalseMBB) {
1416 insertInstrBefore(I, R600::ELSE);
1417 MBB->splice(I, FalseMBB, FalseMBB->begin(),
1418 FalseMBB->end());
1419 MBB->removeSuccessor(FalseMBB, true);
1420 if (LandMBB && FalseMBB->succ_size() != 0)
1421 FalseMBB->removeSuccessor(LandMBB, true);
1422 retireBlock(FalseMBB);
1423 MLI->removeBlock(FalseMBB);
1425 insertInstrBefore(I, R600::ENDIF);
1427 BranchMI->eraseFromParent();
1429 if (LandMBB && TrueMBB && FalseMBB)
1430 MBB->addSuccessor(LandMBB);
1433 void AMDGPUCFGStructurizer::mergeLooplandBlock(MachineBasicBlock *DstBlk,
1434 MachineBasicBlock *LandMBB) {
1435 LLVM_DEBUG(dbgs() << "loopPattern header = BB" << DstBlk->getNumber()
1436 << " land = BB" << LandMBB->getNumber() << "\n";);
1438 insertInstrBefore(DstBlk, R600::WHILELOOP, DebugLoc());
1439 insertInstrEnd(DstBlk, R600::ENDLOOP, DebugLoc());
1440 DstBlk->replaceSuccessor(DstBlk, LandMBB);
1443 void AMDGPUCFGStructurizer::mergeLoopbreakBlock(MachineBasicBlock *ExitingMBB,
1444 MachineBasicBlock *LandMBB) {
1445 LLVM_DEBUG(dbgs() << "loopbreakPattern exiting = BB"
1446 << ExitingMBB->getNumber() << " land = BB"
1447 << LandMBB->getNumber() << "\n";);
1448 MachineInstr *BranchMI = getLoopendBlockBranchInstr(ExitingMBB);
1449 assert(BranchMI && isCondBranch(BranchMI));
1450 DebugLoc DL = BranchMI->getDebugLoc();
1451 MachineBasicBlock *TrueBranch = getTrueBranch(BranchMI);
1452 MachineBasicBlock::iterator I = BranchMI;
1453 if (TrueBranch != LandMBB)
1454 reversePredicateSetter(I, *I->getParent());
1455 insertCondBranchBefore(ExitingMBB, I, R600::IF_PREDICATE_SET, R600::PREDICATE_BIT, DL);
1456 insertInstrBefore(I, R600::BREAK);
1457 insertInstrBefore(I, R600::ENDIF);
1458 //now branchInst can be erase safely
1459 BranchMI->eraseFromParent();
1460 //now take care of successors, retire blocks
1461 ExitingMBB->removeSuccessor(LandMBB, true);
1464 void AMDGPUCFGStructurizer::settleLoopcontBlock(MachineBasicBlock *ContingMBB,
1465 MachineBasicBlock *ContMBB) {
1466 LLVM_DEBUG(dbgs() << "settleLoopcontBlock conting = BB"
1467 << ContingMBB->getNumber() << ", cont = BB"
1468 << ContMBB->getNumber() << "\n";);
1470 MachineInstr *MI = getLoopendBlockBranchInstr(ContingMBB);
1471 if (MI) {
1472 assert(isCondBranch(MI));
1473 MachineBasicBlock::iterator I = MI;
1474 MachineBasicBlock *TrueBranch = getTrueBranch(MI);
1475 int OldOpcode = MI->getOpcode();
1476 DebugLoc DL = MI->getDebugLoc();
1478 bool UseContinueLogical = ((&*ContingMBB->rbegin()) == MI);
1480 if (!UseContinueLogical) {
1481 int BranchOpcode =
1482 TrueBranch == ContMBB ? getBranchNzeroOpcode(OldOpcode) :
1483 getBranchZeroOpcode(OldOpcode);
1484 insertCondBranchBefore(I, BranchOpcode, DL);
1485 // insertEnd to ensure phi-moves, if exist, go before the continue-instr.
1486 insertInstrEnd(ContingMBB, R600::CONTINUE, DL);
1487 insertInstrEnd(ContingMBB, R600::ENDIF, DL);
1488 } else {
1489 int BranchOpcode =
1490 TrueBranch == ContMBB ? getContinueNzeroOpcode(OldOpcode) :
1491 getContinueZeroOpcode(OldOpcode);
1492 insertCondBranchBefore(I, BranchOpcode, DL);
1495 MI->eraseFromParent();
1496 } else {
1497 // if we've arrived here then we've already erased the branch instruction
1498 // travel back up the basic block to see the last reference of our debug
1499 // location we've just inserted that reference here so it should be
1500 // representative insertEnd to ensure phi-moves, if exist, go before the
1501 // continue-instr.
1502 insertInstrEnd(ContingMBB, R600::CONTINUE,
1503 getLastDebugLocInBB(ContingMBB));
1507 int AMDGPUCFGStructurizer::cloneOnSideEntryTo(MachineBasicBlock *PreMBB,
1508 MachineBasicBlock *SrcMBB, MachineBasicBlock *DstMBB) {
1509 int Cloned = 0;
1510 assert(PreMBB->isSuccessor(SrcMBB));
1511 while (SrcMBB && SrcMBB != DstMBB) {
1512 assert(SrcMBB->succ_size() == 1);
1513 if (SrcMBB->pred_size() > 1) {
1514 SrcMBB = cloneBlockForPredecessor(SrcMBB, PreMBB);
1515 ++Cloned;
1518 PreMBB = SrcMBB;
1519 SrcMBB = *SrcMBB->succ_begin();
1522 return Cloned;
1525 MachineBasicBlock *
1526 AMDGPUCFGStructurizer::cloneBlockForPredecessor(MachineBasicBlock *MBB,
1527 MachineBasicBlock *PredMBB) {
1528 assert(PredMBB->isSuccessor(MBB) &&
1529 "succBlk is not a prececessor of curBlk");
1531 MachineBasicBlock *CloneMBB = clone(MBB); //clone instructions
1532 replaceInstrUseOfBlockWith(PredMBB, MBB, CloneMBB);
1533 //srcBlk, oldBlk, newBlk
1535 PredMBB->replaceSuccessor(MBB, CloneMBB);
1537 // add all successor to cloneBlk
1538 cloneSuccessorList(CloneMBB, MBB);
1540 numClonedInstr += MBB->size();
1542 LLVM_DEBUG(dbgs() << "Cloned block: "
1543 << "BB" << MBB->getNumber() << "size " << MBB->size()
1544 << "\n";);
1546 SHOWNEWBLK(CloneMBB, "result of Cloned block: ");
1548 return CloneMBB;
1551 void AMDGPUCFGStructurizer::migrateInstruction(MachineBasicBlock *SrcMBB,
1552 MachineBasicBlock *DstMBB, MachineBasicBlock::iterator I) {
1553 MachineBasicBlock::iterator SpliceEnd;
1554 //look for the input branchinstr, not the AMDGPU branchinstr
1555 MachineInstr *BranchMI = getNormalBlockBranchInstr(SrcMBB);
1556 if (!BranchMI) {
1557 LLVM_DEBUG(dbgs() << "migrateInstruction don't see branch instr\n";);
1558 SpliceEnd = SrcMBB->end();
1559 } else {
1560 LLVM_DEBUG(dbgs() << "migrateInstruction see branch instr: " << *BranchMI);
1561 SpliceEnd = BranchMI;
1563 LLVM_DEBUG(dbgs() << "migrateInstruction before splice dstSize = "
1564 << DstMBB->size() << "srcSize = " << SrcMBB->size()
1565 << "\n";);
1567 //splice insert before insertPos
1568 DstMBB->splice(I, SrcMBB, SrcMBB->begin(), SpliceEnd);
1570 LLVM_DEBUG(dbgs() << "migrateInstruction after splice dstSize = "
1571 << DstMBB->size() << "srcSize = " << SrcMBB->size()
1572 << '\n';);
1575 MachineBasicBlock *
1576 AMDGPUCFGStructurizer::normalizeInfiniteLoopExit(MachineLoop* LoopRep) {
1577 MachineBasicBlock *LoopHeader = LoopRep->getHeader();
1578 MachineBasicBlock *LoopLatch = LoopRep->getLoopLatch();
1580 if (!LoopHeader || !LoopLatch)
1581 return nullptr;
1582 MachineInstr *BranchMI = getLoopendBlockBranchInstr(LoopLatch);
1583 // Is LoopRep an infinite loop ?
1584 if (!BranchMI || !isUncondBranch(BranchMI))
1585 return nullptr;
1587 MachineBasicBlock *DummyExitBlk = FuncRep->CreateMachineBasicBlock();
1588 FuncRep->push_back(DummyExitBlk); //insert to function
1589 SHOWNEWBLK(DummyExitBlk, "DummyExitBlock to normalize infiniteLoop: ");
1590 LLVM_DEBUG(dbgs() << "Old branch instr: " << *BranchMI << "\n";);
1591 LLVMContext &Ctx = LoopHeader->getParent()->getFunction().getContext();
1592 Ctx.emitError("Extra register needed to handle CFG");
1593 return nullptr;
1596 void AMDGPUCFGStructurizer::removeUnconditionalBranch(MachineBasicBlock *MBB) {
1597 MachineInstr *BranchMI;
1599 // I saw two unconditional branch in one basic block in example
1600 // test_fc_do_while_or.c need to fix the upstream on this to remove the loop.
1601 while ((BranchMI = getLoopendBlockBranchInstr(MBB))
1602 && isUncondBranch(BranchMI)) {
1603 LLVM_DEBUG(dbgs() << "Removing uncond branch instr: " << *BranchMI);
1604 BranchMI->eraseFromParent();
1608 void AMDGPUCFGStructurizer::removeRedundantConditionalBranch(
1609 MachineBasicBlock *MBB) {
1610 if (MBB->succ_size() != 2)
1611 return;
1612 MachineBasicBlock *MBB1 = *MBB->succ_begin();
1613 MachineBasicBlock *MBB2 = *std::next(MBB->succ_begin());
1614 if (MBB1 != MBB2)
1615 return;
1617 MachineInstr *BranchMI = getNormalBlockBranchInstr(MBB);
1618 assert(BranchMI && isCondBranch(BranchMI));
1619 LLVM_DEBUG(dbgs() << "Removing unneeded cond branch instr: " << *BranchMI);
1620 BranchMI->eraseFromParent();
1621 SHOWNEWBLK(MBB1, "Removing redundant successor");
1622 MBB->removeSuccessor(MBB1, true);
1625 void AMDGPUCFGStructurizer::addDummyExitBlock(
1626 SmallVectorImpl<MachineBasicBlock*> &RetMBB) {
1627 MachineBasicBlock *DummyExitBlk = FuncRep->CreateMachineBasicBlock();
1628 FuncRep->push_back(DummyExitBlk); //insert to function
1629 insertInstrEnd(DummyExitBlk, R600::RETURN);
1631 for (SmallVectorImpl<MachineBasicBlock *>::iterator It = RetMBB.begin(),
1632 E = RetMBB.end(); It != E; ++It) {
1633 MachineBasicBlock *MBB = *It;
1634 MachineInstr *MI = getReturnInstr(MBB);
1635 if (MI)
1636 MI->eraseFromParent();
1637 MBB->addSuccessor(DummyExitBlk);
1638 LLVM_DEBUG(dbgs() << "Add dummyExitBlock to BB" << MBB->getNumber()
1639 << " successors\n";);
1641 SHOWNEWBLK(DummyExitBlk, "DummyExitBlock: ");
1644 void AMDGPUCFGStructurizer::removeSuccessor(MachineBasicBlock *MBB) {
1645 while (MBB->succ_size())
1646 MBB->removeSuccessor(*MBB->succ_begin());
1649 void AMDGPUCFGStructurizer::recordSccnum(MachineBasicBlock *MBB,
1650 int SccNum) {
1651 BlockInformation *&srcBlkInfo = BlockInfoMap[MBB];
1652 if (!srcBlkInfo)
1653 srcBlkInfo = new BlockInformation();
1654 srcBlkInfo->SccNum = SccNum;
1657 void AMDGPUCFGStructurizer::retireBlock(MachineBasicBlock *MBB) {
1658 LLVM_DEBUG(dbgs() << "Retiring BB" << MBB->getNumber() << "\n";);
1660 BlockInformation *&SrcBlkInfo = BlockInfoMap[MBB];
1662 if (!SrcBlkInfo)
1663 SrcBlkInfo = new BlockInformation();
1665 SrcBlkInfo->IsRetired = true;
1666 assert(MBB->succ_size() == 0 && MBB->pred_size() == 0
1667 && "can't retire block yet");
1670 INITIALIZE_PASS_BEGIN(AMDGPUCFGStructurizer, "amdgpustructurizer",
1671 "AMDGPU CFG Structurizer", false, false)
1672 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
1673 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
1674 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
1675 INITIALIZE_PASS_END(AMDGPUCFGStructurizer, "amdgpustructurizer",
1676 "AMDGPU CFG Structurizer", false, false)
1678 FunctionPass *llvm::createAMDGPUCFGStructurizerPass() {
1679 return new AMDGPUCFGStructurizer();