1 //===- StrongPhiElimination.cpp - Eliminate PHI nodes by inserting copies -===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass eliminates machine instruction PHI nodes by inserting copy
11 // instructions, using an intelligent copy-folding technique based on
12 // dominator information. This is technique is derived from:
14 // Budimlic, et al. Fast copy coalescing and live-range identification.
15 // In Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language
16 // Design and Implementation (Berlin, Germany, June 17 - 19, 2002).
17 // PLDI '02. ACM, New York, NY, 25-32.
18 // DOI= http://doi.acm.org/10.1145/512529.512534
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "strongphielim"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/CodeGen/LiveVariables.h"
25 #include "llvm/CodeGen/MachineDominators.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/Target/TargetInstrInfo.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/ADT/DepthFirstIterator.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Support/Compiler.h"
38 struct VISIBILITY_HIDDEN StrongPHIElimination : public MachineFunctionPass {
39 static char ID; // Pass identification, replacement for typeid
40 StrongPHIElimination() : MachineFunctionPass((intptr_t)&ID) {}
42 DenseMap<MachineBasicBlock*,
43 std::map<unsigned, unsigned> > Waiting;
45 std::map<unsigned, std::vector<unsigned> > Stacks;
46 std::set<unsigned> UsedByAnother;
48 bool runOnMachineFunction(MachineFunction &Fn);
50 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
51 AU.addRequired<MachineDominatorTree>();
52 AU.addRequired<LiveVariables>();
53 MachineFunctionPass::getAnalysisUsage(AU);
56 virtual void releaseMemory() {
64 struct DomForestNode {
66 std::vector<DomForestNode*> children;
69 void addChild(DomForestNode* DFN) { children.push_back(DFN); }
72 typedef std::vector<DomForestNode*>::iterator iterator;
74 DomForestNode(unsigned r, DomForestNode* parent) : reg(r) {
76 parent->addChild(this);
80 for (iterator I = begin(), E = end(); I != E; ++I)
84 inline unsigned getReg() { return reg; }
86 inline DomForestNode::iterator begin() { return children.begin(); }
87 inline DomForestNode::iterator end() { return children.end(); }
90 DenseMap<MachineBasicBlock*, unsigned> preorder;
91 DenseMap<MachineBasicBlock*, unsigned> maxpreorder;
94 void computeDFS(MachineFunction& MF);
95 void processBlock(MachineBasicBlock* MBB);
97 std::vector<DomForestNode*> computeDomForest(std::set<unsigned>& instrs);
98 void processPHIUnion(MachineInstr* Inst,
99 std::set<unsigned>& PHIUnion,
100 std::vector<StrongPHIElimination::DomForestNode*>& DF,
101 std::vector<std::pair<unsigned, unsigned> >& locals);
102 void ScheduleCopies(MachineBasicBlock* MBB, std::set<unsigned>& pushed);
103 void InsertCopies(MachineBasicBlock* MBB);
106 char StrongPHIElimination::ID = 0;
107 RegisterPass<StrongPHIElimination> X("strong-phi-node-elimination",
108 "Eliminate PHI nodes for register allocation, intelligently");
111 const PassInfo *llvm::StrongPHIEliminationID = X.getPassInfo();
113 /// computeDFS - Computes the DFS-in and DFS-out numbers of the dominator tree
114 /// of the given MachineFunction. These numbers are then used in other parts
115 /// of the PHI elimination process.
116 void StrongPHIElimination::computeDFS(MachineFunction& MF) {
117 SmallPtrSet<MachineDomTreeNode*, 8> frontier;
118 SmallPtrSet<MachineDomTreeNode*, 8> visited;
122 MachineDominatorTree& DT = getAnalysis<MachineDominatorTree>();
124 MachineDomTreeNode* node = DT.getRootNode();
126 std::vector<MachineDomTreeNode*> worklist;
127 worklist.push_back(node);
129 while (!worklist.empty()) {
130 MachineDomTreeNode* currNode = worklist.back();
132 if (!frontier.count(currNode)) {
133 frontier.insert(currNode);
135 preorder.insert(std::make_pair(currNode->getBlock(), time));
138 bool inserted = false;
139 for (MachineDomTreeNode::iterator I = node->begin(), E = node->end();
141 if (!frontier.count(*I) && !visited.count(*I)) {
142 worklist.push_back(*I);
148 frontier.erase(currNode);
149 visited.insert(currNode);
150 maxpreorder.insert(std::make_pair(currNode->getBlock(), time));
157 /// PreorderSorter - a helper class that is used to sort registers
158 /// according to the preorder number of their defining blocks
159 class PreorderSorter {
161 DenseMap<MachineBasicBlock*, unsigned>& preorder;
165 PreorderSorter(DenseMap<MachineBasicBlock*, unsigned>& p,
166 LiveVariables& L) : preorder(p), LV(L) { }
168 bool operator()(unsigned A, unsigned B) {
172 MachineBasicBlock* ABlock = LV.getVarInfo(A).DefInst->getParent();
173 MachineBasicBlock* BBlock = LV.getVarInfo(A).DefInst->getParent();
175 if (preorder[ABlock] < preorder[BBlock])
177 else if (preorder[ABlock] > preorder[BBlock])
180 assert(0 && "Error sorting by dominance!");
185 /// computeDomForest - compute the subforest of the DomTree corresponding
186 /// to the defining blocks of the registers in question
187 std::vector<StrongPHIElimination::DomForestNode*>
188 StrongPHIElimination::computeDomForest(std::set<unsigned>& regs) {
189 LiveVariables& LV = getAnalysis<LiveVariables>();
191 DomForestNode* VirtualRoot = new DomForestNode(0, 0);
192 maxpreorder.insert(std::make_pair((MachineBasicBlock*)0, ~0UL));
194 std::vector<unsigned> worklist;
195 worklist.reserve(regs.size());
196 for (std::set<unsigned>::iterator I = regs.begin(), E = regs.end();
198 worklist.push_back(*I);
200 PreorderSorter PS(preorder, LV);
201 std::sort(worklist.begin(), worklist.end(), PS);
203 DomForestNode* CurrentParent = VirtualRoot;
204 std::vector<DomForestNode*> stack;
205 stack.push_back(VirtualRoot);
207 for (std::vector<unsigned>::iterator I = worklist.begin(), E = worklist.end();
209 unsigned pre = preorder[LV.getVarInfo(*I).DefInst->getParent()];
210 MachineBasicBlock* parentBlock =
211 LV.getVarInfo(CurrentParent->getReg()).DefInst->getParent();
213 while (pre > maxpreorder[parentBlock]) {
215 CurrentParent = stack.back();
217 parentBlock = LV.getVarInfo(CurrentParent->getReg()).DefInst->getParent();
220 DomForestNode* child = new DomForestNode(*I, CurrentParent);
221 stack.push_back(child);
222 CurrentParent = child;
225 std::vector<DomForestNode*> ret;
226 ret.insert(ret.end(), VirtualRoot->begin(), VirtualRoot->end());
230 /// isLiveIn - helper method that determines, from a VarInfo, if a register
231 /// is live into a block
232 static bool isLiveIn(LiveVariables::VarInfo& V, MachineBasicBlock* MBB) {
233 if (V.AliveBlocks.test(MBB->getNumber()))
236 if (V.DefInst->getParent() != MBB &&
237 V.UsedBlocks.test(MBB->getNumber()))
243 /// isLiveOut - help method that determines, from a VarInfo, if a register is
244 /// live out of a block.
245 static bool isLiveOut(LiveVariables::VarInfo& V, MachineBasicBlock* MBB) {
246 if (MBB == V.DefInst->getParent() ||
247 V.UsedBlocks.test(MBB->getNumber())) {
248 for (std::vector<MachineInstr*>::iterator I = V.Kills.begin(),
249 E = V.Kills.end(); I != E; ++I)
250 if ((*I)->getParent() == MBB)
259 /// isKillInst - helper method that determines, from a VarInfo, if an
260 /// instruction kills a given register
261 static bool isKillInst(LiveVariables::VarInfo& V, MachineInstr* MI) {
262 return std::find(V.Kills.begin(), V.Kills.end(), MI) != V.Kills.end();
265 /// interferes - checks for local interferences by scanning a block. The only
266 /// trick parameter is 'mode' which tells it the relationship of the two
267 /// registers. 0 - defined in the same block, 1 - first properly dominates
268 /// second, 2 - second properly dominates first
269 static bool interferes(LiveVariables::VarInfo& First, LiveVariables::VarInfo& Second,
270 MachineBasicBlock* scan, unsigned mode) {
271 MachineInstr* def = 0;
272 MachineInstr* kill = 0;
274 bool interference = false;
276 // Wallk the block, checking for interferences
277 for (MachineBasicBlock::iterator MBI = scan->begin(), MBE = scan->end();
279 MachineInstr* curr = MBI;
281 // Same defining block...
283 if (curr == First.DefInst) {
284 // If we find our first DefInst, save it
287 // If there's already an unkilled DefInst, then
288 // this is an interference
292 // If there's a DefInst followed by a KillInst, then
293 // they can't interfere
295 interference = false;
298 // Symmetric with the above
299 } else if (curr == Second.DefInst ) {
306 interference = false;
309 // Store KillInsts if they match up with the DefInst
310 } else if (isKillInst(First, curr)) {
311 if (def == First.DefInst) {
313 } else if (isKillInst(Second, curr)) {
314 if (def == Second.DefInst) {
319 // First properly dominates second...
320 } else if (mode == 1) {
321 if (curr == Second.DefInst) {
322 // DefInst of second without kill of first is an interference
326 // DefInst after a kill is a non-interference
328 interference = false;
331 // Save KillInsts of First
332 } else if (isKillInst(First, curr)) {
335 // Symmetric with the above
336 } else if (mode == 2) {
337 if (curr == First.DefInst) {
342 interference = false;
345 } else if (isKillInst(Second, curr)) {
354 /// processBlock - Eliminate PHIs in the given block
355 void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
356 LiveVariables& LV = getAnalysis<LiveVariables>();
358 // Holds names that have been added to a set in any PHI within this block
359 // before the current one.
360 std::set<unsigned> ProcessedNames;
362 MachineBasicBlock::iterator P = MBB->begin();
363 while (P->getOpcode() == TargetInstrInfo::PHI) {
364 LiveVariables::VarInfo& PHIInfo = LV.getVarInfo(P->getOperand(0).getReg());
366 unsigned DestReg = P->getOperand(0).getReg();
368 // Hold the names that are currently in the candidate set.
369 std::set<unsigned> PHIUnion;
370 std::set<MachineBasicBlock*> UnionedBlocks;
372 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
373 unsigned SrcReg = P->getOperand(i-1).getReg();
374 LiveVariables::VarInfo& SrcInfo = LV.getVarInfo(SrcReg);
376 // Check for trivial interferences
377 if (isLiveIn(SrcInfo, P->getParent()) ||
378 isLiveOut(PHIInfo, SrcInfo.DefInst->getParent()) ||
379 ( PHIInfo.DefInst->getOpcode() == TargetInstrInfo::PHI &&
380 isLiveIn(PHIInfo, SrcInfo.DefInst->getParent()) ) ||
381 ProcessedNames.count(SrcReg) ||
382 UnionedBlocks.count(SrcInfo.DefInst->getParent())) {
384 // add a copy from a_i to p in Waiting[From[a_i]]
385 MachineBasicBlock* From = P->getOperand(i).getMBB();
386 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
387 UsedByAnother.insert(SrcReg);
389 PHIUnion.insert(SrcReg);
390 UnionedBlocks.insert(SrcInfo.DefInst->getParent());
394 std::vector<StrongPHIElimination::DomForestNode*> DF =
395 computeDomForest(PHIUnion);
397 // Walk DomForest to resolve interferences
398 std::vector<std::pair<unsigned, unsigned> > localInterferences;
399 processPHIUnion(P, PHIUnion, DF, localInterferences);
401 // Check for local interferences
402 for (std::vector<std::pair<unsigned, unsigned> >::iterator I =
403 localInterferences.begin(), E = localInterferences.end(); I != E; ++I) {
404 std::pair<unsigned, unsigned> p = *I;
406 LiveVariables::VarInfo& FirstInfo = LV.getVarInfo(p.first);
407 LiveVariables::VarInfo& SecondInfo = LV.getVarInfo(p.second);
409 MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
411 // Determine the block we need to scan and the relationship between
413 MachineBasicBlock* scan = 0;
415 if (FirstInfo.DefInst->getParent() == SecondInfo.DefInst->getParent()) {
416 scan = FirstInfo.DefInst->getParent();
417 mode = 0; // Same block
418 } else if (MDT.dominates(FirstInfo.DefInst->getParent(),
419 SecondInfo.DefInst->getParent())) {
420 scan = SecondInfo.DefInst->getParent();
421 mode = 1; // First dominates second
423 scan = FirstInfo.DefInst->getParent();
424 mode = 2; // Second dominates first
427 // If there's an interference, we need to insert copies
428 if (interferes(FirstInfo, SecondInfo, scan, mode)) {
429 // Insert copies for First
430 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
431 if (P->getOperand(i-1).getReg() == p.first) {
432 unsigned SrcReg = p.first;
433 MachineBasicBlock* From = P->getOperand(i).getMBB();
435 Waiting[From].insert(std::make_pair(SrcReg,
436 P->getOperand(0).getReg()));
437 UsedByAnother.insert(SrcReg);
439 PHIUnion.erase(SrcReg);
445 // FIXME: Cache renaming information
447 ProcessedNames.insert(PHIUnion.begin(), PHIUnion.end());
452 /// processPHIUnion - Take a set of candidate registers to be coallesced when
453 /// decomposing the PHI instruction. Use the DominanceForest to remove the ones
454 /// that are known to interfere, and flag others that need to be checked for
455 /// local interferences.
456 void StrongPHIElimination::processPHIUnion(MachineInstr* Inst,
457 std::set<unsigned>& PHIUnion,
458 std::vector<StrongPHIElimination::DomForestNode*>& DF,
459 std::vector<std::pair<unsigned, unsigned> >& locals) {
461 std::vector<DomForestNode*> worklist(DF.begin(), DF.end());
462 SmallPtrSet<DomForestNode*, 4> visited;
464 LiveVariables& LV = getAnalysis<LiveVariables>();
465 unsigned DestReg = Inst->getOperand(0).getReg();
467 // DF walk on the DomForest
468 while (!worklist.empty()) {
469 DomForestNode* DFNode = worklist.back();
471 LiveVariables::VarInfo& Info = LV.getVarInfo(DFNode->getReg());
472 visited.insert(DFNode);
474 bool inserted = false;
475 for (DomForestNode::iterator CI = DFNode->begin(), CE = DFNode->end();
477 DomForestNode* child = *CI;
478 LiveVariables::VarInfo& CInfo = LV.getVarInfo(child->getReg());
480 if (isLiveOut(Info, CInfo.DefInst->getParent())) {
481 // Insert copies for parent
482 for (int i = Inst->getNumOperands() - 1; i >= 2; i-=2) {
483 if (Inst->getOperand(i-1).getReg() == DFNode->getReg()) {
484 unsigned SrcReg = DFNode->getReg();
485 MachineBasicBlock* From = Inst->getOperand(i).getMBB();
487 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
488 UsedByAnother.insert(SrcReg);
490 PHIUnion.erase(SrcReg);
493 } else if (isLiveIn(Info, CInfo.DefInst->getParent()) ||
494 Info.DefInst->getParent() == CInfo.DefInst->getParent()) {
495 // Add (p, c) to possible local interferences
496 locals.push_back(std::make_pair(DFNode->getReg(), child->getReg()));
499 if (!visited.count(child)) {
500 worklist.push_back(child);
505 if (!inserted) worklist.pop_back();
509 /// ScheduleCopies - Insert copies into predecessor blocks, scheduling
510 /// them properly so as to avoid the 'lost copy' and the 'virtual swap'
513 /// Based on "Practical Improvements to the Construction and Destruction
514 /// of Static Single Assignment Form" by Briggs, et al.
515 void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
516 std::set<unsigned>& pushed) {
517 // FIXME: This function needs to update LiveVariables
518 std::map<unsigned, unsigned>& copy_set= Waiting[MBB];
520 std::map<unsigned, unsigned> worklist;
521 std::map<unsigned, unsigned> map;
523 // Setup worklist of initial copies
524 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
525 E = copy_set.end(); I != E; ) {
526 map.insert(std::make_pair(I->first, I->first));
527 map.insert(std::make_pair(I->second, I->second));
529 if (!UsedByAnother.count(I->first)) {
532 // Avoid iterator invalidation
533 unsigned first = I->first;
535 copy_set.erase(first);
541 LiveVariables& LV = getAnalysis<LiveVariables>();
542 MachineFunction* MF = MBB->getParent();
543 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
545 // Iterate over the worklist, inserting copies
546 while (!worklist.empty() || !copy_set.empty()) {
547 while (!worklist.empty()) {
548 std::pair<unsigned, unsigned> curr = *worklist.begin();
549 worklist.erase(curr.first);
551 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(curr.first);
553 if (isLiveOut(LV.getVarInfo(curr.second), MBB)) {
554 // Create a temporary
555 unsigned t = MF->getRegInfo().createVirtualRegister(RC);
557 // Insert copy from curr.second to a temporary at
558 // the Phi defining curr.second
559 LiveVariables::VarInfo VI = LV.getVarInfo(curr.second);
560 MachineBasicBlock::iterator PI = VI.DefInst;
561 TII->copyRegToReg(*VI.DefInst->getParent(), PI, t,
562 curr.second, RC, RC);
564 // Push temporary on Stacks
565 Stacks[curr.second].push_back(t);
567 // Insert curr.second in pushed
568 pushed.insert(curr.second);
571 // Insert copy from map[curr.first] to curr.second
572 TII->copyRegToReg(*MBB, MBB->end(), curr.second,
573 map[curr.first], RC, RC);
574 map[curr.first] = curr.second;
576 // If curr.first is a destination in copy_set...
577 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
578 E = copy_set.end(); I != E; )
579 if (curr.first == I->second) {
580 std::pair<unsigned, unsigned> temp = *I;
582 // Avoid iterator invalidation
584 copy_set.erase(temp.first);
585 worklist.insert(temp);
593 if (!copy_set.empty()) {
594 std::pair<unsigned, unsigned> curr = *copy_set.begin();
595 copy_set.erase(curr.first);
597 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(curr.first);
599 // Insert a copy from dest to a new temporary t at the end of b
600 unsigned t = MF->getRegInfo().createVirtualRegister(RC);
601 TII->copyRegToReg(*MBB, MBB->end(), t,
602 curr.second, RC, RC);
603 map[curr.second] = t;
605 worklist.insert(curr);
610 /// InsertCopies - insert copies into MBB and all of its successors
611 void StrongPHIElimination::InsertCopies(MachineBasicBlock* MBB) {
612 std::set<unsigned> pushed;
614 // Rewrite register uses from Stacks
615 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
617 for (unsigned i = 0; i < I->getNumOperands(); ++i)
618 if (I->getOperand(i).isRegister() &&
619 Stacks[I->getOperand(i).getReg()].size()) {
620 I->getOperand(i).setReg(Stacks[I->getOperand(i).getReg()].back());
623 // Schedule the copies for this block
624 ScheduleCopies(MBB, pushed);
626 // Recur to our successors
627 for (GraphTraits<MachineBasicBlock*>::ChildIteratorType I =
628 GraphTraits<MachineBasicBlock*>::child_begin(MBB), E =
629 GraphTraits<MachineBasicBlock*>::child_end(MBB); I != E; ++I)
632 // As we exit this block, pop the names we pushed while processing it
633 for (std::set<unsigned>::iterator I = pushed.begin(),
634 E = pushed.end(); I != E; ++I)
635 Stacks[*I].pop_back();
638 bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
639 // Compute DFS numbers of each block
642 // Determine which phi node operands need copies
643 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
645 I->begin()->getOpcode() == TargetInstrInfo::PHI)
649 // FIXME: This process should probably preserve LiveVariables
650 InsertCopies(Fn.begin());
652 // FIXME: Perform renaming
653 // FIXME: Remove Phi instrs