1 //===- StrongPhiElimination.cpp - Eliminate PHI nodes by inserting copies -===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Owen Anderson and is distributed under
6 // the University of Illinois Open Source 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/SSARegMap.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.addPreserved<LiveVariables>();
52 AU.addPreservedID(PHIEliminationID);
53 AU.addRequired<MachineDominatorTree>();
54 AU.addRequired<LiveVariables>();
56 MachineFunctionPass::getAnalysisUsage(AU);
59 virtual void releaseMemory() {
67 struct DomForestNode {
69 std::vector<DomForestNode*> children;
72 void addChild(DomForestNode* DFN) { children.push_back(DFN); }
75 typedef std::vector<DomForestNode*>::iterator iterator;
77 DomForestNode(unsigned r, DomForestNode* parent) : reg(r) {
79 parent->addChild(this);
83 for (iterator I = begin(), E = end(); I != E; ++I)
87 inline unsigned getReg() { return reg; }
89 inline DomForestNode::iterator begin() { return children.begin(); }
90 inline DomForestNode::iterator end() { return children.end(); }
93 DenseMap<MachineBasicBlock*, unsigned> preorder;
94 DenseMap<MachineBasicBlock*, unsigned> maxpreorder;
97 void computeDFS(MachineFunction& MF);
98 void processBlock(MachineBasicBlock* MBB);
100 std::vector<DomForestNode*> computeDomForest(std::set<unsigned>& instrs);
101 void processPHIUnion(MachineInstr* Inst,
102 std::set<unsigned>& PHIUnion,
103 std::vector<StrongPHIElimination::DomForestNode*>& DF,
104 std::vector<std::pair<unsigned, unsigned> >& locals);
105 void ScheduleCopies(MachineBasicBlock* MBB, std::set<unsigned>& pushed);
106 void InsertCopies(MachineBasicBlock* MBB);
109 char StrongPHIElimination::ID = 0;
110 RegisterPass<StrongPHIElimination> X("strong-phi-node-elimination",
111 "Eliminate PHI nodes for register allocation, intelligently");
114 const PassInfo *llvm::StrongPHIEliminationID = X.getPassInfo();
116 /// computeDFS - Computes the DFS-in and DFS-out numbers of the dominator tree
117 /// of the given MachineFunction. These numbers are then used in other parts
118 /// of the PHI elimination process.
119 void StrongPHIElimination::computeDFS(MachineFunction& MF) {
120 SmallPtrSet<MachineDomTreeNode*, 8> frontier;
121 SmallPtrSet<MachineDomTreeNode*, 8> visited;
125 MachineDominatorTree& DT = getAnalysis<MachineDominatorTree>();
127 MachineDomTreeNode* node = DT.getRootNode();
129 std::vector<MachineDomTreeNode*> worklist;
130 worklist.push_back(node);
132 while (!worklist.empty()) {
133 MachineDomTreeNode* currNode = worklist.back();
135 if (!frontier.count(currNode)) {
136 frontier.insert(currNode);
138 preorder.insert(std::make_pair(currNode->getBlock(), time));
141 bool inserted = false;
142 for (MachineDomTreeNode::iterator I = node->begin(), E = node->end();
144 if (!frontier.count(*I) && !visited.count(*I)) {
145 worklist.push_back(*I);
151 frontier.erase(currNode);
152 visited.insert(currNode);
153 maxpreorder.insert(std::make_pair(currNode->getBlock(), time));
160 /// PreorderSorter - a helper class that is used to sort registers
161 /// according to the preorder number of their defining blocks
162 class PreorderSorter {
164 DenseMap<MachineBasicBlock*, unsigned>& preorder;
168 PreorderSorter(DenseMap<MachineBasicBlock*, unsigned>& p,
169 LiveVariables& L) : preorder(p), LV(L) { }
171 bool operator()(unsigned A, unsigned B) {
175 MachineBasicBlock* ABlock = LV.getVarInfo(A).DefInst->getParent();
176 MachineBasicBlock* BBlock = LV.getVarInfo(A).DefInst->getParent();
178 if (preorder[ABlock] < preorder[BBlock])
180 else if (preorder[ABlock] > preorder[BBlock])
183 assert(0 && "Error sorting by dominance!");
188 /// computeDomForest - compute the subforest of the DomTree corresponding
189 /// to the defining blocks of the registers in question
190 std::vector<StrongPHIElimination::DomForestNode*>
191 StrongPHIElimination::computeDomForest(std::set<unsigned>& regs) {
192 LiveVariables& LV = getAnalysis<LiveVariables>();
194 DomForestNode* VirtualRoot = new DomForestNode(0, 0);
195 maxpreorder.insert(std::make_pair((MachineBasicBlock*)0, ~0UL));
197 std::vector<unsigned> worklist;
198 worklist.reserve(regs.size());
199 for (std::set<unsigned>::iterator I = regs.begin(), E = regs.end();
201 worklist.push_back(*I);
203 PreorderSorter PS(preorder, LV);
204 std::sort(worklist.begin(), worklist.end(), PS);
206 DomForestNode* CurrentParent = VirtualRoot;
207 std::vector<DomForestNode*> stack;
208 stack.push_back(VirtualRoot);
210 for (std::vector<unsigned>::iterator I = worklist.begin(), E = worklist.end();
212 unsigned pre = preorder[LV.getVarInfo(*I).DefInst->getParent()];
213 MachineBasicBlock* parentBlock =
214 LV.getVarInfo(CurrentParent->getReg()).DefInst->getParent();
216 while (pre > maxpreorder[parentBlock]) {
218 CurrentParent = stack.back();
220 parentBlock = LV.getVarInfo(CurrentParent->getReg()).DefInst->getParent();
223 DomForestNode* child = new DomForestNode(*I, CurrentParent);
224 stack.push_back(child);
225 CurrentParent = child;
228 std::vector<DomForestNode*> ret;
229 ret.insert(ret.end(), VirtualRoot->begin(), VirtualRoot->end());
233 /// isLiveIn - helper method that determines, from a VarInfo, if a register
234 /// is live into a block
235 bool isLiveIn(LiveVariables::VarInfo& V, MachineBasicBlock* MBB) {
236 if (V.AliveBlocks.test(MBB->getNumber()))
239 if (V.DefInst->getParent() != MBB &&
240 V.UsedBlocks.test(MBB->getNumber()))
246 /// isLiveOut - help method that determines, from a VarInfo, if a register is
247 /// live out of a block.
248 bool isLiveOut(LiveVariables::VarInfo& V, MachineBasicBlock* MBB) {
249 if (MBB == V.DefInst->getParent() ||
250 V.UsedBlocks.test(MBB->getNumber())) {
251 for (std::vector<MachineInstr*>::iterator I = V.Kills.begin(),
252 E = V.Kills.end(); I != E; ++I)
253 if ((*I)->getParent() == MBB)
262 /// isKillInst - helper method that determines, from a VarInfo, if an
263 /// instruction kills a given register
264 bool isKillInst(LiveVariables::VarInfo& V, MachineInstr* MI) {
265 return std::find(V.Kills.begin(), V.Kills.end(), MI) != V.Kills.end();
268 /// interferes - checks for local interferences by scanning a block. The only
269 /// trick parameter is 'mode' which tells it the relationship of the two
270 /// registers. 0 - defined in the same block, 1 - first properly dominates
271 /// second, 2 - second properly dominates first
272 bool interferes(LiveVariables::VarInfo& First, LiveVariables::VarInfo& Second,
273 MachineBasicBlock* scan, unsigned mode) {
274 MachineInstr* def = 0;
275 MachineInstr* kill = 0;
277 bool interference = false;
279 // Wallk the block, checking for interferences
280 for (MachineBasicBlock::iterator MBI = scan->begin(), MBE = scan->end();
282 MachineInstr* curr = MBI;
284 // Same defining block...
286 if (curr == First.DefInst) {
287 // If we find our first DefInst, save it
290 // If there's already an unkilled DefInst, then
291 // this is an interference
295 // If there's a DefInst followed by a KillInst, then
296 // they can't interfere
298 interference = false;
301 // Symmetric with the above
302 } else if (curr == Second.DefInst ) {
309 interference = false;
312 // Store KillInsts if they match up with the DefInst
313 } else if (isKillInst(First, curr)) {
314 if (def == First.DefInst) {
316 } else if (isKillInst(Second, curr)) {
317 if (def == Second.DefInst) {
322 // First properly dominates second...
323 } else if (mode == 1) {
324 if (curr == Second.DefInst) {
325 // DefInst of second without kill of first is an interference
329 // DefInst after a kill is a non-interference
331 interference = false;
334 // Save KillInsts of First
335 } else if (isKillInst(First, curr)) {
338 // Symmetric with the above
339 } else if (mode == 2) {
340 if (curr == First.DefInst) {
345 interference = false;
348 } else if (isKillInst(Second, curr)) {
357 /// processBlock - Eliminate PHIs in the given block
358 void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
359 LiveVariables& LV = getAnalysis<LiveVariables>();
361 // Holds names that have been added to a set in any PHI within this block
362 // before the current one.
363 std::set<unsigned> ProcessedNames;
365 MachineBasicBlock::iterator P = MBB->begin();
366 while (P->getOpcode() == TargetInstrInfo::PHI) {
367 LiveVariables::VarInfo& PHIInfo = LV.getVarInfo(P->getOperand(0).getReg());
369 unsigned DestReg = P->getOperand(0).getReg();
371 // Hold the names that are currently in the candidate set.
372 std::set<unsigned> PHIUnion;
373 std::set<MachineBasicBlock*> UnionedBlocks;
375 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
376 unsigned SrcReg = P->getOperand(i-1).getReg();
377 LiveVariables::VarInfo& SrcInfo = LV.getVarInfo(SrcReg);
379 // Check for trivial interferences
380 if (isLiveIn(SrcInfo, P->getParent()) ||
381 isLiveOut(PHIInfo, SrcInfo.DefInst->getParent()) ||
382 ( PHIInfo.DefInst->getOpcode() == TargetInstrInfo::PHI &&
383 isLiveIn(PHIInfo, SrcInfo.DefInst->getParent()) ) ||
384 ProcessedNames.count(SrcReg) ||
385 UnionedBlocks.count(SrcInfo.DefInst->getParent())) {
387 // add a copy from a_i to p in Waiting[From[a_i]]
388 MachineBasicBlock* From = P->getOperand(i).getMachineBasicBlock();
389 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
390 UsedByAnother.insert(SrcReg);
392 PHIUnion.insert(SrcReg);
393 UnionedBlocks.insert(SrcInfo.DefInst->getParent());
397 std::vector<StrongPHIElimination::DomForestNode*> DF =
398 computeDomForest(PHIUnion);
400 // Walk DomForest to resolve interferences
401 std::vector<std::pair<unsigned, unsigned> > localInterferences;
402 processPHIUnion(P, PHIUnion, DF, localInterferences);
404 // Check for local interferences
405 for (std::vector<std::pair<unsigned, unsigned> >::iterator I =
406 localInterferences.begin(), E = localInterferences.end(); I != E; ++I) {
407 std::pair<unsigned, unsigned> p = *I;
409 LiveVariables::VarInfo& FirstInfo = LV.getVarInfo(p.first);
410 LiveVariables::VarInfo& SecondInfo = LV.getVarInfo(p.second);
412 MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
414 // Determine the block we need to scan and the relationship between
416 MachineBasicBlock* scan = 0;
418 if (FirstInfo.DefInst->getParent() == SecondInfo.DefInst->getParent()) {
419 scan = FirstInfo.DefInst->getParent();
420 mode = 0; // Same block
421 } else if (MDT.dominates(FirstInfo.DefInst->getParent(),
422 SecondInfo.DefInst->getParent())) {
423 scan = SecondInfo.DefInst->getParent();
424 mode = 1; // First dominates second
426 scan = FirstInfo.DefInst->getParent();
427 mode = 2; // Second dominates first
430 // If there's an interference, we need to insert copies
431 if (interferes(FirstInfo, SecondInfo, scan, mode)) {
432 // Insert copies for First
433 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
434 if (P->getOperand(i-1).getReg() == p.first) {
435 unsigned SrcReg = p.first;
436 MachineBasicBlock* From = P->getOperand(i).getMBB();
438 Waiting[From].insert(std::make_pair(SrcReg,
439 P->getOperand(0).getReg()));
440 UsedByAnother.insert(SrcReg);
442 PHIUnion.erase(SrcReg);
448 // FIXME: Cache renaming information
450 ProcessedNames.insert(PHIUnion.begin(), PHIUnion.end());
455 /// processPHIUnion - Take a set of candidate registers to be coallesced when
456 /// decomposing the PHI instruction. Use the DominanceForest to remove the ones
457 /// that are known to interfere, and flag others that need to be checked for
458 /// local interferences.
459 void StrongPHIElimination::processPHIUnion(MachineInstr* Inst,
460 std::set<unsigned>& PHIUnion,
461 std::vector<StrongPHIElimination::DomForestNode*>& DF,
462 std::vector<std::pair<unsigned, unsigned> >& locals) {
464 std::vector<DomForestNode*> worklist(DF.begin(), DF.end());
465 SmallPtrSet<DomForestNode*, 4> visited;
467 LiveVariables& LV = getAnalysis<LiveVariables>();
468 unsigned DestReg = Inst->getOperand(0).getReg();
470 // DF walk on the DomForest
471 while (!worklist.empty()) {
472 DomForestNode* DFNode = worklist.back();
474 LiveVariables::VarInfo& Info = LV.getVarInfo(DFNode->getReg());
475 visited.insert(DFNode);
477 bool inserted = false;
478 for (DomForestNode::iterator CI = DFNode->begin(), CE = DFNode->end();
480 DomForestNode* child = *CI;
481 LiveVariables::VarInfo& CInfo = LV.getVarInfo(child->getReg());
483 if (isLiveOut(Info, CInfo.DefInst->getParent())) {
484 // Insert copies for parent
485 for (int i = Inst->getNumOperands() - 1; i >= 2; i-=2) {
486 if (Inst->getOperand(i-1).getReg() == DFNode->getReg()) {
487 unsigned SrcReg = DFNode->getReg();
488 MachineBasicBlock* From = Inst->getOperand(i).getMBB();
490 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
491 UsedByAnother.insert(SrcReg);
493 PHIUnion.erase(SrcReg);
496 } else if (isLiveIn(Info, CInfo.DefInst->getParent()) ||
497 Info.DefInst->getParent() == CInfo.DefInst->getParent()) {
498 // Add (p, c) to possible local interferences
499 locals.push_back(std::make_pair(DFNode->getReg(), child->getReg()));
502 if (!visited.count(child)) {
503 worklist.push_back(child);
508 if (!inserted) worklist.pop_back();
512 /// ScheduleCopies - Insert copies into predecessor blocks, scheduling
513 /// them properly so as to avoid the 'lost copy' and the 'virtual swap'
516 /// Based on "Practical Improvements to the Construction and Destruction
517 /// of Static Single Assignment Form" by Briggs, et al.
518 void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
519 std::set<unsigned>& pushed) {
520 std::map<unsigned, unsigned>& copy_set= Waiting[MBB];
522 std::map<unsigned, unsigned> worklist;
523 std::map<unsigned, unsigned> map;
525 // Setup worklist of initial copies
526 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
527 E = copy_set.end(); I != E; ) {
528 map.insert(std::make_pair(I->first, I->first));
529 map.insert(std::make_pair(I->second, I->second));
531 if (!UsedByAnother.count(I->first)) {
534 // Avoid iterator invalidation
535 unsigned first = I->first;
537 copy_set.erase(first);
543 LiveVariables& LV = getAnalysis<LiveVariables>();
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 if (isLiveOut(LV.getVarInfo(curr.second), MBB)) {
552 // Insert copy from curr.second to a temporary
553 // Push temporary on Stacks
554 // Insert temporary in pushed
557 // Insert copy from map[curr.first] to curr.second
558 map[curr.first] = curr.second;
560 // If curr.first is a destination in copy_set...
561 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
562 E = copy_set.end(); I != E; )
563 if (curr.first == I->second) {
564 std::pair<unsigned, unsigned> temp = *I;
566 // Avoid iterator invalidation
568 copy_set.erase(temp.first);
569 worklist.insert(temp);
577 if (!copy_set.empty()) {
578 std::pair<unsigned, unsigned> curr = *copy_set.begin();
579 copy_set.erase(curr.first);
581 // Insert a copy from dest to a new temporary t at the end of b
582 // map[curr.second] = t;
584 worklist.insert(curr);
589 /// InsertCopies - insert copies into MBB and all of its successors
590 void StrongPHIElimination::InsertCopies(MachineBasicBlock* MBB) {
591 std::set<unsigned> pushed;
593 // Rewrite register uses from Stacks
594 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
596 for (unsigned i = 0; i < I->getNumOperands(); ++i)
597 if (I->getOperand(i).isRegister() &&
598 Stacks[I->getOperand(i).getReg()].size()) {
599 I->getOperand(i).setReg(Stacks[I->getOperand(i).getReg()].back());
602 // Schedule the copies for this block
603 ScheduleCopies(MBB, pushed);
605 // Recur to our successors
606 for (GraphTraits<MachineBasicBlock*>::ChildIteratorType I =
607 GraphTraits<MachineBasicBlock*>::child_begin(MBB), E =
608 GraphTraits<MachineBasicBlock*>::child_end(MBB); I != E; ++I)
611 // As we exit this block, pop the names we pushed while processing it
612 for (std::set<unsigned>::iterator I = pushed.begin(),
613 E = pushed.end(); I != E; ++I)
614 Stacks[*I].pop_back();
617 bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
618 // Compute DFS numbers of each block
621 // Determine which phi node operands need copies
622 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
624 I->begin()->getOpcode() == TargetInstrInfo::PHI)
628 // FIXME: This process should probably preserve LiveVariables
629 InsertCopies(Fn.begin());
631 // FIXME: Perform renaming