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/LiveIntervalAnalysis.h"
25 #include "llvm/CodeGen/MachineDominators.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineLoopInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/Target/TargetInstrInfo.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/ADT/DepthFirstIterator.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Support/Compiler.h"
39 struct VISIBILITY_HIDDEN StrongPHIElimination : public MachineFunctionPass {
40 static char ID; // Pass identification, replacement for typeid
41 StrongPHIElimination() : MachineFunctionPass((intptr_t)&ID) {}
43 // Waiting stores, for each MBB, the set of copies that need to
44 // be inserted into that MBB
45 DenseMap<MachineBasicBlock*,
46 std::map<unsigned, unsigned> > Waiting;
48 // Stacks holds the renaming stack for each register
49 std::map<unsigned, std::vector<unsigned> > Stacks;
51 // Registers in UsedByAnother are PHI nodes that are themselves
52 // used as operands to another another PHI node
53 std::set<unsigned> UsedByAnother;
55 // RenameSets are the sets of operands (and their VNInfo IDs) to a PHI
56 // (the defining instruction of the key) that can be renamed without copies.
57 std::map<unsigned, std::map<unsigned, unsigned> > RenameSets;
59 // PhiValueNumber holds the ID numbers of the VNs for each phi that we're
60 // eliminating, indexed by the register defined by that phi.
61 std::map<unsigned, unsigned> PhiValueNumber;
63 // Store the DFS-in number of each block
64 DenseMap<MachineBasicBlock*, unsigned> preorder;
66 // Store the DFS-out number of each block
67 DenseMap<MachineBasicBlock*, unsigned> maxpreorder;
69 bool runOnMachineFunction(MachineFunction &Fn);
71 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
72 AU.addRequired<MachineDominatorTree>();
73 AU.addRequired<LiveIntervals>();
75 // TODO: Actually make this true.
76 AU.addPreserved<LiveIntervals>();
77 MachineFunctionPass::getAnalysisUsage(AU);
80 virtual void releaseMemory() {
86 UsedByAnother.clear();
92 /// DomForestNode - Represents a node in the "dominator forest". This is
93 /// a forest in which the nodes represent registers and the edges
94 /// represent a dominance relation in the block defining those registers.
95 struct DomForestNode {
97 // Store references to our children
98 std::vector<DomForestNode*> children;
99 // The register we represent
102 // Add another node as our child
103 void addChild(DomForestNode* DFN) { children.push_back(DFN); }
106 typedef std::vector<DomForestNode*>::iterator iterator;
108 // Create a DomForestNode by providing the register it represents, and
109 // the node to be its parent. The virtual root node has register 0
110 // and a null parent.
111 DomForestNode(unsigned r, DomForestNode* parent) : reg(r) {
113 parent->addChild(this);
117 for (iterator I = begin(), E = end(); I != E; ++I)
121 /// getReg - Return the regiser that this node represents
122 inline unsigned getReg() { return reg; }
124 // Provide iterator access to our children
125 inline DomForestNode::iterator begin() { return children.begin(); }
126 inline DomForestNode::iterator end() { return children.end(); }
129 void computeDFS(MachineFunction& MF);
130 void processBlock(MachineBasicBlock* MBB);
132 std::vector<DomForestNode*> computeDomForest(std::map<unsigned, unsigned>& instrs,
133 MachineRegisterInfo& MRI);
134 void processPHIUnion(MachineInstr* Inst,
135 std::map<unsigned, unsigned>& PHIUnion,
136 std::vector<StrongPHIElimination::DomForestNode*>& DF,
137 std::vector<std::pair<unsigned, unsigned> >& locals);
138 void ScheduleCopies(MachineBasicBlock* MBB, std::set<unsigned>& pushed);
139 void InsertCopies(MachineBasicBlock* MBB,
140 SmallPtrSet<MachineBasicBlock*, 16>& v);
141 void mergeLiveIntervals(unsigned primary, unsigned secondary, unsigned VN);
145 char StrongPHIElimination::ID = 0;
146 static RegisterPass<StrongPHIElimination>
147 X("strong-phi-node-elimination",
148 "Eliminate PHI nodes for register allocation, intelligently");
150 const PassInfo *const llvm::StrongPHIEliminationID = &X;
152 /// computeDFS - Computes the DFS-in and DFS-out numbers of the dominator tree
153 /// of the given MachineFunction. These numbers are then used in other parts
154 /// of the PHI elimination process.
155 void StrongPHIElimination::computeDFS(MachineFunction& MF) {
156 SmallPtrSet<MachineDomTreeNode*, 8> frontier;
157 SmallPtrSet<MachineDomTreeNode*, 8> visited;
161 MachineDominatorTree& DT = getAnalysis<MachineDominatorTree>();
163 MachineDomTreeNode* node = DT.getRootNode();
165 std::vector<MachineDomTreeNode*> worklist;
166 worklist.push_back(node);
168 while (!worklist.empty()) {
169 MachineDomTreeNode* currNode = worklist.back();
171 if (!frontier.count(currNode)) {
172 frontier.insert(currNode);
174 preorder.insert(std::make_pair(currNode->getBlock(), time));
177 bool inserted = false;
178 for (MachineDomTreeNode::iterator I = currNode->begin(), E = currNode->end();
180 if (!frontier.count(*I) && !visited.count(*I)) {
181 worklist.push_back(*I);
187 frontier.erase(currNode);
188 visited.insert(currNode);
189 maxpreorder.insert(std::make_pair(currNode->getBlock(), time));
198 /// PreorderSorter - a helper class that is used to sort registers
199 /// according to the preorder number of their defining blocks
200 class PreorderSorter {
202 DenseMap<MachineBasicBlock*, unsigned>& preorder;
203 MachineRegisterInfo& MRI;
206 PreorderSorter(DenseMap<MachineBasicBlock*, unsigned>& p,
207 MachineRegisterInfo& M) : preorder(p), MRI(M) { }
209 bool operator()(unsigned A, unsigned B) {
213 MachineBasicBlock* ABlock = MRI.getVRegDef(A)->getParent();
214 MachineBasicBlock* BBlock = MRI.getVRegDef(B)->getParent();
216 if (preorder[ABlock] < preorder[BBlock])
218 else if (preorder[ABlock] > preorder[BBlock])
227 /// computeDomForest - compute the subforest of the DomTree corresponding
228 /// to the defining blocks of the registers in question
229 std::vector<StrongPHIElimination::DomForestNode*>
230 StrongPHIElimination::computeDomForest(std::map<unsigned, unsigned>& regs,
231 MachineRegisterInfo& MRI) {
232 // Begin by creating a virtual root node, since the actual results
233 // may well be a forest. Assume this node has maximum DFS-out number.
234 DomForestNode* VirtualRoot = new DomForestNode(0, 0);
235 maxpreorder.insert(std::make_pair((MachineBasicBlock*)0, ~0UL));
237 // Populate a worklist with the registers
238 std::vector<unsigned> worklist;
239 worklist.reserve(regs.size());
240 for (std::map<unsigned, unsigned>::iterator I = regs.begin(), E = regs.end();
242 worklist.push_back(I->first);
244 // Sort the registers by the DFS-in number of their defining block
245 PreorderSorter PS(preorder, MRI);
246 std::sort(worklist.begin(), worklist.end(), PS);
248 // Create a "current parent" stack, and put the virtual root on top of it
249 DomForestNode* CurrentParent = VirtualRoot;
250 std::vector<DomForestNode*> stack;
251 stack.push_back(VirtualRoot);
253 // Iterate over all the registers in the previously computed order
254 for (std::vector<unsigned>::iterator I = worklist.begin(), E = worklist.end();
256 unsigned pre = preorder[MRI.getVRegDef(*I)->getParent()];
257 MachineBasicBlock* parentBlock = CurrentParent->getReg() ?
258 MRI.getVRegDef(CurrentParent->getReg())->getParent() :
261 // If the DFS-in number of the register is greater than the DFS-out number
262 // of the current parent, repeatedly pop the parent stack until it isn't.
263 while (pre > maxpreorder[parentBlock]) {
265 CurrentParent = stack.back();
267 parentBlock = CurrentParent->getReg() ?
268 MRI.getVRegDef(CurrentParent->getReg())->getParent() :
272 // Now that we've found the appropriate parent, create a DomForestNode for
273 // this register and attach it to the forest
274 DomForestNode* child = new DomForestNode(*I, CurrentParent);
276 // Push this new node on the "current parent" stack
277 stack.push_back(child);
278 CurrentParent = child;
281 // Return a vector containing the children of the virtual root node
282 std::vector<DomForestNode*> ret;
283 ret.insert(ret.end(), VirtualRoot->begin(), VirtualRoot->end());
287 /// isLiveIn - helper method that determines, from a regno, if a register
288 /// is live into a block
289 static bool isLiveIn(unsigned r, MachineBasicBlock* MBB,
291 LiveInterval& I = LI.getOrCreateInterval(r);
292 unsigned idx = LI.getMBBStartIdx(MBB);
293 return I.liveBeforeAndAt(idx);
296 /// isLiveOut - help method that determines, from a regno, if a register is
297 /// live out of a block.
298 static bool isLiveOut(unsigned r, MachineBasicBlock* MBB,
300 for (MachineBasicBlock::succ_iterator PI = MBB->succ_begin(),
301 E = MBB->succ_end(); PI != E; ++PI) {
302 if (isLiveIn(r, *PI, LI))
309 /// interferes - checks for local interferences by scanning a block. The only
310 /// trick parameter is 'mode' which tells it the relationship of the two
311 /// registers. 0 - defined in the same block, 1 - first properly dominates
312 /// second, 2 - second properly dominates first
313 static bool interferes(unsigned a, unsigned b, MachineBasicBlock* scan,
314 LiveIntervals& LV, unsigned mode) {
315 MachineInstr* def = 0;
316 MachineInstr* kill = 0;
318 // The code is still in SSA form at this point, so there is only one
319 // definition per VReg. Thus we can safely use MRI->getVRegDef().
320 const MachineRegisterInfo* MRI = &scan->getParent()->getRegInfo();
322 bool interference = false;
324 // Wallk the block, checking for interferences
325 for (MachineBasicBlock::iterator MBI = scan->begin(), MBE = scan->end();
327 MachineInstr* curr = MBI;
329 // Same defining block...
331 if (curr == MRI->getVRegDef(a)) {
332 // If we find our first definition, save it
335 // If there's already an unkilled definition, then
336 // this is an interference
340 // If there's a definition followed by a KillInst, then
341 // they can't interfere
343 interference = false;
346 // Symmetric with the above
347 } else if (curr == MRI->getVRegDef(b)) {
354 interference = false;
357 // Store KillInsts if they match up with the definition
358 } else if (curr->killsRegister(a)) {
359 if (def == MRI->getVRegDef(a)) {
361 } else if (curr->killsRegister(b)) {
362 if (def == MRI->getVRegDef(b)) {
367 // First properly dominates second...
368 } else if (mode == 1) {
369 if (curr == MRI->getVRegDef(b)) {
370 // Definition of second without kill of first is an interference
374 // Definition after a kill is a non-interference
376 interference = false;
379 // Save KillInsts of First
380 } else if (curr->killsRegister(a)) {
383 // Symmetric with the above
384 } else if (mode == 2) {
385 if (curr == MRI->getVRegDef(a)) {
390 interference = false;
393 } else if (curr->killsRegister(b)) {
402 /// processBlock - Determine how to break up PHIs in the current block. Each
403 /// PHI is broken up by some combination of renaming its operands and inserting
404 /// copies. This method is responsible for determining which operands receive
406 void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
407 LiveIntervals& LI = getAnalysis<LiveIntervals>();
408 MachineRegisterInfo& MRI = MBB->getParent()->getRegInfo();
410 // Holds names that have been added to a set in any PHI within this block
411 // before the current one.
412 std::set<unsigned> ProcessedNames;
414 // Iterate over all the PHI nodes in this block
415 MachineBasicBlock::iterator P = MBB->begin();
416 while (P != MBB->end() && P->getOpcode() == TargetInstrInfo::PHI) {
417 unsigned DestReg = P->getOperand(0).getReg();
419 // Don't both doing PHI elimination for dead PHI's.
420 if (P->registerDefIsDead(DestReg)) {
425 LiveInterval& PI = LI.getOrCreateInterval(DestReg);
426 unsigned pIdx = LI.getDefIndex(LI.getInstructionIndex(P));
427 VNInfo* PVN = PI.getLiveRangeContaining(pIdx)->valno;
428 PhiValueNumber.insert(std::make_pair(DestReg, PVN->id));
430 // PHIUnion is the set of incoming registers to the PHI node that
431 // are going to be renames rather than having copies inserted. This set
432 // is refinded over the course of this function. UnionedBlocks is the set
433 // of corresponding MBBs.
434 std::map<unsigned, unsigned> PHIUnion;
435 SmallPtrSet<MachineBasicBlock*, 8> UnionedBlocks;
437 // Iterate over the operands of the PHI node
438 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
439 unsigned SrcReg = P->getOperand(i-1).getReg();
441 // Check for trivial interferences via liveness information, allowing us
442 // to avoid extra work later. Any registers that interfere cannot both
443 // be in the renaming set, so choose one and add copies for it instead.
444 // The conditions are:
445 // 1) if the operand is live into the PHI node's block OR
446 // 2) if the PHI node is live out of the operand's defining block OR
447 // 3) if the operand is itself a PHI node and the original PHI is
448 // live into the operand's defining block OR
449 // 4) if the operand is already being renamed for another PHI node
451 // 5) if any two operands are defined in the same block, insert copies
453 if (isLiveIn(SrcReg, P->getParent(), LI) ||
454 isLiveOut(P->getOperand(0).getReg(),
455 MRI.getVRegDef(SrcReg)->getParent(), LI) ||
456 ( MRI.getVRegDef(SrcReg)->getOpcode() == TargetInstrInfo::PHI &&
457 isLiveIn(P->getOperand(0).getReg(),
458 MRI.getVRegDef(SrcReg)->getParent(), LI) ) ||
459 ProcessedNames.count(SrcReg) ||
460 UnionedBlocks.count(MRI.getVRegDef(SrcReg)->getParent())) {
462 // Add a copy for the selected register
463 MachineBasicBlock* From = P->getOperand(i).getMBB();
464 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
465 UsedByAnother.insert(SrcReg);
467 // Otherwise, add it to the renaming set
468 LiveInterval& I = LI.getOrCreateInterval(SrcReg);
469 unsigned idx = LI.getMBBEndIdx(P->getOperand(i).getMBB());
470 VNInfo* VN = I.getLiveRangeContaining(idx)->valno;
472 assert(VN && "No VNInfo for register?");
474 PHIUnion.insert(std::make_pair(SrcReg, VN->id));
475 UnionedBlocks.insert(MRI.getVRegDef(SrcReg)->getParent());
479 // Compute the dominator forest for the renaming set. This is a forest
480 // where the nodes are the registers and the edges represent dominance
481 // relations between the defining blocks of the registers
482 std::vector<StrongPHIElimination::DomForestNode*> DF =
483 computeDomForest(PHIUnion, MRI);
485 // Walk DomForest to resolve interferences at an inter-block level. This
486 // will remove registers from the renaming set (and insert copies for them)
487 // if interferences are found.
488 std::vector<std::pair<unsigned, unsigned> > localInterferences;
489 processPHIUnion(P, PHIUnion, DF, localInterferences);
491 // If one of the inputs is defined in the same block as the current PHI
492 // then we need to check for a local interference between that input and
494 for (std::map<unsigned, unsigned>::iterator I = PHIUnion.begin(),
495 E = PHIUnion.end(); I != E; ++I)
496 if (MRI.getVRegDef(I->first)->getParent() == P->getParent())
497 localInterferences.push_back(std::make_pair(I->first,
498 P->getOperand(0).getReg()));
500 // The dominator forest walk may have returned some register pairs whose
501 // interference cannot be determined from dominator analysis. We now
502 // examine these pairs for local interferences.
503 for (std::vector<std::pair<unsigned, unsigned> >::iterator I =
504 localInterferences.begin(), E = localInterferences.end(); I != E; ++I) {
505 std::pair<unsigned, unsigned> p = *I;
507 MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
509 // Determine the block we need to scan and the relationship between
511 MachineBasicBlock* scan = 0;
513 if (MRI.getVRegDef(p.first)->getParent() ==
514 MRI.getVRegDef(p.second)->getParent()) {
515 scan = MRI.getVRegDef(p.first)->getParent();
516 mode = 0; // Same block
517 } else if (MDT.dominates(MRI.getVRegDef(p.first)->getParent(),
518 MRI.getVRegDef(p.second)->getParent())) {
519 scan = MRI.getVRegDef(p.second)->getParent();
520 mode = 1; // First dominates second
522 scan = MRI.getVRegDef(p.first)->getParent();
523 mode = 2; // Second dominates first
526 // If there's an interference, we need to insert copies
527 if (interferes(p.first, p.second, scan, LI, mode)) {
528 // Insert copies for First
529 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
530 if (P->getOperand(i-1).getReg() == p.first) {
531 unsigned SrcReg = p.first;
532 MachineBasicBlock* From = P->getOperand(i).getMBB();
534 Waiting[From].insert(std::make_pair(SrcReg,
535 P->getOperand(0).getReg()));
536 UsedByAnother.insert(SrcReg);
538 PHIUnion.erase(SrcReg);
544 // Add the renaming set for this PHI node to our overall renaming information
545 RenameSets.insert(std::make_pair(P->getOperand(0).getReg(), PHIUnion));
547 // Remember which registers are already renamed, so that we don't try to
548 // rename them for another PHI node in this block
549 for (std::map<unsigned, unsigned>::iterator I = PHIUnion.begin(),
550 E = PHIUnion.end(); I != E; ++I)
551 ProcessedNames.insert(I->first);
557 /// processPHIUnion - Take a set of candidate registers to be coalesced when
558 /// decomposing the PHI instruction. Use the DominanceForest to remove the ones
559 /// that are known to interfere, and flag others that need to be checked for
560 /// local interferences.
561 void StrongPHIElimination::processPHIUnion(MachineInstr* Inst,
562 std::map<unsigned, unsigned>& PHIUnion,
563 std::vector<StrongPHIElimination::DomForestNode*>& DF,
564 std::vector<std::pair<unsigned, unsigned> >& locals) {
566 std::vector<DomForestNode*> worklist(DF.begin(), DF.end());
567 SmallPtrSet<DomForestNode*, 4> visited;
569 // Code is still in SSA form, so we can use MRI::getVRegDef()
570 MachineRegisterInfo& MRI = Inst->getParent()->getParent()->getRegInfo();
572 LiveIntervals& LI = getAnalysis<LiveIntervals>();
573 unsigned DestReg = Inst->getOperand(0).getReg();
575 // DF walk on the DomForest
576 while (!worklist.empty()) {
577 DomForestNode* DFNode = worklist.back();
579 visited.insert(DFNode);
581 bool inserted = false;
582 for (DomForestNode::iterator CI = DFNode->begin(), CE = DFNode->end();
584 DomForestNode* child = *CI;
586 // If the current node is live-out of the defining block of one of its
587 // children, insert a copy for it. NOTE: The paper actually calls for
588 // a more elaborate heuristic for determining whether to insert copies
589 // for the child or the parent. In the interest of simplicity, we're
590 // just always choosing the parent.
591 if (isLiveOut(DFNode->getReg(),
592 MRI.getVRegDef(child->getReg())->getParent(), LI)) {
593 // Insert copies for parent
594 for (int i = Inst->getNumOperands() - 1; i >= 2; i-=2) {
595 if (Inst->getOperand(i-1).getReg() == DFNode->getReg()) {
596 unsigned SrcReg = DFNode->getReg();
597 MachineBasicBlock* From = Inst->getOperand(i).getMBB();
599 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
600 UsedByAnother.insert(SrcReg);
602 PHIUnion.erase(SrcReg);
606 // If a node is live-in to the defining block of one of its children, but
607 // not live-out, then we need to scan that block for local interferences.
608 } else if (isLiveIn(DFNode->getReg(),
609 MRI.getVRegDef(child->getReg())->getParent(), LI) ||
610 MRI.getVRegDef(DFNode->getReg())->getParent() ==
611 MRI.getVRegDef(child->getReg())->getParent()) {
612 // Add (p, c) to possible local interferences
613 locals.push_back(std::make_pair(DFNode->getReg(), child->getReg()));
616 if (!visited.count(child)) {
617 worklist.push_back(child);
622 if (!inserted) worklist.pop_back();
626 /// ScheduleCopies - Insert copies into predecessor blocks, scheduling
627 /// them properly so as to avoid the 'lost copy' and the 'virtual swap'
630 /// Based on "Practical Improvements to the Construction and Destruction
631 /// of Static Single Assignment Form" by Briggs, et al.
632 void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
633 std::set<unsigned>& pushed) {
634 // FIXME: This function needs to update LiveVariables
635 std::map<unsigned, unsigned>& copy_set= Waiting[MBB];
637 std::map<unsigned, unsigned> worklist;
638 std::map<unsigned, unsigned> map;
640 // Setup worklist of initial copies
641 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
642 E = copy_set.end(); I != E; ) {
643 map.insert(std::make_pair(I->first, I->first));
644 map.insert(std::make_pair(I->second, I->second));
646 if (!UsedByAnother.count(I->second)) {
649 // Avoid iterator invalidation
650 unsigned first = I->first;
652 copy_set.erase(first);
658 LiveIntervals& LI = getAnalysis<LiveIntervals>();
659 MachineFunction* MF = MBB->getParent();
660 MachineRegisterInfo& MRI = MF->getRegInfo();
661 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
663 // Iterate over the worklist, inserting copies
664 while (!worklist.empty() || !copy_set.empty()) {
665 while (!worklist.empty()) {
666 std::pair<unsigned, unsigned> curr = *worklist.begin();
667 worklist.erase(curr.first);
669 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(curr.first);
671 if (isLiveOut(curr.second, MBB, LI)) {
672 // Create a temporary
673 unsigned t = MF->getRegInfo().createVirtualRegister(RC);
675 // Insert copy from curr.second to a temporary at
676 // the Phi defining curr.second
677 MachineBasicBlock::iterator PI = MRI.getVRegDef(curr.second);
678 TII->copyRegToReg(*PI->getParent(), PI, t,
679 curr.second, RC, RC);
681 // Push temporary on Stacks
682 Stacks[curr.second].push_back(t);
684 // Insert curr.second in pushed
685 pushed.insert(curr.second);
688 // Insert copy from map[curr.first] to curr.second
689 TII->copyRegToReg(*MBB, MBB->getFirstTerminator(), curr.second,
690 map[curr.first], RC, RC);
691 map[curr.first] = curr.second;
693 // If curr.first is a destination in copy_set...
694 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
695 E = copy_set.end(); I != E; )
696 if (curr.first == I->second) {
697 std::pair<unsigned, unsigned> temp = *I;
699 // Avoid iterator invalidation
701 copy_set.erase(temp.first);
702 worklist.insert(temp);
710 if (!copy_set.empty()) {
711 std::pair<unsigned, unsigned> curr = *copy_set.begin();
712 copy_set.erase(curr.first);
714 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(curr.first);
716 // Insert a copy from dest to a new temporary t at the end of b
717 unsigned t = MF->getRegInfo().createVirtualRegister(RC);
718 TII->copyRegToReg(*MBB, MBB->getFirstTerminator(), t,
719 curr.second, RC, RC);
720 map[curr.second] = t;
722 worklist.insert(curr);
727 /// InsertCopies - insert copies into MBB and all of its successors
728 void StrongPHIElimination::InsertCopies(MachineBasicBlock* MBB,
729 SmallPtrSet<MachineBasicBlock*, 16>& visited) {
732 std::set<unsigned> pushed;
734 // Rewrite register uses from Stacks
735 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
737 for (unsigned i = 0; i < I->getNumOperands(); ++i)
738 if (I->getOperand(i).isRegister() &&
739 Stacks[I->getOperand(i).getReg()].size()) {
740 I->getOperand(i).setReg(Stacks[I->getOperand(i).getReg()].back());
743 // Schedule the copies for this block
744 ScheduleCopies(MBB, pushed);
746 // Recur to our successors
747 for (GraphTraits<MachineBasicBlock*>::ChildIteratorType I =
748 GraphTraits<MachineBasicBlock*>::child_begin(MBB), E =
749 GraphTraits<MachineBasicBlock*>::child_end(MBB); I != E; ++I)
750 if (!visited.count(*I))
751 InsertCopies(*I, visited);
753 // As we exit this block, pop the names we pushed while processing it
754 for (std::set<unsigned>::iterator I = pushed.begin(),
755 E = pushed.end(); I != E; ++I)
756 Stacks[*I].pop_back();
759 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
760 /// compute what the resultant value numbers for each value in the input two
761 /// ranges will be. This is complicated by copies between the two which can
762 /// and will commonly cause multiple value numbers to be merged into one.
764 /// VN is the value number that we're trying to resolve. InstDefiningValue
765 /// keeps track of the new InstDefiningValue assignment for the result
766 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
767 /// whether a value in this or other is a copy from the opposite set.
768 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
769 /// already been assigned.
771 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
772 /// contains the value number the copy is from.
774 static unsigned ComputeUltimateVN(VNInfo *VNI,
775 SmallVector<VNInfo*, 16> &NewVNInfo,
776 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
777 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
778 SmallVector<int, 16> &ThisValNoAssignments,
779 SmallVector<int, 16> &OtherValNoAssignments) {
780 unsigned VN = VNI->id;
782 // If the VN has already been computed, just return it.
783 if (ThisValNoAssignments[VN] >= 0)
784 return ThisValNoAssignments[VN];
785 // assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?");
787 // If this val is not a copy from the other val, then it must be a new value
788 // number in the destination.
789 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
790 if (I == ThisFromOther.end()) {
791 NewVNInfo.push_back(VNI);
792 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
794 VNInfo *OtherValNo = I->second;
796 // Otherwise, this *is* a copy from the RHS. If the other side has already
797 // been computed, return it.
798 if (OtherValNoAssignments[OtherValNo->id] >= 0)
799 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
801 // Mark this value number as currently being computed, then ask what the
802 // ultimate value # of the other value is.
803 ThisValNoAssignments[VN] = -2;
804 unsigned UltimateVN =
805 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
806 OtherValNoAssignments, ThisValNoAssignments);
807 return ThisValNoAssignments[VN] = UltimateVN;
810 void StrongPHIElimination::mergeLiveIntervals(unsigned primary,
812 unsigned secondaryVN) {
814 LiveIntervals& LI = getAnalysis<LiveIntervals>();
815 LiveInterval& LHS = LI.getOrCreateInterval(primary);
816 LiveInterval& RHS = LI.getOrCreateInterval(secondary);
818 // Compute the final value assignment, assuming that the live ranges can be
820 SmallVector<int, 16> LHSValNoAssignments;
821 SmallVector<int, 16> RHSValNoAssignments;
822 SmallVector<VNInfo*, 16> NewVNInfo;
824 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
825 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
826 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
828 for (LiveInterval::vni_iterator I = LHS.vni_begin(), E = LHS.vni_end();
831 unsigned VN = VNI->id;
832 if (LHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
835 NewVNInfo.push_back(VNI);
836 LHSValNoAssignments[VN] = NewVNInfo.size()-1;
839 for (LiveInterval::vni_iterator I = RHS.vni_begin(), E = RHS.vni_end();
842 unsigned VN = VNI->id;
843 if (RHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
846 NewVNInfo.push_back(VNI);
847 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
850 // If we get here, we know that we can coalesce the live ranges. Ask the
851 // intervals to coalesce themselves now.
853 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo);
854 LI.removeInterval(secondary);
856 // The valno that was previously the input to the PHI node
857 // now has a PHIKill.
858 LHS.getValNumInfo(RHSValNoAssignments[secondaryVN])->hasPHIKill = true;
861 bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
862 LiveIntervals& LI = getAnalysis<LiveIntervals>();
864 // Compute DFS numbers of each block
867 // Determine which phi node operands need copies
868 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
870 I->begin()->getOpcode() == TargetInstrInfo::PHI)
874 // FIXME: This process should probably preserve LiveVariables
875 SmallPtrSet<MachineBasicBlock*, 16> visited;
876 InsertCopies(Fn.begin(), visited);
879 typedef std::map<unsigned, std::map<unsigned, unsigned> > RenameSetType;
880 for (RenameSetType::iterator I = RenameSets.begin(), E = RenameSets.end();
882 for (std::map<unsigned, unsigned>::iterator SI = I->second.begin(),
883 SE = I->second.end(); SI != SE; ++SI) {
884 mergeLiveIntervals(I->first, SI->first, SI->second);
885 Fn.getRegInfo().replaceRegWith(SI->first, I->first);
888 // FIXME: Insert last-minute copies
891 std::vector<MachineInstr*> phis;
892 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
893 for (MachineBasicBlock::iterator BI = I->begin(), BE = I->end();
895 if (BI->getOpcode() == TargetInstrInfo::PHI)
899 for (std::vector<MachineInstr*>::iterator I = phis.begin(), E = phis.end();
901 MachineInstr* PInstr = *(I++);
903 // If this is a dead PHI node, then remove it from LiveIntervals.
904 unsigned DestReg = PInstr->getOperand(0).getReg();
905 LiveInterval& PI = LI.getInterval(DestReg);
906 if (PInstr->registerDefIsDead(DestReg)) {
907 if (PI.containsOneValue()) {
908 LI.removeInterval(DestReg);
910 unsigned idx = LI.getDefIndex(LI.getInstructionIndex(PInstr));
911 PI.removeRange(*PI.getLiveRangeContaining(idx), true);
914 // If the PHI is not dead, then the valno defined by the PHI
915 // now has an unknown def.
916 unsigned idx = LI.getDefIndex(LI.getInstructionIndex(PInstr));
917 PI.getLiveRangeContaining(idx)->valno->def = ~0U;
920 LI.RemoveMachineInstrFromMaps(PInstr);
921 PInstr->eraseFromParent();