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);
144 char StrongPHIElimination::ID = 0;
145 RegisterPass<StrongPHIElimination> X("strong-phi-node-elimination",
146 "Eliminate PHI nodes for register allocation, intelligently");
149 const PassInfo *llvm::StrongPHIEliminationID = X.getPassInfo();
151 /// computeDFS - Computes the DFS-in and DFS-out numbers of the dominator tree
152 /// of the given MachineFunction. These numbers are then used in other parts
153 /// of the PHI elimination process.
154 void StrongPHIElimination::computeDFS(MachineFunction& MF) {
155 SmallPtrSet<MachineDomTreeNode*, 8> frontier;
156 SmallPtrSet<MachineDomTreeNode*, 8> visited;
160 MachineDominatorTree& DT = getAnalysis<MachineDominatorTree>();
162 MachineDomTreeNode* node = DT.getRootNode();
164 std::vector<MachineDomTreeNode*> worklist;
165 worklist.push_back(node);
167 while (!worklist.empty()) {
168 MachineDomTreeNode* currNode = worklist.back();
170 if (!frontier.count(currNode)) {
171 frontier.insert(currNode);
173 preorder.insert(std::make_pair(currNode->getBlock(), time));
176 bool inserted = false;
177 for (MachineDomTreeNode::iterator I = currNode->begin(), E = currNode->end();
179 if (!frontier.count(*I) && !visited.count(*I)) {
180 worklist.push_back(*I);
186 frontier.erase(currNode);
187 visited.insert(currNode);
188 maxpreorder.insert(std::make_pair(currNode->getBlock(), time));
195 /// PreorderSorter - a helper class that is used to sort registers
196 /// according to the preorder number of their defining blocks
197 class PreorderSorter {
199 DenseMap<MachineBasicBlock*, unsigned>& preorder;
200 MachineRegisterInfo& MRI;
203 PreorderSorter(DenseMap<MachineBasicBlock*, unsigned>& p,
204 MachineRegisterInfo& M) : preorder(p), MRI(M) { }
206 bool operator()(unsigned A, unsigned B) {
210 MachineBasicBlock* ABlock = MRI.getVRegDef(A)->getParent();
211 MachineBasicBlock* BBlock = MRI.getVRegDef(B)->getParent();
213 if (preorder[ABlock] < preorder[BBlock])
215 else if (preorder[ABlock] > preorder[BBlock])
222 /// computeDomForest - compute the subforest of the DomTree corresponding
223 /// to the defining blocks of the registers in question
224 std::vector<StrongPHIElimination::DomForestNode*>
225 StrongPHIElimination::computeDomForest(std::map<unsigned, unsigned>& regs,
226 MachineRegisterInfo& MRI) {
227 // Begin by creating a virtual root node, since the actual results
228 // may well be a forest. Assume this node has maximum DFS-out number.
229 DomForestNode* VirtualRoot = new DomForestNode(0, 0);
230 maxpreorder.insert(std::make_pair((MachineBasicBlock*)0, ~0UL));
232 // Populate a worklist with the registers
233 std::vector<unsigned> worklist;
234 worklist.reserve(regs.size());
235 for (std::map<unsigned, unsigned>::iterator I = regs.begin(), E = regs.end();
237 worklist.push_back(I->first);
239 // Sort the registers by the DFS-in number of their defining block
240 PreorderSorter PS(preorder, MRI);
241 std::sort(worklist.begin(), worklist.end(), PS);
243 // Create a "current parent" stack, and put the virtual root on top of it
244 DomForestNode* CurrentParent = VirtualRoot;
245 std::vector<DomForestNode*> stack;
246 stack.push_back(VirtualRoot);
248 // Iterate over all the registers in the previously computed order
249 for (std::vector<unsigned>::iterator I = worklist.begin(), E = worklist.end();
251 unsigned pre = preorder[MRI.getVRegDef(*I)->getParent()];
252 MachineBasicBlock* parentBlock = CurrentParent->getReg() ?
253 MRI.getVRegDef(CurrentParent->getReg())->getParent() :
256 // If the DFS-in number of the register is greater than the DFS-out number
257 // of the current parent, repeatedly pop the parent stack until it isn't.
258 while (pre > maxpreorder[parentBlock]) {
260 CurrentParent = stack.back();
262 parentBlock = CurrentParent->getReg() ?
263 MRI.getVRegDef(CurrentParent->getReg())->getParent() :
267 // Now that we've found the appropriate parent, create a DomForestNode for
268 // this register and attach it to the forest
269 DomForestNode* child = new DomForestNode(*I, CurrentParent);
271 // Push this new node on the "current parent" stack
272 stack.push_back(child);
273 CurrentParent = child;
276 // Return a vector containing the children of the virtual root node
277 std::vector<DomForestNode*> ret;
278 ret.insert(ret.end(), VirtualRoot->begin(), VirtualRoot->end());
282 /// isLiveIn - helper method that determines, from a regno, if a register
283 /// is live into a block
284 static bool isLiveIn(unsigned r, MachineBasicBlock* MBB,
286 LiveInterval& I = LI.getOrCreateInterval(r);
287 unsigned idx = LI.getMBBStartIdx(MBB);
288 return I.liveBeforeAndAt(idx);
291 /// isLiveOut - help method that determines, from a regno, if a register is
292 /// live out of a block.
293 static bool isLiveOut(unsigned r, MachineBasicBlock* MBB,
295 for (MachineBasicBlock::succ_iterator PI = MBB->succ_begin(),
296 E = MBB->succ_end(); PI != E; ++PI) {
297 if (isLiveIn(r, *PI, LI))
304 /// interferes - checks for local interferences by scanning a block. The only
305 /// trick parameter is 'mode' which tells it the relationship of the two
306 /// registers. 0 - defined in the same block, 1 - first properly dominates
307 /// second, 2 - second properly dominates first
308 static bool interferes(unsigned a, unsigned b, MachineBasicBlock* scan,
309 LiveIntervals& LV, unsigned mode) {
310 MachineInstr* def = 0;
311 MachineInstr* kill = 0;
313 // The code is still in SSA form at this point, so there is only one
314 // definition per VReg. Thus we can safely use MRI->getVRegDef().
315 const MachineRegisterInfo* MRI = &scan->getParent()->getRegInfo();
317 bool interference = false;
319 // Wallk the block, checking for interferences
320 for (MachineBasicBlock::iterator MBI = scan->begin(), MBE = scan->end();
322 MachineInstr* curr = MBI;
324 // Same defining block...
326 if (curr == MRI->getVRegDef(a)) {
327 // If we find our first definition, save it
330 // If there's already an unkilled definition, then
331 // this is an interference
335 // If there's a definition followed by a KillInst, then
336 // they can't interfere
338 interference = false;
341 // Symmetric with the above
342 } else if (curr == MRI->getVRegDef(b)) {
349 interference = false;
352 // Store KillInsts if they match up with the definition
353 } else if (curr->killsRegister(a)) {
354 if (def == MRI->getVRegDef(a)) {
356 } else if (curr->killsRegister(b)) {
357 if (def == MRI->getVRegDef(b)) {
362 // First properly dominates second...
363 } else if (mode == 1) {
364 if (curr == MRI->getVRegDef(b)) {
365 // Definition of second without kill of first is an interference
369 // Definition after a kill is a non-interference
371 interference = false;
374 // Save KillInsts of First
375 } else if (curr->killsRegister(a)) {
378 // Symmetric with the above
379 } else if (mode == 2) {
380 if (curr == MRI->getVRegDef(a)) {
385 interference = false;
388 } else if (curr->killsRegister(b)) {
397 /// processBlock - Determine how to break up PHIs in the current block. Each
398 /// PHI is broken up by some combination of renaming its operands and inserting
399 /// copies. This method is responsible for determining which operands receive
401 void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
402 LiveIntervals& LI = getAnalysis<LiveIntervals>();
403 MachineRegisterInfo& MRI = MBB->getParent()->getRegInfo();
405 // Holds names that have been added to a set in any PHI within this block
406 // before the current one.
407 std::set<unsigned> ProcessedNames;
409 // Iterate over all the PHI nodes in this block
410 MachineBasicBlock::iterator P = MBB->begin();
411 while (P != MBB->end() && P->getOpcode() == TargetInstrInfo::PHI) {
412 unsigned DestReg = P->getOperand(0).getReg();
414 // Don't both doing PHI elimination for dead PHI's.
415 if (P->registerDefIsDead(DestReg)) {
420 LiveInterval& PI = LI.getOrCreateInterval(DestReg);
421 unsigned pIdx = LI.getDefIndex(LI.getInstructionIndex(P));
422 VNInfo* PVN = PI.getLiveRangeContaining(pIdx)->valno;
423 PhiValueNumber.insert(std::make_pair(DestReg, PVN->id));
425 // PHIUnion is the set of incoming registers to the PHI node that
426 // are going to be renames rather than having copies inserted. This set
427 // is refinded over the course of this function. UnionedBlocks is the set
428 // of corresponding MBBs.
429 std::map<unsigned, unsigned> PHIUnion;
430 SmallPtrSet<MachineBasicBlock*, 8> UnionedBlocks;
432 // Iterate over the operands of the PHI node
433 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
434 unsigned SrcReg = P->getOperand(i-1).getReg();
436 // Check for trivial interferences via liveness information, allowing us
437 // to avoid extra work later. Any registers that interfere cannot both
438 // be in the renaming set, so choose one and add copies for it instead.
439 // The conditions are:
440 // 1) if the operand is live into the PHI node's block OR
441 // 2) if the PHI node is live out of the operand's defining block OR
442 // 3) if the operand is itself a PHI node and the original PHI is
443 // live into the operand's defining block OR
444 // 4) if the operand is already being renamed for another PHI node
446 // 5) if any two operands are defined in the same block, insert copies
448 if (isLiveIn(SrcReg, P->getParent(), LI) ||
449 isLiveOut(P->getOperand(0).getReg(),
450 MRI.getVRegDef(SrcReg)->getParent(), LI) ||
451 ( MRI.getVRegDef(SrcReg)->getOpcode() == TargetInstrInfo::PHI &&
452 isLiveIn(P->getOperand(0).getReg(),
453 MRI.getVRegDef(SrcReg)->getParent(), LI) ) ||
454 ProcessedNames.count(SrcReg) ||
455 UnionedBlocks.count(MRI.getVRegDef(SrcReg)->getParent())) {
457 // Add a copy for the selected register
458 MachineBasicBlock* From = P->getOperand(i).getMBB();
459 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
460 UsedByAnother.insert(SrcReg);
462 // Otherwise, add it to the renaming set
463 LiveInterval& I = LI.getOrCreateInterval(SrcReg);
464 unsigned idx = LI.getMBBEndIdx(P->getOperand(i).getMBB());
465 VNInfo* VN = I.getLiveRangeContaining(idx)->valno;
467 assert(VN && "No VNInfo for register?");
469 PHIUnion.insert(std::make_pair(SrcReg, VN->id));
470 UnionedBlocks.insert(MRI.getVRegDef(SrcReg)->getParent());
474 // Compute the dominator forest for the renaming set. This is a forest
475 // where the nodes are the registers and the edges represent dominance
476 // relations between the defining blocks of the registers
477 std::vector<StrongPHIElimination::DomForestNode*> DF =
478 computeDomForest(PHIUnion, MRI);
480 // Walk DomForest to resolve interferences at an inter-block level. This
481 // will remove registers from the renaming set (and insert copies for them)
482 // if interferences are found.
483 std::vector<std::pair<unsigned, unsigned> > localInterferences;
484 processPHIUnion(P, PHIUnion, DF, localInterferences);
486 // If one of the inputs is defined in the same block as the current PHI
487 // then we need to check for a local interference between that input and
489 for (std::map<unsigned, unsigned>::iterator I = PHIUnion.begin(),
490 E = PHIUnion.end(); I != E; ++I)
491 if (MRI.getVRegDef(I->first)->getParent() == P->getParent())
492 localInterferences.push_back(std::make_pair(I->first,
493 P->getOperand(0).getReg()));
495 // The dominator forest walk may have returned some register pairs whose
496 // interference cannot be determined from dominator analysis. We now
497 // examine these pairs for local interferences.
498 for (std::vector<std::pair<unsigned, unsigned> >::iterator I =
499 localInterferences.begin(), E = localInterferences.end(); I != E; ++I) {
500 std::pair<unsigned, unsigned> p = *I;
502 MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
504 // Determine the block we need to scan and the relationship between
506 MachineBasicBlock* scan = 0;
508 if (MRI.getVRegDef(p.first)->getParent() ==
509 MRI.getVRegDef(p.second)->getParent()) {
510 scan = MRI.getVRegDef(p.first)->getParent();
511 mode = 0; // Same block
512 } else if (MDT.dominates(MRI.getVRegDef(p.first)->getParent(),
513 MRI.getVRegDef(p.second)->getParent())) {
514 scan = MRI.getVRegDef(p.second)->getParent();
515 mode = 1; // First dominates second
517 scan = MRI.getVRegDef(p.first)->getParent();
518 mode = 2; // Second dominates first
521 // If there's an interference, we need to insert copies
522 if (interferes(p.first, p.second, scan, LI, mode)) {
523 // Insert copies for First
524 for (int i = P->getNumOperands() - 1; i >= 2; i-=2) {
525 if (P->getOperand(i-1).getReg() == p.first) {
526 unsigned SrcReg = p.first;
527 MachineBasicBlock* From = P->getOperand(i).getMBB();
529 Waiting[From].insert(std::make_pair(SrcReg,
530 P->getOperand(0).getReg()));
531 UsedByAnother.insert(SrcReg);
533 PHIUnion.erase(SrcReg);
539 // Add the renaming set for this PHI node to our overall renaming information
540 RenameSets.insert(std::make_pair(P->getOperand(0).getReg(), PHIUnion));
542 // Remember which registers are already renamed, so that we don't try to
543 // rename them for another PHI node in this block
544 for (std::map<unsigned, unsigned>::iterator I = PHIUnion.begin(),
545 E = PHIUnion.end(); I != E; ++I)
546 ProcessedNames.insert(I->first);
552 /// processPHIUnion - Take a set of candidate registers to be coalesced when
553 /// decomposing the PHI instruction. Use the DominanceForest to remove the ones
554 /// that are known to interfere, and flag others that need to be checked for
555 /// local interferences.
556 void StrongPHIElimination::processPHIUnion(MachineInstr* Inst,
557 std::map<unsigned, unsigned>& PHIUnion,
558 std::vector<StrongPHIElimination::DomForestNode*>& DF,
559 std::vector<std::pair<unsigned, unsigned> >& locals) {
561 std::vector<DomForestNode*> worklist(DF.begin(), DF.end());
562 SmallPtrSet<DomForestNode*, 4> visited;
564 // Code is still in SSA form, so we can use MRI::getVRegDef()
565 MachineRegisterInfo& MRI = Inst->getParent()->getParent()->getRegInfo();
567 LiveIntervals& LI = getAnalysis<LiveIntervals>();
568 unsigned DestReg = Inst->getOperand(0).getReg();
570 // DF walk on the DomForest
571 while (!worklist.empty()) {
572 DomForestNode* DFNode = worklist.back();
574 visited.insert(DFNode);
576 bool inserted = false;
577 for (DomForestNode::iterator CI = DFNode->begin(), CE = DFNode->end();
579 DomForestNode* child = *CI;
581 // If the current node is live-out of the defining block of one of its
582 // children, insert a copy for it. NOTE: The paper actually calls for
583 // a more elaborate heuristic for determining whether to insert copies
584 // for the child or the parent. In the interest of simplicity, we're
585 // just always choosing the parent.
586 if (isLiveOut(DFNode->getReg(),
587 MRI.getVRegDef(child->getReg())->getParent(), LI)) {
588 // Insert copies for parent
589 for (int i = Inst->getNumOperands() - 1; i >= 2; i-=2) {
590 if (Inst->getOperand(i-1).getReg() == DFNode->getReg()) {
591 unsigned SrcReg = DFNode->getReg();
592 MachineBasicBlock* From = Inst->getOperand(i).getMBB();
594 Waiting[From].insert(std::make_pair(SrcReg, DestReg));
595 UsedByAnother.insert(SrcReg);
597 PHIUnion.erase(SrcReg);
601 // If a node is live-in to the defining block of one of its children, but
602 // not live-out, then we need to scan that block for local interferences.
603 } else if (isLiveIn(DFNode->getReg(),
604 MRI.getVRegDef(child->getReg())->getParent(), LI) ||
605 MRI.getVRegDef(DFNode->getReg())->getParent() ==
606 MRI.getVRegDef(child->getReg())->getParent()) {
607 // Add (p, c) to possible local interferences
608 locals.push_back(std::make_pair(DFNode->getReg(), child->getReg()));
611 if (!visited.count(child)) {
612 worklist.push_back(child);
617 if (!inserted) worklist.pop_back();
621 /// ScheduleCopies - Insert copies into predecessor blocks, scheduling
622 /// them properly so as to avoid the 'lost copy' and the 'virtual swap'
625 /// Based on "Practical Improvements to the Construction and Destruction
626 /// of Static Single Assignment Form" by Briggs, et al.
627 void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
628 std::set<unsigned>& pushed) {
629 // FIXME: This function needs to update LiveVariables
630 std::map<unsigned, unsigned>& copy_set= Waiting[MBB];
632 std::map<unsigned, unsigned> worklist;
633 std::map<unsigned, unsigned> map;
635 // Setup worklist of initial copies
636 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
637 E = copy_set.end(); I != E; ) {
638 map.insert(std::make_pair(I->first, I->first));
639 map.insert(std::make_pair(I->second, I->second));
641 if (!UsedByAnother.count(I->second)) {
644 // Avoid iterator invalidation
645 unsigned first = I->first;
647 copy_set.erase(first);
653 LiveIntervals& LI = getAnalysis<LiveIntervals>();
654 MachineFunction* MF = MBB->getParent();
655 MachineRegisterInfo& MRI = MF->getRegInfo();
656 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
658 // Iterate over the worklist, inserting copies
659 while (!worklist.empty() || !copy_set.empty()) {
660 while (!worklist.empty()) {
661 std::pair<unsigned, unsigned> curr = *worklist.begin();
662 worklist.erase(curr.first);
664 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(curr.first);
666 if (isLiveOut(curr.second, MBB, LI)) {
667 // Create a temporary
668 unsigned t = MF->getRegInfo().createVirtualRegister(RC);
670 // Insert copy from curr.second to a temporary at
671 // the Phi defining curr.second
672 MachineBasicBlock::iterator PI = MRI.getVRegDef(curr.second);
673 TII->copyRegToReg(*PI->getParent(), PI, t,
674 curr.second, RC, RC);
676 // Push temporary on Stacks
677 Stacks[curr.second].push_back(t);
679 // Insert curr.second in pushed
680 pushed.insert(curr.second);
683 // Insert copy from map[curr.first] to curr.second
684 TII->copyRegToReg(*MBB, MBB->getFirstTerminator(), curr.second,
685 map[curr.first], RC, RC);
686 map[curr.first] = curr.second;
688 // If curr.first is a destination in copy_set...
689 for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
690 E = copy_set.end(); I != E; )
691 if (curr.first == I->second) {
692 std::pair<unsigned, unsigned> temp = *I;
694 // Avoid iterator invalidation
696 copy_set.erase(temp.first);
697 worklist.insert(temp);
705 if (!copy_set.empty()) {
706 std::pair<unsigned, unsigned> curr = *copy_set.begin();
707 copy_set.erase(curr.first);
709 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(curr.first);
711 // Insert a copy from dest to a new temporary t at the end of b
712 unsigned t = MF->getRegInfo().createVirtualRegister(RC);
713 TII->copyRegToReg(*MBB, MBB->getFirstTerminator(), t,
714 curr.second, RC, RC);
715 map[curr.second] = t;
717 worklist.insert(curr);
722 /// InsertCopies - insert copies into MBB and all of its successors
723 void StrongPHIElimination::InsertCopies(MachineBasicBlock* MBB,
724 SmallPtrSet<MachineBasicBlock*, 16>& visited) {
727 std::set<unsigned> pushed;
729 // Rewrite register uses from Stacks
730 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
732 for (unsigned i = 0; i < I->getNumOperands(); ++i)
733 if (I->getOperand(i).isRegister() &&
734 Stacks[I->getOperand(i).getReg()].size()) {
735 I->getOperand(i).setReg(Stacks[I->getOperand(i).getReg()].back());
738 // Schedule the copies for this block
739 ScheduleCopies(MBB, pushed);
741 // Recur to our successors
742 for (GraphTraits<MachineBasicBlock*>::ChildIteratorType I =
743 GraphTraits<MachineBasicBlock*>::child_begin(MBB), E =
744 GraphTraits<MachineBasicBlock*>::child_end(MBB); I != E; ++I)
745 if (!visited.count(*I))
746 InsertCopies(*I, visited);
748 // As we exit this block, pop the names we pushed while processing it
749 for (std::set<unsigned>::iterator I = pushed.begin(),
750 E = pushed.end(); I != E; ++I)
751 Stacks[*I].pop_back();
754 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
755 /// compute what the resultant value numbers for each value in the input two
756 /// ranges will be. This is complicated by copies between the two which can
757 /// and will commonly cause multiple value numbers to be merged into one.
759 /// VN is the value number that we're trying to resolve. InstDefiningValue
760 /// keeps track of the new InstDefiningValue assignment for the result
761 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
762 /// whether a value in this or other is a copy from the opposite set.
763 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
764 /// already been assigned.
766 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
767 /// contains the value number the copy is from.
769 static unsigned ComputeUltimateVN(VNInfo *VNI,
770 SmallVector<VNInfo*, 16> &NewVNInfo,
771 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
772 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
773 SmallVector<int, 16> &ThisValNoAssignments,
774 SmallVector<int, 16> &OtherValNoAssignments) {
775 unsigned VN = VNI->id;
777 // If the VN has already been computed, just return it.
778 if (ThisValNoAssignments[VN] >= 0)
779 return ThisValNoAssignments[VN];
780 // assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?");
782 // If this val is not a copy from the other val, then it must be a new value
783 // number in the destination.
784 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
785 if (I == ThisFromOther.end()) {
786 NewVNInfo.push_back(VNI);
787 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
789 VNInfo *OtherValNo = I->second;
791 // Otherwise, this *is* a copy from the RHS. If the other side has already
792 // been computed, return it.
793 if (OtherValNoAssignments[OtherValNo->id] >= 0)
794 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
796 // Mark this value number as currently being computed, then ask what the
797 // ultimate value # of the other value is.
798 ThisValNoAssignments[VN] = -2;
799 unsigned UltimateVN =
800 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
801 OtherValNoAssignments, ThisValNoAssignments);
802 return ThisValNoAssignments[VN] = UltimateVN;
805 void StrongPHIElimination::mergeLiveIntervals(unsigned primary,
807 unsigned secondaryVN) {
809 LiveIntervals& LI = getAnalysis<LiveIntervals>();
810 LiveInterval& LHS = LI.getOrCreateInterval(primary);
811 LiveInterval& RHS = LI.getOrCreateInterval(secondary);
813 // Compute the final value assignment, assuming that the live ranges can be
815 SmallVector<int, 16> LHSValNoAssignments;
816 SmallVector<int, 16> RHSValNoAssignments;
817 SmallVector<VNInfo*, 16> NewVNInfo;
819 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
820 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
821 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
823 for (LiveInterval::vni_iterator I = LHS.vni_begin(), E = LHS.vni_end();
826 unsigned VN = VNI->id;
827 if (LHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
830 NewVNInfo.push_back(VNI);
831 LHSValNoAssignments[VN] = NewVNInfo.size()-1;
834 for (LiveInterval::vni_iterator I = RHS.vni_begin(), E = RHS.vni_end();
837 unsigned VN = VNI->id;
838 if (RHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
841 NewVNInfo.push_back(VNI);
842 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
845 // If we get here, we know that we can coalesce the live ranges. Ask the
846 // intervals to coalesce themselves now.
848 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo);
849 LI.removeInterval(secondary);
851 // The valno that was previously the input to the PHI node
852 // now has a PHIKill.
853 LHS.getValNumInfo(RHSValNoAssignments[secondaryVN])->hasPHIKill = true;
856 bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
857 LiveIntervals& LI = getAnalysis<LiveIntervals>();
859 // Compute DFS numbers of each block
862 // Determine which phi node operands need copies
863 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
865 I->begin()->getOpcode() == TargetInstrInfo::PHI)
869 // FIXME: This process should probably preserve LiveVariables
870 SmallPtrSet<MachineBasicBlock*, 16> visited;
871 InsertCopies(Fn.begin(), visited);
874 typedef std::map<unsigned, std::map<unsigned, unsigned> > RenameSetType;
875 for (RenameSetType::iterator I = RenameSets.begin(), E = RenameSets.end();
877 for (std::map<unsigned, unsigned>::iterator SI = I->second.begin(),
878 SE = I->second.end(); SI != SE; ++SI) {
879 mergeLiveIntervals(I->first, SI->first, SI->second);
880 Fn.getRegInfo().replaceRegWith(SI->first, I->first);
883 // FIXME: Insert last-minute copies
886 std::vector<MachineInstr*> phis;
887 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
888 for (MachineBasicBlock::iterator BI = I->begin(), BE = I->end();
890 if (BI->getOpcode() == TargetInstrInfo::PHI)
894 for (std::vector<MachineInstr*>::iterator I = phis.begin(), E = phis.end();
896 MachineInstr* PInstr = *(I++);
898 // If this is a dead PHI node, then remove it from LiveIntervals.
899 unsigned DestReg = PInstr->getOperand(0).getReg();
900 LiveInterval& PI = LI.getInterval(DestReg);
901 if (PInstr->registerDefIsDead(DestReg)) {
902 if (PI.containsOneValue()) {
903 LI.removeInterval(DestReg);
905 unsigned idx = LI.getDefIndex(LI.getInstructionIndex(PInstr));
906 PI.removeRange(*PI.getLiveRangeContaining(idx), true);
909 // If the PHI is not dead, then the valno defined by the PHI
910 // now has an unknown def.
911 unsigned idx = LI.getDefIndex(LI.getInstructionIndex(PInstr));
912 PI.getLiveRangeContaining(idx)->valno->def = ~0U;
915 LI.RemoveMachineInstrFromMaps(PInstr);
916 PInstr->eraseFromParent();