1 /*
2  * Copyright 2011 Christoph Bumiller
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20  * SOFTWARE.
21  */
22 
23 #include "nv50_ir.h"
24 #include "nv50_ir_target.h"
25 
26 #include <stack>
27 #include <limits>
28 
29 namespace nv50_ir {
30 
31 #define MAX_REGISTER_FILE_SIZE 256
32 
33 class RegisterSet
34 {
35 public:
36    RegisterSet(const Target *);
37 
38    void init(const Target *);
39    void reset(DataFile, bool resetMax = false);
40 
41    void periodicMask(DataFile f, uint32_t lock, uint32_t unlock);
42    void intersect(DataFile f, const RegisterSet *);
43 
44    bool assign(int32_t& reg, DataFile f, unsigned int size);
45    void release(DataFile f, int32_t reg, unsigned int size);
46    bool occupy(DataFile f, int32_t reg, unsigned int size);
47    bool occupy(const Value *);
48    void occupyMask(DataFile f, int32_t reg, uint8_t mask);
49 
getMaxAssigned(DataFile f) const50    inline int getMaxAssigned(DataFile f) const { return fill[f]; }
51 
getFileSize(DataFile f,uint8_t regSize) const52    inline unsigned int getFileSize(DataFile f, uint8_t regSize) const
53    {
54       if (restrictedGPR16Range && f == FILE_GPR && regSize == 2)
55          return (last[f] + 1) / 2;
56       return last[f] + 1;
57    }
58 
units(DataFile f,unsigned int size) const59    inline unsigned int units(DataFile f, unsigned int size) const
60    {
61       return size >> unit[f];
62    }
63    // for regs of size >= 4, id is counted in 4-byte words (like nv50/c0 binary)
idToBytes(Value * v) const64    inline unsigned int idToBytes(Value *v) const
65    {
66       return v->reg.data.id * MIN2(v->reg.size, 4);
67    }
idToUnits(Value * v) const68    inline unsigned int idToUnits(Value *v) const
69    {
70       return units(v->reg.file, idToBytes(v));
71    }
bytesToId(Value * v,unsigned int bytes) const72    inline int bytesToId(Value *v, unsigned int bytes) const
73    {
74       if (v->reg.size < 4)
75          return units(v->reg.file, bytes);
76       return bytes / 4;
77    }
unitsToId(DataFile f,int u,uint8_t size) const78    inline int unitsToId(DataFile f, int u, uint8_t size) const
79    {
80       if (u < 0)
81          return -1;
82       return (size < 4) ? u : ((u << unit[f]) / 4);
83    }
84 
85    void print() const;
86 
87 private:
88    BitSet bits[LAST_REGISTER_FILE + 1];
89 
90    int unit[LAST_REGISTER_FILE + 1]; // log2 of allocation granularity
91 
92    int last[LAST_REGISTER_FILE + 1];
93    int fill[LAST_REGISTER_FILE + 1];
94 
95    const bool restrictedGPR16Range;
96 };
97 
98 void
reset(DataFile f,bool resetMax)99 RegisterSet::reset(DataFile f, bool resetMax)
100 {
101    bits[f].fill(0);
102    if (resetMax)
103       fill[f] = -1;
104 }
105 
106 void
init(const Target * targ)107 RegisterSet::init(const Target *targ)
108 {
109    for (unsigned int rf = 0; rf <= FILE_ADDRESS; ++rf) {
110       DataFile f = static_cast<DataFile>(rf);
111       last[rf] = targ->getFileSize(f) - 1;
112       unit[rf] = targ->getFileUnit(f);
113       fill[rf] = -1;
114       assert(last[rf] < MAX_REGISTER_FILE_SIZE);
115       bits[rf].allocate(last[rf] + 1, true);
116    }
117 }
118 
RegisterSet(const Target * targ)119 RegisterSet::RegisterSet(const Target *targ)
120   : restrictedGPR16Range(targ->getChipset() < 0xc0)
121 {
122    init(targ);
123    for (unsigned int i = 0; i <= LAST_REGISTER_FILE; ++i)
124       reset(static_cast<DataFile>(i));
125 }
126 
127 void
periodicMask(DataFile f,uint32_t lock,uint32_t unlock)128 RegisterSet::periodicMask(DataFile f, uint32_t lock, uint32_t unlock)
129 {
130    bits[f].periodicMask32(lock, unlock);
131 }
132 
133 void
intersect(DataFile f,const RegisterSet * set)134 RegisterSet::intersect(DataFile f, const RegisterSet *set)
135 {
136    bits[f] |= set->bits[f];
137 }
138 
139 void
print() const140 RegisterSet::print() const
141 {
142    INFO("GPR:");
143    bits[FILE_GPR].print();
144    INFO("\n");
145 }
146 
147 bool
assign(int32_t & reg,DataFile f,unsigned int size)148 RegisterSet::assign(int32_t& reg, DataFile f, unsigned int size)
149 {
150    reg = bits[f].findFreeRange(size);
151    if (reg < 0)
152       return false;
153    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
154    return true;
155 }
156 
157 bool
occupy(const Value * v)158 RegisterSet::occupy(const Value *v)
159 {
160    return occupy(v->reg.file, v->reg.data.id, v->reg.size >> unit[v->reg.file]);
161 }
162 
163 void
occupyMask(DataFile f,int32_t reg,uint8_t mask)164 RegisterSet::occupyMask(DataFile f, int32_t reg, uint8_t mask)
165 {
166    bits[f].setMask(reg & ~31, static_cast<uint32_t>(mask) << (reg % 32));
167 }
168 
169 bool
occupy(DataFile f,int32_t reg,unsigned int size)170 RegisterSet::occupy(DataFile f, int32_t reg, unsigned int size)
171 {
172    if (bits[f].testRange(reg, size))
173       return false;
174 
175    bits[f].setRange(reg, size);
176 
177    INFO_DBG(0, REG_ALLOC, "reg occupy: %u[%i] %u\n", f, reg, size);
178 
179    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
180 
181    return true;
182 }
183 
184 void
release(DataFile f,int32_t reg,unsigned int size)185 RegisterSet::release(DataFile f, int32_t reg, unsigned int size)
186 {
187    bits[f].clrRange(reg, size);
188 
189    INFO_DBG(0, REG_ALLOC, "reg release: %u[%i] %u\n", f, reg, size);
190 }
191 
192 class RegAlloc
193 {
194 public:
RegAlloc(Program * program)195    RegAlloc(Program *program) : prog(program), sequence(0) { }
196 
197    bool exec();
198    bool execFunc();
199 
200 private:
201    class PhiMovesPass : public Pass {
202    private:
203       virtual bool visit(BasicBlock *);
204       inline bool needNewElseBlock(BasicBlock *b, BasicBlock *p);
205    };
206 
207    class ArgumentMovesPass : public Pass {
208    private:
209       virtual bool visit(BasicBlock *);
210    };
211 
212    class BuildIntervalsPass : public Pass {
213    private:
214       virtual bool visit(BasicBlock *);
215       void collectLiveValues(BasicBlock *);
216       void addLiveRange(Value *, const BasicBlock *, int end);
217    };
218 
219    class InsertConstraintsPass : public Pass {
220    public:
221       bool exec(Function *func);
222    private:
223       virtual bool visit(BasicBlock *);
224 
225       bool insertConstraintMoves();
226 
227       void condenseDefs(Instruction *);
228       void condenseSrcs(Instruction *, const int first, const int last);
229 
230       void addHazard(Instruction *i, const ValueRef *src);
231       void textureMask(TexInstruction *);
232       void addConstraint(Instruction *, int s, int n);
233       bool detectConflict(Instruction *, int s);
234 
235       // target specific functions, TODO: put in subclass or Target
236       void texConstraintNV50(TexInstruction *);
237       void texConstraintNVC0(TexInstruction *);
238       void texConstraintNVE0(TexInstruction *);
239 
240       std::list<Instruction *> constrList;
241 
242       const Target *targ;
243    };
244 
245    bool buildLiveSets(BasicBlock *);
246 
247 private:
248    Program *prog;
249    Function *func;
250 
251    // instructions in control flow / chronological order
252    ArrayList insns;
253 
254    int sequence; // for manual passes through CFG
255 };
256 
257 typedef std::pair<Value *, Value *> ValuePair;
258 
259 class SpillCodeInserter
260 {
261 public:
SpillCodeInserter(Function * fn)262    SpillCodeInserter(Function *fn) : func(fn), stackSize(0), stackBase(0) { }
263 
264    bool run(const std::list<ValuePair>&);
265 
266    Symbol *assignSlot(const Interval&, unsigned int size);
getStackSize() const267    inline int32_t getStackSize() const { return stackSize; }
268 
269 private:
270    Function *func;
271 
272    struct SpillSlot
273    {
274       Interval occup;
275       std::list<Value *> residents; // needed to recalculate occup
276       Symbol *sym;
277       int32_t offset;
sizenv50_ir::SpillCodeInserter::SpillSlot278       inline uint8_t size() const { return sym->reg.size; }
279    };
280    std::list<SpillSlot> slots;
281    int32_t stackSize;
282    int32_t stackBase;
283 
284    LValue *unspill(Instruction *usei, LValue *, Value *slot);
285    void spill(Instruction *defi, Value *slot, LValue *);
286 };
287 
288 void
addLiveRange(Value * val,const BasicBlock * bb,int end)289 RegAlloc::BuildIntervalsPass::addLiveRange(Value *val,
290                                            const BasicBlock *bb,
291                                            int end)
292 {
293    Instruction *insn = val->getUniqueInsn();
294 
295    if (!insn)
296       insn = bb->getFirst();
297 
298    assert(bb->getFirst()->serial <= bb->getExit()->serial);
299    assert(bb->getExit()->serial + 1 >= end);
300 
301    int begin = insn->serial;
302    if (begin < bb->getEntry()->serial || begin > bb->getExit()->serial)
303       begin = bb->getEntry()->serial;
304 
305    INFO_DBG(prog->dbgFlags, REG_ALLOC, "%%%i <- live range [%i(%i), %i)\n",
306             val->id, begin, insn->serial, end);
307 
308    if (begin != end) // empty ranges are only added as hazards for fixed regs
309       val->livei.extend(begin, end);
310 }
311 
312 bool
needNewElseBlock(BasicBlock * b,BasicBlock * p)313 RegAlloc::PhiMovesPass::needNewElseBlock(BasicBlock *b, BasicBlock *p)
314 {
315    if (b->cfg.incidentCount() <= 1)
316       return false;
317 
318    int n = 0;
319    for (Graph::EdgeIterator ei = p->cfg.outgoing(); !ei.end(); ei.next())
320       if (ei.getType() == Graph::Edge::TREE ||
321           ei.getType() == Graph::Edge::FORWARD)
322          ++n;
323    return (n == 2);
324 }
325 
326 // For each operand of each PHI in b, generate a new value by inserting a MOV
327 // at the end of the block it is coming from and replace the operand with its
328 // result. This eliminates liveness conflicts and enables us to let values be
329 // copied to the right register if such a conflict exists nonetheless.
330 //
331 // These MOVs are also crucial in making sure the live intervals of phi srces
332 // are extended until the end of the loop, since they are not included in the
333 // live-in sets.
334 bool
visit(BasicBlock * bb)335 RegAlloc::PhiMovesPass::visit(BasicBlock *bb)
336 {
337    Instruction *phi, *mov;
338    BasicBlock *pb, *pn;
339 
340    std::stack<BasicBlock *> stack;
341 
342    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
343       pb = BasicBlock::get(ei.getNode());
344       assert(pb);
345       if (needNewElseBlock(bb, pb))
346          stack.push(pb);
347    }
348    while (!stack.empty()) {
349       pb = stack.top();
350       pn = new BasicBlock(func);
351       stack.pop();
352 
353       pb->cfg.detach(&bb->cfg);
354       pb->cfg.attach(&pn->cfg, Graph::Edge::TREE);
355       pn->cfg.attach(&bb->cfg, Graph::Edge::FORWARD);
356 
357       assert(pb->getExit()->op != OP_CALL);
358       if (pb->getExit()->asFlow()->target.bb == bb)
359          pb->getExit()->asFlow()->target.bb = pn;
360    }
361 
362    // insert MOVs (phi->src(j) should stem from j-th in-BB)
363    int j = 0;
364    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
365       pb = BasicBlock::get(ei.getNode());
366       if (!pb->isTerminated())
367          pb->insertTail(new_FlowInstruction(func, OP_BRA, bb));
368 
369       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
370          mov = new_Instruction(func, OP_MOV, TYPE_U32);
371 
372          mov->setSrc(0, phi->getSrc(j));
373          mov->setDef(0, new_LValue(func, phi->getDef(0)->asLValue()));
374          phi->setSrc(j, mov->getDef(0));
375 
376          pb->insertBefore(pb->getExit(), mov);
377       }
378       ++j;
379    }
380 
381    return true;
382 }
383 
384 bool
visit(BasicBlock * bb)385 RegAlloc::ArgumentMovesPass::visit(BasicBlock *bb)
386 {
387    // Bind function call inputs/outputs to the same physical register
388    // the callee uses, inserting moves as appropriate for the case a
389    // conflict arises.
390    for (Instruction *i = bb->getEntry(); i; i = i->next) {
391       FlowInstruction *cal = i->asFlow();
392       if (!cal || cal->op != OP_CALL || cal->builtin)
393          continue;
394       RegisterSet clobberSet(prog->getTarget());
395 
396       // Bind input values.
397       for (int s = 0; cal->srcExists(s); ++s) {
398          LValue *tmp = new_LValue(func, cal->getSrc(s)->asLValue());
399          tmp->reg.data.id = cal->target.fn->ins[s].rep()->reg.data.id;
400 
401          Instruction *mov =
402             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
403          mov->setDef(0, tmp);
404          mov->setSrc(0, cal->getSrc(s));
405          cal->setSrc(s, tmp);
406 
407          bb->insertBefore(cal, mov);
408       }
409 
410       // Bind output values.
411       for (int d = 0; cal->defExists(d); ++d) {
412          LValue *tmp = new_LValue(func, cal->getDef(d)->asLValue());
413          tmp->reg.data.id = cal->target.fn->outs[d].rep()->reg.data.id;
414 
415          Instruction *mov =
416             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
417          mov->setSrc(0, tmp);
418          mov->setDef(0, cal->getDef(d));
419          cal->setDef(d, tmp);
420 
421          bb->insertAfter(cal, mov);
422          clobberSet.occupy(tmp);
423       }
424 
425       // Bind clobbered values.
426       for (std::deque<Value *>::iterator it = cal->target.fn->clobbers.begin();
427            it != cal->target.fn->clobbers.end();
428            ++it) {
429          if (clobberSet.occupy(*it)) {
430             Value *tmp = new_LValue(func, (*it)->asLValue());
431             tmp->reg.data.id = (*it)->reg.data.id;
432             cal->setDef(cal->defCount(), tmp);
433          }
434       }
435    }
436 
437    // Update the clobber set of the function.
438    if (BasicBlock::get(func->cfgExit) == bb) {
439       func->buildDefSets();
440       for (unsigned int i = 0; i < bb->defSet.getSize(); ++i)
441          if (bb->defSet.test(i))
442             func->clobbers.push_back(func->getLValue(i));
443    }
444 
445    return true;
446 }
447 
448 // Build the set of live-in variables of bb.
449 bool
buildLiveSets(BasicBlock * bb)450 RegAlloc::buildLiveSets(BasicBlock *bb)
451 {
452    Function *f = bb->getFunction();
453    BasicBlock *bn;
454    Instruction *i;
455    unsigned int s, d;
456 
457    INFO_DBG(prog->dbgFlags, REG_ALLOC, "buildLiveSets(BB:%i)\n", bb->getId());
458 
459    bb->liveSet.allocate(func->allLValues.getSize(), false);
460 
461    int n = 0;
462    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
463       bn = BasicBlock::get(ei.getNode());
464       if (bn == bb)
465          continue;
466       if (bn->cfg.visit(sequence))
467          if (!buildLiveSets(bn))
468             return false;
469       if (n++ || bb->liveSet.marker)
470          bb->liveSet |= bn->liveSet;
471       else
472          bb->liveSet = bn->liveSet;
473    }
474    if (!n && !bb->liveSet.marker)
475       bb->liveSet.fill(0);
476    bb->liveSet.marker = true;
477 
478    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
479       INFO("BB:%i live set of out blocks:\n", bb->getId());
480       bb->liveSet.print();
481    }
482 
483    // if (!bb->getEntry())
484    //   return true;
485 
486    if (bb == BasicBlock::get(f->cfgExit)) {
487       for (std::deque<ValueRef>::iterator it = f->outs.begin();
488            it != f->outs.end(); ++it) {
489          assert(it->get()->asLValue());
490          bb->liveSet.set(it->get()->id);
491       }
492    }
493 
494    for (i = bb->getExit(); i && i != bb->getEntry()->prev; i = i->prev) {
495       for (d = 0; i->defExists(d); ++d)
496          bb->liveSet.clr(i->getDef(d)->id);
497       for (s = 0; i->srcExists(s); ++s)
498          if (i->getSrc(s)->asLValue())
499             bb->liveSet.set(i->getSrc(s)->id);
500    }
501    for (i = bb->getPhi(); i && i->op == OP_PHI; i = i->next)
502       bb->liveSet.clr(i->getDef(0)->id);
503 
504    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
505       INFO("BB:%i live set after propagation:\n", bb->getId());
506       bb->liveSet.print();
507    }
508 
509    return true;
510 }
511 
512 void
collectLiveValues(BasicBlock * bb)513 RegAlloc::BuildIntervalsPass::collectLiveValues(BasicBlock *bb)
514 {
515    BasicBlock *bbA = NULL, *bbB = NULL;
516 
517    if (bb->cfg.outgoingCount()) {
518       // trickery to save a loop of OR'ing liveSets
519       // aliasing works fine with BitSet::setOr
520       for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
521          if (ei.getType() == Graph::Edge::DUMMY)
522             continue;
523          if (bbA) {
524             bb->liveSet.setOr(&bbA->liveSet, &bbB->liveSet);
525             bbA = bb;
526          } else {
527             bbA = bbB;
528          }
529          bbB = BasicBlock::get(ei.getNode());
530       }
531       bb->liveSet.setOr(&bbB->liveSet, bbA ? &bbA->liveSet : NULL);
532    } else
533    if (bb->cfg.incidentCount()) {
534       bb->liveSet.fill(0);
535    }
536 }
537 
538 bool
visit(BasicBlock * bb)539 RegAlloc::BuildIntervalsPass::visit(BasicBlock *bb)
540 {
541    collectLiveValues(bb);
542 
543    INFO_DBG(prog->dbgFlags, REG_ALLOC, "BuildIntervals(BB:%i)\n", bb->getId());
544 
545    // go through out blocks and delete phi sources that do not originate from
546    // the current block from the live set
547    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
548       BasicBlock *out = BasicBlock::get(ei.getNode());
549 
550       for (Instruction *i = out->getPhi(); i && i->op == OP_PHI; i = i->next) {
551          bb->liveSet.clr(i->getDef(0)->id);
552 
553          for (int s = 0; i->srcExists(s); ++s) {
554             assert(i->src(s).getInsn());
555             if (i->getSrc(s)->getUniqueInsn()->bb == bb) // XXX: reachableBy ?
556                bb->liveSet.set(i->getSrc(s)->id);
557             else
558                bb->liveSet.clr(i->getSrc(s)->id);
559          }
560       }
561    }
562 
563    // remaining live-outs are live until end
564    if (bb->getExit()) {
565       for (unsigned int j = 0; j < bb->liveSet.getSize(); ++j)
566          if (bb->liveSet.test(j))
567             addLiveRange(func->getLValue(j), bb, bb->getExit()->serial + 1);
568    }
569 
570    for (Instruction *i = bb->getExit(); i && i->op != OP_PHI; i = i->prev) {
571       for (int d = 0; i->defExists(d); ++d) {
572          bb->liveSet.clr(i->getDef(d)->id);
573          if (i->getDef(d)->reg.data.id >= 0) // add hazard for fixed regs
574             i->getDef(d)->livei.extend(i->serial, i->serial);
575       }
576 
577       for (int s = 0; i->srcExists(s); ++s) {
578          if (!i->getSrc(s)->asLValue())
579             continue;
580          if (!bb->liveSet.test(i->getSrc(s)->id)) {
581             bb->liveSet.set(i->getSrc(s)->id);
582             addLiveRange(i->getSrc(s), bb, i->serial);
583          }
584       }
585    }
586 
587    if (bb == BasicBlock::get(func->cfg.getRoot())) {
588       for (std::deque<ValueDef>::iterator it = func->ins.begin();
589            it != func->ins.end(); ++it) {
590          if (it->get()->reg.data.id >= 0) // add hazard for fixed regs
591             it->get()->livei.extend(0, 1);
592       }
593    }
594 
595    return true;
596 }
597 
598 
599 #define JOIN_MASK_PHI        (1 << 0)
600 #define JOIN_MASK_UNION      (1 << 1)
601 #define JOIN_MASK_MOV        (1 << 2)
602 #define JOIN_MASK_TEX        (1 << 3)
603 
604 class GCRA
605 {
606 public:
607    GCRA(Function *, SpillCodeInserter&);
608    ~GCRA();
609 
610    bool allocateRegisters(ArrayList& insns);
611 
612    void printNodeInfo() const;
613 
614 private:
615    class RIG_Node : public Graph::Node
616    {
617    public:
618       RIG_Node();
619 
620       void init(const RegisterSet&, LValue *);
621 
622       void addInterference(RIG_Node *);
623       void addRegPreference(RIG_Node *);
624 
getValue() const625       inline LValue *getValue() const
626       {
627          return reinterpret_cast<LValue *>(data);
628       }
setValue(LValue * lval)629       inline void setValue(LValue *lval) { data = lval; }
630 
getCompMask() const631       inline uint8_t getCompMask() const
632       {
633          return ((1 << colors) - 1) << (reg & 7);
634       }
635 
get(const Graph::EdgeIterator & ei)636       static inline RIG_Node *get(const Graph::EdgeIterator& ei)
637       {
638          return static_cast<RIG_Node *>(ei.getNode());
639       }
640 
641    public:
642       uint32_t degree;
643       uint16_t degreeLimit; // if deg < degLimit, node is trivially colourable
644       uint16_t colors;
645 
646       DataFile f;
647       int32_t reg;
648 
649       float weight;
650 
651       // list pointers for simplify() phase
652       RIG_Node *next;
653       RIG_Node *prev;
654 
655       // union of the live intervals of all coalesced values (we want to retain
656       //  the separate intervals for testing interference of compound values)
657       Interval livei;
658 
659       std::list<RIG_Node *> prefRegs;
660    };
661 
662 private:
getNode(const LValue * v) const663    inline RIG_Node *getNode(const LValue *v) const { return &nodes[v->id]; }
664 
665    void buildRIG(ArrayList&);
666    bool coalesce(ArrayList&);
667    bool doCoalesce(ArrayList&, unsigned int mask);
668    void calculateSpillWeights();
669    void simplify();
670    bool selectRegisters();
671    void cleanup(const bool success);
672 
673    void simplifyEdge(RIG_Node *, RIG_Node *);
674    void simplifyNode(RIG_Node *);
675 
676    bool coalesceValues(Value *, Value *, bool force);
677    void resolveSplitsAndMerges();
678    void makeCompound(Instruction *, bool isSplit);
679 
680    inline void checkInterference(const RIG_Node *, Graph::EdgeIterator&);
681 
682    inline void insertOrderedTail(std::list<RIG_Node *>&, RIG_Node *);
683    void checkList(std::list<RIG_Node *>&);
684 
685 private:
686    std::stack<uint32_t> stack;
687 
688    // list headers for simplify() phase
689    RIG_Node lo[2];
690    RIG_Node hi;
691 
692    Graph RIG;
693    RIG_Node *nodes;
694    unsigned int nodeCount;
695 
696    Function *func;
697    Program *prog;
698 
699    static uint8_t relDegree[17][17];
700 
701    RegisterSet regs;
702 
703    // need to fixup register id for participants of OP_MERGE/SPLIT
704    std::list<Instruction *> merges;
705    std::list<Instruction *> splits;
706 
707    SpillCodeInserter& spill;
708    std::list<ValuePair> mustSpill;
709 };
710 
711 uint8_t GCRA::relDegree[17][17];
712 
RIG_Node()713 GCRA::RIG_Node::RIG_Node() : Node(NULL), next(this), prev(this)
714 {
715    colors = 0;
716 }
717 
718 void
printNodeInfo() const719 GCRA::printNodeInfo() const
720 {
721    for (unsigned int i = 0; i < nodeCount; ++i) {
722       if (!nodes[i].colors)
723          continue;
724       INFO("RIG_Node[%%%i]($[%u]%i): %u colors, weight %f, deg %u/%u\n X",
725            i,
726            nodes[i].f,nodes[i].reg,nodes[i].colors,
727            nodes[i].weight,
728            nodes[i].degree, nodes[i].degreeLimit);
729 
730       for (Graph::EdgeIterator ei = nodes[i].outgoing(); !ei.end(); ei.next())
731          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
732       for (Graph::EdgeIterator ei = nodes[i].incident(); !ei.end(); ei.next())
733          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
734       INFO("\n");
735    }
736 }
737 
738 void
init(const RegisterSet & regs,LValue * lval)739 GCRA::RIG_Node::init(const RegisterSet& regs, LValue *lval)
740 {
741    setValue(lval);
742    if (lval->reg.data.id >= 0)
743       lval->noSpill = lval->fixedReg = 1;
744 
745    colors = regs.units(lval->reg.file, lval->reg.size);
746    f = lval->reg.file;
747    reg = -1;
748    if (lval->reg.data.id >= 0)
749       reg = regs.idToUnits(lval);
750 
751    weight = std::numeric_limits<float>::infinity();
752    degree = 0;
753    degreeLimit = regs.getFileSize(f, lval->reg.size);
754 
755    livei.insert(lval->livei);
756 }
757 
758 bool
coalesceValues(Value * dst,Value * src,bool force)759 GCRA::coalesceValues(Value *dst, Value *src, bool force)
760 {
761    LValue *rep = dst->join->asLValue();
762    LValue *val = src->join->asLValue();
763 
764    if (!force && val->reg.data.id >= 0) {
765       rep = src->join->asLValue();
766       val = dst->join->asLValue();
767    }
768    RIG_Node *nRep = &nodes[rep->id];
769    RIG_Node *nVal = &nodes[val->id];
770 
771    if (src->reg.file != dst->reg.file) {
772       if (!force)
773          return false;
774       WARN("forced coalescing of values in different files !\n");
775    }
776    if (!force && dst->reg.size != src->reg.size)
777       return false;
778 
779    if ((rep->reg.data.id >= 0) && (rep->reg.data.id != val->reg.data.id)) {
780       if (force) {
781          if (val->reg.data.id >= 0)
782             WARN("forced coalescing of values in different fixed regs !\n");
783       } else {
784          if (val->reg.data.id >= 0)
785             return false;
786          // make sure that there is no overlap with the fixed register of rep
787          for (ArrayList::Iterator it = func->allLValues.iterator();
788               !it.end(); it.next()) {
789             Value *reg = reinterpret_cast<Value *>(it.get())->asLValue();
790             assert(reg);
791             if (reg->interfers(rep) && reg->livei.overlaps(nVal->livei))
792                return false;
793          }
794       }
795    }
796 
797    if (!force && nRep->livei.overlaps(nVal->livei))
798       return false;
799 
800    INFO_DBG(prog->dbgFlags, REG_ALLOC, "joining %%%i($%i) <- %%%i\n",
801             rep->id, rep->reg.data.id, val->id);
802 
803    // set join pointer of all values joined with val
804    for (Value::DefIterator def = val->defs.begin(); def != val->defs.end();
805         ++def)
806       (*def)->get()->join = rep;
807    assert(rep->join == rep && val->join == rep);
808 
809    // add val's definitions to rep and extend the live interval of its RIG node
810    rep->defs.insert(rep->defs.end(), val->defs.begin(), val->defs.end());
811    nRep->livei.unify(nVal->livei);
812    return true;
813 }
814 
815 bool
coalesce(ArrayList & insns)816 GCRA::coalesce(ArrayList& insns)
817 {
818    bool ret = doCoalesce(insns, JOIN_MASK_PHI);
819    if (!ret)
820       return false;
821    switch (func->getProgram()->getTarget()->getChipset() & ~0xf) {
822    case 0x50:
823    case 0x80:
824    case 0x90:
825    case 0xa0:
826       ret = doCoalesce(insns, JOIN_MASK_UNION | JOIN_MASK_TEX);
827       break;
828    case 0xc0:
829    case 0xd0:
830    case 0xe0:
831       ret = doCoalesce(insns, JOIN_MASK_UNION);
832       break;
833    default:
834       break;
835    }
836    if (!ret)
837       return false;
838    return doCoalesce(insns, JOIN_MASK_MOV);
839 }
840 
makeCompMask(int compSize,int base,int size)841 static inline uint8_t makeCompMask(int compSize, int base, int size)
842 {
843    uint8_t m = ((1 << size) - 1) << base;
844 
845    switch (compSize) {
846    case 1:
847       return 0xff;
848    case 2:
849       m |= (m << 2);
850       return (m << 4) | m;
851    case 3:
852    case 4:
853       return (m << 4) | m;
854    default:
855       assert(compSize <= 8);
856       return m;
857    }
858 }
859 
copyCompound(Value * dst,Value * src)860 static inline void copyCompound(Value *dst, Value *src)
861 {
862    LValue *ldst = dst->asLValue();
863    LValue *lsrc = src->asLValue();
864 
865    ldst->compound = lsrc->compound;
866    ldst->compMask = lsrc->compMask;
867 }
868 
869 void
makeCompound(Instruction * insn,bool split)870 GCRA::makeCompound(Instruction *insn, bool split)
871 {
872    LValue *rep = (split ? insn->getSrc(0) : insn->getDef(0))->asLValue();
873 
874    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
875       INFO("makeCompound(split = %i): ", split);
876       insn->print();
877    }
878 
879    const unsigned int size = getNode(rep)->colors;
880    unsigned int base = 0;
881 
882    if (!rep->compound)
883       rep->compMask = 0xff;
884    rep->compound = 1;
885 
886    for (int c = 0; split ? insn->defExists(c) : insn->srcExists(c); ++c) {
887       LValue *val = (split ? insn->getDef(c) : insn->getSrc(c))->asLValue();
888 
889       val->compound = 1;
890       if (!val->compMask)
891          val->compMask = 0xff;
892       val->compMask &= makeCompMask(size, base, getNode(val)->colors);
893       assert(val->compMask);
894 
895       INFO_DBG(prog->dbgFlags, REG_ALLOC, "compound: %%%i:%02x <- %%%i:%02x\n",
896            rep->id, rep->compMask, val->id, val->compMask);
897 
898       base += getNode(val)->colors;
899    }
900    assert(base == size);
901 }
902 
903 bool
doCoalesce(ArrayList & insns,unsigned int mask)904 GCRA::doCoalesce(ArrayList& insns, unsigned int mask)
905 {
906    int c, n;
907 
908    for (n = 0; n < insns.getSize(); ++n) {
909       Instruction *i;
910       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(n));
911 
912       switch (insn->op) {
913       case OP_PHI:
914          if (!(mask & JOIN_MASK_PHI))
915             break;
916          for (c = 0; insn->srcExists(c); ++c)
917             if (!coalesceValues(insn->getDef(0), insn->getSrc(c), false)) {
918                // this is bad
919                ERROR("failed to coalesce phi operands\n");
920                return false;
921             }
922          break;
923       case OP_UNION:
924       case OP_MERGE:
925          if (!(mask & JOIN_MASK_UNION))
926             break;
927          for (c = 0; insn->srcExists(c); ++c)
928             coalesceValues(insn->getDef(0), insn->getSrc(c), true);
929          if (insn->op == OP_MERGE) {
930             merges.push_back(insn);
931             if (insn->srcExists(1))
932                makeCompound(insn, false);
933          }
934          break;
935       case OP_SPLIT:
936          if (!(mask & JOIN_MASK_UNION))
937             break;
938          splits.push_back(insn);
939          for (c = 0; insn->defExists(c); ++c)
940             coalesceValues(insn->getSrc(0), insn->getDef(c), true);
941          makeCompound(insn, true);
942          break;
943       case OP_MOV:
944          if (!(mask & JOIN_MASK_MOV))
945             break;
946          i = NULL;
947          if (!insn->getDef(0)->uses.empty())
948             i = insn->getDef(0)->uses.front()->getInsn();
949          // if this is a contraint-move there will only be a single use
950          if (i && i->op == OP_MERGE) // do we really still need this ?
951             break;
952          i = insn->getSrc(0)->getUniqueInsn();
953          if (i && !i->constrainedDefs()) {
954             if (coalesceValues(insn->getDef(0), insn->getSrc(0), false))
955                copyCompound(insn->getSrc(0), insn->getDef(0));
956          }
957          break;
958       case OP_TEX:
959       case OP_TXB:
960       case OP_TXL:
961       case OP_TXF:
962       case OP_TXQ:
963       case OP_TXD:
964       case OP_TXG:
965       case OP_TEXCSAA:
966          if (!(mask & JOIN_MASK_TEX))
967             break;
968          for (c = 0; insn->srcExists(c) && c != insn->predSrc; ++c)
969             coalesceValues(insn->getDef(c), insn->getSrc(c), true);
970          break;
971       default:
972          break;
973       }
974    }
975    return true;
976 }
977 
978 void
addInterference(RIG_Node * node)979 GCRA::RIG_Node::addInterference(RIG_Node *node)
980 {
981    this->degree += relDegree[node->colors][colors];
982    node->degree += relDegree[colors][node->colors];
983 
984    this->attach(node, Graph::Edge::CROSS);
985 }
986 
987 void
addRegPreference(RIG_Node * node)988 GCRA::RIG_Node::addRegPreference(RIG_Node *node)
989 {
990    prefRegs.push_back(node);
991 }
992 
GCRA(Function * fn,SpillCodeInserter & spill)993 GCRA::GCRA(Function *fn, SpillCodeInserter& spill) :
994    func(fn),
995    regs(fn->getProgram()->getTarget()),
996    spill(spill)
997 {
998    prog = func->getProgram();
999 
1000    // initialize relative degrees array - i takes away from j
1001    for (int i = 1; i <= 16; ++i)
1002       for (int j = 1; j <= 16; ++j)
1003          relDegree[i][j] = j * ((i + j - 1) / j);
1004 }
1005 
~GCRA()1006 GCRA::~GCRA()
1007 {
1008    if (nodes)
1009       delete[] nodes;
1010 }
1011 
1012 void
checkList(std::list<RIG_Node * > & lst)1013 GCRA::checkList(std::list<RIG_Node *>& lst)
1014 {
1015    GCRA::RIG_Node *prev = NULL;
1016 
1017    for (std::list<RIG_Node *>::iterator it = lst.begin();
1018         it != lst.end();
1019         ++it) {
1020       assert((*it)->getValue()->join == (*it)->getValue());
1021       if (prev)
1022          assert(prev->livei.begin() <= (*it)->livei.begin());
1023       prev = *it;
1024    }
1025 }
1026 
1027 void
insertOrderedTail(std::list<RIG_Node * > & list,RIG_Node * node)1028 GCRA::insertOrderedTail(std::list<RIG_Node *>& list, RIG_Node *node)
1029 {
1030    if (node->livei.isEmpty())
1031       return;
1032    // only the intervals of joined values don't necessarily arrive in order
1033    std::list<RIG_Node *>::iterator prev, it;
1034    for (it = list.end(); it != list.begin(); it = prev) {
1035       prev = it;
1036       --prev;
1037       if ((*prev)->livei.begin() <= node->livei.begin())
1038          break;
1039    }
1040    list.insert(it, node);
1041 }
1042 
1043 void
buildRIG(ArrayList & insns)1044 GCRA::buildRIG(ArrayList& insns)
1045 {
1046    std::list<RIG_Node *> values, active;
1047 
1048    for (std::deque<ValueDef>::iterator it = func->ins.begin();
1049         it != func->ins.end(); ++it)
1050       insertOrderedTail(values, getNode(it->get()->asLValue()));
1051 
1052    for (int i = 0; i < insns.getSize(); ++i) {
1053       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(i));
1054       for (int d = 0; insn->defExists(d); ++d)
1055          if (insn->getDef(d)->rep() == insn->getDef(d))
1056             insertOrderedTail(values, getNode(insn->getDef(d)->asLValue()));
1057    }
1058    checkList(values);
1059 
1060    while (!values.empty()) {
1061       RIG_Node *cur = values.front();
1062 
1063       for (std::list<RIG_Node *>::iterator it = active.begin();
1064            it != active.end();
1065            ++it) {
1066          RIG_Node *node = *it;
1067 
1068          if (node->livei.end() <= cur->livei.begin()) {
1069             it = active.erase(it);
1070             --it;
1071          } else
1072          if (node->f == cur->f && node->livei.overlaps(cur->livei)) {
1073             cur->addInterference(node);
1074          }
1075       }
1076       values.pop_front();
1077       active.push_back(cur);
1078    }
1079 }
1080 
1081 void
calculateSpillWeights()1082 GCRA::calculateSpillWeights()
1083 {
1084    for (unsigned int i = 0; i < nodeCount; ++i) {
1085       RIG_Node *const n = &nodes[i];
1086       if (!nodes[i].colors || nodes[i].livei.isEmpty())
1087          continue;
1088       if (nodes[i].reg >= 0) {
1089          // update max reg
1090          regs.occupy(n->f, n->reg, n->colors);
1091          continue;
1092       }
1093       LValue *val = nodes[i].getValue();
1094 
1095       if (!val->noSpill) {
1096          int rc = 0;
1097          for (Value::DefIterator it = val->defs.begin();
1098               it != val->defs.end();
1099               ++it)
1100             rc += (*it)->get()->refCount();
1101 
1102          nodes[i].weight =
1103             (float)rc * (float)rc / (float)nodes[i].livei.extent();
1104       }
1105 
1106       if (nodes[i].degree < nodes[i].degreeLimit) {
1107          int l = 0;
1108          if (val->reg.size > 4)
1109             l = 1;
1110          DLLIST_ADDHEAD(&lo[l], &nodes[i]);
1111       } else {
1112          DLLIST_ADDHEAD(&hi, &nodes[i]);
1113       }
1114    }
1115    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1116       printNodeInfo();
1117 }
1118 
1119 void
simplifyEdge(RIG_Node * a,RIG_Node * b)1120 GCRA::simplifyEdge(RIG_Node *a, RIG_Node *b)
1121 {
1122    bool move = b->degree >= b->degreeLimit;
1123 
1124    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1125             "edge: (%%%i, deg %u/%u) >-< (%%%i, deg %u/%u)\n",
1126             a->getValue()->id, a->degree, a->degreeLimit,
1127             b->getValue()->id, b->degree, b->degreeLimit);
1128 
1129    b->degree -= relDegree[a->colors][b->colors];
1130 
1131    move = move && b->degree < b->degreeLimit;
1132    if (move && !DLLIST_EMPTY(b)) {
1133       int l = (b->getValue()->reg.size > 4) ? 1 : 0;
1134       DLLIST_DEL(b);
1135       DLLIST_ADDTAIL(&lo[l], b);
1136    }
1137 }
1138 
1139 void
simplifyNode(RIG_Node * node)1140 GCRA::simplifyNode(RIG_Node *node)
1141 {
1142    for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1143       simplifyEdge(node, RIG_Node::get(ei));
1144 
1145    for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1146       simplifyEdge(node, RIG_Node::get(ei));
1147 
1148    DLLIST_DEL(node);
1149    stack.push(node->getValue()->id);
1150 
1151    INFO_DBG(prog->dbgFlags, REG_ALLOC, "SIMPLIFY: pushed %%%i%s\n",
1152             node->getValue()->id,
1153             (node->degree < node->degreeLimit) ? "" : "(spill)");
1154 }
1155 
1156 void
simplify()1157 GCRA::simplify()
1158 {
1159    for (;;) {
1160       if (!DLLIST_EMPTY(&lo[0])) {
1161          do {
1162             simplifyNode(lo[0].next);
1163          } while (!DLLIST_EMPTY(&lo[0]));
1164       } else
1165       if (!DLLIST_EMPTY(&lo[1])) {
1166          simplifyNode(lo[1].next);
1167       } else
1168       if (!DLLIST_EMPTY(&hi)) {
1169          RIG_Node *best = hi.next;
1170          float bestScore = best->weight / (float)best->degree;
1171          // spill candidate
1172          for (RIG_Node *it = best->next; it != &hi; it = it->next) {
1173             float score = it->weight / (float)it->degree;
1174             if (score < bestScore) {
1175                best = it;
1176                bestScore = score;
1177             }
1178          }
1179          if (isinf(bestScore)) {
1180             ERROR("no viable spill candidates left\n");
1181             break;
1182          }
1183          simplifyNode(best);
1184       } else {
1185          break;
1186       }
1187    }
1188 }
1189 
1190 void
checkInterference(const RIG_Node * node,Graph::EdgeIterator & ei)1191 GCRA::checkInterference(const RIG_Node *node, Graph::EdgeIterator& ei)
1192 {
1193    const RIG_Node *intf = RIG_Node::get(ei);
1194 
1195    if (intf->reg < 0)
1196       return;
1197    const LValue *vA = node->getValue();
1198    const LValue *vB = intf->getValue();
1199 
1200    const uint8_t intfMask = ((1 << intf->colors) - 1) << (intf->reg & 7);
1201 
1202    if (vA->compound | vB->compound) {
1203       // NOTE: this only works for >aligned< register tuples !
1204       for (Value::DefCIterator D = vA->defs.begin(); D != vA->defs.end(); ++D) {
1205       for (Value::DefCIterator d = vB->defs.begin(); d != vB->defs.end(); ++d) {
1206          const LValue *vD = (*D)->get()->asLValue();
1207          const LValue *vd = (*d)->get()->asLValue();
1208 
1209          if (!vD->livei.overlaps(vd->livei)) {
1210             INFO_DBG(prog->dbgFlags, REG_ALLOC, "(%%%i) X (%%%i): no overlap\n",
1211                      vD->id, vd->id);
1212             continue;
1213          }
1214 
1215          uint8_t mask = vD->compound ? vD->compMask : ~0;
1216          if (vd->compound) {
1217             assert(vB->compound);
1218             mask &= vd->compMask & vB->compMask;
1219          } else {
1220             mask &= intfMask;
1221          }
1222 
1223          INFO_DBG(prog->dbgFlags, REG_ALLOC,
1224                   "(%%%i)%02x X (%%%i)%02x & %02x: $r%i.%02x\n",
1225                   vD->id,
1226                   vD->compound ? vD->compMask : 0xff,
1227                   vd->id,
1228                   vd->compound ? vd->compMask : intfMask,
1229                   vB->compMask, intf->reg & ~7, mask);
1230          if (mask)
1231             regs.occupyMask(node->f, intf->reg & ~7, mask);
1232       }
1233       }
1234    } else {
1235       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1236                "(%%%i) X (%%%i): $r%i + %u\n",
1237                vA->id, vB->id, intf->reg, intf->colors);
1238       regs.occupy(node->f, intf->reg, intf->colors);
1239    }
1240 }
1241 
1242 bool
selectRegisters()1243 GCRA::selectRegisters()
1244 {
1245    INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nSELECT phase\n");
1246 
1247    while (!stack.empty()) {
1248       RIG_Node *node = &nodes[stack.top()];
1249       stack.pop();
1250 
1251       regs.reset(node->f);
1252 
1253       INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nNODE[%%%i, %u colors]\n",
1254                node->getValue()->id, node->colors);
1255 
1256       for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1257          checkInterference(node, ei);
1258       for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1259          checkInterference(node, ei);
1260 
1261       if (!node->prefRegs.empty()) {
1262          for (std::list<RIG_Node *>::const_iterator it = node->prefRegs.begin();
1263               it != node->prefRegs.end();
1264               ++it) {
1265             if ((*it)->reg >= 0 &&
1266                 regs.occupy(node->f, (*it)->reg, node->colors)) {
1267                node->reg = (*it)->reg;
1268                break;
1269             }
1270          }
1271       }
1272       if (node->reg >= 0)
1273          continue;
1274       LValue *lval = node->getValue();
1275       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1276          regs.print();
1277       bool ret = regs.assign(node->reg, node->f, node->colors);
1278       if (ret) {
1279          INFO_DBG(prog->dbgFlags, REG_ALLOC, "assigned reg %i\n", node->reg);
1280          lval->compMask = node->getCompMask();
1281       } else {
1282          INFO_DBG(prog->dbgFlags, REG_ALLOC, "must spill: %%%i (size %u)\n",
1283                   lval->id, lval->reg.size);
1284          Symbol *slot = NULL;
1285          if (lval->reg.file == FILE_GPR)
1286             slot = spill.assignSlot(node->livei, lval->reg.size);
1287          mustSpill.push_back(ValuePair(lval, slot));
1288       }
1289    }
1290    if (!mustSpill.empty())
1291       return false;
1292    for (unsigned int i = 0; i < nodeCount; ++i) {
1293       LValue *lval = nodes[i].getValue();
1294       if (nodes[i].reg >= 0 && nodes[i].colors > 0)
1295          lval->reg.data.id =
1296             regs.unitsToId(nodes[i].f, nodes[i].reg, lval->reg.size);
1297    }
1298    return true;
1299 }
1300 
1301 bool
allocateRegisters(ArrayList & insns)1302 GCRA::allocateRegisters(ArrayList& insns)
1303 {
1304    bool ret;
1305 
1306    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1307             "allocateRegisters to %u instructions\n", insns.getSize());
1308 
1309    nodeCount = func->allLValues.getSize();
1310    nodes = new RIG_Node[nodeCount];
1311    if (!nodes)
1312       return false;
1313    for (unsigned int i = 0; i < nodeCount; ++i) {
1314       LValue *lval = reinterpret_cast<LValue *>(func->allLValues.get(i));
1315       if (lval) {
1316          nodes[i].init(regs, lval);
1317          RIG.insert(&nodes[i]);
1318       }
1319    }
1320 
1321    // coalesce first, we use only 1 RIG node for a group of joined values
1322    ret = coalesce(insns);
1323    if (!ret)
1324       goto out;
1325 
1326    if (func->getProgram()->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1327       func->printLiveIntervals();
1328 
1329    buildRIG(insns);
1330    calculateSpillWeights();
1331    simplify();
1332 
1333    ret = selectRegisters();
1334    if (!ret) {
1335       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1336                "selectRegisters failed, inserting spill code ...\n");
1337       regs.reset(FILE_GPR, true);
1338       spill.run(mustSpill);
1339       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1340          func->print();
1341    } else {
1342       prog->maxGPR = regs.getMaxAssigned(FILE_GPR);
1343    }
1344 
1345 out:
1346    cleanup(ret);
1347    return ret;
1348 }
1349 
1350 void
cleanup(const bool success)1351 GCRA::cleanup(const bool success)
1352 {
1353    mustSpill.clear();
1354 
1355    for (ArrayList::Iterator it = func->allLValues.iterator();
1356         !it.end(); it.next()) {
1357       LValue *lval =  reinterpret_cast<LValue *>(it.get());
1358 
1359       lval->livei.clear();
1360 
1361       lval->compound = 0;
1362       lval->compMask = 0;
1363 
1364       if (lval->join == lval)
1365          continue;
1366 
1367       if (success) {
1368          lval->reg.data.id = lval->join->reg.data.id;
1369       } else {
1370          for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1371               ++d)
1372             lval->join->defs.remove(*d);
1373          lval->join = lval;
1374       }
1375    }
1376 
1377    if (success)
1378       resolveSplitsAndMerges();
1379    splits.clear(); // avoid duplicate entries on next coalesce pass
1380    merges.clear();
1381 
1382    delete[] nodes;
1383    nodes = NULL;
1384 }
1385 
1386 Symbol *
assignSlot(const Interval & livei,unsigned int size)1387 SpillCodeInserter::assignSlot(const Interval &livei, unsigned int size)
1388 {
1389    SpillSlot slot;
1390    int32_t offsetBase = stackSize;
1391    int32_t offset;
1392    std::list<SpillSlot>::iterator pos = slots.end(), it = slots.begin();
1393 
1394    if (offsetBase % size)
1395       offsetBase += size - (offsetBase % size);
1396 
1397    slot.sym = NULL;
1398 
1399    for (offset = offsetBase; offset < stackSize; offset += size) {
1400       while (it != slots.end() && it->offset < offset)
1401          ++it;
1402       if (it == slots.end()) // no slots left
1403          break;
1404       std::list<SpillSlot>::iterator bgn = it;
1405 
1406       while (it != slots.end() && it->offset < (offset + size)) {
1407          it->occup.print();
1408          if (it->occup.overlaps(livei))
1409             break;
1410          ++it;
1411       }
1412       if (it == slots.end() || it->offset >= (offset + size)) {
1413          // fits
1414          for (; bgn != slots.end() && bgn->offset < (offset + size); ++bgn) {
1415             bgn->occup.insert(livei);
1416             if (bgn->size() == size)
1417                slot.sym = bgn->sym;
1418          }
1419          break;
1420       }
1421    }
1422    if (!slot.sym) {
1423       stackSize = offset + size;
1424       slot.offset = offset;
1425       slot.sym = new_Symbol(func->getProgram(), FILE_MEMORY_LOCAL);
1426       if (!func->stackPtr)
1427          offset += func->tlsBase;
1428       slot.sym->setAddress(NULL, offset);
1429       slot.sym->reg.size = size;
1430       slots.insert(pos, slot)->occup.insert(livei);
1431    }
1432    return slot.sym;
1433 }
1434 
1435 void
spill(Instruction * defi,Value * slot,LValue * lval)1436 SpillCodeInserter::spill(Instruction *defi, Value *slot, LValue *lval)
1437 {
1438    const DataType ty = typeOfSize(slot->reg.size);
1439 
1440    Instruction *st;
1441    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1442       st = new_Instruction(func, OP_STORE, ty);
1443       st->setSrc(0, slot);
1444       st->setSrc(1, lval);
1445       lval->noSpill = 1;
1446    } else {
1447       st = new_Instruction(func, OP_CVT, ty);
1448       st->setDef(0, slot);
1449       st->setSrc(0, lval);
1450    }
1451    defi->bb->insertAfter(defi, st);
1452 }
1453 
1454 LValue *
unspill(Instruction * usei,LValue * lval,Value * slot)1455 SpillCodeInserter::unspill(Instruction *usei, LValue *lval, Value *slot)
1456 {
1457    const DataType ty = typeOfSize(slot->reg.size);
1458 
1459    lval = cloneShallow(func, lval);
1460 
1461    Instruction *ld;
1462    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1463       lval->noSpill = 1;
1464       ld = new_Instruction(func, OP_LOAD, ty);
1465    } else {
1466       ld = new_Instruction(func, OP_CVT, ty);
1467    }
1468    ld->setDef(0, lval);
1469    ld->setSrc(0, slot);
1470 
1471    usei->bb->insertBefore(usei, ld);
1472    return lval;
1473 }
1474 
1475 bool
run(const std::list<ValuePair> & lst)1476 SpillCodeInserter::run(const std::list<ValuePair>& lst)
1477 {
1478    for (std::list<ValuePair>::const_iterator it = lst.begin(); it != lst.end();
1479         ++it) {
1480       LValue *lval = it->first->asLValue();
1481       Symbol *mem = it->second ? it->second->asSym() : NULL;
1482 
1483       for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1484            ++d) {
1485          Value *slot = mem ?
1486             static_cast<Value *>(mem) : new_LValue(func, FILE_GPR);
1487          Value *tmp = NULL;
1488          Instruction *last = NULL;
1489 
1490          LValue *dval = (*d)->get()->asLValue();
1491          Instruction *defi = (*d)->getInsn();
1492 
1493          // handle uses first or they'll contain the spill stores
1494          while (!dval->uses.empty()) {
1495             ValueRef *u = dval->uses.front();
1496             Instruction *usei = u->getInsn();
1497             assert(usei);
1498             if (usei->op == OP_PHI) {
1499                tmp = (slot->reg.file == FILE_MEMORY_LOCAL) ? NULL : slot;
1500                last = NULL;
1501             } else
1502             if (!last || usei != last->next) { // TODO: sort uses
1503                tmp = unspill(usei, dval, slot);
1504                last = usei;
1505             }
1506             u->set(tmp);
1507          }
1508 
1509          assert(defi);
1510          if (defi->op == OP_PHI) {
1511             d = lval->defs.erase(d);
1512             --d;
1513             if (slot->reg.file == FILE_MEMORY_LOCAL)
1514                delete_Instruction(func->getProgram(), defi);
1515             else
1516                defi->setDef(0, slot);
1517          } else {
1518             spill(defi, slot, dval);
1519          }
1520       }
1521 
1522    }
1523 
1524    // TODO: We're not trying to reuse old slots in a potential next iteration.
1525    //  We have to update the slots' livei intervals to be able to do that.
1526    stackBase = stackSize;
1527    slots.clear();
1528    return true;
1529 }
1530 
1531 bool
exec()1532 RegAlloc::exec()
1533 {
1534    for (IteratorRef it = prog->calls.iteratorDFS(false);
1535         !it->end(); it->next()) {
1536       func = Function::get(reinterpret_cast<Graph::Node *>(it->get()));
1537 
1538       func->tlsBase = prog->tlsSize;
1539       if (!execFunc())
1540          return false;
1541       prog->tlsSize += func->tlsSize;
1542    }
1543    return true;
1544 }
1545 
1546 bool
execFunc()1547 RegAlloc::execFunc()
1548 {
1549    InsertConstraintsPass insertConstr;
1550    PhiMovesPass insertPhiMoves;
1551    ArgumentMovesPass insertArgMoves;
1552    BuildIntervalsPass buildIntervals;
1553    SpillCodeInserter insertSpills(func);
1554 
1555    GCRA gcra(func, insertSpills);
1556 
1557    unsigned int i, retries;
1558    bool ret;
1559 
1560    ret = insertConstr.exec(func);
1561    if (!ret)
1562       goto out;
1563 
1564    ret = insertPhiMoves.run(func);
1565    if (!ret)
1566       goto out;
1567 
1568    ret = insertArgMoves.run(func);
1569    if (!ret)
1570       goto out;
1571 
1572    // TODO: need to fix up spill slot usage ranges to support > 1 retry
1573    for (retries = 0; retries < 3; ++retries) {
1574       if (retries && (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC))
1575          INFO("Retry: %i\n", retries);
1576       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1577          func->print();
1578 
1579       // spilling to registers may add live ranges, need to rebuild everything
1580       ret = true;
1581       for (sequence = func->cfg.nextSequence(), i = 0;
1582            ret && i <= func->loopNestingBound;
1583            sequence = func->cfg.nextSequence(), ++i)
1584          ret = buildLiveSets(BasicBlock::get(func->cfg.getRoot()));
1585       if (!ret)
1586          break;
1587       func->orderInstructions(this->insns);
1588 
1589       ret = buildIntervals.run(func);
1590       if (!ret)
1591          break;
1592       ret = gcra.allocateRegisters(insns);
1593       if (ret)
1594          break; // success
1595    }
1596    INFO_DBG(prog->dbgFlags, REG_ALLOC, "RegAlloc done: %i\n", ret);
1597 
1598    func->tlsSize = insertSpills.getStackSize();
1599 out:
1600    return ret;
1601 }
1602 
1603 // TODO: check if modifying Instruction::join here breaks anything
1604 void
resolveSplitsAndMerges()1605 GCRA::resolveSplitsAndMerges()
1606 {
1607    for (std::list<Instruction *>::iterator it = splits.begin();
1608         it != splits.end();
1609         ++it) {
1610       Instruction *split = *it;
1611       unsigned int reg = regs.idToBytes(split->getSrc(0));
1612       for (int d = 0; split->defExists(d); ++d) {
1613          Value *v = split->getDef(d);
1614          v->reg.data.id = regs.bytesToId(v, reg);
1615          v->join = v;
1616          reg += v->reg.size;
1617       }
1618    }
1619    splits.clear();
1620 
1621    for (std::list<Instruction *>::iterator it = merges.begin();
1622         it != merges.end();
1623         ++it) {
1624       Instruction *merge = *it;
1625       unsigned int reg = regs.idToBytes(merge->getDef(0));
1626       for (int s = 0; merge->srcExists(s); ++s) {
1627          Value *v = merge->getSrc(s);
1628          v->reg.data.id = regs.bytesToId(v, reg);
1629          v->join = v;
1630          reg += v->reg.size;
1631       }
1632    }
1633    merges.clear();
1634 }
1635 
registerAllocation()1636 bool Program::registerAllocation()
1637 {
1638    RegAlloc ra(this);
1639    return ra.exec();
1640 }
1641 
1642 bool
exec(Function * ir)1643 RegAlloc::InsertConstraintsPass::exec(Function *ir)
1644 {
1645    constrList.clear();
1646 
1647    bool ret = run(ir, true, true);
1648    if (ret)
1649       ret = insertConstraintMoves();
1650    return ret;
1651 }
1652 
1653 // TODO: make part of texture insn
1654 void
textureMask(TexInstruction * tex)1655 RegAlloc::InsertConstraintsPass::textureMask(TexInstruction *tex)
1656 {
1657    Value *def[4];
1658    int c, k, d;
1659    uint8_t mask = 0;
1660 
1661    for (d = 0, k = 0, c = 0; c < 4; ++c) {
1662       if (!(tex->tex.mask & (1 << c)))
1663          continue;
1664       if (tex->getDef(k)->refCount()) {
1665          mask |= 1 << c;
1666          def[d++] = tex->getDef(k);
1667       }
1668       ++k;
1669    }
1670    tex->tex.mask = mask;
1671 
1672    for (c = 0; c < d; ++c)
1673       tex->setDef(c, def[c]);
1674    for (; c < 4; ++c)
1675       tex->setDef(c, NULL);
1676 }
1677 
1678 bool
detectConflict(Instruction * cst,int s)1679 RegAlloc::InsertConstraintsPass::detectConflict(Instruction *cst, int s)
1680 {
1681    Value *v = cst->getSrc(s);
1682 
1683    // current register allocation can't handle it if a value participates in
1684    // multiple constraints
1685    for (Value::UseIterator it = v->uses.begin(); it != v->uses.end(); ++it) {
1686       if (cst != (*it)->getInsn())
1687          return true;
1688    }
1689 
1690    // can start at s + 1 because detectConflict is called on all sources
1691    for (int c = s + 1; cst->srcExists(c); ++c)
1692       if (v == cst->getSrc(c))
1693          return true;
1694 
1695    Instruction *defi = v->getInsn();
1696 
1697    return (!defi || defi->constrainedDefs());
1698 }
1699 
1700 void
addConstraint(Instruction * i,int s,int n)1701 RegAlloc::InsertConstraintsPass::addConstraint(Instruction *i, int s, int n)
1702 {
1703    Instruction *cst;
1704    int d;
1705 
1706    // first, look for an existing identical constraint op
1707    for (std::list<Instruction *>::iterator it = constrList.begin();
1708         it != constrList.end();
1709         ++it) {
1710       cst = (*it);
1711       if (!i->bb->dominatedBy(cst->bb))
1712          break;
1713       for (d = 0; d < n; ++d)
1714          if (cst->getSrc(d) != i->getSrc(d + s))
1715             break;
1716       if (d >= n) {
1717          for (d = 0; d < n; ++d, ++s)
1718             i->setSrc(s, cst->getDef(d));
1719          return;
1720       }
1721    }
1722    cst = new_Instruction(func, OP_CONSTRAINT, i->dType);
1723 
1724    for (d = 0; d < n; ++s, ++d) {
1725       cst->setDef(d, new_LValue(func, FILE_GPR));
1726       cst->setSrc(d, i->getSrc(s));
1727       i->setSrc(s, cst->getDef(d));
1728    }
1729    i->bb->insertBefore(i, cst);
1730 
1731    constrList.push_back(cst);
1732 }
1733 
1734 // Add a dummy use of the pointer source of >= 8 byte loads after the load
1735 // to prevent it from being assigned a register which overlapping the load's
1736 // destination, which would produce random corruptions.
1737 void
addHazard(Instruction * i,const ValueRef * src)1738 RegAlloc::InsertConstraintsPass::addHazard(Instruction *i, const ValueRef *src)
1739 {
1740    Instruction *hzd = new_Instruction(func, OP_NOP, TYPE_NONE);
1741    hzd->setSrc(0, src->get());
1742    i->bb->insertAfter(i, hzd);
1743 
1744 }
1745 
1746 // b32 { %r0 %r1 %r2 %r3 } -> b128 %r0q
1747 void
condenseDefs(Instruction * insn)1748 RegAlloc::InsertConstraintsPass::condenseDefs(Instruction *insn)
1749 {
1750    uint8_t size = 0;
1751    int n;
1752    for (n = 0; insn->defExists(n) && insn->def(n).getFile() == FILE_GPR; ++n)
1753       size += insn->getDef(n)->reg.size;
1754    if (n < 2)
1755       return;
1756    LValue *lval = new_LValue(func, FILE_GPR);
1757    lval->reg.size = size;
1758 
1759    Instruction *split = new_Instruction(func, OP_SPLIT, typeOfSize(size));
1760    split->setSrc(0, lval);
1761    for (int d = 0; d < n; ++d) {
1762       split->setDef(d, insn->getDef(d));
1763       insn->setDef(d, NULL);
1764    }
1765    insn->setDef(0, lval);
1766 
1767    for (int k = 1, d = n; insn->defExists(d); ++d, ++k) {
1768       insn->setDef(k, insn->getDef(d));
1769       insn->setDef(d, NULL);
1770    }
1771    // carry over predicate if any (mainly for OP_UNION uses)
1772    split->setPredicate(insn->cc, insn->getPredicate());
1773 
1774    insn->bb->insertAfter(insn, split);
1775    constrList.push_back(split);
1776 }
1777 void
condenseSrcs(Instruction * insn,const int a,const int b)1778 RegAlloc::InsertConstraintsPass::condenseSrcs(Instruction *insn,
1779                                               const int a, const int b)
1780 {
1781    uint8_t size = 0;
1782    if (a >= b)
1783       return;
1784    for (int s = a; s <= b; ++s)
1785       size += insn->getSrc(s)->reg.size;
1786    if (!size)
1787       return;
1788    LValue *lval = new_LValue(func, FILE_GPR);
1789    lval->reg.size = size;
1790 
1791    Value *save[3];
1792    insn->takeExtraSources(0, save);
1793 
1794    Instruction *merge = new_Instruction(func, OP_MERGE, typeOfSize(size));
1795    merge->setDef(0, lval);
1796    for (int s = a, i = 0; s <= b; ++s, ++i) {
1797       merge->setSrc(i, insn->getSrc(s));
1798       insn->setSrc(s, NULL);
1799    }
1800    insn->setSrc(a, lval);
1801 
1802    for (int k = a + 1, s = b + 1; insn->srcExists(s); ++s, ++k) {
1803       insn->setSrc(k, insn->getSrc(s));
1804       insn->setSrc(s, NULL);
1805    }
1806    insn->bb->insertBefore(insn, merge);
1807 
1808    insn->putExtraSources(0, save);
1809 
1810    constrList.push_back(merge);
1811 }
1812 
1813 void
texConstraintNVE0(TexInstruction * tex)1814 RegAlloc::InsertConstraintsPass::texConstraintNVE0(TexInstruction *tex)
1815 {
1816    textureMask(tex);
1817    condenseDefs(tex);
1818 
1819    int n = tex->srcCount(0xff, true);
1820    if (n > 4) {
1821       condenseSrcs(tex, 0, 3);
1822       if (n > 5) // NOTE: first call modified positions already
1823          condenseSrcs(tex, 4 - (4 - 1), n - 1 - (4 - 1));
1824    } else
1825    if (n > 1) {
1826       condenseSrcs(tex, 0, n - 1);
1827    }
1828 }
1829 
1830 void
texConstraintNVC0(TexInstruction * tex)1831 RegAlloc::InsertConstraintsPass::texConstraintNVC0(TexInstruction *tex)
1832 {
1833    int n, s;
1834 
1835    textureMask(tex);
1836 
1837    if (tex->op == OP_TXQ) {
1838       s = tex->srcCount(0xff);
1839       n = 0;
1840    } else {
1841       s = tex->tex.target.getArgCount();
1842       if (!tex->tex.target.isArray() &&
1843           (tex->tex.rIndirectSrc >= 0 || tex->tex.sIndirectSrc >= 0))
1844          ++s;
1845       if (tex->op == OP_TXD && tex->tex.useOffsets)
1846          ++s;
1847       n = tex->srcCount(0xff) - s;
1848       assert(n <= 4);
1849    }
1850 
1851    if (s > 1)
1852       condenseSrcs(tex, 0, s - 1);
1853    if (n > 1) // NOTE: first call modified positions already
1854       condenseSrcs(tex, 1, n);
1855 
1856    condenseDefs(tex);
1857 }
1858 
1859 void
texConstraintNV50(TexInstruction * tex)1860 RegAlloc::InsertConstraintsPass::texConstraintNV50(TexInstruction *tex)
1861 {
1862    Value *pred = tex->getPredicate();
1863    if (pred)
1864       tex->setPredicate(tex->cc, NULL);
1865 
1866    textureMask(tex);
1867 
1868    assert(tex->defExists(0) && tex->srcExists(0));
1869    // make src and def count match
1870    int c;
1871    for (c = 0; tex->srcExists(c) || tex->defExists(c); ++c) {
1872       if (!tex->srcExists(c))
1873          tex->setSrc(c, new_LValue(func, tex->getSrc(0)->asLValue()));
1874       if (!tex->defExists(c))
1875          tex->setDef(c, new_LValue(func, tex->getDef(0)->asLValue()));
1876    }
1877    if (pred)
1878       tex->setPredicate(tex->cc, pred);
1879    condenseDefs(tex);
1880    condenseSrcs(tex, 0, c - 1);
1881 }
1882 
1883 // Insert constraint markers for instructions whose multiple sources must be
1884 // located in consecutive registers.
1885 bool
visit(BasicBlock * bb)1886 RegAlloc::InsertConstraintsPass::visit(BasicBlock *bb)
1887 {
1888    TexInstruction *tex;
1889    Instruction *next;
1890    int s, size;
1891 
1892    targ = bb->getProgram()->getTarget();
1893 
1894    for (Instruction *i = bb->getEntry(); i; i = next) {
1895       next = i->next;
1896 
1897       if ((tex = i->asTex())) {
1898          switch (targ->getChipset() & ~0xf) {
1899          case 0x50:
1900          case 0x80:
1901          case 0x90:
1902          case 0xa0:
1903             texConstraintNV50(tex);
1904             break;
1905          case 0xc0:
1906          case 0xd0:
1907             texConstraintNVC0(tex);
1908             break;
1909          case 0xe0:
1910             texConstraintNVE0(tex);
1911             break;
1912          default:
1913             break;
1914          }
1915       } else
1916       if (i->op == OP_EXPORT || i->op == OP_STORE) {
1917          for (size = typeSizeof(i->dType), s = 1; size > 0; ++s) {
1918             assert(i->srcExists(s));
1919             size -= i->getSrc(s)->reg.size;
1920          }
1921          condenseSrcs(i, 1, s - 1);
1922       } else
1923       if (i->op == OP_LOAD || i->op == OP_VFETCH) {
1924          condenseDefs(i);
1925          if (i->src(0).isIndirect(0) && typeSizeof(i->dType) >= 8)
1926             addHazard(i, i->src(0).getIndirect(0));
1927       } else
1928       if (i->op == OP_UNION) {
1929          constrList.push_back(i);
1930       }
1931    }
1932    return true;
1933 }
1934 
1935 // Insert extra moves so that, if multiple register constraints on a value are
1936 // in conflict, these conflicts can be resolved.
1937 bool
insertConstraintMoves()1938 RegAlloc::InsertConstraintsPass::insertConstraintMoves()
1939 {
1940    for (std::list<Instruction *>::iterator it = constrList.begin();
1941         it != constrList.end();
1942         ++it) {
1943       Instruction *cst = *it;
1944       Instruction *mov;
1945 
1946       if (cst->op == OP_SPLIT && 0) {
1947          // spilling splits is annoying, just make sure they're separate
1948          for (int d = 0; cst->defExists(d); ++d) {
1949             if (!cst->getDef(d)->refCount())
1950                continue;
1951             LValue *lval = new_LValue(func, cst->def(d).getFile());
1952             const uint8_t size = cst->def(d).getSize();
1953             lval->reg.size = size;
1954 
1955             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
1956             mov->setSrc(0, lval);
1957             mov->setDef(0, cst->getDef(d));
1958             cst->setDef(d, mov->getSrc(0));
1959             cst->bb->insertAfter(cst, mov);
1960 
1961             cst->getSrc(0)->asLValue()->noSpill = 1;
1962             mov->getSrc(0)->asLValue()->noSpill = 1;
1963          }
1964       } else
1965       if (cst->op == OP_MERGE || cst->op == OP_UNION) {
1966          for (int s = 0; cst->srcExists(s); ++s) {
1967             const uint8_t size = cst->src(s).getSize();
1968 
1969             if (!cst->getSrc(s)->defs.size()) {
1970                mov = new_Instruction(func, OP_NOP, typeOfSize(size));
1971                mov->setDef(0, cst->getSrc(s));
1972                cst->bb->insertBefore(cst, mov);
1973                continue;
1974             }
1975             assert(cst->getSrc(s)->defs.size() == 1); // still SSA
1976 
1977             Instruction *defi = cst->getSrc(s)->defs.front()->getInsn();
1978             // catch some cases where don't really need MOVs
1979             if (cst->getSrc(s)->refCount() == 1 && !defi->constrainedDefs())
1980                continue;
1981 
1982             LValue *lval = new_LValue(func, cst->src(s).getFile());
1983             lval->reg.size = size;
1984 
1985             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
1986             mov->setDef(0, lval);
1987             mov->setSrc(0, cst->getSrc(s));
1988             cst->setSrc(s, mov->getDef(0));
1989             cst->bb->insertBefore(cst, mov);
1990 
1991             cst->getDef(0)->asLValue()->noSpill = 1; // doesn't help
1992 
1993             if (cst->op == OP_UNION)
1994                mov->setPredicate(defi->cc, defi->getPredicate());
1995          }
1996       }
1997    }
1998 
1999    return true;
2000 }
2001 
2002 } // namespace nv50_ir
2003