src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp

   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 OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include "codegen/nv50_ir.h"
  24 #include "codegen/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    void occupy(DataFile f, int32_t reg, unsigned int size);
  47    void occupy(const Value *);
  48    void occupyMask(DataFile f, int32_t reg, uint8_t mask);
  49    bool isOccupied(DataFile f, int32_t reg, unsigned int size) const;
  50    bool testOccupy(const Value *);
  51    bool testOccupy(DataFile f, int32_t reg, unsigned int size);
  52
  53    inline int getMaxAssigned(DataFile f) const { return fill[f]; }
  54
  55    inline unsigned int getFileSize(DataFile f, uint8_t regSize) const
  56    {
  57       if (restrictedGPR16Range && f == FILE_GPR && regSize == 2)
  58          return (last[f] + 1) / 2;
  59       return last[f] + 1;
  60    }
  61
  62    inline unsigned int units(DataFile f, unsigned int size) const
  63    {
  64       return size >> unit[f];
  65    }
  66    // for regs of size >= 4, id is counted in 4-byte words (like nv50/c0 binary)
  67    inline unsigned int idToBytes(const Value *v) const
  68    {
  69       return v->reg.data.id * MIN2(v->reg.size, 4);
  70    }
  71    inline unsigned int idToUnits(const Value *v) const
  72    {
  73       return units(v->reg.file, idToBytes(v));
  74    }
  75    inline int bytesToId(Value *v, unsigned int bytes) const
  76    {
  77       if (v->reg.size < 4)
  78          return units(v->reg.file, bytes);
  79       return bytes / 4;
  80    }
  81    inline int unitsToId(DataFile f, int u, uint8_t size) const
  82    {
  83       if (u < 0)
  84          return -1;
  85       return (size < 4) ? u : ((u << unit[f]) / 4);
  86    }
  87
  88    void print() const;
  89
  90 private:
  91    BitSet bits[LAST_REGISTER_FILE + 1];
  92
  93    int unit[LAST_REGISTER_FILE + 1]; // log2 of allocation granularity
  94
  95    int last[LAST_REGISTER_FILE + 1];
  96    int fill[LAST_REGISTER_FILE + 1];
  97
  98    const bool restrictedGPR16Range;
  99 };
 100
 101 void
 102 RegisterSet::reset(DataFile f, bool resetMax)
 103 {
 104    bits[f].fill(0);
 105    if (resetMax)
 106       fill[f] = -1;
 107 }
 108
 109 void
 110 RegisterSet::init(const Target *targ)
 111 {
 112    for (unsigned int rf = 0; rf <= FILE_ADDRESS; ++rf) {
 113       DataFile f = static_cast<DataFile>(rf);
 114       last[rf] = targ->getFileSize(f) - 1;
 115       unit[rf] = targ->getFileUnit(f);
 116       fill[rf] = -1;
 117       assert(last[rf] < MAX_REGISTER_FILE_SIZE);
 118       bits[rf].allocate(last[rf] + 1, true);
 119    }
 120 }
 121
 122 RegisterSet::RegisterSet(const Target *targ)
 123   : restrictedGPR16Range(targ->getChipset() < 0xc0)
 124 {
 125    init(targ);
 126    for (unsigned int i = 0; i <= LAST_REGISTER_FILE; ++i)
 127       reset(static_cast<DataFile>(i));
 128 }
 129
 130 void
 131 RegisterSet::periodicMask(DataFile f, uint32_t lock, uint32_t unlock)
 132 {
 133    bits[f].periodicMask32(lock, unlock);
 134 }
 135
 136 void
 137 RegisterSet::intersect(DataFile f, const RegisterSet *set)
 138 {
 139    bits[f] |= set->bits[f];
 140 }
 141
 142 void
 143 RegisterSet::print() const
 144 {
 145    INFO("GPR:");
 146    bits[FILE_GPR].print();
 147    INFO("\n");
 148 }
 149
 150 bool
 151 RegisterSet::assign(int32_t& reg, DataFile f, unsigned int size)
 152 {
 153    reg = bits[f].findFreeRange(size);
 154    if (reg < 0)
 155       return false;
 156    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
 157    return true;
 158 }
 159
 160 bool
 161 RegisterSet::isOccupied(DataFile f, int32_t reg, unsigned int size) const
 162 {
 163    return bits[f].testRange(reg, size);
 164 }
 165
 166 void
 167 RegisterSet::occupy(const Value *v)
 168 {
 169    occupy(v->reg.file, idToUnits(v), v->reg.size >> unit[v->reg.file]);
 170 }
 171
 172 void
 173 RegisterSet::occupyMask(DataFile f, int32_t reg, uint8_t mask)
 174 {
 175    bits[f].setMask(reg & ~31, static_cast<uint32_t>(mask) << (reg % 32));
 176 }
 177
 178 void
 179 RegisterSet::occupy(DataFile f, int32_t reg, unsigned int size)
 180 {
 181    bits[f].setRange(reg, size);
 182
 183    INFO_DBG(0, REG_ALLOC, "reg occupy: %u[%i] %u\n", f, reg, size);
 184
 185    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
 186 }
 187
 188 bool
 189 RegisterSet::testOccupy(const Value *v)
 190 {
 191    return testOccupy(v->reg.file,
 192                      idToUnits(v), v->reg.size >> unit[v->reg.file]);
 193 }
 194
 195 bool
 196 RegisterSet::testOccupy(DataFile f, int32_t reg, unsigned int size)
 197 {
 198    if (isOccupied(f, reg, size))
 199       return false;
 200    occupy(f, reg, size);
 201    return true;
 202 }
 203
 204 void
 205 RegisterSet::release(DataFile f, int32_t reg, unsigned int size)
 206 {
 207    bits[f].clrRange(reg, size);
 208
 209    INFO_DBG(0, REG_ALLOC, "reg release: %u[%i] %u\n", f, reg, size);
 210 }
 211
 212 class RegAlloc
 213 {
 214 public:
 215    RegAlloc(Program *program) : prog(program), sequence(0) { }
 216
 217    bool exec();
 218    bool execFunc();
 219
 220 private:
 221    class PhiMovesPass : public Pass {
 222    private:
 223       virtual bool visit(BasicBlock *);
 224       inline bool needNewElseBlock(BasicBlock *b, BasicBlock *p);
 225    };
 226
 227    class ArgumentMovesPass : public Pass {
 228    private:
 229       virtual bool visit(BasicBlock *);
 230    };
 231
 232    class BuildIntervalsPass : public Pass {
 233    private:
 234       virtual bool visit(BasicBlock *);
 235       void collectLiveValues(BasicBlock *);
 236       void addLiveRange(Value *, const BasicBlock *, int end);
 237    };
 238
 239    class InsertConstraintsPass : public Pass {
 240    public:
 241       bool exec(Function *func);
 242    private:
 243       virtual bool visit(BasicBlock *);
 244
 245       bool insertConstraintMoves();
 246
 247       void condenseDefs(Instruction *);
 248       void condenseSrcs(Instruction *, const int first, const int last);
 249
 250       void addHazard(Instruction *i, const ValueRef *src);
 251       void textureMask(TexInstruction *);
 252       void addConstraint(Instruction *, int s, int n);
 253       bool detectConflict(Instruction *, int s);
 254
 255       // target specific functions, TODO: put in subclass or Target
 256       void texConstraintNV50(TexInstruction *);
 257       void texConstraintNVC0(TexInstruction *);
 258       void texConstraintNVE0(TexInstruction *);
 259       void texConstraintGM107(TexInstruction *);
 260
 261       std::list<Instruction *> constrList;
 262
 263       const Target *targ;
 264    };
 265
 266    bool buildLiveSets(BasicBlock *);
 267
 268 private:
 269    Program *prog;
 270    Function *func;
 271
 272    // instructions in control flow / chronological order
 273    ArrayList insns;
 274
 275    int sequence; // for manual passes through CFG
 276 };
 277
 278 typedef std::pair<Value *, Value *> ValuePair;
 279
 280 class SpillCodeInserter
 281 {
 282 public:
 283    SpillCodeInserter(Function *fn) : func(fn), stackSize(0), stackBase(0) { }
 284
 285    bool run(const std::list<ValuePair>&);
 286
 287    Symbol *assignSlot(const Interval&, const unsigned int size);
 288    Value *offsetSlot(Value *, const LValue *);
 289    inline int32_t getStackSize() const { return stackSize; }
 290
 291 private:
 292    Function *func;
 293
 294    struct SpillSlot
 295    {
 296       Interval occup;
 297       std::list<Value *> residents; // needed to recalculate occup
 298       Symbol *sym;
 299       int32_t offset;
 300       inline uint8_t size() const { return sym->reg.size; }
 301    };
 302    std::list<SpillSlot> slots;
 303    int32_t stackSize;
 304    int32_t stackBase;
 305
 306    LValue *unspill(Instruction *usei, LValue *, Value *slot);
 307    void spill(Instruction *defi, Value *slot, LValue *);
 308 };
 309
 310 void
 311 RegAlloc::BuildIntervalsPass::addLiveRange(Value *val,
 312                                            const BasicBlock *bb,
 313                                            int end)
 314 {
 315    Instruction *insn = val->getUniqueInsn();
 316
 317    if (!insn)
 318       insn = bb->getFirst();
 319
 320    assert(bb->getFirst()->serial <= bb->getExit()->serial);
 321    assert(bb->getExit()->serial + 1 >= end);
 322
 323    int begin = insn->serial;
 324    if (begin < bb->getEntry()->serial || begin > bb->getExit()->serial)
 325       begin = bb->getEntry()->serial;
 326
 327    INFO_DBG(prog->dbgFlags, REG_ALLOC, "%%%i <- live range [%i(%i), %i)\n",
 328             val->id, begin, insn->serial, end);
 329
 330    if (begin != end) // empty ranges are only added as hazards for fixed regs
 331       val->livei.extend(begin, end);
 332 }
 333
 334 bool
 335 RegAlloc::PhiMovesPass::needNewElseBlock(BasicBlock *b, BasicBlock *p)
 336 {
 337    if (b->cfg.incidentCount() <= 1)
 338       return false;
 339
 340    int n = 0;
 341    for (Graph::EdgeIterator ei = p->cfg.outgoing(); !ei.end(); ei.next())
 342       if (ei.getType() == Graph::Edge::TREE ||
 343           ei.getType() == Graph::Edge::FORWARD)
 344          ++n;
 345    return (n == 2);
 346 }
 347
 348 // For each operand of each PHI in b, generate a new value by inserting a MOV
 349 // at the end of the block it is coming from and replace the operand with its
 350 // result. This eliminates liveness conflicts and enables us to let values be
 351 // copied to the right register if such a conflict exists nonetheless.
 352 //
 353 // These MOVs are also crucial in making sure the live intervals of phi srces
 354 // are extended until the end of the loop, since they are not included in the
 355 // live-in sets.
 356 bool
 357 RegAlloc::PhiMovesPass::visit(BasicBlock *bb)
 358 {
 359    Instruction *phi, *mov;
 360    BasicBlock *pb, *pn;
 361
 362    std::stack<BasicBlock *> stack;
 363
 364    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
 365       pb = BasicBlock::get(ei.getNode());
 366       assert(pb);
 367       if (needNewElseBlock(bb, pb))
 368          stack.push(pb);
 369    }
 370    while (!stack.empty()) {
 371       pb = stack.top();
 372       pn = new BasicBlock(func);
 373       stack.pop();
 374
 375       pb->cfg.detach(&bb->cfg);
 376       pb->cfg.attach(&pn->cfg, Graph::Edge::TREE);
 377       pn->cfg.attach(&bb->cfg, Graph::Edge::FORWARD);
 378
 379       assert(pb->getExit()->op != OP_CALL);
 380       if (pb->getExit()->asFlow()->target.bb == bb)
 381          pb->getExit()->asFlow()->target.bb = pn;
 382    }
 383
 384    // insert MOVs (phi->src(j) should stem from j-th in-BB)
 385    int j = 0;
 386    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
 387       pb = BasicBlock::get(ei.getNode());
 388       if (!pb->isTerminated())
 389          pb->insertTail(new_FlowInstruction(func, OP_BRA, bb));
 390
 391       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 392          mov = new_Instruction(func, OP_MOV, TYPE_U32);
 393
 394          mov->setSrc(0, phi->getSrc(j));
 395          mov->setDef(0, new_LValue(func, phi->getDef(0)->asLValue()));
 396          phi->setSrc(j, mov->getDef(0));
 397
 398          pb->insertBefore(pb->getExit(), mov);
 399       }
 400       ++j;
 401    }
 402
 403    return true;
 404 }
 405
 406 bool
 407 RegAlloc::ArgumentMovesPass::visit(BasicBlock *bb)
 408 {
 409    // Bind function call inputs/outputs to the same physical register
 410    // the callee uses, inserting moves as appropriate for the case a
 411    // conflict arises.
 412    for (Instruction *i = bb->getEntry(); i; i = i->next) {
 413       FlowInstruction *cal = i->asFlow();
 414       // TODO: Handle indirect calls.
 415       // Right now they should only be generated for builtins.
 416       if (!cal || cal->op != OP_CALL || cal->builtin || cal->indirect)
 417          continue;
 418       RegisterSet clobberSet(prog->getTarget());
 419
 420       // Bind input values.
 421       for (int s = cal->indirect ? 1 : 0; cal->srcExists(s); ++s) {
 422          const int t = cal->indirect ? (s - 1) : s;
 423          LValue *tmp = new_LValue(func, cal->getSrc(s)->asLValue());
 424          tmp->reg.data.id = cal->target.fn->ins[t].rep()->reg.data.id;
 425
 426          Instruction *mov =
 427             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 428          mov->setDef(0, tmp);
 429          mov->setSrc(0, cal->getSrc(s));
 430          cal->setSrc(s, tmp);
 431
 432          bb->insertBefore(cal, mov);
 433       }
 434
 435       // Bind output values.
 436       for (int d = 0; cal->defExists(d); ++d) {
 437          LValue *tmp = new_LValue(func, cal->getDef(d)->asLValue());
 438          tmp->reg.data.id = cal->target.fn->outs[d].rep()->reg.data.id;
 439
 440          Instruction *mov =
 441             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 442          mov->setSrc(0, tmp);
 443          mov->setDef(0, cal->getDef(d));
 444          cal->setDef(d, tmp);
 445
 446          bb->insertAfter(cal, mov);
 447          clobberSet.occupy(tmp);
 448       }
 449
 450       // Bind clobbered values.
 451       for (std::deque<Value *>::iterator it = cal->target.fn->clobbers.begin();
 452            it != cal->target.fn->clobbers.end();
 453            ++it) {
 454          if (clobberSet.testOccupy(*it)) {
 455             Value *tmp = new_LValue(func, (*it)->asLValue());
 456             tmp->reg.data.id = (*it)->reg.data.id;
 457             cal->setDef(cal->defCount(), tmp);
 458          }
 459       }
 460    }
 461
 462    // Update the clobber set of the function.
 463    if (BasicBlock::get(func->cfgExit) == bb) {
 464       func->buildDefSets();
 465       for (unsigned int i = 0; i < bb->defSet.getSize(); ++i)
 466          if (bb->defSet.test(i))
 467             func->clobbers.push_back(func->getLValue(i));
 468    }
 469
 470    return true;
 471 }
 472
 473 // Build the set of live-in variables of bb.
 474 bool
 475 RegAlloc::buildLiveSets(BasicBlock *bb)
 476 {
 477    Function *f = bb->getFunction();
 478    BasicBlock *bn;
 479    Instruction *i;
 480    unsigned int s, d;
 481
 482    INFO_DBG(prog->dbgFlags, REG_ALLOC, "buildLiveSets(BB:%i)\n", bb->getId());
 483
 484    bb->liveSet.allocate(func->allLValues.getSize(), false);
 485
 486    int n = 0;
 487    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 488       bn = BasicBlock::get(ei.getNode());
 489       if (bn == bb)
 490          continue;
 491       if (bn->cfg.visit(sequence))
 492          if (!buildLiveSets(bn))
 493             return false;
 494       if (n++ || bb->liveSet.marker)
 495          bb->liveSet |= bn->liveSet;
 496       else
 497          bb->liveSet = bn->liveSet;
 498    }
 499    if (!n && !bb->liveSet.marker)
 500       bb->liveSet.fill(0);
 501    bb->liveSet.marker = true;
 502
 503    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 504       INFO("BB:%i live set of out blocks:\n", bb->getId());
 505       bb->liveSet.print();
 506    }
 507
 508    // if (!bb->getEntry())
 509    //   return true;
 510
 511    if (bb == BasicBlock::get(f->cfgExit)) {
 512       for (std::deque<ValueRef>::iterator it = f->outs.begin();
 513            it != f->outs.end(); ++it) {
 514          assert(it->get()->asLValue());
 515          bb->liveSet.set(it->get()->id);
 516       }
 517    }
 518
 519    for (i = bb->getExit(); i && i != bb->getEntry()->prev; i = i->prev) {
 520       for (d = 0; i->defExists(d); ++d)
 521          bb->liveSet.clr(i->getDef(d)->id);
 522       for (s = 0; i->srcExists(s); ++s)
 523          if (i->getSrc(s)->asLValue())
 524             bb->liveSet.set(i->getSrc(s)->id);
 525    }
 526    for (i = bb->getPhi(); i && i->op == OP_PHI; i = i->next)
 527       bb->liveSet.clr(i->getDef(0)->id);
 528
 529    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 530       INFO("BB:%i live set after propagation:\n", bb->getId());
 531       bb->liveSet.print();
 532    }
 533
 534    return true;
 535 }
 536
 537 void
 538 RegAlloc::BuildIntervalsPass::collectLiveValues(BasicBlock *bb)
 539 {
 540    BasicBlock *bbA = NULL, *bbB = NULL;
 541
 542    if (bb->cfg.outgoingCount()) {
 543       // trickery to save a loop of OR'ing liveSets
 544       // aliasing works fine with BitSet::setOr
 545       for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 546          if (ei.getType() == Graph::Edge::DUMMY)
 547             continue;
 548          if (bbA) {
 549             bb->liveSet.setOr(&bbA->liveSet, &bbB->liveSet);
 550             bbA = bb;
 551          } else {
 552             bbA = bbB;
 553          }
 554          bbB = BasicBlock::get(ei.getNode());
 555       }
 556       bb->liveSet.setOr(&bbB->liveSet, bbA ? &bbA->liveSet : NULL);
 557    } else
 558    if (bb->cfg.incidentCount()) {
 559       bb->liveSet.fill(0);
 560    }
 561 }
 562
 563 bool
 564 RegAlloc::BuildIntervalsPass::visit(BasicBlock *bb)
 565 {
 566    collectLiveValues(bb);
 567
 568    INFO_DBG(prog->dbgFlags, REG_ALLOC, "BuildIntervals(BB:%i)\n", bb->getId());
 569
 570    // go through out blocks and delete phi sources that do not originate from
 571    // the current block from the live set
 572    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 573       BasicBlock *out = BasicBlock::get(ei.getNode());
 574
 575       for (Instruction *i = out->getPhi(); i && i->op == OP_PHI; i = i->next) {
 576          bb->liveSet.clr(i->getDef(0)->id);
 577
 578          for (int s = 0; i->srcExists(s); ++s) {
 579             assert(i->src(s).getInsn());
 580             if (i->getSrc(s)->getUniqueInsn()->bb == bb) // XXX: reachableBy ?
 581                bb->liveSet.set(i->getSrc(s)->id);
 582             else
 583                bb->liveSet.clr(i->getSrc(s)->id);
 584          }
 585       }
 586    }
 587
 588    // remaining live-outs are live until end
 589    if (bb->getExit()) {
 590       for (unsigned int j = 0; j < bb->liveSet.getSize(); ++j)
 591          if (bb->liveSet.test(j))
 592             addLiveRange(func->getLValue(j), bb, bb->getExit()->serial + 1);
 593    }
 594
 595    for (Instruction *i = bb->getExit(); i && i->op != OP_PHI; i = i->prev) {
 596       for (int d = 0; i->defExists(d); ++d) {
 597          bb->liveSet.clr(i->getDef(d)->id);
 598          if (i->getDef(d)->reg.data.id >= 0) // add hazard for fixed regs
 599             i->getDef(d)->livei.extend(i->serial, i->serial);
 600       }
 601
 602       for (int s = 0; i->srcExists(s); ++s) {
 603          if (!i->getSrc(s)->asLValue())
 604             continue;
 605          if (!bb->liveSet.test(i->getSrc(s)->id)) {
 606             bb->liveSet.set(i->getSrc(s)->id);
 607             addLiveRange(i->getSrc(s), bb, i->serial);
 608          }
 609       }
 610    }
 611
 612    if (bb == BasicBlock::get(func->cfg.getRoot())) {
 613       for (std::deque<ValueDef>::iterator it = func->ins.begin();
 614            it != func->ins.end(); ++it) {
 615          if (it->get()->reg.data.id >= 0) // add hazard for fixed regs
 616             it->get()->livei.extend(0, 1);
 617       }
 618    }
 619
 620    return true;
 621 }
 622
 623
 624 #define JOIN_MASK_PHI        (1 << 0)
 625 #define JOIN_MASK_UNION      (1 << 1)
 626 #define JOIN_MASK_MOV        (1 << 2)
 627 #define JOIN_MASK_TEX        (1 << 3)
 628
 629 class GCRA
 630 {
 631 public:
 632    GCRA(Function *, SpillCodeInserter&);
 633    ~GCRA();
 634
 635    bool allocateRegisters(ArrayList& insns);
 636
 637    void printNodeInfo() const;
 638
 639 private:
 640    class RIG_Node : public Graph::Node
 641    {
 642    public:
 643       RIG_Node();
 644
 645       void init(const RegisterSet&, LValue *);
 646
 647       void addInterference(RIG_Node *);
 648       void addRegPreference(RIG_Node *);
 649
 650       inline LValue *getValue() const
 651       {
 652          return reinterpret_cast<LValue *>(data);
 653       }
 654       inline void setValue(LValue *lval) { data = lval; }
 655
 656       inline uint8_t getCompMask() const
 657       {
 658          return ((1 << colors) - 1) << (reg & 7);
 659       }
 660
 661       static inline RIG_Node *get(const Graph::EdgeIterator& ei)
 662       {
 663          return static_cast<RIG_Node *>(ei.getNode());
 664       }
 665
 666    public:
 667       uint32_t degree;
 668       uint16_t degreeLimit; // if deg < degLimit, node is trivially colourable
 669       uint16_t colors;
 670
 671       DataFile f;
 672       int32_t reg;
 673
 674       float weight;
 675
 676       // list pointers for simplify() phase
 677       RIG_Node *next;
 678       RIG_Node *prev;
 679
 680       // union of the live intervals of all coalesced values (we want to retain
 681       //  the separate intervals for testing interference of compound values)
 682       Interval livei;
 683
 684       std::list<RIG_Node *> prefRegs;
 685    };
 686
 687 private:
 688    inline RIG_Node *getNode(const LValue *v) const { return &nodes[v->id]; }
 689
 690    void buildRIG(ArrayList&);
 691    bool coalesce(ArrayList&);
 692    bool doCoalesce(ArrayList&, unsigned int mask);
 693    void calculateSpillWeights();
 694    void simplify();
 695    bool selectRegisters();
 696    void cleanup(const bool success);
 697
 698    void simplifyEdge(RIG_Node *, RIG_Node *);
 699    void simplifyNode(RIG_Node *);
 700
 701    bool coalesceValues(Value *, Value *, bool force);
 702    void resolveSplitsAndMerges();
 703    void makeCompound(Instruction *, bool isSplit);
 704
 705    inline void checkInterference(const RIG_Node *, Graph::EdgeIterator&);
 706
 707    inline void insertOrderedTail(std::list<RIG_Node *>&, RIG_Node *);
 708    void checkList(std::list<RIG_Node *>&);
 709
 710 private:
 711    std::stack<uint32_t> stack;
 712
 713    // list headers for simplify() phase
 714    RIG_Node lo[2];
 715    RIG_Node hi;
 716
 717    Graph RIG;
 718    RIG_Node *nodes;
 719    unsigned int nodeCount;
 720
 721    Function *func;
 722    Program *prog;
 723
 724    static uint8_t relDegree[17][17];
 725
 726    RegisterSet regs;
 727
 728    // need to fixup register id for participants of OP_MERGE/SPLIT
 729    std::list<Instruction *> merges;
 730    std::list<Instruction *> splits;
 731
 732    SpillCodeInserter& spill;
 733    std::list<ValuePair> mustSpill;
 734 };
 735
 736 uint8_t GCRA::relDegree[17][17];
 737
 738 GCRA::RIG_Node::RIG_Node() : Node(NULL), next(this), prev(this)
 739 {
 740    colors = 0;
 741 }
 742
 743 void
 744 GCRA::printNodeInfo() const
 745 {
 746    for (unsigned int i = 0; i < nodeCount; ++i) {
 747       if (!nodes[i].colors)
 748          continue;
 749       INFO("RIG_Node[%%%i]($[%u]%i): %u colors, weight %f, deg %u/%u\n X",
 750            i,
 751            nodes[i].f,nodes[i].reg,nodes[i].colors,
 752            nodes[i].weight,
 753            nodes[i].degree, nodes[i].degreeLimit);
 754
 755       for (Graph::EdgeIterator ei = nodes[i].outgoing(); !ei.end(); ei.next())
 756          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
 757       for (Graph::EdgeIterator ei = nodes[i].incident(); !ei.end(); ei.next())
 758          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
 759       INFO("\n");
 760    }
 761 }
 762
 763 void
 764 GCRA::RIG_Node::init(const RegisterSet& regs, LValue *lval)
 765 {
 766    setValue(lval);
 767    if (lval->reg.data.id >= 0)
 768       lval->noSpill = lval->fixedReg = 1;
 769
 770    colors = regs.units(lval->reg.file, lval->reg.size);
 771    f = lval->reg.file;
 772    reg = -1;
 773    if (lval->reg.data.id >= 0)
 774       reg = regs.idToUnits(lval);
 775
 776    weight = std::numeric_limits<float>::infinity();
 777    degree = 0;
 778    degreeLimit = regs.getFileSize(f, lval->reg.size);
 779    degreeLimit -= relDegree[1][colors] - 1;
 780
 781    livei.insert(lval->livei);
 782 }
 783
 784 bool
 785 GCRA::coalesceValues(Value *dst, Value *src, bool force)
 786 {
 787    LValue *rep = dst->join->asLValue();
 788    LValue *val = src->join->asLValue();
 789
 790    if (!force && val->reg.data.id >= 0) {
 791       rep = src->join->asLValue();
 792       val = dst->join->asLValue();
 793    }
 794    RIG_Node *nRep = &nodes[rep->id];
 795    RIG_Node *nVal = &nodes[val->id];
 796
 797    if (src->reg.file != dst->reg.file) {
 798       if (!force)
 799          return false;
 800       WARN("forced coalescing of values in different files !\n");
 801    }
 802    if (!force && dst->reg.size != src->reg.size)
 803       return false;
 804
 805    if ((rep->reg.data.id >= 0) && (rep->reg.data.id != val->reg.data.id)) {
 806       if (force) {
 807          if (val->reg.data.id >= 0)
 808             WARN("forced coalescing of values in different fixed regs !\n");
 809       } else {
 810          if (val->reg.data.id >= 0)
 811             return false;
 812          // make sure that there is no overlap with the fixed register of rep
 813          for (ArrayList::Iterator it = func->allLValues.iterator();
 814               !it.end(); it.next()) {
 815             Value *reg = reinterpret_cast<Value *>(it.get())->asLValue();
 816             assert(reg);
 817             if (reg->interfers(rep) && reg->livei.overlaps(nVal->livei))
 818                return false;
 819          }
 820       }
 821    }
 822
 823    if (!force && nRep->livei.overlaps(nVal->livei))
 824       return false;
 825
 826    INFO_DBG(prog->dbgFlags, REG_ALLOC, "joining %%%i($%i) <- %%%i\n",
 827             rep->id, rep->reg.data.id, val->id);
 828
 829    // set join pointer of all values joined with val
 830    for (Value::DefIterator def = val->defs.begin(); def != val->defs.end();
 831         ++def)
 832       (*def)->get()->join = rep;
 833    assert(rep->join == rep && val->join == rep);
 834
 835    // add val's definitions to rep and extend the live interval of its RIG node
 836    rep->defs.insert(rep->defs.end(), val->defs.begin(), val->defs.end());
 837    nRep->livei.unify(nVal->livei);
 838    return true;
 839 }
 840
 841 bool
 842 GCRA::coalesce(ArrayList& insns)
 843 {
 844    bool ret = doCoalesce(insns, JOIN_MASK_PHI);
 845    if (!ret)
 846       return false;
 847    switch (func->getProgram()->getTarget()->getChipset() & ~0xf) {
 848    case 0x50:
 849    case 0x80:
 850    case 0x90:
 851    case 0xa0:
 852       ret = doCoalesce(insns, JOIN_MASK_UNION | JOIN_MASK_TEX);
 853       break;
 854    case 0xc0:
 855    case 0xd0:
 856    case 0xe0:
 857    case 0xf0:
 858    case 0x100:
 859    case 0x110:
 860       ret = doCoalesce(insns, JOIN_MASK_UNION);
 861       break;
 862    default:
 863       break;
 864    }
 865    if (!ret)
 866       return false;
 867    return doCoalesce(insns, JOIN_MASK_MOV);
 868 }
 869
 870 static inline uint8_t makeCompMask(int compSize, int base, int size)
 871 {
 872    uint8_t m = ((1 << size) - 1) << base;
 873
 874    switch (compSize) {
 875    case 1:
 876       return 0xff;
 877    case 2:
 878       m |= (m << 2);
 879       return (m << 4) | m;
 880    case 3:
 881    case 4:
 882       return (m << 4) | m;
 883    default:
 884       assert(compSize <= 8);
 885       return m;
 886    }
 887 }
 888
 889 // Used when coalescing moves. The non-compound value will become one, e.g.:
 890 // mov b32 $r0 $r2            / merge b64 $r0d { $r0 $r1 }
 891 // split b64 { $r0 $r1 } $r0d / mov b64 $r0d f64 $r2d
 892 static inline void copyCompound(Value *dst, Value *src)
 893 {
 894    LValue *ldst = dst->asLValue();
 895    LValue *lsrc = src->asLValue();
 896
 897    if (ldst->compound && !lsrc->compound) {
 898       LValue *swap = lsrc;
 899       lsrc = ldst;
 900       ldst = swap;
 901    }
 902
 903    ldst->compound = lsrc->compound;
 904    ldst->compMask = lsrc->compMask;
 905 }
 906
 907 void
 908 GCRA::makeCompound(Instruction *insn, bool split)
 909 {
 910    LValue *rep = (split ? insn->getSrc(0) : insn->getDef(0))->asLValue();
 911
 912    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 913       INFO("makeCompound(split = %i): ", split);
 914       insn->print();
 915    }
 916
 917    const unsigned int size = getNode(rep)->colors;
 918    unsigned int base = 0;
 919
 920    if (!rep->compound)
 921       rep->compMask = 0xff;
 922    rep->compound = 1;
 923
 924    for (int c = 0; split ? insn->defExists(c) : insn->srcExists(c); ++c) {
 925       LValue *val = (split ? insn->getDef(c) : insn->getSrc(c))->asLValue();
 926
 927       val->compound = 1;
 928       if (!val->compMask)
 929          val->compMask = 0xff;
 930       val->compMask &= makeCompMask(size, base, getNode(val)->colors);
 931       assert(val->compMask);
 932
 933       INFO_DBG(prog->dbgFlags, REG_ALLOC, "compound: %%%i:%02x <- %%%i:%02x\n",
 934            rep->id, rep->compMask, val->id, val->compMask);
 935
 936       base += getNode(val)->colors;
 937    }
 938    assert(base == size);
 939 }
 940
 941 bool
 942 GCRA::doCoalesce(ArrayList& insns, unsigned int mask)
 943 {
 944    int c, n;
 945
 946    for (n = 0; n < insns.getSize(); ++n) {
 947       Instruction *i;
 948       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(n));
 949
 950       switch (insn->op) {
 951       case OP_PHI:
 952          if (!(mask & JOIN_MASK_PHI))
 953             break;
 954          for (c = 0; insn->srcExists(c); ++c)
 955             if (!coalesceValues(insn->getDef(0), insn->getSrc(c), false)) {
 956                // this is bad
 957                ERROR("failed to coalesce phi operands\n");
 958                return false;
 959             }
 960          break;
 961       case OP_UNION:
 962       case OP_MERGE:
 963          if (!(mask & JOIN_MASK_UNION))
 964             break;
 965          for (c = 0; insn->srcExists(c); ++c)
 966             coalesceValues(insn->getDef(0), insn->getSrc(c), true);
 967          if (insn->op == OP_MERGE) {
 968             merges.push_back(insn);
 969             if (insn->srcExists(1))
 970                makeCompound(insn, false);
 971          }
 972          break;
 973       case OP_SPLIT:
 974          if (!(mask & JOIN_MASK_UNION))
 975             break;
 976          splits.push_back(insn);
 977          for (c = 0; insn->defExists(c); ++c)
 978             coalesceValues(insn->getSrc(0), insn->getDef(c), true);
 979          makeCompound(insn, true);
 980          break;
 981       case OP_MOV:
 982          if (!(mask & JOIN_MASK_MOV))
 983             break;
 984          i = NULL;
 985          if (!insn->getDef(0)->uses.empty())
 986             i = (*insn->getDef(0)->uses.begin())->getInsn();
 987          // if this is a contraint-move there will only be a single use
 988          if (i && i->op == OP_MERGE) // do we really still need this ?
 989             break;
 990          i = insn->getSrc(0)->getUniqueInsn();
 991          if (i && !i->constrainedDefs()) {
 992             if (coalesceValues(insn->getDef(0), insn->getSrc(0), false))
 993                copyCompound(insn->getSrc(0), insn->getDef(0));
 994          }
 995          break;
 996       case OP_TEX:
 997       case OP_TXB:
 998       case OP_TXL:
 999       case OP_TXF:
1000       case OP_TXQ:
1001       case OP_TXD:
1002       case OP_TXG:
1003       case OP_TXLQ:
1004       case OP_TEXCSAA:
1005       case OP_TEXPREP:
1006          if (!(mask & JOIN_MASK_TEX))
1007             break;
1008          for (c = 0; insn->srcExists(c) && c != insn->predSrc; ++c)
1009             coalesceValues(insn->getDef(c), insn->getSrc(c), true);
1010          break;
1011       default:
1012          break;
1013       }
1014    }
1015    return true;
1016 }
1017
1018 void
1019 GCRA::RIG_Node::addInterference(RIG_Node *node)
1020 {
1021    this->degree += relDegree[node->colors][colors];
1022    node->degree += relDegree[colors][node->colors];
1023
1024    this->attach(node, Graph::Edge::CROSS);
1025 }
1026
1027 void
1028 GCRA::RIG_Node::addRegPreference(RIG_Node *node)
1029 {
1030    prefRegs.push_back(node);
1031 }
1032
1033 GCRA::GCRA(Function *fn, SpillCodeInserter& spill) :
1034    func(fn),
1035    regs(fn->getProgram()->getTarget()),
1036    spill(spill)
1037 {
1038    prog = func->getProgram();
1039
1040    // initialize relative degrees array - i takes away from j
1041    for (int i = 1; i <= 16; ++i)
1042       for (int j = 1; j <= 16; ++j)
1043          relDegree[i][j] = j * ((i + j - 1) / j);
1044 }
1045
1046 GCRA::~GCRA()
1047 {
1048    if (nodes)
1049       delete[] nodes;
1050 }
1051
1052 void
1053 GCRA::checkList(std::list<RIG_Node *>& lst)
1054 {
1055    GCRA::RIG_Node *prev = NULL;
1056
1057    for (std::list<RIG_Node *>::iterator it = lst.begin();
1058         it != lst.end();
1059         ++it) {
1060       assert((*it)->getValue()->join == (*it)->getValue());
1061       if (prev)
1062          assert(prev->livei.begin() <= (*it)->livei.begin());
1063       prev = *it;
1064    }
1065 }
1066
1067 void
1068 GCRA::insertOrderedTail(std::list<RIG_Node *>& list, RIG_Node *node)
1069 {
1070    if (node->livei.isEmpty())
1071       return;
1072    // only the intervals of joined values don't necessarily arrive in order
1073    std::list<RIG_Node *>::iterator prev, it;
1074    for (it = list.end(); it != list.begin(); it = prev) {
1075       prev = it;
1076       --prev;
1077       if ((*prev)->livei.begin() <= node->livei.begin())
1078          break;
1079    }
1080    list.insert(it, node);
1081 }
1082
1083 void
1084 GCRA::buildRIG(ArrayList& insns)
1085 {
1086    std::list<RIG_Node *> values, active;
1087
1088    for (std::deque<ValueDef>::iterator it = func->ins.begin();
1089         it != func->ins.end(); ++it)
1090       insertOrderedTail(values, getNode(it->get()->asLValue()));
1091
1092    for (int i = 0; i < insns.getSize(); ++i) {
1093       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(i));
1094       for (int d = 0; insn->defExists(d); ++d)
1095          if (insn->getDef(d)->rep() == insn->getDef(d))
1096             insertOrderedTail(values, getNode(insn->getDef(d)->asLValue()));
1097    }
1098    checkList(values);
1099
1100    while (!values.empty()) {
1101       RIG_Node *cur = values.front();
1102
1103       for (std::list<RIG_Node *>::iterator it = active.begin();
1104            it != active.end();) {
1105          RIG_Node *node = *it;
1106
1107          if (node->livei.end() <= cur->livei.begin()) {
1108             it = active.erase(it);
1109          } else {
1110             if (node->f == cur->f && node->livei.overlaps(cur->livei))
1111                cur->addInterference(node);
1112             ++it;
1113          }
1114       }
1115       values.pop_front();
1116       active.push_back(cur);
1117    }
1118 }
1119
1120 void
1121 GCRA::calculateSpillWeights()
1122 {
1123    for (unsigned int i = 0; i < nodeCount; ++i) {
1124       RIG_Node *const n = &nodes[i];
1125       if (!nodes[i].colors || nodes[i].livei.isEmpty())
1126          continue;
1127       if (nodes[i].reg >= 0) {
1128          // update max reg
1129          regs.occupy(n->f, n->reg, n->colors);
1130          continue;
1131       }
1132       LValue *val = nodes[i].getValue();
1133
1134       if (!val->noSpill) {
1135          int rc = 0;
1136          for (Value::DefIterator it = val->defs.begin();
1137               it != val->defs.end();
1138               ++it)
1139             rc += (*it)->get()->refCount();
1140
1141          nodes[i].weight =
1142             (float)rc * (float)rc / (float)nodes[i].livei.extent();
1143       }
1144
1145       if (nodes[i].degree < nodes[i].degreeLimit) {
1146          int l = 0;
1147          if (val->reg.size > 4)
1148             l = 1;
1149          DLLIST_ADDHEAD(&lo[l], &nodes[i]);
1150       } else {
1151          DLLIST_ADDHEAD(&hi, &nodes[i]);
1152       }
1153    }
1154    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1155       printNodeInfo();
1156 }
1157
1158 void
1159 GCRA::simplifyEdge(RIG_Node *a, RIG_Node *b)
1160 {
1161    bool move = b->degree >= b->degreeLimit;
1162
1163    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1164             "edge: (%%%i, deg %u/%u) >-< (%%%i, deg %u/%u)\n",
1165             a->getValue()->id, a->degree, a->degreeLimit,
1166             b->getValue()->id, b->degree, b->degreeLimit);
1167
1168    b->degree -= relDegree[a->colors][b->colors];
1169
1170    move = move && b->degree < b->degreeLimit;
1171    if (move && !DLLIST_EMPTY(b)) {
1172       int l = (b->getValue()->reg.size > 4) ? 1 : 0;
1173       DLLIST_DEL(b);
1174       DLLIST_ADDTAIL(&lo[l], b);
1175    }
1176 }
1177
1178 void
1179 GCRA::simplifyNode(RIG_Node *node)
1180 {
1181    for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1182       simplifyEdge(node, RIG_Node::get(ei));
1183
1184    for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1185       simplifyEdge(node, RIG_Node::get(ei));
1186
1187    DLLIST_DEL(node);
1188    stack.push(node->getValue()->id);
1189
1190    INFO_DBG(prog->dbgFlags, REG_ALLOC, "SIMPLIFY: pushed %%%i%s\n",
1191             node->getValue()->id,
1192             (node->degree < node->degreeLimit) ? "" : "(spill)");
1193 }
1194
1195 void
1196 GCRA::simplify()
1197 {
1198    for (;;) {
1199       if (!DLLIST_EMPTY(&lo[0])) {
1200          do {
1201             simplifyNode(lo[0].next);
1202          } while (!DLLIST_EMPTY(&lo[0]));
1203       } else
1204       if (!DLLIST_EMPTY(&lo[1])) {
1205          simplifyNode(lo[1].next);
1206       } else
1207       if (!DLLIST_EMPTY(&hi)) {
1208          RIG_Node *best = hi.next;
1209          float bestScore = best->weight / (float)best->degree;
1210          // spill candidate
1211          for (RIG_Node *it = best->next; it != &hi; it = it->next) {
1212             float score = it->weight / (float)it->degree;
1213             if (score < bestScore) {
1214                best = it;
1215                bestScore = score;
1216             }
1217          }
1218          if (isinf(bestScore)) {
1219             ERROR("no viable spill candidates left\n");
1220             break;
1221          }
1222          simplifyNode(best);
1223       } else {
1224          break;
1225       }
1226    }
1227 }
1228
1229 void
1230 GCRA::checkInterference(const RIG_Node *node, Graph::EdgeIterator& ei)
1231 {
1232    const RIG_Node *intf = RIG_Node::get(ei);
1233
1234    if (intf->reg < 0)
1235       return;
1236    const LValue *vA = node->getValue();
1237    const LValue *vB = intf->getValue();
1238
1239    const uint8_t intfMask = ((1 << intf->colors) - 1) << (intf->reg & 7);
1240
1241    if (vA->compound | vB->compound) {
1242       // NOTE: this only works for >aligned< register tuples !
1243       for (Value::DefCIterator D = vA->defs.begin(); D != vA->defs.end(); ++D) {
1244       for (Value::DefCIterator d = vB->defs.begin(); d != vB->defs.end(); ++d) {
1245          const LValue *vD = (*D)->get()->asLValue();
1246          const LValue *vd = (*d)->get()->asLValue();
1247
1248          if (!vD->livei.overlaps(vd->livei)) {
1249             INFO_DBG(prog->dbgFlags, REG_ALLOC, "(%%%i) X (%%%i): no overlap\n",
1250                      vD->id, vd->id);
1251             continue;
1252          }
1253
1254          uint8_t mask = vD->compound ? vD->compMask : ~0;
1255          if (vd->compound) {
1256             assert(vB->compound);
1257             mask &= vd->compMask & vB->compMask;
1258          } else {
1259             mask &= intfMask;
1260          }
1261
1262          INFO_DBG(prog->dbgFlags, REG_ALLOC,
1263                   "(%%%i)%02x X (%%%i)%02x & %02x: $r%i.%02x\n",
1264                   vD->id,
1265                   vD->compound ? vD->compMask : 0xff,
1266                   vd->id,
1267                   vd->compound ? vd->compMask : intfMask,
1268                   vB->compMask, intf->reg & ~7, mask);
1269          if (mask)
1270             regs.occupyMask(node->f, intf->reg & ~7, mask);
1271       }
1272       }
1273    } else {
1274       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1275                "(%%%i) X (%%%i): $r%i + %u\n",
1276                vA->id, vB->id, intf->reg, intf->colors);
1277       regs.occupy(node->f, intf->reg, intf->colors);
1278    }
1279 }
1280
1281 bool
1282 GCRA::selectRegisters()
1283 {
1284    INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nSELECT phase\n");
1285
1286    while (!stack.empty()) {
1287       RIG_Node *node = &nodes[stack.top()];
1288       stack.pop();
1289
1290       regs.reset(node->f);
1291
1292       INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nNODE[%%%i, %u colors]\n",
1293                node->getValue()->id, node->colors);
1294
1295       for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1296          checkInterference(node, ei);
1297       for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1298          checkInterference(node, ei);
1299
1300       if (!node->prefRegs.empty()) {
1301          for (std::list<RIG_Node *>::const_iterator it = node->prefRegs.begin();
1302               it != node->prefRegs.end();
1303               ++it) {
1304             if ((*it)->reg >= 0 &&
1305                 regs.testOccupy(node->f, (*it)->reg, node->colors)) {
1306                node->reg = (*it)->reg;
1307                break;
1308             }
1309          }
1310       }
1311       if (node->reg >= 0)
1312          continue;
1313       LValue *lval = node->getValue();
1314       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1315          regs.print();
1316       bool ret = regs.assign(node->reg, node->f, node->colors);
1317       if (ret) {
1318          INFO_DBG(prog->dbgFlags, REG_ALLOC, "assigned reg %i\n", node->reg);
1319          lval->compMask = node->getCompMask();
1320       } else {
1321          INFO_DBG(prog->dbgFlags, REG_ALLOC, "must spill: %%%i (size %u)\n",
1322                   lval->id, lval->reg.size);
1323          Symbol *slot = NULL;
1324          if (lval->reg.file == FILE_GPR)
1325             slot = spill.assignSlot(node->livei, lval->reg.size);
1326          mustSpill.push_back(ValuePair(lval, slot));
1327       }
1328    }
1329    if (!mustSpill.empty())
1330       return false;
1331    for (unsigned int i = 0; i < nodeCount; ++i) {
1332       LValue *lval = nodes[i].getValue();
1333       if (nodes[i].reg >= 0 && nodes[i].colors > 0)
1334          lval->reg.data.id =
1335             regs.unitsToId(nodes[i].f, nodes[i].reg, lval->reg.size);
1336    }
1337    return true;
1338 }
1339
1340 bool
1341 GCRA::allocateRegisters(ArrayList& insns)
1342 {
1343    bool ret;
1344
1345    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1346             "allocateRegisters to %u instructions\n", insns.getSize());
1347
1348    nodeCount = func->allLValues.getSize();
1349    nodes = new RIG_Node[nodeCount];
1350    if (!nodes)
1351       return false;
1352    for (unsigned int i = 0; i < nodeCount; ++i) {
1353       LValue *lval = reinterpret_cast<LValue *>(func->allLValues.get(i));
1354       if (lval) {
1355          nodes[i].init(regs, lval);
1356          RIG.insert(&nodes[i]);
1357       }
1358    }
1359
1360    // coalesce first, we use only 1 RIG node for a group of joined values
1361    ret = coalesce(insns);
1362    if (!ret)
1363       goto out;
1364
1365    if (func->getProgram()->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1366       func->printLiveIntervals();
1367
1368    buildRIG(insns);
1369    calculateSpillWeights();
1370    simplify();
1371
1372    ret = selectRegisters();
1373    if (!ret) {
1374       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1375                "selectRegisters failed, inserting spill code ...\n");
1376       regs.reset(FILE_GPR, true);
1377       spill.run(mustSpill);
1378       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1379          func->print();
1380    } else {
1381       prog->maxGPR = std::max(prog->maxGPR, regs.getMaxAssigned(FILE_GPR));
1382    }
1383
1384 out:
1385    cleanup(ret);
1386    return ret;
1387 }
1388
1389 void
1390 GCRA::cleanup(const bool success)
1391 {
1392    mustSpill.clear();
1393
1394    for (ArrayList::Iterator it = func->allLValues.iterator();
1395         !it.end(); it.next()) {
1396       LValue *lval =  reinterpret_cast<LValue *>(it.get());
1397
1398       lval->livei.clear();
1399
1400       lval->compound = 0;
1401       lval->compMask = 0;
1402
1403       if (lval->join == lval)
1404          continue;
1405
1406       if (success) {
1407          lval->reg.data.id = lval->join->reg.data.id;
1408       } else {
1409          for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1410               ++d)
1411             lval->join->defs.remove(*d);
1412          lval->join = lval;
1413       }
1414    }
1415
1416    if (success)
1417       resolveSplitsAndMerges();
1418    splits.clear(); // avoid duplicate entries on next coalesce pass
1419    merges.clear();
1420
1421    delete[] nodes;
1422    nodes = NULL;
1423 }
1424
1425 Symbol *
1426 SpillCodeInserter::assignSlot(const Interval &livei, const unsigned int size)
1427 {
1428    SpillSlot slot;
1429    int32_t offsetBase = stackSize;
1430    int32_t offset;
1431    std::list<SpillSlot>::iterator pos = slots.end(), it = slots.begin();
1432
1433    if (offsetBase % size)
1434       offsetBase += size - (offsetBase % size);
1435
1436    slot.sym = NULL;
1437
1438    for (offset = offsetBase; offset < stackSize; offset += size) {
1439       const int32_t entryEnd = offset + size;
1440       while (it != slots.end() && it->offset < offset)
1441          ++it;
1442       if (it == slots.end()) // no slots left
1443          break;
1444       std::list<SpillSlot>::iterator bgn = it;
1445
1446       while (it != slots.end() && it->offset < entryEnd) {
1447          it->occup.print();
1448          if (it->occup.overlaps(livei))
1449             break;
1450          ++it;
1451       }
1452       if (it == slots.end() || it->offset >= entryEnd) {
1453          // fits
1454          for (; bgn != slots.end() && bgn->offset < entryEnd; ++bgn) {
1455             bgn->occup.insert(livei);
1456             if (bgn->size() == size)
1457                slot.sym = bgn->sym;
1458          }
1459          break;
1460       }
1461    }
1462    if (!slot.sym) {
1463       stackSize = offset + size;
1464       slot.offset = offset;
1465       slot.sym = new_Symbol(func->getProgram(), FILE_MEMORY_LOCAL);
1466       if (!func->stackPtr)
1467          offset += func->tlsBase;
1468       slot.sym->setAddress(NULL, offset);
1469       slot.sym->reg.size = size;
1470       slots.insert(pos, slot)->occup.insert(livei);
1471    }
1472    return slot.sym;
1473 }
1474
1475 Value *
1476 SpillCodeInserter::offsetSlot(Value *base, const LValue *lval)
1477 {
1478    if (!lval->compound || (lval->compMask & 0x1))
1479       return base;
1480    Value *slot = cloneShallow(func, base);
1481
1482    slot->reg.data.offset += (ffs(lval->compMask) - 1) * lval->reg.size;
1483    slot->reg.size = lval->reg.size;
1484
1485    return slot;
1486 }
1487
1488 void
1489 SpillCodeInserter::spill(Instruction *defi, Value *slot, LValue *lval)
1490 {
1491    const DataType ty = typeOfSize(lval->reg.size);
1492
1493    slot = offsetSlot(slot, lval);
1494
1495    Instruction *st;
1496    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1497       st = new_Instruction(func, OP_STORE, ty);
1498       st->setSrc(0, slot);
1499       st->setSrc(1, lval);
1500       lval->noSpill = 1;
1501    } else {
1502       st = new_Instruction(func, OP_CVT, ty);
1503       st->setDef(0, slot);
1504       st->setSrc(0, lval);
1505    }
1506    defi->bb->insertAfter(defi, st);
1507 }
1508
1509 LValue *
1510 SpillCodeInserter::unspill(Instruction *usei, LValue *lval, Value *slot)
1511 {
1512    const DataType ty = typeOfSize(lval->reg.size);
1513
1514    slot = offsetSlot(slot, lval);
1515    lval = cloneShallow(func, lval);
1516
1517    Instruction *ld;
1518    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1519       lval->noSpill = 1;
1520       ld = new_Instruction(func, OP_LOAD, ty);
1521    } else {
1522       ld = new_Instruction(func, OP_CVT, ty);
1523    }
1524    ld->setDef(0, lval);
1525    ld->setSrc(0, slot);
1526
1527    usei->bb->insertBefore(usei, ld);
1528    return lval;
1529 }
1530
1531
1532 // For each value that is to be spilled, go through all its definitions.
1533 // A value can have multiple definitions if it has been coalesced before.
1534 // For each definition, first go through all its uses and insert an unspill
1535 // instruction before it, then replace the use with the temporary register.
1536 // Unspill can be either a load from memory or simply a move to another
1537 // register file.
1538 // For "Pseudo" instructions (like PHI, SPLIT, MERGE) we can erase the use
1539 // if we have spilled to a memory location, or simply with the new register.
1540 // No load or conversion instruction should be needed.
1541 bool
1542 SpillCodeInserter::run(const std::list<ValuePair>& lst)
1543 {
1544    for (std::list<ValuePair>::const_iterator it = lst.begin(); it != lst.end();
1545         ++it) {
1546       LValue *lval = it->first->asLValue();
1547       Symbol *mem = it->second ? it->second->asSym() : NULL;
1548
1549       for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1550            ++d) {
1551          Value *slot = mem ?
1552             static_cast<Value *>(mem) : new_LValue(func, FILE_GPR);
1553          Value *tmp = NULL;
1554          Instruction *last = NULL;
1555
1556          LValue *dval = (*d)->get()->asLValue();
1557          Instruction *defi = (*d)->getInsn();
1558
1559          // Unspill at each use *before* inserting spill instructions,
1560          // we don't want to have the spill instructions in the use list here.
1561          while (!dval->uses.empty()) {
1562             ValueRef *u = *dval->uses.begin();
1563             Instruction *usei = u->getInsn();
1564             assert(usei);
1565             if (usei->isPseudo()) {
1566                tmp = (slot->reg.file == FILE_MEMORY_LOCAL) ? NULL : slot;
1567                last = NULL;
1568             } else
1569             if (!last || usei != last->next) { // TODO: sort uses
1570                tmp = unspill(usei, dval, slot);
1571                last = usei;
1572             }
1573             u->set(tmp);
1574          }
1575
1576          assert(defi);
1577          if (defi->isPseudo()) {
1578             d = lval->defs.erase(d);
1579             --d;
1580             if (slot->reg.file == FILE_MEMORY_LOCAL)
1581                delete_Instruction(func->getProgram(), defi);
1582             else
1583                defi->setDef(0, slot);
1584          } else {
1585             spill(defi, slot, dval);
1586          }
1587       }
1588
1589    }
1590
1591    // TODO: We're not trying to reuse old slots in a potential next iteration.
1592    //  We have to update the slots' livei intervals to be able to do that.
1593    stackBase = stackSize;
1594    slots.clear();
1595    return true;
1596 }
1597
1598 bool
1599 RegAlloc::exec()
1600 {
1601    for (IteratorRef it = prog->calls.iteratorDFS(false);
1602         !it->end(); it->next()) {
1603       func = Function::get(reinterpret_cast<Graph::Node *>(it->get()));
1604
1605       func->tlsBase = prog->tlsSize;
1606       if (!execFunc())
1607          return false;
1608       prog->tlsSize += func->tlsSize;
1609    }
1610    return true;
1611 }
1612
1613 bool
1614 RegAlloc::execFunc()
1615 {
1616    InsertConstraintsPass insertConstr;
1617    PhiMovesPass insertPhiMoves;
1618    ArgumentMovesPass insertArgMoves;
1619    BuildIntervalsPass buildIntervals;
1620    SpillCodeInserter insertSpills(func);
1621
1622    GCRA gcra(func, insertSpills);
1623
1624    unsigned int i, retries;
1625    bool ret;
1626
1627    if (!func->ins.empty()) {
1628       // Insert a nop at the entry so inputs only used by the first instruction
1629       // don't count as having an empty live range.
1630       Instruction *nop = new_Instruction(func, OP_NOP, TYPE_NONE);
1631       BasicBlock::get(func->cfg.getRoot())->insertHead(nop);
1632    }
1633
1634    ret = insertConstr.exec(func);
1635    if (!ret)
1636       goto out;
1637
1638    ret = insertPhiMoves.run(func);
1639    if (!ret)
1640       goto out;
1641
1642    ret = insertArgMoves.run(func);
1643    if (!ret)
1644       goto out;
1645
1646    // TODO: need to fix up spill slot usage ranges to support > 1 retry
1647    for (retries = 0; retries < 3; ++retries) {
1648       if (retries && (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC))
1649          INFO("Retry: %i\n", retries);
1650       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1651          func->print();
1652
1653       // spilling to registers may add live ranges, need to rebuild everything
1654       ret = true;
1655       for (sequence = func->cfg.nextSequence(), i = 0;
1656            ret && i <= func->loopNestingBound;
1657            sequence = func->cfg.nextSequence(), ++i)
1658          ret = buildLiveSets(BasicBlock::get(func->cfg.getRoot()));
1659       if (!ret)
1660          break;
1661       func->orderInstructions(this->insns);
1662
1663       ret = buildIntervals.run(func);
1664       if (!ret)
1665          break;
1666       ret = gcra.allocateRegisters(insns);
1667       if (ret)
1668          break; // success
1669    }
1670    INFO_DBG(prog->dbgFlags, REG_ALLOC, "RegAlloc done: %i\n", ret);
1671
1672    func->tlsSize = insertSpills.getStackSize();
1673 out:
1674    return ret;
1675 }
1676
1677 // TODO: check if modifying Instruction::join here breaks anything
1678 void
1679 GCRA::resolveSplitsAndMerges()
1680 {
1681    for (std::list<Instruction *>::iterator it = splits.begin();
1682         it != splits.end();
1683         ++it) {
1684       Instruction *split = *it;
1685       unsigned int reg = regs.idToBytes(split->getSrc(0));
1686       for (int d = 0; split->defExists(d); ++d) {
1687          Value *v = split->getDef(d);
1688          v->reg.data.id = regs.bytesToId(v, reg);
1689          v->join = v;
1690          reg += v->reg.size;
1691       }
1692    }
1693    splits.clear();
1694
1695    for (std::list<Instruction *>::iterator it = merges.begin();
1696         it != merges.end();
1697         ++it) {
1698       Instruction *merge = *it;
1699       unsigned int reg = regs.idToBytes(merge->getDef(0));
1700       for (int s = 0; merge->srcExists(s); ++s) {
1701          Value *v = merge->getSrc(s);
1702          v->reg.data.id = regs.bytesToId(v, reg);
1703          v->join = v;
1704          reg += v->reg.size;
1705       }
1706    }
1707    merges.clear();
1708 }
1709
1710 bool Program::registerAllocation()
1711 {
1712    RegAlloc ra(this);
1713    return ra.exec();
1714 }
1715
1716 bool
1717 RegAlloc::InsertConstraintsPass::exec(Function *ir)
1718 {
1719    constrList.clear();
1720
1721    bool ret = run(ir, true, true);
1722    if (ret)
1723       ret = insertConstraintMoves();
1724    return ret;
1725 }
1726
1727 // TODO: make part of texture insn
1728 void
1729 RegAlloc::InsertConstraintsPass::textureMask(TexInstruction *tex)
1730 {
1731    Value *def[4];
1732    int c, k, d;
1733    uint8_t mask = 0;
1734
1735    for (d = 0, k = 0, c = 0; c < 4; ++c) {
1736       if (!(tex->tex.mask & (1 << c)))
1737          continue;
1738       if (tex->getDef(k)->refCount()) {
1739          mask |= 1 << c;
1740          def[d++] = tex->getDef(k);
1741       }
1742       ++k;
1743    }
1744    tex->tex.mask = mask;
1745
1746    for (c = 0; c < d; ++c)
1747       tex->setDef(c, def[c]);
1748    for (; c < 4; ++c)
1749       tex->setDef(c, NULL);
1750 }
1751
1752 bool
1753 RegAlloc::InsertConstraintsPass::detectConflict(Instruction *cst, int s)
1754 {
1755    Value *v = cst->getSrc(s);
1756
1757    // current register allocation can't handle it if a value participates in
1758    // multiple constraints
1759    for (Value::UseIterator it = v->uses.begin(); it != v->uses.end(); ++it) {
1760       if (cst != (*it)->getInsn())
1761          return true;
1762    }
1763
1764    // can start at s + 1 because detectConflict is called on all sources
1765    for (int c = s + 1; cst->srcExists(c); ++c)
1766       if (v == cst->getSrc(c))
1767          return true;
1768
1769    Instruction *defi = v->getInsn();
1770
1771    return (!defi || defi->constrainedDefs());
1772 }
1773
1774 void
1775 RegAlloc::InsertConstraintsPass::addConstraint(Instruction *i, int s, int n)
1776 {
1777    Instruction *cst;
1778    int d;
1779
1780    // first, look for an existing identical constraint op
1781    for (std::list<Instruction *>::iterator it = constrList.begin();
1782         it != constrList.end();
1783         ++it) {
1784       cst = (*it);
1785       if (!i->bb->dominatedBy(cst->bb))
1786          break;
1787       for (d = 0; d < n; ++d)
1788          if (cst->getSrc(d) != i->getSrc(d + s))
1789             break;
1790       if (d >= n) {
1791          for (d = 0; d < n; ++d, ++s)
1792             i->setSrc(s, cst->getDef(d));
1793          return;
1794       }
1795    }
1796    cst = new_Instruction(func, OP_CONSTRAINT, i->dType);
1797
1798    for (d = 0; d < n; ++s, ++d) {
1799       cst->setDef(d, new_LValue(func, FILE_GPR));
1800       cst->setSrc(d, i->getSrc(s));
1801       i->setSrc(s, cst->getDef(d));
1802    }
1803    i->bb->insertBefore(i, cst);
1804
1805    constrList.push_back(cst);
1806 }
1807
1808 // Add a dummy use of the pointer source of >= 8 byte loads after the load
1809 // to prevent it from being assigned a register which overlapping the load's
1810 // destination, which would produce random corruptions.
1811 void
1812 RegAlloc::InsertConstraintsPass::addHazard(Instruction *i, const ValueRef *src)
1813 {
1814    Instruction *hzd = new_Instruction(func, OP_NOP, TYPE_NONE);
1815    hzd->setSrc(0, src->get());
1816    i->bb->insertAfter(i, hzd);
1817
1818 }
1819
1820 // b32 { %r0 %r1 %r2 %r3 } -> b128 %r0q
1821 void
1822 RegAlloc::InsertConstraintsPass::condenseDefs(Instruction *insn)
1823 {
1824    uint8_t size = 0;
1825    int n;
1826    for (n = 0; insn->defExists(n) && insn->def(n).getFile() == FILE_GPR; ++n)
1827       size += insn->getDef(n)->reg.size;
1828    if (n < 2)
1829       return;
1830    LValue *lval = new_LValue(func, FILE_GPR);
1831    lval->reg.size = size;
1832
1833    Instruction *split = new_Instruction(func, OP_SPLIT, typeOfSize(size));
1834    split->setSrc(0, lval);
1835    for (int d = 0; d < n; ++d) {
1836       split->setDef(d, insn->getDef(d));
1837       insn->setDef(d, NULL);
1838    }
1839    insn->setDef(0, lval);
1840
1841    for (int k = 1, d = n; insn->defExists(d); ++d, ++k) {
1842       insn->setDef(k, insn->getDef(d));
1843       insn->setDef(d, NULL);
1844    }
1845    // carry over predicate if any (mainly for OP_UNION uses)
1846    split->setPredicate(insn->cc, insn->getPredicate());
1847
1848    insn->bb->insertAfter(insn, split);
1849    constrList.push_back(split);
1850 }
1851 void
1852 RegAlloc::InsertConstraintsPass::condenseSrcs(Instruction *insn,
1853                                               const int a, const int b)
1854 {
1855    uint8_t size = 0;
1856    if (a >= b)
1857       return;
1858    for (int s = a; s <= b; ++s)
1859       size += insn->getSrc(s)->reg.size;
1860    if (!size)
1861       return;
1862    LValue *lval = new_LValue(func, FILE_GPR);
1863    lval->reg.size = size;
1864
1865    Value *save[3];
1866    insn->takeExtraSources(0, save);
1867
1868    Instruction *merge = new_Instruction(func, OP_MERGE, typeOfSize(size));
1869    merge->setDef(0, lval);
1870    for (int s = a, i = 0; s <= b; ++s, ++i) {
1871       merge->setSrc(i, insn->getSrc(s));
1872       insn->setSrc(s, NULL);
1873    }
1874    insn->setSrc(a, lval);
1875
1876    for (int k = a + 1, s = b + 1; insn->srcExists(s); ++s, ++k) {
1877       insn->setSrc(k, insn->getSrc(s));
1878       insn->setSrc(s, NULL);
1879    }
1880    insn->bb->insertBefore(insn, merge);
1881
1882    insn->putExtraSources(0, save);
1883
1884    constrList.push_back(merge);
1885 }
1886
1887 void
1888 RegAlloc::InsertConstraintsPass::texConstraintGM107(TexInstruction *tex)
1889 {
1890    int n, s;
1891
1892    if (isTextureOp(tex->op))
1893       textureMask(tex);
1894    condenseDefs(tex);
1895
1896    if (tex->op == OP_SUSTB || tex->op == OP_SUSTP) {
1897       condenseSrcs(tex, 3, (3 + typeSizeof(tex->dType) / 4) - 1);
1898    } else
1899    if (isTextureOp(tex->op)) {
1900       if (tex->op != OP_TXQ) {
1901          s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
1902          n = tex->srcCount(0xff) - s;
1903       } else {
1904          s = tex->srcCount(0xff);
1905          n = 0;
1906       }
1907
1908       if (s > 1)
1909          condenseSrcs(tex, 0, s - 1);
1910       if (n > 1) // NOTE: first call modified positions already
1911          condenseSrcs(tex, 1, n);
1912    }
1913 }
1914
1915 void
1916 RegAlloc::InsertConstraintsPass::texConstraintNVE0(TexInstruction *tex)
1917 {
1918    if (isTextureOp(tex->op))
1919       textureMask(tex);
1920    condenseDefs(tex);
1921
1922    if (tex->op == OP_SUSTB || tex->op == OP_SUSTP) {
1923       condenseSrcs(tex, 3, (3 + typeSizeof(tex->dType) / 4) - 1);
1924    } else
1925    if (isTextureOp(tex->op)) {
1926       int n = tex->srcCount(0xff, true);
1927       if (n > 4) {
1928          condenseSrcs(tex, 0, 3);
1929          if (n > 5) // NOTE: first call modified positions already
1930             condenseSrcs(tex, 4 - (4 - 1), n - 1 - (4 - 1));
1931       } else
1932       if (n > 1) {
1933          condenseSrcs(tex, 0, n - 1);
1934       }
1935    }
1936 }
1937
1938 void
1939 RegAlloc::InsertConstraintsPass::texConstraintNVC0(TexInstruction *tex)
1940 {
1941    int n, s;
1942
1943    textureMask(tex);
1944
1945    if (tex->op == OP_TXQ) {
1946       s = tex->srcCount(0xff);
1947       n = 0;
1948    } else {
1949       s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
1950       if (!tex->tex.target.isArray() &&
1951           (tex->tex.rIndirectSrc >= 0 || tex->tex.sIndirectSrc >= 0))
1952          ++s;
1953       if (tex->op == OP_TXD && tex->tex.useOffsets)
1954          ++s;
1955       n = tex->srcCount(0xff) - s;
1956       assert(n <= 4);
1957    }
1958
1959    if (s > 1)
1960       condenseSrcs(tex, 0, s - 1);
1961    if (n > 1) // NOTE: first call modified positions already
1962       condenseSrcs(tex, 1, n);
1963
1964    condenseDefs(tex);
1965 }
1966
1967 void
1968 RegAlloc::InsertConstraintsPass::texConstraintNV50(TexInstruction *tex)
1969 {
1970    Value *pred = tex->getPredicate();
1971    if (pred)
1972       tex->setPredicate(tex->cc, NULL);
1973
1974    textureMask(tex);
1975
1976    assert(tex->defExists(0) && tex->srcExists(0));
1977    // make src and def count match
1978    int c;
1979    for (c = 0; tex->srcExists(c) || tex->defExists(c); ++c) {
1980       if (!tex->srcExists(c))
1981          tex->setSrc(c, new_LValue(func, tex->getSrc(0)->asLValue()));
1982       if (!tex->defExists(c))
1983          tex->setDef(c, new_LValue(func, tex->getDef(0)->asLValue()));
1984    }
1985    if (pred)
1986       tex->setPredicate(tex->cc, pred);
1987    condenseDefs(tex);
1988    condenseSrcs(tex, 0, c - 1);
1989 }
1990
1991 // Insert constraint markers for instructions whose multiple sources must be
1992 // located in consecutive registers.
1993 bool
1994 RegAlloc::InsertConstraintsPass::visit(BasicBlock *bb)
1995 {
1996    TexInstruction *tex;
1997    Instruction *next;
1998    int s, size;
1999
2000    targ = bb->getProgram()->getTarget();
2001
2002    for (Instruction *i = bb->getEntry(); i; i = next) {
2003       next = i->next;
2004
2005       if ((tex = i->asTex())) {
2006          switch (targ->getChipset() & ~0xf) {
2007          case 0x50:
2008          case 0x80:
2009          case 0x90:
2010          case 0xa0:
2011             texConstraintNV50(tex);
2012             break;
2013          case 0xc0:
2014          case 0xd0:
2015             texConstraintNVC0(tex);
2016             break;
2017          case 0xe0:
2018          case 0xf0:
2019          case 0x100:
2020             texConstraintNVE0(tex);
2021             break;
2022          case 0x110:
2023             texConstraintGM107(tex);
2024             break;
2025          default:
2026             break;
2027          }
2028       } else
2029       if (i->op == OP_EXPORT || i->op == OP_STORE) {
2030          for (size = typeSizeof(i->dType), s = 1; size > 0; ++s) {
2031             assert(i->srcExists(s));
2032             size -= i->getSrc(s)->reg.size;
2033          }
2034          condenseSrcs(i, 1, s - 1);
2035       } else
2036       if (i->op == OP_LOAD || i->op == OP_VFETCH) {
2037          condenseDefs(i);
2038          if (i->src(0).isIndirect(0) && typeSizeof(i->dType) >= 8)
2039             addHazard(i, i->src(0).getIndirect(0));
2040       } else
2041       if (i->op == OP_UNION ||
2042           i->op == OP_MERGE ||
2043           i->op == OP_SPLIT) {
2044          constrList.push_back(i);
2045       }
2046    }
2047    return true;
2048 }
2049
2050 // Insert extra moves so that, if multiple register constraints on a value are
2051 // in conflict, these conflicts can be resolved.
2052 bool
2053 RegAlloc::InsertConstraintsPass::insertConstraintMoves()
2054 {
2055    for (std::list<Instruction *>::iterator it = constrList.begin();
2056         it != constrList.end();
2057         ++it) {
2058       Instruction *cst = *it;
2059       Instruction *mov;
2060
2061       if (cst->op == OP_SPLIT && 0) {
2062          // spilling splits is annoying, just make sure they're separate
2063          for (int d = 0; cst->defExists(d); ++d) {
2064             if (!cst->getDef(d)->refCount())
2065                continue;
2066             LValue *lval = new_LValue(func, cst->def(d).getFile());
2067             const uint8_t size = cst->def(d).getSize();
2068             lval->reg.size = size;
2069
2070             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
2071             mov->setSrc(0, lval);
2072             mov->setDef(0, cst->getDef(d));
2073             cst->setDef(d, mov->getSrc(0));
2074             cst->bb->insertAfter(cst, mov);
2075
2076             cst->getSrc(0)->asLValue()->noSpill = 1;
2077             mov->getSrc(0)->asLValue()->noSpill = 1;
2078          }
2079       } else
2080       if (cst->op == OP_MERGE || cst->op == OP_UNION) {
2081          for (int s = 0; cst->srcExists(s); ++s) {
2082             const uint8_t size = cst->src(s).getSize();
2083
2084             if (!cst->getSrc(s)->defs.size()) {
2085                mov = new_Instruction(func, OP_NOP, typeOfSize(size));
2086                mov->setDef(0, cst->getSrc(s));
2087                cst->bb->insertBefore(cst, mov);
2088                continue;
2089             }
2090             assert(cst->getSrc(s)->defs.size() == 1); // still SSA
2091
2092             Instruction *defi = cst->getSrc(s)->defs.front()->getInsn();
2093             // catch some cases where don't really need MOVs
2094             if (cst->getSrc(s)->refCount() == 1 && !defi->constrainedDefs())
2095                continue;
2096
2097             LValue *lval = new_LValue(func, cst->src(s).getFile());
2098             lval->reg.size = size;
2099
2100             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
2101             mov->setDef(0, lval);
2102             mov->setSrc(0, cst->getSrc(s));
2103             cst->setSrc(s, mov->getDef(0));
2104             cst->bb->insertBefore(cst, mov);
2105
2106             cst->getDef(0)->asLValue()->noSpill = 1; // doesn't help
2107
2108             if (cst->op == OP_UNION)
2109                mov->setPredicate(defi->cc, defi->getPredicate());
2110          }
2111       }
2112    }
2113
2114    return true;
2115 }
2116
2117 } // namespace nv50_ir