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 <algorithm>
  27 #include <stack>
  28 #include <limits>
  29 #if __cplusplus >= 201103L
  30 #include <unordered_map>
  31 #else
  32 #include <tr1/unordered_map>
  33 #endif
  34
  35 namespace nv50_ir {
  36
  37 #if __cplusplus >= 201103L
  38 using std::hash;
  39 using std::unordered_map;
  40 #elif !defined(ANDROID)
  41 using std::tr1::hash;
  42 using std::tr1::unordered_map;
  43 #else
  44 #error Android release before Lollipop is not supported!
  45 #endif
  46
  47 #define MAX_REGISTER_FILE_SIZE 256
  48
  49 class RegisterSet
  50 {
  51 public:
  52    RegisterSet(const Target *);
  53
  54    void init(const Target *);
  55    void reset(DataFile, bool resetMax = false);
  56
  57    void periodicMask(DataFile f, uint32_t lock, uint32_t unlock);
  58    void intersect(DataFile f, const RegisterSet *);
  59
  60    bool assign(int32_t& reg, DataFile f, unsigned int size);
  61    void release(DataFile f, int32_t reg, unsigned int size);
  62    void occupy(DataFile f, int32_t reg, unsigned int size);
  63    void occupy(const Value *);
  64    void occupyMask(DataFile f, int32_t reg, uint8_t mask);
  65    bool isOccupied(DataFile f, int32_t reg, unsigned int size) const;
  66    bool testOccupy(const Value *);
  67    bool testOccupy(DataFile f, int32_t reg, unsigned int size);
  68
  69    inline int getMaxAssigned(DataFile f) const { return fill[f]; }
  70
  71    inline unsigned int getFileSize(DataFile f, uint8_t regSize) const
  72    {
  73       if (restrictedGPR16Range && f == FILE_GPR && regSize == 2)
  74          return (last[f] + 1) / 2;
  75       return last[f] + 1;
  76    }
  77
  78    inline unsigned int units(DataFile f, unsigned int size) const
  79    {
  80       return size >> unit[f];
  81    }
  82    // for regs of size >= 4, id is counted in 4-byte words (like nv50/c0 binary)
  83    inline unsigned int idToBytes(const Value *v) const
  84    {
  85       return v->reg.data.id * MIN2(v->reg.size, 4);
  86    }
  87    inline unsigned int idToUnits(const Value *v) const
  88    {
  89       return units(v->reg.file, idToBytes(v));
  90    }
  91    inline int bytesToId(Value *v, unsigned int bytes) const
  92    {
  93       if (v->reg.size < 4)
  94          return units(v->reg.file, bytes);
  95       return bytes / 4;
  96    }
  97    inline int unitsToId(DataFile f, int u, uint8_t size) const
  98    {
  99       if (u < 0)
 100          return -1;
 101       return (size < 4) ? u : ((u << unit[f]) / 4);
 102    }
 103
 104    void print(DataFile f) const;
 105
 106    const bool restrictedGPR16Range;
 107
 108 private:
 109    BitSet bits[LAST_REGISTER_FILE + 1];
 110
 111    int unit[LAST_REGISTER_FILE + 1]; // log2 of allocation granularity
 112
 113    int last[LAST_REGISTER_FILE + 1];
 114    int fill[LAST_REGISTER_FILE + 1];
 115 };
 116
 117 void
 118 RegisterSet::reset(DataFile f, bool resetMax)
 119 {
 120    bits[f].fill(0);
 121    if (resetMax)
 122       fill[f] = -1;
 123 }
 124
 125 void
 126 RegisterSet::init(const Target *targ)
 127 {
 128    for (unsigned int rf = 0; rf <= FILE_ADDRESS; ++rf) {
 129       DataFile f = static_cast<DataFile>(rf);
 130       last[rf] = targ->getFileSize(f) - 1;
 131       unit[rf] = targ->getFileUnit(f);
 132       fill[rf] = -1;
 133       assert(last[rf] < MAX_REGISTER_FILE_SIZE);
 134       bits[rf].allocate(last[rf] + 1, true);
 135    }
 136 }
 137
 138 RegisterSet::RegisterSet(const Target *targ)
 139   : restrictedGPR16Range(targ->getChipset() < 0xc0)
 140 {
 141    init(targ);
 142    for (unsigned int i = 0; i <= LAST_REGISTER_FILE; ++i)
 143       reset(static_cast<DataFile>(i));
 144 }
 145
 146 void
 147 RegisterSet::periodicMask(DataFile f, uint32_t lock, uint32_t unlock)
 148 {
 149    bits[f].periodicMask32(lock, unlock);
 150 }
 151
 152 void
 153 RegisterSet::intersect(DataFile f, const RegisterSet *set)
 154 {
 155    bits[f] |= set->bits[f];
 156 }
 157
 158 void
 159 RegisterSet::print(DataFile f) const
 160 {
 161    INFO("GPR:");
 162    bits[f].print();
 163    INFO("\n");
 164 }
 165
 166 bool
 167 RegisterSet::assign(int32_t& reg, DataFile f, unsigned int size)
 168 {
 169    reg = bits[f].findFreeRange(size);
 170    if (reg < 0)
 171       return false;
 172    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
 173    return true;
 174 }
 175
 176 bool
 177 RegisterSet::isOccupied(DataFile f, int32_t reg, unsigned int size) const
 178 {
 179    return bits[f].testRange(reg, size);
 180 }
 181
 182 void
 183 RegisterSet::occupy(const Value *v)
 184 {
 185    occupy(v->reg.file, idToUnits(v), v->reg.size >> unit[v->reg.file]);
 186 }
 187
 188 void
 189 RegisterSet::occupyMask(DataFile f, int32_t reg, uint8_t mask)
 190 {
 191    bits[f].setMask(reg & ~31, static_cast<uint32_t>(mask) << (reg % 32));
 192 }
 193
 194 void
 195 RegisterSet::occupy(DataFile f, int32_t reg, unsigned int size)
 196 {
 197    bits[f].setRange(reg, size);
 198
 199    INFO_DBG(0, REG_ALLOC, "reg occupy: %u[%i] %u\n", f, reg, size);
 200
 201    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
 202 }
 203
 204 bool
 205 RegisterSet::testOccupy(const Value *v)
 206 {
 207    return testOccupy(v->reg.file,
 208                      idToUnits(v), v->reg.size >> unit[v->reg.file]);
 209 }
 210
 211 bool
 212 RegisterSet::testOccupy(DataFile f, int32_t reg, unsigned int size)
 213 {
 214    if (isOccupied(f, reg, size))
 215       return false;
 216    occupy(f, reg, size);
 217    return true;
 218 }
 219
 220 void
 221 RegisterSet::release(DataFile f, int32_t reg, unsigned int size)
 222 {
 223    bits[f].clrRange(reg, size);
 224
 225    INFO_DBG(0, REG_ALLOC, "reg release: %u[%i] %u\n", f, reg, size);
 226 }
 227
 228 class RegAlloc
 229 {
 230 public:
 231    RegAlloc(Program *program) : prog(program), sequence(0) { }
 232
 233    bool exec();
 234    bool execFunc();
 235
 236 private:
 237    class PhiMovesPass : public Pass {
 238    private:
 239       virtual bool visit(BasicBlock *);
 240       inline bool needNewElseBlock(BasicBlock *b, BasicBlock *p);
 241       inline void splitEdges(BasicBlock *b);
 242    };
 243
 244    class ArgumentMovesPass : public Pass {
 245    private:
 246       virtual bool visit(BasicBlock *);
 247    };
 248
 249    class BuildIntervalsPass : public Pass {
 250    private:
 251       virtual bool visit(BasicBlock *);
 252       void collectLiveValues(BasicBlock *);
 253       void addLiveRange(Value *, const BasicBlock *, int end);
 254    };
 255
 256    class InsertConstraintsPass : public Pass {
 257    public:
 258       bool exec(Function *func);
 259    private:
 260       virtual bool visit(BasicBlock *);
 261
 262       bool insertConstraintMoves();
 263
 264       void condenseDefs(Instruction *);
 265       void condenseSrcs(Instruction *, const int first, const int last);
 266
 267       void addHazard(Instruction *i, const ValueRef *src);
 268       void textureMask(TexInstruction *);
 269       void addConstraint(Instruction *, int s, int n);
 270       bool detectConflict(Instruction *, int s);
 271
 272       // target specific functions, TODO: put in subclass or Target
 273       void texConstraintNV50(TexInstruction *);
 274       void texConstraintNVC0(TexInstruction *);
 275       void texConstraintNVE0(TexInstruction *);
 276       void texConstraintGM107(TexInstruction *);
 277
 278       std::list<Instruction *> constrList;
 279
 280       const Target *targ;
 281    };
 282
 283    bool buildLiveSets(BasicBlock *);
 284
 285 private:
 286    Program *prog;
 287    Function *func;
 288
 289    // instructions in control flow / chronological order
 290    ArrayList insns;
 291
 292    int sequence; // for manual passes through CFG
 293 };
 294
 295 typedef std::pair<Value *, Value *> ValuePair;
 296
 297 class SpillCodeInserter
 298 {
 299 public:
 300    SpillCodeInserter(Function *fn) : func(fn), stackSize(0), stackBase(0) { }
 301
 302    bool run(const std::list<ValuePair>&);
 303
 304    Symbol *assignSlot(const Interval&, const unsigned int size);
 305    Value *offsetSlot(Value *, const LValue *);
 306    inline int32_t getStackSize() const { return stackSize; }
 307
 308 private:
 309    Function *func;
 310
 311    struct SpillSlot
 312    {
 313       Interval occup;
 314       std::list<Value *> residents; // needed to recalculate occup
 315       Symbol *sym;
 316       int32_t offset;
 317       inline uint8_t size() const { return sym->reg.size; }
 318    };
 319    std::list<SpillSlot> slots;
 320    int32_t stackSize;
 321    int32_t stackBase;
 322
 323    LValue *unspill(Instruction *usei, LValue *, Value *slot);
 324    void spill(Instruction *defi, Value *slot, LValue *);
 325 };
 326
 327 void
 328 RegAlloc::BuildIntervalsPass::addLiveRange(Value *val,
 329                                            const BasicBlock *bb,
 330                                            int end)
 331 {
 332    Instruction *insn = val->getUniqueInsn();
 333
 334    if (!insn)
 335       insn = bb->getFirst();
 336
 337    assert(bb->getFirst()->serial <= bb->getExit()->serial);
 338    assert(bb->getExit()->serial + 1 >= end);
 339
 340    int begin = insn->serial;
 341    if (begin < bb->getEntry()->serial || begin > bb->getExit()->serial)
 342       begin = bb->getEntry()->serial;
 343
 344    INFO_DBG(prog->dbgFlags, REG_ALLOC, "%%%i <- live range [%i(%i), %i)\n",
 345             val->id, begin, insn->serial, end);
 346
 347    if (begin != end) // empty ranges are only added as hazards for fixed regs
 348       val->livei.extend(begin, end);
 349 }
 350
 351 bool
 352 RegAlloc::PhiMovesPass::needNewElseBlock(BasicBlock *b, BasicBlock *p)
 353 {
 354    if (b->cfg.incidentCount() <= 1)
 355       return false;
 356
 357    int n = 0;
 358    for (Graph::EdgeIterator ei = p->cfg.outgoing(); !ei.end(); ei.next())
 359       if (ei.getType() == Graph::Edge::TREE ||
 360           ei.getType() == Graph::Edge::FORWARD)
 361          ++n;
 362    return (n == 2);
 363 }
 364
 365 struct PhiMapHash {
 366    size_t operator()(const std::pair<Instruction *, BasicBlock *>& val) const {
 367       return hash<Instruction*>()(val.first) * 31 +
 368          hash<BasicBlock*>()(val.second);
 369    }
 370 };
 371
 372 typedef unordered_map<
 373    std::pair<Instruction *, BasicBlock *>, Value *, PhiMapHash> PhiMap;
 374
 375 // Critical edges need to be split up so that work can be inserted along
 376 // specific edge transitions. Unfortunately manipulating incident edges into a
 377 // BB invalidates all the PHI nodes since their sources are implicitly ordered
 378 // by incident edge order.
 379 //
 380 // TODO: Make it so that that is not the case, and PHI nodes store pointers to
 381 // the original BBs.
 382 void
 383 RegAlloc::PhiMovesPass::splitEdges(BasicBlock *bb)
 384 {
 385    BasicBlock *pb, *pn;
 386    Instruction *phi;
 387    Graph::EdgeIterator ei;
 388    std::stack<BasicBlock *> stack;
 389    int j = 0;
 390
 391    for (ei = bb->cfg.incident(); !ei.end(); ei.next()) {
 392       pb = BasicBlock::get(ei.getNode());
 393       assert(pb);
 394       if (needNewElseBlock(bb, pb))
 395          stack.push(pb);
 396    }
 397
 398    // No critical edges were found, no need to perform any work.
 399    if (stack.empty())
 400       return;
 401
 402    // We're about to, potentially, reorder the inbound edges. This means that
 403    // we need to hold on to the (phi, bb) -> src mapping, and fix up the phi
 404    // nodes after the graph has been modified.
 405    PhiMap phis;
 406
 407    j = 0;
 408    for (ei = bb->cfg.incident(); !ei.end(); ei.next(), j++) {
 409       pb = BasicBlock::get(ei.getNode());
 410       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next)
 411          phis.insert(std::make_pair(std::make_pair(phi, pb), phi->getSrc(j)));
 412    }
 413
 414    while (!stack.empty()) {
 415       pb = stack.top();
 416       pn = new BasicBlock(func);
 417       stack.pop();
 418
 419       pb->cfg.detach(&bb->cfg);
 420       pb->cfg.attach(&pn->cfg, Graph::Edge::TREE);
 421       pn->cfg.attach(&bb->cfg, Graph::Edge::FORWARD);
 422
 423       assert(pb->getExit()->op != OP_CALL);
 424       if (pb->getExit()->asFlow()->target.bb == bb)
 425          pb->getExit()->asFlow()->target.bb = pn;
 426
 427       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 428          PhiMap::iterator it = phis.find(std::make_pair(phi, pb));
 429          assert(it != phis.end());
 430          phis.insert(std::make_pair(std::make_pair(phi, pn), it->second));
 431          phis.erase(it);
 432       }
 433    }
 434
 435    // Now go through and fix up all of the phi node sources.
 436    j = 0;
 437    for (ei = bb->cfg.incident(); !ei.end(); ei.next(), j++) {
 438       pb = BasicBlock::get(ei.getNode());
 439       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 440          PhiMap::const_iterator it = phis.find(std::make_pair(phi, pb));
 441          assert(it != phis.end());
 442
 443          phi->setSrc(j, it->second);
 444       }
 445    }
 446 }
 447
 448 // For each operand of each PHI in b, generate a new value by inserting a MOV
 449 // at the end of the block it is coming from and replace the operand with its
 450 // result. This eliminates liveness conflicts and enables us to let values be
 451 // copied to the right register if such a conflict exists nonetheless.
 452 //
 453 // These MOVs are also crucial in making sure the live intervals of phi srces
 454 // are extended until the end of the loop, since they are not included in the
 455 // live-in sets.
 456 bool
 457 RegAlloc::PhiMovesPass::visit(BasicBlock *bb)
 458 {
 459    Instruction *phi, *mov;
 460
 461    splitEdges(bb);
 462
 463    // insert MOVs (phi->src(j) should stem from j-th in-BB)
 464    int j = 0;
 465    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
 466       BasicBlock *pb = BasicBlock::get(ei.getNode());
 467       if (!pb->isTerminated())
 468          pb->insertTail(new_FlowInstruction(func, OP_BRA, bb));
 469
 470       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 471          LValue *tmp = new_LValue(func, phi->getDef(0)->asLValue());
 472          mov = new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 473
 474          mov->setSrc(0, phi->getSrc(j));
 475          mov->setDef(0, tmp);
 476          phi->setSrc(j, tmp);
 477
 478          pb->insertBefore(pb->getExit(), mov);
 479       }
 480       ++j;
 481    }
 482
 483    return true;
 484 }
 485
 486 bool
 487 RegAlloc::ArgumentMovesPass::visit(BasicBlock *bb)
 488 {
 489    // Bind function call inputs/outputs to the same physical register
 490    // the callee uses, inserting moves as appropriate for the case a
 491    // conflict arises.
 492    for (Instruction *i = bb->getEntry(); i; i = i->next) {
 493       FlowInstruction *cal = i->asFlow();
 494       // TODO: Handle indirect calls.
 495       // Right now they should only be generated for builtins.
 496       if (!cal || cal->op != OP_CALL || cal->builtin || cal->indirect)
 497          continue;
 498       RegisterSet clobberSet(prog->getTarget());
 499
 500       // Bind input values.
 501       for (int s = cal->indirect ? 1 : 0; cal->srcExists(s); ++s) {
 502          const int t = cal->indirect ? (s - 1) : s;
 503          LValue *tmp = new_LValue(func, cal->getSrc(s)->asLValue());
 504          tmp->reg.data.id = cal->target.fn->ins[t].rep()->reg.data.id;
 505
 506          Instruction *mov =
 507             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 508          mov->setDef(0, tmp);
 509          mov->setSrc(0, cal->getSrc(s));
 510          cal->setSrc(s, tmp);
 511
 512          bb->insertBefore(cal, mov);
 513       }
 514
 515       // Bind output values.
 516       for (int d = 0; cal->defExists(d); ++d) {
 517          LValue *tmp = new_LValue(func, cal->getDef(d)->asLValue());
 518          tmp->reg.data.id = cal->target.fn->outs[d].rep()->reg.data.id;
 519
 520          Instruction *mov =
 521             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 522          mov->setSrc(0, tmp);
 523          mov->setDef(0, cal->getDef(d));
 524          cal->setDef(d, tmp);
 525
 526          bb->insertAfter(cal, mov);
 527          clobberSet.occupy(tmp);
 528       }
 529
 530       // Bind clobbered values.
 531       for (std::deque<Value *>::iterator it = cal->target.fn->clobbers.begin();
 532            it != cal->target.fn->clobbers.end();
 533            ++it) {
 534          if (clobberSet.testOccupy(*it)) {
 535             Value *tmp = new_LValue(func, (*it)->asLValue());
 536             tmp->reg.data.id = (*it)->reg.data.id;
 537             cal->setDef(cal->defCount(), tmp);
 538          }
 539       }
 540    }
 541
 542    // Update the clobber set of the function.
 543    if (BasicBlock::get(func->cfgExit) == bb) {
 544       func->buildDefSets();
 545       for (unsigned int i = 0; i < bb->defSet.getSize(); ++i)
 546          if (bb->defSet.test(i))
 547             func->clobbers.push_back(func->getLValue(i));
 548    }
 549
 550    return true;
 551 }
 552
 553 // Build the set of live-in variables of bb.
 554 bool
 555 RegAlloc::buildLiveSets(BasicBlock *bb)
 556 {
 557    Function *f = bb->getFunction();
 558    BasicBlock *bn;
 559    Instruction *i;
 560    unsigned int s, d;
 561
 562    INFO_DBG(prog->dbgFlags, REG_ALLOC, "buildLiveSets(BB:%i)\n", bb->getId());
 563
 564    bb->liveSet.allocate(func->allLValues.getSize(), false);
 565
 566    int n = 0;
 567    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 568       bn = BasicBlock::get(ei.getNode());
 569       if (bn == bb)
 570          continue;
 571       if (bn->cfg.visit(sequence))
 572          if (!buildLiveSets(bn))
 573             return false;
 574       if (n++ || bb->liveSet.marker)
 575          bb->liveSet |= bn->liveSet;
 576       else
 577          bb->liveSet = bn->liveSet;
 578    }
 579    if (!n && !bb->liveSet.marker)
 580       bb->liveSet.fill(0);
 581    bb->liveSet.marker = true;
 582
 583    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 584       INFO("BB:%i live set of out blocks:\n", bb->getId());
 585       bb->liveSet.print();
 586    }
 587
 588    // if (!bb->getEntry())
 589    //   return true;
 590
 591    if (bb == BasicBlock::get(f->cfgExit)) {
 592       for (std::deque<ValueRef>::iterator it = f->outs.begin();
 593            it != f->outs.end(); ++it) {
 594          assert(it->get()->asLValue());
 595          bb->liveSet.set(it->get()->id);
 596       }
 597    }
 598
 599    for (i = bb->getExit(); i && i != bb->getEntry()->prev; i = i->prev) {
 600       for (d = 0; i->defExists(d); ++d)
 601          bb->liveSet.clr(i->getDef(d)->id);
 602       for (s = 0; i->srcExists(s); ++s)
 603          if (i->getSrc(s)->asLValue())
 604             bb->liveSet.set(i->getSrc(s)->id);
 605    }
 606    for (i = bb->getPhi(); i && i->op == OP_PHI; i = i->next)
 607       bb->liveSet.clr(i->getDef(0)->id);
 608
 609    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 610       INFO("BB:%i live set after propagation:\n", bb->getId());
 611       bb->liveSet.print();
 612    }
 613
 614    return true;
 615 }
 616
 617 void
 618 RegAlloc::BuildIntervalsPass::collectLiveValues(BasicBlock *bb)
 619 {
 620    BasicBlock *bbA = NULL, *bbB = NULL;
 621
 622    if (bb->cfg.outgoingCount()) {
 623       // trickery to save a loop of OR'ing liveSets
 624       // aliasing works fine with BitSet::setOr
 625       for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 626          if (ei.getType() == Graph::Edge::DUMMY)
 627             continue;
 628          if (bbA) {
 629             bb->liveSet.setOr(&bbA->liveSet, &bbB->liveSet);
 630             bbA = bb;
 631          } else {
 632             bbA = bbB;
 633          }
 634          bbB = BasicBlock::get(ei.getNode());
 635       }
 636       bb->liveSet.setOr(&bbB->liveSet, bbA ? &bbA->liveSet : NULL);
 637    } else
 638    if (bb->cfg.incidentCount()) {
 639       bb->liveSet.fill(0);
 640    }
 641 }
 642
 643 bool
 644 RegAlloc::BuildIntervalsPass::visit(BasicBlock *bb)
 645 {
 646    collectLiveValues(bb);
 647
 648    INFO_DBG(prog->dbgFlags, REG_ALLOC, "BuildIntervals(BB:%i)\n", bb->getId());
 649
 650    // go through out blocks and delete phi sources that do not originate from
 651    // the current block from the live set
 652    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 653       BasicBlock *out = BasicBlock::get(ei.getNode());
 654
 655       for (Instruction *i = out->getPhi(); i && i->op == OP_PHI; i = i->next) {
 656          bb->liveSet.clr(i->getDef(0)->id);
 657
 658          for (int s = 0; i->srcExists(s); ++s) {
 659             assert(i->src(s).getInsn());
 660             if (i->getSrc(s)->getUniqueInsn()->bb == bb) // XXX: reachableBy ?
 661                bb->liveSet.set(i->getSrc(s)->id);
 662             else
 663                bb->liveSet.clr(i->getSrc(s)->id);
 664          }
 665       }
 666    }
 667
 668    // remaining live-outs are live until end
 669    if (bb->getExit()) {
 670       for (unsigned int j = 0; j < bb->liveSet.getSize(); ++j)
 671          if (bb->liveSet.test(j))
 672             addLiveRange(func->getLValue(j), bb, bb->getExit()->serial + 1);
 673    }
 674
 675    for (Instruction *i = bb->getExit(); i && i->op != OP_PHI; i = i->prev) {
 676       for (int d = 0; i->defExists(d); ++d) {
 677          bb->liveSet.clr(i->getDef(d)->id);
 678          if (i->getDef(d)->reg.data.id >= 0) // add hazard for fixed regs
 679             i->getDef(d)->livei.extend(i->serial, i->serial);
 680       }
 681
 682       for (int s = 0; i->srcExists(s); ++s) {
 683          if (!i->getSrc(s)->asLValue())
 684             continue;
 685          if (!bb->liveSet.test(i->getSrc(s)->id)) {
 686             bb->liveSet.set(i->getSrc(s)->id);
 687             addLiveRange(i->getSrc(s), bb, i->serial);
 688          }
 689       }
 690    }
 691
 692    if (bb == BasicBlock::get(func->cfg.getRoot())) {
 693       for (std::deque<ValueDef>::iterator it = func->ins.begin();
 694            it != func->ins.end(); ++it) {
 695          if (it->get()->reg.data.id >= 0) // add hazard for fixed regs
 696             it->get()->livei.extend(0, 1);
 697       }
 698    }
 699
 700    return true;
 701 }
 702
 703
 704 #define JOIN_MASK_PHI        (1 << 0)
 705 #define JOIN_MASK_UNION      (1 << 1)
 706 #define JOIN_MASK_MOV        (1 << 2)
 707 #define JOIN_MASK_TEX        (1 << 3)
 708
 709 class GCRA
 710 {
 711 public:
 712    GCRA(Function *, SpillCodeInserter&);
 713    ~GCRA();
 714
 715    bool allocateRegisters(ArrayList& insns);
 716
 717    void printNodeInfo() const;
 718
 719 private:
 720    class RIG_Node : public Graph::Node
 721    {
 722    public:
 723       RIG_Node();
 724
 725       void init(const RegisterSet&, LValue *);
 726
 727       void addInterference(RIG_Node *);
 728       void addRegPreference(RIG_Node *);
 729
 730       inline LValue *getValue() const
 731       {
 732          return reinterpret_cast<LValue *>(data);
 733       }
 734       inline void setValue(LValue *lval) { data = lval; }
 735
 736       inline uint8_t getCompMask() const
 737       {
 738          return ((1 << colors) - 1) << (reg & 7);
 739       }
 740
 741       static inline RIG_Node *get(const Graph::EdgeIterator& ei)
 742       {
 743          return static_cast<RIG_Node *>(ei.getNode());
 744       }
 745
 746    public:
 747       uint32_t degree;
 748       uint16_t degreeLimit; // if deg < degLimit, node is trivially colourable
 749       uint16_t colors;
 750
 751       DataFile f;
 752       int32_t reg;
 753
 754       float weight;
 755
 756       // list pointers for simplify() phase
 757       RIG_Node *next;
 758       RIG_Node *prev;
 759
 760       // union of the live intervals of all coalesced values (we want to retain
 761       //  the separate intervals for testing interference of compound values)
 762       Interval livei;
 763
 764       std::list<RIG_Node *> prefRegs;
 765    };
 766
 767 private:
 768    inline RIG_Node *getNode(const LValue *v) const { return &nodes[v->id]; }
 769
 770    void buildRIG(ArrayList&);
 771    bool coalesce(ArrayList&);
 772    bool doCoalesce(ArrayList&, unsigned int mask);
 773    void calculateSpillWeights();
 774    void simplify();
 775    bool selectRegisters();
 776    void cleanup(const bool success);
 777
 778    void simplifyEdge(RIG_Node *, RIG_Node *);
 779    void simplifyNode(RIG_Node *);
 780
 781    bool coalesceValues(Value *, Value *, bool force);
 782    void resolveSplitsAndMerges();
 783    void makeCompound(Instruction *, bool isSplit);
 784
 785    inline void checkInterference(const RIG_Node *, Graph::EdgeIterator&);
 786
 787    inline void insertOrderedTail(std::list<RIG_Node *>&, RIG_Node *);
 788    void checkList(std::list<RIG_Node *>&);
 789
 790 private:
 791    std::stack<uint32_t> stack;
 792
 793    // list headers for simplify() phase
 794    RIG_Node lo[2];
 795    RIG_Node hi;
 796
 797    Graph RIG;
 798    RIG_Node *nodes;
 799    unsigned int nodeCount;
 800
 801    Function *func;
 802    Program *prog;
 803
 804    static uint8_t relDegree[17][17];
 805
 806    RegisterSet regs;
 807
 808    // need to fixup register id for participants of OP_MERGE/SPLIT
 809    std::list<Instruction *> merges;
 810    std::list<Instruction *> splits;
 811
 812    SpillCodeInserter& spill;
 813    std::list<ValuePair> mustSpill;
 814 };
 815
 816 uint8_t GCRA::relDegree[17][17];
 817
 818 GCRA::RIG_Node::RIG_Node() : Node(NULL), next(this), prev(this)
 819 {
 820    colors = 0;
 821 }
 822
 823 void
 824 GCRA::printNodeInfo() const
 825 {
 826    for (unsigned int i = 0; i < nodeCount; ++i) {
 827       if (!nodes[i].colors)
 828          continue;
 829       INFO("RIG_Node[%%%i]($[%u]%i): %u colors, weight %f, deg %u/%u\n X",
 830            i,
 831            nodes[i].f,nodes[i].reg,nodes[i].colors,
 832            nodes[i].weight,
 833            nodes[i].degree, nodes[i].degreeLimit);
 834
 835       for (Graph::EdgeIterator ei = nodes[i].outgoing(); !ei.end(); ei.next())
 836          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
 837       for (Graph::EdgeIterator ei = nodes[i].incident(); !ei.end(); ei.next())
 838          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
 839       INFO("\n");
 840    }
 841 }
 842
 843 static bool
 844 isShortRegOp(Instruction *insn)
 845 {
 846    // Immediates are always in src1. Every other situation can be resolved by
 847    // using a long encoding.
 848    return insn->srcExists(1) && insn->src(1).getFile() == FILE_IMMEDIATE;
 849 }
 850
 851 // Check if this LValue is ever used in an instruction that can't be encoded
 852 // with long registers (i.e. > r63)
 853 static bool
 854 isShortRegVal(LValue *lval)
 855 {
 856    if (lval->getInsn() == NULL)
 857       return false;
 858    for (Value::DefCIterator def = lval->defs.begin();
 859         def != lval->defs.end(); ++def)
 860       if (isShortRegOp((*def)->getInsn()))
 861          return true;
 862    for (Value::UseCIterator use = lval->uses.begin();
 863         use != lval->uses.end(); ++use)
 864       if (isShortRegOp((*use)->getInsn()))
 865          return true;
 866    return false;
 867 }
 868
 869 void
 870 GCRA::RIG_Node::init(const RegisterSet& regs, LValue *lval)
 871 {
 872    setValue(lval);
 873    if (lval->reg.data.id >= 0)
 874       lval->noSpill = lval->fixedReg = 1;
 875
 876    colors = regs.units(lval->reg.file, lval->reg.size);
 877    f = lval->reg.file;
 878    reg = -1;
 879    if (lval->reg.data.id >= 0)
 880       reg = regs.idToUnits(lval);
 881
 882    weight = std::numeric_limits<float>::infinity();
 883    degree = 0;
 884    int size = regs.getFileSize(f, lval->reg.size);
 885    // On nv50, we lose a bit of gpr encoding when there's an embedded
 886    // immediate.
 887    if (regs.restrictedGPR16Range && f == FILE_GPR && isShortRegVal(lval))
 888       size /= 2;
 889    degreeLimit = size;
 890    degreeLimit -= relDegree[1][colors] - 1;
 891
 892    livei.insert(lval->livei);
 893 }
 894
 895 bool
 896 GCRA::coalesceValues(Value *dst, Value *src, bool force)
 897 {
 898    LValue *rep = dst->join->asLValue();
 899    LValue *val = src->join->asLValue();
 900
 901    if (!force && val->reg.data.id >= 0) {
 902       rep = src->join->asLValue();
 903       val = dst->join->asLValue();
 904    }
 905    RIG_Node *nRep = &nodes[rep->id];
 906    RIG_Node *nVal = &nodes[val->id];
 907
 908    if (src->reg.file != dst->reg.file) {
 909       if (!force)
 910          return false;
 911       WARN("forced coalescing of values in different files !\n");
 912    }
 913    if (!force && dst->reg.size != src->reg.size)
 914       return false;
 915
 916    if ((rep->reg.data.id >= 0) && (rep->reg.data.id != val->reg.data.id)) {
 917       if (force) {
 918          if (val->reg.data.id >= 0)
 919             WARN("forced coalescing of values in different fixed regs !\n");
 920       } else {
 921          if (val->reg.data.id >= 0)
 922             return false;
 923          // make sure that there is no overlap with the fixed register of rep
 924          for (ArrayList::Iterator it = func->allLValues.iterator();
 925               !it.end(); it.next()) {
 926             Value *reg = reinterpret_cast<Value *>(it.get())->asLValue();
 927             assert(reg);
 928             if (reg->interfers(rep) && reg->livei.overlaps(nVal->livei))
 929                return false;
 930          }
 931       }
 932    }
 933
 934    if (!force && nRep->livei.overlaps(nVal->livei))
 935       return false;
 936
 937    INFO_DBG(prog->dbgFlags, REG_ALLOC, "joining %%%i($%i) <- %%%i\n",
 938             rep->id, rep->reg.data.id, val->id);
 939
 940    // set join pointer of all values joined with val
 941    for (Value::DefIterator def = val->defs.begin(); def != val->defs.end();
 942         ++def)
 943       (*def)->get()->join = rep;
 944    assert(rep->join == rep && val->join == rep);
 945
 946    // add val's definitions to rep and extend the live interval of its RIG node
 947    rep->defs.insert(rep->defs.end(), val->defs.begin(), val->defs.end());
 948    nRep->livei.unify(nVal->livei);
 949    return true;
 950 }
 951
 952 bool
 953 GCRA::coalesce(ArrayList& insns)
 954 {
 955    bool ret = doCoalesce(insns, JOIN_MASK_PHI);
 956    if (!ret)
 957       return false;
 958    switch (func->getProgram()->getTarget()->getChipset() & ~0xf) {
 959    case 0x50:
 960    case 0x80:
 961    case 0x90:
 962    case 0xa0:
 963       ret = doCoalesce(insns, JOIN_MASK_UNION | JOIN_MASK_TEX);
 964       break;
 965    case 0xc0:
 966    case 0xd0:
 967    case 0xe0:
 968    case 0xf0:
 969    case 0x100:
 970    case 0x110:
 971    case 0x120:
 972    case 0x130:
 973       ret = doCoalesce(insns, JOIN_MASK_UNION);
 974       break;
 975    default:
 976       break;
 977    }
 978    if (!ret)
 979       return false;
 980    return doCoalesce(insns, JOIN_MASK_MOV);
 981 }
 982
 983 static inline uint8_t makeCompMask(int compSize, int base, int size)
 984 {
 985    uint8_t m = ((1 << size) - 1) << base;
 986
 987    switch (compSize) {
 988    case 1:
 989       return 0xff;
 990    case 2:
 991       m |= (m << 2);
 992       return (m << 4) | m;
 993    case 3:
 994    case 4:
 995       return (m << 4) | m;
 996    default:
 997       assert(compSize <= 8);
 998       return m;
 999    }
1000 }
1001
1002 // Used when coalescing moves. The non-compound value will become one, e.g.:
1003 // mov b32 $r0 $r2            / merge b64 $r0d { $r0 $r1 }
1004 // split b64 { $r0 $r1 } $r0d / mov b64 $r0d f64 $r2d
1005 static inline void copyCompound(Value *dst, Value *src)
1006 {
1007    LValue *ldst = dst->asLValue();
1008    LValue *lsrc = src->asLValue();
1009
1010    if (ldst->compound && !lsrc->compound) {
1011       LValue *swap = lsrc;
1012       lsrc = ldst;
1013       ldst = swap;
1014    }
1015
1016    ldst->compound = lsrc->compound;
1017    ldst->compMask = lsrc->compMask;
1018 }
1019
1020 void
1021 GCRA::makeCompound(Instruction *insn, bool split)
1022 {
1023    LValue *rep = (split ? insn->getSrc(0) : insn->getDef(0))->asLValue();
1024
1025    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
1026       INFO("makeCompound(split = %i): ", split);
1027       insn->print();
1028    }
1029
1030    const unsigned int size = getNode(rep)->colors;
1031    unsigned int base = 0;
1032
1033    if (!rep->compound)
1034       rep->compMask = 0xff;
1035    rep->compound = 1;
1036
1037    for (int c = 0; split ? insn->defExists(c) : insn->srcExists(c); ++c) {
1038       LValue *val = (split ? insn->getDef(c) : insn->getSrc(c))->asLValue();
1039
1040       val->compound = 1;
1041       if (!val->compMask)
1042          val->compMask = 0xff;
1043       val->compMask &= makeCompMask(size, base, getNode(val)->colors);
1044       assert(val->compMask);
1045
1046       INFO_DBG(prog->dbgFlags, REG_ALLOC, "compound: %%%i:%02x <- %%%i:%02x\n",
1047            rep->id, rep->compMask, val->id, val->compMask);
1048
1049       base += getNode(val)->colors;
1050    }
1051    assert(base == size);
1052 }
1053
1054 bool
1055 GCRA::doCoalesce(ArrayList& insns, unsigned int mask)
1056 {
1057    int c, n;
1058
1059    for (n = 0; n < insns.getSize(); ++n) {
1060       Instruction *i;
1061       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(n));
1062
1063       switch (insn->op) {
1064       case OP_PHI:
1065          if (!(mask & JOIN_MASK_PHI))
1066             break;
1067          for (c = 0; insn->srcExists(c); ++c)
1068             if (!coalesceValues(insn->getDef(0), insn->getSrc(c), false)) {
1069                // this is bad
1070                ERROR("failed to coalesce phi operands\n");
1071                return false;
1072             }
1073          break;
1074       case OP_UNION:
1075       case OP_MERGE:
1076          if (!(mask & JOIN_MASK_UNION))
1077             break;
1078          for (c = 0; insn->srcExists(c); ++c)
1079             coalesceValues(insn->getDef(0), insn->getSrc(c), true);
1080          if (insn->op == OP_MERGE) {
1081             merges.push_back(insn);
1082             if (insn->srcExists(1))
1083                makeCompound(insn, false);
1084          }
1085          break;
1086       case OP_SPLIT:
1087          if (!(mask & JOIN_MASK_UNION))
1088             break;
1089          splits.push_back(insn);
1090          for (c = 0; insn->defExists(c); ++c)
1091             coalesceValues(insn->getSrc(0), insn->getDef(c), true);
1092          makeCompound(insn, true);
1093          break;
1094       case OP_MOV:
1095          if (!(mask & JOIN_MASK_MOV))
1096             break;
1097          i = NULL;
1098          if (!insn->getDef(0)->uses.empty())
1099             i = (*insn->getDef(0)->uses.begin())->getInsn();
1100          // if this is a contraint-move there will only be a single use
1101          if (i && i->op == OP_MERGE) // do we really still need this ?
1102             break;
1103          i = insn->getSrc(0)->getUniqueInsn();
1104          if (i && !i->constrainedDefs()) {
1105             if (coalesceValues(insn->getDef(0), insn->getSrc(0), false))
1106                copyCompound(insn->getSrc(0), insn->getDef(0));
1107          }
1108          break;
1109       case OP_TEX:
1110       case OP_TXB:
1111       case OP_TXL:
1112       case OP_TXF:
1113       case OP_TXQ:
1114       case OP_TXD:
1115       case OP_TXG:
1116       case OP_TXLQ:
1117       case OP_TEXCSAA:
1118       case OP_TEXPREP:
1119          if (!(mask & JOIN_MASK_TEX))
1120             break;
1121          for (c = 0; insn->srcExists(c) && c != insn->predSrc; ++c)
1122             coalesceValues(insn->getDef(c), insn->getSrc(c), true);
1123          break;
1124       default:
1125          break;
1126       }
1127    }
1128    return true;
1129 }
1130
1131 void
1132 GCRA::RIG_Node::addInterference(RIG_Node *node)
1133 {
1134    this->degree += relDegree[node->colors][colors];
1135    node->degree += relDegree[colors][node->colors];
1136
1137    this->attach(node, Graph::Edge::CROSS);
1138 }
1139
1140 void
1141 GCRA::RIG_Node::addRegPreference(RIG_Node *node)
1142 {
1143    prefRegs.push_back(node);
1144 }
1145
1146 GCRA::GCRA(Function *fn, SpillCodeInserter& spill) :
1147    func(fn),
1148    regs(fn->getProgram()->getTarget()),
1149    spill(spill)
1150 {
1151    prog = func->getProgram();
1152
1153    // initialize relative degrees array - i takes away from j
1154    for (int i = 1; i <= 16; ++i)
1155       for (int j = 1; j <= 16; ++j)
1156          relDegree[i][j] = j * ((i + j - 1) / j);
1157 }
1158
1159 GCRA::~GCRA()
1160 {
1161    if (nodes)
1162       delete[] nodes;
1163 }
1164
1165 void
1166 GCRA::checkList(std::list<RIG_Node *>& lst)
1167 {
1168    GCRA::RIG_Node *prev = NULL;
1169
1170    for (std::list<RIG_Node *>::iterator it = lst.begin();
1171         it != lst.end();
1172         ++it) {
1173       assert((*it)->getValue()->join == (*it)->getValue());
1174       if (prev)
1175          assert(prev->livei.begin() <= (*it)->livei.begin());
1176       prev = *it;
1177    }
1178 }
1179
1180 void
1181 GCRA::insertOrderedTail(std::list<RIG_Node *>& list, RIG_Node *node)
1182 {
1183    if (node->livei.isEmpty())
1184       return;
1185    // only the intervals of joined values don't necessarily arrive in order
1186    std::list<RIG_Node *>::iterator prev, it;
1187    for (it = list.end(); it != list.begin(); it = prev) {
1188       prev = it;
1189       --prev;
1190       if ((*prev)->livei.begin() <= node->livei.begin())
1191          break;
1192    }
1193    list.insert(it, node);
1194 }
1195
1196 void
1197 GCRA::buildRIG(ArrayList& insns)
1198 {
1199    std::list<RIG_Node *> values, active;
1200
1201    for (std::deque<ValueDef>::iterator it = func->ins.begin();
1202         it != func->ins.end(); ++it)
1203       insertOrderedTail(values, getNode(it->get()->asLValue()));
1204
1205    for (int i = 0; i < insns.getSize(); ++i) {
1206       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(i));
1207       for (int d = 0; insn->defExists(d); ++d)
1208          if (insn->getDef(d)->rep() == insn->getDef(d))
1209             insertOrderedTail(values, getNode(insn->getDef(d)->asLValue()));
1210    }
1211    checkList(values);
1212
1213    while (!values.empty()) {
1214       RIG_Node *cur = values.front();
1215
1216       for (std::list<RIG_Node *>::iterator it = active.begin();
1217            it != active.end();) {
1218          RIG_Node *node = *it;
1219
1220          if (node->livei.end() <= cur->livei.begin()) {
1221             it = active.erase(it);
1222          } else {
1223             if (node->f == cur->f && node->livei.overlaps(cur->livei))
1224                cur->addInterference(node);
1225             ++it;
1226          }
1227       }
1228       values.pop_front();
1229       active.push_back(cur);
1230    }
1231 }
1232
1233 void
1234 GCRA::calculateSpillWeights()
1235 {
1236    for (unsigned int i = 0; i < nodeCount; ++i) {
1237       RIG_Node *const n = &nodes[i];
1238       if (!nodes[i].colors || nodes[i].livei.isEmpty())
1239          continue;
1240       if (nodes[i].reg >= 0) {
1241          // update max reg
1242          regs.occupy(n->f, n->reg, n->colors);
1243          continue;
1244       }
1245       LValue *val = nodes[i].getValue();
1246
1247       if (!val->noSpill) {
1248          int rc = 0;
1249          for (Value::DefIterator it = val->defs.begin();
1250               it != val->defs.end();
1251               ++it)
1252             rc += (*it)->get()->refCount();
1253
1254          nodes[i].weight =
1255             (float)rc * (float)rc / (float)nodes[i].livei.extent();
1256       }
1257
1258       if (nodes[i].degree < nodes[i].degreeLimit) {
1259          int l = 0;
1260          if (val->reg.size > 4)
1261             l = 1;
1262          DLLIST_ADDHEAD(&lo[l], &nodes[i]);
1263       } else {
1264          DLLIST_ADDHEAD(&hi, &nodes[i]);
1265       }
1266    }
1267    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1268       printNodeInfo();
1269 }
1270
1271 void
1272 GCRA::simplifyEdge(RIG_Node *a, RIG_Node *b)
1273 {
1274    bool move = b->degree >= b->degreeLimit;
1275
1276    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1277             "edge: (%%%i, deg %u/%u) >-< (%%%i, deg %u/%u)\n",
1278             a->getValue()->id, a->degree, a->degreeLimit,
1279             b->getValue()->id, b->degree, b->degreeLimit);
1280
1281    b->degree -= relDegree[a->colors][b->colors];
1282
1283    move = move && b->degree < b->degreeLimit;
1284    if (move && !DLLIST_EMPTY(b)) {
1285       int l = (b->getValue()->reg.size > 4) ? 1 : 0;
1286       DLLIST_DEL(b);
1287       DLLIST_ADDTAIL(&lo[l], b);
1288    }
1289 }
1290
1291 void
1292 GCRA::simplifyNode(RIG_Node *node)
1293 {
1294    for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1295       simplifyEdge(node, RIG_Node::get(ei));
1296
1297    for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1298       simplifyEdge(node, RIG_Node::get(ei));
1299
1300    DLLIST_DEL(node);
1301    stack.push(node->getValue()->id);
1302
1303    INFO_DBG(prog->dbgFlags, REG_ALLOC, "SIMPLIFY: pushed %%%i%s\n",
1304             node->getValue()->id,
1305             (node->degree < node->degreeLimit) ? "" : "(spill)");
1306 }
1307
1308 void
1309 GCRA::simplify()
1310 {
1311    for (;;) {
1312       if (!DLLIST_EMPTY(&lo[0])) {
1313          do {
1314             simplifyNode(lo[0].next);
1315          } while (!DLLIST_EMPTY(&lo[0]));
1316       } else
1317       if (!DLLIST_EMPTY(&lo[1])) {
1318          simplifyNode(lo[1].next);
1319       } else
1320       if (!DLLIST_EMPTY(&hi)) {
1321          RIG_Node *best = hi.next;
1322          float bestScore = best->weight / (float)best->degree;
1323          // spill candidate
1324          for (RIG_Node *it = best->next; it != &hi; it = it->next) {
1325             float score = it->weight / (float)it->degree;
1326             if (score < bestScore) {
1327                best = it;
1328                bestScore = score;
1329             }
1330          }
1331          if (isinf(bestScore)) {
1332             ERROR("no viable spill candidates left\n");
1333             break;
1334          }
1335          simplifyNode(best);
1336       } else {
1337          break;
1338       }
1339    }
1340 }
1341
1342 void
1343 GCRA::checkInterference(const RIG_Node *node, Graph::EdgeIterator& ei)
1344 {
1345    const RIG_Node *intf = RIG_Node::get(ei);
1346
1347    if (intf->reg < 0)
1348       return;
1349    const LValue *vA = node->getValue();
1350    const LValue *vB = intf->getValue();
1351
1352    const uint8_t intfMask = ((1 << intf->colors) - 1) << (intf->reg & 7);
1353
1354    if (vA->compound | vB->compound) {
1355       // NOTE: this only works for >aligned< register tuples !
1356       for (Value::DefCIterator D = vA->defs.begin(); D != vA->defs.end(); ++D) {
1357       for (Value::DefCIterator d = vB->defs.begin(); d != vB->defs.end(); ++d) {
1358          const LValue *vD = (*D)->get()->asLValue();
1359          const LValue *vd = (*d)->get()->asLValue();
1360
1361          if (!vD->livei.overlaps(vd->livei)) {
1362             INFO_DBG(prog->dbgFlags, REG_ALLOC, "(%%%i) X (%%%i): no overlap\n",
1363                      vD->id, vd->id);
1364             continue;
1365          }
1366
1367          uint8_t mask = vD->compound ? vD->compMask : ~0;
1368          if (vd->compound) {
1369             assert(vB->compound);
1370             mask &= vd->compMask & vB->compMask;
1371          } else {
1372             mask &= intfMask;
1373          }
1374
1375          INFO_DBG(prog->dbgFlags, REG_ALLOC,
1376                   "(%%%i)%02x X (%%%i)%02x & %02x: $r%i.%02x\n",
1377                   vD->id,
1378                   vD->compound ? vD->compMask : 0xff,
1379                   vd->id,
1380                   vd->compound ? vd->compMask : intfMask,
1381                   vB->compMask, intf->reg & ~7, mask);
1382          if (mask)
1383             regs.occupyMask(node->f, intf->reg & ~7, mask);
1384       }
1385       }
1386    } else {
1387       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1388                "(%%%i) X (%%%i): $r%i + %u\n",
1389                vA->id, vB->id, intf->reg, intf->colors);
1390       regs.occupy(node->f, intf->reg, intf->colors);
1391    }
1392 }
1393
1394 bool
1395 GCRA::selectRegisters()
1396 {
1397    INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nSELECT phase\n");
1398
1399    while (!stack.empty()) {
1400       RIG_Node *node = &nodes[stack.top()];
1401       stack.pop();
1402
1403       regs.reset(node->f);
1404
1405       INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nNODE[%%%i, %u colors]\n",
1406                node->getValue()->id, node->colors);
1407
1408       for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1409          checkInterference(node, ei);
1410       for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1411          checkInterference(node, ei);
1412
1413       if (!node->prefRegs.empty()) {
1414          for (std::list<RIG_Node *>::const_iterator it = node->prefRegs.begin();
1415               it != node->prefRegs.end();
1416               ++it) {
1417             if ((*it)->reg >= 0 &&
1418                 regs.testOccupy(node->f, (*it)->reg, node->colors)) {
1419                node->reg = (*it)->reg;
1420                break;
1421             }
1422          }
1423       }
1424       if (node->reg >= 0)
1425          continue;
1426       LValue *lval = node->getValue();
1427       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1428          regs.print(node->f);
1429       bool ret = regs.assign(node->reg, node->f, node->colors);
1430       if (ret) {
1431          INFO_DBG(prog->dbgFlags, REG_ALLOC, "assigned reg %i\n", node->reg);
1432          lval->compMask = node->getCompMask();
1433       } else {
1434          INFO_DBG(prog->dbgFlags, REG_ALLOC, "must spill: %%%i (size %u)\n",
1435                   lval->id, lval->reg.size);
1436          Symbol *slot = NULL;
1437          if (lval->reg.file == FILE_GPR)
1438             slot = spill.assignSlot(node->livei, lval->reg.size);
1439          mustSpill.push_back(ValuePair(lval, slot));
1440       }
1441    }
1442    if (!mustSpill.empty())
1443       return false;
1444    for (unsigned int i = 0; i < nodeCount; ++i) {
1445       LValue *lval = nodes[i].getValue();
1446       if (nodes[i].reg >= 0 && nodes[i].colors > 0)
1447          lval->reg.data.id =
1448             regs.unitsToId(nodes[i].f, nodes[i].reg, lval->reg.size);
1449    }
1450    return true;
1451 }
1452
1453 bool
1454 GCRA::allocateRegisters(ArrayList& insns)
1455 {
1456    bool ret;
1457
1458    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1459             "allocateRegisters to %u instructions\n", insns.getSize());
1460
1461    nodeCount = func->allLValues.getSize();
1462    nodes = new RIG_Node[nodeCount];
1463    if (!nodes)
1464       return false;
1465    for (unsigned int i = 0; i < nodeCount; ++i) {
1466       LValue *lval = reinterpret_cast<LValue *>(func->allLValues.get(i));
1467       if (lval) {
1468          nodes[i].init(regs, lval);
1469          RIG.insert(&nodes[i]);
1470
1471          if (lval->inFile(FILE_GPR) && lval->getInsn() != NULL &&
1472              prog->getTarget()->getChipset() < 0xc0) {
1473             Instruction *insn = lval->getInsn();
1474             if (insn->op == OP_MAD || insn->op == OP_SAD)
1475                // Short encoding only possible if they're all GPRs, no need to
1476                // affect them otherwise.
1477                if (insn->flagsDef < 0 &&
1478                    insn->src(0).getFile() == FILE_GPR &&
1479                    insn->src(1).getFile() == FILE_GPR &&
1480                    insn->src(2).getFile() == FILE_GPR)
1481                   nodes[i].addRegPreference(getNode(insn->getSrc(2)->asLValue()));
1482          }
1483       }
1484    }
1485
1486    // coalesce first, we use only 1 RIG node for a group of joined values
1487    ret = coalesce(insns);
1488    if (!ret)
1489       goto out;
1490
1491    if (func->getProgram()->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1492       func->printLiveIntervals();
1493
1494    buildRIG(insns);
1495    calculateSpillWeights();
1496    simplify();
1497
1498    ret = selectRegisters();
1499    if (!ret) {
1500       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1501                "selectRegisters failed, inserting spill code ...\n");
1502       regs.reset(FILE_GPR, true);
1503       spill.run(mustSpill);
1504       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1505          func->print();
1506    } else {
1507       prog->maxGPR = std::max(prog->maxGPR, regs.getMaxAssigned(FILE_GPR));
1508    }
1509
1510 out:
1511    cleanup(ret);
1512    return ret;
1513 }
1514
1515 void
1516 GCRA::cleanup(const bool success)
1517 {
1518    mustSpill.clear();
1519
1520    for (ArrayList::Iterator it = func->allLValues.iterator();
1521         !it.end(); it.next()) {
1522       LValue *lval =  reinterpret_cast<LValue *>(it.get());
1523
1524       lval->livei.clear();
1525
1526       lval->compound = 0;
1527       lval->compMask = 0;
1528
1529       if (lval->join == lval)
1530          continue;
1531
1532       if (success) {
1533          lval->reg.data.id = lval->join->reg.data.id;
1534       } else {
1535          for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1536               ++d)
1537             lval->join->defs.remove(*d);
1538          lval->join = lval;
1539       }
1540    }
1541
1542    if (success)
1543       resolveSplitsAndMerges();
1544    splits.clear(); // avoid duplicate entries on next coalesce pass
1545    merges.clear();
1546
1547    delete[] nodes;
1548    nodes = NULL;
1549    hi.next = hi.prev = &hi;
1550    lo[0].next = lo[0].prev = &lo[0];
1551    lo[1].next = lo[1].prev = &lo[1];
1552 }
1553
1554 Symbol *
1555 SpillCodeInserter::assignSlot(const Interval &livei, const unsigned int size)
1556 {
1557    SpillSlot slot;
1558    int32_t offsetBase = stackSize;
1559    int32_t offset;
1560    std::list<SpillSlot>::iterator pos = slots.end(), it = slots.begin();
1561
1562    if (offsetBase % size)
1563       offsetBase += size - (offsetBase % size);
1564
1565    slot.sym = NULL;
1566
1567    for (offset = offsetBase; offset < stackSize; offset += size) {
1568       const int32_t entryEnd = offset + size;
1569       while (it != slots.end() && it->offset < offset)
1570          ++it;
1571       if (it == slots.end()) // no slots left
1572          break;
1573       std::list<SpillSlot>::iterator bgn = it;
1574
1575       while (it != slots.end() && it->offset < entryEnd) {
1576          it->occup.print();
1577          if (it->occup.overlaps(livei))
1578             break;
1579          ++it;
1580       }
1581       if (it == slots.end() || it->offset >= entryEnd) {
1582          // fits
1583          for (; bgn != slots.end() && bgn->offset < entryEnd; ++bgn) {
1584             bgn->occup.insert(livei);
1585             if (bgn->size() == size)
1586                slot.sym = bgn->sym;
1587          }
1588          break;
1589       }
1590    }
1591    if (!slot.sym) {
1592       stackSize = offset + size;
1593       slot.offset = offset;
1594       slot.sym = new_Symbol(func->getProgram(), FILE_MEMORY_LOCAL);
1595       if (!func->stackPtr)
1596          offset += func->tlsBase;
1597       slot.sym->setAddress(NULL, offset);
1598       slot.sym->reg.size = size;
1599       slots.insert(pos, slot)->occup.insert(livei);
1600    }
1601    return slot.sym;
1602 }
1603
1604 Value *
1605 SpillCodeInserter::offsetSlot(Value *base, const LValue *lval)
1606 {
1607    if (!lval->compound || (lval->compMask & 0x1))
1608       return base;
1609    Value *slot = cloneShallow(func, base);
1610
1611    slot->reg.data.offset += (ffs(lval->compMask) - 1) * lval->reg.size;
1612    slot->reg.size = lval->reg.size;
1613
1614    return slot;
1615 }
1616
1617 void
1618 SpillCodeInserter::spill(Instruction *defi, Value *slot, LValue *lval)
1619 {
1620    const DataType ty = typeOfSize(lval->reg.size);
1621
1622    slot = offsetSlot(slot, lval);
1623
1624    Instruction *st;
1625    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1626       lval->noSpill = 1;
1627       if (ty != TYPE_B96) {
1628          st = new_Instruction(func, OP_STORE, ty);
1629          st->setSrc(0, slot);
1630          st->setSrc(1, lval);
1631       } else {
1632          st = new_Instruction(func, OP_SPLIT, ty);
1633          st->setSrc(0, lval);
1634          for (int d = 0; d < lval->reg.size / 4; ++d)
1635             st->setDef(d, new_LValue(func, FILE_GPR));
1636
1637          for (int d = lval->reg.size / 4 - 1; d >= 0; --d) {
1638             Value *tmp = cloneShallow(func, slot);
1639             tmp->reg.size = 4;
1640             tmp->reg.data.offset += 4 * d;
1641
1642             Instruction *s = new_Instruction(func, OP_STORE, TYPE_U32);
1643             s->setSrc(0, tmp);
1644             s->setSrc(1, st->getDef(d));
1645             defi->bb->insertAfter(defi, s);
1646          }
1647       }
1648    } else {
1649       st = new_Instruction(func, OP_CVT, ty);
1650       st->setDef(0, slot);
1651       st->setSrc(0, lval);
1652       if (lval->reg.file == FILE_FLAGS)
1653          st->flagsSrc = 0;
1654    }
1655    defi->bb->insertAfter(defi, st);
1656 }
1657
1658 LValue *
1659 SpillCodeInserter::unspill(Instruction *usei, LValue *lval, Value *slot)
1660 {
1661    const DataType ty = typeOfSize(lval->reg.size);
1662
1663    slot = offsetSlot(slot, lval);
1664    lval = cloneShallow(func, lval);
1665
1666    Instruction *ld;
1667    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1668       lval->noSpill = 1;
1669       if (ty != TYPE_B96) {
1670          ld = new_Instruction(func, OP_LOAD, ty);
1671       } else {
1672          ld = new_Instruction(func, OP_MERGE, ty);
1673          for (int d = 0; d < lval->reg.size / 4; ++d) {
1674             Value *tmp = cloneShallow(func, slot);
1675             LValue *val;
1676             tmp->reg.size = 4;
1677             tmp->reg.data.offset += 4 * d;
1678
1679             Instruction *l = new_Instruction(func, OP_LOAD, TYPE_U32);
1680             l->setDef(0, (val = new_LValue(func, FILE_GPR)));
1681             l->setSrc(0, tmp);
1682             usei->bb->insertBefore(usei, l);
1683             ld->setSrc(d, val);
1684             val->noSpill = 1;
1685          }
1686          ld->setDef(0, lval);
1687          usei->bb->insertBefore(usei, ld);
1688          return lval;
1689       }
1690    } else {
1691       ld = new_Instruction(func, OP_CVT, ty);
1692    }
1693    ld->setDef(0, lval);
1694    ld->setSrc(0, slot);
1695    if (lval->reg.file == FILE_FLAGS)
1696       ld->flagsDef = 0;
1697
1698    usei->bb->insertBefore(usei, ld);
1699    return lval;
1700 }
1701
1702 static bool
1703 value_cmp(ValueRef *a, ValueRef *b) {
1704    Instruction *ai = a->getInsn(), *bi = b->getInsn();
1705    if (ai->bb != bi->bb)
1706       return ai->bb->getId() < bi->bb->getId();
1707    return ai->serial < bi->serial;
1708 }
1709
1710 // For each value that is to be spilled, go through all its definitions.
1711 // A value can have multiple definitions if it has been coalesced before.
1712 // For each definition, first go through all its uses and insert an unspill
1713 // instruction before it, then replace the use with the temporary register.
1714 // Unspill can be either a load from memory or simply a move to another
1715 // register file.
1716 // For "Pseudo" instructions (like PHI, SPLIT, MERGE) we can erase the use
1717 // if we have spilled to a memory location, or simply with the new register.
1718 // No load or conversion instruction should be needed.
1719 bool
1720 SpillCodeInserter::run(const std::list<ValuePair>& lst)
1721 {
1722    for (std::list<ValuePair>::const_iterator it = lst.begin(); it != lst.end();
1723         ++it) {
1724       LValue *lval = it->first->asLValue();
1725       Symbol *mem = it->second ? it->second->asSym() : NULL;
1726
1727       // Keep track of which instructions to delete later. Deleting them
1728       // inside the loop is unsafe since a single instruction may have
1729       // multiple destinations that all need to be spilled (like OP_SPLIT).
1730       unordered_set<Instruction *> to_del;
1731
1732       for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1733            ++d) {
1734          Value *slot = mem ?
1735             static_cast<Value *>(mem) : new_LValue(func, FILE_GPR);
1736          Value *tmp = NULL;
1737          Instruction *last = NULL;
1738
1739          LValue *dval = (*d)->get()->asLValue();
1740          Instruction *defi = (*d)->getInsn();
1741
1742          // Sort all the uses by BB/instruction so that we don't unspill
1743          // multiple times in a row, and also remove a source of
1744          // non-determinism.
1745          std::vector<ValueRef *> refs(dval->uses.begin(), dval->uses.end());
1746          std::sort(refs.begin(), refs.end(), value_cmp);
1747
1748          // Unspill at each use *before* inserting spill instructions,
1749          // we don't want to have the spill instructions in the use list here.
1750          for (std::vector<ValueRef*>::const_iterator it = refs.begin();
1751               it != refs.end(); ++it) {
1752             ValueRef *u = *it;
1753             Instruction *usei = u->getInsn();
1754             assert(usei);
1755             if (usei->isPseudo()) {
1756                tmp = (slot->reg.file == FILE_MEMORY_LOCAL) ? NULL : slot;
1757                last = NULL;
1758             } else {
1759                if (!last || (usei != last->next && usei != last))
1760                   tmp = unspill(usei, dval, slot);
1761                last = usei;
1762             }
1763             u->set(tmp);
1764          }
1765
1766          assert(defi);
1767          if (defi->isPseudo()) {
1768             d = lval->defs.erase(d);
1769             --d;
1770             if (slot->reg.file == FILE_MEMORY_LOCAL)
1771                to_del.insert(defi);
1772             else
1773                defi->setDef(0, slot);
1774          } else {
1775             spill(defi, slot, dval);
1776          }
1777       }
1778
1779       for (unordered_set<Instruction *>::const_iterator it = to_del.begin();
1780            it != to_del.end(); ++it)
1781          delete_Instruction(func->getProgram(), *it);
1782    }
1783
1784    // TODO: We're not trying to reuse old slots in a potential next iteration.
1785    //  We have to update the slots' livei intervals to be able to do that.
1786    stackBase = stackSize;
1787    slots.clear();
1788    return true;
1789 }
1790
1791 bool
1792 RegAlloc::exec()
1793 {
1794    for (IteratorRef it = prog->calls.iteratorDFS(false);
1795         !it->end(); it->next()) {
1796       func = Function::get(reinterpret_cast<Graph::Node *>(it->get()));
1797
1798       func->tlsBase = prog->tlsSize;
1799       if (!execFunc())
1800          return false;
1801       prog->tlsSize += func->tlsSize;
1802    }
1803    return true;
1804 }
1805
1806 bool
1807 RegAlloc::execFunc()
1808 {
1809    InsertConstraintsPass insertConstr;
1810    PhiMovesPass insertPhiMoves;
1811    ArgumentMovesPass insertArgMoves;
1812    BuildIntervalsPass buildIntervals;
1813    SpillCodeInserter insertSpills(func);
1814
1815    GCRA gcra(func, insertSpills);
1816
1817    unsigned int i, retries;
1818    bool ret;
1819
1820    if (!func->ins.empty()) {
1821       // Insert a nop at the entry so inputs only used by the first instruction
1822       // don't count as having an empty live range.
1823       Instruction *nop = new_Instruction(func, OP_NOP, TYPE_NONE);
1824       BasicBlock::get(func->cfg.getRoot())->insertHead(nop);
1825    }
1826
1827    ret = insertConstr.exec(func);
1828    if (!ret)
1829       goto out;
1830
1831    ret = insertPhiMoves.run(func);
1832    if (!ret)
1833       goto out;
1834
1835    ret = insertArgMoves.run(func);
1836    if (!ret)
1837       goto out;
1838
1839    // TODO: need to fix up spill slot usage ranges to support > 1 retry
1840    for (retries = 0; retries < 3; ++retries) {
1841       if (retries && (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC))
1842          INFO("Retry: %i\n", retries);
1843       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1844          func->print();
1845
1846       // spilling to registers may add live ranges, need to rebuild everything
1847       ret = true;
1848       for (sequence = func->cfg.nextSequence(), i = 0;
1849            ret && i <= func->loopNestingBound;
1850            sequence = func->cfg.nextSequence(), ++i)
1851          ret = buildLiveSets(BasicBlock::get(func->cfg.getRoot()));
1852       // reset marker
1853       for (ArrayList::Iterator bi = func->allBBlocks.iterator();
1854            !bi.end(); bi.next())
1855          BasicBlock::get(bi)->liveSet.marker = false;
1856       if (!ret)
1857          break;
1858       func->orderInstructions(this->insns);
1859
1860       ret = buildIntervals.run(func);
1861       if (!ret)
1862          break;
1863       ret = gcra.allocateRegisters(insns);
1864       if (ret)
1865          break; // success
1866    }
1867    INFO_DBG(prog->dbgFlags, REG_ALLOC, "RegAlloc done: %i\n", ret);
1868
1869    func->tlsSize = insertSpills.getStackSize();
1870 out:
1871    return ret;
1872 }
1873
1874 // TODO: check if modifying Instruction::join here breaks anything
1875 void
1876 GCRA::resolveSplitsAndMerges()
1877 {
1878    for (std::list<Instruction *>::iterator it = splits.begin();
1879         it != splits.end();
1880         ++it) {
1881       Instruction *split = *it;
1882       unsigned int reg = regs.idToBytes(split->getSrc(0));
1883       for (int d = 0; split->defExists(d); ++d) {
1884          Value *v = split->getDef(d);
1885          v->reg.data.id = regs.bytesToId(v, reg);
1886          v->join = v;
1887          reg += v->reg.size;
1888       }
1889    }
1890    splits.clear();
1891
1892    for (std::list<Instruction *>::iterator it = merges.begin();
1893         it != merges.end();
1894         ++it) {
1895       Instruction *merge = *it;
1896       unsigned int reg = regs.idToBytes(merge->getDef(0));
1897       for (int s = 0; merge->srcExists(s); ++s) {
1898          Value *v = merge->getSrc(s);
1899          v->reg.data.id = regs.bytesToId(v, reg);
1900          v->join = v;
1901          // If the value is defined by a phi/union node, we also need to
1902          // perform the same fixup on that node's sources, since after RA
1903          // their registers should be identical.
1904          if (v->getInsn()->op == OP_PHI || v->getInsn()->op == OP_UNION) {
1905             Instruction *phi = v->getInsn();
1906             for (int phis = 0; phi->srcExists(phis); ++phis)
1907                phi->getSrc(phis)->join = v;
1908          }
1909          reg += v->reg.size;
1910       }
1911    }
1912    merges.clear();
1913 }
1914
1915 bool Program::registerAllocation()
1916 {
1917    RegAlloc ra(this);
1918    return ra.exec();
1919 }
1920
1921 bool
1922 RegAlloc::InsertConstraintsPass::exec(Function *ir)
1923 {
1924    constrList.clear();
1925
1926    bool ret = run(ir, true, true);
1927    if (ret)
1928       ret = insertConstraintMoves();
1929    return ret;
1930 }
1931
1932 // TODO: make part of texture insn
1933 void
1934 RegAlloc::InsertConstraintsPass::textureMask(TexInstruction *tex)
1935 {
1936    Value *def[4];
1937    int c, k, d;
1938    uint8_t mask = 0;
1939
1940    for (d = 0, k = 0, c = 0; c < 4; ++c) {
1941       if (!(tex->tex.mask & (1 << c)))
1942          continue;
1943       if (tex->getDef(k)->refCount()) {
1944          mask |= 1 << c;
1945          def[d++] = tex->getDef(k);
1946       }
1947       ++k;
1948    }
1949    tex->tex.mask = mask;
1950
1951    for (c = 0; c < d; ++c)
1952       tex->setDef(c, def[c]);
1953    for (; c < 4; ++c)
1954       tex->setDef(c, NULL);
1955 }
1956
1957 bool
1958 RegAlloc::InsertConstraintsPass::detectConflict(Instruction *cst, int s)
1959 {
1960    Value *v = cst->getSrc(s);
1961
1962    // current register allocation can't handle it if a value participates in
1963    // multiple constraints
1964    for (Value::UseIterator it = v->uses.begin(); it != v->uses.end(); ++it) {
1965       if (cst != (*it)->getInsn())
1966          return true;
1967    }
1968
1969    // can start at s + 1 because detectConflict is called on all sources
1970    for (int c = s + 1; cst->srcExists(c); ++c)
1971       if (v == cst->getSrc(c))
1972          return true;
1973
1974    Instruction *defi = v->getInsn();
1975
1976    return (!defi || defi->constrainedDefs());
1977 }
1978
1979 void
1980 RegAlloc::InsertConstraintsPass::addConstraint(Instruction *i, int s, int n)
1981 {
1982    Instruction *cst;
1983    int d;
1984
1985    // first, look for an existing identical constraint op
1986    for (std::list<Instruction *>::iterator it = constrList.begin();
1987         it != constrList.end();
1988         ++it) {
1989       cst = (*it);
1990       if (!i->bb->dominatedBy(cst->bb))
1991          break;
1992       for (d = 0; d < n; ++d)
1993          if (cst->getSrc(d) != i->getSrc(d + s))
1994             break;
1995       if (d >= n) {
1996          for (d = 0; d < n; ++d, ++s)
1997             i->setSrc(s, cst->getDef(d));
1998          return;
1999       }
2000    }
2001    cst = new_Instruction(func, OP_CONSTRAINT, i->dType);
2002
2003    for (d = 0; d < n; ++s, ++d) {
2004       cst->setDef(d, new_LValue(func, FILE_GPR));
2005       cst->setSrc(d, i->getSrc(s));
2006       i->setSrc(s, cst->getDef(d));
2007    }
2008    i->bb->insertBefore(i, cst);
2009
2010    constrList.push_back(cst);
2011 }
2012
2013 // Add a dummy use of the pointer source of >= 8 byte loads after the load
2014 // to prevent it from being assigned a register which overlapping the load's
2015 // destination, which would produce random corruptions.
2016 void
2017 RegAlloc::InsertConstraintsPass::addHazard(Instruction *i, const ValueRef *src)
2018 {
2019    Instruction *hzd = new_Instruction(func, OP_NOP, TYPE_NONE);
2020    hzd->setSrc(0, src->get());
2021    i->bb->insertAfter(i, hzd);
2022
2023 }
2024
2025 // b32 { %r0 %r1 %r2 %r3 } -> b128 %r0q
2026 void
2027 RegAlloc::InsertConstraintsPass::condenseDefs(Instruction *insn)
2028 {
2029    uint8_t size = 0;
2030    int n;
2031    for (n = 0; insn->defExists(n) && insn->def(n).getFile() == FILE_GPR; ++n)
2032       size += insn->getDef(n)->reg.size;
2033    if (n < 2)
2034       return;
2035    LValue *lval = new_LValue(func, FILE_GPR);
2036    lval->reg.size = size;
2037
2038    Instruction *split = new_Instruction(func, OP_SPLIT, typeOfSize(size));
2039    split->setSrc(0, lval);
2040    for (int d = 0; d < n; ++d) {
2041       split->setDef(d, insn->getDef(d));
2042       insn->setDef(d, NULL);
2043    }
2044    insn->setDef(0, lval);
2045
2046    for (int k = 1, d = n; insn->defExists(d); ++d, ++k) {
2047       insn->setDef(k, insn->getDef(d));
2048       insn->setDef(d, NULL);
2049    }
2050    // carry over predicate if any (mainly for OP_UNION uses)
2051    split->setPredicate(insn->cc, insn->getPredicate());
2052
2053    insn->bb->insertAfter(insn, split);
2054    constrList.push_back(split);
2055 }
2056 void
2057 RegAlloc::InsertConstraintsPass::condenseSrcs(Instruction *insn,
2058                                               const int a, const int b)
2059 {
2060    uint8_t size = 0;
2061    if (a >= b)
2062       return;
2063    for (int s = a; s <= b; ++s)
2064       size += insn->getSrc(s)->reg.size;
2065    if (!size)
2066       return;
2067    LValue *lval = new_LValue(func, FILE_GPR);
2068    lval->reg.size = size;
2069
2070    Value *save[3];
2071    insn->takeExtraSources(0, save);
2072
2073    Instruction *merge = new_Instruction(func, OP_MERGE, typeOfSize(size));
2074    merge->setDef(0, lval);
2075    for (int s = a, i = 0; s <= b; ++s, ++i) {
2076       merge->setSrc(i, insn->getSrc(s));
2077    }
2078    insn->moveSources(b + 1, a - b);
2079    insn->setSrc(a, lval);
2080    insn->bb->insertBefore(insn, merge);
2081
2082    insn->putExtraSources(0, save);
2083
2084    constrList.push_back(merge);
2085 }
2086
2087 void
2088 RegAlloc::InsertConstraintsPass::texConstraintGM107(TexInstruction *tex)
2089 {
2090    int n, s;
2091
2092    if (isTextureOp(tex->op))
2093       textureMask(tex);
2094    condenseDefs(tex);
2095
2096    if (isSurfaceOp(tex->op)) {
2097       int s = tex->tex.target.getDim() +
2098          (tex->tex.target.isArray() || tex->tex.target.isCube());
2099       int n = 0;
2100
2101       switch (tex->op) {
2102       case OP_SUSTB:
2103       case OP_SUSTP:
2104          n = 4;
2105          break;
2106       case OP_SUREDB:
2107       case OP_SUREDP:
2108          if (tex->subOp == NV50_IR_SUBOP_ATOM_CAS)
2109             n = 2;
2110          break;
2111       default:
2112          break;
2113       }
2114
2115       if (s > 1)
2116          condenseSrcs(tex, 0, s - 1);
2117       if (n > 1)
2118          condenseSrcs(tex, 1, n); // do not condense the tex handle
2119    } else
2120    if (isTextureOp(tex->op)) {
2121       if (tex->op != OP_TXQ) {
2122          s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
2123          if (tex->op == OP_TXD) {
2124             // Indirect handle belongs in the first arg
2125             if (tex->tex.rIndirectSrc >= 0)
2126                s++;
2127             if (!tex->tex.target.isArray() && tex->tex.useOffsets)
2128                s++;
2129          }
2130          n = tex->srcCount(0xff) - s;
2131       } else {
2132          s = tex->srcCount(0xff);
2133          n = 0;
2134       }
2135
2136       if (s > 1)
2137          condenseSrcs(tex, 0, s - 1);
2138       if (n > 1) // NOTE: first call modified positions already
2139          condenseSrcs(tex, 1, n);
2140    }
2141 }
2142
2143 void
2144 RegAlloc::InsertConstraintsPass::texConstraintNVE0(TexInstruction *tex)
2145 {
2146    if (isTextureOp(tex->op))
2147       textureMask(tex);
2148    condenseDefs(tex);
2149
2150    if (tex->op == OP_SUSTB || tex->op == OP_SUSTP) {
2151       condenseSrcs(tex, 3, 6);
2152    } else
2153    if (isTextureOp(tex->op)) {
2154       int n = tex->srcCount(0xff, true);
2155       if (n > 4) {
2156          condenseSrcs(tex, 0, 3);
2157          if (n > 5) // NOTE: first call modified positions already
2158             condenseSrcs(tex, 4 - (4 - 1), n - 1 - (4 - 1));
2159       } else
2160       if (n > 1) {
2161          condenseSrcs(tex, 0, n - 1);
2162       }
2163    }
2164 }
2165
2166 void
2167 RegAlloc::InsertConstraintsPass::texConstraintNVC0(TexInstruction *tex)
2168 {
2169    int n, s;
2170
2171    if (isTextureOp(tex->op))
2172       textureMask(tex);
2173
2174    if (tex->op == OP_TXQ) {
2175       s = tex->srcCount(0xff);
2176       n = 0;
2177    } else if (isSurfaceOp(tex->op)) {
2178       s = tex->tex.target.getDim() + (tex->tex.target.isArray() || tex->tex.target.isCube());
2179       if (tex->op == OP_SUSTB || tex->op == OP_SUSTP)
2180          n = 4;
2181       else
2182          n = 0;
2183    } else {
2184       s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
2185       if (!tex->tex.target.isArray() &&
2186           (tex->tex.rIndirectSrc >= 0 || tex->tex.sIndirectSrc >= 0))
2187          ++s;
2188       if (tex->op == OP_TXD && tex->tex.useOffsets)
2189          ++s;
2190       n = tex->srcCount(0xff) - s;
2191       assert(n <= 4);
2192    }
2193
2194    if (s > 1)
2195       condenseSrcs(tex, 0, s - 1);
2196    if (n > 1) // NOTE: first call modified positions already
2197       condenseSrcs(tex, 1, n);
2198
2199    condenseDefs(tex);
2200 }
2201
2202 void
2203 RegAlloc::InsertConstraintsPass::texConstraintNV50(TexInstruction *tex)
2204 {
2205    Value *pred = tex->getPredicate();
2206    if (pred)
2207       tex->setPredicate(tex->cc, NULL);
2208
2209    textureMask(tex);
2210
2211    assert(tex->defExists(0) && tex->srcExists(0));
2212    // make src and def count match
2213    int c;
2214    for (c = 0; tex->srcExists(c) || tex->defExists(c); ++c) {
2215       if (!tex->srcExists(c))
2216          tex->setSrc(c, new_LValue(func, tex->getSrc(0)->asLValue()));
2217       if (!tex->defExists(c))
2218          tex->setDef(c, new_LValue(func, tex->getDef(0)->asLValue()));
2219    }
2220    if (pred)
2221       tex->setPredicate(tex->cc, pred);
2222    condenseDefs(tex);
2223    condenseSrcs(tex, 0, c - 1);
2224 }
2225
2226 // Insert constraint markers for instructions whose multiple sources must be
2227 // located in consecutive registers.
2228 bool
2229 RegAlloc::InsertConstraintsPass::visit(BasicBlock *bb)
2230 {
2231    TexInstruction *tex;
2232    Instruction *next;
2233    int s, size;
2234
2235    targ = bb->getProgram()->getTarget();
2236
2237    for (Instruction *i = bb->getEntry(); i; i = next) {
2238       next = i->next;
2239
2240       if ((tex = i->asTex())) {
2241          switch (targ->getChipset() & ~0xf) {
2242          case 0x50:
2243          case 0x80:
2244          case 0x90:
2245          case 0xa0:
2246             texConstraintNV50(tex);
2247             break;
2248          case 0xc0:
2249          case 0xd0:
2250             texConstraintNVC0(tex);
2251             break;
2252          case 0xe0:
2253          case 0xf0:
2254          case 0x100:
2255             texConstraintNVE0(tex);
2256             break;
2257          case 0x110:
2258          case 0x120:
2259          case 0x130:
2260             texConstraintGM107(tex);
2261             break;
2262          default:
2263             break;
2264          }
2265       } else
2266       if (i->op == OP_EXPORT || i->op == OP_STORE) {
2267          for (size = typeSizeof(i->dType), s = 1; size > 0; ++s) {
2268             assert(i->srcExists(s));
2269             size -= i->getSrc(s)->reg.size;
2270          }
2271          condenseSrcs(i, 1, s - 1);
2272       } else
2273       if (i->op == OP_LOAD || i->op == OP_VFETCH) {
2274          condenseDefs(i);
2275          if (i->src(0).isIndirect(0) && typeSizeof(i->dType) >= 8)
2276             addHazard(i, i->src(0).getIndirect(0));
2277          if (i->src(0).isIndirect(1) && typeSizeof(i->dType) >= 8)
2278             addHazard(i, i->src(0).getIndirect(1));
2279       } else
2280       if (i->op == OP_UNION ||
2281           i->op == OP_MERGE ||
2282           i->op == OP_SPLIT) {
2283          constrList.push_back(i);
2284       }
2285    }
2286    return true;
2287 }
2288
2289 // Insert extra moves so that, if multiple register constraints on a value are
2290 // in conflict, these conflicts can be resolved.
2291 bool
2292 RegAlloc::InsertConstraintsPass::insertConstraintMoves()
2293 {
2294    for (std::list<Instruction *>::iterator it = constrList.begin();
2295         it != constrList.end();
2296         ++it) {
2297       Instruction *cst = *it;
2298       Instruction *mov;
2299
2300       if (cst->op == OP_SPLIT && 0) {
2301          // spilling splits is annoying, just make sure they're separate
2302          for (int d = 0; cst->defExists(d); ++d) {
2303             if (!cst->getDef(d)->refCount())
2304                continue;
2305             LValue *lval = new_LValue(func, cst->def(d).getFile());
2306             const uint8_t size = cst->def(d).getSize();
2307             lval->reg.size = size;
2308
2309             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
2310             mov->setSrc(0, lval);
2311             mov->setDef(0, cst->getDef(d));
2312             cst->setDef(d, mov->getSrc(0));
2313             cst->bb->insertAfter(cst, mov);
2314
2315             cst->getSrc(0)->asLValue()->noSpill = 1;
2316             mov->getSrc(0)->asLValue()->noSpill = 1;
2317          }
2318       } else
2319       if (cst->op == OP_MERGE || cst->op == OP_UNION) {
2320          for (int s = 0; cst->srcExists(s); ++s) {
2321             const uint8_t size = cst->src(s).getSize();
2322
2323             if (!cst->getSrc(s)->defs.size()) {
2324                mov = new_Instruction(func, OP_NOP, typeOfSize(size));
2325                mov->setDef(0, cst->getSrc(s));
2326                cst->bb->insertBefore(cst, mov);
2327                continue;
2328             }
2329             assert(cst->getSrc(s)->defs.size() == 1); // still SSA
2330
2331             Instruction *defi = cst->getSrc(s)->defs.front()->getInsn();
2332             // catch some cases where don't really need MOVs
2333             if (cst->getSrc(s)->refCount() == 1 && !defi->constrainedDefs())
2334                continue;
2335
2336             LValue *lval = new_LValue(func, cst->src(s).getFile());
2337             lval->reg.size = size;
2338
2339             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
2340             mov->setDef(0, lval);
2341             mov->setSrc(0, cst->getSrc(s));
2342             cst->setSrc(s, mov->getDef(0));
2343             cst->bb->insertBefore(cst, mov);
2344
2345             cst->getDef(0)->asLValue()->noSpill = 1; // doesn't help
2346
2347             if (cst->op == OP_UNION)
2348                mov->setPredicate(defi->cc, defi->getPredicate());
2349          }
2350       }
2351    }
2352
2353    return true;
2354 }
2355
2356 } // namespace nv50_ir