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 #else
  41 using std::tr1::hash;
  42 using std::tr1::unordered_map;
  43 #endif
  44
  45 #define MAX_REGISTER_FILE_SIZE 256
  46
  47 class RegisterSet
  48 {
  49 public:
  50    RegisterSet(const Target *);
  51
  52    void init(const Target *);
  53    void reset(DataFile, bool resetMax = false);
  54
  55    void periodicMask(DataFile f, uint32_t lock, uint32_t unlock);
  56    void intersect(DataFile f, const RegisterSet *);
  57
  58    bool assign(int32_t& reg, DataFile f, unsigned int size, unsigned int maxReg);
  59    void release(DataFile f, int32_t reg, unsigned int size);
  60    void occupy(DataFile f, int32_t reg, unsigned int size);
  61    void occupy(const Value *);
  62    void occupyMask(DataFile f, int32_t reg, uint8_t mask);
  63    bool isOccupied(DataFile f, int32_t reg, unsigned int size) const;
  64    bool testOccupy(const Value *);
  65    bool testOccupy(DataFile f, int32_t reg, unsigned int size);
  66
  67    inline int getMaxAssigned(DataFile f) const { return fill[f]; }
  68
  69    inline unsigned int getFileSize(DataFile f) const
  70    {
  71       return last[f] + 1;
  72    }
  73
  74    inline unsigned int units(DataFile f, unsigned int size) const
  75    {
  76       return size >> unit[f];
  77    }
  78    // for regs of size >= 4, id is counted in 4-byte words (like nv50/c0 binary)
  79    inline unsigned int idToBytes(const Value *v) const
  80    {
  81       return v->reg.data.id * MIN2(v->reg.size, 4);
  82    }
  83    inline unsigned int idToUnits(const Value *v) const
  84    {
  85       return units(v->reg.file, idToBytes(v));
  86    }
  87    inline int bytesToId(Value *v, unsigned int bytes) const
  88    {
  89       if (v->reg.size < 4)
  90          return units(v->reg.file, bytes);
  91       return bytes / 4;
  92    }
  93    inline int unitsToId(DataFile f, int u, uint8_t size) const
  94    {
  95       if (u < 0)
  96          return -1;
  97       return (size < 4) ? u : ((u << unit[f]) / 4);
  98    }
  99
 100    void print(DataFile f) const;
 101
 102    const bool restrictedGPR16Range;
 103
 104 private:
 105    BitSet bits[LAST_REGISTER_FILE + 1];
 106
 107    int unit[LAST_REGISTER_FILE + 1]; // log2 of allocation granularity
 108
 109    int last[LAST_REGISTER_FILE + 1];
 110    int fill[LAST_REGISTER_FILE + 1];
 111 };
 112
 113 void
 114 RegisterSet::reset(DataFile f, bool resetMax)
 115 {
 116    bits[f].fill(0);
 117    if (resetMax)
 118       fill[f] = -1;
 119 }
 120
 121 void
 122 RegisterSet::init(const Target *targ)
 123 {
 124    for (unsigned int rf = 0; rf <= FILE_ADDRESS; ++rf) {
 125       DataFile f = static_cast<DataFile>(rf);
 126       last[rf] = targ->getFileSize(f) - 1;
 127       unit[rf] = targ->getFileUnit(f);
 128       fill[rf] = -1;
 129       assert(last[rf] < MAX_REGISTER_FILE_SIZE);
 130       bits[rf].allocate(last[rf] + 1, true);
 131    }
 132 }
 133
 134 RegisterSet::RegisterSet(const Target *targ)
 135   : restrictedGPR16Range(targ->getChipset() < 0xc0)
 136 {
 137    init(targ);
 138    for (unsigned int i = 0; i <= LAST_REGISTER_FILE; ++i)
 139       reset(static_cast<DataFile>(i));
 140 }
 141
 142 void
 143 RegisterSet::periodicMask(DataFile f, uint32_t lock, uint32_t unlock)
 144 {
 145    bits[f].periodicMask32(lock, unlock);
 146 }
 147
 148 void
 149 RegisterSet::intersect(DataFile f, const RegisterSet *set)
 150 {
 151    bits[f] |= set->bits[f];
 152 }
 153
 154 void
 155 RegisterSet::print(DataFile f) const
 156 {
 157    INFO("GPR:");
 158    bits[f].print();
 159    INFO("\n");
 160 }
 161
 162 bool
 163 RegisterSet::assign(int32_t& reg, DataFile f, unsigned int size, unsigned int maxReg)
 164 {
 165    reg = bits[f].findFreeRange(size, maxReg);
 166    if (reg < 0)
 167       return false;
 168    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
 169    return true;
 170 }
 171
 172 bool
 173 RegisterSet::isOccupied(DataFile f, int32_t reg, unsigned int size) const
 174 {
 175    return bits[f].testRange(reg, size);
 176 }
 177
 178 void
 179 RegisterSet::occupy(const Value *v)
 180 {
 181    occupy(v->reg.file, idToUnits(v), v->reg.size >> unit[v->reg.file]);
 182 }
 183
 184 void
 185 RegisterSet::occupyMask(DataFile f, int32_t reg, uint8_t mask)
 186 {
 187    bits[f].setMask(reg & ~31, static_cast<uint32_t>(mask) << (reg % 32));
 188 }
 189
 190 void
 191 RegisterSet::occupy(DataFile f, int32_t reg, unsigned int size)
 192 {
 193    bits[f].setRange(reg, size);
 194
 195    INFO_DBG(0, REG_ALLOC, "reg occupy: %u[%i] %u\n", f, reg, size);
 196
 197    fill[f] = MAX2(fill[f], (int32_t)(reg + size - 1));
 198 }
 199
 200 bool
 201 RegisterSet::testOccupy(const Value *v)
 202 {
 203    return testOccupy(v->reg.file,
 204                      idToUnits(v), v->reg.size >> unit[v->reg.file]);
 205 }
 206
 207 bool
 208 RegisterSet::testOccupy(DataFile f, int32_t reg, unsigned int size)
 209 {
 210    if (isOccupied(f, reg, size))
 211       return false;
 212    occupy(f, reg, size);
 213    return true;
 214 }
 215
 216 void
 217 RegisterSet::release(DataFile f, int32_t reg, unsigned int size)
 218 {
 219    bits[f].clrRange(reg, size);
 220
 221    INFO_DBG(0, REG_ALLOC, "reg release: %u[%i] %u\n", f, reg, size);
 222 }
 223
 224 class RegAlloc
 225 {
 226 public:
 227    RegAlloc(Program *program) : prog(program), sequence(0) { }
 228
 229    bool exec();
 230    bool execFunc();
 231
 232 private:
 233    class PhiMovesPass : public Pass {
 234    private:
 235       virtual bool visit(BasicBlock *);
 236       inline bool needNewElseBlock(BasicBlock *b, BasicBlock *p);
 237       inline void splitEdges(BasicBlock *b);
 238    };
 239
 240    class ArgumentMovesPass : public Pass {
 241    private:
 242       virtual bool visit(BasicBlock *);
 243    };
 244
 245    class BuildIntervalsPass : public Pass {
 246    private:
 247       virtual bool visit(BasicBlock *);
 248       void collectLiveValues(BasicBlock *);
 249       void addLiveRange(Value *, const BasicBlock *, int end);
 250    };
 251
 252    class InsertConstraintsPass : public Pass {
 253    public:
 254       bool exec(Function *func);
 255    private:
 256       virtual bool visit(BasicBlock *);
 257
 258       void insertConstraintMove(Instruction *, int s);
 259       bool insertConstraintMoves();
 260
 261       void condenseDefs(Instruction *);
 262       void condenseDefs(Instruction *, const int first, const int last);
 263       void condenseSrcs(Instruction *, const int first, const int last);
 264
 265       void addHazard(Instruction *i, const ValueRef *src);
 266       void textureMask(TexInstruction *);
 267       void addConstraint(Instruction *, int s, int n);
 268       bool detectConflict(Instruction *, int s);
 269
 270       // target specific functions, TODO: put in subclass or Target
 271       void texConstraintNV50(TexInstruction *);
 272       void texConstraintNVC0(TexInstruction *);
 273       void texConstraintNVE0(TexInstruction *);
 274       void texConstraintGM107(TexInstruction *);
 275
 276       bool isScalarTexGM107(TexInstruction *);
 277       void handleScalarTexGM107(TexInstruction *);
 278
 279       std::list<Instruction *> constrList;
 280
 281       const Target *targ;
 282    };
 283
 284    bool buildLiveSets(BasicBlock *);
 285
 286 private:
 287    Program *prog;
 288    Function *func;
 289
 290    // instructions in control flow / chronological order
 291    ArrayList insns;
 292
 293    int sequence; // for manual passes through CFG
 294 };
 295
 296 typedef std::pair<Value *, Value *> ValuePair;
 297
 298 class SpillCodeInserter
 299 {
 300 public:
 301    SpillCodeInserter(Function *fn) : func(fn), stackSize(0), stackBase(0) { }
 302
 303    bool run(const std::list<ValuePair>&);
 304
 305    Symbol *assignSlot(const Interval&, const unsigned int size);
 306    Value *offsetSlot(Value *, const LValue *);
 307    inline int32_t getStackSize() const { return stackSize; }
 308
 309 private:
 310    Function *func;
 311
 312    struct SpillSlot
 313    {
 314       Interval occup;
 315       std::list<Value *> residents; // needed to recalculate occup
 316       Symbol *sym;
 317       int32_t offset;
 318       inline uint8_t size() const { return sym->reg.size; }
 319    };
 320    std::list<SpillSlot> slots;
 321    int32_t stackSize;
 322    int32_t stackBase;
 323
 324    LValue *unspill(Instruction *usei, LValue *, Value *slot);
 325    void spill(Instruction *defi, Value *slot, LValue *);
 326 };
 327
 328 void
 329 RegAlloc::BuildIntervalsPass::addLiveRange(Value *val,
 330                                            const BasicBlock *bb,
 331                                            int end)
 332 {
 333    Instruction *insn = val->getUniqueInsn();
 334
 335    if (!insn)
 336       insn = bb->getFirst();
 337
 338    assert(bb->getFirst()->serial <= bb->getExit()->serial);
 339    assert(bb->getExit()->serial + 1 >= end);
 340
 341    int begin = insn->serial;
 342    if (begin < bb->getEntry()->serial || begin > bb->getExit()->serial)
 343       begin = bb->getEntry()->serial;
 344
 345    INFO_DBG(prog->dbgFlags, REG_ALLOC, "%%%i <- live range [%i(%i), %i)\n",
 346             val->id, begin, insn->serial, end);
 347
 348    if (begin != end) // empty ranges are only added as hazards for fixed regs
 349       val->livei.extend(begin, end);
 350 }
 351
 352 bool
 353 RegAlloc::PhiMovesPass::needNewElseBlock(BasicBlock *b, BasicBlock *p)
 354 {
 355    if (b->cfg.incidentCount() <= 1)
 356       return false;
 357
 358    int n = 0;
 359    for (Graph::EdgeIterator ei = p->cfg.outgoing(); !ei.end(); ei.next())
 360       if (ei.getType() == Graph::Edge::TREE ||
 361           ei.getType() == Graph::Edge::FORWARD)
 362          ++n;
 363    return (n == 2);
 364 }
 365
 366 struct PhiMapHash {
 367    size_t operator()(const std::pair<Instruction *, BasicBlock *>& val) const {
 368       return hash<Instruction*>()(val.first) * 31 +
 369          hash<BasicBlock*>()(val.second);
 370    }
 371 };
 372
 373 typedef unordered_map<
 374    std::pair<Instruction *, BasicBlock *>, Value *, PhiMapHash> PhiMap;
 375
 376 // Critical edges need to be split up so that work can be inserted along
 377 // specific edge transitions. Unfortunately manipulating incident edges into a
 378 // BB invalidates all the PHI nodes since their sources are implicitly ordered
 379 // by incident edge order.
 380 //
 381 // TODO: Make it so that that is not the case, and PHI nodes store pointers to
 382 // the original BBs.
 383 void
 384 RegAlloc::PhiMovesPass::splitEdges(BasicBlock *bb)
 385 {
 386    BasicBlock *pb, *pn;
 387    Instruction *phi;
 388    Graph::EdgeIterator ei;
 389    std::stack<BasicBlock *> stack;
 390    int j = 0;
 391
 392    for (ei = bb->cfg.incident(); !ei.end(); ei.next()) {
 393       pb = BasicBlock::get(ei.getNode());
 394       assert(pb);
 395       if (needNewElseBlock(bb, pb))
 396          stack.push(pb);
 397    }
 398
 399    // No critical edges were found, no need to perform any work.
 400    if (stack.empty())
 401       return;
 402
 403    // We're about to, potentially, reorder the inbound edges. This means that
 404    // we need to hold on to the (phi, bb) -> src mapping, and fix up the phi
 405    // nodes after the graph has been modified.
 406    PhiMap phis;
 407
 408    j = 0;
 409    for (ei = bb->cfg.incident(); !ei.end(); ei.next(), j++) {
 410       pb = BasicBlock::get(ei.getNode());
 411       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next)
 412          phis.insert(std::make_pair(std::make_pair(phi, pb), phi->getSrc(j)));
 413    }
 414
 415    while (!stack.empty()) {
 416       pb = stack.top();
 417       pn = new BasicBlock(func);
 418       stack.pop();
 419
 420       pb->cfg.detach(&bb->cfg);
 421       pb->cfg.attach(&pn->cfg, Graph::Edge::TREE);
 422       pn->cfg.attach(&bb->cfg, Graph::Edge::FORWARD);
 423
 424       assert(pb->getExit()->op != OP_CALL);
 425       if (pb->getExit()->asFlow()->target.bb == bb)
 426          pb->getExit()->asFlow()->target.bb = pn;
 427
 428       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 429          PhiMap::iterator it = phis.find(std::make_pair(phi, pb));
 430          assert(it != phis.end());
 431          phis.insert(std::make_pair(std::make_pair(phi, pn), it->second));
 432          phis.erase(it);
 433       }
 434    }
 435
 436    // Now go through and fix up all of the phi node sources.
 437    j = 0;
 438    for (ei = bb->cfg.incident(); !ei.end(); ei.next(), j++) {
 439       pb = BasicBlock::get(ei.getNode());
 440       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 441          PhiMap::const_iterator it = phis.find(std::make_pair(phi, pb));
 442          assert(it != phis.end());
 443
 444          phi->setSrc(j, it->second);
 445       }
 446    }
 447 }
 448
 449 // For each operand of each PHI in b, generate a new value by inserting a MOV
 450 // at the end of the block it is coming from and replace the operand with its
 451 // result. This eliminates liveness conflicts and enables us to let values be
 452 // copied to the right register if such a conflict exists nonetheless.
 453 //
 454 // These MOVs are also crucial in making sure the live intervals of phi srces
 455 // are extended until the end of the loop, since they are not included in the
 456 // live-in sets.
 457 bool
 458 RegAlloc::PhiMovesPass::visit(BasicBlock *bb)
 459 {
 460    Instruction *phi, *mov;
 461
 462    splitEdges(bb);
 463
 464    // insert MOVs (phi->src(j) should stem from j-th in-BB)
 465    int j = 0;
 466    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
 467       BasicBlock *pb = BasicBlock::get(ei.getNode());
 468       if (!pb->isTerminated())
 469          pb->insertTail(new_FlowInstruction(func, OP_BRA, bb));
 470
 471       for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) {
 472          LValue *tmp = new_LValue(func, phi->getDef(0)->asLValue());
 473          mov = new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 474
 475          mov->setSrc(0, phi->getSrc(j));
 476          mov->setDef(0, tmp);
 477          phi->setSrc(j, tmp);
 478
 479          pb->insertBefore(pb->getExit(), mov);
 480       }
 481       ++j;
 482    }
 483
 484    return true;
 485 }
 486
 487 bool
 488 RegAlloc::ArgumentMovesPass::visit(BasicBlock *bb)
 489 {
 490    // Bind function call inputs/outputs to the same physical register
 491    // the callee uses, inserting moves as appropriate for the case a
 492    // conflict arises.
 493    for (Instruction *i = bb->getEntry(); i; i = i->next) {
 494       FlowInstruction *cal = i->asFlow();
 495       // TODO: Handle indirect calls.
 496       // Right now they should only be generated for builtins.
 497       if (!cal || cal->op != OP_CALL || cal->builtin || cal->indirect)
 498          continue;
 499       RegisterSet clobberSet(prog->getTarget());
 500
 501       // Bind input values.
 502       for (int s = cal->indirect ? 1 : 0; cal->srcExists(s); ++s) {
 503          const int t = cal->indirect ? (s - 1) : s;
 504          LValue *tmp = new_LValue(func, cal->getSrc(s)->asLValue());
 505          tmp->reg.data.id = cal->target.fn->ins[t].rep()->reg.data.id;
 506
 507          Instruction *mov =
 508             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 509          mov->setDef(0, tmp);
 510          mov->setSrc(0, cal->getSrc(s));
 511          cal->setSrc(s, tmp);
 512
 513          bb->insertBefore(cal, mov);
 514       }
 515
 516       // Bind output values.
 517       for (int d = 0; cal->defExists(d); ++d) {
 518          LValue *tmp = new_LValue(func, cal->getDef(d)->asLValue());
 519          tmp->reg.data.id = cal->target.fn->outs[d].rep()->reg.data.id;
 520
 521          Instruction *mov =
 522             new_Instruction(func, OP_MOV, typeOfSize(tmp->reg.size));
 523          mov->setSrc(0, tmp);
 524          mov->setDef(0, cal->getDef(d));
 525          cal->setDef(d, tmp);
 526
 527          bb->insertAfter(cal, mov);
 528          clobberSet.occupy(tmp);
 529       }
 530
 531       // Bind clobbered values.
 532       for (std::deque<Value *>::iterator it = cal->target.fn->clobbers.begin();
 533            it != cal->target.fn->clobbers.end();
 534            ++it) {
 535          if (clobberSet.testOccupy(*it)) {
 536             Value *tmp = new_LValue(func, (*it)->asLValue());
 537             tmp->reg.data.id = (*it)->reg.data.id;
 538             cal->setDef(cal->defCount(), tmp);
 539          }
 540       }
 541    }
 542
 543    // Update the clobber set of the function.
 544    if (BasicBlock::get(func->cfgExit) == bb) {
 545       func->buildDefSets();
 546       for (unsigned int i = 0; i < bb->defSet.getSize(); ++i)
 547          if (bb->defSet.test(i))
 548             func->clobbers.push_back(func->getLValue(i));
 549    }
 550
 551    return true;
 552 }
 553
 554 // Build the set of live-in variables of bb.
 555 bool
 556 RegAlloc::buildLiveSets(BasicBlock *bb)
 557 {
 558    Function *f = bb->getFunction();
 559    BasicBlock *bn;
 560    Instruction *i;
 561    unsigned int s, d;
 562
 563    INFO_DBG(prog->dbgFlags, REG_ALLOC, "buildLiveSets(BB:%i)\n", bb->getId());
 564
 565    bb->liveSet.allocate(func->allLValues.getSize(), false);
 566
 567    int n = 0;
 568    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 569       bn = BasicBlock::get(ei.getNode());
 570       if (bn == bb)
 571          continue;
 572       if (bn->cfg.visit(sequence))
 573          if (!buildLiveSets(bn))
 574             return false;
 575       if (n++ || bb->liveSet.marker)
 576          bb->liveSet |= bn->liveSet;
 577       else
 578          bb->liveSet = bn->liveSet;
 579    }
 580    if (!n && !bb->liveSet.marker)
 581       bb->liveSet.fill(0);
 582    bb->liveSet.marker = true;
 583
 584    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 585       INFO("BB:%i live set of out blocks:\n", bb->getId());
 586       bb->liveSet.print();
 587    }
 588
 589    // if (!bb->getEntry())
 590    //   return true;
 591
 592    if (bb == BasicBlock::get(f->cfgExit)) {
 593       for (std::deque<ValueRef>::iterator it = f->outs.begin();
 594            it != f->outs.end(); ++it) {
 595          assert(it->get()->asLValue());
 596          bb->liveSet.set(it->get()->id);
 597       }
 598    }
 599
 600    for (i = bb->getExit(); i && i != bb->getEntry()->prev; i = i->prev) {
 601       for (d = 0; i->defExists(d); ++d)
 602          bb->liveSet.clr(i->getDef(d)->id);
 603       for (s = 0; i->srcExists(s); ++s)
 604          if (i->getSrc(s)->asLValue())
 605             bb->liveSet.set(i->getSrc(s)->id);
 606    }
 607    for (i = bb->getPhi(); i && i->op == OP_PHI; i = i->next)
 608       bb->liveSet.clr(i->getDef(0)->id);
 609
 610    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
 611       INFO("BB:%i live set after propagation:\n", bb->getId());
 612       bb->liveSet.print();
 613    }
 614
 615    return true;
 616 }
 617
 618 void
 619 RegAlloc::BuildIntervalsPass::collectLiveValues(BasicBlock *bb)
 620 {
 621    BasicBlock *bbA = NULL, *bbB = NULL;
 622
 623    if (bb->cfg.outgoingCount()) {
 624       // trickery to save a loop of OR'ing liveSets
 625       // aliasing works fine with BitSet::setOr
 626       for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 627          if (bbA) {
 628             bb->liveSet.setOr(&bbA->liveSet, &bbB->liveSet);
 629             bbA = bb;
 630          } else {
 631             bbA = bbB;
 632          }
 633          bbB = BasicBlock::get(ei.getNode());
 634       }
 635       bb->liveSet.setOr(&bbB->liveSet, bbA ? &bbA->liveSet : NULL);
 636    } else
 637    if (bb->cfg.incidentCount()) {
 638       bb->liveSet.fill(0);
 639    }
 640 }
 641
 642 bool
 643 RegAlloc::BuildIntervalsPass::visit(BasicBlock *bb)
 644 {
 645    collectLiveValues(bb);
 646
 647    INFO_DBG(prog->dbgFlags, REG_ALLOC, "BuildIntervals(BB:%i)\n", bb->getId());
 648
 649    // go through out blocks and delete phi sources that do not originate from
 650    // the current block from the live set
 651    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
 652       BasicBlock *out = BasicBlock::get(ei.getNode());
 653
 654       for (Instruction *i = out->getPhi(); i && i->op == OP_PHI; i = i->next) {
 655          bb->liveSet.clr(i->getDef(0)->id);
 656
 657          for (int s = 0; i->srcExists(s); ++s) {
 658             assert(i->src(s).getInsn());
 659             if (i->getSrc(s)->getUniqueInsn()->bb == bb) // XXX: reachableBy ?
 660                bb->liveSet.set(i->getSrc(s)->id);
 661             else
 662                bb->liveSet.clr(i->getSrc(s)->id);
 663          }
 664       }
 665    }
 666
 667    // remaining live-outs are live until end
 668    if (bb->getExit()) {
 669       for (unsigned int j = 0; j < bb->liveSet.getSize(); ++j)
 670          if (bb->liveSet.test(j))
 671             addLiveRange(func->getLValue(j), bb, bb->getExit()->serial + 1);
 672    }
 673
 674    for (Instruction *i = bb->getExit(); i && i->op != OP_PHI; i = i->prev) {
 675       for (int d = 0; i->defExists(d); ++d) {
 676          bb->liveSet.clr(i->getDef(d)->id);
 677          if (i->getDef(d)->reg.data.id >= 0) // add hazard for fixed regs
 678             i->getDef(d)->livei.extend(i->serial, i->serial);
 679       }
 680
 681       for (int s = 0; i->srcExists(s); ++s) {
 682          if (!i->getSrc(s)->asLValue())
 683             continue;
 684          if (!bb->liveSet.test(i->getSrc(s)->id)) {
 685             bb->liveSet.set(i->getSrc(s)->id);
 686             addLiveRange(i->getSrc(s), bb, i->serial);
 687          }
 688       }
 689    }
 690
 691    if (bb == BasicBlock::get(func->cfg.getRoot())) {
 692       for (std::deque<ValueDef>::iterator it = func->ins.begin();
 693            it != func->ins.end(); ++it) {
 694          if (it->get()->reg.data.id >= 0) // add hazard for fixed regs
 695             it->get()->livei.extend(0, 1);
 696       }
 697    }
 698
 699    return true;
 700 }
 701
 702
 703 #define JOIN_MASK_PHI        (1 << 0)
 704 #define JOIN_MASK_UNION      (1 << 1)
 705 #define JOIN_MASK_MOV        (1 << 2)
 706 #define JOIN_MASK_TEX        (1 << 3)
 707
 708 class GCRA
 709 {
 710 public:
 711    GCRA(Function *, SpillCodeInserter&);
 712    ~GCRA();
 713
 714    bool allocateRegisters(ArrayList& insns);
 715
 716    void printNodeInfo() const;
 717
 718 private:
 719    class RIG_Node : public Graph::Node
 720    {
 721    public:
 722       RIG_Node();
 723
 724       void init(const RegisterSet&, LValue *);
 725
 726       void addInterference(RIG_Node *);
 727       void addRegPreference(RIG_Node *);
 728
 729       inline LValue *getValue() const
 730       {
 731          return reinterpret_cast<LValue *>(data);
 732       }
 733       inline void setValue(LValue *lval) { data = lval; }
 734
 735       inline uint8_t getCompMask() const
 736       {
 737          return ((1 << colors) - 1) << (reg & 7);
 738       }
 739
 740       static inline RIG_Node *get(const Graph::EdgeIterator& ei)
 741       {
 742          return static_cast<RIG_Node *>(ei.getNode());
 743       }
 744
 745    public:
 746       uint32_t degree;
 747       uint16_t degreeLimit; // if deg < degLimit, node is trivially colourable
 748       uint16_t maxReg;
 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    bool 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    struct RelDegree {
 805       uint8_t data[17][17];
 806
 807       RelDegree() {
 808          for (int i = 1; i <= 16; ++i)
 809             for (int j = 1; j <= 16; ++j)
 810                data[i][j] = j * ((i + j - 1) / j);
 811       }
 812
 813       const uint8_t* operator[](std::size_t i) const {
 814          return data[i];
 815       }
 816    };
 817
 818    static const RelDegree relDegree;
 819
 820    RegisterSet regs;
 821
 822    // need to fixup register id for participants of OP_MERGE/SPLIT
 823    std::list<Instruction *> merges;
 824    std::list<Instruction *> splits;
 825
 826    SpillCodeInserter& spill;
 827    std::list<ValuePair> mustSpill;
 828 };
 829
 830 const GCRA::RelDegree GCRA::relDegree;
 831
 832 GCRA::RIG_Node::RIG_Node() : Node(NULL), next(this), prev(this)
 833 {
 834    colors = 0;
 835 }
 836
 837 void
 838 GCRA::printNodeInfo() const
 839 {
 840    for (unsigned int i = 0; i < nodeCount; ++i) {
 841       if (!nodes[i].colors)
 842          continue;
 843       INFO("RIG_Node[%%%i]($[%u]%i): %u colors, weight %f, deg %u/%u\n X",
 844            i,
 845            nodes[i].f,nodes[i].reg,nodes[i].colors,
 846            nodes[i].weight,
 847            nodes[i].degree, nodes[i].degreeLimit);
 848
 849       for (Graph::EdgeIterator ei = nodes[i].outgoing(); !ei.end(); ei.next())
 850          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
 851       for (Graph::EdgeIterator ei = nodes[i].incident(); !ei.end(); ei.next())
 852          INFO(" %%%i", RIG_Node::get(ei)->getValue()->id);
 853       INFO("\n");
 854    }
 855 }
 856
 857 static bool
 858 isShortRegOp(Instruction *insn)
 859 {
 860    // Immediates are always in src1 (except zeroes, which end up getting
 861    // replaced with a zero reg). Every other situation can be resolved by
 862    // using a long encoding.
 863    return insn->srcExists(1) && insn->src(1).getFile() == FILE_IMMEDIATE &&
 864       insn->getSrc(1)->reg.data.u64;
 865 }
 866
 867 // Check if this LValue is ever used in an instruction that can't be encoded
 868 // with long registers (i.e. > r63)
 869 static bool
 870 isShortRegVal(LValue *lval)
 871 {
 872    if (lval->getInsn() == NULL)
 873       return false;
 874    for (Value::DefCIterator def = lval->defs.begin();
 875         def != lval->defs.end(); ++def)
 876       if (isShortRegOp((*def)->getInsn()))
 877          return true;
 878    for (Value::UseCIterator use = lval->uses.begin();
 879         use != lval->uses.end(); ++use)
 880       if (isShortRegOp((*use)->getInsn()))
 881          return true;
 882    return false;
 883 }
 884
 885 void
 886 GCRA::RIG_Node::init(const RegisterSet& regs, LValue *lval)
 887 {
 888    setValue(lval);
 889    if (lval->reg.data.id >= 0)
 890       lval->noSpill = lval->fixedReg = 1;
 891
 892    colors = regs.units(lval->reg.file, lval->reg.size);
 893    f = lval->reg.file;
 894    reg = -1;
 895    if (lval->reg.data.id >= 0)
 896       reg = regs.idToUnits(lval);
 897
 898    weight = std::numeric_limits<float>::infinity();
 899    degree = 0;
 900    maxReg = regs.getFileSize(f);
 901    // On nv50, we lose a bit of gpr encoding when there's an embedded
 902    // immediate.
 903    if (regs.restrictedGPR16Range && f == FILE_GPR && (lval->reg.size == 2 || isShortRegVal(lval)))
 904       maxReg /= 2;
 905    degreeLimit = maxReg;
 906    degreeLimit -= relDegree[1][colors] - 1;
 907
 908    livei.insert(lval->livei);
 909 }
 910
 911 bool
 912 GCRA::coalesceValues(Value *dst, Value *src, bool force)
 913 {
 914    LValue *rep = dst->join->asLValue();
 915    LValue *val = src->join->asLValue();
 916
 917    if (!force && val->reg.data.id >= 0) {
 918       rep = src->join->asLValue();
 919       val = dst->join->asLValue();
 920    }
 921    RIG_Node *nRep = &nodes[rep->id];
 922    RIG_Node *nVal = &nodes[val->id];
 923
 924    if (src->reg.file != dst->reg.file) {
 925       if (!force)
 926          return false;
 927       WARN("forced coalescing of values in different files !\n");
 928    }
 929    if (!force && dst->reg.size != src->reg.size)
 930       return false;
 931
 932    if ((rep->reg.data.id >= 0) && (rep->reg.data.id != val->reg.data.id)) {
 933       if (force) {
 934          if (val->reg.data.id >= 0)
 935             WARN("forced coalescing of values in different fixed regs !\n");
 936       } else {
 937          if (val->reg.data.id >= 0)
 938             return false;
 939          // make sure that there is no overlap with the fixed register of rep
 940          for (ArrayList::Iterator it = func->allLValues.iterator();
 941               !it.end(); it.next()) {
 942             Value *reg = reinterpret_cast<Value *>(it.get())->asLValue();
 943             assert(reg);
 944             if (reg->interfers(rep) && reg->livei.overlaps(nVal->livei))
 945                return false;
 946          }
 947       }
 948    }
 949
 950    if (!force && nRep->livei.overlaps(nVal->livei))
 951       return false;
 952
 953    INFO_DBG(prog->dbgFlags, REG_ALLOC, "joining %%%i($%i) <- %%%i\n",
 954             rep->id, rep->reg.data.id, val->id);
 955
 956    // set join pointer of all values joined with val
 957    for (Value::DefIterator def = val->defs.begin(); def != val->defs.end();
 958         ++def)
 959       (*def)->get()->join = rep;
 960    assert(rep->join == rep && val->join == rep);
 961
 962    // add val's definitions to rep and extend the live interval of its RIG node
 963    rep->defs.insert(rep->defs.end(), val->defs.begin(), val->defs.end());
 964    nRep->livei.unify(nVal->livei);
 965    nRep->degreeLimit = MIN2(nRep->degreeLimit, nVal->degreeLimit);
 966    nRep->maxReg = MIN2(nRep->maxReg, nVal->maxReg);
 967    return true;
 968 }
 969
 970 bool
 971 GCRA::coalesce(ArrayList& insns)
 972 {
 973    bool ret = doCoalesce(insns, JOIN_MASK_PHI);
 974    if (!ret)
 975       return false;
 976    switch (func->getProgram()->getTarget()->getChipset() & ~0xf) {
 977    case 0x50:
 978    case 0x80:
 979    case 0x90:
 980    case 0xa0:
 981       ret = doCoalesce(insns, JOIN_MASK_UNION | JOIN_MASK_TEX);
 982       break;
 983    case 0xc0:
 984    case 0xd0:
 985    case 0xe0:
 986    case 0xf0:
 987    case 0x100:
 988    case 0x110:
 989    case 0x120:
 990    case 0x130:
 991    case 0x140:
 992       ret = doCoalesce(insns, JOIN_MASK_UNION);
 993       break;
 994    default:
 995       break;
 996    }
 997    if (!ret)
 998       return false;
 999    return doCoalesce(insns, JOIN_MASK_MOV);
1000 }
1001
1002 static inline uint8_t makeCompMask(int compSize, int base, int size)
1003 {
1004    uint8_t m = ((1 << size) - 1) << base;
1005
1006    switch (compSize) {
1007    case 1:
1008       return 0xff;
1009    case 2:
1010       m |= (m << 2);
1011       return (m << 4) | m;
1012    case 3:
1013    case 4:
1014       return (m << 4) | m;
1015    default:
1016       assert(compSize <= 8);
1017       return m;
1018    }
1019 }
1020
1021 // Used when coalescing moves. The non-compound value will become one, e.g.:
1022 // mov b32 $r0 $r2            / merge b64 $r0d { $r0 $r1 }
1023 // split b64 { $r0 $r1 } $r0d / mov b64 $r0d f64 $r2d
1024 static inline void copyCompound(Value *dst, Value *src)
1025 {
1026    LValue *ldst = dst->asLValue();
1027    LValue *lsrc = src->asLValue();
1028
1029    if (ldst->compound && !lsrc->compound) {
1030       LValue *swap = lsrc;
1031       lsrc = ldst;
1032       ldst = swap;
1033    }
1034
1035    ldst->compound = lsrc->compound;
1036    ldst->compMask = lsrc->compMask;
1037 }
1038
1039 void
1040 GCRA::makeCompound(Instruction *insn, bool split)
1041 {
1042    LValue *rep = (split ? insn->getSrc(0) : insn->getDef(0))->asLValue();
1043
1044    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC) {
1045       INFO("makeCompound(split = %i): ", split);
1046       insn->print();
1047    }
1048
1049    const unsigned int size = getNode(rep)->colors;
1050    unsigned int base = 0;
1051
1052    if (!rep->compound)
1053       rep->compMask = 0xff;
1054    rep->compound = 1;
1055
1056    for (int c = 0; split ? insn->defExists(c) : insn->srcExists(c); ++c) {
1057       LValue *val = (split ? insn->getDef(c) : insn->getSrc(c))->asLValue();
1058
1059       val->compound = 1;
1060       if (!val->compMask)
1061          val->compMask = 0xff;
1062       val->compMask &= makeCompMask(size, base, getNode(val)->colors);
1063       assert(val->compMask);
1064
1065       INFO_DBG(prog->dbgFlags, REG_ALLOC, "compound: %%%i:%02x <- %%%i:%02x\n",
1066            rep->id, rep->compMask, val->id, val->compMask);
1067
1068       base += getNode(val)->colors;
1069    }
1070    assert(base == size);
1071 }
1072
1073 bool
1074 GCRA::doCoalesce(ArrayList& insns, unsigned int mask)
1075 {
1076    int c, n;
1077
1078    for (n = 0; n < insns.getSize(); ++n) {
1079       Instruction *i;
1080       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(n));
1081
1082       switch (insn->op) {
1083       case OP_PHI:
1084          if (!(mask & JOIN_MASK_PHI))
1085             break;
1086          for (c = 0; insn->srcExists(c); ++c)
1087             if (!coalesceValues(insn->getDef(0), insn->getSrc(c), false)) {
1088                // this is bad
1089                ERROR("failed to coalesce phi operands\n");
1090                return false;
1091             }
1092          break;
1093       case OP_UNION:
1094       case OP_MERGE:
1095          if (!(mask & JOIN_MASK_UNION))
1096             break;
1097          for (c = 0; insn->srcExists(c); ++c)
1098             coalesceValues(insn->getDef(0), insn->getSrc(c), true);
1099          if (insn->op == OP_MERGE) {
1100             merges.push_back(insn);
1101             if (insn->srcExists(1))
1102                makeCompound(insn, false);
1103          }
1104          break;
1105       case OP_SPLIT:
1106          if (!(mask & JOIN_MASK_UNION))
1107             break;
1108          splits.push_back(insn);
1109          for (c = 0; insn->defExists(c); ++c)
1110             coalesceValues(insn->getSrc(0), insn->getDef(c), true);
1111          makeCompound(insn, true);
1112          break;
1113       case OP_MOV:
1114          if (!(mask & JOIN_MASK_MOV))
1115             break;
1116          i = NULL;
1117          if (!insn->getDef(0)->uses.empty())
1118             i = (*insn->getDef(0)->uses.begin())->getInsn();
1119          // if this is a contraint-move there will only be a single use
1120          if (i && i->op == OP_MERGE) // do we really still need this ?
1121             break;
1122          i = insn->getSrc(0)->getUniqueInsn();
1123          if (i && !i->constrainedDefs()) {
1124             if (coalesceValues(insn->getDef(0), insn->getSrc(0), false))
1125                copyCompound(insn->getSrc(0), insn->getDef(0));
1126          }
1127          break;
1128       case OP_TEX:
1129       case OP_TXB:
1130       case OP_TXL:
1131       case OP_TXF:
1132       case OP_TXQ:
1133       case OP_TXD:
1134       case OP_TXG:
1135       case OP_TXLQ:
1136       case OP_TEXCSAA:
1137       case OP_TEXPREP:
1138          if (!(mask & JOIN_MASK_TEX))
1139             break;
1140          for (c = 0; insn->srcExists(c) && c != insn->predSrc; ++c)
1141             coalesceValues(insn->getDef(c), insn->getSrc(c), true);
1142          break;
1143       default:
1144          break;
1145       }
1146    }
1147    return true;
1148 }
1149
1150 void
1151 GCRA::RIG_Node::addInterference(RIG_Node *node)
1152 {
1153    this->degree += relDegree[node->colors][colors];
1154    node->degree += relDegree[colors][node->colors];
1155
1156    this->attach(node, Graph::Edge::CROSS);
1157 }
1158
1159 void
1160 GCRA::RIG_Node::addRegPreference(RIG_Node *node)
1161 {
1162    prefRegs.push_back(node);
1163 }
1164
1165 GCRA::GCRA(Function *fn, SpillCodeInserter& spill) :
1166    func(fn),
1167    regs(fn->getProgram()->getTarget()),
1168    spill(spill)
1169 {
1170    prog = func->getProgram();
1171 }
1172
1173 GCRA::~GCRA()
1174 {
1175    if (nodes)
1176       delete[] nodes;
1177 }
1178
1179 void
1180 GCRA::checkList(std::list<RIG_Node *>& lst)
1181 {
1182    GCRA::RIG_Node *prev = NULL;
1183
1184    for (std::list<RIG_Node *>::iterator it = lst.begin();
1185         it != lst.end();
1186         ++it) {
1187       assert((*it)->getValue()->join == (*it)->getValue());
1188       if (prev)
1189          assert(prev->livei.begin() <= (*it)->livei.begin());
1190       prev = *it;
1191    }
1192 }
1193
1194 void
1195 GCRA::insertOrderedTail(std::list<RIG_Node *>& list, RIG_Node *node)
1196 {
1197    if (node->livei.isEmpty())
1198       return;
1199    // only the intervals of joined values don't necessarily arrive in order
1200    std::list<RIG_Node *>::iterator prev, it;
1201    for (it = list.end(); it != list.begin(); it = prev) {
1202       prev = it;
1203       --prev;
1204       if ((*prev)->livei.begin() <= node->livei.begin())
1205          break;
1206    }
1207    list.insert(it, node);
1208 }
1209
1210 void
1211 GCRA::buildRIG(ArrayList& insns)
1212 {
1213    std::list<RIG_Node *> values, active;
1214
1215    for (std::deque<ValueDef>::iterator it = func->ins.begin();
1216         it != func->ins.end(); ++it)
1217       insertOrderedTail(values, getNode(it->get()->asLValue()));
1218
1219    for (int i = 0; i < insns.getSize(); ++i) {
1220       Instruction *insn = reinterpret_cast<Instruction *>(insns.get(i));
1221       for (int d = 0; insn->defExists(d); ++d)
1222          if (insn->getDef(d)->rep() == insn->getDef(d))
1223             insertOrderedTail(values, getNode(insn->getDef(d)->asLValue()));
1224    }
1225    checkList(values);
1226
1227    while (!values.empty()) {
1228       RIG_Node *cur = values.front();
1229
1230       for (std::list<RIG_Node *>::iterator it = active.begin();
1231            it != active.end();) {
1232          RIG_Node *node = *it;
1233
1234          if (node->livei.end() <= cur->livei.begin()) {
1235             it = active.erase(it);
1236          } else {
1237             if (node->f == cur->f && node->livei.overlaps(cur->livei))
1238                cur->addInterference(node);
1239             ++it;
1240          }
1241       }
1242       values.pop_front();
1243       active.push_back(cur);
1244    }
1245 }
1246
1247 void
1248 GCRA::calculateSpillWeights()
1249 {
1250    for (unsigned int i = 0; i < nodeCount; ++i) {
1251       RIG_Node *const n = &nodes[i];
1252       if (!nodes[i].colors || nodes[i].livei.isEmpty())
1253          continue;
1254       if (nodes[i].reg >= 0) {
1255          // update max reg
1256          regs.occupy(n->f, n->reg, n->colors);
1257          continue;
1258       }
1259       LValue *val = nodes[i].getValue();
1260
1261       if (!val->noSpill) {
1262          int rc = 0;
1263          for (Value::DefIterator it = val->defs.begin();
1264               it != val->defs.end();
1265               ++it)
1266             rc += (*it)->get()->refCount();
1267
1268          nodes[i].weight =
1269             (float)rc * (float)rc / (float)nodes[i].livei.extent();
1270       }
1271
1272       if (nodes[i].degree < nodes[i].degreeLimit) {
1273          int l = 0;
1274          if (val->reg.size > 4)
1275             l = 1;
1276          DLLIST_ADDHEAD(&lo[l], &nodes[i]);
1277       } else {
1278          DLLIST_ADDHEAD(&hi, &nodes[i]);
1279       }
1280    }
1281    if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1282       printNodeInfo();
1283 }
1284
1285 void
1286 GCRA::simplifyEdge(RIG_Node *a, RIG_Node *b)
1287 {
1288    bool move = b->degree >= b->degreeLimit;
1289
1290    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1291             "edge: (%%%i, deg %u/%u) >-< (%%%i, deg %u/%u)\n",
1292             a->getValue()->id, a->degree, a->degreeLimit,
1293             b->getValue()->id, b->degree, b->degreeLimit);
1294
1295    b->degree -= relDegree[a->colors][b->colors];
1296
1297    move = move && b->degree < b->degreeLimit;
1298    if (move && !DLLIST_EMPTY(b)) {
1299       int l = (b->getValue()->reg.size > 4) ? 1 : 0;
1300       DLLIST_DEL(b);
1301       DLLIST_ADDTAIL(&lo[l], b);
1302    }
1303 }
1304
1305 void
1306 GCRA::simplifyNode(RIG_Node *node)
1307 {
1308    for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1309       simplifyEdge(node, RIG_Node::get(ei));
1310
1311    for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1312       simplifyEdge(node, RIG_Node::get(ei));
1313
1314    DLLIST_DEL(node);
1315    stack.push(node->getValue()->id);
1316
1317    INFO_DBG(prog->dbgFlags, REG_ALLOC, "SIMPLIFY: pushed %%%i%s\n",
1318             node->getValue()->id,
1319             (node->degree < node->degreeLimit) ? "" : "(spill)");
1320 }
1321
1322 bool
1323 GCRA::simplify()
1324 {
1325    for (;;) {
1326       if (!DLLIST_EMPTY(&lo[0])) {
1327          do {
1328             simplifyNode(lo[0].next);
1329          } while (!DLLIST_EMPTY(&lo[0]));
1330       } else
1331       if (!DLLIST_EMPTY(&lo[1])) {
1332          simplifyNode(lo[1].next);
1333       } else
1334       if (!DLLIST_EMPTY(&hi)) {
1335          RIG_Node *best = hi.next;
1336          unsigned bestMaxReg = best->maxReg;
1337          float bestScore = best->weight / (float)best->degree;
1338          // Spill candidate. First go through the ones with the highest max
1339          // register, then the ones with lower. That way the ones with the
1340          // lowest requirement will be allocated first, since it's a stack.
1341          for (RIG_Node *it = best->next; it != &hi; it = it->next) {
1342             float score = it->weight / (float)it->degree;
1343             if (score < bestScore || it->maxReg > bestMaxReg) {
1344                best = it;
1345                bestScore = score;
1346                bestMaxReg = it->maxReg;
1347             }
1348          }
1349          if (isinf(bestScore)) {
1350             ERROR("no viable spill candidates left\n");
1351             return false;
1352          }
1353          simplifyNode(best);
1354       } else {
1355          return true;
1356       }
1357    }
1358 }
1359
1360 void
1361 GCRA::checkInterference(const RIG_Node *node, Graph::EdgeIterator& ei)
1362 {
1363    const RIG_Node *intf = RIG_Node::get(ei);
1364
1365    if (intf->reg < 0)
1366       return;
1367    const LValue *vA = node->getValue();
1368    const LValue *vB = intf->getValue();
1369
1370    const uint8_t intfMask = ((1 << intf->colors) - 1) << (intf->reg & 7);
1371
1372    if (vA->compound | vB->compound) {
1373       // NOTE: this only works for >aligned< register tuples !
1374       for (Value::DefCIterator D = vA->defs.begin(); D != vA->defs.end(); ++D) {
1375       for (Value::DefCIterator d = vB->defs.begin(); d != vB->defs.end(); ++d) {
1376          const LValue *vD = (*D)->get()->asLValue();
1377          const LValue *vd = (*d)->get()->asLValue();
1378
1379          if (!vD->livei.overlaps(vd->livei)) {
1380             INFO_DBG(prog->dbgFlags, REG_ALLOC, "(%%%i) X (%%%i): no overlap\n",
1381                      vD->id, vd->id);
1382             continue;
1383          }
1384
1385          uint8_t mask = vD->compound ? vD->compMask : ~0;
1386          if (vd->compound) {
1387             assert(vB->compound);
1388             mask &= vd->compMask & vB->compMask;
1389          } else {
1390             mask &= intfMask;
1391          }
1392
1393          INFO_DBG(prog->dbgFlags, REG_ALLOC,
1394                   "(%%%i)%02x X (%%%i)%02x & %02x: $r%i.%02x\n",
1395                   vD->id,
1396                   vD->compound ? vD->compMask : 0xff,
1397                   vd->id,
1398                   vd->compound ? vd->compMask : intfMask,
1399                   vB->compMask, intf->reg & ~7, mask);
1400          if (mask)
1401             regs.occupyMask(node->f, intf->reg & ~7, mask);
1402       }
1403       }
1404    } else {
1405       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1406                "(%%%i) X (%%%i): $r%i + %u\n",
1407                vA->id, vB->id, intf->reg, intf->colors);
1408       regs.occupy(node->f, intf->reg, intf->colors);
1409    }
1410 }
1411
1412 bool
1413 GCRA::selectRegisters()
1414 {
1415    INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nSELECT phase\n");
1416
1417    while (!stack.empty()) {
1418       RIG_Node *node = &nodes[stack.top()];
1419       stack.pop();
1420
1421       regs.reset(node->f);
1422
1423       INFO_DBG(prog->dbgFlags, REG_ALLOC, "\nNODE[%%%i, %u colors]\n",
1424                node->getValue()->id, node->colors);
1425
1426       for (Graph::EdgeIterator ei = node->outgoing(); !ei.end(); ei.next())
1427          checkInterference(node, ei);
1428       for (Graph::EdgeIterator ei = node->incident(); !ei.end(); ei.next())
1429          checkInterference(node, ei);
1430
1431       if (!node->prefRegs.empty()) {
1432          for (std::list<RIG_Node *>::const_iterator it = node->prefRegs.begin();
1433               it != node->prefRegs.end();
1434               ++it) {
1435             if ((*it)->reg >= 0 &&
1436                 regs.testOccupy(node->f, (*it)->reg, node->colors)) {
1437                node->reg = (*it)->reg;
1438                break;
1439             }
1440          }
1441       }
1442       if (node->reg >= 0)
1443          continue;
1444       LValue *lval = node->getValue();
1445       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1446          regs.print(node->f);
1447       bool ret = regs.assign(node->reg, node->f, node->colors, node->maxReg);
1448       if (ret) {
1449          INFO_DBG(prog->dbgFlags, REG_ALLOC, "assigned reg %i\n", node->reg);
1450          lval->compMask = node->getCompMask();
1451       } else {
1452          INFO_DBG(prog->dbgFlags, REG_ALLOC, "must spill: %%%i (size %u)\n",
1453                   lval->id, lval->reg.size);
1454          Symbol *slot = NULL;
1455          if (lval->reg.file == FILE_GPR)
1456             slot = spill.assignSlot(node->livei, lval->reg.size);
1457          mustSpill.push_back(ValuePair(lval, slot));
1458       }
1459    }
1460    if (!mustSpill.empty())
1461       return false;
1462    for (unsigned int i = 0; i < nodeCount; ++i) {
1463       LValue *lval = nodes[i].getValue();
1464       if (nodes[i].reg >= 0 && nodes[i].colors > 0)
1465          lval->reg.data.id =
1466             regs.unitsToId(nodes[i].f, nodes[i].reg, lval->reg.size);
1467    }
1468    return true;
1469 }
1470
1471 bool
1472 GCRA::allocateRegisters(ArrayList& insns)
1473 {
1474    bool ret;
1475
1476    INFO_DBG(prog->dbgFlags, REG_ALLOC,
1477             "allocateRegisters to %u instructions\n", insns.getSize());
1478
1479    nodeCount = func->allLValues.getSize();
1480    nodes = new RIG_Node[nodeCount];
1481    if (!nodes)
1482       return false;
1483    for (unsigned int i = 0; i < nodeCount; ++i) {
1484       LValue *lval = reinterpret_cast<LValue *>(func->allLValues.get(i));
1485       if (lval) {
1486          nodes[i].init(regs, lval);
1487          RIG.insert(&nodes[i]);
1488
1489          if (lval->inFile(FILE_GPR) && lval->getInsn() != NULL) {
1490             Instruction *insn = lval->getInsn();
1491             if (insn->op != OP_MAD && insn->op != OP_FMA && insn->op != OP_SAD)
1492                continue;
1493             // For both of the cases below, we only want to add the preference
1494             // if all arguments are in registers.
1495             if (insn->src(0).getFile() != FILE_GPR ||
1496                 insn->src(1).getFile() != FILE_GPR ||
1497                 insn->src(2).getFile() != FILE_GPR)
1498                continue;
1499             if (prog->getTarget()->getChipset() < 0xc0) {
1500                // Outputting a flag is not supported with short encodings nor
1501                // with immediate arguments.
1502                // See handleMADforNV50.
1503                if (insn->flagsDef >= 0)
1504                   continue;
1505             } else {
1506                // We can only fold immediate arguments if dst == src2. This
1507                // only matters if one of the first two arguments is an
1508                // immediate. This form is also only supported for floats.
1509                // See handleMADforNVC0.
1510                ImmediateValue imm;
1511                if (insn->dType != TYPE_F32)
1512                   continue;
1513                if (!insn->src(0).getImmediate(imm) &&
1514                    !insn->src(1).getImmediate(imm))
1515                   continue;
1516             }
1517
1518             nodes[i].addRegPreference(getNode(insn->getSrc(2)->asLValue()));
1519          }
1520       }
1521    }
1522
1523    // coalesce first, we use only 1 RIG node for a group of joined values
1524    ret = coalesce(insns);
1525    if (!ret)
1526       goto out;
1527
1528    if (func->getProgram()->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1529       func->printLiveIntervals();
1530
1531    buildRIG(insns);
1532    calculateSpillWeights();
1533    ret = simplify();
1534    if (!ret)
1535       goto out;
1536
1537    ret = selectRegisters();
1538    if (!ret) {
1539       INFO_DBG(prog->dbgFlags, REG_ALLOC,
1540                "selectRegisters failed, inserting spill code ...\n");
1541       regs.reset(FILE_GPR, true);
1542       spill.run(mustSpill);
1543       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1544          func->print();
1545    } else {
1546       prog->maxGPR = std::max(prog->maxGPR, regs.getMaxAssigned(FILE_GPR));
1547    }
1548
1549 out:
1550    cleanup(ret);
1551    return ret;
1552 }
1553
1554 void
1555 GCRA::cleanup(const bool success)
1556 {
1557    mustSpill.clear();
1558
1559    for (ArrayList::Iterator it = func->allLValues.iterator();
1560         !it.end(); it.next()) {
1561       LValue *lval =  reinterpret_cast<LValue *>(it.get());
1562
1563       lval->livei.clear();
1564
1565       lval->compound = 0;
1566       lval->compMask = 0;
1567
1568       if (lval->join == lval)
1569          continue;
1570
1571       if (success) {
1572          lval->reg.data.id = lval->join->reg.data.id;
1573       } else {
1574          for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1575               ++d)
1576             lval->join->defs.remove(*d);
1577          lval->join = lval;
1578       }
1579    }
1580
1581    if (success)
1582       resolveSplitsAndMerges();
1583    splits.clear(); // avoid duplicate entries on next coalesce pass
1584    merges.clear();
1585
1586    delete[] nodes;
1587    nodes = NULL;
1588    hi.next = hi.prev = &hi;
1589    lo[0].next = lo[0].prev = &lo[0];
1590    lo[1].next = lo[1].prev = &lo[1];
1591 }
1592
1593 Symbol *
1594 SpillCodeInserter::assignSlot(const Interval &livei, const unsigned int size)
1595 {
1596    SpillSlot slot;
1597    int32_t offsetBase = stackSize;
1598    int32_t offset;
1599    std::list<SpillSlot>::iterator pos = slots.end(), it = slots.begin();
1600
1601    if (offsetBase % size)
1602       offsetBase += size - (offsetBase % size);
1603
1604    slot.sym = NULL;
1605
1606    for (offset = offsetBase; offset < stackSize; offset += size) {
1607       const int32_t entryEnd = offset + size;
1608       while (it != slots.end() && it->offset < offset)
1609          ++it;
1610       if (it == slots.end()) // no slots left
1611          break;
1612       std::list<SpillSlot>::iterator bgn = it;
1613
1614       while (it != slots.end() && it->offset < entryEnd) {
1615          it->occup.print();
1616          if (it->occup.overlaps(livei))
1617             break;
1618          ++it;
1619       }
1620       if (it == slots.end() || it->offset >= entryEnd) {
1621          // fits
1622          for (; bgn != slots.end() && bgn->offset < entryEnd; ++bgn) {
1623             bgn->occup.insert(livei);
1624             if (bgn->size() == size)
1625                slot.sym = bgn->sym;
1626          }
1627          break;
1628       }
1629    }
1630    if (!slot.sym) {
1631       stackSize = offset + size;
1632       slot.offset = offset;
1633       slot.sym = new_Symbol(func->getProgram(), FILE_MEMORY_LOCAL);
1634       if (!func->stackPtr)
1635          offset += func->tlsBase;
1636       slot.sym->setAddress(NULL, offset);
1637       slot.sym->reg.size = size;
1638       slots.insert(pos, slot)->occup.insert(livei);
1639    }
1640    return slot.sym;
1641 }
1642
1643 Value *
1644 SpillCodeInserter::offsetSlot(Value *base, const LValue *lval)
1645 {
1646    if (!lval->compound || (lval->compMask & 0x1))
1647       return base;
1648    Value *slot = cloneShallow(func, base);
1649
1650    slot->reg.data.offset += (ffs(lval->compMask) - 1) * lval->reg.size;
1651    slot->reg.size = lval->reg.size;
1652
1653    return slot;
1654 }
1655
1656 void
1657 SpillCodeInserter::spill(Instruction *defi, Value *slot, LValue *lval)
1658 {
1659    const DataType ty = typeOfSize(lval->reg.size);
1660
1661    slot = offsetSlot(slot, lval);
1662
1663    Instruction *st;
1664    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1665       lval->noSpill = 1;
1666       if (ty != TYPE_B96) {
1667          st = new_Instruction(func, OP_STORE, ty);
1668          st->setSrc(0, slot);
1669          st->setSrc(1, lval);
1670       } else {
1671          st = new_Instruction(func, OP_SPLIT, ty);
1672          st->setSrc(0, lval);
1673          for (int d = 0; d < lval->reg.size / 4; ++d)
1674             st->setDef(d, new_LValue(func, FILE_GPR));
1675
1676          for (int d = lval->reg.size / 4 - 1; d >= 0; --d) {
1677             Value *tmp = cloneShallow(func, slot);
1678             tmp->reg.size = 4;
1679             tmp->reg.data.offset += 4 * d;
1680
1681             Instruction *s = new_Instruction(func, OP_STORE, TYPE_U32);
1682             s->setSrc(0, tmp);
1683             s->setSrc(1, st->getDef(d));
1684             defi->bb->insertAfter(defi, s);
1685          }
1686       }
1687    } else {
1688       st = new_Instruction(func, OP_CVT, ty);
1689       st->setDef(0, slot);
1690       st->setSrc(0, lval);
1691       if (lval->reg.file == FILE_FLAGS)
1692          st->flagsSrc = 0;
1693    }
1694    defi->bb->insertAfter(defi, st);
1695 }
1696
1697 LValue *
1698 SpillCodeInserter::unspill(Instruction *usei, LValue *lval, Value *slot)
1699 {
1700    const DataType ty = typeOfSize(lval->reg.size);
1701
1702    slot = offsetSlot(slot, lval);
1703    lval = cloneShallow(func, lval);
1704
1705    Instruction *ld;
1706    if (slot->reg.file == FILE_MEMORY_LOCAL) {
1707       lval->noSpill = 1;
1708       if (ty != TYPE_B96) {
1709          ld = new_Instruction(func, OP_LOAD, ty);
1710       } else {
1711          ld = new_Instruction(func, OP_MERGE, ty);
1712          for (int d = 0; d < lval->reg.size / 4; ++d) {
1713             Value *tmp = cloneShallow(func, slot);
1714             LValue *val;
1715             tmp->reg.size = 4;
1716             tmp->reg.data.offset += 4 * d;
1717
1718             Instruction *l = new_Instruction(func, OP_LOAD, TYPE_U32);
1719             l->setDef(0, (val = new_LValue(func, FILE_GPR)));
1720             l->setSrc(0, tmp);
1721             usei->bb->insertBefore(usei, l);
1722             ld->setSrc(d, val);
1723             val->noSpill = 1;
1724          }
1725          ld->setDef(0, lval);
1726          usei->bb->insertBefore(usei, ld);
1727          return lval;
1728       }
1729    } else {
1730       ld = new_Instruction(func, OP_CVT, ty);
1731    }
1732    ld->setDef(0, lval);
1733    ld->setSrc(0, slot);
1734    if (lval->reg.file == FILE_FLAGS)
1735       ld->flagsDef = 0;
1736
1737    usei->bb->insertBefore(usei, ld);
1738    return lval;
1739 }
1740
1741 static bool
1742 value_cmp(ValueRef *a, ValueRef *b) {
1743    Instruction *ai = a->getInsn(), *bi = b->getInsn();
1744    if (ai->bb != bi->bb)
1745       return ai->bb->getId() < bi->bb->getId();
1746    return ai->serial < bi->serial;
1747 }
1748
1749 // For each value that is to be spilled, go through all its definitions.
1750 // A value can have multiple definitions if it has been coalesced before.
1751 // For each definition, first go through all its uses and insert an unspill
1752 // instruction before it, then replace the use with the temporary register.
1753 // Unspill can be either a load from memory or simply a move to another
1754 // register file.
1755 // For "Pseudo" instructions (like PHI, SPLIT, MERGE) we can erase the use
1756 // if we have spilled to a memory location, or simply with the new register.
1757 // No load or conversion instruction should be needed.
1758 bool
1759 SpillCodeInserter::run(const std::list<ValuePair>& lst)
1760 {
1761    for (std::list<ValuePair>::const_iterator it = lst.begin(); it != lst.end();
1762         ++it) {
1763       LValue *lval = it->first->asLValue();
1764       Symbol *mem = it->second ? it->second->asSym() : NULL;
1765
1766       // Keep track of which instructions to delete later. Deleting them
1767       // inside the loop is unsafe since a single instruction may have
1768       // multiple destinations that all need to be spilled (like OP_SPLIT).
1769       unordered_set<Instruction *> to_del;
1770
1771       for (Value::DefIterator d = lval->defs.begin(); d != lval->defs.end();
1772            ++d) {
1773          Value *slot = mem ?
1774             static_cast<Value *>(mem) : new_LValue(func, FILE_GPR);
1775          Value *tmp = NULL;
1776          Instruction *last = NULL;
1777
1778          LValue *dval = (*d)->get()->asLValue();
1779          Instruction *defi = (*d)->getInsn();
1780
1781          // Sort all the uses by BB/instruction so that we don't unspill
1782          // multiple times in a row, and also remove a source of
1783          // non-determinism.
1784          std::vector<ValueRef *> refs(dval->uses.begin(), dval->uses.end());
1785          std::sort(refs.begin(), refs.end(), value_cmp);
1786
1787          // Unspill at each use *before* inserting spill instructions,
1788          // we don't want to have the spill instructions in the use list here.
1789          for (std::vector<ValueRef*>::const_iterator it = refs.begin();
1790               it != refs.end(); ++it) {
1791             ValueRef *u = *it;
1792             Instruction *usei = u->getInsn();
1793             assert(usei);
1794             if (usei->isPseudo()) {
1795                tmp = (slot->reg.file == FILE_MEMORY_LOCAL) ? NULL : slot;
1796                last = NULL;
1797             } else {
1798                if (!last || (usei != last->next && usei != last))
1799                   tmp = unspill(usei, dval, slot);
1800                last = usei;
1801             }
1802             u->set(tmp);
1803          }
1804
1805          assert(defi);
1806          if (defi->isPseudo()) {
1807             d = lval->defs.erase(d);
1808             --d;
1809             if (slot->reg.file == FILE_MEMORY_LOCAL)
1810                to_del.insert(defi);
1811             else
1812                defi->setDef(0, slot);
1813          } else {
1814             spill(defi, slot, dval);
1815          }
1816       }
1817
1818       for (unordered_set<Instruction *>::const_iterator it = to_del.begin();
1819            it != to_del.end(); ++it)
1820          delete_Instruction(func->getProgram(), *it);
1821    }
1822
1823    // TODO: We're not trying to reuse old slots in a potential next iteration.
1824    //  We have to update the slots' livei intervals to be able to do that.
1825    stackBase = stackSize;
1826    slots.clear();
1827    return true;
1828 }
1829
1830 bool
1831 RegAlloc::exec()
1832 {
1833    for (IteratorRef it = prog->calls.iteratorDFS(false);
1834         !it->end(); it->next()) {
1835       func = Function::get(reinterpret_cast<Graph::Node *>(it->get()));
1836
1837       func->tlsBase = prog->tlsSize;
1838       if (!execFunc())
1839          return false;
1840       prog->tlsSize += func->tlsSize;
1841    }
1842    return true;
1843 }
1844
1845 bool
1846 RegAlloc::execFunc()
1847 {
1848    InsertConstraintsPass insertConstr;
1849    PhiMovesPass insertPhiMoves;
1850    ArgumentMovesPass insertArgMoves;
1851    BuildIntervalsPass buildIntervals;
1852    SpillCodeInserter insertSpills(func);
1853
1854    GCRA gcra(func, insertSpills);
1855
1856    unsigned int i, retries;
1857    bool ret;
1858
1859    if (!func->ins.empty()) {
1860       // Insert a nop at the entry so inputs only used by the first instruction
1861       // don't count as having an empty live range.
1862       Instruction *nop = new_Instruction(func, OP_NOP, TYPE_NONE);
1863       BasicBlock::get(func->cfg.getRoot())->insertHead(nop);
1864    }
1865
1866    ret = insertConstr.exec(func);
1867    if (!ret)
1868       goto out;
1869
1870    ret = insertPhiMoves.run(func);
1871    if (!ret)
1872       goto out;
1873
1874    ret = insertArgMoves.run(func);
1875    if (!ret)
1876       goto out;
1877
1878    // TODO: need to fix up spill slot usage ranges to support > 1 retry
1879    for (retries = 0; retries < 3; ++retries) {
1880       if (retries && (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC))
1881          INFO("Retry: %i\n", retries);
1882       if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
1883          func->print();
1884
1885       // spilling to registers may add live ranges, need to rebuild everything
1886       ret = true;
1887       for (sequence = func->cfg.nextSequence(), i = 0;
1888            ret && i <= func->loopNestingBound;
1889            sequence = func->cfg.nextSequence(), ++i)
1890          ret = buildLiveSets(BasicBlock::get(func->cfg.getRoot()));
1891       // reset marker
1892       for (ArrayList::Iterator bi = func->allBBlocks.iterator();
1893            !bi.end(); bi.next())
1894          BasicBlock::get(bi)->liveSet.marker = false;
1895       if (!ret)
1896          break;
1897       func->orderInstructions(this->insns);
1898
1899       ret = buildIntervals.run(func);
1900       if (!ret)
1901          break;
1902       ret = gcra.allocateRegisters(insns);
1903       if (ret)
1904          break; // success
1905    }
1906    INFO_DBG(prog->dbgFlags, REG_ALLOC, "RegAlloc done: %i\n", ret);
1907
1908    func->tlsSize = insertSpills.getStackSize();
1909 out:
1910    return ret;
1911 }
1912
1913 // TODO: check if modifying Instruction::join here breaks anything
1914 void
1915 GCRA::resolveSplitsAndMerges()
1916 {
1917    for (std::list<Instruction *>::iterator it = splits.begin();
1918         it != splits.end();
1919         ++it) {
1920       Instruction *split = *it;
1921       unsigned int reg = regs.idToBytes(split->getSrc(0));
1922       for (int d = 0; split->defExists(d); ++d) {
1923          Value *v = split->getDef(d);
1924          v->reg.data.id = regs.bytesToId(v, reg);
1925          v->join = v;
1926          reg += v->reg.size;
1927       }
1928    }
1929    splits.clear();
1930
1931    for (std::list<Instruction *>::iterator it = merges.begin();
1932         it != merges.end();
1933         ++it) {
1934       Instruction *merge = *it;
1935       unsigned int reg = regs.idToBytes(merge->getDef(0));
1936       for (int s = 0; merge->srcExists(s); ++s) {
1937          Value *v = merge->getSrc(s);
1938          v->reg.data.id = regs.bytesToId(v, reg);
1939          v->join = v;
1940          // If the value is defined by a phi/union node, we also need to
1941          // perform the same fixup on that node's sources, since after RA
1942          // their registers should be identical.
1943          if (v->getInsn()->op == OP_PHI || v->getInsn()->op == OP_UNION) {
1944             Instruction *phi = v->getInsn();
1945             for (int phis = 0; phi->srcExists(phis); ++phis) {
1946                phi->getSrc(phis)->join = v;
1947                phi->getSrc(phis)->reg.data.id = v->reg.data.id;
1948             }
1949          }
1950          reg += v->reg.size;
1951       }
1952    }
1953    merges.clear();
1954 }
1955
1956 bool Program::registerAllocation()
1957 {
1958    RegAlloc ra(this);
1959    return ra.exec();
1960 }
1961
1962 bool
1963 RegAlloc::InsertConstraintsPass::exec(Function *ir)
1964 {
1965    constrList.clear();
1966
1967    bool ret = run(ir, true, true);
1968    if (ret)
1969       ret = insertConstraintMoves();
1970    return ret;
1971 }
1972
1973 // TODO: make part of texture insn
1974 void
1975 RegAlloc::InsertConstraintsPass::textureMask(TexInstruction *tex)
1976 {
1977    Value *def[4];
1978    int c, k, d;
1979    uint8_t mask = 0;
1980
1981    for (d = 0, k = 0, c = 0; c < 4; ++c) {
1982       if (!(tex->tex.mask & (1 << c)))
1983          continue;
1984       if (tex->getDef(k)->refCount()) {
1985          mask |= 1 << c;
1986          def[d++] = tex->getDef(k);
1987       }
1988       ++k;
1989    }
1990    tex->tex.mask = mask;
1991
1992    for (c = 0; c < d; ++c)
1993       tex->setDef(c, def[c]);
1994    for (; c < 4; ++c)
1995       tex->setDef(c, NULL);
1996 }
1997
1998 bool
1999 RegAlloc::InsertConstraintsPass::detectConflict(Instruction *cst, int s)
2000 {
2001    Value *v = cst->getSrc(s);
2002
2003    // current register allocation can't handle it if a value participates in
2004    // multiple constraints
2005    for (Value::UseIterator it = v->uses.begin(); it != v->uses.end(); ++it) {
2006       if (cst != (*it)->getInsn())
2007          return true;
2008    }
2009
2010    // can start at s + 1 because detectConflict is called on all sources
2011    for (int c = s + 1; cst->srcExists(c); ++c)
2012       if (v == cst->getSrc(c))
2013          return true;
2014
2015    Instruction *defi = v->getInsn();
2016
2017    return (!defi || defi->constrainedDefs());
2018 }
2019
2020 void
2021 RegAlloc::InsertConstraintsPass::addConstraint(Instruction *i, int s, int n)
2022 {
2023    Instruction *cst;
2024    int d;
2025
2026    // first, look for an existing identical constraint op
2027    for (std::list<Instruction *>::iterator it = constrList.begin();
2028         it != constrList.end();
2029         ++it) {
2030       cst = (*it);
2031       if (!i->bb->dominatedBy(cst->bb))
2032          break;
2033       for (d = 0; d < n; ++d)
2034          if (cst->getSrc(d) != i->getSrc(d + s))
2035             break;
2036       if (d >= n) {
2037          for (d = 0; d < n; ++d, ++s)
2038             i->setSrc(s, cst->getDef(d));
2039          return;
2040       }
2041    }
2042    cst = new_Instruction(func, OP_CONSTRAINT, i->dType);
2043
2044    for (d = 0; d < n; ++s, ++d) {
2045       cst->setDef(d, new_LValue(func, FILE_GPR));
2046       cst->setSrc(d, i->getSrc(s));
2047       i->setSrc(s, cst->getDef(d));
2048    }
2049    i->bb->insertBefore(i, cst);
2050
2051    constrList.push_back(cst);
2052 }
2053
2054 // Add a dummy use of the pointer source of >= 8 byte loads after the load
2055 // to prevent it from being assigned a register which overlapping the load's
2056 // destination, which would produce random corruptions.
2057 void
2058 RegAlloc::InsertConstraintsPass::addHazard(Instruction *i, const ValueRef *src)
2059 {
2060    Instruction *hzd = new_Instruction(func, OP_NOP, TYPE_NONE);
2061    hzd->setSrc(0, src->get());
2062    i->bb->insertAfter(i, hzd);
2063
2064 }
2065
2066 // b32 { %r0 %r1 %r2 %r3 } -> b128 %r0q
2067 void
2068 RegAlloc::InsertConstraintsPass::condenseDefs(Instruction *insn)
2069 {
2070    int n;
2071    for (n = 0; insn->defExists(n) && insn->def(n).getFile() == FILE_GPR; ++n);
2072    condenseDefs(insn, 0, n - 1);
2073 }
2074
2075 void
2076 RegAlloc::InsertConstraintsPass::condenseDefs(Instruction *insn,
2077                                               const int a, const int b)
2078 {
2079    uint8_t size = 0;
2080    if (a >= b)
2081       return;
2082    for (int s = a; s <= b; ++s)
2083       size += insn->getDef(s)->reg.size;
2084    if (!size)
2085       return;
2086
2087    LValue *lval = new_LValue(func, FILE_GPR);
2088    lval->reg.size = size;
2089
2090    Instruction *split = new_Instruction(func, OP_SPLIT, typeOfSize(size));
2091    split->setSrc(0, lval);
2092    for (int d = a; d <= b; ++d) {
2093       split->setDef(d - a, insn->getDef(d));
2094       insn->setDef(d, NULL);
2095    }
2096    insn->setDef(a, lval);
2097
2098    for (int k = a + 1, d = b + 1; insn->defExists(d); ++d, ++k) {
2099       insn->setDef(k, insn->getDef(d));
2100       insn->setDef(d, NULL);
2101    }
2102    // carry over predicate if any (mainly for OP_UNION uses)
2103    split->setPredicate(insn->cc, insn->getPredicate());
2104
2105    insn->bb->insertAfter(insn, split);
2106    constrList.push_back(split);
2107 }
2108
2109 void
2110 RegAlloc::InsertConstraintsPass::condenseSrcs(Instruction *insn,
2111                                               const int a, const int b)
2112 {
2113    uint8_t size = 0;
2114    if (a >= b)
2115       return;
2116    for (int s = a; s <= b; ++s)
2117       size += insn->getSrc(s)->reg.size;
2118    if (!size)
2119       return;
2120    LValue *lval = new_LValue(func, FILE_GPR);
2121    lval->reg.size = size;
2122
2123    Value *save[3];
2124    insn->takeExtraSources(0, save);
2125
2126    Instruction *merge = new_Instruction(func, OP_MERGE, typeOfSize(size));
2127    merge->setDef(0, lval);
2128    for (int s = a, i = 0; s <= b; ++s, ++i) {
2129       merge->setSrc(i, insn->getSrc(s));
2130    }
2131    insn->moveSources(b + 1, a - b);
2132    insn->setSrc(a, lval);
2133    insn->bb->insertBefore(insn, merge);
2134
2135    insn->putExtraSources(0, save);
2136
2137    constrList.push_back(merge);
2138 }
2139
2140 bool
2141 RegAlloc::InsertConstraintsPass::isScalarTexGM107(TexInstruction *tex)
2142 {
2143    if (tex->tex.sIndirectSrc >= 0 ||
2144        tex->tex.rIndirectSrc >= 0 ||
2145        tex->tex.derivAll)
2146       return false;
2147
2148    if (tex->tex.mask == 5 || tex->tex.mask == 6)
2149       return false;
2150
2151    switch (tex->op) {
2152    case OP_TEX:
2153    case OP_TXF:
2154    case OP_TXG:
2155    case OP_TXL:
2156       break;
2157    default:
2158       return false;
2159    }
2160
2161    // legal variants:
2162    // TEXS.1D.LZ
2163    // TEXS.2D
2164    // TEXS.2D.LZ
2165    // TEXS.2D.LL
2166    // TEXS.2D.DC
2167    // TEXS.2D.LL.DC
2168    // TEXS.2D.LZ.DC
2169    // TEXS.A2D
2170    // TEXS.A2D.LZ
2171    // TEXS.A2D.LZ.DC
2172    // TEXS.3D
2173    // TEXS.3D.LZ
2174    // TEXS.CUBE
2175    // TEXS.CUBE.LL
2176
2177    // TLDS.1D.LZ
2178    // TLDS.1D.LL
2179    // TLDS.2D.LZ
2180    // TLSD.2D.LZ.AOFFI
2181    // TLDS.2D.LZ.MZ
2182    // TLDS.2D.LL
2183    // TLDS.2D.LL.AOFFI
2184    // TLDS.A2D.LZ
2185    // TLDS.3D.LZ
2186
2187    // TLD4S: all 2D/RECT variants and only offset
2188
2189    switch (tex->op) {
2190    case OP_TEX:
2191       if (tex->tex.useOffsets)
2192          return false;
2193
2194       switch (tex->tex.target.getEnum()) {
2195       case TEX_TARGET_1D:
2196       case TEX_TARGET_2D_ARRAY_SHADOW:
2197          return tex->tex.levelZero;
2198       case TEX_TARGET_CUBE:
2199          return !tex->tex.levelZero;
2200       case TEX_TARGET_2D:
2201       case TEX_TARGET_2D_ARRAY:
2202       case TEX_TARGET_2D_SHADOW:
2203       case TEX_TARGET_3D:
2204       case TEX_TARGET_RECT:
2205       case TEX_TARGET_RECT_SHADOW:
2206          return true;
2207       default:
2208          return false;
2209       }
2210
2211    case OP_TXL:
2212       if (tex->tex.useOffsets)
2213          return false;
2214
2215       switch (tex->tex.target.getEnum()) {
2216       case TEX_TARGET_2D:
2217       case TEX_TARGET_2D_SHADOW:
2218       case TEX_TARGET_RECT:
2219       case TEX_TARGET_RECT_SHADOW:
2220       case TEX_TARGET_CUBE:
2221          return true;
2222       default:
2223          return false;
2224       }
2225
2226    case OP_TXF:
2227       switch (tex->tex.target.getEnum()) {
2228       case TEX_TARGET_1D:
2229          return !tex->tex.useOffsets;
2230       case TEX_TARGET_2D:
2231       case TEX_TARGET_RECT:
2232          return true;
2233       case TEX_TARGET_2D_ARRAY:
2234       case TEX_TARGET_2D_MS:
2235       case TEX_TARGET_3D:
2236          return !tex->tex.useOffsets && tex->tex.levelZero;
2237       default:
2238          return false;
2239       }
2240
2241    case OP_TXG:
2242       if (tex->tex.useOffsets > 1)
2243          return false;
2244       if (tex->tex.mask != 0x3 && tex->tex.mask != 0xf)
2245          return false;
2246
2247       switch (tex->tex.target.getEnum()) {
2248       case TEX_TARGET_2D:
2249       case TEX_TARGET_2D_MS:
2250       case TEX_TARGET_2D_SHADOW:
2251       case TEX_TARGET_RECT:
2252       case TEX_TARGET_RECT_SHADOW:
2253          return true;
2254       default:
2255          return false;
2256       }
2257
2258    default:
2259       return false;
2260    }
2261 }
2262
2263 void
2264 RegAlloc::InsertConstraintsPass::handleScalarTexGM107(TexInstruction *tex)
2265 {
2266    int defCount = tex->defCount(0xff);
2267    int srcCount = tex->srcCount(0xff);
2268
2269    tex->tex.scalar = true;
2270
2271    // 1. handle defs
2272    if (defCount > 3)
2273       condenseDefs(tex, 2, 3);
2274    if (defCount > 1)
2275       condenseDefs(tex, 0, 1);
2276
2277    // 2. handle srcs
2278    // special case for TXF.A2D
2279    if (tex->op == OP_TXF && tex->tex.target == TEX_TARGET_2D_ARRAY) {
2280       assert(srcCount >= 3);
2281       condenseSrcs(tex, 1, 2);
2282    } else {
2283       if (srcCount > 3)
2284          condenseSrcs(tex, 2, 3);
2285       // only if we have more than 2 sources
2286       if (srcCount > 2)
2287          condenseSrcs(tex, 0, 1);
2288    }
2289
2290    assert(!tex->defExists(2) && !tex->srcExists(2));
2291 }
2292
2293 void
2294 RegAlloc::InsertConstraintsPass::texConstraintGM107(TexInstruction *tex)
2295 {
2296    int n, s;
2297
2298    if (isTextureOp(tex->op))
2299       textureMask(tex);
2300
2301    if (targ->getChipset() < NVISA_GV100_CHIPSET) {
2302       if (isScalarTexGM107(tex)) {
2303          handleScalarTexGM107(tex);
2304          return;
2305       }
2306
2307       assert(!tex->tex.scalar);
2308       condenseDefs(tex);
2309    } else {
2310       if (isTextureOp(tex->op)) {
2311          int defCount = tex->defCount(0xff);
2312          if (defCount > 3)
2313             condenseDefs(tex, 2, 3);
2314          if (defCount > 1)
2315             condenseDefs(tex, 0, 1);
2316       } else {
2317          condenseDefs(tex);
2318       }
2319    }
2320
2321    if (isSurfaceOp(tex->op)) {
2322       int s = tex->tex.target.getDim() +
2323          (tex->tex.target.isArray() || tex->tex.target.isCube());
2324       int n = 0;
2325
2326       switch (tex->op) {
2327       case OP_SUSTB:
2328       case OP_SUSTP:
2329          n = 4;
2330          break;
2331       case OP_SUREDB:
2332       case OP_SUREDP:
2333          if (tex->subOp == NV50_IR_SUBOP_ATOM_CAS)
2334             n = 2;
2335          break;
2336       default:
2337          break;
2338       }
2339
2340       if (s > 1)
2341          condenseSrcs(tex, 0, s - 1);
2342       if (n > 1)
2343          condenseSrcs(tex, 1, n); // do not condense the tex handle
2344    } else
2345    if (isTextureOp(tex->op)) {
2346       if (tex->op != OP_TXQ) {
2347          s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
2348          if (tex->op == OP_TXD) {
2349             // Indirect handle belongs in the first arg
2350             if (tex->tex.rIndirectSrc >= 0)
2351                s++;
2352             if (!tex->tex.target.isArray() && tex->tex.useOffsets)
2353                s++;
2354          }
2355          n = tex->srcCount(0xff, true) - s;
2356          // TODO: Is this necessary? Perhaps just has to be aligned to the
2357          // level that the first arg is, not necessarily to 4. This
2358          // requirement has not been rigorously verified, as it has been on
2359          // Kepler.
2360          if (n > 0 && n < 3) {
2361             if (tex->srcExists(n + s)) // move potential predicate out of the way
2362                tex->moveSources(n + s, 3 - n);
2363             while (n < 3)
2364                tex->setSrc(s + n++, new_LValue(func, FILE_GPR));
2365          }
2366       } else {
2367          s = tex->srcCount(0xff, true);
2368          n = 0;
2369       }
2370
2371       if (s > 1)
2372          condenseSrcs(tex, 0, s - 1);
2373       if (n > 1) // NOTE: first call modified positions already
2374          condenseSrcs(tex, 1, n);
2375    }
2376 }
2377
2378 void
2379 RegAlloc::InsertConstraintsPass::texConstraintNVE0(TexInstruction *tex)
2380 {
2381    if (isTextureOp(tex->op))
2382       textureMask(tex);
2383    condenseDefs(tex);
2384
2385    if (tex->op == OP_SUSTB || tex->op == OP_SUSTP) {
2386       condenseSrcs(tex, 3, 6);
2387    } else
2388    if (isTextureOp(tex->op)) {
2389       int n = tex->srcCount(0xff, true);
2390       int s = n > 4 ? 4 : n;
2391       if (n > 4 && n < 7) {
2392          if (tex->srcExists(n)) // move potential predicate out of the way
2393             tex->moveSources(n, 7 - n);
2394
2395          while (n < 7)
2396             tex->setSrc(n++, new_LValue(func, FILE_GPR));
2397       }
2398       if (s > 1)
2399          condenseSrcs(tex, 0, s - 1);
2400       if (n > 4)
2401          condenseSrcs(tex, 1, n - s);
2402    }
2403 }
2404
2405 void
2406 RegAlloc::InsertConstraintsPass::texConstraintNVC0(TexInstruction *tex)
2407 {
2408    int n, s;
2409
2410    if (isTextureOp(tex->op))
2411       textureMask(tex);
2412
2413    if (tex->op == OP_TXQ) {
2414       s = tex->srcCount(0xff);
2415       n = 0;
2416    } else if (isSurfaceOp(tex->op)) {
2417       s = tex->tex.target.getDim() + (tex->tex.target.isArray() || tex->tex.target.isCube());
2418       if (tex->op == OP_SUSTB || tex->op == OP_SUSTP)
2419          n = 4;
2420       else
2421          n = 0;
2422    } else {
2423       s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
2424       if (!tex->tex.target.isArray() &&
2425           (tex->tex.rIndirectSrc >= 0 || tex->tex.sIndirectSrc >= 0))
2426          ++s;
2427       if (tex->op == OP_TXD && tex->tex.useOffsets)
2428          ++s;
2429       n = tex->srcCount(0xff) - s;
2430       assert(n <= 4);
2431    }
2432
2433    if (s > 1)
2434       condenseSrcs(tex, 0, s - 1);
2435    if (n > 1) // NOTE: first call modified positions already
2436       condenseSrcs(tex, 1, n);
2437
2438    condenseDefs(tex);
2439 }
2440
2441 void
2442 RegAlloc::InsertConstraintsPass::texConstraintNV50(TexInstruction *tex)
2443 {
2444    Value *pred = tex->getPredicate();
2445    if (pred)
2446       tex->setPredicate(tex->cc, NULL);
2447
2448    textureMask(tex);
2449
2450    assert(tex->defExists(0) && tex->srcExists(0));
2451    // make src and def count match
2452    int c;
2453    for (c = 0; tex->srcExists(c) || tex->defExists(c); ++c) {
2454       if (!tex->srcExists(c))
2455          tex->setSrc(c, new_LValue(func, tex->getSrc(0)->asLValue()));
2456       else
2457          insertConstraintMove(tex, c);
2458       if (!tex->defExists(c))
2459          tex->setDef(c, new_LValue(func, tex->getDef(0)->asLValue()));
2460    }
2461    if (pred)
2462       tex->setPredicate(tex->cc, pred);
2463    condenseDefs(tex);
2464    condenseSrcs(tex, 0, c - 1);
2465 }
2466
2467 // Insert constraint markers for instructions whose multiple sources must be
2468 // located in consecutive registers.
2469 bool
2470 RegAlloc::InsertConstraintsPass::visit(BasicBlock *bb)
2471 {
2472    TexInstruction *tex;
2473    Instruction *next;
2474    int s, size;
2475
2476    targ = bb->getProgram()->getTarget();
2477
2478    for (Instruction *i = bb->getEntry(); i; i = next) {
2479       next = i->next;
2480
2481       if ((tex = i->asTex())) {
2482          switch (targ->getChipset() & ~0xf) {
2483          case 0x50:
2484          case 0x80:
2485          case 0x90:
2486          case 0xa0:
2487             texConstraintNV50(tex);
2488             break;
2489          case 0xc0:
2490          case 0xd0:
2491             texConstraintNVC0(tex);
2492             break;
2493          case 0xe0:
2494          case 0xf0:
2495          case 0x100:
2496             texConstraintNVE0(tex);
2497             break;
2498          case 0x110:
2499          case 0x120:
2500          case 0x130:
2501          case 0x140:
2502             texConstraintGM107(tex);
2503             break;
2504          default:
2505             break;
2506          }
2507       } else
2508       if (i->op == OP_EXPORT || i->op == OP_STORE) {
2509          for (size = typeSizeof(i->dType), s = 1; size > 0; ++s) {
2510             assert(i->srcExists(s));
2511             size -= i->getSrc(s)->reg.size;
2512          }
2513          condenseSrcs(i, 1, s - 1);
2514       } else
2515       if (i->op == OP_LOAD || i->op == OP_VFETCH) {
2516          condenseDefs(i);
2517          if (i->src(0).isIndirect(0) && typeSizeof(i->dType) >= 8)
2518             addHazard(i, i->src(0).getIndirect(0));
2519          if (i->src(0).isIndirect(1) && typeSizeof(i->dType) >= 8)
2520             addHazard(i, i->src(0).getIndirect(1));
2521       } else
2522       if (i->op == OP_UNION ||
2523           i->op == OP_MERGE ||
2524           i->op == OP_SPLIT) {
2525          constrList.push_back(i);
2526       }
2527    }
2528    return true;
2529 }
2530
2531 void
2532 RegAlloc::InsertConstraintsPass::insertConstraintMove(Instruction *cst, int s)
2533 {
2534    const uint8_t size = cst->src(s).getSize();
2535
2536    assert(cst->getSrc(s)->defs.size() == 1); // still SSA
2537
2538    Instruction *defi = cst->getSrc(s)->defs.front()->getInsn();
2539
2540    bool imm = defi->op == OP_MOV &&
2541       defi->src(0).getFile() == FILE_IMMEDIATE;
2542    bool load = defi->op == OP_LOAD &&
2543       defi->src(0).getFile() == FILE_MEMORY_CONST &&
2544       !defi->src(0).isIndirect(0);
2545    // catch some cases where don't really need MOVs
2546    if (cst->getSrc(s)->refCount() == 1 && !defi->constrainedDefs()) {
2547       if (imm || load) {
2548          // Move the defi right before the cst. No point in expanding
2549          // the range.
2550          defi->bb->remove(defi);
2551          cst->bb->insertBefore(cst, defi);
2552       }
2553       return;
2554    }
2555
2556    LValue *lval = new_LValue(func, cst->src(s).getFile());
2557    lval->reg.size = size;
2558
2559    Instruction *mov = new_Instruction(func, OP_MOV, typeOfSize(size));
2560    mov->setDef(0, lval);
2561    mov->setSrc(0, cst->getSrc(s));
2562
2563    if (load) {
2564       mov->op = OP_LOAD;
2565       mov->setSrc(0, defi->getSrc(0));
2566    } else if (imm) {
2567       mov->setSrc(0, defi->getSrc(0));
2568    }
2569
2570    if (defi->getPredicate())
2571       mov->setPredicate(defi->cc, defi->getPredicate());
2572
2573    cst->setSrc(s, mov->getDef(0));
2574    cst->bb->insertBefore(cst, mov);
2575
2576    cst->getDef(0)->asLValue()->noSpill = 1; // doesn't help
2577 }
2578
2579 // Insert extra moves so that, if multiple register constraints on a value are
2580 // in conflict, these conflicts can be resolved.
2581 bool
2582 RegAlloc::InsertConstraintsPass::insertConstraintMoves()
2583 {
2584    for (std::list<Instruction *>::iterator it = constrList.begin();
2585         it != constrList.end();
2586         ++it) {
2587       Instruction *cst = *it;
2588       Instruction *mov;
2589
2590       if (cst->op == OP_SPLIT && 0) {
2591          // spilling splits is annoying, just make sure they're separate
2592          for (int d = 0; cst->defExists(d); ++d) {
2593             if (!cst->getDef(d)->refCount())
2594                continue;
2595             LValue *lval = new_LValue(func, cst->def(d).getFile());
2596             const uint8_t size = cst->def(d).getSize();
2597             lval->reg.size = size;
2598
2599             mov = new_Instruction(func, OP_MOV, typeOfSize(size));
2600             mov->setSrc(0, lval);
2601             mov->setDef(0, cst->getDef(d));
2602             cst->setDef(d, mov->getSrc(0));
2603             cst->bb->insertAfter(cst, mov);
2604
2605             cst->getSrc(0)->asLValue()->noSpill = 1;
2606             mov->getSrc(0)->asLValue()->noSpill = 1;
2607          }
2608       } else
2609       if (cst->op == OP_MERGE || cst->op == OP_UNION) {
2610          for (int s = 0; cst->srcExists(s); ++s) {
2611             const uint8_t size = cst->src(s).getSize();
2612
2613             if (!cst->getSrc(s)->defs.size()) {
2614                mov = new_Instruction(func, OP_NOP, typeOfSize(size));
2615                mov->setDef(0, cst->getSrc(s));
2616                cst->bb->insertBefore(cst, mov);
2617                continue;
2618             }
2619
2620             insertConstraintMove(cst, s);
2621          }
2622       }
2623    }
2624
2625    return true;
2626 }
2627
2628 } // namespace nv50_ir