src/gallium/drivers/nouveau/codegen/nv50_ir.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 #include "codegen/nv50_ir_driver.h"
  26
  27 extern "C" {
  28 #include "nouveau_debug.h"
  29 #include "nv50/nv50_program.h"
  30 }
  31
  32 namespace nv50_ir {
  33
  34 Modifier::Modifier(operation op)
  35 {
  36    switch (op) {
  37    case OP_NEG: bits = NV50_IR_MOD_NEG; break;
  38    case OP_ABS: bits = NV50_IR_MOD_ABS; break;
  39    case OP_SAT: bits = NV50_IR_MOD_SAT; break;
  40    case OP_NOT: bits = NV50_IR_MOD_NOT; break;
  41    default:
  42       bits = 0;
  43       break;
  44    }
  45 }
  46
  47 Modifier Modifier::operator*(const Modifier m) const
  48 {
  49    unsigned int a, b, c;
  50
  51    b = m.bits;
  52    if (this->bits & NV50_IR_MOD_ABS)
  53       b &= ~NV50_IR_MOD_NEG;
  54
  55    a = (this->bits ^ b)      & (NV50_IR_MOD_NOT | NV50_IR_MOD_NEG);
  56    c = (this->bits | m.bits) & (NV50_IR_MOD_ABS | NV50_IR_MOD_SAT);
  57
  58    return Modifier(a | c);
  59 }
  60
  61 ValueRef::ValueRef(Value *v) : value(NULL), insn(NULL)
  62 {
  63    indirect[0] = -1;
  64    indirect[1] = -1;
  65    usedAsPtr = false;
  66    set(v);
  67 }
  68
  69 ValueRef::ValueRef(const ValueRef& ref) : value(NULL), insn(ref.insn)
  70 {
  71    set(ref);
  72    usedAsPtr = ref.usedAsPtr;
  73 }
  74
  75 ValueRef::~ValueRef()
  76 {
  77    this->set(NULL);
  78 }
  79
  80 bool ValueRef::getImmediate(ImmediateValue &imm) const
  81 {
  82    const ValueRef *src = this;
  83    Modifier m;
  84    DataType type = src->insn->sType;
  85
  86    while (src) {
  87       if (src->mod) {
  88          if (src->insn->sType != type)
  89             break;
  90          m *= src->mod;
  91       }
  92       if (src->getFile() == FILE_IMMEDIATE) {
  93          imm = *(src->value->asImm());
  94          // The immediate's type isn't required to match its use, it's
  95          // more of a hint; applying a modifier makes use of that hint.
  96          imm.reg.type = type;
  97          m.applyTo(imm);
  98          return true;
  99       }
 100
 101       Instruction *insn = src->value->getUniqueInsn();
 102
 103       if (insn && insn->op == OP_MOV) {
 104          src = &insn->src(0);
 105          if (src->mod)
 106             WARN("OP_MOV with modifier encountered !\n");
 107       } else {
 108          src = NULL;
 109       }
 110    }
 111    return false;
 112 }
 113
 114 ValueDef::ValueDef(Value *v) : value(NULL), insn(NULL)
 115 {
 116    set(v);
 117 }
 118
 119 ValueDef::ValueDef(const ValueDef& def) : value(NULL), insn(NULL)
 120 {
 121    set(def.get());
 122 }
 123
 124 ValueDef::~ValueDef()
 125 {
 126    this->set(NULL);
 127 }
 128
 129 void
 130 ValueRef::set(const ValueRef &ref)
 131 {
 132    this->set(ref.get());
 133    mod = ref.mod;
 134    indirect[0] = ref.indirect[0];
 135    indirect[1] = ref.indirect[1];
 136 }
 137
 138 void
 139 ValueRef::set(Value *refVal)
 140 {
 141    if (value == refVal)
 142       return;
 143    if (value)
 144       value->uses.erase(this);
 145    if (refVal)
 146       refVal->uses.insert(this);
 147
 148    value = refVal;
 149 }
 150
 151 void
 152 ValueDef::set(Value *defVal)
 153 {
 154    if (value == defVal)
 155       return;
 156    if (value)
 157       value->defs.remove(this);
 158    if (defVal)
 159       defVal->defs.push_back(this);
 160
 161    value = defVal;
 162 }
 163
 164 // Check if we can replace this definition's value by the value in @rep,
 165 // including the source modifiers, i.e. make sure that all uses support
 166 // @rep.mod.
 167 bool
 168 ValueDef::mayReplace(const ValueRef &rep)
 169 {
 170    if (!rep.mod)
 171       return true;
 172
 173    if (!insn || !insn->bb) // Unbound instruction ?
 174       return false;
 175
 176    const Target *target = insn->bb->getProgram()->getTarget();
 177
 178    for (Value::UseIterator it = value->uses.begin(); it != value->uses.end();
 179         ++it) {
 180       Instruction *insn = (*it)->getInsn();
 181       int s = -1;
 182
 183       for (int i = 0; insn->srcExists(i); ++i) {
 184          if (insn->src(i).get() == value) {
 185             // If there are multiple references to us we'd have to check if the
 186             // combination of mods is still supported, but just bail for now.
 187             if (&insn->src(i) != (*it))
 188                return false;
 189             s = i;
 190          }
 191       }
 192       assert(s >= 0); // integrity of uses list
 193
 194       if (!target->isModSupported(insn, s, rep.mod))
 195          return false;
 196    }
 197    return true;
 198 }
 199
 200 void
 201 ValueDef::replace(const ValueRef &repVal, bool doSet)
 202 {
 203    assert(mayReplace(repVal));
 204
 205    if (value == repVal.get())
 206       return;
 207
 208    while (!value->uses.empty()) {
 209       ValueRef *ref = *value->uses.begin();
 210       ref->set(repVal.get());
 211       ref->mod *= repVal.mod;
 212    }
 213
 214    if (doSet)
 215       set(repVal.get());
 216 }
 217
 218 Value::Value()
 219 {
 220   join = this;
 221   memset(&reg, 0, sizeof(reg));
 222   reg.size = 4;
 223 }
 224
 225 LValue::LValue(Function *fn, DataFile file)
 226 {
 227    reg.file = file;
 228    reg.size = (file != FILE_PREDICATE) ? 4 : 1;
 229    reg.data.id = -1;
 230
 231    compMask = 0;
 232    compound = 0;
 233    ssa = 0;
 234    fixedReg = 0;
 235    noSpill = 0;
 236
 237    fn->add(this, this->id);
 238 }
 239
 240 LValue::LValue(Function *fn, LValue *lval)
 241 {
 242    assert(lval);
 243
 244    reg.file = lval->reg.file;
 245    reg.size = lval->reg.size;
 246    reg.data.id = -1;
 247
 248    compMask = 0;
 249    compound = 0;
 250    ssa = 0;
 251    fixedReg = 0;
 252    noSpill = 0;
 253
 254    fn->add(this, this->id);
 255 }
 256
 257 LValue *
 258 LValue::clone(ClonePolicy<Function>& pol) const
 259 {
 260    LValue *that = new_LValue(pol.context(), reg.file);
 261
 262    pol.set<Value>(this, that);
 263
 264    that->reg.size = this->reg.size;
 265    that->reg.type = this->reg.type;
 266    that->reg.data = this->reg.data;
 267
 268    return that;
 269 }
 270
 271 bool
 272 LValue::isUniform() const
 273 {
 274    if (defs.size() > 1)
 275       return false;
 276    Instruction *insn = getInsn();
 277    // let's not try too hard here for now ...
 278    return !insn->srcExists(1) && insn->getSrc(0)->isUniform();
 279 }
 280
 281 Symbol::Symbol(Program *prog, DataFile f, ubyte fidx)
 282 {
 283    baseSym = NULL;
 284
 285    reg.file = f;
 286    reg.fileIndex = fidx;
 287    reg.data.offset = 0;
 288
 289    prog->add(this, this->id);
 290 }
 291
 292 Symbol *
 293 Symbol::clone(ClonePolicy<Function>& pol) const
 294 {
 295    Program *prog = pol.context()->getProgram();
 296
 297    Symbol *that = new_Symbol(prog, reg.file, reg.fileIndex);
 298
 299    pol.set<Value>(this, that);
 300
 301    that->reg.size = this->reg.size;
 302    that->reg.type = this->reg.type;
 303    that->reg.data = this->reg.data;
 304
 305    that->baseSym = this->baseSym;
 306
 307    return that;
 308 }
 309
 310 bool
 311 Symbol::isUniform() const
 312 {
 313    return
 314       reg.file != FILE_SYSTEM_VALUE &&
 315       reg.file != FILE_MEMORY_LOCAL &&
 316       reg.file != FILE_SHADER_INPUT;
 317 }
 318
 319 ImmediateValue::ImmediateValue(Program *prog, uint32_t uval)
 320 {
 321    memset(&reg, 0, sizeof(reg));
 322
 323    reg.file = FILE_IMMEDIATE;
 324    reg.size = 4;
 325    reg.type = TYPE_U32;
 326
 327    reg.data.u32 = uval;
 328
 329    prog->add(this, this->id);
 330 }
 331
 332 ImmediateValue::ImmediateValue(Program *prog, float fval)
 333 {
 334    memset(&reg, 0, sizeof(reg));
 335
 336    reg.file = FILE_IMMEDIATE;
 337    reg.size = 4;
 338    reg.type = TYPE_F32;
 339
 340    reg.data.f32 = fval;
 341
 342    prog->add(this, this->id);
 343 }
 344
 345 ImmediateValue::ImmediateValue(Program *prog, double dval)
 346 {
 347    memset(&reg, 0, sizeof(reg));
 348
 349    reg.file = FILE_IMMEDIATE;
 350    reg.size = 8;
 351    reg.type = TYPE_F64;
 352
 353    reg.data.f64 = dval;
 354
 355    prog->add(this, this->id);
 356 }
 357
 358 ImmediateValue::ImmediateValue(const ImmediateValue *proto, DataType ty)
 359 {
 360    reg = proto->reg;
 361
 362    reg.type = ty;
 363    reg.size = typeSizeof(ty);
 364 }
 365
 366 ImmediateValue *
 367 ImmediateValue::clone(ClonePolicy<Function>& pol) const
 368 {
 369    Program *prog = pol.context()->getProgram();
 370    ImmediateValue *that = new_ImmediateValue(prog, 0u);
 371
 372    pol.set<Value>(this, that);
 373
 374    that->reg.size = this->reg.size;
 375    that->reg.type = this->reg.type;
 376    that->reg.data = this->reg.data;
 377
 378    return that;
 379 }
 380
 381 bool
 382 ImmediateValue::isInteger(const int i) const
 383 {
 384    switch (reg.type) {
 385    case TYPE_S8:
 386       return reg.data.s8 == i;
 387    case TYPE_U8:
 388       return reg.data.u8 == i;
 389    case TYPE_S16:
 390       return reg.data.s16 == i;
 391    case TYPE_U16:
 392       return reg.data.u16 == i;
 393    case TYPE_S32:
 394    case TYPE_U32:
 395       return reg.data.s32 == i; // as if ...
 396    case TYPE_S64:
 397    case TYPE_U64:
 398       return reg.data.s64 == i; // as if ...
 399    case TYPE_F32:
 400       return reg.data.f32 == static_cast<float>(i);
 401    case TYPE_F64:
 402       return reg.data.f64 == static_cast<double>(i);
 403    default:
 404       return false;
 405    }
 406 }
 407
 408 bool
 409 ImmediateValue::isNegative() const
 410 {
 411    switch (reg.type) {
 412    case TYPE_S8:  return reg.data.s8 < 0;
 413    case TYPE_S16: return reg.data.s16 < 0;
 414    case TYPE_S32:
 415    case TYPE_U32: return reg.data.s32 < 0;
 416    case TYPE_F32: return reg.data.u32 & (1 << 31);
 417    case TYPE_F64: return reg.data.u64 & (1ULL << 63);
 418    default:
 419       return false;
 420    }
 421 }
 422
 423 bool
 424 ImmediateValue::isPow2() const
 425 {
 426    switch (reg.type) {
 427    case TYPE_U8:
 428    case TYPE_U16:
 429    case TYPE_U32: return util_is_power_of_two(reg.data.u32);
 430    default:
 431       return false;
 432    }
 433 }
 434
 435 void
 436 ImmediateValue::applyLog2()
 437 {
 438    switch (reg.type) {
 439    case TYPE_S8:
 440    case TYPE_S16:
 441    case TYPE_S32:
 442       assert(!this->isNegative());
 443       // fall through
 444    case TYPE_U8:
 445    case TYPE_U16:
 446    case TYPE_U32:
 447       reg.data.u32 = util_logbase2(reg.data.u32);
 448       break;
 449    case TYPE_F32:
 450       reg.data.f32 = log2f(reg.data.f32);
 451       break;
 452    case TYPE_F64:
 453       reg.data.f64 = log2(reg.data.f64);
 454       break;
 455    default:
 456       assert(0);
 457       break;
 458    }
 459 }
 460
 461 bool
 462 ImmediateValue::compare(CondCode cc, float fval) const
 463 {
 464    if (reg.type != TYPE_F32)
 465       ERROR("immediate value is not of type f32");
 466
 467    switch (static_cast<CondCode>(cc & 7)) {
 468    case CC_TR: return true;
 469    case CC_FL: return false;
 470    case CC_LT: return reg.data.f32 <  fval;
 471    case CC_LE: return reg.data.f32 <= fval;
 472    case CC_GT: return reg.data.f32 >  fval;
 473    case CC_GE: return reg.data.f32 >= fval;
 474    case CC_EQ: return reg.data.f32 == fval;
 475    case CC_NE: return reg.data.f32 != fval;
 476    default:
 477       assert(0);
 478       return false;
 479    }
 480 }
 481
 482 ImmediateValue&
 483 ImmediateValue::operator=(const ImmediateValue &that)
 484 {
 485    this->reg = that.reg;
 486    return (*this);
 487 }
 488
 489 bool
 490 Value::interfers(const Value *that) const
 491 {
 492    uint32_t idA, idB;
 493
 494    if (that->reg.file != reg.file || that->reg.fileIndex != reg.fileIndex)
 495       return false;
 496    if (this->asImm())
 497       return false;
 498
 499    if (this->asSym()) {
 500       idA = this->join->reg.data.offset;
 501       idB = that->join->reg.data.offset;
 502    } else {
 503       idA = this->join->reg.data.id * MIN2(this->reg.size, 4);
 504       idB = that->join->reg.data.id * MIN2(that->reg.size, 4);
 505    }
 506
 507    if (idA < idB)
 508       return (idA + this->reg.size > idB);
 509    else
 510    if (idA > idB)
 511       return (idB + that->reg.size > idA);
 512    else
 513       return (idA == idB);
 514 }
 515
 516 bool
 517 Value::equals(const Value *that, bool strict) const
 518 {
 519    if (strict)
 520       return this == that;
 521
 522    if (that->reg.file != reg.file || that->reg.fileIndex != reg.fileIndex)
 523       return false;
 524    if (that->reg.size != this->reg.size)
 525       return false;
 526
 527    if (that->reg.data.id != this->reg.data.id)
 528       return false;
 529
 530    return true;
 531 }
 532
 533 bool
 534 ImmediateValue::equals(const Value *that, bool strict) const
 535 {
 536    const ImmediateValue *imm = that->asImm();
 537    if (!imm)
 538       return false;
 539    return reg.data.u64 == imm->reg.data.u64;
 540 }
 541
 542 bool
 543 Symbol::equals(const Value *that, bool strict) const
 544 {
 545    if (reg.file != that->reg.file || reg.fileIndex != that->reg.fileIndex)
 546       return false;
 547    assert(that->asSym());
 548
 549    if (this->baseSym != that->asSym()->baseSym)
 550       return false;
 551
 552    if (reg.file == FILE_SYSTEM_VALUE)
 553       return (this->reg.data.sv.sv    == that->reg.data.sv.sv &&
 554               this->reg.data.sv.index == that->reg.data.sv.index);
 555    return this->reg.data.offset == that->reg.data.offset;
 556 }
 557
 558 void Instruction::init()
 559 {
 560    next = prev = 0;
 561
 562    cc = CC_ALWAYS;
 563    rnd = ROUND_N;
 564    cache = CACHE_CA;
 565    subOp = 0;
 566
 567    saturate = 0;
 568    join = 0;
 569    exit = 0;
 570    terminator = 0;
 571    ftz = 0;
 572    dnz = 0;
 573    perPatch = 0;
 574    fixed = 0;
 575    encSize = 0;
 576    ipa = 0;
 577    mask = 0;
 578
 579    lanes = 0xf;
 580
 581    postFactor = 0;
 582
 583    predSrc = -1;
 584    flagsDef = -1;
 585    flagsSrc = -1;
 586 }
 587
 588 Instruction::Instruction()
 589 {
 590    init();
 591
 592    op = OP_NOP;
 593    dType = sType = TYPE_F32;
 594
 595    id = -1;
 596    bb = 0;
 597 }
 598
 599 Instruction::Instruction(Function *fn, operation opr, DataType ty)
 600 {
 601    init();
 602
 603    op = opr;
 604    dType = sType = ty;
 605
 606    fn->add(this, id);
 607 }
 608
 609 Instruction::~Instruction()
 610 {
 611    if (bb) {
 612       Function *fn = bb->getFunction();
 613       bb->remove(this);
 614       fn->allInsns.remove(id);
 615    }
 616
 617    for (int s = 0; srcExists(s); ++s)
 618       setSrc(s, NULL);
 619    // must unlink defs too since the list pointers will get deallocated
 620    for (int d = 0; defExists(d); ++d)
 621       setDef(d, NULL);
 622 }
 623
 624 void
 625 Instruction::setDef(int i, Value *val)
 626 {
 627    int size = defs.size();
 628    if (i >= size) {
 629       defs.resize(i + 1);
 630       while (size <= i)
 631          defs[size++].setInsn(this);
 632    }
 633    defs[i].set(val);
 634 }
 635
 636 void
 637 Instruction::setSrc(int s, Value *val)
 638 {
 639    int size = srcs.size();
 640    if (s >= size) {
 641       srcs.resize(s + 1);
 642       while (size <= s)
 643          srcs[size++].setInsn(this);
 644    }
 645    srcs[s].set(val);
 646 }
 647
 648 void
 649 Instruction::setSrc(int s, const ValueRef& ref)
 650 {
 651    setSrc(s, ref.get());
 652    srcs[s].mod = ref.mod;
 653 }
 654
 655 void
 656 Instruction::swapSources(int a, int b)
 657 {
 658    Value *value = srcs[a].get();
 659    Modifier m = srcs[a].mod;
 660
 661    setSrc(a, srcs[b]);
 662
 663    srcs[b].set(value);
 664    srcs[b].mod = m;
 665 }
 666
 667 static inline void moveSourcesAdjustIndex(int8_t &index, int s, int delta)
 668 {
 669    if (index >= s)
 670       index += delta;
 671    else
 672    if ((delta < 0) && (index >= (s + delta)))
 673       index = -1;
 674 }
 675
 676 // Moves sources [@s,last_source] by @delta.
 677 // If @delta < 0, sources [@s - abs(@delta), @s) are erased.
 678 void
 679 Instruction::moveSources(const int s, const int delta)
 680 {
 681    if (delta == 0)
 682       return;
 683    assert(s + delta >= 0);
 684
 685    int k;
 686
 687    for (k = 0; srcExists(k); ++k) {
 688       for (int i = 0; i < 2; ++i)
 689          moveSourcesAdjustIndex(src(k).indirect[i], s, delta);
 690    }
 691    moveSourcesAdjustIndex(predSrc, s, delta);
 692    moveSourcesAdjustIndex(flagsSrc, s, delta);
 693    if (asTex()) {
 694       TexInstruction *tex = asTex();
 695       moveSourcesAdjustIndex(tex->tex.rIndirectSrc, s, delta);
 696       moveSourcesAdjustIndex(tex->tex.sIndirectSrc, s, delta);
 697    }
 698
 699    if (delta > 0) {
 700       --k;
 701       for (int p = k + delta; k >= s; --k, --p)
 702          setSrc(p, src(k));
 703    } else {
 704       int p;
 705       for (p = s; p < k; ++p)
 706          setSrc(p + delta, src(p));
 707       for (; (p + delta) < k; ++p)
 708          setSrc(p + delta, NULL);
 709    }
 710 }
 711
 712 void
 713 Instruction::takeExtraSources(int s, Value *values[3])
 714 {
 715    values[0] = getIndirect(s, 0);
 716    if (values[0])
 717       setIndirect(s, 0, NULL);
 718
 719    values[1] = getIndirect(s, 1);
 720    if (values[1])
 721       setIndirect(s, 1, NULL);
 722
 723    values[2] = getPredicate();
 724    if (values[2])
 725       setPredicate(cc, NULL);
 726 }
 727
 728 void
 729 Instruction::putExtraSources(int s, Value *values[3])
 730 {
 731    if (values[0])
 732       setIndirect(s, 0, values[0]);
 733    if (values[1])
 734       setIndirect(s, 1, values[1]);
 735    if (values[2])
 736       setPredicate(cc, values[2]);
 737 }
 738
 739 Instruction *
 740 Instruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 741 {
 742    if (!i)
 743       i = new_Instruction(pol.context(), op, dType);
 744 #ifndef NDEBUG // non-conformant assert, so this is required
 745    assert(typeid(*i) == typeid(*this));
 746 #endif
 747
 748    pol.set<Instruction>(this, i);
 749
 750    i->sType = sType;
 751
 752    i->rnd = rnd;
 753    i->cache = cache;
 754    i->subOp = subOp;
 755
 756    i->saturate = saturate;
 757    i->join = join;
 758    i->exit = exit;
 759    i->mask = mask;
 760    i->ftz = ftz;
 761    i->dnz = dnz;
 762    i->ipa = ipa;
 763    i->lanes = lanes;
 764    i->perPatch = perPatch;
 765
 766    i->postFactor = postFactor;
 767
 768    for (int d = 0; defExists(d); ++d)
 769       i->setDef(d, pol.get(getDef(d)));
 770
 771    for (int s = 0; srcExists(s); ++s) {
 772       i->setSrc(s, pol.get(getSrc(s)));
 773       i->src(s).mod = src(s).mod;
 774    }
 775
 776    i->cc = cc;
 777    i->predSrc = predSrc;
 778    i->flagsDef = flagsDef;
 779    i->flagsSrc = flagsSrc;
 780
 781    return i;
 782 }
 783
 784 unsigned int
 785 Instruction::defCount(unsigned int mask, bool singleFile) const
 786 {
 787    unsigned int i, n;
 788
 789    if (singleFile) {
 790       unsigned int d = ffs(mask);
 791       if (!d)
 792          return 0;
 793       for (i = d--; defExists(i); ++i)
 794          if (getDef(i)->reg.file != getDef(d)->reg.file)
 795             mask &= ~(1 << i);
 796    }
 797
 798    for (n = 0, i = 0; this->defExists(i); ++i, mask >>= 1)
 799       n += mask & 1;
 800    return n;
 801 }
 802
 803 unsigned int
 804 Instruction::srcCount(unsigned int mask, bool singleFile) const
 805 {
 806    unsigned int i, n;
 807
 808    if (singleFile) {
 809       unsigned int s = ffs(mask);
 810       if (!s)
 811          return 0;
 812       for (i = s--; srcExists(i); ++i)
 813          if (getSrc(i)->reg.file != getSrc(s)->reg.file)
 814             mask &= ~(1 << i);
 815    }
 816
 817    for (n = 0, i = 0; this->srcExists(i); ++i, mask >>= 1)
 818       n += mask & 1;
 819    return n;
 820 }
 821
 822 bool
 823 Instruction::setIndirect(int s, int dim, Value *value)
 824 {
 825    assert(this->srcExists(s));
 826
 827    int p = srcs[s].indirect[dim];
 828    if (p < 0) {
 829       if (!value)
 830          return true;
 831       p = srcs.size();
 832       while (p > 0 && !srcExists(p - 1))
 833          --p;
 834    }
 835    setSrc(p, value);
 836    srcs[p].usedAsPtr = (value != 0);
 837    srcs[s].indirect[dim] = value ? p : -1;
 838    return true;
 839 }
 840
 841 bool
 842 Instruction::setPredicate(CondCode ccode, Value *value)
 843 {
 844    cc = ccode;
 845
 846    if (!value) {
 847       if (predSrc >= 0) {
 848          srcs[predSrc].set(NULL);
 849          predSrc = -1;
 850       }
 851       return true;
 852    }
 853
 854    if (predSrc < 0) {
 855       predSrc = srcs.size();
 856       while (predSrc > 0 && !srcExists(predSrc - 1))
 857          --predSrc;
 858    }
 859
 860    setSrc(predSrc, value);
 861    return true;
 862 }
 863
 864 bool
 865 Instruction::writesPredicate() const
 866 {
 867    for (int d = 0; defExists(d); ++d)
 868       if (getDef(d)->inFile(FILE_PREDICATE) || getDef(d)->inFile(FILE_FLAGS))
 869          return true;
 870    return false;
 871 }
 872
 873 bool
 874 Instruction::canCommuteDefSrc(const Instruction *i) const
 875 {
 876    for (int d = 0; defExists(d); ++d)
 877       for (int s = 0; i->srcExists(s); ++s)
 878          if (getDef(d)->interfers(i->getSrc(s)))
 879             return false;
 880    return true;
 881 }
 882
 883 bool
 884 Instruction::canCommuteDefDef(const Instruction *i) const
 885 {
 886    for (int d = 0; defExists(d); ++d)
 887       for (int c = 0; i->defExists(c); ++c)
 888          if (getDef(d)->interfers(i->getDef(c)))
 889             return false;
 890    return true;
 891 }
 892
 893 bool
 894 Instruction::isCommutationLegal(const Instruction *i) const
 895 {
 896    return canCommuteDefDef(i) &&
 897       canCommuteDefSrc(i) &&
 898       i->canCommuteDefSrc(this);
 899 }
 900
 901 TexInstruction::TexInstruction(Function *fn, operation op)
 902    : Instruction(fn, op, TYPE_F32)
 903 {
 904    memset(&tex, 0, sizeof(tex));
 905
 906    tex.rIndirectSrc = -1;
 907    tex.sIndirectSrc = -1;
 908 }
 909
 910 TexInstruction::~TexInstruction()
 911 {
 912    for (int c = 0; c < 3; ++c) {
 913       dPdx[c].set(NULL);
 914       dPdy[c].set(NULL);
 915    }
 916    for (int n = 0; n < 4; ++n)
 917       for (int c = 0; c < 3; ++c)
 918          offset[n][c].set(NULL);
 919 }
 920
 921 TexInstruction *
 922 TexInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 923 {
 924    TexInstruction *tex = (i ? static_cast<TexInstruction *>(i) :
 925                           new_TexInstruction(pol.context(), op));
 926
 927    Instruction::clone(pol, tex);
 928
 929    tex->tex = this->tex;
 930
 931    if (op == OP_TXD) {
 932       for (unsigned int c = 0; c < tex->tex.target.getDim(); ++c) {
 933          tex->dPdx[c].set(dPdx[c]);
 934          tex->dPdy[c].set(dPdy[c]);
 935       }
 936    }
 937
 938    for (int n = 0; n < tex->tex.useOffsets; ++n)
 939       for (int c = 0; c < 3; ++c)
 940          tex->offset[n][c].set(offset[n][c]);
 941
 942    return tex;
 943 }
 944
 945 const struct TexInstruction::Target::Desc TexInstruction::Target::descTable[] =
 946 {
 947    { "1D",                1, 1, false, false, false },
 948    { "2D",                2, 2, false, false, false },
 949    { "2D_MS",             2, 3, false, false, false },
 950    { "3D",                3, 3, false, false, false },
 951    { "CUBE",              2, 3, false, true,  false },
 952    { "1D_SHADOW",         1, 1, false, false, true  },
 953    { "2D_SHADOW",         2, 2, false, false, true  },
 954    { "CUBE_SHADOW",       2, 3, false, true,  true  },
 955    { "1D_ARRAY",          1, 2, true,  false, false },
 956    { "2D_ARRAY",          2, 3, true,  false, false },
 957    { "2D_MS_ARRAY",       2, 4, true,  false, false },
 958    { "CUBE_ARRAY",        2, 4, true,  true,  false },
 959    { "1D_ARRAY_SHADOW",   1, 2, true,  false, true  },
 960    { "2D_ARRAY_SHADOW",   2, 3, true,  false, true  },
 961    { "RECT",              2, 2, false, false, false },
 962    { "RECT_SHADOW",       2, 2, false, false, true  },
 963    { "CUBE_ARRAY_SHADOW", 2, 4, true,  true,  true  },
 964    { "BUFFER",            1, 1, false, false, false },
 965 };
 966
 967 const struct TexInstruction::ImgFormatDesc TexInstruction::formatTable[] =
 968 {
 969    { "NONE",         0, {  0,  0,  0,  0 },  UINT },
 970
 971    { "RGBA32F",      4, { 32, 32, 32, 32 }, FLOAT },
 972    { "RGBA16F",      4, { 16, 16, 16, 16 }, FLOAT },
 973    { "RG32F",        2, { 32, 32,  0,  0 }, FLOAT },
 974    { "RG16F",        2, { 16, 16,  0,  0 }, FLOAT },
 975    { "R11G11B10F",   3, { 11, 11, 10,  0 }, FLOAT },
 976    { "R32F",         1, { 32,  0,  0,  0 }, FLOAT },
 977    { "R16F",         1, { 16,  0,  0,  0 }, FLOAT },
 978
 979    { "RGBA32UI",     4, { 32, 32, 32, 32 },  UINT },
 980    { "RGBA16UI",     4, { 16, 16, 16, 16 },  UINT },
 981    { "RGB10A2UI",    4, { 10, 10, 10,  2 },  UINT },
 982    { "RGBA8UI",      4, {  8,  8,  8,  8 },  UINT },
 983    { "RG32UI",       2, { 32, 32,  0,  0 },  UINT },
 984    { "RG16UI",       2, { 16, 16,  0,  0 },  UINT },
 985    { "RG8UI",        2, {  8,  8,  0,  0 },  UINT },
 986    { "R32UI",        1, { 32,  0,  0,  0 },  UINT },
 987    { "R16UI",        1, { 16,  0,  0,  0 },  UINT },
 988    { "R8UI",         1, {  8,  0,  0,  0 },  UINT },
 989
 990    { "RGBA32I",      4, { 32, 32, 32, 32 },  SINT },
 991    { "RGBA16I",      4, { 16, 16, 16, 16 },  SINT },
 992    { "RGBA8I",       4, {  8,  8,  8,  8 },  SINT },
 993    { "RG32I",        2, { 32, 32,  0,  0 },  SINT },
 994    { "RG16I",        2, { 16, 16,  0,  0 },  SINT },
 995    { "RG8I",         2, {  8,  8,  0,  0 },  SINT },
 996    { "R32I",         1, { 32,  0,  0,  0 },  SINT },
 997    { "R16I",         1, { 16,  0,  0,  0 },  SINT },
 998    { "R8I",          1, {  8,  0,  0,  0 },  SINT },
 999
1000    { "RGBA16",       4, { 16, 16, 16, 16 }, UNORM },
1001    { "RGB10A2",      4, { 10, 10, 10,  2 }, UNORM },
1002    { "RGBA8",        4, {  8,  8,  8,  8 }, UNORM },
1003    { "RG16",         2, { 16, 16,  0,  0 }, UNORM },
1004    { "RG8",          2, {  8,  8,  0,  0 }, UNORM },
1005    { "R16",          1, { 16,  0,  0,  0 }, UNORM },
1006    { "R8",           1, {  8,  0,  0,  0 }, UNORM },
1007
1008    { "RGBA16_SNORM", 4, { 16, 16, 16, 16 }, SNORM },
1009    { "RGBA8_SNORM",  4, {  8,  8,  8,  8 }, SNORM },
1010    { "RG16_SNORM",   2, { 16, 16,  0,  0 }, SNORM },
1011    { "RG8_SNORM",    2, {  8,  8,  0,  0 }, SNORM },
1012    { "R16_SNORM",    1, { 16,  0,  0,  0 }, SNORM },
1013    { "R8_SNORM",     1, {  8,  0,  0,  0 }, SNORM },
1014
1015    { "BGRA8",        4, {  8,  8,  8,  8 }, UNORM, true },
1016 };
1017
1018 void
1019 TexInstruction::setIndirectR(Value *v)
1020 {
1021    int p = ((tex.rIndirectSrc < 0) && v) ? srcs.size() : tex.rIndirectSrc;
1022    if (p >= 0) {
1023       tex.rIndirectSrc = p;
1024       setSrc(p, v);
1025       srcs[p].usedAsPtr = !!v;
1026    }
1027 }
1028
1029 void
1030 TexInstruction::setIndirectS(Value *v)
1031 {
1032    int p = ((tex.sIndirectSrc < 0) && v) ? srcs.size() : tex.sIndirectSrc;
1033    if (p >= 0) {
1034       tex.sIndirectSrc = p;
1035       setSrc(p, v);
1036       srcs[p].usedAsPtr = !!v;
1037    }
1038 }
1039
1040 CmpInstruction::CmpInstruction(Function *fn, operation op)
1041    : Instruction(fn, op, TYPE_F32)
1042 {
1043    setCond = CC_ALWAYS;
1044 }
1045
1046 CmpInstruction *
1047 CmpInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
1048 {
1049    CmpInstruction *cmp = (i ? static_cast<CmpInstruction *>(i) :
1050                           new_CmpInstruction(pol.context(), op));
1051    cmp->dType = dType;
1052    Instruction::clone(pol, cmp);
1053    cmp->setCond = setCond;
1054    return cmp;
1055 }
1056
1057 FlowInstruction::FlowInstruction(Function *fn, operation op, void *targ)
1058    : Instruction(fn, op, TYPE_NONE)
1059 {
1060    if (op == OP_CALL)
1061       target.fn = reinterpret_cast<Function *>(targ);
1062    else
1063       target.bb = reinterpret_cast<BasicBlock *>(targ);
1064
1065    if (op == OP_BRA ||
1066        op == OP_CONT || op == OP_BREAK ||
1067        op == OP_RET || op == OP_EXIT)
1068       terminator = 1;
1069    else
1070    if (op == OP_JOIN)
1071       terminator = targ ? 1 : 0;
1072
1073    allWarp = absolute = limit = builtin = indirect = 0;
1074 }
1075
1076 FlowInstruction *
1077 FlowInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
1078 {
1079    FlowInstruction *flow = (i ? static_cast<FlowInstruction *>(i) :
1080                             new_FlowInstruction(pol.context(), op, NULL));
1081
1082    Instruction::clone(pol, flow);
1083    flow->allWarp = allWarp;
1084    flow->absolute = absolute;
1085    flow->limit = limit;
1086    flow->builtin = builtin;
1087
1088    if (builtin)
1089       flow->target.builtin = target.builtin;
1090    else
1091    if (op == OP_CALL)
1092       flow->target.fn = target.fn;
1093    else
1094    if (target.bb)
1095       flow->target.bb = pol.get<BasicBlock>(target.bb);
1096
1097    return flow;
1098 }
1099
1100 Program::Program(Type type, Target *arch)
1101    : progType(type),
1102      target(arch),
1103      mem_Instruction(sizeof(Instruction), 6),
1104      mem_CmpInstruction(sizeof(CmpInstruction), 4),
1105      mem_TexInstruction(sizeof(TexInstruction), 4),
1106      mem_FlowInstruction(sizeof(FlowInstruction), 4),
1107      mem_LValue(sizeof(LValue), 8),
1108      mem_Symbol(sizeof(Symbol), 7),
1109      mem_ImmediateValue(sizeof(ImmediateValue), 7)
1110 {
1111    code = NULL;
1112    binSize = 0;
1113
1114    maxGPR = -1;
1115
1116    main = new Function(this, "MAIN", ~0);
1117    calls.insert(&main->call);
1118
1119    dbgFlags = 0;
1120    optLevel = 0;
1121
1122    targetPriv = NULL;
1123 }
1124
1125 Program::~Program()
1126 {
1127    for (ArrayList::Iterator it = allFuncs.iterator(); !it.end(); it.next())
1128       delete reinterpret_cast<Function *>(it.get());
1129
1130    for (ArrayList::Iterator it = allRValues.iterator(); !it.end(); it.next())
1131       releaseValue(reinterpret_cast<Value *>(it.get()));
1132 }
1133
1134 void Program::releaseInstruction(Instruction *insn)
1135 {
1136    // TODO: make this not suck so much
1137
1138    insn->~Instruction();
1139
1140    if (insn->asCmp())
1141       mem_CmpInstruction.release(insn);
1142    else
1143    if (insn->asTex())
1144       mem_TexInstruction.release(insn);
1145    else
1146    if (insn->asFlow())
1147       mem_FlowInstruction.release(insn);
1148    else
1149       mem_Instruction.release(insn);
1150 }
1151
1152 void Program::releaseValue(Value *value)
1153 {
1154    value->~Value();
1155
1156    if (value->asLValue())
1157       mem_LValue.release(value);
1158    else
1159    if (value->asImm())
1160       mem_ImmediateValue.release(value);
1161    else
1162    if (value->asSym())
1163       mem_Symbol.release(value);
1164 }
1165
1166
1167 } // namespace nv50_ir
1168
1169 extern "C" {
1170
1171 static void
1172 nv50_ir_init_prog_info(struct nv50_ir_prog_info *info)
1173 {
1174    if (info->type == PIPE_SHADER_TESS_CTRL || info->type == PIPE_SHADER_TESS_EVAL) {
1175       info->prop.tp.domain = PIPE_PRIM_MAX;
1176       info->prop.tp.outputPrim = PIPE_PRIM_MAX;
1177    }
1178    if (info->type == PIPE_SHADER_GEOMETRY) {
1179       info->prop.gp.instanceCount = 1;
1180       info->prop.gp.maxVertices = 1;
1181    }
1182    if (info->type == PIPE_SHADER_COMPUTE) {
1183       info->prop.cp.numThreads[0] =
1184       info->prop.cp.numThreads[1] =
1185       info->prop.cp.numThreads[2] = 1;
1186    }
1187    info->io.pointSize = 0xff;
1188    info->io.instanceId = 0xff;
1189    info->io.vertexId = 0xff;
1190    info->io.edgeFlagIn = 0xff;
1191    info->io.edgeFlagOut = 0xff;
1192    info->io.fragDepth = 0xff;
1193    info->io.sampleMask = 0xff;
1194    info->io.backFaceColor[0] = info->io.backFaceColor[1] = 0xff;
1195 }
1196
1197 int
1198 nv50_ir_generate_code(struct nv50_ir_prog_info *info)
1199 {
1200    int ret = 0;
1201
1202    nv50_ir::Program::Type type;
1203
1204    nv50_ir_init_prog_info(info);
1205
1206 #define PROG_TYPE_CASE(a, b)                                      \
1207    case PIPE_SHADER_##a: type = nv50_ir::Program::TYPE_##b; break
1208
1209    switch (info->type) {
1210    PROG_TYPE_CASE(VERTEX, VERTEX);
1211    PROG_TYPE_CASE(TESS_CTRL, TESSELLATION_CONTROL);
1212    PROG_TYPE_CASE(TESS_EVAL, TESSELLATION_EVAL);
1213    PROG_TYPE_CASE(GEOMETRY, GEOMETRY);
1214    PROG_TYPE_CASE(FRAGMENT, FRAGMENT);
1215    PROG_TYPE_CASE(COMPUTE, COMPUTE);
1216    default:
1217       INFO_DBG(info->dbgFlags, VERBOSE, "unsupported program type %u\n", info->type);
1218       return -1;
1219    }
1220    INFO_DBG(info->dbgFlags, VERBOSE, "translating program of type %u\n", type);
1221
1222    nv50_ir::Target *targ = nv50_ir::Target::create(info->target);
1223    if (!targ)
1224       return -1;
1225
1226    nv50_ir::Program *prog = new nv50_ir::Program(type, targ);
1227    if (!prog) {
1228       nv50_ir::Target::destroy(targ);
1229       return -1;
1230    }
1231    prog->driver = info;
1232    prog->dbgFlags = info->dbgFlags;
1233    prog->optLevel = info->optLevel;
1234
1235    switch (info->bin.sourceRep) {
1236    case PIPE_SHADER_IR_TGSI:
1237       ret = prog->makeFromTGSI(info) ? 0 : -2;
1238       break;
1239    default:
1240       ret = -1;
1241       break;
1242    }
1243    if (ret < 0)
1244       goto out;
1245    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
1246       prog->print();
1247
1248    targ->parseDriverInfo(info);
1249    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_PRE_SSA);
1250
1251    prog->convertToSSA();
1252
1253    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
1254       prog->print();
1255
1256    prog->optimizeSSA(info->optLevel);
1257    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_SSA);
1258
1259    if (prog->dbgFlags & NV50_IR_DEBUG_BASIC)
1260       prog->print();
1261
1262    if (!prog->registerAllocation()) {
1263       ret = -4;
1264       goto out;
1265    }
1266    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_POST_RA);
1267
1268    prog->optimizePostRA(info->optLevel);
1269
1270    if (!prog->emitBinary(info)) {
1271       ret = -5;
1272       goto out;
1273    }
1274
1275 out:
1276    INFO_DBG(prog->dbgFlags, VERBOSE, "nv50_ir_generate_code: ret = %i\n", ret);
1277
1278    info->bin.maxGPR = prog->maxGPR;
1279    info->bin.code = prog->code;
1280    info->bin.codeSize = prog->binSize;
1281    info->bin.tlsSpace = prog->tlsSize;
1282
1283    delete prog;
1284    nv50_ir::Target::destroy(targ);
1285
1286    return ret;
1287 }
1288
1289 } // extern "C"