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_F32:
 397       return reg.data.f32 == static_cast<float>(i);
 398    case TYPE_F64:
 399       return reg.data.f64 == static_cast<double>(i);
 400    default:
 401       return false;
 402    }
 403 }
 404
 405 bool
 406 ImmediateValue::isNegative() const
 407 {
 408    switch (reg.type) {
 409    case TYPE_S8:  return reg.data.s8 < 0;
 410    case TYPE_S16: return reg.data.s16 < 0;
 411    case TYPE_S32:
 412    case TYPE_U32: return reg.data.s32 < 0;
 413    case TYPE_F32: return reg.data.u32 & (1 << 31);
 414    case TYPE_F64: return reg.data.u64 & (1ULL << 63);
 415    default:
 416       return false;
 417    }
 418 }
 419
 420 bool
 421 ImmediateValue::isPow2() const
 422 {
 423    switch (reg.type) {
 424    case TYPE_U8:
 425    case TYPE_U16:
 426    case TYPE_U32: return util_is_power_of_two(reg.data.u32);
 427    default:
 428       return false;
 429    }
 430 }
 431
 432 void
 433 ImmediateValue::applyLog2()
 434 {
 435    switch (reg.type) {
 436    case TYPE_S8:
 437    case TYPE_S16:
 438    case TYPE_S32:
 439       assert(!this->isNegative());
 440       // fall through
 441    case TYPE_U8:
 442    case TYPE_U16:
 443    case TYPE_U32:
 444       reg.data.u32 = util_logbase2(reg.data.u32);
 445       break;
 446    case TYPE_F32:
 447       reg.data.f32 = log2f(reg.data.f32);
 448       break;
 449    case TYPE_F64:
 450       reg.data.f64 = log2(reg.data.f64);
 451       break;
 452    default:
 453       assert(0);
 454       break;
 455    }
 456 }
 457
 458 bool
 459 ImmediateValue::compare(CondCode cc, float fval) const
 460 {
 461    if (reg.type != TYPE_F32)
 462       ERROR("immediate value is not of type f32");
 463
 464    switch (static_cast<CondCode>(cc & 7)) {
 465    case CC_TR: return true;
 466    case CC_FL: return false;
 467    case CC_LT: return reg.data.f32 <  fval;
 468    case CC_LE: return reg.data.f32 <= fval;
 469    case CC_GT: return reg.data.f32 >  fval;
 470    case CC_GE: return reg.data.f32 >= fval;
 471    case CC_EQ: return reg.data.f32 == fval;
 472    case CC_NE: return reg.data.f32 != fval;
 473    default:
 474       assert(0);
 475       return false;
 476    }
 477 }
 478
 479 ImmediateValue&
 480 ImmediateValue::operator=(const ImmediateValue &that)
 481 {
 482    this->reg = that.reg;
 483    return (*this);
 484 }
 485
 486 bool
 487 Value::interfers(const Value *that) const
 488 {
 489    uint32_t idA, idB;
 490
 491    if (that->reg.file != reg.file || that->reg.fileIndex != reg.fileIndex)
 492       return false;
 493    if (this->asImm())
 494       return false;
 495
 496    if (this->asSym()) {
 497       idA = this->join->reg.data.offset;
 498       idB = that->join->reg.data.offset;
 499    } else {
 500       idA = this->join->reg.data.id * MIN2(this->reg.size, 4);
 501       idB = that->join->reg.data.id * MIN2(that->reg.size, 4);
 502    }
 503
 504    if (idA < idB)
 505       return (idA + this->reg.size > idB);
 506    else
 507    if (idA > idB)
 508       return (idB + that->reg.size > idA);
 509    else
 510       return (idA == idB);
 511 }
 512
 513 bool
 514 Value::equals(const Value *that, bool strict) const
 515 {
 516    if (strict)
 517       return this == that;
 518
 519    if (that->reg.file != reg.file || that->reg.fileIndex != reg.fileIndex)
 520       return false;
 521    if (that->reg.size != this->reg.size)
 522       return false;
 523
 524    if (that->reg.data.id != this->reg.data.id)
 525       return false;
 526
 527    return true;
 528 }
 529
 530 bool
 531 ImmediateValue::equals(const Value *that, bool strict) const
 532 {
 533    const ImmediateValue *imm = that->asImm();
 534    if (!imm)
 535       return false;
 536    return reg.data.u64 == imm->reg.data.u64;
 537 }
 538
 539 bool
 540 Symbol::equals(const Value *that, bool strict) const
 541 {
 542    if (reg.file != that->reg.file || reg.fileIndex != that->reg.fileIndex)
 543       return false;
 544    assert(that->asSym());
 545
 546    if (this->baseSym != that->asSym()->baseSym)
 547       return false;
 548
 549    if (reg.file == FILE_SYSTEM_VALUE)
 550       return (this->reg.data.sv.sv    == that->reg.data.sv.sv &&
 551               this->reg.data.sv.index == that->reg.data.sv.index);
 552    return this->reg.data.offset == that->reg.data.offset;
 553 }
 554
 555 void Instruction::init()
 556 {
 557    next = prev = 0;
 558
 559    cc = CC_ALWAYS;
 560    rnd = ROUND_N;
 561    cache = CACHE_CA;
 562    subOp = 0;
 563
 564    saturate = 0;
 565    join = 0;
 566    exit = 0;
 567    terminator = 0;
 568    ftz = 0;
 569    dnz = 0;
 570    perPatch = 0;
 571    fixed = 0;
 572    encSize = 0;
 573    ipa = 0;
 574    mask = 0;
 575
 576    lanes = 0xf;
 577
 578    postFactor = 0;
 579
 580    predSrc = -1;
 581    flagsDef = -1;
 582    flagsSrc = -1;
 583 }
 584
 585 Instruction::Instruction()
 586 {
 587    init();
 588
 589    op = OP_NOP;
 590    dType = sType = TYPE_F32;
 591
 592    id = -1;
 593    bb = 0;
 594 }
 595
 596 Instruction::Instruction(Function *fn, operation opr, DataType ty)
 597 {
 598    init();
 599
 600    op = opr;
 601    dType = sType = ty;
 602
 603    fn->add(this, id);
 604 }
 605
 606 Instruction::~Instruction()
 607 {
 608    if (bb) {
 609       Function *fn = bb->getFunction();
 610       bb->remove(this);
 611       fn->allInsns.remove(id);
 612    }
 613
 614    for (int s = 0; srcExists(s); ++s)
 615       setSrc(s, NULL);
 616    // must unlink defs too since the list pointers will get deallocated
 617    for (int d = 0; defExists(d); ++d)
 618       setDef(d, NULL);
 619 }
 620
 621 void
 622 Instruction::setDef(int i, Value *val)
 623 {
 624    int size = defs.size();
 625    if (i >= size) {
 626       defs.resize(i + 1);
 627       while (size <= i)
 628          defs[size++].setInsn(this);
 629    }
 630    defs[i].set(val);
 631 }
 632
 633 void
 634 Instruction::setSrc(int s, Value *val)
 635 {
 636    int size = srcs.size();
 637    if (s >= size) {
 638       srcs.resize(s + 1);
 639       while (size <= s)
 640          srcs[size++].setInsn(this);
 641    }
 642    srcs[s].set(val);
 643 }
 644
 645 void
 646 Instruction::setSrc(int s, const ValueRef& ref)
 647 {
 648    setSrc(s, ref.get());
 649    srcs[s].mod = ref.mod;
 650 }
 651
 652 void
 653 Instruction::swapSources(int a, int b)
 654 {
 655    Value *value = srcs[a].get();
 656    Modifier m = srcs[a].mod;
 657
 658    setSrc(a, srcs[b]);
 659
 660    srcs[b].set(value);
 661    srcs[b].mod = m;
 662 }
 663
 664 static inline void moveSourcesAdjustIndex(int8_t &index, int s, int delta)
 665 {
 666    if (index >= s)
 667       index += delta;
 668    else
 669    if ((delta < 0) && (index >= (s + delta)))
 670       index = -1;
 671 }
 672
 673 // Moves sources [@s,last_source] by @delta.
 674 // If @delta < 0, sources [@s - abs(@delta), @s) are erased.
 675 void
 676 Instruction::moveSources(const int s, const int delta)
 677 {
 678    if (delta == 0)
 679       return;
 680    assert(s + delta >= 0);
 681
 682    int k;
 683
 684    for (k = 0; srcExists(k); ++k) {
 685       for (int i = 0; i < 2; ++i)
 686          moveSourcesAdjustIndex(src(k).indirect[i], s, delta);
 687    }
 688    moveSourcesAdjustIndex(predSrc, s, delta);
 689    moveSourcesAdjustIndex(flagsSrc, s, delta);
 690    if (asTex()) {
 691       TexInstruction *tex = asTex();
 692       moveSourcesAdjustIndex(tex->tex.rIndirectSrc, s, delta);
 693       moveSourcesAdjustIndex(tex->tex.sIndirectSrc, s, delta);
 694    }
 695
 696    if (delta > 0) {
 697       --k;
 698       for (int p = k + delta; k >= s; --k, --p)
 699          setSrc(p, src(k));
 700    } else {
 701       int p;
 702       for (p = s; p < k; ++p)
 703          setSrc(p + delta, src(p));
 704       for (; (p + delta) < k; ++p)
 705          setSrc(p + delta, NULL);
 706    }
 707 }
 708
 709 void
 710 Instruction::takeExtraSources(int s, Value *values[3])
 711 {
 712    values[0] = getIndirect(s, 0);
 713    if (values[0])
 714       setIndirect(s, 0, NULL);
 715
 716    values[1] = getIndirect(s, 1);
 717    if (values[1])
 718       setIndirect(s, 1, NULL);
 719
 720    values[2] = getPredicate();
 721    if (values[2])
 722       setPredicate(cc, NULL);
 723 }
 724
 725 void
 726 Instruction::putExtraSources(int s, Value *values[3])
 727 {
 728    if (values[0])
 729       setIndirect(s, 0, values[0]);
 730    if (values[1])
 731       setIndirect(s, 1, values[1]);
 732    if (values[2])
 733       setPredicate(cc, values[2]);
 734 }
 735
 736 Instruction *
 737 Instruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 738 {
 739    if (!i)
 740       i = new_Instruction(pol.context(), op, dType);
 741 #ifndef NDEBUG // non-conformant assert, so this is required
 742    assert(typeid(*i) == typeid(*this));
 743 #endif
 744
 745    pol.set<Instruction>(this, i);
 746
 747    i->sType = sType;
 748
 749    i->rnd = rnd;
 750    i->cache = cache;
 751    i->subOp = subOp;
 752
 753    i->saturate = saturate;
 754    i->join = join;
 755    i->exit = exit;
 756    i->mask = mask;
 757    i->ftz = ftz;
 758    i->dnz = dnz;
 759    i->ipa = ipa;
 760    i->lanes = lanes;
 761    i->perPatch = perPatch;
 762
 763    i->postFactor = postFactor;
 764
 765    for (int d = 0; defExists(d); ++d)
 766       i->setDef(d, pol.get(getDef(d)));
 767
 768    for (int s = 0; srcExists(s); ++s) {
 769       i->setSrc(s, pol.get(getSrc(s)));
 770       i->src(s).mod = src(s).mod;
 771    }
 772
 773    i->cc = cc;
 774    i->predSrc = predSrc;
 775    i->flagsDef = flagsDef;
 776    i->flagsSrc = flagsSrc;
 777
 778    return i;
 779 }
 780
 781 unsigned int
 782 Instruction::defCount(unsigned int mask, bool singleFile) const
 783 {
 784    unsigned int i, n;
 785
 786    if (singleFile) {
 787       unsigned int d = ffs(mask);
 788       if (!d)
 789          return 0;
 790       for (i = d--; defExists(i); ++i)
 791          if (getDef(i)->reg.file != getDef(d)->reg.file)
 792             mask &= ~(1 << i);
 793    }
 794
 795    for (n = 0, i = 0; this->defExists(i); ++i, mask >>= 1)
 796       n += mask & 1;
 797    return n;
 798 }
 799
 800 unsigned int
 801 Instruction::srcCount(unsigned int mask, bool singleFile) const
 802 {
 803    unsigned int i, n;
 804
 805    if (singleFile) {
 806       unsigned int s = ffs(mask);
 807       if (!s)
 808          return 0;
 809       for (i = s--; srcExists(i); ++i)
 810          if (getSrc(i)->reg.file != getSrc(s)->reg.file)
 811             mask &= ~(1 << i);
 812    }
 813
 814    for (n = 0, i = 0; this->srcExists(i); ++i, mask >>= 1)
 815       n += mask & 1;
 816    return n;
 817 }
 818
 819 bool
 820 Instruction::setIndirect(int s, int dim, Value *value)
 821 {
 822    assert(this->srcExists(s));
 823
 824    int p = srcs[s].indirect[dim];
 825    if (p < 0) {
 826       if (!value)
 827          return true;
 828       p = srcs.size();
 829       while (p > 0 && !srcExists(p - 1))
 830          --p;
 831    }
 832    setSrc(p, value);
 833    srcs[p].usedAsPtr = (value != 0);
 834    srcs[s].indirect[dim] = value ? p : -1;
 835    return true;
 836 }
 837
 838 bool
 839 Instruction::setPredicate(CondCode ccode, Value *value)
 840 {
 841    cc = ccode;
 842
 843    if (!value) {
 844       if (predSrc >= 0) {
 845          srcs[predSrc].set(NULL);
 846          predSrc = -1;
 847       }
 848       return true;
 849    }
 850
 851    if (predSrc < 0) {
 852       predSrc = srcs.size();
 853       while (predSrc > 0 && !srcExists(predSrc - 1))
 854          --predSrc;
 855    }
 856
 857    setSrc(predSrc, value);
 858    return true;
 859 }
 860
 861 bool
 862 Instruction::writesPredicate() const
 863 {
 864    for (int d = 0; defExists(d); ++d)
 865       if (getDef(d)->inFile(FILE_PREDICATE) || getDef(d)->inFile(FILE_FLAGS))
 866          return true;
 867    return false;
 868 }
 869
 870 static bool
 871 insnCheckCommutationDefSrc(const Instruction *a, const Instruction *b)
 872 {
 873    for (int d = 0; a->defExists(d); ++d)
 874       for (int s = 0; b->srcExists(s); ++s)
 875          if (a->getDef(d)->interfers(b->getSrc(s)))
 876             return false;
 877    return true;
 878 }
 879
 880 static bool
 881 insnCheckCommutationDefDef(const Instruction *a, const Instruction *b)
 882 {
 883    for (int d = 0; a->defExists(d); ++d)
 884       for (int c = 0; b->defExists(c); ++c)
 885          if (a->getDef(d)->interfers(b->getDef(c)))
 886             return false;
 887    return true;
 888 }
 889
 890 bool
 891 Instruction::isCommutationLegal(const Instruction *i) const
 892 {
 893    bool ret = insnCheckCommutationDefDef(this, i);
 894    ret = ret && insnCheckCommutationDefSrc(this, i);
 895    ret = ret && insnCheckCommutationDefSrc(i, this);
 896    return ret;
 897 }
 898
 899 TexInstruction::TexInstruction(Function *fn, operation op)
 900    : Instruction(fn, op, TYPE_F32)
 901 {
 902    memset(&tex, 0, sizeof(tex));
 903
 904    tex.rIndirectSrc = -1;
 905    tex.sIndirectSrc = -1;
 906 }
 907
 908 TexInstruction::~TexInstruction()
 909 {
 910    for (int c = 0; c < 3; ++c) {
 911       dPdx[c].set(NULL);
 912       dPdy[c].set(NULL);
 913    }
 914    for (int n = 0; n < 4; ++n)
 915       for (int c = 0; c < 3; ++c)
 916          offset[n][c].set(NULL);
 917 }
 918
 919 TexInstruction *
 920 TexInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 921 {
 922    TexInstruction *tex = (i ? static_cast<TexInstruction *>(i) :
 923                           new_TexInstruction(pol.context(), op));
 924
 925    Instruction::clone(pol, tex);
 926
 927    tex->tex = this->tex;
 928
 929    if (op == OP_TXD) {
 930       for (unsigned int c = 0; c < tex->tex.target.getDim(); ++c) {
 931          tex->dPdx[c].set(dPdx[c]);
 932          tex->dPdy[c].set(dPdy[c]);
 933       }
 934    }
 935
 936    for (int n = 0; n < tex->tex.useOffsets; ++n)
 937       for (int c = 0; c < 3; ++c)
 938          tex->offset[n][c].set(offset[n][c]);
 939
 940    return tex;
 941 }
 942
 943 const struct TexInstruction::Target::Desc TexInstruction::Target::descTable[] =
 944 {
 945    { "1D",                1, 1, false, false, false },
 946    { "2D",                2, 2, false, false, false },
 947    { "2D_MS",             2, 3, false, false, false },
 948    { "3D",                3, 3, false, false, false },
 949    { "CUBE",              2, 3, false, true,  false },
 950    { "1D_SHADOW",         1, 1, false, false, true  },
 951    { "2D_SHADOW",         2, 2, false, false, true  },
 952    { "CUBE_SHADOW",       2, 3, false, true,  true  },
 953    { "1D_ARRAY",          1, 2, true,  false, false },
 954    { "2D_ARRAY",          2, 3, true,  false, false },
 955    { "2D_MS_ARRAY",       2, 4, true,  false, false },
 956    { "CUBE_ARRAY",        2, 4, true,  true,  false },
 957    { "1D_ARRAY_SHADOW",   1, 2, true,  false, true  },
 958    { "2D_ARRAY_SHADOW",   2, 3, true,  false, true  },
 959    { "RECT",              2, 2, false, false, false },
 960    { "RECT_SHADOW",       2, 2, false, false, true  },
 961    { "CUBE_ARRAY_SHADOW", 2, 4, true,  true,  true  },
 962    { "BUFFER",            1, 1, false, false, false },
 963 };
 964
 965 void
 966 TexInstruction::setIndirectR(Value *v)
 967 {
 968    int p = ((tex.rIndirectSrc < 0) && v) ? srcs.size() : tex.rIndirectSrc;
 969    if (p >= 0) {
 970       tex.rIndirectSrc = p;
 971       setSrc(p, v);
 972       srcs[p].usedAsPtr = !!v;
 973    }
 974 }
 975
 976 void
 977 TexInstruction::setIndirectS(Value *v)
 978 {
 979    int p = ((tex.sIndirectSrc < 0) && v) ? srcs.size() : tex.sIndirectSrc;
 980    if (p >= 0) {
 981       tex.sIndirectSrc = p;
 982       setSrc(p, v);
 983       srcs[p].usedAsPtr = !!v;
 984    }
 985 }
 986
 987 CmpInstruction::CmpInstruction(Function *fn, operation op)
 988    : Instruction(fn, op, TYPE_F32)
 989 {
 990    setCond = CC_ALWAYS;
 991 }
 992
 993 CmpInstruction *
 994 CmpInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 995 {
 996    CmpInstruction *cmp = (i ? static_cast<CmpInstruction *>(i) :
 997                           new_CmpInstruction(pol.context(), op));
 998    cmp->dType = dType;
 999    Instruction::clone(pol, cmp);
1000    cmp->setCond = setCond;
1001    return cmp;
1002 }
1003
1004 FlowInstruction::FlowInstruction(Function *fn, operation op, void *targ)
1005    : Instruction(fn, op, TYPE_NONE)
1006 {
1007    if (op == OP_CALL)
1008       target.fn = reinterpret_cast<Function *>(targ);
1009    else
1010       target.bb = reinterpret_cast<BasicBlock *>(targ);
1011
1012    if (op == OP_BRA ||
1013        op == OP_CONT || op == OP_BREAK ||
1014        op == OP_RET || op == OP_EXIT)
1015       terminator = 1;
1016    else
1017    if (op == OP_JOIN)
1018       terminator = targ ? 1 : 0;
1019
1020    allWarp = absolute = limit = builtin = indirect = 0;
1021 }
1022
1023 FlowInstruction *
1024 FlowInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
1025 {
1026    FlowInstruction *flow = (i ? static_cast<FlowInstruction *>(i) :
1027                             new_FlowInstruction(pol.context(), op, NULL));
1028
1029    Instruction::clone(pol, flow);
1030    flow->allWarp = allWarp;
1031    flow->absolute = absolute;
1032    flow->limit = limit;
1033    flow->builtin = builtin;
1034
1035    if (builtin)
1036       flow->target.builtin = target.builtin;
1037    else
1038    if (op == OP_CALL)
1039       flow->target.fn = target.fn;
1040    else
1041    if (target.bb)
1042       flow->target.bb = pol.get<BasicBlock>(target.bb);
1043
1044    return flow;
1045 }
1046
1047 Program::Program(Type type, Target *arch)
1048    : progType(type),
1049      target(arch),
1050      mem_Instruction(sizeof(Instruction), 6),
1051      mem_CmpInstruction(sizeof(CmpInstruction), 4),
1052      mem_TexInstruction(sizeof(TexInstruction), 4),
1053      mem_FlowInstruction(sizeof(FlowInstruction), 4),
1054      mem_LValue(sizeof(LValue), 8),
1055      mem_Symbol(sizeof(Symbol), 7),
1056      mem_ImmediateValue(sizeof(ImmediateValue), 7)
1057 {
1058    code = NULL;
1059    binSize = 0;
1060
1061    maxGPR = -1;
1062
1063    main = new Function(this, "MAIN", ~0);
1064    calls.insert(&main->call);
1065
1066    dbgFlags = 0;
1067    optLevel = 0;
1068
1069    targetPriv = NULL;
1070 }
1071
1072 Program::~Program()
1073 {
1074    for (ArrayList::Iterator it = allFuncs.iterator(); !it.end(); it.next())
1075       delete reinterpret_cast<Function *>(it.get());
1076
1077    for (ArrayList::Iterator it = allRValues.iterator(); !it.end(); it.next())
1078       releaseValue(reinterpret_cast<Value *>(it.get()));
1079 }
1080
1081 void Program::releaseInstruction(Instruction *insn)
1082 {
1083    // TODO: make this not suck so much
1084
1085    insn->~Instruction();
1086
1087    if (insn->asCmp())
1088       mem_CmpInstruction.release(insn);
1089    else
1090    if (insn->asTex())
1091       mem_TexInstruction.release(insn);
1092    else
1093    if (insn->asFlow())
1094       mem_FlowInstruction.release(insn);
1095    else
1096       mem_Instruction.release(insn);
1097 }
1098
1099 void Program::releaseValue(Value *value)
1100 {
1101    value->~Value();
1102
1103    if (value->asLValue())
1104       mem_LValue.release(value);
1105    else
1106    if (value->asImm())
1107       mem_ImmediateValue.release(value);
1108    else
1109    if (value->asSym())
1110       mem_Symbol.release(value);
1111 }
1112
1113
1114 } // namespace nv50_ir
1115
1116 extern "C" {
1117
1118 static void
1119 nv50_ir_init_prog_info(struct nv50_ir_prog_info *info)
1120 {
1121 #if defined(PIPE_SHADER_HULL) && defined(PIPE_SHADER_DOMAIN)
1122    if (info->type == PIPE_SHADER_HULL || info->type == PIPE_SHADER_DOMAIN) {
1123       info->prop.tp.domain = PIPE_PRIM_MAX;
1124       info->prop.tp.outputPrim = PIPE_PRIM_MAX;
1125    }
1126 #endif
1127    if (info->type == PIPE_SHADER_GEOMETRY) {
1128       info->prop.gp.instanceCount = 1;
1129       info->prop.gp.maxVertices = 1;
1130    }
1131    info->io.pointSize = 0xff;
1132    info->io.instanceId = 0xff;
1133    info->io.vertexId = 0xff;
1134    info->io.edgeFlagIn = 0xff;
1135    info->io.edgeFlagOut = 0xff;
1136    info->io.fragDepth = 0xff;
1137    info->io.sampleMask = 0xff;
1138    info->io.backFaceColor[0] = info->io.backFaceColor[1] = 0xff;
1139 }
1140
1141 int
1142 nv50_ir_generate_code(struct nv50_ir_prog_info *info)
1143 {
1144    int ret = 0;
1145
1146    nv50_ir::Program::Type type;
1147
1148    nv50_ir_init_prog_info(info);
1149
1150 #define PROG_TYPE_CASE(a, b)                                      \
1151    case PIPE_SHADER_##a: type = nv50_ir::Program::TYPE_##b; break
1152
1153    switch (info->type) {
1154    PROG_TYPE_CASE(VERTEX, VERTEX);
1155    PROG_TYPE_CASE(TESS_CTRL, TESSELLATION_CONTROL);
1156    PROG_TYPE_CASE(TESS_EVAL, TESSELLATION_EVAL);
1157    PROG_TYPE_CASE(GEOMETRY, GEOMETRY);
1158    PROG_TYPE_CASE(FRAGMENT, FRAGMENT);
1159    PROG_TYPE_CASE(COMPUTE, COMPUTE);
1160    default:
1161       type = nv50_ir::Program::TYPE_COMPUTE;
1162       break;
1163    }
1164    INFO_DBG(info->dbgFlags, VERBOSE, "translating program of type %u\n", type);
1165
1166    nv50_ir::Target *targ = nv50_ir::Target::create(info->target);
1167    if (!targ)
1168       return -1;
1169
1170    nv50_ir::Program *prog = new nv50_ir::Program(type, targ);
1171    if (!prog)
1172       return -1;
1173    prog->driver = info;
1174    prog->dbgFlags = info->dbgFlags;
1175    prog->optLevel = info->optLevel;
1176
1177    switch (info->bin.sourceRep) {
1178 #if 0
1179    case PIPE_IR_LLVM:
1180    case PIPE_IR_GLSL:
1181       return -1;
1182    case PIPE_IR_SM4:
1183       ret = prog->makeFromSM4(info) ? 0 : -2;
1184       break;
1185    case PIPE_IR_TGSI:
1186 #endif
1187    default:
1188       ret = prog->makeFromTGSI(info) ? 0 : -2;
1189       break;
1190    }
1191    if (ret < 0)
1192       goto out;
1193    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
1194       prog->print();
1195
1196    targ->parseDriverInfo(info);
1197    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_PRE_SSA);
1198
1199    prog->convertToSSA();
1200
1201    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
1202       prog->print();
1203
1204    prog->optimizeSSA(info->optLevel);
1205    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_SSA);
1206
1207    if (prog->dbgFlags & NV50_IR_DEBUG_BASIC)
1208       prog->print();
1209
1210    if (!prog->registerAllocation()) {
1211       ret = -4;
1212       goto out;
1213    }
1214    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_POST_RA);
1215
1216    prog->optimizePostRA(info->optLevel);
1217
1218    if (!prog->emitBinary(info)) {
1219       ret = -5;
1220       goto out;
1221    }
1222
1223 out:
1224    INFO_DBG(prog->dbgFlags, VERBOSE, "nv50_ir_generate_code: ret = %i\n", ret);
1225
1226    info->bin.maxGPR = prog->maxGPR;
1227    info->bin.code = prog->code;
1228    info->bin.codeSize = prog->binSize;
1229    info->bin.tlsSpace = prog->tlsSize;
1230
1231    delete prog;
1232    nv50_ir::Target::destroy(targ);
1233
1234    return ret;
1235 }
1236
1237 } // extern "C"