src/gallium/drivers/nv50/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 BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #include "nv50_ir.h"
  24 #include "nv50_ir_target.h"
  25 #include "nv50_ir_driver.h"
  26
  27 extern "C" {
  28 #include "nv50/nv50_program.h"
  29 #include "nv50/nv50_debug.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.remove(this);
 145    if (refVal)
 146       refVal->uses.push_back(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.front();
 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    return this->reg.data.offset == that->reg.data.offset;
 550 }
 551
 552 void Instruction::init()
 553 {
 554    next = prev = 0;
 555
 556    cc = CC_ALWAYS;
 557    rnd = ROUND_N;
 558    cache = CACHE_CA;
 559    subOp = 0;
 560
 561    saturate = 0;
 562    join = 0;
 563    exit = 0;
 564    terminator = 0;
 565    ftz = 0;
 566    dnz = 0;
 567    atomic = 0;
 568    perPatch = 0;
 569    fixed = 0;
 570    encSize = 0;
 571    ipa = 0;
 572
 573    lanes = 0xf;
 574
 575    postFactor = 0;
 576
 577    predSrc = -1;
 578    flagsDef = -1;
 579    flagsSrc = -1;
 580 }
 581
 582 Instruction::Instruction()
 583 {
 584    init();
 585
 586    op = OP_NOP;
 587    dType = sType = TYPE_F32;
 588
 589    id = -1;
 590    bb = 0;
 591 }
 592
 593 Instruction::Instruction(Function *fn, operation opr, DataType ty)
 594 {
 595    init();
 596
 597    op = opr;
 598    dType = sType = ty;
 599
 600    fn->add(this, id);
 601 }
 602
 603 Instruction::~Instruction()
 604 {
 605    if (bb) {
 606       Function *fn = bb->getFunction();
 607       bb->remove(this);
 608       fn->allInsns.remove(id);
 609    }
 610
 611    for (int s = 0; srcExists(s); ++s)
 612       setSrc(s, NULL);
 613    // must unlink defs too since the list pointers will get deallocated
 614    for (int d = 0; defExists(d); ++d)
 615       setDef(d, NULL);
 616 }
 617
 618 void
 619 Instruction::setDef(int i, Value *val)
 620 {
 621    int size = defs.size();
 622    if (i >= size) {
 623       defs.resize(i + 1);
 624       while (size <= i)
 625          defs[size++].setInsn(this);
 626    }
 627    defs[i].set(val);
 628 }
 629
 630 void
 631 Instruction::setSrc(int s, Value *val)
 632 {
 633    int size = srcs.size();
 634    if (s >= size) {
 635       srcs.resize(s + 1);
 636       while (size <= s)
 637          srcs[size++].setInsn(this);
 638    }
 639    srcs[s].set(val);
 640 }
 641
 642 void
 643 Instruction::setSrc(int s, const ValueRef& ref)
 644 {
 645    setSrc(s, ref.get());
 646    srcs[s].mod = ref.mod;
 647 }
 648
 649 void
 650 Instruction::swapSources(int a, int b)
 651 {
 652    Value *value = srcs[a].get();
 653    Modifier m = srcs[a].mod;
 654
 655    setSrc(a, srcs[b]);
 656
 657    srcs[b].set(value);
 658    srcs[b].mod = m;
 659 }
 660
 661 void
 662 Instruction::takeExtraSources(int s, Value *values[3])
 663 {
 664    values[0] = getIndirect(s, 0);
 665    if (values[0])
 666       setIndirect(s, 0, NULL);
 667
 668    values[1] = getIndirect(s, 1);
 669    if (values[1])
 670       setIndirect(s, 1, NULL);
 671
 672    values[2] = getPredicate();
 673    if (values[2])
 674       setPredicate(cc, NULL);
 675 }
 676
 677 void
 678 Instruction::putExtraSources(int s, Value *values[3])
 679 {
 680    if (values[0])
 681       setIndirect(s, 0, values[0]);
 682    if (values[1])
 683       setIndirect(s, 1, values[1]);
 684    if (values[2])
 685       setPredicate(cc, values[2]);
 686 }
 687
 688 Instruction *
 689 Instruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 690 {
 691    if (!i)
 692       i = new_Instruction(pol.context(), op, dType);
 693    assert(typeid(*i) == typeid(*this));
 694
 695    pol.set<Instruction>(this, i);
 696
 697    i->sType = sType;
 698
 699    i->rnd = rnd;
 700    i->cache = cache;
 701    i->subOp = subOp;
 702
 703    i->saturate = saturate;
 704    i->join = join;
 705    i->exit = exit;
 706    i->atomic = atomic;
 707    i->ftz = ftz;
 708    i->dnz = dnz;
 709    i->ipa = ipa;
 710    i->lanes = lanes;
 711    i->perPatch = perPatch;
 712
 713    i->postFactor = postFactor;
 714
 715    for (int d = 0; defExists(d); ++d)
 716       i->setDef(d, pol.get(getDef(d)));
 717
 718    for (int s = 0; srcExists(s); ++s) {
 719       i->setSrc(s, pol.get(getSrc(s)));
 720       i->src(s).mod = src(s).mod;
 721    }
 722
 723    i->cc = cc;
 724    i->predSrc = predSrc;
 725    i->flagsDef = flagsDef;
 726    i->flagsSrc = flagsSrc;
 727
 728    return i;
 729 }
 730
 731 unsigned int
 732 Instruction::defCount(unsigned int mask, bool singleFile) const
 733 {
 734    unsigned int i, n;
 735
 736    if (singleFile) {
 737       unsigned int d = ffs(mask);
 738       if (!d)
 739          return 0;
 740       for (i = d--; defExists(i); ++i)
 741          if (getDef(i)->reg.file != getDef(d)->reg.file)
 742             mask &= ~(1 << i);
 743    }
 744
 745    for (n = 0, i = 0; this->defExists(i); ++i, mask >>= 1)
 746       n += mask & 1;
 747    return n;
 748 }
 749
 750 unsigned int
 751 Instruction::srcCount(unsigned int mask, bool singleFile) const
 752 {
 753    unsigned int i, n;
 754
 755    if (singleFile) {
 756       unsigned int s = ffs(mask);
 757       if (!s)
 758          return 0;
 759       for (i = s--; srcExists(i); ++i)
 760          if (getSrc(i)->reg.file != getSrc(s)->reg.file)
 761             mask &= ~(1 << i);
 762    }
 763
 764    for (n = 0, i = 0; this->srcExists(i); ++i, mask >>= 1)
 765       n += mask & 1;
 766    return n;
 767 }
 768
 769 bool
 770 Instruction::setIndirect(int s, int dim, Value *value)
 771 {
 772    assert(this->srcExists(s));
 773
 774    int p = srcs[s].indirect[dim];
 775    if (p < 0) {
 776       if (!value)
 777          return true;
 778       p = srcs.size();
 779       while (p > 0 && !srcExists(p - 1))
 780          --p;
 781    }
 782    setSrc(p, value);
 783    srcs[p].usedAsPtr = (value != 0);
 784    srcs[s].indirect[dim] = value ? p : -1;
 785    return true;
 786 }
 787
 788 bool
 789 Instruction::setPredicate(CondCode ccode, Value *value)
 790 {
 791    cc = ccode;
 792
 793    if (!value) {
 794       if (predSrc >= 0) {
 795          srcs[predSrc].set(NULL);
 796          predSrc = -1;
 797       }
 798       return true;
 799    }
 800
 801    if (predSrc < 0) {
 802       predSrc = srcs.size();
 803       while (predSrc > 0 && !srcExists(predSrc - 1))
 804          --predSrc;
 805    }
 806
 807    setSrc(predSrc, value);
 808    return true;
 809 }
 810
 811 bool
 812 Instruction::writesPredicate() const
 813 {
 814    for (int d = 0; defExists(d); ++d)
 815       if (getDef(d)->inFile(FILE_PREDICATE) || getDef(d)->inFile(FILE_FLAGS))
 816          return true;
 817    return false;
 818 }
 819
 820 static bool
 821 insnCheckCommutationDefSrc(const Instruction *a, const Instruction *b)
 822 {
 823    for (int d = 0; a->defExists(d); ++d)
 824       for (int s = 0; b->srcExists(s); ++s)
 825          if (a->getDef(d)->interfers(b->getSrc(s)))
 826             return false;
 827    return true;
 828 }
 829
 830 static bool
 831 insnCheckCommutationDefDef(const Instruction *a, const Instruction *b)
 832 {
 833    for (int d = 0; a->defExists(d); ++d)
 834       for (int c = 0; b->defExists(c); ++c)
 835          if (a->getDef(d)->interfers(b->getDef(c)))
 836             return false;
 837    return true;
 838 }
 839
 840 bool
 841 Instruction::isCommutationLegal(const Instruction *i) const
 842 {
 843    bool ret = insnCheckCommutationDefDef(this, i);
 844    ret = ret && insnCheckCommutationDefSrc(this, i);
 845    ret = ret && insnCheckCommutationDefSrc(i, this);
 846    return ret;
 847 }
 848
 849 TexInstruction::TexInstruction(Function *fn, operation op)
 850    : Instruction(fn, op, TYPE_F32)
 851 {
 852    memset(&tex, 0, sizeof(tex));
 853
 854    tex.rIndirectSrc = -1;
 855    tex.sIndirectSrc = -1;
 856 }
 857
 858 TexInstruction::~TexInstruction()
 859 {
 860    for (int c = 0; c < 3; ++c) {
 861       dPdx[c].set(NULL);
 862       dPdy[c].set(NULL);
 863    }
 864 }
 865
 866 TexInstruction *
 867 TexInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 868 {
 869    TexInstruction *tex = (i ? static_cast<TexInstruction *>(i) :
 870                           new_TexInstruction(pol.context(), op));
 871
 872    Instruction::clone(pol, tex);
 873
 874    tex->tex = this->tex;
 875
 876    if (op == OP_TXD) {
 877       for (unsigned int c = 0; c < tex->tex.target.getDim(); ++c) {
 878          tex->dPdx[c].set(dPdx[c]);
 879          tex->dPdy[c].set(dPdy[c]);
 880       }
 881    }
 882
 883    return tex;
 884 }
 885
 886 const struct TexInstruction::Target::Desc TexInstruction::Target::descTable[] =
 887 {
 888    { "1D",                1, 1, false, false, false },
 889    { "2D",                2, 2, false, false, false },
 890    { "2D_MS",             2, 2, false, false, false },
 891    { "3D",                3, 3, false, false, false },
 892    { "CUBE",              2, 3, false, true,  false },
 893    { "1D_SHADOW",         1, 1, false, false, true  },
 894    { "2D_SHADOW",         2, 2, false, false, true  },
 895    { "CUBE_SHADOW",       2, 3, false, true,  true  },
 896    { "1D_ARRAY",          1, 2, true,  false, false },
 897    { "2D_ARRAY",          2, 3, true,  false, false },
 898    { "2D_MS_ARRAY",       2, 3, true,  false, false },
 899    { "CUBE_ARRAY",        2, 4, true,  true,  false },
 900    { "1D_ARRAY_SHADOW",   1, 2, true,  false, true  },
 901    { "2D_ARRAY_SHADOW",   2, 3, true,  false, true  },
 902    { "RECT",              2, 2, false, false, false },
 903    { "RECT_SHADOW",       2, 2, false, false, true  },
 904    { "CUBE_ARRAY_SHADOW", 2, 4, true,  true,  true  },
 905    { "BUFFER",            1, 1, false, false, false },
 906 };
 907
 908 CmpInstruction::CmpInstruction(Function *fn, operation op)
 909    : Instruction(fn, op, TYPE_F32)
 910 {
 911    setCond = CC_ALWAYS;
 912 }
 913
 914 CmpInstruction *
 915 CmpInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 916 {
 917    CmpInstruction *cmp = (i ? static_cast<CmpInstruction *>(i) :
 918                           new_CmpInstruction(pol.context(), op));
 919    cmp->dType = dType;
 920    Instruction::clone(pol, cmp);
 921    cmp->setCond = setCond;
 922    return cmp;
 923 }
 924
 925 FlowInstruction::FlowInstruction(Function *fn, operation op, void *targ)
 926    : Instruction(fn, op, TYPE_NONE)
 927 {
 928    if (op == OP_CALL)
 929       target.fn = reinterpret_cast<Function *>(targ);
 930    else
 931       target.bb = reinterpret_cast<BasicBlock *>(targ);
 932
 933    if (op == OP_BRA ||
 934        op == OP_CONT || op == OP_BREAK ||
 935        op == OP_RET || op == OP_EXIT)
 936       terminator = 1;
 937    else
 938    if (op == OP_JOIN)
 939       terminator = targ ? 1 : 0;
 940
 941    allWarp = absolute = limit = builtin = 0;
 942 }
 943
 944 FlowInstruction *
 945 FlowInstruction::clone(ClonePolicy<Function>& pol, Instruction *i) const
 946 {
 947    FlowInstruction *flow = (i ? static_cast<FlowInstruction *>(i) :
 948                             new_FlowInstruction(pol.context(), op, NULL));
 949
 950    Instruction::clone(pol, flow);
 951    flow->allWarp = allWarp;
 952    flow->absolute = absolute;
 953    flow->limit = limit;
 954    flow->builtin = builtin;
 955
 956    if (builtin)
 957       flow->target.builtin = target.builtin;
 958    else
 959    if (op == OP_CALL)
 960       flow->target.fn = target.fn;
 961    else
 962    if (target.bb)
 963       flow->target.bb = pol.get<BasicBlock>(target.bb);
 964
 965    return flow;
 966 }
 967
 968 Program::Program(Type type, Target *arch)
 969    : progType(type),
 970      target(arch),
 971      mem_Instruction(sizeof(Instruction), 6),
 972      mem_CmpInstruction(sizeof(CmpInstruction), 4),
 973      mem_TexInstruction(sizeof(TexInstruction), 4),
 974      mem_FlowInstruction(sizeof(FlowInstruction), 4),
 975      mem_LValue(sizeof(LValue), 8),
 976      mem_Symbol(sizeof(Symbol), 7),
 977      mem_ImmediateValue(sizeof(ImmediateValue), 7)
 978 {
 979    code = NULL;
 980    binSize = 0;
 981
 982    maxGPR = -1;
 983
 984    main = new Function(this, "MAIN", ~0);
 985    calls.insert(&main->call);
 986
 987    dbgFlags = 0;
 988    optLevel = 0;
 989
 990    targetPriv = NULL;
 991 }
 992
 993 Program::~Program()
 994 {
 995    for (ArrayList::Iterator it = allFuncs.iterator(); !it.end(); it.next())
 996       delete reinterpret_cast<Function *>(it.get());
 997
 998    for (ArrayList::Iterator it = allRValues.iterator(); !it.end(); it.next())
 999       releaseValue(reinterpret_cast<Value *>(it.get()));
1000 }
1001
1002 void Program::releaseInstruction(Instruction *insn)
1003 {
1004    // TODO: make this not suck so much
1005
1006    insn->~Instruction();
1007
1008    if (insn->asCmp())
1009       mem_CmpInstruction.release(insn);
1010    else
1011    if (insn->asTex())
1012       mem_TexInstruction.release(insn);
1013    else
1014    if (insn->asFlow())
1015       mem_FlowInstruction.release(insn);
1016    else
1017       mem_Instruction.release(insn);
1018 }
1019
1020 void Program::releaseValue(Value *value)
1021 {
1022    value->~Value();
1023
1024    if (value->asLValue())
1025       mem_LValue.release(value);
1026    else
1027    if (value->asImm())
1028       mem_ImmediateValue.release(value);
1029    else
1030    if (value->asSym())
1031       mem_Symbol.release(value);
1032 }
1033
1034
1035 } // namespace nv50_ir
1036
1037 extern "C" {
1038
1039 static void
1040 nv50_ir_init_prog_info(struct nv50_ir_prog_info *info)
1041 {
1042 #if defined(PIPE_SHADER_HULL) && defined(PIPE_SHADER_DOMAIN)
1043    if (info->type == PIPE_SHADER_HULL || info->type == PIPE_SHADER_DOMAIN) {
1044       info->prop.tp.domain = PIPE_PRIM_MAX;
1045       info->prop.tp.outputPrim = PIPE_PRIM_MAX;
1046    }
1047 #endif
1048    if (info->type == PIPE_SHADER_GEOMETRY) {
1049       info->prop.gp.instanceCount = 1;
1050       info->prop.gp.maxVertices = 1;
1051    }
1052    info->io.clipDistance = 0xff;
1053    info->io.pointSize = 0xff;
1054    info->io.vertexId = 0xff;
1055    info->io.edgeFlagIn = 0xff;
1056    info->io.edgeFlagOut = 0xff;
1057    info->io.fragDepth = 0xff;
1058    info->io.sampleMask = 0xff;
1059    info->io.backFaceColor[0] = info->io.backFaceColor[1] = 0xff;
1060 }
1061
1062 int
1063 nv50_ir_generate_code(struct nv50_ir_prog_info *info)
1064 {
1065    int ret = 0;
1066
1067    nv50_ir::Program::Type type;
1068
1069    nv50_ir_init_prog_info(info);
1070
1071 #define PROG_TYPE_CASE(a, b)                                      \
1072    case PIPE_SHADER_##a: type = nv50_ir::Program::TYPE_##b; break
1073
1074    switch (info->type) {
1075    PROG_TYPE_CASE(VERTEX, VERTEX);
1076 // PROG_TYPE_CASE(HULL, TESSELLATION_CONTROL);
1077 // PROG_TYPE_CASE(DOMAIN, TESSELLATION_EVAL);
1078    PROG_TYPE_CASE(GEOMETRY, GEOMETRY);
1079    PROG_TYPE_CASE(FRAGMENT, FRAGMENT);
1080    default:
1081       type = nv50_ir::Program::TYPE_COMPUTE;
1082       break;
1083    }
1084    INFO_DBG(info->dbgFlags, VERBOSE, "translating program of type %u\n", type);
1085
1086    nv50_ir::Target *targ = nv50_ir::Target::create(info->target);
1087    if (!targ)
1088       return -1;
1089
1090    nv50_ir::Program *prog = new nv50_ir::Program(type, targ);
1091    if (!prog)
1092       return -1;
1093    prog->dbgFlags = info->dbgFlags;
1094    prog->optLevel = info->optLevel;
1095
1096    switch (info->bin.sourceRep) {
1097 #if 0
1098    case PIPE_IR_LLVM:
1099    case PIPE_IR_GLSL:
1100       return -1;
1101    case PIPE_IR_SM4:
1102       ret = prog->makeFromSM4(info) ? 0 : -2;
1103       break;
1104    case PIPE_IR_TGSI:
1105 #endif
1106    default:
1107       ret = prog->makeFromTGSI(info) ? 0 : -2;
1108       break;
1109    }
1110    if (ret < 0)
1111       goto out;
1112    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
1113       prog->print();
1114
1115    targ->parseDriverInfo(info);
1116    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_PRE_SSA);
1117
1118    prog->convertToSSA();
1119
1120    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
1121       prog->print();
1122
1123    prog->optimizeSSA(info->optLevel);
1124    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_SSA);
1125
1126    if (prog->dbgFlags & NV50_IR_DEBUG_BASIC)
1127       prog->print();
1128
1129    if (!prog->registerAllocation()) {
1130       ret = -4;
1131       goto out;
1132    }
1133    prog->getTarget()->runLegalizePass(prog, nv50_ir::CG_STAGE_POST_RA);
1134
1135    prog->optimizePostRA(info->optLevel);
1136
1137    if (!prog->emitBinary(info)) {
1138       ret = -5;
1139       goto out;
1140    }
1141
1142 out:
1143    INFO_DBG(prog->dbgFlags, VERBOSE, "nv50_ir_generate_code: ret = %i\n", ret);
1144
1145    info->bin.maxGPR = prog->maxGPR;
1146    info->bin.code = prog->code;
1147    info->bin.codeSize = prog->binSize;
1148    info->bin.tlsSpace = prog->tlsSize;
1149
1150    delete prog;
1151    nv50_ir::Target::destroy(targ);
1152
1153    return ret;
1154 }
1155
1156 } // extern "C"