src/gallium/drivers/nouveau/codegen/nv50_ir_target.cpp

   1 /*
   2  * Copyright 2011 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include "codegen/nv50_ir.h"
  24 #include "codegen/nv50_ir_target.h"
  25
  26 namespace nv50_ir {
  27
  28 const uint8_t Target::operationSrcNr[] =
  29 {
  30    0, 0,                   // NOP, PHI
  31    0, 0, 0, 0,             // UNION, SPLIT, MERGE, CONSTRAINT
  32    1, 1, 2,                // MOV, LOAD, STORE
  33    2, 2, 2, 2, 2, 3, 3, 3, // ADD, SUB, MUL, DIV, MOD, MAD, FMA, SAD
  34    1, 1, 1,                // ABS, NEG, NOT
  35    2, 2, 2, 2, 2,          // AND, OR, XOR, SHL, SHR
  36    2, 2, 1,                // MAX, MIN, SAT
  37    1, 1, 1, 1,             // CEIL, FLOOR, TRUNC, CVT
  38    3, 3, 3, 2, 3, 3,       // SET_AND,OR,XOR, SET, SELP, SLCT
  39    1, 1, 1, 1, 1, 1,       // RCP, RSQ, LG2, SIN, COS, EX2
  40    1, 1, 1, 1, 1, 2,       // EXP, LOG, PRESIN, PREEX2, SQRT, POW
  41    0, 0, 0, 0, 0,          // BRA, CALL, RET, CONT, BREAK,
  42    0, 0, 0,                // PRERET,CONT,BREAK
  43    0, 0, 0, 0, 0, 0,       // BRKPT, JOINAT, JOIN, DISCARD, EXIT, MEMBAR
  44    1, 1, 1, 2, 1, 2,       // VFETCH, PFETCH, AFETCH, EXPORT, LINTERP, PINTERP
  45    1, 1,                   // EMIT, RESTART
  46    1, 1, 1,                // TEX, TXB, TXL,
  47    1, 1, 1, 1, 1, 1, 2,    // TXF, TXQ, TXD, TXG, TXLQ, TEXCSAA, TEXPREP
  48    1, 1, 2, 2, 2, 2, 2,    // SULDB, SULDP, SUSTB, SUSTP, SUREDB, SUREDP, SULEA
  49    3, 3, 3, 1, 3,          // SUBFM, SUCLAMP, SUEAU, SUQ, MADSP
  50    0,                      // TEXBAR
  51    1, 1,                   // DFDX, DFDY
  52    1, 2, 1, 2, 0, 0,       // RDSV, WRSV, PIXLD, QUADOP, QUADON, QUADPOP
  53    2, 3, 2, 1, 3,          // POPCNT, INSBF, EXTBF, BFIND, PERMT
  54    2, 2,                   // ATOM, BAR
  55    2, 2, 2, 2, 3, 2,       // VADD, VAVG, VMIN, VMAX, VSAD, VSET,
  56    2, 2, 2, 1,             // VSHR, VSHL, VSEL, CCTL
  57    3,                      // SHFL
  58    0
  59 };
  60
  61 const OpClass Target::operationClass[] =
  62 {
  63    // NOP; PHI; UNION, SPLIT, MERGE, CONSTRAINT
  64    OPCLASS_OTHER,
  65    OPCLASS_PSEUDO,
  66    OPCLASS_PSEUDO, OPCLASS_PSEUDO, OPCLASS_PSEUDO, OPCLASS_PSEUDO,
  67    // MOV; LOAD; STORE
  68    OPCLASS_MOVE,
  69    OPCLASS_LOAD,
  70    OPCLASS_STORE,
  71    // ADD, SUB, MUL; DIV, MOD; MAD, FMA, SAD
  72    OPCLASS_ARITH, OPCLASS_ARITH, OPCLASS_ARITH,
  73    OPCLASS_ARITH, OPCLASS_ARITH,
  74    OPCLASS_ARITH, OPCLASS_ARITH, OPCLASS_ARITH,
  75    // ABS, NEG; NOT, AND, OR, XOR; SHL, SHR
  76    OPCLASS_CONVERT, OPCLASS_CONVERT,
  77    OPCLASS_LOGIC, OPCLASS_LOGIC, OPCLASS_LOGIC, OPCLASS_LOGIC,
  78    OPCLASS_SHIFT, OPCLASS_SHIFT,
  79    // MAX, MIN
  80    OPCLASS_COMPARE, OPCLASS_COMPARE,
  81    // SAT, CEIL, FLOOR, TRUNC; CVT
  82    OPCLASS_CONVERT, OPCLASS_CONVERT, OPCLASS_CONVERT, OPCLASS_CONVERT,
  83    OPCLASS_CONVERT,
  84    // SET(AND,OR,XOR); SELP, SLCT
  85    OPCLASS_COMPARE, OPCLASS_COMPARE, OPCLASS_COMPARE, OPCLASS_COMPARE,
  86    OPCLASS_COMPARE, OPCLASS_COMPARE,
  87    // RCP, RSQ, LG2, SIN, COS; EX2, EXP, LOG, PRESIN, PREEX2; SQRT, POW
  88    OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU,
  89    OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU,
  90    OPCLASS_SFU, OPCLASS_SFU,
  91    // BRA, CALL, RET; CONT, BREAK, PRE(RET,CONT,BREAK); BRKPT, JOINAT, JOIN
  92    OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW,
  93    OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW,
  94    OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW,
  95    // DISCARD, EXIT
  96    OPCLASS_FLOW, OPCLASS_FLOW,
  97    // MEMBAR
  98    OPCLASS_CONTROL,
  99    // VFETCH, PFETCH, AFETCH, EXPORT
 100    OPCLASS_LOAD, OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_STORE,
 101    // LINTERP, PINTERP
 102    OPCLASS_SFU, OPCLASS_SFU,
 103    // EMIT, RESTART
 104    OPCLASS_CONTROL, OPCLASS_CONTROL,
 105    // TEX, TXB, TXL, TXF; TXQ, TXD, TXG, TXLQ; TEXCSAA, TEXPREP
 106    OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE,
 107    OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE,
 108    OPCLASS_TEXTURE, OPCLASS_TEXTURE,
 109    // SULDB, SULDP, SUSTB, SUSTP; SUREDB, SUREDP, SULEA
 110    OPCLASS_SURFACE, OPCLASS_SURFACE, OPCLASS_ATOMIC, OPCLASS_SURFACE,
 111    OPCLASS_SURFACE, OPCLASS_SURFACE, OPCLASS_SURFACE,
 112    // SUBFM, SUCLAMP, SUEAU, SUQ, MADSP
 113    OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_ARITH,
 114    // TEXBAR
 115    OPCLASS_OTHER,
 116    // DFDX, DFDY, RDSV, WRSV; PIXLD, QUADOP, QUADON, QUADPOP
 117    OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER,
 118    OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_CONTROL, OPCLASS_CONTROL,
 119    // POPCNT, INSBF, EXTBF, BFIND; PERMT
 120    OPCLASS_BITFIELD, OPCLASS_BITFIELD, OPCLASS_BITFIELD, OPCLASS_BITFIELD,
 121    OPCLASS_BITFIELD,
 122    // ATOM, BAR
 123    OPCLASS_ATOMIC, OPCLASS_CONTROL,
 124    // VADD, VAVG, VMIN, VMAX
 125    OPCLASS_VECTOR, OPCLASS_VECTOR, OPCLASS_VECTOR, OPCLASS_VECTOR,
 126    // VSAD, VSET, VSHR, VSHL
 127    OPCLASS_VECTOR, OPCLASS_VECTOR, OPCLASS_VECTOR, OPCLASS_VECTOR,
 128    // VSEL, CCTL
 129    OPCLASS_VECTOR, OPCLASS_CONTROL,
 130    // SHFL
 131    OPCLASS_OTHER,
 132    OPCLASS_PSEUDO // LAST
 133 };
 134
 135
 136 extern Target *getTargetGM107(unsigned int chipset);
 137 extern Target *getTargetNVC0(unsigned int chipset);
 138 extern Target *getTargetNV50(unsigned int chipset);
 139
 140 Target *Target::create(unsigned int chipset)
 141 {
 142    STATIC_ASSERT(Elements(operationSrcNr) == OP_LAST + 1);
 143    STATIC_ASSERT(Elements(operationClass) == OP_LAST + 1);
 144    switch (chipset & ~0xf) {
 145    case 0x110:
 146    case 0x120:
 147       return getTargetGM107(chipset);
 148    case 0xc0:
 149    case 0xd0:
 150    case 0xe0:
 151    case 0xf0:
 152    case 0x100:
 153       return getTargetNVC0(chipset);
 154    case 0x50:
 155    case 0x80:
 156    case 0x90:
 157    case 0xa0:
 158       return getTargetNV50(chipset);
 159    default:
 160       ERROR("unsupported target: NV%x\n", chipset);
 161       return 0;
 162    }
 163 }
 164
 165 void Target::destroy(Target *targ)
 166 {
 167    delete targ;
 168 }
 169
 170 CodeEmitter::CodeEmitter(const Target *target) : targ(target), interpInfo(NULL)
 171 {
 172 }
 173
 174 void
 175 CodeEmitter::setCodeLocation(void *ptr, uint32_t size)
 176 {
 177    code = reinterpret_cast<uint32_t *>(ptr);
 178    codeSize = 0;
 179    codeSizeLimit = size;
 180 }
 181
 182 void
 183 CodeEmitter::printBinary() const
 184 {
 185    uint32_t *bin = code - codeSize / 4;
 186    INFO("program binary (%u bytes)", codeSize);
 187    for (unsigned int pos = 0; pos < codeSize / 4; ++pos) {
 188       if ((pos % 8) == 0)
 189          INFO("\n");
 190       INFO("%08x ", bin[pos]);
 191    }
 192    INFO("\n");
 193 }
 194
 195 static inline uint32_t sizeToBundlesNVE4(uint32_t size)
 196 {
 197    return (size + 55) / 56;
 198 }
 199
 200 void
 201 CodeEmitter::prepareEmission(Program *prog)
 202 {
 203    for (ArrayList::Iterator fi = prog->allFuncs.iterator();
 204         !fi.end(); fi.next()) {
 205       Function *func = reinterpret_cast<Function *>(fi.get());
 206       func->binPos = prog->binSize;
 207       prepareEmission(func);
 208
 209       // adjust sizes & positions for schedulding info:
 210       if (prog->getTarget()->hasSWSched) {
 211          uint32_t adjPos = func->binPos;
 212          BasicBlock *bb = NULL;
 213          for (int i = 0; i < func->bbCount; ++i) {
 214             bb = func->bbArray[i];
 215             int32_t adjSize = bb->binSize;
 216             if (adjPos % 64) {
 217                adjSize -= 64 - adjPos % 64;
 218                if (adjSize < 0)
 219                   adjSize = 0;
 220             }
 221             adjSize = bb->binSize + sizeToBundlesNVE4(adjSize) * 8;
 222             bb->binPos = adjPos;
 223             bb->binSize = adjSize;
 224             adjPos += adjSize;
 225          }
 226          if (bb)
 227             func->binSize = adjPos - func->binPos;
 228       }
 229
 230       prog->binSize += func->binSize;
 231    }
 232 }
 233
 234 void
 235 CodeEmitter::prepareEmission(Function *func)
 236 {
 237    func->bbCount = 0;
 238    func->bbArray = new BasicBlock * [func->cfg.getSize()];
 239
 240    BasicBlock::get(func->cfg.getRoot())->binPos = func->binPos;
 241
 242    for (IteratorRef it = func->cfg.iteratorCFG(); !it->end(); it->next())
 243       prepareEmission(BasicBlock::get(*it));
 244 }
 245
 246 void
 247 CodeEmitter::prepareEmission(BasicBlock *bb)
 248 {
 249    Instruction *i, *next;
 250    Function *func = bb->getFunction();
 251    int j;
 252    unsigned int nShort;
 253
 254    for (j = func->bbCount - 1; j >= 0 && !func->bbArray[j]->binSize; --j);
 255
 256    for (; j >= 0; --j) {
 257       BasicBlock *in = func->bbArray[j];
 258       Instruction *exit = in->getExit();
 259
 260       if (exit && exit->op == OP_BRA && exit->asFlow()->target.bb == bb) {
 261          in->binSize -= 8;
 262          func->binSize -= 8;
 263
 264          for (++j; j < func->bbCount; ++j)
 265             func->bbArray[j]->binPos -= 8;
 266
 267          in->remove(exit);
 268       }
 269       bb->binPos = in->binPos + in->binSize;
 270       if (in->binSize) // no more no-op branches to bb
 271          break;
 272    }
 273    func->bbArray[func->bbCount++] = bb;
 274
 275    if (!bb->getExit())
 276       return;
 277
 278    // determine encoding size, try to group short instructions
 279    nShort = 0;
 280    for (i = bb->getEntry(); i; i = next) {
 281       next = i->next;
 282
 283       if (i->op == OP_MEMBAR && !targ->isOpSupported(OP_MEMBAR, TYPE_NONE)) {
 284          bb->remove(i);
 285          continue;
 286       }
 287
 288       i->encSize = getMinEncodingSize(i);
 289       if (next && i->encSize < 8)
 290          ++nShort;
 291       else
 292       if ((nShort & 1) && next && getMinEncodingSize(next) == 4) {
 293          if (i->isCommutationLegal(i->next)) {
 294             bb->permuteAdjacent(i, next);
 295             next->encSize = 4;
 296             next = i;
 297             i = i->prev;
 298             ++nShort;
 299          } else
 300          if (i->isCommutationLegal(i->prev) && next->next) {
 301             bb->permuteAdjacent(i->prev, i);
 302             next->encSize = 4;
 303             next = next->next;
 304             bb->binSize += 4;
 305             ++nShort;
 306          } else {
 307             i->encSize = 8;
 308             i->prev->encSize = 8;
 309             bb->binSize += 4;
 310             nShort = 0;
 311          }
 312       } else {
 313          i->encSize = 8;
 314          if (nShort & 1) {
 315             i->prev->encSize = 8;
 316             bb->binSize += 4;
 317          }
 318          nShort = 0;
 319       }
 320       bb->binSize += i->encSize;
 321    }
 322
 323    if (bb->getExit()->encSize == 4) {
 324       assert(nShort);
 325       bb->getExit()->encSize = 8;
 326       bb->binSize += 4;
 327
 328       if ((bb->getExit()->prev->encSize == 4) && !(nShort & 1)) {
 329          bb->binSize += 8;
 330          bb->getExit()->prev->encSize = 8;
 331       }
 332    }
 333    assert(!bb->getEntry() || (bb->getExit() && bb->getExit()->encSize == 8));
 334
 335    func->binSize += bb->binSize;
 336 }
 337
 338 void
 339 Program::emitSymbolTable(struct nv50_ir_prog_info *info)
 340 {
 341    unsigned int n = 0, nMax = allFuncs.getSize();
 342
 343    info->bin.syms =
 344       (struct nv50_ir_prog_symbol *)MALLOC(nMax * sizeof(*info->bin.syms));
 345
 346    for (ArrayList::Iterator fi = allFuncs.iterator();
 347         !fi.end();
 348         fi.next(), ++n) {
 349       Function *f = (Function *)fi.get();
 350       assert(n < nMax);
 351
 352       info->bin.syms[n].label = f->getLabel();
 353       info->bin.syms[n].offset = f->binPos;
 354    }
 355
 356    info->bin.numSyms = n;
 357 }
 358
 359 bool
 360 Program::emitBinary(struct nv50_ir_prog_info *info)
 361 {
 362    CodeEmitter *emit = target->getCodeEmitter(progType);
 363
 364    emit->prepareEmission(this);
 365
 366    if (dbgFlags & NV50_IR_DEBUG_BASIC)
 367       this->print();
 368
 369    if (!binSize) {
 370       code = NULL;
 371       return false;
 372    }
 373    code = reinterpret_cast<uint32_t *>(MALLOC(binSize));
 374    if (!code)
 375       return false;
 376    emit->setCodeLocation(code, binSize);
 377    info->bin.instructions = 0;
 378
 379    for (ArrayList::Iterator fi = allFuncs.iterator(); !fi.end(); fi.next()) {
 380       Function *fn = reinterpret_cast<Function *>(fi.get());
 381
 382       assert(emit->getCodeSize() == fn->binPos);
 383
 384       for (int b = 0; b < fn->bbCount; ++b) {
 385          for (Instruction *i = fn->bbArray[b]->getEntry(); i; i = i->next) {
 386             emit->emitInstruction(i);
 387             info->bin.instructions++;
 388             if (i->sType == TYPE_F64 || i->dType == TYPE_F64)
 389                info->io.fp64 = true;
 390          }
 391       }
 392    }
 393    info->bin.relocData = emit->getRelocInfo();
 394    info->bin.interpData = emit->getInterpInfo();
 395
 396    emitSymbolTable(info);
 397
 398    // the nvc0 driver will print the binary iself together with the header
 399    if ((dbgFlags & NV50_IR_DEBUG_BASIC) && getTarget()->getChipset() < 0xc0)
 400       emit->printBinary();
 401
 402    delete emit;
 403    return true;
 404 }
 405
 406 #define RELOC_ALLOC_INCREMENT 8
 407
 408 bool
 409 CodeEmitter::addReloc(RelocEntry::Type ty, int w, uint32_t data, uint32_t m,
 410                       int s)
 411 {
 412    unsigned int n = relocInfo ? relocInfo->count : 0;
 413
 414    if (!(n % RELOC_ALLOC_INCREMENT)) {
 415       size_t size = sizeof(RelocInfo) + n * sizeof(RelocEntry);
 416       relocInfo = reinterpret_cast<RelocInfo *>(
 417          REALLOC(relocInfo, n ? size : 0,
 418                  size + RELOC_ALLOC_INCREMENT * sizeof(RelocEntry)));
 419       if (!relocInfo)
 420          return false;
 421       if (n == 0)
 422          memset(relocInfo, 0, sizeof(RelocInfo));
 423    }
 424    ++relocInfo->count;
 425
 426    relocInfo->entry[n].data = data;
 427    relocInfo->entry[n].mask = m;
 428    relocInfo->entry[n].offset = codeSize + w * 4;
 429    relocInfo->entry[n].bitPos = s;
 430    relocInfo->entry[n].type = ty;
 431
 432    return true;
 433 }
 434
 435 bool
 436 CodeEmitter::addInterp(int ipa, int reg, InterpApply apply)
 437 {
 438    unsigned int n = interpInfo ? interpInfo->count : 0;
 439
 440    if (!(n % RELOC_ALLOC_INCREMENT)) {
 441       size_t size = sizeof(InterpInfo) + n * sizeof(InterpEntry);
 442       interpInfo = reinterpret_cast<InterpInfo *>(
 443          REALLOC(interpInfo, n ? size : 0,
 444                  size + RELOC_ALLOC_INCREMENT * sizeof(InterpEntry)));
 445       if (!interpInfo)
 446          return false;
 447       if (n == 0)
 448          memset(interpInfo, 0, sizeof(InterpInfo));
 449    }
 450    ++interpInfo->count;
 451
 452    interpInfo->entry[n] = InterpEntry(ipa, reg, codeSize >> 2);
 453    interpInfo->apply = apply;
 454
 455    return true;
 456 }
 457
 458 void
 459 RelocEntry::apply(uint32_t *binary, const RelocInfo *info) const
 460 {
 461    uint32_t value = 0;
 462
 463    switch (type) {
 464    case TYPE_CODE: value = info->codePos; break;
 465    case TYPE_BUILTIN: value = info->libPos; break;
 466    case TYPE_DATA: value = info->dataPos; break;
 467    default:
 468       assert(0);
 469       break;
 470    }
 471    value += data;
 472    value = (bitPos < 0) ? (value >> -bitPos) : (value << bitPos);
 473
 474    binary[offset / 4] &= ~mask;
 475    binary[offset / 4] |= value & mask;
 476 }
 477
 478 } // namespace nv50_ir
 479
 480
 481 #include "codegen/nv50_ir_driver.h"
 482
 483 extern "C" {
 484
 485 void
 486 nv50_ir_relocate_code(void *relocData, uint32_t *code,
 487                       uint32_t codePos,
 488                       uint32_t libPos,
 489                       uint32_t dataPos)
 490 {
 491    nv50_ir::RelocInfo *info = reinterpret_cast<nv50_ir::RelocInfo *>(relocData);
 492
 493    info->codePos = codePos;
 494    info->libPos = libPos;
 495    info->dataPos = dataPos;
 496
 497    for (unsigned int i = 0; i < info->count; ++i)
 498       info->entry[i].apply(code, info);
 499 }
 500
 501 void
 502 nv50_ir_change_interp(void *interpData, uint32_t *code,
 503                       bool force_persample_interp, bool flatshade)
 504 {
 505    nv50_ir::InterpInfo *info = reinterpret_cast<nv50_ir::InterpInfo *>(
 506       interpData);
 507
 508    // force_persample_interp: all non-flat -> per-sample
 509    // flatshade: all color -> flat
 510    for (unsigned i = 0; i < info->count; ++i)
 511       info->apply(&info->entry[i], code, force_persample_interp, flatshade);
 512 }
 513
 514 void
 515 nv50_ir_get_target_library(uint32_t chipset,
 516                            const uint32_t **code, uint32_t *size)
 517 {
 518    nv50_ir::Target *targ = nv50_ir::Target::create(chipset);
 519    targ->getBuiltinCode(code, size);
 520    nv50_ir::Target::destroy(targ);
 521 }
 522
 523 }