src/mesa/program/program.c

   1 /*
   2  * Mesa 3-D graphics library
   3  * Version:  6.5.3
   4  *
   5  * Copyright (C) 1999-2007  Brian Paul   All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included
  15  * in all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  21  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  22  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 /**
  26  * \file program.c
  27  * Vertex and fragment program support functions.
  28  * \author Brian Paul
  29  */
  30
  31
  32 #include "main/glheader.h"
  33 #include "main/context.h"
  34 #include "main/hash.h"
  35 #include "program.h"
  36 #include "prog_cache.h"
  37 #include "prog_parameter.h"
  38 #include "prog_instruction.h"
  39
  40
  41 /**
  42  * A pointer to this dummy program is put into the hash table when
  43  * glGenPrograms is called.
  44  */
  45 struct gl_program _mesa_DummyProgram;
  46
  47
  48 /**
  49  * Init context's vertex/fragment program state
  50  */
  51 void
  52 _mesa_init_program(struct gl_context *ctx)
  53 {
  54    /*
  55     * If this assertion fails, we need to increase the field
  56     * size for register indexes (see INST_INDEX_BITS).
  57     */
  58    ASSERT(ctx->Const.VertexProgram.MaxUniformComponents / 4
  59           <= (1 << INST_INDEX_BITS));
  60    ASSERT(ctx->Const.FragmentProgram.MaxUniformComponents / 4
  61           <= (1 << INST_INDEX_BITS));
  62
  63    ASSERT(ctx->Const.VertexProgram.MaxTemps <= (1 << INST_INDEX_BITS));
  64    ASSERT(ctx->Const.VertexProgram.MaxLocalParams <= (1 << INST_INDEX_BITS));
  65    ASSERT(ctx->Const.FragmentProgram.MaxTemps <= (1 << INST_INDEX_BITS));
  66    ASSERT(ctx->Const.FragmentProgram.MaxLocalParams <= (1 << INST_INDEX_BITS));
  67
  68    ASSERT(ctx->Const.VertexProgram.MaxUniformComponents <= 4 * MAX_UNIFORMS);
  69    ASSERT(ctx->Const.FragmentProgram.MaxUniformComponents <= 4 * MAX_UNIFORMS);
  70
  71    ASSERT(ctx->Const.VertexProgram.MaxAddressOffset <= (1 << INST_INDEX_BITS));
  72    ASSERT(ctx->Const.FragmentProgram.MaxAddressOffset <= (1 << INST_INDEX_BITS));
  73
  74    /* If this fails, increase prog_instruction::TexSrcUnit size */
  75    ASSERT(MAX_TEXTURE_UNITS <= (1 << 5));
  76
  77    /* If this fails, increase prog_instruction::TexSrcTarget size */
  78    ASSERT(NUM_TEXTURE_TARGETS <= (1 << 4));
  79
  80    ctx->Program.ErrorPos = -1;
  81    ctx->Program.ErrorString = _mesa_strdup("");
  82
  83    ctx->VertexProgram.Enabled = GL_FALSE;
  84    ctx->VertexProgram.PointSizeEnabled =
  85       (ctx->API == API_OPENGLES2) ? GL_TRUE : GL_FALSE;
  86    ctx->VertexProgram.TwoSideEnabled = GL_FALSE;
  87    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
  88                             ctx->Shared->DefaultVertexProgram);
  89    assert(ctx->VertexProgram.Current);
  90    ctx->VertexProgram.Cache = _mesa_new_program_cache();
  91
  92    ctx->FragmentProgram.Enabled = GL_FALSE;
  93    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
  94                             ctx->Shared->DefaultFragmentProgram);
  95    assert(ctx->FragmentProgram.Current);
  96    ctx->FragmentProgram.Cache = _mesa_new_program_cache();
  97
  98    ctx->GeometryProgram.Enabled = GL_FALSE;
  99    /* right now by default we don't have a geometry program */
 100    _mesa_reference_geomprog(ctx, &ctx->GeometryProgram.Current,
 101                             NULL);
 102    ctx->GeometryProgram.Cache = _mesa_new_program_cache();
 103
 104    /* XXX probably move this stuff */
 105    ctx->ATIFragmentShader.Enabled = GL_FALSE;
 106    ctx->ATIFragmentShader.Current = ctx->Shared->DefaultFragmentShader;
 107    assert(ctx->ATIFragmentShader.Current);
 108    ctx->ATIFragmentShader.Current->RefCount++;
 109 }
 110
 111
 112 /**
 113  * Free a context's vertex/fragment program state
 114  */
 115 void
 116 _mesa_free_program_data(struct gl_context *ctx)
 117 {
 118    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current, NULL);
 119    _mesa_delete_program_cache(ctx, ctx->VertexProgram.Cache);
 120    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current, NULL);
 121    _mesa_delete_shader_cache(ctx, ctx->FragmentProgram.Cache);
 122    _mesa_reference_geomprog(ctx, &ctx->GeometryProgram.Current, NULL);
 123    _mesa_delete_program_cache(ctx, ctx->GeometryProgram.Cache);
 124
 125    /* XXX probably move this stuff */
 126    if (ctx->ATIFragmentShader.Current) {
 127       ctx->ATIFragmentShader.Current->RefCount--;
 128       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
 129          free(ctx->ATIFragmentShader.Current);
 130       }
 131    }
 132
 133    free((void *) ctx->Program.ErrorString);
 134 }
 135
 136
 137 /**
 138  * Update the default program objects in the given context to reference those
 139  * specified in the shared state and release those referencing the old
 140  * shared state.
 141  */
 142 void
 143 _mesa_update_default_objects_program(struct gl_context *ctx)
 144 {
 145    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
 146                             ctx->Shared->DefaultVertexProgram);
 147    assert(ctx->VertexProgram.Current);
 148
 149    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
 150                             ctx->Shared->DefaultFragmentProgram);
 151    assert(ctx->FragmentProgram.Current);
 152
 153    _mesa_reference_geomprog(ctx, &ctx->GeometryProgram.Current,
 154                       ctx->Shared->DefaultGeometryProgram);
 155
 156    /* XXX probably move this stuff */
 157    if (ctx->ATIFragmentShader.Current) {
 158       ctx->ATIFragmentShader.Current->RefCount--;
 159       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
 160          free(ctx->ATIFragmentShader.Current);
 161       }
 162    }
 163    ctx->ATIFragmentShader.Current = (struct ati_fragment_shader *) ctx->Shared->DefaultFragmentShader;
 164    assert(ctx->ATIFragmentShader.Current);
 165    ctx->ATIFragmentShader.Current->RefCount++;
 166 }
 167
 168
 169 /**
 170  * Set the vertex/fragment program error state (position and error string).
 171  * This is generally called from within the parsers.
 172  */
 173 void
 174 _mesa_set_program_error(struct gl_context *ctx, GLint pos, const char *string)
 175 {
 176    ctx->Program.ErrorPos = pos;
 177    free((void *) ctx->Program.ErrorString);
 178    if (!string)
 179       string = "";
 180    ctx->Program.ErrorString = _mesa_strdup(string);
 181 }
 182
 183
 184 /**
 185  * Find the line number and column for 'pos' within 'string'.
 186  * Return a copy of the line which contains 'pos'.  Free the line with
 187  * free().
 188  * \param string  the program string
 189  * \param pos     the position within the string
 190  * \param line    returns the line number corresponding to 'pos'.
 191  * \param col     returns the column number corresponding to 'pos'.
 192  * \return copy of the line containing 'pos'.
 193  */
 194 const GLubyte *
 195 _mesa_find_line_column(const GLubyte *string, const GLubyte *pos,
 196                        GLint *line, GLint *col)
 197 {
 198    const GLubyte *lineStart = string;
 199    const GLubyte *p = string;
 200    GLubyte *s;
 201    int len;
 202
 203    *line = 1;
 204
 205    while (p != pos) {
 206       if (*p == (GLubyte) '\n') {
 207          (*line)++;
 208          lineStart = p + 1;
 209       }
 210       p++;
 211    }
 212
 213    *col = (pos - lineStart) + 1;
 214
 215    /* return copy of this line */
 216    while (*p != 0 && *p != '\n')
 217       p++;
 218    len = p - lineStart;
 219    s = malloc(len + 1);
 220    memcpy(s, lineStart, len);
 221    s[len] = 0;
 222
 223    return s;
 224 }
 225
 226
 227 /**
 228  * Initialize a new vertex/fragment program object.
 229  */
 230 static struct gl_program *
 231 _mesa_init_program_struct( struct gl_context *ctx, struct gl_program *prog,
 232                            GLenum target, GLuint id)
 233 {
 234    (void) ctx;
 235    if (prog) {
 236       GLuint i;
 237       memset(prog, 0, sizeof(*prog));
 238       prog->Id = id;
 239       prog->Target = target;
 240       prog->RefCount = 1;
 241       prog->Format = GL_PROGRAM_FORMAT_ASCII_ARB;
 242
 243       /* default mapping from samplers to texture units */
 244       for (i = 0; i < MAX_SAMPLERS; i++)
 245          prog->SamplerUnits[i] = i;
 246    }
 247
 248    return prog;
 249 }
 250
 251
 252 /**
 253  * Initialize a new fragment program object.
 254  */
 255 struct gl_program *
 256 _mesa_init_fragment_program( struct gl_context *ctx, struct gl_fragment_program *prog,
 257                              GLenum target, GLuint id)
 258 {
 259    if (prog)
 260       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 261    else
 262       return NULL;
 263 }
 264
 265
 266 /**
 267  * Initialize a new vertex program object.
 268  */
 269 struct gl_program *
 270 _mesa_init_vertex_program( struct gl_context *ctx, struct gl_vertex_program *prog,
 271                            GLenum target, GLuint id)
 272 {
 273    if (prog)
 274       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 275    else
 276       return NULL;
 277 }
 278
 279
 280 /**
 281  * Initialize a new geometry program object.
 282  */
 283 struct gl_program *
 284 _mesa_init_geometry_program( struct gl_context *ctx, struct gl_geometry_program *prog,
 285                              GLenum target, GLuint id)
 286 {
 287    if (prog)
 288       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 289    else
 290       return NULL;
 291 }
 292
 293
 294 /**
 295  * Allocate and initialize a new fragment/vertex program object but
 296  * don't put it into the program hash table.  Called via
 297  * ctx->Driver.NewProgram.  May be overridden (ie. replaced) by a
 298  * device driver function to implement OO deriviation with additional
 299  * types not understood by this function.
 300  *
 301  * \param ctx  context
 302  * \param id   program id/number
 303  * \param target  program target/type
 304  * \return  pointer to new program object
 305  */
 306 struct gl_program *
 307 _mesa_new_program(struct gl_context *ctx, GLenum target, GLuint id)
 308 {
 309    struct gl_program *prog;
 310    switch (target) {
 311    case GL_VERTEX_PROGRAM_ARB: /* == GL_VERTEX_PROGRAM_NV */
 312       prog = _mesa_init_vertex_program(ctx, CALLOC_STRUCT(gl_vertex_program),
 313                                        target, id );
 314       break;
 315    case GL_FRAGMENT_PROGRAM_NV:
 316    case GL_FRAGMENT_PROGRAM_ARB:
 317       prog =_mesa_init_fragment_program(ctx,
 318                                          CALLOC_STRUCT(gl_fragment_program),
 319                                          target, id );
 320       break;
 321    case MESA_GEOMETRY_PROGRAM:
 322       prog = _mesa_init_geometry_program(ctx,
 323                                          CALLOC_STRUCT(gl_geometry_program),
 324                                          target, id);
 325       break;
 326    default:
 327       _mesa_problem(ctx, "bad target in _mesa_new_program");
 328       prog = NULL;
 329    }
 330    return prog;
 331 }
 332
 333
 334 /**
 335  * Delete a program and remove it from the hash table, ignoring the
 336  * reference count.
 337  * Called via ctx->Driver.DeleteProgram.  May be wrapped (OO deriviation)
 338  * by a device driver function.
 339  */
 340 void
 341 _mesa_delete_program(struct gl_context *ctx, struct gl_program *prog)
 342 {
 343    (void) ctx;
 344    ASSERT(prog);
 345    ASSERT(prog->RefCount==0);
 346
 347    if (prog == &_mesa_DummyProgram)
 348       return;
 349
 350    free(prog->String);
 351
 352    if (prog->Instructions) {
 353       _mesa_free_instructions(prog->Instructions, prog->NumInstructions);
 354    }
 355    if (prog->Parameters) {
 356       _mesa_free_parameter_list(prog->Parameters);
 357    }
 358
 359    free(prog);
 360 }
 361
 362
 363 /**
 364  * Return the gl_program object for a given ID.
 365  * Basically just a wrapper for _mesa_HashLookup() to avoid a lot of
 366  * casts elsewhere.
 367  */
 368 struct gl_program *
 369 _mesa_lookup_program(struct gl_context *ctx, GLuint id)
 370 {
 371    if (id)
 372       return (struct gl_program *) _mesa_HashLookup(ctx->Shared->Programs, id);
 373    else
 374       return NULL;
 375 }
 376
 377
 378 /**
 379  * Reference counting for vertex/fragment programs
 380  * This is normally only called from the _mesa_reference_program() macro
 381  * when there's a real pointer change.
 382  */
 383 void
 384 _mesa_reference_program_(struct gl_context *ctx,
 385                          struct gl_program **ptr,
 386                          struct gl_program *prog)
 387 {
 388 #ifndef NDEBUG
 389    assert(ptr);
 390    if (*ptr && prog) {
 391       /* sanity check */
 392       if ((*ptr)->Target == GL_VERTEX_PROGRAM_ARB)
 393          ASSERT(prog->Target == GL_VERTEX_PROGRAM_ARB);
 394       else if ((*ptr)->Target == GL_FRAGMENT_PROGRAM_ARB)
 395          ASSERT(prog->Target == GL_FRAGMENT_PROGRAM_ARB ||
 396                 prog->Target == GL_FRAGMENT_PROGRAM_NV);
 397       else if ((*ptr)->Target == MESA_GEOMETRY_PROGRAM)
 398          ASSERT(prog->Target == MESA_GEOMETRY_PROGRAM);
 399    }
 400 #endif
 401
 402    if (*ptr) {
 403       GLboolean deleteFlag;
 404
 405       /*_glthread_LOCK_MUTEX((*ptr)->Mutex);*/
 406 #if 0
 407       printf("Program %p ID=%u Target=%s  Refcount-- to %d\n",
 408              *ptr, (*ptr)->Id,
 409              ((*ptr)->Target == GL_VERTEX_PROGRAM_ARB ? "VP" :
 410               ((*ptr)->Target == MESA_GEOMETRY_PROGRAM ? "GP" : "FP")),
 411              (*ptr)->RefCount - 1);
 412 #endif
 413       ASSERT((*ptr)->RefCount > 0);
 414       (*ptr)->RefCount--;
 415
 416       deleteFlag = ((*ptr)->RefCount == 0);
 417       /*_glthread_UNLOCK_MUTEX((*ptr)->Mutex);*/
 418
 419       if (deleteFlag) {
 420          ASSERT(ctx);
 421          ctx->Driver.DeleteProgram(ctx, *ptr);
 422       }
 423
 424       *ptr = NULL;
 425    }
 426
 427    assert(!*ptr);
 428    if (prog) {
 429       /*_glthread_LOCK_MUTEX(prog->Mutex);*/
 430       prog->RefCount++;
 431 #if 0
 432       printf("Program %p ID=%u Target=%s  Refcount++ to %d\n",
 433              prog, prog->Id,
 434              (prog->Target == GL_VERTEX_PROGRAM_ARB ? "VP" :
 435               (prog->Target == MESA_GEOMETRY_PROGRAM ? "GP" : "FP")),
 436              prog->RefCount);
 437 #endif
 438       /*_glthread_UNLOCK_MUTEX(prog->Mutex);*/
 439    }
 440
 441    *ptr = prog;
 442 }
 443
 444
 445 /**
 446  * Return a copy of a program.
 447  * XXX Problem here if the program object is actually OO-derivation
 448  * made by a device driver.
 449  */
 450 struct gl_program *
 451 _mesa_clone_program(struct gl_context *ctx, const struct gl_program *prog)
 452 {
 453    struct gl_program *clone;
 454
 455    clone = ctx->Driver.NewProgram(ctx, prog->Target, prog->Id);
 456    if (!clone)
 457       return NULL;
 458
 459    assert(clone->Target == prog->Target);
 460    assert(clone->RefCount == 1);
 461
 462    clone->String = (GLubyte *) _mesa_strdup((char *) prog->String);
 463    clone->Format = prog->Format;
 464    clone->Instructions = _mesa_alloc_instructions(prog->NumInstructions);
 465    if (!clone->Instructions) {
 466       _mesa_reference_program(ctx, &clone, NULL);
 467       return NULL;
 468    }
 469    _mesa_copy_instructions(clone->Instructions, prog->Instructions,
 470                            prog->NumInstructions);
 471    clone->InputsRead = prog->InputsRead;
 472    clone->OutputsWritten = prog->OutputsWritten;
 473    clone->SamplersUsed = prog->SamplersUsed;
 474    clone->ShadowSamplers = prog->ShadowSamplers;
 475    memcpy(clone->TexturesUsed, prog->TexturesUsed, sizeof(prog->TexturesUsed));
 476
 477    if (prog->Parameters)
 478       clone->Parameters = _mesa_clone_parameter_list(prog->Parameters);
 479    memcpy(clone->LocalParams, prog->LocalParams, sizeof(clone->LocalParams));
 480    memcpy(clone->LocalParams, prog->LocalParams, sizeof(clone->LocalParams));
 481    clone->IndirectRegisterFiles = prog->IndirectRegisterFiles;
 482    clone->NumInstructions = prog->NumInstructions;
 483    clone->NumTemporaries = prog->NumTemporaries;
 484    clone->NumParameters = prog->NumParameters;
 485    clone->NumAttributes = prog->NumAttributes;
 486    clone->NumAddressRegs = prog->NumAddressRegs;
 487    clone->NumNativeInstructions = prog->NumNativeInstructions;
 488    clone->NumNativeTemporaries = prog->NumNativeTemporaries;
 489    clone->NumNativeParameters = prog->NumNativeParameters;
 490    clone->NumNativeAttributes = prog->NumNativeAttributes;
 491    clone->NumNativeAddressRegs = prog->NumNativeAddressRegs;
 492    clone->NumAluInstructions = prog->NumAluInstructions;
 493    clone->NumTexInstructions = prog->NumTexInstructions;
 494    clone->NumTexIndirections = prog->NumTexIndirections;
 495    clone->NumNativeAluInstructions = prog->NumNativeAluInstructions;
 496    clone->NumNativeTexInstructions = prog->NumNativeTexInstructions;
 497    clone->NumNativeTexIndirections = prog->NumNativeTexIndirections;
 498
 499    switch (prog->Target) {
 500    case GL_VERTEX_PROGRAM_ARB:
 501       {
 502          const struct gl_vertex_program *vp = gl_vertex_program_const(prog);
 503          struct gl_vertex_program *vpc = gl_vertex_program(clone);
 504          vpc->IsPositionInvariant = vp->IsPositionInvariant;
 505       }
 506       break;
 507    case GL_FRAGMENT_PROGRAM_ARB:
 508       {
 509          const struct gl_fragment_program *fp = gl_fragment_program_const(prog);
 510          struct gl_fragment_program *fpc = gl_fragment_program(clone);
 511          fpc->UsesKill = fp->UsesKill;
 512          fpc->UsesDFdy = fp->UsesDFdy;
 513          fpc->OriginUpperLeft = fp->OriginUpperLeft;
 514          fpc->PixelCenterInteger = fp->PixelCenterInteger;
 515       }
 516       break;
 517    case MESA_GEOMETRY_PROGRAM:
 518       {
 519          const struct gl_geometry_program *gp = gl_geometry_program_const(prog);
 520          struct gl_geometry_program *gpc = gl_geometry_program(clone);
 521          gpc->VerticesOut = gp->VerticesOut;
 522          gpc->InputType = gp->InputType;
 523          gpc->OutputType = gp->OutputType;
 524       }
 525       break;
 526    default:
 527       _mesa_problem(NULL, "Unexpected target in _mesa_clone_program");
 528    }
 529
 530    return clone;
 531 }
 532
 533
 534 /**
 535  * Insert 'count' NOP instructions at 'start' in the given program.
 536  * Adjust branch targets accordingly.
 537  */
 538 GLboolean
 539 _mesa_insert_instructions(struct gl_program *prog, GLuint start, GLuint count)
 540 {
 541    const GLuint origLen = prog->NumInstructions;
 542    const GLuint newLen = origLen + count;
 543    struct prog_instruction *newInst;
 544    GLuint i;
 545
 546    /* adjust branches */
 547    for (i = 0; i < prog->NumInstructions; i++) {
 548       struct prog_instruction *inst = prog->Instructions + i;
 549       if (inst->BranchTarget > 0) {
 550          if ((GLuint)inst->BranchTarget >= start) {
 551             inst->BranchTarget += count;
 552          }
 553       }
 554    }
 555
 556    /* Alloc storage for new instructions */
 557    newInst = _mesa_alloc_instructions(newLen);
 558    if (!newInst) {
 559       return GL_FALSE;
 560    }
 561
 562    /* Copy 'start' instructions into new instruction buffer */
 563    _mesa_copy_instructions(newInst, prog->Instructions, start);
 564
 565    /* init the new instructions */
 566    _mesa_init_instructions(newInst + start, count);
 567
 568    /* Copy the remaining/tail instructions to new inst buffer */
 569    _mesa_copy_instructions(newInst + start + count,
 570                            prog->Instructions + start,
 571                            origLen - start);
 572
 573    /* free old instructions */
 574    _mesa_free_instructions(prog->Instructions, origLen);
 575
 576    /* install new instructions */
 577    prog->Instructions = newInst;
 578    prog->NumInstructions = newLen;
 579
 580    return GL_TRUE;
 581 }
 582
 583 /**
 584  * Delete 'count' instructions at 'start' in the given program.
 585  * Adjust branch targets accordingly.
 586  */
 587 GLboolean
 588 _mesa_delete_instructions(struct gl_program *prog, GLuint start, GLuint count)
 589 {
 590    const GLuint origLen = prog->NumInstructions;
 591    const GLuint newLen = origLen - count;
 592    struct prog_instruction *newInst;
 593    GLuint i;
 594
 595    /* adjust branches */
 596    for (i = 0; i < prog->NumInstructions; i++) {
 597       struct prog_instruction *inst = prog->Instructions + i;
 598       if (inst->BranchTarget > 0) {
 599          if (inst->BranchTarget > (GLint) start) {
 600             inst->BranchTarget -= count;
 601          }
 602       }
 603    }
 604
 605    /* Alloc storage for new instructions */
 606    newInst = _mesa_alloc_instructions(newLen);
 607    if (!newInst) {
 608       return GL_FALSE;
 609    }
 610
 611    /* Copy 'start' instructions into new instruction buffer */
 612    _mesa_copy_instructions(newInst, prog->Instructions, start);
 613
 614    /* Copy the remaining/tail instructions to new inst buffer */
 615    _mesa_copy_instructions(newInst + start,
 616                            prog->Instructions + start + count,
 617                            newLen - start);
 618
 619    /* free old instructions */
 620    _mesa_free_instructions(prog->Instructions, origLen);
 621
 622    /* install new instructions */
 623    prog->Instructions = newInst;
 624    prog->NumInstructions = newLen;
 625
 626    return GL_TRUE;
 627 }
 628
 629
 630 /**
 631  * Search instructions for registers that match (oldFile, oldIndex),
 632  * replacing them with (newFile, newIndex).
 633  */
 634 static void
 635 replace_registers(struct prog_instruction *inst, GLuint numInst,
 636                   GLuint oldFile, GLuint oldIndex,
 637                   GLuint newFile, GLuint newIndex)
 638 {
 639    GLuint i, j;
 640    for (i = 0; i < numInst; i++) {
 641       /* src regs */
 642       for (j = 0; j < _mesa_num_inst_src_regs(inst[i].Opcode); j++) {
 643          if (inst[i].SrcReg[j].File == oldFile &&
 644              inst[i].SrcReg[j].Index == oldIndex) {
 645             inst[i].SrcReg[j].File = newFile;
 646             inst[i].SrcReg[j].Index = newIndex;
 647          }
 648       }
 649       /* dst reg */
 650       if (inst[i].DstReg.File == oldFile && inst[i].DstReg.Index == oldIndex) {
 651          inst[i].DstReg.File = newFile;
 652          inst[i].DstReg.Index = newIndex;
 653       }
 654    }
 655 }
 656
 657
 658 /**
 659  * Search instructions for references to program parameters.  When found,
 660  * increment the parameter index by 'offset'.
 661  * Used when combining programs.
 662  */
 663 static void
 664 adjust_param_indexes(struct prog_instruction *inst, GLuint numInst,
 665                      GLuint offset)
 666 {
 667    GLuint i, j;
 668    for (i = 0; i < numInst; i++) {
 669       for (j = 0; j < _mesa_num_inst_src_regs(inst[i].Opcode); j++) {
 670          GLuint f = inst[i].SrcReg[j].File;
 671          if (f == PROGRAM_CONSTANT ||
 672              f == PROGRAM_UNIFORM ||
 673              f == PROGRAM_STATE_VAR) {
 674             inst[i].SrcReg[j].Index += offset;
 675          }
 676       }
 677    }
 678 }
 679
 680
 681 /**
 682  * Combine two programs into one.  Fix instructions so the outputs of
 683  * the first program go to the inputs of the second program.
 684  */
 685 struct gl_program *
 686 _mesa_combine_programs(struct gl_context *ctx,
 687                        const struct gl_program *progA,
 688                        const struct gl_program *progB)
 689 {
 690    struct prog_instruction *newInst;
 691    struct gl_program *newProg;
 692    const GLuint lenA = progA->NumInstructions - 1; /* omit END instr */
 693    const GLuint lenB = progB->NumInstructions;
 694    const GLuint numParamsA = _mesa_num_parameters(progA->Parameters);
 695    const GLuint newLength = lenA + lenB;
 696    GLboolean usedTemps[MAX_PROGRAM_TEMPS];
 697    GLuint firstTemp = 0;
 698    GLbitfield64 inputsB;
 699    GLuint i;
 700
 701    ASSERT(progA->Target == progB->Target);
 702
 703    newInst = _mesa_alloc_instructions(newLength);
 704    if (!newInst)
 705       return GL_FALSE;
 706
 707    _mesa_copy_instructions(newInst, progA->Instructions, lenA);
 708    _mesa_copy_instructions(newInst + lenA, progB->Instructions, lenB);
 709
 710    /* adjust branch / instruction addresses for B's instructions */
 711    for (i = 0; i < lenB; i++) {
 712       newInst[lenA + i].BranchTarget += lenA;
 713    }
 714
 715    newProg = ctx->Driver.NewProgram(ctx, progA->Target, 0);
 716    newProg->Instructions = newInst;
 717    newProg->NumInstructions = newLength;
 718
 719    /* find used temp regs (we may need new temps below) */
 720    _mesa_find_used_registers(newProg, PROGRAM_TEMPORARY,
 721                              usedTemps, MAX_PROGRAM_TEMPS);
 722
 723    if (newProg->Target == GL_FRAGMENT_PROGRAM_ARB) {
 724       const struct gl_fragment_program *fprogA, *fprogB;
 725       struct gl_fragment_program *newFprog;
 726       GLbitfield64 progB_inputsRead = progB->InputsRead;
 727       GLint progB_colorFile, progB_colorIndex;
 728
 729       fprogA = gl_fragment_program_const(progA);
 730       fprogB = gl_fragment_program_const(progB);
 731       newFprog = gl_fragment_program(newProg);
 732
 733       newFprog->UsesKill = fprogA->UsesKill || fprogB->UsesKill;
 734       newFprog->UsesDFdy = fprogA->UsesDFdy || fprogB->UsesDFdy;
 735
 736       /* We'll do a search and replace for instances
 737        * of progB_colorFile/progB_colorIndex below...
 738        */
 739       progB_colorFile = PROGRAM_INPUT;
 740       progB_colorIndex = VARYING_SLOT_COL0;
 741
 742       /*
 743        * The fragment program may get color from a state var rather than
 744        * a fragment input (vertex output) if it's constant.
 745        * See the texenvprogram.c code.
 746        * So, search the program's parameter list now to see if the program
 747        * gets color from a state var instead of a conventional fragment
 748        * input register.
 749        */
 750       for (i = 0; i < progB->Parameters->NumParameters; i++) {
 751          struct gl_program_parameter *p = &progB->Parameters->Parameters[i];
 752          if (p->Type == PROGRAM_STATE_VAR &&
 753              p->StateIndexes[0] == STATE_INTERNAL &&
 754              p->StateIndexes[1] == STATE_CURRENT_ATTRIB &&
 755              (int) p->StateIndexes[2] == (int) VERT_ATTRIB_COLOR0) {
 756             progB_inputsRead |= VARYING_BIT_COL0;
 757             progB_colorFile = PROGRAM_STATE_VAR;
 758             progB_colorIndex = i;
 759             break;
 760          }
 761       }
 762
 763       /* Connect color outputs of fprogA to color inputs of fprogB, via a
 764        * new temporary register.
 765        */
 766       if ((progA->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_COLOR)) &&
 767           (progB_inputsRead & VARYING_BIT_COL0)) {
 768          GLint tempReg = _mesa_find_free_register(usedTemps, MAX_PROGRAM_TEMPS,
 769                                                   firstTemp);
 770          if (tempReg < 0) {
 771             _mesa_problem(ctx, "No free temp regs found in "
 772                           "_mesa_combine_programs(), using 31");
 773             tempReg = 31;
 774          }
 775          firstTemp = tempReg + 1;
 776
 777          /* replace writes to result.color[0] with tempReg */
 778          replace_registers(newInst, lenA,
 779                            PROGRAM_OUTPUT, FRAG_RESULT_COLOR,
 780                            PROGRAM_TEMPORARY, tempReg);
 781          /* replace reads from the input color with tempReg */
 782          replace_registers(newInst + lenA, lenB,
 783                            progB_colorFile, progB_colorIndex, /* search for */
 784                            PROGRAM_TEMPORARY, tempReg  /* replace with */ );
 785       }
 786
 787       /* compute combined program's InputsRead */
 788       inputsB = progB_inputsRead;
 789       if (progA->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_COLOR)) {
 790          inputsB &= ~(1 << VARYING_SLOT_COL0);
 791       }
 792       newProg->InputsRead = progA->InputsRead | inputsB;
 793       newProg->OutputsWritten = progB->OutputsWritten;
 794       newProg->SamplersUsed = progA->SamplersUsed | progB->SamplersUsed;
 795    }
 796    else {
 797       /* vertex program */
 798       assert(0);      /* XXX todo */
 799    }
 800
 801    /*
 802     * Merge parameters (uniforms, constants, etc)
 803     */
 804    newProg->Parameters = _mesa_combine_parameter_lists(progA->Parameters,
 805                                                        progB->Parameters);
 806
 807    adjust_param_indexes(newInst + lenA, lenB, numParamsA);
 808
 809
 810    return newProg;
 811 }
 812
 813
 814 /**
 815  * Populate the 'used' array with flags indicating which registers (TEMPs,
 816  * INPUTs, OUTPUTs, etc, are used by the given program.
 817  * \param file  type of register to scan for
 818  * \param used  returns true/false flags for in use / free
 819  * \param usedSize  size of the 'used' array
 820  */
 821 void
 822 _mesa_find_used_registers(const struct gl_program *prog,
 823                           gl_register_file file,
 824                           GLboolean used[], GLuint usedSize)
 825 {
 826    GLuint i, j;
 827
 828    memset(used, 0, usedSize);
 829
 830    for (i = 0; i < prog->NumInstructions; i++) {
 831       const struct prog_instruction *inst = prog->Instructions + i;
 832       const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
 833
 834       if (inst->DstReg.File == file) {
 835          ASSERT(inst->DstReg.Index < usedSize);
 836          if(inst->DstReg.Index < usedSize)
 837             used[inst->DstReg.Index] = GL_TRUE;
 838       }
 839
 840       for (j = 0; j < n; j++) {
 841          if (inst->SrcReg[j].File == file) {
 842             ASSERT(inst->SrcReg[j].Index < (GLint) usedSize);
 843             if (inst->SrcReg[j].Index < (GLint) usedSize)
 844                used[inst->SrcReg[j].Index] = GL_TRUE;
 845          }
 846       }
 847    }
 848 }
 849
 850
 851 /**
 852  * Scan the given 'used' register flag array for the first entry
 853  * that's >= firstReg.
 854  * \param used  vector of flags indicating registers in use (as returned
 855  *              by _mesa_find_used_registers())
 856  * \param usedSize  size of the 'used' array
 857  * \param firstReg  first register to start searching at
 858  * \return index of unused register, or -1 if none.
 859  */
 860 GLint
 861 _mesa_find_free_register(const GLboolean used[],
 862                          GLuint usedSize, GLuint firstReg)
 863 {
 864    GLuint i;
 865
 866    assert(firstReg < usedSize);
 867
 868    for (i = firstReg; i < usedSize; i++)
 869       if (!used[i])
 870          return i;
 871
 872    return -1;
 873 }
 874
 875
 876
 877 /**
 878  * Check if the given register index is valid (doesn't exceed implementation-
 879  * dependent limits).
 880  * \return GL_TRUE if OK, GL_FALSE if bad index
 881  */
 882 GLboolean
 883 _mesa_valid_register_index(const struct gl_context *ctx,
 884                            gl_shader_type shaderType,
 885                            gl_register_file file, GLint index)
 886 {
 887    const struct gl_program_constants *c;
 888
 889    switch (shaderType) {
 890    case MESA_SHADER_VERTEX:
 891       c = &ctx->Const.VertexProgram;
 892       break;
 893    case MESA_SHADER_FRAGMENT:
 894       c = &ctx->Const.FragmentProgram;
 895       break;
 896    case MESA_SHADER_GEOMETRY:
 897       c = &ctx->Const.GeometryProgram;
 898       break;
 899    default:
 900       _mesa_problem(ctx,
 901                     "unexpected shader type in _mesa_valid_register_index()");
 902       return GL_FALSE;
 903    }
 904
 905    switch (file) {
 906    case PROGRAM_UNDEFINED:
 907       return GL_TRUE;  /* XXX or maybe false? */
 908
 909    case PROGRAM_TEMPORARY:
 910       return index >= 0 && index < (GLint) c->MaxTemps;
 911
 912    case PROGRAM_ENV_PARAM:
 913       return index >= 0 && index < (GLint) c->MaxEnvParams;
 914
 915    case PROGRAM_LOCAL_PARAM:
 916       return index >= 0 && index < (GLint) c->MaxLocalParams;
 917
 918    case PROGRAM_UNIFORM:
 919    case PROGRAM_STATE_VAR:
 920       /* aka constant buffer */
 921       return index >= 0 && index < (GLint) c->MaxUniformComponents / 4;
 922
 923    case PROGRAM_CONSTANT:
 924       /* constant buffer w/ possible relative negative addressing */
 925       return (index > (int) c->MaxUniformComponents / -4 &&
 926               index < (int) c->MaxUniformComponents / 4);
 927
 928    case PROGRAM_INPUT:
 929       if (index < 0)
 930          return GL_FALSE;
 931
 932       switch (shaderType) {
 933       case MESA_SHADER_VERTEX:
 934          return index < VERT_ATTRIB_GENERIC0 + (GLint) c->MaxAttribs;
 935       case MESA_SHADER_FRAGMENT:
 936          return index < VARYING_SLOT_VAR0 + (GLint) ctx->Const.MaxVarying;
 937       case MESA_SHADER_GEOMETRY:
 938          return index < VARYING_SLOT_VAR0 + (GLint) ctx->Const.MaxVarying;
 939       default:
 940          return GL_FALSE;
 941       }
 942
 943    case PROGRAM_OUTPUT:
 944       if (index < 0)
 945          return GL_FALSE;
 946
 947       switch (shaderType) {
 948       case MESA_SHADER_VERTEX:
 949          return index < VARYING_SLOT_VAR0 + (GLint) ctx->Const.MaxVarying;
 950       case MESA_SHADER_FRAGMENT:
 951          return index < FRAG_RESULT_DATA0 + (GLint) ctx->Const.MaxDrawBuffers;
 952       case MESA_SHADER_GEOMETRY:
 953          return index < VARYING_SLOT_VAR0 + (GLint) ctx->Const.MaxVarying;
 954       default:
 955          return GL_FALSE;
 956       }
 957
 958    case PROGRAM_ADDRESS:
 959       return index >= 0 && index < (GLint) c->MaxAddressRegs;
 960
 961    default:
 962       _mesa_problem(ctx,
 963                     "unexpected register file in _mesa_valid_register_index()");
 964       return GL_FALSE;
 965    }
 966 }
 967
 968
 969
 970 /**
 971  * "Post-process" a GPU program.  This is intended to be used for debugging.
 972  * Example actions include no-op'ing instructions or changing instruction
 973  * behaviour.
 974  */
 975 void
 976 _mesa_postprocess_program(struct gl_context *ctx, struct gl_program *prog)
 977 {
 978    static const GLfloat white[4] = { 0.5, 0.5, 0.5, 0.5 };
 979    GLuint i;
 980    GLuint whiteSwizzle;
 981    GLint whiteIndex = _mesa_add_unnamed_constant(prog->Parameters,
 982                                                  (gl_constant_value *) white,
 983                                                  4, &whiteSwizzle);
 984
 985    (void) whiteIndex;
 986
 987    for (i = 0; i < prog->NumInstructions; i++) {
 988       struct prog_instruction *inst = prog->Instructions + i;
 989       const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
 990
 991       (void) n;
 992
 993       if (_mesa_is_tex_instruction(inst->Opcode)) {
 994 #if 0
 995          /* replace TEX/TXP/TXB with MOV */
 996          inst->Opcode = OPCODE_MOV;
 997          inst->DstReg.WriteMask = WRITEMASK_XYZW;
 998          inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
 999          inst->SrcReg[0].Negate = NEGATE_NONE;
1000 #endif
1001
1002 #if 0
1003          /* disable shadow texture mode */
1004          inst->TexShadow = 0;
1005 #endif
1006       }
1007
1008       if (inst->Opcode == OPCODE_TXP) {
1009 #if 0
1010          inst->Opcode = OPCODE_MOV;
1011          inst->DstReg.WriteMask = WRITEMASK_XYZW;
1012          inst->SrcReg[0].File = PROGRAM_CONSTANT;
1013          inst->SrcReg[0].Index = whiteIndex;
1014          inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
1015          inst->SrcReg[0].Negate = NEGATE_NONE;
1016 #endif
1017 #if 0
1018          inst->TexShadow = 0;
1019 #endif
1020 #if 0
1021          inst->Opcode = OPCODE_TEX;
1022          inst->TexShadow = 0;
1023 #endif
1024       }
1025
1026    }
1027 }