src/mesa/main/texenvprogram.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "glheader.h"
  29 #include "macros.h"
  30 #include "enums.h"
  31 #include "shader/program.h"
  32 #include "shader/prog_parameter.h"
  33 #include "shader/prog_cache.h"
  34 #include "shader/prog_instruction.h"
  35 #include "shader/prog_print.h"
  36 #include "shader/prog_statevars.h"
  37 #include "shader/programopt.h"
  38 #include "texenvprogram.h"
  39
  40
  41 struct texenvprog_cache_item
  42 {
  43    GLuint hash;
  44    void *key;
  45    struct gl_fragment_program *data;
  46    struct texenvprog_cache_item *next;
  47 };
  48
  49
  50 /**
  51  * Up to nine instructions per tex unit, plus fog, specular color.
  52  */
  53 #define MAX_INSTRUCTIONS ((MAX_TEXTURE_UNITS * 9) + 12)
  54
  55 #define DISASSEM (MESA_VERBOSE & VERBOSE_DISASSEM)
  56
  57 struct mode_opt {
  58    GLubyte Source:4;
  59    GLubyte Operand:3;
  60 };
  61
  62 struct state_key {
  63    GLuint nr_enabled_units:8;
  64    GLuint enabled_units:8;
  65    GLuint separate_specular:1;
  66    GLuint fog_enabled:1;
  67    GLuint fog_mode:2;
  68    GLuint inputs_available:12;
  69
  70    struct {
  71       GLuint enabled:1;
  72       GLuint source_index:3;   /* one of TEXTURE_1D/2D/3D/CUBE/RECT_INDEX */
  73       GLuint shadow:1;
  74       GLuint ScaleShiftRGB:2;
  75       GLuint ScaleShiftA:2;
  76
  77       GLuint NumArgsRGB:2;
  78       GLuint ModeRGB:4;
  79       GLuint NumArgsA:2;
  80       GLuint ModeA:4;
  81
  82       struct mode_opt OptRGB[3];
  83       struct mode_opt OptA[3];
  84    } unit[8];
  85 };
  86
  87 #define FOG_LINEAR  0
  88 #define FOG_EXP     1
  89 #define FOG_EXP2    2
  90 #define FOG_UNKNOWN 3
  91
  92 static GLuint translate_fog_mode( GLenum mode )
  93 {
  94    switch (mode) {
  95    case GL_LINEAR: return FOG_LINEAR;
  96    case GL_EXP: return FOG_EXP;
  97    case GL_EXP2: return FOG_EXP2;
  98    default: return FOG_UNKNOWN;
  99    }
 100 }
 101
 102 #define OPR_SRC_COLOR           0
 103 #define OPR_ONE_MINUS_SRC_COLOR 1
 104 #define OPR_SRC_ALPHA           2
 105 #define OPR_ONE_MINUS_SRC_ALPHA 3
 106 #define OPR_ZERO                4
 107 #define OPR_ONE                 5
 108 #define OPR_UNKNOWN             7
 109
 110 static GLuint translate_operand( GLenum operand )
 111 {
 112    switch (operand) {
 113    case GL_SRC_COLOR: return OPR_SRC_COLOR;
 114    case GL_ONE_MINUS_SRC_COLOR: return OPR_ONE_MINUS_SRC_COLOR;
 115    case GL_SRC_ALPHA: return OPR_SRC_ALPHA;
 116    case GL_ONE_MINUS_SRC_ALPHA: return OPR_ONE_MINUS_SRC_ALPHA;
 117    case GL_ZERO: return OPR_ZERO;
 118    case GL_ONE: return OPR_ONE;
 119    default:     return OPR_UNKNOWN;
 120    }
 121 }
 122
 123 #define SRC_TEXTURE  0
 124 #define SRC_TEXTURE0 1
 125 #define SRC_TEXTURE1 2
 126 #define SRC_TEXTURE2 3
 127 #define SRC_TEXTURE3 4
 128 #define SRC_TEXTURE4 5
 129 #define SRC_TEXTURE5 6
 130 #define SRC_TEXTURE6 7
 131 #define SRC_TEXTURE7 8
 132 #define SRC_CONSTANT 9
 133 #define SRC_PRIMARY_COLOR 10
 134 #define SRC_PREVIOUS 11
 135 #define SRC_UNKNOWN  15
 136
 137 static GLuint translate_source( GLenum src )
 138 {
 139    switch (src) {
 140    case GL_TEXTURE: return SRC_TEXTURE;
 141    case GL_TEXTURE0:
 142    case GL_TEXTURE1:
 143    case GL_TEXTURE2:
 144    case GL_TEXTURE3:
 145    case GL_TEXTURE4:
 146    case GL_TEXTURE5:
 147    case GL_TEXTURE6:
 148    case GL_TEXTURE7: return SRC_TEXTURE0 + (src - GL_TEXTURE0);
 149    case GL_CONSTANT: return SRC_CONSTANT;
 150    case GL_PRIMARY_COLOR: return SRC_PRIMARY_COLOR;
 151    case GL_PREVIOUS: return SRC_PREVIOUS;
 152    default: return SRC_UNKNOWN;
 153    }
 154 }
 155
 156 #define MODE_REPLACE       0
 157 #define MODE_MODULATE      1
 158 #define MODE_ADD           2
 159 #define MODE_ADD_SIGNED    3
 160 #define MODE_INTERPOLATE   4
 161 #define MODE_SUBTRACT      5
 162 #define MODE_DOT3_RGB      6
 163 #define MODE_DOT3_RGB_EXT  7
 164 #define MODE_DOT3_RGBA     8
 165 #define MODE_DOT3_RGBA_EXT 9
 166 #define MODE_MODULATE_ADD_ATI           10
 167 #define MODE_MODULATE_SIGNED_ADD_ATI    11
 168 #define MODE_MODULATE_SUBTRACT_ATI      12
 169 #define MODE_UNKNOWN       15
 170
 171 static GLuint translate_mode( GLenum mode )
 172 {
 173    switch (mode) {
 174    case GL_REPLACE: return MODE_REPLACE;
 175    case GL_MODULATE: return MODE_MODULATE;
 176    case GL_ADD: return MODE_ADD;
 177    case GL_ADD_SIGNED: return MODE_ADD_SIGNED;
 178    case GL_INTERPOLATE: return MODE_INTERPOLATE;
 179    case GL_SUBTRACT: return MODE_SUBTRACT;
 180    case GL_DOT3_RGB: return MODE_DOT3_RGB;
 181    case GL_DOT3_RGB_EXT: return MODE_DOT3_RGB_EXT;
 182    case GL_DOT3_RGBA: return MODE_DOT3_RGBA;
 183    case GL_DOT3_RGBA_EXT: return MODE_DOT3_RGBA_EXT;
 184    case GL_MODULATE_ADD_ATI: return MODE_MODULATE_ADD_ATI;
 185    case GL_MODULATE_SIGNED_ADD_ATI: return MODE_MODULATE_SIGNED_ADD_ATI;
 186    case GL_MODULATE_SUBTRACT_ATI: return MODE_MODULATE_SUBTRACT_ATI;
 187    default: return MODE_UNKNOWN;
 188    }
 189 }
 190
 191 #define TEXTURE_UNKNOWN_INDEX 7
 192 static GLuint translate_tex_src_bit( GLbitfield bit )
 193 {
 194    switch (bit) {
 195    case TEXTURE_1D_BIT:   return TEXTURE_1D_INDEX;
 196    case TEXTURE_2D_BIT:   return TEXTURE_2D_INDEX;
 197    case TEXTURE_RECT_BIT: return TEXTURE_RECT_INDEX;
 198    case TEXTURE_3D_BIT:   return TEXTURE_3D_INDEX;
 199    case TEXTURE_CUBE_BIT: return TEXTURE_CUBE_INDEX;
 200    default: return TEXTURE_UNKNOWN_INDEX;
 201    }
 202 }
 203
 204 #define VERT_BIT_TEX_ANY    (0xff << VERT_ATTRIB_TEX0)
 205 #define VERT_RESULT_TEX_ANY (0xff << VERT_RESULT_TEX0)
 206
 207 /**
 208  * Identify all possible varying inputs.  The fragment program will
 209  * never reference non-varying inputs, but will track them via state
 210  * constants instead.
 211  *
 212  * This function figures out all the inputs that the fragment program
 213  * has access to.  The bitmask is later reduced to just those which
 214  * are actually referenced.
 215  */
 216 static GLbitfield get_fp_input_mask( GLcontext *ctx )
 217 {
 218    GLbitfield fp_inputs = 0x0;
 219
 220    if (ctx->VertexProgram._Overriden) {
 221       /* Somebody's messing with the vertex program and we don't have
 222        * a clue what's happening.  Assume that it could be producing
 223        * all possible outputs.
 224        */
 225       fp_inputs = ~0;
 226    }
 227    else if (ctx->RenderMode == GL_FEEDBACK) {
 228       fp_inputs = (FRAG_BIT_COL0 | FRAG_BIT_TEX0);
 229    }
 230    else if (!ctx->VertexProgram._Enabled ||
 231             !ctx->VertexProgram._Current) {
 232
 233       /* Fixed function logic */
 234       GLbitfield varying_inputs = ctx->varying_vp_inputs;
 235
 236       /* First look at what values may be computed by the generated
 237        * vertex program:
 238        */
 239       if (ctx->Light.Enabled) {
 240          fp_inputs |= FRAG_BIT_COL0;
 241
 242          if (ctx->_TriangleCaps & DD_SEPARATE_SPECULAR)
 243             fp_inputs |= FRAG_BIT_COL1;
 244       }
 245
 246       fp_inputs |= (ctx->Texture._TexGenEnabled |
 247                     ctx->Texture._TexMatEnabled) << FRAG_ATTRIB_TEX0;
 248
 249       /* Then look at what might be varying as a result of enabled
 250        * arrays, etc:
 251        */
 252       if (varying_inputs & VERT_BIT_COLOR0) fp_inputs |= FRAG_BIT_COL0;
 253       if (varying_inputs & VERT_BIT_COLOR1) fp_inputs |= FRAG_BIT_COL1;
 254
 255       fp_inputs |= (((varying_inputs & VERT_BIT_TEX_ANY) >> VERT_ATTRIB_TEX0)
 256                     << FRAG_ATTRIB_TEX0);
 257
 258    }
 259    else {
 260       /* calculate from vp->outputs */
 261       GLbitfield vp_outputs = ctx->VertexProgram._Current->Base.OutputsWritten;
 262
 263       if (vp_outputs & (1 << VERT_RESULT_COL0)) fp_inputs |= FRAG_BIT_COL0;
 264       if (vp_outputs & (1 << VERT_RESULT_COL1)) fp_inputs |= FRAG_BIT_COL1;
 265
 266       fp_inputs |= (((vp_outputs & VERT_RESULT_TEX_ANY) >> VERT_RESULT_TEX0)
 267                     << FRAG_ATTRIB_TEX0);
 268    }
 269
 270    return fp_inputs;
 271 }
 272
 273
 274 /**
 275  * Examine current texture environment state and generate a unique
 276  * key to identify it.
 277  */
 278 static void make_state_key( GLcontext *ctx,  struct state_key *key )
 279 {
 280    GLuint i, j;
 281    GLbitfield inputs_referenced = FRAG_BIT_COL0;
 282    GLbitfield inputs_available = get_fp_input_mask( ctx );
 283
 284    memset(key, 0, sizeof(*key));
 285
 286    for (i=0;i<MAX_TEXTURE_UNITS;i++) {
 287       const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[i];
 288
 289       if (!texUnit->_ReallyEnabled || !texUnit->Enabled)
 290          continue;
 291
 292       key->unit[i].enabled = 1;
 293       key->enabled_units |= (1<<i);
 294       key->nr_enabled_units = i+1;
 295       inputs_referenced |= FRAG_BIT_TEX(i);
 296
 297       key->unit[i].source_index =
 298          translate_tex_src_bit(texUnit->_ReallyEnabled);
 299       key->unit[i].shadow = texUnit->_Current->CompareMode == GL_COMPARE_R_TO_TEXTURE;
 300
 301       key->unit[i].NumArgsRGB = texUnit->_CurrentCombine->_NumArgsRGB;
 302       key->unit[i].NumArgsA = texUnit->_CurrentCombine->_NumArgsA;
 303
 304       key->unit[i].ModeRGB =
 305          translate_mode(texUnit->_CurrentCombine->ModeRGB);
 306       key->unit[i].ModeA =
 307          translate_mode(texUnit->_CurrentCombine->ModeA);
 308
 309       key->unit[i].ScaleShiftRGB = texUnit->_CurrentCombine->ScaleShiftRGB;
 310       key->unit[i].ScaleShiftA = texUnit->_CurrentCombine->ScaleShiftA;
 311
 312       for (j=0;j<3;j++) {
 313          key->unit[i].OptRGB[j].Operand =
 314             translate_operand(texUnit->_CurrentCombine->OperandRGB[j]);
 315          key->unit[i].OptA[j].Operand =
 316             translate_operand(texUnit->_CurrentCombine->OperandA[j]);
 317          key->unit[i].OptRGB[j].Source =
 318             translate_source(texUnit->_CurrentCombine->SourceRGB[j]);
 319          key->unit[i].OptA[j].Source =
 320             translate_source(texUnit->_CurrentCombine->SourceA[j]);
 321       }
 322    }
 323
 324    if (ctx->_TriangleCaps & DD_SEPARATE_SPECULAR) {
 325       key->separate_specular = 1;
 326       inputs_referenced |= FRAG_BIT_COL1;
 327    }
 328
 329    if (ctx->Fog.Enabled) {
 330       key->fog_enabled = 1;
 331       key->fog_mode = translate_fog_mode(ctx->Fog.Mode);
 332       inputs_referenced |= FRAG_BIT_FOGC; /* maybe */
 333    }
 334
 335    key->inputs_available = (inputs_available & inputs_referenced);
 336 }
 337
 338 /**
 339  * Use uregs to represent registers internally, translate to Mesa's
 340  * expected formats on emit.
 341  *
 342  * NOTE: These are passed by value extensively in this file rather
 343  * than as usual by pointer reference.  If this disturbs you, try
 344  * remembering they are just 32bits in size.
 345  *
 346  * GCC is smart enough to deal with these dword-sized structures in
 347  * much the same way as if I had defined them as dwords and was using
 348  * macros to access and set the fields.  This is much nicer and easier
 349  * to evolve.
 350  */
 351 struct ureg {
 352    GLuint file:4;
 353    GLuint idx:8;
 354    GLuint negatebase:1;
 355    GLuint abs:1;
 356    GLuint negateabs:1;
 357    GLuint swz:12;
 358    GLuint pad:5;
 359 };
 360
 361 static const struct ureg undef = {
 362    PROGRAM_UNDEFINED,
 363    ~0,
 364    0,
 365    0,
 366    0,
 367    0,
 368    0
 369 };
 370
 371
 372 /** State used to build the fragment program:
 373  */
 374 struct texenv_fragment_program {
 375    struct gl_fragment_program *program;
 376    GLcontext *ctx;
 377    struct state_key *state;
 378
 379    GLbitfield alu_temps;        /**< Track texture indirections, see spec. */
 380    GLbitfield temps_output;     /**< Track texture indirections, see spec. */
 381    GLbitfield temp_in_use;      /**< Tracks temporary regs which are in use. */
 382    GLboolean error;
 383
 384    struct ureg src_texture[MAX_TEXTURE_UNITS];
 385    /* Reg containing each texture unit's sampled texture color,
 386     * else undef.
 387     */
 388
 389    struct ureg src_previous;    /**< Reg containing color from previous
 390                                  * stage.  May need to be decl'd.
 391                                  */
 392
 393    GLuint last_tex_stage;       /**< Number of last enabled texture unit */
 394
 395    struct ureg half;
 396    struct ureg one;
 397    struct ureg zero;
 398 };
 399
 400
 401
 402 static struct ureg make_ureg(GLuint file, GLuint idx)
 403 {
 404    struct ureg reg;
 405    reg.file = file;
 406    reg.idx = idx;
 407    reg.negatebase = 0;
 408    reg.abs = 0;
 409    reg.negateabs = 0;
 410    reg.swz = SWIZZLE_NOOP;
 411    reg.pad = 0;
 412    return reg;
 413 }
 414
 415 static struct ureg swizzle( struct ureg reg, int x, int y, int z, int w )
 416 {
 417    reg.swz = MAKE_SWIZZLE4(GET_SWZ(reg.swz, x),
 418                            GET_SWZ(reg.swz, y),
 419                            GET_SWZ(reg.swz, z),
 420                            GET_SWZ(reg.swz, w));
 421
 422    return reg;
 423 }
 424
 425 static struct ureg swizzle1( struct ureg reg, int x )
 426 {
 427    return swizzle(reg, x, x, x, x);
 428 }
 429
 430 static struct ureg negate( struct ureg reg )
 431 {
 432    reg.negatebase ^= 1;
 433    return reg;
 434 }
 435
 436 static GLboolean is_undef( struct ureg reg )
 437 {
 438    return reg.file == PROGRAM_UNDEFINED;
 439 }
 440
 441
 442 static struct ureg get_temp( struct texenv_fragment_program *p )
 443 {
 444    GLint bit;
 445
 446    /* First try and reuse temps which have been used already:
 447     */
 448    bit = _mesa_ffs( ~p->temp_in_use & p->alu_temps );
 449
 450    /* Then any unused temporary:
 451     */
 452    if (!bit)
 453       bit = _mesa_ffs( ~p->temp_in_use );
 454
 455    if (!bit) {
 456       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 457       _mesa_exit(1);
 458    }
 459
 460    if ((GLuint) bit > p->program->Base.NumTemporaries)
 461       p->program->Base.NumTemporaries = bit;
 462
 463    p->temp_in_use |= 1<<(bit-1);
 464    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 465 }
 466
 467 static struct ureg get_tex_temp( struct texenv_fragment_program *p )
 468 {
 469    int bit;
 470
 471    /* First try to find available temp not previously used (to avoid
 472     * starting a new texture indirection).  According to the spec, the
 473     * ~p->temps_output isn't necessary, but will keep it there for
 474     * now:
 475     */
 476    bit = _mesa_ffs( ~p->temp_in_use & ~p->alu_temps & ~p->temps_output );
 477
 478    /* Then any unused temporary:
 479     */
 480    if (!bit)
 481       bit = _mesa_ffs( ~p->temp_in_use );
 482
 483    if (!bit) {
 484       _mesa_problem(NULL, "%s: out of temporaries\n", __FILE__);
 485       _mesa_exit(1);
 486    }
 487
 488    if ((GLuint) bit > p->program->Base.NumTemporaries)
 489       p->program->Base.NumTemporaries = bit;
 490
 491    p->temp_in_use |= 1<<(bit-1);
 492    return make_ureg(PROGRAM_TEMPORARY, (bit-1));
 493 }
 494
 495
 496 /** Mark a temp reg as being no longer allocatable. */
 497 static void reserve_temp( struct texenv_fragment_program *p, struct ureg r )
 498 {
 499    if (r.file == PROGRAM_TEMPORARY)
 500       p->temps_output |= (1 << r.idx);
 501 }
 502
 503
 504 static void release_temps(GLcontext *ctx, struct texenv_fragment_program *p )
 505 {
 506    GLuint max_temp = ctx->Const.FragmentProgram.MaxTemps;
 507
 508    /* KW: To support tex_env_crossbar, don't release the registers in
 509     * temps_output.
 510     */
 511    if (max_temp >= sizeof(int) * 8)
 512       p->temp_in_use = p->temps_output;
 513    else
 514       p->temp_in_use = ~((1<<max_temp)-1) | p->temps_output;
 515 }
 516
 517
 518 static struct ureg register_param5( struct texenv_fragment_program *p,
 519                                     GLint s0,
 520                                     GLint s1,
 521                                     GLint s2,
 522                                     GLint s3,
 523                                     GLint s4)
 524 {
 525    gl_state_index tokens[STATE_LENGTH];
 526    GLuint idx;
 527    tokens[0] = s0;
 528    tokens[1] = s1;
 529    tokens[2] = s2;
 530    tokens[3] = s3;
 531    tokens[4] = s4;
 532    idx = _mesa_add_state_reference( p->program->Base.Parameters, tokens );
 533    return make_ureg(PROGRAM_STATE_VAR, idx);
 534 }
 535
 536
 537 #define register_param1(p,s0)          register_param5(p,s0,0,0,0,0)
 538 #define register_param2(p,s0,s1)       register_param5(p,s0,s1,0,0,0)
 539 #define register_param3(p,s0,s1,s2)    register_param5(p,s0,s1,s2,0,0)
 540 #define register_param4(p,s0,s1,s2,s3) register_param5(p,s0,s1,s2,s3,0)
 541
 542 static GLuint frag_to_vert_attrib( GLuint attrib )
 543 {
 544    switch (attrib) {
 545    case FRAG_ATTRIB_COL0: return VERT_ATTRIB_COLOR0;
 546    case FRAG_ATTRIB_COL1: return VERT_ATTRIB_COLOR1;
 547    default:
 548       assert(attrib >= FRAG_ATTRIB_TEX0);
 549       assert(attrib <= FRAG_ATTRIB_TEX7);
 550       return attrib - FRAG_ATTRIB_TEX0 + VERT_ATTRIB_TEX0;
 551    }
 552 }
 553
 554
 555 static struct ureg register_input( struct texenv_fragment_program *p, GLuint input )
 556 {
 557    if (p->state->inputs_available & (1<<input)) {
 558       p->program->Base.InputsRead |= (1 << input);
 559       return make_ureg(PROGRAM_INPUT, input);
 560    }
 561    else {
 562       GLuint idx = frag_to_vert_attrib( input );
 563       return register_param3( p, STATE_INTERNAL, STATE_CURRENT_ATTRIB, idx );
 564    }
 565 }
 566
 567
 568 static void emit_arg( struct prog_src_register *reg,
 569                       struct ureg ureg )
 570 {
 571    reg->File = ureg.file;
 572    reg->Index = ureg.idx;
 573    reg->Swizzle = ureg.swz;
 574    reg->NegateBase = ureg.negatebase ? 0xf : 0x0;
 575    reg->Abs = ureg.abs;
 576    reg->NegateAbs = ureg.negateabs;
 577 }
 578
 579 static void emit_dst( struct prog_dst_register *dst,
 580                       struct ureg ureg, GLuint mask )
 581 {
 582    dst->File = ureg.file;
 583    dst->Index = ureg.idx;
 584    dst->WriteMask = mask;
 585    dst->CondMask = COND_TR;  /* always pass cond test */
 586    dst->CondSwizzle = SWIZZLE_NOOP;
 587 }
 588
 589 static struct prog_instruction *
 590 emit_op(struct texenv_fragment_program *p,
 591         enum prog_opcode op,
 592         struct ureg dest,
 593         GLuint mask,
 594         GLboolean saturate,
 595         struct ureg src0,
 596         struct ureg src1,
 597         struct ureg src2 )
 598 {
 599    GLuint nr = p->program->Base.NumInstructions++;
 600    struct prog_instruction *inst = &p->program->Base.Instructions[nr];
 601
 602    assert(nr < MAX_INSTRUCTIONS);
 603
 604    _mesa_init_instructions(inst, 1);
 605    inst->Opcode = op;
 606
 607    emit_arg( &inst->SrcReg[0], src0 );
 608    emit_arg( &inst->SrcReg[1], src1 );
 609    emit_arg( &inst->SrcReg[2], src2 );
 610
 611    inst->SaturateMode = saturate ? SATURATE_ZERO_ONE : SATURATE_OFF;
 612
 613    emit_dst( &inst->DstReg, dest, mask );
 614
 615 #if 0
 616    /* Accounting for indirection tracking:
 617     */
 618    if (dest.file == PROGRAM_TEMPORARY)
 619       p->temps_output |= 1 << dest.idx;
 620 #endif
 621
 622    return inst;
 623 }
 624
 625
 626 static struct ureg emit_arith( struct texenv_fragment_program *p,
 627                                enum prog_opcode op,
 628                                struct ureg dest,
 629                                GLuint mask,
 630                                GLboolean saturate,
 631                                struct ureg src0,
 632                                struct ureg src1,
 633                                struct ureg src2 )
 634 {
 635    emit_op(p, op, dest, mask, saturate, src0, src1, src2);
 636
 637    /* Accounting for indirection tracking:
 638     */
 639    if (src0.file == PROGRAM_TEMPORARY)
 640       p->alu_temps |= 1 << src0.idx;
 641
 642    if (!is_undef(src1) && src1.file == PROGRAM_TEMPORARY)
 643       p->alu_temps |= 1 << src1.idx;
 644
 645    if (!is_undef(src2) && src2.file == PROGRAM_TEMPORARY)
 646       p->alu_temps |= 1 << src2.idx;
 647
 648    if (dest.file == PROGRAM_TEMPORARY)
 649       p->alu_temps |= 1 << dest.idx;
 650
 651    p->program->Base.NumAluInstructions++;
 652    return dest;
 653 }
 654
 655 static struct ureg emit_texld( struct texenv_fragment_program *p,
 656                                enum prog_opcode op,
 657                                struct ureg dest,
 658                                GLuint destmask,
 659                                GLuint tex_unit,
 660                                GLuint tex_idx,
 661                                struct ureg coord )
 662 {
 663    struct prog_instruction *inst = emit_op( p, op,
 664                                           dest, destmask,
 665                                           GL_FALSE,     /* don't saturate? */
 666                                           coord,        /* arg 0? */
 667                                           undef,
 668                                           undef);
 669
 670    inst->TexSrcTarget = tex_idx;
 671    inst->TexSrcUnit = tex_unit;
 672
 673    p->program->Base.NumTexInstructions++;
 674
 675    /* Accounting for indirection tracking:
 676     */
 677    reserve_temp(p, dest);
 678
 679    /* Is this a texture indirection?
 680     */
 681    if ((coord.file == PROGRAM_TEMPORARY &&
 682         (p->temps_output & (1<<coord.idx))) ||
 683        (dest.file == PROGRAM_TEMPORARY &&
 684         (p->alu_temps & (1<<dest.idx)))) {
 685       p->program->Base.NumTexIndirections++;
 686       p->temps_output = 1<<coord.idx;
 687       p->alu_temps = 0;
 688       assert(0);                /* KW: texture env crossbar */
 689    }
 690
 691    return dest;
 692 }
 693
 694
 695 static struct ureg register_const4f( struct texenv_fragment_program *p,
 696                                      GLfloat s0,
 697                                      GLfloat s1,
 698                                      GLfloat s2,
 699                                      GLfloat s3)
 700 {
 701    GLfloat values[4];
 702    GLuint idx, swizzle;
 703    struct ureg r;
 704    values[0] = s0;
 705    values[1] = s1;
 706    values[2] = s2;
 707    values[3] = s3;
 708    idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values, 4,
 709                                      &swizzle );
 710    r = make_ureg(PROGRAM_CONSTANT, idx);
 711    r.swz = swizzle;
 712    return r;
 713 }
 714
 715 #define register_scalar_const(p, s0)    register_const4f(p, s0, s0, s0, s0)
 716 #define register_const1f(p, s0)         register_const4f(p, s0, 0, 0, 1)
 717 #define register_const2f(p, s0, s1)     register_const4f(p, s0, s1, 0, 1)
 718 #define register_const3f(p, s0, s1, s2) register_const4f(p, s0, s1, s2, 1)
 719
 720
 721 static struct ureg get_one( struct texenv_fragment_program *p )
 722 {
 723    if (is_undef(p->one))
 724       p->one = register_scalar_const(p, 1.0);
 725    return p->one;
 726 }
 727
 728 static struct ureg get_half( struct texenv_fragment_program *p )
 729 {
 730    if (is_undef(p->half))
 731       p->half = register_scalar_const(p, 0.5);
 732    return p->half;
 733 }
 734
 735 static struct ureg get_zero( struct texenv_fragment_program *p )
 736 {
 737    if (is_undef(p->zero))
 738       p->zero = register_scalar_const(p, 0.0);
 739    return p->zero;
 740 }
 741
 742
 743 static void program_error( struct texenv_fragment_program *p, const char *msg )
 744 {
 745    _mesa_problem(NULL, msg);
 746    p->error = 1;
 747 }
 748
 749 static struct ureg get_source( struct texenv_fragment_program *p,
 750                                GLuint src, GLuint unit )
 751 {
 752    switch (src) {
 753    case SRC_TEXTURE:
 754       assert(!is_undef(p->src_texture[unit]));
 755       return p->src_texture[unit];
 756
 757    case SRC_TEXTURE0:
 758    case SRC_TEXTURE1:
 759    case SRC_TEXTURE2:
 760    case SRC_TEXTURE3:
 761    case SRC_TEXTURE4:
 762    case SRC_TEXTURE5:
 763    case SRC_TEXTURE6:
 764    case SRC_TEXTURE7:
 765       assert(!is_undef(p->src_texture[src - SRC_TEXTURE0]));
 766       return p->src_texture[src - SRC_TEXTURE0];
 767
 768    case SRC_CONSTANT:
 769       return register_param2(p, STATE_TEXENV_COLOR, unit);
 770
 771    case SRC_PRIMARY_COLOR:
 772       return register_input(p, FRAG_ATTRIB_COL0);
 773
 774    case SRC_PREVIOUS:
 775    default:
 776       if (is_undef(p->src_previous))
 777          return register_input(p, FRAG_ATTRIB_COL0);
 778       else
 779          return p->src_previous;
 780    }
 781 }
 782
 783 static struct ureg emit_combine_source( struct texenv_fragment_program *p,
 784                                         GLuint mask,
 785                                         GLuint unit,
 786                                         GLuint source,
 787                                         GLuint operand )
 788 {
 789    struct ureg arg, src, one;
 790
 791    src = get_source(p, source, unit);
 792
 793    switch (operand) {
 794    case OPR_ONE_MINUS_SRC_COLOR:
 795       /* Get unused tmp,
 796        * Emit tmp = 1.0 - arg.xyzw
 797        */
 798       arg = get_temp( p );
 799       one = get_one( p );
 800       return emit_arith( p, OPCODE_SUB, arg, mask, 0, one, src, undef);
 801
 802    case OPR_SRC_ALPHA:
 803       if (mask == WRITEMASK_W)
 804          return src;
 805       else
 806          return swizzle1( src, SWIZZLE_W );
 807    case OPR_ONE_MINUS_SRC_ALPHA:
 808       /* Get unused tmp,
 809        * Emit tmp = 1.0 - arg.wwww
 810        */
 811       arg = get_temp(p);
 812       one = get_one(p);
 813       return emit_arith(p, OPCODE_SUB, arg, mask, 0,
 814                         one, swizzle1(src, SWIZZLE_W), undef);
 815    case OPR_ZERO:
 816       return get_zero(p);
 817    case OPR_ONE:
 818       return get_one(p);
 819    case OPR_SRC_COLOR:
 820    default:
 821       return src;
 822    }
 823 }
 824
 825 static GLboolean args_match( struct state_key *key, GLuint unit )
 826 {
 827    GLuint i, nr = key->unit[unit].NumArgsRGB;
 828
 829    for (i = 0 ; i < nr ; i++) {
 830       if (key->unit[unit].OptA[i].Source != key->unit[unit].OptRGB[i].Source)
 831          return GL_FALSE;
 832
 833       switch(key->unit[unit].OptA[i].Operand) {
 834       case OPR_SRC_ALPHA:
 835          switch(key->unit[unit].OptRGB[i].Operand) {
 836          case OPR_SRC_COLOR:
 837          case OPR_SRC_ALPHA:
 838             break;
 839          default:
 840             return GL_FALSE;
 841          }
 842          break;
 843       case OPR_ONE_MINUS_SRC_ALPHA:
 844          switch(key->unit[unit].OptRGB[i].Operand) {
 845          case OPR_ONE_MINUS_SRC_COLOR:
 846          case OPR_ONE_MINUS_SRC_ALPHA:
 847             break;
 848          default:
 849             return GL_FALSE;
 850          }
 851          break;
 852       default:
 853          return GL_FALSE;       /* impossible */
 854       }
 855    }
 856
 857    return GL_TRUE;
 858 }
 859
 860 static struct ureg emit_combine( struct texenv_fragment_program *p,
 861                                  struct ureg dest,
 862                                  GLuint mask,
 863                                  GLboolean saturate,
 864                                  GLuint unit,
 865                                  GLuint nr,
 866                                  GLuint mode,
 867                                  const struct mode_opt *opt)
 868 {
 869    struct ureg src[3];
 870    struct ureg tmp, half;
 871    GLuint i;
 872
 873    tmp = undef; /* silence warning (bug 5318) */
 874
 875    for (i = 0; i < nr; i++)
 876       src[i] = emit_combine_source( p, mask, unit, opt[i].Source, opt[i].Operand );
 877
 878    switch (mode) {
 879    case MODE_REPLACE:
 880       if (mask == WRITEMASK_XYZW && !saturate)
 881          return src[0];
 882       else
 883          return emit_arith( p, OPCODE_MOV, dest, mask, saturate, src[0], undef, undef );
 884    case MODE_MODULATE:
 885       return emit_arith( p, OPCODE_MUL, dest, mask, saturate,
 886                          src[0], src[1], undef );
 887    case MODE_ADD:
 888       return emit_arith( p, OPCODE_ADD, dest, mask, saturate,
 889                          src[0], src[1], undef );
 890    case MODE_ADD_SIGNED:
 891       /* tmp = arg0 + arg1
 892        * result = tmp - .5
 893        */
 894       half = get_half(p);
 895       tmp = get_temp( p );
 896       emit_arith( p, OPCODE_ADD, tmp, mask, 0, src[0], src[1], undef );
 897       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp, half, undef );
 898       return dest;
 899    case MODE_INTERPOLATE:
 900       /* Arg0 * (Arg2) + Arg1 * (1-Arg2) -- note arguments are reordered:
 901        */
 902       return emit_arith( p, OPCODE_LRP, dest, mask, saturate, src[2], src[0], src[1] );
 903
 904    case MODE_SUBTRACT:
 905       return emit_arith( p, OPCODE_SUB, dest, mask, saturate, src[0], src[1], undef );
 906
 907    case MODE_DOT3_RGBA:
 908    case MODE_DOT3_RGBA_EXT:
 909    case MODE_DOT3_RGB_EXT:
 910    case MODE_DOT3_RGB: {
 911       struct ureg tmp0 = get_temp( p );
 912       struct ureg tmp1 = get_temp( p );
 913       struct ureg neg1 = register_scalar_const(p, -1);
 914       struct ureg two  = register_scalar_const(p, 2);
 915
 916       /* tmp0 = 2*src0 - 1
 917        * tmp1 = 2*src1 - 1
 918        *
 919        * dst = tmp0 dot3 tmp1
 920        */
 921       emit_arith( p, OPCODE_MAD, tmp0, WRITEMASK_XYZW, 0,
 922                   two, src[0], neg1);
 923
 924       if (_mesa_memcmp(&src[0], &src[1], sizeof(struct ureg)) == 0)
 925          tmp1 = tmp0;
 926       else
 927          emit_arith( p, OPCODE_MAD, tmp1, WRITEMASK_XYZW, 0,
 928                      two, src[1], neg1);
 929       emit_arith( p, OPCODE_DP3, dest, mask, saturate, tmp0, tmp1, undef);
 930       return dest;
 931    }
 932    case MODE_MODULATE_ADD_ATI:
 933       /* Arg0 * Arg2 + Arg1 */
 934       return emit_arith( p, OPCODE_MAD, dest, mask, saturate,
 935                          src[0], src[2], src[1] );
 936    case MODE_MODULATE_SIGNED_ADD_ATI: {
 937       /* Arg0 * Arg2 + Arg1 - 0.5 */
 938       struct ureg tmp0 = get_temp(p);
 939       half = get_half(p);
 940       emit_arith( p, OPCODE_MAD, tmp0, mask, 0, src[0], src[2], src[1] );
 941       emit_arith( p, OPCODE_SUB, dest, mask, saturate, tmp0, half, undef );
 942       return dest;
 943    }
 944    case MODE_MODULATE_SUBTRACT_ATI:
 945       /* Arg0 * Arg2 - Arg1 */
 946       emit_arith( p, OPCODE_MAD, dest, mask, 0, src[0], src[2], negate(src[1]) );
 947       return dest;
 948    default:
 949       return src[0];
 950    }
 951 }
 952
 953
 954 /**
 955  * Generate instructions for one texture unit's env/combiner mode.
 956  */
 957 static struct ureg
 958 emit_texenv(struct texenv_fragment_program *p, GLuint unit)
 959 {
 960    struct state_key *key = p->state;
 961    GLboolean saturate = (unit < p->last_tex_stage);
 962    GLuint rgb_shift, alpha_shift;
 963    struct ureg out, shift;
 964    struct ureg dest;
 965
 966    if (!key->unit[unit].enabled) {
 967       return get_source(p, SRC_PREVIOUS, 0);
 968    }
 969
 970    switch (key->unit[unit].ModeRGB) {
 971    case MODE_DOT3_RGB_EXT:
 972       alpha_shift = key->unit[unit].ScaleShiftA;
 973       rgb_shift = 0;
 974       break;
 975    case MODE_DOT3_RGBA_EXT:
 976       alpha_shift = 0;
 977       rgb_shift = 0;
 978       break;
 979    default:
 980       rgb_shift = key->unit[unit].ScaleShiftRGB;
 981       alpha_shift = key->unit[unit].ScaleShiftA;
 982       break;
 983    }
 984
 985    /* If this is the very last calculation, emit direct to output reg:
 986     */
 987    if (key->separate_specular ||
 988        unit != p->last_tex_stage ||
 989        alpha_shift ||
 990        rgb_shift)
 991       dest = get_temp( p );
 992    else
 993       dest = make_ureg(PROGRAM_OUTPUT, FRAG_RESULT_COLR);
 994
 995    /* Emit the RGB and A combine ops
 996     */
 997    if (key->unit[unit].ModeRGB == key->unit[unit].ModeA &&
 998        args_match(key, unit)) {
 999       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
1000                           unit,
1001                           key->unit[unit].NumArgsRGB,
1002                           key->unit[unit].ModeRGB,
1003                           key->unit[unit].OptRGB);
1004    }
1005    else if (key->unit[unit].ModeRGB == MODE_DOT3_RGBA_EXT ||
1006             key->unit[unit].ModeRGB == MODE_DOT3_RGBA) {
1007
1008       out = emit_combine( p, dest, WRITEMASK_XYZW, saturate,
1009                           unit,
1010                           key->unit[unit].NumArgsRGB,
1011                           key->unit[unit].ModeRGB,
1012                           key->unit[unit].OptRGB);
1013    }
1014    else {
1015       /* Need to do something to stop from re-emitting identical
1016        * argument calculations here:
1017        */
1018       out = emit_combine( p, dest, WRITEMASK_XYZ, saturate,
1019                           unit,
1020                           key->unit[unit].NumArgsRGB,
1021                           key->unit[unit].ModeRGB,
1022                           key->unit[unit].OptRGB);
1023       out = emit_combine( p, dest, WRITEMASK_W, saturate,
1024                           unit,
1025                           key->unit[unit].NumArgsA,
1026                           key->unit[unit].ModeA,
1027                           key->unit[unit].OptA);
1028    }
1029
1030    /* Deal with the final shift:
1031     */
1032    if (alpha_shift || rgb_shift) {
1033       if (rgb_shift == alpha_shift) {
1034          shift = register_scalar_const(p, (GLfloat)(1<<rgb_shift));
1035       }
1036       else {
1037          shift = register_const4f(p,
1038                                   (GLfloat)(1<<rgb_shift),
1039                                   (GLfloat)(1<<rgb_shift),
1040                                   (GLfloat)(1<<rgb_shift),
1041                                   (GLfloat)(1<<alpha_shift));
1042       }
1043       return emit_arith( p, OPCODE_MUL, dest, WRITEMASK_XYZW,
1044                          saturate, out, shift, undef );
1045    }
1046    else
1047       return out;
1048 }
1049
1050
1051 /**
1052  * Generate instruction for getting a texture source term.
1053  */
1054 static void load_texture( struct texenv_fragment_program *p, GLuint unit )
1055 {
1056    if (is_undef(p->src_texture[unit])) {
1057       GLuint dim = p->state->unit[unit].source_index;
1058       struct ureg texcoord = register_input(p, FRAG_ATTRIB_TEX0+unit);
1059       struct ureg tmp = get_tex_temp( p );
1060
1061       if (dim == TEXTURE_UNKNOWN_INDEX)
1062          program_error(p, "TexSrcBit");
1063
1064       /* TODO: Use D0_MASK_XY where possible.
1065        */
1066       if (p->state->unit[unit].enabled) {
1067          p->src_texture[unit] = emit_texld( p, OPCODE_TXP,
1068                                             tmp, WRITEMASK_XYZW,
1069                                             unit, dim, texcoord );
1070
1071          if (p->state->unit[unit].shadow)
1072             p->program->Base.ShadowSamplers |= 1 << unit;
1073
1074          p->program->Base.SamplersUsed |= (1 << unit);
1075          /* This identity mapping should already be in place
1076           * (see _mesa_init_program_struct()) but let's be safe.
1077           */
1078          p->program->Base.SamplerUnits[unit] = unit;
1079       }
1080       else
1081          p->src_texture[unit] = get_zero(p);
1082    }
1083 }
1084
1085 static GLboolean load_texenv_source( struct texenv_fragment_program *p,
1086                                      GLuint src, GLuint unit )
1087 {
1088    switch (src) {
1089    case SRC_TEXTURE:
1090       load_texture(p, unit);
1091       break;
1092
1093    case SRC_TEXTURE0:
1094    case SRC_TEXTURE1:
1095    case SRC_TEXTURE2:
1096    case SRC_TEXTURE3:
1097    case SRC_TEXTURE4:
1098    case SRC_TEXTURE5:
1099    case SRC_TEXTURE6:
1100    case SRC_TEXTURE7:
1101       load_texture(p, src - SRC_TEXTURE0);
1102       break;
1103
1104    default:
1105       break;
1106    }
1107
1108    return GL_TRUE;
1109 }
1110
1111
1112 /**
1113  * Generate instructions for loading all texture source terms.
1114  */
1115 static GLboolean
1116 load_texunit_sources( struct texenv_fragment_program *p, int unit )
1117 {
1118    struct state_key *key = p->state;
1119    GLuint i;
1120
1121    for (i = 0; i < key->unit[unit].NumArgsRGB; i++) {
1122       load_texenv_source( p, key->unit[unit].OptRGB[i].Source, unit);
1123    }
1124
1125    for (i = 0; i < key->unit[unit].NumArgsA; i++) {
1126       load_texenv_source( p, key->unit[unit].OptA[i].Source, unit );
1127    }
1128
1129    return GL_TRUE;
1130 }
1131
1132
1133 /**
1134  * Generate a new fragment program which implements the context's
1135  * current texture env/combine mode.
1136  */
1137 static void
1138 create_new_program(GLcontext *ctx, struct state_key *key,
1139                    struct gl_fragment_program *program)
1140 {
1141    struct prog_instruction instBuffer[MAX_INSTRUCTIONS];
1142    struct texenv_fragment_program p;
1143    GLuint unit;
1144    struct ureg cf, out;
1145
1146    _mesa_memset(&p, 0, sizeof(p));
1147    p.ctx = ctx;
1148    p.state = key;
1149    p.program = program;
1150
1151    /* During code generation, use locally-allocated instruction buffer,
1152     * then alloc dynamic storage below.
1153     */
1154    p.program->Base.Instructions = instBuffer;
1155    p.program->Base.Target = GL_FRAGMENT_PROGRAM_ARB;
1156    p.program->Base.NumTexIndirections = 1;      /* correct? */
1157    p.program->Base.NumTexInstructions = 0;
1158    p.program->Base.NumAluInstructions = 0;
1159    p.program->Base.String = NULL;
1160    p.program->Base.NumInstructions =
1161    p.program->Base.NumTemporaries =
1162    p.program->Base.NumParameters =
1163    p.program->Base.NumAttributes = p.program->Base.NumAddressRegs = 0;
1164    p.program->Base.Parameters = _mesa_new_parameter_list();
1165
1166    p.program->Base.InputsRead = 0;
1167    p.program->Base.OutputsWritten = 1 << FRAG_RESULT_COLR;
1168
1169    for (unit = 0; unit < MAX_TEXTURE_UNITS; unit++)
1170       p.src_texture[unit] = undef;
1171
1172    p.src_previous = undef;
1173    p.half = undef;
1174    p.zero = undef;
1175    p.one = undef;
1176
1177    p.last_tex_stage = 0;
1178    release_temps(ctx, &p);
1179
1180    if (key->enabled_units) {
1181       /* First pass - to support texture_env_crossbar, first identify
1182        * all referenced texture sources and emit texld instructions
1183        * for each:
1184        */
1185       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits ; unit++)
1186          if (key->unit[unit].enabled) {
1187             load_texunit_sources( &p, unit );
1188             p.last_tex_stage = unit;
1189          }
1190
1191       /* Second pass - emit combine instructions to build final color:
1192        */
1193       for (unit = 0 ; unit < ctx->Const.MaxTextureUnits; unit++)
1194          if (key->enabled_units & (1<<unit)) {
1195             p.src_previous = emit_texenv( &p, unit );
1196             reserve_temp(&p, p.src_previous); /* don't re-use this temp reg */
1197             release_temps(ctx, &p);     /* release all temps */
1198          }
1199    }
1200
1201    cf = get_source( &p, SRC_PREVIOUS, 0 );
1202    out = make_ureg( PROGRAM_OUTPUT, FRAG_RESULT_COLR );
1203
1204    if (key->separate_specular) {
1205       /* Emit specular add.
1206        */
1207       struct ureg s = register_input(&p, FRAG_ATTRIB_COL1);
1208       emit_arith( &p, OPCODE_ADD, out, WRITEMASK_XYZ, 0, cf, s, undef );
1209       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_W, 0, cf, undef, undef );
1210    }
1211    else if (_mesa_memcmp(&cf, &out, sizeof(cf)) != 0) {
1212       /* Will wind up in here if no texture enabled or a couple of
1213        * other scenarios (GL_REPLACE for instance).
1214        */
1215       emit_arith( &p, OPCODE_MOV, out, WRITEMASK_XYZW, 0, cf, undef, undef );
1216    }
1217
1218    /* Finish up:
1219     */
1220    emit_arith( &p, OPCODE_END, undef, WRITEMASK_XYZW, 0, undef, undef, undef);
1221
1222    if (key->fog_enabled) {
1223       /* Pull fog mode from GLcontext, the value in the state key is
1224        * a reduced value and not what is expected in FogOption
1225        */
1226       p.program->FogOption = ctx->Fog.Mode;
1227       p.program->Base.InputsRead |= FRAG_BIT_FOGC; /* XXX new */
1228    } else
1229       p.program->FogOption = GL_NONE;
1230
1231    if (p.program->Base.NumTexIndirections > ctx->Const.FragmentProgram.MaxTexIndirections)
1232       program_error(&p, "Exceeded max nr indirect texture lookups");
1233
1234    if (p.program->Base.NumTexInstructions > ctx->Const.FragmentProgram.MaxTexInstructions)
1235       program_error(&p, "Exceeded max TEX instructions");
1236
1237    if (p.program->Base.NumAluInstructions > ctx->Const.FragmentProgram.MaxAluInstructions)
1238       program_error(&p, "Exceeded max ALU instructions");
1239
1240    ASSERT(p.program->Base.NumInstructions <= MAX_INSTRUCTIONS);
1241
1242    /* Allocate final instruction array */
1243    p.program->Base.Instructions
1244       = _mesa_alloc_instructions(p.program->Base.NumInstructions);
1245    if (!p.program->Base.Instructions) {
1246       _mesa_error(ctx, GL_OUT_OF_MEMORY,
1247                   "generating tex env program");
1248       return;
1249    }
1250    _mesa_copy_instructions(p.program->Base.Instructions, instBuffer,
1251                            p.program->Base.NumInstructions);
1252
1253    if (p.program->FogOption) {
1254       _mesa_append_fog_code(ctx, p.program);
1255       p.program->FogOption = GL_NONE;
1256    }
1257
1258
1259    /* Notify driver the fragment program has (actually) changed.
1260     */
1261    if (ctx->Driver.ProgramStringNotify) {
1262       ctx->Driver.ProgramStringNotify( ctx, GL_FRAGMENT_PROGRAM_ARB,
1263                                        &p.program->Base );
1264    }
1265
1266    if (DISASSEM) {
1267       _mesa_print_program(&p.program->Base);
1268       _mesa_printf("\n");
1269    }
1270 }
1271
1272
1273 /**
1274  * Return a fragment program which implements the current
1275  * fixed-function texture, fog and color-sum operations.
1276  */
1277 struct gl_fragment_program *
1278 _mesa_get_fixed_func_fragment_program(GLcontext *ctx)
1279 {
1280    struct gl_fragment_program *prog;
1281    struct state_key key;
1282
1283    make_state_key(ctx, &key);
1284
1285    prog = (struct gl_fragment_program *)
1286       _mesa_search_program_cache(ctx->FragmentProgram.Cache,
1287                                  &key, sizeof(key));
1288
1289    if (!prog) {
1290       prog = (struct gl_fragment_program *)
1291          ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
1292
1293       create_new_program(ctx, &key, prog);
1294
1295       _mesa_program_cache_insert(ctx, ctx->FragmentProgram.Cache,
1296                                  &key, sizeof(key), &prog->Base);
1297    }
1298
1299    return prog;
1300 }
1301
1302
1303
1304 /**
1305  * If _MaintainTexEnvProgram is set we'll generate a fragment program that
1306  * implements the current texture env/combine mode.
1307  * This function generates that program and puts it into effect.
1308  *
1309  * XXX: remove this function.  currently only called by some drivers,
1310  * not by mesa core.  We now handle this properly from inside mesa.
1311  */
1312 void
1313 _mesa_UpdateTexEnvProgram( GLcontext *ctx )
1314 {
1315    const struct gl_fragment_program *prev = ctx->FragmentProgram._Current;
1316
1317    ASSERT(ctx->FragmentProgram._MaintainTexEnvProgram);
1318
1319    /* If a conventional fragment program/shader isn't in effect... */
1320    if (!ctx->FragmentProgram._Enabled &&
1321        (!ctx->Shader.CurrentProgram ||
1322         !ctx->Shader.CurrentProgram->FragmentProgram) )
1323    {
1324       struct gl_fragment_program *newProg;
1325
1326       newProg = _mesa_get_fixed_func_fragment_program(ctx);
1327
1328       _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current, newProg);
1329       _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._TexEnvProgram, newProg);
1330    }
1331
1332    /* Tell the driver about the change.  Could define a new target for
1333     * this?
1334     */
1335    if (ctx->FragmentProgram._Current != prev && ctx->Driver.BindProgram) {
1336       ctx->Driver.BindProgram(ctx, GL_FRAGMENT_PROGRAM_ARB,
1337                          (struct gl_program *) ctx->FragmentProgram._Current);
1338    }
1339 }