src/mesa/shader/programopt.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  * BRIAN PAUL 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  programopt.c
  27  * Vertex/Fragment program optimizations and transformations for program
  28  * options, etc.
  29  *
  30  * \author Brian Paul
  31  */
  32
  33
  34 #include "main/glheader.h"
  35 #include "main/context.h"
  36 #include "prog_parameter.h"
  37 #include "prog_statevars.h"
  38 #include "program.h"
  39 #include "programopt.h"
  40 #include "prog_instruction.h"
  41
  42
  43 /**
  44  * This function inserts instructions for coordinate modelview * projection
  45  * into a vertex program.
  46  * May be used to implement the position_invariant option.
  47  */
  48 static void
  49 _mesa_insert_mvp_dp4_code(GLcontext *ctx, struct gl_vertex_program *vprog)
  50 {
  51    struct prog_instruction *newInst;
  52    const GLuint origLen = vprog->Base.NumInstructions;
  53    const GLuint newLen = origLen + 4;
  54    GLuint i;
  55
  56    /*
  57     * Setup state references for the modelview/projection matrix.
  58     * XXX we should check if these state vars are already declared.
  59     */
  60    static const gl_state_index mvpState[4][STATE_LENGTH] = {
  61       { STATE_MVP_MATRIX, 0, 0, 0, 0 },  /* state.matrix.mvp.row[0] */
  62       { STATE_MVP_MATRIX, 0, 1, 1, 0 },  /* state.matrix.mvp.row[1] */
  63       { STATE_MVP_MATRIX, 0, 2, 2, 0 },  /* state.matrix.mvp.row[2] */
  64       { STATE_MVP_MATRIX, 0, 3, 3, 0 },  /* state.matrix.mvp.row[3] */
  65    };
  66    GLint mvpRef[4];
  67
  68    for (i = 0; i < 4; i++) {
  69       mvpRef[i] = _mesa_add_state_reference(vprog->Base.Parameters,
  70                                             mvpState[i]);
  71    }
  72
  73    /* Alloc storage for new instructions */
  74    newInst = _mesa_alloc_instructions(newLen);
  75    if (!newInst) {
  76       _mesa_error(ctx, GL_OUT_OF_MEMORY,
  77                   "glProgramString(inserting position_invariant code)");
  78       return;
  79    }
  80
  81    /*
  82     * Generated instructions:
  83     * newInst[0] = DP4 result.position.x, mvp.row[0], vertex.position;
  84     * newInst[1] = DP4 result.position.y, mvp.row[1], vertex.position;
  85     * newInst[2] = DP4 result.position.z, mvp.row[2], vertex.position;
  86     * newInst[3] = DP4 result.position.w, mvp.row[3], vertex.position;
  87     */
  88    _mesa_init_instructions(newInst, 4);
  89    for (i = 0; i < 4; i++) {
  90       newInst[i].Opcode = OPCODE_DP4;
  91       newInst[i].DstReg.File = PROGRAM_OUTPUT;
  92       newInst[i].DstReg.Index = VERT_RESULT_HPOS;
  93       newInst[i].DstReg.WriteMask = (WRITEMASK_X << i);
  94       newInst[i].SrcReg[0].File = PROGRAM_STATE_VAR;
  95       newInst[i].SrcReg[0].Index = mvpRef[i];
  96       newInst[i].SrcReg[0].Swizzle = SWIZZLE_NOOP;
  97       newInst[i].SrcReg[1].File = PROGRAM_INPUT;
  98       newInst[i].SrcReg[1].Index = VERT_ATTRIB_POS;
  99       newInst[i].SrcReg[1].Swizzle = SWIZZLE_NOOP;
 100    }
 101
 102    /* Append original instructions after new instructions */
 103    _mesa_copy_instructions (newInst + 4, vprog->Base.Instructions, origLen);
 104
 105    /* free old instructions */
 106    _mesa_free_instructions(vprog->Base.Instructions, origLen);
 107
 108    /* install new instructions */
 109    vprog->Base.Instructions = newInst;
 110    vprog->Base.NumInstructions = newLen;
 111    vprog->Base.InputsRead |= VERT_BIT_POS;
 112    vprog->Base.OutputsWritten |= BITFIELD64_BIT(VERT_RESULT_HPOS);
 113 }
 114
 115
 116 static void
 117 _mesa_insert_mvp_mad_code(GLcontext *ctx, struct gl_vertex_program *vprog)
 118 {
 119    struct prog_instruction *newInst;
 120    const GLuint origLen = vprog->Base.NumInstructions;
 121    const GLuint newLen = origLen + 4;
 122    GLuint hposTemp;
 123    GLuint i;
 124
 125    /*
 126     * Setup state references for the modelview/projection matrix.
 127     * XXX we should check if these state vars are already declared.
 128     */
 129    static const gl_state_index mvpState[4][STATE_LENGTH] = {
 130       { STATE_MVP_MATRIX, 0, 0, 0, STATE_MATRIX_TRANSPOSE },
 131       { STATE_MVP_MATRIX, 0, 1, 1, STATE_MATRIX_TRANSPOSE },
 132       { STATE_MVP_MATRIX, 0, 2, 2, STATE_MATRIX_TRANSPOSE },
 133       { STATE_MVP_MATRIX, 0, 3, 3, STATE_MATRIX_TRANSPOSE },
 134    };
 135    GLint mvpRef[4];
 136
 137    for (i = 0; i < 4; i++) {
 138       mvpRef[i] = _mesa_add_state_reference(vprog->Base.Parameters,
 139                                             mvpState[i]);
 140    }
 141
 142    /* Alloc storage for new instructions */
 143    newInst = _mesa_alloc_instructions(newLen);
 144    if (!newInst) {
 145       _mesa_error(ctx, GL_OUT_OF_MEMORY,
 146                   "glProgramString(inserting position_invariant code)");
 147       return;
 148    }
 149
 150    /* TEMP hposTemp; */
 151    hposTemp = vprog->Base.NumTemporaries++;
 152
 153    /*
 154     * Generated instructions:
 155     *    emit_op2(p, OPCODE_MUL, tmp, 0, swizzle1(src,X), mat[0]);
 156     *    emit_op3(p, OPCODE_MAD, tmp, 0, swizzle1(src,Y), mat[1], tmp);
 157     *    emit_op3(p, OPCODE_MAD, tmp, 0, swizzle1(src,Z), mat[2], tmp);
 158     *    emit_op3(p, OPCODE_MAD, dest, 0, swizzle1(src,W), mat[3], tmp);
 159     */
 160    _mesa_init_instructions(newInst, 4);
 161
 162    newInst[0].Opcode = OPCODE_MUL;
 163    newInst[0].DstReg.File = PROGRAM_TEMPORARY;
 164    newInst[0].DstReg.Index = hposTemp;
 165    newInst[0].DstReg.WriteMask = WRITEMASK_XYZW;
 166    newInst[0].SrcReg[0].File = PROGRAM_INPUT;
 167    newInst[0].SrcReg[0].Index = VERT_ATTRIB_POS;
 168    newInst[0].SrcReg[0].Swizzle = SWIZZLE_XXXX;
 169    newInst[0].SrcReg[1].File = PROGRAM_STATE_VAR;
 170    newInst[0].SrcReg[1].Index = mvpRef[0];
 171    newInst[0].SrcReg[1].Swizzle = SWIZZLE_NOOP;
 172
 173    for (i = 1; i <= 2; i++) {
 174       newInst[i].Opcode = OPCODE_MAD;
 175       newInst[i].DstReg.File = PROGRAM_TEMPORARY;
 176       newInst[i].DstReg.Index = hposTemp;
 177       newInst[i].DstReg.WriteMask = WRITEMASK_XYZW;
 178       newInst[i].SrcReg[0].File = PROGRAM_INPUT;
 179       newInst[i].SrcReg[0].Index = VERT_ATTRIB_POS;
 180       newInst[i].SrcReg[0].Swizzle = MAKE_SWIZZLE4(i,i,i,i);
 181       newInst[i].SrcReg[1].File = PROGRAM_STATE_VAR;
 182       newInst[i].SrcReg[1].Index = mvpRef[i];
 183       newInst[i].SrcReg[1].Swizzle = SWIZZLE_NOOP;
 184       newInst[i].SrcReg[2].File = PROGRAM_TEMPORARY;
 185       newInst[i].SrcReg[2].Index = hposTemp;
 186       newInst[1].SrcReg[2].Swizzle = SWIZZLE_NOOP;
 187    }
 188
 189    newInst[3].Opcode = OPCODE_MAD;
 190    newInst[3].DstReg.File = PROGRAM_OUTPUT;
 191    newInst[3].DstReg.Index = VERT_RESULT_HPOS;
 192    newInst[3].DstReg.WriteMask = WRITEMASK_XYZW;
 193    newInst[3].SrcReg[0].File = PROGRAM_INPUT;
 194    newInst[3].SrcReg[0].Index = VERT_ATTRIB_POS;
 195    newInst[3].SrcReg[0].Swizzle = SWIZZLE_WWWW;
 196    newInst[3].SrcReg[1].File = PROGRAM_STATE_VAR;
 197    newInst[3].SrcReg[1].Index = mvpRef[3];
 198    newInst[3].SrcReg[1].Swizzle = SWIZZLE_NOOP;
 199    newInst[3].SrcReg[2].File = PROGRAM_TEMPORARY;
 200    newInst[3].SrcReg[2].Index = hposTemp;
 201    newInst[3].SrcReg[2].Swizzle = SWIZZLE_NOOP;
 202
 203
 204    /* Append original instructions after new instructions */
 205    _mesa_copy_instructions (newInst + 4, vprog->Base.Instructions, origLen);
 206
 207    /* free old instructions */
 208    _mesa_free_instructions(vprog->Base.Instructions, origLen);
 209
 210    /* install new instructions */
 211    vprog->Base.Instructions = newInst;
 212    vprog->Base.NumInstructions = newLen;
 213    vprog->Base.InputsRead |= VERT_BIT_POS;
 214    vprog->Base.OutputsWritten |= BITFIELD64_BIT(VERT_RESULT_HPOS);
 215 }
 216
 217
 218 void
 219 _mesa_insert_mvp_code(GLcontext *ctx, struct gl_vertex_program *vprog)
 220 {
 221    if (ctx->mvp_with_dp4)
 222       _mesa_insert_mvp_dp4_code( ctx, vprog );
 223    else
 224       _mesa_insert_mvp_mad_code( ctx, vprog );
 225 }
 226
 227
 228
 229
 230
 231
 232 /**
 233  * Append extra instructions onto the given fragment program to implement
 234  * the fog mode specified by fprog->FogOption.
 235  * The fragment.fogcoord input is used to compute the fog blend factor.
 236  *
 237  * XXX with a little work, this function could be adapted to add fog code
 238  * to vertex programs too.
 239  */
 240 void
 241 _mesa_append_fog_code(GLcontext *ctx, struct gl_fragment_program *fprog)
 242 {
 243    static const gl_state_index fogPStateOpt[STATE_LENGTH]
 244       = { STATE_INTERNAL, STATE_FOG_PARAMS_OPTIMIZED, 0, 0, 0 };
 245    static const gl_state_index fogColorState[STATE_LENGTH]
 246       = { STATE_FOG_COLOR, 0, 0, 0, 0};
 247    struct prog_instruction *newInst, *inst;
 248    const GLuint origLen = fprog->Base.NumInstructions;
 249    const GLuint newLen = origLen + 5;
 250    GLuint i;
 251    GLint fogPRefOpt, fogColorRef; /* state references */
 252    GLuint colorTemp, fogFactorTemp; /* temporary registerss */
 253
 254    if (fprog->FogOption == GL_NONE) {
 255       _mesa_problem(ctx, "_mesa_append_fog_code() called for fragment program"
 256                     " with FogOption == GL_NONE");
 257       return;
 258    }
 259
 260    /* Alloc storage for new instructions */
 261    newInst = _mesa_alloc_instructions(newLen);
 262    if (!newInst) {
 263       _mesa_error(ctx, GL_OUT_OF_MEMORY,
 264                   "glProgramString(inserting fog_option code)");
 265       return;
 266    }
 267
 268    /* Copy orig instructions into new instruction buffer */
 269    _mesa_copy_instructions(newInst, fprog->Base.Instructions, origLen);
 270
 271    /* PARAM fogParamsRefOpt = internal optimized fog params; */
 272    fogPRefOpt
 273       = _mesa_add_state_reference(fprog->Base.Parameters, fogPStateOpt);
 274    /* PARAM fogColorRef = state.fog.color; */
 275    fogColorRef
 276       = _mesa_add_state_reference(fprog->Base.Parameters, fogColorState);
 277
 278    /* TEMP colorTemp; */
 279    colorTemp = fprog->Base.NumTemporaries++;
 280    /* TEMP fogFactorTemp; */
 281    fogFactorTemp = fprog->Base.NumTemporaries++;
 282
 283    /* Scan program to find where result.color is written */
 284    inst = newInst;
 285    for (i = 0; i < fprog->Base.NumInstructions; i++) {
 286       if (inst->Opcode == OPCODE_END)
 287          break;
 288       if (inst->DstReg.File == PROGRAM_OUTPUT &&
 289           inst->DstReg.Index == FRAG_RESULT_COLOR) {
 290          /* change the instruction to write to colorTemp w/ clamping */
 291          inst->DstReg.File = PROGRAM_TEMPORARY;
 292          inst->DstReg.Index = colorTemp;
 293          inst->SaturateMode = SATURATE_ZERO_ONE;
 294          /* don't break (may be several writes to result.color) */
 295       }
 296       inst++;
 297    }
 298    assert(inst->Opcode == OPCODE_END); /* we'll overwrite this inst */
 299
 300    _mesa_init_instructions(inst, 5);
 301
 302    /* emit instructions to compute fog blending factor */
 303    if (fprog->FogOption == GL_LINEAR) {
 304       /* MAD fogFactorTemp.x, fragment.fogcoord.x, fogPRefOpt.x, fogPRefOpt.y; */
 305       inst->Opcode = OPCODE_MAD;
 306       inst->DstReg.File = PROGRAM_TEMPORARY;
 307       inst->DstReg.Index = fogFactorTemp;
 308       inst->DstReg.WriteMask = WRITEMASK_X;
 309       inst->SrcReg[0].File = PROGRAM_INPUT;
 310       inst->SrcReg[0].Index = FRAG_ATTRIB_FOGC;
 311       inst->SrcReg[0].Swizzle = SWIZZLE_XXXX;
 312       inst->SrcReg[1].File = PROGRAM_STATE_VAR;
 313       inst->SrcReg[1].Index = fogPRefOpt;
 314       inst->SrcReg[1].Swizzle = SWIZZLE_XXXX;
 315       inst->SrcReg[2].File = PROGRAM_STATE_VAR;
 316       inst->SrcReg[2].Index = fogPRefOpt;
 317       inst->SrcReg[2].Swizzle = SWIZZLE_YYYY;
 318       inst->SaturateMode = SATURATE_ZERO_ONE;
 319       inst++;
 320    }
 321    else {
 322       ASSERT(fprog->FogOption == GL_EXP || fprog->FogOption == GL_EXP2);
 323       /* fogPRefOpt.z = d/ln(2), fogPRefOpt.w = d/sqrt(ln(2) */
 324       /* EXP: MUL fogFactorTemp.x, fogPRefOpt.z, fragment.fogcoord.x; */
 325       /* EXP2: MUL fogFactorTemp.x, fogPRefOpt.w, fragment.fogcoord.x; */
 326       inst->Opcode = OPCODE_MUL;
 327       inst->DstReg.File = PROGRAM_TEMPORARY;
 328       inst->DstReg.Index = fogFactorTemp;
 329       inst->DstReg.WriteMask = WRITEMASK_X;
 330       inst->SrcReg[0].File = PROGRAM_STATE_VAR;
 331       inst->SrcReg[0].Index = fogPRefOpt;
 332       inst->SrcReg[0].Swizzle
 333          = (fprog->FogOption == GL_EXP) ? SWIZZLE_ZZZZ : SWIZZLE_WWWW;
 334       inst->SrcReg[1].File = PROGRAM_INPUT;
 335       inst->SrcReg[1].Index = FRAG_ATTRIB_FOGC;
 336       inst->SrcReg[1].Swizzle = SWIZZLE_XXXX;
 337       inst++;
 338       if (fprog->FogOption == GL_EXP2) {
 339          /* MUL fogFactorTemp.x, fogFactorTemp.x, fogFactorTemp.x; */
 340          inst->Opcode = OPCODE_MUL;
 341          inst->DstReg.File = PROGRAM_TEMPORARY;
 342          inst->DstReg.Index = fogFactorTemp;
 343          inst->DstReg.WriteMask = WRITEMASK_X;
 344          inst->SrcReg[0].File = PROGRAM_TEMPORARY;
 345          inst->SrcReg[0].Index = fogFactorTemp;
 346          inst->SrcReg[0].Swizzle = SWIZZLE_XXXX;
 347          inst->SrcReg[1].File = PROGRAM_TEMPORARY;
 348          inst->SrcReg[1].Index = fogFactorTemp;
 349          inst->SrcReg[1].Swizzle = SWIZZLE_XXXX;
 350          inst++;
 351       }
 352       /* EX2_SAT fogFactorTemp.x, -fogFactorTemp.x; */
 353       inst->Opcode = OPCODE_EX2;
 354       inst->DstReg.File = PROGRAM_TEMPORARY;
 355       inst->DstReg.Index = fogFactorTemp;
 356       inst->DstReg.WriteMask = WRITEMASK_X;
 357       inst->SrcReg[0].File = PROGRAM_TEMPORARY;
 358       inst->SrcReg[0].Index = fogFactorTemp;
 359       inst->SrcReg[0].Negate = NEGATE_XYZW;
 360       inst->SrcReg[0].Swizzle = SWIZZLE_XXXX;
 361       inst->SaturateMode = SATURATE_ZERO_ONE;
 362       inst++;
 363    }
 364    /* LRP result.color.xyz, fogFactorTemp.xxxx, colorTemp, fogColorRef; */
 365    inst->Opcode = OPCODE_LRP;
 366    inst->DstReg.File = PROGRAM_OUTPUT;
 367    inst->DstReg.Index = FRAG_RESULT_COLOR;
 368    inst->DstReg.WriteMask = WRITEMASK_XYZ;
 369    inst->SrcReg[0].File = PROGRAM_TEMPORARY;
 370    inst->SrcReg[0].Index = fogFactorTemp;
 371    inst->SrcReg[0].Swizzle = SWIZZLE_XXXX;
 372    inst->SrcReg[1].File = PROGRAM_TEMPORARY;
 373    inst->SrcReg[1].Index = colorTemp;
 374    inst->SrcReg[1].Swizzle = SWIZZLE_NOOP;
 375    inst->SrcReg[2].File = PROGRAM_STATE_VAR;
 376    inst->SrcReg[2].Index = fogColorRef;
 377    inst->SrcReg[2].Swizzle = SWIZZLE_NOOP;
 378    inst++;
 379    /* MOV result.color.w, colorTemp.x;  # copy alpha */
 380    inst->Opcode = OPCODE_MOV;
 381    inst->DstReg.File = PROGRAM_OUTPUT;
 382    inst->DstReg.Index = FRAG_RESULT_COLOR;
 383    inst->DstReg.WriteMask = WRITEMASK_W;
 384    inst->SrcReg[0].File = PROGRAM_TEMPORARY;
 385    inst->SrcReg[0].Index = colorTemp;
 386    inst->SrcReg[0].Swizzle = SWIZZLE_NOOP;
 387    inst++;
 388    /* END; */
 389    inst->Opcode = OPCODE_END;
 390    inst++;
 391
 392    /* free old instructions */
 393    _mesa_free_instructions(fprog->Base.Instructions, origLen);
 394
 395    /* install new instructions */
 396    fprog->Base.Instructions = newInst;
 397    fprog->Base.NumInstructions = inst - newInst;
 398    fprog->Base.InputsRead |= FRAG_BIT_FOGC;
 399    /* XXX do this?  fprog->FogOption = GL_NONE; */
 400 }
 401
 402
 403
 404 static GLboolean
 405 is_texture_instruction(const struct prog_instruction *inst)
 406 {
 407    switch (inst->Opcode) {
 408    case OPCODE_TEX:
 409    case OPCODE_TXB:
 410    case OPCODE_TXD:
 411    case OPCODE_TXL:
 412    case OPCODE_TXP:
 413    case OPCODE_TXP_NV:
 414       return GL_TRUE;
 415    default:
 416       return GL_FALSE;
 417    }
 418 }
 419
 420
 421 /**
 422  * Count the number of texure indirections in the given program.
 423  * The program's NumTexIndirections field will be updated.
 424  * See the GL_ARB_fragment_program spec (issue 24) for details.
 425  * XXX we count texture indirections in texenvprogram.c (maybe use this code
 426  * instead and elsewhere).
 427  */
 428 void
 429 _mesa_count_texture_indirections(struct gl_program *prog)
 430 {
 431    GLuint indirections = 1;
 432    GLbitfield tempsOutput = 0x0;
 433    GLbitfield aluTemps = 0x0;
 434    GLuint i;
 435
 436    for (i = 0; i < prog->NumInstructions; i++) {
 437       const struct prog_instruction *inst = prog->Instructions + i;
 438
 439       if (is_texture_instruction(inst)) {
 440          if (((inst->SrcReg[0].File == PROGRAM_TEMPORARY) &&
 441               (tempsOutput & (1 << inst->SrcReg[0].Index))) ||
 442              ((inst->Opcode != OPCODE_KIL) &&
 443               (inst->DstReg.File == PROGRAM_TEMPORARY) &&
 444               (aluTemps & (1 << inst->DstReg.Index))))
 445             {
 446                indirections++;
 447                tempsOutput = 0x0;
 448                aluTemps = 0x0;
 449             }
 450       }
 451       else {
 452          GLuint j;
 453          for (j = 0; j < 3; j++) {
 454             if (inst->SrcReg[j].File == PROGRAM_TEMPORARY)
 455                aluTemps |= (1 << inst->SrcReg[j].Index);
 456          }
 457          if (inst->DstReg.File == PROGRAM_TEMPORARY)
 458             aluTemps |= (1 << inst->DstReg.Index);
 459       }
 460
 461       if ((inst->Opcode != OPCODE_KIL) && (inst->DstReg.File == PROGRAM_TEMPORARY))
 462          tempsOutput |= (1 << inst->DstReg.Index);
 463    }
 464
 465    prog->NumTexIndirections = indirections;
 466 }
 467
 468
 469 /**
 470  * Count number of texture instructions in given program and update the
 471  * program's NumTexInstructions field.
 472  */
 473 void
 474 _mesa_count_texture_instructions(struct gl_program *prog)
 475 {
 476    GLuint i;
 477    prog->NumTexInstructions = 0;
 478    for (i = 0; i < prog->NumInstructions; i++) {
 479       prog->NumTexInstructions += is_texture_instruction(prog->Instructions + i);
 480    }
 481 }
 482
 483
 484 /**
 485  * Scan/rewrite program to remove reads of custom (output) registers.
 486  * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
 487  * (for vertex shaders).
 488  * In GLSL shaders, varying vars can be read and written.
 489  * On some hardware, trying to read an output register causes trouble.
 490  * So, rewrite the program to use a temporary register in this case.
 491  */
 492 void
 493 _mesa_remove_output_reads(struct gl_program *prog, gl_register_file type)
 494 {
 495    GLuint i;
 496    GLint outputMap[VERT_RESULT_MAX];
 497    GLuint numVaryingReads = 0;
 498    GLboolean usedTemps[MAX_PROGRAM_TEMPS];
 499    GLuint firstTemp = 0;
 500
 501    _mesa_find_used_registers(prog, PROGRAM_TEMPORARY,
 502                              usedTemps, MAX_PROGRAM_TEMPS);
 503
 504    assert(type == PROGRAM_VARYING || type == PROGRAM_OUTPUT);
 505    assert(prog->Target == GL_VERTEX_PROGRAM_ARB || type != PROGRAM_VARYING);
 506
 507    for (i = 0; i < VERT_RESULT_MAX; i++)
 508       outputMap[i] = -1;
 509
 510    /* look for instructions which read from varying vars */
 511    for (i = 0; i < prog->NumInstructions; i++) {
 512       struct prog_instruction *inst = prog->Instructions + i;
 513       const GLuint numSrc = _mesa_num_inst_src_regs(inst->Opcode);
 514       GLuint j;
 515       for (j = 0; j < numSrc; j++) {
 516          if (inst->SrcReg[j].File == type) {
 517             /* replace the read with a temp reg */
 518             const GLuint var = inst->SrcReg[j].Index;
 519             if (outputMap[var] == -1) {
 520                numVaryingReads++;
 521                outputMap[var] = _mesa_find_free_register(usedTemps,
 522                                                          MAX_PROGRAM_TEMPS,
 523                                                          firstTemp);
 524                firstTemp = outputMap[var] + 1;
 525             }
 526             inst->SrcReg[j].File = PROGRAM_TEMPORARY;
 527             inst->SrcReg[j].Index = outputMap[var];
 528          }
 529       }
 530    }
 531
 532    if (numVaryingReads == 0)
 533       return; /* nothing to be done */
 534
 535    /* look for instructions which write to the varying vars identified above */
 536    for (i = 0; i < prog->NumInstructions; i++) {
 537       struct prog_instruction *inst = prog->Instructions + i;
 538       if (inst->DstReg.File == type &&
 539           outputMap[inst->DstReg.Index] >= 0) {
 540          /* change inst to write to the temp reg, instead of the varying */
 541          inst->DstReg.File = PROGRAM_TEMPORARY;
 542          inst->DstReg.Index = outputMap[inst->DstReg.Index];
 543       }
 544    }
 545
 546    /* insert new instructions to copy the temp vars to the varying vars */
 547    {
 548       struct prog_instruction *inst;
 549       GLint endPos, var;
 550
 551       /* Look for END instruction and insert the new varying writes */
 552       endPos = -1;
 553       for (i = 0; i < prog->NumInstructions; i++) {
 554          struct prog_instruction *inst = prog->Instructions + i;
 555          if (inst->Opcode == OPCODE_END) {
 556             endPos = i;
 557             _mesa_insert_instructions(prog, i, numVaryingReads);
 558             break;
 559          }
 560       }
 561
 562       assert(endPos >= 0);
 563
 564       /* insert new MOV instructions here */
 565       inst = prog->Instructions + endPos;
 566       for (var = 0; var < VERT_RESULT_MAX; var++) {
 567          if (outputMap[var] >= 0) {
 568             /* MOV VAR[var], TEMP[tmp]; */
 569             inst->Opcode = OPCODE_MOV;
 570             inst->DstReg.File = type;
 571             inst->DstReg.Index = var;
 572             inst->SrcReg[0].File = PROGRAM_TEMPORARY;
 573             inst->SrcReg[0].Index = outputMap[var];
 574             inst++;
 575          }
 576       }
 577    }
 578 }
 579
 580
 581 /**
 582  * Make the given fragment program into a "no-op" shader.
 583  * Actually, just copy the incoming fragment color (or texcoord)
 584  * to the output color.
 585  * This is for debug/test purposes.
 586  */
 587 void
 588 _mesa_nop_fragment_program(GLcontext *ctx, struct gl_fragment_program *prog)
 589 {
 590    struct prog_instruction *inst;
 591    GLuint inputAttr;
 592
 593    inst = _mesa_alloc_instructions(2);
 594    if (!inst) {
 595       _mesa_error(ctx, GL_OUT_OF_MEMORY, "_mesa_nop_fragment_program");
 596       return;
 597    }
 598
 599    _mesa_init_instructions(inst, 2);
 600
 601    inst[0].Opcode = OPCODE_MOV;
 602    inst[0].DstReg.File = PROGRAM_OUTPUT;
 603    inst[0].DstReg.Index = FRAG_RESULT_COLOR;
 604    inst[0].SrcReg[0].File = PROGRAM_INPUT;
 605    if (prog->Base.InputsRead & FRAG_BIT_COL0)
 606       inputAttr = FRAG_ATTRIB_COL0;
 607    else
 608       inputAttr = FRAG_ATTRIB_TEX0;
 609    inst[0].SrcReg[0].Index = inputAttr;
 610
 611    inst[1].Opcode = OPCODE_END;
 612
 613    _mesa_free_instructions(prog->Base.Instructions,
 614                            prog->Base.NumInstructions);
 615
 616    prog->Base.Instructions = inst;
 617    prog->Base.NumInstructions = 2;
 618    prog->Base.InputsRead = 1 << inputAttr;
 619    prog->Base.OutputsWritten = BITFIELD64_BIT(FRAG_RESULT_COLOR);
 620 }
 621
 622
 623 /**
 624  * \sa _mesa_nop_fragment_program
 625  * Replace the given vertex program with a "no-op" program that just
 626  * transforms vertex position and emits color.
 627  */
 628 void
 629 _mesa_nop_vertex_program(GLcontext *ctx, struct gl_vertex_program *prog)
 630 {
 631    struct prog_instruction *inst;
 632    GLuint inputAttr;
 633
 634    /*
 635     * Start with a simple vertex program that emits color.
 636     */
 637    inst = _mesa_alloc_instructions(2);
 638    if (!inst) {
 639       _mesa_error(ctx, GL_OUT_OF_MEMORY, "_mesa_nop_vertex_program");
 640       return;
 641    }
 642
 643    _mesa_init_instructions(inst, 2);
 644
 645    inst[0].Opcode = OPCODE_MOV;
 646    inst[0].DstReg.File = PROGRAM_OUTPUT;
 647    inst[0].DstReg.Index = VERT_RESULT_COL0;
 648    inst[0].SrcReg[0].File = PROGRAM_INPUT;
 649    if (prog->Base.InputsRead & VERT_BIT_COLOR0)
 650       inputAttr = VERT_ATTRIB_COLOR0;
 651    else
 652       inputAttr = VERT_ATTRIB_TEX0;
 653    inst[0].SrcReg[0].Index = inputAttr;
 654
 655    inst[1].Opcode = OPCODE_END;
 656
 657    _mesa_free_instructions(prog->Base.Instructions,
 658                            prog->Base.NumInstructions);
 659
 660    prog->Base.Instructions = inst;
 661    prog->Base.NumInstructions = 2;
 662    prog->Base.InputsRead = 1 << inputAttr;
 663    prog->Base.OutputsWritten = BITFIELD64_BIT(VERT_RESULT_COL0);
 664
 665    /*
 666     * Now insert code to do standard modelview/projection transformation.
 667     */
 668    _mesa_insert_mvp_code(ctx, prog);
 669 }