src/gallium/drivers/r300/r300_fs.c

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  *                Joakim Sindholt <opensource@zhasha.com>
   4  * Copyright 2009 Marek Olšák <maraeo@gmail.com>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * on the rights to use, copy, modify, merge, publish, distribute, sub
  10  * license, and/or sell copies of the Software, and to permit persons to whom
  11  * the Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  21  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  22  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  23  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  24
  25 #include "util/u_format.h"
  26 #include "util/u_math.h"
  27 #include "util/u_memory.h"
  28
  29 #include "tgsi/tgsi_dump.h"
  30 #include "tgsi/tgsi_ureg.h"
  31
  32 #include "r300_cb.h"
  33 #include "r300_context.h"
  34 #include "r300_emit.h"
  35 #include "r300_screen.h"
  36 #include "r300_fs.h"
  37 #include "r300_reg.h"
  38 #include "r300_texture.h"
  39 #include "r300_tgsi_to_rc.h"
  40
  41 #include "compiler/radeon_compiler.h"
  42
  43 /* Convert info about FS input semantics to r300_shader_semantics. */
  44 void r300_shader_read_fs_inputs(struct tgsi_shader_info* info,
  45                                 struct r300_shader_semantics* fs_inputs)
  46 {
  47     int i;
  48     unsigned index;
  49
  50     r300_shader_semantics_reset(fs_inputs);
  51
  52     for (i = 0; i < info->num_inputs; i++) {
  53         index = info->input_semantic_index[i];
  54
  55         switch (info->input_semantic_name[i]) {
  56             case TGSI_SEMANTIC_COLOR:
  57                 assert(index < ATTR_COLOR_COUNT);
  58                 fs_inputs->color[index] = i;
  59                 break;
  60
  61             case TGSI_SEMANTIC_GENERIC:
  62                 assert(index < ATTR_GENERIC_COUNT);
  63                 fs_inputs->generic[index] = i;
  64                 break;
  65
  66             case TGSI_SEMANTIC_FOG:
  67                 assert(index == 0);
  68                 fs_inputs->fog = i;
  69                 break;
  70
  71             case TGSI_SEMANTIC_POSITION:
  72                 assert(index == 0);
  73                 fs_inputs->wpos = i;
  74                 break;
  75
  76             case TGSI_SEMANTIC_FACE:
  77                 assert(index == 0);
  78                 fs_inputs->face = i;
  79                 break;
  80
  81             default:
  82                 fprintf(stderr, "r300: FP: Unknown input semantic: %i\n",
  83                         info->input_semantic_name[i]);
  84         }
  85     }
  86 }
  87
  88 static void find_output_registers(struct r300_fragment_program_compiler * compiler,
  89                                   struct r300_fragment_shader_code *shader)
  90 {
  91     unsigned i, colorbuf_count = 0;
  92
  93     /* Mark the outputs as not present initially */
  94     compiler->OutputColor[0] = shader->info.num_outputs;
  95     compiler->OutputColor[1] = shader->info.num_outputs;
  96     compiler->OutputColor[2] = shader->info.num_outputs;
  97     compiler->OutputColor[3] = shader->info.num_outputs;
  98     compiler->OutputDepth = shader->info.num_outputs;
  99
 100     /* Now see where they really are. */
 101     for(i = 0; i < shader->info.num_outputs; ++i) {
 102         switch(shader->info.output_semantic_name[i]) {
 103             case TGSI_SEMANTIC_COLOR:
 104                 compiler->OutputColor[colorbuf_count] = i;
 105                 colorbuf_count++;
 106                 break;
 107             case TGSI_SEMANTIC_POSITION:
 108                 compiler->OutputDepth = i;
 109                 break;
 110         }
 111     }
 112 }
 113
 114 static void allocate_hardware_inputs(
 115     struct r300_fragment_program_compiler * c,
 116     void (*allocate)(void * data, unsigned input, unsigned hwreg),
 117     void * mydata)
 118 {
 119     struct r300_shader_semantics* inputs =
 120         (struct r300_shader_semantics*)c->UserData;
 121     int i, reg = 0;
 122
 123     /* Allocate input registers. */
 124     for (i = 0; i < ATTR_COLOR_COUNT; i++) {
 125         if (inputs->color[i] != ATTR_UNUSED) {
 126             allocate(mydata, inputs->color[i], reg++);
 127         }
 128     }
 129     if (inputs->face != ATTR_UNUSED) {
 130         allocate(mydata, inputs->face, reg++);
 131     }
 132     for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
 133         if (inputs->generic[i] != ATTR_UNUSED) {
 134             allocate(mydata, inputs->generic[i], reg++);
 135         }
 136     }
 137     if (inputs->fog != ATTR_UNUSED) {
 138         allocate(mydata, inputs->fog, reg++);
 139     }
 140     if (inputs->wpos != ATTR_UNUSED) {
 141         allocate(mydata, inputs->wpos, reg++);
 142     }
 143 }
 144
 145 static void get_external_state(
 146     struct r300_context* r300,
 147     struct r300_fragment_program_external_state* state)
 148 {
 149     struct r300_textures_state *texstate = r300->textures_state.state;
 150     unsigned i;
 151
 152     for (i = 0; i < texstate->sampler_state_count; i++) {
 153         struct r300_sampler_state *s = texstate->sampler_states[i];
 154         struct r300_sampler_view *v = texstate->sampler_views[i];
 155         struct r300_resource *t;
 156
 157         if (!s || !v) {
 158             continue;
 159         }
 160
 161         t = r300_resource(v->base.texture);
 162
 163         if (s->state.compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
 164             state->unit[i].compare_mode_enabled = 1;
 165
 166             /* Fortunately, no need to translate this. */
 167             state->unit[i].texture_compare_func = s->state.compare_func;
 168         }
 169
 170         state->unit[i].non_normalized_coords = !s->state.normalized_coords;
 171         state->unit[i].convert_unorm_to_snorm =
 172                 v->base.format == PIPE_FORMAT_RGTC1_SNORM ||
 173                 v->base.format == PIPE_FORMAT_LATC1_SNORM;
 174
 175         /* Pass texture swizzling to the compiler, some lowering passes need it. */
 176         if (v->base.format == PIPE_FORMAT_RGTC1_SNORM ||
 177             v->base.format == PIPE_FORMAT_LATC1_SNORM) {
 178             unsigned char swizzle[4];
 179
 180             util_format_compose_swizzles(
 181                             util_format_description(v->base.format)->swizzle,
 182                             v->swizzle,
 183                             swizzle);
 184
 185             state->unit[i].texture_swizzle =
 186                     RC_MAKE_SWIZZLE(swizzle[0], swizzle[1],
 187                                     swizzle[2], swizzle[3]);
 188         } else if (state->unit[i].compare_mode_enabled) {
 189             state->unit[i].texture_swizzle =
 190                 RC_MAKE_SWIZZLE(v->swizzle[0], v->swizzle[1],
 191                                 v->swizzle[2], v->swizzle[3]);
 192         }
 193
 194         /* XXX this should probably take into account STR, not just S. */
 195         if (t->tex.is_npot) {
 196             switch (s->state.wrap_s) {
 197             case PIPE_TEX_WRAP_REPEAT:
 198                 state->unit[i].wrap_mode = RC_WRAP_REPEAT;
 199                 break;
 200
 201             case PIPE_TEX_WRAP_MIRROR_REPEAT:
 202                 state->unit[i].wrap_mode = RC_WRAP_MIRRORED_REPEAT;
 203                 break;
 204
 205             case PIPE_TEX_WRAP_MIRROR_CLAMP:
 206             case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 207             case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 208                 state->unit[i].wrap_mode = RC_WRAP_MIRRORED_CLAMP;
 209                 break;
 210
 211             default:
 212                 state->unit[i].wrap_mode = RC_WRAP_NONE;
 213             }
 214
 215             if (t->b.b.target == PIPE_TEXTURE_3D)
 216                 state->unit[i].clamp_and_scale_before_fetch = TRUE;
 217         }
 218     }
 219 }
 220
 221 static void r300_translate_fragment_shader(
 222     struct r300_context* r300,
 223     struct r300_fragment_shader_code* shader,
 224     const struct tgsi_token *tokens);
 225
 226 static void r300_dummy_fragment_shader(
 227     struct r300_context* r300,
 228     struct r300_fragment_shader_code* shader)
 229 {
 230     struct pipe_shader_state state;
 231     struct ureg_program *ureg;
 232     struct ureg_dst out;
 233     struct ureg_src imm;
 234
 235     /* Make a simple fragment shader which outputs (0, 0, 0, 1) */
 236     ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 237     out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 238     imm = ureg_imm4f(ureg, 0, 0, 0, 1);
 239
 240     ureg_MOV(ureg, out, imm);
 241     ureg_END(ureg);
 242
 243     state.tokens = ureg_finalize(ureg);
 244
 245     shader->dummy = TRUE;
 246     r300_translate_fragment_shader(r300, shader, state.tokens);
 247
 248     ureg_destroy(ureg);
 249 }
 250
 251 static void r300_emit_fs_code_to_buffer(
 252     struct r300_context *r300,
 253     struct r300_fragment_shader_code *shader)
 254 {
 255     struct rX00_fragment_program_code *generic_code = &shader->code;
 256     unsigned imm_count = shader->immediates_count;
 257     unsigned imm_first = shader->externals_count;
 258     unsigned imm_end = generic_code->constants.Count;
 259     struct rc_constant *constants = generic_code->constants.Constants;
 260     unsigned i;
 261     CB_LOCALS;
 262
 263     if (r300->screen->caps.is_r500) {
 264         struct r500_fragment_program_code *code = &generic_code->code.r500;
 265
 266         shader->cb_code_size = 19 +
 267                                ((code->inst_end + 1) * 6) +
 268                                imm_count * 7 +
 269                                code->int_constant_count * 2;
 270
 271         NEW_CB(shader->cb_code, shader->cb_code_size);
 272         OUT_CB_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO);
 273         OUT_CB_REG(R500_US_PIXSIZE, code->max_temp_idx);
 274         OUT_CB_REG(R500_US_FC_CTRL, code->us_fc_ctrl);
 275         for(i = 0; i < code->int_constant_count; i++){
 276                 OUT_CB_REG(R500_US_FC_INT_CONST_0 + (i * 4),
 277                                                 code->int_constants[i]);
 278         }
 279         OUT_CB_REG(R500_US_CODE_RANGE,
 280                    R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(code->inst_end));
 281         OUT_CB_REG(R500_US_CODE_OFFSET, 0);
 282         OUT_CB_REG(R500_US_CODE_ADDR,
 283                    R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(code->inst_end));
 284
 285         OUT_CB_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_INSTR);
 286         OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, (code->inst_end + 1) * 6);
 287         for (i = 0; i <= code->inst_end; i++) {
 288             OUT_CB(code->inst[i].inst0);
 289             OUT_CB(code->inst[i].inst1);
 290             OUT_CB(code->inst[i].inst2);
 291             OUT_CB(code->inst[i].inst3);
 292             OUT_CB(code->inst[i].inst4);
 293             OUT_CB(code->inst[i].inst5);
 294         }
 295
 296         /* Emit immediates. */
 297         if (imm_count) {
 298             for(i = imm_first; i < imm_end; ++i) {
 299                 if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
 300                     const float *data = constants[i].u.Immediate;
 301
 302                     OUT_CB_REG(R500_GA_US_VECTOR_INDEX,
 303                                R500_GA_US_VECTOR_INDEX_TYPE_CONST |
 304                                (i & R500_GA_US_VECTOR_INDEX_MASK));
 305                     OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, 4);
 306                     OUT_CB_TABLE(data, 4);
 307                 }
 308             }
 309         }
 310     } else { /* r300 */
 311         struct r300_fragment_program_code *code = &generic_code->code.r300;
 312         unsigned int alu_length = code->alu.length;
 313         unsigned int alu_iterations = ((alu_length - 1) / 64) + 1;
 314         unsigned int tex_length = code->tex.length;
 315         unsigned int tex_iterations =
 316             tex_length > 0 ? ((tex_length - 1) / 32) + 1 : 0;
 317         unsigned int iterations =
 318             alu_iterations > tex_iterations ? alu_iterations : tex_iterations;
 319         unsigned int bank = 0;
 320
 321         shader->cb_code_size = 15 +
 322             /* R400_US_CODE_BANK */
 323             (r300->screen->caps.is_r400 ? 2 * (iterations + 1): 0) +
 324             /* R400_US_CODE_EXT */
 325             (r300->screen->caps.is_r400 ? 2 : 0) +
 326             /* R300_US_ALU_{RGB,ALPHA}_{INST,ADDR}_0, R400_US_ALU_EXT_ADDR_0 */
 327             (code->r390_mode ? (5 * alu_iterations) : 4) +
 328             /* R400_US_ALU_EXT_ADDR_[0-63] */
 329             (code->r390_mode ? (code->alu.length) : 0) +
 330             /* R300_US_ALU_{RGB,ALPHA}_{INST,ADDR}_0 */
 331             code->alu.length * 4 +
 332             /* R300_US_TEX_INST_0, R300_US_TEX_INST_[0-31] */
 333             (code->tex.length > 0 ? code->tex.length + tex_iterations : 0) +
 334             imm_count * 5;
 335
 336         NEW_CB(shader->cb_code, shader->cb_code_size);
 337
 338         OUT_CB_REG(R300_US_CONFIG, code->config);
 339         OUT_CB_REG(R300_US_PIXSIZE, code->pixsize);
 340         OUT_CB_REG(R300_US_CODE_OFFSET, code->code_offset);
 341
 342         if (code->r390_mode) {
 343             OUT_CB_REG(R400_US_CODE_EXT, code->r400_code_offset_ext);
 344         } else if (r300->screen->caps.is_r400) {
 345             /* This register appears to affect shaders even if r390_mode is
 346              * disabled, so it needs to be set to 0 for shaders that
 347              * don't use r390_mode. */
 348             OUT_CB_REG(R400_US_CODE_EXT, 0);
 349         }
 350
 351         OUT_CB_REG_SEQ(R300_US_CODE_ADDR_0, 4);
 352         OUT_CB_TABLE(code->code_addr, 4);
 353
 354         do {
 355             unsigned int bank_alu_length = (alu_length < 64 ? alu_length : 64);
 356             unsigned int bank_alu_offset = bank * 64;
 357             unsigned int bank_tex_length = (tex_length < 32 ? tex_length : 32);
 358             unsigned int bank_tex_offset = bank * 32;
 359
 360             if (r300->screen->caps.is_r400) {
 361                 OUT_CB_REG(R400_US_CODE_BANK, code->r390_mode ?
 362                                 (bank << R400_BANK_SHIFT) | R400_R390_MODE_ENABLE : 0);//2
 363             }
 364
 365             if (bank_alu_length > 0) {
 366                 OUT_CB_REG_SEQ(R300_US_ALU_RGB_INST_0, bank_alu_length);
 367                 for (i = 0; i < bank_alu_length; i++)
 368                     OUT_CB(code->alu.inst[i + bank_alu_offset].rgb_inst);
 369
 370                 OUT_CB_REG_SEQ(R300_US_ALU_RGB_ADDR_0, bank_alu_length);
 371                 for (i = 0; i < bank_alu_length; i++)
 372                     OUT_CB(code->alu.inst[i + bank_alu_offset].rgb_addr);
 373
 374                 OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_INST_0, bank_alu_length);
 375                 for (i = 0; i < bank_alu_length; i++)
 376                     OUT_CB(code->alu.inst[i + bank_alu_offset].alpha_inst);
 377
 378                 OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_ADDR_0, bank_alu_length);
 379                 for (i = 0; i < bank_alu_length; i++)
 380                     OUT_CB(code->alu.inst[i + bank_alu_offset].alpha_addr);
 381
 382                 if (code->r390_mode) {
 383                     OUT_CB_REG_SEQ(R400_US_ALU_EXT_ADDR_0, bank_alu_length);
 384                     for (i = 0; i < bank_alu_length; i++)
 385                         OUT_CB(code->alu.inst[i + bank_alu_offset].r400_ext_addr);
 386                 }
 387             }
 388
 389             if (bank_tex_length > 0) {
 390                 OUT_CB_REG_SEQ(R300_US_TEX_INST_0, bank_tex_length);
 391                 OUT_CB_TABLE(code->tex.inst + bank_tex_offset, bank_tex_length);
 392             }
 393
 394             alu_length -= bank_alu_length;
 395             tex_length -= bank_tex_length;
 396             bank++;
 397         } while(code->r390_mode && (alu_length > 0 || tex_length > 0));
 398
 399         /* R400_US_CODE_BANK needs to be reset to 0, otherwise some shaders
 400          * will be rendered incorrectly. */
 401         if (r300->screen->caps.is_r400) {
 402             OUT_CB_REG(R400_US_CODE_BANK,
 403                 code->r390_mode ? R400_R390_MODE_ENABLE : 0);
 404         }
 405
 406         /* Emit immediates. */
 407         if (imm_count) {
 408             for(i = imm_first; i < imm_end; ++i) {
 409                 if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
 410                     const float *data = constants[i].u.Immediate;
 411
 412                     OUT_CB_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4);
 413                     OUT_CB(pack_float24(data[0]));
 414                     OUT_CB(pack_float24(data[1]));
 415                     OUT_CB(pack_float24(data[2]));
 416                     OUT_CB(pack_float24(data[3]));
 417                 }
 418             }
 419         }
 420     }
 421
 422     OUT_CB_REG(R300_FG_DEPTH_SRC, shader->fg_depth_src);
 423     OUT_CB_REG(R300_US_W_FMT, shader->us_out_w);
 424     END_CB;
 425 }
 426
 427 static void r300_translate_fragment_shader(
 428     struct r300_context* r300,
 429     struct r300_fragment_shader_code* shader,
 430     const struct tgsi_token *tokens)
 431 {
 432     struct r300_fragment_program_compiler compiler;
 433     struct tgsi_to_rc ttr;
 434     int wpos, face;
 435     unsigned i;
 436
 437     tgsi_scan_shader(tokens, &shader->info);
 438     r300_shader_read_fs_inputs(&shader->info, &shader->inputs);
 439
 440     wpos = shader->inputs.wpos;
 441     face = shader->inputs.face;
 442
 443     /* Setup the compiler. */
 444     memset(&compiler, 0, sizeof(compiler));
 445     rc_init(&compiler.Base, &r300->fs_regalloc_state);
 446     DBG_ON(r300, DBG_FP) ? compiler.Base.Debug |= RC_DBG_LOG : 0;
 447     DBG_ON(r300, DBG_P_STAT) ? compiler.Base.Debug |= RC_DBG_STATS : 0;
 448
 449     compiler.code = &shader->code;
 450     compiler.state = shader->compare_state;
 451     compiler.Base.is_r500 = r300->screen->caps.is_r500;
 452     compiler.Base.is_r400 = r300->screen->caps.is_r400;
 453     compiler.Base.disable_optimizations = DBG_ON(r300, DBG_NO_OPT);
 454     compiler.Base.has_half_swizzles = TRUE;
 455     compiler.Base.has_presub = TRUE;
 456     compiler.Base.has_omod = TRUE;
 457     compiler.Base.max_temp_regs =
 458         compiler.Base.is_r500 ? 128 : (compiler.Base.is_r400 ? 64 : 32);
 459     compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;
 460     compiler.Base.max_alu_insts =
 461         (compiler.Base.is_r500 || compiler.Base.is_r400) ? 512 : 64;
 462     compiler.Base.max_tex_insts =
 463         (compiler.Base.is_r500 || compiler.Base.is_r400) ? 512 : 32;
 464     compiler.AllocateHwInputs = &allocate_hardware_inputs;
 465     compiler.UserData = &shader->inputs;
 466
 467     find_output_registers(&compiler, shader);
 468
 469     shader->write_all = FALSE;
 470     for (i = 0; i < shader->info.num_properties; i++) {
 471         if (shader->info.properties[i].name == TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
 472             shader->write_all = TRUE;
 473         }
 474     }
 475
 476     if (compiler.Base.Debug & RC_DBG_LOG) {
 477         DBG(r300, DBG_FP, "r300: Initial fragment program\n");
 478         tgsi_dump(tokens, 0);
 479     }
 480
 481     /* Translate TGSI to our internal representation */
 482     ttr.compiler = &compiler.Base;
 483     ttr.info = &shader->info;
 484     ttr.use_half_swizzles = TRUE;
 485
 486     r300_tgsi_to_rc(&ttr, tokens);
 487
 488     if (ttr.error) {
 489         fprintf(stderr, "r300 FP: Cannot translate a shader. "
 490                 "Using a dummy shader instead.\n");
 491         r300_dummy_fragment_shader(r300, shader);
 492         return;
 493     }
 494
 495     if (!r300->screen->caps.is_r500 ||
 496         compiler.Base.Program.Constants.Count > 200) {
 497         compiler.Base.remove_unused_constants = TRUE;
 498     }
 499
 500     /**
 501      * Transform the program to support WPOS.
 502      *
 503      * Introduce a small fragment at the start of the program that will be
 504      * the only code that directly reads the WPOS input.
 505      * All other code pieces that reference that input will be rewritten
 506      * to read from a newly allocated temporary. */
 507     if (wpos != ATTR_UNUSED) {
 508         /* Moving the input to some other reg is not really necessary. */
 509         rc_transform_fragment_wpos(&compiler.Base, wpos, wpos, TRUE);
 510     }
 511
 512     if (face != ATTR_UNUSED) {
 513         rc_transform_fragment_face(&compiler.Base, face);
 514     }
 515
 516     /* Invoke the compiler */
 517     r3xx_compile_fragment_program(&compiler);
 518
 519     if (compiler.Base.Error) {
 520         fprintf(stderr, "r300 FP: Compiler Error:\n%sUsing a dummy shader"
 521                 " instead.\n", compiler.Base.ErrorMsg);
 522
 523         if (shader->dummy) {
 524             fprintf(stderr, "r300 FP: Cannot compile the dummy shader! "
 525                     "Giving up...\n");
 526             abort();
 527         }
 528
 529         rc_destroy(&compiler.Base);
 530         r300_dummy_fragment_shader(r300, shader);
 531         return;
 532     }
 533
 534     /* Shaders with zero instructions are invalid,
 535      * use the dummy shader instead. */
 536     if (shader->code.code.r500.inst_end == -1) {
 537         rc_destroy(&compiler.Base);
 538         r300_dummy_fragment_shader(r300, shader);
 539         return;
 540     }
 541
 542     /* Initialize numbers of constants for each type. */
 543     shader->externals_count = 0;
 544     for (i = 0;
 545          i < shader->code.constants.Count &&
 546          shader->code.constants.Constants[i].Type == RC_CONSTANT_EXTERNAL; i++) {
 547         shader->externals_count = i+1;
 548     }
 549     shader->immediates_count = 0;
 550     shader->rc_state_count = 0;
 551
 552     for (i = shader->externals_count; i < shader->code.constants.Count; i++) {
 553         switch (shader->code.constants.Constants[i].Type) {
 554             case RC_CONSTANT_IMMEDIATE:
 555                 ++shader->immediates_count;
 556                 break;
 557             case RC_CONSTANT_STATE:
 558                 ++shader->rc_state_count;
 559                 break;
 560             default:
 561                 assert(0);
 562         }
 563     }
 564
 565     /* Setup shader depth output. */
 566     if (shader->code.writes_depth) {
 567         shader->fg_depth_src = R300_FG_DEPTH_SRC_SHADER;
 568         shader->us_out_w = R300_W_FMT_W24 | R300_W_SRC_US;
 569     } else {
 570         shader->fg_depth_src = R300_FG_DEPTH_SRC_SCAN;
 571         shader->us_out_w = R300_W_FMT_W0 | R300_W_SRC_US;
 572     }
 573
 574     /* And, finally... */
 575     rc_destroy(&compiler.Base);
 576
 577     /* Build the command buffer. */
 578     r300_emit_fs_code_to_buffer(r300, shader);
 579 }
 580
 581 boolean r300_pick_fragment_shader(struct r300_context* r300)
 582 {
 583     struct r300_fragment_shader* fs = r300_fs(r300);
 584     struct r300_fragment_program_external_state state = {{{ 0 }}};
 585     struct r300_fragment_shader_code* ptr;
 586
 587     get_external_state(r300, &state);
 588
 589     if (!fs->first) {
 590         /* Build the fragment shader for the first time. */
 591         fs->first = fs->shader = CALLOC_STRUCT(r300_fragment_shader_code);
 592
 593         memcpy(&fs->shader->compare_state, &state,
 594             sizeof(struct r300_fragment_program_external_state));
 595         r300_translate_fragment_shader(r300, fs->shader, fs->state.tokens);
 596         return TRUE;
 597
 598     } else {
 599         /* Check if the currently-bound shader has been compiled
 600          * with the texture-compare state we need. */
 601         if (memcmp(&fs->shader->compare_state, &state, sizeof(state)) != 0) {
 602             /* Search for the right shader. */
 603             ptr = fs->first;
 604             while (ptr) {
 605                 if (memcmp(&ptr->compare_state, &state, sizeof(state)) == 0) {
 606                     if (fs->shader != ptr) {
 607                         fs->shader = ptr;
 608                         return TRUE;
 609                     }
 610                     /* The currently-bound one is OK. */
 611                     return FALSE;
 612                 }
 613                 ptr = ptr->next;
 614             }
 615
 616             /* Not found, gotta compile a new one. */
 617             ptr = CALLOC_STRUCT(r300_fragment_shader_code);
 618             ptr->next = fs->first;
 619             fs->first = fs->shader = ptr;
 620
 621             ptr->compare_state = state;
 622             r300_translate_fragment_shader(r300, ptr, fs->state.tokens);
 623             return TRUE;
 624         }
 625     }
 626
 627     return FALSE;
 628 }