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     struct r300_rs_state *rs = r300->rs_state.state;
 151     unsigned i;
 152
 153     state->frag_clamp = rs ? rs->rs.clamp_fragment_color : 0;
 154
 155     for (i = 0; i < texstate->sampler_state_count; i++) {
 156         struct r300_sampler_state *s = texstate->sampler_states[i];
 157         struct r300_sampler_view *v = texstate->sampler_views[i];
 158         struct r300_resource *t;
 159
 160         if (!s || !v) {
 161             continue;
 162         }
 163
 164         t = r300_resource(v->base.texture);
 165
 166         if (s->state.compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
 167             state->unit[i].compare_mode_enabled = 1;
 168
 169             /* Fortunately, no need to translate this. */
 170             state->unit[i].texture_compare_func = s->state.compare_func;
 171         }
 172
 173         state->unit[i].non_normalized_coords = !s->state.normalized_coords;
 174         state->unit[i].convert_unorm_to_snorm =
 175                 v->base.format == PIPE_FORMAT_RGTC1_SNORM ||
 176                 v->base.format == PIPE_FORMAT_LATC1_SNORM;
 177
 178         /* Pass texture swizzling to the compiler, some lowering passes need it. */
 179         if (v->base.format == PIPE_FORMAT_RGTC1_SNORM ||
 180             v->base.format == PIPE_FORMAT_LATC1_SNORM) {
 181             unsigned char swizzle[4];
 182
 183             util_format_compose_swizzles(
 184                             util_format_description(v->base.format)->swizzle,
 185                             v->swizzle,
 186                             swizzle);
 187
 188             state->unit[i].texture_swizzle =
 189                     RC_MAKE_SWIZZLE(swizzle[0], swizzle[1],
 190                                     swizzle[2], swizzle[3]);
 191         } else if (state->unit[i].compare_mode_enabled) {
 192             state->unit[i].texture_swizzle =
 193                 RC_MAKE_SWIZZLE(v->swizzle[0], v->swizzle[1],
 194                                 v->swizzle[2], v->swizzle[3]);
 195         }
 196
 197         /* XXX this should probably take into account STR, not just S. */
 198         if (t->tex.is_npot) {
 199             switch (s->state.wrap_s) {
 200             case PIPE_TEX_WRAP_REPEAT:
 201                 state->unit[i].wrap_mode = RC_WRAP_REPEAT;
 202                 break;
 203
 204             case PIPE_TEX_WRAP_MIRROR_REPEAT:
 205                 state->unit[i].wrap_mode = RC_WRAP_MIRRORED_REPEAT;
 206                 break;
 207
 208             case PIPE_TEX_WRAP_MIRROR_CLAMP:
 209             case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 210             case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 211                 state->unit[i].wrap_mode = RC_WRAP_MIRRORED_CLAMP;
 212                 break;
 213
 214             default:
 215                 state->unit[i].wrap_mode = RC_WRAP_NONE;
 216             }
 217
 218             if (t->b.b.b.target == PIPE_TEXTURE_3D)
 219                 state->unit[i].clamp_and_scale_before_fetch = TRUE;
 220         }
 221     }
 222 }
 223
 224 static void r300_translate_fragment_shader(
 225     struct r300_context* r300,
 226     struct r300_fragment_shader_code* shader,
 227     const struct tgsi_token *tokens);
 228
 229 static void r300_dummy_fragment_shader(
 230     struct r300_context* r300,
 231     struct r300_fragment_shader_code* shader)
 232 {
 233     struct pipe_shader_state state;
 234     struct ureg_program *ureg;
 235     struct ureg_dst out;
 236     struct ureg_src imm;
 237
 238     /* Make a simple fragment shader which outputs (0, 0, 0, 1) */
 239     ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 240     out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 241     imm = ureg_imm4f(ureg, 0, 0, 0, 1);
 242
 243     ureg_MOV(ureg, out, imm);
 244     ureg_END(ureg);
 245
 246     state.tokens = ureg_finalize(ureg);
 247
 248     shader->dummy = TRUE;
 249     r300_translate_fragment_shader(r300, shader, state.tokens);
 250
 251     ureg_destroy(ureg);
 252 }
 253
 254 static void r300_emit_fs_code_to_buffer(
 255     struct r300_context *r300,
 256     struct r300_fragment_shader_code *shader)
 257 {
 258     struct rX00_fragment_program_code *generic_code = &shader->code;
 259     unsigned imm_count = shader->immediates_count;
 260     unsigned imm_first = shader->externals_count;
 261     unsigned imm_end = generic_code->constants.Count;
 262     struct rc_constant *constants = generic_code->constants.Constants;
 263     unsigned i;
 264     CB_LOCALS;
 265
 266     if (r300->screen->caps.is_r500) {
 267         struct r500_fragment_program_code *code = &generic_code->code.r500;
 268
 269         shader->cb_code_size = 19 +
 270                                ((code->inst_end + 1) * 6) +
 271                                imm_count * 7 +
 272                                code->int_constant_count * 2;
 273
 274         NEW_CB(shader->cb_code, shader->cb_code_size);
 275         OUT_CB_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO);
 276         OUT_CB_REG(R500_US_PIXSIZE, code->max_temp_idx);
 277         OUT_CB_REG(R500_US_FC_CTRL, code->us_fc_ctrl);
 278         for(i = 0; i < code->int_constant_count; i++){
 279                 OUT_CB_REG(R500_US_FC_INT_CONST_0 + (i * 4),
 280                                                 code->int_constants[i]);
 281         }
 282         OUT_CB_REG(R500_US_CODE_RANGE,
 283                    R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(code->inst_end));
 284         OUT_CB_REG(R500_US_CODE_OFFSET, 0);
 285         OUT_CB_REG(R500_US_CODE_ADDR,
 286                    R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(code->inst_end));
 287
 288         OUT_CB_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_INSTR);
 289         OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, (code->inst_end + 1) * 6);
 290         for (i = 0; i <= code->inst_end; i++) {
 291             OUT_CB(code->inst[i].inst0);
 292             OUT_CB(code->inst[i].inst1);
 293             OUT_CB(code->inst[i].inst2);
 294             OUT_CB(code->inst[i].inst3);
 295             OUT_CB(code->inst[i].inst4);
 296             OUT_CB(code->inst[i].inst5);
 297         }
 298
 299         /* Emit immediates. */
 300         if (imm_count) {
 301             for(i = imm_first; i < imm_end; ++i) {
 302                 if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
 303                     const float *data = constants[i].u.Immediate;
 304
 305                     OUT_CB_REG(R500_GA_US_VECTOR_INDEX,
 306                                R500_GA_US_VECTOR_INDEX_TYPE_CONST |
 307                                (i & R500_GA_US_VECTOR_INDEX_MASK));
 308                     OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, 4);
 309                     OUT_CB_TABLE(data, 4);
 310                 }
 311             }
 312         }
 313     } else { /* r300 */
 314         struct r300_fragment_program_code *code = &generic_code->code.r300;
 315         unsigned int alu_length = code->alu.length;
 316         unsigned int alu_iterations = ((alu_length - 1) / 64) + 1;
 317         unsigned int tex_length = code->tex.length;
 318         unsigned int tex_iterations =
 319             tex_length > 0 ? ((tex_length - 1) / 32) + 1 : 0;
 320         unsigned int iterations =
 321             alu_iterations > tex_iterations ? alu_iterations : tex_iterations;
 322         unsigned int bank = 0;
 323
 324         shader->cb_code_size = 15 +
 325             /* R400_US_CODE_BANK */
 326             (r300->screen->caps.is_r400 ? 2 * (iterations + 1): 0) +
 327             /* R400_US_CODE_EXT */
 328             (r300->screen->caps.is_r400 ? 2 : 0) +
 329             /* R300_US_ALU_{RGB,ALPHA}_{INST,ADDR}_0, R400_US_ALU_EXT_ADDR_0 */
 330             (code->r390_mode ? (5 * alu_iterations) : 4) +
 331             /* R400_US_ALU_EXT_ADDR_[0-63] */
 332             (code->r390_mode ? (code->alu.length) : 0) +
 333             /* R300_US_ALU_{RGB,ALPHA}_{INST,ADDR}_0 */
 334             code->alu.length * 4 +
 335             /* R300_US_TEX_INST_0, R300_US_TEX_INST_[0-31] */
 336             (code->tex.length > 0 ? code->tex.length + tex_iterations : 0) +
 337             imm_count * 5;
 338
 339         NEW_CB(shader->cb_code, shader->cb_code_size);
 340
 341         OUT_CB_REG(R300_US_CONFIG, code->config);
 342         OUT_CB_REG(R300_US_PIXSIZE, code->pixsize);
 343         OUT_CB_REG(R300_US_CODE_OFFSET, code->code_offset);
 344
 345         if (code->r390_mode) {
 346             OUT_CB_REG(R400_US_CODE_EXT, code->r400_code_offset_ext);
 347         } else if (r300->screen->caps.is_r400) {
 348             /* This register appears to affect shaders even if r390_mode is
 349              * disabled, so it needs to be set to 0 for shaders that
 350              * don't use r390_mode. */
 351             OUT_CB_REG(R400_US_CODE_EXT, 0);
 352         }
 353
 354         OUT_CB_REG_SEQ(R300_US_CODE_ADDR_0, 4);
 355         OUT_CB_TABLE(code->code_addr, 4);
 356
 357         do {
 358             unsigned int bank_alu_length = (alu_length < 64 ? alu_length : 64);
 359             unsigned int bank_alu_offset = bank * 64;
 360             unsigned int bank_tex_length = (tex_length < 32 ? tex_length : 32);
 361             unsigned int bank_tex_offset = bank * 32;
 362
 363             if (r300->screen->caps.is_r400) {
 364                 OUT_CB_REG(R400_US_CODE_BANK, code->r390_mode ?
 365                                 (bank << R400_BANK_SHIFT) | R400_R390_MODE_ENABLE : 0);//2
 366             }
 367
 368             if (bank_alu_length > 0) {
 369                 OUT_CB_REG_SEQ(R300_US_ALU_RGB_INST_0, bank_alu_length);
 370                 for (i = 0; i < bank_alu_length; i++)
 371                     OUT_CB(code->alu.inst[i + bank_alu_offset].rgb_inst);
 372
 373                 OUT_CB_REG_SEQ(R300_US_ALU_RGB_ADDR_0, bank_alu_length);
 374                 for (i = 0; i < bank_alu_length; i++)
 375                     OUT_CB(code->alu.inst[i + bank_alu_offset].rgb_addr);
 376
 377                 OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_INST_0, bank_alu_length);
 378                 for (i = 0; i < bank_alu_length; i++)
 379                     OUT_CB(code->alu.inst[i + bank_alu_offset].alpha_inst);
 380
 381                 OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_ADDR_0, bank_alu_length);
 382                 for (i = 0; i < bank_alu_length; i++)
 383                     OUT_CB(code->alu.inst[i + bank_alu_offset].alpha_addr);
 384
 385                 if (code->r390_mode) {
 386                     OUT_CB_REG_SEQ(R400_US_ALU_EXT_ADDR_0, bank_alu_length);
 387                     for (i = 0; i < bank_alu_length; i++)
 388                         OUT_CB(code->alu.inst[i + bank_alu_offset].r400_ext_addr);
 389                 }
 390             }
 391
 392             if (bank_tex_length > 0) {
 393                 OUT_CB_REG_SEQ(R300_US_TEX_INST_0, bank_tex_length);
 394                 OUT_CB_TABLE(code->tex.inst + bank_tex_offset, bank_tex_length);
 395             }
 396
 397             alu_length -= bank_alu_length;
 398             tex_length -= bank_tex_length;
 399             bank++;
 400         } while(code->r390_mode && (alu_length > 0 || tex_length > 0));
 401
 402         /* R400_US_CODE_BANK needs to be reset to 0, otherwise some shaders
 403          * will be rendered incorrectly. */
 404         if (r300->screen->caps.is_r400) {
 405             OUT_CB_REG(R400_US_CODE_BANK,
 406                 code->r390_mode ? R400_R390_MODE_ENABLE : 0);
 407         }
 408
 409         /* Emit immediates. */
 410         if (imm_count) {
 411             for(i = imm_first; i < imm_end; ++i) {
 412                 if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
 413                     const float *data = constants[i].u.Immediate;
 414
 415                     OUT_CB_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4);
 416                     OUT_CB(pack_float24(data[0]));
 417                     OUT_CB(pack_float24(data[1]));
 418                     OUT_CB(pack_float24(data[2]));
 419                     OUT_CB(pack_float24(data[3]));
 420                 }
 421             }
 422         }
 423     }
 424
 425     OUT_CB_REG(R300_FG_DEPTH_SRC, shader->fg_depth_src);
 426     OUT_CB_REG(R300_US_W_FMT, shader->us_out_w);
 427     END_CB;
 428 }
 429
 430 static void r300_translate_fragment_shader(
 431     struct r300_context* r300,
 432     struct r300_fragment_shader_code* shader,
 433     const struct tgsi_token *tokens)
 434 {
 435     struct r300_fragment_program_compiler compiler;
 436     struct tgsi_to_rc ttr;
 437     int wpos, face;
 438     unsigned i;
 439
 440     tgsi_scan_shader(tokens, &shader->info);
 441     r300_shader_read_fs_inputs(&shader->info, &shader->inputs);
 442
 443     wpos = shader->inputs.wpos;
 444     face = shader->inputs.face;
 445
 446     /* Setup the compiler. */
 447     memset(&compiler, 0, sizeof(compiler));
 448     rc_init(&compiler.Base);
 449     DBG_ON(r300, DBG_FP) ? compiler.Base.Debug |= RC_DBG_LOG : 0;
 450     DBG_ON(r300, DBG_P_STAT) ? compiler.Base.Debug |= RC_DBG_STATS : 0;
 451
 452     compiler.code = &shader->code;
 453     compiler.state = shader->compare_state;
 454     compiler.Base.is_r500 = r300->screen->caps.is_r500;
 455     compiler.Base.is_r400 = r300->screen->caps.is_r400;
 456     compiler.Base.disable_optimizations = DBG_ON(r300, DBG_NO_OPT);
 457     compiler.Base.has_half_swizzles = TRUE;
 458     compiler.Base.has_presub = TRUE;
 459     compiler.Base.has_omod = TRUE;
 460     compiler.Base.max_temp_regs =
 461         compiler.Base.is_r500 ? 128 : (compiler.Base.is_r400 ? 64 : 32);
 462     compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;
 463     compiler.Base.max_alu_insts =
 464         (compiler.Base.is_r500 || compiler.Base.is_r400) ? 512 : 64;
 465     compiler.Base.max_tex_insts =
 466         (compiler.Base.is_r500 || compiler.Base.is_r400) ? 512 : 32;
 467     compiler.AllocateHwInputs = &allocate_hardware_inputs;
 468     compiler.UserData = &shader->inputs;
 469
 470     find_output_registers(&compiler, shader);
 471
 472     shader->write_all = FALSE;
 473     for (i = 0; i < shader->info.num_properties; i++) {
 474         if (shader->info.properties[i].name == TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
 475             shader->write_all = TRUE;
 476         }
 477     }
 478
 479     if (compiler.Base.Debug & RC_DBG_LOG) {
 480         DBG(r300, DBG_FP, "r300: Initial fragment program\n");
 481         tgsi_dump(tokens, 0);
 482     }
 483
 484     /* Translate TGSI to our internal representation */
 485     ttr.compiler = &compiler.Base;
 486     ttr.info = &shader->info;
 487     ttr.use_half_swizzles = TRUE;
 488
 489     r300_tgsi_to_rc(&ttr, tokens);
 490
 491     if (ttr.error) {
 492         fprintf(stderr, "r300 FP: Cannot translate a shader. "
 493                 "Using a dummy shader instead.\n");
 494         r300_dummy_fragment_shader(r300, shader);
 495         return;
 496     }
 497
 498     if (!r300->screen->caps.is_r500 ||
 499         compiler.Base.Program.Constants.Count > 200) {
 500         compiler.Base.remove_unused_constants = TRUE;
 501     }
 502
 503     /**
 504      * Transform the program to support WPOS.
 505      *
 506      * Introduce a small fragment at the start of the program that will be
 507      * the only code that directly reads the WPOS input.
 508      * All other code pieces that reference that input will be rewritten
 509      * to read from a newly allocated temporary. */
 510     if (wpos != ATTR_UNUSED) {
 511         /* Moving the input to some other reg is not really necessary. */
 512         rc_transform_fragment_wpos(&compiler.Base, wpos, wpos, TRUE);
 513     }
 514
 515     if (face != ATTR_UNUSED) {
 516         rc_transform_fragment_face(&compiler.Base, face);
 517     }
 518
 519     /* Invoke the compiler */
 520     r3xx_compile_fragment_program(&compiler);
 521
 522     if (compiler.Base.Error) {
 523         fprintf(stderr, "r300 FP: Compiler Error:\n%sUsing a dummy shader"
 524                 " instead.\n", compiler.Base.ErrorMsg);
 525
 526         if (shader->dummy) {
 527             fprintf(stderr, "r300 FP: Cannot compile the dummy shader! "
 528                     "Giving up...\n");
 529             abort();
 530         }
 531
 532         rc_destroy(&compiler.Base);
 533         r300_dummy_fragment_shader(r300, shader);
 534         return;
 535     }
 536
 537     /* Shaders with zero instructions are invalid,
 538      * use the dummy shader instead. */
 539     if (shader->code.code.r500.inst_end == -1) {
 540         rc_destroy(&compiler.Base);
 541         r300_dummy_fragment_shader(r300, shader);
 542         return;
 543     }
 544
 545     /* Initialize numbers of constants for each type. */
 546     shader->externals_count = 0;
 547     for (i = 0;
 548          i < shader->code.constants.Count &&
 549          shader->code.constants.Constants[i].Type == RC_CONSTANT_EXTERNAL; i++) {
 550         shader->externals_count = i+1;
 551     }
 552     shader->immediates_count = 0;
 553     shader->rc_state_count = 0;
 554
 555     for (i = shader->externals_count; i < shader->code.constants.Count; i++) {
 556         switch (shader->code.constants.Constants[i].Type) {
 557             case RC_CONSTANT_IMMEDIATE:
 558                 ++shader->immediates_count;
 559                 break;
 560             case RC_CONSTANT_STATE:
 561                 ++shader->rc_state_count;
 562                 break;
 563             default:
 564                 assert(0);
 565         }
 566     }
 567
 568     /* Setup shader depth output. */
 569     if (shader->code.writes_depth) {
 570         shader->fg_depth_src = R300_FG_DEPTH_SRC_SHADER;
 571         shader->us_out_w = R300_W_FMT_W24 | R300_W_SRC_US;
 572     } else {
 573         shader->fg_depth_src = R300_FG_DEPTH_SRC_SCAN;
 574         shader->us_out_w = R300_W_FMT_W0 | R300_W_SRC_US;
 575     }
 576
 577     /* And, finally... */
 578     rc_destroy(&compiler.Base);
 579
 580     /* Build the command buffer. */
 581     r300_emit_fs_code_to_buffer(r300, shader);
 582 }
 583
 584 boolean r300_pick_fragment_shader(struct r300_context* r300)
 585 {
 586     struct r300_fragment_shader* fs = r300_fs(r300);
 587     struct r300_fragment_program_external_state state = {{{ 0 }}};
 588     struct r300_fragment_shader_code* ptr;
 589
 590     get_external_state(r300, &state);
 591
 592     if (!fs->first) {
 593         /* Build the fragment shader for the first time. */
 594         fs->first = fs->shader = CALLOC_STRUCT(r300_fragment_shader_code);
 595
 596         memcpy(&fs->shader->compare_state, &state,
 597             sizeof(struct r300_fragment_program_external_state));
 598         r300_translate_fragment_shader(r300, fs->shader, fs->state.tokens);
 599         return TRUE;
 600
 601     } else {
 602         /* Check if the currently-bound shader has been compiled
 603          * with the texture-compare state we need. */
 604         if (memcmp(&fs->shader->compare_state, &state, sizeof(state)) != 0) {
 605             /* Search for the right shader. */
 606             ptr = fs->first;
 607             while (ptr) {
 608                 if (memcmp(&ptr->compare_state, &state, sizeof(state)) == 0) {
 609                     if (fs->shader != ptr) {
 610                         fs->shader = ptr;
 611                         return TRUE;
 612                     }
 613                     /* The currently-bound one is OK. */
 614                     return FALSE;
 615                 }
 616                 ptr = ptr->next;
 617             }
 618
 619             /* Not found, gotta compile a new one. */
 620             ptr = CALLOC_STRUCT(r300_fragment_shader_code);
 621             ptr->next = fs->first;
 622             fs->first = fs->shader = ptr;
 623
 624             ptr->compare_state = state;
 625             r300_translate_fragment_shader(r300, ptr, fs->state.tokens);
 626             return TRUE;
 627         }
 628     }
 629
 630     return FALSE;
 631 }