git.oblomov.eu Git - wine/blob - dlls/wined3d/baseshader.c

   1 /*
   2  * shaders implementation
   3  *
   4  * Copyright 2002-2003 Jason Edmeades
   5  * Copyright 2002-2003 Raphael Junqueira
   6  * Copyright 2005 Oliver Stieber
   7  * Copyright 2006 Ivan Gyurdiev
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  22  */
  23
  24 #include "config.h"
  25 #include <string.h>
  26 #include <stdio.h>
  27 #include "wined3d_private.h"
  28
  29 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  30
  31 #define GLNAME_REQUIRE_GLSL  ((const char *)1)
  32
  33 inline static BOOL shader_is_version_token(DWORD token) {
  34     return shader_is_pshader_version(token) ||
  35            shader_is_vshader_version(token);
  36 }
  37
  38 int shader_addline(
  39     SHADER_BUFFER* buffer,
  40     const char *format, ...) {
  41
  42     char* base = buffer->buffer + buffer->bsize;
  43     int rc;
  44
  45     va_list args;
  46     va_start(args, format);
  47     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
  48     va_end(args);
  49
  50     if (rc < 0 ||                                   /* C89 */
  51         rc > SHADER_PGMSIZE - 1 - buffer->bsize) {  /* C99 */
  52
  53         ERR("The buffer allocated for the shader program string "
  54             "is too small at %d bytes.\n", SHADER_PGMSIZE);
  55         buffer->bsize = SHADER_PGMSIZE - 1;
  56         return -1;
  57     }
  58
  59     buffer->bsize += rc;
  60     buffer->lineNo++;
  61     TRACE("GL HW (%u, %u) : %s", buffer->lineNo, buffer->bsize, base);
  62     return 0;
  63 }
  64
  65 const SHADER_OPCODE* shader_get_opcode(
  66     IWineD3DBaseShader *iface, const DWORD code) {
  67
  68     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl*) iface;
  69
  70     DWORD i = 0;
  71     DWORD hex_version = This->baseShader.hex_version;
  72     const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
  73
  74     /** TODO: use dichotomic search */
  75     while (NULL != shader_ins[i].name) {
  76         if (((code & D3DSI_OPCODE_MASK) == shader_ins[i].opcode) &&
  77             (((hex_version >= shader_ins[i].min_version) && (hex_version <= shader_ins[i].max_version)) ||
  78             ((shader_ins[i].min_version == 0) && (shader_ins[i].max_version == 0)))) {
  79             return &shader_ins[i];
  80         }
  81         ++i;
  82     }
  83     FIXME("Unsupported opcode %#lx(%ld) masked %#lx, shader version %#lx\n",
  84        code, code, code & D3DSI_OPCODE_MASK, hex_version);
  85     return NULL;
  86 }
  87
  88 /* Read a parameter opcode from the input stream,
  89  * and possibly a relative addressing token.
  90  * Return the number of tokens read */
  91 int shader_get_param(
  92     IWineD3DBaseShader* iface,
  93     const DWORD* pToken,
  94     DWORD* param,
  95     DWORD* addr_token) {
  96
  97     /* PS >= 3.0 have relative addressing (with token)
  98      * VS >= 2.0 have relative addressing (with token)
  99      * VS >= 1.0 < 2.0 have relative addressing (without token)
 100      * The version check below should work in general */
 101
 102     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 103     char rel_token = D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
 104         ((*pToken & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
 105
 106     *param = *pToken;
 107     *addr_token = rel_token? *(pToken + 1): 0;
 108     return rel_token? 2:1;
 109 }
 110
 111 /* Return the number of parameters to skip for an opcode */
 112 static inline int shader_skip_opcode(
 113     IWineD3DBaseShaderImpl* This,
 114     const SHADER_OPCODE* curOpcode,
 115     DWORD opcode_token) {
 116
 117    /* Shaders >= 2.0 may contain address tokens, but fortunately they
 118     * have a useful legnth mask - use it here. Shaders 1.0 contain no such tokens */
 119
 120     return (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
 121         ((opcode_token & D3DSI_INSTLENGTH_MASK) >> D3DSI_INSTLENGTH_SHIFT):
 122         curOpcode->num_params;
 123 }
 124
 125 /* Read the parameters of an unrecognized opcode from the input stream
 126  * Return the number of tokens read.
 127  *
 128  * Note: This function assumes source or destination token format.
 129  * It will not work with specially-formatted tokens like DEF or DCL,
 130  * but hopefully those would be recognized */
 131
 132 int shader_skip_unrecognized(
 133     IWineD3DBaseShader* iface,
 134     const DWORD* pToken) {
 135
 136     int tokens_read = 0;
 137     int i = 0;
 138
 139     /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
 140     while (*pToken & 0x80000000) {
 141
 142         DWORD param, addr_token;
 143         tokens_read += shader_get_param(iface, pToken, &param, &addr_token);
 144         pToken += tokens_read;
 145
 146         FIXME("Unrecognized opcode param: token=%08lX "
 147             "addr_token=%08lX name=", param, addr_token);
 148         shader_dump_param(iface, param, addr_token, i);
 149         FIXME("\n");
 150         ++i;
 151     }
 152     return tokens_read;
 153 }
 154
 155 /* Convert floating point offset relative
 156  * to a register file to an absolute offset for float constants */
 157
 158 unsigned int shader_get_float_offset(const DWORD reg) {
 159
 160      unsigned int regnum = reg & D3DSP_REGNUM_MASK;
 161      int regtype = shader_get_regtype(reg);
 162
 163      switch (regtype) {
 164         case D3DSPR_CONST: return regnum;
 165         case D3DSPR_CONST2: return 2048 + regnum;
 166         case D3DSPR_CONST3: return 4096 + regnum;
 167         case D3DSPR_CONST4: return 6144 + regnum;
 168         default:
 169             FIXME("Unsupported register type: %d\n", regtype);
 170             return regnum;
 171      }
 172 }
 173
 174 static void shader_parse_decl_usage(
 175     DWORD *semantics_map,
 176     DWORD usage_token, DWORD param) {
 177
 178     unsigned int usage = (usage_token & D3DSP_DCL_USAGE_MASK) >> D3DSP_DCL_USAGE_SHIFT;
 179     unsigned int usage_idx = (usage_token & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
 180     unsigned int regnum = param & D3DSP_REGNUM_MASK;
 181
 182     switch(usage) {
 183         case D3DDECLUSAGE_POSITION:
 184             if (usage_idx == 0) { /* tween data */
 185                 TRACE("Setting position to %d\n", regnum);
 186                 semantics_map[WINED3DSHADERDECLUSAGE_POSITION] = param;
 187             } else {
 188                 /* TODO: position indexes go from 0-8!!*/
 189                 TRACE("Setting position 2 to %d because usage_idx = %d\n", regnum, usage_idx);
 190                 /* robots uses positions up to 8, the position arrays are just packed.*/
 191                 if (usage_idx > 1) {
 192                     TRACE("Loaded for position %d (greater than 2)\n", usage_idx);
 193                 }
 194                 semantics_map[WINED3DSHADERDECLUSAGE_POSITION2 + usage_idx-1] = param;
 195             }
 196             break;
 197
 198         case D3DDECLUSAGE_BLENDINDICES:
 199             TRACE("Setting BLENDINDICES to %d\n", regnum);
 200             semantics_map[WINED3DSHADERDECLUSAGE_BLENDINDICES] = param;
 201             if (usage_idx != 0) FIXME("Extended BLENDINDICES\n");
 202             break;
 203
 204         case D3DDECLUSAGE_BLENDWEIGHT:
 205             TRACE("Setting BLENDWEIGHT to %d\n", regnum);
 206             semantics_map[WINED3DSHADERDECLUSAGE_BLENDWEIGHT] = param;
 207             if (usage_idx != 0) FIXME("Extended blend weights\n");
 208             break;
 209
 210         case D3DDECLUSAGE_NORMAL:
 211             if (usage_idx == 0) { /* tween data */
 212                 TRACE("Setting normal to %d\n", regnum);
 213                 semantics_map[WINED3DSHADERDECLUSAGE_NORMAL] = param;
 214             } else {
 215                 TRACE("Setting normal 2 to %d because usage = %d\n", regnum, usage_idx);
 216                 semantics_map[WINED3DSHADERDECLUSAGE_NORMAL2] = param;
 217             }
 218             break;
 219
 220         case D3DDECLUSAGE_PSIZE:
 221             TRACE("Setting PSIZE to %d\n", regnum);
 222             semantics_map[WINED3DSHADERDECLUSAGE_PSIZE] = param;
 223             if (usage_idx != 0) FIXME("Extended PSIZE\n");
 224             break;
 225
 226         case D3DDECLUSAGE_COLOR:
 227             if (usage_idx == 0)  {
 228                 TRACE("Setting DIFFUSE to %d\n", regnum);
 229                 semantics_map[WINED3DSHADERDECLUSAGE_DIFFUSE] = param;
 230             } else {
 231                 TRACE("Setting SPECULAR to %d\n", regnum);
 232                 semantics_map[WINED3DSHADERDECLUSAGE_SPECULAR] = param;
 233             }
 234             break;
 235
 236         case D3DDECLUSAGE_TEXCOORD:
 237             if (usage_idx > 7) {
 238                 FIXME("Program uses texture coordinate %d but only 0-7 have been "
 239                     "implemented\n", usage_idx);
 240             } else {
 241                 TRACE("Setting TEXCOORD %d  to %d\n", usage_idx, regnum);
 242                 semantics_map[WINED3DSHADERDECLUSAGE_TEXCOORD0 + usage_idx] = param;
 243             }
 244             break;
 245
 246         case D3DDECLUSAGE_TANGENT:
 247             TRACE("Setting TANGENT to %d\n", regnum);
 248             semantics_map[WINED3DSHADERDECLUSAGE_TANGENT] = param;
 249             break;
 250
 251         case D3DDECLUSAGE_BINORMAL:
 252             TRACE("Setting BINORMAL to %d\n", regnum);
 253             semantics_map[WINED3DSHADERDECLUSAGE_BINORMAL] = param;
 254             break;
 255
 256         case D3DDECLUSAGE_TESSFACTOR:
 257             TRACE("Setting TESSFACTOR to %d\n", regnum);
 258             semantics_map[WINED3DSHADERDECLUSAGE_TESSFACTOR] = param;
 259             break;
 260
 261         case D3DDECLUSAGE_POSITIONT:
 262             if (usage_idx == 0) { /* tween data */
 263                 FIXME("Setting positiont to %d\n", regnum);
 264                 semantics_map[WINED3DSHADERDECLUSAGE_POSITIONT] = param;
 265             } else {
 266                 FIXME("Setting positiont 2 to %d because usage = %d\n", regnum, usage_idx);
 267                 semantics_map[WINED3DSHADERDECLUSAGE_POSITIONT2] = param;
 268                 if (usage_idx != 0) FIXME("Extended positiont\n");
 269             }
 270             break;
 271
 272         case D3DDECLUSAGE_FOG:
 273             TRACE("Setting FOG to %d\n", regnum);
 274             semantics_map[WINED3DSHADERDECLUSAGE_FOG] = param;
 275             break;
 276
 277         case D3DDECLUSAGE_DEPTH:
 278             TRACE("Setting DEPTH to %d\n", regnum);
 279             semantics_map[WINED3DSHADERDECLUSAGE_DEPTH] = param;
 280             break;
 281
 282         case D3DDECLUSAGE_SAMPLE:
 283             TRACE("Setting SAMPLE to %d\n", regnum);
 284             semantics_map[WINED3DSHADERDECLUSAGE_SAMPLE] = param;
 285             break;
 286
 287         default:
 288             FIXME("Unrecognised dcl %#x", usage);
 289     }
 290 }
 291
 292 /* Note that this does not count the loop register
 293  * as an address register. */
 294
 295 void shader_get_registers_used(
 296     IWineD3DBaseShader *iface,
 297     shader_reg_maps* reg_maps,
 298     CONST DWORD* pToken) {
 299
 300     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 301
 302     /* There are some minor differences between pixel and vertex shaders */
 303     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 304
 305     if (pToken == NULL)
 306         return;
 307
 308     while (D3DVS_END() != *pToken) {
 309         CONST SHADER_OPCODE* curOpcode;
 310         DWORD opcode_token;
 311
 312         /* Skip version */
 313         if (shader_is_version_token(*pToken)) {
 314              ++pToken;
 315              continue;
 316
 317         /* Skip comments */
 318         } else if (shader_is_comment(*pToken)) {
 319              DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 320              ++pToken;
 321              pToken += comment_len;
 322              continue;
 323         }
 324
 325         /* Fetch opcode */
 326         opcode_token = *pToken++;
 327         curOpcode = shader_get_opcode(iface, opcode_token);
 328
 329         /* Unhandled opcode, and its parameters */
 330         if (NULL == curOpcode) {
 331            while (*pToken & 0x80000000)
 332                ++pToken;
 333
 334         /* Handle declarations */
 335         } else if (D3DSIO_DCL == curOpcode->opcode) {
 336
 337             DWORD usage = *pToken++;
 338             DWORD param = *pToken++;
 339             DWORD regtype = shader_get_regtype(param);
 340             unsigned int regnum = param & D3DSP_REGNUM_MASK;
 341
 342             /* Vshader: mark attributes used
 343                Pshader: mark 3.0 input registers used, save token */
 344             if (D3DSPR_INPUT == regtype) {
 345
 346                 if (!pshader)
 347                     reg_maps->attributes[regnum] = 1;
 348                 else
 349                     reg_maps->packed_input[regnum] = 1;
 350
 351                 shader_parse_decl_usage(reg_maps->semantics_in, usage, param);
 352
 353             /* Vshader: mark 3.0 output registers used, save token */
 354             } else if (D3DSPR_OUTPUT == regtype) {
 355                 reg_maps->packed_output[regnum] = 1;
 356                 shader_parse_decl_usage(reg_maps->semantics_out, usage, param);
 357
 358             /* Save sampler usage token */
 359             } else if (D3DSPR_SAMPLER == regtype)
 360                 reg_maps->samplers[regnum] = usage;
 361
 362         /* Skip definitions (for now) */
 363         } else if (D3DSIO_DEF == curOpcode->opcode) {
 364             pToken += curOpcode->num_params;
 365
 366         /* If there's a loop in the shader */
 367         } else if (D3DSIO_LOOP == curOpcode->opcode) {
 368             reg_maps->loop = 1;
 369             pToken += curOpcode->num_params;
 370
 371         /* Set texture, address, temporary registers */
 372         } else {
 373             int i, limit;
 374
 375             /* Declare 1.X samplers implicitly, based on the destination reg. number */
 376             if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 377                 (D3DSIO_TEX == curOpcode->opcode ||
 378                  D3DSIO_TEXM3x2TEX == curOpcode->opcode ||
 379                  D3DSIO_TEXM3x3TEX == curOpcode->opcode ||
 380                  D3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
 381                  D3DSIO_TEXM3x3VSPEC == curOpcode->opcode)) {
 382
 383                 /* Fake sampler usage, only set reserved bit and ttype */
 384                 DWORD sampler_code = *pToken & D3DSP_REGNUM_MASK;
 385                 reg_maps->samplers[sampler_code] = (0x1 << 31) | D3DSTT_2D;
 386             }
 387
 388             /* This will loop over all the registers and try to
 389              * make a bitmask of the ones we're interested in.
 390              *
 391              * Relative addressing tokens are ignored, but that's
 392              * okay, since we'll catch any address registers when
 393              * they are initialized (required by spec) */
 394
 395             limit = (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)?
 396                 curOpcode->num_params + 1: curOpcode->num_params;
 397
 398             for (i = 0; i < limit; ++i) {
 399
 400                 DWORD param, addr_token, reg, regtype;
 401                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
 402
 403                 regtype = (param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT;
 404                 reg = param & D3DSP_REGNUM_MASK;
 405
 406                 if (D3DSPR_TEXTURE == regtype) { /* vs: D3DSPR_ADDR */
 407
 408                     if (pshader)
 409                         reg_maps->texcoord[reg] = 1;
 410                     else
 411                         reg_maps->address[reg] = 1;
 412                 }
 413
 414                 else if (D3DSPR_TEMP == regtype)
 415                     reg_maps->temporary[reg] = 1;
 416
 417                 else if (D3DSPR_INPUT == regtype && !pshader)
 418                     reg_maps->attributes[reg] = 1;
 419              }
 420         }
 421     }
 422 }
 423
 424 static void shader_dump_decl_usage(
 425     DWORD decl,
 426     DWORD param) {
 427
 428     DWORD regtype = shader_get_regtype(param);
 429     TRACE("dcl_");
 430
 431     if (regtype == D3DSPR_SAMPLER) {
 432         DWORD ttype = decl & D3DSP_TEXTURETYPE_MASK;
 433
 434         switch (ttype) {
 435             case D3DSTT_2D: TRACE("2d"); break;
 436             case D3DSTT_CUBE: TRACE("cube"); break;
 437             case D3DSTT_VOLUME: TRACE("volume"); break;
 438             default: TRACE("unknown_ttype(%08lx)", ttype);
 439        }
 440
 441     } else {
 442
 443         DWORD usage = decl & D3DSP_DCL_USAGE_MASK;
 444         DWORD idx = (decl & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
 445
 446         switch(usage) {
 447         case D3DDECLUSAGE_POSITION:
 448             TRACE("%s%ld", "position", idx);
 449             break;
 450         case D3DDECLUSAGE_BLENDINDICES:
 451             TRACE("%s", "blend");
 452             break;
 453         case D3DDECLUSAGE_BLENDWEIGHT:
 454             TRACE("%s", "weight");
 455             break;
 456         case D3DDECLUSAGE_NORMAL:
 457             TRACE("%s%ld", "normal", idx);
 458             break;
 459         case D3DDECLUSAGE_PSIZE:
 460             TRACE("%s", "psize");
 461             break;
 462         case D3DDECLUSAGE_COLOR:
 463             if(idx == 0)  {
 464                 TRACE("%s", "color");
 465             } else {
 466                 TRACE("%s%ld", "specular", (idx - 1));
 467             }
 468             break;
 469         case D3DDECLUSAGE_TEXCOORD:
 470             TRACE("%s%ld", "texture", idx);
 471             break;
 472         case D3DDECLUSAGE_TANGENT:
 473             TRACE("%s", "tangent");
 474             break;
 475         case D3DDECLUSAGE_BINORMAL:
 476             TRACE("%s", "binormal");
 477             break;
 478         case D3DDECLUSAGE_TESSFACTOR:
 479             TRACE("%s", "tessfactor");
 480             break;
 481         case D3DDECLUSAGE_POSITIONT:
 482             TRACE("%s%ld", "positionT", idx);
 483             break;
 484         case D3DDECLUSAGE_FOG:
 485             TRACE("%s", "fog");
 486             break;
 487         case D3DDECLUSAGE_DEPTH:
 488             TRACE("%s", "depth");
 489             break;
 490         case D3DDECLUSAGE_SAMPLE:
 491             TRACE("%s", "sample");
 492             break;
 493         default:
 494             FIXME("unknown_semantics(%08lx)", usage);
 495         }
 496     }
 497 }
 498
 499 static void shader_dump_arr_entry(
 500     IWineD3DBaseShader *iface,
 501     const DWORD param,
 502     const DWORD addr_token,
 503     unsigned int reg,
 504     int input) {
 505
 506     char relative =
 507         ((param & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
 508
 509     if (relative) {
 510         TRACE("[");
 511         if (addr_token)
 512             shader_dump_param(iface, addr_token, 0, input);
 513         else
 514             TRACE("a0.x");
 515         TRACE(" + ");
 516      }
 517      TRACE("%u", reg);
 518      if (relative)
 519          TRACE("]");
 520 }
 521
 522 void shader_dump_param(
 523     IWineD3DBaseShader *iface,
 524     const DWORD param,
 525     const DWORD addr_token,
 526     int input) {
 527
 528     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 529     static const char* rastout_reg_names[] = { "oPos", "oFog", "oPts" };
 530     char swizzle_reg_chars[4];
 531
 532     DWORD reg = param & D3DSP_REGNUM_MASK;
 533     DWORD regtype = shader_get_regtype(param);
 534
 535     /* There are some minor differences between pixel and vertex shaders */
 536     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 537
 538     /* For one, we'd prefer color components to be shown for pshaders.
 539      * FIXME: use the swizzle function for this */
 540
 541     swizzle_reg_chars[0] = pshader? 'r': 'x';
 542     swizzle_reg_chars[1] = pshader? 'g': 'y';
 543     swizzle_reg_chars[2] = pshader? 'b': 'z';
 544     swizzle_reg_chars[3] = pshader? 'a': 'w';
 545
 546     if (input) {
 547         if ( ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG) ||
 548              ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_BIASNEG) ||
 549              ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_SIGNNEG) ||
 550              ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_X2NEG) )
 551             TRACE("-");
 552         else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_COMP)
 553             TRACE("1-");
 554         else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NOT)
 555             TRACE("!");
 556     }
 557
 558     switch (regtype) {
 559         case D3DSPR_TEMP:
 560             TRACE("r%lu", reg);
 561             break;
 562         case D3DSPR_INPUT:
 563             TRACE("v");
 564             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 565             break;
 566         case D3DSPR_CONST:
 567         case D3DSPR_CONST2:
 568         case D3DSPR_CONST3:
 569         case D3DSPR_CONST4:
 570             TRACE("c");
 571             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
 572             break;
 573         case D3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
 574             TRACE("%c%lu", (pshader? 't':'a'), reg);
 575             break;
 576         case D3DSPR_RASTOUT:
 577             TRACE("%s", rastout_reg_names[reg]);
 578             break;
 579         case D3DSPR_COLOROUT:
 580             TRACE("oC%lu", reg);
 581             break;
 582         case D3DSPR_DEPTHOUT:
 583             TRACE("oDepth");
 584             break;
 585         case D3DSPR_ATTROUT:
 586             TRACE("oD%lu", reg);
 587             break;
 588         case D3DSPR_TEXCRDOUT:
 589
 590             /* Vertex shaders >= 3.0 use general purpose output registers
 591              * (D3DSPR_OUTPUT), which can include an address token */
 592
 593             if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
 594                 TRACE("o");
 595                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
 596             }
 597             else
 598                TRACE("oT%lu", reg);
 599             break;
 600         case D3DSPR_CONSTINT:
 601             TRACE("i");
 602             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 603             break;
 604         case D3DSPR_CONSTBOOL:
 605             TRACE("b");
 606             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 607             break;
 608         case D3DSPR_LABEL:
 609             TRACE("l%lu", reg);
 610             break;
 611         case D3DSPR_LOOP:
 612             TRACE("aL");
 613             break;
 614         case D3DSPR_SAMPLER:
 615             TRACE("s%lu", reg);
 616             break;
 617         case D3DSPR_PREDICATE:
 618             TRACE("p%lu", reg);
 619             break;
 620         default:
 621             TRACE("unhandled_rtype(%#lx)", regtype);
 622             break;
 623    }
 624
 625    if (!input) {
 626        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
 627
 628        if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
 629            TRACE(".");
 630            if (param & D3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
 631            if (param & D3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
 632            if (param & D3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
 633            if (param & D3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
 634        }
 635
 636    } else {
 637         /** operand input */
 638         DWORD swizzle = (param & D3DSP_SWIZZLE_MASK) >> D3DSP_SWIZZLE_SHIFT;
 639         DWORD swizzle_r = swizzle & 0x03;
 640         DWORD swizzle_g = (swizzle >> 2) & 0x03;
 641         DWORD swizzle_b = (swizzle >> 4) & 0x03;
 642         DWORD swizzle_a = (swizzle >> 6) & 0x03;
 643
 644         if (0 != (param & D3DSP_SRCMOD_MASK)) {
 645             DWORD mask = param & D3DSP_SRCMOD_MASK;
 646             switch (mask) {
 647                 case D3DSPSM_NONE:    break;
 648                 case D3DSPSM_NEG:     break;
 649                 case D3DSPSM_NOT:     break;
 650                 case D3DSPSM_BIAS:    TRACE("_bias"); break;
 651                 case D3DSPSM_BIASNEG: TRACE("_bias"); break;
 652                 case D3DSPSM_SIGN:    TRACE("_bx2"); break;
 653                 case D3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 654                 case D3DSPSM_COMP:    break;
 655                 case D3DSPSM_X2:      TRACE("_x2"); break;
 656                 case D3DSPSM_X2NEG:   TRACE("_x2"); break;
 657                 case D3DSPSM_DZ:      TRACE("_dz"); break;
 658                 case D3DSPSM_DW:      TRACE("_dw"); break;
 659                 default:
 660                     TRACE("_unknown_modifier(%#lx)", mask >> D3DSP_SRCMOD_SHIFT);
 661             }
 662         }
 663
 664         /**
 665         * swizzle bits fields:
 666         *  RRGGBBAA
 667         */
 668         if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
 669             if (swizzle_r == swizzle_g &&
 670                 swizzle_r == swizzle_b &&
 671                 swizzle_r == swizzle_a) {
 672                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
 673             } else {
 674                 TRACE(".%c%c%c%c",
 675                 swizzle_reg_chars[swizzle_r],
 676                 swizzle_reg_chars[swizzle_g],
 677                 swizzle_reg_chars[swizzle_b],
 678                 swizzle_reg_chars[swizzle_a]);
 679             }
 680         }
 681     }
 682 }
 683
 684 /** Shared code in order to generate the bulk of the shader string.
 685     Use the shader_header_fct & shader_footer_fct to add strings
 686     that are specific to pixel or vertex functions
 687     NOTE: A description of how to parse tokens can be found at:
 688           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
 689 void shader_generate_main(
 690     IWineD3DBaseShader *iface,
 691     SHADER_BUFFER* buffer,
 692     shader_reg_maps* reg_maps,
 693     CONST DWORD* pFunction) {
 694
 695     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 696     const DWORD *pToken = pFunction;
 697     const SHADER_OPCODE *curOpcode = NULL;
 698     SHADER_HANDLER hw_fct = NULL;
 699     DWORD opcode_token;
 700     DWORD i;
 701     SHADER_OPCODE_ARG hw_arg;
 702
 703     /* Initialize current parsing state */
 704     hw_arg.shader = iface;
 705     hw_arg.buffer = buffer;
 706     hw_arg.reg_maps = reg_maps;
 707     This->baseShader.parse_state.current_row = 0;
 708
 709     /* Second pass, process opcodes */
 710     if (NULL != pToken) {
 711         while (D3DPS_END() != *pToken) {
 712
 713             /* Skip version token */
 714             if (shader_is_version_token(*pToken)) {
 715                 ++pToken;
 716                 continue;
 717             }
 718
 719             /* Skip comment tokens */
 720             if (shader_is_comment(*pToken)) {
 721                 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 722                 ++pToken;
 723                 TRACE("#%s\n", (char*)pToken);
 724                 pToken += comment_len;
 725                 continue;
 726             }
 727
 728             /* Read opcode */
 729             opcode_token = *pToken++;
 730             curOpcode = shader_get_opcode(iface, opcode_token);
 731             hw_fct = (curOpcode == NULL)? NULL:
 732                      (wined3d_settings.shader_mode == SHADER_GLSL)?
 733                          curOpcode->hw_glsl_fct : curOpcode->hw_fct;
 734
 735             /* Unknown opcode and its parameters */
 736             if (NULL == curOpcode) {
 737                 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
 738                 pToken += shader_skip_unrecognized(iface, pToken);
 739
 740             /* Nothing to do */
 741             } else if (D3DSIO_DCL == curOpcode->opcode ||
 742                        D3DSIO_NOP == curOpcode->opcode) {
 743
 744                 pToken += shader_skip_opcode(This, curOpcode, opcode_token);
 745
 746             /* If a generator function is set for current shader target, use it */
 747             } else if (hw_fct != NULL) {
 748
 749                 hw_arg.opcode = curOpcode;
 750
 751                 /* Destination token */
 752                 if (curOpcode->dst_token) {
 753
 754                     DWORD param, addr_token = 0;
 755                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 756                     hw_arg.dst = param;
 757                     hw_arg.dst_addr = addr_token;
 758                 }
 759
 760                 /* Predication token */
 761                 if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)
 762                     hw_arg.predicate = *pToken++;
 763
 764                 /* Other source tokens */
 765                 for (i = curOpcode->dst_token; i < curOpcode->num_params; i++) {
 766
 767                     DWORD param, addr_token = 0;
 768
 769                     /* DEF* instructions have constant src parameters, not registers */
 770                     if (curOpcode->opcode == D3DSIO_DEF ||
 771                         curOpcode->opcode == D3DSIO_DEFI ||
 772                         curOpcode->opcode == D3DSIO_DEFB) {
 773                         param = *pToken++;
 774
 775                     } else
 776                         pToken += shader_get_param(iface, pToken, &param, &addr_token);
 777
 778                     hw_arg.src[i-1] = param;
 779                     hw_arg.src_addr[i-1] = addr_token;
 780                 }
 781
 782                 /* Call appropriate function for output target */
 783                 hw_fct(&hw_arg);
 784
 785                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
 786                 if (wined3d_settings.shader_mode == SHADER_GLSL)
 787                     shader_glsl_add_instruction_modifiers(&hw_arg);
 788
 789             /* Unhandled opcode */
 790             } else {
 791
 792                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
 793                 pToken += shader_skip_opcode(This, curOpcode, opcode_token);
 794             }
 795         }
 796         /* TODO: What about result.depth? */
 797
 798     }
 799 }
 800
 801 void shader_dump_ins_modifiers(const DWORD output) {
 802
 803     DWORD shift = (output & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
 804     DWORD mmask = output & D3DSP_DSTMOD_MASK;
 805
 806     switch (shift) {
 807         case 0: break;
 808         case 13: TRACE("_d8"); break;
 809         case 14: TRACE("_d4"); break;
 810         case 15: TRACE("_d2"); break;
 811         case 1: TRACE("_x2"); break;
 812         case 2: TRACE("_x4"); break;
 813         case 3: TRACE("_x8"); break;
 814         default: TRACE("_unhandled_shift(%ld)", shift); break;
 815     }
 816
 817     if (mmask & D3DSPDM_SATURATE)         TRACE("_sat");
 818     if (mmask & D3DSPDM_PARTIALPRECISION) TRACE("_pp");
 819     if (mmask & D3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
 820
 821     mmask &= ~(D3DSPDM_SATURATE | D3DSPDM_PARTIALPRECISION | D3DSPDM_MSAMPCENTROID);
 822     if (mmask)
 823         FIXME("_unrecognized_modifier(%#lx)", mmask >> D3DSP_DSTMOD_SHIFT);
 824 }
 825
 826 /* First pass: trace shader, initialize length and version */
 827 void shader_trace_init(
 828     IWineD3DBaseShader *iface,
 829     const DWORD* pFunction) {
 830
 831     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
 832
 833     const DWORD* pToken = pFunction;
 834     const SHADER_OPCODE* curOpcode = NULL;
 835     DWORD opcode_token;
 836     unsigned int len = 0;
 837     DWORD i;
 838
 839     TRACE("(%p) : Parsing programme\n", This);
 840
 841     if (NULL != pToken) {
 842         while (D3DVS_END() != *pToken) {
 843             if (shader_is_version_token(*pToken)) { /** version */
 844                 This->baseShader.hex_version = *pToken;
 845                 TRACE("%s_%lu_%lu\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
 846                     D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
 847                     D3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
 848                 ++pToken;
 849                 ++len;
 850                 continue;
 851             }
 852             if (shader_is_comment(*pToken)) { /** comment */
 853                 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 854                 ++pToken;
 855                 TRACE("//%s\n", (char*)pToken);
 856                 pToken += comment_len;
 857                 len += comment_len + 1;
 858                 continue;
 859             }
 860             opcode_token = *pToken++;
 861             curOpcode = shader_get_opcode(iface, opcode_token);
 862             len++;
 863
 864             if (NULL == curOpcode) {
 865                 int tokens_read;
 866                 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
 867                 tokens_read = shader_skip_unrecognized(iface, pToken);
 868                 pToken += tokens_read;
 869                 len += tokens_read;
 870
 871             } else {
 872                 if (curOpcode->opcode == D3DSIO_DCL) {
 873
 874                     DWORD usage = *pToken;
 875                     DWORD param = *(pToken + 1);
 876
 877                     shader_dump_decl_usage(usage, param);
 878                     shader_dump_ins_modifiers(param);
 879                     TRACE(" ");
 880                     shader_dump_param(iface, param, 0, 0);
 881                     pToken += 2;
 882                     len += 2;
 883
 884                 } else if (curOpcode->opcode == D3DSIO_DEF) {
 885
 886                         unsigned int offset = shader_get_float_offset(*pToken);
 887
 888                         TRACE("def c%u = %f, %f, %f, %f", offset,
 889                             *(float *)(pToken + 1),
 890                             *(float *)(pToken + 2),
 891                             *(float *)(pToken + 3),
 892                             *(float *)(pToken + 4));
 893
 894                         pToken += 5;
 895                         len += 5;
 896                 } else if (curOpcode->opcode == D3DSIO_DEFI) {
 897
 898                         TRACE("defi i%lu = %ld, %ld, %ld, %ld", *pToken & D3DSP_REGNUM_MASK,
 899                             (long) *(pToken + 1),
 900                             (long) *(pToken + 2),
 901                             (long) *(pToken + 3),
 902                             (long) *(pToken + 4));
 903
 904                         pToken += 5;
 905                         len += 5;
 906
 907                 } else if (curOpcode->opcode == D3DSIO_DEFB) {
 908
 909                         TRACE("defb b%lu = %s", *pToken & D3DSP_REGNUM_MASK,
 910                             *(pToken + 1)? "true": "false");
 911
 912                         pToken += 2;
 913                         len += 2;
 914
 915                 } else {
 916
 917                     DWORD param, addr_token;
 918                     int tokens_read;
 919
 920                     /* Print out predication source token first - it follows
 921                      * the destination token. */
 922                     if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
 923                         TRACE("(");
 924                         shader_dump_param(iface, *(pToken + 2), 0, 1);
 925                         TRACE(") ");
 926                     }
 927
 928                     TRACE("%s", curOpcode->name);
 929
 930                     /* Destination token */
 931                     if (curOpcode->dst_token) {
 932
 933                         /* Destination token */
 934                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 935                         pToken += tokens_read;
 936                         len += tokens_read;
 937
 938                         shader_dump_ins_modifiers(param);
 939                         TRACE(" ");
 940                         shader_dump_param(iface, param, addr_token, 0);
 941                     }
 942
 943                     /* Predication token - already printed out, just skip it */
 944                     if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
 945                         pToken++;
 946                         len++;
 947                     }
 948
 949                     /* Other source tokens */
 950                     for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
 951
 952                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 953                         pToken += tokens_read;
 954                         len += tokens_read;
 955
 956                         TRACE((i == 0)? " " : ", ");
 957                         shader_dump_param(iface, param, addr_token, 1);
 958                     }
 959                 }
 960                 TRACE("\n");
 961             }
 962         }
 963         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
 964     } else {
 965         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
 966     }
 967 }
 968
 969 /* TODO: Move other shared code here */