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             if (D3DSPR_INPUT == regtype) {
 343
 344                 if (!pshader)
 345                     reg_maps->attributes[regnum] = 1;
 346
 347                 shader_parse_decl_usage(reg_maps->semantics_in, usage, param);
 348
 349             } else if (D3DSPR_OUTPUT == regtype) {
 350                 shader_parse_decl_usage(reg_maps->semantics_out, usage, param);
 351             }
 352
 353             /* Handle samplers here */
 354
 355         /* Skip definitions (for now) */
 356         } else if (D3DSIO_DEF == curOpcode->opcode) {
 357             pToken += curOpcode->num_params;
 358
 359         /* Set texture registers, and temporary registers */
 360         } else {
 361             int i, limit;
 362
 363             /* This will loop over all the registers and try to
 364              * make a bitmask of the ones we're interested in.
 365              *
 366              * Relative addressing tokens are ignored, but that's
 367              * okay, since we'll catch any address registers when
 368              * they are initialized (required by spec) */
 369
 370             limit = (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)?
 371                 curOpcode->num_params + 1: curOpcode->num_params;
 372
 373             for (i = 0; i < limit; ++i) {
 374
 375                 DWORD param, addr_token, reg, regtype;
 376                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
 377
 378                 regtype = (param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT;
 379                 reg = param & D3DSP_REGNUM_MASK;
 380
 381                 if (D3DSPR_TEXTURE == regtype) { /* vs: D3DSPR_ADDR */
 382
 383                     if (pshader)
 384                         reg_maps->texcoord[reg] = 1;
 385                     else
 386                         reg_maps->address[reg] = 1;
 387                 }
 388
 389                 else if (D3DSPR_TEMP == regtype)
 390                     reg_maps->temporary[reg] = 1;
 391
 392                 else if (D3DSPR_INPUT == regtype && !pshader)
 393                     reg_maps->attributes[reg] = 1;
 394              }
 395         }
 396     }
 397 }
 398
 399 static void shader_dump_decl_usage(
 400     DWORD decl,
 401     DWORD param) {
 402
 403     DWORD regtype = shader_get_regtype(param);
 404     TRACE("dcl_");
 405
 406     if (regtype == D3DSPR_SAMPLER) {
 407         DWORD ttype = decl & D3DSP_TEXTURETYPE_MASK;
 408
 409         switch (ttype) {
 410             case D3DSTT_2D: TRACE("2d"); break;
 411             case D3DSTT_CUBE: TRACE("cube"); break;
 412             case D3DSTT_VOLUME: TRACE("volume"); break;
 413             default: TRACE("unknown_ttype(%08lx)", ttype);
 414        }
 415
 416     } else {
 417
 418         DWORD usage = decl & D3DSP_DCL_USAGE_MASK;
 419         DWORD idx = (decl & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
 420
 421         switch(usage) {
 422         case D3DDECLUSAGE_POSITION:
 423             TRACE("%s%ld", "position", idx);
 424             break;
 425         case D3DDECLUSAGE_BLENDINDICES:
 426             TRACE("%s", "blend");
 427             break;
 428         case D3DDECLUSAGE_BLENDWEIGHT:
 429             TRACE("%s", "weight");
 430             break;
 431         case D3DDECLUSAGE_NORMAL:
 432             TRACE("%s%ld", "normal", idx);
 433             break;
 434         case D3DDECLUSAGE_PSIZE:
 435             TRACE("%s", "psize");
 436             break;
 437         case D3DDECLUSAGE_COLOR:
 438             if(idx == 0)  {
 439                 TRACE("%s", "color");
 440             } else {
 441                 TRACE("%s%ld", "specular", (idx - 1));
 442             }
 443             break;
 444         case D3DDECLUSAGE_TEXCOORD:
 445             TRACE("%s%ld", "texture", idx);
 446             break;
 447         case D3DDECLUSAGE_TANGENT:
 448             TRACE("%s", "tangent");
 449             break;
 450         case D3DDECLUSAGE_BINORMAL:
 451             TRACE("%s", "binormal");
 452             break;
 453         case D3DDECLUSAGE_TESSFACTOR:
 454             TRACE("%s", "tessfactor");
 455             break;
 456         case D3DDECLUSAGE_POSITIONT:
 457             TRACE("%s%ld", "positionT", idx);
 458             break;
 459         case D3DDECLUSAGE_FOG:
 460             TRACE("%s", "fog");
 461             break;
 462         case D3DDECLUSAGE_DEPTH:
 463             TRACE("%s", "depth");
 464             break;
 465         case D3DDECLUSAGE_SAMPLE:
 466             TRACE("%s", "sample");
 467             break;
 468         default:
 469             FIXME("unknown_semantics(%08lx)", usage);
 470         }
 471     }
 472 }
 473
 474 static void shader_dump_arr_entry(
 475     IWineD3DBaseShader *iface,
 476     const DWORD param,
 477     const DWORD addr_token,
 478     unsigned int reg,
 479     int input) {
 480
 481     char relative =
 482         ((param & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
 483
 484     if (relative) {
 485         TRACE("[");
 486         if (addr_token)
 487             shader_dump_param(iface, addr_token, 0, input);
 488         else
 489             TRACE("a0.x");
 490         TRACE(" + ");
 491      }
 492      TRACE("%u", reg);
 493      if (relative)
 494          TRACE("]");
 495 }
 496
 497 void shader_dump_param(
 498     IWineD3DBaseShader *iface,
 499     const DWORD param,
 500     const DWORD addr_token,
 501     int input) {
 502
 503     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 504     static const char* rastout_reg_names[] = { "oPos", "oFog", "oPts" };
 505     char swizzle_reg_chars[4];
 506
 507     DWORD reg = param & D3DSP_REGNUM_MASK;
 508     DWORD regtype = shader_get_regtype(param);
 509
 510     /* There are some minor differences between pixel and vertex shaders */
 511     BOOL pshader = shader_is_pshader_version(This->baseShader.hex_version);
 512
 513     /* For one, we'd prefer color components to be shown for pshaders.
 514      * FIXME: use the swizzle function for this */
 515
 516     swizzle_reg_chars[0] = pshader? 'r': 'x';
 517     swizzle_reg_chars[1] = pshader? 'g': 'y';
 518     swizzle_reg_chars[2] = pshader? 'b': 'z';
 519     swizzle_reg_chars[3] = pshader? 'a': 'w';
 520
 521     if (input) {
 522         if ( ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG) ||
 523              ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_BIASNEG) ||
 524              ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_SIGNNEG) ||
 525              ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_X2NEG) )
 526             TRACE("-");
 527         else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_COMP)
 528             TRACE("1-");
 529         else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NOT)
 530             TRACE("!");
 531     }
 532
 533     switch (regtype) {
 534         case D3DSPR_TEMP:
 535             TRACE("r%lu", reg);
 536             break;
 537         case D3DSPR_INPUT:
 538             TRACE("v");
 539             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 540             break;
 541         case D3DSPR_CONST:
 542         case D3DSPR_CONST2:
 543         case D3DSPR_CONST3:
 544         case D3DSPR_CONST4:
 545             TRACE("c");
 546             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
 547             break;
 548         case D3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
 549             TRACE("%c%lu", (pshader? 't':'a'), reg);
 550             break;
 551         case D3DSPR_RASTOUT:
 552             TRACE("%s", rastout_reg_names[reg]);
 553             break;
 554         case D3DSPR_COLOROUT:
 555             TRACE("oC%lu", reg);
 556             break;
 557         case D3DSPR_DEPTHOUT:
 558             TRACE("oDepth");
 559             break;
 560         case D3DSPR_ATTROUT:
 561             TRACE("oD%lu", reg);
 562             break;
 563         case D3DSPR_TEXCRDOUT:
 564
 565             /* Vertex shaders >= 3.0 use general purpose output registers
 566              * (D3DSPR_OUTPUT), which can include an address token */
 567
 568             if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
 569                 TRACE("o");
 570                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
 571             }
 572             else
 573                TRACE("oT%lu", reg);
 574             break;
 575         case D3DSPR_CONSTINT:
 576             TRACE("i");
 577             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 578             break;
 579         case D3DSPR_CONSTBOOL:
 580             TRACE("b");
 581             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 582             break;
 583         case D3DSPR_LABEL:
 584             TRACE("l%lu", reg);
 585             break;
 586         case D3DSPR_LOOP:
 587             TRACE("aL");
 588             break;
 589         case D3DSPR_SAMPLER:
 590             TRACE("s%lu", reg);
 591             break;
 592         case D3DSPR_PREDICATE:
 593             TRACE("p%lu", reg);
 594             break;
 595         default:
 596             TRACE("unhandled_rtype(%#lx)", regtype);
 597             break;
 598    }
 599
 600    if (!input) {
 601        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
 602
 603        if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
 604            TRACE(".");
 605            if (param & D3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
 606            if (param & D3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
 607            if (param & D3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
 608            if (param & D3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
 609        }
 610
 611    } else {
 612         /** operand input */
 613         DWORD swizzle = (param & D3DSP_SWIZZLE_MASK) >> D3DSP_SWIZZLE_SHIFT;
 614         DWORD swizzle_r = swizzle & 0x03;
 615         DWORD swizzle_g = (swizzle >> 2) & 0x03;
 616         DWORD swizzle_b = (swizzle >> 4) & 0x03;
 617         DWORD swizzle_a = (swizzle >> 6) & 0x03;
 618
 619         if (0 != (param & D3DSP_SRCMOD_MASK)) {
 620             DWORD mask = param & D3DSP_SRCMOD_MASK;
 621             switch (mask) {
 622                 case D3DSPSM_NONE:    break;
 623                 case D3DSPSM_NEG:     break;
 624                 case D3DSPSM_NOT:     break;
 625                 case D3DSPSM_BIAS:    TRACE("_bias"); break;
 626                 case D3DSPSM_BIASNEG: TRACE("_bias"); break;
 627                 case D3DSPSM_SIGN:    TRACE("_bx2"); break;
 628                 case D3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 629                 case D3DSPSM_COMP:    break;
 630                 case D3DSPSM_X2:      TRACE("_x2"); break;
 631                 case D3DSPSM_X2NEG:   TRACE("_x2"); break;
 632                 case D3DSPSM_DZ:      TRACE("_dz"); break;
 633                 case D3DSPSM_DW:      TRACE("_dw"); break;
 634                 default:
 635                     TRACE("_unknown_modifier(%#lx)", mask >> D3DSP_SRCMOD_SHIFT);
 636             }
 637         }
 638
 639         /**
 640         * swizzle bits fields:
 641         *  RRGGBBAA
 642         */
 643         if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
 644             if (swizzle_r == swizzle_g &&
 645                 swizzle_r == swizzle_b &&
 646                 swizzle_r == swizzle_a) {
 647                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
 648             } else {
 649                 TRACE(".%c%c%c%c",
 650                 swizzle_reg_chars[swizzle_r],
 651                 swizzle_reg_chars[swizzle_g],
 652                 swizzle_reg_chars[swizzle_b],
 653                 swizzle_reg_chars[swizzle_a]);
 654             }
 655         }
 656     }
 657 }
 658
 659 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 660 void shader_generate_arb_declarations(
 661     IWineD3DBaseShader *iface,
 662     shader_reg_maps* reg_maps,
 663     SHADER_BUFFER* buffer) {
 664
 665     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 666     DWORD i;
 667
 668     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 669         if (reg_maps->temporary[i])
 670             shader_addline(buffer, "TEMP R%lu;\n", i);
 671     }
 672
 673     for (i = 0; i < This->baseShader.limits.address; i++) {
 674         if (reg_maps->address[i])
 675             shader_addline(buffer, "ADDRESS A%ld;\n", i);
 676     }
 677
 678     for(i = 0; i < This->baseShader.limits.texture; i++) {
 679         if (reg_maps->texcoord[i])
 680             shader_addline(buffer,"TEMP T%lu;\n", i);
 681     }
 682
 683     /* Texture coordinate registers must be pre-loaded */
 684     for (i = 0; i < This->baseShader.limits.texture; i++) {
 685         if (reg_maps->texcoord[i])
 686             shader_addline(buffer, "MOV T%lu, fragment.texcoord[%lu];\n", i, i);
 687     }
 688
 689     /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 690     shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 691                    This->baseShader.limits.constant_float,
 692                    This->baseShader.limits.constant_float - 1);
 693 }
 694
 695 /** Generate the variable & register declarations for the GLSL output target */
 696 void shader_generate_glsl_declarations(
 697     IWineD3DBaseShader *iface,
 698     shader_reg_maps* reg_maps,
 699     SHADER_BUFFER* buffer) {
 700
 701     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 702     int i;
 703
 704     FIXME("GLSL not fully implemented yet.\n");
 705
 706     /* Declare the constants (aka uniforms) */
 707     shader_addline(buffer, "uniform vec4 C[%u];\n", This->baseShader.limits.constant_float);
 708
 709     /* Declare texture samplers
 710      * TODO: Make this work for textures other than 2D */
 711     for (i = 0; i < This->baseShader.limits.texture; i++) {
 712         shader_addline(buffer, "uniform sampler2D mytex%lu;\n", i);
 713     }
 714
 715     /* Declare address variables */
 716     for (i = 0; i < This->baseShader.limits.address; i++) {
 717         if (reg_maps->address[i])
 718             shader_addline(buffer, "ivec4 A%ld;\n", i);
 719     }
 720
 721     /* Declare texture coordinate temporaries and initialize them */
 722     for (i = 0; i < This->baseShader.limits.texture; i++) {
 723         shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
 724     }
 725
 726     /* Declare temporary variables */
 727     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 728         if (reg_maps->temporary[i])
 729             shader_addline(buffer, "vec4 R%lu;\n", i);
 730     }
 731
 732     /* Declare attributes */
 733     for (i = 0; i < This->baseShader.limits.attributes; i++) {
 734         if (reg_maps->attributes[i])
 735             shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
 736     }
 737
 738     /* Temporary variables for matrix operations */
 739     shader_addline(buffer, "vec4 tmp0;\n");
 740     shader_addline(buffer, "vec4 tmp1;\n");
 741
 742     /* Start the main program */
 743     shader_addline(buffer, "void main() {\n");
 744 }
 745
 746 /** Shared code in order to generate the bulk of the shader string.
 747     Use the shader_header_fct & shader_footer_fct to add strings
 748     that are specific to pixel or vertex functions
 749     NOTE: A description of how to parse tokens can be found at:
 750           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
 751 void shader_generate_main(
 752     IWineD3DBaseShader *iface,
 753     SHADER_BUFFER* buffer,
 754     shader_reg_maps* reg_maps,
 755     CONST DWORD* pFunction) {
 756
 757     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 758     const DWORD *pToken = pFunction;
 759     const SHADER_OPCODE *curOpcode = NULL;
 760     SHADER_HANDLER hw_fct = NULL;
 761     DWORD opcode_token;
 762     DWORD i;
 763     SHADER_OPCODE_ARG hw_arg;
 764
 765     /* Initialize current parsing state */
 766     hw_arg.shader = iface;
 767     hw_arg.buffer = buffer;
 768     hw_arg.reg_maps = reg_maps;
 769     This->baseShader.parse_state.current_row = 0;
 770
 771     /* Second pass, process opcodes */
 772     if (NULL != pToken) {
 773         while (D3DPS_END() != *pToken) {
 774
 775             /* Skip version token */
 776             if (shader_is_version_token(*pToken)) {
 777                 ++pToken;
 778                 continue;
 779             }
 780
 781             /* Skip comment tokens */
 782             if (shader_is_comment(*pToken)) {
 783                 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 784                 ++pToken;
 785                 TRACE("#%s\n", (char*)pToken);
 786                 pToken += comment_len;
 787                 continue;
 788             }
 789
 790             /* Read opcode */
 791             opcode_token = *pToken++;
 792             curOpcode = shader_get_opcode(iface, opcode_token);
 793             hw_fct = (curOpcode == NULL)? NULL:
 794                      (wined3d_settings.shader_mode == SHADER_GLSL)?
 795                          curOpcode->hw_glsl_fct : curOpcode->hw_fct;
 796
 797             /* Unknown opcode and its parameters */
 798             if (NULL == curOpcode) {
 799                 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
 800                 pToken += shader_skip_unrecognized(iface, pToken);
 801
 802             /* Nothing to do */
 803             } else if (D3DSIO_DCL == curOpcode->opcode ||
 804                        D3DSIO_NOP == curOpcode->opcode) {
 805
 806                 pToken += shader_skip_opcode(This, curOpcode, opcode_token);
 807
 808             /* If a generator function is set for current shader target, use it */
 809             } else if (hw_fct != NULL) {
 810
 811                 hw_arg.opcode = curOpcode;
 812
 813                 if (curOpcode->num_params > 0) {
 814
 815                     DWORD param, addr_token = 0;
 816
 817                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 818                     hw_arg.dst = param;
 819                     hw_arg.dst_addr = addr_token;
 820
 821                     if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)
 822                         hw_arg.predicate = *pToken++;
 823
 824                     for (i = 1; i < curOpcode->num_params; i++) {
 825                         /* DEF* instructions have constant src parameters, not registers */
 826                         if (curOpcode->opcode == D3DSIO_DEF ||
 827                             curOpcode->opcode == D3DSIO_DEFI ||
 828                             curOpcode->opcode == D3DSIO_DEFB) {
 829                             param = *pToken++;
 830
 831                         } else
 832                             pToken += shader_get_param(iface, pToken, &param, &addr_token);
 833
 834                         hw_arg.src[i-1] = param;
 835                         hw_arg.src_addr[i-1] = addr_token;
 836                     }
 837                 }
 838
 839                 /* Call appropriate function for output target */
 840                 hw_fct(&hw_arg);
 841
 842             /* Unhandled opcode */
 843             } else {
 844
 845                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
 846                 pToken += shader_skip_opcode(This, curOpcode, opcode_token);
 847             }
 848         }
 849         /* TODO: What about result.depth? */
 850
 851     }
 852 }
 853
 854 void shader_dump_ins_modifiers(const DWORD output) {
 855
 856     DWORD shift = (output & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
 857     DWORD mmask = output & D3DSP_DSTMOD_MASK;
 858
 859     switch (shift) {
 860         case 0: break;
 861         case 13: TRACE("_d8"); break;
 862         case 14: TRACE("_d4"); break;
 863         case 15: TRACE("_d2"); break;
 864         case 1: TRACE("_x2"); break;
 865         case 2: TRACE("_x4"); break;
 866         case 3: TRACE("_x8"); break;
 867         default: TRACE("_unhandled_shift(%ld)", shift); break;
 868     }
 869
 870     if (mmask & D3DSPDM_SATURATE)         TRACE("_sat");
 871     if (mmask & D3DSPDM_PARTIALPRECISION) TRACE("_pp");
 872     if (mmask & D3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
 873
 874     mmask &= ~(D3DSPDM_SATURATE | D3DSPDM_PARTIALPRECISION | D3DSPDM_MSAMPCENTROID);
 875     if (mmask)
 876         FIXME("_unrecognized_modifier(%#lx)", mmask >> D3DSP_DSTMOD_SHIFT);
 877 }
 878
 879 /** Process the D3DSIO_DEF opcode into an ARB string - creates a local vec4
 880  * float constant, and stores it's usage on the regmaps. */
 881 void shader_hw_def(SHADER_OPCODE_ARG* arg) {
 882
 883     DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
 884
 885     shader_addline(arg->buffer,
 886                    "PARAM C%lu = { %f, %f, %f, %f };\n", reg,
 887                    *((const float *)(arg->src + 0)),
 888                    *((const float *)(arg->src + 1)),
 889                    *((const float *)(arg->src + 2)),
 890                    *((const float *)(arg->src + 3)) );
 891
 892     arg->reg_maps->constantsF[reg] = 1;
 893 }
 894
 895 /* First pass: trace shader, initialize length and version */
 896 void shader_trace_init(
 897     IWineD3DBaseShader *iface,
 898     const DWORD* pFunction) {
 899
 900     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
 901
 902     const DWORD* pToken = pFunction;
 903     const SHADER_OPCODE* curOpcode = NULL;
 904     DWORD opcode_token;
 905     unsigned int len = 0;
 906     DWORD i;
 907
 908     TRACE("(%p) : Parsing programme\n", This);
 909
 910     if (NULL != pToken) {
 911         while (D3DVS_END() != *pToken) {
 912             if (shader_is_version_token(*pToken)) { /** version */
 913                 This->baseShader.hex_version = *pToken;
 914                 TRACE("%s_%lu_%lu\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
 915                     D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
 916                     D3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
 917                 ++pToken;
 918                 ++len;
 919                 continue;
 920             }
 921             if (shader_is_comment(*pToken)) { /** comment */
 922                 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 923                 ++pToken;
 924                 TRACE("//%s\n", (char*)pToken);
 925                 pToken += comment_len;
 926                 len += comment_len + 1;
 927                 continue;
 928             }
 929             opcode_token = *pToken++;
 930             curOpcode = shader_get_opcode(iface, opcode_token);
 931             len++;
 932
 933             if (NULL == curOpcode) {
 934                 int tokens_read;
 935                 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
 936                 tokens_read = shader_skip_unrecognized(iface, pToken);
 937                 pToken += tokens_read;
 938                 len += tokens_read;
 939
 940             } else {
 941                 if (curOpcode->opcode == D3DSIO_DCL) {
 942
 943                     DWORD usage = *pToken;
 944                     DWORD param = *(pToken + 1);
 945
 946                     shader_dump_decl_usage(usage, param);
 947                     shader_dump_ins_modifiers(param);
 948                     TRACE(" ");
 949                     shader_dump_param(iface, param, 0, 0);
 950                     pToken += 2;
 951                     len += 2;
 952
 953                 } else if (curOpcode->opcode == D3DSIO_DEF) {
 954
 955                         unsigned int offset = shader_get_float_offset(*pToken);
 956
 957                         TRACE("def c%u = %f, %f, %f, %f", offset,
 958                             *(float *)(pToken + 1),
 959                             *(float *)(pToken + 2),
 960                             *(float *)(pToken + 3),
 961                             *(float *)(pToken + 4));
 962
 963                         pToken += 5;
 964                         len += 5;
 965                 } else if (curOpcode->opcode == D3DSIO_DEFI) {
 966
 967                         TRACE("defi i%lu = %ld, %ld, %ld, %ld", *pToken & D3DSP_REGNUM_MASK,
 968                             (long) *(pToken + 1),
 969                             (long) *(pToken + 2),
 970                             (long) *(pToken + 3),
 971                             (long) *(pToken + 4));
 972
 973                         pToken += 5;
 974                         len += 5;
 975
 976                 } else if (curOpcode->opcode == D3DSIO_DEFB) {
 977
 978                         TRACE("defb b%lu = %s", *pToken & D3DSP_REGNUM_MASK,
 979                             *(pToken + 1)? "true": "false");
 980
 981                         pToken += 2;
 982                         len += 2;
 983
 984                 } else {
 985
 986                     DWORD param, addr_token;
 987                     int tokens_read;
 988
 989                     /* Print out predication source token first - it follows
 990                      * the destination token. */
 991                     if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
 992                         TRACE("(");
 993                         shader_dump_param(iface, *(pToken + 2), 0, 1);
 994                         TRACE(") ");
 995                     }
 996
 997                     TRACE("%s", curOpcode->name);
 998                     if (curOpcode->num_params > 0) {
 999
1000                         /* Destination token */
1001                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1002                         pToken += tokens_read;
1003                         len += tokens_read;
1004
1005                         shader_dump_ins_modifiers(param);
1006                         TRACE(" ");
1007                         shader_dump_param(iface, param, addr_token, 0);
1008
1009                         /* Predication token - already printed out, just skip it */
1010                         if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
1011                             pToken++;
1012                             len++;
1013                         }
1014                         /* Other source tokens */
1015                         for (i = 1; i < curOpcode->num_params; ++i) {
1016
1017                             tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1018                             pToken += tokens_read;
1019                             len += tokens_read;
1020
1021                             TRACE(", ");
1022                             shader_dump_param(iface, param, addr_token, 1);
1023                         }
1024                     }
1025                 }
1026                 TRACE("\n");
1027             }
1028         }
1029         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1030     } else {
1031         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1032     }
1033 }
1034
1035 /* TODO: Move other shared code here */