dlls/wined3d/baseshader.c

   1 /*
   2  * shaders implementation
   3  *
   4  * Copyright 2002-2003 Jason Edmeades
   5  * Copyright 2002-2003 Raphael Junqueira
   6  * Copyright 2004 Christian Costa
   7  * Copyright 2005 Oliver Stieber
   8  * Copyright 2006 Ivan Gyurdiev
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #include "config.h"
  26 #include <string.h>
  27 #include <stdio.h>
  28 #include "wined3d_private.h"
  29
  30 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  31
  32 #define GLNAME_REQUIRE_GLSL  ((const char *)1)
  33
  34 static inline BOOL shader_is_version_token(DWORD token) {
  35     return shader_is_pshader_version(token) ||
  36            shader_is_vshader_version(token);
  37 }
  38
  39 int shader_addline(
  40     SHADER_BUFFER* buffer,
  41     const char *format, ...) {
  42
  43     char* base = buffer->buffer + buffer->bsize;
  44     int rc;
  45
  46     va_list args;
  47     va_start(args, format);
  48     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
  49     va_end(args);
  50
  51     if (rc < 0 ||                                   /* C89 */
  52         rc > SHADER_PGMSIZE - 1 - buffer->bsize) {  /* C99 */
  53
  54         ERR("The buffer allocated for the shader program string "
  55             "is too small at %d bytes.\n", SHADER_PGMSIZE);
  56         buffer->bsize = SHADER_PGMSIZE - 1;
  57         return -1;
  58     }
  59
  60     if (buffer->newline) {
  61         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
  62         buffer->newline = FALSE;
  63     } else {
  64         TRACE("%s", base);
  65     }
  66
  67     buffer->bsize += rc;
  68     if (buffer->buffer[buffer->bsize-1] == '\n') {
  69         buffer->lineNo++;
  70         buffer->newline = TRUE;
  71     }
  72     return 0;
  73 }
  74
  75 const SHADER_OPCODE* shader_get_opcode(
  76     IWineD3DBaseShader *iface, const DWORD code) {
  77
  78     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl*) iface;
  79
  80     DWORD i = 0;
  81     DWORD hex_version = This->baseShader.hex_version;
  82     const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
  83
  84     /** TODO: use dichotomic search */
  85     while (NULL != shader_ins[i].name) {
  86         if (((code & WINED3DSI_OPCODE_MASK) == shader_ins[i].opcode) &&
  87             (((hex_version >= shader_ins[i].min_version) && (hex_version <= shader_ins[i].max_version)) ||
  88             ((shader_ins[i].min_version == 0) && (shader_ins[i].max_version == 0)))) {
  89             return &shader_ins[i];
  90         }
  91         ++i;
  92     }
  93     FIXME("Unsupported opcode %#x(%d) masked %#x, shader version %#x\n",
  94        code, code, code & WINED3DSI_OPCODE_MASK, hex_version);
  95     return NULL;
  96 }
  97
  98 /* Read a parameter opcode from the input stream,
  99  * and possibly a relative addressing token.
 100  * Return the number of tokens read */
 101 int shader_get_param(
 102     IWineD3DBaseShader* iface,
 103     const DWORD* pToken,
 104     DWORD* param,
 105     DWORD* addr_token) {
 106
 107     /* PS >= 3.0 have relative addressing (with token)
 108      * VS >= 2.0 have relative addressing (with token)
 109      * VS >= 1.0 < 2.0 have relative addressing (without token)
 110      * The version check below should work in general */
 111
 112     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 113     char rel_token = WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
 114         ((*pToken & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
 115
 116     *param = *pToken;
 117     *addr_token = rel_token? *(pToken + 1): 0;
 118     return rel_token? 2:1;
 119 }
 120
 121 /* Return the number of parameters to skip for an opcode */
 122 static inline int shader_skip_opcode(
 123     IWineD3DBaseShaderImpl* This,
 124     const SHADER_OPCODE* curOpcode,
 125     DWORD opcode_token) {
 126
 127    /* Shaders >= 2.0 may contain address tokens, but fortunately they
 128     * have a useful legnth mask - use it here. Shaders 1.0 contain no such tokens */
 129
 130     return (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
 131         ((opcode_token & WINED3DSI_INSTLENGTH_MASK) >> WINED3DSI_INSTLENGTH_SHIFT):
 132         curOpcode->num_params;
 133 }
 134
 135 /* Read the parameters of an unrecognized opcode from the input stream
 136  * Return the number of tokens read.
 137  *
 138  * Note: This function assumes source or destination token format.
 139  * It will not work with specially-formatted tokens like DEF or DCL,
 140  * but hopefully those would be recognized */
 141
 142 int shader_skip_unrecognized(
 143     IWineD3DBaseShader* iface,
 144     const DWORD* pToken) {
 145
 146     int tokens_read = 0;
 147     int i = 0;
 148
 149     /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
 150     while (*pToken & 0x80000000) {
 151
 152         DWORD param, addr_token;
 153         tokens_read += shader_get_param(iface, pToken, &param, &addr_token);
 154         pToken += tokens_read;
 155
 156         FIXME("Unrecognized opcode param: token=%08x "
 157             "addr_token=%08x name=", param, addr_token);
 158         shader_dump_param(iface, param, addr_token, i);
 159         FIXME("\n");
 160         ++i;
 161     }
 162     return tokens_read;
 163 }
 164
 165 /* Convert floating point offset relative
 166  * to a register file to an absolute offset for float constants */
 167
 168 unsigned int shader_get_float_offset(const DWORD reg) {
 169
 170      unsigned int regnum = reg & WINED3DSP_REGNUM_MASK;
 171      int regtype = shader_get_regtype(reg);
 172
 173      switch (regtype) {
 174         case WINED3DSPR_CONST: return regnum;
 175         case WINED3DSPR_CONST2: return 2048 + regnum;
 176         case WINED3DSPR_CONST3: return 4096 + regnum;
 177         case WINED3DSPR_CONST4: return 6144 + regnum;
 178         default:
 179             FIXME("Unsupported register type: %d\n", regtype);
 180             return regnum;
 181      }
 182 }
 183
 184 /* Note that this does not count the loop register
 185  * as an address register. */
 186
 187 HRESULT shader_get_registers_used(
 188     IWineD3DBaseShader *iface,
 189     shader_reg_maps* reg_maps,
 190     semantic* semantics_in,
 191     semantic* semantics_out,
 192     CONST DWORD* pToken,
 193     IWineD3DStateBlockImpl *stateBlock) {
 194
 195     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 196
 197     /* There are some minor differences between pixel and vertex shaders */
 198     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 199
 200     reg_maps->bumpmat = -1;
 201     reg_maps->luminanceparams = -1;
 202
 203     if (pToken == NULL)
 204         return WINED3D_OK;
 205
 206     while (WINED3DVS_END() != *pToken) {
 207         CONST SHADER_OPCODE* curOpcode;
 208         DWORD opcode_token;
 209
 210         /* Skip version */
 211         if (shader_is_version_token(*pToken)) {
 212              ++pToken;
 213              continue;
 214
 215         /* Skip comments */
 216         } else if (shader_is_comment(*pToken)) {
 217              DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 218              ++pToken;
 219              pToken += comment_len;
 220              continue;
 221         }
 222
 223         /* Fetch opcode */
 224         opcode_token = *pToken++;
 225         curOpcode = shader_get_opcode(iface, opcode_token);
 226
 227         /* Unhandled opcode, and its parameters */
 228         if (NULL == curOpcode) {
 229            while (*pToken & 0x80000000)
 230                ++pToken;
 231
 232         /* Handle declarations */
 233         } else if (WINED3DSIO_DCL == curOpcode->opcode) {
 234
 235             DWORD usage = *pToken++;
 236             DWORD param = *pToken++;
 237             DWORD regtype = shader_get_regtype(param);
 238             unsigned int regnum = param & WINED3DSP_REGNUM_MASK;
 239
 240             /* Vshader: mark attributes used
 241                Pshader: mark 3.0 input registers used, save token */
 242             if (WINED3DSPR_INPUT == regtype) {
 243
 244                 if (!pshader)
 245                     reg_maps->attributes[regnum] = 1;
 246                 else
 247                     reg_maps->packed_input[regnum] = 1;
 248
 249                 semantics_in[regnum].usage = usage;
 250                 semantics_in[regnum].reg = param;
 251
 252             /* Vshader: mark 3.0 output registers used, save token */
 253             } else if (WINED3DSPR_OUTPUT == regtype) {
 254                 reg_maps->packed_output[regnum] = 1;
 255                 semantics_out[regnum].usage = usage;
 256                 semantics_out[regnum].reg = param;
 257                 if (usage & (WINED3DDECLUSAGE_FOG << WINED3DSP_DCL_USAGE_SHIFT))
 258                     reg_maps->fog = 1;
 259
 260             /* Save sampler usage token */
 261             } else if (WINED3DSPR_SAMPLER == regtype)
 262                 reg_maps->samplers[regnum] = usage;
 263
 264         } else if (WINED3DSIO_DEF == curOpcode->opcode) {
 265
 266             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 267             if (!lconst) return E_OUTOFMEMORY;
 268             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 269             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 270             list_add_head(&This->baseShader.constantsF, &lconst->entry);
 271             pToken += curOpcode->num_params;
 272
 273         } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
 274
 275             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 276             if (!lconst) return E_OUTOFMEMORY;
 277             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 278             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 279             list_add_head(&This->baseShader.constantsI, &lconst->entry);
 280             pToken += curOpcode->num_params;
 281
 282         } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
 283
 284             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 285             if (!lconst) return E_OUTOFMEMORY;
 286             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 287             memcpy(&lconst->value, pToken + 1, 1 * sizeof(DWORD));
 288             list_add_head(&This->baseShader.constantsB, &lconst->entry);
 289             pToken += curOpcode->num_params;
 290
 291         /* If there's a loop in the shader */
 292         } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
 293                    WINED3DSIO_REP == curOpcode->opcode) {
 294             reg_maps->loop = 1;
 295             pToken += curOpcode->num_params;
 296
 297         /* For subroutine prototypes */
 298         } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
 299
 300             DWORD snum = *pToken & WINED3DSP_REGNUM_MASK;
 301             reg_maps->labels[snum] = 1;
 302             pToken += curOpcode->num_params;
 303
 304         } else if(WINED3DSIO_BEM == curOpcode->opcode) {
 305             DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
 306             if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != regnum) {
 307                 FIXME("Pixel shader uses bem or texbem instruction on more than 1 sampler\n");
 308             } else {
 309                 reg_maps->bumpmat = regnum;
 310             }
 311
 312         /* Set texture, address, temporary registers */
 313         } else {
 314             int i, limit;
 315
 316             /* Declare 1.X samplers implicitly, based on the destination reg. number */
 317             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 318                 (WINED3DSIO_TEX == curOpcode->opcode ||
 319                  WINED3DSIO_TEXBEM == curOpcode->opcode ||
 320                  WINED3DSIO_TEXBEML == curOpcode->opcode ||
 321                  WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
 322                  WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
 323                  WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
 324                  WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
 325                  WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
 326                  WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
 327                  WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
 328                  WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
 329
 330                 /* Fake sampler usage, only set reserved bit and ttype */
 331                 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
 332
 333                 if(!stateBlock->textures[sampler_code]) {
 334                     ERR("No texture bound to sampler %d\n", sampler_code);
 335                     reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 336                 } else {
 337                     int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
 338                     switch(texType) {
 339                         case GL_TEXTURE_2D:
 340                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 341                             break;
 342
 343                         case GL_TEXTURE_3D:
 344                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
 345                             break;
 346
 347                         case GL_TEXTURE_CUBE_MAP_ARB:
 348                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
 349                             break;
 350
 351                         default:
 352                             ERR("Unexpected gl texture type found: %d\n", texType);
 353                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 354                     }
 355                 }
 356
 357                 /* texbem is only valid with < 1.4 pixel shaders */
 358                 if(WINED3DSIO_TEXBEM  == curOpcode->opcode ||
 359                     WINED3DSIO_TEXBEML == curOpcode->opcode) {
 360                     if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != sampler_code) {
 361                         FIXME("Pixel shader uses texbem instruction on more than 1 sampler\n");
 362                     } else {
 363                         reg_maps->bumpmat = sampler_code;
 364                         if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
 365                             reg_maps->luminanceparams = sampler_code;
 366                         }
 367                     }
 368                 }
 369             }
 370
 371             /* This will loop over all the registers and try to
 372              * make a bitmask of the ones we're interested in.
 373              *
 374              * Relative addressing tokens are ignored, but that's
 375              * okay, since we'll catch any address registers when
 376              * they are initialized (required by spec) */
 377
 378             limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
 379                 curOpcode->num_params + 1: curOpcode->num_params;
 380
 381             for (i = 0; i < limit; ++i) {
 382
 383                 DWORD param, addr_token, reg, regtype;
 384                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
 385
 386                 regtype = shader_get_regtype(param);
 387                 reg = param & WINED3DSP_REGNUM_MASK;
 388
 389                 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
 390
 391                     if (pshader)
 392                         reg_maps->texcoord[reg] = 1;
 393                     else
 394                         reg_maps->address[reg] = 1;
 395                 }
 396
 397                 else if (WINED3DSPR_TEMP == regtype)
 398                     reg_maps->temporary[reg] = 1;
 399
 400                 else if (WINED3DSPR_INPUT == regtype && !pshader)
 401                     reg_maps->attributes[reg] = 1;
 402
 403                 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
 404                     reg_maps->fog = 1;
 405              }
 406         }
 407     }
 408
 409     return WINED3D_OK;
 410 }
 411
 412 static void shader_dump_decl_usage(
 413     IWineD3DBaseShaderImpl* This,
 414     DWORD decl,
 415     DWORD param) {
 416
 417     DWORD regtype = shader_get_regtype(param);
 418
 419     TRACE("dcl");
 420
 421     if (regtype == WINED3DSPR_SAMPLER) {
 422         DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
 423
 424         switch (ttype) {
 425             case WINED3DSTT_2D: TRACE("_2d"); break;
 426             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 427             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 428             default: TRACE("_unknown_ttype(%08x)", ttype);
 429        }
 430
 431     } else {
 432
 433         DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
 434         DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
 435
 436         /* Pixel shaders 3.0 don't have usage semantics */
 437         char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 438         if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
 439             return;
 440         else
 441             TRACE("_");
 442
 443         switch(usage) {
 444         case WINED3DDECLUSAGE_POSITION:
 445             TRACE("position%d", idx);
 446             break;
 447         case WINED3DDECLUSAGE_BLENDINDICES:
 448             TRACE("blend");
 449             break;
 450         case WINED3DDECLUSAGE_BLENDWEIGHT:
 451             TRACE("weight");
 452             break;
 453         case WINED3DDECLUSAGE_NORMAL:
 454             TRACE("normal%d", idx);
 455             break;
 456         case WINED3DDECLUSAGE_PSIZE:
 457             TRACE("psize");
 458             break;
 459         case WINED3DDECLUSAGE_COLOR:
 460             if(idx == 0)  {
 461                 TRACE("color");
 462             } else {
 463                 TRACE("specular%d", (idx - 1));
 464             }
 465             break;
 466         case WINED3DDECLUSAGE_TEXCOORD:
 467             TRACE("texture%d", idx);
 468             break;
 469         case WINED3DDECLUSAGE_TANGENT:
 470             TRACE("tangent");
 471             break;
 472         case WINED3DDECLUSAGE_BINORMAL:
 473             TRACE("binormal");
 474             break;
 475         case WINED3DDECLUSAGE_TESSFACTOR:
 476             TRACE("tessfactor");
 477             break;
 478         case WINED3DDECLUSAGE_POSITIONT:
 479             TRACE("positionT%d", idx);
 480             break;
 481         case WINED3DDECLUSAGE_FOG:
 482             TRACE("fog");
 483             break;
 484         case WINED3DDECLUSAGE_DEPTH:
 485             TRACE("depth");
 486             break;
 487         case WINED3DDECLUSAGE_SAMPLE:
 488             TRACE("sample");
 489             break;
 490         default:
 491             FIXME("unknown_semantics(%08x)", usage);
 492         }
 493     }
 494 }
 495
 496 static void shader_dump_arr_entry(
 497     IWineD3DBaseShader *iface,
 498     const DWORD param,
 499     const DWORD addr_token,
 500     unsigned int reg,
 501     int input) {
 502
 503     char relative =
 504         ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
 505
 506     if (relative) {
 507         TRACE("[");
 508         if (addr_token)
 509             shader_dump_param(iface, addr_token, 0, input);
 510         else
 511             TRACE("a0.x");
 512         TRACE(" + ");
 513      }
 514      TRACE("%u", reg);
 515      if (relative)
 516          TRACE("]");
 517 }
 518
 519 void shader_dump_param(
 520     IWineD3DBaseShader *iface,
 521     const DWORD param,
 522     const DWORD addr_token,
 523     int input) {
 524
 525     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 526     static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
 527     static const char * const misctype_reg_names[] = { "vPos", "vFace"};
 528     char swizzle_reg_chars[4];
 529
 530     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 531     DWORD regtype = shader_get_regtype(param);
 532     DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
 533
 534     /* There are some minor differences between pixel and vertex shaders */
 535     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 536
 537     /* For one, we'd prefer color components to be shown for pshaders.
 538      * FIXME: use the swizzle function for this */
 539
 540     swizzle_reg_chars[0] = pshader? 'r': 'x';
 541     swizzle_reg_chars[1] = pshader? 'g': 'y';
 542     swizzle_reg_chars[2] = pshader? 'b': 'z';
 543     swizzle_reg_chars[3] = pshader? 'a': 'w';
 544
 545     if (input) {
 546         if ( (modifier == WINED3DSPSM_NEG) ||
 547              (modifier == WINED3DSPSM_BIASNEG) ||
 548              (modifier == WINED3DSPSM_SIGNNEG) ||
 549              (modifier == WINED3DSPSM_X2NEG) ||
 550              (modifier == WINED3DSPSM_ABSNEG) )
 551             TRACE("-");
 552         else if (modifier == WINED3DSPSM_COMP)
 553             TRACE("1-");
 554         else if (modifier == WINED3DSPSM_NOT)
 555             TRACE("!");
 556
 557         if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
 558             TRACE("abs(");
 559     }
 560
 561     switch (regtype) {
 562         case WINED3DSPR_TEMP:
 563             TRACE("r%u", reg);
 564             break;
 565         case WINED3DSPR_INPUT:
 566             TRACE("v");
 567             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 568             break;
 569         case WINED3DSPR_CONST:
 570         case WINED3DSPR_CONST2:
 571         case WINED3DSPR_CONST3:
 572         case WINED3DSPR_CONST4:
 573             TRACE("c");
 574             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
 575             break;
 576         case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
 577             TRACE("%c%u", (pshader? 't':'a'), reg);
 578             break;
 579         case WINED3DSPR_RASTOUT:
 580             TRACE("%s", rastout_reg_names[reg]);
 581             break;
 582         case WINED3DSPR_COLOROUT:
 583             TRACE("oC%u", reg);
 584             break;
 585         case WINED3DSPR_DEPTHOUT:
 586             TRACE("oDepth");
 587             break;
 588         case WINED3DSPR_ATTROUT:
 589             TRACE("oD%u", reg);
 590             break;
 591         case WINED3DSPR_TEXCRDOUT:
 592
 593             /* Vertex shaders >= 3.0 use general purpose output registers
 594              * (WINED3DSPR_OUTPUT), which can include an address token */
 595
 596             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
 597                 TRACE("o");
 598                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
 599             }
 600             else
 601                TRACE("oT%u", reg);
 602             break;
 603         case WINED3DSPR_CONSTINT:
 604             TRACE("i");
 605             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 606             break;
 607         case WINED3DSPR_CONSTBOOL:
 608             TRACE("b");
 609             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 610             break;
 611         case WINED3DSPR_LABEL:
 612             TRACE("l%u", reg);
 613             break;
 614         case WINED3DSPR_LOOP:
 615             TRACE("aL");
 616             break;
 617         case WINED3DSPR_SAMPLER:
 618             TRACE("s%u", reg);
 619             break;
 620         case WINED3DSPR_MISCTYPE:
 621             if (reg > 1) {
 622                 FIXME("Unhandled misctype register %d\n", reg);
 623             } else {
 624                 TRACE("%s", misctype_reg_names[reg]);
 625             }
 626             break;
 627         case WINED3DSPR_PREDICATE:
 628             TRACE("p%u", reg);
 629             break;
 630         default:
 631             TRACE("unhandled_rtype(%#x)", regtype);
 632             break;
 633    }
 634
 635    if (!input) {
 636        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
 637
 638        if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 639            TRACE(".");
 640            if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
 641            if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
 642            if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
 643            if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
 644        }
 645
 646    } else {
 647         /** operand input */
 648         DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 649         DWORD swizzle_r = swizzle & 0x03;
 650         DWORD swizzle_g = (swizzle >> 2) & 0x03;
 651         DWORD swizzle_b = (swizzle >> 4) & 0x03;
 652         DWORD swizzle_a = (swizzle >> 6) & 0x03;
 653
 654         if (0 != modifier) {
 655             switch (modifier) {
 656                 case WINED3DSPSM_NONE:    break;
 657                 case WINED3DSPSM_NEG:     break;
 658                 case WINED3DSPSM_NOT:     break;
 659                 case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
 660                 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
 661                 case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
 662                 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 663                 case WINED3DSPSM_COMP:    break;
 664                 case WINED3DSPSM_X2:      TRACE("_x2"); break;
 665                 case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
 666                 case WINED3DSPSM_DZ:      TRACE("_dz"); break;
 667                 case WINED3DSPSM_DW:      TRACE("_dw"); break;
 668                 case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
 669                 case WINED3DSPSM_ABS:     TRACE(")"); break;
 670                 default:
 671                     TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
 672             }
 673         }
 674
 675         /**
 676         * swizzle bits fields:
 677         *  RRGGBBAA
 678         */
 679         if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
 680             if (swizzle_r == swizzle_g &&
 681                 swizzle_r == swizzle_b &&
 682                 swizzle_r == swizzle_a) {
 683                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
 684             } else {
 685                 TRACE(".%c%c%c%c",
 686                 swizzle_reg_chars[swizzle_r],
 687                 swizzle_reg_chars[swizzle_g],
 688                 swizzle_reg_chars[swizzle_b],
 689                 swizzle_reg_chars[swizzle_a]);
 690             }
 691         }
 692     }
 693 }
 694
 695 /** Shared code in order to generate the bulk of the shader string.
 696     Use the shader_header_fct & shader_footer_fct to add strings
 697     that are specific to pixel or vertex functions
 698     NOTE: A description of how to parse tokens can be found at:
 699           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
 700 void shader_generate_main(
 701     IWineD3DBaseShader *iface,
 702     SHADER_BUFFER* buffer,
 703     shader_reg_maps* reg_maps,
 704     CONST DWORD* pFunction) {
 705
 706     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 707     const DWORD *pToken = pFunction;
 708     const SHADER_OPCODE *curOpcode = NULL;
 709     SHADER_HANDLER hw_fct = NULL;
 710     DWORD i;
 711     SHADER_OPCODE_ARG hw_arg;
 712
 713     /* Initialize current parsing state */
 714     hw_arg.shader = iface;
 715     hw_arg.buffer = buffer;
 716     hw_arg.reg_maps = reg_maps;
 717     This->baseShader.parse_state.current_row = 0;
 718
 719     /* Second pass, process opcodes */
 720     if (NULL != pToken) {
 721         while (WINED3DPS_END() != *pToken) {
 722
 723             /* Skip version token */
 724             if (shader_is_version_token(*pToken)) {
 725                 ++pToken;
 726                 continue;
 727             }
 728
 729             /* Skip comment tokens */
 730             if (shader_is_comment(*pToken)) {
 731                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 732                 ++pToken;
 733                 TRACE("#%s\n", (const char*)pToken);
 734                 pToken += comment_len;
 735                 continue;
 736             }
 737
 738             /* Read opcode */
 739             hw_arg.opcode_token = *pToken++;
 740             curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
 741
 742             /* Select handler */
 743             if (curOpcode == NULL)
 744                 hw_fct = NULL;
 745             else if (This->baseShader.shader_mode == SHADER_GLSL)
 746                 hw_fct = curOpcode->hw_glsl_fct;
 747             else if (This->baseShader.shader_mode == SHADER_ARB)
 748                 hw_fct = curOpcode->hw_fct;
 749
 750             /* Unknown opcode and its parameters */
 751             if (NULL == curOpcode) {
 752                 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
 753                 pToken += shader_skip_unrecognized(iface, pToken);
 754
 755             /* Nothing to do */
 756             } else if (WINED3DSIO_DCL == curOpcode->opcode ||
 757                        WINED3DSIO_NOP == curOpcode->opcode ||
 758                        WINED3DSIO_DEF == curOpcode->opcode ||
 759                        WINED3DSIO_DEFI == curOpcode->opcode ||
 760                        WINED3DSIO_DEFB == curOpcode->opcode ||
 761                        WINED3DSIO_PHASE == curOpcode->opcode ||
 762                        WINED3DSIO_RET == curOpcode->opcode) {
 763
 764                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 765
 766             /* If a generator function is set for current shader target, use it */
 767             } else if (hw_fct != NULL) {
 768
 769                 hw_arg.opcode = curOpcode;
 770
 771                 /* Destination token */
 772                 if (curOpcode->dst_token) {
 773
 774                     DWORD param, addr_token = 0;
 775                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 776                     hw_arg.dst = param;
 777                     hw_arg.dst_addr = addr_token;
 778                 }
 779
 780                 /* Predication token */
 781                 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
 782                     hw_arg.predicate = *pToken++;
 783
 784                 /* Other source tokens */
 785                 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
 786
 787                     DWORD param, addr_token = 0;
 788                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 789                     hw_arg.src[i] = param;
 790                     hw_arg.src_addr[i] = addr_token;
 791                 }
 792
 793                 /* Call appropriate function for output target */
 794                 hw_fct(&hw_arg);
 795
 796                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
 797                 if (This->baseShader.shader_mode == SHADER_GLSL)
 798                     shader_glsl_add_instruction_modifiers(&hw_arg);
 799
 800             /* Unhandled opcode */
 801             } else {
 802
 803                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
 804                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 805             }
 806         }
 807         /* TODO: What about result.depth? */
 808
 809     }
 810 }
 811
 812 void shader_dump_ins_modifiers(const DWORD output) {
 813
 814     DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 815     DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
 816
 817     switch (shift) {
 818         case 0: break;
 819         case 13: TRACE("_d8"); break;
 820         case 14: TRACE("_d4"); break;
 821         case 15: TRACE("_d2"); break;
 822         case 1: TRACE("_x2"); break;
 823         case 2: TRACE("_x4"); break;
 824         case 3: TRACE("_x8"); break;
 825         default: TRACE("_unhandled_shift(%d)", shift); break;
 826     }
 827
 828     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
 829     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
 830     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
 831
 832     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
 833     if (mmask)
 834         FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
 835 }
 836
 837 /* First pass: trace shader, initialize length and version */
 838 void shader_trace_init(
 839     IWineD3DBaseShader *iface,
 840     const DWORD* pFunction) {
 841
 842     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
 843
 844     const DWORD* pToken = pFunction;
 845     const SHADER_OPCODE* curOpcode = NULL;
 846     DWORD opcode_token;
 847     unsigned int len = 0;
 848     DWORD i;
 849
 850     TRACE("(%p) : Parsing programme\n", This);
 851
 852     if (NULL != pToken) {
 853         while (WINED3DVS_END() != *pToken) {
 854             if (shader_is_version_token(*pToken)) { /** version */
 855                 This->baseShader.hex_version = *pToken;
 856                 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
 857                     WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
 858                     WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
 859                 ++pToken;
 860                 ++len;
 861                 continue;
 862             }
 863             if (shader_is_comment(*pToken)) { /** comment */
 864                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 865                 ++pToken;
 866                 TRACE("//%s\n", (const char*)pToken);
 867                 pToken += comment_len;
 868                 len += comment_len + 1;
 869                 continue;
 870             }
 871             opcode_token = *pToken++;
 872             curOpcode = shader_get_opcode(iface, opcode_token);
 873             len++;
 874
 875             if (NULL == curOpcode) {
 876                 int tokens_read;
 877                 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
 878                 tokens_read = shader_skip_unrecognized(iface, pToken);
 879                 pToken += tokens_read;
 880                 len += tokens_read;
 881
 882             } else {
 883                 if (curOpcode->opcode == WINED3DSIO_DCL) {
 884
 885                     DWORD usage = *pToken;
 886                     DWORD param = *(pToken + 1);
 887
 888                     shader_dump_decl_usage(This, usage, param);
 889                     shader_dump_ins_modifiers(param);
 890                     TRACE(" ");
 891                     shader_dump_param(iface, param, 0, 0);
 892                     pToken += 2;
 893                     len += 2;
 894
 895                 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
 896
 897                         unsigned int offset = shader_get_float_offset(*pToken);
 898
 899                         TRACE("def c%u = %f, %f, %f, %f", offset,
 900                             *(const float *)(pToken + 1),
 901                             *(const float *)(pToken + 2),
 902                             *(const float *)(pToken + 3),
 903                             *(const float *)(pToken + 4));
 904
 905                         pToken += 5;
 906                         len += 5;
 907                 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
 908
 909                         TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
 910                             *(pToken + 1),
 911                             *(pToken + 2),
 912                             *(pToken + 3),
 913                             *(pToken + 4));
 914
 915                         pToken += 5;
 916                         len += 5;
 917
 918                 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
 919
 920                         TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
 921                             *(pToken + 1)? "true": "false");
 922
 923                         pToken += 2;
 924                         len += 2;
 925
 926                 } else {
 927
 928                     DWORD param, addr_token;
 929                     int tokens_read;
 930
 931                     /* Print out predication source token first - it follows
 932                      * the destination token. */
 933                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
 934                         TRACE("(");
 935                         shader_dump_param(iface, *(pToken + 2), 0, 1);
 936                         TRACE(") ");
 937                     }
 938
 939                     TRACE("%s", curOpcode->name);
 940
 941                     if (curOpcode->opcode == WINED3DSIO_IFC ||
 942                         curOpcode->opcode == WINED3DSIO_BREAKC) {
 943
 944                         DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
 945                         switch (op) {
 946                             case COMPARISON_GT: TRACE("_gt"); break;
 947                             case COMPARISON_EQ: TRACE("_eq"); break;
 948                             case COMPARISON_GE: TRACE("_ge"); break;
 949                             case COMPARISON_LT: TRACE("_lt"); break;
 950                             case COMPARISON_NE: TRACE("_ne"); break;
 951                             case COMPARISON_LE: TRACE("_le"); break;
 952                             default:
 953                                 TRACE("_(%u)", op);
 954                         }
 955                     } else if (curOpcode->opcode == WINED3DSIO_TEX &&
 956                                This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
 957                         if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
 958                     }
 959
 960                     /* Destination token */
 961                     if (curOpcode->dst_token) {
 962
 963                         /* Destination token */
 964                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 965                         pToken += tokens_read;
 966                         len += tokens_read;
 967
 968                         shader_dump_ins_modifiers(param);
 969                         TRACE(" ");
 970                         shader_dump_param(iface, param, addr_token, 0);
 971                     }
 972
 973                     /* Predication token - already printed out, just skip it */
 974                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
 975                         pToken++;
 976                         len++;
 977                     }
 978
 979                     /* Other source tokens */
 980                     for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
 981
 982                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 983                         pToken += tokens_read;
 984                         len += tokens_read;
 985
 986                         TRACE((i == 0)? " " : ", ");
 987                         shader_dump_param(iface, param, addr_token, 1);
 988                     }
 989                 }
 990                 TRACE("\n");
 991             }
 992         }
 993         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
 994     } else {
 995         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
 996     }
 997 }
 998
 999 void shader_delete_constant_list(
1000     struct list* clist) {
1001
1002     struct list *ptr;
1003     struct local_constant* constant;
1004
1005     ptr = list_head(clist);
1006     while (ptr) {
1007         constant = LIST_ENTRY(ptr, struct local_constant, entry);
1008         ptr = list_next(clist, ptr);
1009         HeapFree(GetProcessHeap(), 0, constant);
1010     }
1011 }
1012
1013 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1014 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1015 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1016 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1017
1018 const shader_backend_t none_shader_backend = {
1019     &shader_none_select,
1020     &shader_none_select_depth_blt,
1021     &shader_none_load_constants,
1022     &shader_none_cleanup
1023 };