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 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     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 197
 198     /* There are some minor differences between pixel and vertex shaders */
 199     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 200
 201     reg_maps->bumpmat = -1;
 202     reg_maps->luminanceparams = -1;
 203
 204     if (pToken == NULL)
 205         return WINED3D_OK;
 206
 207     while (WINED3DVS_END() != *pToken) {
 208         CONST SHADER_OPCODE* curOpcode;
 209         DWORD opcode_token;
 210
 211         /* Skip version */
 212         if (shader_is_version_token(*pToken)) {
 213              ++pToken;
 214              continue;
 215
 216         /* Skip comments */
 217         } else if (shader_is_comment(*pToken)) {
 218              DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 219              ++pToken;
 220              pToken += comment_len;
 221              continue;
 222         }
 223
 224         /* Fetch opcode */
 225         opcode_token = *pToken++;
 226         curOpcode = shader_get_opcode(iface, opcode_token);
 227
 228         /* Unhandled opcode, and its parameters */
 229         if (NULL == curOpcode) {
 230            while (*pToken & 0x80000000)
 231                ++pToken;
 232
 233         /* Handle declarations */
 234         } else if (WINED3DSIO_DCL == curOpcode->opcode) {
 235
 236             DWORD usage = *pToken++;
 237             DWORD param = *pToken++;
 238             DWORD regtype = shader_get_regtype(param);
 239             unsigned int regnum = param & WINED3DSP_REGNUM_MASK;
 240
 241             /* Vshader: mark attributes used
 242                Pshader: mark 3.0 input registers used, save token */
 243             if (WINED3DSPR_INPUT == regtype) {
 244
 245                 if (!pshader)
 246                     reg_maps->attributes[regnum] = 1;
 247                 else
 248                     reg_maps->packed_input[regnum] = 1;
 249
 250                 semantics_in[regnum].usage = usage;
 251                 semantics_in[regnum].reg = param;
 252
 253             /* Vshader: mark 3.0 output registers used, save token */
 254             } else if (WINED3DSPR_OUTPUT == regtype) {
 255                 reg_maps->packed_output[regnum] = 1;
 256                 semantics_out[regnum].usage = usage;
 257                 semantics_out[regnum].reg = param;
 258                 if (usage & (WINED3DDECLUSAGE_FOG << WINED3DSP_DCL_USAGE_SHIFT))
 259                     reg_maps->fog = 1;
 260
 261             /* Save sampler usage token */
 262             } else if (WINED3DSPR_SAMPLER == regtype)
 263                 reg_maps->samplers[regnum] = usage;
 264
 265         } else if (WINED3DSIO_DEF == curOpcode->opcode) {
 266
 267             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 268             if (!lconst) return E_OUTOFMEMORY;
 269             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 270             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 271
 272             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 273             if(WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 && pshader) {
 274                 float *value = (float *) lconst->value;
 275                 if(value[0] < -1.0) value[0] = -1.0;
 276                 else if(value[0] >  1.0) value[0] =  1.0;
 277                 if(value[1] < -1.0) value[1] = -1.0;
 278                 else if(value[1] >  1.0) value[1] =  1.0;
 279                 if(value[2] < -1.0) value[2] = -1.0;
 280                 else if(value[2] >  1.0) value[2] =  1.0;
 281                 if(value[3] < -1.0) value[3] = -1.0;
 282                 else if(value[3] >  1.0) value[3] =  1.0;
 283             }
 284
 285             list_add_head(&This->baseShader.constantsF, &lconst->entry);
 286             pToken += curOpcode->num_params;
 287
 288         } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
 289
 290             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 291             if (!lconst) return E_OUTOFMEMORY;
 292             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 293             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 294             list_add_head(&This->baseShader.constantsI, &lconst->entry);
 295             pToken += curOpcode->num_params;
 296
 297         } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
 298
 299             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 300             if (!lconst) return E_OUTOFMEMORY;
 301             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 302             memcpy(&lconst->value, pToken + 1, 1 * sizeof(DWORD));
 303             list_add_head(&This->baseShader.constantsB, &lconst->entry);
 304             pToken += curOpcode->num_params;
 305
 306         /* If there's a loop in the shader */
 307         } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
 308                    WINED3DSIO_REP == curOpcode->opcode) {
 309             cur_loop_depth++;
 310             if(cur_loop_depth > max_loop_depth)
 311                 max_loop_depth = cur_loop_depth;
 312             pToken += curOpcode->num_params;
 313
 314         } else if (WINED3DSIO_ENDLOOP == curOpcode->opcode ||
 315                    WINED3DSIO_ENDREP == curOpcode->opcode) {
 316             cur_loop_depth--;
 317
 318         /* For subroutine prototypes */
 319         } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
 320
 321             DWORD snum = *pToken & WINED3DSP_REGNUM_MASK;
 322             reg_maps->labels[snum] = 1;
 323             pToken += curOpcode->num_params;
 324
 325         /* Set texture, address, temporary registers */
 326         } else {
 327             int i, limit;
 328
 329             /* Declare 1.X samplers implicitly, based on the destination reg. number */
 330             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 331                 (WINED3DSIO_TEX == curOpcode->opcode ||
 332                  WINED3DSIO_TEXBEM == curOpcode->opcode ||
 333                  WINED3DSIO_TEXBEML == curOpcode->opcode ||
 334                  WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
 335                  WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
 336                  WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
 337                  WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
 338                  WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
 339                  WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
 340                  WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
 341                  WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
 342
 343                 /* Fake sampler usage, only set reserved bit and ttype */
 344                 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
 345
 346                 if(!stateBlock->textures[sampler_code]) {
 347                     ERR("No texture bound to sampler %d\n", sampler_code);
 348                     reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 349                 } else {
 350                     int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
 351                     switch(texType) {
 352                         case GL_TEXTURE_2D:
 353                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 354                             break;
 355
 356                         case GL_TEXTURE_3D:
 357                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
 358                             break;
 359
 360                         case GL_TEXTURE_CUBE_MAP_ARB:
 361                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
 362                             break;
 363
 364                         default:
 365                             ERR("Unexpected gl texture type found: %d\n", texType);
 366                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 367                     }
 368                 }
 369
 370                 /* texbem is only valid with < 1.4 pixel shaders */
 371                 if(WINED3DSIO_TEXBEM  == curOpcode->opcode ||
 372                     WINED3DSIO_TEXBEML == curOpcode->opcode) {
 373                     if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != sampler_code) {
 374                         FIXME("Pixel shader uses texbem instruction on more than 1 sampler\n");
 375                     } else {
 376                         reg_maps->bumpmat = sampler_code;
 377                         if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
 378                             reg_maps->luminanceparams = sampler_code;
 379                         }
 380                     }
 381                 }
 382             }
 383             if(WINED3DSIO_NRM  == curOpcode->opcode) {
 384                 reg_maps->usesnrm = 1;
 385             } else if(WINED3DSIO_BEM == curOpcode->opcode) {
 386                 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
 387                 if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != regnum) {
 388                     FIXME("Pixel shader uses bem or texbem instruction on more than 1 sampler\n");
 389                 } else {
 390                     reg_maps->bumpmat = regnum;
 391                 }
 392             }
 393
 394             /* This will loop over all the registers and try to
 395              * make a bitmask of the ones we're interested in.
 396              *
 397              * Relative addressing tokens are ignored, but that's
 398              * okay, since we'll catch any address registers when
 399              * they are initialized (required by spec) */
 400
 401             limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
 402                 curOpcode->num_params + 1: curOpcode->num_params;
 403
 404             for (i = 0; i < limit; ++i) {
 405
 406                 DWORD param, addr_token, reg, regtype;
 407                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
 408
 409                 regtype = shader_get_regtype(param);
 410                 reg = param & WINED3DSP_REGNUM_MASK;
 411
 412                 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
 413
 414                     if (pshader)
 415                         reg_maps->texcoord[reg] = 1;
 416                     else
 417                         reg_maps->address[reg] = 1;
 418                 }
 419
 420                 else if (WINED3DSPR_TEMP == regtype)
 421                     reg_maps->temporary[reg] = 1;
 422
 423                 else if (WINED3DSPR_INPUT == regtype && !pshader)
 424                     reg_maps->attributes[reg] = 1;
 425
 426                 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
 427                     reg_maps->fog = 1;
 428
 429                 else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
 430                     reg_maps->vpos = 1;
 431             }
 432         }
 433     }
 434     reg_maps->loop_depth = max_loop_depth;
 435
 436     return WINED3D_OK;
 437 }
 438
 439 static void shader_dump_decl_usage(
 440     IWineD3DBaseShaderImpl* This,
 441     DWORD decl,
 442     DWORD param) {
 443
 444     DWORD regtype = shader_get_regtype(param);
 445
 446     TRACE("dcl");
 447
 448     if (regtype == WINED3DSPR_SAMPLER) {
 449         DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
 450
 451         switch (ttype) {
 452             case WINED3DSTT_2D: TRACE("_2d"); break;
 453             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 454             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 455             default: TRACE("_unknown_ttype(%08x)", ttype);
 456        }
 457
 458     } else {
 459
 460         DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
 461         DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
 462
 463         /* Pixel shaders 3.0 don't have usage semantics */
 464         char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 465         if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
 466             return;
 467         else
 468             TRACE("_");
 469
 470         switch(usage) {
 471         case WINED3DDECLUSAGE_POSITION:
 472             TRACE("position%d", idx);
 473             break;
 474         case WINED3DDECLUSAGE_BLENDINDICES:
 475             TRACE("blend");
 476             break;
 477         case WINED3DDECLUSAGE_BLENDWEIGHT:
 478             TRACE("weight");
 479             break;
 480         case WINED3DDECLUSAGE_NORMAL:
 481             TRACE("normal%d", idx);
 482             break;
 483         case WINED3DDECLUSAGE_PSIZE:
 484             TRACE("psize");
 485             break;
 486         case WINED3DDECLUSAGE_COLOR:
 487             if(idx == 0)  {
 488                 TRACE("color");
 489             } else {
 490                 TRACE("specular%d", (idx - 1));
 491             }
 492             break;
 493         case WINED3DDECLUSAGE_TEXCOORD:
 494             TRACE("texture%d", idx);
 495             break;
 496         case WINED3DDECLUSAGE_TANGENT:
 497             TRACE("tangent");
 498             break;
 499         case WINED3DDECLUSAGE_BINORMAL:
 500             TRACE("binormal");
 501             break;
 502         case WINED3DDECLUSAGE_TESSFACTOR:
 503             TRACE("tessfactor");
 504             break;
 505         case WINED3DDECLUSAGE_POSITIONT:
 506             TRACE("positionT%d", idx);
 507             break;
 508         case WINED3DDECLUSAGE_FOG:
 509             TRACE("fog");
 510             break;
 511         case WINED3DDECLUSAGE_DEPTH:
 512             TRACE("depth");
 513             break;
 514         case WINED3DDECLUSAGE_SAMPLE:
 515             TRACE("sample");
 516             break;
 517         default:
 518             FIXME("unknown_semantics(%08x)", usage);
 519         }
 520     }
 521 }
 522
 523 static void shader_dump_arr_entry(
 524     IWineD3DBaseShader *iface,
 525     const DWORD param,
 526     const DWORD addr_token,
 527     unsigned int reg,
 528     int input) {
 529
 530     char relative =
 531         ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
 532
 533     if (relative) {
 534         TRACE("[");
 535         if (addr_token)
 536             shader_dump_param(iface, addr_token, 0, input);
 537         else
 538             TRACE("a0.x");
 539         TRACE(" + ");
 540      }
 541      TRACE("%u", reg);
 542      if (relative)
 543          TRACE("]");
 544 }
 545
 546 void shader_dump_param(
 547     IWineD3DBaseShader *iface,
 548     const DWORD param,
 549     const DWORD addr_token,
 550     int input) {
 551
 552     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 553     static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
 554     static const char * const misctype_reg_names[] = { "vPos", "vFace"};
 555     char swizzle_reg_chars[4];
 556
 557     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 558     DWORD regtype = shader_get_regtype(param);
 559     DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
 560
 561     /* There are some minor differences between pixel and vertex shaders */
 562     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 563
 564     /* For one, we'd prefer color components to be shown for pshaders.
 565      * FIXME: use the swizzle function for this */
 566
 567     swizzle_reg_chars[0] = pshader? 'r': 'x';
 568     swizzle_reg_chars[1] = pshader? 'g': 'y';
 569     swizzle_reg_chars[2] = pshader? 'b': 'z';
 570     swizzle_reg_chars[3] = pshader? 'a': 'w';
 571
 572     if (input) {
 573         if ( (modifier == WINED3DSPSM_NEG) ||
 574              (modifier == WINED3DSPSM_BIASNEG) ||
 575              (modifier == WINED3DSPSM_SIGNNEG) ||
 576              (modifier == WINED3DSPSM_X2NEG) ||
 577              (modifier == WINED3DSPSM_ABSNEG) )
 578             TRACE("-");
 579         else if (modifier == WINED3DSPSM_COMP)
 580             TRACE("1-");
 581         else if (modifier == WINED3DSPSM_NOT)
 582             TRACE("!");
 583
 584         if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
 585             TRACE("abs(");
 586     }
 587
 588     switch (regtype) {
 589         case WINED3DSPR_TEMP:
 590             TRACE("r%u", reg);
 591             break;
 592         case WINED3DSPR_INPUT:
 593             TRACE("v");
 594             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 595             break;
 596         case WINED3DSPR_CONST:
 597         case WINED3DSPR_CONST2:
 598         case WINED3DSPR_CONST3:
 599         case WINED3DSPR_CONST4:
 600             TRACE("c");
 601             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
 602             break;
 603         case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
 604             TRACE("%c%u", (pshader? 't':'a'), reg);
 605             break;
 606         case WINED3DSPR_RASTOUT:
 607             TRACE("%s", rastout_reg_names[reg]);
 608             break;
 609         case WINED3DSPR_COLOROUT:
 610             TRACE("oC%u", reg);
 611             break;
 612         case WINED3DSPR_DEPTHOUT:
 613             TRACE("oDepth");
 614             break;
 615         case WINED3DSPR_ATTROUT:
 616             TRACE("oD%u", reg);
 617             break;
 618         case WINED3DSPR_TEXCRDOUT:
 619
 620             /* Vertex shaders >= 3.0 use general purpose output registers
 621              * (WINED3DSPR_OUTPUT), which can include an address token */
 622
 623             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
 624                 TRACE("o");
 625                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
 626             }
 627             else
 628                TRACE("oT%u", reg);
 629             break;
 630         case WINED3DSPR_CONSTINT:
 631             TRACE("i");
 632             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 633             break;
 634         case WINED3DSPR_CONSTBOOL:
 635             TRACE("b");
 636             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 637             break;
 638         case WINED3DSPR_LABEL:
 639             TRACE("l%u", reg);
 640             break;
 641         case WINED3DSPR_LOOP:
 642             TRACE("aL");
 643             break;
 644         case WINED3DSPR_SAMPLER:
 645             TRACE("s%u", reg);
 646             break;
 647         case WINED3DSPR_MISCTYPE:
 648             if (reg > 1) {
 649                 FIXME("Unhandled misctype register %d\n", reg);
 650             } else {
 651                 TRACE("%s", misctype_reg_names[reg]);
 652             }
 653             break;
 654         case WINED3DSPR_PREDICATE:
 655             TRACE("p%u", reg);
 656             break;
 657         default:
 658             TRACE("unhandled_rtype(%#x)", regtype);
 659             break;
 660    }
 661
 662    if (!input) {
 663        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
 664
 665        if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 666            TRACE(".");
 667            if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
 668            if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
 669            if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
 670            if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
 671        }
 672
 673    } else {
 674         /** operand input */
 675         DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 676         DWORD swizzle_r = swizzle & 0x03;
 677         DWORD swizzle_g = (swizzle >> 2) & 0x03;
 678         DWORD swizzle_b = (swizzle >> 4) & 0x03;
 679         DWORD swizzle_a = (swizzle >> 6) & 0x03;
 680
 681         if (0 != modifier) {
 682             switch (modifier) {
 683                 case WINED3DSPSM_NONE:    break;
 684                 case WINED3DSPSM_NEG:     break;
 685                 case WINED3DSPSM_NOT:     break;
 686                 case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
 687                 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
 688                 case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
 689                 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 690                 case WINED3DSPSM_COMP:    break;
 691                 case WINED3DSPSM_X2:      TRACE("_x2"); break;
 692                 case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
 693                 case WINED3DSPSM_DZ:      TRACE("_dz"); break;
 694                 case WINED3DSPSM_DW:      TRACE("_dw"); break;
 695                 case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
 696                 case WINED3DSPSM_ABS:     TRACE(")"); break;
 697                 default:
 698                     TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
 699             }
 700         }
 701
 702         /**
 703         * swizzle bits fields:
 704         *  RRGGBBAA
 705         */
 706         if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
 707             if (swizzle_r == swizzle_g &&
 708                 swizzle_r == swizzle_b &&
 709                 swizzle_r == swizzle_a) {
 710                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
 711             } else {
 712                 TRACE(".%c%c%c%c",
 713                 swizzle_reg_chars[swizzle_r],
 714                 swizzle_reg_chars[swizzle_g],
 715                 swizzle_reg_chars[swizzle_b],
 716                 swizzle_reg_chars[swizzle_a]);
 717             }
 718         }
 719     }
 720 }
 721
 722 /** Shared code in order to generate the bulk of the shader string.
 723     Use the shader_header_fct & shader_footer_fct to add strings
 724     that are specific to pixel or vertex functions
 725     NOTE: A description of how to parse tokens can be found at:
 726           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
 727 void shader_generate_main(
 728     IWineD3DBaseShader *iface,
 729     SHADER_BUFFER* buffer,
 730     shader_reg_maps* reg_maps,
 731     CONST DWORD* pFunction) {
 732
 733     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 734     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
 735     const DWORD *pToken = pFunction;
 736     const SHADER_OPCODE *curOpcode = NULL;
 737     SHADER_HANDLER hw_fct = NULL;
 738     DWORD i;
 739     SHADER_OPCODE_ARG hw_arg;
 740
 741     /* Initialize current parsing state */
 742     hw_arg.shader = iface;
 743     hw_arg.buffer = buffer;
 744     hw_arg.reg_maps = reg_maps;
 745     This->baseShader.parse_state.current_row = 0;
 746
 747     /* Second pass, process opcodes */
 748     if (NULL != pToken) {
 749         while (WINED3DPS_END() != *pToken) {
 750
 751             /* Skip version token */
 752             if (shader_is_version_token(*pToken)) {
 753                 ++pToken;
 754                 continue;
 755             }
 756
 757             /* Skip comment tokens */
 758             if (shader_is_comment(*pToken)) {
 759                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 760                 ++pToken;
 761                 TRACE("#%s\n", (const char*)pToken);
 762                 pToken += comment_len;
 763                 continue;
 764             }
 765
 766             /* Read opcode */
 767             hw_arg.opcode_token = *pToken++;
 768             curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
 769
 770             /* Select handler */
 771             if (curOpcode == NULL)
 772                 hw_fct = NULL;
 773             else if (This->baseShader.shader_mode == SHADER_GLSL)
 774                 hw_fct = curOpcode->hw_glsl_fct;
 775             else if (This->baseShader.shader_mode == SHADER_ARB)
 776                 hw_fct = curOpcode->hw_fct;
 777
 778             /* Unknown opcode and its parameters */
 779             if (NULL == curOpcode) {
 780                 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
 781                 pToken += shader_skip_unrecognized(iface, pToken);
 782
 783             /* Nothing to do */
 784             } else if (WINED3DSIO_DCL == curOpcode->opcode ||
 785                        WINED3DSIO_NOP == curOpcode->opcode ||
 786                        WINED3DSIO_DEF == curOpcode->opcode ||
 787                        WINED3DSIO_DEFI == curOpcode->opcode ||
 788                        WINED3DSIO_DEFB == curOpcode->opcode ||
 789                        WINED3DSIO_PHASE == curOpcode->opcode ||
 790                        WINED3DSIO_RET == curOpcode->opcode) {
 791
 792                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 793
 794             /* If a generator function is set for current shader target, use it */
 795             } else if (hw_fct != NULL) {
 796
 797                 hw_arg.opcode = curOpcode;
 798
 799                 /* Destination token */
 800                 if (curOpcode->dst_token) {
 801
 802                     DWORD param, addr_token = 0;
 803                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 804                     hw_arg.dst = param;
 805                     hw_arg.dst_addr = addr_token;
 806                 }
 807
 808                 /* Predication token */
 809                 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
 810                     hw_arg.predicate = *pToken++;
 811
 812                 /* Other source tokens */
 813                 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
 814
 815                     DWORD param, addr_token = 0;
 816                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 817                     hw_arg.src[i] = param;
 818                     hw_arg.src_addr[i] = addr_token;
 819                 }
 820
 821                 /* Call appropriate function for output target */
 822                 hw_fct(&hw_arg);
 823
 824                 /* Add color correction if needed */
 825                 device->shader_backend->shader_color_correction(&hw_arg);
 826
 827                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
 828                 if (This->baseShader.shader_mode == SHADER_GLSL)
 829                     shader_glsl_add_instruction_modifiers(&hw_arg);
 830
 831             /* Unhandled opcode */
 832             } else {
 833
 834                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
 835                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 836             }
 837         }
 838         /* TODO: What about result.depth? */
 839
 840     }
 841 }
 842
 843 void shader_dump_ins_modifiers(const DWORD output) {
 844
 845     DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 846     DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
 847
 848     switch (shift) {
 849         case 0: break;
 850         case 13: TRACE("_d8"); break;
 851         case 14: TRACE("_d4"); break;
 852         case 15: TRACE("_d2"); break;
 853         case 1: TRACE("_x2"); break;
 854         case 2: TRACE("_x4"); break;
 855         case 3: TRACE("_x8"); break;
 856         default: TRACE("_unhandled_shift(%d)", shift); break;
 857     }
 858
 859     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
 860     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
 861     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
 862
 863     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
 864     if (mmask)
 865         FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
 866 }
 867
 868 /* First pass: trace shader, initialize length and version */
 869 void shader_trace_init(
 870     IWineD3DBaseShader *iface,
 871     const DWORD* pFunction) {
 872
 873     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
 874
 875     const DWORD* pToken = pFunction;
 876     const SHADER_OPCODE* curOpcode = NULL;
 877     DWORD opcode_token;
 878     unsigned int len = 0;
 879     DWORD i;
 880
 881     TRACE("(%p) : Parsing programme\n", This);
 882
 883     if (NULL != pToken) {
 884         while (WINED3DVS_END() != *pToken) {
 885             if (shader_is_version_token(*pToken)) { /** version */
 886                 This->baseShader.hex_version = *pToken;
 887                 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
 888                     WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
 889                     WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
 890                 ++pToken;
 891                 ++len;
 892                 continue;
 893             }
 894             if (shader_is_comment(*pToken)) { /** comment */
 895                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 896                 ++pToken;
 897                 TRACE("//%s\n", (const char*)pToken);
 898                 pToken += comment_len;
 899                 len += comment_len + 1;
 900                 continue;
 901             }
 902             opcode_token = *pToken++;
 903             curOpcode = shader_get_opcode(iface, opcode_token);
 904             len++;
 905
 906             if (NULL == curOpcode) {
 907                 int tokens_read;
 908                 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
 909                 tokens_read = shader_skip_unrecognized(iface, pToken);
 910                 pToken += tokens_read;
 911                 len += tokens_read;
 912
 913             } else {
 914                 if (curOpcode->opcode == WINED3DSIO_DCL) {
 915
 916                     DWORD usage = *pToken;
 917                     DWORD param = *(pToken + 1);
 918
 919                     shader_dump_decl_usage(This, usage, param);
 920                     shader_dump_ins_modifiers(param);
 921                     TRACE(" ");
 922                     shader_dump_param(iface, param, 0, 0);
 923                     pToken += 2;
 924                     len += 2;
 925
 926                 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
 927
 928                         unsigned int offset = shader_get_float_offset(*pToken);
 929
 930                         TRACE("def c%u = %f, %f, %f, %f", offset,
 931                             *(const float *)(pToken + 1),
 932                             *(const float *)(pToken + 2),
 933                             *(const float *)(pToken + 3),
 934                             *(const float *)(pToken + 4));
 935
 936                         pToken += 5;
 937                         len += 5;
 938                 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
 939
 940                         TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
 941                             *(pToken + 1),
 942                             *(pToken + 2),
 943                             *(pToken + 3),
 944                             *(pToken + 4));
 945
 946                         pToken += 5;
 947                         len += 5;
 948
 949                 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
 950
 951                         TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
 952                             *(pToken + 1)? "true": "false");
 953
 954                         pToken += 2;
 955                         len += 2;
 956
 957                 } else {
 958
 959                     DWORD param, addr_token;
 960                     int tokens_read;
 961
 962                     /* Print out predication source token first - it follows
 963                      * the destination token. */
 964                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
 965                         TRACE("(");
 966                         shader_dump_param(iface, *(pToken + 2), 0, 1);
 967                         TRACE(") ");
 968                     }
 969                     if (opcode_token & WINED3DSI_COISSUE) {
 970                         /* PixWin marks instructions with the coissue flag with a '+' */
 971                         TRACE("+");
 972                     }
 973
 974                     TRACE("%s", curOpcode->name);
 975
 976                     if (curOpcode->opcode == WINED3DSIO_IFC ||
 977                         curOpcode->opcode == WINED3DSIO_BREAKC) {
 978
 979                         DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
 980                         switch (op) {
 981                             case COMPARISON_GT: TRACE("_gt"); break;
 982                             case COMPARISON_EQ: TRACE("_eq"); break;
 983                             case COMPARISON_GE: TRACE("_ge"); break;
 984                             case COMPARISON_LT: TRACE("_lt"); break;
 985                             case COMPARISON_NE: TRACE("_ne"); break;
 986                             case COMPARISON_LE: TRACE("_le"); break;
 987                             default:
 988                                 TRACE("_(%u)", op);
 989                         }
 990                     } else if (curOpcode->opcode == WINED3DSIO_TEX &&
 991                                This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
 992                         if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
 993                     }
 994
 995                     /* Destination token */
 996                     if (curOpcode->dst_token) {
 997
 998                         /* Destination token */
 999                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1000                         pToken += tokens_read;
1001                         len += tokens_read;
1002
1003                         shader_dump_ins_modifiers(param);
1004                         TRACE(" ");
1005                         shader_dump_param(iface, param, addr_token, 0);
1006                     }
1007
1008                     /* Predication token - already printed out, just skip it */
1009                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1010                         pToken++;
1011                         len++;
1012                     }
1013
1014                     /* Other source tokens */
1015                     for (i = curOpcode->dst_token; 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((i == 0)? " " : ", ");
1022                         shader_dump_param(iface, param, addr_token, 1);
1023                     }
1024                 }
1025                 TRACE("\n");
1026             }
1027         }
1028         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1029     } else {
1030         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1031     }
1032 }
1033
1034 void shader_delete_constant_list(
1035     struct list* clist) {
1036
1037     struct list *ptr;
1038     struct local_constant* constant;
1039
1040     ptr = list_head(clist);
1041     while (ptr) {
1042         constant = LIST_ENTRY(ptr, struct local_constant, entry);
1043         ptr = list_next(clist, ptr);
1044         HeapFree(GetProcessHeap(), 0, constant);
1045     }
1046 }
1047
1048 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1049 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1050 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1051 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1052
1053 const shader_backend_t none_shader_backend = {
1054     &shader_none_select,
1055     &shader_none_select_depth_blt,
1056     &shader_none_load_constants,
1057     &shader_none_cleanup
1058 };