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
 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
 271             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 272             if(WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 && pshader) {
 273                 float *value = (float *) lconst->value;
 274                 if(value[0] < -1.0) value[0] = -1.0;
 275                 else if(value[0] >  1.0) value[0] =  1.0;
 276                 if(value[1] < -1.0) value[1] = -1.0;
 277                 else if(value[1] >  1.0) value[1] =  1.0;
 278                 if(value[2] < -1.0) value[2] = -1.0;
 279                 else if(value[2] >  1.0) value[2] =  1.0;
 280                 if(value[3] < -1.0) value[3] = -1.0;
 281                 else if(value[3] >  1.0) value[3] =  1.0;
 282             }
 283
 284             list_add_head(&This->baseShader.constantsF, &lconst->entry);
 285             pToken += curOpcode->num_params;
 286
 287         } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
 288
 289             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 290             if (!lconst) return E_OUTOFMEMORY;
 291             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 292             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 293             list_add_head(&This->baseShader.constantsI, &lconst->entry);
 294             pToken += curOpcode->num_params;
 295
 296         } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
 297
 298             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 299             if (!lconst) return E_OUTOFMEMORY;
 300             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
 301             memcpy(&lconst->value, pToken + 1, 1 * sizeof(DWORD));
 302             list_add_head(&This->baseShader.constantsB, &lconst->entry);
 303             pToken += curOpcode->num_params;
 304
 305         /* If there's a loop in the shader */
 306         } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
 307                    WINED3DSIO_REP == curOpcode->opcode) {
 308             reg_maps->loop = 1;
 309             pToken += curOpcode->num_params;
 310
 311         /* For subroutine prototypes */
 312         } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
 313
 314             DWORD snum = *pToken & WINED3DSP_REGNUM_MASK;
 315             reg_maps->labels[snum] = 1;
 316             pToken += curOpcode->num_params;
 317
 318         /* Set texture, address, temporary registers */
 319         } else {
 320             int i, limit;
 321
 322             /* Declare 1.X samplers implicitly, based on the destination reg. number */
 323             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 324                 (WINED3DSIO_TEX == curOpcode->opcode ||
 325                  WINED3DSIO_TEXBEM == curOpcode->opcode ||
 326                  WINED3DSIO_TEXBEML == curOpcode->opcode ||
 327                  WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
 328                  WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
 329                  WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
 330                  WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
 331                  WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
 332                  WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
 333                  WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
 334                  WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
 335
 336                 /* Fake sampler usage, only set reserved bit and ttype */
 337                 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
 338
 339                 if(!stateBlock->textures[sampler_code]) {
 340                     ERR("No texture bound to sampler %d\n", sampler_code);
 341                     reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 342                 } else {
 343                     int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
 344                     switch(texType) {
 345                         case GL_TEXTURE_2D:
 346                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 347                             break;
 348
 349                         case GL_TEXTURE_3D:
 350                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
 351                             break;
 352
 353                         case GL_TEXTURE_CUBE_MAP_ARB:
 354                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
 355                             break;
 356
 357                         default:
 358                             ERR("Unexpected gl texture type found: %d\n", texType);
 359                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 360                     }
 361                 }
 362
 363                 /* texbem is only valid with < 1.4 pixel shaders */
 364                 if(WINED3DSIO_TEXBEM  == curOpcode->opcode ||
 365                     WINED3DSIO_TEXBEML == curOpcode->opcode) {
 366                     if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != sampler_code) {
 367                         FIXME("Pixel shader uses texbem instruction on more than 1 sampler\n");
 368                     } else {
 369                         reg_maps->bumpmat = sampler_code;
 370                         if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
 371                             reg_maps->luminanceparams = sampler_code;
 372                         }
 373                     }
 374                 }
 375             }
 376             if(WINED3DSIO_NRM  == curOpcode->opcode) {
 377                 reg_maps->usesnrm = 1;
 378             } else if(WINED3DSIO_BEM == curOpcode->opcode) {
 379                 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
 380                 if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != regnum) {
 381                     FIXME("Pixel shader uses bem or texbem instruction on more than 1 sampler\n");
 382                 } else {
 383                     reg_maps->bumpmat = regnum;
 384                 }
 385             }
 386
 387             /* This will loop over all the registers and try to
 388              * make a bitmask of the ones we're interested in.
 389              *
 390              * Relative addressing tokens are ignored, but that's
 391              * okay, since we'll catch any address registers when
 392              * they are initialized (required by spec) */
 393
 394             limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
 395                 curOpcode->num_params + 1: curOpcode->num_params;
 396
 397             for (i = 0; i < limit; ++i) {
 398
 399                 DWORD param, addr_token, reg, regtype;
 400                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
 401
 402                 regtype = shader_get_regtype(param);
 403                 reg = param & WINED3DSP_REGNUM_MASK;
 404
 405                 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
 406
 407                     if (pshader)
 408                         reg_maps->texcoord[reg] = 1;
 409                     else
 410                         reg_maps->address[reg] = 1;
 411                 }
 412
 413                 else if (WINED3DSPR_TEMP == regtype)
 414                     reg_maps->temporary[reg] = 1;
 415
 416                 else if (WINED3DSPR_INPUT == regtype && !pshader)
 417                     reg_maps->attributes[reg] = 1;
 418
 419                 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
 420                     reg_maps->fog = 1;
 421             }
 422         }
 423     }
 424
 425     return WINED3D_OK;
 426 }
 427
 428 static void shader_dump_decl_usage(
 429     IWineD3DBaseShaderImpl* This,
 430     DWORD decl,
 431     DWORD param) {
 432
 433     DWORD regtype = shader_get_regtype(param);
 434
 435     TRACE("dcl");
 436
 437     if (regtype == WINED3DSPR_SAMPLER) {
 438         DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
 439
 440         switch (ttype) {
 441             case WINED3DSTT_2D: TRACE("_2d"); break;
 442             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 443             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 444             default: TRACE("_unknown_ttype(%08x)", ttype);
 445        }
 446
 447     } else {
 448
 449         DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
 450         DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
 451
 452         /* Pixel shaders 3.0 don't have usage semantics */
 453         char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 454         if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
 455             return;
 456         else
 457             TRACE("_");
 458
 459         switch(usage) {
 460         case WINED3DDECLUSAGE_POSITION:
 461             TRACE("position%d", idx);
 462             break;
 463         case WINED3DDECLUSAGE_BLENDINDICES:
 464             TRACE("blend");
 465             break;
 466         case WINED3DDECLUSAGE_BLENDWEIGHT:
 467             TRACE("weight");
 468             break;
 469         case WINED3DDECLUSAGE_NORMAL:
 470             TRACE("normal%d", idx);
 471             break;
 472         case WINED3DDECLUSAGE_PSIZE:
 473             TRACE("psize");
 474             break;
 475         case WINED3DDECLUSAGE_COLOR:
 476             if(idx == 0)  {
 477                 TRACE("color");
 478             } else {
 479                 TRACE("specular%d", (idx - 1));
 480             }
 481             break;
 482         case WINED3DDECLUSAGE_TEXCOORD:
 483             TRACE("texture%d", idx);
 484             break;
 485         case WINED3DDECLUSAGE_TANGENT:
 486             TRACE("tangent");
 487             break;
 488         case WINED3DDECLUSAGE_BINORMAL:
 489             TRACE("binormal");
 490             break;
 491         case WINED3DDECLUSAGE_TESSFACTOR:
 492             TRACE("tessfactor");
 493             break;
 494         case WINED3DDECLUSAGE_POSITIONT:
 495             TRACE("positionT%d", idx);
 496             break;
 497         case WINED3DDECLUSAGE_FOG:
 498             TRACE("fog");
 499             break;
 500         case WINED3DDECLUSAGE_DEPTH:
 501             TRACE("depth");
 502             break;
 503         case WINED3DDECLUSAGE_SAMPLE:
 504             TRACE("sample");
 505             break;
 506         default:
 507             FIXME("unknown_semantics(%08x)", usage);
 508         }
 509     }
 510 }
 511
 512 static void shader_dump_arr_entry(
 513     IWineD3DBaseShader *iface,
 514     const DWORD param,
 515     const DWORD addr_token,
 516     unsigned int reg,
 517     int input) {
 518
 519     char relative =
 520         ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
 521
 522     if (relative) {
 523         TRACE("[");
 524         if (addr_token)
 525             shader_dump_param(iface, addr_token, 0, input);
 526         else
 527             TRACE("a0.x");
 528         TRACE(" + ");
 529      }
 530      TRACE("%u", reg);
 531      if (relative)
 532          TRACE("]");
 533 }
 534
 535 void shader_dump_param(
 536     IWineD3DBaseShader *iface,
 537     const DWORD param,
 538     const DWORD addr_token,
 539     int input) {
 540
 541     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 542     static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
 543     static const char * const misctype_reg_names[] = { "vPos", "vFace"};
 544     char swizzle_reg_chars[4];
 545
 546     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 547     DWORD regtype = shader_get_regtype(param);
 548     DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
 549
 550     /* There are some minor differences between pixel and vertex shaders */
 551     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 552
 553     /* For one, we'd prefer color components to be shown for pshaders.
 554      * FIXME: use the swizzle function for this */
 555
 556     swizzle_reg_chars[0] = pshader? 'r': 'x';
 557     swizzle_reg_chars[1] = pshader? 'g': 'y';
 558     swizzle_reg_chars[2] = pshader? 'b': 'z';
 559     swizzle_reg_chars[3] = pshader? 'a': 'w';
 560
 561     if (input) {
 562         if ( (modifier == WINED3DSPSM_NEG) ||
 563              (modifier == WINED3DSPSM_BIASNEG) ||
 564              (modifier == WINED3DSPSM_SIGNNEG) ||
 565              (modifier == WINED3DSPSM_X2NEG) ||
 566              (modifier == WINED3DSPSM_ABSNEG) )
 567             TRACE("-");
 568         else if (modifier == WINED3DSPSM_COMP)
 569             TRACE("1-");
 570         else if (modifier == WINED3DSPSM_NOT)
 571             TRACE("!");
 572
 573         if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
 574             TRACE("abs(");
 575     }
 576
 577     switch (regtype) {
 578         case WINED3DSPR_TEMP:
 579             TRACE("r%u", reg);
 580             break;
 581         case WINED3DSPR_INPUT:
 582             TRACE("v");
 583             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 584             break;
 585         case WINED3DSPR_CONST:
 586         case WINED3DSPR_CONST2:
 587         case WINED3DSPR_CONST3:
 588         case WINED3DSPR_CONST4:
 589             TRACE("c");
 590             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
 591             break;
 592         case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
 593             TRACE("%c%u", (pshader? 't':'a'), reg);
 594             break;
 595         case WINED3DSPR_RASTOUT:
 596             TRACE("%s", rastout_reg_names[reg]);
 597             break;
 598         case WINED3DSPR_COLOROUT:
 599             TRACE("oC%u", reg);
 600             break;
 601         case WINED3DSPR_DEPTHOUT:
 602             TRACE("oDepth");
 603             break;
 604         case WINED3DSPR_ATTROUT:
 605             TRACE("oD%u", reg);
 606             break;
 607         case WINED3DSPR_TEXCRDOUT:
 608
 609             /* Vertex shaders >= 3.0 use general purpose output registers
 610              * (WINED3DSPR_OUTPUT), which can include an address token */
 611
 612             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
 613                 TRACE("o");
 614                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
 615             }
 616             else
 617                TRACE("oT%u", reg);
 618             break;
 619         case WINED3DSPR_CONSTINT:
 620             TRACE("i");
 621             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 622             break;
 623         case WINED3DSPR_CONSTBOOL:
 624             TRACE("b");
 625             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 626             break;
 627         case WINED3DSPR_LABEL:
 628             TRACE("l%u", reg);
 629             break;
 630         case WINED3DSPR_LOOP:
 631             TRACE("aL");
 632             break;
 633         case WINED3DSPR_SAMPLER:
 634             TRACE("s%u", reg);
 635             break;
 636         case WINED3DSPR_MISCTYPE:
 637             if (reg > 1) {
 638                 FIXME("Unhandled misctype register %d\n", reg);
 639             } else {
 640                 TRACE("%s", misctype_reg_names[reg]);
 641             }
 642             break;
 643         case WINED3DSPR_PREDICATE:
 644             TRACE("p%u", reg);
 645             break;
 646         default:
 647             TRACE("unhandled_rtype(%#x)", regtype);
 648             break;
 649    }
 650
 651    if (!input) {
 652        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
 653
 654        if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 655            TRACE(".");
 656            if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
 657            if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
 658            if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
 659            if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
 660        }
 661
 662    } else {
 663         /** operand input */
 664         DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 665         DWORD swizzle_r = swizzle & 0x03;
 666         DWORD swizzle_g = (swizzle >> 2) & 0x03;
 667         DWORD swizzle_b = (swizzle >> 4) & 0x03;
 668         DWORD swizzle_a = (swizzle >> 6) & 0x03;
 669
 670         if (0 != modifier) {
 671             switch (modifier) {
 672                 case WINED3DSPSM_NONE:    break;
 673                 case WINED3DSPSM_NEG:     break;
 674                 case WINED3DSPSM_NOT:     break;
 675                 case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
 676                 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
 677                 case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
 678                 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 679                 case WINED3DSPSM_COMP:    break;
 680                 case WINED3DSPSM_X2:      TRACE("_x2"); break;
 681                 case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
 682                 case WINED3DSPSM_DZ:      TRACE("_dz"); break;
 683                 case WINED3DSPSM_DW:      TRACE("_dw"); break;
 684                 case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
 685                 case WINED3DSPSM_ABS:     TRACE(")"); break;
 686                 default:
 687                     TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
 688             }
 689         }
 690
 691         /**
 692         * swizzle bits fields:
 693         *  RRGGBBAA
 694         */
 695         if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
 696             if (swizzle_r == swizzle_g &&
 697                 swizzle_r == swizzle_b &&
 698                 swizzle_r == swizzle_a) {
 699                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
 700             } else {
 701                 TRACE(".%c%c%c%c",
 702                 swizzle_reg_chars[swizzle_r],
 703                 swizzle_reg_chars[swizzle_g],
 704                 swizzle_reg_chars[swizzle_b],
 705                 swizzle_reg_chars[swizzle_a]);
 706             }
 707         }
 708     }
 709 }
 710
 711 /** Shared code in order to generate the bulk of the shader string.
 712     Use the shader_header_fct & shader_footer_fct to add strings
 713     that are specific to pixel or vertex functions
 714     NOTE: A description of how to parse tokens can be found at:
 715           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
 716 void shader_generate_main(
 717     IWineD3DBaseShader *iface,
 718     SHADER_BUFFER* buffer,
 719     shader_reg_maps* reg_maps,
 720     CONST DWORD* pFunction) {
 721
 722     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 723     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
 724     const DWORD *pToken = pFunction;
 725     const SHADER_OPCODE *curOpcode = NULL;
 726     SHADER_HANDLER hw_fct = NULL;
 727     DWORD i;
 728     SHADER_OPCODE_ARG hw_arg;
 729
 730     /* Initialize current parsing state */
 731     hw_arg.shader = iface;
 732     hw_arg.buffer = buffer;
 733     hw_arg.reg_maps = reg_maps;
 734     This->baseShader.parse_state.current_row = 0;
 735
 736     /* Second pass, process opcodes */
 737     if (NULL != pToken) {
 738         while (WINED3DPS_END() != *pToken) {
 739
 740             /* Skip version token */
 741             if (shader_is_version_token(*pToken)) {
 742                 ++pToken;
 743                 continue;
 744             }
 745
 746             /* Skip comment tokens */
 747             if (shader_is_comment(*pToken)) {
 748                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 749                 ++pToken;
 750                 TRACE("#%s\n", (const char*)pToken);
 751                 pToken += comment_len;
 752                 continue;
 753             }
 754
 755             /* Read opcode */
 756             hw_arg.opcode_token = *pToken++;
 757             curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
 758
 759             /* Select handler */
 760             if (curOpcode == NULL)
 761                 hw_fct = NULL;
 762             else if (This->baseShader.shader_mode == SHADER_GLSL)
 763                 hw_fct = curOpcode->hw_glsl_fct;
 764             else if (This->baseShader.shader_mode == SHADER_ARB)
 765                 hw_fct = curOpcode->hw_fct;
 766
 767             /* Unknown opcode and its parameters */
 768             if (NULL == curOpcode) {
 769                 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
 770                 pToken += shader_skip_unrecognized(iface, pToken);
 771
 772             /* Nothing to do */
 773             } else if (WINED3DSIO_DCL == curOpcode->opcode ||
 774                        WINED3DSIO_NOP == curOpcode->opcode ||
 775                        WINED3DSIO_DEF == curOpcode->opcode ||
 776                        WINED3DSIO_DEFI == curOpcode->opcode ||
 777                        WINED3DSIO_DEFB == curOpcode->opcode ||
 778                        WINED3DSIO_PHASE == curOpcode->opcode ||
 779                        WINED3DSIO_RET == curOpcode->opcode) {
 780
 781                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 782
 783             /* If a generator function is set for current shader target, use it */
 784             } else if (hw_fct != NULL) {
 785
 786                 hw_arg.opcode = curOpcode;
 787
 788                 /* Destination token */
 789                 if (curOpcode->dst_token) {
 790
 791                     DWORD param, addr_token = 0;
 792                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 793                     hw_arg.dst = param;
 794                     hw_arg.dst_addr = addr_token;
 795                 }
 796
 797                 /* Predication token */
 798                 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
 799                     hw_arg.predicate = *pToken++;
 800
 801                 /* Other source tokens */
 802                 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
 803
 804                     DWORD param, addr_token = 0;
 805                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 806                     hw_arg.src[i] = param;
 807                     hw_arg.src_addr[i] = addr_token;
 808                 }
 809
 810                 /* Call appropriate function for output target */
 811                 hw_fct(&hw_arg);
 812
 813                 /* Add color correction if needed */
 814                 device->shader_backend->shader_color_correction(&hw_arg);
 815
 816                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
 817                 if (This->baseShader.shader_mode == SHADER_GLSL)
 818                     shader_glsl_add_instruction_modifiers(&hw_arg);
 819
 820             /* Unhandled opcode */
 821             } else {
 822
 823                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
 824                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 825             }
 826         }
 827         /* TODO: What about result.depth? */
 828
 829     }
 830 }
 831
 832 void shader_dump_ins_modifiers(const DWORD output) {
 833
 834     DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 835     DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
 836
 837     switch (shift) {
 838         case 0: break;
 839         case 13: TRACE("_d8"); break;
 840         case 14: TRACE("_d4"); break;
 841         case 15: TRACE("_d2"); break;
 842         case 1: TRACE("_x2"); break;
 843         case 2: TRACE("_x4"); break;
 844         case 3: TRACE("_x8"); break;
 845         default: TRACE("_unhandled_shift(%d)", shift); break;
 846     }
 847
 848     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
 849     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
 850     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
 851
 852     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
 853     if (mmask)
 854         FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
 855 }
 856
 857 /* First pass: trace shader, initialize length and version */
 858 void shader_trace_init(
 859     IWineD3DBaseShader *iface,
 860     const DWORD* pFunction) {
 861
 862     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
 863
 864     const DWORD* pToken = pFunction;
 865     const SHADER_OPCODE* curOpcode = NULL;
 866     DWORD opcode_token;
 867     unsigned int len = 0;
 868     DWORD i;
 869
 870     TRACE("(%p) : Parsing programme\n", This);
 871
 872     if (NULL != pToken) {
 873         while (WINED3DVS_END() != *pToken) {
 874             if (shader_is_version_token(*pToken)) { /** version */
 875                 This->baseShader.hex_version = *pToken;
 876                 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
 877                     WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
 878                     WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
 879                 ++pToken;
 880                 ++len;
 881                 continue;
 882             }
 883             if (shader_is_comment(*pToken)) { /** comment */
 884                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
 885                 ++pToken;
 886                 TRACE("//%s\n", (const char*)pToken);
 887                 pToken += comment_len;
 888                 len += comment_len + 1;
 889                 continue;
 890             }
 891             opcode_token = *pToken++;
 892             curOpcode = shader_get_opcode(iface, opcode_token);
 893             len++;
 894
 895             if (NULL == curOpcode) {
 896                 int tokens_read;
 897                 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
 898                 tokens_read = shader_skip_unrecognized(iface, pToken);
 899                 pToken += tokens_read;
 900                 len += tokens_read;
 901
 902             } else {
 903                 if (curOpcode->opcode == WINED3DSIO_DCL) {
 904
 905                     DWORD usage = *pToken;
 906                     DWORD param = *(pToken + 1);
 907
 908                     shader_dump_decl_usage(This, usage, param);
 909                     shader_dump_ins_modifiers(param);
 910                     TRACE(" ");
 911                     shader_dump_param(iface, param, 0, 0);
 912                     pToken += 2;
 913                     len += 2;
 914
 915                 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
 916
 917                         unsigned int offset = shader_get_float_offset(*pToken);
 918
 919                         TRACE("def c%u = %f, %f, %f, %f", offset,
 920                             *(const float *)(pToken + 1),
 921                             *(const float *)(pToken + 2),
 922                             *(const float *)(pToken + 3),
 923                             *(const float *)(pToken + 4));
 924
 925                         pToken += 5;
 926                         len += 5;
 927                 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
 928
 929                         TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
 930                             *(pToken + 1),
 931                             *(pToken + 2),
 932                             *(pToken + 3),
 933                             *(pToken + 4));
 934
 935                         pToken += 5;
 936                         len += 5;
 937
 938                 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
 939
 940                         TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
 941                             *(pToken + 1)? "true": "false");
 942
 943                         pToken += 2;
 944                         len += 2;
 945
 946                 } else {
 947
 948                     DWORD param, addr_token;
 949                     int tokens_read;
 950
 951                     /* Print out predication source token first - it follows
 952                      * the destination token. */
 953                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
 954                         TRACE("(");
 955                         shader_dump_param(iface, *(pToken + 2), 0, 1);
 956                         TRACE(") ");
 957                     }
 958                     if (opcode_token & WINED3DSI_COISSUE) {
 959                         /* PixWin marks instructions with the coissue flag with a '+' */
 960                         TRACE("+");
 961                     }
 962
 963                     TRACE("%s", curOpcode->name);
 964
 965                     if (curOpcode->opcode == WINED3DSIO_IFC ||
 966                         curOpcode->opcode == WINED3DSIO_BREAKC) {
 967
 968                         DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
 969                         switch (op) {
 970                             case COMPARISON_GT: TRACE("_gt"); break;
 971                             case COMPARISON_EQ: TRACE("_eq"); break;
 972                             case COMPARISON_GE: TRACE("_ge"); break;
 973                             case COMPARISON_LT: TRACE("_lt"); break;
 974                             case COMPARISON_NE: TRACE("_ne"); break;
 975                             case COMPARISON_LE: TRACE("_le"); break;
 976                             default:
 977                                 TRACE("_(%u)", op);
 978                         }
 979                     } else if (curOpcode->opcode == WINED3DSIO_TEX &&
 980                                This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
 981                         if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
 982                     }
 983
 984                     /* Destination token */
 985                     if (curOpcode->dst_token) {
 986
 987                         /* Destination token */
 988                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 989                         pToken += tokens_read;
 990                         len += tokens_read;
 991
 992                         shader_dump_ins_modifiers(param);
 993                         TRACE(" ");
 994                         shader_dump_param(iface, param, addr_token, 0);
 995                     }
 996
 997                     /* Predication token - already printed out, just skip it */
 998                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
 999                         pToken++;
1000                         len++;
1001                     }
1002
1003                     /* Other source tokens */
1004                     for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
1005
1006                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1007                         pToken += tokens_read;
1008                         len += tokens_read;
1009
1010                         TRACE((i == 0)? " " : ", ");
1011                         shader_dump_param(iface, param, addr_token, 1);
1012                     }
1013                 }
1014                 TRACE("\n");
1015             }
1016         }
1017         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1018     } else {
1019         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1020     }
1021 }
1022
1023 void shader_delete_constant_list(
1024     struct list* clist) {
1025
1026     struct list *ptr;
1027     struct local_constant* constant;
1028
1029     ptr = list_head(clist);
1030     while (ptr) {
1031         constant = LIST_ENTRY(ptr, struct local_constant, entry);
1032         ptr = list_next(clist, ptr);
1033         HeapFree(GetProcessHeap(), 0, constant);
1034     }
1035 }
1036
1037 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1038 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1039 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1040 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1041
1042 const shader_backend_t none_shader_backend = {
1043     &shader_none_select,
1044     &shader_none_select_depth_blt,
1045     &shader_none_load_constants,
1046     &shader_none_cleanup
1047 };