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 inline static 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     buffer->bsize += rc;
  61     buffer->lineNo++;
  62     TRACE("GL HW (%u, %u) : %s", buffer->lineNo, buffer->bsize, base);
  63     return 0;
  64 }
  65
  66 const SHADER_OPCODE* shader_get_opcode(
  67     IWineD3DBaseShader *iface, const DWORD code) {
  68
  69     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl*) iface;
  70
  71     DWORD i = 0;
  72     DWORD hex_version = This->baseShader.hex_version;
  73     const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
  74
  75     /** TODO: use dichotomic search */
  76     while (NULL != shader_ins[i].name) {
  77         if (((code & D3DSI_OPCODE_MASK) == shader_ins[i].opcode) &&
  78             (((hex_version >= shader_ins[i].min_version) && (hex_version <= shader_ins[i].max_version)) ||
  79             ((shader_ins[i].min_version == 0) && (shader_ins[i].max_version == 0)))) {
  80             return &shader_ins[i];
  81         }
  82         ++i;
  83     }
  84     FIXME("Unsupported opcode %#lx(%ld) masked %#lx, shader version %#lx\n",
  85        code, code, code & D3DSI_OPCODE_MASK, hex_version);
  86     return NULL;
  87 }
  88
  89 /* Read a parameter opcode from the input stream,
  90  * and possibly a relative addressing token.
  91  * Return the number of tokens read */
  92 int shader_get_param(
  93     IWineD3DBaseShader* iface,
  94     const DWORD* pToken,
  95     DWORD* param,
  96     DWORD* addr_token) {
  97
  98     /* PS >= 3.0 have relative addressing (with token)
  99      * VS >= 2.0 have relative addressing (with token)
 100      * VS >= 1.0 < 2.0 have relative addressing (without token)
 101      * The version check below should work in general */
 102
 103     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 104     char rel_token = D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
 105         ((*pToken & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
 106
 107     *param = *pToken;
 108     *addr_token = rel_token? *(pToken + 1): 0;
 109     return rel_token? 2:1;
 110 }
 111
 112 /* Return the number of parameters to skip for an opcode */
 113 static inline int shader_skip_opcode(
 114     IWineD3DBaseShaderImpl* This,
 115     const SHADER_OPCODE* curOpcode,
 116     DWORD opcode_token) {
 117
 118    /* Shaders >= 2.0 may contain address tokens, but fortunately they
 119     * have a useful legnth mask - use it here. Shaders 1.0 contain no such tokens */
 120
 121     return (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
 122         ((opcode_token & D3DSI_INSTLENGTH_MASK) >> D3DSI_INSTLENGTH_SHIFT):
 123         curOpcode->num_params;
 124 }
 125
 126 /* Read the parameters of an unrecognized opcode from the input stream
 127  * Return the number of tokens read.
 128  *
 129  * Note: This function assumes source or destination token format.
 130  * It will not work with specially-formatted tokens like DEF or DCL,
 131  * but hopefully those would be recognized */
 132
 133 int shader_skip_unrecognized(
 134     IWineD3DBaseShader* iface,
 135     const DWORD* pToken) {
 136
 137     int tokens_read = 0;
 138     int i = 0;
 139
 140     /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
 141     while (*pToken & 0x80000000) {
 142
 143         DWORD param, addr_token;
 144         tokens_read += shader_get_param(iface, pToken, &param, &addr_token);
 145         pToken += tokens_read;
 146
 147         FIXME("Unrecognized opcode param: token=%08lX "
 148             "addr_token=%08lX name=", param, addr_token);
 149         shader_dump_param(iface, param, addr_token, i);
 150         FIXME("\n");
 151         ++i;
 152     }
 153     return tokens_read;
 154 }
 155
 156 /* Convert floating point offset relative
 157  * to a register file to an absolute offset for float constants */
 158
 159 unsigned int shader_get_float_offset(const DWORD reg) {
 160
 161      unsigned int regnum = reg & D3DSP_REGNUM_MASK;
 162      int regtype = shader_get_regtype(reg);
 163
 164      switch (regtype) {
 165         case D3DSPR_CONST: return regnum;
 166         case D3DSPR_CONST2: return 2048 + regnum;
 167         case D3DSPR_CONST3: return 4096 + regnum;
 168         case D3DSPR_CONST4: return 6144 + regnum;
 169         default:
 170             FIXME("Unsupported register type: %d\n", regtype);
 171             return regnum;
 172      }
 173 }
 174
 175 /* Note that this does not count the loop register
 176  * as an address register. */
 177
 178 HRESULT shader_get_registers_used(
 179     IWineD3DBaseShader *iface,
 180     shader_reg_maps* reg_maps,
 181     semantic* semantics_in,
 182     semantic* semantics_out,
 183     CONST DWORD* pToken) {
 184
 185     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 186
 187     /* There are some minor differences between pixel and vertex shaders */
 188     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 189
 190     if (pToken == NULL)
 191         return WINED3D_OK;
 192
 193     while (D3DVS_END() != *pToken) {
 194         CONST SHADER_OPCODE* curOpcode;
 195         DWORD opcode_token;
 196
 197         /* Skip version */
 198         if (shader_is_version_token(*pToken)) {
 199              ++pToken;
 200              continue;
 201
 202         /* Skip comments */
 203         } else if (shader_is_comment(*pToken)) {
 204              DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 205              ++pToken;
 206              pToken += comment_len;
 207              continue;
 208         }
 209
 210         /* Fetch opcode */
 211         opcode_token = *pToken++;
 212         curOpcode = shader_get_opcode(iface, opcode_token);
 213
 214         /* Unhandled opcode, and its parameters */
 215         if (NULL == curOpcode) {
 216            while (*pToken & 0x80000000)
 217                ++pToken;
 218
 219         /* Handle declarations */
 220         } else if (D3DSIO_DCL == curOpcode->opcode) {
 221
 222             DWORD usage = *pToken++;
 223             DWORD param = *pToken++;
 224             DWORD regtype = shader_get_regtype(param);
 225             unsigned int regnum = param & D3DSP_REGNUM_MASK;
 226
 227             /* Vshader: mark attributes used
 228                Pshader: mark 3.0 input registers used, save token */
 229             if (D3DSPR_INPUT == regtype) {
 230
 231                 if (!pshader)
 232                     reg_maps->attributes[regnum] = 1;
 233                 else
 234                     reg_maps->packed_input[regnum] = 1;
 235
 236                 semantics_in[regnum].usage = usage;
 237                 semantics_in[regnum].reg = param;
 238
 239             /* Vshader: mark 3.0 output registers used, save token */
 240             } else if (D3DSPR_OUTPUT == regtype) {
 241                 reg_maps->packed_output[regnum] = 1;
 242                 semantics_out[regnum].usage = usage;
 243                 semantics_out[regnum].reg = param;
 244
 245             /* Save sampler usage token */
 246             } else if (D3DSPR_SAMPLER == regtype)
 247                 reg_maps->samplers[regnum] = usage;
 248
 249         } else if (D3DSIO_DEF == curOpcode->opcode) {
 250
 251             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 252             if (!lconst) return E_OUTOFMEMORY;
 253             lconst->idx = *pToken & D3DSP_REGNUM_MASK;
 254             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 255             list_add_head(&This->baseShader.constantsF, &lconst->entry);
 256             pToken += curOpcode->num_params;
 257
 258         } else if (D3DSIO_DEFI == curOpcode->opcode) {
 259
 260             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 261             if (!lconst) return E_OUTOFMEMORY;
 262             lconst->idx = *pToken & D3DSP_REGNUM_MASK;
 263             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
 264             list_add_head(&This->baseShader.constantsI, &lconst->entry);
 265             pToken += curOpcode->num_params;
 266
 267         } else if (D3DSIO_DEFB == curOpcode->opcode) {
 268
 269             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 270             if (!lconst) return E_OUTOFMEMORY;
 271             lconst->idx = *pToken & D3DSP_REGNUM_MASK;
 272             memcpy(&lconst->value, pToken + 1, 1 * sizeof(DWORD));
 273             list_add_head(&This->baseShader.constantsB, &lconst->entry);
 274             pToken += curOpcode->num_params;
 275
 276         /* If there's a loop in the shader */
 277         } else if (D3DSIO_LOOP == curOpcode->opcode ||
 278                    D3DSIO_REP == curOpcode->opcode) {
 279             reg_maps->loop = 1;
 280             pToken += curOpcode->num_params;
 281
 282         /* For subroutine prototypes */
 283         } else if (D3DSIO_LABEL == curOpcode->opcode) {
 284
 285             DWORD snum = *pToken & D3DSP_REGNUM_MASK;
 286             reg_maps->labels[snum] = 1;
 287             pToken += curOpcode->num_params;
 288
 289         /* Set texture, address, temporary registers */
 290         } else {
 291             int i, limit;
 292
 293             /* Declare 1.X samplers implicitly, based on the destination reg. number */
 294             if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 295                 (D3DSIO_TEX == curOpcode->opcode ||
 296                  D3DSIO_TEXBEM == curOpcode->opcode ||
 297                  D3DSIO_TEXM3x2TEX == curOpcode->opcode ||
 298                  D3DSIO_TEXM3x3TEX == curOpcode->opcode)) {
 299
 300                 /* Fake sampler usage, only set reserved bit and ttype */
 301                 DWORD sampler_code = *pToken & D3DSP_REGNUM_MASK;
 302                 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
 303
 304             } else if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 305                 (D3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
 306                  D3DSIO_TEXM3x3VSPEC == curOpcode->opcode)) {
 307
 308                 /* 3D sampler usage, only set reserved bit and ttype
 309                  * FIXME: This could be either Cube or Volume, but we wouldn't know unless
 310                  * we waited to generate the shader until the textures were all bound.
 311                  * For now, use Cube textures because they are more common. */
 312                 DWORD sampler_code = *pToken & D3DSP_REGNUM_MASK;
 313                 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
 314             } else if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
 315                 (D3DSIO_TEXDP3TEX == curOpcode->opcode)) {
 316
 317                 /* 1D Sampler usage */
 318                 DWORD sampler_code = *pToken & D3DSP_REGNUM_MASK;
 319                 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_1D;
 320             }
 321
 322             /* This will loop over all the registers and try to
 323              * make a bitmask of the ones we're interested in.
 324              *
 325              * Relative addressing tokens are ignored, but that's
 326              * okay, since we'll catch any address registers when
 327              * they are initialized (required by spec) */
 328
 329             limit = (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)?
 330                 curOpcode->num_params + 1: curOpcode->num_params;
 331
 332             for (i = 0; i < limit; ++i) {
 333
 334                 DWORD param, addr_token, reg, regtype;
 335                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
 336
 337                 regtype = shader_get_regtype(param);
 338                 reg = param & D3DSP_REGNUM_MASK;
 339
 340                 if (D3DSPR_TEXTURE == regtype) { /* vs: D3DSPR_ADDR */
 341
 342                     if (pshader)
 343                         reg_maps->texcoord[reg] = 1;
 344                     else
 345                         reg_maps->address[reg] = 1;
 346                 }
 347
 348                 else if (D3DSPR_TEMP == regtype)
 349                     reg_maps->temporary[reg] = 1;
 350
 351                 else if (D3DSPR_INPUT == regtype && !pshader)
 352                     reg_maps->attributes[reg] = 1;
 353
 354                 else if (D3DSPR_RASTOUT == regtype && reg == 1)
 355                     reg_maps->fog = 1;
 356              }
 357         }
 358     }
 359
 360     return WINED3D_OK;
 361 }
 362
 363 static void shader_dump_decl_usage(
 364     IWineD3DBaseShaderImpl* This,
 365     DWORD decl,
 366     DWORD param) {
 367
 368     DWORD regtype = shader_get_regtype(param);
 369
 370     TRACE("dcl");
 371
 372     if (regtype == D3DSPR_SAMPLER) {
 373         DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
 374
 375         switch (ttype) {
 376             case WINED3DSTT_2D: TRACE("_2d"); break;
 377             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 378             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 379             default: TRACE("_unknown_ttype(%08lx)", ttype);
 380        }
 381
 382     } else {
 383
 384         DWORD usage = decl & D3DSP_DCL_USAGE_MASK;
 385         DWORD idx = (decl & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
 386
 387         /* Pixel shaders 3.0 don't have usage semantics */
 388         char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 389         if (pshader && This->baseShader.hex_version < D3DPS_VERSION(3,0))
 390             return;
 391         else
 392             TRACE("_");
 393
 394         switch(usage) {
 395         case D3DDECLUSAGE_POSITION:
 396             TRACE("%s%ld", "position", idx);
 397             break;
 398         case D3DDECLUSAGE_BLENDINDICES:
 399             TRACE("%s", "blend");
 400             break;
 401         case D3DDECLUSAGE_BLENDWEIGHT:
 402             TRACE("%s", "weight");
 403             break;
 404         case D3DDECLUSAGE_NORMAL:
 405             TRACE("%s%ld", "normal", idx);
 406             break;
 407         case D3DDECLUSAGE_PSIZE:
 408             TRACE("%s", "psize");
 409             break;
 410         case D3DDECLUSAGE_COLOR:
 411             if(idx == 0)  {
 412                 TRACE("%s", "color");
 413             } else {
 414                 TRACE("%s%ld", "specular", (idx - 1));
 415             }
 416             break;
 417         case D3DDECLUSAGE_TEXCOORD:
 418             TRACE("%s%ld", "texture", idx);
 419             break;
 420         case D3DDECLUSAGE_TANGENT:
 421             TRACE("%s", "tangent");
 422             break;
 423         case D3DDECLUSAGE_BINORMAL:
 424             TRACE("%s", "binormal");
 425             break;
 426         case D3DDECLUSAGE_TESSFACTOR:
 427             TRACE("%s", "tessfactor");
 428             break;
 429         case D3DDECLUSAGE_POSITIONT:
 430             TRACE("%s%ld", "positionT", idx);
 431             break;
 432         case D3DDECLUSAGE_FOG:
 433             TRACE("%s", "fog");
 434             break;
 435         case D3DDECLUSAGE_DEPTH:
 436             TRACE("%s", "depth");
 437             break;
 438         case D3DDECLUSAGE_SAMPLE:
 439             TRACE("%s", "sample");
 440             break;
 441         default:
 442             FIXME("unknown_semantics(%08lx)", usage);
 443         }
 444     }
 445 }
 446
 447 static void shader_dump_arr_entry(
 448     IWineD3DBaseShader *iface,
 449     const DWORD param,
 450     const DWORD addr_token,
 451     unsigned int reg,
 452     int input) {
 453
 454     char relative =
 455         ((param & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
 456
 457     if (relative) {
 458         TRACE("[");
 459         if (addr_token)
 460             shader_dump_param(iface, addr_token, 0, input);
 461         else
 462             TRACE("a0.x");
 463         TRACE(" + ");
 464      }
 465      TRACE("%u", reg);
 466      if (relative)
 467          TRACE("]");
 468 }
 469
 470 void shader_dump_param(
 471     IWineD3DBaseShader *iface,
 472     const DWORD param,
 473     const DWORD addr_token,
 474     int input) {
 475
 476     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 477     static const char* rastout_reg_names[] = { "oPos", "oFog", "oPts" };
 478     char swizzle_reg_chars[4];
 479
 480     DWORD reg = param & D3DSP_REGNUM_MASK;
 481     DWORD regtype = shader_get_regtype(param);
 482     DWORD modifier = param & D3DSP_SRCMOD_MASK;
 483
 484     /* There are some minor differences between pixel and vertex shaders */
 485     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 486
 487     /* For one, we'd prefer color components to be shown for pshaders.
 488      * FIXME: use the swizzle function for this */
 489
 490     swizzle_reg_chars[0] = pshader? 'r': 'x';
 491     swizzle_reg_chars[1] = pshader? 'g': 'y';
 492     swizzle_reg_chars[2] = pshader? 'b': 'z';
 493     swizzle_reg_chars[3] = pshader? 'a': 'w';
 494
 495     if (input) {
 496         if ( (modifier == D3DSPSM_NEG) ||
 497              (modifier == D3DSPSM_BIASNEG) ||
 498              (modifier == D3DSPSM_SIGNNEG) ||
 499              (modifier == D3DSPSM_X2NEG) ||
 500              (modifier == D3DSPSM_ABSNEG) )
 501             TRACE("-");
 502         else if (modifier == D3DSPSM_COMP)
 503             TRACE("1-");
 504         else if (modifier == D3DSPSM_NOT)
 505             TRACE("!");
 506
 507         if (modifier == D3DSPSM_ABS || modifier == D3DSPSM_ABSNEG)
 508             TRACE("abs(");
 509     }
 510
 511     switch (regtype) {
 512         case D3DSPR_TEMP:
 513             TRACE("r%lu", reg);
 514             break;
 515         case D3DSPR_INPUT:
 516             TRACE("v");
 517             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 518             break;
 519         case D3DSPR_CONST:
 520         case D3DSPR_CONST2:
 521         case D3DSPR_CONST3:
 522         case D3DSPR_CONST4:
 523             TRACE("c");
 524             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
 525             break;
 526         case D3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
 527             TRACE("%c%lu", (pshader? 't':'a'), reg);
 528             break;
 529         case D3DSPR_RASTOUT:
 530             TRACE("%s", rastout_reg_names[reg]);
 531             break;
 532         case D3DSPR_COLOROUT:
 533             TRACE("oC%lu", reg);
 534             break;
 535         case D3DSPR_DEPTHOUT:
 536             TRACE("oDepth");
 537             break;
 538         case D3DSPR_ATTROUT:
 539             TRACE("oD%lu", reg);
 540             break;
 541         case D3DSPR_TEXCRDOUT:
 542
 543             /* Vertex shaders >= 3.0 use general purpose output registers
 544              * (D3DSPR_OUTPUT), which can include an address token */
 545
 546             if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
 547                 TRACE("o");
 548                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
 549             }
 550             else
 551                TRACE("oT%lu", reg);
 552             break;
 553         case D3DSPR_CONSTINT:
 554             TRACE("i");
 555             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 556             break;
 557         case D3DSPR_CONSTBOOL:
 558             TRACE("b");
 559             shader_dump_arr_entry(iface, param, addr_token, reg, input);
 560             break;
 561         case D3DSPR_LABEL:
 562             TRACE("l%lu", reg);
 563             break;
 564         case D3DSPR_LOOP:
 565             TRACE("aL");
 566             break;
 567         case D3DSPR_SAMPLER:
 568             TRACE("s%lu", reg);
 569             break;
 570         case D3DSPR_PREDICATE:
 571             TRACE("p%lu", reg);
 572             break;
 573         default:
 574             TRACE("unhandled_rtype(%#lx)", regtype);
 575             break;
 576    }
 577
 578    if (!input) {
 579        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
 580
 581        if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
 582            TRACE(".");
 583            if (param & D3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
 584            if (param & D3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
 585            if (param & D3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
 586            if (param & D3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
 587        }
 588
 589    } else {
 590         /** operand input */
 591         DWORD swizzle = (param & D3DSP_SWIZZLE_MASK) >> D3DSP_SWIZZLE_SHIFT;
 592         DWORD swizzle_r = swizzle & 0x03;
 593         DWORD swizzle_g = (swizzle >> 2) & 0x03;
 594         DWORD swizzle_b = (swizzle >> 4) & 0x03;
 595         DWORD swizzle_a = (swizzle >> 6) & 0x03;
 596
 597         if (0 != modifier) {
 598             switch (modifier) {
 599                 case D3DSPSM_NONE:    break;
 600                 case D3DSPSM_NEG:     break;
 601                 case D3DSPSM_NOT:     break;
 602                 case D3DSPSM_BIAS:    TRACE("_bias"); break;
 603                 case D3DSPSM_BIASNEG: TRACE("_bias"); break;
 604                 case D3DSPSM_SIGN:    TRACE("_bx2"); break;
 605                 case D3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 606                 case D3DSPSM_COMP:    break;
 607                 case D3DSPSM_X2:      TRACE("_x2"); break;
 608                 case D3DSPSM_X2NEG:   TRACE("_x2"); break;
 609                 case D3DSPSM_DZ:      TRACE("_dz"); break;
 610                 case D3DSPSM_DW:      TRACE("_dw"); break;
 611                 case D3DSPSM_ABSNEG:  TRACE(")"); break;
 612                 case D3DSPSM_ABS:     TRACE(")"); break;
 613                 default:
 614                     TRACE("_unknown_modifier(%#lx)", modifier >> D3DSP_SRCMOD_SHIFT);
 615             }
 616         }
 617
 618         /**
 619         * swizzle bits fields:
 620         *  RRGGBBAA
 621         */
 622         if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
 623             if (swizzle_r == swizzle_g &&
 624                 swizzle_r == swizzle_b &&
 625                 swizzle_r == swizzle_a) {
 626                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
 627             } else {
 628                 TRACE(".%c%c%c%c",
 629                 swizzle_reg_chars[swizzle_r],
 630                 swizzle_reg_chars[swizzle_g],
 631                 swizzle_reg_chars[swizzle_b],
 632                 swizzle_reg_chars[swizzle_a]);
 633             }
 634         }
 635     }
 636 }
 637
 638 /** Shared code in order to generate the bulk of the shader string.
 639     Use the shader_header_fct & shader_footer_fct to add strings
 640     that are specific to pixel or vertex functions
 641     NOTE: A description of how to parse tokens can be found at:
 642           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
 643 void shader_generate_main(
 644     IWineD3DBaseShader *iface,
 645     SHADER_BUFFER* buffer,
 646     shader_reg_maps* reg_maps,
 647     CONST DWORD* pFunction) {
 648
 649     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 650     const DWORD *pToken = pFunction;
 651     const SHADER_OPCODE *curOpcode = NULL;
 652     SHADER_HANDLER hw_fct = NULL;
 653     DWORD i;
 654     SHADER_OPCODE_ARG hw_arg;
 655
 656     /* Initialize current parsing state */
 657     hw_arg.shader = iface;
 658     hw_arg.buffer = buffer;
 659     hw_arg.reg_maps = reg_maps;
 660     This->baseShader.parse_state.current_row = 0;
 661
 662     /* Second pass, process opcodes */
 663     if (NULL != pToken) {
 664         while (D3DPS_END() != *pToken) {
 665
 666             /* Skip version token */
 667             if (shader_is_version_token(*pToken)) {
 668                 ++pToken;
 669                 continue;
 670             }
 671
 672             /* Skip comment tokens */
 673             if (shader_is_comment(*pToken)) {
 674                 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 675                 ++pToken;
 676                 TRACE("#%s\n", (char*)pToken);
 677                 pToken += comment_len;
 678                 continue;
 679             }
 680
 681             /* Read opcode */
 682             hw_arg.opcode_token = *pToken++;
 683             curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
 684
 685             /* Select handler */
 686             if (curOpcode == NULL)
 687                 hw_fct = NULL;
 688             else if (This->baseShader.shader_mode == SHADER_GLSL)
 689                 hw_fct = curOpcode->hw_glsl_fct;
 690             else if (This->baseShader.shader_mode == SHADER_ARB)
 691                 hw_fct = curOpcode->hw_fct;
 692
 693             /* Unknown opcode and its parameters */
 694             if (NULL == curOpcode) {
 695                 FIXME("Unrecognized opcode: token=%08lX\n", hw_arg.opcode_token);
 696                 pToken += shader_skip_unrecognized(iface, pToken);
 697
 698             /* Nothing to do */
 699             } else if (D3DSIO_DCL == curOpcode->opcode ||
 700                        D3DSIO_NOP == curOpcode->opcode ||
 701                        D3DSIO_DEF == curOpcode->opcode ||
 702                        D3DSIO_DEFI == curOpcode->opcode ||
 703                        D3DSIO_DEFB == curOpcode->opcode ||
 704                        D3DSIO_PHASE == curOpcode->opcode ||
 705                        D3DSIO_RET == curOpcode->opcode) {
 706
 707                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 708
 709             /* If a generator function is set for current shader target, use it */
 710             } else if (hw_fct != NULL) {
 711
 712                 hw_arg.opcode = curOpcode;
 713
 714                 /* Destination token */
 715                 if (curOpcode->dst_token) {
 716
 717                     DWORD param, addr_token = 0;
 718                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 719                     hw_arg.dst = param;
 720                     hw_arg.dst_addr = addr_token;
 721                 }
 722
 723                 /* Predication token */
 724                 if (hw_arg.opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)
 725                     hw_arg.predicate = *pToken++;
 726
 727                 /* Other source tokens */
 728                 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
 729
 730                     DWORD param, addr_token = 0;
 731                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
 732                     hw_arg.src[i] = param;
 733                     hw_arg.src_addr[i] = addr_token;
 734                 }
 735
 736                 /* Call appropriate function for output target */
 737                 hw_fct(&hw_arg);
 738
 739                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
 740                 if (This->baseShader.shader_mode == SHADER_GLSL)
 741                     shader_glsl_add_instruction_modifiers(&hw_arg);
 742
 743             /* Unhandled opcode */
 744             } else {
 745
 746                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
 747                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
 748             }
 749         }
 750         /* TODO: What about result.depth? */
 751
 752     }
 753 }
 754
 755 void shader_dump_ins_modifiers(const DWORD output) {
 756
 757     DWORD shift = (output & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
 758     DWORD mmask = output & D3DSP_DSTMOD_MASK;
 759
 760     switch (shift) {
 761         case 0: break;
 762         case 13: TRACE("_d8"); break;
 763         case 14: TRACE("_d4"); break;
 764         case 15: TRACE("_d2"); break;
 765         case 1: TRACE("_x2"); break;
 766         case 2: TRACE("_x4"); break;
 767         case 3: TRACE("_x8"); break;
 768         default: TRACE("_unhandled_shift(%ld)", shift); break;
 769     }
 770
 771     if (mmask & D3DSPDM_SATURATE)         TRACE("_sat");
 772     if (mmask & D3DSPDM_PARTIALPRECISION) TRACE("_pp");
 773     if (mmask & D3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
 774
 775     mmask &= ~(D3DSPDM_SATURATE | D3DSPDM_PARTIALPRECISION | D3DSPDM_MSAMPCENTROID);
 776     if (mmask)
 777         FIXME("_unrecognized_modifier(%#lx)", mmask >> D3DSP_DSTMOD_SHIFT);
 778 }
 779
 780 /* First pass: trace shader, initialize length and version */
 781 void shader_trace_init(
 782     IWineD3DBaseShader *iface,
 783     const DWORD* pFunction) {
 784
 785     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
 786
 787     const DWORD* pToken = pFunction;
 788     const SHADER_OPCODE* curOpcode = NULL;
 789     DWORD opcode_token;
 790     unsigned int len = 0;
 791     DWORD i;
 792
 793     TRACE("(%p) : Parsing programme\n", This);
 794
 795     if (NULL != pToken) {
 796         while (D3DVS_END() != *pToken) {
 797             if (shader_is_version_token(*pToken)) { /** version */
 798                 This->baseShader.hex_version = *pToken;
 799                 TRACE("%s_%lu_%lu\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
 800                     D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
 801                     D3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
 802                 ++pToken;
 803                 ++len;
 804                 continue;
 805             }
 806             if (shader_is_comment(*pToken)) { /** comment */
 807                 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
 808                 ++pToken;
 809                 TRACE("//%s\n", (char*)pToken);
 810                 pToken += comment_len;
 811                 len += comment_len + 1;
 812                 continue;
 813             }
 814             opcode_token = *pToken++;
 815             curOpcode = shader_get_opcode(iface, opcode_token);
 816             len++;
 817
 818             if (NULL == curOpcode) {
 819                 int tokens_read;
 820                 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
 821                 tokens_read = shader_skip_unrecognized(iface, pToken);
 822                 pToken += tokens_read;
 823                 len += tokens_read;
 824
 825             } else {
 826                 if (curOpcode->opcode == D3DSIO_DCL) {
 827
 828                     DWORD usage = *pToken;
 829                     DWORD param = *(pToken + 1);
 830
 831                     shader_dump_decl_usage(This, usage, param);
 832                     shader_dump_ins_modifiers(param);
 833                     TRACE(" ");
 834                     shader_dump_param(iface, param, 0, 0);
 835                     pToken += 2;
 836                     len += 2;
 837
 838                 } else if (curOpcode->opcode == D3DSIO_DEF) {
 839
 840                         unsigned int offset = shader_get_float_offset(*pToken);
 841
 842                         TRACE("def c%u = %f, %f, %f, %f", offset,
 843                             *(float *)(pToken + 1),
 844                             *(float *)(pToken + 2),
 845                             *(float *)(pToken + 3),
 846                             *(float *)(pToken + 4));
 847
 848                         pToken += 5;
 849                         len += 5;
 850                 } else if (curOpcode->opcode == D3DSIO_DEFI) {
 851
 852                         TRACE("defi i%lu = %ld, %ld, %ld, %ld", *pToken & D3DSP_REGNUM_MASK,
 853                             (long) *(pToken + 1),
 854                             (long) *(pToken + 2),
 855                             (long) *(pToken + 3),
 856                             (long) *(pToken + 4));
 857
 858                         pToken += 5;
 859                         len += 5;
 860
 861                 } else if (curOpcode->opcode == D3DSIO_DEFB) {
 862
 863                         TRACE("defb b%lu = %s", *pToken & D3DSP_REGNUM_MASK,
 864                             *(pToken + 1)? "true": "false");
 865
 866                         pToken += 2;
 867                         len += 2;
 868
 869                 } else {
 870
 871                     DWORD param, addr_token;
 872                     int tokens_read;
 873
 874                     /* Print out predication source token first - it follows
 875                      * the destination token. */
 876                     if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
 877                         TRACE("(");
 878                         shader_dump_param(iface, *(pToken + 2), 0, 1);
 879                         TRACE(") ");
 880                     }
 881
 882                     TRACE("%s", curOpcode->name);
 883
 884                     if (curOpcode->opcode == D3DSIO_IFC ||
 885                         curOpcode->opcode == D3DSIO_BREAKC) {
 886
 887                         DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
 888                         switch (op) {
 889                             case COMPARISON_GT: TRACE("_gt"); break;
 890                             case COMPARISON_EQ: TRACE("_eq"); break;
 891                             case COMPARISON_GE: TRACE("_ge"); break;
 892                             case COMPARISON_LT: TRACE("_lt"); break;
 893                             case COMPARISON_NE: TRACE("_ne"); break;
 894                             case COMPARISON_LE: TRACE("_le"); break;
 895                             default:
 896                                 TRACE("_(%lu)", op);
 897                         }
 898                     }
 899
 900                     /* Destination token */
 901                     if (curOpcode->dst_token) {
 902
 903                         /* Destination token */
 904                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 905                         pToken += tokens_read;
 906                         len += tokens_read;
 907
 908                         shader_dump_ins_modifiers(param);
 909                         TRACE(" ");
 910                         shader_dump_param(iface, param, addr_token, 0);
 911                     }
 912
 913                     /* Predication token - already printed out, just skip it */
 914                     if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
 915                         pToken++;
 916                         len++;
 917                     }
 918
 919                     /* Other source tokens */
 920                     for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
 921
 922                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
 923                         pToken += tokens_read;
 924                         len += tokens_read;
 925
 926                         TRACE((i == 0)? " " : ", ");
 927                         shader_dump_param(iface, param, addr_token, 1);
 928                     }
 929                 }
 930                 TRACE("\n");
 931             }
 932         }
 933         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
 934     } else {
 935         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
 936     }
 937 }
 938
 939 void shader_delete_constant_list(
 940     struct list* clist) {
 941
 942     struct list *ptr;
 943     struct local_constant* constant;
 944
 945     ptr = list_head(clist);
 946     while (ptr) {
 947         constant = LIST_ENTRY(ptr, struct local_constant, entry);
 948         ptr = list_next(clist, ptr);
 949         HeapFree(GetProcessHeap(), 0, constant);
 950     }
 951 }