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  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
  10  * Copyright 2009 Henri Verbeet for CodeWeavers
  11  *
  12  * This library is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public
  14  * License as published by the Free Software Foundation; either
  15  * version 2.1 of the License, or (at your option) any later version.
  16  *
  17  * This library is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with this library; if not, write to the Free Software
  24  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  25  */
  26
  27 #include "config.h"
  28 #include <string.h>
  29 #include <stdio.h>
  30 #include "wined3d_private.h"
  31
  32 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  33 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  34
  35 static const char *shader_opcode_names[] =
  36 {
  37     /* WINED3DSIH_ABS           */ "abs",
  38     /* WINED3DSIH_ADD           */ "add",
  39     /* WINED3DSIH_BEM           */ "bem",
  40     /* WINED3DSIH_BREAK         */ "break",
  41     /* WINED3DSIH_BREAKC        */ "breakc",
  42     /* WINED3DSIH_BREAKP        */ "breakp",
  43     /* WINED3DSIH_CALL          */ "call",
  44     /* WINED3DSIH_CALLNZ        */ "callnz",
  45     /* WINED3DSIH_CMP           */ "cmp",
  46     /* WINED3DSIH_CND           */ "cnd",
  47     /* WINED3DSIH_CRS           */ "crs",
  48     /* WINED3DSIH_DCL           */ "dcl",
  49     /* WINED3DSIH_DEF           */ "def",
  50     /* WINED3DSIH_DEFB          */ "defb",
  51     /* WINED3DSIH_DEFI          */ "defi",
  52     /* WINED3DSIH_DP2ADD        */ "dp2add",
  53     /* WINED3DSIH_DP3           */ "dp3",
  54     /* WINED3DSIH_DP4           */ "dp4",
  55     /* WINED3DSIH_DST           */ "dst",
  56     /* WINED3DSIH_DSX           */ "dsx",
  57     /* WINED3DSIH_DSY           */ "dsy",
  58     /* WINED3DSIH_ELSE          */ "else",
  59     /* WINED3DSIH_ENDIF         */ "endif",
  60     /* WINED3DSIH_ENDLOOP       */ "endloop",
  61     /* WINED3DSIH_ENDREP        */ "endrep",
  62     /* WINED3DSIH_EXP           */ "exp",
  63     /* WINED3DSIH_EXPP          */ "expp",
  64     /* WINED3DSIH_FRC           */ "frc",
  65     /* WINED3DSIH_IF            */ "if",
  66     /* WINED3DSIH_IFC           */ "ifc",
  67     /* WINED3DSIH_LABEL         */ "label",
  68     /* WINED3DSIH_LIT           */ "lit",
  69     /* WINED3DSIH_LOG           */ "log",
  70     /* WINED3DSIH_LOGP          */ "logp",
  71     /* WINED3DSIH_LOOP          */ "loop",
  72     /* WINED3DSIH_LRP           */ "lrp",
  73     /* WINED3DSIH_M3x2          */ "m3x2",
  74     /* WINED3DSIH_M3x3          */ "m3x3",
  75     /* WINED3DSIH_M3x4          */ "m3x4",
  76     /* WINED3DSIH_M4x3          */ "m4x3",
  77     /* WINED3DSIH_M4x4          */ "m4x4",
  78     /* WINED3DSIH_MAD           */ "mad",
  79     /* WINED3DSIH_MAX           */ "max",
  80     /* WINED3DSIH_MIN           */ "min",
  81     /* WINED3DSIH_MOV           */ "mov",
  82     /* WINED3DSIH_MOVA          */ "mova",
  83     /* WINED3DSIH_MUL           */ "mul",
  84     /* WINED3DSIH_NOP           */ "nop",
  85     /* WINED3DSIH_NRM           */ "nrm",
  86     /* WINED3DSIH_PHASE         */ "phase",
  87     /* WINED3DSIH_POW           */ "pow",
  88     /* WINED3DSIH_RCP           */ "rcp",
  89     /* WINED3DSIH_REP           */ "rep",
  90     /* WINED3DSIH_RET           */ "ret",
  91     /* WINED3DSIH_RSQ           */ "rsq",
  92     /* WINED3DSIH_SETP          */ "setp",
  93     /* WINED3DSIH_SGE           */ "sge",
  94     /* WINED3DSIH_SGN           */ "sgn",
  95     /* WINED3DSIH_SINCOS        */ "sincos",
  96     /* WINED3DSIH_SLT           */ "slt",
  97     /* WINED3DSIH_SUB           */ "sub",
  98     /* WINED3DSIH_TEX           */ "texld",
  99     /* WINED3DSIH_TEXBEM        */ "texbem",
 100     /* WINED3DSIH_TEXBEML       */ "texbeml",
 101     /* WINED3DSIH_TEXCOORD      */ "texcrd",
 102     /* WINED3DSIH_TEXDEPTH      */ "texdepth",
 103     /* WINED3DSIH_TEXDP3        */ "texdp3",
 104     /* WINED3DSIH_TEXDP3TEX     */ "texdp3tex",
 105     /* WINED3DSIH_TEXKILL       */ "texkill",
 106     /* WINED3DSIH_TEXLDD        */ "texldd",
 107     /* WINED3DSIH_TEXLDL        */ "texldl",
 108     /* WINED3DSIH_TEXM3x2DEPTH  */ "texm3x2depth",
 109     /* WINED3DSIH_TEXM3x2PAD    */ "texm3x2pad",
 110     /* WINED3DSIH_TEXM3x2TEX    */ "texm3x2tex",
 111     /* WINED3DSIH_TEXM3x3       */ "texm3x3",
 112     /* WINED3DSIH_TEXM3x3DIFF   */ "texm3x3diff",
 113     /* WINED3DSIH_TEXM3x3PAD    */ "texm3x3pad",
 114     /* WINED3DSIH_TEXM3x3SPEC   */ "texm3x3spec",
 115     /* WINED3DSIH_TEXM3x3TEX    */ "texm3x3tex",
 116     /* WINED3DSIH_TEXM3x3VSPEC  */ "texm3x3vspec",
 117     /* WINED3DSIH_TEXREG2AR     */ "texreg2ar",
 118     /* WINED3DSIH_TEXREG2GB     */ "texreg2gb",
 119     /* WINED3DSIH_TEXREG2RGB    */ "texreg2rgb",
 120 };
 121
 122 const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 123 {
 124     switch (version_token >> 16)
 125     {
 126         case WINED3D_SM1_VS:
 127         case WINED3D_SM1_PS:
 128             return &sm1_shader_frontend;
 129
 130         case WINED3D_SM4_PS:
 131         case WINED3D_SM4_VS:
 132         case WINED3D_SM4_GS:
 133             return &sm4_shader_frontend;
 134
 135         default:
 136             FIXME("Unrecognised version token %#x\n", version_token);
 137             return NULL;
 138     }
 139 }
 140
 141 void shader_buffer_init(struct SHADER_BUFFER *buffer)
 142 {
 143     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 144     buffer->buffer[0] = '\0';
 145     buffer->bsize = 0;
 146     buffer->lineNo = 0;
 147     buffer->newline = TRUE;
 148 }
 149
 150 void shader_buffer_free(struct SHADER_BUFFER *buffer)
 151 {
 152     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 153 }
 154
 155 int shader_vaddline(SHADER_BUFFER* buffer, const char *format, va_list args)
 156 {
 157     char* base = buffer->buffer + buffer->bsize;
 158     int rc;
 159
 160     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 161
 162     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 163     {
 164         ERR("The buffer allocated for the shader program string "
 165             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 166         buffer->bsize = SHADER_PGMSIZE - 1;
 167         return -1;
 168     }
 169
 170     if (buffer->newline) {
 171         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 172         buffer->newline = FALSE;
 173     } else {
 174         TRACE("%s", base);
 175     }
 176
 177     buffer->bsize += rc;
 178     if (buffer->buffer[buffer->bsize-1] == '\n') {
 179         buffer->lineNo++;
 180         buffer->newline = TRUE;
 181     }
 182     return 0;
 183 }
 184
 185 int shader_addline(SHADER_BUFFER* buffer, const char *format, ...)
 186 {
 187     int ret;
 188     va_list args;
 189
 190     va_start(args, format);
 191     ret = shader_vaddline(buffer, format, args);
 192     va_end(args);
 193
 194     return ret;
 195 }
 196
 197 void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDevice *device)
 198 {
 199     shader->ref = 1;
 200     shader->device = device;
 201     list_init(&shader->linked_programs);
 202 }
 203
 204 /* Convert floating point offset relative
 205  * to a register file to an absolute offset for float constants */
 206 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
 207 {
 208     switch (register_type)
 209     {
 210         case WINED3DSPR_CONST: return register_idx;
 211         case WINED3DSPR_CONST2: return 2048 + register_idx;
 212         case WINED3DSPR_CONST3: return 4096 + register_idx;
 213         case WINED3DSPR_CONST4: return 6144 + register_idx;
 214         default:
 215             FIXME("Unsupported register type: %d\n", register_type);
 216             return register_idx;
 217     }
 218 }
 219
 220 static void shader_delete_constant_list(struct list* clist) {
 221
 222     struct list *ptr;
 223     struct local_constant* constant;
 224
 225     ptr = list_head(clist);
 226     while (ptr) {
 227         constant = LIST_ENTRY(ptr, struct local_constant, entry);
 228         ptr = list_next(clist, ptr);
 229         HeapFree(GetProcessHeap(), 0, constant);
 230     }
 231     list_init(clist);
 232 }
 233
 234 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 235 {
 236     DWORD idx, shift;
 237     idx = bit >> 5;
 238     shift = bit & 0x1f;
 239     bitmap[idx] |= (1 << shift);
 240 }
 241
 242 static void shader_record_register_usage(IWineD3DBaseShaderImpl *This, struct shader_reg_maps *reg_maps,
 243         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 244 {
 245     switch (reg->type)
 246     {
 247         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 248             if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord[reg->idx] = 1;
 249             else reg_maps->address[reg->idx] = 1;
 250             break;
 251
 252         case WINED3DSPR_TEMP:
 253             reg_maps->temporary[reg->idx] = 1;
 254             break;
 255
 256         case WINED3DSPR_INPUT:
 257             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 258             {
 259                 if (reg->rel_addr)
 260                 {
 261                     /* If relative addressing is used, we must assume that all registers
 262                      * are used. Even if it is a construct like v3[aL], we can't assume
 263                      * that v0, v1 and v2 aren't read because aL can be negative */
 264                     unsigned int i;
 265                     for (i = 0; i < MAX_REG_INPUT; ++i)
 266                     {
 267                         ((IWineD3DPixelShaderImpl *)This)->input_reg_used[i] = TRUE;
 268                     }
 269                 }
 270                 else
 271                 {
 272                     ((IWineD3DPixelShaderImpl *)This)->input_reg_used[reg->idx] = TRUE;
 273                 }
 274             }
 275             else reg_maps->input_registers |= 1 << reg->idx;
 276             break;
 277
 278         case WINED3DSPR_RASTOUT:
 279             if (reg->idx == 1) reg_maps->fog = 1;
 280             break;
 281
 282         case WINED3DSPR_MISCTYPE:
 283             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 284             {
 285                 if (reg->idx == 0) reg_maps->vpos = 1;
 286                 else if (reg->idx == 1) reg_maps->usesfacing = 1;
 287             }
 288             break;
 289
 290         case WINED3DSPR_CONST:
 291             if (reg->rel_addr)
 292             {
 293                 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
 294                 {
 295                     if (reg->idx <= ((IWineD3DVertexShaderImpl *)This)->min_rel_offset)
 296                         ((IWineD3DVertexShaderImpl *)This)->min_rel_offset = reg->idx;
 297                     else if (reg->idx >= ((IWineD3DVertexShaderImpl *)This)->max_rel_offset)
 298                         ((IWineD3DVertexShaderImpl *)This)->max_rel_offset = reg->idx;
 299                 }
 300                 reg_maps->usesrelconstF = TRUE;
 301             }
 302             else
 303             {
 304                 set_bitmap_bit(reg_maps->constf, reg->idx);
 305             }
 306             break;
 307
 308         case WINED3DSPR_CONSTINT:
 309             reg_maps->integer_constants |= (1 << reg->idx);
 310             break;
 311
 312         case WINED3DSPR_CONSTBOOL:
 313             reg_maps->boolean_constants |= (1 << reg->idx);
 314             break;
 315
 316         case WINED3DSPR_COLOROUT:
 317             reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
 318             break;
 319
 320         default:
 321             TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
 322             break;
 323     }
 324 }
 325
 326 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 327 {
 328     switch(instr)
 329     {
 330         case WINED3DSIH_M4x4:
 331         case WINED3DSIH_M3x4:
 332             return param == 1 ? 3 : 0;
 333
 334         case WINED3DSIH_M4x3:
 335         case WINED3DSIH_M3x3:
 336             return param == 1 ? 2 : 0;
 337
 338         case WINED3DSIH_M3x2:
 339             return param == 1 ? 1 : 0;
 340
 341         default:
 342             return 0;
 343     }
 344 }
 345
 346 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
 347 {
 348     static const char *semantic_names[] =
 349     {
 350         /* WINED3DDECLUSAGE_POSITION        */ "SV_POSITION",
 351         /* WINED3DDECLUSAGE_BLENDWEIGHT     */ "BLENDWEIGHT",
 352         /* WINED3DDECLUSAGE_BLENDINDICES    */ "BLENDINDICES",
 353         /* WINED3DDECLUSAGE_NORMAL          */ "NORMAL",
 354         /* WINED3DDECLUSAGE_PSIZE           */ "PSIZE",
 355         /* WINED3DDECLUSAGE_TEXCOORD        */ "TEXCOORD",
 356         /* WINED3DDECLUSAGE_TANGENT         */ "TANGENT",
 357         /* WINED3DDECLUSAGE_BINORMAL        */ "BINORMAL",
 358         /* WINED3DDECLUSAGE_TESSFACTOR      */ "TESSFACTOR",
 359         /* WINED3DDECLUSAGE_POSITIONT       */ "POSITIONT",
 360         /* WINED3DDECLUSAGE_COLOR           */ "COLOR",
 361         /* WINED3DDECLUSAGE_FOG             */ "FOG",
 362         /* WINED3DDECLUSAGE_DEPTH           */ "DEPTH",
 363         /* WINED3DDECLUSAGE_SAMPLE          */ "SAMPLE",
 364     };
 365
 366     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 367     {
 368         FIXME("Unrecognized usage %#x\n", usage);
 369         return "UNRECOGNIZED";
 370     }
 371
 372     return semantic_names[usage];
 373 }
 374
 375 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
 376 {
 377     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 378 }
 379
 380 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 381         const struct wined3d_shader_semantic *s)
 382 {
 383     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 384     e->semantic_idx = s->usage_idx;
 385     e->sysval_semantic = 0;
 386     e->component_type = 0;
 387     e->register_idx = s->reg.reg.idx;
 388     e->mask = s->reg.write_mask;
 389 }
 390
 391 /* Note that this does not count the loop register
 392  * as an address register. */
 393
 394 HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
 395         struct shader_reg_maps *reg_maps, struct wined3d_shader_attribute *attributes,
 396         struct wined3d_shader_signature_element *input_signature,
 397         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 398 {
 399     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 400     void *fe_data = This->baseShader.frontend_data;
 401     struct wined3d_shader_version shader_version;
 402     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 403     const DWORD* pToken = byte_code;
 404
 405     /* There are some minor differences between pixel and vertex shaders */
 406
 407     memset(reg_maps, 0, sizeof(*reg_maps));
 408
 409     /* get_registers_used is called on every compile on some 1.x shaders, which can result
 410      * in stacking up a collection of local constants. Delete the old constants if existing
 411      */
 412     shader_delete_constant_list(&This->baseShader.constantsF);
 413     shader_delete_constant_list(&This->baseShader.constantsB);
 414     shader_delete_constant_list(&This->baseShader.constantsI);
 415
 416     fe->shader_read_header(fe_data, &pToken, &shader_version);
 417     reg_maps->shader_version = shader_version;
 418
 419     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 420                                  sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 421     if(!reg_maps->constf) {
 422         ERR("Out of memory\n");
 423         return E_OUTOFMEMORY;
 424     }
 425
 426     while (!fe->shader_is_end(fe_data, &pToken))
 427     {
 428         struct wined3d_shader_instruction ins;
 429         const char *comment;
 430         UINT param_size;
 431
 432         /* Skip comments */
 433         fe->shader_read_comment(&pToken, &comment);
 434         if (comment) continue;
 435
 436         /* Fetch opcode */
 437         fe->shader_read_opcode(fe_data, &pToken, &ins, &param_size);
 438
 439         /* Unhandled opcode, and its parameters */
 440         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 441         {
 442             TRACE("Skipping unrecognized instruction.\n");
 443             pToken += param_size;
 444             continue;
 445         }
 446
 447         /* Handle declarations */
 448         if (ins.handler_idx == WINED3DSIH_DCL)
 449         {
 450             struct wined3d_shader_semantic semantic;
 451
 452             fe->shader_read_semantic(&pToken, &semantic);
 453
 454             switch (semantic.reg.reg.type)
 455             {
 456                 /* Vshader: mark attributes used
 457                  * Pshader: mark 3.0 input registers used, save token */
 458                 case WINED3DSPR_INPUT:
 459                     reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
 460                     if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
 461                     {
 462                         attributes[semantic.reg.reg.idx].usage = semantic.usage;
 463                         attributes[semantic.reg.reg.idx].usage_idx = semantic.usage_idx;
 464                     }
 465                     else
 466                     {
 467                         shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
 468                     }
 469                     break;
 470
 471                 /* Vshader: mark 3.0 output registers used, save token */
 472                 case WINED3DSPR_OUTPUT:
 473                     reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
 474                     shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
 475                     if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
 476                     break;
 477
 478                 /* Save sampler usage token */
 479                 case WINED3DSPR_SAMPLER:
 480                     reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
 481                     break;
 482
 483                 default:
 484                     TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
 485                     break;
 486             }
 487         }
 488         else if (ins.handler_idx == WINED3DSIH_DEF)
 489         {
 490             struct wined3d_shader_dst_param dst;
 491             struct wined3d_shader_src_param rel_addr;
 492
 493             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 494             if (!lconst) return E_OUTOFMEMORY;
 495
 496             fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
 497             lconst->idx = dst.reg.idx;
 498
 499             memcpy(lconst->value, pToken, 4 * sizeof(DWORD));
 500             pToken += 4;
 501
 502             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 503             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 504             {
 505                 float *value = (float *) lconst->value;
 506                 if (value[0] < -1.0f) value[0] = -1.0f;
 507                 else if (value[0] > 1.0f) value[0] = 1.0f;
 508                 if (value[1] < -1.0f) value[1] = -1.0f;
 509                 else if (value[1] > 1.0f) value[1] = 1.0f;
 510                 if (value[2] < -1.0f) value[2] = -1.0f;
 511                 else if (value[2] > 1.0f) value[2] = 1.0f;
 512                 if (value[3] < -1.0f) value[3] = -1.0f;
 513                 else if (value[3] > 1.0f) value[3] = 1.0f;
 514             }
 515
 516             list_add_head(&This->baseShader.constantsF, &lconst->entry);
 517         }
 518         else if (ins.handler_idx == WINED3DSIH_DEFI)
 519         {
 520             struct wined3d_shader_dst_param dst;
 521             struct wined3d_shader_src_param rel_addr;
 522
 523             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 524             if (!lconst) return E_OUTOFMEMORY;
 525
 526             fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
 527             lconst->idx = dst.reg.idx;
 528
 529             memcpy(lconst->value, pToken, 4 * sizeof(DWORD));
 530             pToken += 4;
 531
 532             list_add_head(&This->baseShader.constantsI, &lconst->entry);
 533             reg_maps->local_int_consts |= (1 << dst.reg.idx);
 534         }
 535         else if (ins.handler_idx == WINED3DSIH_DEFB)
 536         {
 537             struct wined3d_shader_dst_param dst;
 538             struct wined3d_shader_src_param rel_addr;
 539
 540             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 541             if (!lconst) return E_OUTOFMEMORY;
 542
 543             fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
 544             lconst->idx = dst.reg.idx;
 545
 546             memcpy(lconst->value, pToken, sizeof(DWORD));
 547             ++pToken;
 548
 549             list_add_head(&This->baseShader.constantsB, &lconst->entry);
 550             reg_maps->local_bool_consts |= (1 << dst.reg.idx);
 551         }
 552         /* If there's a loop in the shader */
 553         else if (ins.handler_idx == WINED3DSIH_LOOP
 554                 || ins.handler_idx == WINED3DSIH_REP)
 555         {
 556             struct wined3d_shader_src_param src, rel_addr;
 557
 558             fe->shader_read_src_param(fe_data, &pToken, &src, &rel_addr);
 559
 560             /* Rep and Loop always use an integer constant for the control parameters */
 561             if (ins.handler_idx == WINED3DSIH_REP)
 562             {
 563                 reg_maps->integer_constants |= 1 << src.reg.idx;
 564             }
 565             else
 566             {
 567                 fe->shader_read_src_param(fe_data, &pToken, &src, &rel_addr);
 568                 reg_maps->integer_constants |= 1 << src.reg.idx;
 569             }
 570
 571             cur_loop_depth++;
 572             if(cur_loop_depth > max_loop_depth)
 573                 max_loop_depth = cur_loop_depth;
 574         }
 575         else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 576                 || ins.handler_idx == WINED3DSIH_ENDREP)
 577         {
 578             cur_loop_depth--;
 579         }
 580         /* For subroutine prototypes */
 581         else if (ins.handler_idx == WINED3DSIH_LABEL)
 582         {
 583             struct wined3d_shader_src_param src, rel_addr;
 584
 585             fe->shader_read_src_param(fe_data, &pToken, &src, &rel_addr);
 586             reg_maps->labels[src.reg.idx] = 1;
 587         }
 588         /* Set texture, address, temporary registers */
 589         else
 590         {
 591             int i, limit;
 592             BOOL color0_mov = FALSE;
 593
 594             /* This will loop over all the registers and try to
 595              * make a bitmask of the ones we're interested in.
 596              *
 597              * Relative addressing tokens are ignored, but that's
 598              * okay, since we'll catch any address registers when
 599              * they are initialized (required by spec) */
 600
 601             if (ins.dst_count)
 602             {
 603                 struct wined3d_shader_dst_param dst_param;
 604                 struct wined3d_shader_src_param dst_rel_addr;
 605
 606                 fe->shader_read_dst_param(fe_data, &pToken, &dst_param, &dst_rel_addr);
 607
 608                 shader_record_register_usage(This, reg_maps, &dst_param.reg, shader_version.type);
 609
 610                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 611                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 612                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 613                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
 614                         && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
 615                 {
 616                     reg_maps->texcoord_mask[dst_param.reg.idx] |= dst_param.write_mask;
 617                 }
 618
 619                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
 620                         && dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
 621                 {
 622                     /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 623                      * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 624                      * the mov and perform the sRGB write correction from the source register.
 625                      *
 626                      * However, if the mov is only partial, we can't do this, and if the write
 627                      * comes from an instruction other than MOV it is hard to do as well. If
 628                      * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 629                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)This;
 630
 631                     ps->color0_mov = FALSE;
 632                     if (ins.handler_idx == WINED3DSIH_MOV)
 633                     {
 634                         /* Used later when the source register is read. */
 635                         color0_mov = TRUE;
 636                     }
 637                 }
 638
 639                 /* Declare 1.X samplers implicitly, based on the destination reg. number */
 640                 if (shader_version.major == 1
 641                         && (ins.handler_idx == WINED3DSIH_TEX
 642                             || ins.handler_idx == WINED3DSIH_TEXBEM
 643                             || ins.handler_idx == WINED3DSIH_TEXBEML
 644                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 645                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 646                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 647                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 648                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 649                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 650                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 651                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 652                 {
 653                     /* Fake sampler usage, only set reserved bit and ttype */
 654                     DWORD sampler_code = dst_param.reg.idx;
 655
 656                     TRACE("Setting fake 2D sampler for 1.x pixelshader\n");
 657                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 658
 659                     /* texbem is only valid with < 1.4 pixel shaders */
 660                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 661                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 662                     {
 663                         reg_maps->bumpmat[sampler_code] = TRUE;
 664                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 665                         {
 666                             reg_maps->luminanceparams[sampler_code] = TRUE;
 667                         }
 668                     }
 669                 }
 670                 else if (ins.handler_idx == WINED3DSIH_BEM)
 671                 {
 672                     reg_maps->bumpmat[dst_param.reg.idx] = TRUE;
 673                 }
 674             }
 675
 676             if (ins.handler_idx == WINED3DSIH_NRM)
 677             {
 678                 reg_maps->usesnrm = 1;
 679             }
 680             else if (ins.handler_idx == WINED3DSIH_DSY)
 681             {
 682                 reg_maps->usesdsy = 1;
 683             }
 684             else if (ins.handler_idx == WINED3DSIH_DSX)
 685             {
 686                 reg_maps->usesdsx = 1;
 687             }
 688             else if(ins.handler_idx == WINED3DSIH_TEXLDD)
 689             {
 690                 reg_maps->usestexldd = 1;
 691             }
 692             else if(ins.handler_idx == WINED3DSIH_TEXLDL)
 693             {
 694                 reg_maps->usestexldl = 1;
 695             }
 696             else if(ins.handler_idx == WINED3DSIH_MOVA)
 697             {
 698                 reg_maps->usesmova = 1;
 699             }
 700             else if(ins.handler_idx == WINED3DSIH_IFC)
 701             {
 702                 reg_maps->usesifc = 1;
 703             }
 704             else if(ins.handler_idx == WINED3DSIH_CALL)
 705             {
 706                 reg_maps->usescall = 1;
 707             }
 708
 709             limit = ins.src_count + (ins.predicate ? 1 : 0);
 710             for (i = 0; i < limit; ++i)
 711             {
 712                 struct wined3d_shader_src_param src_param, src_rel_addr;
 713                 unsigned int count;
 714
 715                 fe->shader_read_src_param(fe_data, &pToken, &src_param, &src_rel_addr);
 716                 count = get_instr_extra_regcount(ins.handler_idx, i);
 717
 718                 shader_record_register_usage(This, reg_maps, &src_param.reg, shader_version.type);
 719                 while (count)
 720                 {
 721                     ++src_param.reg.idx;
 722                     shader_record_register_usage(This, reg_maps, &src_param.reg, shader_version.type);
 723                     --count;
 724                 }
 725
 726                 if(color0_mov)
 727                 {
 728                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
 729                     if(src_param.reg.type == WINED3DSPR_TEMP &&
 730                        src_param.swizzle == WINED3DSP_NOSWIZZLE)
 731                     {
 732                         ps->color0_mov = TRUE;
 733                         ps->color0_reg = src_param.reg.idx;
 734                     }
 735                 }
 736             }
 737         }
 738     }
 739     reg_maps->loop_depth = max_loop_depth;
 740
 741     This->baseShader.functionLength = ((char *)pToken - (char *)byte_code);
 742
 743     return WINED3D_OK;
 744 }
 745
 746 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 747         const struct wined3d_shader_version *shader_version)
 748 {
 749     TRACE("dcl");
 750
 751     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 752     {
 753         switch (semantic->sampler_type)
 754         {
 755             case WINED3DSTT_2D: TRACE("_2d"); break;
 756             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 757             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 758             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 759         }
 760     }
 761     else
 762     {
 763         /* Pixel shaders 3.0 don't have usage semantics */
 764         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
 765             return;
 766         else
 767             TRACE("_");
 768
 769         switch (semantic->usage)
 770         {
 771             case WINED3DDECLUSAGE_POSITION:
 772                 TRACE("position%d", semantic->usage_idx);
 773                 break;
 774             case WINED3DDECLUSAGE_BLENDINDICES:
 775                 TRACE("blend");
 776                 break;
 777             case WINED3DDECLUSAGE_BLENDWEIGHT:
 778                 TRACE("weight");
 779                 break;
 780             case WINED3DDECLUSAGE_NORMAL:
 781                 TRACE("normal%d", semantic->usage_idx);
 782                 break;
 783             case WINED3DDECLUSAGE_PSIZE:
 784                 TRACE("psize");
 785                 break;
 786             case WINED3DDECLUSAGE_COLOR:
 787                 if (semantic->usage_idx == 0) TRACE("color");
 788                 else TRACE("specular%d", (semantic->usage_idx - 1));
 789                 break;
 790             case WINED3DDECLUSAGE_TEXCOORD:
 791                 TRACE("texture%d", semantic->usage_idx);
 792                 break;
 793             case WINED3DDECLUSAGE_TANGENT:
 794                 TRACE("tangent");
 795                 break;
 796             case WINED3DDECLUSAGE_BINORMAL:
 797                 TRACE("binormal");
 798                 break;
 799             case WINED3DDECLUSAGE_TESSFACTOR:
 800                 TRACE("tessfactor");
 801                 break;
 802             case WINED3DDECLUSAGE_POSITIONT:
 803                 TRACE("positionT%d", semantic->usage_idx);
 804                 break;
 805             case WINED3DDECLUSAGE_FOG:
 806                 TRACE("fog");
 807                 break;
 808             case WINED3DDECLUSAGE_DEPTH:
 809                 TRACE("depth");
 810                 break;
 811             case WINED3DDECLUSAGE_SAMPLE:
 812                 TRACE("sample");
 813                 break;
 814             default:
 815                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 816         }
 817     }
 818 }
 819
 820 static void shader_dump_register(const struct wined3d_shader_register *reg,
 821         const struct wined3d_shader_version *shader_version)
 822 {
 823     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 824     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 825     UINT offset = reg->idx;
 826
 827     switch (reg->type)
 828     {
 829         case WINED3DSPR_TEMP:
 830             TRACE("r");
 831             break;
 832
 833         case WINED3DSPR_INPUT:
 834             TRACE("v");
 835             break;
 836
 837         case WINED3DSPR_CONST:
 838         case WINED3DSPR_CONST2:
 839         case WINED3DSPR_CONST3:
 840         case WINED3DSPR_CONST4:
 841             TRACE("c");
 842             offset = shader_get_float_offset(reg->type, reg->idx);
 843             break;
 844
 845         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 846             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 847             break;
 848
 849         case WINED3DSPR_RASTOUT:
 850             TRACE("%s", rastout_reg_names[reg->idx]);
 851             break;
 852
 853         case WINED3DSPR_COLOROUT:
 854             TRACE("oC");
 855             break;
 856
 857         case WINED3DSPR_DEPTHOUT:
 858             TRACE("oDepth");
 859             break;
 860
 861         case WINED3DSPR_ATTROUT:
 862             TRACE("oD");
 863             break;
 864
 865         case WINED3DSPR_TEXCRDOUT:
 866             /* Vertex shaders >= 3.0 use general purpose output registers
 867              * (WINED3DSPR_OUTPUT), which can include an address token */
 868             if (shader_version->major >= 3) TRACE("o");
 869             else TRACE("oT");
 870             break;
 871
 872         case WINED3DSPR_CONSTINT:
 873             TRACE("i");
 874             break;
 875
 876         case WINED3DSPR_CONSTBOOL:
 877             TRACE("b");
 878             break;
 879
 880         case WINED3DSPR_LABEL:
 881             TRACE("l");
 882             break;
 883
 884         case WINED3DSPR_LOOP:
 885             TRACE("aL");
 886             break;
 887
 888         case WINED3DSPR_SAMPLER:
 889             TRACE("s");
 890             break;
 891
 892         case WINED3DSPR_MISCTYPE:
 893             if (reg->idx > 1) FIXME("Unhandled misctype register %d\n", reg->idx);
 894             else TRACE("%s", misctype_reg_names[reg->idx]);
 895             break;
 896
 897         case WINED3DSPR_PREDICATE:
 898             TRACE("p");
 899             break;
 900
 901         case WINED3DSPR_IMMCONST:
 902             TRACE("l");
 903             break;
 904
 905         default:
 906             TRACE("unhandled_rtype(%#x)", reg->type);
 907             break;
 908     }
 909
 910     if (reg->type == WINED3DSPR_IMMCONST)
 911     {
 912         TRACE("(");
 913         switch (reg->immconst_type)
 914         {
 915             case WINED3D_IMMCONST_FLOAT:
 916                 TRACE("%.8e", *(float *)reg->immconst_data);
 917                 break;
 918
 919             case WINED3D_IMMCONST_FLOAT4:
 920                 TRACE("%.8e, %.8e, %.8e, %.8e",
 921                         *(float *)&reg->immconst_data[0], *(float *)&reg->immconst_data[1],
 922                         *(float *)&reg->immconst_data[2], *(float *)&reg->immconst_data[3]);
 923                 break;
 924
 925             default:
 926                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
 927                 break;
 928         }
 929         TRACE(")");
 930     }
 931     else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
 932     {
 933         if (reg->rel_addr)
 934         {
 935             TRACE("[");
 936             shader_dump_src_param(reg->rel_addr, shader_version);
 937             TRACE(" + ");
 938         }
 939         TRACE("%u", offset);
 940         if (reg->rel_addr) TRACE("]");
 941     }
 942 }
 943
 944 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
 945         const struct wined3d_shader_version *shader_version)
 946 {
 947     DWORD write_mask = param->write_mask;
 948
 949     shader_dump_register(&param->reg, shader_version);
 950
 951     if (write_mask != WINED3DSP_WRITEMASK_ALL)
 952     {
 953         static const char *write_mask_chars = "xyzw";
 954
 955         TRACE(".");
 956         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
 957         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
 958         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
 959         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
 960     }
 961 }
 962
 963 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
 964         const struct wined3d_shader_version *shader_version)
 965 {
 966     DWORD src_modifier = param->modifiers;
 967     DWORD swizzle = param->swizzle;
 968
 969     if (src_modifier == WINED3DSPSM_NEG
 970             || src_modifier == WINED3DSPSM_BIASNEG
 971             || src_modifier == WINED3DSPSM_SIGNNEG
 972             || src_modifier == WINED3DSPSM_X2NEG
 973             || src_modifier == WINED3DSPSM_ABSNEG)
 974         TRACE("-");
 975     else if (src_modifier == WINED3DSPSM_COMP)
 976         TRACE("1-");
 977     else if (src_modifier == WINED3DSPSM_NOT)
 978         TRACE("!");
 979
 980     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
 981         TRACE("abs(");
 982
 983     shader_dump_register(&param->reg, shader_version);
 984
 985     if (src_modifier)
 986     {
 987         switch (src_modifier)
 988         {
 989             case WINED3DSPSM_NONE:    break;
 990             case WINED3DSPSM_NEG:     break;
 991             case WINED3DSPSM_NOT:     break;
 992             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
 993             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
 994             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
 995             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
 996             case WINED3DSPSM_COMP:    break;
 997             case WINED3DSPSM_X2:      TRACE("_x2"); break;
 998             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
 999             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1000             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1001             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1002             case WINED3DSPSM_ABS:     TRACE(")"); break;
1003             default:
1004                                       TRACE("_unknown_modifier(%#x)", src_modifier);
1005         }
1006     }
1007
1008     if (swizzle != WINED3DSP_NOSWIZZLE)
1009     {
1010         static const char *swizzle_chars = "xyzw";
1011         DWORD swizzle_x = swizzle & 0x03;
1012         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1013         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1014         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1015
1016         if (swizzle_x == swizzle_y
1017                 && swizzle_x == swizzle_z
1018                 && swizzle_x == swizzle_w)
1019         {
1020             TRACE(".%c", swizzle_chars[swizzle_x]);
1021         }
1022         else
1023         {
1024             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1025                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1026         }
1027     }
1028 }
1029
1030 /* Shared code in order to generate the bulk of the shader string.
1031  * NOTE: A description of how to parse tokens can be found on msdn */
1032 void shader_generate_main(IWineD3DBaseShader *iface, SHADER_BUFFER *buffer,
1033         const shader_reg_maps *reg_maps, const DWORD *pFunction, void *backend_ctx)
1034 {
1035     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
1036     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
1037     const struct wined3d_shader_frontend *fe = This->baseShader.frontend;
1038     void *fe_data = This->baseShader.frontend_data;
1039     struct wined3d_shader_src_param src_rel_addr[4];
1040     struct wined3d_shader_src_param src_param[4];
1041     struct wined3d_shader_version shader_version;
1042     struct wined3d_shader_src_param dst_rel_addr;
1043     struct wined3d_shader_dst_param dst_param;
1044     struct wined3d_shader_instruction ins;
1045     struct wined3d_shader_context ctx;
1046     const DWORD *pToken = pFunction;
1047     DWORD i;
1048
1049     /* Initialize current parsing state */
1050     ctx.shader = iface;
1051     ctx.reg_maps = reg_maps;
1052     ctx.buffer = buffer;
1053     ctx.backend_data = backend_ctx;
1054
1055     ins.ctx = &ctx;
1056     ins.dst = &dst_param;
1057     ins.src = src_param;
1058     This->baseShader.parse_state.current_row = 0;
1059
1060     fe->shader_read_header(fe_data, &pToken, &shader_version);
1061
1062     while (!fe->shader_is_end(fe_data, &pToken))
1063     {
1064         const char *comment;
1065         UINT param_size;
1066
1067         /* Skip comment tokens */
1068         fe->shader_read_comment(&pToken, &comment);
1069         if (comment) continue;
1070
1071         /* Read opcode */
1072         fe->shader_read_opcode(fe_data, &pToken, &ins, &param_size);
1073
1074         /* Unknown opcode and its parameters */
1075         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1076         {
1077             TRACE("Skipping unrecognized instruction.\n");
1078             pToken += param_size;
1079             continue;
1080         }
1081
1082         /* Nothing to do */
1083         if (ins.handler_idx == WINED3DSIH_DCL
1084                 || ins.handler_idx == WINED3DSIH_NOP
1085                 || ins.handler_idx == WINED3DSIH_DEF
1086                 || ins.handler_idx == WINED3DSIH_DEFI
1087                 || ins.handler_idx == WINED3DSIH_DEFB
1088                 || ins.handler_idx == WINED3DSIH_PHASE)
1089         {
1090             pToken += param_size;
1091             continue;
1092         }
1093
1094         /* Destination token */
1095         if (ins.dst_count) fe->shader_read_dst_param(fe_data, &pToken, &dst_param, &dst_rel_addr);
1096
1097         /* Predication token */
1098         if (ins.predicate) ins.predicate = *pToken++;
1099
1100         /* Other source tokens */
1101         for (i = 0; i < ins.src_count; ++i)
1102         {
1103             fe->shader_read_src_param(fe_data, &pToken, &src_param[i], &src_rel_addr[i]);
1104         }
1105
1106         /* Call appropriate function for output target */
1107         device->shader_backend->shader_handle_instruction(&ins);
1108     }
1109 }
1110
1111 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1112 {
1113     DWORD mmask = dst->modifiers;
1114
1115     switch (dst->shift)
1116     {
1117         case 0: break;
1118         case 13: TRACE("_d8"); break;
1119         case 14: TRACE("_d4"); break;
1120         case 15: TRACE("_d2"); break;
1121         case 1: TRACE("_x2"); break;
1122         case 2: TRACE("_x4"); break;
1123         case 3: TRACE("_x8"); break;
1124         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1125     }
1126
1127     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1128     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1129     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1130
1131     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1132     if (mmask)
1133         FIXME("_unrecognized_modifier(%#x)", mmask);
1134 }
1135
1136 void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *pFunction)
1137 {
1138     struct wined3d_shader_version shader_version;
1139     const DWORD* pToken = pFunction;
1140     const char *type_prefix;
1141     DWORD i;
1142
1143     TRACE("Parsing %p\n", pFunction);
1144
1145     fe->shader_read_header(fe_data, &pToken, &shader_version);
1146
1147     switch (shader_version.type)
1148     {
1149         case WINED3D_SHADER_TYPE_VERTEX:
1150             type_prefix = "vs";
1151             break;
1152
1153         case WINED3D_SHADER_TYPE_GEOMETRY:
1154             type_prefix = "gs";
1155             break;
1156
1157         case WINED3D_SHADER_TYPE_PIXEL:
1158             type_prefix = "ps";
1159             break;
1160
1161         default:
1162             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1163             type_prefix = "unknown";
1164             break;
1165     }
1166
1167     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1168
1169     while (!fe->shader_is_end(fe_data, &pToken))
1170     {
1171         struct wined3d_shader_instruction ins;
1172         const char *comment;
1173         UINT param_size;
1174
1175         /* comment */
1176         fe->shader_read_comment(&pToken, &comment);
1177         if (comment)
1178         {
1179             TRACE("//%s\n", comment);
1180             continue;
1181         }
1182
1183         fe->shader_read_opcode(fe_data, &pToken, &ins, &param_size);
1184         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1185         {
1186             TRACE("Skipping unrecognized instruction.\n");
1187             pToken += param_size;
1188             continue;
1189         }
1190
1191         if (ins.handler_idx == WINED3DSIH_DCL)
1192         {
1193             struct wined3d_shader_semantic semantic;
1194
1195             fe->shader_read_semantic(&pToken, &semantic);
1196
1197             shader_dump_decl_usage(&semantic, &shader_version);
1198             shader_dump_ins_modifiers(&semantic.reg);
1199             TRACE(" ");
1200             shader_dump_dst_param(&semantic.reg, &shader_version);
1201         }
1202         else if (ins.handler_idx == WINED3DSIH_DEF)
1203         {
1204             struct wined3d_shader_dst_param dst;
1205             struct wined3d_shader_src_param rel_addr;
1206
1207             fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
1208
1209             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1210                     *(const float *)(pToken),
1211                     *(const float *)(pToken + 1),
1212                     *(const float *)(pToken + 2),
1213                     *(const float *)(pToken + 3));
1214             pToken += 4;
1215         }
1216         else if (ins.handler_idx == WINED3DSIH_DEFI)
1217         {
1218             struct wined3d_shader_dst_param dst;
1219             struct wined3d_shader_src_param rel_addr;
1220
1221             fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
1222
1223             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1224                     *(pToken),
1225                     *(pToken + 1),
1226                     *(pToken + 2),
1227                     *(pToken + 3));
1228             pToken += 4;
1229         }
1230         else if (ins.handler_idx == WINED3DSIH_DEFB)
1231         {
1232             struct wined3d_shader_dst_param dst;
1233             struct wined3d_shader_src_param rel_addr;
1234
1235             fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
1236
1237             TRACE("defb b%u = %s", dst.reg.idx, *pToken ? "true" : "false");
1238             ++pToken;
1239         }
1240         else
1241         {
1242             struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1243             struct wined3d_shader_dst_param dst_param;
1244             struct wined3d_shader_src_param src_param;
1245
1246             if (ins.dst_count)
1247             {
1248                 fe->shader_read_dst_param(fe_data, &pToken, &dst_param, &dst_rel_addr);
1249             }
1250
1251             /* Print out predication source token first - it follows
1252              * the destination token. */
1253             if (ins.predicate)
1254             {
1255                 fe->shader_read_src_param(fe_data, &pToken, &src_param, &src_rel_addr);
1256                 TRACE("(");
1257                 shader_dump_src_param(&src_param, &shader_version);
1258                 TRACE(") ");
1259             }
1260
1261             /* PixWin marks instructions with the coissue flag with a '+' */
1262             if (ins.coissue) TRACE("+");
1263
1264             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1265
1266             if (ins.handler_idx == WINED3DSIH_IFC
1267                     || ins.handler_idx == WINED3DSIH_BREAKC)
1268             {
1269                 switch (ins.flags)
1270                 {
1271                     case COMPARISON_GT: TRACE("_gt"); break;
1272                     case COMPARISON_EQ: TRACE("_eq"); break;
1273                     case COMPARISON_GE: TRACE("_ge"); break;
1274                     case COMPARISON_LT: TRACE("_lt"); break;
1275                     case COMPARISON_NE: TRACE("_ne"); break;
1276                     case COMPARISON_LE: TRACE("_le"); break;
1277                     default: TRACE("_(%u)", ins.flags);
1278                 }
1279             }
1280             else if (ins.handler_idx == WINED3DSIH_TEX
1281                     && shader_version.major >= 2
1282                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1283             {
1284                 TRACE("p");
1285             }
1286
1287             /* We already read the destination token, print it. */
1288             if (ins.dst_count)
1289             {
1290                 shader_dump_ins_modifiers(&dst_param);
1291                 TRACE(" ");
1292                 shader_dump_dst_param(&dst_param, &shader_version);
1293             }
1294
1295             /* Other source tokens */
1296             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1297             {
1298                 fe->shader_read_src_param(fe_data, &pToken, &src_param, &src_rel_addr);
1299                 TRACE(!i ? " " : ", ");
1300                 shader_dump_src_param(&src_param, &shader_version);
1301             }
1302         }
1303         TRACE("\n");
1304     }
1305 }
1306
1307 void shader_cleanup(IWineD3DBaseShader *iface)
1308 {
1309     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1310
1311     ((IWineD3DDeviceImpl *)This->baseShader.device)->shader_backend->shader_destroy(iface);
1312     HeapFree(GetProcessHeap(), 0, This->baseShader.reg_maps.constf);
1313     HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1314     shader_delete_constant_list(&This->baseShader.constantsF);
1315     shader_delete_constant_list(&This->baseShader.constantsB);
1316     shader_delete_constant_list(&This->baseShader.constantsI);
1317     list_remove(&This->baseShader.shader_list_entry);
1318
1319     if (This->baseShader.frontend && This->baseShader.frontend_data)
1320     {
1321         This->baseShader.frontend->shader_free(This->baseShader.frontend_data);
1322     }
1323 }
1324
1325 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1326 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1327 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1328 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1329 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1330 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1331 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1332 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1333 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1334 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1335 static void shader_none_free(IWineD3DDevice *iface) {}
1336 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1337
1338 #define GLINFO_LOCATION      (*gl_info)
1339 static void shader_none_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
1340 {
1341     /* Set the shader caps to 0 for the none shader backend */
1342     pCaps->VertexShaderVersion  = 0;
1343     pCaps->PixelShaderVersion    = 0;
1344     pCaps->PixelShader1xMaxValue = 0.0f;
1345 }
1346 #undef GLINFO_LOCATION
1347 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1348 {
1349     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1350     {
1351         TRACE("Checking support for fixup:\n");
1352         dump_color_fixup_desc(fixup);
1353     }
1354
1355     /* Faked to make some apps happy. */
1356     if (!is_yuv_fixup(fixup))
1357     {
1358         TRACE("[OK]\n");
1359         return TRUE;
1360     }
1361
1362     TRACE("[FAILED]\n");
1363     return FALSE;
1364 }
1365
1366 const shader_backend_t none_shader_backend = {
1367     shader_none_handle_instruction,
1368     shader_none_select,
1369     shader_none_select_depth_blt,
1370     shader_none_deselect_depth_blt,
1371     shader_none_update_float_vertex_constants,
1372     shader_none_update_float_pixel_constants,
1373     shader_none_load_constants,
1374     shader_none_load_np2fixup_constants,
1375     shader_none_destroy,
1376     shader_none_alloc,
1377     shader_none_free,
1378     shader_none_dirty_const,
1379     shader_none_get_caps,
1380     shader_none_color_fixup_supported,
1381 };