git.oblomov.eu Git - wine/blob - dlls/wined3d/shader.c

   1 /*
   2  * Copyright 2002-2003 Jason Edmeades
   3  * Copyright 2002-2003 Raphael Junqueira
   4  * Copyright 2004 Christian Costa
   5  * Copyright 2005 Oliver Stieber
   6  * Copyright 2006 Ivan Gyurdiev
   7  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
   8  * Copyright 2009 Henri Verbeet for CodeWeavers
   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
  27 #include <math.h>
  28 #include <stdio.h>
  29 #include <string.h>
  30
  31 #include "wined3d_private.h"
  32
  33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  34 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  35
  36 static const char * const shader_opcode_names[] =
  37 {
  38     /* WINED3DSIH_ABS           */ "abs",
  39     /* WINED3DSIH_ADD           */ "add",
  40     /* WINED3DSIH_AND           */ "and",
  41     /* WINED3DSIH_BEM           */ "bem",
  42     /* WINED3DSIH_BREAK         */ "break",
  43     /* WINED3DSIH_BREAKC        */ "breakc",
  44     /* WINED3DSIH_BREAKP        */ "breakp",
  45     /* WINED3DSIH_CALL          */ "call",
  46     /* WINED3DSIH_CALLNZ        */ "callnz",
  47     /* WINED3DSIH_CMP           */ "cmp",
  48     /* WINED3DSIH_CND           */ "cnd",
  49     /* WINED3DSIH_CRS           */ "crs",
  50     /* WINED3DSIH_CUT           */ "cut",
  51     /* WINED3DSIH_DCL           */ "dcl",
  52     /* WINED3DSIH_DEF           */ "def",
  53     /* WINED3DSIH_DEFB          */ "defb",
  54     /* WINED3DSIH_DEFI          */ "defi",
  55     /* WINED3DSIH_DIV           */ "div",
  56     /* WINED3DSIH_DP2ADD        */ "dp2add",
  57     /* WINED3DSIH_DP3           */ "dp3",
  58     /* WINED3DSIH_DP4           */ "dp4",
  59     /* WINED3DSIH_DST           */ "dst",
  60     /* WINED3DSIH_DSX           */ "dsx",
  61     /* WINED3DSIH_DSY           */ "dsy",
  62     /* WINED3DSIH_ELSE          */ "else",
  63     /* WINED3DSIH_EMIT          */ "emit",
  64     /* WINED3DSIH_ENDIF         */ "endif",
  65     /* WINED3DSIH_ENDLOOP       */ "endloop",
  66     /* WINED3DSIH_ENDREP        */ "endrep",
  67     /* WINED3DSIH_EXP           */ "exp",
  68     /* WINED3DSIH_EXPP          */ "expp",
  69     /* WINED3DSIH_FRC           */ "frc",
  70     /* WINED3DSIH_FTOI          */ "ftoi",
  71     /* WINED3DSIH_IADD          */ "iadd",
  72     /* WINED3DSIH_IEQ           */ "ieq",
  73     /* WINED3DSIH_IF            */ "if",
  74     /* WINED3DSIH_IFC           */ "ifc",
  75     /* WINED3DSIH_IGE           */ "ige",
  76     /* WINED3DSIH_IMUL          */ "imul",
  77     /* WINED3DSIH_ITOF          */ "itof",
  78     /* WINED3DSIH_LABEL         */ "label",
  79     /* WINED3DSIH_LD            */ "ld",
  80     /* WINED3DSIH_LIT           */ "lit",
  81     /* WINED3DSIH_LOG           */ "log",
  82     /* WINED3DSIH_LOGP          */ "logp",
  83     /* WINED3DSIH_LOOP          */ "loop",
  84     /* WINED3DSIH_LRP           */ "lrp",
  85     /* WINED3DSIH_LT            */ "lt",
  86     /* WINED3DSIH_M3x2          */ "m3x2",
  87     /* WINED3DSIH_M3x3          */ "m3x3",
  88     /* WINED3DSIH_M3x4          */ "m3x4",
  89     /* WINED3DSIH_M4x3          */ "m4x3",
  90     /* WINED3DSIH_M4x4          */ "m4x4",
  91     /* WINED3DSIH_MAD           */ "mad",
  92     /* WINED3DSIH_MAX           */ "max",
  93     /* WINED3DSIH_MIN           */ "min",
  94     /* WINED3DSIH_MOV           */ "mov",
  95     /* WINED3DSIH_MOVA          */ "mova",
  96     /* WINED3DSIH_MOVC          */ "movc",
  97     /* WINED3DSIH_MUL           */ "mul",
  98     /* WINED3DSIH_NOP           */ "nop",
  99     /* WINED3DSIH_NRM           */ "nrm",
 100     /* WINED3DSIH_PHASE         */ "phase",
 101     /* WINED3DSIH_POW           */ "pow",
 102     /* WINED3DSIH_RCP           */ "rcp",
 103     /* WINED3DSIH_REP           */ "rep",
 104     /* WINED3DSIH_RET           */ "ret",
 105     /* WINED3DSIH_RSQ           */ "rsq",
 106     /* WINED3DSIH_SAMPLE        */ "sample",
 107     /* WINED3DSIH_SAMPLE_GRAD   */ "sample_d",
 108     /* WINED3DSIH_SAMPLE_LOD    */ "sample_l",
 109     /* WINED3DSIH_SETP          */ "setp",
 110     /* WINED3DSIH_SGE           */ "sge",
 111     /* WINED3DSIH_SGN           */ "sgn",
 112     /* WINED3DSIH_SINCOS        */ "sincos",
 113     /* WINED3DSIH_SLT           */ "slt",
 114     /* WINED3DSIH_SUB           */ "sub",
 115     /* WINED3DSIH_TEX           */ "texld",
 116     /* WINED3DSIH_TEXBEM        */ "texbem",
 117     /* WINED3DSIH_TEXBEML       */ "texbeml",
 118     /* WINED3DSIH_TEXCOORD      */ "texcrd",
 119     /* WINED3DSIH_TEXDEPTH      */ "texdepth",
 120     /* WINED3DSIH_TEXDP3        */ "texdp3",
 121     /* WINED3DSIH_TEXDP3TEX     */ "texdp3tex",
 122     /* WINED3DSIH_TEXKILL       */ "texkill",
 123     /* WINED3DSIH_TEXLDD        */ "texldd",
 124     /* WINED3DSIH_TEXLDL        */ "texldl",
 125     /* WINED3DSIH_TEXM3x2DEPTH  */ "texm3x2depth",
 126     /* WINED3DSIH_TEXM3x2PAD    */ "texm3x2pad",
 127     /* WINED3DSIH_TEXM3x2TEX    */ "texm3x2tex",
 128     /* WINED3DSIH_TEXM3x3       */ "texm3x3",
 129     /* WINED3DSIH_TEXM3x3DIFF   */ "texm3x3diff",
 130     /* WINED3DSIH_TEXM3x3PAD    */ "texm3x3pad",
 131     /* WINED3DSIH_TEXM3x3SPEC   */ "texm3x3spec",
 132     /* WINED3DSIH_TEXM3x3TEX    */ "texm3x3tex",
 133     /* WINED3DSIH_TEXM3x3VSPEC  */ "texm3x3vspec",
 134     /* WINED3DSIH_TEXREG2AR     */ "texreg2ar",
 135     /* WINED3DSIH_TEXREG2GB     */ "texreg2gb",
 136     /* WINED3DSIH_TEXREG2RGB    */ "texreg2rgb",
 137     /* WINED3DSIH_UTOF          */ "utof",
 138 };
 139
 140 static const char * const semantic_names[] =
 141 {
 142     /* WINED3DDECLUSAGE_POSITION        */ "SV_POSITION",
 143     /* WINED3DDECLUSAGE_BLENDWEIGHT     */ "BLENDWEIGHT",
 144     /* WINED3DDECLUSAGE_BLENDINDICES    */ "BLENDINDICES",
 145     /* WINED3DDECLUSAGE_NORMAL          */ "NORMAL",
 146     /* WINED3DDECLUSAGE_PSIZE           */ "PSIZE",
 147     /* WINED3DDECLUSAGE_TEXCOORD        */ "TEXCOORD",
 148     /* WINED3DDECLUSAGE_TANGENT         */ "TANGENT",
 149     /* WINED3DDECLUSAGE_BINORMAL        */ "BINORMAL",
 150     /* WINED3DDECLUSAGE_TESSFACTOR      */ "TESSFACTOR",
 151     /* WINED3DDECLUSAGE_POSITIONT       */ "POSITIONT",
 152     /* WINED3DDECLUSAGE_COLOR           */ "COLOR",
 153     /* WINED3DDECLUSAGE_FOG             */ "FOG",
 154     /* WINED3DDECLUSAGE_DEPTH           */ "DEPTH",
 155     /* WINED3DDECLUSAGE_SAMPLE          */ "SAMPLE",
 156 };
 157
 158 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
 159 {
 160     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 161     {
 162         FIXME("Unrecognized usage %#x.\n", usage);
 163         return "UNRECOGNIZED";
 164     }
 165
 166     return semantic_names[usage];
 167 }
 168
 169 static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
 170 {
 171     unsigned int i;
 172
 173     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 174     {
 175         if (!strcmp(name, semantic_names[i])) return i;
 176     }
 177
 178     return ~0U;
 179 }
 180
 181 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
 182 {
 183     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 184 }
 185
 186 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 187         const struct wined3d_shader_semantic *s)
 188 {
 189     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 190     e->semantic_idx = s->usage_idx;
 191     e->sysval_semantic = 0;
 192     e->component_type = 0;
 193     e->register_idx = s->reg.reg.idx;
 194     e->mask = s->reg.write_mask;
 195 }
 196
 197 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
 198         WINED3DDECLUSAGE usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
 199 {
 200     e->semantic_name = shader_semantic_name_from_usage(usage);
 201     e->semantic_idx = usage_idx;
 202     e->sysval_semantic = 0;
 203     e->component_type = 0;
 204     e->register_idx = reg_idx;
 205     e->mask = write_mask;
 206 }
 207
 208 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 209 {
 210     switch (version_token >> 16)
 211     {
 212         case WINED3D_SM1_VS:
 213         case WINED3D_SM1_PS:
 214             return &sm1_shader_frontend;
 215
 216         case WINED3D_SM4_PS:
 217         case WINED3D_SM4_VS:
 218         case WINED3D_SM4_GS:
 219             return &sm4_shader_frontend;
 220
 221         default:
 222             FIXME("Unrecognised version token %#x\n", version_token);
 223             return NULL;
 224     }
 225 }
 226
 227 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 228 {
 229     buffer->buffer[0] = '\0';
 230     buffer->bsize = 0;
 231     buffer->lineNo = 0;
 232     buffer->newline = TRUE;
 233 }
 234
 235 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 236 {
 237     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 238     if (!buffer->buffer)
 239     {
 240         ERR("Failed to allocate shader buffer memory.\n");
 241         return FALSE;
 242     }
 243
 244     shader_buffer_clear(buffer);
 245     return TRUE;
 246 }
 247
 248 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 249 {
 250     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 251 }
 252
 253 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 254 {
 255     char *base = buffer->buffer + buffer->bsize;
 256     int rc;
 257
 258     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 259
 260     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 261     {
 262         ERR("The buffer allocated for the shader program string "
 263             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 264         buffer->bsize = SHADER_PGMSIZE - 1;
 265         return -1;
 266     }
 267
 268     if (buffer->newline)
 269     {
 270         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 271         buffer->newline = FALSE;
 272     }
 273     else
 274     {
 275         TRACE("%s", base);
 276     }
 277
 278     buffer->bsize += rc;
 279     if (buffer->buffer[buffer->bsize-1] == '\n')
 280     {
 281         ++buffer->lineNo;
 282         buffer->newline = TRUE;
 283     }
 284
 285     return 0;
 286 }
 287
 288 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 289 {
 290     va_list args;
 291     int ret;
 292
 293     va_start(args, format);
 294     ret = shader_vaddline(buffer, format, args);
 295     va_end(args);
 296
 297     return ret;
 298 }
 299
 300 static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
 301         void *parent, const struct wined3d_parent_ops *parent_ops)
 302 {
 303     shader->ref = 1;
 304     shader->device = (IWineD3DDevice *)device;
 305     shader->parent = parent;
 306     shader->parent_ops = parent_ops;
 307     list_init(&shader->linked_programs);
 308     list_add_head(&device->shaders, &shader->shader_list_entry);
 309 }
 310
 311 /* Convert floating point offset relative to a register file to an absolute
 312  * offset for float constants. */
 313 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
 314 {
 315     switch (register_type)
 316     {
 317         case WINED3DSPR_CONST: return register_idx;
 318         case WINED3DSPR_CONST2: return 2048 + register_idx;
 319         case WINED3DSPR_CONST3: return 4096 + register_idx;
 320         case WINED3DSPR_CONST4: return 6144 + register_idx;
 321         default:
 322             FIXME("Unsupported register type: %u.\n", register_type);
 323             return register_idx;
 324     }
 325 }
 326
 327 static void shader_delete_constant_list(struct list *clist)
 328 {
 329     struct local_constant *constant;
 330     struct list *ptr;
 331
 332     ptr = list_head(clist);
 333     while (ptr)
 334     {
 335         constant = LIST_ENTRY(ptr, struct local_constant, entry);
 336         ptr = list_next(clist, ptr);
 337         HeapFree(GetProcessHeap(), 0, constant);
 338     }
 339     list_init(clist);
 340 }
 341
 342 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 343 {
 344     DWORD idx, shift;
 345     idx = bit >> 5;
 346     shift = bit & 0x1f;
 347     bitmap[idx] |= (1 << shift);
 348 }
 349
 350 static void shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct shader_reg_maps *reg_maps,
 351         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 352 {
 353     switch (reg->type)
 354     {
 355         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 356             if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
 357             else reg_maps->address |= 1 << reg->idx;
 358             break;
 359
 360         case WINED3DSPR_TEMP:
 361             reg_maps->temporary |= 1 << reg->idx;
 362             break;
 363
 364         case WINED3DSPR_INPUT:
 365             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 366             {
 367                 if (reg->rel_addr)
 368                 {
 369                     /* If relative addressing is used, we must assume that all registers
 370                      * are used. Even if it is a construct like v3[aL], we can't assume
 371                      * that v0, v1 and v2 aren't read because aL can be negative */
 372                     unsigned int i;
 373                     for (i = 0; i < MAX_REG_INPUT; ++i)
 374                     {
 375                         ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
 376                     }
 377                 }
 378                 else
 379                 {
 380                     ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg->idx] = TRUE;
 381                 }
 382             }
 383             else reg_maps->input_registers |= 1 << reg->idx;
 384             break;
 385
 386         case WINED3DSPR_RASTOUT:
 387             if (reg->idx == 1) reg_maps->fog = 1;
 388             break;
 389
 390         case WINED3DSPR_MISCTYPE:
 391             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 392             {
 393                 if (!reg->idx) reg_maps->vpos = 1;
 394                 else if (reg->idx == 1) reg_maps->usesfacing = 1;
 395             }
 396             break;
 397
 398         case WINED3DSPR_CONST:
 399             if (reg->rel_addr)
 400             {
 401                 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
 402                 {
 403                     if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
 404                     {
 405                         ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
 406                     }
 407                     if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
 408                     {
 409                         ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
 410                     }
 411                 }
 412                 reg_maps->usesrelconstF = TRUE;
 413             }
 414             else
 415             {
 416                 set_bitmap_bit(reg_maps->constf, reg->idx);
 417             }
 418             break;
 419
 420         case WINED3DSPR_CONSTINT:
 421             reg_maps->integer_constants |= (1 << reg->idx);
 422             break;
 423
 424         case WINED3DSPR_CONSTBOOL:
 425             reg_maps->boolean_constants |= (1 << reg->idx);
 426             break;
 427
 428         case WINED3DSPR_COLOROUT:
 429             reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
 430             break;
 431
 432         default:
 433             TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
 434             break;
 435     }
 436 }
 437
 438 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 439 {
 440     switch (instr)
 441     {
 442         case WINED3DSIH_M4x4:
 443         case WINED3DSIH_M3x4:
 444             return param == 1 ? 3 : 0;
 445
 446         case WINED3DSIH_M4x3:
 447         case WINED3DSIH_M3x3:
 448             return param == 1 ? 2 : 0;
 449
 450         case WINED3DSIH_M3x2:
 451             return param == 1 ? 1 : 0;
 452
 453         default:
 454             return 0;
 455     }
 456 }
 457
 458 /* Note that this does not count the loop register as an address register. */
 459 static HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
 460         struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 461         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 462 {
 463     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
 464     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 465     void *fe_data = shader->baseShader.frontend_data;
 466     struct wined3d_shader_version shader_version;
 467     const DWORD *ptr = byte_code;
 468
 469     memset(reg_maps, 0, sizeof(*reg_maps));
 470
 471     /* get_registers_used() is called on every compile on some 1.x shaders,
 472      * which can result in stacking up a collection of local constants.
 473      * Delete the old constants if existing. */
 474     shader_delete_constant_list(&shader->baseShader.constantsF);
 475     shader_delete_constant_list(&shader->baseShader.constantsB);
 476     shader_delete_constant_list(&shader->baseShader.constantsI);
 477
 478     fe->shader_read_header(fe_data, &ptr, &shader_version);
 479     reg_maps->shader_version = shader_version;
 480
 481     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 482             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 483     if (!reg_maps->constf)
 484     {
 485         ERR("Failed to allocate constant map memory.\n");
 486         return E_OUTOFMEMORY;
 487     }
 488
 489     while (!fe->shader_is_end(fe_data, &ptr))
 490     {
 491         struct wined3d_shader_instruction ins;
 492         const char *comment;
 493         UINT comment_size;
 494         UINT param_size;
 495
 496         /* Skip comments. */
 497         fe->shader_read_comment(&ptr, &comment, &comment_size);
 498         if (comment) continue;
 499
 500         /* Fetch opcode. */
 501         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
 502
 503         /* Unhandled opcode, and its parameters. */
 504         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 505         {
 506             TRACE("Skipping unrecognized instruction.\n");
 507             ptr += param_size;
 508             continue;
 509         }
 510
 511         /* Handle declarations. */
 512         if (ins.handler_idx == WINED3DSIH_DCL)
 513         {
 514             struct wined3d_shader_semantic semantic;
 515
 516             fe->shader_read_semantic(&ptr, &semantic);
 517
 518             switch (semantic.reg.reg.type)
 519             {
 520                 /* Mark input registers used. */
 521                 case WINED3DSPR_INPUT:
 522                     reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
 523                     shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
 524                     break;
 525
 526                 /* Vertex shader: mark 3.0 output registers used, save token. */
 527                 case WINED3DSPR_OUTPUT:
 528                     reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
 529                     shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
 530                     if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
 531                     break;
 532
 533                 /* Save sampler usage token. */
 534                 case WINED3DSPR_SAMPLER:
 535                     reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
 536                     break;
 537
 538                 default:
 539                     TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
 540                     break;
 541             }
 542         }
 543         else if (ins.handler_idx == WINED3DSIH_DEF)
 544         {
 545             struct wined3d_shader_src_param rel_addr;
 546             struct wined3d_shader_dst_param dst;
 547
 548             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 549             if (!lconst) return E_OUTOFMEMORY;
 550
 551             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 552             lconst->idx = dst.reg.idx;
 553
 554             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 555             ptr += 4;
 556
 557             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 558             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 559             {
 560                 float *value = (float *)lconst->value;
 561                 if (value[0] < -1.0f) value[0] = -1.0f;
 562                 else if (value[0] > 1.0f) value[0] = 1.0f;
 563                 if (value[1] < -1.0f) value[1] = -1.0f;
 564                 else if (value[1] > 1.0f) value[1] = 1.0f;
 565                 if (value[2] < -1.0f) value[2] = -1.0f;
 566                 else if (value[2] > 1.0f) value[2] = 1.0f;
 567                 if (value[3] < -1.0f) value[3] = -1.0f;
 568                 else if (value[3] > 1.0f) value[3] = 1.0f;
 569             }
 570
 571             list_add_head(&shader->baseShader.constantsF, &lconst->entry);
 572         }
 573         else if (ins.handler_idx == WINED3DSIH_DEFI)
 574         {
 575             struct wined3d_shader_src_param rel_addr;
 576             struct wined3d_shader_dst_param dst;
 577
 578             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 579             if (!lconst) return E_OUTOFMEMORY;
 580
 581             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 582             lconst->idx = dst.reg.idx;
 583
 584             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 585             ptr += 4;
 586
 587             list_add_head(&shader->baseShader.constantsI, &lconst->entry);
 588             reg_maps->local_int_consts |= (1 << dst.reg.idx);
 589         }
 590         else if (ins.handler_idx == WINED3DSIH_DEFB)
 591         {
 592             struct wined3d_shader_src_param rel_addr;
 593             struct wined3d_shader_dst_param dst;
 594
 595             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 596             if (!lconst) return E_OUTOFMEMORY;
 597
 598             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 599             lconst->idx = dst.reg.idx;
 600
 601             memcpy(lconst->value, ptr, sizeof(DWORD));
 602             ++ptr;
 603
 604             list_add_head(&shader->baseShader.constantsB, &lconst->entry);
 605             reg_maps->local_bool_consts |= (1 << dst.reg.idx);
 606         }
 607         /* If there's a loop in the shader. */
 608         else if (ins.handler_idx == WINED3DSIH_LOOP
 609                 || ins.handler_idx == WINED3DSIH_REP)
 610         {
 611             struct wined3d_shader_src_param src, rel_addr;
 612
 613             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 614
 615             /* Rep and Loop always use an integer constant for the control parameters. */
 616             if (ins.handler_idx == WINED3DSIH_REP)
 617             {
 618                 reg_maps->integer_constants |= 1 << src.reg.idx;
 619             }
 620             else
 621             {
 622                 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 623                 reg_maps->integer_constants |= 1 << src.reg.idx;
 624             }
 625
 626             cur_loop_depth++;
 627             if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
 628         }
 629         else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 630                 || ins.handler_idx == WINED3DSIH_ENDREP)
 631         {
 632             cur_loop_depth--;
 633         }
 634         /* For subroutine prototypes. */
 635         else if (ins.handler_idx == WINED3DSIH_LABEL)
 636         {
 637             struct wined3d_shader_src_param src, rel_addr;
 638
 639             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 640             reg_maps->labels |= 1 << src.reg.idx;
 641         }
 642         /* Set texture, address, temporary registers. */
 643         else
 644         {
 645             BOOL color0_mov = FALSE;
 646             int i, limit;
 647
 648             /* This will loop over all the registers and try to
 649              * make a bitmask of the ones we're interested in.
 650              *
 651              * Relative addressing tokens are ignored, but that's
 652              * okay, since we'll catch any address registers when
 653              * they are initialized (required by spec). */
 654             for (i = 0; i < ins.dst_count; ++i)
 655             {
 656                 struct wined3d_shader_src_param dst_rel_addr;
 657                 struct wined3d_shader_dst_param dst_param;
 658
 659                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
 660
 661                 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
 662
 663                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 664                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 665                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 666                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 667                 {
 668                     UINT idx = dst_param.reg.idx;
 669
 670                     switch (dst_param.reg.type)
 671                     {
 672                         case WINED3DSPR_RASTOUT:
 673                             switch (idx)
 674                             {
 675                                 case 0: /* oPos */
 676                                     reg_maps->output_registers |= 1 << 10;
 677                                     shader_signature_from_usage(&output_signature[10],
 678                                             WINED3DDECLUSAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 679                                     break;
 680
 681                                 case 1: /* oFog */
 682                                     reg_maps->output_registers |= 1 << 11;
 683                                     shader_signature_from_usage(&output_signature[11],
 684                                             WINED3DDECLUSAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 685                                     break;
 686
 687                                 case 2: /* oPts */
 688                                     reg_maps->output_registers |= 1 << 11;
 689                                     shader_signature_from_usage(&output_signature[11],
 690                                             WINED3DDECLUSAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 691                                     break;
 692                             }
 693                             break;
 694
 695                         case WINED3DSPR_ATTROUT:
 696                             if (idx < 2)
 697                             {
 698                                 idx += 8;
 699                                 if (reg_maps->output_registers & (1 << idx))
 700                                 {
 701                                     output_signature[idx].mask |= dst_param.write_mask;
 702                                 }
 703                                 else
 704                                 {
 705                                     reg_maps->output_registers |= 1 << idx;
 706                                     shader_signature_from_usage(&output_signature[idx],
 707                                             WINED3DDECLUSAGE_COLOR, idx - 8, idx, dst_param.write_mask);
 708                                 }
 709                             }
 710                             break;
 711
 712                         case WINED3DSPR_TEXCRDOUT:
 713
 714                             reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
 715                             if (reg_maps->output_registers & (1 << idx))
 716                             {
 717                                 output_signature[idx].mask |= dst_param.write_mask;
 718                             }
 719                             else
 720                             {
 721                                 reg_maps->output_registers |= 1 << idx;
 722                                 shader_signature_from_usage(&output_signature[idx],
 723                                         WINED3DDECLUSAGE_TEXCOORD, idx, idx, dst_param.write_mask);
 724                             }
 725                             break;
 726
 727                         default:
 728                             break;
 729                     }
 730                 }
 731
 732                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 733                 {
 734                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 735
 736                     if (dst_param.reg.type == WINED3DSPR_COLOROUT && !dst_param.reg.idx)
 737                     {
 738                     /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 739                      * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 740                      * the mov and perform the sRGB write correction from the source register.
 741                      *
 742                      * However, if the mov is only partial, we can't do this, and if the write
 743                      * comes from an instruction other than MOV it is hard to do as well. If
 744                      * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 745
 746                         ps->color0_mov = FALSE;
 747                         if (ins.handler_idx == WINED3DSIH_MOV
 748                                 && dst_param.write_mask == WINED3DSP_WRITEMASK_ALL)
 749                         {
 750                             /* Used later when the source register is read. */
 751                             color0_mov = TRUE;
 752                         }
 753                     }
 754                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 755                      * end
 756                      */
 757                     else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
 758                     {
 759                         ps->color0_mov = FALSE;
 760                     }
 761                 }
 762
 763                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 764                 if (shader_version.major == 1
 765                         && (ins.handler_idx == WINED3DSIH_TEX
 766                             || ins.handler_idx == WINED3DSIH_TEXBEM
 767                             || ins.handler_idx == WINED3DSIH_TEXBEML
 768                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 769                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 770                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 771                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 772                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 773                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 774                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 775                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 776                 {
 777                     /* Fake sampler usage, only set reserved bit and type. */
 778                     DWORD sampler_code = dst_param.reg.idx;
 779
 780                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 781                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 782
 783                     /* texbem is only valid with < 1.4 pixel shaders */
 784                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 785                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 786                     {
 787                         reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 788                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 789                         {
 790                             reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
 791                         }
 792                     }
 793                 }
 794                 else if (ins.handler_idx == WINED3DSIH_BEM)
 795                 {
 796                     reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 797                 }
 798             }
 799
 800             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 801             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 802             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 803             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 804             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 805             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 806             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 807             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 808             else if (ins.handler_idx == WINED3DSIH_RCP) reg_maps->usesrcp = 1;
 809
 810             limit = ins.src_count + (ins.predicate ? 1 : 0);
 811             for (i = 0; i < limit; ++i)
 812             {
 813                 struct wined3d_shader_src_param src_param, src_rel_addr;
 814                 unsigned int count;
 815
 816                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
 817                 count = get_instr_extra_regcount(ins.handler_idx, i);
 818
 819                 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 820                 while (count)
 821                 {
 822                     ++src_param.reg.idx;
 823                     shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 824                     --count;
 825                 }
 826
 827                 if (color0_mov)
 828                 {
 829                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 830                     if (src_param.reg.type == WINED3DSPR_TEMP
 831                             && src_param.swizzle == WINED3DSP_NOSWIZZLE)
 832                     {
 833                         ps->color0_mov = TRUE;
 834                         ps->color0_reg = src_param.reg.idx;
 835                     }
 836                 }
 837             }
 838         }
 839     }
 840     reg_maps->loop_depth = max_loop_depth;
 841
 842     shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
 843
 844     return WINED3D_OK;
 845 }
 846
 847 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
 848 {
 849     DWORD map = 1 << max;
 850     map |= map - 1;
 851     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 852
 853     return wined3d_log2i(map);
 854 }
 855
 856 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 857         const struct wined3d_shader_version *shader_version)
 858 {
 859     TRACE("dcl");
 860
 861     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 862     {
 863         switch (semantic->sampler_type)
 864         {
 865             case WINED3DSTT_2D: TRACE("_2d"); break;
 866             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 867             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 868             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 869         }
 870     }
 871     else
 872     {
 873         /* Pixel shaders 3.0 don't have usage semantics. */
 874         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 875         else TRACE("_");
 876
 877         switch (semantic->usage)
 878         {
 879             case WINED3DDECLUSAGE_POSITION:
 880                 TRACE("position%u", semantic->usage_idx);
 881                 break;
 882
 883             case WINED3DDECLUSAGE_BLENDINDICES:
 884                 TRACE("blend");
 885                 break;
 886
 887             case WINED3DDECLUSAGE_BLENDWEIGHT:
 888                 TRACE("weight");
 889                 break;
 890
 891             case WINED3DDECLUSAGE_NORMAL:
 892                 TRACE("normal%u", semantic->usage_idx);
 893                 break;
 894
 895             case WINED3DDECLUSAGE_PSIZE:
 896                 TRACE("psize");
 897                 break;
 898
 899             case WINED3DDECLUSAGE_COLOR:
 900                 if (!semantic->usage_idx) TRACE("color");
 901                 else TRACE("specular%u", (semantic->usage_idx - 1));
 902                 break;
 903
 904             case WINED3DDECLUSAGE_TEXCOORD:
 905                 TRACE("texture%u", semantic->usage_idx);
 906                 break;
 907
 908             case WINED3DDECLUSAGE_TANGENT:
 909                 TRACE("tangent");
 910                 break;
 911
 912             case WINED3DDECLUSAGE_BINORMAL:
 913                 TRACE("binormal");
 914                 break;
 915
 916             case WINED3DDECLUSAGE_TESSFACTOR:
 917                 TRACE("tessfactor");
 918                 break;
 919
 920             case WINED3DDECLUSAGE_POSITIONT:
 921                 TRACE("positionT%u", semantic->usage_idx);
 922                 break;
 923
 924             case WINED3DDECLUSAGE_FOG:
 925                 TRACE("fog");
 926                 break;
 927
 928             case WINED3DDECLUSAGE_DEPTH:
 929                 TRACE("depth");
 930                 break;
 931
 932             case WINED3DDECLUSAGE_SAMPLE:
 933                 TRACE("sample");
 934                 break;
 935
 936             default:
 937                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 938         }
 939     }
 940 }
 941
 942 static void shader_dump_register(const struct wined3d_shader_register *reg,
 943         const struct wined3d_shader_version *shader_version)
 944 {
 945     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 946     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 947     UINT offset = reg->idx;
 948
 949     switch (reg->type)
 950     {
 951         case WINED3DSPR_TEMP:
 952             TRACE("r");
 953             break;
 954
 955         case WINED3DSPR_INPUT:
 956             TRACE("v");
 957             break;
 958
 959         case WINED3DSPR_CONST:
 960         case WINED3DSPR_CONST2:
 961         case WINED3DSPR_CONST3:
 962         case WINED3DSPR_CONST4:
 963             TRACE("c");
 964             offset = shader_get_float_offset(reg->type, reg->idx);
 965             break;
 966
 967         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 968             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 969             break;
 970
 971         case WINED3DSPR_RASTOUT:
 972             TRACE("%s", rastout_reg_names[reg->idx]);
 973             break;
 974
 975         case WINED3DSPR_COLOROUT:
 976             TRACE("oC");
 977             break;
 978
 979         case WINED3DSPR_DEPTHOUT:
 980             TRACE("oDepth");
 981             break;
 982
 983         case WINED3DSPR_ATTROUT:
 984             TRACE("oD");
 985             break;
 986
 987         case WINED3DSPR_TEXCRDOUT:
 988             /* Vertex shaders >= 3.0 use general purpose output registers
 989              * (WINED3DSPR_OUTPUT), which can include an address token. */
 990             if (shader_version->major >= 3) TRACE("o");
 991             else TRACE("oT");
 992             break;
 993
 994         case WINED3DSPR_CONSTINT:
 995             TRACE("i");
 996             break;
 997
 998         case WINED3DSPR_CONSTBOOL:
 999             TRACE("b");
1000             break;
1001
1002         case WINED3DSPR_LABEL:
1003             TRACE("l");
1004             break;
1005
1006         case WINED3DSPR_LOOP:
1007             TRACE("aL");
1008             break;
1009
1010         case WINED3DSPR_SAMPLER:
1011             TRACE("s");
1012             break;
1013
1014         case WINED3DSPR_MISCTYPE:
1015             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
1016             else TRACE("%s", misctype_reg_names[reg->idx]);
1017             break;
1018
1019         case WINED3DSPR_PREDICATE:
1020             TRACE("p");
1021             break;
1022
1023         case WINED3DSPR_IMMCONST:
1024             TRACE("l");
1025             break;
1026
1027         case WINED3DSPR_CONSTBUFFER:
1028             TRACE("cb");
1029             break;
1030
1031         case WINED3DSPR_NULL:
1032             TRACE("null");
1033             break;
1034
1035         case WINED3DSPR_RESOURCE:
1036             TRACE("t");
1037             break;
1038
1039         default:
1040             TRACE("unhandled_rtype(%#x)", reg->type);
1041             break;
1042     }
1043
1044     if (reg->type == WINED3DSPR_IMMCONST)
1045     {
1046         TRACE("(");
1047         switch (reg->immconst_type)
1048         {
1049             case WINED3D_IMMCONST_SCALAR:
1050                 TRACE("%.8e", *(const float *)reg->immconst_data);
1051                 break;
1052
1053             case WINED3D_IMMCONST_VEC4:
1054                 TRACE("%.8e, %.8e, %.8e, %.8e",
1055                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1056                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1057                 break;
1058
1059             default:
1060                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1061                 break;
1062         }
1063         TRACE(")");
1064     }
1065     else if (reg->type != WINED3DSPR_RASTOUT
1066             && reg->type != WINED3DSPR_MISCTYPE
1067             && reg->type != WINED3DSPR_NULL)
1068     {
1069         if (reg->array_idx != ~0U)
1070         {
1071             TRACE("%u[%u", offset, reg->array_idx);
1072             if (reg->rel_addr)
1073             {
1074                 TRACE(" + ");
1075                 shader_dump_src_param(reg->rel_addr, shader_version);
1076             }
1077             TRACE("]");
1078         }
1079         else
1080         {
1081             if (reg->rel_addr)
1082             {
1083                 TRACE("[");
1084                 shader_dump_src_param(reg->rel_addr, shader_version);
1085                 TRACE(" + ");
1086             }
1087             TRACE("%u", offset);
1088             if (reg->rel_addr) TRACE("]");
1089         }
1090     }
1091 }
1092
1093 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1094         const struct wined3d_shader_version *shader_version)
1095 {
1096     DWORD write_mask = param->write_mask;
1097
1098     shader_dump_register(&param->reg, shader_version);
1099
1100     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1101     {
1102         static const char *write_mask_chars = "xyzw";
1103
1104         TRACE(".");
1105         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1106         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1107         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1108         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1109     }
1110 }
1111
1112 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1113         const struct wined3d_shader_version *shader_version)
1114 {
1115     DWORD src_modifier = param->modifiers;
1116     DWORD swizzle = param->swizzle;
1117
1118     if (src_modifier == WINED3DSPSM_NEG
1119             || src_modifier == WINED3DSPSM_BIASNEG
1120             || src_modifier == WINED3DSPSM_SIGNNEG
1121             || src_modifier == WINED3DSPSM_X2NEG
1122             || src_modifier == WINED3DSPSM_ABSNEG)
1123         TRACE("-");
1124     else if (src_modifier == WINED3DSPSM_COMP)
1125         TRACE("1-");
1126     else if (src_modifier == WINED3DSPSM_NOT)
1127         TRACE("!");
1128
1129     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1130         TRACE("abs(");
1131
1132     shader_dump_register(&param->reg, shader_version);
1133
1134     if (src_modifier)
1135     {
1136         switch (src_modifier)
1137         {
1138             case WINED3DSPSM_NONE:    break;
1139             case WINED3DSPSM_NEG:     break;
1140             case WINED3DSPSM_NOT:     break;
1141             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1142             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1143             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1144             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1145             case WINED3DSPSM_COMP:    break;
1146             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1147             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1148             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1149             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1150             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1151             case WINED3DSPSM_ABS:     TRACE(")"); break;
1152             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1153         }
1154     }
1155
1156     if (swizzle != WINED3DSP_NOSWIZZLE)
1157     {
1158         static const char *swizzle_chars = "xyzw";
1159         DWORD swizzle_x = swizzle & 0x03;
1160         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1161         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1162         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1163
1164         if (swizzle_x == swizzle_y
1165                 && swizzle_x == swizzle_z
1166                 && swizzle_x == swizzle_w)
1167         {
1168             TRACE(".%c", swizzle_chars[swizzle_x]);
1169         }
1170         else
1171         {
1172             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1173                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1174         }
1175     }
1176 }
1177
1178 /* Shared code in order to generate the bulk of the shader string.
1179  * NOTE: A description of how to parse tokens can be found on MSDN. */
1180 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1181         const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1182 {
1183     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1184     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1185     const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1186     void *fe_data = shader->baseShader.frontend_data;
1187     struct wined3d_shader_src_param dst_rel_addr[2];
1188     struct wined3d_shader_src_param src_rel_addr[4];
1189     struct wined3d_shader_dst_param dst_param[2];
1190     struct wined3d_shader_src_param src_param[4];
1191     struct wined3d_shader_version shader_version;
1192     struct wined3d_shader_instruction ins;
1193     struct wined3d_shader_context ctx;
1194     const DWORD *ptr = byte_code;
1195     DWORD i;
1196
1197     /* Initialize current parsing state. */
1198     ctx.shader = iface;
1199     ctx.gl_info = &device->adapter->gl_info;
1200     ctx.reg_maps = reg_maps;
1201     ctx.buffer = buffer;
1202     ctx.backend_data = backend_ctx;
1203
1204     ins.ctx = &ctx;
1205     ins.dst = dst_param;
1206     ins.src = src_param;
1207     shader->baseShader.parse_state.current_row = 0;
1208
1209     fe->shader_read_header(fe_data, &ptr, &shader_version);
1210
1211     while (!fe->shader_is_end(fe_data, &ptr))
1212     {
1213         const char *comment;
1214         UINT comment_size;
1215         UINT param_size;
1216
1217         /* Skip comment tokens. */
1218         fe->shader_read_comment(&ptr, &comment, &comment_size);
1219         if (comment) continue;
1220
1221         /* Read opcode. */
1222         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1223
1224         /* Unknown opcode and its parameters. */
1225         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1226         {
1227             TRACE("Skipping unrecognized instruction.\n");
1228             ptr += param_size;
1229             continue;
1230         }
1231
1232         /* Nothing to do. */
1233         if (ins.handler_idx == WINED3DSIH_DCL
1234                 || ins.handler_idx == WINED3DSIH_NOP
1235                 || ins.handler_idx == WINED3DSIH_DEF
1236                 || ins.handler_idx == WINED3DSIH_DEFI
1237                 || ins.handler_idx == WINED3DSIH_DEFB
1238                 || ins.handler_idx == WINED3DSIH_PHASE)
1239         {
1240             ptr += param_size;
1241             continue;
1242         }
1243
1244         /* Destination tokens */
1245         for (i = 0; i < ins.dst_count; ++i)
1246         {
1247             fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1248         }
1249
1250         /* Predication token */
1251         if (ins.predicate)
1252         {
1253             FIXME("Predicates not implemented.\n");
1254             ins.predicate = *ptr++;
1255         }
1256
1257         /* Other source tokens */
1258         for (i = 0; i < ins.src_count; ++i)
1259         {
1260             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1261         }
1262
1263         /* Call appropriate function for output target */
1264         device->shader_backend->shader_handle_instruction(&ins);
1265     }
1266 }
1267
1268 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1269 {
1270     DWORD mmask = dst->modifiers;
1271
1272     switch (dst->shift)
1273     {
1274         case 0: break;
1275         case 13: TRACE("_d8"); break;
1276         case 14: TRACE("_d4"); break;
1277         case 15: TRACE("_d2"); break;
1278         case 1: TRACE("_x2"); break;
1279         case 2: TRACE("_x4"); break;
1280         case 3: TRACE("_x8"); break;
1281         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1282     }
1283
1284     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1285     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1286     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1287
1288     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1289     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1290 }
1291
1292 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1293 {
1294     struct wined3d_shader_version shader_version;
1295     const DWORD *ptr = byte_code;
1296     const char *type_prefix;
1297     DWORD i;
1298
1299     TRACE("Parsing %p.\n", byte_code);
1300
1301     fe->shader_read_header(fe_data, &ptr, &shader_version);
1302
1303     switch (shader_version.type)
1304     {
1305         case WINED3D_SHADER_TYPE_VERTEX:
1306             type_prefix = "vs";
1307             break;
1308
1309         case WINED3D_SHADER_TYPE_GEOMETRY:
1310             type_prefix = "gs";
1311             break;
1312
1313         case WINED3D_SHADER_TYPE_PIXEL:
1314             type_prefix = "ps";
1315             break;
1316
1317         default:
1318             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1319             type_prefix = "unknown";
1320             break;
1321     }
1322
1323     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1324
1325     while (!fe->shader_is_end(fe_data, &ptr))
1326     {
1327         struct wined3d_shader_instruction ins;
1328         const char *comment;
1329         UINT comment_size;
1330         UINT param_size;
1331
1332         /* comment */
1333         fe->shader_read_comment(&ptr, &comment, &comment_size);
1334         if (comment)
1335         {
1336             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1337             {
1338                 const char *end = comment + comment_size;
1339                 const char *ptr = comment + 4;
1340                 const char *line = ptr;
1341
1342                 TRACE("// TEXT\n");
1343                 while (ptr != end)
1344                 {
1345                     if (*ptr == '\n')
1346                     {
1347                         UINT len = ptr - line;
1348                         if (len && *(ptr - 1) == '\r') --len;
1349                         TRACE("// %s\n", debugstr_an(line, len));
1350                         line = ++ptr;
1351                     }
1352                     else ++ptr;
1353                 }
1354                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1355             }
1356             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1357             continue;
1358         }
1359
1360         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1361         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1362         {
1363             TRACE("Skipping unrecognized instruction.\n");
1364             ptr += param_size;
1365             continue;
1366         }
1367
1368         if (ins.handler_idx == WINED3DSIH_DCL)
1369         {
1370             struct wined3d_shader_semantic semantic;
1371
1372             fe->shader_read_semantic(&ptr, &semantic);
1373
1374             shader_dump_decl_usage(&semantic, &shader_version);
1375             shader_dump_ins_modifiers(&semantic.reg);
1376             TRACE(" ");
1377             shader_dump_dst_param(&semantic.reg, &shader_version);
1378         }
1379         else if (ins.handler_idx == WINED3DSIH_DEF)
1380         {
1381             struct wined3d_shader_src_param rel_addr;
1382             struct wined3d_shader_dst_param dst;
1383
1384             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1385
1386             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1387                     *(const float *)(ptr),
1388                     *(const float *)(ptr + 1),
1389                     *(const float *)(ptr + 2),
1390                     *(const float *)(ptr + 3));
1391             ptr += 4;
1392         }
1393         else if (ins.handler_idx == WINED3DSIH_DEFI)
1394         {
1395             struct wined3d_shader_src_param rel_addr;
1396             struct wined3d_shader_dst_param dst;
1397
1398             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1399
1400             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1401                     *(ptr),
1402                     *(ptr + 1),
1403                     *(ptr + 2),
1404                     *(ptr + 3));
1405             ptr += 4;
1406         }
1407         else if (ins.handler_idx == WINED3DSIH_DEFB)
1408         {
1409             struct wined3d_shader_src_param rel_addr;
1410             struct wined3d_shader_dst_param dst;
1411
1412             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1413
1414             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1415             ++ptr;
1416         }
1417         else
1418         {
1419             struct wined3d_shader_src_param dst_rel_addr[2];
1420             struct wined3d_shader_src_param src_rel_addr;
1421             struct wined3d_shader_dst_param dst_param[2];
1422             struct wined3d_shader_src_param src_param;
1423
1424             for (i = 0; i < ins.dst_count; ++i)
1425             {
1426                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1427             }
1428
1429             /* Print out predication source token first - it follows
1430              * the destination token. */
1431             if (ins.predicate)
1432             {
1433                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1434                 TRACE("(");
1435                 shader_dump_src_param(&src_param, &shader_version);
1436                 TRACE(") ");
1437             }
1438
1439             /* PixWin marks instructions with the coissue flag with a '+' */
1440             if (ins.coissue) TRACE("+");
1441
1442             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1443
1444             if (ins.handler_idx == WINED3DSIH_IFC
1445                     || ins.handler_idx == WINED3DSIH_BREAKC)
1446             {
1447                 switch (ins.flags)
1448                 {
1449                     case COMPARISON_GT: TRACE("_gt"); break;
1450                     case COMPARISON_EQ: TRACE("_eq"); break;
1451                     case COMPARISON_GE: TRACE("_ge"); break;
1452                     case COMPARISON_LT: TRACE("_lt"); break;
1453                     case COMPARISON_NE: TRACE("_ne"); break;
1454                     case COMPARISON_LE: TRACE("_le"); break;
1455                     default: TRACE("_(%u)", ins.flags);
1456                 }
1457             }
1458             else if (ins.handler_idx == WINED3DSIH_TEX
1459                     && shader_version.major >= 2
1460                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1461             {
1462                 TRACE("p");
1463             }
1464
1465             /* We already read the destination tokens, print them. */
1466             for (i = 0; i < ins.dst_count; ++i)
1467             {
1468                 shader_dump_ins_modifiers(&dst_param[i]);
1469                 TRACE(!i ? " " : ", ");
1470                 shader_dump_dst_param(&dst_param[i], &shader_version);
1471             }
1472
1473             /* Other source tokens */
1474             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1475             {
1476                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1477                 TRACE(!i ? " " : ", ");
1478                 shader_dump_src_param(&src_param, &shader_version);
1479             }
1480         }
1481         TRACE("\n");
1482     }
1483 }
1484
1485 static void shader_cleanup(IWineD3DBaseShader *iface)
1486 {
1487     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1488
1489     ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1490     HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1491     HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1492     shader_delete_constant_list(&shader->baseShader.constantsF);
1493     shader_delete_constant_list(&shader->baseShader.constantsB);
1494     shader_delete_constant_list(&shader->baseShader.constantsI);
1495     list_remove(&shader->baseShader.shader_list_entry);
1496
1497     if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1498     {
1499         shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1500     }
1501 }
1502
1503 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1504 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1505 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type, const SIZE *ds_mask_size) {}
1506 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1507 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1508 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1509 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1510 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1511 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1512 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1513 static void shader_none_free(IWineD3DDevice *iface) {}
1514 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1515
1516 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1517 {
1518     /* Set the shader caps to 0 for the none shader backend */
1519     caps->VertexShaderVersion = 0;
1520     caps->MaxVertexShaderConst = 0;
1521     caps->PixelShaderVersion = 0;
1522     caps->PixelShader1xMaxValue = 0.0f;
1523     caps->MaxPixelShaderConst = 0;
1524     caps->VSClipping = FALSE;
1525 }
1526
1527 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1528 {
1529     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1530     {
1531         TRACE("Checking support for fixup:\n");
1532         dump_color_fixup_desc(fixup);
1533     }
1534
1535     /* Faked to make some apps happy. */
1536     if (!is_complex_fixup(fixup))
1537     {
1538         TRACE("[OK]\n");
1539         return TRUE;
1540     }
1541
1542     TRACE("[FAILED]\n");
1543     return FALSE;
1544 }
1545
1546 const shader_backend_t none_shader_backend = {
1547     shader_none_handle_instruction,
1548     shader_none_select,
1549     shader_none_select_depth_blt,
1550     shader_none_deselect_depth_blt,
1551     shader_none_update_float_vertex_constants,
1552     shader_none_update_float_pixel_constants,
1553     shader_none_load_constants,
1554     shader_none_load_np2fixup_constants,
1555     shader_none_destroy,
1556     shader_none_alloc,
1557     shader_none_free,
1558     shader_none_dirty_const,
1559     shader_none_get_caps,
1560     shader_none_color_fixup_supported,
1561 };
1562
1563 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1564 {
1565     if (!data)
1566     {
1567         *data_size = shader->baseShader.functionLength;
1568         return WINED3D_OK;
1569     }
1570
1571     if (*data_size < shader->baseShader.functionLength)
1572     {
1573         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1574          * than the required size we should write the required size and
1575          * return D3DERR_MOREDATA. That's not actually true. */
1576         return WINED3DERR_INVALIDCALL;
1577     }
1578
1579     memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1580
1581     return WINED3D_OK;
1582 }
1583
1584 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1585         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1586 {
1587     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1588     const struct wined3d_shader_frontend *fe;
1589     HRESULT hr;
1590
1591     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1592             shader, byte_code, output_signature, float_const_count);
1593
1594     fe = shader_select_frontend(*byte_code);
1595     if (!fe)
1596     {
1597         FIXME("Unable to find frontend for shader.\n");
1598         return WINED3DERR_INVALIDCALL;
1599     }
1600     shader->baseShader.frontend = fe;
1601     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1602     if (!shader->baseShader.frontend_data)
1603     {
1604         FIXME("Failed to initialize frontend.\n");
1605         return WINED3DERR_INVALIDCALL;
1606     }
1607
1608     /* First pass: trace shader. */
1609     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1610
1611     /* Initialize immediate constant lists. */
1612     list_init(&shader->baseShader.constantsF);
1613     list_init(&shader->baseShader.constantsB);
1614     list_init(&shader->baseShader.constantsI);
1615
1616     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1617     hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1618             reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1619             byte_code, float_const_count);
1620     if (FAILED(hr)) return hr;
1621
1622     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1623     if (!shader->baseShader.function) return E_OUTOFMEMORY;
1624     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1625
1626     return WINED3D_OK;
1627 }
1628
1629 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1630 {
1631     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1632
1633     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1634             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1635             || IsEqualGUID(riid, &IID_IWineD3DBase)
1636             || IsEqualGUID(riid, &IID_IUnknown))
1637     {
1638         IUnknown_AddRef(iface);
1639         *object = iface;
1640         return S_OK;
1641     }
1642
1643     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1644
1645     *object = NULL;
1646     return E_NOINTERFACE;
1647 }
1648
1649 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1650 {
1651     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1652     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1653
1654     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1655
1656     return refcount;
1657 }
1658
1659 /* Do not call while under the GL lock. */
1660 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1661 {
1662     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1663     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1664
1665     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1666
1667     if (!refcount)
1668     {
1669         shader_cleanup((IWineD3DBaseShader *)iface);
1670         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1671         HeapFree(GetProcessHeap(), 0, shader);
1672     }
1673
1674     return refcount;
1675 }
1676
1677 static void * STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface)
1678 {
1679     TRACE("iface %p.\n", iface);
1680
1681     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1682 }
1683
1684 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1685 {
1686     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1687
1688     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1689 }
1690
1691 /* Set local constants for d3d8 shaders. */
1692 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1693         UINT start_idx, const float *src_data, UINT count)
1694 {
1695     IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1696     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1697     UINT i, end_idx;
1698
1699     TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1700
1701     end_idx = start_idx + count;
1702     if (end_idx > device->d3d_vshader_constantF)
1703     {
1704         WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1705         end_idx = device->d3d_vshader_constantF;
1706     }
1707
1708     for (i = start_idx; i < end_idx; ++i)
1709     {
1710         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1711         if (!lconst) return E_OUTOFMEMORY;
1712
1713         lconst->idx = i;
1714         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1715         list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1716     }
1717
1718     return WINED3D_OK;
1719 }
1720
1721 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1722 {
1723     /* IUnknown methods */
1724     vertexshader_QueryInterface,
1725     vertexshader_AddRef,
1726     vertexshader_Release,
1727     /* IWineD3DBase methods */
1728     vertexshader_GetParent,
1729     /* IWineD3DBaseShader methods */
1730     vertexshader_GetFunction,
1731     /* IWineD3DVertexShader methods */
1732     vertexshader_SetLocalConstantsF,
1733 };
1734
1735 void find_vs_compile_args(const struct wined3d_state *state,
1736         IWineD3DVertexShaderImpl *shader, struct vs_compile_args *args)
1737 {
1738     args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1739             == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1740     args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1741             && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1742     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1743 }
1744
1745 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1746 {
1747     if (usage_idx1 != usage_idx2) return FALSE;
1748     if (usage1 == usage2) return TRUE;
1749     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1750     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1751
1752     return FALSE;
1753 }
1754
1755 BOOL vshader_get_input(struct IWineD3DVertexShaderImpl *shader,
1756         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1757 {
1758     WORD map = shader->baseShader.reg_maps.input_registers;
1759     unsigned int i;
1760
1761     for (i = 0; map; map >>= 1, ++i)
1762     {
1763         if (!(map & 1)) continue;
1764
1765         if (match_usage(shader->attributes[i].usage,
1766                 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1767         {
1768             *regnum = i;
1769             return TRUE;
1770         }
1771     }
1772     return FALSE;
1773 }
1774
1775 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1776 {
1777     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1778             shader->baseShader.reg_maps.shader_version.minor);
1779     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1780
1781     shader->baseShader.limits.texcoord = 0;
1782     shader->baseShader.limits.attributes = 16;
1783     shader->baseShader.limits.packed_input = 0;
1784
1785     switch (shader_version)
1786     {
1787         case WINED3D_SHADER_VERSION(1, 0):
1788         case WINED3D_SHADER_VERSION(1, 1):
1789             shader->baseShader.limits.temporary = 12;
1790             shader->baseShader.limits.constant_bool = 0;
1791             shader->baseShader.limits.constant_int = 0;
1792             shader->baseShader.limits.address = 1;
1793             shader->baseShader.limits.packed_output = 12;
1794             shader->baseShader.limits.sampler = 0;
1795             shader->baseShader.limits.label = 0;
1796             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1797              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1798              * constants? */
1799             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1800             break;
1801
1802         case WINED3D_SHADER_VERSION(2, 0):
1803         case WINED3D_SHADER_VERSION(2, 1):
1804             shader->baseShader.limits.temporary = 12;
1805             shader->baseShader.limits.constant_bool = 16;
1806             shader->baseShader.limits.constant_int = 16;
1807             shader->baseShader.limits.address = 1;
1808             shader->baseShader.limits.packed_output = 12;
1809             shader->baseShader.limits.sampler = 0;
1810             shader->baseShader.limits.label = 16;
1811             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1812             break;
1813
1814         case WINED3D_SHADER_VERSION(4, 0):
1815             FIXME("Using 3.0 limits for 4.0 shader.\n");
1816             /* Fall through. */
1817
1818         case WINED3D_SHADER_VERSION(3, 0):
1819             shader->baseShader.limits.temporary = 32;
1820             shader->baseShader.limits.constant_bool = 32;
1821             shader->baseShader.limits.constant_int = 32;
1822             shader->baseShader.limits.address = 1;
1823             shader->baseShader.limits.packed_output = 12;
1824             shader->baseShader.limits.sampler = 4;
1825             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1826             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1827              * even though they are capable of supporting much more (GL
1828              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1829              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1830              * shaders to 256. */
1831             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1832             break;
1833
1834         default:
1835             shader->baseShader.limits.temporary = 12;
1836             shader->baseShader.limits.constant_bool = 16;
1837             shader->baseShader.limits.constant_int = 16;
1838             shader->baseShader.limits.address = 1;
1839             shader->baseShader.limits.packed_output = 12;
1840             shader->baseShader.limits.sampler = 0;
1841             shader->baseShader.limits.label = 16;
1842             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1843             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1844                     shader->baseShader.reg_maps.shader_version.major,
1845                     shader->baseShader.reg_maps.shader_version.minor);
1846     }
1847 }
1848
1849 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1850         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1851         void *parent, const struct wined3d_parent_ops *parent_ops)
1852 {
1853     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1854     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1855     unsigned int i;
1856     HRESULT hr;
1857     WORD map;
1858
1859     if (!byte_code) return WINED3DERR_INVALIDCALL;
1860
1861     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1862     shader_init(&shader->baseShader, device, parent, parent_ops);
1863
1864     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1865             output_signature, device->d3d_vshader_constantF);
1866     if (FAILED(hr))
1867     {
1868         WARN("Failed to set function, hr %#x.\n", hr);
1869         shader_cleanup((IWineD3DBaseShader *)shader);
1870         return hr;
1871     }
1872
1873     map = shader->baseShader.reg_maps.input_registers;
1874     for (i = 0; map; map >>= 1, ++i)
1875     {
1876         if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1877
1878         shader->attributes[i].usage =
1879                 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1880         shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1881     }
1882
1883     if (output_signature)
1884     {
1885         for (i = 0; i < output_signature->element_count; ++i)
1886         {
1887             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1888             reg_maps->output_registers |= 1 << e->register_idx;
1889             shader->baseShader.output_signature[e->register_idx] = *e;
1890         }
1891     }
1892
1893     vertexshader_set_limits(shader);
1894
1895     if (device->vs_selected_mode == SHADER_ARB
1896             && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1897             && shader->min_rel_offset <= shader->max_rel_offset)
1898     {
1899         if (shader->max_rel_offset - shader->min_rel_offset > 127)
1900         {
1901             FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1902             FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1903             FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1904         }
1905         else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1906         {
1907             shader->rel_offset = shader->min_rel_offset + 63;
1908         }
1909         else if (shader->max_rel_offset > 63)
1910         {
1911             shader->rel_offset = shader->min_rel_offset;
1912         }
1913         else
1914         {
1915             shader->rel_offset = 0;
1916         }
1917     }
1918
1919     shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1920             && !list_empty(&shader->baseShader.constantsF);
1921
1922     return WINED3D_OK;
1923 }
1924
1925 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1926         REFIID riid, void **object)
1927 {
1928     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1929
1930     if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1931             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1932             || IsEqualGUID(riid, &IID_IWineD3DBase)
1933             || IsEqualGUID(riid, &IID_IUnknown))
1934     {
1935         IUnknown_AddRef(iface);
1936         *object = iface;
1937         return S_OK;
1938     }
1939
1940     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1941
1942     *object = NULL;
1943     return E_NOINTERFACE;
1944 }
1945
1946 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1947 {
1948     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1949     ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1950
1951     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1952
1953     return refcount;
1954 }
1955
1956 /* Do not call while under the GL lock. */
1957 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1958 {
1959     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1960     ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1961
1962     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1963
1964     if (!refcount)
1965     {
1966         shader_cleanup((IWineD3DBaseShader *)iface);
1967         shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1968         HeapFree(GetProcessHeap(), 0, shader);
1969     }
1970
1971     return refcount;
1972 }
1973
1974 static void * STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface)
1975 {
1976     TRACE("iface %p.\n", iface);
1977
1978     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1979 }
1980
1981 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1982 {
1983     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1984
1985     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1986 }
1987
1988 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1989 {
1990     /* IUnknown methods */
1991     geometryshader_QueryInterface,
1992     geometryshader_AddRef,
1993     geometryshader_Release,
1994     /* IWineD3DBase methods */
1995     geometryshader_GetParent,
1996     /* IWineD3DBaseShader methods */
1997     geometryshader_GetFunction,
1998 };
1999
2000 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
2001         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2002         void *parent, const struct wined3d_parent_ops *parent_ops)
2003 {
2004     HRESULT hr;
2005
2006     shader->vtbl = &wined3d_geometryshader_vtbl;
2007     shader_init(&shader->base_shader, device, parent, parent_ops);
2008
2009     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
2010     if (FAILED(hr))
2011     {
2012         WARN("Failed to set function, hr %#x.\n", hr);
2013         shader_cleanup((IWineD3DBaseShader *)shader);
2014         return hr;
2015     }
2016
2017     shader->base_shader.load_local_constsF = FALSE;
2018
2019     return WINED3D_OK;
2020 }
2021
2022 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
2023 {
2024     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
2025
2026     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
2027             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
2028             || IsEqualGUID(riid, &IID_IWineD3DBase)
2029             || IsEqualGUID(riid, &IID_IUnknown))
2030     {
2031         IUnknown_AddRef(iface);
2032         *object = iface;
2033         return S_OK;
2034     }
2035
2036     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
2037
2038     *object = NULL;
2039     return E_NOINTERFACE;
2040 }
2041
2042 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
2043 {
2044     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2045     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
2046
2047     TRACE("%p increasing refcount to %u.\n", shader, refcount);
2048
2049     return refcount;
2050 }
2051
2052 /* Do not call while under the GL lock. */
2053 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
2054 {
2055     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2056     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
2057
2058     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2059
2060     if (!refcount)
2061     {
2062         shader_cleanup((IWineD3DBaseShader *)iface);
2063         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
2064         HeapFree(GetProcessHeap(), 0, shader);
2065     }
2066
2067     return refcount;
2068 }
2069
2070 static void * STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface)
2071 {
2072     TRACE("iface %p.\n", iface);
2073
2074     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
2075 }
2076
2077 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
2078 {
2079     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
2080
2081     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
2082 }
2083
2084 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2085 {
2086     /* IUnknown methods */
2087     pixelshader_QueryInterface,
2088     pixelshader_AddRef,
2089     pixelshader_Release,
2090     /* IWineD3DBase methods */
2091     pixelshader_GetParent,
2092     /* IWineD3DBaseShader methods */
2093     pixelshader_GetFunction
2094 };
2095
2096 void find_ps_compile_args(const struct wined3d_state *state,
2097         IWineD3DPixelShaderImpl *shader, struct ps_compile_args *args)
2098 {
2099     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
2100     IWineD3DBaseTextureImpl *texture;
2101     UINT i;
2102
2103     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2104     if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
2105     {
2106         IWineD3DSurfaceImpl *rt = device->render_targets[0];
2107         if(rt->resource.format->Flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
2108     }
2109
2110     if (shader->baseShader.reg_maps.shader_version.major == 1
2111             && shader->baseShader.reg_maps.shader_version.minor <= 3)
2112     {
2113         for (i = 0; i < 4; ++i)
2114         {
2115             DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
2116             DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
2117             if (flags & WINED3DTTFF_PROJECTED)
2118                 tex_transform |= WINED3D_PSARGS_PROJECTED;
2119             args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2120         }
2121     }
2122
2123     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2124     {
2125         if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2126         texture = state->textures[i];
2127         if (!texture)
2128         {
2129             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2130             continue;
2131         }
2132         args->color_fixup[i] = texture->resource.format->color_fixup;
2133
2134         if (texture->resource.format->Flags & WINED3DFMT_FLAG_SHADOW)
2135             args->shadow |= 1 << i;
2136
2137         /* Flag samplers that need NP2 texcoord fixup. */
2138         if (!texture->baseTexture.pow2Matrix_identity)
2139         {
2140             args->np2_fixup |= (1 << i);
2141         }
2142     }
2143     if (shader->baseShader.reg_maps.shader_version.major >= 3)
2144     {
2145         if (device->strided_streams.position_transformed)
2146         {
2147             args->vp_mode = pretransformed;
2148         }
2149         else if (use_vs(state))
2150         {
2151             args->vp_mode = vertexshader;
2152         }
2153         else
2154         {
2155             args->vp_mode = fixedfunction;
2156         }
2157         args->fog = FOG_OFF;
2158     }
2159     else
2160     {
2161         args->vp_mode = vertexshader;
2162         if (state->render_states[WINED3DRS_FOGENABLE])
2163         {
2164             switch (state->render_states[WINED3DRS_FOGTABLEMODE])
2165             {
2166                 case WINED3DFOG_NONE:
2167                     if (device->strided_streams.position_transformed || use_vs(state))
2168                     {
2169                         args->fog = FOG_LINEAR;
2170                         break;
2171                     }
2172
2173                     switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
2174                     {
2175                         case WINED3DFOG_NONE: /* Fall through. */
2176                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2177                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2178                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2179                     }
2180                     break;
2181
2182                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2183                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2184                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2185             }
2186         }
2187         else
2188         {
2189             args->fog = FOG_OFF;
2190         }
2191     }
2192 }
2193
2194 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2195 {
2196     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2197             shader->baseShader.reg_maps.shader_version.minor);
2198
2199     shader->baseShader.limits.attributes = 0;
2200     shader->baseShader.limits.address = 0;
2201     shader->baseShader.limits.packed_output = 0;
2202
2203     switch (shader_version)
2204     {
2205         case WINED3D_SHADER_VERSION(1, 0):
2206         case WINED3D_SHADER_VERSION(1, 1):
2207         case WINED3D_SHADER_VERSION(1, 2):
2208         case WINED3D_SHADER_VERSION(1, 3):
2209             shader->baseShader.limits.temporary = 2;
2210             shader->baseShader.limits.constant_float = 8;
2211             shader->baseShader.limits.constant_int = 0;
2212             shader->baseShader.limits.constant_bool = 0;
2213             shader->baseShader.limits.texcoord = 4;
2214             shader->baseShader.limits.sampler = 4;
2215             shader->baseShader.limits.packed_input = 0;
2216             shader->baseShader.limits.label = 0;
2217             break;
2218
2219         case WINED3D_SHADER_VERSION(1, 4):
2220             shader->baseShader.limits.temporary = 6;
2221             shader->baseShader.limits.constant_float = 8;
2222             shader->baseShader.limits.constant_int = 0;
2223             shader->baseShader.limits.constant_bool = 0;
2224             shader->baseShader.limits.texcoord = 6;
2225             shader->baseShader.limits.sampler = 6;
2226             shader->baseShader.limits.packed_input = 0;
2227             shader->baseShader.limits.label = 0;
2228             break;
2229
2230         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2231         case WINED3D_SHADER_VERSION(2, 0):
2232             shader->baseShader.limits.temporary = 32;
2233             shader->baseShader.limits.constant_float = 32;
2234             shader->baseShader.limits.constant_int = 16;
2235             shader->baseShader.limits.constant_bool = 16;
2236             shader->baseShader.limits.texcoord = 8;
2237             shader->baseShader.limits.sampler = 16;
2238             shader->baseShader.limits.packed_input = 0;
2239             break;
2240
2241         case WINED3D_SHADER_VERSION(2, 1):
2242             shader->baseShader.limits.temporary = 32;
2243             shader->baseShader.limits.constant_float = 32;
2244             shader->baseShader.limits.constant_int = 16;
2245             shader->baseShader.limits.constant_bool = 16;
2246             shader->baseShader.limits.texcoord = 8;
2247             shader->baseShader.limits.sampler = 16;
2248             shader->baseShader.limits.packed_input = 0;
2249             shader->baseShader.limits.label = 16;
2250             break;
2251
2252         case WINED3D_SHADER_VERSION(4, 0):
2253             FIXME("Using 3.0 limits for 4.0 shader.\n");
2254             /* Fall through. */
2255
2256         case WINED3D_SHADER_VERSION(3, 0):
2257             shader->baseShader.limits.temporary = 32;
2258             shader->baseShader.limits.constant_float = 224;
2259             shader->baseShader.limits.constant_int = 16;
2260             shader->baseShader.limits.constant_bool = 16;
2261             shader->baseShader.limits.texcoord = 0;
2262             shader->baseShader.limits.sampler = 16;
2263             shader->baseShader.limits.packed_input = 12;
2264             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2265             break;
2266
2267         default:
2268             shader->baseShader.limits.temporary = 32;
2269             shader->baseShader.limits.constant_float = 32;
2270             shader->baseShader.limits.constant_int = 16;
2271             shader->baseShader.limits.constant_bool = 16;
2272             shader->baseShader.limits.texcoord = 8;
2273             shader->baseShader.limits.sampler = 16;
2274             shader->baseShader.limits.packed_input = 0;
2275             shader->baseShader.limits.label = 0;
2276             FIXME("Unrecognized pixel shader version %u.%u\n",
2277                     shader->baseShader.reg_maps.shader_version.major,
2278                     shader->baseShader.reg_maps.shader_version.minor);
2279     }
2280 }
2281
2282 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2283         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2284         void *parent, const struct wined3d_parent_ops *parent_ops)
2285 {
2286     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2287     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2288     HRESULT hr;
2289
2290     if (!byte_code) return WINED3DERR_INVALIDCALL;
2291
2292     shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2293     shader_init(&shader->baseShader, device, parent, parent_ops);
2294
2295     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2296             output_signature, device->d3d_pshader_constantF);
2297     if (FAILED(hr))
2298     {
2299         WARN("Failed to set function, hr %#x.\n", hr);
2300         shader_cleanup((IWineD3DBaseShader *)shader);
2301         return hr;
2302     }
2303
2304     pixelshader_set_limits(shader);
2305
2306     for (i = 0; i < MAX_REG_INPUT; ++i)
2307     {
2308         if (shader->input_reg_used[i])
2309         {
2310             ++num_regs_used;
2311             highest_reg_used = i;
2312         }
2313     }
2314
2315     /* Don't do any register mapping magic if it is not needed, or if we can't
2316      * achieve anything anyway */
2317     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2318             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2319     {
2320         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2321         {
2322             /* This happens with relative addressing. The input mapper function
2323              * warns about this if the higher registers are declared too, so
2324              * don't write a FIXME here */
2325             WARN("More varying registers used than supported\n");
2326         }
2327
2328         for (i = 0; i < MAX_REG_INPUT; ++i)
2329         {
2330             shader->input_reg_map[i] = i;
2331         }
2332
2333         shader->declared_in_count = highest_reg_used + 1;
2334     }
2335     else
2336     {
2337         shader->declared_in_count = 0;
2338         for (i = 0; i < MAX_REG_INPUT; ++i)
2339         {
2340             if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2341             else shader->input_reg_map[i] = ~0U;
2342         }
2343     }
2344
2345     shader->baseShader.load_local_constsF = FALSE;
2346
2347     return WINED3D_OK;
2348 }
2349
2350 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2351 {
2352     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2353     unsigned int i;
2354
2355     if (reg_maps->shader_version.major != 1) return;
2356
2357     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2358     {
2359         /* We don't sample from this sampler. */
2360         if (!sampler_type[i]) continue;
2361
2362         if (!textures[i])
2363         {
2364             WARN("No texture bound to sampler %u, using 2D.\n", i);
2365             sampler_type[i] = WINED3DSTT_2D;
2366             continue;
2367         }
2368
2369         switch (((IWineD3DBaseTextureImpl *)textures[i])->baseTexture.target)
2370         {
2371             case GL_TEXTURE_RECTANGLE_ARB:
2372             case GL_TEXTURE_2D:
2373                 /* We have to select between texture rectangles and 2D
2374                  * textures later because 2.0 and 3.0 shaders only have
2375                  * WINED3DSTT_2D as well. */
2376                 sampler_type[i] = WINED3DSTT_2D;
2377                 break;
2378
2379             case GL_TEXTURE_3D:
2380                 sampler_type[i] = WINED3DSTT_VOLUME;
2381                 break;
2382
2383             case GL_TEXTURE_CUBE_MAP_ARB:
2384                 sampler_type[i] = WINED3DSTT_CUBE;
2385                 break;
2386
2387             default:
2388                 FIXME("Unrecognized texture type %#x, using 2D.\n",
2389                         ((IWineD3DBaseTextureImpl *)textures[i])->baseTexture.target);
2390                 sampler_type[i] = WINED3DSTT_2D;
2391         }
2392     }
2393 }