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_SQRT          */ "sqrt",
 115     /* WINED3DSIH_SUB           */ "sub",
 116     /* WINED3DSIH_TEX           */ "texld",
 117     /* WINED3DSIH_TEXBEM        */ "texbem",
 118     /* WINED3DSIH_TEXBEML       */ "texbeml",
 119     /* WINED3DSIH_TEXCOORD      */ "texcrd",
 120     /* WINED3DSIH_TEXDEPTH      */ "texdepth",
 121     /* WINED3DSIH_TEXDP3        */ "texdp3",
 122     /* WINED3DSIH_TEXDP3TEX     */ "texdp3tex",
 123     /* WINED3DSIH_TEXKILL       */ "texkill",
 124     /* WINED3DSIH_TEXLDD        */ "texldd",
 125     /* WINED3DSIH_TEXLDL        */ "texldl",
 126     /* WINED3DSIH_TEXM3x2DEPTH  */ "texm3x2depth",
 127     /* WINED3DSIH_TEXM3x2PAD    */ "texm3x2pad",
 128     /* WINED3DSIH_TEXM3x2TEX    */ "texm3x2tex",
 129     /* WINED3DSIH_TEXM3x3       */ "texm3x3",
 130     /* WINED3DSIH_TEXM3x3DIFF   */ "texm3x3diff",
 131     /* WINED3DSIH_TEXM3x3PAD    */ "texm3x3pad",
 132     /* WINED3DSIH_TEXM3x3SPEC   */ "texm3x3spec",
 133     /* WINED3DSIH_TEXM3x3TEX    */ "texm3x3tex",
 134     /* WINED3DSIH_TEXM3x3VSPEC  */ "texm3x3vspec",
 135     /* WINED3DSIH_TEXREG2AR     */ "texreg2ar",
 136     /* WINED3DSIH_TEXREG2GB     */ "texreg2gb",
 137     /* WINED3DSIH_TEXREG2RGB    */ "texreg2rgb",
 138     /* WINED3DSIH_UTOF          */ "utof",
 139 };
 140
 141 static const char * const semantic_names[] =
 142 {
 143     /* WINED3DDECLUSAGE_POSITION        */ "SV_POSITION",
 144     /* WINED3DDECLUSAGE_BLENDWEIGHT     */ "BLENDWEIGHT",
 145     /* WINED3DDECLUSAGE_BLENDINDICES    */ "BLENDINDICES",
 146     /* WINED3DDECLUSAGE_NORMAL          */ "NORMAL",
 147     /* WINED3DDECLUSAGE_PSIZE           */ "PSIZE",
 148     /* WINED3DDECLUSAGE_TEXCOORD        */ "TEXCOORD",
 149     /* WINED3DDECLUSAGE_TANGENT         */ "TANGENT",
 150     /* WINED3DDECLUSAGE_BINORMAL        */ "BINORMAL",
 151     /* WINED3DDECLUSAGE_TESSFACTOR      */ "TESSFACTOR",
 152     /* WINED3DDECLUSAGE_POSITIONT       */ "POSITIONT",
 153     /* WINED3DDECLUSAGE_COLOR           */ "COLOR",
 154     /* WINED3DDECLUSAGE_FOG             */ "FOG",
 155     /* WINED3DDECLUSAGE_DEPTH           */ "DEPTH",
 156     /* WINED3DDECLUSAGE_SAMPLE          */ "SAMPLE",
 157 };
 158
 159 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
 160 {
 161     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 162     {
 163         FIXME("Unrecognized usage %#x.\n", usage);
 164         return "UNRECOGNIZED";
 165     }
 166
 167     return semantic_names[usage];
 168 }
 169
 170 static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
 171 {
 172     unsigned int i;
 173
 174     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 175     {
 176         if (!strcmp(name, semantic_names[i])) return i;
 177     }
 178
 179     return ~0U;
 180 }
 181
 182 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
 183 {
 184     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 185 }
 186
 187 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 188         const struct wined3d_shader_semantic *s)
 189 {
 190     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 191     e->semantic_idx = s->usage_idx;
 192     e->sysval_semantic = 0;
 193     e->component_type = 0;
 194     e->register_idx = s->reg.reg.idx;
 195     e->mask = s->reg.write_mask;
 196 }
 197
 198 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
 199         WINED3DDECLUSAGE usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
 200 {
 201     e->semantic_name = shader_semantic_name_from_usage(usage);
 202     e->semantic_idx = usage_idx;
 203     e->sysval_semantic = 0;
 204     e->component_type = 0;
 205     e->register_idx = reg_idx;
 206     e->mask = write_mask;
 207 }
 208
 209 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 210 {
 211     switch (version_token >> 16)
 212     {
 213         case WINED3D_SM1_VS:
 214         case WINED3D_SM1_PS:
 215             return &sm1_shader_frontend;
 216
 217         case WINED3D_SM4_PS:
 218         case WINED3D_SM4_VS:
 219         case WINED3D_SM4_GS:
 220             return &sm4_shader_frontend;
 221
 222         default:
 223             FIXME("Unrecognised version token %#x\n", version_token);
 224             return NULL;
 225     }
 226 }
 227
 228 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 229 {
 230     buffer->buffer[0] = '\0';
 231     buffer->bsize = 0;
 232     buffer->lineNo = 0;
 233     buffer->newline = TRUE;
 234 }
 235
 236 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 237 {
 238     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 239     if (!buffer->buffer)
 240     {
 241         ERR("Failed to allocate shader buffer memory.\n");
 242         return FALSE;
 243     }
 244
 245     shader_buffer_clear(buffer);
 246     return TRUE;
 247 }
 248
 249 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 250 {
 251     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 252 }
 253
 254 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 255 {
 256     char *base = buffer->buffer + buffer->bsize;
 257     int rc;
 258
 259     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 260
 261     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 262     {
 263         ERR("The buffer allocated for the shader program string "
 264             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 265         buffer->bsize = SHADER_PGMSIZE - 1;
 266         return -1;
 267     }
 268
 269     if (buffer->newline)
 270     {
 271         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 272         buffer->newline = FALSE;
 273     }
 274     else
 275     {
 276         TRACE("%s", base);
 277     }
 278
 279     buffer->bsize += rc;
 280     if (buffer->buffer[buffer->bsize-1] == '\n')
 281     {
 282         ++buffer->lineNo;
 283         buffer->newline = TRUE;
 284     }
 285
 286     return 0;
 287 }
 288
 289 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 290 {
 291     va_list args;
 292     int ret;
 293
 294     va_start(args, format);
 295     ret = shader_vaddline(buffer, format, args);
 296     va_end(args);
 297
 298     return ret;
 299 }
 300
 301 static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
 302         void *parent, const struct wined3d_parent_ops *parent_ops)
 303 {
 304     shader->ref = 1;
 305     shader->device = (IWineD3DDevice *)device;
 306     shader->parent = parent;
 307     shader->parent_ops = parent_ops;
 308     list_init(&shader->linked_programs);
 309     list_add_head(&device->shaders, &shader->shader_list_entry);
 310 }
 311
 312 /* Convert floating point offset relative to a register file to an absolute
 313  * offset for float constants. */
 314 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
 315 {
 316     switch (register_type)
 317     {
 318         case WINED3DSPR_CONST: return register_idx;
 319         case WINED3DSPR_CONST2: return 2048 + register_idx;
 320         case WINED3DSPR_CONST3: return 4096 + register_idx;
 321         case WINED3DSPR_CONST4: return 6144 + register_idx;
 322         default:
 323             FIXME("Unsupported register type: %u.\n", register_type);
 324             return register_idx;
 325     }
 326 }
 327
 328 static void shader_delete_constant_list(struct list *clist)
 329 {
 330     struct local_constant *constant;
 331     struct list *ptr;
 332
 333     ptr = list_head(clist);
 334     while (ptr)
 335     {
 336         constant = LIST_ENTRY(ptr, struct local_constant, entry);
 337         ptr = list_next(clist, ptr);
 338         HeapFree(GetProcessHeap(), 0, constant);
 339     }
 340     list_init(clist);
 341 }
 342
 343 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 344 {
 345     DWORD idx, shift;
 346     idx = bit >> 5;
 347     shift = bit & 0x1f;
 348     bitmap[idx] |= (1 << shift);
 349 }
 350
 351 static void shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct shader_reg_maps *reg_maps,
 352         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 353 {
 354     switch (reg->type)
 355     {
 356         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 357             if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
 358             else reg_maps->address |= 1 << reg->idx;
 359             break;
 360
 361         case WINED3DSPR_TEMP:
 362             reg_maps->temporary |= 1 << reg->idx;
 363             break;
 364
 365         case WINED3DSPR_INPUT:
 366             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 367             {
 368                 if (reg->rel_addr)
 369                 {
 370                     /* If relative addressing is used, we must assume that all registers
 371                      * are used. Even if it is a construct like v3[aL], we can't assume
 372                      * that v0, v1 and v2 aren't read because aL can be negative */
 373                     unsigned int i;
 374                     for (i = 0; i < MAX_REG_INPUT; ++i)
 375                     {
 376                         ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
 377                     }
 378                 }
 379                 else
 380                 {
 381                     ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg->idx] = TRUE;
 382                 }
 383             }
 384             else reg_maps->input_registers |= 1 << reg->idx;
 385             break;
 386
 387         case WINED3DSPR_RASTOUT:
 388             if (reg->idx == 1) reg_maps->fog = 1;
 389             break;
 390
 391         case WINED3DSPR_MISCTYPE:
 392             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 393             {
 394                 if (!reg->idx) reg_maps->vpos = 1;
 395                 else if (reg->idx == 1) reg_maps->usesfacing = 1;
 396             }
 397             break;
 398
 399         case WINED3DSPR_CONST:
 400             if (reg->rel_addr)
 401             {
 402                 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
 403                 {
 404                     if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
 405                     {
 406                         ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
 407                     }
 408                     if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
 409                     {
 410                         ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
 411                     }
 412                 }
 413                 reg_maps->usesrelconstF = TRUE;
 414             }
 415             else
 416             {
 417                 set_bitmap_bit(reg_maps->constf, reg->idx);
 418             }
 419             break;
 420
 421         case WINED3DSPR_CONSTINT:
 422             reg_maps->integer_constants |= (1 << reg->idx);
 423             break;
 424
 425         case WINED3DSPR_CONSTBOOL:
 426             reg_maps->boolean_constants |= (1 << reg->idx);
 427             break;
 428
 429         case WINED3DSPR_COLOROUT:
 430             reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
 431             break;
 432
 433         default:
 434             TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
 435             break;
 436     }
 437 }
 438
 439 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 440 {
 441     switch (instr)
 442     {
 443         case WINED3DSIH_M4x4:
 444         case WINED3DSIH_M3x4:
 445             return param == 1 ? 3 : 0;
 446
 447         case WINED3DSIH_M4x3:
 448         case WINED3DSIH_M3x3:
 449             return param == 1 ? 2 : 0;
 450
 451         case WINED3DSIH_M3x2:
 452             return param == 1 ? 1 : 0;
 453
 454         default:
 455             return 0;
 456     }
 457 }
 458
 459 /* Note that this does not count the loop register as an address register. */
 460 static HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
 461         struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 462         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 463 {
 464     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
 465     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 466     void *fe_data = shader->baseShader.frontend_data;
 467     struct wined3d_shader_version shader_version;
 468     const DWORD *ptr = byte_code;
 469
 470     memset(reg_maps, 0, sizeof(*reg_maps));
 471
 472     /* get_registers_used() is called on every compile on some 1.x shaders,
 473      * which can result in stacking up a collection of local constants.
 474      * Delete the old constants if existing. */
 475     shader_delete_constant_list(&shader->baseShader.constantsF);
 476     shader_delete_constant_list(&shader->baseShader.constantsB);
 477     shader_delete_constant_list(&shader->baseShader.constantsI);
 478
 479     fe->shader_read_header(fe_data, &ptr, &shader_version);
 480     reg_maps->shader_version = shader_version;
 481
 482     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 483             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 484     if (!reg_maps->constf)
 485     {
 486         ERR("Failed to allocate constant map memory.\n");
 487         return E_OUTOFMEMORY;
 488     }
 489
 490     while (!fe->shader_is_end(fe_data, &ptr))
 491     {
 492         struct wined3d_shader_instruction ins;
 493         const char *comment;
 494         UINT comment_size;
 495         UINT param_size;
 496
 497         /* Skip comments. */
 498         fe->shader_read_comment(&ptr, &comment, &comment_size);
 499         if (comment) continue;
 500
 501         /* Fetch opcode. */
 502         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
 503
 504         /* Unhandled opcode, and its parameters. */
 505         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 506         {
 507             TRACE("Skipping unrecognized instruction.\n");
 508             ptr += param_size;
 509             continue;
 510         }
 511
 512         /* Handle declarations. */
 513         if (ins.handler_idx == WINED3DSIH_DCL)
 514         {
 515             struct wined3d_shader_semantic semantic;
 516
 517             fe->shader_read_semantic(&ptr, &semantic);
 518
 519             switch (semantic.reg.reg.type)
 520             {
 521                 /* Mark input registers used. */
 522                 case WINED3DSPR_INPUT:
 523                     reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
 524                     shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
 525                     break;
 526
 527                 /* Vertex shader: mark 3.0 output registers used, save token. */
 528                 case WINED3DSPR_OUTPUT:
 529                     reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
 530                     shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
 531                     if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
 532                     break;
 533
 534                 /* Save sampler usage token. */
 535                 case WINED3DSPR_SAMPLER:
 536                     reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
 537                     break;
 538
 539                 default:
 540                     TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
 541                     break;
 542             }
 543         }
 544         else if (ins.handler_idx == WINED3DSIH_DEF)
 545         {
 546             struct wined3d_shader_src_param rel_addr;
 547             struct wined3d_shader_dst_param dst;
 548
 549             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 550             if (!lconst) return E_OUTOFMEMORY;
 551
 552             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 553             lconst->idx = dst.reg.idx;
 554
 555             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 556             ptr += 4;
 557
 558             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 559             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 560             {
 561                 float *value = (float *)lconst->value;
 562                 if (value[0] < -1.0f) value[0] = -1.0f;
 563                 else if (value[0] > 1.0f) value[0] = 1.0f;
 564                 if (value[1] < -1.0f) value[1] = -1.0f;
 565                 else if (value[1] > 1.0f) value[1] = 1.0f;
 566                 if (value[2] < -1.0f) value[2] = -1.0f;
 567                 else if (value[2] > 1.0f) value[2] = 1.0f;
 568                 if (value[3] < -1.0f) value[3] = -1.0f;
 569                 else if (value[3] > 1.0f) value[3] = 1.0f;
 570             }
 571
 572             list_add_head(&shader->baseShader.constantsF, &lconst->entry);
 573         }
 574         else if (ins.handler_idx == WINED3DSIH_DEFI)
 575         {
 576             struct wined3d_shader_src_param rel_addr;
 577             struct wined3d_shader_dst_param dst;
 578
 579             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 580             if (!lconst) return E_OUTOFMEMORY;
 581
 582             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 583             lconst->idx = dst.reg.idx;
 584
 585             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 586             ptr += 4;
 587
 588             list_add_head(&shader->baseShader.constantsI, &lconst->entry);
 589             reg_maps->local_int_consts |= (1 << dst.reg.idx);
 590         }
 591         else if (ins.handler_idx == WINED3DSIH_DEFB)
 592         {
 593             struct wined3d_shader_src_param rel_addr;
 594             struct wined3d_shader_dst_param dst;
 595
 596             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 597             if (!lconst) return E_OUTOFMEMORY;
 598
 599             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 600             lconst->idx = dst.reg.idx;
 601
 602             memcpy(lconst->value, ptr, sizeof(DWORD));
 603             ++ptr;
 604
 605             list_add_head(&shader->baseShader.constantsB, &lconst->entry);
 606             reg_maps->local_bool_consts |= (1 << dst.reg.idx);
 607         }
 608         /* If there's a loop in the shader. */
 609         else if (ins.handler_idx == WINED3DSIH_LOOP
 610                 || ins.handler_idx == WINED3DSIH_REP)
 611         {
 612             struct wined3d_shader_src_param src, rel_addr;
 613
 614             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 615
 616             /* Rep and Loop always use an integer constant for the control parameters. */
 617             if (ins.handler_idx == WINED3DSIH_REP)
 618             {
 619                 reg_maps->integer_constants |= 1 << src.reg.idx;
 620             }
 621             else
 622             {
 623                 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 624                 reg_maps->integer_constants |= 1 << src.reg.idx;
 625             }
 626
 627             cur_loop_depth++;
 628             if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
 629         }
 630         else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 631                 || ins.handler_idx == WINED3DSIH_ENDREP)
 632         {
 633             cur_loop_depth--;
 634         }
 635         /* For subroutine prototypes. */
 636         else if (ins.handler_idx == WINED3DSIH_LABEL)
 637         {
 638             struct wined3d_shader_src_param src, rel_addr;
 639
 640             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 641             reg_maps->labels |= 1 << src.reg.idx;
 642         }
 643         /* Set texture, address, temporary registers. */
 644         else
 645         {
 646             BOOL color0_mov = FALSE;
 647             int i, limit;
 648
 649             /* This will loop over all the registers and try to
 650              * make a bitmask of the ones we're interested in.
 651              *
 652              * Relative addressing tokens are ignored, but that's
 653              * okay, since we'll catch any address registers when
 654              * they are initialized (required by spec). */
 655             for (i = 0; i < ins.dst_count; ++i)
 656             {
 657                 struct wined3d_shader_src_param dst_rel_addr;
 658                 struct wined3d_shader_dst_param dst_param;
 659
 660                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
 661
 662                 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
 663
 664                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 665                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 666                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 667                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 668                 {
 669                     UINT idx = dst_param.reg.idx;
 670
 671                     switch (dst_param.reg.type)
 672                     {
 673                         case WINED3DSPR_RASTOUT:
 674                             switch (idx)
 675                             {
 676                                 case 0: /* oPos */
 677                                     reg_maps->output_registers |= 1 << 10;
 678                                     shader_signature_from_usage(&output_signature[10],
 679                                             WINED3DDECLUSAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 680                                     break;
 681
 682                                 case 1: /* oFog */
 683                                     reg_maps->output_registers |= 1 << 11;
 684                                     shader_signature_from_usage(&output_signature[11],
 685                                             WINED3DDECLUSAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 686                                     break;
 687
 688                                 case 2: /* oPts */
 689                                     reg_maps->output_registers |= 1 << 11;
 690                                     shader_signature_from_usage(&output_signature[11],
 691                                             WINED3DDECLUSAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 692                                     break;
 693                             }
 694                             break;
 695
 696                         case WINED3DSPR_ATTROUT:
 697                             if (idx < 2)
 698                             {
 699                                 idx += 8;
 700                                 if (reg_maps->output_registers & (1 << idx))
 701                                 {
 702                                     output_signature[idx].mask |= dst_param.write_mask;
 703                                 }
 704                                 else
 705                                 {
 706                                     reg_maps->output_registers |= 1 << idx;
 707                                     shader_signature_from_usage(&output_signature[idx],
 708                                             WINED3DDECLUSAGE_COLOR, idx - 8, idx, dst_param.write_mask);
 709                                 }
 710                             }
 711                             break;
 712
 713                         case WINED3DSPR_TEXCRDOUT:
 714
 715                             reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
 716                             if (reg_maps->output_registers & (1 << idx))
 717                             {
 718                                 output_signature[idx].mask |= dst_param.write_mask;
 719                             }
 720                             else
 721                             {
 722                                 reg_maps->output_registers |= 1 << idx;
 723                                 shader_signature_from_usage(&output_signature[idx],
 724                                         WINED3DDECLUSAGE_TEXCOORD, idx, idx, dst_param.write_mask);
 725                             }
 726                             break;
 727
 728                         default:
 729                             break;
 730                     }
 731                 }
 732
 733                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 734                 {
 735                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 736
 737                     if (dst_param.reg.type == WINED3DSPR_COLOROUT && !dst_param.reg.idx)
 738                     {
 739                     /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 740                      * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 741                      * the mov and perform the sRGB write correction from the source register.
 742                      *
 743                      * However, if the mov is only partial, we can't do this, and if the write
 744                      * comes from an instruction other than MOV it is hard to do as well. If
 745                      * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 746
 747                         ps->color0_mov = FALSE;
 748                         if (ins.handler_idx == WINED3DSIH_MOV
 749                                 && dst_param.write_mask == WINED3DSP_WRITEMASK_ALL)
 750                         {
 751                             /* Used later when the source register is read. */
 752                             color0_mov = TRUE;
 753                         }
 754                     }
 755                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 756                      * end
 757                      */
 758                     else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
 759                     {
 760                         ps->color0_mov = FALSE;
 761                     }
 762                 }
 763
 764                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 765                 if (shader_version.major == 1
 766                         && (ins.handler_idx == WINED3DSIH_TEX
 767                             || ins.handler_idx == WINED3DSIH_TEXBEM
 768                             || ins.handler_idx == WINED3DSIH_TEXBEML
 769                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 770                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 771                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 772                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 773                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 774                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 775                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 776                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 777                 {
 778                     /* Fake sampler usage, only set reserved bit and type. */
 779                     DWORD sampler_code = dst_param.reg.idx;
 780
 781                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 782                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 783
 784                     /* texbem is only valid with < 1.4 pixel shaders */
 785                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 786                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 787                     {
 788                         reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 789                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 790                         {
 791                             reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
 792                         }
 793                     }
 794                 }
 795                 else if (ins.handler_idx == WINED3DSIH_BEM)
 796                 {
 797                     reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 798                 }
 799             }
 800
 801             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 802             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 803             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 804             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 805             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 806             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 807             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 808             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 809             else if (ins.handler_idx == WINED3DSIH_RCP) reg_maps->usesrcp = 1;
 810
 811             limit = ins.src_count + (ins.predicate ? 1 : 0);
 812             for (i = 0; i < limit; ++i)
 813             {
 814                 struct wined3d_shader_src_param src_param, src_rel_addr;
 815                 unsigned int count;
 816
 817                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
 818                 count = get_instr_extra_regcount(ins.handler_idx, i);
 819
 820                 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 821                 while (count)
 822                 {
 823                     ++src_param.reg.idx;
 824                     shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 825                     --count;
 826                 }
 827
 828                 if (color0_mov)
 829                 {
 830                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 831                     if (src_param.reg.type == WINED3DSPR_TEMP
 832                             && src_param.swizzle == WINED3DSP_NOSWIZZLE)
 833                     {
 834                         ps->color0_mov = TRUE;
 835                         ps->color0_reg = src_param.reg.idx;
 836                     }
 837                 }
 838             }
 839         }
 840     }
 841     reg_maps->loop_depth = max_loop_depth;
 842
 843     shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
 844
 845     return WINED3D_OK;
 846 }
 847
 848 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
 849 {
 850     DWORD map = 1 << max;
 851     map |= map - 1;
 852     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 853
 854     return wined3d_log2i(map);
 855 }
 856
 857 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 858         const struct wined3d_shader_version *shader_version)
 859 {
 860     TRACE("dcl");
 861
 862     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 863     {
 864         switch (semantic->sampler_type)
 865         {
 866             case WINED3DSTT_2D: TRACE("_2d"); break;
 867             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 868             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 869             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 870         }
 871     }
 872     else
 873     {
 874         /* Pixel shaders 3.0 don't have usage semantics. */
 875         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 876         else TRACE("_");
 877
 878         switch (semantic->usage)
 879         {
 880             case WINED3DDECLUSAGE_POSITION:
 881                 TRACE("position%u", semantic->usage_idx);
 882                 break;
 883
 884             case WINED3DDECLUSAGE_BLENDINDICES:
 885                 TRACE("blend");
 886                 break;
 887
 888             case WINED3DDECLUSAGE_BLENDWEIGHT:
 889                 TRACE("weight");
 890                 break;
 891
 892             case WINED3DDECLUSAGE_NORMAL:
 893                 TRACE("normal%u", semantic->usage_idx);
 894                 break;
 895
 896             case WINED3DDECLUSAGE_PSIZE:
 897                 TRACE("psize");
 898                 break;
 899
 900             case WINED3DDECLUSAGE_COLOR:
 901                 if (!semantic->usage_idx) TRACE("color");
 902                 else TRACE("specular%u", (semantic->usage_idx - 1));
 903                 break;
 904
 905             case WINED3DDECLUSAGE_TEXCOORD:
 906                 TRACE("texture%u", semantic->usage_idx);
 907                 break;
 908
 909             case WINED3DDECLUSAGE_TANGENT:
 910                 TRACE("tangent");
 911                 break;
 912
 913             case WINED3DDECLUSAGE_BINORMAL:
 914                 TRACE("binormal");
 915                 break;
 916
 917             case WINED3DDECLUSAGE_TESSFACTOR:
 918                 TRACE("tessfactor");
 919                 break;
 920
 921             case WINED3DDECLUSAGE_POSITIONT:
 922                 TRACE("positionT%u", semantic->usage_idx);
 923                 break;
 924
 925             case WINED3DDECLUSAGE_FOG:
 926                 TRACE("fog");
 927                 break;
 928
 929             case WINED3DDECLUSAGE_DEPTH:
 930                 TRACE("depth");
 931                 break;
 932
 933             case WINED3DDECLUSAGE_SAMPLE:
 934                 TRACE("sample");
 935                 break;
 936
 937             default:
 938                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 939         }
 940     }
 941 }
 942
 943 static void shader_dump_register(const struct wined3d_shader_register *reg,
 944         const struct wined3d_shader_version *shader_version)
 945 {
 946     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 947     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 948     UINT offset = reg->idx;
 949
 950     switch (reg->type)
 951     {
 952         case WINED3DSPR_TEMP:
 953             TRACE("r");
 954             break;
 955
 956         case WINED3DSPR_INPUT:
 957             TRACE("v");
 958             break;
 959
 960         case WINED3DSPR_CONST:
 961         case WINED3DSPR_CONST2:
 962         case WINED3DSPR_CONST3:
 963         case WINED3DSPR_CONST4:
 964             TRACE("c");
 965             offset = shader_get_float_offset(reg->type, reg->idx);
 966             break;
 967
 968         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 969             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 970             break;
 971
 972         case WINED3DSPR_RASTOUT:
 973             TRACE("%s", rastout_reg_names[reg->idx]);
 974             break;
 975
 976         case WINED3DSPR_COLOROUT:
 977             TRACE("oC");
 978             break;
 979
 980         case WINED3DSPR_DEPTHOUT:
 981             TRACE("oDepth");
 982             break;
 983
 984         case WINED3DSPR_ATTROUT:
 985             TRACE("oD");
 986             break;
 987
 988         case WINED3DSPR_TEXCRDOUT:
 989             /* Vertex shaders >= 3.0 use general purpose output registers
 990              * (WINED3DSPR_OUTPUT), which can include an address token. */
 991             if (shader_version->major >= 3) TRACE("o");
 992             else TRACE("oT");
 993             break;
 994
 995         case WINED3DSPR_CONSTINT:
 996             TRACE("i");
 997             break;
 998
 999         case WINED3DSPR_CONSTBOOL:
1000             TRACE("b");
1001             break;
1002
1003         case WINED3DSPR_LABEL:
1004             TRACE("l");
1005             break;
1006
1007         case WINED3DSPR_LOOP:
1008             TRACE("aL");
1009             break;
1010
1011         case WINED3DSPR_SAMPLER:
1012             TRACE("s");
1013             break;
1014
1015         case WINED3DSPR_MISCTYPE:
1016             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
1017             else TRACE("%s", misctype_reg_names[reg->idx]);
1018             break;
1019
1020         case WINED3DSPR_PREDICATE:
1021             TRACE("p");
1022             break;
1023
1024         case WINED3DSPR_IMMCONST:
1025             TRACE("l");
1026             break;
1027
1028         case WINED3DSPR_CONSTBUFFER:
1029             TRACE("cb");
1030             break;
1031
1032         case WINED3DSPR_NULL:
1033             TRACE("null");
1034             break;
1035
1036         case WINED3DSPR_RESOURCE:
1037             TRACE("t");
1038             break;
1039
1040         default:
1041             TRACE("unhandled_rtype(%#x)", reg->type);
1042             break;
1043     }
1044
1045     if (reg->type == WINED3DSPR_IMMCONST)
1046     {
1047         TRACE("(");
1048         switch (reg->immconst_type)
1049         {
1050             case WINED3D_IMMCONST_SCALAR:
1051                 TRACE("%.8e", *(const float *)reg->immconst_data);
1052                 break;
1053
1054             case WINED3D_IMMCONST_VEC4:
1055                 TRACE("%.8e, %.8e, %.8e, %.8e",
1056                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1057                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1058                 break;
1059
1060             default:
1061                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1062                 break;
1063         }
1064         TRACE(")");
1065     }
1066     else if (reg->type != WINED3DSPR_RASTOUT
1067             && reg->type != WINED3DSPR_MISCTYPE
1068             && reg->type != WINED3DSPR_NULL)
1069     {
1070         if (reg->array_idx != ~0U)
1071         {
1072             TRACE("%u[%u", offset, reg->array_idx);
1073             if (reg->rel_addr)
1074             {
1075                 TRACE(" + ");
1076                 shader_dump_src_param(reg->rel_addr, shader_version);
1077             }
1078             TRACE("]");
1079         }
1080         else
1081         {
1082             if (reg->rel_addr)
1083             {
1084                 TRACE("[");
1085                 shader_dump_src_param(reg->rel_addr, shader_version);
1086                 TRACE(" + ");
1087             }
1088             TRACE("%u", offset);
1089             if (reg->rel_addr) TRACE("]");
1090         }
1091     }
1092 }
1093
1094 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1095         const struct wined3d_shader_version *shader_version)
1096 {
1097     DWORD write_mask = param->write_mask;
1098
1099     shader_dump_register(&param->reg, shader_version);
1100
1101     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1102     {
1103         static const char *write_mask_chars = "xyzw";
1104
1105         TRACE(".");
1106         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1107         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1108         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1109         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1110     }
1111 }
1112
1113 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1114         const struct wined3d_shader_version *shader_version)
1115 {
1116     DWORD src_modifier = param->modifiers;
1117     DWORD swizzle = param->swizzle;
1118
1119     if (src_modifier == WINED3DSPSM_NEG
1120             || src_modifier == WINED3DSPSM_BIASNEG
1121             || src_modifier == WINED3DSPSM_SIGNNEG
1122             || src_modifier == WINED3DSPSM_X2NEG
1123             || src_modifier == WINED3DSPSM_ABSNEG)
1124         TRACE("-");
1125     else if (src_modifier == WINED3DSPSM_COMP)
1126         TRACE("1-");
1127     else if (src_modifier == WINED3DSPSM_NOT)
1128         TRACE("!");
1129
1130     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1131         TRACE("abs(");
1132
1133     shader_dump_register(&param->reg, shader_version);
1134
1135     if (src_modifier)
1136     {
1137         switch (src_modifier)
1138         {
1139             case WINED3DSPSM_NONE:    break;
1140             case WINED3DSPSM_NEG:     break;
1141             case WINED3DSPSM_NOT:     break;
1142             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1143             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1144             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1145             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1146             case WINED3DSPSM_COMP:    break;
1147             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1148             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1149             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1150             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1151             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1152             case WINED3DSPSM_ABS:     TRACE(")"); break;
1153             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1154         }
1155     }
1156
1157     if (swizzle != WINED3DSP_NOSWIZZLE)
1158     {
1159         static const char *swizzle_chars = "xyzw";
1160         DWORD swizzle_x = swizzle & 0x03;
1161         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1162         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1163         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1164
1165         if (swizzle_x == swizzle_y
1166                 && swizzle_x == swizzle_z
1167                 && swizzle_x == swizzle_w)
1168         {
1169             TRACE(".%c", swizzle_chars[swizzle_x]);
1170         }
1171         else
1172         {
1173             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1174                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1175         }
1176     }
1177 }
1178
1179 /* Shared code in order to generate the bulk of the shader string.
1180  * NOTE: A description of how to parse tokens can be found on MSDN. */
1181 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1182         const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1183 {
1184     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1185     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1186     const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1187     void *fe_data = shader->baseShader.frontend_data;
1188     struct wined3d_shader_src_param dst_rel_addr[2];
1189     struct wined3d_shader_src_param src_rel_addr[4];
1190     struct wined3d_shader_dst_param dst_param[2];
1191     struct wined3d_shader_src_param src_param[4];
1192     struct wined3d_shader_version shader_version;
1193     struct wined3d_shader_loop_state loop_state;
1194     struct wined3d_shader_instruction ins;
1195     struct wined3d_shader_tex_mx tex_mx;
1196     struct wined3d_shader_context ctx;
1197     const DWORD *ptr = byte_code;
1198     DWORD i;
1199
1200     /* Initialize current parsing state. */
1201     tex_mx.current_row = 0;
1202     loop_state.current_depth = 0;
1203     loop_state.current_reg = 0;
1204
1205     ctx.shader = iface;
1206     ctx.gl_info = &device->adapter->gl_info;
1207     ctx.reg_maps = reg_maps;
1208     ctx.buffer = buffer;
1209     ctx.tex_mx = &tex_mx;
1210     ctx.loop_state = &loop_state;
1211     ctx.backend_data = backend_ctx;
1212
1213     ins.ctx = &ctx;
1214     ins.dst = dst_param;
1215     ins.src = src_param;
1216
1217     fe->shader_read_header(fe_data, &ptr, &shader_version);
1218
1219     while (!fe->shader_is_end(fe_data, &ptr))
1220     {
1221         const char *comment;
1222         UINT comment_size;
1223         UINT param_size;
1224
1225         /* Skip comment tokens. */
1226         fe->shader_read_comment(&ptr, &comment, &comment_size);
1227         if (comment) continue;
1228
1229         /* Read opcode. */
1230         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1231
1232         /* Unknown opcode and its parameters. */
1233         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1234         {
1235             TRACE("Skipping unrecognized instruction.\n");
1236             ptr += param_size;
1237             continue;
1238         }
1239
1240         /* Nothing to do. */
1241         if (ins.handler_idx == WINED3DSIH_DCL
1242                 || ins.handler_idx == WINED3DSIH_NOP
1243                 || ins.handler_idx == WINED3DSIH_DEF
1244                 || ins.handler_idx == WINED3DSIH_DEFI
1245                 || ins.handler_idx == WINED3DSIH_DEFB
1246                 || ins.handler_idx == WINED3DSIH_PHASE)
1247         {
1248             ptr += param_size;
1249             continue;
1250         }
1251
1252         /* Destination tokens */
1253         for (i = 0; i < ins.dst_count; ++i)
1254         {
1255             fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1256         }
1257
1258         /* Predication token */
1259         if (ins.predicate)
1260         {
1261             FIXME("Predicates not implemented.\n");
1262             ins.predicate = *ptr++;
1263         }
1264
1265         /* Other source tokens */
1266         for (i = 0; i < ins.src_count; ++i)
1267         {
1268             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1269         }
1270
1271         /* Call appropriate function for output target */
1272         device->shader_backend->shader_handle_instruction(&ins);
1273     }
1274 }
1275
1276 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1277 {
1278     DWORD mmask = dst->modifiers;
1279
1280     switch (dst->shift)
1281     {
1282         case 0: break;
1283         case 13: TRACE("_d8"); break;
1284         case 14: TRACE("_d4"); break;
1285         case 15: TRACE("_d2"); break;
1286         case 1: TRACE("_x2"); break;
1287         case 2: TRACE("_x4"); break;
1288         case 3: TRACE("_x8"); break;
1289         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1290     }
1291
1292     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1293     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1294     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1295
1296     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1297     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1298 }
1299
1300 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1301 {
1302     struct wined3d_shader_version shader_version;
1303     const DWORD *ptr = byte_code;
1304     const char *type_prefix;
1305     DWORD i;
1306
1307     TRACE("Parsing %p.\n", byte_code);
1308
1309     fe->shader_read_header(fe_data, &ptr, &shader_version);
1310
1311     switch (shader_version.type)
1312     {
1313         case WINED3D_SHADER_TYPE_VERTEX:
1314             type_prefix = "vs";
1315             break;
1316
1317         case WINED3D_SHADER_TYPE_GEOMETRY:
1318             type_prefix = "gs";
1319             break;
1320
1321         case WINED3D_SHADER_TYPE_PIXEL:
1322             type_prefix = "ps";
1323             break;
1324
1325         default:
1326             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1327             type_prefix = "unknown";
1328             break;
1329     }
1330
1331     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1332
1333     while (!fe->shader_is_end(fe_data, &ptr))
1334     {
1335         struct wined3d_shader_instruction ins;
1336         const char *comment;
1337         UINT comment_size;
1338         UINT param_size;
1339
1340         /* comment */
1341         fe->shader_read_comment(&ptr, &comment, &comment_size);
1342         if (comment)
1343         {
1344             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1345             {
1346                 const char *end = comment + comment_size;
1347                 const char *ptr = comment + 4;
1348                 const char *line = ptr;
1349
1350                 TRACE("// TEXT\n");
1351                 while (ptr != end)
1352                 {
1353                     if (*ptr == '\n')
1354                     {
1355                         UINT len = ptr - line;
1356                         if (len && *(ptr - 1) == '\r') --len;
1357                         TRACE("// %s\n", debugstr_an(line, len));
1358                         line = ++ptr;
1359                     }
1360                     else ++ptr;
1361                 }
1362                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1363             }
1364             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1365             continue;
1366         }
1367
1368         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1369         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1370         {
1371             TRACE("Skipping unrecognized instruction.\n");
1372             ptr += param_size;
1373             continue;
1374         }
1375
1376         if (ins.handler_idx == WINED3DSIH_DCL)
1377         {
1378             struct wined3d_shader_semantic semantic;
1379
1380             fe->shader_read_semantic(&ptr, &semantic);
1381
1382             shader_dump_decl_usage(&semantic, &shader_version);
1383             shader_dump_ins_modifiers(&semantic.reg);
1384             TRACE(" ");
1385             shader_dump_dst_param(&semantic.reg, &shader_version);
1386         }
1387         else if (ins.handler_idx == WINED3DSIH_DEF)
1388         {
1389             struct wined3d_shader_src_param rel_addr;
1390             struct wined3d_shader_dst_param dst;
1391
1392             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1393
1394             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1395                     *(const float *)(ptr),
1396                     *(const float *)(ptr + 1),
1397                     *(const float *)(ptr + 2),
1398                     *(const float *)(ptr + 3));
1399             ptr += 4;
1400         }
1401         else if (ins.handler_idx == WINED3DSIH_DEFI)
1402         {
1403             struct wined3d_shader_src_param rel_addr;
1404             struct wined3d_shader_dst_param dst;
1405
1406             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1407
1408             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1409                     *(ptr),
1410                     *(ptr + 1),
1411                     *(ptr + 2),
1412                     *(ptr + 3));
1413             ptr += 4;
1414         }
1415         else if (ins.handler_idx == WINED3DSIH_DEFB)
1416         {
1417             struct wined3d_shader_src_param rel_addr;
1418             struct wined3d_shader_dst_param dst;
1419
1420             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1421
1422             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1423             ++ptr;
1424         }
1425         else
1426         {
1427             struct wined3d_shader_src_param dst_rel_addr[2];
1428             struct wined3d_shader_src_param src_rel_addr;
1429             struct wined3d_shader_dst_param dst_param[2];
1430             struct wined3d_shader_src_param src_param;
1431
1432             for (i = 0; i < ins.dst_count; ++i)
1433             {
1434                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1435             }
1436
1437             /* Print out predication source token first - it follows
1438              * the destination token. */
1439             if (ins.predicate)
1440             {
1441                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1442                 TRACE("(");
1443                 shader_dump_src_param(&src_param, &shader_version);
1444                 TRACE(") ");
1445             }
1446
1447             /* PixWin marks instructions with the coissue flag with a '+' */
1448             if (ins.coissue) TRACE("+");
1449
1450             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1451
1452             if (ins.handler_idx == WINED3DSIH_IFC
1453                     || ins.handler_idx == WINED3DSIH_BREAKC)
1454             {
1455                 switch (ins.flags)
1456                 {
1457                     case COMPARISON_GT: TRACE("_gt"); break;
1458                     case COMPARISON_EQ: TRACE("_eq"); break;
1459                     case COMPARISON_GE: TRACE("_ge"); break;
1460                     case COMPARISON_LT: TRACE("_lt"); break;
1461                     case COMPARISON_NE: TRACE("_ne"); break;
1462                     case COMPARISON_LE: TRACE("_le"); break;
1463                     default: TRACE("_(%u)", ins.flags);
1464                 }
1465             }
1466             else if (ins.handler_idx == WINED3DSIH_TEX
1467                     && shader_version.major >= 2
1468                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1469             {
1470                 TRACE("p");
1471             }
1472
1473             /* We already read the destination tokens, print them. */
1474             for (i = 0; i < ins.dst_count; ++i)
1475             {
1476                 shader_dump_ins_modifiers(&dst_param[i]);
1477                 TRACE(!i ? " " : ", ");
1478                 shader_dump_dst_param(&dst_param[i], &shader_version);
1479             }
1480
1481             /* Other source tokens */
1482             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1483             {
1484                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1485                 TRACE(!i ? " " : ", ");
1486                 shader_dump_src_param(&src_param, &shader_version);
1487             }
1488         }
1489         TRACE("\n");
1490     }
1491 }
1492
1493 static void shader_cleanup(IWineD3DBaseShader *iface)
1494 {
1495     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1496
1497     ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1498     HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1499     HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1500     shader_delete_constant_list(&shader->baseShader.constantsF);
1501     shader_delete_constant_list(&shader->baseShader.constantsB);
1502     shader_delete_constant_list(&shader->baseShader.constantsI);
1503     list_remove(&shader->baseShader.shader_list_entry);
1504
1505     if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1506     {
1507         shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1508     }
1509 }
1510
1511 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1512 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1513 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type, const SIZE *ds_mask_size) {}
1514 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1515 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1516 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1517 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1518 static void shader_none_load_np2fixup_constants(void *shader_priv,
1519         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1520 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1521 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1522 static void shader_none_free(IWineD3DDevice *iface) {}
1523 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1524
1525 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1526 {
1527     /* Set the shader caps to 0 for the none shader backend */
1528     caps->VertexShaderVersion = 0;
1529     caps->MaxVertexShaderConst = 0;
1530     caps->PixelShaderVersion = 0;
1531     caps->PixelShader1xMaxValue = 0.0f;
1532     caps->MaxPixelShaderConst = 0;
1533     caps->VSClipping = FALSE;
1534 }
1535
1536 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1537 {
1538     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1539     {
1540         TRACE("Checking support for fixup:\n");
1541         dump_color_fixup_desc(fixup);
1542     }
1543
1544     /* Faked to make some apps happy. */
1545     if (!is_complex_fixup(fixup))
1546     {
1547         TRACE("[OK]\n");
1548         return TRUE;
1549     }
1550
1551     TRACE("[FAILED]\n");
1552     return FALSE;
1553 }
1554
1555 const shader_backend_t none_shader_backend = {
1556     shader_none_handle_instruction,
1557     shader_none_select,
1558     shader_none_select_depth_blt,
1559     shader_none_deselect_depth_blt,
1560     shader_none_update_float_vertex_constants,
1561     shader_none_update_float_pixel_constants,
1562     shader_none_load_constants,
1563     shader_none_load_np2fixup_constants,
1564     shader_none_destroy,
1565     shader_none_alloc,
1566     shader_none_free,
1567     shader_none_dirty_const,
1568     shader_none_get_caps,
1569     shader_none_color_fixup_supported,
1570 };
1571
1572 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1573 {
1574     if (!data)
1575     {
1576         *data_size = shader->baseShader.functionLength;
1577         return WINED3D_OK;
1578     }
1579
1580     if (*data_size < shader->baseShader.functionLength)
1581     {
1582         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1583          * than the required size we should write the required size and
1584          * return D3DERR_MOREDATA. That's not actually true. */
1585         return WINED3DERR_INVALIDCALL;
1586     }
1587
1588     memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1589
1590     return WINED3D_OK;
1591 }
1592
1593 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1594         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1595 {
1596     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1597     const struct wined3d_shader_frontend *fe;
1598     HRESULT hr;
1599
1600     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1601             shader, byte_code, output_signature, float_const_count);
1602
1603     fe = shader_select_frontend(*byte_code);
1604     if (!fe)
1605     {
1606         FIXME("Unable to find frontend for shader.\n");
1607         return WINED3DERR_INVALIDCALL;
1608     }
1609     shader->baseShader.frontend = fe;
1610     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1611     if (!shader->baseShader.frontend_data)
1612     {
1613         FIXME("Failed to initialize frontend.\n");
1614         return WINED3DERR_INVALIDCALL;
1615     }
1616
1617     /* First pass: trace shader. */
1618     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1619
1620     /* Initialize immediate constant lists. */
1621     list_init(&shader->baseShader.constantsF);
1622     list_init(&shader->baseShader.constantsB);
1623     list_init(&shader->baseShader.constantsI);
1624
1625     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1626     hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1627             reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1628             byte_code, float_const_count);
1629     if (FAILED(hr)) return hr;
1630
1631     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1632     if (!shader->baseShader.function) return E_OUTOFMEMORY;
1633     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1634
1635     return WINED3D_OK;
1636 }
1637
1638 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1639 {
1640     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1641
1642     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1643             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1644             || IsEqualGUID(riid, &IID_IWineD3DBase)
1645             || IsEqualGUID(riid, &IID_IUnknown))
1646     {
1647         IUnknown_AddRef(iface);
1648         *object = iface;
1649         return S_OK;
1650     }
1651
1652     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1653
1654     *object = NULL;
1655     return E_NOINTERFACE;
1656 }
1657
1658 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1659 {
1660     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1661     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1662
1663     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1664
1665     return refcount;
1666 }
1667
1668 /* Do not call while under the GL lock. */
1669 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1670 {
1671     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1672     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1673
1674     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1675
1676     if (!refcount)
1677     {
1678         shader_cleanup((IWineD3DBaseShader *)iface);
1679         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1680         HeapFree(GetProcessHeap(), 0, shader);
1681     }
1682
1683     return refcount;
1684 }
1685
1686 static void * STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface)
1687 {
1688     TRACE("iface %p.\n", iface);
1689
1690     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1691 }
1692
1693 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1694 {
1695     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1696
1697     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1698 }
1699
1700 /* Set local constants for d3d8 shaders. */
1701 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1702         UINT start_idx, const float *src_data, UINT count)
1703 {
1704     IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1705     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1706     UINT i, end_idx;
1707
1708     TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1709
1710     end_idx = start_idx + count;
1711     if (end_idx > device->d3d_vshader_constantF)
1712     {
1713         WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1714         end_idx = device->d3d_vshader_constantF;
1715     }
1716
1717     for (i = start_idx; i < end_idx; ++i)
1718     {
1719         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1720         if (!lconst) return E_OUTOFMEMORY;
1721
1722         lconst->idx = i;
1723         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1724         list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1725     }
1726
1727     return WINED3D_OK;
1728 }
1729
1730 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1731 {
1732     /* IUnknown methods */
1733     vertexshader_QueryInterface,
1734     vertexshader_AddRef,
1735     vertexshader_Release,
1736     /* IWineD3DBase methods */
1737     vertexshader_GetParent,
1738     /* IWineD3DBaseShader methods */
1739     vertexshader_GetFunction,
1740     /* IWineD3DVertexShader methods */
1741     vertexshader_SetLocalConstantsF,
1742 };
1743
1744 void find_vs_compile_args(const struct wined3d_state *state,
1745         IWineD3DVertexShaderImpl *shader, struct vs_compile_args *args)
1746 {
1747     args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1748             == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1749     args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1750             && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1751     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1752 }
1753
1754 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1755 {
1756     if (usage_idx1 != usage_idx2) return FALSE;
1757     if (usage1 == usage2) return TRUE;
1758     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1759     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1760
1761     return FALSE;
1762 }
1763
1764 BOOL vshader_get_input(struct IWineD3DVertexShaderImpl *shader,
1765         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1766 {
1767     WORD map = shader->baseShader.reg_maps.input_registers;
1768     unsigned int i;
1769
1770     for (i = 0; map; map >>= 1, ++i)
1771     {
1772         if (!(map & 1)) continue;
1773
1774         if (match_usage(shader->attributes[i].usage,
1775                 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1776         {
1777             *regnum = i;
1778             return TRUE;
1779         }
1780     }
1781     return FALSE;
1782 }
1783
1784 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1785 {
1786     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1787             shader->baseShader.reg_maps.shader_version.minor);
1788     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1789
1790     shader->baseShader.limits.texcoord = 0;
1791     shader->baseShader.limits.attributes = 16;
1792     shader->baseShader.limits.packed_input = 0;
1793
1794     switch (shader_version)
1795     {
1796         case WINED3D_SHADER_VERSION(1, 0):
1797         case WINED3D_SHADER_VERSION(1, 1):
1798             shader->baseShader.limits.temporary = 12;
1799             shader->baseShader.limits.constant_bool = 0;
1800             shader->baseShader.limits.constant_int = 0;
1801             shader->baseShader.limits.address = 1;
1802             shader->baseShader.limits.packed_output = 12;
1803             shader->baseShader.limits.sampler = 0;
1804             shader->baseShader.limits.label = 0;
1805             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1806              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1807              * constants? */
1808             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1809             break;
1810
1811         case WINED3D_SHADER_VERSION(2, 0):
1812         case WINED3D_SHADER_VERSION(2, 1):
1813             shader->baseShader.limits.temporary = 12;
1814             shader->baseShader.limits.constant_bool = 16;
1815             shader->baseShader.limits.constant_int = 16;
1816             shader->baseShader.limits.address = 1;
1817             shader->baseShader.limits.packed_output = 12;
1818             shader->baseShader.limits.sampler = 0;
1819             shader->baseShader.limits.label = 16;
1820             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1821             break;
1822
1823         case WINED3D_SHADER_VERSION(4, 0):
1824             FIXME("Using 3.0 limits for 4.0 shader.\n");
1825             /* Fall through. */
1826
1827         case WINED3D_SHADER_VERSION(3, 0):
1828             shader->baseShader.limits.temporary = 32;
1829             shader->baseShader.limits.constant_bool = 32;
1830             shader->baseShader.limits.constant_int = 32;
1831             shader->baseShader.limits.address = 1;
1832             shader->baseShader.limits.packed_output = 12;
1833             shader->baseShader.limits.sampler = 4;
1834             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1835             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1836              * even though they are capable of supporting much more (GL
1837              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1838              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1839              * shaders to 256. */
1840             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1841             break;
1842
1843         default:
1844             shader->baseShader.limits.temporary = 12;
1845             shader->baseShader.limits.constant_bool = 16;
1846             shader->baseShader.limits.constant_int = 16;
1847             shader->baseShader.limits.address = 1;
1848             shader->baseShader.limits.packed_output = 12;
1849             shader->baseShader.limits.sampler = 0;
1850             shader->baseShader.limits.label = 16;
1851             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1852             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1853                     shader->baseShader.reg_maps.shader_version.major,
1854                     shader->baseShader.reg_maps.shader_version.minor);
1855     }
1856 }
1857
1858 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1859         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1860         void *parent, const struct wined3d_parent_ops *parent_ops)
1861 {
1862     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1863     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1864     unsigned int i;
1865     HRESULT hr;
1866     WORD map;
1867
1868     if (!byte_code) return WINED3DERR_INVALIDCALL;
1869
1870     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1871     shader_init(&shader->baseShader, device, parent, parent_ops);
1872
1873     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1874             output_signature, device->d3d_vshader_constantF);
1875     if (FAILED(hr))
1876     {
1877         WARN("Failed to set function, hr %#x.\n", hr);
1878         shader_cleanup((IWineD3DBaseShader *)shader);
1879         return hr;
1880     }
1881
1882     map = shader->baseShader.reg_maps.input_registers;
1883     for (i = 0; map; map >>= 1, ++i)
1884     {
1885         if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1886
1887         shader->attributes[i].usage =
1888                 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1889         shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1890     }
1891
1892     if (output_signature)
1893     {
1894         for (i = 0; i < output_signature->element_count; ++i)
1895         {
1896             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1897             reg_maps->output_registers |= 1 << e->register_idx;
1898             shader->baseShader.output_signature[e->register_idx] = *e;
1899         }
1900     }
1901
1902     vertexshader_set_limits(shader);
1903
1904     if (device->vs_selected_mode == SHADER_ARB
1905             && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1906             && shader->min_rel_offset <= shader->max_rel_offset)
1907     {
1908         if (shader->max_rel_offset - shader->min_rel_offset > 127)
1909         {
1910             FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1911             FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1912             FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1913         }
1914         else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1915         {
1916             shader->rel_offset = shader->min_rel_offset + 63;
1917         }
1918         else if (shader->max_rel_offset > 63)
1919         {
1920             shader->rel_offset = shader->min_rel_offset;
1921         }
1922         else
1923         {
1924             shader->rel_offset = 0;
1925         }
1926     }
1927
1928     shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1929             && !list_empty(&shader->baseShader.constantsF);
1930
1931     return WINED3D_OK;
1932 }
1933
1934 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1935         REFIID riid, void **object)
1936 {
1937     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1938
1939     if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1940             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1941             || IsEqualGUID(riid, &IID_IWineD3DBase)
1942             || IsEqualGUID(riid, &IID_IUnknown))
1943     {
1944         IUnknown_AddRef(iface);
1945         *object = iface;
1946         return S_OK;
1947     }
1948
1949     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1950
1951     *object = NULL;
1952     return E_NOINTERFACE;
1953 }
1954
1955 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1956 {
1957     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1958     ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1959
1960     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1961
1962     return refcount;
1963 }
1964
1965 /* Do not call while under the GL lock. */
1966 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1967 {
1968     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1969     ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1970
1971     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1972
1973     if (!refcount)
1974     {
1975         shader_cleanup((IWineD3DBaseShader *)iface);
1976         shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1977         HeapFree(GetProcessHeap(), 0, shader);
1978     }
1979
1980     return refcount;
1981 }
1982
1983 static void * STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface)
1984 {
1985     TRACE("iface %p.\n", iface);
1986
1987     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1988 }
1989
1990 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1991 {
1992     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1993
1994     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1995 }
1996
1997 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1998 {
1999     /* IUnknown methods */
2000     geometryshader_QueryInterface,
2001     geometryshader_AddRef,
2002     geometryshader_Release,
2003     /* IWineD3DBase methods */
2004     geometryshader_GetParent,
2005     /* IWineD3DBaseShader methods */
2006     geometryshader_GetFunction,
2007 };
2008
2009 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
2010         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2011         void *parent, const struct wined3d_parent_ops *parent_ops)
2012 {
2013     HRESULT hr;
2014
2015     shader->vtbl = &wined3d_geometryshader_vtbl;
2016     shader_init(&shader->base_shader, device, parent, parent_ops);
2017
2018     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
2019     if (FAILED(hr))
2020     {
2021         WARN("Failed to set function, hr %#x.\n", hr);
2022         shader_cleanup((IWineD3DBaseShader *)shader);
2023         return hr;
2024     }
2025
2026     shader->base_shader.load_local_constsF = FALSE;
2027
2028     return WINED3D_OK;
2029 }
2030
2031 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
2032 {
2033     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
2034
2035     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
2036             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
2037             || IsEqualGUID(riid, &IID_IWineD3DBase)
2038             || IsEqualGUID(riid, &IID_IUnknown))
2039     {
2040         IUnknown_AddRef(iface);
2041         *object = iface;
2042         return S_OK;
2043     }
2044
2045     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
2046
2047     *object = NULL;
2048     return E_NOINTERFACE;
2049 }
2050
2051 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
2052 {
2053     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2054     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
2055
2056     TRACE("%p increasing refcount to %u.\n", shader, refcount);
2057
2058     return refcount;
2059 }
2060
2061 /* Do not call while under the GL lock. */
2062 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
2063 {
2064     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2065     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
2066
2067     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2068
2069     if (!refcount)
2070     {
2071         shader_cleanup((IWineD3DBaseShader *)iface);
2072         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
2073         HeapFree(GetProcessHeap(), 0, shader);
2074     }
2075
2076     return refcount;
2077 }
2078
2079 static void * STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface)
2080 {
2081     TRACE("iface %p.\n", iface);
2082
2083     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
2084 }
2085
2086 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
2087 {
2088     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
2089
2090     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
2091 }
2092
2093 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2094 {
2095     /* IUnknown methods */
2096     pixelshader_QueryInterface,
2097     pixelshader_AddRef,
2098     pixelshader_Release,
2099     /* IWineD3DBase methods */
2100     pixelshader_GetParent,
2101     /* IWineD3DBaseShader methods */
2102     pixelshader_GetFunction
2103 };
2104
2105 void find_ps_compile_args(const struct wined3d_state *state,
2106         IWineD3DPixelShaderImpl *shader, struct ps_compile_args *args)
2107 {
2108     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
2109     IWineD3DBaseTextureImpl *texture;
2110     UINT i;
2111
2112     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2113     if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
2114     {
2115         IWineD3DSurfaceImpl *rt = device->render_targets[0];
2116         if(rt->resource.format->Flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
2117     }
2118
2119     if (shader->baseShader.reg_maps.shader_version.major == 1
2120             && shader->baseShader.reg_maps.shader_version.minor <= 3)
2121     {
2122         for (i = 0; i < 4; ++i)
2123         {
2124             DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
2125             DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
2126             if (flags & WINED3DTTFF_PROJECTED)
2127                 tex_transform |= WINED3D_PSARGS_PROJECTED;
2128             args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2129         }
2130     }
2131
2132     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2133     {
2134         if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2135         texture = state->textures[i];
2136         if (!texture)
2137         {
2138             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2139             continue;
2140         }
2141         args->color_fixup[i] = texture->resource.format->color_fixup;
2142
2143         if (texture->resource.format->Flags & WINED3DFMT_FLAG_SHADOW)
2144             args->shadow |= 1 << i;
2145
2146         /* Flag samplers that need NP2 texcoord fixup. */
2147         if (!texture->baseTexture.pow2Matrix_identity)
2148         {
2149             args->np2_fixup |= (1 << i);
2150         }
2151     }
2152     if (shader->baseShader.reg_maps.shader_version.major >= 3)
2153     {
2154         if (device->strided_streams.position_transformed)
2155         {
2156             args->vp_mode = pretransformed;
2157         }
2158         else if (use_vs(state))
2159         {
2160             args->vp_mode = vertexshader;
2161         }
2162         else
2163         {
2164             args->vp_mode = fixedfunction;
2165         }
2166         args->fog = FOG_OFF;
2167     }
2168     else
2169     {
2170         args->vp_mode = vertexshader;
2171         if (state->render_states[WINED3DRS_FOGENABLE])
2172         {
2173             switch (state->render_states[WINED3DRS_FOGTABLEMODE])
2174             {
2175                 case WINED3DFOG_NONE:
2176                     if (device->strided_streams.position_transformed || use_vs(state))
2177                     {
2178                         args->fog = FOG_LINEAR;
2179                         break;
2180                     }
2181
2182                     switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
2183                     {
2184                         case WINED3DFOG_NONE: /* Fall through. */
2185                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2186                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2187                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2188                     }
2189                     break;
2190
2191                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2192                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2193                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2194             }
2195         }
2196         else
2197         {
2198             args->fog = FOG_OFF;
2199         }
2200     }
2201 }
2202
2203 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2204 {
2205     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2206             shader->baseShader.reg_maps.shader_version.minor);
2207
2208     shader->baseShader.limits.attributes = 0;
2209     shader->baseShader.limits.address = 0;
2210     shader->baseShader.limits.packed_output = 0;
2211
2212     switch (shader_version)
2213     {
2214         case WINED3D_SHADER_VERSION(1, 0):
2215         case WINED3D_SHADER_VERSION(1, 1):
2216         case WINED3D_SHADER_VERSION(1, 2):
2217         case WINED3D_SHADER_VERSION(1, 3):
2218             shader->baseShader.limits.temporary = 2;
2219             shader->baseShader.limits.constant_float = 8;
2220             shader->baseShader.limits.constant_int = 0;
2221             shader->baseShader.limits.constant_bool = 0;
2222             shader->baseShader.limits.texcoord = 4;
2223             shader->baseShader.limits.sampler = 4;
2224             shader->baseShader.limits.packed_input = 0;
2225             shader->baseShader.limits.label = 0;
2226             break;
2227
2228         case WINED3D_SHADER_VERSION(1, 4):
2229             shader->baseShader.limits.temporary = 6;
2230             shader->baseShader.limits.constant_float = 8;
2231             shader->baseShader.limits.constant_int = 0;
2232             shader->baseShader.limits.constant_bool = 0;
2233             shader->baseShader.limits.texcoord = 6;
2234             shader->baseShader.limits.sampler = 6;
2235             shader->baseShader.limits.packed_input = 0;
2236             shader->baseShader.limits.label = 0;
2237             break;
2238
2239         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2240         case WINED3D_SHADER_VERSION(2, 0):
2241             shader->baseShader.limits.temporary = 32;
2242             shader->baseShader.limits.constant_float = 32;
2243             shader->baseShader.limits.constant_int = 16;
2244             shader->baseShader.limits.constant_bool = 16;
2245             shader->baseShader.limits.texcoord = 8;
2246             shader->baseShader.limits.sampler = 16;
2247             shader->baseShader.limits.packed_input = 0;
2248             break;
2249
2250         case WINED3D_SHADER_VERSION(2, 1):
2251             shader->baseShader.limits.temporary = 32;
2252             shader->baseShader.limits.constant_float = 32;
2253             shader->baseShader.limits.constant_int = 16;
2254             shader->baseShader.limits.constant_bool = 16;
2255             shader->baseShader.limits.texcoord = 8;
2256             shader->baseShader.limits.sampler = 16;
2257             shader->baseShader.limits.packed_input = 0;
2258             shader->baseShader.limits.label = 16;
2259             break;
2260
2261         case WINED3D_SHADER_VERSION(4, 0):
2262             FIXME("Using 3.0 limits for 4.0 shader.\n");
2263             /* Fall through. */
2264
2265         case WINED3D_SHADER_VERSION(3, 0):
2266             shader->baseShader.limits.temporary = 32;
2267             shader->baseShader.limits.constant_float = 224;
2268             shader->baseShader.limits.constant_int = 16;
2269             shader->baseShader.limits.constant_bool = 16;
2270             shader->baseShader.limits.texcoord = 0;
2271             shader->baseShader.limits.sampler = 16;
2272             shader->baseShader.limits.packed_input = 12;
2273             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2274             break;
2275
2276         default:
2277             shader->baseShader.limits.temporary = 32;
2278             shader->baseShader.limits.constant_float = 32;
2279             shader->baseShader.limits.constant_int = 16;
2280             shader->baseShader.limits.constant_bool = 16;
2281             shader->baseShader.limits.texcoord = 8;
2282             shader->baseShader.limits.sampler = 16;
2283             shader->baseShader.limits.packed_input = 0;
2284             shader->baseShader.limits.label = 0;
2285             FIXME("Unrecognized pixel shader version %u.%u\n",
2286                     shader->baseShader.reg_maps.shader_version.major,
2287                     shader->baseShader.reg_maps.shader_version.minor);
2288     }
2289 }
2290
2291 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2292         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2293         void *parent, const struct wined3d_parent_ops *parent_ops)
2294 {
2295     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2296     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2297     HRESULT hr;
2298
2299     if (!byte_code) return WINED3DERR_INVALIDCALL;
2300
2301     shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2302     shader_init(&shader->baseShader, device, parent, parent_ops);
2303
2304     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2305             output_signature, device->d3d_pshader_constantF);
2306     if (FAILED(hr))
2307     {
2308         WARN("Failed to set function, hr %#x.\n", hr);
2309         shader_cleanup((IWineD3DBaseShader *)shader);
2310         return hr;
2311     }
2312
2313     pixelshader_set_limits(shader);
2314
2315     for (i = 0; i < MAX_REG_INPUT; ++i)
2316     {
2317         if (shader->input_reg_used[i])
2318         {
2319             ++num_regs_used;
2320             highest_reg_used = i;
2321         }
2322     }
2323
2324     /* Don't do any register mapping magic if it is not needed, or if we can't
2325      * achieve anything anyway */
2326     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2327             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2328     {
2329         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2330         {
2331             /* This happens with relative addressing. The input mapper function
2332              * warns about this if the higher registers are declared too, so
2333              * don't write a FIXME here */
2334             WARN("More varying registers used than supported\n");
2335         }
2336
2337         for (i = 0; i < MAX_REG_INPUT; ++i)
2338         {
2339             shader->input_reg_map[i] = i;
2340         }
2341
2342         shader->declared_in_count = highest_reg_used + 1;
2343     }
2344     else
2345     {
2346         shader->declared_in_count = 0;
2347         for (i = 0; i < MAX_REG_INPUT; ++i)
2348         {
2349             if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2350             else shader->input_reg_map[i] = ~0U;
2351         }
2352     }
2353
2354     shader->baseShader.load_local_constsF = FALSE;
2355
2356     return WINED3D_OK;
2357 }
2358
2359 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2360 {
2361     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2362     unsigned int i;
2363
2364     if (reg_maps->shader_version.major != 1) return;
2365
2366     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2367     {
2368         /* We don't sample from this sampler. */
2369         if (!sampler_type[i]) continue;
2370
2371         if (!textures[i])
2372         {
2373             WARN("No texture bound to sampler %u, using 2D.\n", i);
2374             sampler_type[i] = WINED3DSTT_2D;
2375             continue;
2376         }
2377
2378         switch (((IWineD3DBaseTextureImpl *)textures[i])->baseTexture.target)
2379         {
2380             case GL_TEXTURE_RECTANGLE_ARB:
2381             case GL_TEXTURE_2D:
2382                 /* We have to select between texture rectangles and 2D
2383                  * textures later because 2.0 and 3.0 shaders only have
2384                  * WINED3DSTT_2D as well. */
2385                 sampler_type[i] = WINED3DSTT_2D;
2386                 break;
2387
2388             case GL_TEXTURE_3D:
2389                 sampler_type[i] = WINED3DSTT_VOLUME;
2390                 break;
2391
2392             case GL_TEXTURE_CUBE_MAP_ARB:
2393                 sampler_type[i] = WINED3DSTT_CUBE;
2394                 break;
2395
2396             default:
2397                 FIXME("Unrecognized texture type %#x, using 2D.\n",
2398                         ((IWineD3DBaseTextureImpl *)textures[i])->baseTexture.target);
2399                 sampler_type[i] = WINED3DSTT_2D;
2400         }
2401     }
2402 }