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_instruction ins;
1194     struct wined3d_shader_context ctx;
1195     const DWORD *ptr = byte_code;
1196     DWORD i;
1197
1198     /* Initialize current parsing state. */
1199     ctx.shader = iface;
1200     ctx.gl_info = &device->adapter->gl_info;
1201     ctx.reg_maps = reg_maps;
1202     ctx.buffer = buffer;
1203     ctx.backend_data = backend_ctx;
1204
1205     ins.ctx = &ctx;
1206     ins.dst = dst_param;
1207     ins.src = src_param;
1208     shader->baseShader.parse_state.current_row = 0;
1209
1210     fe->shader_read_header(fe_data, &ptr, &shader_version);
1211
1212     while (!fe->shader_is_end(fe_data, &ptr))
1213     {
1214         const char *comment;
1215         UINT comment_size;
1216         UINT param_size;
1217
1218         /* Skip comment tokens. */
1219         fe->shader_read_comment(&ptr, &comment, &comment_size);
1220         if (comment) continue;
1221
1222         /* Read opcode. */
1223         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1224
1225         /* Unknown opcode and its parameters. */
1226         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1227         {
1228             TRACE("Skipping unrecognized instruction.\n");
1229             ptr += param_size;
1230             continue;
1231         }
1232
1233         /* Nothing to do. */
1234         if (ins.handler_idx == WINED3DSIH_DCL
1235                 || ins.handler_idx == WINED3DSIH_NOP
1236                 || ins.handler_idx == WINED3DSIH_DEF
1237                 || ins.handler_idx == WINED3DSIH_DEFI
1238                 || ins.handler_idx == WINED3DSIH_DEFB
1239                 || ins.handler_idx == WINED3DSIH_PHASE)
1240         {
1241             ptr += param_size;
1242             continue;
1243         }
1244
1245         /* Destination tokens */
1246         for (i = 0; i < ins.dst_count; ++i)
1247         {
1248             fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1249         }
1250
1251         /* Predication token */
1252         if (ins.predicate)
1253         {
1254             FIXME("Predicates not implemented.\n");
1255             ins.predicate = *ptr++;
1256         }
1257
1258         /* Other source tokens */
1259         for (i = 0; i < ins.src_count; ++i)
1260         {
1261             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1262         }
1263
1264         /* Call appropriate function for output target */
1265         device->shader_backend->shader_handle_instruction(&ins);
1266     }
1267 }
1268
1269 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1270 {
1271     DWORD mmask = dst->modifiers;
1272
1273     switch (dst->shift)
1274     {
1275         case 0: break;
1276         case 13: TRACE("_d8"); break;
1277         case 14: TRACE("_d4"); break;
1278         case 15: TRACE("_d2"); break;
1279         case 1: TRACE("_x2"); break;
1280         case 2: TRACE("_x4"); break;
1281         case 3: TRACE("_x8"); break;
1282         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1283     }
1284
1285     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1286     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1287     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1288
1289     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1290     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1291 }
1292
1293 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1294 {
1295     struct wined3d_shader_version shader_version;
1296     const DWORD *ptr = byte_code;
1297     const char *type_prefix;
1298     DWORD i;
1299
1300     TRACE("Parsing %p.\n", byte_code);
1301
1302     fe->shader_read_header(fe_data, &ptr, &shader_version);
1303
1304     switch (shader_version.type)
1305     {
1306         case WINED3D_SHADER_TYPE_VERTEX:
1307             type_prefix = "vs";
1308             break;
1309
1310         case WINED3D_SHADER_TYPE_GEOMETRY:
1311             type_prefix = "gs";
1312             break;
1313
1314         case WINED3D_SHADER_TYPE_PIXEL:
1315             type_prefix = "ps";
1316             break;
1317
1318         default:
1319             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1320             type_prefix = "unknown";
1321             break;
1322     }
1323
1324     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1325
1326     while (!fe->shader_is_end(fe_data, &ptr))
1327     {
1328         struct wined3d_shader_instruction ins;
1329         const char *comment;
1330         UINT comment_size;
1331         UINT param_size;
1332
1333         /* comment */
1334         fe->shader_read_comment(&ptr, &comment, &comment_size);
1335         if (comment)
1336         {
1337             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1338             {
1339                 const char *end = comment + comment_size;
1340                 const char *ptr = comment + 4;
1341                 const char *line = ptr;
1342
1343                 TRACE("// TEXT\n");
1344                 while (ptr != end)
1345                 {
1346                     if (*ptr == '\n')
1347                     {
1348                         UINT len = ptr - line;
1349                         if (len && *(ptr - 1) == '\r') --len;
1350                         TRACE("// %s\n", debugstr_an(line, len));
1351                         line = ++ptr;
1352                     }
1353                     else ++ptr;
1354                 }
1355                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1356             }
1357             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1358             continue;
1359         }
1360
1361         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1362         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1363         {
1364             TRACE("Skipping unrecognized instruction.\n");
1365             ptr += param_size;
1366             continue;
1367         }
1368
1369         if (ins.handler_idx == WINED3DSIH_DCL)
1370         {
1371             struct wined3d_shader_semantic semantic;
1372
1373             fe->shader_read_semantic(&ptr, &semantic);
1374
1375             shader_dump_decl_usage(&semantic, &shader_version);
1376             shader_dump_ins_modifiers(&semantic.reg);
1377             TRACE(" ");
1378             shader_dump_dst_param(&semantic.reg, &shader_version);
1379         }
1380         else if (ins.handler_idx == WINED3DSIH_DEF)
1381         {
1382             struct wined3d_shader_src_param rel_addr;
1383             struct wined3d_shader_dst_param dst;
1384
1385             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1386
1387             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1388                     *(const float *)(ptr),
1389                     *(const float *)(ptr + 1),
1390                     *(const float *)(ptr + 2),
1391                     *(const float *)(ptr + 3));
1392             ptr += 4;
1393         }
1394         else if (ins.handler_idx == WINED3DSIH_DEFI)
1395         {
1396             struct wined3d_shader_src_param rel_addr;
1397             struct wined3d_shader_dst_param dst;
1398
1399             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1400
1401             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1402                     *(ptr),
1403                     *(ptr + 1),
1404                     *(ptr + 2),
1405                     *(ptr + 3));
1406             ptr += 4;
1407         }
1408         else if (ins.handler_idx == WINED3DSIH_DEFB)
1409         {
1410             struct wined3d_shader_src_param rel_addr;
1411             struct wined3d_shader_dst_param dst;
1412
1413             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1414
1415             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1416             ++ptr;
1417         }
1418         else
1419         {
1420             struct wined3d_shader_src_param dst_rel_addr[2];
1421             struct wined3d_shader_src_param src_rel_addr;
1422             struct wined3d_shader_dst_param dst_param[2];
1423             struct wined3d_shader_src_param src_param;
1424
1425             for (i = 0; i < ins.dst_count; ++i)
1426             {
1427                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1428             }
1429
1430             /* Print out predication source token first - it follows
1431              * the destination token. */
1432             if (ins.predicate)
1433             {
1434                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1435                 TRACE("(");
1436                 shader_dump_src_param(&src_param, &shader_version);
1437                 TRACE(") ");
1438             }
1439
1440             /* PixWin marks instructions with the coissue flag with a '+' */
1441             if (ins.coissue) TRACE("+");
1442
1443             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1444
1445             if (ins.handler_idx == WINED3DSIH_IFC
1446                     || ins.handler_idx == WINED3DSIH_BREAKC)
1447             {
1448                 switch (ins.flags)
1449                 {
1450                     case COMPARISON_GT: TRACE("_gt"); break;
1451                     case COMPARISON_EQ: TRACE("_eq"); break;
1452                     case COMPARISON_GE: TRACE("_ge"); break;
1453                     case COMPARISON_LT: TRACE("_lt"); break;
1454                     case COMPARISON_NE: TRACE("_ne"); break;
1455                     case COMPARISON_LE: TRACE("_le"); break;
1456                     default: TRACE("_(%u)", ins.flags);
1457                 }
1458             }
1459             else if (ins.handler_idx == WINED3DSIH_TEX
1460                     && shader_version.major >= 2
1461                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1462             {
1463                 TRACE("p");
1464             }
1465
1466             /* We already read the destination tokens, print them. */
1467             for (i = 0; i < ins.dst_count; ++i)
1468             {
1469                 shader_dump_ins_modifiers(&dst_param[i]);
1470                 TRACE(!i ? " " : ", ");
1471                 shader_dump_dst_param(&dst_param[i], &shader_version);
1472             }
1473
1474             /* Other source tokens */
1475             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1476             {
1477                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1478                 TRACE(!i ? " " : ", ");
1479                 shader_dump_src_param(&src_param, &shader_version);
1480             }
1481         }
1482         TRACE("\n");
1483     }
1484 }
1485
1486 static void shader_cleanup(IWineD3DBaseShader *iface)
1487 {
1488     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1489
1490     ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1491     HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1492     HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1493     shader_delete_constant_list(&shader->baseShader.constantsF);
1494     shader_delete_constant_list(&shader->baseShader.constantsB);
1495     shader_delete_constant_list(&shader->baseShader.constantsI);
1496     list_remove(&shader->baseShader.shader_list_entry);
1497
1498     if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1499     {
1500         shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1501     }
1502 }
1503
1504 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1505 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1506 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type, const SIZE *ds_mask_size) {}
1507 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1508 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1509 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1510 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1511 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1512 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1513 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1514 static void shader_none_free(IWineD3DDevice *iface) {}
1515 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1516
1517 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1518 {
1519     /* Set the shader caps to 0 for the none shader backend */
1520     caps->VertexShaderVersion = 0;
1521     caps->MaxVertexShaderConst = 0;
1522     caps->PixelShaderVersion = 0;
1523     caps->PixelShader1xMaxValue = 0.0f;
1524     caps->MaxPixelShaderConst = 0;
1525     caps->VSClipping = FALSE;
1526 }
1527
1528 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1529 {
1530     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1531     {
1532         TRACE("Checking support for fixup:\n");
1533         dump_color_fixup_desc(fixup);
1534     }
1535
1536     /* Faked to make some apps happy. */
1537     if (!is_complex_fixup(fixup))
1538     {
1539         TRACE("[OK]\n");
1540         return TRUE;
1541     }
1542
1543     TRACE("[FAILED]\n");
1544     return FALSE;
1545 }
1546
1547 const shader_backend_t none_shader_backend = {
1548     shader_none_handle_instruction,
1549     shader_none_select,
1550     shader_none_select_depth_blt,
1551     shader_none_deselect_depth_blt,
1552     shader_none_update_float_vertex_constants,
1553     shader_none_update_float_pixel_constants,
1554     shader_none_load_constants,
1555     shader_none_load_np2fixup_constants,
1556     shader_none_destroy,
1557     shader_none_alloc,
1558     shader_none_free,
1559     shader_none_dirty_const,
1560     shader_none_get_caps,
1561     shader_none_color_fixup_supported,
1562 };
1563
1564 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1565 {
1566     if (!data)
1567     {
1568         *data_size = shader->baseShader.functionLength;
1569         return WINED3D_OK;
1570     }
1571
1572     if (*data_size < shader->baseShader.functionLength)
1573     {
1574         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1575          * than the required size we should write the required size and
1576          * return D3DERR_MOREDATA. That's not actually true. */
1577         return WINED3DERR_INVALIDCALL;
1578     }
1579
1580     memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1581
1582     return WINED3D_OK;
1583 }
1584
1585 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1586         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1587 {
1588     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1589     const struct wined3d_shader_frontend *fe;
1590     HRESULT hr;
1591
1592     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1593             shader, byte_code, output_signature, float_const_count);
1594
1595     fe = shader_select_frontend(*byte_code);
1596     if (!fe)
1597     {
1598         FIXME("Unable to find frontend for shader.\n");
1599         return WINED3DERR_INVALIDCALL;
1600     }
1601     shader->baseShader.frontend = fe;
1602     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1603     if (!shader->baseShader.frontend_data)
1604     {
1605         FIXME("Failed to initialize frontend.\n");
1606         return WINED3DERR_INVALIDCALL;
1607     }
1608
1609     /* First pass: trace shader. */
1610     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1611
1612     /* Initialize immediate constant lists. */
1613     list_init(&shader->baseShader.constantsF);
1614     list_init(&shader->baseShader.constantsB);
1615     list_init(&shader->baseShader.constantsI);
1616
1617     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1618     hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1619             reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1620             byte_code, float_const_count);
1621     if (FAILED(hr)) return hr;
1622
1623     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1624     if (!shader->baseShader.function) return E_OUTOFMEMORY;
1625     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1626
1627     return WINED3D_OK;
1628 }
1629
1630 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1631 {
1632     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1633
1634     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1635             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1636             || IsEqualGUID(riid, &IID_IWineD3DBase)
1637             || IsEqualGUID(riid, &IID_IUnknown))
1638     {
1639         IUnknown_AddRef(iface);
1640         *object = iface;
1641         return S_OK;
1642     }
1643
1644     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1645
1646     *object = NULL;
1647     return E_NOINTERFACE;
1648 }
1649
1650 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1651 {
1652     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1653     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1654
1655     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1656
1657     return refcount;
1658 }
1659
1660 /* Do not call while under the GL lock. */
1661 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1662 {
1663     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1664     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1665
1666     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1667
1668     if (!refcount)
1669     {
1670         shader_cleanup((IWineD3DBaseShader *)iface);
1671         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1672         HeapFree(GetProcessHeap(), 0, shader);
1673     }
1674
1675     return refcount;
1676 }
1677
1678 static void * STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface)
1679 {
1680     TRACE("iface %p.\n", iface);
1681
1682     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1683 }
1684
1685 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1686 {
1687     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1688
1689     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1690 }
1691
1692 /* Set local constants for d3d8 shaders. */
1693 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1694         UINT start_idx, const float *src_data, UINT count)
1695 {
1696     IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1697     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1698     UINT i, end_idx;
1699
1700     TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1701
1702     end_idx = start_idx + count;
1703     if (end_idx > device->d3d_vshader_constantF)
1704     {
1705         WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1706         end_idx = device->d3d_vshader_constantF;
1707     }
1708
1709     for (i = start_idx; i < end_idx; ++i)
1710     {
1711         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1712         if (!lconst) return E_OUTOFMEMORY;
1713
1714         lconst->idx = i;
1715         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1716         list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1717     }
1718
1719     return WINED3D_OK;
1720 }
1721
1722 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1723 {
1724     /* IUnknown methods */
1725     vertexshader_QueryInterface,
1726     vertexshader_AddRef,
1727     vertexshader_Release,
1728     /* IWineD3DBase methods */
1729     vertexshader_GetParent,
1730     /* IWineD3DBaseShader methods */
1731     vertexshader_GetFunction,
1732     /* IWineD3DVertexShader methods */
1733     vertexshader_SetLocalConstantsF,
1734 };
1735
1736 void find_vs_compile_args(const struct wined3d_state *state,
1737         IWineD3DVertexShaderImpl *shader, struct vs_compile_args *args)
1738 {
1739     args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1740             == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1741     args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1742             && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1743     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1744 }
1745
1746 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1747 {
1748     if (usage_idx1 != usage_idx2) return FALSE;
1749     if (usage1 == usage2) return TRUE;
1750     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1751     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1752
1753     return FALSE;
1754 }
1755
1756 BOOL vshader_get_input(struct IWineD3DVertexShaderImpl *shader,
1757         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1758 {
1759     WORD map = shader->baseShader.reg_maps.input_registers;
1760     unsigned int i;
1761
1762     for (i = 0; map; map >>= 1, ++i)
1763     {
1764         if (!(map & 1)) continue;
1765
1766         if (match_usage(shader->attributes[i].usage,
1767                 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1768         {
1769             *regnum = i;
1770             return TRUE;
1771         }
1772     }
1773     return FALSE;
1774 }
1775
1776 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1777 {
1778     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1779             shader->baseShader.reg_maps.shader_version.minor);
1780     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1781
1782     shader->baseShader.limits.texcoord = 0;
1783     shader->baseShader.limits.attributes = 16;
1784     shader->baseShader.limits.packed_input = 0;
1785
1786     switch (shader_version)
1787     {
1788         case WINED3D_SHADER_VERSION(1, 0):
1789         case WINED3D_SHADER_VERSION(1, 1):
1790             shader->baseShader.limits.temporary = 12;
1791             shader->baseShader.limits.constant_bool = 0;
1792             shader->baseShader.limits.constant_int = 0;
1793             shader->baseShader.limits.address = 1;
1794             shader->baseShader.limits.packed_output = 12;
1795             shader->baseShader.limits.sampler = 0;
1796             shader->baseShader.limits.label = 0;
1797             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1798              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1799              * constants? */
1800             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1801             break;
1802
1803         case WINED3D_SHADER_VERSION(2, 0):
1804         case WINED3D_SHADER_VERSION(2, 1):
1805             shader->baseShader.limits.temporary = 12;
1806             shader->baseShader.limits.constant_bool = 16;
1807             shader->baseShader.limits.constant_int = 16;
1808             shader->baseShader.limits.address = 1;
1809             shader->baseShader.limits.packed_output = 12;
1810             shader->baseShader.limits.sampler = 0;
1811             shader->baseShader.limits.label = 16;
1812             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1813             break;
1814
1815         case WINED3D_SHADER_VERSION(4, 0):
1816             FIXME("Using 3.0 limits for 4.0 shader.\n");
1817             /* Fall through. */
1818
1819         case WINED3D_SHADER_VERSION(3, 0):
1820             shader->baseShader.limits.temporary = 32;
1821             shader->baseShader.limits.constant_bool = 32;
1822             shader->baseShader.limits.constant_int = 32;
1823             shader->baseShader.limits.address = 1;
1824             shader->baseShader.limits.packed_output = 12;
1825             shader->baseShader.limits.sampler = 4;
1826             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1827             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1828              * even though they are capable of supporting much more (GL
1829              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1830              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1831              * shaders to 256. */
1832             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1833             break;
1834
1835         default:
1836             shader->baseShader.limits.temporary = 12;
1837             shader->baseShader.limits.constant_bool = 16;
1838             shader->baseShader.limits.constant_int = 16;
1839             shader->baseShader.limits.address = 1;
1840             shader->baseShader.limits.packed_output = 12;
1841             shader->baseShader.limits.sampler = 0;
1842             shader->baseShader.limits.label = 16;
1843             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1844             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1845                     shader->baseShader.reg_maps.shader_version.major,
1846                     shader->baseShader.reg_maps.shader_version.minor);
1847     }
1848 }
1849
1850 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1851         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1852         void *parent, const struct wined3d_parent_ops *parent_ops)
1853 {
1854     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1855     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1856     unsigned int i;
1857     HRESULT hr;
1858     WORD map;
1859
1860     if (!byte_code) return WINED3DERR_INVALIDCALL;
1861
1862     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1863     shader_init(&shader->baseShader, device, parent, parent_ops);
1864
1865     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1866             output_signature, device->d3d_vshader_constantF);
1867     if (FAILED(hr))
1868     {
1869         WARN("Failed to set function, hr %#x.\n", hr);
1870         shader_cleanup((IWineD3DBaseShader *)shader);
1871         return hr;
1872     }
1873
1874     map = shader->baseShader.reg_maps.input_registers;
1875     for (i = 0; map; map >>= 1, ++i)
1876     {
1877         if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1878
1879         shader->attributes[i].usage =
1880                 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1881         shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1882     }
1883
1884     if (output_signature)
1885     {
1886         for (i = 0; i < output_signature->element_count; ++i)
1887         {
1888             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1889             reg_maps->output_registers |= 1 << e->register_idx;
1890             shader->baseShader.output_signature[e->register_idx] = *e;
1891         }
1892     }
1893
1894     vertexshader_set_limits(shader);
1895
1896     if (device->vs_selected_mode == SHADER_ARB
1897             && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1898             && shader->min_rel_offset <= shader->max_rel_offset)
1899     {
1900         if (shader->max_rel_offset - shader->min_rel_offset > 127)
1901         {
1902             FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1903             FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1904             FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1905         }
1906         else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1907         {
1908             shader->rel_offset = shader->min_rel_offset + 63;
1909         }
1910         else if (shader->max_rel_offset > 63)
1911         {
1912             shader->rel_offset = shader->min_rel_offset;
1913         }
1914         else
1915         {
1916             shader->rel_offset = 0;
1917         }
1918     }
1919
1920     shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1921             && !list_empty(&shader->baseShader.constantsF);
1922
1923     return WINED3D_OK;
1924 }
1925
1926 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1927         REFIID riid, void **object)
1928 {
1929     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1930
1931     if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1932             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1933             || IsEqualGUID(riid, &IID_IWineD3DBase)
1934             || IsEqualGUID(riid, &IID_IUnknown))
1935     {
1936         IUnknown_AddRef(iface);
1937         *object = iface;
1938         return S_OK;
1939     }
1940
1941     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1942
1943     *object = NULL;
1944     return E_NOINTERFACE;
1945 }
1946
1947 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1948 {
1949     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1950     ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1951
1952     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1953
1954     return refcount;
1955 }
1956
1957 /* Do not call while under the GL lock. */
1958 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1959 {
1960     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1961     ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1962
1963     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1964
1965     if (!refcount)
1966     {
1967         shader_cleanup((IWineD3DBaseShader *)iface);
1968         shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1969         HeapFree(GetProcessHeap(), 0, shader);
1970     }
1971
1972     return refcount;
1973 }
1974
1975 static void * STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface)
1976 {
1977     TRACE("iface %p.\n", iface);
1978
1979     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1980 }
1981
1982 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1983 {
1984     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1985
1986     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1987 }
1988
1989 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1990 {
1991     /* IUnknown methods */
1992     geometryshader_QueryInterface,
1993     geometryshader_AddRef,
1994     geometryshader_Release,
1995     /* IWineD3DBase methods */
1996     geometryshader_GetParent,
1997     /* IWineD3DBaseShader methods */
1998     geometryshader_GetFunction,
1999 };
2000
2001 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
2002         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2003         void *parent, const struct wined3d_parent_ops *parent_ops)
2004 {
2005     HRESULT hr;
2006
2007     shader->vtbl = &wined3d_geometryshader_vtbl;
2008     shader_init(&shader->base_shader, device, parent, parent_ops);
2009
2010     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
2011     if (FAILED(hr))
2012     {
2013         WARN("Failed to set function, hr %#x.\n", hr);
2014         shader_cleanup((IWineD3DBaseShader *)shader);
2015         return hr;
2016     }
2017
2018     shader->base_shader.load_local_constsF = FALSE;
2019
2020     return WINED3D_OK;
2021 }
2022
2023 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
2024 {
2025     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
2026
2027     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
2028             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
2029             || IsEqualGUID(riid, &IID_IWineD3DBase)
2030             || IsEqualGUID(riid, &IID_IUnknown))
2031     {
2032         IUnknown_AddRef(iface);
2033         *object = iface;
2034         return S_OK;
2035     }
2036
2037     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
2038
2039     *object = NULL;
2040     return E_NOINTERFACE;
2041 }
2042
2043 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
2044 {
2045     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2046     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
2047
2048     TRACE("%p increasing refcount to %u.\n", shader, refcount);
2049
2050     return refcount;
2051 }
2052
2053 /* Do not call while under the GL lock. */
2054 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
2055 {
2056     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2057     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
2058
2059     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2060
2061     if (!refcount)
2062     {
2063         shader_cleanup((IWineD3DBaseShader *)iface);
2064         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
2065         HeapFree(GetProcessHeap(), 0, shader);
2066     }
2067
2068     return refcount;
2069 }
2070
2071 static void * STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface)
2072 {
2073     TRACE("iface %p.\n", iface);
2074
2075     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
2076 }
2077
2078 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
2079 {
2080     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
2081
2082     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
2083 }
2084
2085 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2086 {
2087     /* IUnknown methods */
2088     pixelshader_QueryInterface,
2089     pixelshader_AddRef,
2090     pixelshader_Release,
2091     /* IWineD3DBase methods */
2092     pixelshader_GetParent,
2093     /* IWineD3DBaseShader methods */
2094     pixelshader_GetFunction
2095 };
2096
2097 void find_ps_compile_args(const struct wined3d_state *state,
2098         IWineD3DPixelShaderImpl *shader, struct ps_compile_args *args)
2099 {
2100     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
2101     IWineD3DBaseTextureImpl *texture;
2102     UINT i;
2103
2104     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2105     if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
2106     {
2107         IWineD3DSurfaceImpl *rt = device->render_targets[0];
2108         if(rt->resource.format->Flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
2109     }
2110
2111     if (shader->baseShader.reg_maps.shader_version.major == 1
2112             && shader->baseShader.reg_maps.shader_version.minor <= 3)
2113     {
2114         for (i = 0; i < 4; ++i)
2115         {
2116             DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
2117             DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
2118             if (flags & WINED3DTTFF_PROJECTED)
2119                 tex_transform |= WINED3D_PSARGS_PROJECTED;
2120             args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2121         }
2122     }
2123
2124     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2125     {
2126         if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2127         texture = state->textures[i];
2128         if (!texture)
2129         {
2130             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2131             continue;
2132         }
2133         args->color_fixup[i] = texture->resource.format->color_fixup;
2134
2135         if (texture->resource.format->Flags & WINED3DFMT_FLAG_SHADOW)
2136             args->shadow |= 1 << i;
2137
2138         /* Flag samplers that need NP2 texcoord fixup. */
2139         if (!texture->baseTexture.pow2Matrix_identity)
2140         {
2141             args->np2_fixup |= (1 << i);
2142         }
2143     }
2144     if (shader->baseShader.reg_maps.shader_version.major >= 3)
2145     {
2146         if (device->strided_streams.position_transformed)
2147         {
2148             args->vp_mode = pretransformed;
2149         }
2150         else if (use_vs(state))
2151         {
2152             args->vp_mode = vertexshader;
2153         }
2154         else
2155         {
2156             args->vp_mode = fixedfunction;
2157         }
2158         args->fog = FOG_OFF;
2159     }
2160     else
2161     {
2162         args->vp_mode = vertexshader;
2163         if (state->render_states[WINED3DRS_FOGENABLE])
2164         {
2165             switch (state->render_states[WINED3DRS_FOGTABLEMODE])
2166             {
2167                 case WINED3DFOG_NONE:
2168                     if (device->strided_streams.position_transformed || use_vs(state))
2169                     {
2170                         args->fog = FOG_LINEAR;
2171                         break;
2172                     }
2173
2174                     switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
2175                     {
2176                         case WINED3DFOG_NONE: /* Fall through. */
2177                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2178                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2179                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2180                     }
2181                     break;
2182
2183                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2184                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2185                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2186             }
2187         }
2188         else
2189         {
2190             args->fog = FOG_OFF;
2191         }
2192     }
2193 }
2194
2195 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2196 {
2197     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2198             shader->baseShader.reg_maps.shader_version.minor);
2199
2200     shader->baseShader.limits.attributes = 0;
2201     shader->baseShader.limits.address = 0;
2202     shader->baseShader.limits.packed_output = 0;
2203
2204     switch (shader_version)
2205     {
2206         case WINED3D_SHADER_VERSION(1, 0):
2207         case WINED3D_SHADER_VERSION(1, 1):
2208         case WINED3D_SHADER_VERSION(1, 2):
2209         case WINED3D_SHADER_VERSION(1, 3):
2210             shader->baseShader.limits.temporary = 2;
2211             shader->baseShader.limits.constant_float = 8;
2212             shader->baseShader.limits.constant_int = 0;
2213             shader->baseShader.limits.constant_bool = 0;
2214             shader->baseShader.limits.texcoord = 4;
2215             shader->baseShader.limits.sampler = 4;
2216             shader->baseShader.limits.packed_input = 0;
2217             shader->baseShader.limits.label = 0;
2218             break;
2219
2220         case WINED3D_SHADER_VERSION(1, 4):
2221             shader->baseShader.limits.temporary = 6;
2222             shader->baseShader.limits.constant_float = 8;
2223             shader->baseShader.limits.constant_int = 0;
2224             shader->baseShader.limits.constant_bool = 0;
2225             shader->baseShader.limits.texcoord = 6;
2226             shader->baseShader.limits.sampler = 6;
2227             shader->baseShader.limits.packed_input = 0;
2228             shader->baseShader.limits.label = 0;
2229             break;
2230
2231         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2232         case WINED3D_SHADER_VERSION(2, 0):
2233             shader->baseShader.limits.temporary = 32;
2234             shader->baseShader.limits.constant_float = 32;
2235             shader->baseShader.limits.constant_int = 16;
2236             shader->baseShader.limits.constant_bool = 16;
2237             shader->baseShader.limits.texcoord = 8;
2238             shader->baseShader.limits.sampler = 16;
2239             shader->baseShader.limits.packed_input = 0;
2240             break;
2241
2242         case WINED3D_SHADER_VERSION(2, 1):
2243             shader->baseShader.limits.temporary = 32;
2244             shader->baseShader.limits.constant_float = 32;
2245             shader->baseShader.limits.constant_int = 16;
2246             shader->baseShader.limits.constant_bool = 16;
2247             shader->baseShader.limits.texcoord = 8;
2248             shader->baseShader.limits.sampler = 16;
2249             shader->baseShader.limits.packed_input = 0;
2250             shader->baseShader.limits.label = 16;
2251             break;
2252
2253         case WINED3D_SHADER_VERSION(4, 0):
2254             FIXME("Using 3.0 limits for 4.0 shader.\n");
2255             /* Fall through. */
2256
2257         case WINED3D_SHADER_VERSION(3, 0):
2258             shader->baseShader.limits.temporary = 32;
2259             shader->baseShader.limits.constant_float = 224;
2260             shader->baseShader.limits.constant_int = 16;
2261             shader->baseShader.limits.constant_bool = 16;
2262             shader->baseShader.limits.texcoord = 0;
2263             shader->baseShader.limits.sampler = 16;
2264             shader->baseShader.limits.packed_input = 12;
2265             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2266             break;
2267
2268         default:
2269             shader->baseShader.limits.temporary = 32;
2270             shader->baseShader.limits.constant_float = 32;
2271             shader->baseShader.limits.constant_int = 16;
2272             shader->baseShader.limits.constant_bool = 16;
2273             shader->baseShader.limits.texcoord = 8;
2274             shader->baseShader.limits.sampler = 16;
2275             shader->baseShader.limits.packed_input = 0;
2276             shader->baseShader.limits.label = 0;
2277             FIXME("Unrecognized pixel shader version %u.%u\n",
2278                     shader->baseShader.reg_maps.shader_version.major,
2279                     shader->baseShader.reg_maps.shader_version.minor);
2280     }
2281 }
2282
2283 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2284         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2285         void *parent, const struct wined3d_parent_ops *parent_ops)
2286 {
2287     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2288     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2289     HRESULT hr;
2290
2291     if (!byte_code) return WINED3DERR_INVALIDCALL;
2292
2293     shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2294     shader_init(&shader->baseShader, device, parent, parent_ops);
2295
2296     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2297             output_signature, device->d3d_pshader_constantF);
2298     if (FAILED(hr))
2299     {
2300         WARN("Failed to set function, hr %#x.\n", hr);
2301         shader_cleanup((IWineD3DBaseShader *)shader);
2302         return hr;
2303     }
2304
2305     pixelshader_set_limits(shader);
2306
2307     for (i = 0; i < MAX_REG_INPUT; ++i)
2308     {
2309         if (shader->input_reg_used[i])
2310         {
2311             ++num_regs_used;
2312             highest_reg_used = i;
2313         }
2314     }
2315
2316     /* Don't do any register mapping magic if it is not needed, or if we can't
2317      * achieve anything anyway */
2318     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2319             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2320     {
2321         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2322         {
2323             /* This happens with relative addressing. The input mapper function
2324              * warns about this if the higher registers are declared too, so
2325              * don't write a FIXME here */
2326             WARN("More varying registers used than supported\n");
2327         }
2328
2329         for (i = 0; i < MAX_REG_INPUT; ++i)
2330         {
2331             shader->input_reg_map[i] = i;
2332         }
2333
2334         shader->declared_in_count = highest_reg_used + 1;
2335     }
2336     else
2337     {
2338         shader->declared_in_count = 0;
2339         for (i = 0; i < MAX_REG_INPUT; ++i)
2340         {
2341             if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2342             else shader->input_reg_map[i] = ~0U;
2343         }
2344     }
2345
2346     shader->baseShader.load_local_constsF = FALSE;
2347
2348     return WINED3D_OK;
2349 }
2350
2351 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2352 {
2353     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2354     unsigned int i;
2355
2356     if (reg_maps->shader_version.major != 1) return;
2357
2358     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2359     {
2360         /* We don't sample from this sampler. */
2361         if (!sampler_type[i]) continue;
2362
2363         if (!textures[i])
2364         {
2365             WARN("No texture bound to sampler %u, using 2D.\n", i);
2366             sampler_type[i] = WINED3DSTT_2D;
2367             continue;
2368         }
2369
2370         switch (((IWineD3DBaseTextureImpl *)textures[i])->baseTexture.target)
2371         {
2372             case GL_TEXTURE_RECTANGLE_ARB:
2373             case GL_TEXTURE_2D:
2374                 /* We have to select between texture rectangles and 2D
2375                  * textures later because 2.0 and 3.0 shaders only have
2376                  * WINED3DSTT_2D as well. */
2377                 sampler_type[i] = WINED3DSTT_2D;
2378                 break;
2379
2380             case GL_TEXTURE_3D:
2381                 sampler_type[i] = WINED3DSTT_VOLUME;
2382                 break;
2383
2384             case GL_TEXTURE_CUBE_MAP_ARB:
2385                 sampler_type[i] = WINED3DSTT_CUBE;
2386                 break;
2387
2388             default:
2389                 FIXME("Unrecognized texture type %#x, using 2D.\n",
2390                         ((IWineD3DBaseTextureImpl *)textures[i])->baseTexture.target);
2391                 sampler_type[i] = WINED3DSTT_2D;
2392         }
2393     }
2394 }