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 2009-2011 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 #include "wine/port.h"
  27
  28 #include <math.h>
  29 #include <stdio.h>
  30 #include <string.h>
  31
  32 #include "wined3d_private.h"
  33
  34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  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_DCL_CONSTANT_BUFFER   */ "dcl_constantBuffer",
  53     /* WINED3DSIH_DCL_INPUT_PRIMITIVE   */ "dcl_inputPrimitive",
  54     /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY   */ "dcl_outputTopology",
  55     /* WINED3DSIH_DCL_VERTICES_OUT      */ "dcl_maxOutputVertexCount",
  56     /* WINED3DSIH_DEF                   */ "def",
  57     /* WINED3DSIH_DEFB                  */ "defb",
  58     /* WINED3DSIH_DEFI                  */ "defi",
  59     /* WINED3DSIH_DIV                   */ "div",
  60     /* WINED3DSIH_DP2ADD                */ "dp2add",
  61     /* WINED3DSIH_DP3                   */ "dp3",
  62     /* WINED3DSIH_DP4                   */ "dp4",
  63     /* WINED3DSIH_DST                   */ "dst",
  64     /* WINED3DSIH_DSX                   */ "dsx",
  65     /* WINED3DSIH_DSY                   */ "dsy",
  66     /* WINED3DSIH_ELSE                  */ "else",
  67     /* WINED3DSIH_EMIT                  */ "emit",
  68     /* WINED3DSIH_ENDIF                 */ "endif",
  69     /* WINED3DSIH_ENDLOOP               */ "endloop",
  70     /* WINED3DSIH_ENDREP                */ "endrep",
  71     /* WINED3DSIH_EQ                    */ "eq",
  72     /* WINED3DSIH_EXP                   */ "exp",
  73     /* WINED3DSIH_EXPP                  */ "expp",
  74     /* WINED3DSIH_FRC                   */ "frc",
  75     /* WINED3DSIH_FTOI                  */ "ftoi",
  76     /* WINED3DSIH_GE                    */ "ge",
  77     /* WINED3DSIH_IADD                  */ "iadd",
  78     /* WINED3DSIH_IEQ                   */ "ieq",
  79     /* WINED3DSIH_IF                    */ "if",
  80     /* WINED3DSIH_IFC                   */ "ifc",
  81     /* WINED3DSIH_IGE                   */ "ige",
  82     /* WINED3DSIH_IMUL                  */ "imul",
  83     /* WINED3DSIH_ITOF                  */ "itof",
  84     /* WINED3DSIH_LABEL                 */ "label",
  85     /* WINED3DSIH_LD                    */ "ld",
  86     /* WINED3DSIH_LIT                   */ "lit",
  87     /* WINED3DSIH_LOG                   */ "log",
  88     /* WINED3DSIH_LOGP                  */ "logp",
  89     /* WINED3DSIH_LOOP                  */ "loop",
  90     /* WINED3DSIH_LRP                   */ "lrp",
  91     /* WINED3DSIH_LT                    */ "lt",
  92     /* WINED3DSIH_M3x2                  */ "m3x2",
  93     /* WINED3DSIH_M3x3                  */ "m3x3",
  94     /* WINED3DSIH_M3x4                  */ "m3x4",
  95     /* WINED3DSIH_M4x3                  */ "m4x3",
  96     /* WINED3DSIH_M4x4                  */ "m4x4",
  97     /* WINED3DSIH_MAD                   */ "mad",
  98     /* WINED3DSIH_MAX                   */ "max",
  99     /* WINED3DSIH_MIN                   */ "min",
 100     /* WINED3DSIH_MOV                   */ "mov",
 101     /* WINED3DSIH_MOVA                  */ "mova",
 102     /* WINED3DSIH_MOVC                  */ "movc",
 103     /* WINED3DSIH_MUL                   */ "mul",
 104     /* WINED3DSIH_NOP                   */ "nop",
 105     /* WINED3DSIH_NRM                   */ "nrm",
 106     /* WINED3DSIH_PHASE                 */ "phase",
 107     /* WINED3DSIH_POW                   */ "pow",
 108     /* WINED3DSIH_RCP                   */ "rcp",
 109     /* WINED3DSIH_REP                   */ "rep",
 110     /* WINED3DSIH_RET                   */ "ret",
 111     /* WINED3DSIH_ROUND_NI              */ "round_ni",
 112     /* WINED3DSIH_RSQ                   */ "rsq",
 113     /* WINED3DSIH_SAMPLE                */ "sample",
 114     /* WINED3DSIH_SAMPLE_GRAD           */ "sample_d",
 115     /* WINED3DSIH_SAMPLE_LOD            */ "sample_l",
 116     /* WINED3DSIH_SETP                  */ "setp",
 117     /* WINED3DSIH_SGE                   */ "sge",
 118     /* WINED3DSIH_SGN                   */ "sgn",
 119     /* WINED3DSIH_SINCOS                */ "sincos",
 120     /* WINED3DSIH_SLT                   */ "slt",
 121     /* WINED3DSIH_SQRT                  */ "sqrt",
 122     /* WINED3DSIH_SUB                   */ "sub",
 123     /* WINED3DSIH_TEX                   */ "texld",
 124     /* WINED3DSIH_TEXBEM                */ "texbem",
 125     /* WINED3DSIH_TEXBEML               */ "texbeml",
 126     /* WINED3DSIH_TEXCOORD              */ "texcrd",
 127     /* WINED3DSIH_TEXDEPTH              */ "texdepth",
 128     /* WINED3DSIH_TEXDP3                */ "texdp3",
 129     /* WINED3DSIH_TEXDP3TEX             */ "texdp3tex",
 130     /* WINED3DSIH_TEXKILL               */ "texkill",
 131     /* WINED3DSIH_TEXLDD                */ "texldd",
 132     /* WINED3DSIH_TEXLDL                */ "texldl",
 133     /* WINED3DSIH_TEXM3x2DEPTH          */ "texm3x2depth",
 134     /* WINED3DSIH_TEXM3x2PAD            */ "texm3x2pad",
 135     /* WINED3DSIH_TEXM3x2TEX            */ "texm3x2tex",
 136     /* WINED3DSIH_TEXM3x3               */ "texm3x3",
 137     /* WINED3DSIH_TEXM3x3DIFF           */ "texm3x3diff",
 138     /* WINED3DSIH_TEXM3x3PAD            */ "texm3x3pad",
 139     /* WINED3DSIH_TEXM3x3SPEC           */ "texm3x3spec",
 140     /* WINED3DSIH_TEXM3x3TEX            */ "texm3x3tex",
 141     /* WINED3DSIH_TEXM3x3VSPEC          */ "texm3x3vspec",
 142     /* WINED3DSIH_TEXREG2AR             */ "texreg2ar",
 143     /* WINED3DSIH_TEXREG2GB             */ "texreg2gb",
 144     /* WINED3DSIH_TEXREG2RGB            */ "texreg2rgb",
 145     /* WINED3DSIH_UDIV                  */ "udiv",
 146     /* WINED3DSIH_USHR                  */ "ushr",
 147     /* WINED3DSIH_UTOF                  */ "utof",
 148     /* WINED3DSIH_XOR                   */ "xor",
 149 };
 150
 151 static const char * const semantic_names[] =
 152 {
 153     /* WINED3D_DECL_USAGE_POSITION      */ "SV_POSITION",
 154     /* WINED3D_DECL_USAGE_BLEND_WEIGHT  */ "BLENDWEIGHT",
 155     /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
 156     /* WINED3D_DECL_USAGE_NORMAL        */ "NORMAL",
 157     /* WINED3D_DECL_USAGE_PSIZE         */ "PSIZE",
 158     /* WINED3D_DECL_USAGE_TEXCOORD      */ "TEXCOORD",
 159     /* WINED3D_DECL_USAGE_TANGENT       */ "TANGENT",
 160     /* WINED3D_DECL_USAGE_BINORMAL      */ "BINORMAL",
 161     /* WINED3D_DECL_USAGE_TESS_FACTOR   */ "TESSFACTOR",
 162     /* WINED3D_DECL_USAGE_POSITIONT     */ "POSITIONT",
 163     /* WINED3D_DECL_USAGE_COLOR         */ "COLOR",
 164     /* WINED3D_DECL_USAGE_FOG           */ "FOG",
 165     /* WINED3D_DECL_USAGE_DEPTH         */ "DEPTH",
 166     /* WINED3D_DECL_USAGE_SAMPLE        */ "SAMPLE",
 167 };
 168
 169 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
 170 {
 171     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 172     {
 173         FIXME("Unrecognized usage %#x.\n", usage);
 174         return "UNRECOGNIZED";
 175     }
 176
 177     return semantic_names[usage];
 178 }
 179
 180 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
 181 {
 182     unsigned int i;
 183
 184     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 185     {
 186         if (!strcmp(name, semantic_names[i])) return i;
 187     }
 188
 189     return ~0U;
 190 }
 191
 192 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
 193 {
 194     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 195 }
 196
 197 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 198         const struct wined3d_shader_semantic *s)
 199 {
 200     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 201     e->semantic_idx = s->usage_idx;
 202     e->sysval_semantic = 0;
 203     e->component_type = 0;
 204     e->register_idx = s->reg.reg.idx[0].offset;
 205     e->mask = s->reg.write_mask;
 206 }
 207
 208 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
 209         enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
 210 {
 211     e->semantic_name = shader_semantic_name_from_usage(usage);
 212     e->semantic_idx = usage_idx;
 213     e->sysval_semantic = 0;
 214     e->component_type = 0;
 215     e->register_idx = reg_idx;
 216     e->mask = write_mask;
 217 }
 218
 219 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 220 {
 221     switch (version_token >> 16)
 222     {
 223         case WINED3D_SM1_VS:
 224         case WINED3D_SM1_PS:
 225             return &sm1_shader_frontend;
 226
 227         case WINED3D_SM4_PS:
 228         case WINED3D_SM4_VS:
 229         case WINED3D_SM4_GS:
 230             return &sm4_shader_frontend;
 231
 232         default:
 233             FIXME("Unrecognised version token %#x\n", version_token);
 234             return NULL;
 235     }
 236 }
 237
 238 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 239 {
 240     buffer->buffer[0] = '\0';
 241     buffer->content_size = 0;
 242     buffer->lineNo = 0;
 243     buffer->newline = TRUE;
 244 }
 245
 246 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 247 {
 248     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 249     if (!buffer->buffer)
 250     {
 251         ERR("Failed to allocate shader buffer memory.\n");
 252         return FALSE;
 253     }
 254     buffer->buffer_size = SHADER_PGMSIZE;
 255
 256     shader_buffer_clear(buffer);
 257     return TRUE;
 258 }
 259
 260 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 261 {
 262     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 263 }
 264
 265 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 266 {
 267     char *base = buffer->buffer + buffer->content_size;
 268     int rc;
 269     char *new_buffer;
 270
 271     while(1)
 272     {
 273         rc = vsnprintf(base, buffer->buffer_size - buffer->content_size, format, args);
 274         if (rc < 0 /* C89 */ || (unsigned int)rc >= buffer->buffer_size - buffer->content_size /* C99 */)
 275         {
 276             new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, buffer->buffer_size * 2);
 277             if (!new_buffer)
 278             {
 279                 ERR("The buffer allocated for the shader program string is too small at %d bytes.\n", buffer->buffer_size);
 280                 buffer->content_size = buffer->buffer_size - 1;
 281                 return -1;
 282             }
 283             buffer->buffer = new_buffer;
 284             buffer->buffer_size = buffer->buffer_size * 2;
 285             base = buffer->buffer + buffer->content_size;
 286         }
 287         else
 288         {
 289             break;
 290         }
 291     }
 292
 293     if (buffer->newline)
 294     {
 295         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->content_size, base);
 296         buffer->newline = FALSE;
 297     }
 298     else
 299     {
 300         TRACE("%s", base);
 301     }
 302
 303     buffer->content_size += rc;
 304     if (buffer->buffer[buffer->content_size-1] == '\n')
 305     {
 306         ++buffer->lineNo;
 307         buffer->newline = TRUE;
 308     }
 309
 310     return 0;
 311 }
 312
 313 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 314 {
 315     va_list args;
 316     int ret;
 317
 318     va_start(args, format);
 319     ret = shader_vaddline(buffer, format, args);
 320     va_end(args);
 321
 322     return ret;
 323 }
 324
 325 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
 326         void *parent, const struct wined3d_parent_ops *parent_ops)
 327 {
 328     shader->ref = 1;
 329     shader->device = device;
 330     shader->parent = parent;
 331     shader->parent_ops = parent_ops;
 332     list_init(&shader->linked_programs);
 333     list_add_head(&device->shaders, &shader->shader_list_entry);
 334 }
 335
 336 /* Convert floating point offset relative to a register file to an absolute
 337  * offset for float constants. */
 338 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
 339 {
 340     switch (register_type)
 341     {
 342         case WINED3DSPR_CONST: return register_idx;
 343         case WINED3DSPR_CONST2: return 2048 + register_idx;
 344         case WINED3DSPR_CONST3: return 4096 + register_idx;
 345         case WINED3DSPR_CONST4: return 6144 + register_idx;
 346         default:
 347             FIXME("Unsupported register type: %u.\n", register_type);
 348             return register_idx;
 349     }
 350 }
 351
 352 static void shader_delete_constant_list(struct list *clist)
 353 {
 354     struct wined3d_shader_lconst *constant;
 355     struct list *ptr;
 356
 357     ptr = list_head(clist);
 358     while (ptr)
 359     {
 360         constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
 361         ptr = list_next(clist, ptr);
 362         HeapFree(GetProcessHeap(), 0, constant);
 363     }
 364     list_init(clist);
 365 }
 366
 367 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 368 {
 369     DWORD idx, shift;
 370     idx = bit >> 5;
 371     shift = bit & 0x1f;
 372     bitmap[idx] |= (1 << shift);
 373 }
 374
 375 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
 376         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 377 {
 378     switch (reg->type)
 379     {
 380         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 381             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 382                 reg_maps->texcoord |= 1 << reg->idx[0].offset;
 383             else
 384                 reg_maps->address |= 1 << reg->idx[0].offset;
 385             break;
 386
 387         case WINED3DSPR_TEMP:
 388             reg_maps->temporary |= 1 << reg->idx[0].offset;
 389             break;
 390
 391         case WINED3DSPR_INPUT:
 392             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 393             {
 394                 if (reg->idx[0].rel_addr)
 395                 {
 396                     /* If relative addressing is used, we must assume that all registers
 397                      * are used. Even if it is a construct like v3[aL], we can't assume
 398                      * that v0, v1 and v2 aren't read because aL can be negative */
 399                     unsigned int i;
 400                     for (i = 0; i < MAX_REG_INPUT; ++i)
 401                     {
 402                         shader->u.ps.input_reg_used[i] = TRUE;
 403                     }
 404                 }
 405                 else
 406                 {
 407                     shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
 408                 }
 409             }
 410             else
 411                 reg_maps->input_registers |= 1 << reg->idx[0].offset;
 412             break;
 413
 414         case WINED3DSPR_RASTOUT:
 415             if (reg->idx[0].offset == 1)
 416                 reg_maps->fog = 1;
 417             break;
 418
 419         case WINED3DSPR_MISCTYPE:
 420             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 421             {
 422                 if (!reg->idx[0].offset)
 423                     reg_maps->vpos = 1;
 424                 else if (reg->idx[0].offset == 1)
 425                     reg_maps->usesfacing = 1;
 426             }
 427             break;
 428
 429         case WINED3DSPR_CONST:
 430             if (reg->idx[0].rel_addr)
 431             {
 432                 if (reg->idx[0].offset < reg_maps->min_rel_offset)
 433                     reg_maps->min_rel_offset = reg->idx[0].offset;
 434                 if (reg->idx[0].offset > reg_maps->max_rel_offset)
 435                     reg_maps->max_rel_offset = reg->idx[0].offset;
 436                 reg_maps->usesrelconstF = TRUE;
 437             }
 438             else
 439             {
 440                 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
 441             }
 442             break;
 443
 444         case WINED3DSPR_CONSTINT:
 445             reg_maps->integer_constants |= (1 << reg->idx[0].offset);
 446             break;
 447
 448         case WINED3DSPR_CONSTBOOL:
 449             reg_maps->boolean_constants |= (1 << reg->idx[0].offset);
 450             break;
 451
 452         case WINED3DSPR_COLOROUT:
 453             reg_maps->rt_mask |= (1 << reg->idx[0].offset);
 454             break;
 455
 456         default:
 457             TRACE("Not recording register of type %#x and [%#x][%#x].\n",
 458                     reg->type, reg->idx[0].offset, reg->idx[1].offset);
 459             break;
 460     }
 461 }
 462
 463 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 464 {
 465     switch (instr)
 466     {
 467         case WINED3DSIH_M4x4:
 468         case WINED3DSIH_M3x4:
 469             return param == 1 ? 3 : 0;
 470
 471         case WINED3DSIH_M4x3:
 472         case WINED3DSIH_M3x3:
 473             return param == 1 ? 2 : 0;
 474
 475         case WINED3DSIH_M3x2:
 476             return param == 1 ? 1 : 0;
 477
 478         default:
 479             return 0;
 480     }
 481 }
 482
 483 /* Note that this does not count the loop register as an address register. */
 484 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
 485         struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 486         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 487 {
 488     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 489     void *fe_data = shader->frontend_data;
 490     struct wined3d_shader_version shader_version;
 491     const DWORD *ptr = byte_code;
 492
 493     memset(reg_maps, 0, sizeof(*reg_maps));
 494     reg_maps->min_rel_offset = ~0U;
 495
 496     fe->shader_read_header(fe_data, &ptr, &shader_version);
 497     reg_maps->shader_version = shader_version;
 498
 499     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 500             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 501     if (!reg_maps->constf)
 502     {
 503         ERR("Failed to allocate constant map memory.\n");
 504         return E_OUTOFMEMORY;
 505     }
 506
 507     while (!fe->shader_is_end(fe_data, &ptr))
 508     {
 509         struct wined3d_shader_instruction ins;
 510
 511         /* Fetch opcode. */
 512         fe->shader_read_instruction(fe_data, &ptr, &ins);
 513
 514         /* Unhandled opcode, and its parameters. */
 515         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 516         {
 517             TRACE("Skipping unrecognized instruction.\n");
 518             continue;
 519         }
 520
 521         /* Handle declarations. */
 522         if (ins.handler_idx == WINED3DSIH_DCL)
 523         {
 524             struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
 525
 526             switch (semantic->reg.reg.type)
 527             {
 528                 /* Mark input registers used. */
 529                 case WINED3DSPR_INPUT:
 530                     reg_maps->input_registers |= 1 << semantic->reg.reg.idx[0].offset;
 531                     shader_signature_from_semantic(&input_signature[semantic->reg.reg.idx[0].offset], semantic);
 532                     break;
 533
 534                 /* Vertex shader: mark 3.0 output registers used, save token. */
 535                 case WINED3DSPR_OUTPUT:
 536                     reg_maps->output_registers |= 1 << semantic->reg.reg.idx[0].offset;
 537                     shader_signature_from_semantic(&output_signature[semantic->reg.reg.idx[0].offset], semantic);
 538                     if (semantic->usage == WINED3D_DECL_USAGE_FOG)
 539                         reg_maps->fog = 1;
 540                     break;
 541
 542                 /* Save sampler usage token. */
 543                 case WINED3DSPR_SAMPLER:
 544                     reg_maps->sampler_type[semantic->reg.reg.idx[0].offset] = semantic->sampler_type;
 545                     break;
 546
 547                 default:
 548                     TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
 549                     break;
 550             }
 551         }
 552         else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
 553         {
 554             struct wined3d_shader_register *reg = &ins.declaration.src.reg;
 555             if (reg->idx[0].offset >= WINED3D_MAX_CBS)
 556                 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
 557             else
 558                 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
 559         }
 560         else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
 561         {
 562             if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
 563                 shader->u.gs.input_type = ins.declaration.primitive_type;
 564             else
 565                 FIXME("Invalid instruction %#x for shader type %#x.\n",
 566                         ins.handler_idx, shader_version.type);
 567         }
 568         else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
 569         {
 570             if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
 571                 shader->u.gs.output_type = ins.declaration.primitive_type;
 572             else
 573                 FIXME("Invalid instruction %#x for shader type %#x.\n",
 574                         ins.handler_idx, shader_version.type);
 575         }
 576         else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
 577         {
 578             if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
 579                 shader->u.gs.vertices_out = ins.declaration.count;
 580             else
 581                 FIXME("Invalid instruction %#x for shader type %#x.\n",
 582                         ins.handler_idx, shader_version.type);
 583         }
 584         else if (ins.handler_idx == WINED3DSIH_DEF)
 585         {
 586             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 587             float *value;
 588             if (!lconst) return E_OUTOFMEMORY;
 589
 590             lconst->idx = ins.dst[0].reg.idx[0].offset;
 591             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 592             value = (float *)lconst->value;
 593
 594             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 595             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 596             {
 597                 if (value[0] < -1.0f) value[0] = -1.0f;
 598                 else if (value[0] > 1.0f) value[0] = 1.0f;
 599                 if (value[1] < -1.0f) value[1] = -1.0f;
 600                 else if (value[1] > 1.0f) value[1] = 1.0f;
 601                 if (value[2] < -1.0f) value[2] = -1.0f;
 602                 else if (value[2] > 1.0f) value[2] = 1.0f;
 603                 if (value[3] < -1.0f) value[3] = -1.0f;
 604                 else if (value[3] > 1.0f) value[3] = 1.0f;
 605             }
 606
 607             list_add_head(&shader->constantsF, &lconst->entry);
 608
 609             if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
 610                     || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
 611             {
 612                 shader->lconst_inf_or_nan = TRUE;
 613             }
 614         }
 615         else if (ins.handler_idx == WINED3DSIH_DEFI)
 616         {
 617             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 618             if (!lconst) return E_OUTOFMEMORY;
 619
 620             lconst->idx = ins.dst[0].reg.idx[0].offset;
 621             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 622
 623             list_add_head(&shader->constantsI, &lconst->entry);
 624             reg_maps->local_int_consts |= (1 << lconst->idx);
 625         }
 626         else if (ins.handler_idx == WINED3DSIH_DEFB)
 627         {
 628             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 629             if (!lconst) return E_OUTOFMEMORY;
 630
 631             lconst->idx = ins.dst[0].reg.idx[0].offset;
 632             memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
 633
 634             list_add_head(&shader->constantsB, &lconst->entry);
 635             reg_maps->local_bool_consts |= (1 << lconst->idx);
 636         }
 637         /* For subroutine prototypes. */
 638         else if (ins.handler_idx == WINED3DSIH_LABEL)
 639         {
 640             reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
 641         }
 642         /* Set texture, address, temporary registers. */
 643         else
 644         {
 645             BOOL color0_mov = FALSE;
 646             unsigned int i;
 647
 648             /* This will loop over all the registers and try to
 649              * make a bitmask of the ones we're interested in.
 650              *
 651              * Relative addressing tokens are ignored, but that's
 652              * okay, since we'll catch any address registers when
 653              * they are initialized (required by spec). */
 654             for (i = 0; i < ins.dst_count; ++i)
 655             {
 656                 shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg, shader_version.type);
 657
 658                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 659                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 660                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 661                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 662                 {
 663                     UINT idx = ins.dst[i].reg.idx[0].offset;
 664
 665                     switch (ins.dst[i].reg.type)
 666                     {
 667                         case WINED3DSPR_RASTOUT:
 668                             switch (idx)
 669                             {
 670                                 case 0: /* oPos */
 671                                     reg_maps->output_registers |= 1 << 10;
 672                                     shader_signature_from_usage(&output_signature[10],
 673                                             WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 674                                     break;
 675
 676                                 case 1: /* oFog */
 677                                     reg_maps->output_registers |= 1 << 11;
 678                                     shader_signature_from_usage(&output_signature[11],
 679                                             WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 680                                     break;
 681
 682                                 case 2: /* oPts */
 683                                     reg_maps->output_registers |= 1 << 11;
 684                                     shader_signature_from_usage(&output_signature[11],
 685                                             WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 686                                     break;
 687                             }
 688                             break;
 689
 690                         case WINED3DSPR_ATTROUT:
 691                             if (idx < 2)
 692                             {
 693                                 idx += 8;
 694                                 if (reg_maps->output_registers & (1 << idx))
 695                                 {
 696                                     output_signature[idx].mask |= ins.dst[i].write_mask;
 697                                 }
 698                                 else
 699                                 {
 700                                     reg_maps->output_registers |= 1 << idx;
 701                                     shader_signature_from_usage(&output_signature[idx],
 702                                             WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
 703                                 }
 704                             }
 705                             break;
 706
 707                         case WINED3DSPR_TEXCRDOUT:
 708
 709                             reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
 710                             if (reg_maps->output_registers & (1 << idx))
 711                             {
 712                                 output_signature[idx].mask |= ins.dst[i].write_mask;
 713                             }
 714                             else
 715                             {
 716                                 reg_maps->output_registers |= 1 << idx;
 717                                 shader_signature_from_usage(&output_signature[idx],
 718                                         WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
 719                             }
 720                             break;
 721
 722                         default:
 723                             break;
 724                     }
 725                 }
 726
 727                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 728                 {
 729                     if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
 730                     {
 731                         /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 732                          * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 733                          * the mov and perform the sRGB write correction from the source register.
 734                          *
 735                          * However, if the mov is only partial, we can't do this, and if the write
 736                          * comes from an instruction other than MOV it is hard to do as well. If
 737                          * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 738                         shader->u.ps.color0_mov = FALSE;
 739                         if (ins.handler_idx == WINED3DSIH_MOV
 740                                 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
 741                         {
 742                             /* Used later when the source register is read. */
 743                             color0_mov = TRUE;
 744                         }
 745                     }
 746                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 747                      * end
 748                      */
 749                     else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
 750                             && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
 751                     {
 752                         shader->u.ps.color0_mov = FALSE;
 753                     }
 754                 }
 755
 756                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 757                 if (shader_version.major == 1
 758                         && (ins.handler_idx == WINED3DSIH_TEX
 759                             || ins.handler_idx == WINED3DSIH_TEXBEM
 760                             || ins.handler_idx == WINED3DSIH_TEXBEML
 761                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 762                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 763                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 764                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 765                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 766                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 767                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 768                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 769                 {
 770                     /* Fake sampler usage, only set reserved bit and type. */
 771                     DWORD sampler_code = ins.dst[i].reg.idx[0].offset;
 772
 773                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 774                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 775
 776                     /* texbem is only valid with < 1.4 pixel shaders */
 777                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 778                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 779                     {
 780                         reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
 781                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 782                         {
 783                             reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx[0].offset;
 784                         }
 785                     }
 786                 }
 787                 else if (ins.handler_idx == WINED3DSIH_BEM)
 788                 {
 789                     reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
 790                 }
 791             }
 792
 793             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 794             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 795             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 796             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 797             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 798             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 799             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 800             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 801             else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
 802             else if (ins.handler_idx == WINED3DSIH_LOOP
 803                     || ins.handler_idx == WINED3DSIH_REP)
 804             {
 805                 ++cur_loop_depth;
 806                 if (cur_loop_depth > max_loop_depth)
 807                     max_loop_depth = cur_loop_depth;
 808             }
 809             else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 810                     || ins.handler_idx == WINED3DSIH_ENDREP)
 811                 --cur_loop_depth;
 812
 813             if (ins.predicate)
 814                 shader_record_register_usage(shader, reg_maps, &ins.predicate->reg, shader_version.type);
 815
 816             for (i = 0; i < ins.src_count; ++i)
 817             {
 818                 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
 819                 struct wined3d_shader_register reg = ins.src[i].reg;
 820
 821                 shader_record_register_usage(shader, reg_maps, &ins.src[i].reg, shader_version.type);
 822                 while (count)
 823                 {
 824                     ++reg.idx[0].offset;
 825                     shader_record_register_usage(shader, reg_maps, &reg, shader_version.type);
 826                     --count;
 827                 }
 828
 829                 if (color0_mov)
 830                 {
 831                     if (ins.src[i].reg.type == WINED3DSPR_TEMP
 832                             && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
 833                     {
 834                         shader->u.ps.color0_mov = TRUE;
 835                         shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
 836                     }
 837                 }
 838             }
 839         }
 840     }
 841     reg_maps->loop_depth = max_loop_depth;
 842
 843     /* PS before 2.0 don't have explicit color outputs. Instead the value of
 844      * R0 is written to the render target. */
 845     if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 846         reg_maps->rt_mask |= (1 << 0);
 847
 848     shader->functionLength = ((const char *)ptr - (const char *)byte_code);
 849
 850     return WINED3D_OK;
 851 }
 852
 853 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 854 {
 855     DWORD map = 1 << max;
 856     map |= map - 1;
 857     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 858
 859     return wined3d_log2i(map);
 860 }
 861
 862 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 863         const struct wined3d_shader_version *shader_version)
 864 {
 865     TRACE("dcl");
 866
 867     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 868     {
 869         switch (semantic->sampler_type)
 870         {
 871             case WINED3DSTT_2D: TRACE("_2d"); break;
 872             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 873             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 874             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 875         }
 876     }
 877     else
 878     {
 879         /* Pixel shaders 3.0 don't have usage semantics. */
 880         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 881         else TRACE("_");
 882
 883         switch (semantic->usage)
 884         {
 885             case WINED3D_DECL_USAGE_POSITION:
 886                 TRACE("position%u", semantic->usage_idx);
 887                 break;
 888
 889             case WINED3D_DECL_USAGE_BLEND_INDICES:
 890                 TRACE("blend");
 891                 break;
 892
 893             case WINED3D_DECL_USAGE_BLEND_WEIGHT:
 894                 TRACE("weight");
 895                 break;
 896
 897             case WINED3D_DECL_USAGE_NORMAL:
 898                 TRACE("normal%u", semantic->usage_idx);
 899                 break;
 900
 901             case WINED3D_DECL_USAGE_PSIZE:
 902                 TRACE("psize");
 903                 break;
 904
 905             case WINED3D_DECL_USAGE_COLOR:
 906                 if (!semantic->usage_idx) TRACE("color");
 907                 else TRACE("specular%u", (semantic->usage_idx - 1));
 908                 break;
 909
 910             case WINED3D_DECL_USAGE_TEXCOORD:
 911                 TRACE("texture%u", semantic->usage_idx);
 912                 break;
 913
 914             case WINED3D_DECL_USAGE_TANGENT:
 915                 TRACE("tangent");
 916                 break;
 917
 918             case WINED3D_DECL_USAGE_BINORMAL:
 919                 TRACE("binormal");
 920                 break;
 921
 922             case WINED3D_DECL_USAGE_TESS_FACTOR:
 923                 TRACE("tessfactor");
 924                 break;
 925
 926             case WINED3D_DECL_USAGE_POSITIONT:
 927                 TRACE("positionT%u", semantic->usage_idx);
 928                 break;
 929
 930             case WINED3D_DECL_USAGE_FOG:
 931                 TRACE("fog");
 932                 break;
 933
 934             case WINED3D_DECL_USAGE_DEPTH:
 935                 TRACE("depth");
 936                 break;
 937
 938             case WINED3D_DECL_USAGE_SAMPLE:
 939                 TRACE("sample");
 940                 break;
 941
 942             default:
 943                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 944         }
 945     }
 946 }
 947
 948 static void shader_dump_register(const struct wined3d_shader_register *reg,
 949         const struct wined3d_shader_version *shader_version)
 950 {
 951     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 952     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 953     UINT offset = reg->idx[0].offset;
 954
 955     switch (reg->type)
 956     {
 957         case WINED3DSPR_TEMP:
 958             TRACE("r");
 959             break;
 960
 961         case WINED3DSPR_INPUT:
 962             TRACE("v");
 963             break;
 964
 965         case WINED3DSPR_CONST:
 966         case WINED3DSPR_CONST2:
 967         case WINED3DSPR_CONST3:
 968         case WINED3DSPR_CONST4:
 969             TRACE("c");
 970             offset = shader_get_float_offset(reg->type, offset);
 971             break;
 972
 973         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 974             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 975             break;
 976
 977         case WINED3DSPR_RASTOUT:
 978             TRACE("%s", rastout_reg_names[offset]);
 979             break;
 980
 981         case WINED3DSPR_COLOROUT:
 982             TRACE("oC");
 983             break;
 984
 985         case WINED3DSPR_DEPTHOUT:
 986             TRACE("oDepth");
 987             break;
 988
 989         case WINED3DSPR_ATTROUT:
 990             TRACE("oD");
 991             break;
 992
 993         case WINED3DSPR_TEXCRDOUT:
 994             /* Vertex shaders >= 3.0 use general purpose output registers
 995              * (WINED3DSPR_OUTPUT), which can include an address token. */
 996             if (shader_version->major >= 3) TRACE("o");
 997             else TRACE("oT");
 998             break;
 999
1000         case WINED3DSPR_CONSTINT:
1001             TRACE("i");
1002             break;
1003
1004         case WINED3DSPR_CONSTBOOL:
1005             TRACE("b");
1006             break;
1007
1008         case WINED3DSPR_LABEL:
1009             TRACE("l");
1010             break;
1011
1012         case WINED3DSPR_LOOP:
1013             TRACE("aL");
1014             break;
1015
1016         case WINED3DSPR_SAMPLER:
1017             TRACE("s");
1018             break;
1019
1020         case WINED3DSPR_MISCTYPE:
1021             if (offset > 1)
1022                 FIXME("Unhandled misctype register %u.\n", offset);
1023             else
1024                 TRACE("%s", misctype_reg_names[offset]);
1025             break;
1026
1027         case WINED3DSPR_PREDICATE:
1028             TRACE("p");
1029             break;
1030
1031         case WINED3DSPR_IMMCONST:
1032             TRACE("l");
1033             break;
1034
1035         case WINED3DSPR_CONSTBUFFER:
1036             TRACE("cb");
1037             break;
1038
1039         case WINED3DSPR_PRIMID:
1040             TRACE("primID");
1041             break;
1042
1043         case WINED3DSPR_NULL:
1044             TRACE("null");
1045             break;
1046
1047         case WINED3DSPR_RESOURCE:
1048             TRACE("t");
1049             break;
1050
1051         default:
1052             TRACE("unhandled_rtype(%#x)", reg->type);
1053             break;
1054     }
1055
1056     if (reg->type == WINED3DSPR_IMMCONST)
1057     {
1058         TRACE("(");
1059         switch (reg->immconst_type)
1060         {
1061             case WINED3D_IMMCONST_SCALAR:
1062                 switch (reg->data_type)
1063                 {
1064                     case WINED3D_DATA_FLOAT:
1065                         TRACE("%.8e", *(const float *)reg->immconst_data);
1066                         break;
1067                     case WINED3D_DATA_INT:
1068                         TRACE("%d", reg->immconst_data[0]);
1069                         break;
1070                     case WINED3D_DATA_RESOURCE:
1071                     case WINED3D_DATA_SAMPLER:
1072                     case WINED3D_DATA_UINT:
1073                         TRACE("%u", reg->immconst_data[0]);
1074                         break;
1075                     default:
1076                         TRACE("<unhandled data type %#x>", reg->data_type);
1077                         break;
1078                 }
1079                 break;
1080
1081             case WINED3D_IMMCONST_VEC4:
1082                 switch (reg->data_type)
1083                 {
1084                     case WINED3D_DATA_FLOAT:
1085                         TRACE("%.8e, %.8e, %.8e, %.8e",
1086                                 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1087                                 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1088                         break;
1089                     case WINED3D_DATA_INT:
1090                         TRACE("%d, %d, %d, %d",
1091                                 reg->immconst_data[0], reg->immconst_data[1],
1092                                 reg->immconst_data[2], reg->immconst_data[3]);
1093                         break;
1094                     case WINED3D_DATA_RESOURCE:
1095                     case WINED3D_DATA_SAMPLER:
1096                     case WINED3D_DATA_UINT:
1097                         TRACE("%u, %u, %u, %u",
1098                                 reg->immconst_data[0], reg->immconst_data[1],
1099                                 reg->immconst_data[2], reg->immconst_data[3]);
1100                         break;
1101                     default:
1102                         TRACE("<unhandled data type %#x>", reg->data_type);
1103                         break;
1104                 }
1105                 break;
1106
1107             default:
1108                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1109                 break;
1110         }
1111         TRACE(")");
1112     }
1113     else if (reg->type != WINED3DSPR_RASTOUT
1114             && reg->type != WINED3DSPR_MISCTYPE
1115             && reg->type != WINED3DSPR_NULL)
1116     {
1117         if (offset != ~0U)
1118         {
1119             TRACE("[");
1120             if (reg->idx[0].rel_addr)
1121             {
1122                 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1123                 TRACE(" + ");
1124             }
1125             TRACE("%u]", offset);
1126
1127             if (reg->idx[1].offset != ~0U)
1128             {
1129                 TRACE("[");
1130                 if (reg->idx[1].rel_addr)
1131                 {
1132                     shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1133                     TRACE(" + ");
1134                 }
1135                 TRACE("%u]", reg->idx[1].offset);
1136             }
1137         }
1138     }
1139 }
1140
1141 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1142         const struct wined3d_shader_version *shader_version)
1143 {
1144     DWORD write_mask = param->write_mask;
1145
1146     shader_dump_register(&param->reg, shader_version);
1147
1148     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1149     {
1150         static const char *write_mask_chars = "xyzw";
1151
1152         TRACE(".");
1153         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1154         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1155         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1156         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1157     }
1158 }
1159
1160 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1161         const struct wined3d_shader_version *shader_version)
1162 {
1163     enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1164     DWORD swizzle = param->swizzle;
1165
1166     if (src_modifier == WINED3DSPSM_NEG
1167             || src_modifier == WINED3DSPSM_BIASNEG
1168             || src_modifier == WINED3DSPSM_SIGNNEG
1169             || src_modifier == WINED3DSPSM_X2NEG
1170             || src_modifier == WINED3DSPSM_ABSNEG)
1171         TRACE("-");
1172     else if (src_modifier == WINED3DSPSM_COMP)
1173         TRACE("1-");
1174     else if (src_modifier == WINED3DSPSM_NOT)
1175         TRACE("!");
1176
1177     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1178         TRACE("abs(");
1179
1180     shader_dump_register(&param->reg, shader_version);
1181
1182     if (src_modifier)
1183     {
1184         switch (src_modifier)
1185         {
1186             case WINED3DSPSM_NONE:    break;
1187             case WINED3DSPSM_NEG:     break;
1188             case WINED3DSPSM_NOT:     break;
1189             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1190             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1191             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1192             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1193             case WINED3DSPSM_COMP:    break;
1194             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1195             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1196             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1197             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1198             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1199             case WINED3DSPSM_ABS:     TRACE(")"); break;
1200             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1201         }
1202     }
1203
1204     if (swizzle != WINED3DSP_NOSWIZZLE)
1205     {
1206         static const char *swizzle_chars = "xyzw";
1207         DWORD swizzle_x = swizzle & 0x03;
1208         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1209         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1210         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1211
1212         if (swizzle_x == swizzle_y
1213                 && swizzle_x == swizzle_z
1214                 && swizzle_x == swizzle_w)
1215         {
1216             TRACE(".%c", swizzle_chars[swizzle_x]);
1217         }
1218         else
1219         {
1220             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1221                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1222         }
1223     }
1224 }
1225
1226 /* Shared code in order to generate the bulk of the shader string.
1227  * NOTE: A description of how to parse tokens can be found on MSDN. */
1228 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1229         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1230 {
1231     struct wined3d_device *device = shader->device;
1232     const struct wined3d_shader_frontend *fe = shader->frontend;
1233     void *fe_data = shader->frontend_data;
1234     struct wined3d_shader_version shader_version;
1235     struct wined3d_shader_loop_state loop_state;
1236     struct wined3d_shader_instruction ins;
1237     struct wined3d_shader_tex_mx tex_mx;
1238     struct wined3d_shader_context ctx;
1239     const DWORD *ptr = byte_code;
1240
1241     /* Initialize current parsing state. */
1242     tex_mx.current_row = 0;
1243     loop_state.current_depth = 0;
1244     loop_state.current_reg = 0;
1245
1246     ctx.shader = shader;
1247     ctx.gl_info = &device->adapter->gl_info;
1248     ctx.reg_maps = reg_maps;
1249     ctx.buffer = buffer;
1250     ctx.tex_mx = &tex_mx;
1251     ctx.loop_state = &loop_state;
1252     ctx.backend_data = backend_ctx;
1253     ins.ctx = &ctx;
1254
1255     fe->shader_read_header(fe_data, &ptr, &shader_version);
1256
1257     while (!fe->shader_is_end(fe_data, &ptr))
1258     {
1259         /* Read opcode. */
1260         fe->shader_read_instruction(fe_data, &ptr, &ins);
1261
1262         /* Unknown opcode and its parameters. */
1263         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1264         {
1265             TRACE("Skipping unrecognized instruction.\n");
1266             continue;
1267         }
1268
1269         if (ins.predicate)
1270             FIXME("Predicates not implemented.\n");
1271
1272         /* Call appropriate function for output target */
1273         device->shader_backend->shader_handle_instruction(&ins);
1274     }
1275 }
1276
1277 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1278 {
1279     DWORD mmask = dst->modifiers;
1280
1281     switch (dst->shift)
1282     {
1283         case 0: break;
1284         case 13: TRACE("_d8"); break;
1285         case 14: TRACE("_d4"); break;
1286         case 15: TRACE("_d2"); break;
1287         case 1: TRACE("_x2"); break;
1288         case 2: TRACE("_x4"); break;
1289         case 3: TRACE("_x8"); break;
1290         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1291     }
1292
1293     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1294     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1295     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1296
1297     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1298     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1299 }
1300
1301 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1302 {
1303     switch (primitive_type)
1304     {
1305         case WINED3D_PT_UNDEFINED:
1306             TRACE("undefined");
1307             break;
1308         case WINED3D_PT_POINTLIST:
1309             TRACE("pointlist");
1310             break;
1311         case WINED3D_PT_LINELIST:
1312             TRACE("linelist");
1313             break;
1314         case WINED3D_PT_LINESTRIP:
1315             TRACE("linestrip");
1316             break;
1317         case WINED3D_PT_TRIANGLELIST:
1318             TRACE("trianglelist");
1319             break;
1320         case WINED3D_PT_TRIANGLESTRIP:
1321             TRACE("trianglestrip");
1322             break;
1323         case WINED3D_PT_TRIANGLEFAN:
1324             TRACE("trianglefan");
1325             break;
1326         case WINED3D_PT_LINELIST_ADJ:
1327             TRACE("linelist_adj");
1328             break;
1329         case WINED3D_PT_LINESTRIP_ADJ:
1330             TRACE("linestrip_adj");
1331             break;
1332         case WINED3D_PT_TRIANGLELIST_ADJ:
1333             TRACE("trianglelist_adj");
1334             break;
1335         case WINED3D_PT_TRIANGLESTRIP_ADJ:
1336             TRACE("trianglestrip_adj");
1337             break;
1338         default:
1339             TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1340             break;
1341     }
1342 }
1343
1344 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1345 {
1346     struct wined3d_shader_version shader_version;
1347     const DWORD *ptr = byte_code;
1348     const char *type_prefix;
1349     DWORD i;
1350
1351     TRACE("Parsing %p.\n", byte_code);
1352
1353     fe->shader_read_header(fe_data, &ptr, &shader_version);
1354
1355     switch (shader_version.type)
1356     {
1357         case WINED3D_SHADER_TYPE_VERTEX:
1358             type_prefix = "vs";
1359             break;
1360
1361         case WINED3D_SHADER_TYPE_GEOMETRY:
1362             type_prefix = "gs";
1363             break;
1364
1365         case WINED3D_SHADER_TYPE_PIXEL:
1366             type_prefix = "ps";
1367             break;
1368
1369         default:
1370             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1371             type_prefix = "unknown";
1372             break;
1373     }
1374
1375     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1376
1377     while (!fe->shader_is_end(fe_data, &ptr))
1378     {
1379         struct wined3d_shader_instruction ins;
1380
1381         fe->shader_read_instruction(fe_data, &ptr, &ins);
1382         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1383         {
1384             TRACE("Skipping unrecognized instruction.\n");
1385             continue;
1386         }
1387
1388         if (ins.handler_idx == WINED3DSIH_DCL)
1389         {
1390             shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1391             shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1392             TRACE(" ");
1393             shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1394         }
1395         else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1396         {
1397             TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1398             shader_dump_src_param(&ins.declaration.src, &shader_version);
1399             TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1400         }
1401         else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1402                 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1403         {
1404             TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1405             shader_dump_primitive_type(ins.declaration.primitive_type);
1406         }
1407         else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1408         {
1409             TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1410         }
1411         else if (ins.handler_idx == WINED3DSIH_DEF)
1412         {
1413             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1414                     ins.dst[0].reg.idx[0].offset),
1415                     *(const float *)&ins.src[0].reg.immconst_data[0],
1416                     *(const float *)&ins.src[0].reg.immconst_data[1],
1417                     *(const float *)&ins.src[0].reg.immconst_data[2],
1418                     *(const float *)&ins.src[0].reg.immconst_data[3]);
1419         }
1420         else if (ins.handler_idx == WINED3DSIH_DEFI)
1421         {
1422             TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1423                     ins.src[0].reg.immconst_data[0],
1424                     ins.src[0].reg.immconst_data[1],
1425                     ins.src[0].reg.immconst_data[2],
1426                     ins.src[0].reg.immconst_data[3]);
1427         }
1428         else if (ins.handler_idx == WINED3DSIH_DEFB)
1429         {
1430             TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1431         }
1432         else
1433         {
1434             if (ins.predicate)
1435             {
1436                 TRACE("(");
1437                 shader_dump_src_param(ins.predicate, &shader_version);
1438                 TRACE(") ");
1439             }
1440
1441             /* PixWin marks instructions with the coissue flag with a '+' */
1442             if (ins.coissue) TRACE("+");
1443
1444             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1445
1446             if (ins.handler_idx == WINED3DSIH_IFC
1447                     || ins.handler_idx == WINED3DSIH_BREAKC)
1448             {
1449                 switch (ins.flags)
1450                 {
1451                     case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1452                     case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1453                     case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1454                     case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1455                     case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1456                     case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1457                     default: TRACE("_(%u)", ins.flags);
1458                 }
1459             }
1460             else if (ins.handler_idx == WINED3DSIH_TEX
1461                     && shader_version.major >= 2
1462                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1463             {
1464                 TRACE("p");
1465             }
1466
1467             for (i = 0; i < ins.dst_count; ++i)
1468             {
1469                 shader_dump_ins_modifiers(&ins.dst[i]);
1470                 TRACE(!i ? " " : ", ");
1471                 shader_dump_dst_param(&ins.dst[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                 TRACE(!i ? " " : ", ");
1478                 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1479             }
1480         }
1481         TRACE("\n");
1482     }
1483 }
1484
1485 static void shader_cleanup(struct wined3d_shader *shader)
1486 {
1487     shader->device->shader_backend->shader_destroy(shader);
1488     HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1489     HeapFree(GetProcessHeap(), 0, shader->function);
1490     shader_delete_constant_list(&shader->constantsF);
1491     shader_delete_constant_list(&shader->constantsB);
1492     shader_delete_constant_list(&shader->constantsI);
1493     list_remove(&shader->shader_list_entry);
1494
1495     if (shader->frontend && shader->frontend_data)
1496         shader->frontend->shader_free(shader->frontend_data);
1497 }
1498
1499 struct shader_none_priv
1500 {
1501     const struct fragment_pipeline *fragment_pipe;
1502     BOOL ffp_proj_control;
1503 };
1504
1505 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1506 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1507         enum tex_types tex_type, const SIZE *ds_mask_size) {}
1508 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1509 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1510 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1511 static void shader_none_load_constants(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1512 static void shader_none_load_np2fixup_constants(void *shader_priv,
1513         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1514 static void shader_none_destroy(struct wined3d_shader *shader) {}
1515 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1516
1517 static void shader_none_select(const struct wined3d_context *context, enum wined3d_shader_mode vertex_mode,
1518         enum wined3d_shader_mode fragment_mode)
1519 {
1520     const struct wined3d_gl_info *gl_info = context->gl_info;
1521     struct wined3d_device *device = context->swapchain->device;
1522     struct shader_none_priv *priv = device->shader_priv;
1523
1524     priv->fragment_pipe->enable_extension(gl_info, fragment_mode == WINED3D_SHADER_MODE_FFP);
1525 }
1526
1527 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct fragment_pipeline *fragment_pipe)
1528 {
1529     struct fragment_caps fragment_caps;
1530     struct shader_none_priv *priv;
1531     void *fragment_priv;
1532
1533     if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1534         return E_OUTOFMEMORY;
1535
1536     if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1537     {
1538         ERR("Failed to initialize fragment pipe.\n");
1539         HeapFree(GetProcessHeap(), 0, priv);
1540         return E_FAIL;
1541     }
1542
1543     fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
1544     priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
1545     device->fragment_priv = fragment_priv;
1546     priv->fragment_pipe = fragment_pipe;
1547     device->shader_priv = priv;
1548
1549     return WINED3D_OK;
1550 }
1551
1552 static void shader_none_free(struct wined3d_device *device)
1553 {
1554     struct shader_none_priv *priv = device->shader_priv;
1555
1556     priv->fragment_pipe->free_private(device);
1557     HeapFree(GetProcessHeap(), 0, priv);
1558 }
1559
1560 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1561 {
1562     /* Set the shader caps to 0 for the none shader backend */
1563     caps->vs_version = 0;
1564     caps->gs_version = 0;
1565     caps->ps_version = 0;
1566     caps->vs_uniform_count = 0;
1567     caps->ps_uniform_count = 0;
1568     caps->ps_1x_max_value = 0.0f;
1569     caps->wined3d_caps = 0;
1570 }
1571
1572 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1573 {
1574     /* We "support" every possible fixup, since we don't support any shader
1575      * model, and will never have to actually sample a texture. */
1576     return TRUE;
1577 }
1578
1579 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
1580 {
1581     struct shader_none_priv *priv = shader_priv;
1582
1583     return priv->ffp_proj_control;
1584 }
1585
1586 const struct wined3d_shader_backend_ops none_shader_backend =
1587 {
1588     shader_none_handle_instruction,
1589     shader_none_select,
1590     shader_none_select_depth_blt,
1591     shader_none_deselect_depth_blt,
1592     shader_none_update_float_vertex_constants,
1593     shader_none_update_float_pixel_constants,
1594     shader_none_load_constants,
1595     shader_none_load_np2fixup_constants,
1596     shader_none_destroy,
1597     shader_none_alloc,
1598     shader_none_free,
1599     shader_none_context_destroyed,
1600     shader_none_get_caps,
1601     shader_none_color_fixup_supported,
1602     shader_none_has_ffp_proj_control,
1603 };
1604
1605 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1606         const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1607         enum wined3d_shader_type type, unsigned int max_version)
1608 {
1609     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1610     const struct wined3d_shader_frontend *fe;
1611     HRESULT hr;
1612     unsigned int backend_version;
1613
1614     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1615             shader, byte_code, output_signature, float_const_count);
1616
1617     fe = shader_select_frontend(*byte_code);
1618     if (!fe)
1619     {
1620         FIXME("Unable to find frontend for shader.\n");
1621         return WINED3DERR_INVALIDCALL;
1622     }
1623     shader->frontend = fe;
1624     shader->frontend_data = fe->shader_init(byte_code, output_signature);
1625     if (!shader->frontend_data)
1626     {
1627         FIXME("Failed to initialize frontend.\n");
1628         return WINED3DERR_INVALIDCALL;
1629     }
1630
1631     /* First pass: trace shader. */
1632     if (TRACE_ON(d3d_shader))
1633         shader_trace_init(fe, shader->frontend_data, byte_code);
1634
1635     /* Initialize immediate constant lists. */
1636     list_init(&shader->constantsF);
1637     list_init(&shader->constantsB);
1638     list_init(&shader->constantsI);
1639     shader->lconst_inf_or_nan = FALSE;
1640
1641     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1642     hr = shader_get_registers_used(shader, fe,
1643             reg_maps, shader->input_signature, shader->output_signature,
1644             byte_code, float_const_count);
1645     if (FAILED(hr)) return hr;
1646
1647     if (reg_maps->shader_version.type != type)
1648     {
1649         WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1650         return WINED3DERR_INVALIDCALL;
1651     }
1652     if (reg_maps->shader_version.major > max_version)
1653     {
1654         WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1655         return WINED3DERR_INVALIDCALL;
1656     }
1657     switch (type)
1658     {
1659         case WINED3D_SHADER_TYPE_VERTEX:
1660             backend_version = shader->device->vs_version;
1661             break;
1662         case WINED3D_SHADER_TYPE_GEOMETRY:
1663             backend_version = shader->device->gs_version;
1664             break;
1665         case WINED3D_SHADER_TYPE_PIXEL:
1666             backend_version = shader->device->ps_version;
1667             break;
1668         default:
1669             FIXME("No backend version-checking for this shader type\n");
1670             backend_version = 0;
1671     }
1672     if (reg_maps->shader_version.major > backend_version)
1673     {
1674         WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1675                 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1676         return WINED3DERR_INVALIDCALL;
1677     }
1678
1679     shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1680     if (!shader->function)
1681         return E_OUTOFMEMORY;
1682     memcpy(shader->function, byte_code, shader->functionLength);
1683
1684     return WINED3D_OK;
1685 }
1686
1687 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1688 {
1689     ULONG refcount = InterlockedIncrement(&shader->ref);
1690
1691     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1692
1693     return refcount;
1694 }
1695
1696 /* Do not call while under the GL lock. */
1697 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1698 {
1699     ULONG refcount = InterlockedDecrement(&shader->ref);
1700
1701     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1702
1703     if (!refcount)
1704     {
1705         shader_cleanup(shader);
1706         shader->parent_ops->wined3d_object_destroyed(shader->parent);
1707         HeapFree(GetProcessHeap(), 0, shader);
1708     }
1709
1710     return refcount;
1711 }
1712
1713 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1714 {
1715     TRACE("shader %p.\n", shader);
1716
1717     return shader->parent;
1718 }
1719
1720 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1721         void *byte_code, UINT *byte_code_size)
1722 {
1723     TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1724
1725     if (!byte_code)
1726     {
1727         *byte_code_size = shader->functionLength;
1728         return WINED3D_OK;
1729     }
1730
1731     if (*byte_code_size < shader->functionLength)
1732     {
1733         /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1734          * than the required size we should write the required size and
1735          * return D3DERR_MOREDATA. That's not actually true. */
1736         return WINED3DERR_INVALIDCALL;
1737     }
1738
1739     memcpy(byte_code, shader->function, shader->functionLength);
1740
1741     return WINED3D_OK;
1742 }
1743
1744 /* Set local constants for d3d8 shaders. */
1745 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1746         UINT start_idx, const float *src_data, UINT count)
1747 {
1748     UINT end_idx = start_idx + count;
1749     UINT i;
1750
1751     TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1752
1753     if (end_idx > shader->limits.constant_float)
1754     {
1755         WARN("end_idx %u > float constants limit %u.\n",
1756                 end_idx, shader->limits.constant_float);
1757         end_idx = shader->limits.constant_float;
1758     }
1759
1760     for (i = start_idx; i < end_idx; ++i)
1761     {
1762         struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1763         float *value;
1764         if (!lconst)
1765             return E_OUTOFMEMORY;
1766
1767         lconst->idx = i;
1768         value = (float *)lconst->value;
1769         memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1770         list_add_head(&shader->constantsF, &lconst->entry);
1771
1772         if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1773                 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1774         {
1775             shader->lconst_inf_or_nan = TRUE;
1776         }
1777     }
1778
1779     return WINED3D_OK;
1780 }
1781
1782 void find_vs_compile_args(const struct wined3d_state *state,
1783         const struct wined3d_shader *shader, struct vs_compile_args *args)
1784 {
1785     args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1786             == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1787     args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1788             && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1789     args->swizzle_map = shader->device->stream_info.swizzle_map;
1790 }
1791
1792 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1793 {
1794     if (usage_idx1 != usage_idx2)
1795         return FALSE;
1796     if (usage1 == usage2)
1797         return TRUE;
1798     if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1799         return TRUE;
1800     if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1801         return TRUE;
1802
1803     return FALSE;
1804 }
1805
1806 BOOL vshader_get_input(const struct wined3d_shader *shader,
1807         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1808 {
1809     WORD map = shader->reg_maps.input_registers;
1810     unsigned int i;
1811
1812     for (i = 0; map; map >>= 1, ++i)
1813     {
1814         if (!(map & 1)) continue;
1815
1816         if (match_usage(shader->u.vs.attributes[i].usage,
1817                 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1818         {
1819             *regnum = i;
1820             return TRUE;
1821         }
1822     }
1823     return FALSE;
1824 }
1825
1826 static void vertexshader_set_limits(struct wined3d_shader *shader)
1827 {
1828     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1829             shader->reg_maps.shader_version.minor);
1830     struct wined3d_device *device = shader->device;
1831
1832     shader->limits.packed_input = 0;
1833
1834     switch (shader_version)
1835     {
1836         case WINED3D_SHADER_VERSION(1, 0):
1837         case WINED3D_SHADER_VERSION(1, 1):
1838             shader->limits.constant_bool = 0;
1839             shader->limits.constant_int = 0;
1840             shader->limits.packed_output = 12;
1841             shader->limits.sampler = 0;
1842             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1843              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1844              * constants? */
1845             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1846             break;
1847
1848         case WINED3D_SHADER_VERSION(2, 0):
1849         case WINED3D_SHADER_VERSION(2, 1):
1850             shader->limits.constant_bool = 16;
1851             shader->limits.constant_int = 16;
1852             shader->limits.packed_output = 12;
1853             shader->limits.sampler = 0;
1854             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1855             break;
1856
1857         case WINED3D_SHADER_VERSION(3, 0):
1858             shader->limits.constant_bool = 16;
1859             shader->limits.constant_int = 16;
1860             shader->limits.packed_output = 12;
1861             shader->limits.sampler = 4;
1862             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1863              * even though they are capable of supporting much more (GL
1864              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1865              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1866              * shaders to 256. */
1867             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1868             break;
1869
1870         case WINED3D_SHADER_VERSION(4, 0):
1871             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1872             shader->limits.constant_int = 0;
1873             shader->limits.constant_float = 0;
1874             shader->limits.constant_bool = 0;
1875             shader->limits.packed_output = 16;
1876             shader->limits.packed_input = 0;
1877             break;
1878
1879         default:
1880             shader->limits.constant_bool = 16;
1881             shader->limits.constant_int = 16;
1882             shader->limits.packed_output = 12;
1883             shader->limits.sampler = 0;
1884             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1885             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1886                     shader->reg_maps.shader_version.major,
1887                     shader->reg_maps.shader_version.minor);
1888     }
1889 }
1890
1891 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1892         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1893         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1894 {
1895     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1896     unsigned int i;
1897     HRESULT hr;
1898     WORD map;
1899
1900     if (!byte_code) return WINED3DERR_INVALIDCALL;
1901
1902     shader_init(shader, device, parent, parent_ops);
1903     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1904             WINED3D_SHADER_TYPE_VERTEX, max_version);
1905     if (FAILED(hr))
1906     {
1907         WARN("Failed to set function, hr %#x.\n", hr);
1908         shader_cleanup(shader);
1909         return hr;
1910     }
1911
1912     map = reg_maps->input_registers;
1913     for (i = 0; map; map >>= 1, ++i)
1914     {
1915         if (!(map & 1) || !shader->input_signature[i].semantic_name)
1916             continue;
1917
1918         shader->u.vs.attributes[i].usage =
1919                 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1920         shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1921     }
1922
1923     if (output_signature)
1924     {
1925         for (i = 0; i < output_signature->element_count; ++i)
1926         {
1927             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1928             reg_maps->output_registers |= 1 << e->register_idx;
1929             shader->output_signature[e->register_idx] = *e;
1930         }
1931     }
1932
1933     vertexshader_set_limits(shader);
1934
1935     shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
1936             shader->lconst_inf_or_nan;
1937
1938     return WINED3D_OK;
1939 }
1940
1941 static void geometryshader_set_limits(struct wined3d_shader *shader)
1942 {
1943     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1944             shader->reg_maps.shader_version.minor);
1945
1946     switch (shader_version)
1947     {
1948         case WINED3D_SHADER_VERSION(4, 0):
1949             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1950             shader->limits.constant_int = 0;
1951             shader->limits.constant_float = 0;
1952             shader->limits.constant_bool = 0;
1953             shader->limits.packed_output = 32;
1954             shader->limits.packed_input = 16;
1955             break;
1956
1957         default:
1958             memset(&shader->limits, 0, sizeof(shader->limits));
1959             FIXME("Unhandled geometry shader version \"%u.%u\".\n",
1960                     shader->reg_maps.shader_version.major,
1961                     shader->reg_maps.shader_version.minor);
1962     }
1963 }
1964
1965 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1966         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1967         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1968 {
1969     HRESULT hr;
1970
1971     shader_init(shader, device, parent, parent_ops);
1972     hr = shader_set_function(shader, byte_code, output_signature, 0,
1973             WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1974     if (FAILED(hr))
1975     {
1976         WARN("Failed to set function, hr %#x.\n", hr);
1977         shader_cleanup(shader);
1978         return hr;
1979     }
1980
1981     geometryshader_set_limits(shader);
1982
1983     shader->load_local_constsF = shader->lconst_inf_or_nan;
1984
1985     return WINED3D_OK;
1986 }
1987
1988 void find_ps_compile_args(const struct wined3d_state *state,
1989         const struct wined3d_shader *shader, struct ps_compile_args *args)
1990 {
1991     struct wined3d_device *device = shader->device;
1992     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1993     const struct wined3d_texture *texture;
1994     UINT i;
1995
1996     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1997     if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1998     {
1999         const struct wined3d_surface *rt = state->fb->render_targets[0];
2000         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
2001         {
2002             static unsigned int warned = 0;
2003
2004             args->srgb_correction = 1;
2005             if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2006                 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2007                         "support, expect rendering artifacts.\n");
2008         }
2009     }
2010
2011     if (shader->reg_maps.shader_version.major == 1
2012             && shader->reg_maps.shader_version.minor <= 3)
2013     {
2014         for (i = 0; i < shader->limits.sampler; ++i)
2015         {
2016             DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2017
2018             if (flags & WINED3D_TTFF_PROJECTED)
2019             {
2020                 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2021
2022                 if (!state->vertex_shader)
2023                 {
2024                     unsigned int j;
2025                     unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2026                     DWORD max_valid = WINED3D_TTFF_COUNT4;
2027                     enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
2028
2029                     for (j = 0; j < state->vertex_declaration->element_count; ++j)
2030                     {
2031                         struct wined3d_vertex_declaration_element *element =
2032                                 &state->vertex_declaration->elements[j];
2033
2034                         if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2035                                 && element->usage_idx == index)
2036                         {
2037                             max_valid = element->format->component_count;
2038                             break;
2039                         }
2040                     }
2041                     if (!tex_transform || tex_transform > max_valid)
2042                     {
2043                         WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2044                                 tex_transform, max_valid);
2045                         tex_transform = max_valid;
2046                     }
2047                     if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
2048                             || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
2049                             || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
2050                         tex_transform |= WINED3D_PSARGS_PROJECTED;
2051                     else
2052                     {
2053                         WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2054                         WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2055                                 i, tex_transform, sampler_type);
2056                     }
2057                 }
2058                 else
2059                     tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2060
2061                 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2062             }
2063         }
2064     }
2065     if (shader->reg_maps.shader_version.major == 1
2066             && shader->reg_maps.shader_version.minor <= 4)
2067     {
2068         for (i = 0; i < shader->limits.sampler; ++i)
2069         {
2070             const struct wined3d_texture *texture = state->textures[i];
2071
2072             if (!shader->reg_maps.sampler_type[i])
2073                 continue;
2074
2075             /* Treat unbound textures as 2D. The dummy texture will provide
2076              * the proper sample value. The tex_types bitmap defaults to
2077              * 2D because of the memset. */
2078             if (!texture)
2079                 continue;
2080
2081             switch (texture->target)
2082             {
2083                 /* RECT textures are distinguished from 2D textures via np2_fixup */
2084                 case GL_TEXTURE_RECTANGLE_ARB:
2085                 case GL_TEXTURE_2D:
2086                     break;
2087
2088                 case GL_TEXTURE_3D:
2089                     args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2090                     break;
2091
2092                 case GL_TEXTURE_CUBE_MAP_ARB:
2093                     args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2094                     break;
2095             }
2096         }
2097     }
2098
2099     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2100     {
2101         if (!shader->reg_maps.sampler_type[i])
2102             continue;
2103
2104         texture = state->textures[i];
2105         if (!texture)
2106         {
2107             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2108             continue;
2109         }
2110         args->color_fixup[i] = texture->resource.format->color_fixup;
2111
2112         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2113             args->shadow |= 1 << i;
2114
2115         /* Flag samplers that need NP2 texcoord fixup. */
2116         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2117             args->np2_fixup |= (1 << i);
2118     }
2119     if (shader->reg_maps.shader_version.major >= 3)
2120     {
2121         if (device->stream_info.position_transformed)
2122             args->vp_mode = pretransformed;
2123         else if (use_vs(state))
2124             args->vp_mode = vertexshader;
2125         else
2126             args->vp_mode = fixedfunction;
2127         args->fog = FOG_OFF;
2128     }
2129     else
2130     {
2131         args->vp_mode = vertexshader;
2132         if (state->render_states[WINED3D_RS_FOGENABLE])
2133         {
2134             switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2135             {
2136                 case WINED3D_FOG_NONE:
2137                     if (device->stream_info.position_transformed || use_vs(state))
2138                     {
2139                         args->fog = FOG_LINEAR;
2140                         break;
2141                     }
2142
2143                     switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2144                     {
2145                         case WINED3D_FOG_NONE: /* Fall through. */
2146                         case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2147                         case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
2148                         case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
2149                     }
2150                     break;
2151
2152                 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2153                 case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
2154                 case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
2155             }
2156         }
2157         else
2158         {
2159             args->fog = FOG_OFF;
2160         }
2161     }
2162 }
2163
2164 static void pixelshader_set_limits(struct wined3d_shader *shader)
2165 {
2166     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2167             shader->reg_maps.shader_version.minor);
2168
2169     shader->limits.packed_output = 0;
2170
2171     switch (shader_version)
2172     {
2173         case WINED3D_SHADER_VERSION(1, 0):
2174         case WINED3D_SHADER_VERSION(1, 1):
2175         case WINED3D_SHADER_VERSION(1, 2):
2176         case WINED3D_SHADER_VERSION(1, 3):
2177             shader->limits.constant_float = 8;
2178             shader->limits.constant_int = 0;
2179             shader->limits.constant_bool = 0;
2180             shader->limits.sampler = 4;
2181             shader->limits.packed_input = 0;
2182             break;
2183
2184         case WINED3D_SHADER_VERSION(1, 4):
2185             shader->limits.constant_float = 8;
2186             shader->limits.constant_int = 0;
2187             shader->limits.constant_bool = 0;
2188             shader->limits.sampler = 6;
2189             shader->limits.packed_input = 0;
2190             break;
2191
2192         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2193         case WINED3D_SHADER_VERSION(2, 0):
2194             shader->limits.constant_float = 32;
2195             shader->limits.constant_int = 16;
2196             shader->limits.constant_bool = 16;
2197             shader->limits.sampler = 16;
2198             shader->limits.packed_input = 0;
2199             break;
2200
2201         case WINED3D_SHADER_VERSION(2, 1):
2202             shader->limits.constant_float = 32;
2203             shader->limits.constant_int = 16;
2204             shader->limits.constant_bool = 16;
2205             shader->limits.sampler = 16;
2206             shader->limits.packed_input = 0;
2207             break;
2208
2209         case WINED3D_SHADER_VERSION(3, 0):
2210             shader->limits.constant_float = 224;
2211             shader->limits.constant_int = 16;
2212             shader->limits.constant_bool = 16;
2213             shader->limits.sampler = 16;
2214             shader->limits.packed_input = 12;
2215             break;
2216
2217         case WINED3D_SHADER_VERSION(4, 0):
2218             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
2219             shader->limits.constant_int = 0;
2220             shader->limits.constant_float = 0;
2221             shader->limits.constant_bool = 0;
2222             shader->limits.packed_input = 32;
2223             break;
2224
2225         default:
2226             shader->limits.constant_float = 32;
2227             shader->limits.constant_int = 16;
2228             shader->limits.constant_bool = 16;
2229             shader->limits.sampler = 16;
2230             shader->limits.packed_input = 0;
2231             FIXME("Unrecognized pixel shader version %u.%u\n",
2232                     shader->reg_maps.shader_version.major,
2233                     shader->reg_maps.shader_version.minor);
2234     }
2235 }
2236
2237 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2238         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2239         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2240 {
2241     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2242     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2243     HRESULT hr;
2244
2245     if (!byte_code) return WINED3DERR_INVALIDCALL;
2246
2247     shader_init(shader, device, parent, parent_ops);
2248     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2249             WINED3D_SHADER_TYPE_PIXEL, max_version);
2250     if (FAILED(hr))
2251     {
2252         WARN("Failed to set function, hr %#x.\n", hr);
2253         shader_cleanup(shader);
2254         return hr;
2255     }
2256
2257     pixelshader_set_limits(shader);
2258
2259     for (i = 0; i < MAX_REG_INPUT; ++i)
2260     {
2261         if (shader->u.ps.input_reg_used[i])
2262         {
2263             ++num_regs_used;
2264             highest_reg_used = i;
2265         }
2266     }
2267
2268     /* Don't do any register mapping magic if it is not needed, or if we can't
2269      * achieve anything anyway */
2270     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2271             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2272     {
2273         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2274         {
2275             /* This happens with relative addressing. The input mapper function
2276              * warns about this if the higher registers are declared too, so
2277              * don't write a FIXME here */
2278             WARN("More varying registers used than supported\n");
2279         }
2280
2281         for (i = 0; i < MAX_REG_INPUT; ++i)
2282         {
2283             shader->u.ps.input_reg_map[i] = i;
2284         }
2285
2286         shader->u.ps.declared_in_count = highest_reg_used + 1;
2287     }
2288     else
2289     {
2290         shader->u.ps.declared_in_count = 0;
2291         for (i = 0; i < MAX_REG_INPUT; ++i)
2292         {
2293             if (shader->u.ps.input_reg_used[i])
2294                 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2295             else shader->u.ps.input_reg_map[i] = ~0U;
2296         }
2297     }
2298
2299     shader->load_local_constsF = shader->lconst_inf_or_nan;
2300
2301     return WINED3D_OK;
2302 }
2303
2304 void pixelshader_update_samplers(struct wined3d_shader *shader, WORD tex_types)
2305 {
2306     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2307     enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2308     unsigned int i;
2309
2310     if (reg_maps->shader_version.major != 1) return;
2311
2312     for (i = 0; i < shader->limits.sampler; ++i)
2313     {
2314         /* We don't sample from this sampler. */
2315         if (!sampler_type[i]) continue;
2316
2317         switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
2318         {
2319             case WINED3D_SHADER_TEX_2D:
2320                 sampler_type[i] = WINED3DSTT_2D;
2321                 break;
2322
2323             case WINED3D_SHADER_TEX_3D:
2324                 sampler_type[i] = WINED3DSTT_VOLUME;
2325                 break;
2326
2327             case WINED3D_SHADER_TEX_CUBE:
2328                 sampler_type[i] = WINED3DSTT_CUBE;
2329                 break;
2330         }
2331     }
2332 }
2333
2334 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2335         const struct wined3d_shader_signature *output_signature, void *parent,
2336         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2337 {
2338     struct wined3d_shader *object;
2339     HRESULT hr;
2340
2341     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2342             device, byte_code, output_signature, parent, parent_ops, shader);
2343
2344     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2345     if (!object)
2346         return E_OUTOFMEMORY;
2347
2348     hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2349     if (FAILED(hr))
2350     {
2351         WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2352         HeapFree(GetProcessHeap(), 0, object);
2353         return hr;
2354     }
2355
2356     TRACE("Created geometry shader %p.\n", object);
2357     *shader = object;
2358
2359     return WINED3D_OK;
2360 }
2361
2362 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2363         const struct wined3d_shader_signature *output_signature, void *parent,
2364         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2365 {
2366     struct wined3d_shader *object;
2367     HRESULT hr;
2368
2369     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2370             device, byte_code, output_signature, parent, parent_ops, shader);
2371
2372     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2373     if (!object)
2374         return E_OUTOFMEMORY;
2375
2376     hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2377     if (FAILED(hr))
2378     {
2379         WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2380         HeapFree(GetProcessHeap(), 0, object);
2381         return hr;
2382     }
2383
2384     TRACE("Created pixel shader %p.\n", object);
2385     *shader = object;
2386
2387     return WINED3D_OK;
2388 }
2389
2390 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2391         const struct wined3d_shader_signature *output_signature, void *parent,
2392         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2393 {
2394     struct wined3d_shader *object;
2395     HRESULT hr;
2396
2397     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2398             device, byte_code, output_signature, parent, parent_ops, shader);
2399
2400     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2401     if (!object)
2402         return E_OUTOFMEMORY;
2403
2404     hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2405     if (FAILED(hr))
2406     {
2407         WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2408         HeapFree(GetProcessHeap(), 0, object);
2409         return hr;
2410     }
2411
2412     TRACE("Created vertex shader %p.\n", object);
2413     *shader = object;
2414
2415     return WINED3D_OK;
2416 }