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 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  36
  37 static const char * const shader_opcode_names[] =
  38 {
  39     /* WINED3DSIH_ABS                   */ "abs",
  40     /* WINED3DSIH_ADD                   */ "add",
  41     /* WINED3DSIH_AND                   */ "and",
  42     /* WINED3DSIH_BEM                   */ "bem",
  43     /* WINED3DSIH_BREAK                 */ "break",
  44     /* WINED3DSIH_BREAKC                */ "breakc",
  45     /* WINED3DSIH_BREAKP                */ "breakp",
  46     /* WINED3DSIH_CALL                  */ "call",
  47     /* WINED3DSIH_CALLNZ                */ "callnz",
  48     /* WINED3DSIH_CMP                   */ "cmp",
  49     /* WINED3DSIH_CND                   */ "cnd",
  50     /* WINED3DSIH_CRS                   */ "crs",
  51     /* WINED3DSIH_CUT                   */ "cut",
  52     /* WINED3DSIH_DCL                   */ "dcl",
  53     /* WINED3DSIH_DCL_VERTICES_OUT      */ "dcl_maxOutputVertexCount",
  54     /* WINED3DSIH_DEF                   */ "def",
  55     /* WINED3DSIH_DEFB                  */ "defb",
  56     /* WINED3DSIH_DEFI                  */ "defi",
  57     /* WINED3DSIH_DIV                   */ "div",
  58     /* WINED3DSIH_DP2ADD                */ "dp2add",
  59     /* WINED3DSIH_DP3                   */ "dp3",
  60     /* WINED3DSIH_DP4                   */ "dp4",
  61     /* WINED3DSIH_DST                   */ "dst",
  62     /* WINED3DSIH_DSX                   */ "dsx",
  63     /* WINED3DSIH_DSY                   */ "dsy",
  64     /* WINED3DSIH_ELSE                  */ "else",
  65     /* WINED3DSIH_EMIT                  */ "emit",
  66     /* WINED3DSIH_ENDIF                 */ "endif",
  67     /* WINED3DSIH_ENDLOOP               */ "endloop",
  68     /* WINED3DSIH_ENDREP                */ "endrep",
  69     /* WINED3DSIH_EQ                    */ "eq",
  70     /* WINED3DSIH_EXP                   */ "exp",
  71     /* WINED3DSIH_EXPP                  */ "expp",
  72     /* WINED3DSIH_FRC                   */ "frc",
  73     /* WINED3DSIH_FTOI                  */ "ftoi",
  74     /* WINED3DSIH_GE                    */ "ge",
  75     /* WINED3DSIH_IADD                  */ "iadd",
  76     /* WINED3DSIH_IEQ                   */ "ieq",
  77     /* WINED3DSIH_IF                    */ "if",
  78     /* WINED3DSIH_IFC                   */ "ifc",
  79     /* WINED3DSIH_IGE                   */ "ige",
  80     /* WINED3DSIH_IMUL                  */ "imul",
  81     /* WINED3DSIH_ITOF                  */ "itof",
  82     /* WINED3DSIH_LABEL                 */ "label",
  83     /* WINED3DSIH_LD                    */ "ld",
  84     /* WINED3DSIH_LIT                   */ "lit",
  85     /* WINED3DSIH_LOG                   */ "log",
  86     /* WINED3DSIH_LOGP                  */ "logp",
  87     /* WINED3DSIH_LOOP                  */ "loop",
  88     /* WINED3DSIH_LRP                   */ "lrp",
  89     /* WINED3DSIH_LT                    */ "lt",
  90     /* WINED3DSIH_M3x2                  */ "m3x2",
  91     /* WINED3DSIH_M3x3                  */ "m3x3",
  92     /* WINED3DSIH_M3x4                  */ "m3x4",
  93     /* WINED3DSIH_M4x3                  */ "m4x3",
  94     /* WINED3DSIH_M4x4                  */ "m4x4",
  95     /* WINED3DSIH_MAD                   */ "mad",
  96     /* WINED3DSIH_MAX                   */ "max",
  97     /* WINED3DSIH_MIN                   */ "min",
  98     /* WINED3DSIH_MOV                   */ "mov",
  99     /* WINED3DSIH_MOVA                  */ "mova",
 100     /* WINED3DSIH_MOVC                  */ "movc",
 101     /* WINED3DSIH_MUL                   */ "mul",
 102     /* WINED3DSIH_NOP                   */ "nop",
 103     /* WINED3DSIH_NRM                   */ "nrm",
 104     /* WINED3DSIH_PHASE                 */ "phase",
 105     /* WINED3DSIH_POW                   */ "pow",
 106     /* WINED3DSIH_RCP                   */ "rcp",
 107     /* WINED3DSIH_REP                   */ "rep",
 108     /* WINED3DSIH_RET                   */ "ret",
 109     /* WINED3DSIH_ROUND_NI              */ "round_ni",
 110     /* WINED3DSIH_RSQ                   */ "rsq",
 111     /* WINED3DSIH_SAMPLE                */ "sample",
 112     /* WINED3DSIH_SAMPLE_GRAD           */ "sample_d",
 113     /* WINED3DSIH_SAMPLE_LOD            */ "sample_l",
 114     /* WINED3DSIH_SETP                  */ "setp",
 115     /* WINED3DSIH_SGE                   */ "sge",
 116     /* WINED3DSIH_SGN                   */ "sgn",
 117     /* WINED3DSIH_SINCOS                */ "sincos",
 118     /* WINED3DSIH_SLT                   */ "slt",
 119     /* WINED3DSIH_SQRT                  */ "sqrt",
 120     /* WINED3DSIH_SUB                   */ "sub",
 121     /* WINED3DSIH_TEX                   */ "texld",
 122     /* WINED3DSIH_TEXBEM                */ "texbem",
 123     /* WINED3DSIH_TEXBEML               */ "texbeml",
 124     /* WINED3DSIH_TEXCOORD              */ "texcrd",
 125     /* WINED3DSIH_TEXDEPTH              */ "texdepth",
 126     /* WINED3DSIH_TEXDP3                */ "texdp3",
 127     /* WINED3DSIH_TEXDP3TEX             */ "texdp3tex",
 128     /* WINED3DSIH_TEXKILL               */ "texkill",
 129     /* WINED3DSIH_TEXLDD                */ "texldd",
 130     /* WINED3DSIH_TEXLDL                */ "texldl",
 131     /* WINED3DSIH_TEXM3x2DEPTH          */ "texm3x2depth",
 132     /* WINED3DSIH_TEXM3x2PAD            */ "texm3x2pad",
 133     /* WINED3DSIH_TEXM3x2TEX            */ "texm3x2tex",
 134     /* WINED3DSIH_TEXM3x3               */ "texm3x3",
 135     /* WINED3DSIH_TEXM3x3DIFF           */ "texm3x3diff",
 136     /* WINED3DSIH_TEXM3x3PAD            */ "texm3x3pad",
 137     /* WINED3DSIH_TEXM3x3SPEC           */ "texm3x3spec",
 138     /* WINED3DSIH_TEXM3x3TEX            */ "texm3x3tex",
 139     /* WINED3DSIH_TEXM3x3VSPEC          */ "texm3x3vspec",
 140     /* WINED3DSIH_TEXREG2AR             */ "texreg2ar",
 141     /* WINED3DSIH_TEXREG2GB             */ "texreg2gb",
 142     /* WINED3DSIH_TEXREG2RGB            */ "texreg2rgb",
 143     /* WINED3DSIH_UDIV                  */ "udiv",
 144     /* WINED3DSIH_USHR                  */ "ushr",
 145     /* WINED3DSIH_UTOF                  */ "utof",
 146     /* WINED3DSIH_XOR                   */ "xor",
 147 };
 148
 149 static const char * const semantic_names[] =
 150 {
 151     /* WINED3D_DECL_USAGE_POSITION      */ "SV_POSITION",
 152     /* WINED3D_DECL_USAGE_BLEND_WEIGHT  */ "BLENDWEIGHT",
 153     /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
 154     /* WINED3D_DECL_USAGE_NORMAL        */ "NORMAL",
 155     /* WINED3D_DECL_USAGE_PSIZE         */ "PSIZE",
 156     /* WINED3D_DECL_USAGE_TEXCOORD      */ "TEXCOORD",
 157     /* WINED3D_DECL_USAGE_TANGENT       */ "TANGENT",
 158     /* WINED3D_DECL_USAGE_BINORMAL      */ "BINORMAL",
 159     /* WINED3D_DECL_USAGE_TESS_FACTOR   */ "TESSFACTOR",
 160     /* WINED3D_DECL_USAGE_POSITIONT     */ "POSITIONT",
 161     /* WINED3D_DECL_USAGE_COLOR         */ "COLOR",
 162     /* WINED3D_DECL_USAGE_FOG           */ "FOG",
 163     /* WINED3D_DECL_USAGE_DEPTH         */ "DEPTH",
 164     /* WINED3D_DECL_USAGE_SAMPLE        */ "SAMPLE",
 165 };
 166
 167 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
 168 {
 169     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 170     {
 171         FIXME("Unrecognized usage %#x.\n", usage);
 172         return "UNRECOGNIZED";
 173     }
 174
 175     return semantic_names[usage];
 176 }
 177
 178 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
 179 {
 180     unsigned int i;
 181
 182     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 183     {
 184         if (!strcmp(name, semantic_names[i])) return i;
 185     }
 186
 187     return ~0U;
 188 }
 189
 190 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
 191 {
 192     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 193 }
 194
 195 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 196         const struct wined3d_shader_semantic *s)
 197 {
 198     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 199     e->semantic_idx = s->usage_idx;
 200     e->sysval_semantic = 0;
 201     e->component_type = 0;
 202     e->register_idx = s->reg.reg.idx;
 203     e->mask = s->reg.write_mask;
 204 }
 205
 206 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
 207         enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
 208 {
 209     e->semantic_name = shader_semantic_name_from_usage(usage);
 210     e->semantic_idx = usage_idx;
 211     e->sysval_semantic = 0;
 212     e->component_type = 0;
 213     e->register_idx = reg_idx;
 214     e->mask = write_mask;
 215 }
 216
 217 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 218 {
 219     switch (version_token >> 16)
 220     {
 221         case WINED3D_SM1_VS:
 222         case WINED3D_SM1_PS:
 223             return &sm1_shader_frontend;
 224
 225         case WINED3D_SM4_PS:
 226         case WINED3D_SM4_VS:
 227         case WINED3D_SM4_GS:
 228             return &sm4_shader_frontend;
 229
 230         default:
 231             FIXME("Unrecognised version token %#x\n", version_token);
 232             return NULL;
 233     }
 234 }
 235
 236 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 237 {
 238     buffer->buffer[0] = '\0';
 239     buffer->bsize = 0;
 240     buffer->lineNo = 0;
 241     buffer->newline = TRUE;
 242 }
 243
 244 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 245 {
 246     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 247     if (!buffer->buffer)
 248     {
 249         ERR("Failed to allocate shader buffer memory.\n");
 250         return FALSE;
 251     }
 252
 253     shader_buffer_clear(buffer);
 254     return TRUE;
 255 }
 256
 257 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 258 {
 259     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 260 }
 261
 262 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 263 {
 264     char *base = buffer->buffer + buffer->bsize;
 265     int rc;
 266
 267     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 268
 269     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 270     {
 271         ERR("The buffer allocated for the shader program string "
 272             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 273         buffer->bsize = SHADER_PGMSIZE - 1;
 274         return -1;
 275     }
 276
 277     if (buffer->newline)
 278     {
 279         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 280         buffer->newline = FALSE;
 281     }
 282     else
 283     {
 284         TRACE("%s", base);
 285     }
 286
 287     buffer->bsize += rc;
 288     if (buffer->buffer[buffer->bsize-1] == '\n')
 289     {
 290         ++buffer->lineNo;
 291         buffer->newline = TRUE;
 292     }
 293
 294     return 0;
 295 }
 296
 297 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 298 {
 299     va_list args;
 300     int ret;
 301
 302     va_start(args, format);
 303     ret = shader_vaddline(buffer, format, args);
 304     va_end(args);
 305
 306     return ret;
 307 }
 308
 309 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
 310         void *parent, const struct wined3d_parent_ops *parent_ops)
 311 {
 312     shader->ref = 1;
 313     shader->device = device;
 314     shader->parent = parent;
 315     shader->parent_ops = parent_ops;
 316     list_init(&shader->linked_programs);
 317     list_add_head(&device->shaders, &shader->shader_list_entry);
 318 }
 319
 320 /* Convert floating point offset relative to a register file to an absolute
 321  * offset for float constants. */
 322 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
 323 {
 324     switch (register_type)
 325     {
 326         case WINED3DSPR_CONST: return register_idx;
 327         case WINED3DSPR_CONST2: return 2048 + register_idx;
 328         case WINED3DSPR_CONST3: return 4096 + register_idx;
 329         case WINED3DSPR_CONST4: return 6144 + register_idx;
 330         default:
 331             FIXME("Unsupported register type: %u.\n", register_type);
 332             return register_idx;
 333     }
 334 }
 335
 336 static void shader_delete_constant_list(struct list *clist)
 337 {
 338     struct wined3d_shader_lconst *constant;
 339     struct list *ptr;
 340
 341     ptr = list_head(clist);
 342     while (ptr)
 343     {
 344         constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
 345         ptr = list_next(clist, ptr);
 346         HeapFree(GetProcessHeap(), 0, constant);
 347     }
 348     list_init(clist);
 349 }
 350
 351 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 352 {
 353     DWORD idx, shift;
 354     idx = bit >> 5;
 355     shift = bit & 0x1f;
 356     bitmap[idx] |= (1 << shift);
 357 }
 358
 359 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
 360         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 361 {
 362     switch (reg->type)
 363     {
 364         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 365             if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
 366             else reg_maps->address |= 1 << reg->idx;
 367             break;
 368
 369         case WINED3DSPR_TEMP:
 370             reg_maps->temporary |= 1 << reg->idx;
 371             break;
 372
 373         case WINED3DSPR_INPUT:
 374             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 375             {
 376                 if (reg->rel_addr)
 377                 {
 378                     /* If relative addressing is used, we must assume that all registers
 379                      * are used. Even if it is a construct like v3[aL], we can't assume
 380                      * that v0, v1 and v2 aren't read because aL can be negative */
 381                     unsigned int i;
 382                     for (i = 0; i < MAX_REG_INPUT; ++i)
 383                     {
 384                         shader->u.ps.input_reg_used[i] = TRUE;
 385                     }
 386                 }
 387                 else
 388                 {
 389                     shader->u.ps.input_reg_used[reg->idx] = TRUE;
 390                 }
 391             }
 392             else reg_maps->input_registers |= 1 << reg->idx;
 393             break;
 394
 395         case WINED3DSPR_RASTOUT:
 396             if (reg->idx == 1) reg_maps->fog = 1;
 397             break;
 398
 399         case WINED3DSPR_MISCTYPE:
 400             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 401             {
 402                 if (!reg->idx) reg_maps->vpos = 1;
 403                 else if (reg->idx == 1) reg_maps->usesfacing = 1;
 404             }
 405             break;
 406
 407         case WINED3DSPR_CONST:
 408             if (reg->rel_addr)
 409             {
 410                 if (reg->idx < reg_maps->min_rel_offset) reg_maps->min_rel_offset = reg->idx;
 411                 if (reg->idx > reg_maps->max_rel_offset) reg_maps->max_rel_offset = reg->idx;
 412                 reg_maps->usesrelconstF = TRUE;
 413             }
 414             else
 415             {
 416                 set_bitmap_bit(reg_maps->constf, reg->idx);
 417             }
 418             break;
 419
 420         case WINED3DSPR_CONSTINT:
 421             reg_maps->integer_constants |= (1 << reg->idx);
 422             break;
 423
 424         case WINED3DSPR_CONSTBOOL:
 425             reg_maps->boolean_constants |= (1 << reg->idx);
 426             break;
 427
 428         case WINED3DSPR_COLOROUT:
 429             reg_maps->rt_mask |= (1 << reg->idx);
 430             break;
 431
 432         default:
 433             TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
 434             break;
 435     }
 436 }
 437
 438 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 439 {
 440     switch (instr)
 441     {
 442         case WINED3DSIH_M4x4:
 443         case WINED3DSIH_M3x4:
 444             return param == 1 ? 3 : 0;
 445
 446         case WINED3DSIH_M4x3:
 447         case WINED3DSIH_M3x3:
 448             return param == 1 ? 2 : 0;
 449
 450         case WINED3DSIH_M3x2:
 451             return param == 1 ? 1 : 0;
 452
 453         default:
 454             return 0;
 455     }
 456 }
 457
 458 /* Note that this does not count the loop register as an address register. */
 459 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
 460         struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 461         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 462 {
 463     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 464     void *fe_data = shader->frontend_data;
 465     struct wined3d_shader_version shader_version;
 466     const DWORD *ptr = byte_code;
 467
 468     memset(reg_maps, 0, sizeof(*reg_maps));
 469     reg_maps->min_rel_offset = ~0U;
 470
 471     fe->shader_read_header(fe_data, &ptr, &shader_version);
 472     reg_maps->shader_version = shader_version;
 473
 474     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 475             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 476     if (!reg_maps->constf)
 477     {
 478         ERR("Failed to allocate constant map memory.\n");
 479         return E_OUTOFMEMORY;
 480     }
 481
 482     while (!fe->shader_is_end(fe_data, &ptr))
 483     {
 484         struct wined3d_shader_instruction ins;
 485         const char *comment;
 486         UINT comment_size;
 487
 488         /* Skip comments. */
 489         fe->shader_read_comment(&ptr, &comment, &comment_size);
 490         if (comment) continue;
 491
 492         /* Fetch opcode. */
 493         fe->shader_read_instruction(fe_data, &ptr, &ins);
 494
 495         /* Unhandled opcode, and its parameters. */
 496         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 497         {
 498             TRACE("Skipping unrecognized instruction.\n");
 499             continue;
 500         }
 501
 502         /* Handle declarations. */
 503         if (ins.handler_idx == WINED3DSIH_DCL)
 504         {
 505             struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
 506
 507             switch (semantic->reg.reg.type)
 508             {
 509                 /* Mark input registers used. */
 510                 case WINED3DSPR_INPUT:
 511                     reg_maps->input_registers |= 1 << semantic->reg.reg.idx;
 512                     shader_signature_from_semantic(&input_signature[semantic->reg.reg.idx], semantic);
 513                     break;
 514
 515                 /* Vertex shader: mark 3.0 output registers used, save token. */
 516                 case WINED3DSPR_OUTPUT:
 517                     reg_maps->output_registers |= 1 << semantic->reg.reg.idx;
 518                     shader_signature_from_semantic(&output_signature[semantic->reg.reg.idx], semantic);
 519                     if (semantic->usage == WINED3D_DECL_USAGE_FOG)
 520                         reg_maps->fog = 1;
 521                     break;
 522
 523                 /* Save sampler usage token. */
 524                 case WINED3DSPR_SAMPLER:
 525                     reg_maps->sampler_type[semantic->reg.reg.idx] = semantic->sampler_type;
 526                     break;
 527
 528                 default:
 529                     TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
 530                     break;
 531             }
 532         }
 533         else if (ins.handler_idx == WINED3DSIH_DEF)
 534         {
 535             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 536             if (!lconst) return E_OUTOFMEMORY;
 537
 538             lconst->idx = ins.dst[0].reg.idx;
 539             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 540
 541             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 542             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 543             {
 544                 float *value = (float *)lconst->value;
 545                 if (value[0] < -1.0f) value[0] = -1.0f;
 546                 else if (value[0] > 1.0f) value[0] = 1.0f;
 547                 if (value[1] < -1.0f) value[1] = -1.0f;
 548                 else if (value[1] > 1.0f) value[1] = 1.0f;
 549                 if (value[2] < -1.0f) value[2] = -1.0f;
 550                 else if (value[2] > 1.0f) value[2] = 1.0f;
 551                 if (value[3] < -1.0f) value[3] = -1.0f;
 552                 else if (value[3] > 1.0f) value[3] = 1.0f;
 553             }
 554
 555             list_add_head(&shader->constantsF, &lconst->entry);
 556         }
 557         else if (ins.handler_idx == WINED3DSIH_DEFI)
 558         {
 559             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 560             if (!lconst) return E_OUTOFMEMORY;
 561
 562             lconst->idx = ins.dst[0].reg.idx;
 563             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 564
 565             list_add_head(&shader->constantsI, &lconst->entry);
 566             reg_maps->local_int_consts |= (1 << lconst->idx);
 567         }
 568         else if (ins.handler_idx == WINED3DSIH_DEFB)
 569         {
 570             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 571             if (!lconst) return E_OUTOFMEMORY;
 572
 573             lconst->idx = ins.dst[0].reg.idx;
 574             memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
 575
 576             list_add_head(&shader->constantsB, &lconst->entry);
 577             reg_maps->local_bool_consts |= (1 << lconst->idx);
 578         }
 579         /* For subroutine prototypes. */
 580         else if (ins.handler_idx == WINED3DSIH_LABEL)
 581         {
 582             reg_maps->labels |= 1 << ins.src[0].reg.idx;
 583         }
 584         /* Set texture, address, temporary registers. */
 585         else
 586         {
 587             BOOL color0_mov = FALSE;
 588             unsigned int i;
 589
 590             /* This will loop over all the registers and try to
 591              * make a bitmask of the ones we're interested in.
 592              *
 593              * Relative addressing tokens are ignored, but that's
 594              * okay, since we'll catch any address registers when
 595              * they are initialized (required by spec). */
 596             for (i = 0; i < ins.dst_count; ++i)
 597             {
 598                 shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg, shader_version.type);
 599
 600                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 601                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 602                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 603                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 604                 {
 605                     UINT idx = ins.dst[i].reg.idx;
 606
 607                     switch (ins.dst[i].reg.type)
 608                     {
 609                         case WINED3DSPR_RASTOUT:
 610                             switch (idx)
 611                             {
 612                                 case 0: /* oPos */
 613                                     reg_maps->output_registers |= 1 << 10;
 614                                     shader_signature_from_usage(&output_signature[10],
 615                                             WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 616                                     break;
 617
 618                                 case 1: /* oFog */
 619                                     reg_maps->output_registers |= 1 << 11;
 620                                     shader_signature_from_usage(&output_signature[11],
 621                                             WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 622                                     break;
 623
 624                                 case 2: /* oPts */
 625                                     reg_maps->output_registers |= 1 << 11;
 626                                     shader_signature_from_usage(&output_signature[11],
 627                                             WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 628                                     break;
 629                             }
 630                             break;
 631
 632                         case WINED3DSPR_ATTROUT:
 633                             if (idx < 2)
 634                             {
 635                                 idx += 8;
 636                                 if (reg_maps->output_registers & (1 << idx))
 637                                 {
 638                                     output_signature[idx].mask |= ins.dst[i].write_mask;
 639                                 }
 640                                 else
 641                                 {
 642                                     reg_maps->output_registers |= 1 << idx;
 643                                     shader_signature_from_usage(&output_signature[idx],
 644                                             WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
 645                                 }
 646                             }
 647                             break;
 648
 649                         case WINED3DSPR_TEXCRDOUT:
 650
 651                             reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
 652                             if (reg_maps->output_registers & (1 << idx))
 653                             {
 654                                 output_signature[idx].mask |= ins.dst[i].write_mask;
 655                             }
 656                             else
 657                             {
 658                                 reg_maps->output_registers |= 1 << idx;
 659                                 shader_signature_from_usage(&output_signature[idx],
 660                                         WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
 661                             }
 662                             break;
 663
 664                         default:
 665                             break;
 666                     }
 667                 }
 668
 669                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 670                 {
 671                     if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx)
 672                     {
 673                         /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 674                          * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 675                          * the mov and perform the sRGB write correction from the source register.
 676                          *
 677                          * However, if the mov is only partial, we can't do this, and if the write
 678                          * comes from an instruction other than MOV it is hard to do as well. If
 679                          * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 680                         shader->u.ps.color0_mov = FALSE;
 681                         if (ins.handler_idx == WINED3DSIH_MOV
 682                                 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
 683                         {
 684                             /* Used later when the source register is read. */
 685                             color0_mov = TRUE;
 686                         }
 687                     }
 688                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 689                      * end
 690                      */
 691                     else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
 692                             && ins.dst[i].reg.idx == shader->u.ps.color0_reg)
 693                     {
 694                         shader->u.ps.color0_mov = FALSE;
 695                     }
 696                 }
 697
 698                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 699                 if (shader_version.major == 1
 700                         && (ins.handler_idx == WINED3DSIH_TEX
 701                             || ins.handler_idx == WINED3DSIH_TEXBEM
 702                             || ins.handler_idx == WINED3DSIH_TEXBEML
 703                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 704                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 705                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 706                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 707                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 708                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 709                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 710                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 711                 {
 712                     /* Fake sampler usage, only set reserved bit and type. */
 713                     DWORD sampler_code = ins.dst[i].reg.idx;
 714
 715                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 716                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 717
 718                     /* texbem is only valid with < 1.4 pixel shaders */
 719                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 720                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 721                     {
 722                         reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx;
 723                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 724                         {
 725                             reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx;
 726                         }
 727                     }
 728                 }
 729                 else if (ins.handler_idx == WINED3DSIH_BEM)
 730                 {
 731                     reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx;
 732                 }
 733             }
 734
 735             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 736             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 737             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 738             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 739             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 740             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 741             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 742             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 743             else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
 744             else if (ins.handler_idx == WINED3DSIH_LOOP
 745                     || ins.handler_idx == WINED3DSIH_REP)
 746             {
 747                 ++cur_loop_depth;
 748                 if (cur_loop_depth > max_loop_depth)
 749                     max_loop_depth = cur_loop_depth;
 750             }
 751             else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 752                     || ins.handler_idx == WINED3DSIH_ENDREP)
 753                 --cur_loop_depth;
 754
 755             if (ins.predicate)
 756                 shader_record_register_usage(shader, reg_maps, &ins.predicate->reg, shader_version.type);
 757
 758             for (i = 0; i < ins.src_count; ++i)
 759             {
 760                 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
 761                 struct wined3d_shader_register reg = ins.src[i].reg;
 762
 763                 shader_record_register_usage(shader, reg_maps, &ins.src[i].reg, shader_version.type);
 764                 while (count)
 765                 {
 766                     ++reg.idx;
 767                     shader_record_register_usage(shader, reg_maps, &reg, shader_version.type);
 768                     --count;
 769                 }
 770
 771                 if (color0_mov)
 772                 {
 773                     if (ins.src[i].reg.type == WINED3DSPR_TEMP
 774                             && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
 775                     {
 776                         shader->u.ps.color0_mov = TRUE;
 777                         shader->u.ps.color0_reg = ins.src[i].reg.idx;
 778                     }
 779                 }
 780             }
 781         }
 782     }
 783     reg_maps->loop_depth = max_loop_depth;
 784
 785     /* PS before 2.0 don't have explicit color outputs. Instead the value of
 786      * R0 is written to the render target. */
 787     if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 788         reg_maps->rt_mask |= (1 << 0);
 789
 790     shader->functionLength = ((const char *)ptr - (const char *)byte_code);
 791
 792     return WINED3D_OK;
 793 }
 794
 795 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 796 {
 797     DWORD map = 1 << max;
 798     map |= map - 1;
 799     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 800
 801     return wined3d_log2i(map);
 802 }
 803
 804 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 805         const struct wined3d_shader_version *shader_version)
 806 {
 807     TRACE("dcl");
 808
 809     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 810     {
 811         switch (semantic->sampler_type)
 812         {
 813             case WINED3DSTT_2D: TRACE("_2d"); break;
 814             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 815             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 816             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 817         }
 818     }
 819     else
 820     {
 821         /* Pixel shaders 3.0 don't have usage semantics. */
 822         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 823         else TRACE("_");
 824
 825         switch (semantic->usage)
 826         {
 827             case WINED3D_DECL_USAGE_POSITION:
 828                 TRACE("position%u", semantic->usage_idx);
 829                 break;
 830
 831             case WINED3D_DECL_USAGE_BLEND_INDICES:
 832                 TRACE("blend");
 833                 break;
 834
 835             case WINED3D_DECL_USAGE_BLEND_WEIGHT:
 836                 TRACE("weight");
 837                 break;
 838
 839             case WINED3D_DECL_USAGE_NORMAL:
 840                 TRACE("normal%u", semantic->usage_idx);
 841                 break;
 842
 843             case WINED3D_DECL_USAGE_PSIZE:
 844                 TRACE("psize");
 845                 break;
 846
 847             case WINED3D_DECL_USAGE_COLOR:
 848                 if (!semantic->usage_idx) TRACE("color");
 849                 else TRACE("specular%u", (semantic->usage_idx - 1));
 850                 break;
 851
 852             case WINED3D_DECL_USAGE_TEXCOORD:
 853                 TRACE("texture%u", semantic->usage_idx);
 854                 break;
 855
 856             case WINED3D_DECL_USAGE_TANGENT:
 857                 TRACE("tangent");
 858                 break;
 859
 860             case WINED3D_DECL_USAGE_BINORMAL:
 861                 TRACE("binormal");
 862                 break;
 863
 864             case WINED3D_DECL_USAGE_TESS_FACTOR:
 865                 TRACE("tessfactor");
 866                 break;
 867
 868             case WINED3D_DECL_USAGE_POSITIONT:
 869                 TRACE("positionT%u", semantic->usage_idx);
 870                 break;
 871
 872             case WINED3D_DECL_USAGE_FOG:
 873                 TRACE("fog");
 874                 break;
 875
 876             case WINED3D_DECL_USAGE_DEPTH:
 877                 TRACE("depth");
 878                 break;
 879
 880             case WINED3D_DECL_USAGE_SAMPLE:
 881                 TRACE("sample");
 882                 break;
 883
 884             default:
 885                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 886         }
 887     }
 888 }
 889
 890 static void shader_dump_register(const struct wined3d_shader_register *reg,
 891         const struct wined3d_shader_version *shader_version)
 892 {
 893     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 894     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 895     UINT offset = reg->idx;
 896
 897     switch (reg->type)
 898     {
 899         case WINED3DSPR_TEMP:
 900             TRACE("r");
 901             break;
 902
 903         case WINED3DSPR_INPUT:
 904             TRACE("v");
 905             break;
 906
 907         case WINED3DSPR_CONST:
 908         case WINED3DSPR_CONST2:
 909         case WINED3DSPR_CONST3:
 910         case WINED3DSPR_CONST4:
 911             TRACE("c");
 912             offset = shader_get_float_offset(reg->type, reg->idx);
 913             break;
 914
 915         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 916             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 917             break;
 918
 919         case WINED3DSPR_RASTOUT:
 920             TRACE("%s", rastout_reg_names[reg->idx]);
 921             break;
 922
 923         case WINED3DSPR_COLOROUT:
 924             TRACE("oC");
 925             break;
 926
 927         case WINED3DSPR_DEPTHOUT:
 928             TRACE("oDepth");
 929             break;
 930
 931         case WINED3DSPR_ATTROUT:
 932             TRACE("oD");
 933             break;
 934
 935         case WINED3DSPR_TEXCRDOUT:
 936             /* Vertex shaders >= 3.0 use general purpose output registers
 937              * (WINED3DSPR_OUTPUT), which can include an address token. */
 938             if (shader_version->major >= 3) TRACE("o");
 939             else TRACE("oT");
 940             break;
 941
 942         case WINED3DSPR_CONSTINT:
 943             TRACE("i");
 944             break;
 945
 946         case WINED3DSPR_CONSTBOOL:
 947             TRACE("b");
 948             break;
 949
 950         case WINED3DSPR_LABEL:
 951             TRACE("l");
 952             break;
 953
 954         case WINED3DSPR_LOOP:
 955             TRACE("aL");
 956             break;
 957
 958         case WINED3DSPR_SAMPLER:
 959             TRACE("s");
 960             break;
 961
 962         case WINED3DSPR_MISCTYPE:
 963             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
 964             else TRACE("%s", misctype_reg_names[reg->idx]);
 965             break;
 966
 967         case WINED3DSPR_PREDICATE:
 968             TRACE("p");
 969             break;
 970
 971         case WINED3DSPR_IMMCONST:
 972             TRACE("l");
 973             break;
 974
 975         case WINED3DSPR_CONSTBUFFER:
 976             TRACE("cb");
 977             break;
 978
 979         case WINED3DSPR_NULL:
 980             TRACE("null");
 981             break;
 982
 983         case WINED3DSPR_RESOURCE:
 984             TRACE("t");
 985             break;
 986
 987         default:
 988             TRACE("unhandled_rtype(%#x)", reg->type);
 989             break;
 990     }
 991
 992     if (reg->type == WINED3DSPR_IMMCONST)
 993     {
 994         TRACE("(");
 995         switch (reg->immconst_type)
 996         {
 997             case WINED3D_IMMCONST_SCALAR:
 998                 switch (reg->data_type)
 999                 {
1000                     case WINED3D_DATA_FLOAT:
1001                         TRACE("%.8e", *(const float *)reg->immconst_data);
1002                         break;
1003                     case WINED3D_DATA_INT:
1004                         TRACE("%d", reg->immconst_data[0]);
1005                         break;
1006                     case WINED3D_DATA_RESOURCE:
1007                     case WINED3D_DATA_SAMPLER:
1008                     case WINED3D_DATA_UINT:
1009                         TRACE("%u", reg->immconst_data[0]);
1010                         break;
1011                     default:
1012                         TRACE("<unhandled data type %#x>", reg->data_type);
1013                         break;
1014                 }
1015                 break;
1016
1017             case WINED3D_IMMCONST_VEC4:
1018                 switch (reg->data_type)
1019                 {
1020                     case WINED3D_DATA_FLOAT:
1021                         TRACE("%.8e, %.8e, %.8e, %.8e",
1022                                 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1023                                 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1024                         break;
1025                     case WINED3D_DATA_INT:
1026                         TRACE("%d, %d, %d, %d",
1027                                 reg->immconst_data[0], reg->immconst_data[1],
1028                                 reg->immconst_data[2], reg->immconst_data[3]);
1029                         break;
1030                     case WINED3D_DATA_RESOURCE:
1031                     case WINED3D_DATA_SAMPLER:
1032                     case WINED3D_DATA_UINT:
1033                         TRACE("%u, %u, %u, %u",
1034                                 reg->immconst_data[0], reg->immconst_data[1],
1035                                 reg->immconst_data[2], reg->immconst_data[3]);
1036                         break;
1037                     default:
1038                         TRACE("<unhandled data type %#x>\n", reg->data_type);
1039                         break;
1040                 }
1041                 break;
1042
1043             default:
1044                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1045                 break;
1046         }
1047         TRACE(")");
1048     }
1049     else if (reg->type != WINED3DSPR_RASTOUT
1050             && reg->type != WINED3DSPR_MISCTYPE
1051             && reg->type != WINED3DSPR_NULL)
1052     {
1053         if (reg->array_idx != ~0U)
1054         {
1055             TRACE("%u[%u", offset, reg->array_idx);
1056             if (reg->rel_addr)
1057             {
1058                 TRACE(" + ");
1059                 shader_dump_src_param(reg->rel_addr, shader_version);
1060             }
1061             TRACE("]");
1062         }
1063         else
1064         {
1065             if (reg->rel_addr)
1066             {
1067                 TRACE("[");
1068                 shader_dump_src_param(reg->rel_addr, shader_version);
1069                 TRACE(" + ");
1070             }
1071             TRACE("%u", offset);
1072             if (reg->rel_addr) TRACE("]");
1073         }
1074     }
1075 }
1076
1077 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1078         const struct wined3d_shader_version *shader_version)
1079 {
1080     DWORD write_mask = param->write_mask;
1081
1082     shader_dump_register(&param->reg, shader_version);
1083
1084     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1085     {
1086         static const char *write_mask_chars = "xyzw";
1087
1088         TRACE(".");
1089         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1090         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1091         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1092         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1093     }
1094 }
1095
1096 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1097         const struct wined3d_shader_version *shader_version)
1098 {
1099     enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1100     DWORD swizzle = param->swizzle;
1101
1102     if (src_modifier == WINED3DSPSM_NEG
1103             || src_modifier == WINED3DSPSM_BIASNEG
1104             || src_modifier == WINED3DSPSM_SIGNNEG
1105             || src_modifier == WINED3DSPSM_X2NEG
1106             || src_modifier == WINED3DSPSM_ABSNEG)
1107         TRACE("-");
1108     else if (src_modifier == WINED3DSPSM_COMP)
1109         TRACE("1-");
1110     else if (src_modifier == WINED3DSPSM_NOT)
1111         TRACE("!");
1112
1113     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1114         TRACE("abs(");
1115
1116     shader_dump_register(&param->reg, shader_version);
1117
1118     if (src_modifier)
1119     {
1120         switch (src_modifier)
1121         {
1122             case WINED3DSPSM_NONE:    break;
1123             case WINED3DSPSM_NEG:     break;
1124             case WINED3DSPSM_NOT:     break;
1125             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1126             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1127             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1128             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1129             case WINED3DSPSM_COMP:    break;
1130             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1131             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1132             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1133             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1134             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1135             case WINED3DSPSM_ABS:     TRACE(")"); break;
1136             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1137         }
1138     }
1139
1140     if (swizzle != WINED3DSP_NOSWIZZLE)
1141     {
1142         static const char *swizzle_chars = "xyzw";
1143         DWORD swizzle_x = swizzle & 0x03;
1144         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1145         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1146         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1147
1148         if (swizzle_x == swizzle_y
1149                 && swizzle_x == swizzle_z
1150                 && swizzle_x == swizzle_w)
1151         {
1152             TRACE(".%c", swizzle_chars[swizzle_x]);
1153         }
1154         else
1155         {
1156             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1157                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1158         }
1159     }
1160 }
1161
1162 /* Shared code in order to generate the bulk of the shader string.
1163  * NOTE: A description of how to parse tokens can be found on MSDN. */
1164 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1165         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1166 {
1167     struct wined3d_device *device = shader->device;
1168     const struct wined3d_shader_frontend *fe = shader->frontend;
1169     void *fe_data = shader->frontend_data;
1170     struct wined3d_shader_version shader_version;
1171     struct wined3d_shader_loop_state loop_state;
1172     struct wined3d_shader_instruction ins;
1173     struct wined3d_shader_tex_mx tex_mx;
1174     struct wined3d_shader_context ctx;
1175     const DWORD *ptr = byte_code;
1176
1177     /* Initialize current parsing state. */
1178     tex_mx.current_row = 0;
1179     loop_state.current_depth = 0;
1180     loop_state.current_reg = 0;
1181
1182     ctx.shader = shader;
1183     ctx.gl_info = &device->adapter->gl_info;
1184     ctx.reg_maps = reg_maps;
1185     ctx.buffer = buffer;
1186     ctx.tex_mx = &tex_mx;
1187     ctx.loop_state = &loop_state;
1188     ctx.backend_data = backend_ctx;
1189     ins.ctx = &ctx;
1190
1191     fe->shader_read_header(fe_data, &ptr, &shader_version);
1192
1193     while (!fe->shader_is_end(fe_data, &ptr))
1194     {
1195         const char *comment;
1196         UINT comment_size;
1197
1198         /* Skip comment tokens. */
1199         fe->shader_read_comment(&ptr, &comment, &comment_size);
1200         if (comment) continue;
1201
1202         /* Read opcode. */
1203         fe->shader_read_instruction(fe_data, &ptr, &ins);
1204
1205         /* Unknown opcode and its parameters. */
1206         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1207         {
1208             TRACE("Skipping unrecognized instruction.\n");
1209             continue;
1210         }
1211
1212         if (ins.predicate)
1213             FIXME("Predicates not implemented.\n");
1214
1215         /* Call appropriate function for output target */
1216         device->shader_backend->shader_handle_instruction(&ins);
1217     }
1218 }
1219
1220 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1221 {
1222     DWORD mmask = dst->modifiers;
1223
1224     switch (dst->shift)
1225     {
1226         case 0: break;
1227         case 13: TRACE("_d8"); break;
1228         case 14: TRACE("_d4"); break;
1229         case 15: TRACE("_d2"); break;
1230         case 1: TRACE("_x2"); break;
1231         case 2: TRACE("_x4"); break;
1232         case 3: TRACE("_x8"); break;
1233         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1234     }
1235
1236     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1237     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1238     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1239
1240     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1241     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1242 }
1243
1244 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1245 {
1246     struct wined3d_shader_version shader_version;
1247     const DWORD *ptr = byte_code;
1248     const char *type_prefix;
1249     DWORD i;
1250
1251     TRACE("Parsing %p.\n", byte_code);
1252
1253     fe->shader_read_header(fe_data, &ptr, &shader_version);
1254
1255     switch (shader_version.type)
1256     {
1257         case WINED3D_SHADER_TYPE_VERTEX:
1258             type_prefix = "vs";
1259             break;
1260
1261         case WINED3D_SHADER_TYPE_GEOMETRY:
1262             type_prefix = "gs";
1263             break;
1264
1265         case WINED3D_SHADER_TYPE_PIXEL:
1266             type_prefix = "ps";
1267             break;
1268
1269         default:
1270             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1271             type_prefix = "unknown";
1272             break;
1273     }
1274
1275     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1276
1277     while (!fe->shader_is_end(fe_data, &ptr))
1278     {
1279         struct wined3d_shader_instruction ins;
1280         const char *comment;
1281         UINT comment_size;
1282
1283         /* comment */
1284         fe->shader_read_comment(&ptr, &comment, &comment_size);
1285         if (comment)
1286         {
1287             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1288             {
1289                 const char *end = comment + comment_size;
1290                 const char *ptr = comment + 4;
1291                 const char *line = ptr;
1292
1293                 TRACE("// TEXT\n");
1294                 while (ptr != end)
1295                 {
1296                     if (*ptr == '\n')
1297                     {
1298                         UINT len = ptr - line;
1299                         if (len && *(ptr - 1) == '\r') --len;
1300                         TRACE("// %s\n", debugstr_an(line, len));
1301                         line = ++ptr;
1302                     }
1303                     else ++ptr;
1304                 }
1305                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1306             }
1307             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1308             continue;
1309         }
1310
1311         fe->shader_read_instruction(fe_data, &ptr, &ins);
1312         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1313         {
1314             TRACE("Skipping unrecognized instruction.\n");
1315             continue;
1316         }
1317
1318         if (ins.handler_idx == WINED3DSIH_DCL)
1319         {
1320             shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1321             shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1322             TRACE(" ");
1323             shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1324         }
1325         else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1326         {
1327             TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1328         }
1329         else if (ins.handler_idx == WINED3DSIH_DEF)
1330         {
1331             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type, ins.dst[0].reg.idx),
1332                     *(const float *)&ins.src[0].reg.immconst_data[0],
1333                     *(const float *)&ins.src[0].reg.immconst_data[1],
1334                     *(const float *)&ins.src[0].reg.immconst_data[2],
1335                     *(const float *)&ins.src[0].reg.immconst_data[3]);
1336         }
1337         else if (ins.handler_idx == WINED3DSIH_DEFI)
1338         {
1339             TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx,
1340                     ins.src[0].reg.immconst_data[0],
1341                     ins.src[0].reg.immconst_data[1],
1342                     ins.src[0].reg.immconst_data[2],
1343                     ins.src[0].reg.immconst_data[3]);
1344         }
1345         else if (ins.handler_idx == WINED3DSIH_DEFB)
1346         {
1347             TRACE("defb b%u = %s", ins.dst[0].reg.idx, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1348         }
1349         else
1350         {
1351             if (ins.predicate)
1352             {
1353                 TRACE("(");
1354                 shader_dump_src_param(ins.predicate, &shader_version);
1355                 TRACE(") ");
1356             }
1357
1358             /* PixWin marks instructions with the coissue flag with a '+' */
1359             if (ins.coissue) TRACE("+");
1360
1361             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1362
1363             if (ins.handler_idx == WINED3DSIH_IFC
1364                     || ins.handler_idx == WINED3DSIH_BREAKC)
1365             {
1366                 switch (ins.flags)
1367                 {
1368                     case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1369                     case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1370                     case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1371                     case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1372                     case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1373                     case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1374                     default: TRACE("_(%u)", ins.flags);
1375                 }
1376             }
1377             else if (ins.handler_idx == WINED3DSIH_TEX
1378                     && shader_version.major >= 2
1379                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1380             {
1381                 TRACE("p");
1382             }
1383
1384             for (i = 0; i < ins.dst_count; ++i)
1385             {
1386                 shader_dump_ins_modifiers(&ins.dst[i]);
1387                 TRACE(!i ? " " : ", ");
1388                 shader_dump_dst_param(&ins.dst[i], &shader_version);
1389             }
1390
1391             /* Other source tokens */
1392             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1393             {
1394                 TRACE(!i ? " " : ", ");
1395                 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1396             }
1397         }
1398         TRACE("\n");
1399     }
1400 }
1401
1402 static void shader_cleanup(struct wined3d_shader *shader)
1403 {
1404     shader->device->shader_backend->shader_destroy(shader);
1405     HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1406     HeapFree(GetProcessHeap(), 0, shader->function);
1407     shader_delete_constant_list(&shader->constantsF);
1408     shader_delete_constant_list(&shader->constantsB);
1409     shader_delete_constant_list(&shader->constantsI);
1410     list_remove(&shader->shader_list_entry);
1411
1412     if (shader->frontend && shader->frontend_data)
1413         shader->frontend->shader_free(shader->frontend_data);
1414 }
1415
1416 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1417 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1418 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1419         enum tex_types tex_type, const SIZE *ds_mask_size) {}
1420 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1421 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1422 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1423 static void shader_none_load_constants(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1424 static void shader_none_load_np2fixup_constants(void *shader_priv,
1425         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1426 static void shader_none_destroy(struct wined3d_shader *shader) {}
1427 static HRESULT shader_none_alloc(struct wined3d_device *device) {return WINED3D_OK;}
1428 static void shader_none_free(struct wined3d_device *device) {}
1429 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1430
1431 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1432 {
1433     /* Set the shader caps to 0 for the none shader backend */
1434     caps->VertexShaderVersion = 0;
1435     caps->MaxVertexShaderConst = 0;
1436     caps->PixelShaderVersion = 0;
1437     caps->PixelShader1xMaxValue = 0.0f;
1438     caps->MaxPixelShaderConst = 0;
1439     caps->VSClipping = FALSE;
1440 }
1441
1442 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1443 {
1444     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1445     {
1446         TRACE("Checking support for fixup:\n");
1447         dump_color_fixup_desc(fixup);
1448     }
1449
1450     /* Faked to make some apps happy. */
1451     if (!is_complex_fixup(fixup))
1452     {
1453         TRACE("[OK]\n");
1454         return TRUE;
1455     }
1456
1457     TRACE("[FAILED]\n");
1458     return FALSE;
1459 }
1460
1461 const struct wined3d_shader_backend_ops none_shader_backend =
1462 {
1463     shader_none_handle_instruction,
1464     shader_none_select,
1465     shader_none_select_depth_blt,
1466     shader_none_deselect_depth_blt,
1467     shader_none_update_float_vertex_constants,
1468     shader_none_update_float_pixel_constants,
1469     shader_none_load_constants,
1470     shader_none_load_np2fixup_constants,
1471     shader_none_destroy,
1472     shader_none_alloc,
1473     shader_none_free,
1474     shader_none_context_destroyed,
1475     shader_none_get_caps,
1476     shader_none_color_fixup_supported,
1477 };
1478
1479 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1480         const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1481         enum wined3d_shader_type type, unsigned int max_version)
1482 {
1483     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1484     const struct wined3d_shader_frontend *fe;
1485     HRESULT hr;
1486     unsigned int backend_version;
1487
1488     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1489             shader, byte_code, output_signature, float_const_count);
1490
1491     fe = shader_select_frontend(*byte_code);
1492     if (!fe)
1493     {
1494         FIXME("Unable to find frontend for shader.\n");
1495         return WINED3DERR_INVALIDCALL;
1496     }
1497     shader->frontend = fe;
1498     shader->frontend_data = fe->shader_init(byte_code, output_signature);
1499     if (!shader->frontend_data)
1500     {
1501         FIXME("Failed to initialize frontend.\n");
1502         return WINED3DERR_INVALIDCALL;
1503     }
1504
1505     /* First pass: trace shader. */
1506     if (TRACE_ON(d3d_shader))
1507         shader_trace_init(fe, shader->frontend_data, byte_code);
1508
1509     /* Initialize immediate constant lists. */
1510     list_init(&shader->constantsF);
1511     list_init(&shader->constantsB);
1512     list_init(&shader->constantsI);
1513
1514     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1515     hr = shader_get_registers_used(shader, fe,
1516             reg_maps, shader->input_signature, shader->output_signature,
1517             byte_code, float_const_count);
1518     if (FAILED(hr)) return hr;
1519
1520     if (reg_maps->shader_version.type != type)
1521     {
1522         WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1523         return WINED3DERR_INVALIDCALL;
1524     }
1525     if (reg_maps->shader_version.major > max_version)
1526     {
1527         WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1528         return WINED3DERR_INVALIDCALL;
1529     }
1530     switch (type)
1531     {
1532         case WINED3D_SHADER_TYPE_VERTEX:
1533             backend_version = shader->device->vshader_version;
1534             break;
1535         case WINED3D_SHADER_TYPE_PIXEL:
1536             backend_version = shader->device->pshader_version;
1537             break;
1538         default:
1539             FIXME("No backend version-checking for this shader type\n");
1540             backend_version = 0;
1541     }
1542     if (reg_maps->shader_version.major > backend_version)
1543     {
1544         WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1545                 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1546         return WINED3DERR_INVALIDCALL;
1547     }
1548
1549     shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1550     if (!shader->function)
1551         return E_OUTOFMEMORY;
1552     memcpy(shader->function, byte_code, shader->functionLength);
1553
1554     return WINED3D_OK;
1555 }
1556
1557 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1558 {
1559     ULONG refcount = InterlockedIncrement(&shader->ref);
1560
1561     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1562
1563     return refcount;
1564 }
1565
1566 /* Do not call while under the GL lock. */
1567 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1568 {
1569     ULONG refcount = InterlockedDecrement(&shader->ref);
1570
1571     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1572
1573     if (!refcount)
1574     {
1575         shader_cleanup(shader);
1576         shader->parent_ops->wined3d_object_destroyed(shader->parent);
1577         HeapFree(GetProcessHeap(), 0, shader);
1578     }
1579
1580     return refcount;
1581 }
1582
1583 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1584 {
1585     TRACE("shader %p.\n", shader);
1586
1587     return shader->parent;
1588 }
1589
1590 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1591         void *byte_code, UINT *byte_code_size)
1592 {
1593     TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1594
1595     if (!byte_code)
1596     {
1597         *byte_code_size = shader->functionLength;
1598         return WINED3D_OK;
1599     }
1600
1601     if (*byte_code_size < shader->functionLength)
1602     {
1603         /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1604          * than the required size we should write the required size and
1605          * return D3DERR_MOREDATA. That's not actually true. */
1606         return WINED3DERR_INVALIDCALL;
1607     }
1608
1609     memcpy(byte_code, shader->function, shader->functionLength);
1610
1611     return WINED3D_OK;
1612 }
1613
1614 /* Set local constants for d3d8 shaders. */
1615 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1616         UINT start_idx, const float *src_data, UINT count)
1617 {
1618     UINT end_idx = start_idx + count;
1619     UINT i;
1620
1621     TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1622
1623     if (end_idx > shader->limits.constant_float)
1624     {
1625         WARN("end_idx %u > float constants limit %u.\n",
1626                 end_idx, shader->limits.constant_float);
1627         end_idx = shader->limits.constant_float;
1628     }
1629
1630     for (i = start_idx; i < end_idx; ++i)
1631     {
1632         struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1633         if (!lconst)
1634             return E_OUTOFMEMORY;
1635
1636         lconst->idx = i;
1637         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1638         list_add_head(&shader->constantsF, &lconst->entry);
1639     }
1640
1641     return WINED3D_OK;
1642 }
1643
1644 void find_vs_compile_args(const struct wined3d_state *state,
1645         const struct wined3d_shader *shader, struct vs_compile_args *args)
1646 {
1647     args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1648             == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1649     args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1650             && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1651     args->swizzle_map = shader->device->strided_streams.swizzle_map;
1652 }
1653
1654 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1655 {
1656     if (usage_idx1 != usage_idx2)
1657         return FALSE;
1658     if (usage1 == usage2)
1659         return TRUE;
1660     if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1661         return TRUE;
1662     if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1663         return TRUE;
1664
1665     return FALSE;
1666 }
1667
1668 BOOL vshader_get_input(const struct wined3d_shader *shader,
1669         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1670 {
1671     WORD map = shader->reg_maps.input_registers;
1672     unsigned int i;
1673
1674     for (i = 0; map; map >>= 1, ++i)
1675     {
1676         if (!(map & 1)) continue;
1677
1678         if (match_usage(shader->u.vs.attributes[i].usage,
1679                 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1680         {
1681             *regnum = i;
1682             return TRUE;
1683         }
1684     }
1685     return FALSE;
1686 }
1687
1688 static void vertexshader_set_limits(struct wined3d_shader *shader)
1689 {
1690     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1691             shader->reg_maps.shader_version.minor);
1692     struct wined3d_device *device = shader->device;
1693
1694     shader->limits.texcoord = 0;
1695     shader->limits.attributes = 16;
1696     shader->limits.packed_input = 0;
1697
1698     switch (shader_version)
1699     {
1700         case WINED3D_SHADER_VERSION(1, 0):
1701         case WINED3D_SHADER_VERSION(1, 1):
1702             shader->limits.temporary = 12;
1703             shader->limits.constant_bool = 0;
1704             shader->limits.constant_int = 0;
1705             shader->limits.address = 1;
1706             shader->limits.packed_output = 12;
1707             shader->limits.sampler = 0;
1708             shader->limits.label = 0;
1709             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1710              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1711              * constants? */
1712             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1713             break;
1714
1715         case WINED3D_SHADER_VERSION(2, 0):
1716         case WINED3D_SHADER_VERSION(2, 1):
1717             shader->limits.temporary = 12;
1718             shader->limits.constant_bool = 16;
1719             shader->limits.constant_int = 16;
1720             shader->limits.address = 1;
1721             shader->limits.packed_output = 12;
1722             shader->limits.sampler = 0;
1723             shader->limits.label = 16;
1724             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1725             break;
1726
1727         case WINED3D_SHADER_VERSION(4, 0):
1728             FIXME("Using 3.0 limits for 4.0 shader.\n");
1729             /* Fall through. */
1730
1731         case WINED3D_SHADER_VERSION(3, 0):
1732             shader->limits.temporary = 32;
1733             shader->limits.constant_bool = 32;
1734             shader->limits.constant_int = 32;
1735             shader->limits.address = 1;
1736             shader->limits.packed_output = 12;
1737             shader->limits.sampler = 4;
1738             shader->limits.label = 16; /* FIXME: 2048 */
1739             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1740              * even though they are capable of supporting much more (GL
1741              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1742              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1743              * shaders to 256. */
1744             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1745             break;
1746
1747         default:
1748             shader->limits.temporary = 12;
1749             shader->limits.constant_bool = 16;
1750             shader->limits.constant_int = 16;
1751             shader->limits.address = 1;
1752             shader->limits.packed_output = 12;
1753             shader->limits.sampler = 0;
1754             shader->limits.label = 16;
1755             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1756             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1757                     shader->reg_maps.shader_version.major,
1758                     shader->reg_maps.shader_version.minor);
1759     }
1760 }
1761
1762 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1763         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1764         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1765 {
1766     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1767     unsigned int i;
1768     HRESULT hr;
1769     WORD map;
1770
1771     if (!byte_code) return WINED3DERR_INVALIDCALL;
1772
1773     shader_init(shader, device, parent, parent_ops);
1774     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1775             WINED3D_SHADER_TYPE_VERTEX, max_version);
1776     if (FAILED(hr))
1777     {
1778         WARN("Failed to set function, hr %#x.\n", hr);
1779         shader_cleanup(shader);
1780         return hr;
1781     }
1782
1783     map = reg_maps->input_registers;
1784     for (i = 0; map; map >>= 1, ++i)
1785     {
1786         if (!(map & 1) || !shader->input_signature[i].semantic_name)
1787             continue;
1788
1789         shader->u.vs.attributes[i].usage =
1790                 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1791         shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1792     }
1793
1794     if (output_signature)
1795     {
1796         for (i = 0; i < output_signature->element_count; ++i)
1797         {
1798             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1799             reg_maps->output_registers |= 1 << e->register_idx;
1800             shader->output_signature[e->register_idx] = *e;
1801         }
1802     }
1803
1804     vertexshader_set_limits(shader);
1805
1806     shader->load_local_constsF = reg_maps->usesrelconstF
1807             && !list_empty(&shader->constantsF);
1808
1809     return WINED3D_OK;
1810 }
1811
1812 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1813         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1814         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1815 {
1816     HRESULT hr;
1817
1818     shader_init(shader, device, parent, parent_ops);
1819     hr = shader_set_function(shader, byte_code, output_signature, 0,
1820             WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1821     if (FAILED(hr))
1822     {
1823         WARN("Failed to set function, hr %#x.\n", hr);
1824         shader_cleanup(shader);
1825         return hr;
1826     }
1827
1828     shader->load_local_constsF = FALSE;
1829
1830     return WINED3D_OK;
1831 }
1832
1833 void find_ps_compile_args(const struct wined3d_state *state,
1834         const struct wined3d_shader *shader, struct ps_compile_args *args)
1835 {
1836     struct wined3d_device *device = shader->device;
1837     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1838     const struct wined3d_texture *texture;
1839     UINT i;
1840
1841     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1842     if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1843     {
1844         const struct wined3d_surface *rt = state->fb->render_targets[0];
1845         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
1846         {
1847             static unsigned int warned = 0;
1848
1849             args->srgb_correction = 1;
1850             if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
1851                 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
1852                         "support, expect rendering artifacts.\n");
1853         }
1854     }
1855
1856     if (shader->reg_maps.shader_version.major == 1
1857             && shader->reg_maps.shader_version.minor <= 3)
1858     {
1859         for (i = 0; i < 4; ++i)
1860         {
1861             DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
1862
1863             if (flags & WINED3D_TTFF_PROJECTED)
1864             {
1865                 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
1866
1867                 if (!state->vertex_shader)
1868                 {
1869                     unsigned int j;
1870                     unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
1871                     DWORD max_valid = WINED3D_TTFF_COUNT4;
1872                     enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
1873
1874                     for (j = 0; j < state->vertex_declaration->element_count; ++j)
1875                     {
1876                         struct wined3d_vertex_declaration_element *element =
1877                                 &state->vertex_declaration->elements[j];
1878
1879                         if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
1880                                 && element->usage_idx == index)
1881                         {
1882                             max_valid = element->format->component_count;
1883                             break;
1884                         }
1885                     }
1886                     if (!tex_transform || tex_transform > max_valid)
1887                     {
1888                         WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
1889                                 tex_transform, max_valid);
1890                         tex_transform = max_valid;
1891                     }
1892                     if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
1893                             || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
1894                             || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
1895                         tex_transform |= WINED3D_PSARGS_PROJECTED;
1896                     else
1897                     {
1898                         WARN("Application requested projected texture with unsuitable texture coordinates.\n");
1899                         WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
1900                                 i, tex_transform, sampler_type);
1901                     }
1902                 }
1903                 else
1904                     tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
1905
1906                 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
1907             }
1908         }
1909     }
1910
1911     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1912     {
1913         if (!shader->reg_maps.sampler_type[i])
1914             continue;
1915
1916         texture = state->textures[i];
1917         if (!texture)
1918         {
1919             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1920             continue;
1921         }
1922         args->color_fixup[i] = texture->resource.format->color_fixup;
1923
1924         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
1925             args->shadow |= 1 << i;
1926
1927         /* Flag samplers that need NP2 texcoord fixup. */
1928         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
1929             args->np2_fixup |= (1 << i);
1930     }
1931     if (shader->reg_maps.shader_version.major >= 3)
1932     {
1933         if (device->strided_streams.position_transformed)
1934         {
1935             args->vp_mode = pretransformed;
1936         }
1937         else if (use_vs(state))
1938         {
1939             args->vp_mode = vertexshader;
1940         }
1941         else
1942         {
1943             args->vp_mode = fixedfunction;
1944         }
1945         args->fog = FOG_OFF;
1946     }
1947     else
1948     {
1949         args->vp_mode = vertexshader;
1950         if (state->render_states[WINED3D_RS_FOGENABLE])
1951         {
1952             switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
1953             {
1954                 case WINED3D_FOG_NONE:
1955                     if (device->strided_streams.position_transformed || use_vs(state))
1956                     {
1957                         args->fog = FOG_LINEAR;
1958                         break;
1959                     }
1960
1961                     switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
1962                     {
1963                         case WINED3D_FOG_NONE: /* Fall through. */
1964                         case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
1965                         case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
1966                         case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
1967                     }
1968                     break;
1969
1970                 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
1971                 case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
1972                 case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
1973             }
1974         }
1975         else
1976         {
1977             args->fog = FOG_OFF;
1978         }
1979     }
1980 }
1981
1982 static void pixelshader_set_limits(struct wined3d_shader *shader)
1983 {
1984     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1985             shader->reg_maps.shader_version.minor);
1986
1987     shader->limits.attributes = 0;
1988     shader->limits.address = 0;
1989     shader->limits.packed_output = 0;
1990
1991     switch (shader_version)
1992     {
1993         case WINED3D_SHADER_VERSION(1, 0):
1994         case WINED3D_SHADER_VERSION(1, 1):
1995         case WINED3D_SHADER_VERSION(1, 2):
1996         case WINED3D_SHADER_VERSION(1, 3):
1997             shader->limits.temporary = 2;
1998             shader->limits.constant_float = 8;
1999             shader->limits.constant_int = 0;
2000             shader->limits.constant_bool = 0;
2001             shader->limits.texcoord = 4;
2002             shader->limits.sampler = 4;
2003             shader->limits.packed_input = 0;
2004             shader->limits.label = 0;
2005             break;
2006
2007         case WINED3D_SHADER_VERSION(1, 4):
2008             shader->limits.temporary = 6;
2009             shader->limits.constant_float = 8;
2010             shader->limits.constant_int = 0;
2011             shader->limits.constant_bool = 0;
2012             shader->limits.texcoord = 6;
2013             shader->limits.sampler = 6;
2014             shader->limits.packed_input = 0;
2015             shader->limits.label = 0;
2016             break;
2017
2018         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2019         case WINED3D_SHADER_VERSION(2, 0):
2020             shader->limits.temporary = 32;
2021             shader->limits.constant_float = 32;
2022             shader->limits.constant_int = 16;
2023             shader->limits.constant_bool = 16;
2024             shader->limits.texcoord = 8;
2025             shader->limits.sampler = 16;
2026             shader->limits.packed_input = 0;
2027             break;
2028
2029         case WINED3D_SHADER_VERSION(2, 1):
2030             shader->limits.temporary = 32;
2031             shader->limits.constant_float = 32;
2032             shader->limits.constant_int = 16;
2033             shader->limits.constant_bool = 16;
2034             shader->limits.texcoord = 8;
2035             shader->limits.sampler = 16;
2036             shader->limits.packed_input = 0;
2037             shader->limits.label = 16;
2038             break;
2039
2040         case WINED3D_SHADER_VERSION(4, 0):
2041             FIXME("Using 3.0 limits for 4.0 shader.\n");
2042             /* Fall through. */
2043
2044         case WINED3D_SHADER_VERSION(3, 0):
2045             shader->limits.temporary = 32;
2046             shader->limits.constant_float = 224;
2047             shader->limits.constant_int = 16;
2048             shader->limits.constant_bool = 16;
2049             shader->limits.texcoord = 0;
2050             shader->limits.sampler = 16;
2051             shader->limits.packed_input = 12;
2052             shader->limits.label = 16; /* FIXME: 2048 */
2053             break;
2054
2055         default:
2056             shader->limits.temporary = 32;
2057             shader->limits.constant_float = 32;
2058             shader->limits.constant_int = 16;
2059             shader->limits.constant_bool = 16;
2060             shader->limits.texcoord = 8;
2061             shader->limits.sampler = 16;
2062             shader->limits.packed_input = 0;
2063             shader->limits.label = 0;
2064             FIXME("Unrecognized pixel shader version %u.%u\n",
2065                     shader->reg_maps.shader_version.major,
2066                     shader->reg_maps.shader_version.minor);
2067     }
2068 }
2069
2070 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2071         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2072         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2073 {
2074     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2075     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2076     HRESULT hr;
2077
2078     if (!byte_code) return WINED3DERR_INVALIDCALL;
2079
2080     shader_init(shader, device, parent, parent_ops);
2081     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2082             WINED3D_SHADER_TYPE_PIXEL, max_version);
2083     if (FAILED(hr))
2084     {
2085         WARN("Failed to set function, hr %#x.\n", hr);
2086         shader_cleanup(shader);
2087         return hr;
2088     }
2089
2090     pixelshader_set_limits(shader);
2091
2092     for (i = 0; i < MAX_REG_INPUT; ++i)
2093     {
2094         if (shader->u.ps.input_reg_used[i])
2095         {
2096             ++num_regs_used;
2097             highest_reg_used = i;
2098         }
2099     }
2100
2101     /* Don't do any register mapping magic if it is not needed, or if we can't
2102      * achieve anything anyway */
2103     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2104             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2105     {
2106         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2107         {
2108             /* This happens with relative addressing. The input mapper function
2109              * warns about this if the higher registers are declared too, so
2110              * don't write a FIXME here */
2111             WARN("More varying registers used than supported\n");
2112         }
2113
2114         for (i = 0; i < MAX_REG_INPUT; ++i)
2115         {
2116             shader->u.ps.input_reg_map[i] = i;
2117         }
2118
2119         shader->u.ps.declared_in_count = highest_reg_used + 1;
2120     }
2121     else
2122     {
2123         shader->u.ps.declared_in_count = 0;
2124         for (i = 0; i < MAX_REG_INPUT; ++i)
2125         {
2126             if (shader->u.ps.input_reg_used[i])
2127                 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2128             else shader->u.ps.input_reg_map[i] = ~0U;
2129         }
2130     }
2131
2132     shader->load_local_constsF = FALSE;
2133
2134     return WINED3D_OK;
2135 }
2136
2137 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2138 {
2139     enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2140     unsigned int i;
2141
2142     if (reg_maps->shader_version.major != 1) return;
2143
2144     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2145     {
2146         /* We don't sample from this sampler. */
2147         if (!sampler_type[i]) continue;
2148
2149         if (!textures[i])
2150         {
2151             WARN("No texture bound to sampler %u, using 2D.\n", i);
2152             sampler_type[i] = WINED3DSTT_2D;
2153             continue;
2154         }
2155
2156         switch (textures[i]->target)
2157         {
2158             case GL_TEXTURE_RECTANGLE_ARB:
2159             case GL_TEXTURE_2D:
2160                 /* We have to select between texture rectangles and 2D
2161                  * textures later because 2.0 and 3.0 shaders only have
2162                  * WINED3DSTT_2D as well. */
2163                 sampler_type[i] = WINED3DSTT_2D;
2164                 break;
2165
2166             case GL_TEXTURE_3D:
2167                 sampler_type[i] = WINED3DSTT_VOLUME;
2168                 break;
2169
2170             case GL_TEXTURE_CUBE_MAP_ARB:
2171                 sampler_type[i] = WINED3DSTT_CUBE;
2172                 break;
2173
2174             default:
2175                 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2176                 sampler_type[i] = WINED3DSTT_2D;
2177         }
2178     }
2179 }
2180
2181 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2182         const struct wined3d_shader_signature *output_signature, void *parent,
2183         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2184 {
2185     struct wined3d_shader *object;
2186     HRESULT hr;
2187
2188     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2189             device, byte_code, output_signature, parent, parent_ops, shader);
2190
2191     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2192     if (!object)
2193     {
2194         ERR("Failed to allocate shader memory.\n");
2195         return E_OUTOFMEMORY;
2196     }
2197
2198     hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2199     if (FAILED(hr))
2200     {
2201         WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2202         HeapFree(GetProcessHeap(), 0, object);
2203         return hr;
2204     }
2205
2206     TRACE("Created geometry shader %p.\n", object);
2207     *shader = object;
2208
2209     return WINED3D_OK;
2210 }
2211
2212 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2213         const struct wined3d_shader_signature *output_signature, void *parent,
2214         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2215 {
2216     struct wined3d_shader *object;
2217     HRESULT hr;
2218
2219     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2220             device, byte_code, output_signature, parent, parent_ops, shader);
2221
2222     if (device->ps_selected_mode == SHADER_NONE)
2223         return WINED3DERR_INVALIDCALL;
2224
2225     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2226     if (!object)
2227     {
2228         ERR("Failed to allocate shader memory.\n");
2229         return E_OUTOFMEMORY;
2230     }
2231
2232     hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2233     if (FAILED(hr))
2234     {
2235         WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2236         HeapFree(GetProcessHeap(), 0, object);
2237         return hr;
2238     }
2239
2240     TRACE("Created pixel shader %p.\n", object);
2241     *shader = object;
2242
2243     return WINED3D_OK;
2244 }
2245
2246 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2247         const struct wined3d_shader_signature *output_signature, void *parent,
2248         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2249 {
2250     struct wined3d_shader *object;
2251     HRESULT hr;
2252
2253     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2254             device, byte_code, output_signature, parent, parent_ops, shader);
2255
2256     if (device->vs_selected_mode == SHADER_NONE)
2257         return WINED3DERR_INVALIDCALL;
2258
2259     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2260     if (!object)
2261     {
2262         ERR("Failed to allocate shader memory.\n");
2263         return E_OUTOFMEMORY;
2264     }
2265
2266     hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2267     if (FAILED(hr))
2268     {
2269         WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2270         HeapFree(GetProcessHeap(), 0, object);
2271         return hr;
2272     }
2273
2274     TRACE("Created vertex shader %p.\n", object);
2275     *shader = object;
2276
2277     return WINED3D_OK;
2278 }