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_CONSTANT_BUFFER   */ "dcl_constantBuffer",
  54     /* WINED3DSIH_DCL_INPUT_PRIMITIVE   */ "dcl_inputPrimitive",
  55     /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY   */ "dcl_outputTopology",
  56     /* WINED3DSIH_DCL_VERTICES_OUT      */ "dcl_maxOutputVertexCount",
  57     /* WINED3DSIH_DEF                   */ "def",
  58     /* WINED3DSIH_DEFB                  */ "defb",
  59     /* WINED3DSIH_DEFI                  */ "defi",
  60     /* WINED3DSIH_DIV                   */ "div",
  61     /* WINED3DSIH_DP2ADD                */ "dp2add",
  62     /* WINED3DSIH_DP3                   */ "dp3",
  63     /* WINED3DSIH_DP4                   */ "dp4",
  64     /* WINED3DSIH_DST                   */ "dst",
  65     /* WINED3DSIH_DSX                   */ "dsx",
  66     /* WINED3DSIH_DSY                   */ "dsy",
  67     /* WINED3DSIH_ELSE                  */ "else",
  68     /* WINED3DSIH_EMIT                  */ "emit",
  69     /* WINED3DSIH_ENDIF                 */ "endif",
  70     /* WINED3DSIH_ENDLOOP               */ "endloop",
  71     /* WINED3DSIH_ENDREP                */ "endrep",
  72     /* WINED3DSIH_EQ                    */ "eq",
  73     /* WINED3DSIH_EXP                   */ "exp",
  74     /* WINED3DSIH_EXPP                  */ "expp",
  75     /* WINED3DSIH_FRC                   */ "frc",
  76     /* WINED3DSIH_FTOI                  */ "ftoi",
  77     /* WINED3DSIH_GE                    */ "ge",
  78     /* WINED3DSIH_IADD                  */ "iadd",
  79     /* WINED3DSIH_IEQ                   */ "ieq",
  80     /* WINED3DSIH_IF                    */ "if",
  81     /* WINED3DSIH_IFC                   */ "ifc",
  82     /* WINED3DSIH_IGE                   */ "ige",
  83     /* WINED3DSIH_IMUL                  */ "imul",
  84     /* WINED3DSIH_ITOF                  */ "itof",
  85     /* WINED3DSIH_LABEL                 */ "label",
  86     /* WINED3DSIH_LD                    */ "ld",
  87     /* WINED3DSIH_LIT                   */ "lit",
  88     /* WINED3DSIH_LOG                   */ "log",
  89     /* WINED3DSIH_LOGP                  */ "logp",
  90     /* WINED3DSIH_LOOP                  */ "loop",
  91     /* WINED3DSIH_LRP                   */ "lrp",
  92     /* WINED3DSIH_LT                    */ "lt",
  93     /* WINED3DSIH_M3x2                  */ "m3x2",
  94     /* WINED3DSIH_M3x3                  */ "m3x3",
  95     /* WINED3DSIH_M3x4                  */ "m3x4",
  96     /* WINED3DSIH_M4x3                  */ "m4x3",
  97     /* WINED3DSIH_M4x4                  */ "m4x4",
  98     /* WINED3DSIH_MAD                   */ "mad",
  99     /* WINED3DSIH_MAX                   */ "max",
 100     /* WINED3DSIH_MIN                   */ "min",
 101     /* WINED3DSIH_MOV                   */ "mov",
 102     /* WINED3DSIH_MOVA                  */ "mova",
 103     /* WINED3DSIH_MOVC                  */ "movc",
 104     /* WINED3DSIH_MUL                   */ "mul",
 105     /* WINED3DSIH_NOP                   */ "nop",
 106     /* WINED3DSIH_NRM                   */ "nrm",
 107     /* WINED3DSIH_PHASE                 */ "phase",
 108     /* WINED3DSIH_POW                   */ "pow",
 109     /* WINED3DSIH_RCP                   */ "rcp",
 110     /* WINED3DSIH_REP                   */ "rep",
 111     /* WINED3DSIH_RET                   */ "ret",
 112     /* WINED3DSIH_ROUND_NI              */ "round_ni",
 113     /* WINED3DSIH_RSQ                   */ "rsq",
 114     /* WINED3DSIH_SAMPLE                */ "sample",
 115     /* WINED3DSIH_SAMPLE_GRAD           */ "sample_d",
 116     /* WINED3DSIH_SAMPLE_LOD            */ "sample_l",
 117     /* WINED3DSIH_SETP                  */ "setp",
 118     /* WINED3DSIH_SGE                   */ "sge",
 119     /* WINED3DSIH_SGN                   */ "sgn",
 120     /* WINED3DSIH_SINCOS                */ "sincos",
 121     /* WINED3DSIH_SLT                   */ "slt",
 122     /* WINED3DSIH_SQRT                  */ "sqrt",
 123     /* WINED3DSIH_SUB                   */ "sub",
 124     /* WINED3DSIH_TEX                   */ "texld",
 125     /* WINED3DSIH_TEXBEM                */ "texbem",
 126     /* WINED3DSIH_TEXBEML               */ "texbeml",
 127     /* WINED3DSIH_TEXCOORD              */ "texcrd",
 128     /* WINED3DSIH_TEXDEPTH              */ "texdepth",
 129     /* WINED3DSIH_TEXDP3                */ "texdp3",
 130     /* WINED3DSIH_TEXDP3TEX             */ "texdp3tex",
 131     /* WINED3DSIH_TEXKILL               */ "texkill",
 132     /* WINED3DSIH_TEXLDD                */ "texldd",
 133     /* WINED3DSIH_TEXLDL                */ "texldl",
 134     /* WINED3DSIH_TEXM3x2DEPTH          */ "texm3x2depth",
 135     /* WINED3DSIH_TEXM3x2PAD            */ "texm3x2pad",
 136     /* WINED3DSIH_TEXM3x2TEX            */ "texm3x2tex",
 137     /* WINED3DSIH_TEXM3x3               */ "texm3x3",
 138     /* WINED3DSIH_TEXM3x3DIFF           */ "texm3x3diff",
 139     /* WINED3DSIH_TEXM3x3PAD            */ "texm3x3pad",
 140     /* WINED3DSIH_TEXM3x3SPEC           */ "texm3x3spec",
 141     /* WINED3DSIH_TEXM3x3TEX            */ "texm3x3tex",
 142     /* WINED3DSIH_TEXM3x3VSPEC          */ "texm3x3vspec",
 143     /* WINED3DSIH_TEXREG2AR             */ "texreg2ar",
 144     /* WINED3DSIH_TEXREG2GB             */ "texreg2gb",
 145     /* WINED3DSIH_TEXREG2RGB            */ "texreg2rgb",
 146     /* WINED3DSIH_UDIV                  */ "udiv",
 147     /* WINED3DSIH_USHR                  */ "ushr",
 148     /* WINED3DSIH_UTOF                  */ "utof",
 149     /* WINED3DSIH_XOR                   */ "xor",
 150 };
 151
 152 static const char * const semantic_names[] =
 153 {
 154     /* WINED3D_DECL_USAGE_POSITION      */ "SV_POSITION",
 155     /* WINED3D_DECL_USAGE_BLEND_WEIGHT  */ "BLENDWEIGHT",
 156     /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
 157     /* WINED3D_DECL_USAGE_NORMAL        */ "NORMAL",
 158     /* WINED3D_DECL_USAGE_PSIZE         */ "PSIZE",
 159     /* WINED3D_DECL_USAGE_TEXCOORD      */ "TEXCOORD",
 160     /* WINED3D_DECL_USAGE_TANGENT       */ "TANGENT",
 161     /* WINED3D_DECL_USAGE_BINORMAL      */ "BINORMAL",
 162     /* WINED3D_DECL_USAGE_TESS_FACTOR   */ "TESSFACTOR",
 163     /* WINED3D_DECL_USAGE_POSITIONT     */ "POSITIONT",
 164     /* WINED3D_DECL_USAGE_COLOR         */ "COLOR",
 165     /* WINED3D_DECL_USAGE_FOG           */ "FOG",
 166     /* WINED3D_DECL_USAGE_DEPTH         */ "DEPTH",
 167     /* WINED3D_DECL_USAGE_SAMPLE        */ "SAMPLE",
 168 };
 169
 170 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
 171 {
 172     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 173     {
 174         FIXME("Unrecognized usage %#x.\n", usage);
 175         return "UNRECOGNIZED";
 176     }
 177
 178     return semantic_names[usage];
 179 }
 180
 181 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
 182 {
 183     unsigned int i;
 184
 185     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 186     {
 187         if (!strcmp(name, semantic_names[i])) return i;
 188     }
 189
 190     return ~0U;
 191 }
 192
 193 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
 194 {
 195     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 196 }
 197
 198 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 199         const struct wined3d_shader_semantic *s)
 200 {
 201     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 202     e->semantic_idx = s->usage_idx;
 203     e->sysval_semantic = 0;
 204     e->component_type = 0;
 205     e->register_idx = s->reg.reg.idx[0].offset;
 206     e->mask = s->reg.write_mask;
 207 }
 208
 209 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
 210         enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
 211 {
 212     e->semantic_name = shader_semantic_name_from_usage(usage);
 213     e->semantic_idx = usage_idx;
 214     e->sysval_semantic = 0;
 215     e->component_type = 0;
 216     e->register_idx = reg_idx;
 217     e->mask = write_mask;
 218 }
 219
 220 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 221 {
 222     switch (version_token >> 16)
 223     {
 224         case WINED3D_SM1_VS:
 225         case WINED3D_SM1_PS:
 226             return &sm1_shader_frontend;
 227
 228         case WINED3D_SM4_PS:
 229         case WINED3D_SM4_VS:
 230         case WINED3D_SM4_GS:
 231             return &sm4_shader_frontend;
 232
 233         default:
 234             FIXME("Unrecognised version token %#x\n", version_token);
 235             return NULL;
 236     }
 237 }
 238
 239 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 240 {
 241     buffer->buffer[0] = '\0';
 242     buffer->bsize = 0;
 243     buffer->lineNo = 0;
 244     buffer->newline = TRUE;
 245 }
 246
 247 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 248 {
 249     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 250     if (!buffer->buffer)
 251     {
 252         ERR("Failed to allocate shader buffer memory.\n");
 253         return FALSE;
 254     }
 255
 256     shader_buffer_clear(buffer);
 257     return TRUE;
 258 }
 259
 260 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 261 {
 262     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 263 }
 264
 265 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 266 {
 267     char *base = buffer->buffer + buffer->bsize;
 268     int rc;
 269
 270     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 271
 272     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 273     {
 274         ERR("The buffer allocated for the shader program string "
 275             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 276         buffer->bsize = SHADER_PGMSIZE - 1;
 277         return -1;
 278     }
 279
 280     if (buffer->newline)
 281     {
 282         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 283         buffer->newline = FALSE;
 284     }
 285     else
 286     {
 287         TRACE("%s", base);
 288     }
 289
 290     buffer->bsize += rc;
 291     if (buffer->buffer[buffer->bsize-1] == '\n')
 292     {
 293         ++buffer->lineNo;
 294         buffer->newline = TRUE;
 295     }
 296
 297     return 0;
 298 }
 299
 300 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 301 {
 302     va_list args;
 303     int ret;
 304
 305     va_start(args, format);
 306     ret = shader_vaddline(buffer, format, args);
 307     va_end(args);
 308
 309     return ret;
 310 }
 311
 312 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
 313         void *parent, const struct wined3d_parent_ops *parent_ops)
 314 {
 315     shader->ref = 1;
 316     shader->device = device;
 317     shader->parent = parent;
 318     shader->parent_ops = parent_ops;
 319     list_init(&shader->linked_programs);
 320     list_add_head(&device->shaders, &shader->shader_list_entry);
 321 }
 322
 323 /* Convert floating point offset relative to a register file to an absolute
 324  * offset for float constants. */
 325 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
 326 {
 327     switch (register_type)
 328     {
 329         case WINED3DSPR_CONST: return register_idx;
 330         case WINED3DSPR_CONST2: return 2048 + register_idx;
 331         case WINED3DSPR_CONST3: return 4096 + register_idx;
 332         case WINED3DSPR_CONST4: return 6144 + register_idx;
 333         default:
 334             FIXME("Unsupported register type: %u.\n", register_type);
 335             return register_idx;
 336     }
 337 }
 338
 339 static void shader_delete_constant_list(struct list *clist)
 340 {
 341     struct wined3d_shader_lconst *constant;
 342     struct list *ptr;
 343
 344     ptr = list_head(clist);
 345     while (ptr)
 346     {
 347         constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
 348         ptr = list_next(clist, ptr);
 349         HeapFree(GetProcessHeap(), 0, constant);
 350     }
 351     list_init(clist);
 352 }
 353
 354 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 355 {
 356     DWORD idx, shift;
 357     idx = bit >> 5;
 358     shift = bit & 0x1f;
 359     bitmap[idx] |= (1 << shift);
 360 }
 361
 362 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
 363         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 364 {
 365     switch (reg->type)
 366     {
 367         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 368             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 369                 reg_maps->texcoord |= 1 << reg->idx[0].offset;
 370             else
 371                 reg_maps->address |= 1 << reg->idx[0].offset;
 372             break;
 373
 374         case WINED3DSPR_TEMP:
 375             reg_maps->temporary |= 1 << reg->idx[0].offset;
 376             break;
 377
 378         case WINED3DSPR_INPUT:
 379             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 380             {
 381                 if (reg->idx[0].rel_addr)
 382                 {
 383                     /* If relative addressing is used, we must assume that all registers
 384                      * are used. Even if it is a construct like v3[aL], we can't assume
 385                      * that v0, v1 and v2 aren't read because aL can be negative */
 386                     unsigned int i;
 387                     for (i = 0; i < MAX_REG_INPUT; ++i)
 388                     {
 389                         shader->u.ps.input_reg_used[i] = TRUE;
 390                     }
 391                 }
 392                 else
 393                 {
 394                     shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
 395                 }
 396             }
 397             else
 398                 reg_maps->input_registers |= 1 << reg->idx[0].offset;
 399             break;
 400
 401         case WINED3DSPR_RASTOUT:
 402             if (reg->idx[0].offset == 1)
 403                 reg_maps->fog = 1;
 404             break;
 405
 406         case WINED3DSPR_MISCTYPE:
 407             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 408             {
 409                 if (!reg->idx[0].offset)
 410                     reg_maps->vpos = 1;
 411                 else if (reg->idx[0].offset == 1)
 412                     reg_maps->usesfacing = 1;
 413             }
 414             break;
 415
 416         case WINED3DSPR_CONST:
 417             if (reg->idx[0].rel_addr)
 418             {
 419                 if (reg->idx[0].offset < reg_maps->min_rel_offset)
 420                     reg_maps->min_rel_offset = reg->idx[0].offset;
 421                 if (reg->idx[0].offset > reg_maps->max_rel_offset)
 422                     reg_maps->max_rel_offset = reg->idx[0].offset;
 423                 reg_maps->usesrelconstF = TRUE;
 424             }
 425             else
 426             {
 427                 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
 428             }
 429             break;
 430
 431         case WINED3DSPR_CONSTINT:
 432             reg_maps->integer_constants |= (1 << reg->idx[0].offset);
 433             break;
 434
 435         case WINED3DSPR_CONSTBOOL:
 436             reg_maps->boolean_constants |= (1 << reg->idx[0].offset);
 437             break;
 438
 439         case WINED3DSPR_COLOROUT:
 440             reg_maps->rt_mask |= (1 << reg->idx[0].offset);
 441             break;
 442
 443         default:
 444             TRACE("Not recording register of type %#x and [%#x][%#x].\n",
 445                     reg->type, reg->idx[0].offset, reg->idx[1].offset);
 446             break;
 447     }
 448 }
 449
 450 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 451 {
 452     switch (instr)
 453     {
 454         case WINED3DSIH_M4x4:
 455         case WINED3DSIH_M3x4:
 456             return param == 1 ? 3 : 0;
 457
 458         case WINED3DSIH_M4x3:
 459         case WINED3DSIH_M3x3:
 460             return param == 1 ? 2 : 0;
 461
 462         case WINED3DSIH_M3x2:
 463             return param == 1 ? 1 : 0;
 464
 465         default:
 466             return 0;
 467     }
 468 }
 469
 470 /* Note that this does not count the loop register as an address register. */
 471 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
 472         struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 473         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 474 {
 475     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 476     void *fe_data = shader->frontend_data;
 477     struct wined3d_shader_version shader_version;
 478     const DWORD *ptr = byte_code;
 479
 480     memset(reg_maps, 0, sizeof(*reg_maps));
 481     reg_maps->min_rel_offset = ~0U;
 482
 483     fe->shader_read_header(fe_data, &ptr, &shader_version);
 484     reg_maps->shader_version = shader_version;
 485
 486     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 487             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 488     if (!reg_maps->constf)
 489     {
 490         ERR("Failed to allocate constant map memory.\n");
 491         return E_OUTOFMEMORY;
 492     }
 493
 494     while (!fe->shader_is_end(fe_data, &ptr))
 495     {
 496         struct wined3d_shader_instruction ins;
 497
 498         /* Fetch opcode. */
 499         fe->shader_read_instruction(fe_data, &ptr, &ins);
 500
 501         /* Unhandled opcode, and its parameters. */
 502         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 503         {
 504             TRACE("Skipping unrecognized instruction.\n");
 505             continue;
 506         }
 507
 508         /* Handle declarations. */
 509         if (ins.handler_idx == WINED3DSIH_DCL)
 510         {
 511             struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
 512
 513             switch (semantic->reg.reg.type)
 514             {
 515                 /* Mark input registers used. */
 516                 case WINED3DSPR_INPUT:
 517                     reg_maps->input_registers |= 1 << semantic->reg.reg.idx[0].offset;
 518                     shader_signature_from_semantic(&input_signature[semantic->reg.reg.idx[0].offset], semantic);
 519                     break;
 520
 521                 /* Vertex shader: mark 3.0 output registers used, save token. */
 522                 case WINED3DSPR_OUTPUT:
 523                     reg_maps->output_registers |= 1 << semantic->reg.reg.idx[0].offset;
 524                     shader_signature_from_semantic(&output_signature[semantic->reg.reg.idx[0].offset], semantic);
 525                     if (semantic->usage == WINED3D_DECL_USAGE_FOG)
 526                         reg_maps->fog = 1;
 527                     break;
 528
 529                 /* Save sampler usage token. */
 530                 case WINED3DSPR_SAMPLER:
 531                     reg_maps->sampler_type[semantic->reg.reg.idx[0].offset] = semantic->sampler_type;
 532                     break;
 533
 534                 default:
 535                     TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
 536                     break;
 537             }
 538         }
 539         else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
 540         {
 541             struct wined3d_shader_register *reg = &ins.declaration.src.reg;
 542             if (reg->idx[0].offset >= WINED3D_MAX_CBS)
 543                 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
 544             else
 545                 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
 546         }
 547         else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
 548         {
 549             if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
 550                 shader->u.gs.input_type = ins.declaration.primitive_type;
 551             else
 552                 FIXME("Invalid instruction %#x for shader type %#x.\n",
 553                         ins.handler_idx, shader_version.type);
 554         }
 555         else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
 556         {
 557             if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
 558                 shader->u.gs.output_type = ins.declaration.primitive_type;
 559             else
 560                 FIXME("Invalid instruction %#x for shader type %#x.\n",
 561                         ins.handler_idx, shader_version.type);
 562         }
 563         else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
 564         {
 565             if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
 566                 shader->u.gs.vertices_out = ins.declaration.count;
 567             else
 568                 FIXME("Invalid instruction %#x for shader type %#x.\n",
 569                         ins.handler_idx, shader_version.type);
 570         }
 571         else if (ins.handler_idx == WINED3DSIH_DEF)
 572         {
 573             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 574             if (!lconst) return E_OUTOFMEMORY;
 575
 576             lconst->idx = ins.dst[0].reg.idx[0].offset;
 577             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 578
 579             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 580             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 581             {
 582                 float *value = (float *)lconst->value;
 583                 if (value[0] < -1.0f) value[0] = -1.0f;
 584                 else if (value[0] > 1.0f) value[0] = 1.0f;
 585                 if (value[1] < -1.0f) value[1] = -1.0f;
 586                 else if (value[1] > 1.0f) value[1] = 1.0f;
 587                 if (value[2] < -1.0f) value[2] = -1.0f;
 588                 else if (value[2] > 1.0f) value[2] = 1.0f;
 589                 if (value[3] < -1.0f) value[3] = -1.0f;
 590                 else if (value[3] > 1.0f) value[3] = 1.0f;
 591             }
 592
 593             list_add_head(&shader->constantsF, &lconst->entry);
 594         }
 595         else if (ins.handler_idx == WINED3DSIH_DEFI)
 596         {
 597             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 598             if (!lconst) return E_OUTOFMEMORY;
 599
 600             lconst->idx = ins.dst[0].reg.idx[0].offset;
 601             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 602
 603             list_add_head(&shader->constantsI, &lconst->entry);
 604             reg_maps->local_int_consts |= (1 << lconst->idx);
 605         }
 606         else if (ins.handler_idx == WINED3DSIH_DEFB)
 607         {
 608             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 609             if (!lconst) return E_OUTOFMEMORY;
 610
 611             lconst->idx = ins.dst[0].reg.idx[0].offset;
 612             memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
 613
 614             list_add_head(&shader->constantsB, &lconst->entry);
 615             reg_maps->local_bool_consts |= (1 << lconst->idx);
 616         }
 617         /* For subroutine prototypes. */
 618         else if (ins.handler_idx == WINED3DSIH_LABEL)
 619         {
 620             reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
 621         }
 622         /* Set texture, address, temporary registers. */
 623         else
 624         {
 625             BOOL color0_mov = FALSE;
 626             unsigned int i;
 627
 628             /* This will loop over all the registers and try to
 629              * make a bitmask of the ones we're interested in.
 630              *
 631              * Relative addressing tokens are ignored, but that's
 632              * okay, since we'll catch any address registers when
 633              * they are initialized (required by spec). */
 634             for (i = 0; i < ins.dst_count; ++i)
 635             {
 636                 shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg, shader_version.type);
 637
 638                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 639                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 640                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 641                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 642                 {
 643                     UINT idx = ins.dst[i].reg.idx[0].offset;
 644
 645                     switch (ins.dst[i].reg.type)
 646                     {
 647                         case WINED3DSPR_RASTOUT:
 648                             switch (idx)
 649                             {
 650                                 case 0: /* oPos */
 651                                     reg_maps->output_registers |= 1 << 10;
 652                                     shader_signature_from_usage(&output_signature[10],
 653                                             WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 654                                     break;
 655
 656                                 case 1: /* oFog */
 657                                     reg_maps->output_registers |= 1 << 11;
 658                                     shader_signature_from_usage(&output_signature[11],
 659                                             WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 660                                     break;
 661
 662                                 case 2: /* oPts */
 663                                     reg_maps->output_registers |= 1 << 11;
 664                                     shader_signature_from_usage(&output_signature[11],
 665                                             WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 666                                     break;
 667                             }
 668                             break;
 669
 670                         case WINED3DSPR_ATTROUT:
 671                             if (idx < 2)
 672                             {
 673                                 idx += 8;
 674                                 if (reg_maps->output_registers & (1 << idx))
 675                                 {
 676                                     output_signature[idx].mask |= ins.dst[i].write_mask;
 677                                 }
 678                                 else
 679                                 {
 680                                     reg_maps->output_registers |= 1 << idx;
 681                                     shader_signature_from_usage(&output_signature[idx],
 682                                             WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
 683                                 }
 684                             }
 685                             break;
 686
 687                         case WINED3DSPR_TEXCRDOUT:
 688
 689                             reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
 690                             if (reg_maps->output_registers & (1 << idx))
 691                             {
 692                                 output_signature[idx].mask |= ins.dst[i].write_mask;
 693                             }
 694                             else
 695                             {
 696                                 reg_maps->output_registers |= 1 << idx;
 697                                 shader_signature_from_usage(&output_signature[idx],
 698                                         WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
 699                             }
 700                             break;
 701
 702                         default:
 703                             break;
 704                     }
 705                 }
 706
 707                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 708                 {
 709                     if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
 710                     {
 711                         /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 712                          * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 713                          * the mov and perform the sRGB write correction from the source register.
 714                          *
 715                          * However, if the mov is only partial, we can't do this, and if the write
 716                          * comes from an instruction other than MOV it is hard to do as well. If
 717                          * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 718                         shader->u.ps.color0_mov = FALSE;
 719                         if (ins.handler_idx == WINED3DSIH_MOV
 720                                 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
 721                         {
 722                             /* Used later when the source register is read. */
 723                             color0_mov = TRUE;
 724                         }
 725                     }
 726                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 727                      * end
 728                      */
 729                     else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
 730                             && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
 731                     {
 732                         shader->u.ps.color0_mov = FALSE;
 733                     }
 734                 }
 735
 736                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 737                 if (shader_version.major == 1
 738                         && (ins.handler_idx == WINED3DSIH_TEX
 739                             || ins.handler_idx == WINED3DSIH_TEXBEM
 740                             || ins.handler_idx == WINED3DSIH_TEXBEML
 741                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 742                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 743                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 744                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 745                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 746                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 747                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 748                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 749                 {
 750                     /* Fake sampler usage, only set reserved bit and type. */
 751                     DWORD sampler_code = ins.dst[i].reg.idx[0].offset;
 752
 753                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 754                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 755
 756                     /* texbem is only valid with < 1.4 pixel shaders */
 757                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 758                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 759                     {
 760                         reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
 761                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 762                         {
 763                             reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx[0].offset;
 764                         }
 765                     }
 766                 }
 767                 else if (ins.handler_idx == WINED3DSIH_BEM)
 768                 {
 769                     reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
 770                 }
 771             }
 772
 773             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 774             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 775             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 776             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 777             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 778             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 779             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 780             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 781             else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
 782             else if (ins.handler_idx == WINED3DSIH_LOOP
 783                     || ins.handler_idx == WINED3DSIH_REP)
 784             {
 785                 ++cur_loop_depth;
 786                 if (cur_loop_depth > max_loop_depth)
 787                     max_loop_depth = cur_loop_depth;
 788             }
 789             else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 790                     || ins.handler_idx == WINED3DSIH_ENDREP)
 791                 --cur_loop_depth;
 792
 793             if (ins.predicate)
 794                 shader_record_register_usage(shader, reg_maps, &ins.predicate->reg, shader_version.type);
 795
 796             for (i = 0; i < ins.src_count; ++i)
 797             {
 798                 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
 799                 struct wined3d_shader_register reg = ins.src[i].reg;
 800
 801                 shader_record_register_usage(shader, reg_maps, &ins.src[i].reg, shader_version.type);
 802                 while (count)
 803                 {
 804                     ++reg.idx[0].offset;
 805                     shader_record_register_usage(shader, reg_maps, &reg, shader_version.type);
 806                     --count;
 807                 }
 808
 809                 if (color0_mov)
 810                 {
 811                     if (ins.src[i].reg.type == WINED3DSPR_TEMP
 812                             && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
 813                     {
 814                         shader->u.ps.color0_mov = TRUE;
 815                         shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
 816                     }
 817                 }
 818             }
 819         }
 820     }
 821     reg_maps->loop_depth = max_loop_depth;
 822
 823     /* PS before 2.0 don't have explicit color outputs. Instead the value of
 824      * R0 is written to the render target. */
 825     if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 826         reg_maps->rt_mask |= (1 << 0);
 827
 828     shader->functionLength = ((const char *)ptr - (const char *)byte_code);
 829
 830     return WINED3D_OK;
 831 }
 832
 833 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 834 {
 835     DWORD map = 1 << max;
 836     map |= map - 1;
 837     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 838
 839     return wined3d_log2i(map);
 840 }
 841
 842 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 843         const struct wined3d_shader_version *shader_version)
 844 {
 845     TRACE("dcl");
 846
 847     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 848     {
 849         switch (semantic->sampler_type)
 850         {
 851             case WINED3DSTT_2D: TRACE("_2d"); break;
 852             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 853             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 854             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 855         }
 856     }
 857     else
 858     {
 859         /* Pixel shaders 3.0 don't have usage semantics. */
 860         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 861         else TRACE("_");
 862
 863         switch (semantic->usage)
 864         {
 865             case WINED3D_DECL_USAGE_POSITION:
 866                 TRACE("position%u", semantic->usage_idx);
 867                 break;
 868
 869             case WINED3D_DECL_USAGE_BLEND_INDICES:
 870                 TRACE("blend");
 871                 break;
 872
 873             case WINED3D_DECL_USAGE_BLEND_WEIGHT:
 874                 TRACE("weight");
 875                 break;
 876
 877             case WINED3D_DECL_USAGE_NORMAL:
 878                 TRACE("normal%u", semantic->usage_idx);
 879                 break;
 880
 881             case WINED3D_DECL_USAGE_PSIZE:
 882                 TRACE("psize");
 883                 break;
 884
 885             case WINED3D_DECL_USAGE_COLOR:
 886                 if (!semantic->usage_idx) TRACE("color");
 887                 else TRACE("specular%u", (semantic->usage_idx - 1));
 888                 break;
 889
 890             case WINED3D_DECL_USAGE_TEXCOORD:
 891                 TRACE("texture%u", semantic->usage_idx);
 892                 break;
 893
 894             case WINED3D_DECL_USAGE_TANGENT:
 895                 TRACE("tangent");
 896                 break;
 897
 898             case WINED3D_DECL_USAGE_BINORMAL:
 899                 TRACE("binormal");
 900                 break;
 901
 902             case WINED3D_DECL_USAGE_TESS_FACTOR:
 903                 TRACE("tessfactor");
 904                 break;
 905
 906             case WINED3D_DECL_USAGE_POSITIONT:
 907                 TRACE("positionT%u", semantic->usage_idx);
 908                 break;
 909
 910             case WINED3D_DECL_USAGE_FOG:
 911                 TRACE("fog");
 912                 break;
 913
 914             case WINED3D_DECL_USAGE_DEPTH:
 915                 TRACE("depth");
 916                 break;
 917
 918             case WINED3D_DECL_USAGE_SAMPLE:
 919                 TRACE("sample");
 920                 break;
 921
 922             default:
 923                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 924         }
 925     }
 926 }
 927
 928 static void shader_dump_register(const struct wined3d_shader_register *reg,
 929         const struct wined3d_shader_version *shader_version)
 930 {
 931     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 932     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 933     UINT offset = reg->idx[0].offset;
 934
 935     switch (reg->type)
 936     {
 937         case WINED3DSPR_TEMP:
 938             TRACE("r");
 939             break;
 940
 941         case WINED3DSPR_INPUT:
 942             TRACE("v");
 943             break;
 944
 945         case WINED3DSPR_CONST:
 946         case WINED3DSPR_CONST2:
 947         case WINED3DSPR_CONST3:
 948         case WINED3DSPR_CONST4:
 949             TRACE("c");
 950             offset = shader_get_float_offset(reg->type, offset);
 951             break;
 952
 953         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 954             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 955             break;
 956
 957         case WINED3DSPR_RASTOUT:
 958             TRACE("%s", rastout_reg_names[offset]);
 959             break;
 960
 961         case WINED3DSPR_COLOROUT:
 962             TRACE("oC");
 963             break;
 964
 965         case WINED3DSPR_DEPTHOUT:
 966             TRACE("oDepth");
 967             break;
 968
 969         case WINED3DSPR_ATTROUT:
 970             TRACE("oD");
 971             break;
 972
 973         case WINED3DSPR_TEXCRDOUT:
 974             /* Vertex shaders >= 3.0 use general purpose output registers
 975              * (WINED3DSPR_OUTPUT), which can include an address token. */
 976             if (shader_version->major >= 3) TRACE("o");
 977             else TRACE("oT");
 978             break;
 979
 980         case WINED3DSPR_CONSTINT:
 981             TRACE("i");
 982             break;
 983
 984         case WINED3DSPR_CONSTBOOL:
 985             TRACE("b");
 986             break;
 987
 988         case WINED3DSPR_LABEL:
 989             TRACE("l");
 990             break;
 991
 992         case WINED3DSPR_LOOP:
 993             TRACE("aL");
 994             break;
 995
 996         case WINED3DSPR_SAMPLER:
 997             TRACE("s");
 998             break;
 999
1000         case WINED3DSPR_MISCTYPE:
1001             if (offset > 1)
1002                 FIXME("Unhandled misctype register %u.\n", offset);
1003             else
1004                 TRACE("%s", misctype_reg_names[offset]);
1005             break;
1006
1007         case WINED3DSPR_PREDICATE:
1008             TRACE("p");
1009             break;
1010
1011         case WINED3DSPR_IMMCONST:
1012             TRACE("l");
1013             break;
1014
1015         case WINED3DSPR_CONSTBUFFER:
1016             TRACE("cb");
1017             break;
1018
1019         case WINED3DSPR_PRIMID:
1020             TRACE("primID");
1021             break;
1022
1023         case WINED3DSPR_NULL:
1024             TRACE("null");
1025             break;
1026
1027         case WINED3DSPR_RESOURCE:
1028             TRACE("t");
1029             break;
1030
1031         default:
1032             TRACE("unhandled_rtype(%#x)", reg->type);
1033             break;
1034     }
1035
1036     if (reg->type == WINED3DSPR_IMMCONST)
1037     {
1038         TRACE("(");
1039         switch (reg->immconst_type)
1040         {
1041             case WINED3D_IMMCONST_SCALAR:
1042                 switch (reg->data_type)
1043                 {
1044                     case WINED3D_DATA_FLOAT:
1045                         TRACE("%.8e", *(const float *)reg->immconst_data);
1046                         break;
1047                     case WINED3D_DATA_INT:
1048                         TRACE("%d", reg->immconst_data[0]);
1049                         break;
1050                     case WINED3D_DATA_RESOURCE:
1051                     case WINED3D_DATA_SAMPLER:
1052                     case WINED3D_DATA_UINT:
1053                         TRACE("%u", reg->immconst_data[0]);
1054                         break;
1055                     default:
1056                         TRACE("<unhandled data type %#x>", reg->data_type);
1057                         break;
1058                 }
1059                 break;
1060
1061             case WINED3D_IMMCONST_VEC4:
1062                 switch (reg->data_type)
1063                 {
1064                     case WINED3D_DATA_FLOAT:
1065                         TRACE("%.8e, %.8e, %.8e, %.8e",
1066                                 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1067                                 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1068                         break;
1069                     case WINED3D_DATA_INT:
1070                         TRACE("%d, %d, %d, %d",
1071                                 reg->immconst_data[0], reg->immconst_data[1],
1072                                 reg->immconst_data[2], reg->immconst_data[3]);
1073                         break;
1074                     case WINED3D_DATA_RESOURCE:
1075                     case WINED3D_DATA_SAMPLER:
1076                     case WINED3D_DATA_UINT:
1077                         TRACE("%u, %u, %u, %u",
1078                                 reg->immconst_data[0], reg->immconst_data[1],
1079                                 reg->immconst_data[2], reg->immconst_data[3]);
1080                         break;
1081                     default:
1082                         TRACE("<unhandled data type %#x>", reg->data_type);
1083                         break;
1084                 }
1085                 break;
1086
1087             default:
1088                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1089                 break;
1090         }
1091         TRACE(")");
1092     }
1093     else if (reg->type != WINED3DSPR_RASTOUT
1094             && reg->type != WINED3DSPR_MISCTYPE
1095             && reg->type != WINED3DSPR_NULL)
1096     {
1097         if (offset != ~0U)
1098         {
1099             TRACE("[");
1100             if (reg->idx[0].rel_addr)
1101             {
1102                 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1103                 TRACE(" + ");
1104             }
1105             TRACE("%u]", offset);
1106
1107             if (reg->idx[1].offset != ~0U)
1108             {
1109                 TRACE("[");
1110                 if (reg->idx[1].rel_addr)
1111                 {
1112                     shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1113                     TRACE(" + ");
1114                 }
1115                 TRACE("%u]", reg->idx[1].offset);
1116             }
1117         }
1118     }
1119 }
1120
1121 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1122         const struct wined3d_shader_version *shader_version)
1123 {
1124     DWORD write_mask = param->write_mask;
1125
1126     shader_dump_register(&param->reg, shader_version);
1127
1128     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1129     {
1130         static const char *write_mask_chars = "xyzw";
1131
1132         TRACE(".");
1133         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1134         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1135         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1136         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1137     }
1138 }
1139
1140 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1141         const struct wined3d_shader_version *shader_version)
1142 {
1143     enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1144     DWORD swizzle = param->swizzle;
1145
1146     if (src_modifier == WINED3DSPSM_NEG
1147             || src_modifier == WINED3DSPSM_BIASNEG
1148             || src_modifier == WINED3DSPSM_SIGNNEG
1149             || src_modifier == WINED3DSPSM_X2NEG
1150             || src_modifier == WINED3DSPSM_ABSNEG)
1151         TRACE("-");
1152     else if (src_modifier == WINED3DSPSM_COMP)
1153         TRACE("1-");
1154     else if (src_modifier == WINED3DSPSM_NOT)
1155         TRACE("!");
1156
1157     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1158         TRACE("abs(");
1159
1160     shader_dump_register(&param->reg, shader_version);
1161
1162     if (src_modifier)
1163     {
1164         switch (src_modifier)
1165         {
1166             case WINED3DSPSM_NONE:    break;
1167             case WINED3DSPSM_NEG:     break;
1168             case WINED3DSPSM_NOT:     break;
1169             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1170             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1171             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1172             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1173             case WINED3DSPSM_COMP:    break;
1174             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1175             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1176             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1177             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1178             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1179             case WINED3DSPSM_ABS:     TRACE(")"); break;
1180             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1181         }
1182     }
1183
1184     if (swizzle != WINED3DSP_NOSWIZZLE)
1185     {
1186         static const char *swizzle_chars = "xyzw";
1187         DWORD swizzle_x = swizzle & 0x03;
1188         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1189         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1190         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1191
1192         if (swizzle_x == swizzle_y
1193                 && swizzle_x == swizzle_z
1194                 && swizzle_x == swizzle_w)
1195         {
1196             TRACE(".%c", swizzle_chars[swizzle_x]);
1197         }
1198         else
1199         {
1200             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1201                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1202         }
1203     }
1204 }
1205
1206 /* Shared code in order to generate the bulk of the shader string.
1207  * NOTE: A description of how to parse tokens can be found on MSDN. */
1208 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1209         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1210 {
1211     struct wined3d_device *device = shader->device;
1212     const struct wined3d_shader_frontend *fe = shader->frontend;
1213     void *fe_data = shader->frontend_data;
1214     struct wined3d_shader_version shader_version;
1215     struct wined3d_shader_loop_state loop_state;
1216     struct wined3d_shader_instruction ins;
1217     struct wined3d_shader_tex_mx tex_mx;
1218     struct wined3d_shader_context ctx;
1219     const DWORD *ptr = byte_code;
1220
1221     /* Initialize current parsing state. */
1222     tex_mx.current_row = 0;
1223     loop_state.current_depth = 0;
1224     loop_state.current_reg = 0;
1225
1226     ctx.shader = shader;
1227     ctx.gl_info = &device->adapter->gl_info;
1228     ctx.reg_maps = reg_maps;
1229     ctx.buffer = buffer;
1230     ctx.tex_mx = &tex_mx;
1231     ctx.loop_state = &loop_state;
1232     ctx.backend_data = backend_ctx;
1233     ins.ctx = &ctx;
1234
1235     fe->shader_read_header(fe_data, &ptr, &shader_version);
1236
1237     while (!fe->shader_is_end(fe_data, &ptr))
1238     {
1239         /* Read opcode. */
1240         fe->shader_read_instruction(fe_data, &ptr, &ins);
1241
1242         /* Unknown opcode and its parameters. */
1243         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1244         {
1245             TRACE("Skipping unrecognized instruction.\n");
1246             continue;
1247         }
1248
1249         if (ins.predicate)
1250             FIXME("Predicates not implemented.\n");
1251
1252         /* Call appropriate function for output target */
1253         device->shader_backend->shader_handle_instruction(&ins);
1254     }
1255 }
1256
1257 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1258 {
1259     DWORD mmask = dst->modifiers;
1260
1261     switch (dst->shift)
1262     {
1263         case 0: break;
1264         case 13: TRACE("_d8"); break;
1265         case 14: TRACE("_d4"); break;
1266         case 15: TRACE("_d2"); break;
1267         case 1: TRACE("_x2"); break;
1268         case 2: TRACE("_x4"); break;
1269         case 3: TRACE("_x8"); break;
1270         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1271     }
1272
1273     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1274     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1275     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1276
1277     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1278     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1279 }
1280
1281 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1282 {
1283     switch (primitive_type)
1284     {
1285         case WINED3D_PT_UNDEFINED:
1286             TRACE("undefined");
1287             break;
1288         case WINED3D_PT_POINTLIST:
1289             TRACE("pointlist");
1290             break;
1291         case WINED3D_PT_LINELIST:
1292             TRACE("linelist");
1293             break;
1294         case WINED3D_PT_LINESTRIP:
1295             TRACE("linestrip");
1296             break;
1297         case WINED3D_PT_TRIANGLELIST:
1298             TRACE("trianglelist");
1299             break;
1300         case WINED3D_PT_TRIANGLESTRIP:
1301             TRACE("trianglestrip");
1302             break;
1303         case WINED3D_PT_TRIANGLEFAN:
1304             TRACE("trianglefan");
1305             break;
1306         case WINED3D_PT_LINELIST_ADJ:
1307             TRACE("linelist_adj");
1308             break;
1309         case WINED3D_PT_LINESTRIP_ADJ:
1310             TRACE("linestrip_adj");
1311             break;
1312         case WINED3D_PT_TRIANGLELIST_ADJ:
1313             TRACE("trianglelist_adj");
1314             break;
1315         case WINED3D_PT_TRIANGLESTRIP_ADJ:
1316             TRACE("trianglestrip_adj");
1317             break;
1318         default:
1319             TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1320             break;
1321     }
1322 }
1323
1324 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1325 {
1326     struct wined3d_shader_version shader_version;
1327     const DWORD *ptr = byte_code;
1328     const char *type_prefix;
1329     DWORD i;
1330
1331     TRACE("Parsing %p.\n", byte_code);
1332
1333     fe->shader_read_header(fe_data, &ptr, &shader_version);
1334
1335     switch (shader_version.type)
1336     {
1337         case WINED3D_SHADER_TYPE_VERTEX:
1338             type_prefix = "vs";
1339             break;
1340
1341         case WINED3D_SHADER_TYPE_GEOMETRY:
1342             type_prefix = "gs";
1343             break;
1344
1345         case WINED3D_SHADER_TYPE_PIXEL:
1346             type_prefix = "ps";
1347             break;
1348
1349         default:
1350             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1351             type_prefix = "unknown";
1352             break;
1353     }
1354
1355     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1356
1357     while (!fe->shader_is_end(fe_data, &ptr))
1358     {
1359         struct wined3d_shader_instruction ins;
1360
1361         fe->shader_read_instruction(fe_data, &ptr, &ins);
1362         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1363         {
1364             TRACE("Skipping unrecognized instruction.\n");
1365             continue;
1366         }
1367
1368         if (ins.handler_idx == WINED3DSIH_DCL)
1369         {
1370             shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1371             shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1372             TRACE(" ");
1373             shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1374         }
1375         else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1376         {
1377             TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1378             shader_dump_src_param(&ins.declaration.src, &shader_version);
1379             TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1380         }
1381         else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1382                 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1383         {
1384             TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1385             shader_dump_primitive_type(ins.declaration.primitive_type);
1386         }
1387         else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1388         {
1389             TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1390         }
1391         else if (ins.handler_idx == WINED3DSIH_DEF)
1392         {
1393             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1394                     ins.dst[0].reg.idx[0].offset),
1395                     *(const float *)&ins.src[0].reg.immconst_data[0],
1396                     *(const float *)&ins.src[0].reg.immconst_data[1],
1397                     *(const float *)&ins.src[0].reg.immconst_data[2],
1398                     *(const float *)&ins.src[0].reg.immconst_data[3]);
1399         }
1400         else if (ins.handler_idx == WINED3DSIH_DEFI)
1401         {
1402             TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1403                     ins.src[0].reg.immconst_data[0],
1404                     ins.src[0].reg.immconst_data[1],
1405                     ins.src[0].reg.immconst_data[2],
1406                     ins.src[0].reg.immconst_data[3]);
1407         }
1408         else if (ins.handler_idx == WINED3DSIH_DEFB)
1409         {
1410             TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1411         }
1412         else
1413         {
1414             if (ins.predicate)
1415             {
1416                 TRACE("(");
1417                 shader_dump_src_param(ins.predicate, &shader_version);
1418                 TRACE(") ");
1419             }
1420
1421             /* PixWin marks instructions with the coissue flag with a '+' */
1422             if (ins.coissue) TRACE("+");
1423
1424             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1425
1426             if (ins.handler_idx == WINED3DSIH_IFC
1427                     || ins.handler_idx == WINED3DSIH_BREAKC)
1428             {
1429                 switch (ins.flags)
1430                 {
1431                     case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1432                     case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1433                     case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1434                     case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1435                     case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1436                     case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1437                     default: TRACE("_(%u)", ins.flags);
1438                 }
1439             }
1440             else if (ins.handler_idx == WINED3DSIH_TEX
1441                     && shader_version.major >= 2
1442                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1443             {
1444                 TRACE("p");
1445             }
1446
1447             for (i = 0; i < ins.dst_count; ++i)
1448             {
1449                 shader_dump_ins_modifiers(&ins.dst[i]);
1450                 TRACE(!i ? " " : ", ");
1451                 shader_dump_dst_param(&ins.dst[i], &shader_version);
1452             }
1453
1454             /* Other source tokens */
1455             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1456             {
1457                 TRACE(!i ? " " : ", ");
1458                 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1459             }
1460         }
1461         TRACE("\n");
1462     }
1463 }
1464
1465 static void shader_cleanup(struct wined3d_shader *shader)
1466 {
1467     shader->device->shader_backend->shader_destroy(shader);
1468     HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1469     HeapFree(GetProcessHeap(), 0, shader->function);
1470     shader_delete_constant_list(&shader->constantsF);
1471     shader_delete_constant_list(&shader->constantsB);
1472     shader_delete_constant_list(&shader->constantsI);
1473     list_remove(&shader->shader_list_entry);
1474
1475     if (shader->frontend && shader->frontend_data)
1476         shader->frontend->shader_free(shader->frontend_data);
1477 }
1478
1479 struct shader_none_priv
1480 {
1481     const struct fragment_pipeline *fragment_pipe;
1482 };
1483
1484 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1485 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1486         enum tex_types tex_type, const SIZE *ds_mask_size) {}
1487 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1488 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1489 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1490 static void shader_none_load_constants(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1491 static void shader_none_load_np2fixup_constants(void *shader_priv,
1492         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1493 static void shader_none_destroy(struct wined3d_shader *shader) {}
1494 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1495
1496 static void shader_none_select(const struct wined3d_context *context, enum wined3d_shader_mode vertex_mode,
1497         enum wined3d_shader_mode fragment_mode)
1498 {
1499     const struct wined3d_gl_info *gl_info = context->gl_info;
1500     struct wined3d_device *device = context->swapchain->device;
1501     struct shader_none_priv *priv = device->shader_priv;
1502
1503     priv->fragment_pipe->enable_extension(gl_info, fragment_mode == WINED3D_SHADER_MODE_FFP);
1504 }
1505
1506 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct fragment_pipeline *fragment_pipe)
1507 {
1508     struct shader_none_priv *priv;
1509     void *fragment_priv;
1510
1511     if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1512     {
1513         ERR("Failed to allocate private data.\n");
1514         return E_OUTOFMEMORY;
1515     }
1516
1517     if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1518     {
1519         ERR("Failed to initialize fragment pipe.\n");
1520         HeapFree(GetProcessHeap(), 0, priv);
1521         return E_FAIL;
1522     }
1523
1524     device->fragment_priv = fragment_priv;
1525     priv->fragment_pipe = fragment_pipe;
1526     device->shader_priv = priv;
1527
1528     return WINED3D_OK;
1529 }
1530
1531 static void shader_none_free(struct wined3d_device *device)
1532 {
1533     struct shader_none_priv *priv = device->shader_priv;
1534
1535     priv->fragment_pipe->free_private(device);
1536     HeapFree(GetProcessHeap(), 0, priv);
1537 }
1538
1539 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1540 {
1541     /* Set the shader caps to 0 for the none shader backend */
1542     caps->vs_version = 0;
1543     caps->gs_version = 0;
1544     caps->ps_version = 0;
1545     caps->vs_uniform_count = 0;
1546     caps->ps_uniform_count = 0;
1547     caps->ps_1x_max_value = 0.0f;
1548     caps->vs_clipping = FALSE;
1549 }
1550
1551 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1552 {
1553     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1554     {
1555         TRACE("Checking support for fixup:\n");
1556         dump_color_fixup_desc(fixup);
1557     }
1558
1559     /* Faked to make some apps happy. */
1560     if (!is_complex_fixup(fixup))
1561     {
1562         TRACE("[OK]\n");
1563         return TRUE;
1564     }
1565
1566     TRACE("[FAILED]\n");
1567     return FALSE;
1568 }
1569
1570 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
1571 {
1572     struct shader_none_priv *priv = shader_priv;
1573
1574     return priv->fragment_pipe->ffp_proj_control;
1575 }
1576
1577 const struct wined3d_shader_backend_ops none_shader_backend =
1578 {
1579     shader_none_handle_instruction,
1580     shader_none_select,
1581     shader_none_select_depth_blt,
1582     shader_none_deselect_depth_blt,
1583     shader_none_update_float_vertex_constants,
1584     shader_none_update_float_pixel_constants,
1585     shader_none_load_constants,
1586     shader_none_load_np2fixup_constants,
1587     shader_none_destroy,
1588     shader_none_alloc,
1589     shader_none_free,
1590     shader_none_context_destroyed,
1591     shader_none_get_caps,
1592     shader_none_color_fixup_supported,
1593     shader_none_has_ffp_proj_control,
1594 };
1595
1596 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1597         const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1598         enum wined3d_shader_type type, unsigned int max_version)
1599 {
1600     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1601     const struct wined3d_shader_frontend *fe;
1602     HRESULT hr;
1603     unsigned int backend_version;
1604
1605     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1606             shader, byte_code, output_signature, float_const_count);
1607
1608     fe = shader_select_frontend(*byte_code);
1609     if (!fe)
1610     {
1611         FIXME("Unable to find frontend for shader.\n");
1612         return WINED3DERR_INVALIDCALL;
1613     }
1614     shader->frontend = fe;
1615     shader->frontend_data = fe->shader_init(byte_code, output_signature);
1616     if (!shader->frontend_data)
1617     {
1618         FIXME("Failed to initialize frontend.\n");
1619         return WINED3DERR_INVALIDCALL;
1620     }
1621
1622     /* First pass: trace shader. */
1623     if (TRACE_ON(d3d_shader))
1624         shader_trace_init(fe, shader->frontend_data, byte_code);
1625
1626     /* Initialize immediate constant lists. */
1627     list_init(&shader->constantsF);
1628     list_init(&shader->constantsB);
1629     list_init(&shader->constantsI);
1630
1631     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1632     hr = shader_get_registers_used(shader, fe,
1633             reg_maps, shader->input_signature, shader->output_signature,
1634             byte_code, float_const_count);
1635     if (FAILED(hr)) return hr;
1636
1637     if (reg_maps->shader_version.type != type)
1638     {
1639         WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1640         return WINED3DERR_INVALIDCALL;
1641     }
1642     if (reg_maps->shader_version.major > max_version)
1643     {
1644         WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1645         return WINED3DERR_INVALIDCALL;
1646     }
1647     switch (type)
1648     {
1649         case WINED3D_SHADER_TYPE_VERTEX:
1650             backend_version = shader->device->vs_version;
1651             break;
1652         case WINED3D_SHADER_TYPE_GEOMETRY:
1653             backend_version = shader->device->gs_version;
1654             break;
1655         case WINED3D_SHADER_TYPE_PIXEL:
1656             backend_version = shader->device->ps_version;
1657             break;
1658         default:
1659             FIXME("No backend version-checking for this shader type\n");
1660             backend_version = 0;
1661     }
1662     if (reg_maps->shader_version.major > backend_version)
1663     {
1664         WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1665                 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1666         return WINED3DERR_INVALIDCALL;
1667     }
1668
1669     shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1670     if (!shader->function)
1671         return E_OUTOFMEMORY;
1672     memcpy(shader->function, byte_code, shader->functionLength);
1673
1674     return WINED3D_OK;
1675 }
1676
1677 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1678 {
1679     ULONG refcount = InterlockedIncrement(&shader->ref);
1680
1681     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1682
1683     return refcount;
1684 }
1685
1686 /* Do not call while under the GL lock. */
1687 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1688 {
1689     ULONG refcount = InterlockedDecrement(&shader->ref);
1690
1691     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1692
1693     if (!refcount)
1694     {
1695         shader_cleanup(shader);
1696         shader->parent_ops->wined3d_object_destroyed(shader->parent);
1697         HeapFree(GetProcessHeap(), 0, shader);
1698     }
1699
1700     return refcount;
1701 }
1702
1703 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1704 {
1705     TRACE("shader %p.\n", shader);
1706
1707     return shader->parent;
1708 }
1709
1710 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1711         void *byte_code, UINT *byte_code_size)
1712 {
1713     TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1714
1715     if (!byte_code)
1716     {
1717         *byte_code_size = shader->functionLength;
1718         return WINED3D_OK;
1719     }
1720
1721     if (*byte_code_size < shader->functionLength)
1722     {
1723         /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1724          * than the required size we should write the required size and
1725          * return D3DERR_MOREDATA. That's not actually true. */
1726         return WINED3DERR_INVALIDCALL;
1727     }
1728
1729     memcpy(byte_code, shader->function, shader->functionLength);
1730
1731     return WINED3D_OK;
1732 }
1733
1734 /* Set local constants for d3d8 shaders. */
1735 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1736         UINT start_idx, const float *src_data, UINT count)
1737 {
1738     UINT end_idx = start_idx + count;
1739     UINT i;
1740
1741     TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1742
1743     if (end_idx > shader->limits.constant_float)
1744     {
1745         WARN("end_idx %u > float constants limit %u.\n",
1746                 end_idx, shader->limits.constant_float);
1747         end_idx = shader->limits.constant_float;
1748     }
1749
1750     for (i = start_idx; i < end_idx; ++i)
1751     {
1752         struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1753         if (!lconst)
1754             return E_OUTOFMEMORY;
1755
1756         lconst->idx = i;
1757         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1758         list_add_head(&shader->constantsF, &lconst->entry);
1759     }
1760
1761     return WINED3D_OK;
1762 }
1763
1764 void find_vs_compile_args(const struct wined3d_state *state,
1765         const struct wined3d_shader *shader, struct vs_compile_args *args)
1766 {
1767     args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1768             == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1769     args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1770             && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1771     args->swizzle_map = shader->device->strided_streams.swizzle_map;
1772 }
1773
1774 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1775 {
1776     if (usage_idx1 != usage_idx2)
1777         return FALSE;
1778     if (usage1 == usage2)
1779         return TRUE;
1780     if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1781         return TRUE;
1782     if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1783         return TRUE;
1784
1785     return FALSE;
1786 }
1787
1788 BOOL vshader_get_input(const struct wined3d_shader *shader,
1789         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1790 {
1791     WORD map = shader->reg_maps.input_registers;
1792     unsigned int i;
1793
1794     for (i = 0; map; map >>= 1, ++i)
1795     {
1796         if (!(map & 1)) continue;
1797
1798         if (match_usage(shader->u.vs.attributes[i].usage,
1799                 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1800         {
1801             *regnum = i;
1802             return TRUE;
1803         }
1804     }
1805     return FALSE;
1806 }
1807
1808 static void vertexshader_set_limits(struct wined3d_shader *shader)
1809 {
1810     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1811             shader->reg_maps.shader_version.minor);
1812     struct wined3d_device *device = shader->device;
1813
1814     shader->limits.packed_input = 0;
1815
1816     switch (shader_version)
1817     {
1818         case WINED3D_SHADER_VERSION(1, 0):
1819         case WINED3D_SHADER_VERSION(1, 1):
1820             shader->limits.constant_bool = 0;
1821             shader->limits.constant_int = 0;
1822             shader->limits.packed_output = 12;
1823             shader->limits.sampler = 0;
1824             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1825              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1826              * constants? */
1827             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1828             break;
1829
1830         case WINED3D_SHADER_VERSION(2, 0):
1831         case WINED3D_SHADER_VERSION(2, 1):
1832             shader->limits.constant_bool = 16;
1833             shader->limits.constant_int = 16;
1834             shader->limits.packed_output = 12;
1835             shader->limits.sampler = 0;
1836             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1837             break;
1838
1839         case WINED3D_SHADER_VERSION(3, 0):
1840             shader->limits.constant_bool = 16;
1841             shader->limits.constant_int = 16;
1842             shader->limits.packed_output = 12;
1843             shader->limits.sampler = 4;
1844             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1845              * even though they are capable of supporting much more (GL
1846              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1847              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1848              * shaders to 256. */
1849             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1850             break;
1851
1852         case WINED3D_SHADER_VERSION(4, 0):
1853             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1854             shader->limits.constant_int = 0;
1855             shader->limits.constant_float = 0;
1856             shader->limits.constant_bool = 0;
1857             shader->limits.packed_output = 16;
1858             shader->limits.packed_input = 0;
1859             break;
1860
1861         default:
1862             shader->limits.constant_bool = 16;
1863             shader->limits.constant_int = 16;
1864             shader->limits.packed_output = 12;
1865             shader->limits.sampler = 0;
1866             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1867             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1868                     shader->reg_maps.shader_version.major,
1869                     shader->reg_maps.shader_version.minor);
1870     }
1871 }
1872
1873 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1874         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1875         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1876 {
1877     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1878     unsigned int i;
1879     HRESULT hr;
1880     WORD map;
1881
1882     if (!byte_code) return WINED3DERR_INVALIDCALL;
1883
1884     shader_init(shader, device, parent, parent_ops);
1885     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1886             WINED3D_SHADER_TYPE_VERTEX, max_version);
1887     if (FAILED(hr))
1888     {
1889         WARN("Failed to set function, hr %#x.\n", hr);
1890         shader_cleanup(shader);
1891         return hr;
1892     }
1893
1894     map = reg_maps->input_registers;
1895     for (i = 0; map; map >>= 1, ++i)
1896     {
1897         if (!(map & 1) || !shader->input_signature[i].semantic_name)
1898             continue;
1899
1900         shader->u.vs.attributes[i].usage =
1901                 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1902         shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1903     }
1904
1905     if (output_signature)
1906     {
1907         for (i = 0; i < output_signature->element_count; ++i)
1908         {
1909             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1910             reg_maps->output_registers |= 1 << e->register_idx;
1911             shader->output_signature[e->register_idx] = *e;
1912         }
1913     }
1914
1915     vertexshader_set_limits(shader);
1916
1917     shader->load_local_constsF = reg_maps->usesrelconstF
1918             && !list_empty(&shader->constantsF);
1919
1920     return WINED3D_OK;
1921 }
1922
1923 static void geometryshader_set_limits(struct wined3d_shader *shader)
1924 {
1925     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1926             shader->reg_maps.shader_version.minor);
1927
1928     switch (shader_version)
1929     {
1930         case WINED3D_SHADER_VERSION(4, 0):
1931             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1932             shader->limits.constant_int = 0;
1933             shader->limits.constant_float = 0;
1934             shader->limits.constant_bool = 0;
1935             shader->limits.packed_output = 32;
1936             shader->limits.packed_input = 16;
1937             break;
1938
1939         default:
1940             memset(&shader->limits, 0, sizeof(shader->limits));
1941             FIXME("Unhandled geometry shader version \"%u.%u\".\n",
1942                     shader->reg_maps.shader_version.major,
1943                     shader->reg_maps.shader_version.minor);
1944     }
1945 }
1946
1947 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1948         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1949         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1950 {
1951     HRESULT hr;
1952
1953     shader_init(shader, device, parent, parent_ops);
1954     hr = shader_set_function(shader, byte_code, output_signature, 0,
1955             WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1956     if (FAILED(hr))
1957     {
1958         WARN("Failed to set function, hr %#x.\n", hr);
1959         shader_cleanup(shader);
1960         return hr;
1961     }
1962
1963     geometryshader_set_limits(shader);
1964
1965     shader->load_local_constsF = FALSE;
1966
1967     return WINED3D_OK;
1968 }
1969
1970 void find_ps_compile_args(const struct wined3d_state *state,
1971         const struct wined3d_shader *shader, struct ps_compile_args *args)
1972 {
1973     struct wined3d_device *device = shader->device;
1974     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1975     const struct wined3d_texture *texture;
1976     UINT i;
1977
1978     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1979     if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1980     {
1981         const struct wined3d_surface *rt = state->fb->render_targets[0];
1982         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
1983         {
1984             static unsigned int warned = 0;
1985
1986             args->srgb_correction = 1;
1987             if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
1988                 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
1989                         "support, expect rendering artifacts.\n");
1990         }
1991     }
1992
1993     if (shader->reg_maps.shader_version.major == 1
1994             && shader->reg_maps.shader_version.minor <= 3)
1995     {
1996         for (i = 0; i < 4; ++i)
1997         {
1998             DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
1999
2000             if (flags & WINED3D_TTFF_PROJECTED)
2001             {
2002                 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2003
2004                 if (!state->vertex_shader)
2005                 {
2006                     unsigned int j;
2007                     unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2008                     DWORD max_valid = WINED3D_TTFF_COUNT4;
2009                     enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
2010
2011                     for (j = 0; j < state->vertex_declaration->element_count; ++j)
2012                     {
2013                         struct wined3d_vertex_declaration_element *element =
2014                                 &state->vertex_declaration->elements[j];
2015
2016                         if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2017                                 && element->usage_idx == index)
2018                         {
2019                             max_valid = element->format->component_count;
2020                             break;
2021                         }
2022                     }
2023                     if (!tex_transform || tex_transform > max_valid)
2024                     {
2025                         WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2026                                 tex_transform, max_valid);
2027                         tex_transform = max_valid;
2028                     }
2029                     if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
2030                             || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
2031                             || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
2032                         tex_transform |= WINED3D_PSARGS_PROJECTED;
2033                     else
2034                     {
2035                         WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2036                         WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2037                                 i, tex_transform, sampler_type);
2038                     }
2039                 }
2040                 else
2041                     tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2042
2043                 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2044             }
2045         }
2046     }
2047
2048     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2049     {
2050         if (!shader->reg_maps.sampler_type[i])
2051             continue;
2052
2053         texture = state->textures[i];
2054         if (!texture)
2055         {
2056             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2057             continue;
2058         }
2059         args->color_fixup[i] = texture->resource.format->color_fixup;
2060
2061         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2062             args->shadow |= 1 << i;
2063
2064         /* Flag samplers that need NP2 texcoord fixup. */
2065         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2066             args->np2_fixup |= (1 << i);
2067     }
2068     if (shader->reg_maps.shader_version.major >= 3)
2069     {
2070         if (device->strided_streams.position_transformed)
2071         {
2072             args->vp_mode = pretransformed;
2073         }
2074         else if (use_vs(state))
2075         {
2076             args->vp_mode = vertexshader;
2077         }
2078         else
2079         {
2080             args->vp_mode = fixedfunction;
2081         }
2082         args->fog = FOG_OFF;
2083     }
2084     else
2085     {
2086         args->vp_mode = vertexshader;
2087         if (state->render_states[WINED3D_RS_FOGENABLE])
2088         {
2089             switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2090             {
2091                 case WINED3D_FOG_NONE:
2092                     if (device->strided_streams.position_transformed || use_vs(state))
2093                     {
2094                         args->fog = FOG_LINEAR;
2095                         break;
2096                     }
2097
2098                     switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2099                     {
2100                         case WINED3D_FOG_NONE: /* Fall through. */
2101                         case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2102                         case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
2103                         case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
2104                     }
2105                     break;
2106
2107                 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2108                 case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
2109                 case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
2110             }
2111         }
2112         else
2113         {
2114             args->fog = FOG_OFF;
2115         }
2116     }
2117 }
2118
2119 static void pixelshader_set_limits(struct wined3d_shader *shader)
2120 {
2121     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2122             shader->reg_maps.shader_version.minor);
2123
2124     shader->limits.packed_output = 0;
2125
2126     switch (shader_version)
2127     {
2128         case WINED3D_SHADER_VERSION(1, 0):
2129         case WINED3D_SHADER_VERSION(1, 1):
2130         case WINED3D_SHADER_VERSION(1, 2):
2131         case WINED3D_SHADER_VERSION(1, 3):
2132             shader->limits.constant_float = 8;
2133             shader->limits.constant_int = 0;
2134             shader->limits.constant_bool = 0;
2135             shader->limits.sampler = 4;
2136             shader->limits.packed_input = 0;
2137             break;
2138
2139         case WINED3D_SHADER_VERSION(1, 4):
2140             shader->limits.constant_float = 8;
2141             shader->limits.constant_int = 0;
2142             shader->limits.constant_bool = 0;
2143             shader->limits.sampler = 6;
2144             shader->limits.packed_input = 0;
2145             break;
2146
2147         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2148         case WINED3D_SHADER_VERSION(2, 0):
2149             shader->limits.constant_float = 32;
2150             shader->limits.constant_int = 16;
2151             shader->limits.constant_bool = 16;
2152             shader->limits.sampler = 16;
2153             shader->limits.packed_input = 0;
2154             break;
2155
2156         case WINED3D_SHADER_VERSION(2, 1):
2157             shader->limits.constant_float = 32;
2158             shader->limits.constant_int = 16;
2159             shader->limits.constant_bool = 16;
2160             shader->limits.sampler = 16;
2161             shader->limits.packed_input = 0;
2162             break;
2163
2164         case WINED3D_SHADER_VERSION(3, 0):
2165             shader->limits.constant_float = 224;
2166             shader->limits.constant_int = 16;
2167             shader->limits.constant_bool = 16;
2168             shader->limits.sampler = 16;
2169             shader->limits.packed_input = 12;
2170             break;
2171
2172         case WINED3D_SHADER_VERSION(4, 0):
2173             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
2174             shader->limits.constant_int = 0;
2175             shader->limits.constant_float = 0;
2176             shader->limits.constant_bool = 0;
2177             shader->limits.packed_input = 32;
2178             break;
2179
2180         default:
2181             shader->limits.constant_float = 32;
2182             shader->limits.constant_int = 16;
2183             shader->limits.constant_bool = 16;
2184             shader->limits.sampler = 16;
2185             shader->limits.packed_input = 0;
2186             FIXME("Unrecognized pixel shader version %u.%u\n",
2187                     shader->reg_maps.shader_version.major,
2188                     shader->reg_maps.shader_version.minor);
2189     }
2190 }
2191
2192 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2193         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2194         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2195 {
2196     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2197     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2198     HRESULT hr;
2199
2200     if (!byte_code) return WINED3DERR_INVALIDCALL;
2201
2202     shader_init(shader, device, parent, parent_ops);
2203     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2204             WINED3D_SHADER_TYPE_PIXEL, max_version);
2205     if (FAILED(hr))
2206     {
2207         WARN("Failed to set function, hr %#x.\n", hr);
2208         shader_cleanup(shader);
2209         return hr;
2210     }
2211
2212     pixelshader_set_limits(shader);
2213
2214     for (i = 0; i < MAX_REG_INPUT; ++i)
2215     {
2216         if (shader->u.ps.input_reg_used[i])
2217         {
2218             ++num_regs_used;
2219             highest_reg_used = i;
2220         }
2221     }
2222
2223     /* Don't do any register mapping magic if it is not needed, or if we can't
2224      * achieve anything anyway */
2225     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2226             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2227     {
2228         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2229         {
2230             /* This happens with relative addressing. The input mapper function
2231              * warns about this if the higher registers are declared too, so
2232              * don't write a FIXME here */
2233             WARN("More varying registers used than supported\n");
2234         }
2235
2236         for (i = 0; i < MAX_REG_INPUT; ++i)
2237         {
2238             shader->u.ps.input_reg_map[i] = i;
2239         }
2240
2241         shader->u.ps.declared_in_count = highest_reg_used + 1;
2242     }
2243     else
2244     {
2245         shader->u.ps.declared_in_count = 0;
2246         for (i = 0; i < MAX_REG_INPUT; ++i)
2247         {
2248             if (shader->u.ps.input_reg_used[i])
2249                 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2250             else shader->u.ps.input_reg_map[i] = ~0U;
2251         }
2252     }
2253
2254     shader->load_local_constsF = FALSE;
2255
2256     return WINED3D_OK;
2257 }
2258
2259 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2260 {
2261     enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2262     unsigned int i;
2263
2264     if (reg_maps->shader_version.major != 1) return;
2265
2266     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2267     {
2268         /* We don't sample from this sampler. */
2269         if (!sampler_type[i]) continue;
2270
2271         if (!textures[i])
2272         {
2273             WARN("No texture bound to sampler %u, using 2D.\n", i);
2274             sampler_type[i] = WINED3DSTT_2D;
2275             continue;
2276         }
2277
2278         switch (textures[i]->target)
2279         {
2280             case GL_TEXTURE_RECTANGLE_ARB:
2281             case GL_TEXTURE_2D:
2282                 /* We have to select between texture rectangles and 2D
2283                  * textures later because 2.0 and 3.0 shaders only have
2284                  * WINED3DSTT_2D as well. */
2285                 sampler_type[i] = WINED3DSTT_2D;
2286                 break;
2287
2288             case GL_TEXTURE_3D:
2289                 sampler_type[i] = WINED3DSTT_VOLUME;
2290                 break;
2291
2292             case GL_TEXTURE_CUBE_MAP_ARB:
2293                 sampler_type[i] = WINED3DSTT_CUBE;
2294                 break;
2295
2296             default:
2297                 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2298                 sampler_type[i] = WINED3DSTT_2D;
2299         }
2300     }
2301 }
2302
2303 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2304         const struct wined3d_shader_signature *output_signature, void *parent,
2305         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2306 {
2307     struct wined3d_shader *object;
2308     HRESULT hr;
2309
2310     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2311             device, byte_code, output_signature, parent, parent_ops, shader);
2312
2313     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2314     if (!object)
2315     {
2316         ERR("Failed to allocate shader memory.\n");
2317         return E_OUTOFMEMORY;
2318     }
2319
2320     hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2321     if (FAILED(hr))
2322     {
2323         WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2324         HeapFree(GetProcessHeap(), 0, object);
2325         return hr;
2326     }
2327
2328     TRACE("Created geometry shader %p.\n", object);
2329     *shader = object;
2330
2331     return WINED3D_OK;
2332 }
2333
2334 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2335         const struct wined3d_shader_signature *output_signature, void *parent,
2336         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2337 {
2338     struct wined3d_shader *object;
2339     HRESULT hr;
2340
2341     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2342             device, byte_code, output_signature, parent, parent_ops, shader);
2343
2344     if (device->ps_selected_mode == SHADER_NONE)
2345         return WINED3DERR_INVALIDCALL;
2346
2347     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2348     if (!object)
2349     {
2350         ERR("Failed to allocate shader memory.\n");
2351         return E_OUTOFMEMORY;
2352     }
2353
2354     hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2355     if (FAILED(hr))
2356     {
2357         WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2358         HeapFree(GetProcessHeap(), 0, object);
2359         return hr;
2360     }
2361
2362     TRACE("Created pixel shader %p.\n", object);
2363     *shader = object;
2364
2365     return WINED3D_OK;
2366 }
2367
2368 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2369         const struct wined3d_shader_signature *output_signature, void *parent,
2370         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2371 {
2372     struct wined3d_shader *object;
2373     HRESULT hr;
2374
2375     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2376             device, byte_code, output_signature, parent, parent_ops, shader);
2377
2378     if (device->vs_selected_mode == SHADER_NONE)
2379         return WINED3DERR_INVALIDCALL;
2380
2381     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2382     if (!object)
2383     {
2384         ERR("Failed to allocate shader memory.\n");
2385         return E_OUTOFMEMORY;
2386     }
2387
2388     hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2389     if (FAILED(hr))
2390     {
2391         WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2392         HeapFree(GetProcessHeap(), 0, object);
2393         return hr;
2394     }
2395
2396     TRACE("Created vertex shader %p.\n", object);
2397     *shader = object;
2398
2399     return WINED3D_OK;
2400 }