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_DEF)
 548         {
 549             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 550             if (!lconst) return E_OUTOFMEMORY;
 551
 552             lconst->idx = ins.dst[0].reg.idx[0].offset;
 553             memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
 554
 555             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 556             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 557             {
 558                 float *value = (float *)lconst->value;
 559                 if (value[0] < -1.0f) value[0] = -1.0f;
 560                 else if (value[0] > 1.0f) value[0] = 1.0f;
 561                 if (value[1] < -1.0f) value[1] = -1.0f;
 562                 else if (value[1] > 1.0f) value[1] = 1.0f;
 563                 if (value[2] < -1.0f) value[2] = -1.0f;
 564                 else if (value[2] > 1.0f) value[2] = 1.0f;
 565                 if (value[3] < -1.0f) value[3] = -1.0f;
 566                 else if (value[3] > 1.0f) value[3] = 1.0f;
 567             }
 568
 569             list_add_head(&shader->constantsF, &lconst->entry);
 570         }
 571         else if (ins.handler_idx == WINED3DSIH_DEFI)
 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             list_add_head(&shader->constantsI, &lconst->entry);
 580             reg_maps->local_int_consts |= (1 << lconst->idx);
 581         }
 582         else if (ins.handler_idx == WINED3DSIH_DEFB)
 583         {
 584             struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
 585             if (!lconst) return E_OUTOFMEMORY;
 586
 587             lconst->idx = ins.dst[0].reg.idx[0].offset;
 588             memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
 589
 590             list_add_head(&shader->constantsB, &lconst->entry);
 591             reg_maps->local_bool_consts |= (1 << lconst->idx);
 592         }
 593         /* For subroutine prototypes. */
 594         else if (ins.handler_idx == WINED3DSIH_LABEL)
 595         {
 596             reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
 597         }
 598         /* Set texture, address, temporary registers. */
 599         else
 600         {
 601             BOOL color0_mov = FALSE;
 602             unsigned int i;
 603
 604             /* This will loop over all the registers and try to
 605              * make a bitmask of the ones we're interested in.
 606              *
 607              * Relative addressing tokens are ignored, but that's
 608              * okay, since we'll catch any address registers when
 609              * they are initialized (required by spec). */
 610             for (i = 0; i < ins.dst_count; ++i)
 611             {
 612                 shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg, shader_version.type);
 613
 614                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 615                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 616                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 617                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 618                 {
 619                     UINT idx = ins.dst[i].reg.idx[0].offset;
 620
 621                     switch (ins.dst[i].reg.type)
 622                     {
 623                         case WINED3DSPR_RASTOUT:
 624                             switch (idx)
 625                             {
 626                                 case 0: /* oPos */
 627                                     reg_maps->output_registers |= 1 << 10;
 628                                     shader_signature_from_usage(&output_signature[10],
 629                                             WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 630                                     break;
 631
 632                                 case 1: /* oFog */
 633                                     reg_maps->output_registers |= 1 << 11;
 634                                     shader_signature_from_usage(&output_signature[11],
 635                                             WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 636                                     break;
 637
 638                                 case 2: /* oPts */
 639                                     reg_maps->output_registers |= 1 << 11;
 640                                     shader_signature_from_usage(&output_signature[11],
 641                                             WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 642                                     break;
 643                             }
 644                             break;
 645
 646                         case WINED3DSPR_ATTROUT:
 647                             if (idx < 2)
 648                             {
 649                                 idx += 8;
 650                                 if (reg_maps->output_registers & (1 << idx))
 651                                 {
 652                                     output_signature[idx].mask |= ins.dst[i].write_mask;
 653                                 }
 654                                 else
 655                                 {
 656                                     reg_maps->output_registers |= 1 << idx;
 657                                     shader_signature_from_usage(&output_signature[idx],
 658                                             WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
 659                                 }
 660                             }
 661                             break;
 662
 663                         case WINED3DSPR_TEXCRDOUT:
 664
 665                             reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
 666                             if (reg_maps->output_registers & (1 << idx))
 667                             {
 668                                 output_signature[idx].mask |= ins.dst[i].write_mask;
 669                             }
 670                             else
 671                             {
 672                                 reg_maps->output_registers |= 1 << idx;
 673                                 shader_signature_from_usage(&output_signature[idx],
 674                                         WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
 675                             }
 676                             break;
 677
 678                         default:
 679                             break;
 680                     }
 681                 }
 682
 683                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 684                 {
 685                     if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
 686                     {
 687                         /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 688                          * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 689                          * the mov and perform the sRGB write correction from the source register.
 690                          *
 691                          * However, if the mov is only partial, we can't do this, and if the write
 692                          * comes from an instruction other than MOV it is hard to do as well. If
 693                          * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 694                         shader->u.ps.color0_mov = FALSE;
 695                         if (ins.handler_idx == WINED3DSIH_MOV
 696                                 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
 697                         {
 698                             /* Used later when the source register is read. */
 699                             color0_mov = TRUE;
 700                         }
 701                     }
 702                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 703                      * end
 704                      */
 705                     else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
 706                             && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
 707                     {
 708                         shader->u.ps.color0_mov = FALSE;
 709                     }
 710                 }
 711
 712                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 713                 if (shader_version.major == 1
 714                         && (ins.handler_idx == WINED3DSIH_TEX
 715                             || ins.handler_idx == WINED3DSIH_TEXBEM
 716                             || ins.handler_idx == WINED3DSIH_TEXBEML
 717                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 718                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 719                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 720                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 721                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 722                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 723                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 724                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 725                 {
 726                     /* Fake sampler usage, only set reserved bit and type. */
 727                     DWORD sampler_code = ins.dst[i].reg.idx[0].offset;
 728
 729                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 730                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 731
 732                     /* texbem is only valid with < 1.4 pixel shaders */
 733                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 734                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 735                     {
 736                         reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
 737                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 738                         {
 739                             reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx[0].offset;
 740                         }
 741                     }
 742                 }
 743                 else if (ins.handler_idx == WINED3DSIH_BEM)
 744                 {
 745                     reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
 746                 }
 747             }
 748
 749             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 750             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 751             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 752             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 753             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 754             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 755             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 756             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 757             else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
 758             else if (ins.handler_idx == WINED3DSIH_LOOP
 759                     || ins.handler_idx == WINED3DSIH_REP)
 760             {
 761                 ++cur_loop_depth;
 762                 if (cur_loop_depth > max_loop_depth)
 763                     max_loop_depth = cur_loop_depth;
 764             }
 765             else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 766                     || ins.handler_idx == WINED3DSIH_ENDREP)
 767                 --cur_loop_depth;
 768
 769             if (ins.predicate)
 770                 shader_record_register_usage(shader, reg_maps, &ins.predicate->reg, shader_version.type);
 771
 772             for (i = 0; i < ins.src_count; ++i)
 773             {
 774                 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
 775                 struct wined3d_shader_register reg = ins.src[i].reg;
 776
 777                 shader_record_register_usage(shader, reg_maps, &ins.src[i].reg, shader_version.type);
 778                 while (count)
 779                 {
 780                     ++reg.idx[0].offset;
 781                     shader_record_register_usage(shader, reg_maps, &reg, shader_version.type);
 782                     --count;
 783                 }
 784
 785                 if (color0_mov)
 786                 {
 787                     if (ins.src[i].reg.type == WINED3DSPR_TEMP
 788                             && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
 789                     {
 790                         shader->u.ps.color0_mov = TRUE;
 791                         shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
 792                     }
 793                 }
 794             }
 795         }
 796     }
 797     reg_maps->loop_depth = max_loop_depth;
 798
 799     /* PS before 2.0 don't have explicit color outputs. Instead the value of
 800      * R0 is written to the render target. */
 801     if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 802         reg_maps->rt_mask |= (1 << 0);
 803
 804     shader->functionLength = ((const char *)ptr - (const char *)byte_code);
 805
 806     return WINED3D_OK;
 807 }
 808
 809 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 810 {
 811     DWORD map = 1 << max;
 812     map |= map - 1;
 813     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 814
 815     return wined3d_log2i(map);
 816 }
 817
 818 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 819         const struct wined3d_shader_version *shader_version)
 820 {
 821     TRACE("dcl");
 822
 823     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 824     {
 825         switch (semantic->sampler_type)
 826         {
 827             case WINED3DSTT_2D: TRACE("_2d"); break;
 828             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 829             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 830             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 831         }
 832     }
 833     else
 834     {
 835         /* Pixel shaders 3.0 don't have usage semantics. */
 836         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 837         else TRACE("_");
 838
 839         switch (semantic->usage)
 840         {
 841             case WINED3D_DECL_USAGE_POSITION:
 842                 TRACE("position%u", semantic->usage_idx);
 843                 break;
 844
 845             case WINED3D_DECL_USAGE_BLEND_INDICES:
 846                 TRACE("blend");
 847                 break;
 848
 849             case WINED3D_DECL_USAGE_BLEND_WEIGHT:
 850                 TRACE("weight");
 851                 break;
 852
 853             case WINED3D_DECL_USAGE_NORMAL:
 854                 TRACE("normal%u", semantic->usage_idx);
 855                 break;
 856
 857             case WINED3D_DECL_USAGE_PSIZE:
 858                 TRACE("psize");
 859                 break;
 860
 861             case WINED3D_DECL_USAGE_COLOR:
 862                 if (!semantic->usage_idx) TRACE("color");
 863                 else TRACE("specular%u", (semantic->usage_idx - 1));
 864                 break;
 865
 866             case WINED3D_DECL_USAGE_TEXCOORD:
 867                 TRACE("texture%u", semantic->usage_idx);
 868                 break;
 869
 870             case WINED3D_DECL_USAGE_TANGENT:
 871                 TRACE("tangent");
 872                 break;
 873
 874             case WINED3D_DECL_USAGE_BINORMAL:
 875                 TRACE("binormal");
 876                 break;
 877
 878             case WINED3D_DECL_USAGE_TESS_FACTOR:
 879                 TRACE("tessfactor");
 880                 break;
 881
 882             case WINED3D_DECL_USAGE_POSITIONT:
 883                 TRACE("positionT%u", semantic->usage_idx);
 884                 break;
 885
 886             case WINED3D_DECL_USAGE_FOG:
 887                 TRACE("fog");
 888                 break;
 889
 890             case WINED3D_DECL_USAGE_DEPTH:
 891                 TRACE("depth");
 892                 break;
 893
 894             case WINED3D_DECL_USAGE_SAMPLE:
 895                 TRACE("sample");
 896                 break;
 897
 898             default:
 899                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 900         }
 901     }
 902 }
 903
 904 static void shader_dump_register(const struct wined3d_shader_register *reg,
 905         const struct wined3d_shader_version *shader_version)
 906 {
 907     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 908     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 909     UINT offset = reg->idx[0].offset;
 910
 911     switch (reg->type)
 912     {
 913         case WINED3DSPR_TEMP:
 914             TRACE("r");
 915             break;
 916
 917         case WINED3DSPR_INPUT:
 918             TRACE("v");
 919             break;
 920
 921         case WINED3DSPR_CONST:
 922         case WINED3DSPR_CONST2:
 923         case WINED3DSPR_CONST3:
 924         case WINED3DSPR_CONST4:
 925             TRACE("c");
 926             offset = shader_get_float_offset(reg->type, offset);
 927             break;
 928
 929         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 930             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 931             break;
 932
 933         case WINED3DSPR_RASTOUT:
 934             TRACE("%s", rastout_reg_names[offset]);
 935             break;
 936
 937         case WINED3DSPR_COLOROUT:
 938             TRACE("oC");
 939             break;
 940
 941         case WINED3DSPR_DEPTHOUT:
 942             TRACE("oDepth");
 943             break;
 944
 945         case WINED3DSPR_ATTROUT:
 946             TRACE("oD");
 947             break;
 948
 949         case WINED3DSPR_TEXCRDOUT:
 950             /* Vertex shaders >= 3.0 use general purpose output registers
 951              * (WINED3DSPR_OUTPUT), which can include an address token. */
 952             if (shader_version->major >= 3) TRACE("o");
 953             else TRACE("oT");
 954             break;
 955
 956         case WINED3DSPR_CONSTINT:
 957             TRACE("i");
 958             break;
 959
 960         case WINED3DSPR_CONSTBOOL:
 961             TRACE("b");
 962             break;
 963
 964         case WINED3DSPR_LABEL:
 965             TRACE("l");
 966             break;
 967
 968         case WINED3DSPR_LOOP:
 969             TRACE("aL");
 970             break;
 971
 972         case WINED3DSPR_SAMPLER:
 973             TRACE("s");
 974             break;
 975
 976         case WINED3DSPR_MISCTYPE:
 977             if (offset > 1)
 978                 FIXME("Unhandled misctype register %u.\n", offset);
 979             else
 980                 TRACE("%s", misctype_reg_names[offset]);
 981             break;
 982
 983         case WINED3DSPR_PREDICATE:
 984             TRACE("p");
 985             break;
 986
 987         case WINED3DSPR_IMMCONST:
 988             TRACE("l");
 989             break;
 990
 991         case WINED3DSPR_CONSTBUFFER:
 992             TRACE("cb");
 993             break;
 994
 995         case WINED3DSPR_PRIMID:
 996             TRACE("primID");
 997             break;
 998
 999         case WINED3DSPR_NULL:
1000             TRACE("null");
1001             break;
1002
1003         case WINED3DSPR_RESOURCE:
1004             TRACE("t");
1005             break;
1006
1007         default:
1008             TRACE("unhandled_rtype(%#x)", reg->type);
1009             break;
1010     }
1011
1012     if (reg->type == WINED3DSPR_IMMCONST)
1013     {
1014         TRACE("(");
1015         switch (reg->immconst_type)
1016         {
1017             case WINED3D_IMMCONST_SCALAR:
1018                 switch (reg->data_type)
1019                 {
1020                     case WINED3D_DATA_FLOAT:
1021                         TRACE("%.8e", *(const float *)reg->immconst_data);
1022                         break;
1023                     case WINED3D_DATA_INT:
1024                         TRACE("%d", reg->immconst_data[0]);
1025                         break;
1026                     case WINED3D_DATA_RESOURCE:
1027                     case WINED3D_DATA_SAMPLER:
1028                     case WINED3D_DATA_UINT:
1029                         TRACE("%u", reg->immconst_data[0]);
1030                         break;
1031                     default:
1032                         TRACE("<unhandled data type %#x>", reg->data_type);
1033                         break;
1034                 }
1035                 break;
1036
1037             case WINED3D_IMMCONST_VEC4:
1038                 switch (reg->data_type)
1039                 {
1040                     case WINED3D_DATA_FLOAT:
1041                         TRACE("%.8e, %.8e, %.8e, %.8e",
1042                                 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1043                                 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1044                         break;
1045                     case WINED3D_DATA_INT:
1046                         TRACE("%d, %d, %d, %d",
1047                                 reg->immconst_data[0], reg->immconst_data[1],
1048                                 reg->immconst_data[2], reg->immconst_data[3]);
1049                         break;
1050                     case WINED3D_DATA_RESOURCE:
1051                     case WINED3D_DATA_SAMPLER:
1052                     case WINED3D_DATA_UINT:
1053                         TRACE("%u, %u, %u, %u",
1054                                 reg->immconst_data[0], reg->immconst_data[1],
1055                                 reg->immconst_data[2], reg->immconst_data[3]);
1056                         break;
1057                     default:
1058                         TRACE("<unhandled data type %#x>", reg->data_type);
1059                         break;
1060                 }
1061                 break;
1062
1063             default:
1064                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1065                 break;
1066         }
1067         TRACE(")");
1068     }
1069     else if (reg->type != WINED3DSPR_RASTOUT
1070             && reg->type != WINED3DSPR_MISCTYPE
1071             && reg->type != WINED3DSPR_NULL)
1072     {
1073         if (offset != ~0U)
1074         {
1075             TRACE("[");
1076             if (reg->idx[0].rel_addr)
1077             {
1078                 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1079                 TRACE(" + ");
1080             }
1081             TRACE("%u]", offset);
1082
1083             if (reg->idx[1].offset != ~0U)
1084             {
1085                 TRACE("[");
1086                 if (reg->idx[1].rel_addr)
1087                 {
1088                     shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1089                     TRACE(" + ");
1090                 }
1091                 TRACE("%u]", reg->idx[1].offset);
1092             }
1093         }
1094     }
1095 }
1096
1097 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1098         const struct wined3d_shader_version *shader_version)
1099 {
1100     DWORD write_mask = param->write_mask;
1101
1102     shader_dump_register(&param->reg, shader_version);
1103
1104     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1105     {
1106         static const char *write_mask_chars = "xyzw";
1107
1108         TRACE(".");
1109         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1110         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1111         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1112         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1113     }
1114 }
1115
1116 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1117         const struct wined3d_shader_version *shader_version)
1118 {
1119     enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1120     DWORD swizzle = param->swizzle;
1121
1122     if (src_modifier == WINED3DSPSM_NEG
1123             || src_modifier == WINED3DSPSM_BIASNEG
1124             || src_modifier == WINED3DSPSM_SIGNNEG
1125             || src_modifier == WINED3DSPSM_X2NEG
1126             || src_modifier == WINED3DSPSM_ABSNEG)
1127         TRACE("-");
1128     else if (src_modifier == WINED3DSPSM_COMP)
1129         TRACE("1-");
1130     else if (src_modifier == WINED3DSPSM_NOT)
1131         TRACE("!");
1132
1133     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1134         TRACE("abs(");
1135
1136     shader_dump_register(&param->reg, shader_version);
1137
1138     if (src_modifier)
1139     {
1140         switch (src_modifier)
1141         {
1142             case WINED3DSPSM_NONE:    break;
1143             case WINED3DSPSM_NEG:     break;
1144             case WINED3DSPSM_NOT:     break;
1145             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1146             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1147             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1148             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1149             case WINED3DSPSM_COMP:    break;
1150             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1151             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1152             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1153             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1154             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1155             case WINED3DSPSM_ABS:     TRACE(")"); break;
1156             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1157         }
1158     }
1159
1160     if (swizzle != WINED3DSP_NOSWIZZLE)
1161     {
1162         static const char *swizzle_chars = "xyzw";
1163         DWORD swizzle_x = swizzle & 0x03;
1164         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1165         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1166         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1167
1168         if (swizzle_x == swizzle_y
1169                 && swizzle_x == swizzle_z
1170                 && swizzle_x == swizzle_w)
1171         {
1172             TRACE(".%c", swizzle_chars[swizzle_x]);
1173         }
1174         else
1175         {
1176             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1177                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1178         }
1179     }
1180 }
1181
1182 /* Shared code in order to generate the bulk of the shader string.
1183  * NOTE: A description of how to parse tokens can be found on MSDN. */
1184 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1185         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1186 {
1187     struct wined3d_device *device = shader->device;
1188     const struct wined3d_shader_frontend *fe = shader->frontend;
1189     void *fe_data = shader->frontend_data;
1190     struct wined3d_shader_version shader_version;
1191     struct wined3d_shader_loop_state loop_state;
1192     struct wined3d_shader_instruction ins;
1193     struct wined3d_shader_tex_mx tex_mx;
1194     struct wined3d_shader_context ctx;
1195     const DWORD *ptr = byte_code;
1196
1197     /* Initialize current parsing state. */
1198     tex_mx.current_row = 0;
1199     loop_state.current_depth = 0;
1200     loop_state.current_reg = 0;
1201
1202     ctx.shader = shader;
1203     ctx.gl_info = &device->adapter->gl_info;
1204     ctx.reg_maps = reg_maps;
1205     ctx.buffer = buffer;
1206     ctx.tex_mx = &tex_mx;
1207     ctx.loop_state = &loop_state;
1208     ctx.backend_data = backend_ctx;
1209     ins.ctx = &ctx;
1210
1211     fe->shader_read_header(fe_data, &ptr, &shader_version);
1212
1213     while (!fe->shader_is_end(fe_data, &ptr))
1214     {
1215         /* Read opcode. */
1216         fe->shader_read_instruction(fe_data, &ptr, &ins);
1217
1218         /* Unknown opcode and its parameters. */
1219         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1220         {
1221             TRACE("Skipping unrecognized instruction.\n");
1222             continue;
1223         }
1224
1225         if (ins.predicate)
1226             FIXME("Predicates not implemented.\n");
1227
1228         /* Call appropriate function for output target */
1229         device->shader_backend->shader_handle_instruction(&ins);
1230     }
1231 }
1232
1233 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1234 {
1235     DWORD mmask = dst->modifiers;
1236
1237     switch (dst->shift)
1238     {
1239         case 0: break;
1240         case 13: TRACE("_d8"); break;
1241         case 14: TRACE("_d4"); break;
1242         case 15: TRACE("_d2"); break;
1243         case 1: TRACE("_x2"); break;
1244         case 2: TRACE("_x4"); break;
1245         case 3: TRACE("_x8"); break;
1246         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1247     }
1248
1249     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1250     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1251     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1252
1253     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1254     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1255 }
1256
1257 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1258 {
1259     switch (primitive_type)
1260     {
1261         case WINED3D_PT_UNDEFINED:
1262             TRACE("undefined");
1263             break;
1264         case WINED3D_PT_POINTLIST:
1265             TRACE("pointlist");
1266             break;
1267         case WINED3D_PT_LINELIST:
1268             TRACE("linelist");
1269             break;
1270         case WINED3D_PT_LINESTRIP:
1271             TRACE("linestrip");
1272             break;
1273         case WINED3D_PT_TRIANGLELIST:
1274             TRACE("trianglelist");
1275             break;
1276         case WINED3D_PT_TRIANGLESTRIP:
1277             TRACE("trianglestrip");
1278             break;
1279         case WINED3D_PT_TRIANGLEFAN:
1280             TRACE("trianglefan");
1281             break;
1282         case WINED3D_PT_LINELIST_ADJ:
1283             TRACE("linelist_adj");
1284             break;
1285         case WINED3D_PT_LINESTRIP_ADJ:
1286             TRACE("linestrip_adj");
1287             break;
1288         case WINED3D_PT_TRIANGLELIST_ADJ:
1289             TRACE("trianglelist_adj");
1290             break;
1291         case WINED3D_PT_TRIANGLESTRIP_ADJ:
1292             TRACE("trianglestrip_adj");
1293             break;
1294         default:
1295             TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1296             break;
1297     }
1298 }
1299
1300 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1301 {
1302     struct wined3d_shader_version shader_version;
1303     const DWORD *ptr = byte_code;
1304     const char *type_prefix;
1305     DWORD i;
1306
1307     TRACE("Parsing %p.\n", byte_code);
1308
1309     fe->shader_read_header(fe_data, &ptr, &shader_version);
1310
1311     switch (shader_version.type)
1312     {
1313         case WINED3D_SHADER_TYPE_VERTEX:
1314             type_prefix = "vs";
1315             break;
1316
1317         case WINED3D_SHADER_TYPE_GEOMETRY:
1318             type_prefix = "gs";
1319             break;
1320
1321         case WINED3D_SHADER_TYPE_PIXEL:
1322             type_prefix = "ps";
1323             break;
1324
1325         default:
1326             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1327             type_prefix = "unknown";
1328             break;
1329     }
1330
1331     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1332
1333     while (!fe->shader_is_end(fe_data, &ptr))
1334     {
1335         struct wined3d_shader_instruction ins;
1336
1337         fe->shader_read_instruction(fe_data, &ptr, &ins);
1338         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1339         {
1340             TRACE("Skipping unrecognized instruction.\n");
1341             continue;
1342         }
1343
1344         if (ins.handler_idx == WINED3DSIH_DCL)
1345         {
1346             shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1347             shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1348             TRACE(" ");
1349             shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1350         }
1351         else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1352         {
1353             TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1354             shader_dump_src_param(&ins.declaration.src, &shader_version);
1355             TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1356         }
1357         else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1358                 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1359         {
1360             TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1361             shader_dump_primitive_type(ins.declaration.primitive_type);
1362         }
1363         else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1364         {
1365             TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1366         }
1367         else if (ins.handler_idx == WINED3DSIH_DEF)
1368         {
1369             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1370                     ins.dst[0].reg.idx[0].offset),
1371                     *(const float *)&ins.src[0].reg.immconst_data[0],
1372                     *(const float *)&ins.src[0].reg.immconst_data[1],
1373                     *(const float *)&ins.src[0].reg.immconst_data[2],
1374                     *(const float *)&ins.src[0].reg.immconst_data[3]);
1375         }
1376         else if (ins.handler_idx == WINED3DSIH_DEFI)
1377         {
1378             TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1379                     ins.src[0].reg.immconst_data[0],
1380                     ins.src[0].reg.immconst_data[1],
1381                     ins.src[0].reg.immconst_data[2],
1382                     ins.src[0].reg.immconst_data[3]);
1383         }
1384         else if (ins.handler_idx == WINED3DSIH_DEFB)
1385         {
1386             TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1387         }
1388         else
1389         {
1390             if (ins.predicate)
1391             {
1392                 TRACE("(");
1393                 shader_dump_src_param(ins.predicate, &shader_version);
1394                 TRACE(") ");
1395             }
1396
1397             /* PixWin marks instructions with the coissue flag with a '+' */
1398             if (ins.coissue) TRACE("+");
1399
1400             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1401
1402             if (ins.handler_idx == WINED3DSIH_IFC
1403                     || ins.handler_idx == WINED3DSIH_BREAKC)
1404             {
1405                 switch (ins.flags)
1406                 {
1407                     case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1408                     case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1409                     case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1410                     case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1411                     case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1412                     case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1413                     default: TRACE("_(%u)", ins.flags);
1414                 }
1415             }
1416             else if (ins.handler_idx == WINED3DSIH_TEX
1417                     && shader_version.major >= 2
1418                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1419             {
1420                 TRACE("p");
1421             }
1422
1423             for (i = 0; i < ins.dst_count; ++i)
1424             {
1425                 shader_dump_ins_modifiers(&ins.dst[i]);
1426                 TRACE(!i ? " " : ", ");
1427                 shader_dump_dst_param(&ins.dst[i], &shader_version);
1428             }
1429
1430             /* Other source tokens */
1431             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1432             {
1433                 TRACE(!i ? " " : ", ");
1434                 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1435             }
1436         }
1437         TRACE("\n");
1438     }
1439 }
1440
1441 static void shader_cleanup(struct wined3d_shader *shader)
1442 {
1443     shader->device->shader_backend->shader_destroy(shader);
1444     HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1445     HeapFree(GetProcessHeap(), 0, shader->function);
1446     shader_delete_constant_list(&shader->constantsF);
1447     shader_delete_constant_list(&shader->constantsB);
1448     shader_delete_constant_list(&shader->constantsI);
1449     list_remove(&shader->shader_list_entry);
1450
1451     if (shader->frontend && shader->frontend_data)
1452         shader->frontend->shader_free(shader->frontend_data);
1453 }
1454
1455 struct shader_none_priv
1456 {
1457     const struct fragment_pipeline *fragment_pipe;
1458 };
1459
1460 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1461 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1462         enum tex_types tex_type, const SIZE *ds_mask_size) {}
1463 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1464 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1465 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1466 static void shader_none_load_constants(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1467 static void shader_none_load_np2fixup_constants(void *shader_priv,
1468         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1469 static void shader_none_destroy(struct wined3d_shader *shader) {}
1470 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1471
1472 static void shader_none_select(const struct wined3d_context *context, enum wined3d_shader_mode vertex_mode,
1473         enum wined3d_shader_mode fragment_mode)
1474 {
1475     const struct wined3d_gl_info *gl_info = context->gl_info;
1476     struct wined3d_device *device = context->swapchain->device;
1477     struct shader_none_priv *priv = device->shader_priv;
1478
1479     priv->fragment_pipe->enable_extension(gl_info, fragment_mode == WINED3D_SHADER_MODE_FFP);
1480 }
1481
1482 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct fragment_pipeline *fragment_pipe)
1483 {
1484     struct shader_none_priv *priv;
1485     void *fragment_priv;
1486
1487     if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1488     {
1489         ERR("Failed to allocate private data.\n");
1490         return E_OUTOFMEMORY;
1491     }
1492
1493     if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1494     {
1495         ERR("Failed to initialize fragment pipe.\n");
1496         HeapFree(GetProcessHeap(), 0, priv);
1497         return E_FAIL;
1498     }
1499
1500     device->fragment_priv = fragment_priv;
1501     priv->fragment_pipe = fragment_pipe;
1502     device->shader_priv = priv;
1503
1504     return WINED3D_OK;
1505 }
1506
1507 static void shader_none_free(struct wined3d_device *device)
1508 {
1509     struct shader_none_priv *priv = device->shader_priv;
1510
1511     priv->fragment_pipe->free_private(device);
1512     HeapFree(GetProcessHeap(), 0, priv);
1513 }
1514
1515 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1516 {
1517     /* Set the shader caps to 0 for the none shader backend */
1518     caps->vs_version = 0;
1519     caps->gs_version = 0;
1520     caps->ps_version = 0;
1521     caps->vs_uniform_count = 0;
1522     caps->ps_uniform_count = 0;
1523     caps->ps_1x_max_value = 0.0f;
1524     caps->vs_clipping = FALSE;
1525 }
1526
1527 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1528 {
1529     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1530     {
1531         TRACE("Checking support for fixup:\n");
1532         dump_color_fixup_desc(fixup);
1533     }
1534
1535     /* Faked to make some apps happy. */
1536     if (!is_complex_fixup(fixup))
1537     {
1538         TRACE("[OK]\n");
1539         return TRUE;
1540     }
1541
1542     TRACE("[FAILED]\n");
1543     return FALSE;
1544 }
1545
1546 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
1547 {
1548     struct shader_none_priv *priv = shader_priv;
1549
1550     return priv->fragment_pipe->ffp_proj_control;
1551 }
1552
1553 const struct wined3d_shader_backend_ops none_shader_backend =
1554 {
1555     shader_none_handle_instruction,
1556     shader_none_select,
1557     shader_none_select_depth_blt,
1558     shader_none_deselect_depth_blt,
1559     shader_none_update_float_vertex_constants,
1560     shader_none_update_float_pixel_constants,
1561     shader_none_load_constants,
1562     shader_none_load_np2fixup_constants,
1563     shader_none_destroy,
1564     shader_none_alloc,
1565     shader_none_free,
1566     shader_none_context_destroyed,
1567     shader_none_get_caps,
1568     shader_none_color_fixup_supported,
1569     shader_none_has_ffp_proj_control,
1570 };
1571
1572 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1573         const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1574         enum wined3d_shader_type type, unsigned int max_version)
1575 {
1576     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1577     const struct wined3d_shader_frontend *fe;
1578     HRESULT hr;
1579     unsigned int backend_version;
1580
1581     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1582             shader, byte_code, output_signature, float_const_count);
1583
1584     fe = shader_select_frontend(*byte_code);
1585     if (!fe)
1586     {
1587         FIXME("Unable to find frontend for shader.\n");
1588         return WINED3DERR_INVALIDCALL;
1589     }
1590     shader->frontend = fe;
1591     shader->frontend_data = fe->shader_init(byte_code, output_signature);
1592     if (!shader->frontend_data)
1593     {
1594         FIXME("Failed to initialize frontend.\n");
1595         return WINED3DERR_INVALIDCALL;
1596     }
1597
1598     /* First pass: trace shader. */
1599     if (TRACE_ON(d3d_shader))
1600         shader_trace_init(fe, shader->frontend_data, byte_code);
1601
1602     /* Initialize immediate constant lists. */
1603     list_init(&shader->constantsF);
1604     list_init(&shader->constantsB);
1605     list_init(&shader->constantsI);
1606
1607     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1608     hr = shader_get_registers_used(shader, fe,
1609             reg_maps, shader->input_signature, shader->output_signature,
1610             byte_code, float_const_count);
1611     if (FAILED(hr)) return hr;
1612
1613     if (reg_maps->shader_version.type != type)
1614     {
1615         WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1616         return WINED3DERR_INVALIDCALL;
1617     }
1618     if (reg_maps->shader_version.major > max_version)
1619     {
1620         WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1621         return WINED3DERR_INVALIDCALL;
1622     }
1623     switch (type)
1624     {
1625         case WINED3D_SHADER_TYPE_VERTEX:
1626             backend_version = shader->device->vs_version;
1627             break;
1628         case WINED3D_SHADER_TYPE_GEOMETRY:
1629             backend_version = shader->device->gs_version;
1630             break;
1631         case WINED3D_SHADER_TYPE_PIXEL:
1632             backend_version = shader->device->ps_version;
1633             break;
1634         default:
1635             FIXME("No backend version-checking for this shader type\n");
1636             backend_version = 0;
1637     }
1638     if (reg_maps->shader_version.major > backend_version)
1639     {
1640         WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1641                 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1642         return WINED3DERR_INVALIDCALL;
1643     }
1644
1645     shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1646     if (!shader->function)
1647         return E_OUTOFMEMORY;
1648     memcpy(shader->function, byte_code, shader->functionLength);
1649
1650     return WINED3D_OK;
1651 }
1652
1653 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1654 {
1655     ULONG refcount = InterlockedIncrement(&shader->ref);
1656
1657     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1658
1659     return refcount;
1660 }
1661
1662 /* Do not call while under the GL lock. */
1663 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1664 {
1665     ULONG refcount = InterlockedDecrement(&shader->ref);
1666
1667     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1668
1669     if (!refcount)
1670     {
1671         shader_cleanup(shader);
1672         shader->parent_ops->wined3d_object_destroyed(shader->parent);
1673         HeapFree(GetProcessHeap(), 0, shader);
1674     }
1675
1676     return refcount;
1677 }
1678
1679 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1680 {
1681     TRACE("shader %p.\n", shader);
1682
1683     return shader->parent;
1684 }
1685
1686 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1687         void *byte_code, UINT *byte_code_size)
1688 {
1689     TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1690
1691     if (!byte_code)
1692     {
1693         *byte_code_size = shader->functionLength;
1694         return WINED3D_OK;
1695     }
1696
1697     if (*byte_code_size < shader->functionLength)
1698     {
1699         /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1700          * than the required size we should write the required size and
1701          * return D3DERR_MOREDATA. That's not actually true. */
1702         return WINED3DERR_INVALIDCALL;
1703     }
1704
1705     memcpy(byte_code, shader->function, shader->functionLength);
1706
1707     return WINED3D_OK;
1708 }
1709
1710 /* Set local constants for d3d8 shaders. */
1711 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1712         UINT start_idx, const float *src_data, UINT count)
1713 {
1714     UINT end_idx = start_idx + count;
1715     UINT i;
1716
1717     TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1718
1719     if (end_idx > shader->limits.constant_float)
1720     {
1721         WARN("end_idx %u > float constants limit %u.\n",
1722                 end_idx, shader->limits.constant_float);
1723         end_idx = shader->limits.constant_float;
1724     }
1725
1726     for (i = start_idx; i < end_idx; ++i)
1727     {
1728         struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1729         if (!lconst)
1730             return E_OUTOFMEMORY;
1731
1732         lconst->idx = i;
1733         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1734         list_add_head(&shader->constantsF, &lconst->entry);
1735     }
1736
1737     return WINED3D_OK;
1738 }
1739
1740 void find_vs_compile_args(const struct wined3d_state *state,
1741         const struct wined3d_shader *shader, struct vs_compile_args *args)
1742 {
1743     args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1744             == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1745     args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1746             && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1747     args->swizzle_map = shader->device->strided_streams.swizzle_map;
1748 }
1749
1750 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1751 {
1752     if (usage_idx1 != usage_idx2)
1753         return FALSE;
1754     if (usage1 == usage2)
1755         return TRUE;
1756     if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1757         return TRUE;
1758     if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1759         return TRUE;
1760
1761     return FALSE;
1762 }
1763
1764 BOOL vshader_get_input(const struct wined3d_shader *shader,
1765         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1766 {
1767     WORD map = shader->reg_maps.input_registers;
1768     unsigned int i;
1769
1770     for (i = 0; map; map >>= 1, ++i)
1771     {
1772         if (!(map & 1)) continue;
1773
1774         if (match_usage(shader->u.vs.attributes[i].usage,
1775                 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1776         {
1777             *regnum = i;
1778             return TRUE;
1779         }
1780     }
1781     return FALSE;
1782 }
1783
1784 static void vertexshader_set_limits(struct wined3d_shader *shader)
1785 {
1786     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1787             shader->reg_maps.shader_version.minor);
1788     struct wined3d_device *device = shader->device;
1789
1790     shader->limits.packed_input = 0;
1791
1792     switch (shader_version)
1793     {
1794         case WINED3D_SHADER_VERSION(1, 0):
1795         case WINED3D_SHADER_VERSION(1, 1):
1796             shader->limits.constant_bool = 0;
1797             shader->limits.constant_int = 0;
1798             shader->limits.packed_output = 12;
1799             shader->limits.sampler = 0;
1800             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1801              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1802              * constants? */
1803             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1804             break;
1805
1806         case WINED3D_SHADER_VERSION(2, 0):
1807         case WINED3D_SHADER_VERSION(2, 1):
1808             shader->limits.constant_bool = 16;
1809             shader->limits.constant_int = 16;
1810             shader->limits.packed_output = 12;
1811             shader->limits.sampler = 0;
1812             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1813             break;
1814
1815         case WINED3D_SHADER_VERSION(3, 0):
1816             shader->limits.constant_bool = 16;
1817             shader->limits.constant_int = 16;
1818             shader->limits.packed_output = 12;
1819             shader->limits.sampler = 4;
1820             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1821              * even though they are capable of supporting much more (GL
1822              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1823              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1824              * shaders to 256. */
1825             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1826             break;
1827
1828         case WINED3D_SHADER_VERSION(4, 0):
1829             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1830             shader->limits.constant_int = 0;
1831             shader->limits.constant_float = 0;
1832             shader->limits.constant_bool = 0;
1833             shader->limits.packed_output = 16;
1834             shader->limits.packed_input = 0;
1835             break;
1836
1837         default:
1838             shader->limits.constant_bool = 16;
1839             shader->limits.constant_int = 16;
1840             shader->limits.packed_output = 12;
1841             shader->limits.sampler = 0;
1842             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1843             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1844                     shader->reg_maps.shader_version.major,
1845                     shader->reg_maps.shader_version.minor);
1846     }
1847 }
1848
1849 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1850         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1851         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1852 {
1853     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1854     unsigned int i;
1855     HRESULT hr;
1856     WORD map;
1857
1858     if (!byte_code) return WINED3DERR_INVALIDCALL;
1859
1860     shader_init(shader, device, parent, parent_ops);
1861     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1862             WINED3D_SHADER_TYPE_VERTEX, max_version);
1863     if (FAILED(hr))
1864     {
1865         WARN("Failed to set function, hr %#x.\n", hr);
1866         shader_cleanup(shader);
1867         return hr;
1868     }
1869
1870     map = reg_maps->input_registers;
1871     for (i = 0; map; map >>= 1, ++i)
1872     {
1873         if (!(map & 1) || !shader->input_signature[i].semantic_name)
1874             continue;
1875
1876         shader->u.vs.attributes[i].usage =
1877                 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1878         shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1879     }
1880
1881     if (output_signature)
1882     {
1883         for (i = 0; i < output_signature->element_count; ++i)
1884         {
1885             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1886             reg_maps->output_registers |= 1 << e->register_idx;
1887             shader->output_signature[e->register_idx] = *e;
1888         }
1889     }
1890
1891     vertexshader_set_limits(shader);
1892
1893     shader->load_local_constsF = reg_maps->usesrelconstF
1894             && !list_empty(&shader->constantsF);
1895
1896     return WINED3D_OK;
1897 }
1898
1899 static void geometryshader_set_limits(struct wined3d_shader *shader)
1900 {
1901     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1902             shader->reg_maps.shader_version.minor);
1903
1904     switch (shader_version)
1905     {
1906         case WINED3D_SHADER_VERSION(4, 0):
1907             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1908             shader->limits.constant_int = 0;
1909             shader->limits.constant_float = 0;
1910             shader->limits.constant_bool = 0;
1911             shader->limits.packed_output = 32;
1912             shader->limits.packed_input = 16;
1913             break;
1914
1915         default:
1916             memset(&shader->limits, 0, sizeof(shader->limits));
1917             FIXME("Unhandled geometry shader version \"%u.%u\".\n",
1918                     shader->reg_maps.shader_version.major,
1919                     shader->reg_maps.shader_version.minor);
1920     }
1921 }
1922
1923 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1924         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1925         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1926 {
1927     HRESULT hr;
1928
1929     shader_init(shader, device, parent, parent_ops);
1930     hr = shader_set_function(shader, byte_code, output_signature, 0,
1931             WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1932     if (FAILED(hr))
1933     {
1934         WARN("Failed to set function, hr %#x.\n", hr);
1935         shader_cleanup(shader);
1936         return hr;
1937     }
1938
1939     geometryshader_set_limits(shader);
1940
1941     shader->load_local_constsF = FALSE;
1942
1943     return WINED3D_OK;
1944 }
1945
1946 void find_ps_compile_args(const struct wined3d_state *state,
1947         const struct wined3d_shader *shader, struct ps_compile_args *args)
1948 {
1949     struct wined3d_device *device = shader->device;
1950     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1951     const struct wined3d_texture *texture;
1952     UINT i;
1953
1954     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1955     if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1956     {
1957         const struct wined3d_surface *rt = state->fb->render_targets[0];
1958         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
1959         {
1960             static unsigned int warned = 0;
1961
1962             args->srgb_correction = 1;
1963             if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
1964                 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
1965                         "support, expect rendering artifacts.\n");
1966         }
1967     }
1968
1969     if (shader->reg_maps.shader_version.major == 1
1970             && shader->reg_maps.shader_version.minor <= 3)
1971     {
1972         for (i = 0; i < 4; ++i)
1973         {
1974             DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
1975
1976             if (flags & WINED3D_TTFF_PROJECTED)
1977             {
1978                 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
1979
1980                 if (!state->vertex_shader)
1981                 {
1982                     unsigned int j;
1983                     unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
1984                     DWORD max_valid = WINED3D_TTFF_COUNT4;
1985                     enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
1986
1987                     for (j = 0; j < state->vertex_declaration->element_count; ++j)
1988                     {
1989                         struct wined3d_vertex_declaration_element *element =
1990                                 &state->vertex_declaration->elements[j];
1991
1992                         if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
1993                                 && element->usage_idx == index)
1994                         {
1995                             max_valid = element->format->component_count;
1996                             break;
1997                         }
1998                     }
1999                     if (!tex_transform || tex_transform > max_valid)
2000                     {
2001                         WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2002                                 tex_transform, max_valid);
2003                         tex_transform = max_valid;
2004                     }
2005                     if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
2006                             || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
2007                             || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
2008                         tex_transform |= WINED3D_PSARGS_PROJECTED;
2009                     else
2010                     {
2011                         WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2012                         WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2013                                 i, tex_transform, sampler_type);
2014                     }
2015                 }
2016                 else
2017                     tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2018
2019                 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2020             }
2021         }
2022     }
2023
2024     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2025     {
2026         if (!shader->reg_maps.sampler_type[i])
2027             continue;
2028
2029         texture = state->textures[i];
2030         if (!texture)
2031         {
2032             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2033             continue;
2034         }
2035         args->color_fixup[i] = texture->resource.format->color_fixup;
2036
2037         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2038             args->shadow |= 1 << i;
2039
2040         /* Flag samplers that need NP2 texcoord fixup. */
2041         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2042             args->np2_fixup |= (1 << i);
2043     }
2044     if (shader->reg_maps.shader_version.major >= 3)
2045     {
2046         if (device->strided_streams.position_transformed)
2047         {
2048             args->vp_mode = pretransformed;
2049         }
2050         else if (use_vs(state))
2051         {
2052             args->vp_mode = vertexshader;
2053         }
2054         else
2055         {
2056             args->vp_mode = fixedfunction;
2057         }
2058         args->fog = FOG_OFF;
2059     }
2060     else
2061     {
2062         args->vp_mode = vertexshader;
2063         if (state->render_states[WINED3D_RS_FOGENABLE])
2064         {
2065             switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2066             {
2067                 case WINED3D_FOG_NONE:
2068                     if (device->strided_streams.position_transformed || use_vs(state))
2069                     {
2070                         args->fog = FOG_LINEAR;
2071                         break;
2072                     }
2073
2074                     switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2075                     {
2076                         case WINED3D_FOG_NONE: /* Fall through. */
2077                         case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2078                         case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
2079                         case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
2080                     }
2081                     break;
2082
2083                 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2084                 case WINED3D_FOG_EXP:    args->fog = FOG_EXP;    break;
2085                 case WINED3D_FOG_EXP2:   args->fog = FOG_EXP2;   break;
2086             }
2087         }
2088         else
2089         {
2090             args->fog = FOG_OFF;
2091         }
2092     }
2093 }
2094
2095 static void pixelshader_set_limits(struct wined3d_shader *shader)
2096 {
2097     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2098             shader->reg_maps.shader_version.minor);
2099
2100     shader->limits.packed_output = 0;
2101
2102     switch (shader_version)
2103     {
2104         case WINED3D_SHADER_VERSION(1, 0):
2105         case WINED3D_SHADER_VERSION(1, 1):
2106         case WINED3D_SHADER_VERSION(1, 2):
2107         case WINED3D_SHADER_VERSION(1, 3):
2108             shader->limits.constant_float = 8;
2109             shader->limits.constant_int = 0;
2110             shader->limits.constant_bool = 0;
2111             shader->limits.sampler = 4;
2112             shader->limits.packed_input = 0;
2113             break;
2114
2115         case WINED3D_SHADER_VERSION(1, 4):
2116             shader->limits.constant_float = 8;
2117             shader->limits.constant_int = 0;
2118             shader->limits.constant_bool = 0;
2119             shader->limits.sampler = 6;
2120             shader->limits.packed_input = 0;
2121             break;
2122
2123         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2124         case WINED3D_SHADER_VERSION(2, 0):
2125             shader->limits.constant_float = 32;
2126             shader->limits.constant_int = 16;
2127             shader->limits.constant_bool = 16;
2128             shader->limits.sampler = 16;
2129             shader->limits.packed_input = 0;
2130             break;
2131
2132         case WINED3D_SHADER_VERSION(2, 1):
2133             shader->limits.constant_float = 32;
2134             shader->limits.constant_int = 16;
2135             shader->limits.constant_bool = 16;
2136             shader->limits.sampler = 16;
2137             shader->limits.packed_input = 0;
2138             break;
2139
2140         case WINED3D_SHADER_VERSION(3, 0):
2141             shader->limits.constant_float = 224;
2142             shader->limits.constant_int = 16;
2143             shader->limits.constant_bool = 16;
2144             shader->limits.sampler = 16;
2145             shader->limits.packed_input = 12;
2146             break;
2147
2148         case WINED3D_SHADER_VERSION(4, 0):
2149             shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
2150             shader->limits.constant_int = 0;
2151             shader->limits.constant_float = 0;
2152             shader->limits.constant_bool = 0;
2153             shader->limits.packed_input = 32;
2154             break;
2155
2156         default:
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             FIXME("Unrecognized pixel shader version %u.%u\n",
2163                     shader->reg_maps.shader_version.major,
2164                     shader->reg_maps.shader_version.minor);
2165     }
2166 }
2167
2168 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2169         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2170         void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2171 {
2172     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2173     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2174     HRESULT hr;
2175
2176     if (!byte_code) return WINED3DERR_INVALIDCALL;
2177
2178     shader_init(shader, device, parent, parent_ops);
2179     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2180             WINED3D_SHADER_TYPE_PIXEL, max_version);
2181     if (FAILED(hr))
2182     {
2183         WARN("Failed to set function, hr %#x.\n", hr);
2184         shader_cleanup(shader);
2185         return hr;
2186     }
2187
2188     pixelshader_set_limits(shader);
2189
2190     for (i = 0; i < MAX_REG_INPUT; ++i)
2191     {
2192         if (shader->u.ps.input_reg_used[i])
2193         {
2194             ++num_regs_used;
2195             highest_reg_used = i;
2196         }
2197     }
2198
2199     /* Don't do any register mapping magic if it is not needed, or if we can't
2200      * achieve anything anyway */
2201     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2202             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2203     {
2204         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2205         {
2206             /* This happens with relative addressing. The input mapper function
2207              * warns about this if the higher registers are declared too, so
2208              * don't write a FIXME here */
2209             WARN("More varying registers used than supported\n");
2210         }
2211
2212         for (i = 0; i < MAX_REG_INPUT; ++i)
2213         {
2214             shader->u.ps.input_reg_map[i] = i;
2215         }
2216
2217         shader->u.ps.declared_in_count = highest_reg_used + 1;
2218     }
2219     else
2220     {
2221         shader->u.ps.declared_in_count = 0;
2222         for (i = 0; i < MAX_REG_INPUT; ++i)
2223         {
2224             if (shader->u.ps.input_reg_used[i])
2225                 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2226             else shader->u.ps.input_reg_map[i] = ~0U;
2227         }
2228     }
2229
2230     shader->load_local_constsF = FALSE;
2231
2232     return WINED3D_OK;
2233 }
2234
2235 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2236 {
2237     enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2238     unsigned int i;
2239
2240     if (reg_maps->shader_version.major != 1) return;
2241
2242     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2243     {
2244         /* We don't sample from this sampler. */
2245         if (!sampler_type[i]) continue;
2246
2247         if (!textures[i])
2248         {
2249             WARN("No texture bound to sampler %u, using 2D.\n", i);
2250             sampler_type[i] = WINED3DSTT_2D;
2251             continue;
2252         }
2253
2254         switch (textures[i]->target)
2255         {
2256             case GL_TEXTURE_RECTANGLE_ARB:
2257             case GL_TEXTURE_2D:
2258                 /* We have to select between texture rectangles and 2D
2259                  * textures later because 2.0 and 3.0 shaders only have
2260                  * WINED3DSTT_2D as well. */
2261                 sampler_type[i] = WINED3DSTT_2D;
2262                 break;
2263
2264             case GL_TEXTURE_3D:
2265                 sampler_type[i] = WINED3DSTT_VOLUME;
2266                 break;
2267
2268             case GL_TEXTURE_CUBE_MAP_ARB:
2269                 sampler_type[i] = WINED3DSTT_CUBE;
2270                 break;
2271
2272             default:
2273                 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2274                 sampler_type[i] = WINED3DSTT_2D;
2275         }
2276     }
2277 }
2278
2279 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2280         const struct wined3d_shader_signature *output_signature, void *parent,
2281         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2282 {
2283     struct wined3d_shader *object;
2284     HRESULT hr;
2285
2286     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2287             device, byte_code, output_signature, parent, parent_ops, shader);
2288
2289     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2290     if (!object)
2291     {
2292         ERR("Failed to allocate shader memory.\n");
2293         return E_OUTOFMEMORY;
2294     }
2295
2296     hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2297     if (FAILED(hr))
2298     {
2299         WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2300         HeapFree(GetProcessHeap(), 0, object);
2301         return hr;
2302     }
2303
2304     TRACE("Created geometry shader %p.\n", object);
2305     *shader = object;
2306
2307     return WINED3D_OK;
2308 }
2309
2310 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2311         const struct wined3d_shader_signature *output_signature, void *parent,
2312         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2313 {
2314     struct wined3d_shader *object;
2315     HRESULT hr;
2316
2317     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2318             device, byte_code, output_signature, parent, parent_ops, shader);
2319
2320     if (device->ps_selected_mode == SHADER_NONE)
2321         return WINED3DERR_INVALIDCALL;
2322
2323     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2324     if (!object)
2325     {
2326         ERR("Failed to allocate shader memory.\n");
2327         return E_OUTOFMEMORY;
2328     }
2329
2330     hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2331     if (FAILED(hr))
2332     {
2333         WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2334         HeapFree(GetProcessHeap(), 0, object);
2335         return hr;
2336     }
2337
2338     TRACE("Created pixel shader %p.\n", object);
2339     *shader = object;
2340
2341     return WINED3D_OK;
2342 }
2343
2344 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2345         const struct wined3d_shader_signature *output_signature, void *parent,
2346         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2347 {
2348     struct wined3d_shader *object;
2349     HRESULT hr;
2350
2351     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2352             device, byte_code, output_signature, parent, parent_ops, shader);
2353
2354     if (device->vs_selected_mode == SHADER_NONE)
2355         return WINED3DERR_INVALIDCALL;
2356
2357     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2358     if (!object)
2359     {
2360         ERR("Failed to allocate shader memory.\n");
2361         return E_OUTOFMEMORY;
2362     }
2363
2364     hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2365     if (FAILED(hr))
2366     {
2367         WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2368         HeapFree(GetProcessHeap(), 0, object);
2369         return hr;
2370     }
2371
2372     TRACE("Created vertex shader %p.\n", object);
2373     *shader = object;
2374
2375     return WINED3D_OK;
2376 }