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 2009-2010 Henri Verbeet for CodeWeavers
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #include "config.h"
  26
  27 #include <math.h>
  28 #include <stdio.h>
  29 #include <string.h>
  30
  31 #include "wined3d_private.h"
  32
  33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  34 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  35
  36 static const char * const shader_opcode_names[] =
  37 {
  38     /* WINED3DSIH_ABS           */ "abs",
  39     /* WINED3DSIH_ADD           */ "add",
  40     /* WINED3DSIH_AND           */ "and",
  41     /* WINED3DSIH_BEM           */ "bem",
  42     /* WINED3DSIH_BREAK         */ "break",
  43     /* WINED3DSIH_BREAKC        */ "breakc",
  44     /* WINED3DSIH_BREAKP        */ "breakp",
  45     /* WINED3DSIH_CALL          */ "call",
  46     /* WINED3DSIH_CALLNZ        */ "callnz",
  47     /* WINED3DSIH_CMP           */ "cmp",
  48     /* WINED3DSIH_CND           */ "cnd",
  49     /* WINED3DSIH_CRS           */ "crs",
  50     /* WINED3DSIH_CUT           */ "cut",
  51     /* WINED3DSIH_DCL           */ "dcl",
  52     /* WINED3DSIH_DEF           */ "def",
  53     /* WINED3DSIH_DEFB          */ "defb",
  54     /* WINED3DSIH_DEFI          */ "defi",
  55     /* WINED3DSIH_DIV           */ "div",
  56     /* WINED3DSIH_DP2ADD        */ "dp2add",
  57     /* WINED3DSIH_DP3           */ "dp3",
  58     /* WINED3DSIH_DP4           */ "dp4",
  59     /* WINED3DSIH_DST           */ "dst",
  60     /* WINED3DSIH_DSX           */ "dsx",
  61     /* WINED3DSIH_DSY           */ "dsy",
  62     /* WINED3DSIH_ELSE          */ "else",
  63     /* WINED3DSIH_EMIT          */ "emit",
  64     /* WINED3DSIH_ENDIF         */ "endif",
  65     /* WINED3DSIH_ENDLOOP       */ "endloop",
  66     /* WINED3DSIH_ENDREP        */ "endrep",
  67     /* WINED3DSIH_EXP           */ "exp",
  68     /* WINED3DSIH_EXPP          */ "expp",
  69     /* WINED3DSIH_FRC           */ "frc",
  70     /* WINED3DSIH_FTOI          */ "ftoi",
  71     /* WINED3DSIH_IADD          */ "iadd",
  72     /* WINED3DSIH_IEQ           */ "ieq",
  73     /* WINED3DSIH_IF            */ "if",
  74     /* WINED3DSIH_IFC           */ "ifc",
  75     /* WINED3DSIH_IGE           */ "ige",
  76     /* WINED3DSIH_IMUL          */ "imul",
  77     /* WINED3DSIH_ITOF          */ "itof",
  78     /* WINED3DSIH_LABEL         */ "label",
  79     /* WINED3DSIH_LD            */ "ld",
  80     /* WINED3DSIH_LIT           */ "lit",
  81     /* WINED3DSIH_LOG           */ "log",
  82     /* WINED3DSIH_LOGP          */ "logp",
  83     /* WINED3DSIH_LOOP          */ "loop",
  84     /* WINED3DSIH_LRP           */ "lrp",
  85     /* WINED3DSIH_LT            */ "lt",
  86     /* WINED3DSIH_M3x2          */ "m3x2",
  87     /* WINED3DSIH_M3x3          */ "m3x3",
  88     /* WINED3DSIH_M3x4          */ "m3x4",
  89     /* WINED3DSIH_M4x3          */ "m4x3",
  90     /* WINED3DSIH_M4x4          */ "m4x4",
  91     /* WINED3DSIH_MAD           */ "mad",
  92     /* WINED3DSIH_MAX           */ "max",
  93     /* WINED3DSIH_MIN           */ "min",
  94     /* WINED3DSIH_MOV           */ "mov",
  95     /* WINED3DSIH_MOVA          */ "mova",
  96     /* WINED3DSIH_MOVC          */ "movc",
  97     /* WINED3DSIH_MUL           */ "mul",
  98     /* WINED3DSIH_NOP           */ "nop",
  99     /* WINED3DSIH_NRM           */ "nrm",
 100     /* WINED3DSIH_PHASE         */ "phase",
 101     /* WINED3DSIH_POW           */ "pow",
 102     /* WINED3DSIH_RCP           */ "rcp",
 103     /* WINED3DSIH_REP           */ "rep",
 104     /* WINED3DSIH_RET           */ "ret",
 105     /* WINED3DSIH_RSQ           */ "rsq",
 106     /* WINED3DSIH_SAMPLE        */ "sample",
 107     /* WINED3DSIH_SAMPLE_GRAD   */ "sample_d",
 108     /* WINED3DSIH_SAMPLE_LOD    */ "sample_l",
 109     /* WINED3DSIH_SETP          */ "setp",
 110     /* WINED3DSIH_SGE           */ "sge",
 111     /* WINED3DSIH_SGN           */ "sgn",
 112     /* WINED3DSIH_SINCOS        */ "sincos",
 113     /* WINED3DSIH_SLT           */ "slt",
 114     /* WINED3DSIH_SQRT          */ "sqrt",
 115     /* WINED3DSIH_SUB           */ "sub",
 116     /* WINED3DSIH_TEX           */ "texld",
 117     /* WINED3DSIH_TEXBEM        */ "texbem",
 118     /* WINED3DSIH_TEXBEML       */ "texbeml",
 119     /* WINED3DSIH_TEXCOORD      */ "texcrd",
 120     /* WINED3DSIH_TEXDEPTH      */ "texdepth",
 121     /* WINED3DSIH_TEXDP3        */ "texdp3",
 122     /* WINED3DSIH_TEXDP3TEX     */ "texdp3tex",
 123     /* WINED3DSIH_TEXKILL       */ "texkill",
 124     /* WINED3DSIH_TEXLDD        */ "texldd",
 125     /* WINED3DSIH_TEXLDL        */ "texldl",
 126     /* WINED3DSIH_TEXM3x2DEPTH  */ "texm3x2depth",
 127     /* WINED3DSIH_TEXM3x2PAD    */ "texm3x2pad",
 128     /* WINED3DSIH_TEXM3x2TEX    */ "texm3x2tex",
 129     /* WINED3DSIH_TEXM3x3       */ "texm3x3",
 130     /* WINED3DSIH_TEXM3x3DIFF   */ "texm3x3diff",
 131     /* WINED3DSIH_TEXM3x3PAD    */ "texm3x3pad",
 132     /* WINED3DSIH_TEXM3x3SPEC   */ "texm3x3spec",
 133     /* WINED3DSIH_TEXM3x3TEX    */ "texm3x3tex",
 134     /* WINED3DSIH_TEXM3x3VSPEC  */ "texm3x3vspec",
 135     /* WINED3DSIH_TEXREG2AR     */ "texreg2ar",
 136     /* WINED3DSIH_TEXREG2GB     */ "texreg2gb",
 137     /* WINED3DSIH_TEXREG2RGB    */ "texreg2rgb",
 138     /* WINED3DSIH_UTOF          */ "utof",
 139 };
 140
 141 static const char * const semantic_names[] =
 142 {
 143     /* WINED3DDECLUSAGE_POSITION        */ "SV_POSITION",
 144     /* WINED3DDECLUSAGE_BLENDWEIGHT     */ "BLENDWEIGHT",
 145     /* WINED3DDECLUSAGE_BLENDINDICES    */ "BLENDINDICES",
 146     /* WINED3DDECLUSAGE_NORMAL          */ "NORMAL",
 147     /* WINED3DDECLUSAGE_PSIZE           */ "PSIZE",
 148     /* WINED3DDECLUSAGE_TEXCOORD        */ "TEXCOORD",
 149     /* WINED3DDECLUSAGE_TANGENT         */ "TANGENT",
 150     /* WINED3DDECLUSAGE_BINORMAL        */ "BINORMAL",
 151     /* WINED3DDECLUSAGE_TESSFACTOR      */ "TESSFACTOR",
 152     /* WINED3DDECLUSAGE_POSITIONT       */ "POSITIONT",
 153     /* WINED3DDECLUSAGE_COLOR           */ "COLOR",
 154     /* WINED3DDECLUSAGE_FOG             */ "FOG",
 155     /* WINED3DDECLUSAGE_DEPTH           */ "DEPTH",
 156     /* WINED3DDECLUSAGE_SAMPLE          */ "SAMPLE",
 157 };
 158
 159 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
 160 {
 161     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 162     {
 163         FIXME("Unrecognized usage %#x.\n", usage);
 164         return "UNRECOGNIZED";
 165     }
 166
 167     return semantic_names[usage];
 168 }
 169
 170 static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
 171 {
 172     unsigned int i;
 173
 174     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 175     {
 176         if (!strcmp(name, semantic_names[i])) return i;
 177     }
 178
 179     return ~0U;
 180 }
 181
 182 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
 183 {
 184     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 185 }
 186
 187 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 188         const struct wined3d_shader_semantic *s)
 189 {
 190     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 191     e->semantic_idx = s->usage_idx;
 192     e->sysval_semantic = 0;
 193     e->component_type = 0;
 194     e->register_idx = s->reg.reg.idx;
 195     e->mask = s->reg.write_mask;
 196 }
 197
 198 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
 199         WINED3DDECLUSAGE usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
 200 {
 201     e->semantic_name = shader_semantic_name_from_usage(usage);
 202     e->semantic_idx = usage_idx;
 203     e->sysval_semantic = 0;
 204     e->component_type = 0;
 205     e->register_idx = reg_idx;
 206     e->mask = write_mask;
 207 }
 208
 209 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 210 {
 211     switch (version_token >> 16)
 212     {
 213         case WINED3D_SM1_VS:
 214         case WINED3D_SM1_PS:
 215             return &sm1_shader_frontend;
 216
 217         case WINED3D_SM4_PS:
 218         case WINED3D_SM4_VS:
 219         case WINED3D_SM4_GS:
 220             return &sm4_shader_frontend;
 221
 222         default:
 223             FIXME("Unrecognised version token %#x\n", version_token);
 224             return NULL;
 225     }
 226 }
 227
 228 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 229 {
 230     buffer->buffer[0] = '\0';
 231     buffer->bsize = 0;
 232     buffer->lineNo = 0;
 233     buffer->newline = TRUE;
 234 }
 235
 236 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 237 {
 238     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 239     if (!buffer->buffer)
 240     {
 241         ERR("Failed to allocate shader buffer memory.\n");
 242         return FALSE;
 243     }
 244
 245     shader_buffer_clear(buffer);
 246     return TRUE;
 247 }
 248
 249 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 250 {
 251     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 252 }
 253
 254 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 255 {
 256     char *base = buffer->buffer + buffer->bsize;
 257     int rc;
 258
 259     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 260
 261     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 262     {
 263         ERR("The buffer allocated for the shader program string "
 264             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 265         buffer->bsize = SHADER_PGMSIZE - 1;
 266         return -1;
 267     }
 268
 269     if (buffer->newline)
 270     {
 271         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 272         buffer->newline = FALSE;
 273     }
 274     else
 275     {
 276         TRACE("%s", base);
 277     }
 278
 279     buffer->bsize += rc;
 280     if (buffer->buffer[buffer->bsize-1] == '\n')
 281     {
 282         ++buffer->lineNo;
 283         buffer->newline = TRUE;
 284     }
 285
 286     return 0;
 287 }
 288
 289 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 290 {
 291     va_list args;
 292     int ret;
 293
 294     va_start(args, format);
 295     ret = shader_vaddline(buffer, format, args);
 296     va_end(args);
 297
 298     return ret;
 299 }
 300
 301 static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
 302         void *parent, const struct wined3d_parent_ops *parent_ops)
 303 {
 304     shader->ref = 1;
 305     shader->device = device;
 306     shader->parent = parent;
 307     shader->parent_ops = parent_ops;
 308     list_init(&shader->linked_programs);
 309     list_add_head(&device->shaders, &shader->shader_list_entry);
 310 }
 311
 312 /* Convert floating point offset relative to a register file to an absolute
 313  * offset for float constants. */
 314 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
 315 {
 316     switch (register_type)
 317     {
 318         case WINED3DSPR_CONST: return register_idx;
 319         case WINED3DSPR_CONST2: return 2048 + register_idx;
 320         case WINED3DSPR_CONST3: return 4096 + register_idx;
 321         case WINED3DSPR_CONST4: return 6144 + register_idx;
 322         default:
 323             FIXME("Unsupported register type: %u.\n", register_type);
 324             return register_idx;
 325     }
 326 }
 327
 328 static void shader_delete_constant_list(struct list *clist)
 329 {
 330     struct local_constant *constant;
 331     struct list *ptr;
 332
 333     ptr = list_head(clist);
 334     while (ptr)
 335     {
 336         constant = LIST_ENTRY(ptr, struct local_constant, entry);
 337         ptr = list_next(clist, ptr);
 338         HeapFree(GetProcessHeap(), 0, constant);
 339     }
 340     list_init(clist);
 341 }
 342
 343 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 344 {
 345     DWORD idx, shift;
 346     idx = bit >> 5;
 347     shift = bit & 0x1f;
 348     bitmap[idx] |= (1 << shift);
 349 }
 350
 351 static void shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct wined3d_shader_reg_maps *reg_maps,
 352         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 353 {
 354     switch (reg->type)
 355     {
 356         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 357             if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
 358             else reg_maps->address |= 1 << reg->idx;
 359             break;
 360
 361         case WINED3DSPR_TEMP:
 362             reg_maps->temporary |= 1 << reg->idx;
 363             break;
 364
 365         case WINED3DSPR_INPUT:
 366             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 367             {
 368                 if (reg->rel_addr)
 369                 {
 370                     /* If relative addressing is used, we must assume that all registers
 371                      * are used. Even if it is a construct like v3[aL], we can't assume
 372                      * that v0, v1 and v2 aren't read because aL can be negative */
 373                     unsigned int i;
 374                     for (i = 0; i < MAX_REG_INPUT; ++i)
 375                     {
 376                         ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
 377                     }
 378                 }
 379                 else
 380                 {
 381                     ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg->idx] = TRUE;
 382                 }
 383             }
 384             else reg_maps->input_registers |= 1 << reg->idx;
 385             break;
 386
 387         case WINED3DSPR_RASTOUT:
 388             if (reg->idx == 1) reg_maps->fog = 1;
 389             break;
 390
 391         case WINED3DSPR_MISCTYPE:
 392             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 393             {
 394                 if (!reg->idx) reg_maps->vpos = 1;
 395                 else if (reg->idx == 1) reg_maps->usesfacing = 1;
 396             }
 397             break;
 398
 399         case WINED3DSPR_CONST:
 400             if (reg->rel_addr)
 401             {
 402                 if (reg->idx < reg_maps->min_rel_offset) reg_maps->min_rel_offset = reg->idx;
 403                 if (reg->idx > reg_maps->max_rel_offset) reg_maps->max_rel_offset = reg->idx;
 404                 reg_maps->usesrelconstF = TRUE;
 405             }
 406             else
 407             {
 408                 set_bitmap_bit(reg_maps->constf, reg->idx);
 409             }
 410             break;
 411
 412         case WINED3DSPR_CONSTINT:
 413             reg_maps->integer_constants |= (1 << reg->idx);
 414             break;
 415
 416         case WINED3DSPR_CONSTBOOL:
 417             reg_maps->boolean_constants |= (1 << reg->idx);
 418             break;
 419
 420         case WINED3DSPR_COLOROUT:
 421             reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
 422             break;
 423
 424         default:
 425             TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
 426             break;
 427     }
 428 }
 429
 430 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 431 {
 432     switch (instr)
 433     {
 434         case WINED3DSIH_M4x4:
 435         case WINED3DSIH_M3x4:
 436             return param == 1 ? 3 : 0;
 437
 438         case WINED3DSIH_M4x3:
 439         case WINED3DSIH_M3x3:
 440             return param == 1 ? 2 : 0;
 441
 442         case WINED3DSIH_M3x2:
 443             return param == 1 ? 1 : 0;
 444
 445         default:
 446             return 0;
 447     }
 448 }
 449
 450 /* Note that this does not count the loop register as an address register. */
 451 static HRESULT shader_get_registers_used(IWineD3DBaseShaderImpl *shader, const struct wined3d_shader_frontend *fe,
 452         struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 453         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 454 {
 455     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 456     void *fe_data = shader->baseShader.frontend_data;
 457     struct wined3d_shader_version shader_version;
 458     const DWORD *ptr = byte_code;
 459
 460     memset(reg_maps, 0, sizeof(*reg_maps));
 461     reg_maps->min_rel_offset = ~0U;
 462
 463     fe->shader_read_header(fe_data, &ptr, &shader_version);
 464     reg_maps->shader_version = shader_version;
 465
 466     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 467             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 468     if (!reg_maps->constf)
 469     {
 470         ERR("Failed to allocate constant map memory.\n");
 471         return E_OUTOFMEMORY;
 472     }
 473
 474     while (!fe->shader_is_end(fe_data, &ptr))
 475     {
 476         struct wined3d_shader_instruction ins;
 477         const char *comment;
 478         UINT comment_size;
 479         UINT param_size;
 480
 481         /* Skip comments. */
 482         fe->shader_read_comment(&ptr, &comment, &comment_size);
 483         if (comment) continue;
 484
 485         /* Fetch opcode. */
 486         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
 487
 488         /* Unhandled opcode, and its parameters. */
 489         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 490         {
 491             TRACE("Skipping unrecognized instruction.\n");
 492             ptr += param_size;
 493             continue;
 494         }
 495
 496         /* Handle declarations. */
 497         if (ins.handler_idx == WINED3DSIH_DCL)
 498         {
 499             struct wined3d_shader_semantic semantic;
 500
 501             fe->shader_read_semantic(&ptr, &semantic);
 502
 503             switch (semantic.reg.reg.type)
 504             {
 505                 /* Mark input registers used. */
 506                 case WINED3DSPR_INPUT:
 507                     reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
 508                     shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
 509                     break;
 510
 511                 /* Vertex shader: mark 3.0 output registers used, save token. */
 512                 case WINED3DSPR_OUTPUT:
 513                     reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
 514                     shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
 515                     if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
 516                     break;
 517
 518                 /* Save sampler usage token. */
 519                 case WINED3DSPR_SAMPLER:
 520                     reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
 521                     break;
 522
 523                 default:
 524                     TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
 525                     break;
 526             }
 527         }
 528         else if (ins.handler_idx == WINED3DSIH_DEF)
 529         {
 530             struct wined3d_shader_src_param rel_addr;
 531             struct wined3d_shader_dst_param dst;
 532
 533             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 534             if (!lconst) return E_OUTOFMEMORY;
 535
 536             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 537             lconst->idx = dst.reg.idx;
 538
 539             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 540             ptr += 4;
 541
 542             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 543             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 544             {
 545                 float *value = (float *)lconst->value;
 546                 if (value[0] < -1.0f) value[0] = -1.0f;
 547                 else if (value[0] > 1.0f) value[0] = 1.0f;
 548                 if (value[1] < -1.0f) value[1] = -1.0f;
 549                 else if (value[1] > 1.0f) value[1] = 1.0f;
 550                 if (value[2] < -1.0f) value[2] = -1.0f;
 551                 else if (value[2] > 1.0f) value[2] = 1.0f;
 552                 if (value[3] < -1.0f) value[3] = -1.0f;
 553                 else if (value[3] > 1.0f) value[3] = 1.0f;
 554             }
 555
 556             list_add_head(&shader->baseShader.constantsF, &lconst->entry);
 557         }
 558         else if (ins.handler_idx == WINED3DSIH_DEFI)
 559         {
 560             struct wined3d_shader_src_param rel_addr;
 561             struct wined3d_shader_dst_param dst;
 562
 563             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 564             if (!lconst) return E_OUTOFMEMORY;
 565
 566             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 567             lconst->idx = dst.reg.idx;
 568
 569             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 570             ptr += 4;
 571
 572             list_add_head(&shader->baseShader.constantsI, &lconst->entry);
 573             reg_maps->local_int_consts |= (1 << dst.reg.idx);
 574         }
 575         else if (ins.handler_idx == WINED3DSIH_DEFB)
 576         {
 577             struct wined3d_shader_src_param rel_addr;
 578             struct wined3d_shader_dst_param dst;
 579
 580             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 581             if (!lconst) return E_OUTOFMEMORY;
 582
 583             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 584             lconst->idx = dst.reg.idx;
 585
 586             memcpy(lconst->value, ptr, sizeof(DWORD));
 587             ++ptr;
 588
 589             list_add_head(&shader->baseShader.constantsB, &lconst->entry);
 590             reg_maps->local_bool_consts |= (1 << dst.reg.idx);
 591         }
 592         /* If there's a loop in the shader. */
 593         else if (ins.handler_idx == WINED3DSIH_LOOP
 594                 || ins.handler_idx == WINED3DSIH_REP)
 595         {
 596             struct wined3d_shader_src_param src, rel_addr;
 597
 598             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 599
 600             /* Rep and Loop always use an integer constant for the control parameters. */
 601             if (ins.handler_idx == WINED3DSIH_REP)
 602             {
 603                 reg_maps->integer_constants |= 1 << src.reg.idx;
 604             }
 605             else
 606             {
 607                 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 608                 reg_maps->integer_constants |= 1 << src.reg.idx;
 609             }
 610
 611             cur_loop_depth++;
 612             if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
 613         }
 614         else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 615                 || ins.handler_idx == WINED3DSIH_ENDREP)
 616         {
 617             cur_loop_depth--;
 618         }
 619         /* For subroutine prototypes. */
 620         else if (ins.handler_idx == WINED3DSIH_LABEL)
 621         {
 622             struct wined3d_shader_src_param src, rel_addr;
 623
 624             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 625             reg_maps->labels |= 1 << src.reg.idx;
 626         }
 627         /* Set texture, address, temporary registers. */
 628         else
 629         {
 630             BOOL color0_mov = FALSE;
 631             int i, limit;
 632
 633             /* This will loop over all the registers and try to
 634              * make a bitmask of the ones we're interested in.
 635              *
 636              * Relative addressing tokens are ignored, but that's
 637              * okay, since we'll catch any address registers when
 638              * they are initialized (required by spec). */
 639             for (i = 0; i < ins.dst_count; ++i)
 640             {
 641                 struct wined3d_shader_src_param dst_rel_addr;
 642                 struct wined3d_shader_dst_param dst_param;
 643
 644                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
 645
 646                 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
 647
 648                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 649                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 650                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 651                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
 652                 {
 653                     UINT idx = dst_param.reg.idx;
 654
 655                     switch (dst_param.reg.type)
 656                     {
 657                         case WINED3DSPR_RASTOUT:
 658                             switch (idx)
 659                             {
 660                                 case 0: /* oPos */
 661                                     reg_maps->output_registers |= 1 << 10;
 662                                     shader_signature_from_usage(&output_signature[10],
 663                                             WINED3DDECLUSAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
 664                                     break;
 665
 666                                 case 1: /* oFog */
 667                                     reg_maps->output_registers |= 1 << 11;
 668                                     shader_signature_from_usage(&output_signature[11],
 669                                             WINED3DDECLUSAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
 670                                     break;
 671
 672                                 case 2: /* oPts */
 673                                     reg_maps->output_registers |= 1 << 11;
 674                                     shader_signature_from_usage(&output_signature[11],
 675                                             WINED3DDECLUSAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
 676                                     break;
 677                             }
 678                             break;
 679
 680                         case WINED3DSPR_ATTROUT:
 681                             if (idx < 2)
 682                             {
 683                                 idx += 8;
 684                                 if (reg_maps->output_registers & (1 << idx))
 685                                 {
 686                                     output_signature[idx].mask |= dst_param.write_mask;
 687                                 }
 688                                 else
 689                                 {
 690                                     reg_maps->output_registers |= 1 << idx;
 691                                     shader_signature_from_usage(&output_signature[idx],
 692                                             WINED3DDECLUSAGE_COLOR, idx - 8, idx, dst_param.write_mask);
 693                                 }
 694                             }
 695                             break;
 696
 697                         case WINED3DSPR_TEXCRDOUT:
 698
 699                             reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
 700                             if (reg_maps->output_registers & (1 << idx))
 701                             {
 702                                 output_signature[idx].mask |= dst_param.write_mask;
 703                             }
 704                             else
 705                             {
 706                                 reg_maps->output_registers |= 1 << idx;
 707                                 shader_signature_from_usage(&output_signature[idx],
 708                                         WINED3DDECLUSAGE_TEXCOORD, idx, idx, dst_param.write_mask);
 709                             }
 710                             break;
 711
 712                         default:
 713                             break;
 714                     }
 715                 }
 716
 717                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 718                 {
 719                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 720
 721                     if (dst_param.reg.type == WINED3DSPR_COLOROUT && !dst_param.reg.idx)
 722                     {
 723                     /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 724                      * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 725                      * the mov and perform the sRGB write correction from the source register.
 726                      *
 727                      * However, if the mov is only partial, we can't do this, and if the write
 728                      * comes from an instruction other than MOV it is hard to do as well. If
 729                      * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 730
 731                         ps->color0_mov = FALSE;
 732                         if (ins.handler_idx == WINED3DSIH_MOV
 733                                 && dst_param.write_mask == WINED3DSP_WRITEMASK_ALL)
 734                         {
 735                             /* Used later when the source register is read. */
 736                             color0_mov = TRUE;
 737                         }
 738                     }
 739                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 740                      * end
 741                      */
 742                     else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
 743                     {
 744                         ps->color0_mov = FALSE;
 745                     }
 746                 }
 747
 748                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 749                 if (shader_version.major == 1
 750                         && (ins.handler_idx == WINED3DSIH_TEX
 751                             || ins.handler_idx == WINED3DSIH_TEXBEM
 752                             || ins.handler_idx == WINED3DSIH_TEXBEML
 753                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 754                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 755                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 756                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 757                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 758                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 759                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 760                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 761                 {
 762                     /* Fake sampler usage, only set reserved bit and type. */
 763                     DWORD sampler_code = dst_param.reg.idx;
 764
 765                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 766                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 767
 768                     /* texbem is only valid with < 1.4 pixel shaders */
 769                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 770                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 771                     {
 772                         reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 773                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 774                         {
 775                             reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
 776                         }
 777                     }
 778                 }
 779                 else if (ins.handler_idx == WINED3DSIH_BEM)
 780                 {
 781                     reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 782                 }
 783             }
 784
 785             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 786             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 787             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 788             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 789             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 790             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 791             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 792             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 793             else if (ins.handler_idx == WINED3DSIH_RCP) reg_maps->usesrcp = 1;
 794
 795             limit = ins.src_count + (ins.predicate ? 1 : 0);
 796             for (i = 0; i < limit; ++i)
 797             {
 798                 struct wined3d_shader_src_param src_param, src_rel_addr;
 799                 unsigned int count;
 800
 801                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
 802                 count = get_instr_extra_regcount(ins.handler_idx, i);
 803
 804                 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 805                 while (count)
 806                 {
 807                     ++src_param.reg.idx;
 808                     shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 809                     --count;
 810                 }
 811
 812                 if (color0_mov)
 813                 {
 814                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 815                     if (src_param.reg.type == WINED3DSPR_TEMP
 816                             && src_param.swizzle == WINED3DSP_NOSWIZZLE)
 817                     {
 818                         ps->color0_mov = TRUE;
 819                         ps->color0_reg = src_param.reg.idx;
 820                     }
 821                 }
 822             }
 823         }
 824     }
 825     reg_maps->loop_depth = max_loop_depth;
 826
 827     shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
 828
 829     return WINED3D_OK;
 830 }
 831
 832 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 833 {
 834     DWORD map = 1 << max;
 835     map |= map - 1;
 836     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 837
 838     return wined3d_log2i(map);
 839 }
 840
 841 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 842         const struct wined3d_shader_version *shader_version)
 843 {
 844     TRACE("dcl");
 845
 846     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 847     {
 848         switch (semantic->sampler_type)
 849         {
 850             case WINED3DSTT_2D: TRACE("_2d"); break;
 851             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 852             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 853             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 854         }
 855     }
 856     else
 857     {
 858         /* Pixel shaders 3.0 don't have usage semantics. */
 859         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 860         else TRACE("_");
 861
 862         switch (semantic->usage)
 863         {
 864             case WINED3DDECLUSAGE_POSITION:
 865                 TRACE("position%u", semantic->usage_idx);
 866                 break;
 867
 868             case WINED3DDECLUSAGE_BLENDINDICES:
 869                 TRACE("blend");
 870                 break;
 871
 872             case WINED3DDECLUSAGE_BLENDWEIGHT:
 873                 TRACE("weight");
 874                 break;
 875
 876             case WINED3DDECLUSAGE_NORMAL:
 877                 TRACE("normal%u", semantic->usage_idx);
 878                 break;
 879
 880             case WINED3DDECLUSAGE_PSIZE:
 881                 TRACE("psize");
 882                 break;
 883
 884             case WINED3DDECLUSAGE_COLOR:
 885                 if (!semantic->usage_idx) TRACE("color");
 886                 else TRACE("specular%u", (semantic->usage_idx - 1));
 887                 break;
 888
 889             case WINED3DDECLUSAGE_TEXCOORD:
 890                 TRACE("texture%u", semantic->usage_idx);
 891                 break;
 892
 893             case WINED3DDECLUSAGE_TANGENT:
 894                 TRACE("tangent");
 895                 break;
 896
 897             case WINED3DDECLUSAGE_BINORMAL:
 898                 TRACE("binormal");
 899                 break;
 900
 901             case WINED3DDECLUSAGE_TESSFACTOR:
 902                 TRACE("tessfactor");
 903                 break;
 904
 905             case WINED3DDECLUSAGE_POSITIONT:
 906                 TRACE("positionT%u", semantic->usage_idx);
 907                 break;
 908
 909             case WINED3DDECLUSAGE_FOG:
 910                 TRACE("fog");
 911                 break;
 912
 913             case WINED3DDECLUSAGE_DEPTH:
 914                 TRACE("depth");
 915                 break;
 916
 917             case WINED3DDECLUSAGE_SAMPLE:
 918                 TRACE("sample");
 919                 break;
 920
 921             default:
 922                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 923         }
 924     }
 925 }
 926
 927 static void shader_dump_register(const struct wined3d_shader_register *reg,
 928         const struct wined3d_shader_version *shader_version)
 929 {
 930     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 931     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 932     UINT offset = reg->idx;
 933
 934     switch (reg->type)
 935     {
 936         case WINED3DSPR_TEMP:
 937             TRACE("r");
 938             break;
 939
 940         case WINED3DSPR_INPUT:
 941             TRACE("v");
 942             break;
 943
 944         case WINED3DSPR_CONST:
 945         case WINED3DSPR_CONST2:
 946         case WINED3DSPR_CONST3:
 947         case WINED3DSPR_CONST4:
 948             TRACE("c");
 949             offset = shader_get_float_offset(reg->type, reg->idx);
 950             break;
 951
 952         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 953             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 954             break;
 955
 956         case WINED3DSPR_RASTOUT:
 957             TRACE("%s", rastout_reg_names[reg->idx]);
 958             break;
 959
 960         case WINED3DSPR_COLOROUT:
 961             TRACE("oC");
 962             break;
 963
 964         case WINED3DSPR_DEPTHOUT:
 965             TRACE("oDepth");
 966             break;
 967
 968         case WINED3DSPR_ATTROUT:
 969             TRACE("oD");
 970             break;
 971
 972         case WINED3DSPR_TEXCRDOUT:
 973             /* Vertex shaders >= 3.0 use general purpose output registers
 974              * (WINED3DSPR_OUTPUT), which can include an address token. */
 975             if (shader_version->major >= 3) TRACE("o");
 976             else TRACE("oT");
 977             break;
 978
 979         case WINED3DSPR_CONSTINT:
 980             TRACE("i");
 981             break;
 982
 983         case WINED3DSPR_CONSTBOOL:
 984             TRACE("b");
 985             break;
 986
 987         case WINED3DSPR_LABEL:
 988             TRACE("l");
 989             break;
 990
 991         case WINED3DSPR_LOOP:
 992             TRACE("aL");
 993             break;
 994
 995         case WINED3DSPR_SAMPLER:
 996             TRACE("s");
 997             break;
 998
 999         case WINED3DSPR_MISCTYPE:
1000             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
1001             else TRACE("%s", misctype_reg_names[reg->idx]);
1002             break;
1003
1004         case WINED3DSPR_PREDICATE:
1005             TRACE("p");
1006             break;
1007
1008         case WINED3DSPR_IMMCONST:
1009             TRACE("l");
1010             break;
1011
1012         case WINED3DSPR_CONSTBUFFER:
1013             TRACE("cb");
1014             break;
1015
1016         case WINED3DSPR_NULL:
1017             TRACE("null");
1018             break;
1019
1020         case WINED3DSPR_RESOURCE:
1021             TRACE("t");
1022             break;
1023
1024         default:
1025             TRACE("unhandled_rtype(%#x)", reg->type);
1026             break;
1027     }
1028
1029     if (reg->type == WINED3DSPR_IMMCONST)
1030     {
1031         TRACE("(");
1032         switch (reg->immconst_type)
1033         {
1034             case WINED3D_IMMCONST_SCALAR:
1035                 TRACE("%.8e", *(const float *)reg->immconst_data);
1036                 break;
1037
1038             case WINED3D_IMMCONST_VEC4:
1039                 TRACE("%.8e, %.8e, %.8e, %.8e",
1040                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1041                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1042                 break;
1043
1044             default:
1045                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1046                 break;
1047         }
1048         TRACE(")");
1049     }
1050     else if (reg->type != WINED3DSPR_RASTOUT
1051             && reg->type != WINED3DSPR_MISCTYPE
1052             && reg->type != WINED3DSPR_NULL)
1053     {
1054         if (reg->array_idx != ~0U)
1055         {
1056             TRACE("%u[%u", offset, reg->array_idx);
1057             if (reg->rel_addr)
1058             {
1059                 TRACE(" + ");
1060                 shader_dump_src_param(reg->rel_addr, shader_version);
1061             }
1062             TRACE("]");
1063         }
1064         else
1065         {
1066             if (reg->rel_addr)
1067             {
1068                 TRACE("[");
1069                 shader_dump_src_param(reg->rel_addr, shader_version);
1070                 TRACE(" + ");
1071             }
1072             TRACE("%u", offset);
1073             if (reg->rel_addr) TRACE("]");
1074         }
1075     }
1076 }
1077
1078 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1079         const struct wined3d_shader_version *shader_version)
1080 {
1081     DWORD write_mask = param->write_mask;
1082
1083     shader_dump_register(&param->reg, shader_version);
1084
1085     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1086     {
1087         static const char *write_mask_chars = "xyzw";
1088
1089         TRACE(".");
1090         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1091         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1092         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1093         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1094     }
1095 }
1096
1097 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1098         const struct wined3d_shader_version *shader_version)
1099 {
1100     DWORD src_modifier = param->modifiers;
1101     DWORD swizzle = param->swizzle;
1102
1103     if (src_modifier == WINED3DSPSM_NEG
1104             || src_modifier == WINED3DSPSM_BIASNEG
1105             || src_modifier == WINED3DSPSM_SIGNNEG
1106             || src_modifier == WINED3DSPSM_X2NEG
1107             || src_modifier == WINED3DSPSM_ABSNEG)
1108         TRACE("-");
1109     else if (src_modifier == WINED3DSPSM_COMP)
1110         TRACE("1-");
1111     else if (src_modifier == WINED3DSPSM_NOT)
1112         TRACE("!");
1113
1114     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1115         TRACE("abs(");
1116
1117     shader_dump_register(&param->reg, shader_version);
1118
1119     if (src_modifier)
1120     {
1121         switch (src_modifier)
1122         {
1123             case WINED3DSPSM_NONE:    break;
1124             case WINED3DSPSM_NEG:     break;
1125             case WINED3DSPSM_NOT:     break;
1126             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1127             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1128             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1129             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1130             case WINED3DSPSM_COMP:    break;
1131             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1132             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1133             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1134             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1135             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1136             case WINED3DSPSM_ABS:     TRACE(")"); break;
1137             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1138         }
1139     }
1140
1141     if (swizzle != WINED3DSP_NOSWIZZLE)
1142     {
1143         static const char *swizzle_chars = "xyzw";
1144         DWORD swizzle_x = swizzle & 0x03;
1145         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1146         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1147         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1148
1149         if (swizzle_x == swizzle_y
1150                 && swizzle_x == swizzle_z
1151                 && swizzle_x == swizzle_w)
1152         {
1153             TRACE(".%c", swizzle_chars[swizzle_x]);
1154         }
1155         else
1156         {
1157             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1158                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1159         }
1160     }
1161 }
1162
1163 /* Shared code in order to generate the bulk of the shader string.
1164  * NOTE: A description of how to parse tokens can be found on MSDN. */
1165 void shader_generate_main(IWineD3DBaseShaderImpl *shader, struct wined3d_shader_buffer *buffer,
1166         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1167 {
1168     IWineD3DDeviceImpl *device = shader->baseShader.device;
1169     const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1170     void *fe_data = shader->baseShader.frontend_data;
1171     struct wined3d_shader_src_param dst_rel_addr[2];
1172     struct wined3d_shader_src_param src_rel_addr[4];
1173     struct wined3d_shader_dst_param dst_param[2];
1174     struct wined3d_shader_src_param src_param[4];
1175     struct wined3d_shader_version shader_version;
1176     struct wined3d_shader_loop_state loop_state;
1177     struct wined3d_shader_instruction ins;
1178     struct wined3d_shader_tex_mx tex_mx;
1179     struct wined3d_shader_context ctx;
1180     const DWORD *ptr = byte_code;
1181     DWORD i;
1182
1183     /* Initialize current parsing state. */
1184     tex_mx.current_row = 0;
1185     loop_state.current_depth = 0;
1186     loop_state.current_reg = 0;
1187
1188     ctx.shader = shader;
1189     ctx.gl_info = &device->adapter->gl_info;
1190     ctx.reg_maps = reg_maps;
1191     ctx.buffer = buffer;
1192     ctx.tex_mx = &tex_mx;
1193     ctx.loop_state = &loop_state;
1194     ctx.backend_data = backend_ctx;
1195
1196     ins.ctx = &ctx;
1197     ins.dst = dst_param;
1198     ins.src = src_param;
1199
1200     fe->shader_read_header(fe_data, &ptr, &shader_version);
1201
1202     while (!fe->shader_is_end(fe_data, &ptr))
1203     {
1204         const char *comment;
1205         UINT comment_size;
1206         UINT param_size;
1207
1208         /* Skip comment tokens. */
1209         fe->shader_read_comment(&ptr, &comment, &comment_size);
1210         if (comment) continue;
1211
1212         /* Read opcode. */
1213         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1214
1215         /* Unknown opcode and its parameters. */
1216         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1217         {
1218             TRACE("Skipping unrecognized instruction.\n");
1219             ptr += param_size;
1220             continue;
1221         }
1222
1223         /* Nothing to do. */
1224         if (ins.handler_idx == WINED3DSIH_DCL
1225                 || ins.handler_idx == WINED3DSIH_NOP
1226                 || ins.handler_idx == WINED3DSIH_DEF
1227                 || ins.handler_idx == WINED3DSIH_DEFI
1228                 || ins.handler_idx == WINED3DSIH_DEFB
1229                 || ins.handler_idx == WINED3DSIH_PHASE)
1230         {
1231             ptr += param_size;
1232             continue;
1233         }
1234
1235         /* Destination tokens */
1236         for (i = 0; i < ins.dst_count; ++i)
1237         {
1238             fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1239         }
1240
1241         /* Predication token */
1242         if (ins.predicate)
1243         {
1244             FIXME("Predicates not implemented.\n");
1245             ins.predicate = *ptr++;
1246         }
1247
1248         /* Other source tokens */
1249         for (i = 0; i < ins.src_count; ++i)
1250         {
1251             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1252         }
1253
1254         /* Call appropriate function for output target */
1255         device->shader_backend->shader_handle_instruction(&ins);
1256     }
1257 }
1258
1259 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1260 {
1261     DWORD mmask = dst->modifiers;
1262
1263     switch (dst->shift)
1264     {
1265         case 0: break;
1266         case 13: TRACE("_d8"); break;
1267         case 14: TRACE("_d4"); break;
1268         case 15: TRACE("_d2"); break;
1269         case 1: TRACE("_x2"); break;
1270         case 2: TRACE("_x4"); break;
1271         case 3: TRACE("_x8"); break;
1272         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1273     }
1274
1275     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1276     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1277     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1278
1279     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1280     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1281 }
1282
1283 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1284 {
1285     struct wined3d_shader_version shader_version;
1286     const DWORD *ptr = byte_code;
1287     const char *type_prefix;
1288     DWORD i;
1289
1290     TRACE("Parsing %p.\n", byte_code);
1291
1292     fe->shader_read_header(fe_data, &ptr, &shader_version);
1293
1294     switch (shader_version.type)
1295     {
1296         case WINED3D_SHADER_TYPE_VERTEX:
1297             type_prefix = "vs";
1298             break;
1299
1300         case WINED3D_SHADER_TYPE_GEOMETRY:
1301             type_prefix = "gs";
1302             break;
1303
1304         case WINED3D_SHADER_TYPE_PIXEL:
1305             type_prefix = "ps";
1306             break;
1307
1308         default:
1309             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1310             type_prefix = "unknown";
1311             break;
1312     }
1313
1314     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1315
1316     while (!fe->shader_is_end(fe_data, &ptr))
1317     {
1318         struct wined3d_shader_instruction ins;
1319         const char *comment;
1320         UINT comment_size;
1321         UINT param_size;
1322
1323         /* comment */
1324         fe->shader_read_comment(&ptr, &comment, &comment_size);
1325         if (comment)
1326         {
1327             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1328             {
1329                 const char *end = comment + comment_size;
1330                 const char *ptr = comment + 4;
1331                 const char *line = ptr;
1332
1333                 TRACE("// TEXT\n");
1334                 while (ptr != end)
1335                 {
1336                     if (*ptr == '\n')
1337                     {
1338                         UINT len = ptr - line;
1339                         if (len && *(ptr - 1) == '\r') --len;
1340                         TRACE("// %s\n", debugstr_an(line, len));
1341                         line = ++ptr;
1342                     }
1343                     else ++ptr;
1344                 }
1345                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1346             }
1347             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1348             continue;
1349         }
1350
1351         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1352         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1353         {
1354             TRACE("Skipping unrecognized instruction.\n");
1355             ptr += param_size;
1356             continue;
1357         }
1358
1359         if (ins.handler_idx == WINED3DSIH_DCL)
1360         {
1361             struct wined3d_shader_semantic semantic;
1362
1363             fe->shader_read_semantic(&ptr, &semantic);
1364
1365             shader_dump_decl_usage(&semantic, &shader_version);
1366             shader_dump_ins_modifiers(&semantic.reg);
1367             TRACE(" ");
1368             shader_dump_dst_param(&semantic.reg, &shader_version);
1369         }
1370         else if (ins.handler_idx == WINED3DSIH_DEF)
1371         {
1372             struct wined3d_shader_src_param rel_addr;
1373             struct wined3d_shader_dst_param dst;
1374
1375             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1376
1377             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1378                     *(const float *)(ptr),
1379                     *(const float *)(ptr + 1),
1380                     *(const float *)(ptr + 2),
1381                     *(const float *)(ptr + 3));
1382             ptr += 4;
1383         }
1384         else if (ins.handler_idx == WINED3DSIH_DEFI)
1385         {
1386             struct wined3d_shader_src_param rel_addr;
1387             struct wined3d_shader_dst_param dst;
1388
1389             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1390
1391             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1392                     *(ptr),
1393                     *(ptr + 1),
1394                     *(ptr + 2),
1395                     *(ptr + 3));
1396             ptr += 4;
1397         }
1398         else if (ins.handler_idx == WINED3DSIH_DEFB)
1399         {
1400             struct wined3d_shader_src_param rel_addr;
1401             struct wined3d_shader_dst_param dst;
1402
1403             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1404
1405             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1406             ++ptr;
1407         }
1408         else
1409         {
1410             struct wined3d_shader_src_param dst_rel_addr[2];
1411             struct wined3d_shader_src_param src_rel_addr;
1412             struct wined3d_shader_dst_param dst_param[2];
1413             struct wined3d_shader_src_param src_param;
1414
1415             for (i = 0; i < ins.dst_count; ++i)
1416             {
1417                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1418             }
1419
1420             /* Print out predication source token first - it follows
1421              * the destination token. */
1422             if (ins.predicate)
1423             {
1424                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1425                 TRACE("(");
1426                 shader_dump_src_param(&src_param, &shader_version);
1427                 TRACE(") ");
1428             }
1429
1430             /* PixWin marks instructions with the coissue flag with a '+' */
1431             if (ins.coissue) TRACE("+");
1432
1433             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1434
1435             if (ins.handler_idx == WINED3DSIH_IFC
1436                     || ins.handler_idx == WINED3DSIH_BREAKC)
1437             {
1438                 switch (ins.flags)
1439                 {
1440                     case COMPARISON_GT: TRACE("_gt"); break;
1441                     case COMPARISON_EQ: TRACE("_eq"); break;
1442                     case COMPARISON_GE: TRACE("_ge"); break;
1443                     case COMPARISON_LT: TRACE("_lt"); break;
1444                     case COMPARISON_NE: TRACE("_ne"); break;
1445                     case COMPARISON_LE: TRACE("_le"); break;
1446                     default: TRACE("_(%u)", ins.flags);
1447                 }
1448             }
1449             else if (ins.handler_idx == WINED3DSIH_TEX
1450                     && shader_version.major >= 2
1451                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1452             {
1453                 TRACE("p");
1454             }
1455
1456             /* We already read the destination tokens, print them. */
1457             for (i = 0; i < ins.dst_count; ++i)
1458             {
1459                 shader_dump_ins_modifiers(&dst_param[i]);
1460                 TRACE(!i ? " " : ", ");
1461                 shader_dump_dst_param(&dst_param[i], &shader_version);
1462             }
1463
1464             /* Other source tokens */
1465             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1466             {
1467                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1468                 TRACE(!i ? " " : ", ");
1469                 shader_dump_src_param(&src_param, &shader_version);
1470             }
1471         }
1472         TRACE("\n");
1473     }
1474 }
1475
1476 static void shader_cleanup(IWineD3DBaseShaderImpl *shader)
1477 {
1478     shader->baseShader.device->shader_backend->shader_destroy(shader);
1479     HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1480     HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1481     shader_delete_constant_list(&shader->baseShader.constantsF);
1482     shader_delete_constant_list(&shader->baseShader.constantsB);
1483     shader_delete_constant_list(&shader->baseShader.constantsI);
1484     list_remove(&shader->baseShader.shader_list_entry);
1485
1486     if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1487     {
1488         shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1489     }
1490 }
1491
1492 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1493 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1494 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1495         enum tex_types tex_type, const SIZE *ds_mask_size) {}
1496 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1497 static void shader_none_update_float_vertex_constants(IWineD3DDeviceImpl *device, UINT start, UINT count) {}
1498 static void shader_none_update_float_pixel_constants(IWineD3DDeviceImpl *device, UINT start, UINT count) {}
1499 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1500 static void shader_none_load_np2fixup_constants(void *shader_priv,
1501         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1502 static void shader_none_destroy(IWineD3DBaseShaderImpl *shader) {}
1503 static HRESULT shader_none_alloc(IWineD3DDeviceImpl *device) {return WINED3D_OK;}
1504 static void shader_none_free(IWineD3DDeviceImpl *device) {}
1505 static BOOL shader_none_dirty_const(void) {return FALSE;}
1506
1507 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1508 {
1509     /* Set the shader caps to 0 for the none shader backend */
1510     caps->VertexShaderVersion = 0;
1511     caps->MaxVertexShaderConst = 0;
1512     caps->PixelShaderVersion = 0;
1513     caps->PixelShader1xMaxValue = 0.0f;
1514     caps->MaxPixelShaderConst = 0;
1515     caps->VSClipping = FALSE;
1516 }
1517
1518 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1519 {
1520     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1521     {
1522         TRACE("Checking support for fixup:\n");
1523         dump_color_fixup_desc(fixup);
1524     }
1525
1526     /* Faked to make some apps happy. */
1527     if (!is_complex_fixup(fixup))
1528     {
1529         TRACE("[OK]\n");
1530         return TRUE;
1531     }
1532
1533     TRACE("[FAILED]\n");
1534     return FALSE;
1535 }
1536
1537 const shader_backend_t none_shader_backend = {
1538     shader_none_handle_instruction,
1539     shader_none_select,
1540     shader_none_select_depth_blt,
1541     shader_none_deselect_depth_blt,
1542     shader_none_update_float_vertex_constants,
1543     shader_none_update_float_pixel_constants,
1544     shader_none_load_constants,
1545     shader_none_load_np2fixup_constants,
1546     shader_none_destroy,
1547     shader_none_alloc,
1548     shader_none_free,
1549     shader_none_dirty_const,
1550     shader_none_get_caps,
1551     shader_none_color_fixup_supported,
1552 };
1553
1554 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1555 {
1556     if (!data)
1557     {
1558         *data_size = shader->baseShader.functionLength;
1559         return WINED3D_OK;
1560     }
1561
1562     if (*data_size < shader->baseShader.functionLength)
1563     {
1564         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1565          * than the required size we should write the required size and
1566          * return D3DERR_MOREDATA. That's not actually true. */
1567         return WINED3DERR_INVALIDCALL;
1568     }
1569
1570     memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1571
1572     return WINED3D_OK;
1573 }
1574
1575 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1576         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1577 {
1578     struct wined3d_shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1579     const struct wined3d_shader_frontend *fe;
1580     HRESULT hr;
1581
1582     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1583             shader, byte_code, output_signature, float_const_count);
1584
1585     fe = shader_select_frontend(*byte_code);
1586     if (!fe)
1587     {
1588         FIXME("Unable to find frontend for shader.\n");
1589         return WINED3DERR_INVALIDCALL;
1590     }
1591     shader->baseShader.frontend = fe;
1592     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1593     if (!shader->baseShader.frontend_data)
1594     {
1595         FIXME("Failed to initialize frontend.\n");
1596         return WINED3DERR_INVALIDCALL;
1597     }
1598
1599     /* First pass: trace shader. */
1600     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1601
1602     /* Initialize immediate constant lists. */
1603     list_init(&shader->baseShader.constantsF);
1604     list_init(&shader->baseShader.constantsB);
1605     list_init(&shader->baseShader.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->baseShader.input_signature, shader->baseShader.output_signature,
1610             byte_code, float_const_count);
1611     if (FAILED(hr)) return hr;
1612
1613     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1614     if (!shader->baseShader.function) return E_OUTOFMEMORY;
1615     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1616
1617     return WINED3D_OK;
1618 }
1619
1620 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1621 {
1622     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1623
1624     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1625             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1626             || IsEqualGUID(riid, &IID_IWineD3DBase)
1627             || IsEqualGUID(riid, &IID_IUnknown))
1628     {
1629         IUnknown_AddRef(iface);
1630         *object = iface;
1631         return S_OK;
1632     }
1633
1634     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1635
1636     *object = NULL;
1637     return E_NOINTERFACE;
1638 }
1639
1640 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1641 {
1642     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1643     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1644
1645     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1646
1647     return refcount;
1648 }
1649
1650 /* Do not call while under the GL lock. */
1651 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1652 {
1653     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1654     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1655
1656     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1657
1658     if (!refcount)
1659     {
1660         shader_cleanup((IWineD3DBaseShaderImpl *)shader);
1661         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1662         HeapFree(GetProcessHeap(), 0, shader);
1663     }
1664
1665     return refcount;
1666 }
1667
1668 static void * STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface)
1669 {
1670     TRACE("iface %p.\n", iface);
1671
1672     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1673 }
1674
1675 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1676 {
1677     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1678
1679     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1680 }
1681
1682 /* Set local constants for d3d8 shaders. */
1683 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1684         UINT start_idx, const float *src_data, UINT count)
1685 {
1686     IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1687     IWineD3DDeviceImpl *device = shader->baseShader.device;
1688     UINT i, end_idx;
1689
1690     TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1691
1692     end_idx = start_idx + count;
1693     if (end_idx > device->d3d_vshader_constantF)
1694     {
1695         WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1696         end_idx = device->d3d_vshader_constantF;
1697     }
1698
1699     for (i = start_idx; i < end_idx; ++i)
1700     {
1701         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1702         if (!lconst) return E_OUTOFMEMORY;
1703
1704         lconst->idx = i;
1705         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1706         list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1707     }
1708
1709     return WINED3D_OK;
1710 }
1711
1712 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1713 {
1714     /* IUnknown methods */
1715     vertexshader_QueryInterface,
1716     vertexshader_AddRef,
1717     vertexshader_Release,
1718     /* IWineD3DBase methods */
1719     vertexshader_GetParent,
1720     /* IWineD3DBaseShader methods */
1721     vertexshader_GetFunction,
1722     /* IWineD3DVertexShader methods */
1723     vertexshader_SetLocalConstantsF,
1724 };
1725
1726 void find_vs_compile_args(const struct wined3d_state *state,
1727         IWineD3DVertexShaderImpl *shader, struct vs_compile_args *args)
1728 {
1729     args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1730             == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1731     args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1732             && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1733     args->swizzle_map = shader->baseShader.device->strided_streams.swizzle_map;
1734 }
1735
1736 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1737 {
1738     if (usage_idx1 != usage_idx2) return FALSE;
1739     if (usage1 == usage2) return TRUE;
1740     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1741     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1742
1743     return FALSE;
1744 }
1745
1746 BOOL vshader_get_input(struct IWineD3DVertexShaderImpl *shader,
1747         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1748 {
1749     WORD map = shader->baseShader.reg_maps.input_registers;
1750     unsigned int i;
1751
1752     for (i = 0; map; map >>= 1, ++i)
1753     {
1754         if (!(map & 1)) continue;
1755
1756         if (match_usage(shader->attributes[i].usage,
1757                 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1758         {
1759             *regnum = i;
1760             return TRUE;
1761         }
1762     }
1763     return FALSE;
1764 }
1765
1766 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1767 {
1768     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1769             shader->baseShader.reg_maps.shader_version.minor);
1770     IWineD3DDeviceImpl *device = shader->baseShader.device;
1771
1772     shader->baseShader.limits.texcoord = 0;
1773     shader->baseShader.limits.attributes = 16;
1774     shader->baseShader.limits.packed_input = 0;
1775
1776     switch (shader_version)
1777     {
1778         case WINED3D_SHADER_VERSION(1, 0):
1779         case WINED3D_SHADER_VERSION(1, 1):
1780             shader->baseShader.limits.temporary = 12;
1781             shader->baseShader.limits.constant_bool = 0;
1782             shader->baseShader.limits.constant_int = 0;
1783             shader->baseShader.limits.address = 1;
1784             shader->baseShader.limits.packed_output = 12;
1785             shader->baseShader.limits.sampler = 0;
1786             shader->baseShader.limits.label = 0;
1787             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1788              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1789              * constants? */
1790             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1791             break;
1792
1793         case WINED3D_SHADER_VERSION(2, 0):
1794         case WINED3D_SHADER_VERSION(2, 1):
1795             shader->baseShader.limits.temporary = 12;
1796             shader->baseShader.limits.constant_bool = 16;
1797             shader->baseShader.limits.constant_int = 16;
1798             shader->baseShader.limits.address = 1;
1799             shader->baseShader.limits.packed_output = 12;
1800             shader->baseShader.limits.sampler = 0;
1801             shader->baseShader.limits.label = 16;
1802             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1803             break;
1804
1805         case WINED3D_SHADER_VERSION(4, 0):
1806             FIXME("Using 3.0 limits for 4.0 shader.\n");
1807             /* Fall through. */
1808
1809         case WINED3D_SHADER_VERSION(3, 0):
1810             shader->baseShader.limits.temporary = 32;
1811             shader->baseShader.limits.constant_bool = 32;
1812             shader->baseShader.limits.constant_int = 32;
1813             shader->baseShader.limits.address = 1;
1814             shader->baseShader.limits.packed_output = 12;
1815             shader->baseShader.limits.sampler = 4;
1816             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1817             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1818              * even though they are capable of supporting much more (GL
1819              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1820              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1821              * shaders to 256. */
1822             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1823             break;
1824
1825         default:
1826             shader->baseShader.limits.temporary = 12;
1827             shader->baseShader.limits.constant_bool = 16;
1828             shader->baseShader.limits.constant_int = 16;
1829             shader->baseShader.limits.address = 1;
1830             shader->baseShader.limits.packed_output = 12;
1831             shader->baseShader.limits.sampler = 0;
1832             shader->baseShader.limits.label = 16;
1833             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1834             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1835                     shader->baseShader.reg_maps.shader_version.major,
1836                     shader->baseShader.reg_maps.shader_version.minor);
1837     }
1838 }
1839
1840 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1841         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1842         void *parent, const struct wined3d_parent_ops *parent_ops)
1843 {
1844     struct wined3d_shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1845     unsigned int i;
1846     HRESULT hr;
1847     WORD map;
1848
1849     if (!byte_code) return WINED3DERR_INVALIDCALL;
1850
1851     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1852     shader_init(&shader->baseShader, device, parent, parent_ops);
1853
1854     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1855             output_signature, device->d3d_vshader_constantF);
1856     if (FAILED(hr))
1857     {
1858         WARN("Failed to set function, hr %#x.\n", hr);
1859         shader_cleanup((IWineD3DBaseShaderImpl *)shader);
1860         return hr;
1861     }
1862
1863     map = reg_maps->input_registers;
1864     for (i = 0; map; map >>= 1, ++i)
1865     {
1866         if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1867
1868         shader->attributes[i].usage =
1869                 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1870         shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1871     }
1872
1873     if (output_signature)
1874     {
1875         for (i = 0; i < output_signature->element_count; ++i)
1876         {
1877             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1878             reg_maps->output_registers |= 1 << e->register_idx;
1879             shader->baseShader.output_signature[e->register_idx] = *e;
1880         }
1881     }
1882
1883     vertexshader_set_limits(shader);
1884
1885     shader->baseShader.load_local_constsF = reg_maps->usesrelconstF
1886             && !list_empty(&shader->baseShader.constantsF);
1887
1888     return WINED3D_OK;
1889 }
1890
1891 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1892         REFIID riid, void **object)
1893 {
1894     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1895
1896     if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1897             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1898             || IsEqualGUID(riid, &IID_IWineD3DBase)
1899             || IsEqualGUID(riid, &IID_IUnknown))
1900     {
1901         IUnknown_AddRef(iface);
1902         *object = iface;
1903         return S_OK;
1904     }
1905
1906     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1907
1908     *object = NULL;
1909     return E_NOINTERFACE;
1910 }
1911
1912 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1913 {
1914     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1915     ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1916
1917     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1918
1919     return refcount;
1920 }
1921
1922 /* Do not call while under the GL lock. */
1923 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1924 {
1925     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1926     ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1927
1928     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1929
1930     if (!refcount)
1931     {
1932         shader_cleanup((IWineD3DBaseShaderImpl *)shader);
1933         shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1934         HeapFree(GetProcessHeap(), 0, shader);
1935     }
1936
1937     return refcount;
1938 }
1939
1940 static void * STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface)
1941 {
1942     TRACE("iface %p.\n", iface);
1943
1944     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
1945 }
1946
1947 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1948 {
1949     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1950
1951     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1952 }
1953
1954 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1955 {
1956     /* IUnknown methods */
1957     geometryshader_QueryInterface,
1958     geometryshader_AddRef,
1959     geometryshader_Release,
1960     /* IWineD3DBase methods */
1961     geometryshader_GetParent,
1962     /* IWineD3DBaseShader methods */
1963     geometryshader_GetFunction,
1964 };
1965
1966 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1967         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1968         void *parent, const struct wined3d_parent_ops *parent_ops)
1969 {
1970     HRESULT hr;
1971
1972     shader->vtbl = &wined3d_geometryshader_vtbl;
1973     shader_init(&shader->base_shader, device, parent, parent_ops);
1974
1975     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1976     if (FAILED(hr))
1977     {
1978         WARN("Failed to set function, hr %#x.\n", hr);
1979         shader_cleanup((IWineD3DBaseShaderImpl *)shader);
1980         return hr;
1981     }
1982
1983     shader->base_shader.load_local_constsF = FALSE;
1984
1985     return WINED3D_OK;
1986 }
1987
1988 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1989 {
1990     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1991
1992     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1993             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1994             || IsEqualGUID(riid, &IID_IWineD3DBase)
1995             || IsEqualGUID(riid, &IID_IUnknown))
1996     {
1997         IUnknown_AddRef(iface);
1998         *object = iface;
1999         return S_OK;
2000     }
2001
2002     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
2003
2004     *object = NULL;
2005     return E_NOINTERFACE;
2006 }
2007
2008 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
2009 {
2010     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2011     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
2012
2013     TRACE("%p increasing refcount to %u.\n", shader, refcount);
2014
2015     return refcount;
2016 }
2017
2018 /* Do not call while under the GL lock. */
2019 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
2020 {
2021     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2022     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
2023
2024     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2025
2026     if (!refcount)
2027     {
2028         shader_cleanup((IWineD3DBaseShaderImpl *)shader);
2029         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
2030         HeapFree(GetProcessHeap(), 0, shader);
2031     }
2032
2033     return refcount;
2034 }
2035
2036 static void * STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface)
2037 {
2038     TRACE("iface %p.\n", iface);
2039
2040     return ((IWineD3DBaseShaderImpl *)iface)->baseShader.parent;
2041 }
2042
2043 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
2044 {
2045     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
2046
2047     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
2048 }
2049
2050 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2051 {
2052     /* IUnknown methods */
2053     pixelshader_QueryInterface,
2054     pixelshader_AddRef,
2055     pixelshader_Release,
2056     /* IWineD3DBase methods */
2057     pixelshader_GetParent,
2058     /* IWineD3DBaseShader methods */
2059     pixelshader_GetFunction
2060 };
2061
2062 void find_ps_compile_args(const struct wined3d_state *state,
2063         IWineD3DPixelShaderImpl *shader, struct ps_compile_args *args)
2064 {
2065     IWineD3DDeviceImpl *device = shader->baseShader.device;
2066     const struct wined3d_texture *texture;
2067     UINT i;
2068
2069     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2070     if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
2071     {
2072         IWineD3DSurfaceImpl *rt = device->render_targets[0];
2073         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
2074     }
2075
2076     if (shader->baseShader.reg_maps.shader_version.major == 1
2077             && shader->baseShader.reg_maps.shader_version.minor <= 3)
2078     {
2079         for (i = 0; i < 4; ++i)
2080         {
2081             DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
2082             DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
2083             if (flags & WINED3DTTFF_PROJECTED)
2084                 tex_transform |= WINED3D_PSARGS_PROJECTED;
2085             args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2086         }
2087     }
2088
2089     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2090     {
2091         if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2092         texture = state->textures[i];
2093         if (!texture)
2094         {
2095             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2096             continue;
2097         }
2098         args->color_fixup[i] = texture->resource.format->color_fixup;
2099
2100         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2101             args->shadow |= 1 << i;
2102
2103         /* Flag samplers that need NP2 texcoord fixup. */
2104         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2105             args->np2_fixup |= (1 << i);
2106     }
2107     if (shader->baseShader.reg_maps.shader_version.major >= 3)
2108     {
2109         if (device->strided_streams.position_transformed)
2110         {
2111             args->vp_mode = pretransformed;
2112         }
2113         else if (use_vs(state))
2114         {
2115             args->vp_mode = vertexshader;
2116         }
2117         else
2118         {
2119             args->vp_mode = fixedfunction;
2120         }
2121         args->fog = FOG_OFF;
2122     }
2123     else
2124     {
2125         args->vp_mode = vertexshader;
2126         if (state->render_states[WINED3DRS_FOGENABLE])
2127         {
2128             switch (state->render_states[WINED3DRS_FOGTABLEMODE])
2129             {
2130                 case WINED3DFOG_NONE:
2131                     if (device->strided_streams.position_transformed || use_vs(state))
2132                     {
2133                         args->fog = FOG_LINEAR;
2134                         break;
2135                     }
2136
2137                     switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
2138                     {
2139                         case WINED3DFOG_NONE: /* Fall through. */
2140                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2141                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2142                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2143                     }
2144                     break;
2145
2146                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2147                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2148                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2149             }
2150         }
2151         else
2152         {
2153             args->fog = FOG_OFF;
2154         }
2155     }
2156 }
2157
2158 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2159 {
2160     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2161             shader->baseShader.reg_maps.shader_version.minor);
2162
2163     shader->baseShader.limits.attributes = 0;
2164     shader->baseShader.limits.address = 0;
2165     shader->baseShader.limits.packed_output = 0;
2166
2167     switch (shader_version)
2168     {
2169         case WINED3D_SHADER_VERSION(1, 0):
2170         case WINED3D_SHADER_VERSION(1, 1):
2171         case WINED3D_SHADER_VERSION(1, 2):
2172         case WINED3D_SHADER_VERSION(1, 3):
2173             shader->baseShader.limits.temporary = 2;
2174             shader->baseShader.limits.constant_float = 8;
2175             shader->baseShader.limits.constant_int = 0;
2176             shader->baseShader.limits.constant_bool = 0;
2177             shader->baseShader.limits.texcoord = 4;
2178             shader->baseShader.limits.sampler = 4;
2179             shader->baseShader.limits.packed_input = 0;
2180             shader->baseShader.limits.label = 0;
2181             break;
2182
2183         case WINED3D_SHADER_VERSION(1, 4):
2184             shader->baseShader.limits.temporary = 6;
2185             shader->baseShader.limits.constant_float = 8;
2186             shader->baseShader.limits.constant_int = 0;
2187             shader->baseShader.limits.constant_bool = 0;
2188             shader->baseShader.limits.texcoord = 6;
2189             shader->baseShader.limits.sampler = 6;
2190             shader->baseShader.limits.packed_input = 0;
2191             shader->baseShader.limits.label = 0;
2192             break;
2193
2194         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2195         case WINED3D_SHADER_VERSION(2, 0):
2196             shader->baseShader.limits.temporary = 32;
2197             shader->baseShader.limits.constant_float = 32;
2198             shader->baseShader.limits.constant_int = 16;
2199             shader->baseShader.limits.constant_bool = 16;
2200             shader->baseShader.limits.texcoord = 8;
2201             shader->baseShader.limits.sampler = 16;
2202             shader->baseShader.limits.packed_input = 0;
2203             break;
2204
2205         case WINED3D_SHADER_VERSION(2, 1):
2206             shader->baseShader.limits.temporary = 32;
2207             shader->baseShader.limits.constant_float = 32;
2208             shader->baseShader.limits.constant_int = 16;
2209             shader->baseShader.limits.constant_bool = 16;
2210             shader->baseShader.limits.texcoord = 8;
2211             shader->baseShader.limits.sampler = 16;
2212             shader->baseShader.limits.packed_input = 0;
2213             shader->baseShader.limits.label = 16;
2214             break;
2215
2216         case WINED3D_SHADER_VERSION(4, 0):
2217             FIXME("Using 3.0 limits for 4.0 shader.\n");
2218             /* Fall through. */
2219
2220         case WINED3D_SHADER_VERSION(3, 0):
2221             shader->baseShader.limits.temporary = 32;
2222             shader->baseShader.limits.constant_float = 224;
2223             shader->baseShader.limits.constant_int = 16;
2224             shader->baseShader.limits.constant_bool = 16;
2225             shader->baseShader.limits.texcoord = 0;
2226             shader->baseShader.limits.sampler = 16;
2227             shader->baseShader.limits.packed_input = 12;
2228             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2229             break;
2230
2231         default:
2232             shader->baseShader.limits.temporary = 32;
2233             shader->baseShader.limits.constant_float = 32;
2234             shader->baseShader.limits.constant_int = 16;
2235             shader->baseShader.limits.constant_bool = 16;
2236             shader->baseShader.limits.texcoord = 8;
2237             shader->baseShader.limits.sampler = 16;
2238             shader->baseShader.limits.packed_input = 0;
2239             shader->baseShader.limits.label = 0;
2240             FIXME("Unrecognized pixel shader version %u.%u\n",
2241                     shader->baseShader.reg_maps.shader_version.major,
2242                     shader->baseShader.reg_maps.shader_version.minor);
2243     }
2244 }
2245
2246 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2247         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2248         void *parent, const struct wined3d_parent_ops *parent_ops)
2249 {
2250     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2251     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2252     HRESULT hr;
2253
2254     if (!byte_code) return WINED3DERR_INVALIDCALL;
2255
2256     shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2257     shader_init(&shader->baseShader, device, parent, parent_ops);
2258
2259     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2260             output_signature, device->d3d_pshader_constantF);
2261     if (FAILED(hr))
2262     {
2263         WARN("Failed to set function, hr %#x.\n", hr);
2264         shader_cleanup((IWineD3DBaseShaderImpl *)shader);
2265         return hr;
2266     }
2267
2268     pixelshader_set_limits(shader);
2269
2270     for (i = 0; i < MAX_REG_INPUT; ++i)
2271     {
2272         if (shader->input_reg_used[i])
2273         {
2274             ++num_regs_used;
2275             highest_reg_used = i;
2276         }
2277     }
2278
2279     /* Don't do any register mapping magic if it is not needed, or if we can't
2280      * achieve anything anyway */
2281     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2282             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2283     {
2284         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2285         {
2286             /* This happens with relative addressing. The input mapper function
2287              * warns about this if the higher registers are declared too, so
2288              * don't write a FIXME here */
2289             WARN("More varying registers used than supported\n");
2290         }
2291
2292         for (i = 0; i < MAX_REG_INPUT; ++i)
2293         {
2294             shader->input_reg_map[i] = i;
2295         }
2296
2297         shader->declared_in_count = highest_reg_used + 1;
2298     }
2299     else
2300     {
2301         shader->declared_in_count = 0;
2302         for (i = 0; i < MAX_REG_INPUT; ++i)
2303         {
2304             if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2305             else shader->input_reg_map[i] = ~0U;
2306         }
2307     }
2308
2309     shader->baseShader.load_local_constsF = FALSE;
2310
2311     return WINED3D_OK;
2312 }
2313
2314 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2315 {
2316     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2317     unsigned int i;
2318
2319     if (reg_maps->shader_version.major != 1) return;
2320
2321     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2322     {
2323         /* We don't sample from this sampler. */
2324         if (!sampler_type[i]) continue;
2325
2326         if (!textures[i])
2327         {
2328             WARN("No texture bound to sampler %u, using 2D.\n", i);
2329             sampler_type[i] = WINED3DSTT_2D;
2330             continue;
2331         }
2332
2333         switch (textures[i]->target)
2334         {
2335             case GL_TEXTURE_RECTANGLE_ARB:
2336             case GL_TEXTURE_2D:
2337                 /* We have to select between texture rectangles and 2D
2338                  * textures later because 2.0 and 3.0 shaders only have
2339                  * WINED3DSTT_2D as well. */
2340                 sampler_type[i] = WINED3DSTT_2D;
2341                 break;
2342
2343             case GL_TEXTURE_3D:
2344                 sampler_type[i] = WINED3DSTT_VOLUME;
2345                 break;
2346
2347             case GL_TEXTURE_CUBE_MAP_ARB:
2348                 sampler_type[i] = WINED3DSTT_CUBE;
2349                 break;
2350
2351             default:
2352                 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2353                 sampler_type[i] = WINED3DSTT_2D;
2354         }
2355     }
2356 }