git.oblomov.eu Git - wine/blob - dlls/wined3d/shader.c

   1 /*
   2  * Copyright 2002-2003 Jason Edmeades
   3  * Copyright 2002-2003 Raphael Junqueira
   4  * Copyright 2004 Christian Costa
   5  * Copyright 2005 Oliver Stieber
   6  * Copyright 2006 Ivan Gyurdiev
   7  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
   8  * Copyright 2009-2011 Henri Verbeet for CodeWeavers
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #include "config.h"
  26
  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 wined3d_shader *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(struct wined3d_shader *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                         shader->u.ps.input_reg_used[i] = TRUE;
 377                     }
 378                 }
 379                 else
 380                 {
 381                     shader->u.ps.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(struct wined3d_shader *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->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->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->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->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                     if (dst_param.reg.type == WINED3DSPR_COLOROUT && !dst_param.reg.idx)
 720                     {
 721                         /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 722                          * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 723                          * the mov and perform the sRGB write correction from the source register.
 724                          *
 725                          * However, if the mov is only partial, we can't do this, and if the write
 726                          * comes from an instruction other than MOV it is hard to do as well. If
 727                          * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 728                         shader->u.ps.color0_mov = FALSE;
 729                         if (ins.handler_idx == WINED3DSIH_MOV
 730                                 && dst_param.write_mask == WINED3DSP_WRITEMASK_ALL)
 731                         {
 732                             /* Used later when the source register is read. */
 733                             color0_mov = TRUE;
 734                         }
 735                     }
 736                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 737                      * end
 738                      */
 739                     else if (dst_param.reg.type == WINED3DSPR_TEMP
 740                             && dst_param.reg.idx == shader->u.ps.color0_reg)
 741                     {
 742                         shader->u.ps.color0_mov = FALSE;
 743                     }
 744                 }
 745
 746                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 747                 if (shader_version.major == 1
 748                         && (ins.handler_idx == WINED3DSIH_TEX
 749                             || ins.handler_idx == WINED3DSIH_TEXBEM
 750                             || ins.handler_idx == WINED3DSIH_TEXBEML
 751                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 752                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 753                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 754                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 755                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 756                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 757                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 758                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 759                 {
 760                     /* Fake sampler usage, only set reserved bit and type. */
 761                     DWORD sampler_code = dst_param.reg.idx;
 762
 763                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 764                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 765
 766                     /* texbem is only valid with < 1.4 pixel shaders */
 767                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 768                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 769                     {
 770                         reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 771                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 772                         {
 773                             reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
 774                         }
 775                     }
 776                 }
 777                 else if (ins.handler_idx == WINED3DSIH_BEM)
 778                 {
 779                     reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 780                 }
 781             }
 782
 783             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 784             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 785             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 786             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 787             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 788             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 789             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 790             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 791             else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
 792
 793             limit = ins.src_count + (ins.predicate ? 1 : 0);
 794             for (i = 0; i < limit; ++i)
 795             {
 796                 struct wined3d_shader_src_param src_param, src_rel_addr;
 797                 unsigned int count;
 798
 799                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
 800                 count = get_instr_extra_regcount(ins.handler_idx, i);
 801
 802                 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 803                 while (count)
 804                 {
 805                     ++src_param.reg.idx;
 806                     shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 807                     --count;
 808                 }
 809
 810                 if (color0_mov)
 811                 {
 812                     if (src_param.reg.type == WINED3DSPR_TEMP
 813                             && src_param.swizzle == WINED3DSP_NOSWIZZLE)
 814                     {
 815                         shader->u.ps.color0_mov = TRUE;
 816                         shader->u.ps.color0_reg = src_param.reg.idx;
 817                     }
 818                 }
 819             }
 820         }
 821     }
 822     reg_maps->loop_depth = max_loop_depth;
 823
 824     shader->functionLength = ((const char *)ptr - (const char *)byte_code);
 825
 826     return WINED3D_OK;
 827 }
 828
 829 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 830 {
 831     DWORD map = 1 << max;
 832     map |= map - 1;
 833     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 834
 835     return wined3d_log2i(map);
 836 }
 837
 838 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 839         const struct wined3d_shader_version *shader_version)
 840 {
 841     TRACE("dcl");
 842
 843     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 844     {
 845         switch (semantic->sampler_type)
 846         {
 847             case WINED3DSTT_2D: TRACE("_2d"); break;
 848             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 849             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 850             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 851         }
 852     }
 853     else
 854     {
 855         /* Pixel shaders 3.0 don't have usage semantics. */
 856         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 857         else TRACE("_");
 858
 859         switch (semantic->usage)
 860         {
 861             case WINED3DDECLUSAGE_POSITION:
 862                 TRACE("position%u", semantic->usage_idx);
 863                 break;
 864
 865             case WINED3DDECLUSAGE_BLENDINDICES:
 866                 TRACE("blend");
 867                 break;
 868
 869             case WINED3DDECLUSAGE_BLENDWEIGHT:
 870                 TRACE("weight");
 871                 break;
 872
 873             case WINED3DDECLUSAGE_NORMAL:
 874                 TRACE("normal%u", semantic->usage_idx);
 875                 break;
 876
 877             case WINED3DDECLUSAGE_PSIZE:
 878                 TRACE("psize");
 879                 break;
 880
 881             case WINED3DDECLUSAGE_COLOR:
 882                 if (!semantic->usage_idx) TRACE("color");
 883                 else TRACE("specular%u", (semantic->usage_idx - 1));
 884                 break;
 885
 886             case WINED3DDECLUSAGE_TEXCOORD:
 887                 TRACE("texture%u", semantic->usage_idx);
 888                 break;
 889
 890             case WINED3DDECLUSAGE_TANGENT:
 891                 TRACE("tangent");
 892                 break;
 893
 894             case WINED3DDECLUSAGE_BINORMAL:
 895                 TRACE("binormal");
 896                 break;
 897
 898             case WINED3DDECLUSAGE_TESSFACTOR:
 899                 TRACE("tessfactor");
 900                 break;
 901
 902             case WINED3DDECLUSAGE_POSITIONT:
 903                 TRACE("positionT%u", semantic->usage_idx);
 904                 break;
 905
 906             case WINED3DDECLUSAGE_FOG:
 907                 TRACE("fog");
 908                 break;
 909
 910             case WINED3DDECLUSAGE_DEPTH:
 911                 TRACE("depth");
 912                 break;
 913
 914             case WINED3DDECLUSAGE_SAMPLE:
 915                 TRACE("sample");
 916                 break;
 917
 918             default:
 919                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 920         }
 921     }
 922 }
 923
 924 static void shader_dump_register(const struct wined3d_shader_register *reg,
 925         const struct wined3d_shader_version *shader_version)
 926 {
 927     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 928     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 929     UINT offset = reg->idx;
 930
 931     switch (reg->type)
 932     {
 933         case WINED3DSPR_TEMP:
 934             TRACE("r");
 935             break;
 936
 937         case WINED3DSPR_INPUT:
 938             TRACE("v");
 939             break;
 940
 941         case WINED3DSPR_CONST:
 942         case WINED3DSPR_CONST2:
 943         case WINED3DSPR_CONST3:
 944         case WINED3DSPR_CONST4:
 945             TRACE("c");
 946             offset = shader_get_float_offset(reg->type, reg->idx);
 947             break;
 948
 949         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 950             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 951             break;
 952
 953         case WINED3DSPR_RASTOUT:
 954             TRACE("%s", rastout_reg_names[reg->idx]);
 955             break;
 956
 957         case WINED3DSPR_COLOROUT:
 958             TRACE("oC");
 959             break;
 960
 961         case WINED3DSPR_DEPTHOUT:
 962             TRACE("oDepth");
 963             break;
 964
 965         case WINED3DSPR_ATTROUT:
 966             TRACE("oD");
 967             break;
 968
 969         case WINED3DSPR_TEXCRDOUT:
 970             /* Vertex shaders >= 3.0 use general purpose output registers
 971              * (WINED3DSPR_OUTPUT), which can include an address token. */
 972             if (shader_version->major >= 3) TRACE("o");
 973             else TRACE("oT");
 974             break;
 975
 976         case WINED3DSPR_CONSTINT:
 977             TRACE("i");
 978             break;
 979
 980         case WINED3DSPR_CONSTBOOL:
 981             TRACE("b");
 982             break;
 983
 984         case WINED3DSPR_LABEL:
 985             TRACE("l");
 986             break;
 987
 988         case WINED3DSPR_LOOP:
 989             TRACE("aL");
 990             break;
 991
 992         case WINED3DSPR_SAMPLER:
 993             TRACE("s");
 994             break;
 995
 996         case WINED3DSPR_MISCTYPE:
 997             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
 998             else TRACE("%s", misctype_reg_names[reg->idx]);
 999             break;
1000
1001         case WINED3DSPR_PREDICATE:
1002             TRACE("p");
1003             break;
1004
1005         case WINED3DSPR_IMMCONST:
1006             TRACE("l");
1007             break;
1008
1009         case WINED3DSPR_CONSTBUFFER:
1010             TRACE("cb");
1011             break;
1012
1013         case WINED3DSPR_NULL:
1014             TRACE("null");
1015             break;
1016
1017         case WINED3DSPR_RESOURCE:
1018             TRACE("t");
1019             break;
1020
1021         default:
1022             TRACE("unhandled_rtype(%#x)", reg->type);
1023             break;
1024     }
1025
1026     if (reg->type == WINED3DSPR_IMMCONST)
1027     {
1028         TRACE("(");
1029         switch (reg->immconst_type)
1030         {
1031             case WINED3D_IMMCONST_SCALAR:
1032                 TRACE("%.8e", *(const float *)reg->immconst_data);
1033                 break;
1034
1035             case WINED3D_IMMCONST_VEC4:
1036                 TRACE("%.8e, %.8e, %.8e, %.8e",
1037                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1038                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1039                 break;
1040
1041             default:
1042                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1043                 break;
1044         }
1045         TRACE(")");
1046     }
1047     else if (reg->type != WINED3DSPR_RASTOUT
1048             && reg->type != WINED3DSPR_MISCTYPE
1049             && reg->type != WINED3DSPR_NULL)
1050     {
1051         if (reg->array_idx != ~0U)
1052         {
1053             TRACE("%u[%u", offset, reg->array_idx);
1054             if (reg->rel_addr)
1055             {
1056                 TRACE(" + ");
1057                 shader_dump_src_param(reg->rel_addr, shader_version);
1058             }
1059             TRACE("]");
1060         }
1061         else
1062         {
1063             if (reg->rel_addr)
1064             {
1065                 TRACE("[");
1066                 shader_dump_src_param(reg->rel_addr, shader_version);
1067                 TRACE(" + ");
1068             }
1069             TRACE("%u", offset);
1070             if (reg->rel_addr) TRACE("]");
1071         }
1072     }
1073 }
1074
1075 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1076         const struct wined3d_shader_version *shader_version)
1077 {
1078     DWORD write_mask = param->write_mask;
1079
1080     shader_dump_register(&param->reg, shader_version);
1081
1082     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1083     {
1084         static const char *write_mask_chars = "xyzw";
1085
1086         TRACE(".");
1087         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1088         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1089         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1090         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1091     }
1092 }
1093
1094 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1095         const struct wined3d_shader_version *shader_version)
1096 {
1097     DWORD src_modifier = param->modifiers;
1098     DWORD swizzle = param->swizzle;
1099
1100     if (src_modifier == WINED3DSPSM_NEG
1101             || src_modifier == WINED3DSPSM_BIASNEG
1102             || src_modifier == WINED3DSPSM_SIGNNEG
1103             || src_modifier == WINED3DSPSM_X2NEG
1104             || src_modifier == WINED3DSPSM_ABSNEG)
1105         TRACE("-");
1106     else if (src_modifier == WINED3DSPSM_COMP)
1107         TRACE("1-");
1108     else if (src_modifier == WINED3DSPSM_NOT)
1109         TRACE("!");
1110
1111     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1112         TRACE("abs(");
1113
1114     shader_dump_register(&param->reg, shader_version);
1115
1116     if (src_modifier)
1117     {
1118         switch (src_modifier)
1119         {
1120             case WINED3DSPSM_NONE:    break;
1121             case WINED3DSPSM_NEG:     break;
1122             case WINED3DSPSM_NOT:     break;
1123             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1124             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1125             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1126             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1127             case WINED3DSPSM_COMP:    break;
1128             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1129             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1130             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1131             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1132             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1133             case WINED3DSPSM_ABS:     TRACE(")"); break;
1134             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1135         }
1136     }
1137
1138     if (swizzle != WINED3DSP_NOSWIZZLE)
1139     {
1140         static const char *swizzle_chars = "xyzw";
1141         DWORD swizzle_x = swizzle & 0x03;
1142         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1143         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1144         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1145
1146         if (swizzle_x == swizzle_y
1147                 && swizzle_x == swizzle_z
1148                 && swizzle_x == swizzle_w)
1149         {
1150             TRACE(".%c", swizzle_chars[swizzle_x]);
1151         }
1152         else
1153         {
1154             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1155                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1156         }
1157     }
1158 }
1159
1160 /* Shared code in order to generate the bulk of the shader string.
1161  * NOTE: A description of how to parse tokens can be found on MSDN. */
1162 void shader_generate_main(struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1163         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1164 {
1165     IWineD3DDeviceImpl *device = shader->device;
1166     const struct wined3d_shader_frontend *fe = shader->frontend;
1167     void *fe_data = shader->frontend_data;
1168     struct wined3d_shader_src_param dst_rel_addr[2];
1169     struct wined3d_shader_src_param src_rel_addr[4];
1170     struct wined3d_shader_dst_param dst_param[2];
1171     struct wined3d_shader_src_param src_param[4];
1172     struct wined3d_shader_version shader_version;
1173     struct wined3d_shader_loop_state loop_state;
1174     struct wined3d_shader_instruction ins;
1175     struct wined3d_shader_tex_mx tex_mx;
1176     struct wined3d_shader_context ctx;
1177     const DWORD *ptr = byte_code;
1178     DWORD i;
1179
1180     /* Initialize current parsing state. */
1181     tex_mx.current_row = 0;
1182     loop_state.current_depth = 0;
1183     loop_state.current_reg = 0;
1184
1185     ctx.shader = shader;
1186     ctx.gl_info = &device->adapter->gl_info;
1187     ctx.reg_maps = reg_maps;
1188     ctx.buffer = buffer;
1189     ctx.tex_mx = &tex_mx;
1190     ctx.loop_state = &loop_state;
1191     ctx.backend_data = backend_ctx;
1192
1193     ins.ctx = &ctx;
1194     ins.dst = dst_param;
1195     ins.src = src_param;
1196
1197     fe->shader_read_header(fe_data, &ptr, &shader_version);
1198
1199     while (!fe->shader_is_end(fe_data, &ptr))
1200     {
1201         const char *comment;
1202         UINT comment_size;
1203         UINT param_size;
1204
1205         /* Skip comment tokens. */
1206         fe->shader_read_comment(&ptr, &comment, &comment_size);
1207         if (comment) continue;
1208
1209         /* Read opcode. */
1210         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1211
1212         /* Unknown opcode and its parameters. */
1213         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1214         {
1215             TRACE("Skipping unrecognized instruction.\n");
1216             ptr += param_size;
1217             continue;
1218         }
1219
1220         /* Nothing to do. */
1221         if (ins.handler_idx == WINED3DSIH_DCL
1222                 || ins.handler_idx == WINED3DSIH_NOP
1223                 || ins.handler_idx == WINED3DSIH_DEF
1224                 || ins.handler_idx == WINED3DSIH_DEFI
1225                 || ins.handler_idx == WINED3DSIH_DEFB
1226                 || ins.handler_idx == WINED3DSIH_PHASE)
1227         {
1228             ptr += param_size;
1229             continue;
1230         }
1231
1232         /* Destination tokens */
1233         for (i = 0; i < ins.dst_count; ++i)
1234         {
1235             fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1236         }
1237
1238         /* Predication token */
1239         if (ins.predicate)
1240         {
1241             FIXME("Predicates not implemented.\n");
1242             ins.predicate = *ptr++;
1243         }
1244
1245         /* Other source tokens */
1246         for (i = 0; i < ins.src_count; ++i)
1247         {
1248             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1249         }
1250
1251         /* Call appropriate function for output target */
1252         device->shader_backend->shader_handle_instruction(&ins);
1253     }
1254 }
1255
1256 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1257 {
1258     DWORD mmask = dst->modifiers;
1259
1260     switch (dst->shift)
1261     {
1262         case 0: break;
1263         case 13: TRACE("_d8"); break;
1264         case 14: TRACE("_d4"); break;
1265         case 15: TRACE("_d2"); break;
1266         case 1: TRACE("_x2"); break;
1267         case 2: TRACE("_x4"); break;
1268         case 3: TRACE("_x8"); break;
1269         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1270     }
1271
1272     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1273     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1274     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1275
1276     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1277     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1278 }
1279
1280 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1281 {
1282     struct wined3d_shader_version shader_version;
1283     const DWORD *ptr = byte_code;
1284     const char *type_prefix;
1285     DWORD i;
1286
1287     TRACE("Parsing %p.\n", byte_code);
1288
1289     fe->shader_read_header(fe_data, &ptr, &shader_version);
1290
1291     switch (shader_version.type)
1292     {
1293         case WINED3D_SHADER_TYPE_VERTEX:
1294             type_prefix = "vs";
1295             break;
1296
1297         case WINED3D_SHADER_TYPE_GEOMETRY:
1298             type_prefix = "gs";
1299             break;
1300
1301         case WINED3D_SHADER_TYPE_PIXEL:
1302             type_prefix = "ps";
1303             break;
1304
1305         default:
1306             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1307             type_prefix = "unknown";
1308             break;
1309     }
1310
1311     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1312
1313     while (!fe->shader_is_end(fe_data, &ptr))
1314     {
1315         struct wined3d_shader_instruction ins;
1316         const char *comment;
1317         UINT comment_size;
1318         UINT param_size;
1319
1320         /* comment */
1321         fe->shader_read_comment(&ptr, &comment, &comment_size);
1322         if (comment)
1323         {
1324             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1325             {
1326                 const char *end = comment + comment_size;
1327                 const char *ptr = comment + 4;
1328                 const char *line = ptr;
1329
1330                 TRACE("// TEXT\n");
1331                 while (ptr != end)
1332                 {
1333                     if (*ptr == '\n')
1334                     {
1335                         UINT len = ptr - line;
1336                         if (len && *(ptr - 1) == '\r') --len;
1337                         TRACE("// %s\n", debugstr_an(line, len));
1338                         line = ++ptr;
1339                     }
1340                     else ++ptr;
1341                 }
1342                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1343             }
1344             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1345             continue;
1346         }
1347
1348         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1349         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1350         {
1351             TRACE("Skipping unrecognized instruction.\n");
1352             ptr += param_size;
1353             continue;
1354         }
1355
1356         if (ins.handler_idx == WINED3DSIH_DCL)
1357         {
1358             struct wined3d_shader_semantic semantic;
1359
1360             fe->shader_read_semantic(&ptr, &semantic);
1361
1362             shader_dump_decl_usage(&semantic, &shader_version);
1363             shader_dump_ins_modifiers(&semantic.reg);
1364             TRACE(" ");
1365             shader_dump_dst_param(&semantic.reg, &shader_version);
1366         }
1367         else if (ins.handler_idx == WINED3DSIH_DEF)
1368         {
1369             struct wined3d_shader_src_param rel_addr;
1370             struct wined3d_shader_dst_param dst;
1371
1372             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1373
1374             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1375                     *(const float *)(ptr),
1376                     *(const float *)(ptr + 1),
1377                     *(const float *)(ptr + 2),
1378                     *(const float *)(ptr + 3));
1379             ptr += 4;
1380         }
1381         else if (ins.handler_idx == WINED3DSIH_DEFI)
1382         {
1383             struct wined3d_shader_src_param rel_addr;
1384             struct wined3d_shader_dst_param dst;
1385
1386             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1387
1388             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1389                     *(ptr),
1390                     *(ptr + 1),
1391                     *(ptr + 2),
1392                     *(ptr + 3));
1393             ptr += 4;
1394         }
1395         else if (ins.handler_idx == WINED3DSIH_DEFB)
1396         {
1397             struct wined3d_shader_src_param rel_addr;
1398             struct wined3d_shader_dst_param dst;
1399
1400             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1401
1402             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1403             ++ptr;
1404         }
1405         else
1406         {
1407             struct wined3d_shader_src_param dst_rel_addr[2];
1408             struct wined3d_shader_src_param src_rel_addr;
1409             struct wined3d_shader_dst_param dst_param[2];
1410             struct wined3d_shader_src_param src_param;
1411
1412             for (i = 0; i < ins.dst_count; ++i)
1413             {
1414                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1415             }
1416
1417             /* Print out predication source token first - it follows
1418              * the destination token. */
1419             if (ins.predicate)
1420             {
1421                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1422                 TRACE("(");
1423                 shader_dump_src_param(&src_param, &shader_version);
1424                 TRACE(") ");
1425             }
1426
1427             /* PixWin marks instructions with the coissue flag with a '+' */
1428             if (ins.coissue) TRACE("+");
1429
1430             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1431
1432             if (ins.handler_idx == WINED3DSIH_IFC
1433                     || ins.handler_idx == WINED3DSIH_BREAKC)
1434             {
1435                 switch (ins.flags)
1436                 {
1437                     case COMPARISON_GT: TRACE("_gt"); break;
1438                     case COMPARISON_EQ: TRACE("_eq"); break;
1439                     case COMPARISON_GE: TRACE("_ge"); break;
1440                     case COMPARISON_LT: TRACE("_lt"); break;
1441                     case COMPARISON_NE: TRACE("_ne"); break;
1442                     case COMPARISON_LE: TRACE("_le"); break;
1443                     default: TRACE("_(%u)", ins.flags);
1444                 }
1445             }
1446             else if (ins.handler_idx == WINED3DSIH_TEX
1447                     && shader_version.major >= 2
1448                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1449             {
1450                 TRACE("p");
1451             }
1452
1453             /* We already read the destination tokens, print them. */
1454             for (i = 0; i < ins.dst_count; ++i)
1455             {
1456                 shader_dump_ins_modifiers(&dst_param[i]);
1457                 TRACE(!i ? " " : ", ");
1458                 shader_dump_dst_param(&dst_param[i], &shader_version);
1459             }
1460
1461             /* Other source tokens */
1462             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1463             {
1464                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1465                 TRACE(!i ? " " : ", ");
1466                 shader_dump_src_param(&src_param, &shader_version);
1467             }
1468         }
1469         TRACE("\n");
1470     }
1471 }
1472
1473 static void shader_cleanup(struct wined3d_shader *shader)
1474 {
1475     shader->device->shader_backend->shader_destroy(shader);
1476     HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1477     HeapFree(GetProcessHeap(), 0, shader->function);
1478     shader_delete_constant_list(&shader->constantsF);
1479     shader_delete_constant_list(&shader->constantsB);
1480     shader_delete_constant_list(&shader->constantsI);
1481     list_remove(&shader->shader_list_entry);
1482
1483     if (shader->frontend && shader->frontend_data)
1484         shader->frontend->shader_free(shader->frontend_data);
1485 }
1486
1487 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1488 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1489 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1490         enum tex_types tex_type, const SIZE *ds_mask_size) {}
1491 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1492 static void shader_none_update_float_vertex_constants(IWineD3DDeviceImpl *device, UINT start, UINT count) {}
1493 static void shader_none_update_float_pixel_constants(IWineD3DDeviceImpl *device, UINT start, UINT count) {}
1494 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1495 static void shader_none_load_np2fixup_constants(void *shader_priv,
1496         const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1497 static void shader_none_destroy(struct wined3d_shader *shader) {}
1498 static HRESULT shader_none_alloc(IWineD3DDeviceImpl *device) {return WINED3D_OK;}
1499 static void shader_none_free(IWineD3DDeviceImpl *device) {}
1500 static BOOL shader_none_dirty_const(void) {return FALSE;}
1501
1502 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1503 {
1504     /* Set the shader caps to 0 for the none shader backend */
1505     caps->VertexShaderVersion = 0;
1506     caps->MaxVertexShaderConst = 0;
1507     caps->PixelShaderVersion = 0;
1508     caps->PixelShader1xMaxValue = 0.0f;
1509     caps->MaxPixelShaderConst = 0;
1510     caps->VSClipping = FALSE;
1511 }
1512
1513 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1514 {
1515     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1516     {
1517         TRACE("Checking support for fixup:\n");
1518         dump_color_fixup_desc(fixup);
1519     }
1520
1521     /* Faked to make some apps happy. */
1522     if (!is_complex_fixup(fixup))
1523     {
1524         TRACE("[OK]\n");
1525         return TRUE;
1526     }
1527
1528     TRACE("[FAILED]\n");
1529     return FALSE;
1530 }
1531
1532 const struct wined3d_shader_backend_ops none_shader_backend =
1533 {
1534     shader_none_handle_instruction,
1535     shader_none_select,
1536     shader_none_select_depth_blt,
1537     shader_none_deselect_depth_blt,
1538     shader_none_update_float_vertex_constants,
1539     shader_none_update_float_pixel_constants,
1540     shader_none_load_constants,
1541     shader_none_load_np2fixup_constants,
1542     shader_none_destroy,
1543     shader_none_alloc,
1544     shader_none_free,
1545     shader_none_dirty_const,
1546     shader_none_get_caps,
1547     shader_none_color_fixup_supported,
1548 };
1549
1550 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1551         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1552 {
1553     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1554     const struct wined3d_shader_frontend *fe;
1555     HRESULT hr;
1556
1557     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1558             shader, byte_code, output_signature, float_const_count);
1559
1560     fe = shader_select_frontend(*byte_code);
1561     if (!fe)
1562     {
1563         FIXME("Unable to find frontend for shader.\n");
1564         return WINED3DERR_INVALIDCALL;
1565     }
1566     shader->frontend = fe;
1567     shader->frontend_data = fe->shader_init(byte_code, output_signature);
1568     if (!shader->frontend_data)
1569     {
1570         FIXME("Failed to initialize frontend.\n");
1571         return WINED3DERR_INVALIDCALL;
1572     }
1573
1574     /* First pass: trace shader. */
1575     if (TRACE_ON(d3d_shader))
1576         shader_trace_init(fe, shader->frontend_data, byte_code);
1577
1578     /* Initialize immediate constant lists. */
1579     list_init(&shader->constantsF);
1580     list_init(&shader->constantsB);
1581     list_init(&shader->constantsI);
1582
1583     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1584     hr = shader_get_registers_used(shader, fe,
1585             reg_maps, shader->input_signature, shader->output_signature,
1586             byte_code, float_const_count);
1587     if (FAILED(hr)) return hr;
1588
1589     shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1590     if (!shader->function)
1591         return E_OUTOFMEMORY;
1592     memcpy(shader->function, byte_code, shader->functionLength);
1593
1594     return WINED3D_OK;
1595 }
1596
1597 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1598 {
1599     ULONG refcount = InterlockedIncrement(&shader->ref);
1600
1601     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1602
1603     return refcount;
1604 }
1605
1606 /* Do not call while under the GL lock. */
1607 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1608 {
1609     ULONG refcount = InterlockedDecrement(&shader->ref);
1610
1611     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1612
1613     if (!refcount)
1614     {
1615         shader_cleanup(shader);
1616         shader->parent_ops->wined3d_object_destroyed(shader->parent);
1617         HeapFree(GetProcessHeap(), 0, shader);
1618     }
1619
1620     return refcount;
1621 }
1622
1623 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1624 {
1625     TRACE("shader %p.\n", shader);
1626
1627     return shader->parent;
1628 }
1629
1630 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1631         void *byte_code, UINT *byte_code_size)
1632 {
1633     TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1634
1635     if (!byte_code)
1636     {
1637         *byte_code_size = shader->functionLength;
1638         return WINED3D_OK;
1639     }
1640
1641     if (*byte_code_size < shader->functionLength)
1642     {
1643         /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1644          * than the required size we should write the required size and
1645          * return D3DERR_MOREDATA. That's not actually true. */
1646         return WINED3DERR_INVALIDCALL;
1647     }
1648
1649     memcpy(byte_code, shader->function, shader->functionLength);
1650
1651     return WINED3D_OK;
1652 }
1653
1654 /* Set local constants for d3d8 shaders. */
1655 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1656         UINT start_idx, const float *src_data, UINT count)
1657 {
1658     UINT end_idx = start_idx + count;
1659     UINT i;
1660
1661     TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1662
1663     if (end_idx > shader->limits.constant_float)
1664     {
1665         WARN("end_idx %u > float constants limit %u.\n",
1666                 end_idx, shader->limits.constant_float);
1667         end_idx = shader->limits.constant_float;
1668     }
1669
1670     for (i = start_idx; i < end_idx; ++i)
1671     {
1672         struct local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1673         if (!lconst)
1674             return E_OUTOFMEMORY;
1675
1676         lconst->idx = i;
1677         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1678         list_add_head(&shader->constantsF, &lconst->entry);
1679     }
1680
1681     return WINED3D_OK;
1682 }
1683
1684 void find_vs_compile_args(const struct wined3d_state *state,
1685         const struct wined3d_shader *shader, struct vs_compile_args *args)
1686 {
1687     args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1688             == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1689     args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1690             && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1691     args->swizzle_map = shader->device->strided_streams.swizzle_map;
1692 }
1693
1694 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1695 {
1696     if (usage_idx1 != usage_idx2) return FALSE;
1697     if (usage1 == usage2) return TRUE;
1698     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1699     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1700
1701     return FALSE;
1702 }
1703
1704 BOOL vshader_get_input(struct wined3d_shader *shader,
1705         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1706 {
1707     WORD map = shader->reg_maps.input_registers;
1708     unsigned int i;
1709
1710     for (i = 0; map; map >>= 1, ++i)
1711     {
1712         if (!(map & 1)) continue;
1713
1714         if (match_usage(shader->u.vs.attributes[i].usage,
1715                 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1716         {
1717             *regnum = i;
1718             return TRUE;
1719         }
1720     }
1721     return FALSE;
1722 }
1723
1724 static void vertexshader_set_limits(struct wined3d_shader *shader)
1725 {
1726     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1727             shader->reg_maps.shader_version.minor);
1728     IWineD3DDeviceImpl *device = shader->device;
1729
1730     shader->limits.texcoord = 0;
1731     shader->limits.attributes = 16;
1732     shader->limits.packed_input = 0;
1733
1734     switch (shader_version)
1735     {
1736         case WINED3D_SHADER_VERSION(1, 0):
1737         case WINED3D_SHADER_VERSION(1, 1):
1738             shader->limits.temporary = 12;
1739             shader->limits.constant_bool = 0;
1740             shader->limits.constant_int = 0;
1741             shader->limits.address = 1;
1742             shader->limits.packed_output = 12;
1743             shader->limits.sampler = 0;
1744             shader->limits.label = 0;
1745             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1746              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1747              * constants? */
1748             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1749             break;
1750
1751         case WINED3D_SHADER_VERSION(2, 0):
1752         case WINED3D_SHADER_VERSION(2, 1):
1753             shader->limits.temporary = 12;
1754             shader->limits.constant_bool = 16;
1755             shader->limits.constant_int = 16;
1756             shader->limits.address = 1;
1757             shader->limits.packed_output = 12;
1758             shader->limits.sampler = 0;
1759             shader->limits.label = 16;
1760             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1761             break;
1762
1763         case WINED3D_SHADER_VERSION(4, 0):
1764             FIXME("Using 3.0 limits for 4.0 shader.\n");
1765             /* Fall through. */
1766
1767         case WINED3D_SHADER_VERSION(3, 0):
1768             shader->limits.temporary = 32;
1769             shader->limits.constant_bool = 32;
1770             shader->limits.constant_int = 32;
1771             shader->limits.address = 1;
1772             shader->limits.packed_output = 12;
1773             shader->limits.sampler = 4;
1774             shader->limits.label = 16; /* FIXME: 2048 */
1775             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1776              * even though they are capable of supporting much more (GL
1777              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1778              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1779              * shaders to 256. */
1780             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1781             break;
1782
1783         default:
1784             shader->limits.temporary = 12;
1785             shader->limits.constant_bool = 16;
1786             shader->limits.constant_int = 16;
1787             shader->limits.address = 1;
1788             shader->limits.packed_output = 12;
1789             shader->limits.sampler = 0;
1790             shader->limits.label = 16;
1791             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1792             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1793                     shader->reg_maps.shader_version.major,
1794                     shader->reg_maps.shader_version.minor);
1795     }
1796 }
1797
1798 HRESULT vertexshader_init(struct wined3d_shader *shader, IWineD3DDeviceImpl *device,
1799         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1800         void *parent, const struct wined3d_parent_ops *parent_ops)
1801 {
1802     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1803     unsigned int i;
1804     HRESULT hr;
1805     WORD map;
1806
1807     if (!byte_code) return WINED3DERR_INVALIDCALL;
1808
1809     shader_init(shader, device, parent, parent_ops);
1810     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF);
1811     if (FAILED(hr))
1812     {
1813         WARN("Failed to set function, hr %#x.\n", hr);
1814         shader_cleanup(shader);
1815         return hr;
1816     }
1817
1818     map = reg_maps->input_registers;
1819     for (i = 0; map; map >>= 1, ++i)
1820     {
1821         if (!(map & 1) || !shader->input_signature[i].semantic_name)
1822             continue;
1823
1824         shader->u.vs.attributes[i].usage =
1825                 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1826         shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1827     }
1828
1829     if (output_signature)
1830     {
1831         for (i = 0; i < output_signature->element_count; ++i)
1832         {
1833             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1834             reg_maps->output_registers |= 1 << e->register_idx;
1835             shader->output_signature[e->register_idx] = *e;
1836         }
1837     }
1838
1839     vertexshader_set_limits(shader);
1840
1841     shader->load_local_constsF = reg_maps->usesrelconstF
1842             && !list_empty(&shader->constantsF);
1843
1844     return WINED3D_OK;
1845 }
1846
1847 HRESULT geometryshader_init(struct wined3d_shader *shader, IWineD3DDeviceImpl *device,
1848         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1849         void *parent, const struct wined3d_parent_ops *parent_ops)
1850 {
1851     HRESULT hr;
1852
1853     shader_init(shader, device, parent, parent_ops);
1854     hr = shader_set_function(shader, byte_code, output_signature, 0);
1855     if (FAILED(hr))
1856     {
1857         WARN("Failed to set function, hr %#x.\n", hr);
1858         shader_cleanup(shader);
1859         return hr;
1860     }
1861
1862     shader->load_local_constsF = FALSE;
1863
1864     return WINED3D_OK;
1865 }
1866
1867 void find_ps_compile_args(const struct wined3d_state *state,
1868         const struct wined3d_shader *shader, struct ps_compile_args *args)
1869 {
1870     IWineD3DDeviceImpl *device = shader->device;
1871     const struct wined3d_texture *texture;
1872     UINT i;
1873
1874     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1875     if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
1876     {
1877         struct wined3d_surface *rt = device->render_targets[0];
1878         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
1879     }
1880
1881     if (shader->reg_maps.shader_version.major == 1
1882             && shader->reg_maps.shader_version.minor <= 3)
1883     {
1884         for (i = 0; i < 4; ++i)
1885         {
1886             DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1887             DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
1888             if (flags & WINED3DTTFF_PROJECTED)
1889                 tex_transform |= WINED3D_PSARGS_PROJECTED;
1890             args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
1891         }
1892     }
1893
1894     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1895     {
1896         if (!shader->reg_maps.sampler_type[i])
1897             continue;
1898
1899         texture = state->textures[i];
1900         if (!texture)
1901         {
1902             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1903             continue;
1904         }
1905         args->color_fixup[i] = texture->resource.format->color_fixup;
1906
1907         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
1908             args->shadow |= 1 << i;
1909
1910         /* Flag samplers that need NP2 texcoord fixup. */
1911         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
1912             args->np2_fixup |= (1 << i);
1913     }
1914     if (shader->reg_maps.shader_version.major >= 3)
1915     {
1916         if (device->strided_streams.position_transformed)
1917         {
1918             args->vp_mode = pretransformed;
1919         }
1920         else if (use_vs(state))
1921         {
1922             args->vp_mode = vertexshader;
1923         }
1924         else
1925         {
1926             args->vp_mode = fixedfunction;
1927         }
1928         args->fog = FOG_OFF;
1929     }
1930     else
1931     {
1932         args->vp_mode = vertexshader;
1933         if (state->render_states[WINED3DRS_FOGENABLE])
1934         {
1935             switch (state->render_states[WINED3DRS_FOGTABLEMODE])
1936             {
1937                 case WINED3DFOG_NONE:
1938                     if (device->strided_streams.position_transformed || use_vs(state))
1939                     {
1940                         args->fog = FOG_LINEAR;
1941                         break;
1942                     }
1943
1944                     switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
1945                     {
1946                         case WINED3DFOG_NONE: /* Fall through. */
1947                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
1948                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
1949                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
1950                     }
1951                     break;
1952
1953                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
1954                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
1955                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
1956             }
1957         }
1958         else
1959         {
1960             args->fog = FOG_OFF;
1961         }
1962     }
1963 }
1964
1965 static void pixelshader_set_limits(struct wined3d_shader *shader)
1966 {
1967     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1968             shader->reg_maps.shader_version.minor);
1969
1970     shader->limits.attributes = 0;
1971     shader->limits.address = 0;
1972     shader->limits.packed_output = 0;
1973
1974     switch (shader_version)
1975     {
1976         case WINED3D_SHADER_VERSION(1, 0):
1977         case WINED3D_SHADER_VERSION(1, 1):
1978         case WINED3D_SHADER_VERSION(1, 2):
1979         case WINED3D_SHADER_VERSION(1, 3):
1980             shader->limits.temporary = 2;
1981             shader->limits.constant_float = 8;
1982             shader->limits.constant_int = 0;
1983             shader->limits.constant_bool = 0;
1984             shader->limits.texcoord = 4;
1985             shader->limits.sampler = 4;
1986             shader->limits.packed_input = 0;
1987             shader->limits.label = 0;
1988             break;
1989
1990         case WINED3D_SHADER_VERSION(1, 4):
1991             shader->limits.temporary = 6;
1992             shader->limits.constant_float = 8;
1993             shader->limits.constant_int = 0;
1994             shader->limits.constant_bool = 0;
1995             shader->limits.texcoord = 6;
1996             shader->limits.sampler = 6;
1997             shader->limits.packed_input = 0;
1998             shader->limits.label = 0;
1999             break;
2000
2001         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2002         case WINED3D_SHADER_VERSION(2, 0):
2003             shader->limits.temporary = 32;
2004             shader->limits.constant_float = 32;
2005             shader->limits.constant_int = 16;
2006             shader->limits.constant_bool = 16;
2007             shader->limits.texcoord = 8;
2008             shader->limits.sampler = 16;
2009             shader->limits.packed_input = 0;
2010             break;
2011
2012         case WINED3D_SHADER_VERSION(2, 1):
2013             shader->limits.temporary = 32;
2014             shader->limits.constant_float = 32;
2015             shader->limits.constant_int = 16;
2016             shader->limits.constant_bool = 16;
2017             shader->limits.texcoord = 8;
2018             shader->limits.sampler = 16;
2019             shader->limits.packed_input = 0;
2020             shader->limits.label = 16;
2021             break;
2022
2023         case WINED3D_SHADER_VERSION(4, 0):
2024             FIXME("Using 3.0 limits for 4.0 shader.\n");
2025             /* Fall through. */
2026
2027         case WINED3D_SHADER_VERSION(3, 0):
2028             shader->limits.temporary = 32;
2029             shader->limits.constant_float = 224;
2030             shader->limits.constant_int = 16;
2031             shader->limits.constant_bool = 16;
2032             shader->limits.texcoord = 0;
2033             shader->limits.sampler = 16;
2034             shader->limits.packed_input = 12;
2035             shader->limits.label = 16; /* FIXME: 2048 */
2036             break;
2037
2038         default:
2039             shader->limits.temporary = 32;
2040             shader->limits.constant_float = 32;
2041             shader->limits.constant_int = 16;
2042             shader->limits.constant_bool = 16;
2043             shader->limits.texcoord = 8;
2044             shader->limits.sampler = 16;
2045             shader->limits.packed_input = 0;
2046             shader->limits.label = 0;
2047             FIXME("Unrecognized pixel shader version %u.%u\n",
2048                     shader->reg_maps.shader_version.major,
2049                     shader->reg_maps.shader_version.minor);
2050     }
2051 }
2052
2053 HRESULT pixelshader_init(struct wined3d_shader *shader, IWineD3DDeviceImpl *device,
2054         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2055         void *parent, const struct wined3d_parent_ops *parent_ops)
2056 {
2057     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2058     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2059     HRESULT hr;
2060
2061     if (!byte_code) return WINED3DERR_INVALIDCALL;
2062
2063     shader_init(shader, device, parent, parent_ops);
2064     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF);
2065     if (FAILED(hr))
2066     {
2067         WARN("Failed to set function, hr %#x.\n", hr);
2068         shader_cleanup(shader);
2069         return hr;
2070     }
2071
2072     pixelshader_set_limits(shader);
2073
2074     for (i = 0; i < MAX_REG_INPUT; ++i)
2075     {
2076         if (shader->u.ps.input_reg_used[i])
2077         {
2078             ++num_regs_used;
2079             highest_reg_used = i;
2080         }
2081     }
2082
2083     /* Don't do any register mapping magic if it is not needed, or if we can't
2084      * achieve anything anyway */
2085     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2086             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2087     {
2088         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2089         {
2090             /* This happens with relative addressing. The input mapper function
2091              * warns about this if the higher registers are declared too, so
2092              * don't write a FIXME here */
2093             WARN("More varying registers used than supported\n");
2094         }
2095
2096         for (i = 0; i < MAX_REG_INPUT; ++i)
2097         {
2098             shader->u.ps.input_reg_map[i] = i;
2099         }
2100
2101         shader->u.ps.declared_in_count = highest_reg_used + 1;
2102     }
2103     else
2104     {
2105         shader->u.ps.declared_in_count = 0;
2106         for (i = 0; i < MAX_REG_INPUT; ++i)
2107         {
2108             if (shader->u.ps.input_reg_used[i])
2109                 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2110             else shader->u.ps.input_reg_map[i] = ~0U;
2111         }
2112     }
2113
2114     shader->load_local_constsF = FALSE;
2115
2116     return WINED3D_OK;
2117 }
2118
2119 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2120 {
2121     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2122     unsigned int i;
2123
2124     if (reg_maps->shader_version.major != 1) return;
2125
2126     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2127     {
2128         /* We don't sample from this sampler. */
2129         if (!sampler_type[i]) continue;
2130
2131         if (!textures[i])
2132         {
2133             WARN("No texture bound to sampler %u, using 2D.\n", i);
2134             sampler_type[i] = WINED3DSTT_2D;
2135             continue;
2136         }
2137
2138         switch (textures[i]->target)
2139         {
2140             case GL_TEXTURE_RECTANGLE_ARB:
2141             case GL_TEXTURE_2D:
2142                 /* We have to select between texture rectangles and 2D
2143                  * textures later because 2.0 and 3.0 shaders only have
2144                  * WINED3DSTT_2D as well. */
2145                 sampler_type[i] = WINED3DSTT_2D;
2146                 break;
2147
2148             case GL_TEXTURE_3D:
2149                 sampler_type[i] = WINED3DSTT_VOLUME;
2150                 break;
2151
2152             case GL_TEXTURE_CUBE_MAP_ARB:
2153                 sampler_type[i] = WINED3DSTT_CUBE;
2154                 break;
2155
2156             default:
2157                 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2158                 sampler_type[i] = WINED3DSTT_2D;
2159         }
2160     }
2161 }