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, struct wined3d_device *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->rt_mask |= (1 << 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             unsigned 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     /* PS before 2.0 don't have explicit color outputs. Instead the value of
 825      * R0 is written to the render target. */
 826     if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 827         reg_maps->rt_mask |= (1 << 0);
 828
 829     shader->functionLength = ((const char *)ptr - (const char *)byte_code);
 830
 831     return WINED3D_OK;
 832 }
 833
 834 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
 835 {
 836     DWORD map = 1 << max;
 837     map |= map - 1;
 838     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 839
 840     return wined3d_log2i(map);
 841 }
 842
 843 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 844         const struct wined3d_shader_version *shader_version)
 845 {
 846     TRACE("dcl");
 847
 848     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 849     {
 850         switch (semantic->sampler_type)
 851         {
 852             case WINED3DSTT_2D: TRACE("_2d"); break;
 853             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 854             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 855             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 856         }
 857     }
 858     else
 859     {
 860         /* Pixel shaders 3.0 don't have usage semantics. */
 861         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 862         else TRACE("_");
 863
 864         switch (semantic->usage)
 865         {
 866             case WINED3DDECLUSAGE_POSITION:
 867                 TRACE("position%u", semantic->usage_idx);
 868                 break;
 869
 870             case WINED3DDECLUSAGE_BLENDINDICES:
 871                 TRACE("blend");
 872                 break;
 873
 874             case WINED3DDECLUSAGE_BLENDWEIGHT:
 875                 TRACE("weight");
 876                 break;
 877
 878             case WINED3DDECLUSAGE_NORMAL:
 879                 TRACE("normal%u", semantic->usage_idx);
 880                 break;
 881
 882             case WINED3DDECLUSAGE_PSIZE:
 883                 TRACE("psize");
 884                 break;
 885
 886             case WINED3DDECLUSAGE_COLOR:
 887                 if (!semantic->usage_idx) TRACE("color");
 888                 else TRACE("specular%u", (semantic->usage_idx - 1));
 889                 break;
 890
 891             case WINED3DDECLUSAGE_TEXCOORD:
 892                 TRACE("texture%u", semantic->usage_idx);
 893                 break;
 894
 895             case WINED3DDECLUSAGE_TANGENT:
 896                 TRACE("tangent");
 897                 break;
 898
 899             case WINED3DDECLUSAGE_BINORMAL:
 900                 TRACE("binormal");
 901                 break;
 902
 903             case WINED3DDECLUSAGE_TESSFACTOR:
 904                 TRACE("tessfactor");
 905                 break;
 906
 907             case WINED3DDECLUSAGE_POSITIONT:
 908                 TRACE("positionT%u", semantic->usage_idx);
 909                 break;
 910
 911             case WINED3DDECLUSAGE_FOG:
 912                 TRACE("fog");
 913                 break;
 914
 915             case WINED3DDECLUSAGE_DEPTH:
 916                 TRACE("depth");
 917                 break;
 918
 919             case WINED3DDECLUSAGE_SAMPLE:
 920                 TRACE("sample");
 921                 break;
 922
 923             default:
 924                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 925         }
 926     }
 927 }
 928
 929 static void shader_dump_register(const struct wined3d_shader_register *reg,
 930         const struct wined3d_shader_version *shader_version)
 931 {
 932     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 933     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 934     UINT offset = reg->idx;
 935
 936     switch (reg->type)
 937     {
 938         case WINED3DSPR_TEMP:
 939             TRACE("r");
 940             break;
 941
 942         case WINED3DSPR_INPUT:
 943             TRACE("v");
 944             break;
 945
 946         case WINED3DSPR_CONST:
 947         case WINED3DSPR_CONST2:
 948         case WINED3DSPR_CONST3:
 949         case WINED3DSPR_CONST4:
 950             TRACE("c");
 951             offset = shader_get_float_offset(reg->type, reg->idx);
 952             break;
 953
 954         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 955             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 956             break;
 957
 958         case WINED3DSPR_RASTOUT:
 959             TRACE("%s", rastout_reg_names[reg->idx]);
 960             break;
 961
 962         case WINED3DSPR_COLOROUT:
 963             TRACE("oC");
 964             break;
 965
 966         case WINED3DSPR_DEPTHOUT:
 967             TRACE("oDepth");
 968             break;
 969
 970         case WINED3DSPR_ATTROUT:
 971             TRACE("oD");
 972             break;
 973
 974         case WINED3DSPR_TEXCRDOUT:
 975             /* Vertex shaders >= 3.0 use general purpose output registers
 976              * (WINED3DSPR_OUTPUT), which can include an address token. */
 977             if (shader_version->major >= 3) TRACE("o");
 978             else TRACE("oT");
 979             break;
 980
 981         case WINED3DSPR_CONSTINT:
 982             TRACE("i");
 983             break;
 984
 985         case WINED3DSPR_CONSTBOOL:
 986             TRACE("b");
 987             break;
 988
 989         case WINED3DSPR_LABEL:
 990             TRACE("l");
 991             break;
 992
 993         case WINED3DSPR_LOOP:
 994             TRACE("aL");
 995             break;
 996
 997         case WINED3DSPR_SAMPLER:
 998             TRACE("s");
 999             break;
1000
1001         case WINED3DSPR_MISCTYPE:
1002             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
1003             else TRACE("%s", misctype_reg_names[reg->idx]);
1004             break;
1005
1006         case WINED3DSPR_PREDICATE:
1007             TRACE("p");
1008             break;
1009
1010         case WINED3DSPR_IMMCONST:
1011             TRACE("l");
1012             break;
1013
1014         case WINED3DSPR_CONSTBUFFER:
1015             TRACE("cb");
1016             break;
1017
1018         case WINED3DSPR_NULL:
1019             TRACE("null");
1020             break;
1021
1022         case WINED3DSPR_RESOURCE:
1023             TRACE("t");
1024             break;
1025
1026         default:
1027             TRACE("unhandled_rtype(%#x)", reg->type);
1028             break;
1029     }
1030
1031     if (reg->type == WINED3DSPR_IMMCONST)
1032     {
1033         TRACE("(");
1034         switch (reg->immconst_type)
1035         {
1036             case WINED3D_IMMCONST_SCALAR:
1037                 TRACE("%.8e", *(const float *)reg->immconst_data);
1038                 break;
1039
1040             case WINED3D_IMMCONST_VEC4:
1041                 TRACE("%.8e, %.8e, %.8e, %.8e",
1042                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1043                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1044                 break;
1045
1046             default:
1047                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1048                 break;
1049         }
1050         TRACE(")");
1051     }
1052     else if (reg->type != WINED3DSPR_RASTOUT
1053             && reg->type != WINED3DSPR_MISCTYPE
1054             && reg->type != WINED3DSPR_NULL)
1055     {
1056         if (reg->array_idx != ~0U)
1057         {
1058             TRACE("%u[%u", offset, reg->array_idx);
1059             if (reg->rel_addr)
1060             {
1061                 TRACE(" + ");
1062                 shader_dump_src_param(reg->rel_addr, shader_version);
1063             }
1064             TRACE("]");
1065         }
1066         else
1067         {
1068             if (reg->rel_addr)
1069             {
1070                 TRACE("[");
1071                 shader_dump_src_param(reg->rel_addr, shader_version);
1072                 TRACE(" + ");
1073             }
1074             TRACE("%u", offset);
1075             if (reg->rel_addr) TRACE("]");
1076         }
1077     }
1078 }
1079
1080 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1081         const struct wined3d_shader_version *shader_version)
1082 {
1083     DWORD write_mask = param->write_mask;
1084
1085     shader_dump_register(&param->reg, shader_version);
1086
1087     if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1088     {
1089         static const char *write_mask_chars = "xyzw";
1090
1091         TRACE(".");
1092         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1093         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1094         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1095         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1096     }
1097 }
1098
1099 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1100         const struct wined3d_shader_version *shader_version)
1101 {
1102     DWORD src_modifier = param->modifiers;
1103     DWORD swizzle = param->swizzle;
1104
1105     if (src_modifier == WINED3DSPSM_NEG
1106             || src_modifier == WINED3DSPSM_BIASNEG
1107             || src_modifier == WINED3DSPSM_SIGNNEG
1108             || src_modifier == WINED3DSPSM_X2NEG
1109             || src_modifier == WINED3DSPSM_ABSNEG)
1110         TRACE("-");
1111     else if (src_modifier == WINED3DSPSM_COMP)
1112         TRACE("1-");
1113     else if (src_modifier == WINED3DSPSM_NOT)
1114         TRACE("!");
1115
1116     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1117         TRACE("abs(");
1118
1119     shader_dump_register(&param->reg, shader_version);
1120
1121     if (src_modifier)
1122     {
1123         switch (src_modifier)
1124         {
1125             case WINED3DSPSM_NONE:    break;
1126             case WINED3DSPSM_NEG:     break;
1127             case WINED3DSPSM_NOT:     break;
1128             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1129             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1130             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1131             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1132             case WINED3DSPSM_COMP:    break;
1133             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1134             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1135             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1136             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1137             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1138             case WINED3DSPSM_ABS:     TRACE(")"); break;
1139             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1140         }
1141     }
1142
1143     if (swizzle != WINED3DSP_NOSWIZZLE)
1144     {
1145         static const char *swizzle_chars = "xyzw";
1146         DWORD swizzle_x = swizzle & 0x03;
1147         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1148         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1149         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1150
1151         if (swizzle_x == swizzle_y
1152                 && swizzle_x == swizzle_z
1153                 && swizzle_x == swizzle_w)
1154         {
1155             TRACE(".%c", swizzle_chars[swizzle_x]);
1156         }
1157         else
1158         {
1159             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1160                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1161         }
1162     }
1163 }
1164
1165 /* Shared code in order to generate the bulk of the shader string.
1166  * NOTE: A description of how to parse tokens can be found on MSDN. */
1167 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1168         const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1169 {
1170     struct wined3d_device *device = shader->device;
1171     const struct wined3d_shader_frontend *fe = shader->frontend;
1172     void *fe_data = shader->frontend_data;
1173     struct wined3d_shader_src_param dst_rel_addr[2];
1174     struct wined3d_shader_src_param src_rel_addr[4];
1175     struct wined3d_shader_dst_param dst_param[2];
1176     struct wined3d_shader_src_param src_param[4];
1177     struct wined3d_shader_version shader_version;
1178     struct wined3d_shader_loop_state loop_state;
1179     struct wined3d_shader_instruction ins;
1180     struct wined3d_shader_tex_mx tex_mx;
1181     struct wined3d_shader_context ctx;
1182     const DWORD *ptr = byte_code;
1183     DWORD i;
1184
1185     /* Initialize current parsing state. */
1186     tex_mx.current_row = 0;
1187     loop_state.current_depth = 0;
1188     loop_state.current_reg = 0;
1189
1190     ctx.shader = shader;
1191     ctx.gl_info = &device->adapter->gl_info;
1192     ctx.reg_maps = reg_maps;
1193     ctx.buffer = buffer;
1194     ctx.tex_mx = &tex_mx;
1195     ctx.loop_state = &loop_state;
1196     ctx.backend_data = backend_ctx;
1197
1198     ins.ctx = &ctx;
1199     ins.dst = dst_param;
1200     ins.src = src_param;
1201
1202     fe->shader_read_header(fe_data, &ptr, &shader_version);
1203
1204     while (!fe->shader_is_end(fe_data, &ptr))
1205     {
1206         const char *comment;
1207         UINT comment_size;
1208         UINT param_size;
1209
1210         /* Skip comment tokens. */
1211         fe->shader_read_comment(&ptr, &comment, &comment_size);
1212         if (comment) continue;
1213
1214         /* Read opcode. */
1215         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1216
1217         /* Unknown opcode and its parameters. */
1218         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1219         {
1220             TRACE("Skipping unrecognized instruction.\n");
1221             ptr += param_size;
1222             continue;
1223         }
1224
1225         /* Nothing to do. */
1226         if (ins.handler_idx == WINED3DSIH_DCL
1227                 || ins.handler_idx == WINED3DSIH_NOP
1228                 || ins.handler_idx == WINED3DSIH_DEF
1229                 || ins.handler_idx == WINED3DSIH_DEFI
1230                 || ins.handler_idx == WINED3DSIH_DEFB
1231                 || ins.handler_idx == WINED3DSIH_PHASE)
1232         {
1233             ptr += param_size;
1234             continue;
1235         }
1236
1237         /* Destination tokens */
1238         for (i = 0; i < ins.dst_count; ++i)
1239         {
1240             fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1241         }
1242
1243         /* Predication token */
1244         if (ins.predicate)
1245         {
1246             FIXME("Predicates not implemented.\n");
1247             ins.predicate = *ptr++;
1248         }
1249
1250         /* Other source tokens */
1251         for (i = 0; i < ins.src_count; ++i)
1252         {
1253             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1254         }
1255
1256         /* Call appropriate function for output target */
1257         device->shader_backend->shader_handle_instruction(&ins);
1258     }
1259 }
1260
1261 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1262 {
1263     DWORD mmask = dst->modifiers;
1264
1265     switch (dst->shift)
1266     {
1267         case 0: break;
1268         case 13: TRACE("_d8"); break;
1269         case 14: TRACE("_d4"); break;
1270         case 15: TRACE("_d2"); break;
1271         case 1: TRACE("_x2"); break;
1272         case 2: TRACE("_x4"); break;
1273         case 3: TRACE("_x8"); break;
1274         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1275     }
1276
1277     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1278     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1279     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1280
1281     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1282     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1283 }
1284
1285 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1286 {
1287     struct wined3d_shader_version shader_version;
1288     const DWORD *ptr = byte_code;
1289     const char *type_prefix;
1290     DWORD i;
1291
1292     TRACE("Parsing %p.\n", byte_code);
1293
1294     fe->shader_read_header(fe_data, &ptr, &shader_version);
1295
1296     switch (shader_version.type)
1297     {
1298         case WINED3D_SHADER_TYPE_VERTEX:
1299             type_prefix = "vs";
1300             break;
1301
1302         case WINED3D_SHADER_TYPE_GEOMETRY:
1303             type_prefix = "gs";
1304             break;
1305
1306         case WINED3D_SHADER_TYPE_PIXEL:
1307             type_prefix = "ps";
1308             break;
1309
1310         default:
1311             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1312             type_prefix = "unknown";
1313             break;
1314     }
1315
1316     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1317
1318     while (!fe->shader_is_end(fe_data, &ptr))
1319     {
1320         struct wined3d_shader_instruction ins;
1321         const char *comment;
1322         UINT comment_size;
1323         UINT param_size;
1324
1325         /* comment */
1326         fe->shader_read_comment(&ptr, &comment, &comment_size);
1327         if (comment)
1328         {
1329             if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1330             {
1331                 const char *end = comment + comment_size;
1332                 const char *ptr = comment + 4;
1333                 const char *line = ptr;
1334
1335                 TRACE("// TEXT\n");
1336                 while (ptr != end)
1337                 {
1338                     if (*ptr == '\n')
1339                     {
1340                         UINT len = ptr - line;
1341                         if (len && *(ptr - 1) == '\r') --len;
1342                         TRACE("// %s\n", debugstr_an(line, len));
1343                         line = ++ptr;
1344                     }
1345                     else ++ptr;
1346                 }
1347                 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1348             }
1349             else TRACE("// %s\n", debugstr_an(comment, comment_size));
1350             continue;
1351         }
1352
1353         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1354         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1355         {
1356             TRACE("Skipping unrecognized instruction.\n");
1357             ptr += param_size;
1358             continue;
1359         }
1360
1361         if (ins.handler_idx == WINED3DSIH_DCL)
1362         {
1363             struct wined3d_shader_semantic semantic;
1364
1365             fe->shader_read_semantic(&ptr, &semantic);
1366
1367             shader_dump_decl_usage(&semantic, &shader_version);
1368             shader_dump_ins_modifiers(&semantic.reg);
1369             TRACE(" ");
1370             shader_dump_dst_param(&semantic.reg, &shader_version);
1371         }
1372         else if (ins.handler_idx == WINED3DSIH_DEF)
1373         {
1374             struct wined3d_shader_src_param rel_addr;
1375             struct wined3d_shader_dst_param dst;
1376
1377             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1378
1379             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1380                     *(const float *)(ptr),
1381                     *(const float *)(ptr + 1),
1382                     *(const float *)(ptr + 2),
1383                     *(const float *)(ptr + 3));
1384             ptr += 4;
1385         }
1386         else if (ins.handler_idx == WINED3DSIH_DEFI)
1387         {
1388             struct wined3d_shader_src_param rel_addr;
1389             struct wined3d_shader_dst_param dst;
1390
1391             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1392
1393             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1394                     *(ptr),
1395                     *(ptr + 1),
1396                     *(ptr + 2),
1397                     *(ptr + 3));
1398             ptr += 4;
1399         }
1400         else if (ins.handler_idx == WINED3DSIH_DEFB)
1401         {
1402             struct wined3d_shader_src_param rel_addr;
1403             struct wined3d_shader_dst_param dst;
1404
1405             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1406
1407             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1408             ++ptr;
1409         }
1410         else
1411         {
1412             struct wined3d_shader_src_param dst_rel_addr[2];
1413             struct wined3d_shader_src_param src_rel_addr;
1414             struct wined3d_shader_dst_param dst_param[2];
1415             struct wined3d_shader_src_param src_param;
1416
1417             for (i = 0; i < ins.dst_count; ++i)
1418             {
1419                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1420             }
1421
1422             /* Print out predication source token first - it follows
1423              * the destination token. */
1424             if (ins.predicate)
1425             {
1426                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1427                 TRACE("(");
1428                 shader_dump_src_param(&src_param, &shader_version);
1429                 TRACE(") ");
1430             }
1431
1432             /* PixWin marks instructions with the coissue flag with a '+' */
1433             if (ins.coissue) TRACE("+");
1434
1435             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1436
1437             if (ins.handler_idx == WINED3DSIH_IFC
1438                     || ins.handler_idx == WINED3DSIH_BREAKC)
1439             {
1440                 switch (ins.flags)
1441                 {
1442                     case COMPARISON_GT: TRACE("_gt"); break;
1443                     case COMPARISON_EQ: TRACE("_eq"); break;
1444                     case COMPARISON_GE: TRACE("_ge"); break;
1445                     case COMPARISON_LT: TRACE("_lt"); break;
1446                     case COMPARISON_NE: TRACE("_ne"); break;
1447                     case COMPARISON_LE: TRACE("_le"); break;
1448                     default: TRACE("_(%u)", ins.flags);
1449                 }
1450             }
1451             else if (ins.handler_idx == WINED3DSIH_TEX
1452                     && shader_version.major >= 2
1453                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1454             {
1455                 TRACE("p");
1456             }
1457
1458             /* We already read the destination tokens, print them. */
1459             for (i = 0; i < ins.dst_count; ++i)
1460             {
1461                 shader_dump_ins_modifiers(&dst_param[i]);
1462                 TRACE(!i ? " " : ", ");
1463                 shader_dump_dst_param(&dst_param[i], &shader_version);
1464             }
1465
1466             /* Other source tokens */
1467             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1468             {
1469                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1470                 TRACE(!i ? " " : ", ");
1471                 shader_dump_src_param(&src_param, &shader_version);
1472             }
1473         }
1474         TRACE("\n");
1475     }
1476 }
1477
1478 static void shader_cleanup(struct wined3d_shader *shader)
1479 {
1480     shader->device->shader_backend->shader_destroy(shader);
1481     HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1482     HeapFree(GetProcessHeap(), 0, shader->function);
1483     shader_delete_constant_list(&shader->constantsF);
1484     shader_delete_constant_list(&shader->constantsB);
1485     shader_delete_constant_list(&shader->constantsI);
1486     list_remove(&shader->shader_list_entry);
1487
1488     if (shader->frontend && shader->frontend_data)
1489         shader->frontend->shader_free(shader->frontend_data);
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(struct wined3d_device *device, UINT start, UINT count) {}
1498 static void shader_none_update_float_pixel_constants(struct wined3d_device *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(struct wined3d_shader *shader) {}
1503 static HRESULT shader_none_alloc(struct wined3d_device *device) {return WINED3D_OK;}
1504 static void shader_none_free(struct wined3d_device *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 struct wined3d_shader_backend_ops none_shader_backend =
1538 {
1539     shader_none_handle_instruction,
1540     shader_none_select,
1541     shader_none_select_depth_blt,
1542     shader_none_deselect_depth_blt,
1543     shader_none_update_float_vertex_constants,
1544     shader_none_update_float_pixel_constants,
1545     shader_none_load_constants,
1546     shader_none_load_np2fixup_constants,
1547     shader_none_destroy,
1548     shader_none_alloc,
1549     shader_none_free,
1550     shader_none_dirty_const,
1551     shader_none_get_caps,
1552     shader_none_color_fixup_supported,
1553 };
1554
1555 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1556         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1557 {
1558     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1559     const struct wined3d_shader_frontend *fe;
1560     HRESULT hr;
1561
1562     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1563             shader, byte_code, output_signature, float_const_count);
1564
1565     fe = shader_select_frontend(*byte_code);
1566     if (!fe)
1567     {
1568         FIXME("Unable to find frontend for shader.\n");
1569         return WINED3DERR_INVALIDCALL;
1570     }
1571     shader->frontend = fe;
1572     shader->frontend_data = fe->shader_init(byte_code, output_signature);
1573     if (!shader->frontend_data)
1574     {
1575         FIXME("Failed to initialize frontend.\n");
1576         return WINED3DERR_INVALIDCALL;
1577     }
1578
1579     /* First pass: trace shader. */
1580     if (TRACE_ON(d3d_shader))
1581         shader_trace_init(fe, shader->frontend_data, byte_code);
1582
1583     /* Initialize immediate constant lists. */
1584     list_init(&shader->constantsF);
1585     list_init(&shader->constantsB);
1586     list_init(&shader->constantsI);
1587
1588     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1589     hr = shader_get_registers_used(shader, fe,
1590             reg_maps, shader->input_signature, shader->output_signature,
1591             byte_code, float_const_count);
1592     if (FAILED(hr)) return hr;
1593
1594     shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1595     if (!shader->function)
1596         return E_OUTOFMEMORY;
1597     memcpy(shader->function, byte_code, shader->functionLength);
1598
1599     return WINED3D_OK;
1600 }
1601
1602 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1603 {
1604     ULONG refcount = InterlockedIncrement(&shader->ref);
1605
1606     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1607
1608     return refcount;
1609 }
1610
1611 /* Do not call while under the GL lock. */
1612 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1613 {
1614     ULONG refcount = InterlockedDecrement(&shader->ref);
1615
1616     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1617
1618     if (!refcount)
1619     {
1620         shader_cleanup(shader);
1621         shader->parent_ops->wined3d_object_destroyed(shader->parent);
1622         HeapFree(GetProcessHeap(), 0, shader);
1623     }
1624
1625     return refcount;
1626 }
1627
1628 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1629 {
1630     TRACE("shader %p.\n", shader);
1631
1632     return shader->parent;
1633 }
1634
1635 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1636         void *byte_code, UINT *byte_code_size)
1637 {
1638     TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1639
1640     if (!byte_code)
1641     {
1642         *byte_code_size = shader->functionLength;
1643         return WINED3D_OK;
1644     }
1645
1646     if (*byte_code_size < shader->functionLength)
1647     {
1648         /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1649          * than the required size we should write the required size and
1650          * return D3DERR_MOREDATA. That's not actually true. */
1651         return WINED3DERR_INVALIDCALL;
1652     }
1653
1654     memcpy(byte_code, shader->function, shader->functionLength);
1655
1656     return WINED3D_OK;
1657 }
1658
1659 /* Set local constants for d3d8 shaders. */
1660 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1661         UINT start_idx, const float *src_data, UINT count)
1662 {
1663     UINT end_idx = start_idx + count;
1664     UINT i;
1665
1666     TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1667
1668     if (end_idx > shader->limits.constant_float)
1669     {
1670         WARN("end_idx %u > float constants limit %u.\n",
1671                 end_idx, shader->limits.constant_float);
1672         end_idx = shader->limits.constant_float;
1673     }
1674
1675     for (i = start_idx; i < end_idx; ++i)
1676     {
1677         struct local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1678         if (!lconst)
1679             return E_OUTOFMEMORY;
1680
1681         lconst->idx = i;
1682         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1683         list_add_head(&shader->constantsF, &lconst->entry);
1684     }
1685
1686     return WINED3D_OK;
1687 }
1688
1689 void find_vs_compile_args(const struct wined3d_state *state,
1690         const struct wined3d_shader *shader, struct vs_compile_args *args)
1691 {
1692     args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1693             == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1694     args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1695             && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1696     args->swizzle_map = shader->device->strided_streams.swizzle_map;
1697 }
1698
1699 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1700 {
1701     if (usage_idx1 != usage_idx2) return FALSE;
1702     if (usage1 == usage2) return TRUE;
1703     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1704     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1705
1706     return FALSE;
1707 }
1708
1709 BOOL vshader_get_input(const struct wined3d_shader *shader,
1710         BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1711 {
1712     WORD map = shader->reg_maps.input_registers;
1713     unsigned int i;
1714
1715     for (i = 0; map; map >>= 1, ++i)
1716     {
1717         if (!(map & 1)) continue;
1718
1719         if (match_usage(shader->u.vs.attributes[i].usage,
1720                 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1721         {
1722             *regnum = i;
1723             return TRUE;
1724         }
1725     }
1726     return FALSE;
1727 }
1728
1729 static void vertexshader_set_limits(struct wined3d_shader *shader)
1730 {
1731     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1732             shader->reg_maps.shader_version.minor);
1733     struct wined3d_device *device = shader->device;
1734
1735     shader->limits.texcoord = 0;
1736     shader->limits.attributes = 16;
1737     shader->limits.packed_input = 0;
1738
1739     switch (shader_version)
1740     {
1741         case WINED3D_SHADER_VERSION(1, 0):
1742         case WINED3D_SHADER_VERSION(1, 1):
1743             shader->limits.temporary = 12;
1744             shader->limits.constant_bool = 0;
1745             shader->limits.constant_int = 0;
1746             shader->limits.address = 1;
1747             shader->limits.packed_output = 12;
1748             shader->limits.sampler = 0;
1749             shader->limits.label = 0;
1750             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1751              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1752              * constants? */
1753             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1754             break;
1755
1756         case WINED3D_SHADER_VERSION(2, 0):
1757         case WINED3D_SHADER_VERSION(2, 1):
1758             shader->limits.temporary = 12;
1759             shader->limits.constant_bool = 16;
1760             shader->limits.constant_int = 16;
1761             shader->limits.address = 1;
1762             shader->limits.packed_output = 12;
1763             shader->limits.sampler = 0;
1764             shader->limits.label = 16;
1765             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1766             break;
1767
1768         case WINED3D_SHADER_VERSION(4, 0):
1769             FIXME("Using 3.0 limits for 4.0 shader.\n");
1770             /* Fall through. */
1771
1772         case WINED3D_SHADER_VERSION(3, 0):
1773             shader->limits.temporary = 32;
1774             shader->limits.constant_bool = 32;
1775             shader->limits.constant_int = 32;
1776             shader->limits.address = 1;
1777             shader->limits.packed_output = 12;
1778             shader->limits.sampler = 4;
1779             shader->limits.label = 16; /* FIXME: 2048 */
1780             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1781              * even though they are capable of supporting much more (GL
1782              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1783              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1784              * shaders to 256. */
1785             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1786             break;
1787
1788         default:
1789             shader->limits.temporary = 12;
1790             shader->limits.constant_bool = 16;
1791             shader->limits.constant_int = 16;
1792             shader->limits.address = 1;
1793             shader->limits.packed_output = 12;
1794             shader->limits.sampler = 0;
1795             shader->limits.label = 16;
1796             shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1797             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1798                     shader->reg_maps.shader_version.major,
1799                     shader->reg_maps.shader_version.minor);
1800     }
1801 }
1802
1803 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1804         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1805         void *parent, const struct wined3d_parent_ops *parent_ops)
1806 {
1807     struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1808     unsigned int i;
1809     HRESULT hr;
1810     WORD map;
1811
1812     if (!byte_code) return WINED3DERR_INVALIDCALL;
1813
1814     shader_init(shader, device, parent, parent_ops);
1815     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF);
1816     if (FAILED(hr))
1817     {
1818         WARN("Failed to set function, hr %#x.\n", hr);
1819         shader_cleanup(shader);
1820         return hr;
1821     }
1822
1823     map = reg_maps->input_registers;
1824     for (i = 0; map; map >>= 1, ++i)
1825     {
1826         if (!(map & 1) || !shader->input_signature[i].semantic_name)
1827             continue;
1828
1829         shader->u.vs.attributes[i].usage =
1830                 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1831         shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1832     }
1833
1834     if (output_signature)
1835     {
1836         for (i = 0; i < output_signature->element_count; ++i)
1837         {
1838             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1839             reg_maps->output_registers |= 1 << e->register_idx;
1840             shader->output_signature[e->register_idx] = *e;
1841         }
1842     }
1843
1844     vertexshader_set_limits(shader);
1845
1846     shader->load_local_constsF = reg_maps->usesrelconstF
1847             && !list_empty(&shader->constantsF);
1848
1849     return WINED3D_OK;
1850 }
1851
1852 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1853         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1854         void *parent, const struct wined3d_parent_ops *parent_ops)
1855 {
1856     HRESULT hr;
1857
1858     shader_init(shader, device, parent, parent_ops);
1859     hr = shader_set_function(shader, byte_code, output_signature, 0);
1860     if (FAILED(hr))
1861     {
1862         WARN("Failed to set function, hr %#x.\n", hr);
1863         shader_cleanup(shader);
1864         return hr;
1865     }
1866
1867     shader->load_local_constsF = FALSE;
1868
1869     return WINED3D_OK;
1870 }
1871
1872 void find_ps_compile_args(const struct wined3d_state *state,
1873         const struct wined3d_shader *shader, struct ps_compile_args *args)
1874 {
1875     struct wined3d_device *device = shader->device;
1876     const struct wined3d_texture *texture;
1877     UINT i;
1878
1879     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1880     if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
1881     {
1882         const struct wined3d_surface *rt = state->fb->render_targets[0];
1883         if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
1884     }
1885
1886     if (shader->reg_maps.shader_version.major == 1
1887             && shader->reg_maps.shader_version.minor <= 3)
1888     {
1889         for (i = 0; i < 4; ++i)
1890         {
1891             DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1892             DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
1893             if (flags & WINED3DTTFF_PROJECTED)
1894                 tex_transform |= WINED3D_PSARGS_PROJECTED;
1895             args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
1896         }
1897     }
1898
1899     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1900     {
1901         if (!shader->reg_maps.sampler_type[i])
1902             continue;
1903
1904         texture = state->textures[i];
1905         if (!texture)
1906         {
1907             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1908             continue;
1909         }
1910         args->color_fixup[i] = texture->resource.format->color_fixup;
1911
1912         if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
1913             args->shadow |= 1 << i;
1914
1915         /* Flag samplers that need NP2 texcoord fixup. */
1916         if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
1917             args->np2_fixup |= (1 << i);
1918     }
1919     if (shader->reg_maps.shader_version.major >= 3)
1920     {
1921         if (device->strided_streams.position_transformed)
1922         {
1923             args->vp_mode = pretransformed;
1924         }
1925         else if (use_vs(state))
1926         {
1927             args->vp_mode = vertexshader;
1928         }
1929         else
1930         {
1931             args->vp_mode = fixedfunction;
1932         }
1933         args->fog = FOG_OFF;
1934     }
1935     else
1936     {
1937         args->vp_mode = vertexshader;
1938         if (state->render_states[WINED3DRS_FOGENABLE])
1939         {
1940             switch (state->render_states[WINED3DRS_FOGTABLEMODE])
1941             {
1942                 case WINED3DFOG_NONE:
1943                     if (device->strided_streams.position_transformed || use_vs(state))
1944                     {
1945                         args->fog = FOG_LINEAR;
1946                         break;
1947                     }
1948
1949                     switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
1950                     {
1951                         case WINED3DFOG_NONE: /* Fall through. */
1952                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
1953                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
1954                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
1955                     }
1956                     break;
1957
1958                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
1959                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
1960                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
1961             }
1962         }
1963         else
1964         {
1965             args->fog = FOG_OFF;
1966         }
1967     }
1968 }
1969
1970 static void pixelshader_set_limits(struct wined3d_shader *shader)
1971 {
1972     DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1973             shader->reg_maps.shader_version.minor);
1974
1975     shader->limits.attributes = 0;
1976     shader->limits.address = 0;
1977     shader->limits.packed_output = 0;
1978
1979     switch (shader_version)
1980     {
1981         case WINED3D_SHADER_VERSION(1, 0):
1982         case WINED3D_SHADER_VERSION(1, 1):
1983         case WINED3D_SHADER_VERSION(1, 2):
1984         case WINED3D_SHADER_VERSION(1, 3):
1985             shader->limits.temporary = 2;
1986             shader->limits.constant_float = 8;
1987             shader->limits.constant_int = 0;
1988             shader->limits.constant_bool = 0;
1989             shader->limits.texcoord = 4;
1990             shader->limits.sampler = 4;
1991             shader->limits.packed_input = 0;
1992             shader->limits.label = 0;
1993             break;
1994
1995         case WINED3D_SHADER_VERSION(1, 4):
1996             shader->limits.temporary = 6;
1997             shader->limits.constant_float = 8;
1998             shader->limits.constant_int = 0;
1999             shader->limits.constant_bool = 0;
2000             shader->limits.texcoord = 6;
2001             shader->limits.sampler = 6;
2002             shader->limits.packed_input = 0;
2003             shader->limits.label = 0;
2004             break;
2005
2006         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2007         case WINED3D_SHADER_VERSION(2, 0):
2008             shader->limits.temporary = 32;
2009             shader->limits.constant_float = 32;
2010             shader->limits.constant_int = 16;
2011             shader->limits.constant_bool = 16;
2012             shader->limits.texcoord = 8;
2013             shader->limits.sampler = 16;
2014             shader->limits.packed_input = 0;
2015             break;
2016
2017         case WINED3D_SHADER_VERSION(2, 1):
2018             shader->limits.temporary = 32;
2019             shader->limits.constant_float = 32;
2020             shader->limits.constant_int = 16;
2021             shader->limits.constant_bool = 16;
2022             shader->limits.texcoord = 8;
2023             shader->limits.sampler = 16;
2024             shader->limits.packed_input = 0;
2025             shader->limits.label = 16;
2026             break;
2027
2028         case WINED3D_SHADER_VERSION(4, 0):
2029             FIXME("Using 3.0 limits for 4.0 shader.\n");
2030             /* Fall through. */
2031
2032         case WINED3D_SHADER_VERSION(3, 0):
2033             shader->limits.temporary = 32;
2034             shader->limits.constant_float = 224;
2035             shader->limits.constant_int = 16;
2036             shader->limits.constant_bool = 16;
2037             shader->limits.texcoord = 0;
2038             shader->limits.sampler = 16;
2039             shader->limits.packed_input = 12;
2040             shader->limits.label = 16; /* FIXME: 2048 */
2041             break;
2042
2043         default:
2044             shader->limits.temporary = 32;
2045             shader->limits.constant_float = 32;
2046             shader->limits.constant_int = 16;
2047             shader->limits.constant_bool = 16;
2048             shader->limits.texcoord = 8;
2049             shader->limits.sampler = 16;
2050             shader->limits.packed_input = 0;
2051             shader->limits.label = 0;
2052             FIXME("Unrecognized pixel shader version %u.%u\n",
2053                     shader->reg_maps.shader_version.major,
2054                     shader->reg_maps.shader_version.minor);
2055     }
2056 }
2057
2058 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2059         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2060         void *parent, const struct wined3d_parent_ops *parent_ops)
2061 {
2062     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2063     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2064     HRESULT hr;
2065
2066     if (!byte_code) return WINED3DERR_INVALIDCALL;
2067
2068     shader_init(shader, device, parent, parent_ops);
2069     hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF);
2070     if (FAILED(hr))
2071     {
2072         WARN("Failed to set function, hr %#x.\n", hr);
2073         shader_cleanup(shader);
2074         return hr;
2075     }
2076
2077     pixelshader_set_limits(shader);
2078
2079     for (i = 0; i < MAX_REG_INPUT; ++i)
2080     {
2081         if (shader->u.ps.input_reg_used[i])
2082         {
2083             ++num_regs_used;
2084             highest_reg_used = i;
2085         }
2086     }
2087
2088     /* Don't do any register mapping magic if it is not needed, or if we can't
2089      * achieve anything anyway */
2090     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2091             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2092     {
2093         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2094         {
2095             /* This happens with relative addressing. The input mapper function
2096              * warns about this if the higher registers are declared too, so
2097              * don't write a FIXME here */
2098             WARN("More varying registers used than supported\n");
2099         }
2100
2101         for (i = 0; i < MAX_REG_INPUT; ++i)
2102         {
2103             shader->u.ps.input_reg_map[i] = i;
2104         }
2105
2106         shader->u.ps.declared_in_count = highest_reg_used + 1;
2107     }
2108     else
2109     {
2110         shader->u.ps.declared_in_count = 0;
2111         for (i = 0; i < MAX_REG_INPUT; ++i)
2112         {
2113             if (shader->u.ps.input_reg_used[i])
2114                 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2115             else shader->u.ps.input_reg_map[i] = ~0U;
2116         }
2117     }
2118
2119     shader->load_local_constsF = FALSE;
2120
2121     return WINED3D_OK;
2122 }
2123
2124 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2125 {
2126     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2127     unsigned int i;
2128
2129     if (reg_maps->shader_version.major != 1) return;
2130
2131     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2132     {
2133         /* We don't sample from this sampler. */
2134         if (!sampler_type[i]) continue;
2135
2136         if (!textures[i])
2137         {
2138             WARN("No texture bound to sampler %u, using 2D.\n", i);
2139             sampler_type[i] = WINED3DSTT_2D;
2140             continue;
2141         }
2142
2143         switch (textures[i]->target)
2144         {
2145             case GL_TEXTURE_RECTANGLE_ARB:
2146             case GL_TEXTURE_2D:
2147                 /* We have to select between texture rectangles and 2D
2148                  * textures later because 2.0 and 3.0 shaders only have
2149                  * WINED3DSTT_2D as well. */
2150                 sampler_type[i] = WINED3DSTT_2D;
2151                 break;
2152
2153             case GL_TEXTURE_3D:
2154                 sampler_type[i] = WINED3DSTT_VOLUME;
2155                 break;
2156
2157             case GL_TEXTURE_CUBE_MAP_ARB:
2158                 sampler_type[i] = WINED3DSTT_CUBE;
2159                 break;
2160
2161             default:
2162                 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2163                 sampler_type[i] = WINED3DSTT_2D;
2164         }
2165     }
2166 }
2167
2168 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2169         const struct wined3d_shader_signature *output_signature, void *parent,
2170         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2171 {
2172     struct wined3d_shader *object;
2173     HRESULT hr;
2174
2175     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2176             device, byte_code, output_signature, parent, parent_ops, shader);
2177
2178     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2179     if (!object)
2180     {
2181         ERR("Failed to allocate shader memory.\n");
2182         return E_OUTOFMEMORY;
2183     }
2184
2185     hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops);
2186     if (FAILED(hr))
2187     {
2188         WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2189         HeapFree(GetProcessHeap(), 0, object);
2190         return hr;
2191     }
2192
2193     TRACE("Created geometry shader %p.\n", object);
2194     *shader = object;
2195
2196     return WINED3D_OK;
2197 }
2198
2199 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2200         const struct wined3d_shader_signature *output_signature, void *parent,
2201         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2202 {
2203     struct wined3d_shader *object;
2204     HRESULT hr;
2205
2206     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2207             device, byte_code, output_signature, parent, parent_ops, shader);
2208
2209     if (device->ps_selected_mode == SHADER_NONE)
2210         return WINED3DERR_INVALIDCALL;
2211
2212     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2213     if (!object)
2214     {
2215         ERR("Failed to allocate shader memory.\n");
2216         return E_OUTOFMEMORY;
2217     }
2218
2219     hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops);
2220     if (FAILED(hr))
2221     {
2222         WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2223         HeapFree(GetProcessHeap(), 0, object);
2224         return hr;
2225     }
2226
2227     TRACE("Created pixel shader %p.\n", object);
2228     *shader = object;
2229
2230     return WINED3D_OK;
2231 }
2232
2233 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2234         const struct wined3d_shader_signature *output_signature, void *parent,
2235         const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2236 {
2237     struct wined3d_shader *object;
2238     HRESULT hr;
2239
2240     TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2241             device, byte_code, output_signature, parent, parent_ops, shader);
2242
2243     if (device->vs_selected_mode == SHADER_NONE)
2244         return WINED3DERR_INVALIDCALL;
2245
2246     object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2247     if (!object)
2248     {
2249         ERR("Failed to allocate shader memory.\n");
2250         return E_OUTOFMEMORY;
2251     }
2252
2253     hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops);
2254     if (FAILED(hr))
2255     {
2256         WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2257         HeapFree(GetProcessHeap(), 0, object);
2258         return hr;
2259     }
2260
2261     TRACE("Created vertex shader %p.\n", object);
2262     *shader = object;
2263
2264     return WINED3D_OK;
2265 }