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 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 *shader_opcode_names[] =
  37 {
  38     /* WINED3DSIH_ABS           */ "abs",
  39     /* WINED3DSIH_ADD           */ "add",
  40     /* WINED3DSIH_BEM           */ "bem",
  41     /* WINED3DSIH_BREAK         */ "break",
  42     /* WINED3DSIH_BREAKC        */ "breakc",
  43     /* WINED3DSIH_BREAKP        */ "breakp",
  44     /* WINED3DSIH_CALL          */ "call",
  45     /* WINED3DSIH_CALLNZ        */ "callnz",
  46     /* WINED3DSIH_CMP           */ "cmp",
  47     /* WINED3DSIH_CND           */ "cnd",
  48     /* WINED3DSIH_CRS           */ "crs",
  49     /* WINED3DSIH_CUT           */ "cut",
  50     /* WINED3DSIH_DCL           */ "dcl",
  51     /* WINED3DSIH_DEF           */ "def",
  52     /* WINED3DSIH_DEFB          */ "defb",
  53     /* WINED3DSIH_DEFI          */ "defi",
  54     /* WINED3DSIH_DP2ADD        */ "dp2add",
  55     /* WINED3DSIH_DP3           */ "dp3",
  56     /* WINED3DSIH_DP4           */ "dp4",
  57     /* WINED3DSIH_DST           */ "dst",
  58     /* WINED3DSIH_DSX           */ "dsx",
  59     /* WINED3DSIH_DSY           */ "dsy",
  60     /* WINED3DSIH_ELSE          */ "else",
  61     /* WINED3DSIH_EMIT          */ "emit",
  62     /* WINED3DSIH_ENDIF         */ "endif",
  63     /* WINED3DSIH_ENDLOOP       */ "endloop",
  64     /* WINED3DSIH_ENDREP        */ "endrep",
  65     /* WINED3DSIH_EXP           */ "exp",
  66     /* WINED3DSIH_EXPP          */ "expp",
  67     /* WINED3DSIH_FRC           */ "frc",
  68     /* WINED3DSIH_IADD          */ "iadd",
  69     /* WINED3DSIH_IF            */ "if",
  70     /* WINED3DSIH_IFC           */ "ifc",
  71     /* WINED3DSIH_IGE           */ "ige",
  72     /* WINED3DSIH_LABEL         */ "label",
  73     /* WINED3DSIH_LIT           */ "lit",
  74     /* WINED3DSIH_LOG           */ "log",
  75     /* WINED3DSIH_LOGP          */ "logp",
  76     /* WINED3DSIH_LOOP          */ "loop",
  77     /* WINED3DSIH_LRP           */ "lrp",
  78     /* WINED3DSIH_LT            */ "lt",
  79     /* WINED3DSIH_M3x2          */ "m3x2",
  80     /* WINED3DSIH_M3x3          */ "m3x3",
  81     /* WINED3DSIH_M3x4          */ "m3x4",
  82     /* WINED3DSIH_M4x3          */ "m4x3",
  83     /* WINED3DSIH_M4x4          */ "m4x4",
  84     /* WINED3DSIH_MAD           */ "mad",
  85     /* WINED3DSIH_MAX           */ "max",
  86     /* WINED3DSIH_MIN           */ "min",
  87     /* WINED3DSIH_MOV           */ "mov",
  88     /* WINED3DSIH_MOVA          */ "mova",
  89     /* WINED3DSIH_MUL           */ "mul",
  90     /* WINED3DSIH_NOP           */ "nop",
  91     /* WINED3DSIH_NRM           */ "nrm",
  92     /* WINED3DSIH_PHASE         */ "phase",
  93     /* WINED3DSIH_POW           */ "pow",
  94     /* WINED3DSIH_RCP           */ "rcp",
  95     /* WINED3DSIH_REP           */ "rep",
  96     /* WINED3DSIH_RET           */ "ret",
  97     /* WINED3DSIH_RSQ           */ "rsq",
  98     /* WINED3DSIH_SETP          */ "setp",
  99     /* WINED3DSIH_SGE           */ "sge",
 100     /* WINED3DSIH_SGN           */ "sgn",
 101     /* WINED3DSIH_SINCOS        */ "sincos",
 102     /* WINED3DSIH_SLT           */ "slt",
 103     /* WINED3DSIH_SUB           */ "sub",
 104     /* WINED3DSIH_TEX           */ "texld",
 105     /* WINED3DSIH_TEXBEM        */ "texbem",
 106     /* WINED3DSIH_TEXBEML       */ "texbeml",
 107     /* WINED3DSIH_TEXCOORD      */ "texcrd",
 108     /* WINED3DSIH_TEXDEPTH      */ "texdepth",
 109     /* WINED3DSIH_TEXDP3        */ "texdp3",
 110     /* WINED3DSIH_TEXDP3TEX     */ "texdp3tex",
 111     /* WINED3DSIH_TEXKILL       */ "texkill",
 112     /* WINED3DSIH_TEXLDD        */ "texldd",
 113     /* WINED3DSIH_TEXLDL        */ "texldl",
 114     /* WINED3DSIH_TEXM3x2DEPTH  */ "texm3x2depth",
 115     /* WINED3DSIH_TEXM3x2PAD    */ "texm3x2pad",
 116     /* WINED3DSIH_TEXM3x2TEX    */ "texm3x2tex",
 117     /* WINED3DSIH_TEXM3x3       */ "texm3x3",
 118     /* WINED3DSIH_TEXM3x3DIFF   */ "texm3x3diff",
 119     /* WINED3DSIH_TEXM3x3PAD    */ "texm3x3pad",
 120     /* WINED3DSIH_TEXM3x3SPEC   */ "texm3x3spec",
 121     /* WINED3DSIH_TEXM3x3TEX    */ "texm3x3tex",
 122     /* WINED3DSIH_TEXM3x3VSPEC  */ "texm3x3vspec",
 123     /* WINED3DSIH_TEXREG2AR     */ "texreg2ar",
 124     /* WINED3DSIH_TEXREG2GB     */ "texreg2gb",
 125     /* WINED3DSIH_TEXREG2RGB    */ "texreg2rgb",
 126 };
 127
 128 static const char *semantic_names[] =
 129 {
 130     /* WINED3DDECLUSAGE_POSITION        */ "SV_POSITION",
 131     /* WINED3DDECLUSAGE_BLENDWEIGHT     */ "BLENDWEIGHT",
 132     /* WINED3DDECLUSAGE_BLENDINDICES    */ "BLENDINDICES",
 133     /* WINED3DDECLUSAGE_NORMAL          */ "NORMAL",
 134     /* WINED3DDECLUSAGE_PSIZE           */ "PSIZE",
 135     /* WINED3DDECLUSAGE_TEXCOORD        */ "TEXCOORD",
 136     /* WINED3DDECLUSAGE_TANGENT         */ "TANGENT",
 137     /* WINED3DDECLUSAGE_BINORMAL        */ "BINORMAL",
 138     /* WINED3DDECLUSAGE_TESSFACTOR      */ "TESSFACTOR",
 139     /* WINED3DDECLUSAGE_POSITIONT       */ "POSITIONT",
 140     /* WINED3DDECLUSAGE_COLOR           */ "COLOR",
 141     /* WINED3DDECLUSAGE_FOG             */ "FOG",
 142     /* WINED3DDECLUSAGE_DEPTH           */ "DEPTH",
 143     /* WINED3DDECLUSAGE_SAMPLE          */ "SAMPLE",
 144 };
 145
 146 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
 147 {
 148     if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
 149     {
 150         FIXME("Unrecognized usage %#x.\n", usage);
 151         return "UNRECOGNIZED";
 152     }
 153
 154     return semantic_names[usage];
 155 }
 156
 157 static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
 158 {
 159     unsigned int i;
 160
 161     for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
 162     {
 163         if (!strcmp(name, semantic_names[i])) return i;
 164     }
 165
 166     return ~0U;
 167 }
 168
 169 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
 170 {
 171     return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
 172 }
 173
 174 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
 175         const struct wined3d_shader_semantic *s)
 176 {
 177     e->semantic_name = shader_semantic_name_from_usage(s->usage);
 178     e->semantic_idx = s->usage_idx;
 179     e->sysval_semantic = 0;
 180     e->component_type = 0;
 181     e->register_idx = s->reg.reg.idx;
 182     e->mask = s->reg.write_mask;
 183 }
 184
 185 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
 186 {
 187     switch (version_token >> 16)
 188     {
 189         case WINED3D_SM1_VS:
 190         case WINED3D_SM1_PS:
 191             return &sm1_shader_frontend;
 192
 193         case WINED3D_SM4_PS:
 194         case WINED3D_SM4_VS:
 195         case WINED3D_SM4_GS:
 196             return &sm4_shader_frontend;
 197
 198         default:
 199             FIXME("Unrecognised version token %#x\n", version_token);
 200             return NULL;
 201     }
 202 }
 203
 204 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
 205 {
 206     buffer->buffer[0] = '\0';
 207     buffer->bsize = 0;
 208     buffer->lineNo = 0;
 209     buffer->newline = TRUE;
 210 }
 211
 212 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
 213 {
 214     buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
 215     if (!buffer->buffer)
 216     {
 217         ERR("Failed to allocate shader buffer memory.\n");
 218         return FALSE;
 219     }
 220
 221     shader_buffer_clear(buffer);
 222     return TRUE;
 223 }
 224
 225 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
 226 {
 227     HeapFree(GetProcessHeap(), 0, buffer->buffer);
 228 }
 229
 230 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
 231 {
 232     char *base = buffer->buffer + buffer->bsize;
 233     int rc;
 234
 235     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
 236
 237     if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
 238     {
 239         ERR("The buffer allocated for the shader program string "
 240             "is too small at %d bytes.\n", SHADER_PGMSIZE);
 241         buffer->bsize = SHADER_PGMSIZE - 1;
 242         return -1;
 243     }
 244
 245     if (buffer->newline)
 246     {
 247         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
 248         buffer->newline = FALSE;
 249     }
 250     else
 251     {
 252         TRACE("%s", base);
 253     }
 254
 255     buffer->bsize += rc;
 256     if (buffer->buffer[buffer->bsize-1] == '\n')
 257     {
 258         ++buffer->lineNo;
 259         buffer->newline = TRUE;
 260     }
 261
 262     return 0;
 263 }
 264
 265 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
 266 {
 267     va_list args;
 268     int ret;
 269
 270     va_start(args, format);
 271     ret = shader_vaddline(buffer, format, args);
 272     va_end(args);
 273
 274     return ret;
 275 }
 276
 277 static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
 278         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
 279 {
 280     shader->ref = 1;
 281     shader->device = (IWineD3DDevice *)device;
 282     shader->parent = parent;
 283     shader->parent_ops = parent_ops;
 284     list_init(&shader->linked_programs);
 285     list_add_head(&device->shaders, &shader->shader_list_entry);
 286 }
 287
 288 /* Convert floating point offset relative to a register file to an absolute
 289  * offset for float constants. */
 290 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
 291 {
 292     switch (register_type)
 293     {
 294         case WINED3DSPR_CONST: return register_idx;
 295         case WINED3DSPR_CONST2: return 2048 + register_idx;
 296         case WINED3DSPR_CONST3: return 4096 + register_idx;
 297         case WINED3DSPR_CONST4: return 6144 + register_idx;
 298         default:
 299             FIXME("Unsupported register type: %u.\n", register_type);
 300             return register_idx;
 301     }
 302 }
 303
 304 static void shader_delete_constant_list(struct list *clist)
 305 {
 306     struct local_constant *constant;
 307     struct list *ptr;
 308
 309     ptr = list_head(clist);
 310     while (ptr)
 311     {
 312         constant = LIST_ENTRY(ptr, struct local_constant, entry);
 313         ptr = list_next(clist, ptr);
 314         HeapFree(GetProcessHeap(), 0, constant);
 315     }
 316     list_init(clist);
 317 }
 318
 319 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
 320 {
 321     DWORD idx, shift;
 322     idx = bit >> 5;
 323     shift = bit & 0x1f;
 324     bitmap[idx] |= (1 << shift);
 325 }
 326
 327 static void shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct shader_reg_maps *reg_maps,
 328         const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
 329 {
 330     switch (reg->type)
 331     {
 332         case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
 333             if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
 334             else reg_maps->address |= 1 << reg->idx;
 335             break;
 336
 337         case WINED3DSPR_TEMP:
 338             reg_maps->temporary |= 1 << reg->idx;
 339             break;
 340
 341         case WINED3DSPR_INPUT:
 342             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 343             {
 344                 if (reg->rel_addr)
 345                 {
 346                     /* If relative addressing is used, we must assume that all registers
 347                      * are used. Even if it is a construct like v3[aL], we can't assume
 348                      * that v0, v1 and v2 aren't read because aL can be negative */
 349                     unsigned int i;
 350                     for (i = 0; i < MAX_REG_INPUT; ++i)
 351                     {
 352                         ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
 353                     }
 354                 }
 355                 else
 356                 {
 357                     ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg->idx] = TRUE;
 358                 }
 359             }
 360             else reg_maps->input_registers |= 1 << reg->idx;
 361             break;
 362
 363         case WINED3DSPR_RASTOUT:
 364             if (reg->idx == 1) reg_maps->fog = 1;
 365             break;
 366
 367         case WINED3DSPR_MISCTYPE:
 368             if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
 369             {
 370                 if (reg->idx == 0) reg_maps->vpos = 1;
 371                 else if (reg->idx == 1) reg_maps->usesfacing = 1;
 372             }
 373             break;
 374
 375         case WINED3DSPR_CONST:
 376             if (reg->rel_addr)
 377             {
 378                 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
 379                 {
 380                     if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
 381                     {
 382                         ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
 383                     }
 384                     if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
 385                     {
 386                         ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
 387                     }
 388                 }
 389                 reg_maps->usesrelconstF = TRUE;
 390             }
 391             else
 392             {
 393                 set_bitmap_bit(reg_maps->constf, reg->idx);
 394             }
 395             break;
 396
 397         case WINED3DSPR_CONSTINT:
 398             reg_maps->integer_constants |= (1 << reg->idx);
 399             break;
 400
 401         case WINED3DSPR_CONSTBOOL:
 402             reg_maps->boolean_constants |= (1 << reg->idx);
 403             break;
 404
 405         case WINED3DSPR_COLOROUT:
 406             reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
 407             break;
 408
 409         default:
 410             TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
 411             break;
 412     }
 413 }
 414
 415 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
 416 {
 417     switch (instr)
 418     {
 419         case WINED3DSIH_M4x4:
 420         case WINED3DSIH_M3x4:
 421             return param == 1 ? 3 : 0;
 422
 423         case WINED3DSIH_M4x3:
 424         case WINED3DSIH_M3x3:
 425             return param == 1 ? 2 : 0;
 426
 427         case WINED3DSIH_M3x2:
 428             return param == 1 ? 1 : 0;
 429
 430         default:
 431             return 0;
 432     }
 433 }
 434
 435 /* Note that this does not count the loop register as an address register. */
 436 static HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
 437         struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
 438         struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
 439 {
 440     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
 441     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
 442     void *fe_data = shader->baseShader.frontend_data;
 443     struct wined3d_shader_version shader_version;
 444     const DWORD *ptr = byte_code;
 445
 446     memset(reg_maps, 0, sizeof(*reg_maps));
 447
 448     /* get_registers_used() is called on every compile on some 1.x shaders,
 449      * which can result in stacking up a collection of local constants.
 450      * Delete the old constants if existing. */
 451     shader_delete_constant_list(&shader->baseShader.constantsF);
 452     shader_delete_constant_list(&shader->baseShader.constantsB);
 453     shader_delete_constant_list(&shader->baseShader.constantsI);
 454
 455     fe->shader_read_header(fe_data, &ptr, &shader_version);
 456     reg_maps->shader_version = shader_version;
 457
 458     reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
 459             sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
 460     if (!reg_maps->constf)
 461     {
 462         ERR("Failed to allocate constant map memory.\n");
 463         return E_OUTOFMEMORY;
 464     }
 465
 466     while (!fe->shader_is_end(fe_data, &ptr))
 467     {
 468         struct wined3d_shader_instruction ins;
 469         const char *comment;
 470         UINT comment_size;
 471         UINT param_size;
 472
 473         /* Skip comments. */
 474         fe->shader_read_comment(&ptr, &comment, &comment_size);
 475         if (comment) continue;
 476
 477         /* Fetch opcode. */
 478         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
 479
 480         /* Unhandled opcode, and its parameters. */
 481         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 482         {
 483             TRACE("Skipping unrecognized instruction.\n");
 484             ptr += param_size;
 485             continue;
 486         }
 487
 488         /* Handle declarations. */
 489         if (ins.handler_idx == WINED3DSIH_DCL)
 490         {
 491             struct wined3d_shader_semantic semantic;
 492
 493             fe->shader_read_semantic(&ptr, &semantic);
 494
 495             switch (semantic.reg.reg.type)
 496             {
 497                 /* Mark input registers used. */
 498                 case WINED3DSPR_INPUT:
 499                     reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
 500                     shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
 501                     break;
 502
 503                 /* Vertex shader: mark 3.0 output registers used, save token. */
 504                 case WINED3DSPR_OUTPUT:
 505                     reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
 506                     shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
 507                     if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
 508                     break;
 509
 510                 /* Save sampler usage token. */
 511                 case WINED3DSPR_SAMPLER:
 512                     reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
 513                     break;
 514
 515                 default:
 516                     TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
 517                     break;
 518             }
 519         }
 520         else if (ins.handler_idx == WINED3DSIH_DEF)
 521         {
 522             struct wined3d_shader_src_param rel_addr;
 523             struct wined3d_shader_dst_param dst;
 524
 525             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 526             if (!lconst) return E_OUTOFMEMORY;
 527
 528             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 529             lconst->idx = dst.reg.idx;
 530
 531             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 532             ptr += 4;
 533
 534             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 535             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 536             {
 537                 float *value = (float *)lconst->value;
 538                 if (value[0] < -1.0f) value[0] = -1.0f;
 539                 else if (value[0] > 1.0f) value[0] = 1.0f;
 540                 if (value[1] < -1.0f) value[1] = -1.0f;
 541                 else if (value[1] > 1.0f) value[1] = 1.0f;
 542                 if (value[2] < -1.0f) value[2] = -1.0f;
 543                 else if (value[2] > 1.0f) value[2] = 1.0f;
 544                 if (value[3] < -1.0f) value[3] = -1.0f;
 545                 else if (value[3] > 1.0f) value[3] = 1.0f;
 546             }
 547
 548             list_add_head(&shader->baseShader.constantsF, &lconst->entry);
 549         }
 550         else if (ins.handler_idx == WINED3DSIH_DEFI)
 551         {
 552             struct wined3d_shader_src_param rel_addr;
 553             struct wined3d_shader_dst_param dst;
 554
 555             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 556             if (!lconst) return E_OUTOFMEMORY;
 557
 558             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 559             lconst->idx = dst.reg.idx;
 560
 561             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 562             ptr += 4;
 563
 564             list_add_head(&shader->baseShader.constantsI, &lconst->entry);
 565             reg_maps->local_int_consts |= (1 << dst.reg.idx);
 566         }
 567         else if (ins.handler_idx == WINED3DSIH_DEFB)
 568         {
 569             struct wined3d_shader_src_param rel_addr;
 570             struct wined3d_shader_dst_param dst;
 571
 572             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 573             if (!lconst) return E_OUTOFMEMORY;
 574
 575             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 576             lconst->idx = dst.reg.idx;
 577
 578             memcpy(lconst->value, ptr, sizeof(DWORD));
 579             ++ptr;
 580
 581             list_add_head(&shader->baseShader.constantsB, &lconst->entry);
 582             reg_maps->local_bool_consts |= (1 << dst.reg.idx);
 583         }
 584         /* If there's a loop in the shader. */
 585         else if (ins.handler_idx == WINED3DSIH_LOOP
 586                 || ins.handler_idx == WINED3DSIH_REP)
 587         {
 588             struct wined3d_shader_src_param src, rel_addr;
 589
 590             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 591
 592             /* Rep and Loop always use an integer constant for the control parameters. */
 593             if (ins.handler_idx == WINED3DSIH_REP)
 594             {
 595                 reg_maps->integer_constants |= 1 << src.reg.idx;
 596             }
 597             else
 598             {
 599                 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 600                 reg_maps->integer_constants |= 1 << src.reg.idx;
 601             }
 602
 603             cur_loop_depth++;
 604             if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
 605         }
 606         else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 607                 || ins.handler_idx == WINED3DSIH_ENDREP)
 608         {
 609             cur_loop_depth--;
 610         }
 611         /* For subroutine prototypes. */
 612         else if (ins.handler_idx == WINED3DSIH_LABEL)
 613         {
 614             struct wined3d_shader_src_param src, rel_addr;
 615
 616             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 617             reg_maps->labels |= 1 << src.reg.idx;
 618         }
 619         /* Set texture, address, temporary registers. */
 620         else
 621         {
 622             BOOL color0_mov = FALSE;
 623             int i, limit;
 624
 625             /* This will loop over all the registers and try to
 626              * make a bitmask of the ones we're interested in.
 627              *
 628              * Relative addressing tokens are ignored, but that's
 629              * okay, since we'll catch any address registers when
 630              * they are initialized (required by spec). */
 631             if (ins.dst_count)
 632             {
 633                 struct wined3d_shader_src_param dst_rel_addr;
 634                 struct wined3d_shader_dst_param dst_param;
 635
 636                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
 637
 638                 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
 639
 640                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 641                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 642                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 643                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
 644                         && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
 645                 {
 646                     reg_maps->texcoord_mask[dst_param.reg.idx] |= dst_param.write_mask;
 647                 }
 648
 649                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 650                 {
 651                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 652
 653                     if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
 654                     {
 655                     /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 656                      * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 657                      * the mov and perform the sRGB write correction from the source register.
 658                      *
 659                      * However, if the mov is only partial, we can't do this, and if the write
 660                      * comes from an instruction other than MOV it is hard to do as well. If
 661                      * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 662
 663                         ps->color0_mov = FALSE;
 664                         if (ins.handler_idx == WINED3DSIH_MOV)
 665                         {
 666                             /* Used later when the source register is read. */
 667                             color0_mov = TRUE;
 668                         }
 669                     }
 670                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 671                      * end
 672                      */
 673                     else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
 674                     {
 675                         ps->color0_mov = FALSE;
 676                     }
 677                 }
 678
 679                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 680                 if (shader_version.major == 1
 681                         && (ins.handler_idx == WINED3DSIH_TEX
 682                             || ins.handler_idx == WINED3DSIH_TEXBEM
 683                             || ins.handler_idx == WINED3DSIH_TEXBEML
 684                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 685                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 686                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 687                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 688                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 689                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 690                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 691                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 692                 {
 693                     /* Fake sampler usage, only set reserved bit and type. */
 694                     DWORD sampler_code = dst_param.reg.idx;
 695
 696                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 697                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 698
 699                     /* texbem is only valid with < 1.4 pixel shaders */
 700                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 701                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 702                     {
 703                         reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 704                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 705                         {
 706                             reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
 707                         }
 708                     }
 709                 }
 710                 else if (ins.handler_idx == WINED3DSIH_BEM)
 711                 {
 712                     reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 713                 }
 714             }
 715
 716             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 717             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 718             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 719             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 720             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 721             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 722             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 723             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 724
 725             limit = ins.src_count + (ins.predicate ? 1 : 0);
 726             for (i = 0; i < limit; ++i)
 727             {
 728                 struct wined3d_shader_src_param src_param, src_rel_addr;
 729                 unsigned int count;
 730
 731                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
 732                 count = get_instr_extra_regcount(ins.handler_idx, i);
 733
 734                 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 735                 while (count)
 736                 {
 737                     ++src_param.reg.idx;
 738                     shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 739                     --count;
 740                 }
 741
 742                 if (color0_mov)
 743                 {
 744                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 745                     if (src_param.reg.type == WINED3DSPR_TEMP
 746                             && src_param.swizzle == WINED3DSP_NOSWIZZLE)
 747                     {
 748                         ps->color0_mov = TRUE;
 749                         ps->color0_reg = src_param.reg.idx;
 750                     }
 751                 }
 752             }
 753         }
 754     }
 755     reg_maps->loop_depth = max_loop_depth;
 756
 757     shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
 758
 759     return WINED3D_OK;
 760 }
 761
 762 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
 763 {
 764     DWORD map = 1 << max;
 765     map |= map - 1;
 766     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 767
 768     return wined3d_log2i(map);
 769 }
 770
 771 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 772         const struct wined3d_shader_version *shader_version)
 773 {
 774     TRACE("dcl");
 775
 776     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 777     {
 778         switch (semantic->sampler_type)
 779         {
 780             case WINED3DSTT_2D: TRACE("_2d"); break;
 781             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 782             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 783             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 784         }
 785     }
 786     else
 787     {
 788         /* Pixel shaders 3.0 don't have usage semantics. */
 789         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 790         else TRACE("_");
 791
 792         switch (semantic->usage)
 793         {
 794             case WINED3DDECLUSAGE_POSITION:
 795                 TRACE("position%u", semantic->usage_idx);
 796                 break;
 797
 798             case WINED3DDECLUSAGE_BLENDINDICES:
 799                 TRACE("blend");
 800                 break;
 801
 802             case WINED3DDECLUSAGE_BLENDWEIGHT:
 803                 TRACE("weight");
 804                 break;
 805
 806             case WINED3DDECLUSAGE_NORMAL:
 807                 TRACE("normal%u", semantic->usage_idx);
 808                 break;
 809
 810             case WINED3DDECLUSAGE_PSIZE:
 811                 TRACE("psize");
 812                 break;
 813
 814             case WINED3DDECLUSAGE_COLOR:
 815                 if (semantic->usage_idx == 0) TRACE("color");
 816                 else TRACE("specular%u", (semantic->usage_idx - 1));
 817                 break;
 818
 819             case WINED3DDECLUSAGE_TEXCOORD:
 820                 TRACE("texture%u", semantic->usage_idx);
 821                 break;
 822
 823             case WINED3DDECLUSAGE_TANGENT:
 824                 TRACE("tangent");
 825                 break;
 826
 827             case WINED3DDECLUSAGE_BINORMAL:
 828                 TRACE("binormal");
 829                 break;
 830
 831             case WINED3DDECLUSAGE_TESSFACTOR:
 832                 TRACE("tessfactor");
 833                 break;
 834
 835             case WINED3DDECLUSAGE_POSITIONT:
 836                 TRACE("positionT%u", semantic->usage_idx);
 837                 break;
 838
 839             case WINED3DDECLUSAGE_FOG:
 840                 TRACE("fog");
 841                 break;
 842
 843             case WINED3DDECLUSAGE_DEPTH:
 844                 TRACE("depth");
 845                 break;
 846
 847             case WINED3DDECLUSAGE_SAMPLE:
 848                 TRACE("sample");
 849                 break;
 850
 851             default:
 852                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 853         }
 854     }
 855 }
 856
 857 static void shader_dump_register(const struct wined3d_shader_register *reg,
 858         const struct wined3d_shader_version *shader_version)
 859 {
 860     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 861     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 862     UINT offset = reg->idx;
 863
 864     switch (reg->type)
 865     {
 866         case WINED3DSPR_TEMP:
 867             TRACE("r");
 868             break;
 869
 870         case WINED3DSPR_INPUT:
 871             TRACE("v");
 872             break;
 873
 874         case WINED3DSPR_CONST:
 875         case WINED3DSPR_CONST2:
 876         case WINED3DSPR_CONST3:
 877         case WINED3DSPR_CONST4:
 878             TRACE("c");
 879             offset = shader_get_float_offset(reg->type, reg->idx);
 880             break;
 881
 882         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 883             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 884             break;
 885
 886         case WINED3DSPR_RASTOUT:
 887             TRACE("%s", rastout_reg_names[reg->idx]);
 888             break;
 889
 890         case WINED3DSPR_COLOROUT:
 891             TRACE("oC");
 892             break;
 893
 894         case WINED3DSPR_DEPTHOUT:
 895             TRACE("oDepth");
 896             break;
 897
 898         case WINED3DSPR_ATTROUT:
 899             TRACE("oD");
 900             break;
 901
 902         case WINED3DSPR_TEXCRDOUT:
 903             /* Vertex shaders >= 3.0 use general purpose output registers
 904              * (WINED3DSPR_OUTPUT), which can include an address token. */
 905             if (shader_version->major >= 3) TRACE("o");
 906             else TRACE("oT");
 907             break;
 908
 909         case WINED3DSPR_CONSTINT:
 910             TRACE("i");
 911             break;
 912
 913         case WINED3DSPR_CONSTBOOL:
 914             TRACE("b");
 915             break;
 916
 917         case WINED3DSPR_LABEL:
 918             TRACE("l");
 919             break;
 920
 921         case WINED3DSPR_LOOP:
 922             TRACE("aL");
 923             break;
 924
 925         case WINED3DSPR_SAMPLER:
 926             TRACE("s");
 927             break;
 928
 929         case WINED3DSPR_MISCTYPE:
 930             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
 931             else TRACE("%s", misctype_reg_names[reg->idx]);
 932             break;
 933
 934         case WINED3DSPR_PREDICATE:
 935             TRACE("p");
 936             break;
 937
 938         case WINED3DSPR_IMMCONST:
 939             TRACE("l");
 940             break;
 941
 942         case WINED3DSPR_CONSTBUFFER:
 943             TRACE("cb");
 944             break;
 945
 946         default:
 947             TRACE("unhandled_rtype(%#x)", reg->type);
 948             break;
 949     }
 950
 951     if (reg->type == WINED3DSPR_IMMCONST)
 952     {
 953         TRACE("(");
 954         switch (reg->immconst_type)
 955         {
 956             case WINED3D_IMMCONST_FLOAT:
 957                 TRACE("%.8e", *(const float *)reg->immconst_data);
 958                 break;
 959
 960             case WINED3D_IMMCONST_FLOAT4:
 961                 TRACE("%.8e, %.8e, %.8e, %.8e",
 962                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
 963                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
 964                 break;
 965
 966             default:
 967                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
 968                 break;
 969         }
 970         TRACE(")");
 971     }
 972     else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
 973     {
 974         if (reg->array_idx != ~0U)
 975         {
 976             TRACE("%u[%u", offset, reg->array_idx);
 977             if (reg->rel_addr)
 978             {
 979                 TRACE(" + ");
 980                 shader_dump_src_param(reg->rel_addr, shader_version);
 981             }
 982             TRACE("]");
 983         }
 984         else
 985         {
 986             if (reg->rel_addr)
 987             {
 988                 TRACE("[");
 989                 shader_dump_src_param(reg->rel_addr, shader_version);
 990                 TRACE(" + ");
 991             }
 992             TRACE("%u", offset);
 993             if (reg->rel_addr) TRACE("]");
 994         }
 995     }
 996 }
 997
 998 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
 999         const struct wined3d_shader_version *shader_version)
1000 {
1001     DWORD write_mask = param->write_mask;
1002
1003     shader_dump_register(&param->reg, shader_version);
1004
1005     if (write_mask != WINED3DSP_WRITEMASK_ALL)
1006     {
1007         static const char *write_mask_chars = "xyzw";
1008
1009         TRACE(".");
1010         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1011         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1012         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1013         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1014     }
1015 }
1016
1017 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1018         const struct wined3d_shader_version *shader_version)
1019 {
1020     DWORD src_modifier = param->modifiers;
1021     DWORD swizzle = param->swizzle;
1022
1023     if (src_modifier == WINED3DSPSM_NEG
1024             || src_modifier == WINED3DSPSM_BIASNEG
1025             || src_modifier == WINED3DSPSM_SIGNNEG
1026             || src_modifier == WINED3DSPSM_X2NEG
1027             || src_modifier == WINED3DSPSM_ABSNEG)
1028         TRACE("-");
1029     else if (src_modifier == WINED3DSPSM_COMP)
1030         TRACE("1-");
1031     else if (src_modifier == WINED3DSPSM_NOT)
1032         TRACE("!");
1033
1034     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1035         TRACE("abs(");
1036
1037     shader_dump_register(&param->reg, shader_version);
1038
1039     if (src_modifier)
1040     {
1041         switch (src_modifier)
1042         {
1043             case WINED3DSPSM_NONE:    break;
1044             case WINED3DSPSM_NEG:     break;
1045             case WINED3DSPSM_NOT:     break;
1046             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1047             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1048             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1049             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1050             case WINED3DSPSM_COMP:    break;
1051             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1052             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1053             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1054             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1055             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1056             case WINED3DSPSM_ABS:     TRACE(")"); break;
1057             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1058         }
1059     }
1060
1061     if (swizzle != WINED3DSP_NOSWIZZLE)
1062     {
1063         static const char *swizzle_chars = "xyzw";
1064         DWORD swizzle_x = swizzle & 0x03;
1065         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1066         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1067         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1068
1069         if (swizzle_x == swizzle_y
1070                 && swizzle_x == swizzle_z
1071                 && swizzle_x == swizzle_w)
1072         {
1073             TRACE(".%c", swizzle_chars[swizzle_x]);
1074         }
1075         else
1076         {
1077             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1078                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1079         }
1080     }
1081 }
1082
1083 /* Shared code in order to generate the bulk of the shader string.
1084  * NOTE: A description of how to parse tokens can be found on MSDN. */
1085 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1086         const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1087 {
1088     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1089     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1090     const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1091     void *fe_data = shader->baseShader.frontend_data;
1092     struct wined3d_shader_src_param src_rel_addr[4];
1093     struct wined3d_shader_src_param src_param[4];
1094     struct wined3d_shader_version shader_version;
1095     struct wined3d_shader_src_param dst_rel_addr;
1096     struct wined3d_shader_dst_param dst_param;
1097     struct wined3d_shader_instruction ins;
1098     struct wined3d_shader_context ctx;
1099     const DWORD *ptr = byte_code;
1100     DWORD i;
1101
1102     /* Initialize current parsing state. */
1103     ctx.shader = iface;
1104     ctx.gl_info = &device->adapter->gl_info;
1105     ctx.reg_maps = reg_maps;
1106     ctx.buffer = buffer;
1107     ctx.backend_data = backend_ctx;
1108
1109     ins.ctx = &ctx;
1110     ins.dst = &dst_param;
1111     ins.src = src_param;
1112     shader->baseShader.parse_state.current_row = 0;
1113
1114     fe->shader_read_header(fe_data, &ptr, &shader_version);
1115
1116     while (!fe->shader_is_end(fe_data, &ptr))
1117     {
1118         const char *comment;
1119         UINT comment_size;
1120         UINT param_size;
1121
1122         /* Skip comment tokens. */
1123         fe->shader_read_comment(&ptr, &comment, &comment_size);
1124         if (comment) continue;
1125
1126         /* Read opcode. */
1127         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1128
1129         /* Unknown opcode and its parameters. */
1130         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1131         {
1132             TRACE("Skipping unrecognized instruction.\n");
1133             ptr += param_size;
1134             continue;
1135         }
1136
1137         /* Nothing to do. */
1138         if (ins.handler_idx == WINED3DSIH_DCL
1139                 || ins.handler_idx == WINED3DSIH_NOP
1140                 || ins.handler_idx == WINED3DSIH_DEF
1141                 || ins.handler_idx == WINED3DSIH_DEFI
1142                 || ins.handler_idx == WINED3DSIH_DEFB
1143                 || ins.handler_idx == WINED3DSIH_PHASE)
1144         {
1145             ptr += param_size;
1146             continue;
1147         }
1148
1149         /* Destination token */
1150         if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1151
1152         /* Predication token */
1153         if (ins.predicate) ins.predicate = *ptr++;
1154
1155         /* Other source tokens */
1156         for (i = 0; i < ins.src_count; ++i)
1157         {
1158             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1159         }
1160
1161         /* Call appropriate function for output target */
1162         device->shader_backend->shader_handle_instruction(&ins);
1163     }
1164 }
1165
1166 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1167 {
1168     DWORD mmask = dst->modifiers;
1169
1170     switch (dst->shift)
1171     {
1172         case 0: break;
1173         case 13: TRACE("_d8"); break;
1174         case 14: TRACE("_d4"); break;
1175         case 15: TRACE("_d2"); break;
1176         case 1: TRACE("_x2"); break;
1177         case 2: TRACE("_x4"); break;
1178         case 3: TRACE("_x8"); break;
1179         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1180     }
1181
1182     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1183     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1184     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1185
1186     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1187     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1188 }
1189
1190 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1191 {
1192     struct wined3d_shader_version shader_version;
1193     const DWORD *ptr = byte_code;
1194     const char *type_prefix;
1195     DWORD i;
1196
1197     TRACE("Parsing %p.\n", byte_code);
1198
1199     fe->shader_read_header(fe_data, &ptr, &shader_version);
1200
1201     switch (shader_version.type)
1202     {
1203         case WINED3D_SHADER_TYPE_VERTEX:
1204             type_prefix = "vs";
1205             break;
1206
1207         case WINED3D_SHADER_TYPE_GEOMETRY:
1208             type_prefix = "gs";
1209             break;
1210
1211         case WINED3D_SHADER_TYPE_PIXEL:
1212             type_prefix = "ps";
1213             break;
1214
1215         default:
1216             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1217             type_prefix = "unknown";
1218             break;
1219     }
1220
1221     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1222
1223     while (!fe->shader_is_end(fe_data, &ptr))
1224     {
1225         struct wined3d_shader_instruction ins;
1226         const char *comment;
1227         UINT comment_size;
1228         UINT param_size;
1229
1230         /* comment */
1231         fe->shader_read_comment(&ptr, &comment, &comment_size);
1232         if (comment)
1233         {
1234             TRACE("// %s\n", debugstr_an(comment, comment_size));
1235             continue;
1236         }
1237
1238         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1239         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1240         {
1241             TRACE("Skipping unrecognized instruction.\n");
1242             ptr += param_size;
1243             continue;
1244         }
1245
1246         if (ins.handler_idx == WINED3DSIH_DCL)
1247         {
1248             struct wined3d_shader_semantic semantic;
1249
1250             fe->shader_read_semantic(&ptr, &semantic);
1251
1252             shader_dump_decl_usage(&semantic, &shader_version);
1253             shader_dump_ins_modifiers(&semantic.reg);
1254             TRACE(" ");
1255             shader_dump_dst_param(&semantic.reg, &shader_version);
1256         }
1257         else if (ins.handler_idx == WINED3DSIH_DEF)
1258         {
1259             struct wined3d_shader_src_param rel_addr;
1260             struct wined3d_shader_dst_param dst;
1261
1262             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1263
1264             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1265                     *(const float *)(ptr),
1266                     *(const float *)(ptr + 1),
1267                     *(const float *)(ptr + 2),
1268                     *(const float *)(ptr + 3));
1269             ptr += 4;
1270         }
1271         else if (ins.handler_idx == WINED3DSIH_DEFI)
1272         {
1273             struct wined3d_shader_src_param rel_addr;
1274             struct wined3d_shader_dst_param dst;
1275
1276             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1277
1278             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1279                     *(ptr),
1280                     *(ptr + 1),
1281                     *(ptr + 2),
1282                     *(ptr + 3));
1283             ptr += 4;
1284         }
1285         else if (ins.handler_idx == WINED3DSIH_DEFB)
1286         {
1287             struct wined3d_shader_src_param rel_addr;
1288             struct wined3d_shader_dst_param dst;
1289
1290             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1291
1292             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1293             ++ptr;
1294         }
1295         else
1296         {
1297             struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1298             struct wined3d_shader_dst_param dst_param;
1299             struct wined3d_shader_src_param src_param;
1300
1301             if (ins.dst_count)
1302             {
1303                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1304             }
1305
1306             /* Print out predication source token first - it follows
1307              * the destination token. */
1308             if (ins.predicate)
1309             {
1310                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1311                 TRACE("(");
1312                 shader_dump_src_param(&src_param, &shader_version);
1313                 TRACE(") ");
1314             }
1315
1316             /* PixWin marks instructions with the coissue flag with a '+' */
1317             if (ins.coissue) TRACE("+");
1318
1319             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1320
1321             if (ins.handler_idx == WINED3DSIH_IFC
1322                     || ins.handler_idx == WINED3DSIH_BREAKC)
1323             {
1324                 switch (ins.flags)
1325                 {
1326                     case COMPARISON_GT: TRACE("_gt"); break;
1327                     case COMPARISON_EQ: TRACE("_eq"); break;
1328                     case COMPARISON_GE: TRACE("_ge"); break;
1329                     case COMPARISON_LT: TRACE("_lt"); break;
1330                     case COMPARISON_NE: TRACE("_ne"); break;
1331                     case COMPARISON_LE: TRACE("_le"); break;
1332                     default: TRACE("_(%u)", ins.flags);
1333                 }
1334             }
1335             else if (ins.handler_idx == WINED3DSIH_TEX
1336                     && shader_version.major >= 2
1337                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1338             {
1339                 TRACE("p");
1340             }
1341
1342             /* We already read the destination token, print it. */
1343             if (ins.dst_count)
1344             {
1345                 shader_dump_ins_modifiers(&dst_param);
1346                 TRACE(" ");
1347                 shader_dump_dst_param(&dst_param, &shader_version);
1348             }
1349
1350             /* Other source tokens */
1351             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1352             {
1353                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1354                 TRACE(!i ? " " : ", ");
1355                 shader_dump_src_param(&src_param, &shader_version);
1356             }
1357         }
1358         TRACE("\n");
1359     }
1360 }
1361
1362 static void shader_cleanup(IWineD3DBaseShader *iface)
1363 {
1364     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1365
1366     ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1367     HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1368     HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1369     shader_delete_constant_list(&shader->baseShader.constantsF);
1370     shader_delete_constant_list(&shader->baseShader.constantsB);
1371     shader_delete_constant_list(&shader->baseShader.constantsI);
1372     list_remove(&shader->baseShader.shader_list_entry);
1373
1374     if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1375     {
1376         shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1377     }
1378 }
1379
1380 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1381 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1382 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1383 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1384 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1385 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1386 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1387 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1388 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1389 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1390 static void shader_none_free(IWineD3DDevice *iface) {}
1391 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1392
1393 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1394 {
1395     /* Set the shader caps to 0 for the none shader backend */
1396     caps->VertexShaderVersion = 0;
1397     caps->PixelShaderVersion = 0;
1398     caps->PixelShader1xMaxValue = 0.0f;
1399 }
1400
1401 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1402 {
1403     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1404     {
1405         TRACE("Checking support for fixup:\n");
1406         dump_color_fixup_desc(fixup);
1407     }
1408
1409     /* Faked to make some apps happy. */
1410     if (!is_complex_fixup(fixup))
1411     {
1412         TRACE("[OK]\n");
1413         return TRUE;
1414     }
1415
1416     TRACE("[FAILED]\n");
1417     return FALSE;
1418 }
1419
1420 const shader_backend_t none_shader_backend = {
1421     shader_none_handle_instruction,
1422     shader_none_select,
1423     shader_none_select_depth_blt,
1424     shader_none_deselect_depth_blt,
1425     shader_none_update_float_vertex_constants,
1426     shader_none_update_float_pixel_constants,
1427     shader_none_load_constants,
1428     shader_none_load_np2fixup_constants,
1429     shader_none_destroy,
1430     shader_none_alloc,
1431     shader_none_free,
1432     shader_none_dirty_const,
1433     shader_none_get_caps,
1434     shader_none_color_fixup_supported,
1435 };
1436
1437 static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1438 {
1439     *parent = shader->baseShader.parent;
1440     IUnknown_AddRef(*parent);
1441     TRACE("shader %p, returning %p.\n", shader, *parent);
1442 }
1443
1444 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1445 {
1446     if (!data)
1447     {
1448         *data_size = shader->baseShader.functionLength;
1449         return WINED3D_OK;
1450     }
1451
1452     if (*data_size < shader->baseShader.functionLength)
1453     {
1454         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1455          * than the required size we should write the required size and
1456          * return D3DERR_MOREDATA. That's not actually true. */
1457         return WINED3DERR_INVALIDCALL;
1458     }
1459
1460     memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1461
1462     return WINED3D_OK;
1463 }
1464
1465 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1466         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1467 {
1468     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1469     const struct wined3d_shader_frontend *fe;
1470     HRESULT hr;
1471
1472     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1473             shader, byte_code, output_signature, float_const_count);
1474
1475     fe = shader_select_frontend(*byte_code);
1476     if (!fe)
1477     {
1478         FIXME("Unable to find frontend for shader.\n");
1479         return WINED3DERR_INVALIDCALL;
1480     }
1481     shader->baseShader.frontend = fe;
1482     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1483     if (!shader->baseShader.frontend_data)
1484     {
1485         FIXME("Failed to initialize frontend.\n");
1486         return WINED3DERR_INVALIDCALL;
1487     }
1488
1489     /* First pass: trace shader. */
1490     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1491
1492     /* Initialize immediate constant lists. */
1493     list_init(&shader->baseShader.constantsF);
1494     list_init(&shader->baseShader.constantsB);
1495     list_init(&shader->baseShader.constantsI);
1496
1497     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1498     hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1499             reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1500             byte_code, float_const_count);
1501     if (FAILED(hr)) return hr;
1502
1503     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1504     if (!shader->baseShader.function) return E_OUTOFMEMORY;
1505     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1506
1507     return WINED3D_OK;
1508 }
1509
1510 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1511 {
1512     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1513
1514     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1515             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1516             || IsEqualGUID(riid, &IID_IWineD3DBase)
1517             || IsEqualGUID(riid, &IID_IUnknown))
1518     {
1519         IUnknown_AddRef(iface);
1520         *object = iface;
1521         return S_OK;
1522     }
1523
1524     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1525
1526     *object = NULL;
1527     return E_NOINTERFACE;
1528 }
1529
1530 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1531 {
1532     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1533     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1534
1535     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1536
1537     return refcount;
1538 }
1539
1540 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1541 {
1542     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1543     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1544
1545     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1546
1547     if (!refcount)
1548     {
1549         shader_cleanup((IWineD3DBaseShader *)iface);
1550         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1551         HeapFree(GetProcessHeap(), 0, shader);
1552     }
1553
1554     return refcount;
1555 }
1556
1557 static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1558 {
1559     TRACE("iface %p, parent %p.\n", iface, parent);
1560
1561     shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1562
1563     return WINED3D_OK;
1564 }
1565
1566 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1567 {
1568     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1569
1570     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1571 }
1572
1573 /* Set local constants for d3d8 shaders. */
1574 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1575         UINT start_idx, const float *src_data, UINT count)
1576 {
1577     IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1578     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1579     UINT i, end_idx;
1580
1581     TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1582
1583     end_idx = start_idx + count;
1584     if (end_idx > device->d3d_vshader_constantF)
1585     {
1586         WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1587         end_idx = device->d3d_vshader_constantF;
1588     }
1589
1590     for (i = start_idx; i < end_idx; ++i)
1591     {
1592         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1593         if (!lconst) return E_OUTOFMEMORY;
1594
1595         lconst->idx = i;
1596         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1597         list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1598     }
1599
1600     return WINED3D_OK;
1601 }
1602
1603 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1604 {
1605     /* IUnknown methods */
1606     vertexshader_QueryInterface,
1607     vertexshader_AddRef,
1608     vertexshader_Release,
1609     /* IWineD3DBase methods */
1610     vertexshader_GetParent,
1611     /* IWineD3DBaseShader methods */
1612     vertexshader_GetFunction,
1613     /* IWineD3DVertexShader methods */
1614     vertexshader_SetLocalConstantsF,
1615 };
1616
1617 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1618         IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1619 {
1620     args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1621     args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1622             && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1623     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1624 }
1625
1626 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1627 {
1628     if (usage_idx1 != usage_idx2) return FALSE;
1629     if (usage1 == usage2) return TRUE;
1630     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1631     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1632
1633     return FALSE;
1634 }
1635
1636 BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1637 {
1638     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1639     WORD map = shader->baseShader.reg_maps.input_registers;
1640     unsigned int i;
1641
1642     for (i = 0; map; map >>= 1, ++i)
1643     {
1644         if (!(map & 1)) continue;
1645
1646         if (match_usage(shader->attributes[i].usage,
1647                 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1648         {
1649             *regnum = i;
1650             return TRUE;
1651         }
1652     }
1653     return FALSE;
1654 }
1655
1656 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1657 {
1658     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1659             shader->baseShader.reg_maps.shader_version.minor);
1660     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1661
1662     shader->baseShader.limits.texcoord = 0;
1663     shader->baseShader.limits.attributes = 16;
1664     shader->baseShader.limits.packed_input = 0;
1665
1666     switch (shader_version)
1667     {
1668         case WINED3D_SHADER_VERSION(1, 0):
1669         case WINED3D_SHADER_VERSION(1, 1):
1670             shader->baseShader.limits.temporary = 12;
1671             shader->baseShader.limits.constant_bool = 0;
1672             shader->baseShader.limits.constant_int = 0;
1673             shader->baseShader.limits.address = 1;
1674             shader->baseShader.limits.packed_output = 0;
1675             shader->baseShader.limits.sampler = 0;
1676             shader->baseShader.limits.label = 0;
1677             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1678              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1679              * constants? */
1680             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1681             break;
1682
1683         case WINED3D_SHADER_VERSION(2, 0):
1684         case WINED3D_SHADER_VERSION(2, 1):
1685             shader->baseShader.limits.temporary = 12;
1686             shader->baseShader.limits.constant_bool = 16;
1687             shader->baseShader.limits.constant_int = 16;
1688             shader->baseShader.limits.address = 1;
1689             shader->baseShader.limits.packed_output = 0;
1690             shader->baseShader.limits.sampler = 0;
1691             shader->baseShader.limits.label = 16;
1692             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1693             break;
1694
1695         case WINED3D_SHADER_VERSION(4, 0):
1696             FIXME("Using 3.0 limits for 4.0 shader.\n");
1697             /* Fall through. */
1698
1699         case WINED3D_SHADER_VERSION(3, 0):
1700             shader->baseShader.limits.temporary = 32;
1701             shader->baseShader.limits.constant_bool = 32;
1702             shader->baseShader.limits.constant_int = 32;
1703             shader->baseShader.limits.address = 1;
1704             shader->baseShader.limits.packed_output = 12;
1705             shader->baseShader.limits.sampler = 4;
1706             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1707             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1708              * even though they are capable of supporting much more (GL
1709              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1710              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1711              * shaders to 256. */
1712             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1713             break;
1714
1715         default:
1716             shader->baseShader.limits.temporary = 12;
1717             shader->baseShader.limits.constant_bool = 16;
1718             shader->baseShader.limits.constant_int = 16;
1719             shader->baseShader.limits.address = 1;
1720             shader->baseShader.limits.packed_output = 0;
1721             shader->baseShader.limits.sampler = 0;
1722             shader->baseShader.limits.label = 16;
1723             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1724             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1725                     shader->baseShader.reg_maps.shader_version.major,
1726                     shader->baseShader.reg_maps.shader_version.minor);
1727     }
1728 }
1729
1730 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1731         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1732         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1733 {
1734     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1735     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1736     unsigned int i;
1737     HRESULT hr;
1738     WORD map;
1739
1740     if (!byte_code) return WINED3DERR_INVALIDCALL;
1741
1742     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1743     shader_init(&shader->baseShader, device, parent, parent_ops);
1744
1745     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1746             output_signature, device->d3d_vshader_constantF);
1747     if (FAILED(hr))
1748     {
1749         WARN("Failed to set function, hr %#x.\n", hr);
1750         shader_cleanup((IWineD3DBaseShader *)shader);
1751         return hr;
1752     }
1753
1754     map = shader->baseShader.reg_maps.input_registers;
1755     for (i = 0; map; map >>= 1, ++i)
1756     {
1757         if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1758
1759         shader->attributes[i].usage =
1760                 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1761         shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1762     }
1763
1764     if (output_signature)
1765     {
1766         for (i = 0; i < output_signature->element_count; ++i)
1767         {
1768             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1769             reg_maps->output_registers |= 1 << e->register_idx;
1770             shader->baseShader.output_signature[e->register_idx] = *e;
1771         }
1772     }
1773
1774     vertexshader_set_limits(shader);
1775
1776     if (device->vs_selected_mode == SHADER_ARB
1777             && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1778             && shader->min_rel_offset <= shader->max_rel_offset)
1779     {
1780         if (shader->max_rel_offset - shader->min_rel_offset > 127)
1781         {
1782             FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1783             FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1784             FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1785         }
1786         else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1787         {
1788             shader->rel_offset = shader->min_rel_offset + 63;
1789         }
1790         else if (shader->max_rel_offset > 63)
1791         {
1792             shader->rel_offset = shader->min_rel_offset;
1793         }
1794         else
1795         {
1796             shader->rel_offset = 0;
1797         }
1798     }
1799
1800     shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1801             && !list_empty(&shader->baseShader.constantsF);
1802
1803     return WINED3D_OK;
1804 }
1805
1806 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1807         REFIID riid, void **object)
1808 {
1809     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1810
1811     if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1812             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1813             || IsEqualGUID(riid, &IID_IWineD3DBase)
1814             || IsEqualGUID(riid, &IID_IUnknown))
1815     {
1816         IUnknown_AddRef(iface);
1817         *object = iface;
1818         return S_OK;
1819     }
1820
1821     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1822
1823     *object = NULL;
1824     return E_NOINTERFACE;
1825 }
1826
1827 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1828 {
1829     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1830     ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1831
1832     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1833
1834     return refcount;
1835 }
1836
1837 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1838 {
1839     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1840     ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1841
1842     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1843
1844     if (!refcount)
1845     {
1846         shader_cleanup((IWineD3DBaseShader *)iface);
1847         shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1848         HeapFree(GetProcessHeap(), 0, shader);
1849     }
1850
1851     return refcount;
1852 }
1853
1854 static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1855 {
1856     TRACE("iface %p, parent %p.\n", iface, parent);
1857
1858     shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1859
1860     return WINED3D_OK;
1861 }
1862
1863 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1864 {
1865     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1866
1867     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1868 }
1869
1870 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1871 {
1872     /* IUnknown methods */
1873     geometryshader_QueryInterface,
1874     geometryshader_AddRef,
1875     geometryshader_Release,
1876     /* IWineD3DBase methods */
1877     geometryshader_GetParent,
1878     /* IWineD3DBaseShader methods */
1879     geometryshader_GetFunction,
1880 };
1881
1882 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1883         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1884         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1885 {
1886     HRESULT hr;
1887
1888     shader->vtbl = &wined3d_geometryshader_vtbl;
1889     shader_init(&shader->base_shader, device, parent, parent_ops);
1890
1891     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1892     if (FAILED(hr))
1893     {
1894         WARN("Failed to set function, hr %#x.\n", hr);
1895         shader_cleanup((IWineD3DBaseShader *)shader);
1896         return hr;
1897     }
1898
1899     shader->base_shader.load_local_constsF = FALSE;
1900
1901     return WINED3D_OK;
1902 }
1903
1904 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1905 {
1906     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1907
1908     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1909             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1910             || IsEqualGUID(riid, &IID_IWineD3DBase)
1911             || IsEqualGUID(riid, &IID_IUnknown))
1912     {
1913         IUnknown_AddRef(iface);
1914         *object = iface;
1915         return S_OK;
1916     }
1917
1918     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1919
1920     *object = NULL;
1921     return E_NOINTERFACE;
1922 }
1923
1924 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
1925 {
1926     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1927     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1928
1929     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1930
1931     return refcount;
1932 }
1933
1934 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
1935 {
1936     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1937     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1938
1939     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1940
1941     if (!refcount)
1942     {
1943         shader_cleanup((IWineD3DBaseShader *)iface);
1944         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1945         HeapFree(GetProcessHeap(), 0, shader);
1946     }
1947
1948     return refcount;
1949 }
1950
1951 static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
1952 {
1953     TRACE("iface %p, parent %p.\n", iface, parent);
1954
1955     shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1956
1957     return WINED3D_OK;
1958 }
1959
1960 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
1961 {
1962     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1963
1964     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1965 }
1966
1967 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
1968 {
1969     /* IUnknown methods */
1970     pixelshader_QueryInterface,
1971     pixelshader_AddRef,
1972     pixelshader_Release,
1973     /* IWineD3DBase methods */
1974     pixelshader_GetParent,
1975     /* IWineD3DBaseShader methods */
1976     pixelshader_GetFunction
1977 };
1978
1979 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
1980         IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
1981 {
1982     IWineD3DBaseTextureImpl *texture;
1983     UINT i;
1984
1985     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1986     args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
1987     args->np2_fixup = 0;
1988
1989     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1990     {
1991         if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
1992         texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
1993         if (!texture)
1994         {
1995             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1996             continue;
1997         }
1998         args->color_fixup[i] = texture->resource.format_desc->color_fixup;
1999
2000         /* Flag samplers that need NP2 texcoord fixup. */
2001         if (!texture->baseTexture.pow2Matrix_identity)
2002         {
2003             args->np2_fixup |= (1 << i);
2004         }
2005     }
2006     if (shader->baseShader.reg_maps.shader_version.major >= 3)
2007     {
2008         if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2009         {
2010             args->vp_mode = pretransformed;
2011         }
2012         else if (use_vs(stateblock))
2013         {
2014             args->vp_mode = vertexshader;
2015         }
2016         else
2017         {
2018             args->vp_mode = fixedfunction;
2019         }
2020         args->fog = FOG_OFF;
2021     }
2022     else
2023     {
2024         args->vp_mode = vertexshader;
2025         if (stateblock->renderState[WINED3DRS_FOGENABLE])
2026         {
2027             switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2028             {
2029                 case WINED3DFOG_NONE:
2030                     if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2031                             || use_vs(stateblock))
2032                     {
2033                         args->fog = FOG_LINEAR;
2034                         break;
2035                     }
2036
2037                     switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2038                     {
2039                         case WINED3DFOG_NONE: /* Fall through. */
2040                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2041                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2042                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2043                     }
2044                     break;
2045
2046                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2047                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2048                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2049             }
2050         }
2051         else
2052         {
2053             args->fog = FOG_OFF;
2054         }
2055     }
2056 }
2057
2058 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2059 {
2060     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2061             shader->baseShader.reg_maps.shader_version.minor);
2062
2063     shader->baseShader.limits.attributes = 0;
2064     shader->baseShader.limits.address = 0;
2065     shader->baseShader.limits.packed_output = 0;
2066
2067     switch (shader_version)
2068     {
2069         case WINED3D_SHADER_VERSION(1, 0):
2070         case WINED3D_SHADER_VERSION(1, 1):
2071         case WINED3D_SHADER_VERSION(1, 2):
2072         case WINED3D_SHADER_VERSION(1, 3):
2073             shader->baseShader.limits.temporary = 2;
2074             shader->baseShader.limits.constant_float = 8;
2075             shader->baseShader.limits.constant_int = 0;
2076             shader->baseShader.limits.constant_bool = 0;
2077             shader->baseShader.limits.texcoord = 4;
2078             shader->baseShader.limits.sampler = 4;
2079             shader->baseShader.limits.packed_input = 0;
2080             shader->baseShader.limits.label = 0;
2081             break;
2082
2083         case WINED3D_SHADER_VERSION(1, 4):
2084             shader->baseShader.limits.temporary = 6;
2085             shader->baseShader.limits.constant_float = 8;
2086             shader->baseShader.limits.constant_int = 0;
2087             shader->baseShader.limits.constant_bool = 0;
2088             shader->baseShader.limits.texcoord = 6;
2089             shader->baseShader.limits.sampler = 6;
2090             shader->baseShader.limits.packed_input = 0;
2091             shader->baseShader.limits.label = 0;
2092             break;
2093
2094         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2095         case WINED3D_SHADER_VERSION(2, 0):
2096             shader->baseShader.limits.temporary = 32;
2097             shader->baseShader.limits.constant_float = 32;
2098             shader->baseShader.limits.constant_int = 16;
2099             shader->baseShader.limits.constant_bool = 16;
2100             shader->baseShader.limits.texcoord = 8;
2101             shader->baseShader.limits.sampler = 16;
2102             shader->baseShader.limits.packed_input = 0;
2103             break;
2104
2105         case WINED3D_SHADER_VERSION(2, 1):
2106             shader->baseShader.limits.temporary = 32;
2107             shader->baseShader.limits.constant_float = 32;
2108             shader->baseShader.limits.constant_int = 16;
2109             shader->baseShader.limits.constant_bool = 16;
2110             shader->baseShader.limits.texcoord = 8;
2111             shader->baseShader.limits.sampler = 16;
2112             shader->baseShader.limits.packed_input = 0;
2113             shader->baseShader.limits.label = 16;
2114             break;
2115
2116         case WINED3D_SHADER_VERSION(4, 0):
2117             FIXME("Using 3.0 limits for 4.0 shader.\n");
2118             /* Fall through. */
2119
2120         case WINED3D_SHADER_VERSION(3, 0):
2121             shader->baseShader.limits.temporary = 32;
2122             shader->baseShader.limits.constant_float = 224;
2123             shader->baseShader.limits.constant_int = 16;
2124             shader->baseShader.limits.constant_bool = 16;
2125             shader->baseShader.limits.texcoord = 0;
2126             shader->baseShader.limits.sampler = 16;
2127             shader->baseShader.limits.packed_input = 12;
2128             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2129             break;
2130
2131         default:
2132             shader->baseShader.limits.temporary = 32;
2133             shader->baseShader.limits.constant_float = 32;
2134             shader->baseShader.limits.constant_int = 16;
2135             shader->baseShader.limits.constant_bool = 16;
2136             shader->baseShader.limits.texcoord = 8;
2137             shader->baseShader.limits.sampler = 16;
2138             shader->baseShader.limits.packed_input = 0;
2139             shader->baseShader.limits.label = 0;
2140             FIXME("Unrecognized pixel shader version %u.%u\n",
2141                     shader->baseShader.reg_maps.shader_version.major,
2142                     shader->baseShader.reg_maps.shader_version.minor);
2143     }
2144 }
2145
2146 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2147         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2148         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2149 {
2150     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2151     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2152     HRESULT hr;
2153
2154     if (!byte_code) return WINED3DERR_INVALIDCALL;
2155
2156     shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2157     shader_init(&shader->baseShader, device, parent, parent_ops);
2158
2159     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2160             output_signature, device->d3d_pshader_constantF);
2161     if (FAILED(hr))
2162     {
2163         WARN("Failed to set function, hr %#x.\n", hr);
2164         shader_cleanup((IWineD3DBaseShader *)shader);
2165         return hr;
2166     }
2167
2168     pixelshader_set_limits(shader);
2169
2170     for (i = 0; i < MAX_REG_INPUT; ++i)
2171     {
2172         if (shader->input_reg_used[i])
2173         {
2174             ++num_regs_used;
2175             highest_reg_used = i;
2176         }
2177     }
2178
2179     /* Don't do any register mapping magic if it is not needed, or if we can't
2180      * achieve anything anyway */
2181     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2182             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2183     {
2184         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2185         {
2186             /* This happens with relative addressing. The input mapper function
2187              * warns about this if the higher registers are declared too, so
2188              * don't write a FIXME here */
2189             WARN("More varying registers used than supported\n");
2190         }
2191
2192         for (i = 0; i < MAX_REG_INPUT; ++i)
2193         {
2194             shader->input_reg_map[i] = i;
2195         }
2196
2197         shader->declared_in_count = highest_reg_used + 1;
2198     }
2199     else
2200     {
2201         shader->declared_in_count = 0;
2202         for (i = 0; i < MAX_REG_INPUT; ++i)
2203         {
2204             if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2205             else shader->input_reg_map[i] = ~0U;
2206         }
2207     }
2208
2209     shader->baseShader.load_local_constsF = FALSE;
2210
2211     return WINED3D_OK;
2212 }
2213
2214 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2215 {
2216     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2217     unsigned int i;
2218
2219     if (reg_maps->shader_version.major != 1) return;
2220
2221     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2222     {
2223         /* We don't sample from this sampler. */
2224         if (!sampler_type[i]) continue;
2225
2226         if (!textures[i])
2227         {
2228             WARN("No texture bound to sampler %u, using 2D.\n", i);
2229             sampler_type[i] = WINED3DSTT_2D;
2230             continue;
2231         }
2232
2233         switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2234         {
2235             case GL_TEXTURE_RECTANGLE_ARB:
2236             case GL_TEXTURE_2D:
2237                 /* We have to select between texture rectangles and 2D
2238                  * textures later because 2.0 and 3.0 shaders only have
2239                  * WINED3DSTT_2D as well. */
2240                 sampler_type[i] = WINED3DSTT_2D;
2241                 break;
2242
2243             case GL_TEXTURE_3D:
2244                 sampler_type[i] = WINED3DSTT_VOLUME;
2245                 break;
2246
2247             case GL_TEXTURE_CUBE_MAP_ARB:
2248                 sampler_type[i] = WINED3DSTT_CUBE;
2249                 break;
2250
2251             default:
2252                 FIXME("Unrecognized texture type %#x, using 2D.\n",
2253                         IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2254                 sampler_type[i] = WINED3DSTT_2D;
2255         }
2256     }
2257 }