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 param_size;
 471
 472         /* Skip comments. */
 473         fe->shader_read_comment(&ptr, &comment);
 474         if (comment) continue;
 475
 476         /* Fetch opcode. */
 477         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
 478
 479         /* Unhandled opcode, and its parameters. */
 480         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
 481         {
 482             TRACE("Skipping unrecognized instruction.\n");
 483             ptr += param_size;
 484             continue;
 485         }
 486
 487         /* Handle declarations. */
 488         if (ins.handler_idx == WINED3DSIH_DCL)
 489         {
 490             struct wined3d_shader_semantic semantic;
 491
 492             fe->shader_read_semantic(&ptr, &semantic);
 493
 494             switch (semantic.reg.reg.type)
 495             {
 496                 /* Mark input registers used. */
 497                 case WINED3DSPR_INPUT:
 498                     reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
 499                     shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
 500                     break;
 501
 502                 /* Vertex shader: mark 3.0 output registers used, save token. */
 503                 case WINED3DSPR_OUTPUT:
 504                     reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
 505                     shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
 506                     if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
 507                     break;
 508
 509                 /* Save sampler usage token. */
 510                 case WINED3DSPR_SAMPLER:
 511                     reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
 512                     break;
 513
 514                 default:
 515                     TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
 516                     break;
 517             }
 518         }
 519         else if (ins.handler_idx == WINED3DSIH_DEF)
 520         {
 521             struct wined3d_shader_src_param rel_addr;
 522             struct wined3d_shader_dst_param dst;
 523
 524             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 525             if (!lconst) return E_OUTOFMEMORY;
 526
 527             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 528             lconst->idx = dst.reg.idx;
 529
 530             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 531             ptr += 4;
 532
 533             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
 534             if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 535             {
 536                 float *value = (float *)lconst->value;
 537                 if (value[0] < -1.0f) value[0] = -1.0f;
 538                 else if (value[0] > 1.0f) value[0] = 1.0f;
 539                 if (value[1] < -1.0f) value[1] = -1.0f;
 540                 else if (value[1] > 1.0f) value[1] = 1.0f;
 541                 if (value[2] < -1.0f) value[2] = -1.0f;
 542                 else if (value[2] > 1.0f) value[2] = 1.0f;
 543                 if (value[3] < -1.0f) value[3] = -1.0f;
 544                 else if (value[3] > 1.0f) value[3] = 1.0f;
 545             }
 546
 547             list_add_head(&shader->baseShader.constantsF, &lconst->entry);
 548         }
 549         else if (ins.handler_idx == WINED3DSIH_DEFI)
 550         {
 551             struct wined3d_shader_src_param rel_addr;
 552             struct wined3d_shader_dst_param dst;
 553
 554             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 555             if (!lconst) return E_OUTOFMEMORY;
 556
 557             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 558             lconst->idx = dst.reg.idx;
 559
 560             memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
 561             ptr += 4;
 562
 563             list_add_head(&shader->baseShader.constantsI, &lconst->entry);
 564             reg_maps->local_int_consts |= (1 << dst.reg.idx);
 565         }
 566         else if (ins.handler_idx == WINED3DSIH_DEFB)
 567         {
 568             struct wined3d_shader_src_param rel_addr;
 569             struct wined3d_shader_dst_param dst;
 570
 571             local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 572             if (!lconst) return E_OUTOFMEMORY;
 573
 574             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
 575             lconst->idx = dst.reg.idx;
 576
 577             memcpy(lconst->value, ptr, sizeof(DWORD));
 578             ++ptr;
 579
 580             list_add_head(&shader->baseShader.constantsB, &lconst->entry);
 581             reg_maps->local_bool_consts |= (1 << dst.reg.idx);
 582         }
 583         /* If there's a loop in the shader. */
 584         else if (ins.handler_idx == WINED3DSIH_LOOP
 585                 || ins.handler_idx == WINED3DSIH_REP)
 586         {
 587             struct wined3d_shader_src_param src, rel_addr;
 588
 589             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 590
 591             /* Rep and Loop always use an integer constant for the control parameters. */
 592             if (ins.handler_idx == WINED3DSIH_REP)
 593             {
 594                 reg_maps->integer_constants |= 1 << src.reg.idx;
 595             }
 596             else
 597             {
 598                 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 599                 reg_maps->integer_constants |= 1 << src.reg.idx;
 600             }
 601
 602             cur_loop_depth++;
 603             if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
 604         }
 605         else if (ins.handler_idx == WINED3DSIH_ENDLOOP
 606                 || ins.handler_idx == WINED3DSIH_ENDREP)
 607         {
 608             cur_loop_depth--;
 609         }
 610         /* For subroutine prototypes. */
 611         else if (ins.handler_idx == WINED3DSIH_LABEL)
 612         {
 613             struct wined3d_shader_src_param src, rel_addr;
 614
 615             fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
 616             reg_maps->labels |= 1 << src.reg.idx;
 617         }
 618         /* Set texture, address, temporary registers. */
 619         else
 620         {
 621             BOOL color0_mov = FALSE;
 622             int i, limit;
 623
 624             /* This will loop over all the registers and try to
 625              * make a bitmask of the ones we're interested in.
 626              *
 627              * Relative addressing tokens are ignored, but that's
 628              * okay, since we'll catch any address registers when
 629              * they are initialized (required by spec). */
 630             if (ins.dst_count)
 631             {
 632                 struct wined3d_shader_src_param dst_rel_addr;
 633                 struct wined3d_shader_dst_param dst_param;
 634
 635                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
 636
 637                 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
 638
 639                 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
 640                  * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
 641                  * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
 642                 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
 643                         && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
 644                 {
 645                     reg_maps->texcoord_mask[dst_param.reg.idx] |= dst_param.write_mask;
 646                 }
 647
 648                 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
 649                 {
 650                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 651
 652                     if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
 653                     {
 654                     /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
 655                      * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
 656                      * the mov and perform the sRGB write correction from the source register.
 657                      *
 658                      * However, if the mov is only partial, we can't do this, and if the write
 659                      * comes from an instruction other than MOV it is hard to do as well. If
 660                      * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
 661
 662                         ps->color0_mov = FALSE;
 663                         if (ins.handler_idx == WINED3DSIH_MOV)
 664                         {
 665                             /* Used later when the source register is read. */
 666                             color0_mov = TRUE;
 667                         }
 668                     }
 669                     /* Also drop the MOV marker if the source register is overwritten prior to the shader
 670                      * end
 671                      */
 672                     else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
 673                     {
 674                         ps->color0_mov = FALSE;
 675                     }
 676                 }
 677
 678                 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
 679                 if (shader_version.major == 1
 680                         && (ins.handler_idx == WINED3DSIH_TEX
 681                             || ins.handler_idx == WINED3DSIH_TEXBEM
 682                             || ins.handler_idx == WINED3DSIH_TEXBEML
 683                             || ins.handler_idx == WINED3DSIH_TEXDP3TEX
 684                             || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
 685                             || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
 686                             || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
 687                             || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
 688                             || ins.handler_idx == WINED3DSIH_TEXREG2AR
 689                             || ins.handler_idx == WINED3DSIH_TEXREG2GB
 690                             || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
 691                 {
 692                     /* Fake sampler usage, only set reserved bit and type. */
 693                     DWORD sampler_code = dst_param.reg.idx;
 694
 695                     TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
 696                     reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
 697
 698                     /* texbem is only valid with < 1.4 pixel shaders */
 699                     if (ins.handler_idx == WINED3DSIH_TEXBEM
 700                             || ins.handler_idx == WINED3DSIH_TEXBEML)
 701                     {
 702                         reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 703                         if (ins.handler_idx == WINED3DSIH_TEXBEML)
 704                         {
 705                             reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
 706                         }
 707                     }
 708                 }
 709                 else if (ins.handler_idx == WINED3DSIH_BEM)
 710                 {
 711                     reg_maps->bumpmat |= 1 << dst_param.reg.idx;
 712                 }
 713             }
 714
 715             if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
 716             else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
 717             else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
 718             else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
 719             else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
 720             else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
 721             else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
 722             else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
 723
 724             limit = ins.src_count + (ins.predicate ? 1 : 0);
 725             for (i = 0; i < limit; ++i)
 726             {
 727                 struct wined3d_shader_src_param src_param, src_rel_addr;
 728                 unsigned int count;
 729
 730                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
 731                 count = get_instr_extra_regcount(ins.handler_idx, i);
 732
 733                 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 734                 while (count)
 735                 {
 736                     ++src_param.reg.idx;
 737                     shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
 738                     --count;
 739                 }
 740
 741                 if (color0_mov)
 742                 {
 743                     IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
 744                     if (src_param.reg.type == WINED3DSPR_TEMP
 745                             && src_param.swizzle == WINED3DSP_NOSWIZZLE)
 746                     {
 747                         ps->color0_mov = TRUE;
 748                         ps->color0_reg = src_param.reg.idx;
 749                     }
 750                 }
 751             }
 752         }
 753     }
 754     reg_maps->loop_depth = max_loop_depth;
 755
 756     shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
 757
 758     return WINED3D_OK;
 759 }
 760
 761 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
 762 {
 763     DWORD map = 1 << max;
 764     map |= map - 1;
 765     map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
 766
 767     return wined3d_log2i(map);
 768 }
 769
 770 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
 771         const struct wined3d_shader_version *shader_version)
 772 {
 773     TRACE("dcl");
 774
 775     if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
 776     {
 777         switch (semantic->sampler_type)
 778         {
 779             case WINED3DSTT_2D: TRACE("_2d"); break;
 780             case WINED3DSTT_CUBE: TRACE("_cube"); break;
 781             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
 782             default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
 783         }
 784     }
 785     else
 786     {
 787         /* Pixel shaders 3.0 don't have usage semantics. */
 788         if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
 789         else TRACE("_");
 790
 791         switch (semantic->usage)
 792         {
 793             case WINED3DDECLUSAGE_POSITION:
 794                 TRACE("position%u", semantic->usage_idx);
 795                 break;
 796
 797             case WINED3DDECLUSAGE_BLENDINDICES:
 798                 TRACE("blend");
 799                 break;
 800
 801             case WINED3DDECLUSAGE_BLENDWEIGHT:
 802                 TRACE("weight");
 803                 break;
 804
 805             case WINED3DDECLUSAGE_NORMAL:
 806                 TRACE("normal%u", semantic->usage_idx);
 807                 break;
 808
 809             case WINED3DDECLUSAGE_PSIZE:
 810                 TRACE("psize");
 811                 break;
 812
 813             case WINED3DDECLUSAGE_COLOR:
 814                 if (semantic->usage_idx == 0) TRACE("color");
 815                 else TRACE("specular%u", (semantic->usage_idx - 1));
 816                 break;
 817
 818             case WINED3DDECLUSAGE_TEXCOORD:
 819                 TRACE("texture%u", semantic->usage_idx);
 820                 break;
 821
 822             case WINED3DDECLUSAGE_TANGENT:
 823                 TRACE("tangent");
 824                 break;
 825
 826             case WINED3DDECLUSAGE_BINORMAL:
 827                 TRACE("binormal");
 828                 break;
 829
 830             case WINED3DDECLUSAGE_TESSFACTOR:
 831                 TRACE("tessfactor");
 832                 break;
 833
 834             case WINED3DDECLUSAGE_POSITIONT:
 835                 TRACE("positionT%u", semantic->usage_idx);
 836                 break;
 837
 838             case WINED3DDECLUSAGE_FOG:
 839                 TRACE("fog");
 840                 break;
 841
 842             case WINED3DDECLUSAGE_DEPTH:
 843                 TRACE("depth");
 844                 break;
 845
 846             case WINED3DDECLUSAGE_SAMPLE:
 847                 TRACE("sample");
 848                 break;
 849
 850             default:
 851                 FIXME("unknown_semantics(0x%08x)", semantic->usage);
 852         }
 853     }
 854 }
 855
 856 static void shader_dump_register(const struct wined3d_shader_register *reg,
 857         const struct wined3d_shader_version *shader_version)
 858 {
 859     static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
 860     static const char * const misctype_reg_names[] = {"vPos", "vFace"};
 861     UINT offset = reg->idx;
 862
 863     switch (reg->type)
 864     {
 865         case WINED3DSPR_TEMP:
 866             TRACE("r");
 867             break;
 868
 869         case WINED3DSPR_INPUT:
 870             TRACE("v");
 871             break;
 872
 873         case WINED3DSPR_CONST:
 874         case WINED3DSPR_CONST2:
 875         case WINED3DSPR_CONST3:
 876         case WINED3DSPR_CONST4:
 877             TRACE("c");
 878             offset = shader_get_float_offset(reg->type, reg->idx);
 879             break;
 880
 881         case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
 882             TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
 883             break;
 884
 885         case WINED3DSPR_RASTOUT:
 886             TRACE("%s", rastout_reg_names[reg->idx]);
 887             break;
 888
 889         case WINED3DSPR_COLOROUT:
 890             TRACE("oC");
 891             break;
 892
 893         case WINED3DSPR_DEPTHOUT:
 894             TRACE("oDepth");
 895             break;
 896
 897         case WINED3DSPR_ATTROUT:
 898             TRACE("oD");
 899             break;
 900
 901         case WINED3DSPR_TEXCRDOUT:
 902             /* Vertex shaders >= 3.0 use general purpose output registers
 903              * (WINED3DSPR_OUTPUT), which can include an address token. */
 904             if (shader_version->major >= 3) TRACE("o");
 905             else TRACE("oT");
 906             break;
 907
 908         case WINED3DSPR_CONSTINT:
 909             TRACE("i");
 910             break;
 911
 912         case WINED3DSPR_CONSTBOOL:
 913             TRACE("b");
 914             break;
 915
 916         case WINED3DSPR_LABEL:
 917             TRACE("l");
 918             break;
 919
 920         case WINED3DSPR_LOOP:
 921             TRACE("aL");
 922             break;
 923
 924         case WINED3DSPR_SAMPLER:
 925             TRACE("s");
 926             break;
 927
 928         case WINED3DSPR_MISCTYPE:
 929             if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
 930             else TRACE("%s", misctype_reg_names[reg->idx]);
 931             break;
 932
 933         case WINED3DSPR_PREDICATE:
 934             TRACE("p");
 935             break;
 936
 937         case WINED3DSPR_IMMCONST:
 938             TRACE("l");
 939             break;
 940
 941         case WINED3DSPR_CONSTBUFFER:
 942             TRACE("cb");
 943             break;
 944
 945         default:
 946             TRACE("unhandled_rtype(%#x)", reg->type);
 947             break;
 948     }
 949
 950     if (reg->type == WINED3DSPR_IMMCONST)
 951     {
 952         TRACE("(");
 953         switch (reg->immconst_type)
 954         {
 955             case WINED3D_IMMCONST_FLOAT:
 956                 TRACE("%.8e", *(const float *)reg->immconst_data);
 957                 break;
 958
 959             case WINED3D_IMMCONST_FLOAT4:
 960                 TRACE("%.8e, %.8e, %.8e, %.8e",
 961                         *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
 962                         *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
 963                 break;
 964
 965             default:
 966                 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
 967                 break;
 968         }
 969         TRACE(")");
 970     }
 971     else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
 972     {
 973         if (reg->array_idx != ~0U)
 974         {
 975             TRACE("%u[%u", offset, reg->array_idx);
 976             if (reg->rel_addr)
 977             {
 978                 TRACE(" + ");
 979                 shader_dump_src_param(reg->rel_addr, shader_version);
 980             }
 981             TRACE("]");
 982         }
 983         else
 984         {
 985             if (reg->rel_addr)
 986             {
 987                 TRACE("[");
 988                 shader_dump_src_param(reg->rel_addr, shader_version);
 989                 TRACE(" + ");
 990             }
 991             TRACE("%u", offset);
 992             if (reg->rel_addr) TRACE("]");
 993         }
 994     }
 995 }
 996
 997 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
 998         const struct wined3d_shader_version *shader_version)
 999 {
1000     DWORD write_mask = param->write_mask;
1001
1002     shader_dump_register(&param->reg, shader_version);
1003
1004     if (write_mask != WINED3DSP_WRITEMASK_ALL)
1005     {
1006         static const char *write_mask_chars = "xyzw";
1007
1008         TRACE(".");
1009         if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1010         if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1011         if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1012         if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1013     }
1014 }
1015
1016 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1017         const struct wined3d_shader_version *shader_version)
1018 {
1019     DWORD src_modifier = param->modifiers;
1020     DWORD swizzle = param->swizzle;
1021
1022     if (src_modifier == WINED3DSPSM_NEG
1023             || src_modifier == WINED3DSPSM_BIASNEG
1024             || src_modifier == WINED3DSPSM_SIGNNEG
1025             || src_modifier == WINED3DSPSM_X2NEG
1026             || src_modifier == WINED3DSPSM_ABSNEG)
1027         TRACE("-");
1028     else if (src_modifier == WINED3DSPSM_COMP)
1029         TRACE("1-");
1030     else if (src_modifier == WINED3DSPSM_NOT)
1031         TRACE("!");
1032
1033     if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1034         TRACE("abs(");
1035
1036     shader_dump_register(&param->reg, shader_version);
1037
1038     if (src_modifier)
1039     {
1040         switch (src_modifier)
1041         {
1042             case WINED3DSPSM_NONE:    break;
1043             case WINED3DSPSM_NEG:     break;
1044             case WINED3DSPSM_NOT:     break;
1045             case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
1046             case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1047             case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
1048             case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1049             case WINED3DSPSM_COMP:    break;
1050             case WINED3DSPSM_X2:      TRACE("_x2"); break;
1051             case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
1052             case WINED3DSPSM_DZ:      TRACE("_dz"); break;
1053             case WINED3DSPSM_DW:      TRACE("_dw"); break;
1054             case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
1055             case WINED3DSPSM_ABS:     TRACE(")"); break;
1056             default:                  TRACE("_unknown_modifier(%#x)", src_modifier);
1057         }
1058     }
1059
1060     if (swizzle != WINED3DSP_NOSWIZZLE)
1061     {
1062         static const char *swizzle_chars = "xyzw";
1063         DWORD swizzle_x = swizzle & 0x03;
1064         DWORD swizzle_y = (swizzle >> 2) & 0x03;
1065         DWORD swizzle_z = (swizzle >> 4) & 0x03;
1066         DWORD swizzle_w = (swizzle >> 6) & 0x03;
1067
1068         if (swizzle_x == swizzle_y
1069                 && swizzle_x == swizzle_z
1070                 && swizzle_x == swizzle_w)
1071         {
1072             TRACE(".%c", swizzle_chars[swizzle_x]);
1073         }
1074         else
1075         {
1076             TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1077                     swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1078         }
1079     }
1080 }
1081
1082 /* Shared code in order to generate the bulk of the shader string.
1083  * NOTE: A description of how to parse tokens can be found on MSDN. */
1084 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1085         const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1086 {
1087     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1088     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1089     const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1090     void *fe_data = shader->baseShader.frontend_data;
1091     struct wined3d_shader_src_param src_rel_addr[4];
1092     struct wined3d_shader_src_param src_param[4];
1093     struct wined3d_shader_version shader_version;
1094     struct wined3d_shader_src_param dst_rel_addr;
1095     struct wined3d_shader_dst_param dst_param;
1096     struct wined3d_shader_instruction ins;
1097     struct wined3d_shader_context ctx;
1098     const DWORD *ptr = byte_code;
1099     DWORD i;
1100
1101     /* Initialize current parsing state. */
1102     ctx.shader = iface;
1103     ctx.reg_maps = reg_maps;
1104     ctx.buffer = buffer;
1105     ctx.backend_data = backend_ctx;
1106
1107     ins.ctx = &ctx;
1108     ins.dst = &dst_param;
1109     ins.src = src_param;
1110     shader->baseShader.parse_state.current_row = 0;
1111
1112     fe->shader_read_header(fe_data, &ptr, &shader_version);
1113
1114     while (!fe->shader_is_end(fe_data, &ptr))
1115     {
1116         const char *comment;
1117         UINT param_size;
1118
1119         /* Skip comment tokens. */
1120         fe->shader_read_comment(&ptr, &comment);
1121         if (comment) continue;
1122
1123         /* Read opcode. */
1124         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1125
1126         /* Unknown opcode and its parameters. */
1127         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1128         {
1129             TRACE("Skipping unrecognized instruction.\n");
1130             ptr += param_size;
1131             continue;
1132         }
1133
1134         /* Nothing to do. */
1135         if (ins.handler_idx == WINED3DSIH_DCL
1136                 || ins.handler_idx == WINED3DSIH_NOP
1137                 || ins.handler_idx == WINED3DSIH_DEF
1138                 || ins.handler_idx == WINED3DSIH_DEFI
1139                 || ins.handler_idx == WINED3DSIH_DEFB
1140                 || ins.handler_idx == WINED3DSIH_PHASE)
1141         {
1142             ptr += param_size;
1143             continue;
1144         }
1145
1146         /* Destination token */
1147         if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1148
1149         /* Predication token */
1150         if (ins.predicate) ins.predicate = *ptr++;
1151
1152         /* Other source tokens */
1153         for (i = 0; i < ins.src_count; ++i)
1154         {
1155             fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1156         }
1157
1158         /* Call appropriate function for output target */
1159         device->shader_backend->shader_handle_instruction(&ins);
1160     }
1161 }
1162
1163 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1164 {
1165     DWORD mmask = dst->modifiers;
1166
1167     switch (dst->shift)
1168     {
1169         case 0: break;
1170         case 13: TRACE("_d8"); break;
1171         case 14: TRACE("_d4"); break;
1172         case 15: TRACE("_d2"); break;
1173         case 1: TRACE("_x2"); break;
1174         case 2: TRACE("_x4"); break;
1175         case 3: TRACE("_x8"); break;
1176         default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1177     }
1178
1179     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
1180     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1181     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
1182
1183     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1184     if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1185 }
1186
1187 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1188 {
1189     struct wined3d_shader_version shader_version;
1190     const DWORD *ptr = byte_code;
1191     const char *type_prefix;
1192     DWORD i;
1193
1194     TRACE("Parsing %p.\n", byte_code);
1195
1196     fe->shader_read_header(fe_data, &ptr, &shader_version);
1197
1198     switch (shader_version.type)
1199     {
1200         case WINED3D_SHADER_TYPE_VERTEX:
1201             type_prefix = "vs";
1202             break;
1203
1204         case WINED3D_SHADER_TYPE_GEOMETRY:
1205             type_prefix = "gs";
1206             break;
1207
1208         case WINED3D_SHADER_TYPE_PIXEL:
1209             type_prefix = "ps";
1210             break;
1211
1212         default:
1213             FIXME("Unhandled shader type %#x.\n", shader_version.type);
1214             type_prefix = "unknown";
1215             break;
1216     }
1217
1218     TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1219
1220     while (!fe->shader_is_end(fe_data, &ptr))
1221     {
1222         struct wined3d_shader_instruction ins;
1223         const char *comment;
1224         UINT param_size;
1225
1226         /* comment */
1227         fe->shader_read_comment(&ptr, &comment);
1228         if (comment)
1229         {
1230             TRACE("//%s\n", comment);
1231             continue;
1232         }
1233
1234         fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1235         if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1236         {
1237             TRACE("Skipping unrecognized instruction.\n");
1238             ptr += param_size;
1239             continue;
1240         }
1241
1242         if (ins.handler_idx == WINED3DSIH_DCL)
1243         {
1244             struct wined3d_shader_semantic semantic;
1245
1246             fe->shader_read_semantic(&ptr, &semantic);
1247
1248             shader_dump_decl_usage(&semantic, &shader_version);
1249             shader_dump_ins_modifiers(&semantic.reg);
1250             TRACE(" ");
1251             shader_dump_dst_param(&semantic.reg, &shader_version);
1252         }
1253         else if (ins.handler_idx == WINED3DSIH_DEF)
1254         {
1255             struct wined3d_shader_src_param rel_addr;
1256             struct wined3d_shader_dst_param dst;
1257
1258             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1259
1260             TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1261                     *(const float *)(ptr),
1262                     *(const float *)(ptr + 1),
1263                     *(const float *)(ptr + 2),
1264                     *(const float *)(ptr + 3));
1265             ptr += 4;
1266         }
1267         else if (ins.handler_idx == WINED3DSIH_DEFI)
1268         {
1269             struct wined3d_shader_src_param rel_addr;
1270             struct wined3d_shader_dst_param dst;
1271
1272             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1273
1274             TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1275                     *(ptr),
1276                     *(ptr + 1),
1277                     *(ptr + 2),
1278                     *(ptr + 3));
1279             ptr += 4;
1280         }
1281         else if (ins.handler_idx == WINED3DSIH_DEFB)
1282         {
1283             struct wined3d_shader_src_param rel_addr;
1284             struct wined3d_shader_dst_param dst;
1285
1286             fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1287
1288             TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1289             ++ptr;
1290         }
1291         else
1292         {
1293             struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1294             struct wined3d_shader_dst_param dst_param;
1295             struct wined3d_shader_src_param src_param;
1296
1297             if (ins.dst_count)
1298             {
1299                 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1300             }
1301
1302             /* Print out predication source token first - it follows
1303              * the destination token. */
1304             if (ins.predicate)
1305             {
1306                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1307                 TRACE("(");
1308                 shader_dump_src_param(&src_param, &shader_version);
1309                 TRACE(") ");
1310             }
1311
1312             /* PixWin marks instructions with the coissue flag with a '+' */
1313             if (ins.coissue) TRACE("+");
1314
1315             TRACE("%s", shader_opcode_names[ins.handler_idx]);
1316
1317             if (ins.handler_idx == WINED3DSIH_IFC
1318                     || ins.handler_idx == WINED3DSIH_BREAKC)
1319             {
1320                 switch (ins.flags)
1321                 {
1322                     case COMPARISON_GT: TRACE("_gt"); break;
1323                     case COMPARISON_EQ: TRACE("_eq"); break;
1324                     case COMPARISON_GE: TRACE("_ge"); break;
1325                     case COMPARISON_LT: TRACE("_lt"); break;
1326                     case COMPARISON_NE: TRACE("_ne"); break;
1327                     case COMPARISON_LE: TRACE("_le"); break;
1328                     default: TRACE("_(%u)", ins.flags);
1329                 }
1330             }
1331             else if (ins.handler_idx == WINED3DSIH_TEX
1332                     && shader_version.major >= 2
1333                     && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1334             {
1335                 TRACE("p");
1336             }
1337
1338             /* We already read the destination token, print it. */
1339             if (ins.dst_count)
1340             {
1341                 shader_dump_ins_modifiers(&dst_param);
1342                 TRACE(" ");
1343                 shader_dump_dst_param(&dst_param, &shader_version);
1344             }
1345
1346             /* Other source tokens */
1347             for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1348             {
1349                 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1350                 TRACE(!i ? " " : ", ");
1351                 shader_dump_src_param(&src_param, &shader_version);
1352             }
1353         }
1354         TRACE("\n");
1355     }
1356 }
1357
1358 static void shader_cleanup(IWineD3DBaseShader *iface)
1359 {
1360     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1361
1362     ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1363     HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1364     HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1365     shader_delete_constant_list(&shader->baseShader.constantsF);
1366     shader_delete_constant_list(&shader->baseShader.constantsB);
1367     shader_delete_constant_list(&shader->baseShader.constantsI);
1368     list_remove(&shader->baseShader.shader_list_entry);
1369
1370     if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1371     {
1372         shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1373     }
1374 }
1375
1376 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1377 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1378 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1379 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1380 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1381 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1382 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1383 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1384 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1385 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1386 static void shader_none_free(IWineD3DDevice *iface) {}
1387 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1388
1389 static void shader_none_get_caps(WINED3DDEVTYPE devtype,
1390         const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1391 {
1392     /* Set the shader caps to 0 for the none shader backend */
1393     caps->VertexShaderVersion = 0;
1394     caps->PixelShaderVersion = 0;
1395     caps->PixelShader1xMaxValue = 0.0f;
1396 }
1397
1398 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1399 {
1400     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1401     {
1402         TRACE("Checking support for fixup:\n");
1403         dump_color_fixup_desc(fixup);
1404     }
1405
1406     /* Faked to make some apps happy. */
1407     if (!is_yuv_fixup(fixup))
1408     {
1409         TRACE("[OK]\n");
1410         return TRUE;
1411     }
1412
1413     TRACE("[FAILED]\n");
1414     return FALSE;
1415 }
1416
1417 const shader_backend_t none_shader_backend = {
1418     shader_none_handle_instruction,
1419     shader_none_select,
1420     shader_none_select_depth_blt,
1421     shader_none_deselect_depth_blt,
1422     shader_none_update_float_vertex_constants,
1423     shader_none_update_float_pixel_constants,
1424     shader_none_load_constants,
1425     shader_none_load_np2fixup_constants,
1426     shader_none_destroy,
1427     shader_none_alloc,
1428     shader_none_free,
1429     shader_none_dirty_const,
1430     shader_none_get_caps,
1431     shader_none_color_fixup_supported,
1432 };
1433
1434 static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1435 {
1436     *parent = shader->baseShader.parent;
1437     IUnknown_AddRef(*parent);
1438     TRACE("shader %p, returning %p.\n", shader, *parent);
1439 }
1440
1441 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1442 {
1443     if (!data)
1444     {
1445         *data_size = shader->baseShader.functionLength;
1446         return WINED3D_OK;
1447     }
1448
1449     if (*data_size < shader->baseShader.functionLength)
1450     {
1451         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1452          * than the required size we should write the required size and
1453          * return D3DERR_MOREDATA. That's not actually true. */
1454         return WINED3DERR_INVALIDCALL;
1455     }
1456
1457     memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1458
1459     return WINED3D_OK;
1460 }
1461
1462 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1463         const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1464 {
1465     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1466     const struct wined3d_shader_frontend *fe;
1467     HRESULT hr;
1468
1469     TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1470             shader, byte_code, output_signature, float_const_count);
1471
1472     fe = shader_select_frontend(*byte_code);
1473     if (!fe)
1474     {
1475         FIXME("Unable to find frontend for shader.\n");
1476         return WINED3DERR_INVALIDCALL;
1477     }
1478     shader->baseShader.frontend = fe;
1479     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1480     if (!shader->baseShader.frontend_data)
1481     {
1482         FIXME("Failed to initialize frontend.\n");
1483         return WINED3DERR_INVALIDCALL;
1484     }
1485
1486     /* First pass: trace shader. */
1487     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1488
1489     /* Initialize immediate constant lists. */
1490     list_init(&shader->baseShader.constantsF);
1491     list_init(&shader->baseShader.constantsB);
1492     list_init(&shader->baseShader.constantsI);
1493
1494     /* Second pass: figure out which registers are used, what the semantics are, etc. */
1495     hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1496             reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1497             byte_code, float_const_count);
1498     if (FAILED(hr)) return hr;
1499
1500     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1501     if (!shader->baseShader.function) return E_OUTOFMEMORY;
1502     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1503
1504     return WINED3D_OK;
1505 }
1506
1507 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1508 {
1509     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1510
1511     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1512             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1513             || IsEqualGUID(riid, &IID_IWineD3DBase)
1514             || IsEqualGUID(riid, &IID_IUnknown))
1515     {
1516         IUnknown_AddRef(iface);
1517         *object = iface;
1518         return S_OK;
1519     }
1520
1521     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1522
1523     *object = NULL;
1524     return E_NOINTERFACE;
1525 }
1526
1527 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1528 {
1529     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1530     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1531
1532     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1533
1534     return refcount;
1535 }
1536
1537 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1538 {
1539     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1540     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1541
1542     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1543
1544     if (!refcount)
1545     {
1546         shader_cleanup((IWineD3DBaseShader *)iface);
1547         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1548         HeapFree(GetProcessHeap(), 0, shader);
1549     }
1550
1551     return refcount;
1552 }
1553
1554 static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1555 {
1556     TRACE("iface %p, parent %p.\n", iface, parent);
1557
1558     shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1559
1560     return WINED3D_OK;
1561 }
1562
1563 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1564 {
1565     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1566
1567     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1568 }
1569
1570 /* Set local constants for d3d8 shaders. */
1571 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1572         UINT start_idx, const float *src_data, UINT count)
1573 {
1574     IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1575     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1576     UINT i, end_idx;
1577
1578     TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1579
1580     end_idx = start_idx + count;
1581     if (end_idx > device->d3d_vshader_constantF)
1582     {
1583         WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1584         end_idx = device->d3d_vshader_constantF;
1585     }
1586
1587     for (i = start_idx; i < end_idx; ++i)
1588     {
1589         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1590         if (!lconst) return E_OUTOFMEMORY;
1591
1592         lconst->idx = i;
1593         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1594         list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1595     }
1596
1597     return WINED3D_OK;
1598 }
1599
1600 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1601 {
1602     /* IUnknown methods */
1603     vertexshader_QueryInterface,
1604     vertexshader_AddRef,
1605     vertexshader_Release,
1606     /* IWineD3DBase methods */
1607     vertexshader_GetParent,
1608     /* IWineD3DBaseShader methods */
1609     vertexshader_GetFunction,
1610     /* IWineD3DVertexShader methods */
1611     vertexshader_SetLocalConstantsF,
1612 };
1613
1614 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1615         IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1616 {
1617     args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1618     args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1619             && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1620     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1621 }
1622
1623 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1624 {
1625     if (usage_idx1 != usage_idx2) return FALSE;
1626     if (usage1 == usage2) return TRUE;
1627     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1628     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1629
1630     return FALSE;
1631 }
1632
1633 BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1634 {
1635     IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1636     WORD map = shader->baseShader.reg_maps.input_registers;
1637     unsigned int i;
1638
1639     for (i = 0; map; map >>= 1, ++i)
1640     {
1641         if (!(map & 1)) continue;
1642
1643         if (match_usage(shader->attributes[i].usage,
1644                 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1645         {
1646             *regnum = i;
1647             return TRUE;
1648         }
1649     }
1650     return FALSE;
1651 }
1652
1653 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1654 {
1655     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1656             shader->baseShader.reg_maps.shader_version.minor);
1657     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1658
1659     shader->baseShader.limits.texcoord = 0;
1660     shader->baseShader.limits.attributes = 16;
1661     shader->baseShader.limits.packed_input = 0;
1662
1663     switch (shader_version)
1664     {
1665         case WINED3D_SHADER_VERSION(1, 0):
1666         case WINED3D_SHADER_VERSION(1, 1):
1667             shader->baseShader.limits.temporary = 12;
1668             shader->baseShader.limits.constant_bool = 0;
1669             shader->baseShader.limits.constant_int = 0;
1670             shader->baseShader.limits.address = 1;
1671             shader->baseShader.limits.packed_output = 0;
1672             shader->baseShader.limits.sampler = 0;
1673             shader->baseShader.limits.label = 0;
1674             /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1675              * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1676              * constants? */
1677             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1678             break;
1679
1680         case WINED3D_SHADER_VERSION(2, 0):
1681         case WINED3D_SHADER_VERSION(2, 1):
1682             shader->baseShader.limits.temporary = 12;
1683             shader->baseShader.limits.constant_bool = 16;
1684             shader->baseShader.limits.constant_int = 16;
1685             shader->baseShader.limits.address = 1;
1686             shader->baseShader.limits.packed_output = 0;
1687             shader->baseShader.limits.sampler = 0;
1688             shader->baseShader.limits.label = 16;
1689             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1690             break;
1691
1692         case WINED3D_SHADER_VERSION(4, 0):
1693             FIXME("Using 3.0 limits for 4.0 shader.\n");
1694             /* Fall through. */
1695
1696         case WINED3D_SHADER_VERSION(3, 0):
1697             shader->baseShader.limits.temporary = 32;
1698             shader->baseShader.limits.constant_bool = 32;
1699             shader->baseShader.limits.constant_int = 32;
1700             shader->baseShader.limits.address = 1;
1701             shader->baseShader.limits.packed_output = 12;
1702             shader->baseShader.limits.sampler = 4;
1703             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1704             /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1705              * even though they are capable of supporting much more (GL
1706              * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1707              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1708              * shaders to 256. */
1709             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1710             break;
1711
1712         default:
1713             shader->baseShader.limits.temporary = 12;
1714             shader->baseShader.limits.constant_bool = 16;
1715             shader->baseShader.limits.constant_int = 16;
1716             shader->baseShader.limits.address = 1;
1717             shader->baseShader.limits.packed_output = 0;
1718             shader->baseShader.limits.sampler = 0;
1719             shader->baseShader.limits.label = 16;
1720             shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1721             FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1722                     shader->baseShader.reg_maps.shader_version.major,
1723                     shader->baseShader.reg_maps.shader_version.minor);
1724     }
1725 }
1726
1727 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1728         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1729         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1730 {
1731     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1732     struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1733     unsigned int i;
1734     HRESULT hr;
1735     WORD map;
1736
1737     if (!byte_code) return WINED3DERR_INVALIDCALL;
1738
1739     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1740     shader_init(&shader->baseShader, device, parent, parent_ops);
1741
1742     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1743             output_signature, device->d3d_vshader_constantF);
1744     if (FAILED(hr))
1745     {
1746         WARN("Failed to set function, hr %#x.\n", hr);
1747         shader_cleanup((IWineD3DBaseShader *)shader);
1748         return hr;
1749     }
1750
1751     map = shader->baseShader.reg_maps.input_registers;
1752     for (i = 0; map; map >>= 1, ++i)
1753     {
1754         if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1755
1756         shader->attributes[i].usage =
1757                 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1758         shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1759     }
1760
1761     if (output_signature)
1762     {
1763         for (i = 0; i < output_signature->element_count; ++i)
1764         {
1765             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1766             reg_maps->output_registers |= 1 << e->register_idx;
1767             shader->baseShader.output_signature[e->register_idx] = *e;
1768         }
1769     }
1770
1771     vertexshader_set_limits(shader);
1772
1773     if (device->vs_selected_mode == SHADER_ARB
1774             && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1775             && shader->min_rel_offset <= shader->max_rel_offset)
1776     {
1777         if (shader->max_rel_offset - shader->min_rel_offset > 127)
1778         {
1779             FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1780             FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1781             FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1782         }
1783         else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1784         {
1785             shader->rel_offset = shader->min_rel_offset + 63;
1786         }
1787         else if (shader->max_rel_offset > 63)
1788         {
1789             shader->rel_offset = shader->min_rel_offset;
1790         }
1791         else
1792         {
1793             shader->rel_offset = 0;
1794         }
1795     }
1796
1797     shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1798             && !list_empty(&shader->baseShader.constantsF);
1799
1800     return WINED3D_OK;
1801 }
1802
1803 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1804         REFIID riid, void **object)
1805 {
1806     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1807
1808     if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1809             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1810             || IsEqualGUID(riid, &IID_IWineD3DBase)
1811             || IsEqualGUID(riid, &IID_IUnknown))
1812     {
1813         IUnknown_AddRef(iface);
1814         *object = iface;
1815         return S_OK;
1816     }
1817
1818     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1819
1820     *object = NULL;
1821     return E_NOINTERFACE;
1822 }
1823
1824 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1825 {
1826     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1827     ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1828
1829     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1830
1831     return refcount;
1832 }
1833
1834 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1835 {
1836     struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1837     ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1838
1839     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1840
1841     if (!refcount)
1842     {
1843         shader_cleanup((IWineD3DBaseShader *)iface);
1844         shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1845         HeapFree(GetProcessHeap(), 0, shader);
1846     }
1847
1848     return refcount;
1849 }
1850
1851 static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1852 {
1853     TRACE("iface %p, parent %p.\n", iface, parent);
1854
1855     shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1856
1857     return WINED3D_OK;
1858 }
1859
1860 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1861 {
1862     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1863
1864     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1865 }
1866
1867 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1868 {
1869     /* IUnknown methods */
1870     geometryshader_QueryInterface,
1871     geometryshader_AddRef,
1872     geometryshader_Release,
1873     /* IWineD3DBase methods */
1874     geometryshader_GetParent,
1875     /* IWineD3DBaseShader methods */
1876     geometryshader_GetFunction,
1877 };
1878
1879 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1880         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1881         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1882 {
1883     HRESULT hr;
1884
1885     shader->vtbl = &wined3d_geometryshader_vtbl;
1886     shader_init(&shader->base_shader, device, parent, parent_ops);
1887
1888     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1889     if (FAILED(hr))
1890     {
1891         WARN("Failed to set function, hr %#x.\n", hr);
1892         shader_cleanup((IWineD3DBaseShader *)shader);
1893         return hr;
1894     }
1895
1896     shader->base_shader.load_local_constsF = FALSE;
1897
1898     return WINED3D_OK;
1899 }
1900
1901 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1902 {
1903     TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1904
1905     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1906             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1907             || IsEqualGUID(riid, &IID_IWineD3DBase)
1908             || IsEqualGUID(riid, &IID_IUnknown))
1909     {
1910         IUnknown_AddRef(iface);
1911         *object = iface;
1912         return S_OK;
1913     }
1914
1915     WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1916
1917     *object = NULL;
1918     return E_NOINTERFACE;
1919 }
1920
1921 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
1922 {
1923     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1924     ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1925
1926     TRACE("%p increasing refcount to %u.\n", shader, refcount);
1927
1928     return refcount;
1929 }
1930
1931 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
1932 {
1933     IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1934     ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1935
1936     TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1937
1938     if (!refcount)
1939     {
1940         shader_cleanup((IWineD3DBaseShader *)iface);
1941         shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1942         HeapFree(GetProcessHeap(), 0, shader);
1943     }
1944
1945     return refcount;
1946 }
1947
1948 static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
1949 {
1950     TRACE("iface %p, parent %p.\n", iface, parent);
1951
1952     shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1953
1954     return WINED3D_OK;
1955 }
1956
1957 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
1958 {
1959     TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1960
1961     return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1962 }
1963
1964 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
1965 {
1966     /* IUnknown methods */
1967     pixelshader_QueryInterface,
1968     pixelshader_AddRef,
1969     pixelshader_Release,
1970     /* IWineD3DBase methods */
1971     pixelshader_GetParent,
1972     /* IWineD3DBaseShader methods */
1973     pixelshader_GetFunction
1974 };
1975
1976 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
1977         IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
1978 {
1979     IWineD3DBaseTextureImpl *texture;
1980     UINT i;
1981
1982     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1983     args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
1984     args->np2_fixup = 0;
1985
1986     for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1987     {
1988         if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
1989         texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
1990         if (!texture)
1991         {
1992             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1993             continue;
1994         }
1995         args->color_fixup[i] = texture->resource.format_desc->color_fixup;
1996
1997         /* Flag samplers that need NP2 texcoord fixup. */
1998         if (!texture->baseTexture.pow2Matrix_identity)
1999         {
2000             args->np2_fixup |= (1 << i);
2001         }
2002     }
2003     if (shader->baseShader.reg_maps.shader_version.major >= 3)
2004     {
2005         if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2006         {
2007             args->vp_mode = pretransformed;
2008         }
2009         else if (use_vs(stateblock))
2010         {
2011             args->vp_mode = vertexshader;
2012         }
2013         else
2014         {
2015             args->vp_mode = fixedfunction;
2016         }
2017         args->fog = FOG_OFF;
2018     }
2019     else
2020     {
2021         args->vp_mode = vertexshader;
2022         if (stateblock->renderState[WINED3DRS_FOGENABLE])
2023         {
2024             switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2025             {
2026                 case WINED3DFOG_NONE:
2027                     if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2028                             || use_vs(stateblock))
2029                     {
2030                         args->fog = FOG_LINEAR;
2031                         break;
2032                     }
2033
2034                     switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2035                     {
2036                         case WINED3DFOG_NONE: /* Fall through. */
2037                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2038                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2039                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2040                     }
2041                     break;
2042
2043                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2044                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
2045                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
2046             }
2047         }
2048         else
2049         {
2050             args->fog = FOG_OFF;
2051         }
2052     }
2053 }
2054
2055 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2056 {
2057     DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2058             shader->baseShader.reg_maps.shader_version.minor);
2059
2060     shader->baseShader.limits.attributes = 0;
2061     shader->baseShader.limits.address = 0;
2062     shader->baseShader.limits.packed_output = 0;
2063
2064     switch (shader_version)
2065     {
2066         case WINED3D_SHADER_VERSION(1, 0):
2067         case WINED3D_SHADER_VERSION(1, 1):
2068         case WINED3D_SHADER_VERSION(1, 2):
2069         case WINED3D_SHADER_VERSION(1, 3):
2070             shader->baseShader.limits.temporary = 2;
2071             shader->baseShader.limits.constant_float = 8;
2072             shader->baseShader.limits.constant_int = 0;
2073             shader->baseShader.limits.constant_bool = 0;
2074             shader->baseShader.limits.texcoord = 4;
2075             shader->baseShader.limits.sampler = 4;
2076             shader->baseShader.limits.packed_input = 0;
2077             shader->baseShader.limits.label = 0;
2078             break;
2079
2080         case WINED3D_SHADER_VERSION(1, 4):
2081             shader->baseShader.limits.temporary = 6;
2082             shader->baseShader.limits.constant_float = 8;
2083             shader->baseShader.limits.constant_int = 0;
2084             shader->baseShader.limits.constant_bool = 0;
2085             shader->baseShader.limits.texcoord = 6;
2086             shader->baseShader.limits.sampler = 6;
2087             shader->baseShader.limits.packed_input = 0;
2088             shader->baseShader.limits.label = 0;
2089             break;
2090
2091         /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2092         case WINED3D_SHADER_VERSION(2, 0):
2093             shader->baseShader.limits.temporary = 32;
2094             shader->baseShader.limits.constant_float = 32;
2095             shader->baseShader.limits.constant_int = 16;
2096             shader->baseShader.limits.constant_bool = 16;
2097             shader->baseShader.limits.texcoord = 8;
2098             shader->baseShader.limits.sampler = 16;
2099             shader->baseShader.limits.packed_input = 0;
2100             break;
2101
2102         case WINED3D_SHADER_VERSION(2, 1):
2103             shader->baseShader.limits.temporary = 32;
2104             shader->baseShader.limits.constant_float = 32;
2105             shader->baseShader.limits.constant_int = 16;
2106             shader->baseShader.limits.constant_bool = 16;
2107             shader->baseShader.limits.texcoord = 8;
2108             shader->baseShader.limits.sampler = 16;
2109             shader->baseShader.limits.packed_input = 0;
2110             shader->baseShader.limits.label = 16;
2111             break;
2112
2113         case WINED3D_SHADER_VERSION(4, 0):
2114             FIXME("Using 3.0 limits for 4.0 shader.\n");
2115             /* Fall through. */
2116
2117         case WINED3D_SHADER_VERSION(3, 0):
2118             shader->baseShader.limits.temporary = 32;
2119             shader->baseShader.limits.constant_float = 224;
2120             shader->baseShader.limits.constant_int = 16;
2121             shader->baseShader.limits.constant_bool = 16;
2122             shader->baseShader.limits.texcoord = 0;
2123             shader->baseShader.limits.sampler = 16;
2124             shader->baseShader.limits.packed_input = 12;
2125             shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2126             break;
2127
2128         default:
2129             shader->baseShader.limits.temporary = 32;
2130             shader->baseShader.limits.constant_float = 32;
2131             shader->baseShader.limits.constant_int = 16;
2132             shader->baseShader.limits.constant_bool = 16;
2133             shader->baseShader.limits.texcoord = 8;
2134             shader->baseShader.limits.sampler = 16;
2135             shader->baseShader.limits.packed_input = 0;
2136             shader->baseShader.limits.label = 0;
2137             FIXME("Unrecognized pixel shader version %u.%u\n",
2138                     shader->baseShader.reg_maps.shader_version.major,
2139                     shader->baseShader.reg_maps.shader_version.minor);
2140     }
2141 }
2142
2143 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2144         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2145         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2146 {
2147     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2148     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2149     HRESULT hr;
2150
2151     if (!byte_code) return WINED3DERR_INVALIDCALL;
2152
2153     shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2154     shader_init(&shader->baseShader, device, parent, parent_ops);
2155
2156     hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2157             output_signature, device->d3d_pshader_constantF);
2158     if (FAILED(hr))
2159     {
2160         WARN("Failed to set function, hr %#x.\n", hr);
2161         shader_cleanup((IWineD3DBaseShader *)shader);
2162         return hr;
2163     }
2164
2165     pixelshader_set_limits(shader);
2166
2167     for (i = 0; i < MAX_REG_INPUT; ++i)
2168     {
2169         if (shader->input_reg_used[i])
2170         {
2171             ++num_regs_used;
2172             highest_reg_used = i;
2173         }
2174     }
2175
2176     /* Don't do any register mapping magic if it is not needed, or if we can't
2177      * achieve anything anyway */
2178     if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2179             || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2180     {
2181         if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2182         {
2183             /* This happens with relative addressing. The input mapper function
2184              * warns about this if the higher registers are declared too, so
2185              * don't write a FIXME here */
2186             WARN("More varying registers used than supported\n");
2187         }
2188
2189         for (i = 0; i < MAX_REG_INPUT; ++i)
2190         {
2191             shader->input_reg_map[i] = i;
2192         }
2193
2194         shader->declared_in_count = highest_reg_used + 1;
2195     }
2196     else
2197     {
2198         shader->declared_in_count = 0;
2199         for (i = 0; i < MAX_REG_INPUT; ++i)
2200         {
2201             if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2202             else shader->input_reg_map[i] = ~0U;
2203         }
2204     }
2205
2206     shader->baseShader.load_local_constsF = FALSE;
2207
2208     return WINED3D_OK;
2209 }
2210
2211 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2212 {
2213     WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2214     unsigned int i;
2215
2216     if (reg_maps->shader_version.major != 1) return;
2217
2218     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2219     {
2220         /* We don't sample from this sampler. */
2221         if (!sampler_type[i]) continue;
2222
2223         if (!textures[i])
2224         {
2225             WARN("No texture bound to sampler %u, using 2D.\n", i);
2226             sampler_type[i] = WINED3DSTT_2D;
2227             continue;
2228         }
2229
2230         switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2231         {
2232             case GL_TEXTURE_RECTANGLE_ARB:
2233             case GL_TEXTURE_2D:
2234                 /* We have to select between texture rectangles and 2D
2235                  * textures later because 2.0 and 3.0 shaders only have
2236                  * WINED3DSTT_2D as well. */
2237                 sampler_type[i] = WINED3DSTT_2D;
2238                 break;
2239
2240             case GL_TEXTURE_3D:
2241                 sampler_type[i] = WINED3DSTT_VOLUME;
2242                 break;
2243
2244             case GL_TEXTURE_CUBE_MAP_ARB:
2245                 sampler_type[i] = WINED3DSTT_CUBE;
2246                 break;
2247
2248             default:
2249                 FIXME("Unrecognized texture type %#x, using 2D.\n",
2250                         IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2251                 sampler_type[i] = WINED3DSTT_2D;
2252         }
2253     }
2254 }