?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
shell32: Replaced drive.ico with a Tango compliant icon.
[wine.git] / dlls / wined3d / shader.c
blob683653245daa0a6e6c741ab33ab8aefb3d9087d8
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)
1154 {
1155 FIXME("Predicates not implemented.\n");
1156 ins.predicate = *ptr++;
1157 }
1158
1159 /* Other source tokens */
1160 for (i = 0; i < ins.src_count; ++i)
1161 {
1162 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1163 }
1164
1165 /* Call appropriate function for output target */
1166 device->shader_backend->shader_handle_instruction(&ins);
1167 }
1168 }
1169
1170 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1171 {
1172 DWORD mmask = dst->modifiers;
1173
1174 switch (dst->shift)
1175 {
1176 case 0: break;
1177 case 13: TRACE("_d8"); break;
1178 case 14: TRACE("_d4"); break;
1179 case 15: TRACE("_d2"); break;
1180 case 1: TRACE("_x2"); break;
1181 case 2: TRACE("_x4"); break;
1182 case 3: TRACE("_x8"); break;
1183 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1184 }
1185
1186 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1187 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1188 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1189
1190 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1191 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1192 }
1193
1194 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1195 {
1196 struct wined3d_shader_version shader_version;
1197 const DWORD *ptr = byte_code;
1198 const char *type_prefix;
1199 DWORD i;
1200
1201 TRACE("Parsing %p.\n", byte_code);
1202
1203 fe->shader_read_header(fe_data, &ptr, &shader_version);
1204
1205 switch (shader_version.type)
1206 {
1207 case WINED3D_SHADER_TYPE_VERTEX:
1208 type_prefix = "vs";
1209 break;
1210
1211 case WINED3D_SHADER_TYPE_GEOMETRY:
1212 type_prefix = "gs";
1213 break;
1214
1215 case WINED3D_SHADER_TYPE_PIXEL:
1216 type_prefix = "ps";
1217 break;
1218
1219 default:
1220 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1221 type_prefix = "unknown";
1222 break;
1223 }
1224
1225 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1226
1227 while (!fe->shader_is_end(fe_data, &ptr))
1228 {
1229 struct wined3d_shader_instruction ins;
1230 const char *comment;
1231 UINT comment_size;
1232 UINT param_size;
1233
1234 /* comment */
1235 fe->shader_read_comment(&ptr, &comment, &comment_size);
1236 if (comment)
1237 {
1238 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1239 {
1240 const char *end = comment + comment_size;
1241 const char *ptr = comment + 4;
1242 const char *line = ptr;
1243
1244 TRACE("// TEXT\n");
1245 while (ptr != end)
1246 {
1247 if (*ptr == '\n')
1248 {
1249 UINT len = ptr - line;
1250 if (len && *(ptr - 1) == '\r') --len;
1251 TRACE("// %s\n", debugstr_an(line, len));
1252 line = ++ptr;
1253 }
1254 else ++ptr;
1255 }
1256 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1257 }
1258 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1259 continue;
1260 }
1261
1262 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1263 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1264 {
1265 TRACE("Skipping unrecognized instruction.\n");
1266 ptr += param_size;
1267 continue;
1268 }
1269
1270 if (ins.handler_idx == WINED3DSIH_DCL)
1271 {
1272 struct wined3d_shader_semantic semantic;
1273
1274 fe->shader_read_semantic(&ptr, &semantic);
1275
1276 shader_dump_decl_usage(&semantic, &shader_version);
1277 shader_dump_ins_modifiers(&semantic.reg);
1278 TRACE(" ");
1279 shader_dump_dst_param(&semantic.reg, &shader_version);
1280 }
1281 else if (ins.handler_idx == WINED3DSIH_DEF)
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("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1289 *(const float *)(ptr),
1290 *(const float *)(ptr + 1),
1291 *(const float *)(ptr + 2),
1292 *(const float *)(ptr + 3));
1293 ptr += 4;
1294 }
1295 else if (ins.handler_idx == WINED3DSIH_DEFI)
1296 {
1297 struct wined3d_shader_src_param rel_addr;
1298 struct wined3d_shader_dst_param dst;
1299
1300 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1301
1302 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1303 *(ptr),
1304 *(ptr + 1),
1305 *(ptr + 2),
1306 *(ptr + 3));
1307 ptr += 4;
1308 }
1309 else if (ins.handler_idx == WINED3DSIH_DEFB)
1310 {
1311 struct wined3d_shader_src_param rel_addr;
1312 struct wined3d_shader_dst_param dst;
1313
1314 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1315
1316 TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1317 ++ptr;
1318 }
1319 else
1320 {
1321 struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1322 struct wined3d_shader_dst_param dst_param;
1323 struct wined3d_shader_src_param src_param;
1324
1325 if (ins.dst_count)
1326 {
1327 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1328 }
1329
1330 /* Print out predication source token first - it follows
1331 * the destination token. */
1332 if (ins.predicate)
1333 {
1334 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1335 TRACE("(");
1336 shader_dump_src_param(&src_param, &shader_version);
1337 TRACE(") ");
1338 }
1339
1340 /* PixWin marks instructions with the coissue flag with a '+' */
1341 if (ins.coissue) TRACE("+");
1342
1343 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1344
1345 if (ins.handler_idx == WINED3DSIH_IFC
1346 || ins.handler_idx == WINED3DSIH_BREAKC)
1347 {
1348 switch (ins.flags)
1349 {
1350 case COMPARISON_GT: TRACE("_gt"); break;
1351 case COMPARISON_EQ: TRACE("_eq"); break;
1352 case COMPARISON_GE: TRACE("_ge"); break;
1353 case COMPARISON_LT: TRACE("_lt"); break;
1354 case COMPARISON_NE: TRACE("_ne"); break;
1355 case COMPARISON_LE: TRACE("_le"); break;
1356 default: TRACE("_(%u)", ins.flags);
1357 }
1358 }
1359 else if (ins.handler_idx == WINED3DSIH_TEX
1360 && shader_version.major >= 2
1361 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1362 {
1363 TRACE("p");
1364 }
1365
1366 /* We already read the destination token, print it. */
1367 if (ins.dst_count)
1368 {
1369 shader_dump_ins_modifiers(&dst_param);
1370 TRACE(" ");
1371 shader_dump_dst_param(&dst_param, &shader_version);
1372 }
1373
1374 /* Other source tokens */
1375 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1376 {
1377 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1378 TRACE(!i ? " " : ", ");
1379 shader_dump_src_param(&src_param, &shader_version);
1380 }
1381 }
1382 TRACE("\n");
1383 }
1384 }
1385
1386 static void shader_cleanup(IWineD3DBaseShader *iface)
1387 {
1388 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1389
1390 ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1391 HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1392 HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1393 shader_delete_constant_list(&shader->baseShader.constantsF);
1394 shader_delete_constant_list(&shader->baseShader.constantsB);
1395 shader_delete_constant_list(&shader->baseShader.constantsI);
1396 list_remove(&shader->baseShader.shader_list_entry);
1397
1398 if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1399 {
1400 shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1401 }
1402 }
1403
1404 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1405 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1406 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1407 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1408 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1409 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1410 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1411 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1412 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1413 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1414 static void shader_none_free(IWineD3DDevice *iface) {}
1415 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1416
1417 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1418 {
1419 /* Set the shader caps to 0 for the none shader backend */
1420 caps->VertexShaderVersion = 0;
1421 caps->PixelShaderVersion = 0;
1422 caps->PixelShader1xMaxValue = 0.0f;
1423 }
1424
1425 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1426 {
1427 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1428 {
1429 TRACE("Checking support for fixup:\n");
1430 dump_color_fixup_desc(fixup);
1431 }
1432
1433 /* Faked to make some apps happy. */
1434 if (!is_complex_fixup(fixup))
1435 {
1436 TRACE("[OK]\n");
1437 return TRUE;
1438 }
1439
1440 TRACE("[FAILED]\n");
1441 return FALSE;
1442 }
1443
1444 const shader_backend_t none_shader_backend = {
1445 shader_none_handle_instruction,
1446 shader_none_select,
1447 shader_none_select_depth_blt,
1448 shader_none_deselect_depth_blt,
1449 shader_none_update_float_vertex_constants,
1450 shader_none_update_float_pixel_constants,
1451 shader_none_load_constants,
1452 shader_none_load_np2fixup_constants,
1453 shader_none_destroy,
1454 shader_none_alloc,
1455 shader_none_free,
1456 shader_none_dirty_const,
1457 shader_none_get_caps,
1458 shader_none_color_fixup_supported,
1459 };
1460
1461 static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1462 {
1463 *parent = shader->baseShader.parent;
1464 IUnknown_AddRef(*parent);
1465 TRACE("shader %p, returning %p.\n", shader, *parent);
1466 }
1467
1468 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1469 {
1470 if (!data)
1471 {
1472 *data_size = shader->baseShader.functionLength;
1473 return WINED3D_OK;
1474 }
1475
1476 if (*data_size < shader->baseShader.functionLength)
1477 {
1478 /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1479 * than the required size we should write the required size and
1480 * return D3DERR_MOREDATA. That's not actually true. */
1481 return WINED3DERR_INVALIDCALL;
1482 }
1483
1484 memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1485
1486 return WINED3D_OK;
1487 }
1488
1489 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1490 const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1491 {
1492 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1493 const struct wined3d_shader_frontend *fe;
1494 HRESULT hr;
1495
1496 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1497 shader, byte_code, output_signature, float_const_count);
1498
1499 fe = shader_select_frontend(*byte_code);
1500 if (!fe)
1501 {
1502 FIXME("Unable to find frontend for shader.\n");
1503 return WINED3DERR_INVALIDCALL;
1504 }
1505 shader->baseShader.frontend = fe;
1506 shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1507 if (!shader->baseShader.frontend_data)
1508 {
1509 FIXME("Failed to initialize frontend.\n");
1510 return WINED3DERR_INVALIDCALL;
1511 }
1512
1513 /* First pass: trace shader. */
1514 if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1515
1516 /* Initialize immediate constant lists. */
1517 list_init(&shader->baseShader.constantsF);
1518 list_init(&shader->baseShader.constantsB);
1519 list_init(&shader->baseShader.constantsI);
1520
1521 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1522 hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1523 reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1524 byte_code, float_const_count);
1525 if (FAILED(hr)) return hr;
1526
1527 shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1528 if (!shader->baseShader.function) return E_OUTOFMEMORY;
1529 memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1530
1531 return WINED3D_OK;
1532 }
1533
1534 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1535 {
1536 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1537
1538 if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1539 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1540 || IsEqualGUID(riid, &IID_IWineD3DBase)
1541 || IsEqualGUID(riid, &IID_IUnknown))
1542 {
1543 IUnknown_AddRef(iface);
1544 *object = iface;
1545 return S_OK;
1546 }
1547
1548 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1549
1550 *object = NULL;
1551 return E_NOINTERFACE;
1552 }
1553
1554 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1555 {
1556 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1557 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1558
1559 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1560
1561 return refcount;
1562 }
1563
1564 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1565 {
1566 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1567 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1568
1569 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1570
1571 if (!refcount)
1572 {
1573 shader_cleanup((IWineD3DBaseShader *)iface);
1574 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1575 HeapFree(GetProcessHeap(), 0, shader);
1576 }
1577
1578 return refcount;
1579 }
1580
1581 static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1582 {
1583 TRACE("iface %p, parent %p.\n", iface, parent);
1584
1585 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1586
1587 return WINED3D_OK;
1588 }
1589
1590 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1591 {
1592 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1593
1594 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1595 }
1596
1597 /* Set local constants for d3d8 shaders. */
1598 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1599 UINT start_idx, const float *src_data, UINT count)
1600 {
1601 IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1602 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1603 UINT i, end_idx;
1604
1605 TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1606
1607 end_idx = start_idx + count;
1608 if (end_idx > device->d3d_vshader_constantF)
1609 {
1610 WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1611 end_idx = device->d3d_vshader_constantF;
1612 }
1613
1614 for (i = start_idx; i < end_idx; ++i)
1615 {
1616 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1617 if (!lconst) return E_OUTOFMEMORY;
1618
1619 lconst->idx = i;
1620 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1621 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1622 }
1623
1624 return WINED3D_OK;
1625 }
1626
1627 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1628 {
1629 /* IUnknown methods */
1630 vertexshader_QueryInterface,
1631 vertexshader_AddRef,
1632 vertexshader_Release,
1633 /* IWineD3DBase methods */
1634 vertexshader_GetParent,
1635 /* IWineD3DBaseShader methods */
1636 vertexshader_GetFunction,
1637 /* IWineD3DVertexShader methods */
1638 vertexshader_SetLocalConstantsF,
1639 };
1640
1641 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1642 IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1643 {
1644 args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1645 args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1646 && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1647 args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1648 }
1649
1650 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1651 {
1652 if (usage_idx1 != usage_idx2) return FALSE;
1653 if (usage1 == usage2) return TRUE;
1654 if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1655 if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1656
1657 return FALSE;
1658 }
1659
1660 BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1661 {
1662 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1663 WORD map = shader->baseShader.reg_maps.input_registers;
1664 unsigned int i;
1665
1666 for (i = 0; map; map >>= 1, ++i)
1667 {
1668 if (!(map & 1)) continue;
1669
1670 if (match_usage(shader->attributes[i].usage,
1671 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1672 {
1673 *regnum = i;
1674 return TRUE;
1675 }
1676 }
1677 return FALSE;
1678 }
1679
1680 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1681 {
1682 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1683 shader->baseShader.reg_maps.shader_version.minor);
1684 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1685
1686 shader->baseShader.limits.texcoord = 0;
1687 shader->baseShader.limits.attributes = 16;
1688 shader->baseShader.limits.packed_input = 0;
1689
1690 switch (shader_version)
1691 {
1692 case WINED3D_SHADER_VERSION(1, 0):
1693 case WINED3D_SHADER_VERSION(1, 1):
1694 shader->baseShader.limits.temporary = 12;
1695 shader->baseShader.limits.constant_bool = 0;
1696 shader->baseShader.limits.constant_int = 0;
1697 shader->baseShader.limits.address = 1;
1698 shader->baseShader.limits.packed_output = 0;
1699 shader->baseShader.limits.sampler = 0;
1700 shader->baseShader.limits.label = 0;
1701 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1702 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1703 * constants? */
1704 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1705 break;
1706
1707 case WINED3D_SHADER_VERSION(2, 0):
1708 case WINED3D_SHADER_VERSION(2, 1):
1709 shader->baseShader.limits.temporary = 12;
1710 shader->baseShader.limits.constant_bool = 16;
1711 shader->baseShader.limits.constant_int = 16;
1712 shader->baseShader.limits.address = 1;
1713 shader->baseShader.limits.packed_output = 0;
1714 shader->baseShader.limits.sampler = 0;
1715 shader->baseShader.limits.label = 16;
1716 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1717 break;
1718
1719 case WINED3D_SHADER_VERSION(4, 0):
1720 FIXME("Using 3.0 limits for 4.0 shader.\n");
1721 /* Fall through. */
1722
1723 case WINED3D_SHADER_VERSION(3, 0):
1724 shader->baseShader.limits.temporary = 32;
1725 shader->baseShader.limits.constant_bool = 32;
1726 shader->baseShader.limits.constant_int = 32;
1727 shader->baseShader.limits.address = 1;
1728 shader->baseShader.limits.packed_output = 12;
1729 shader->baseShader.limits.sampler = 4;
1730 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1731 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1732 * even though they are capable of supporting much more (GL
1733 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1734 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1735 * shaders to 256. */
1736 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1737 break;
1738
1739 default:
1740 shader->baseShader.limits.temporary = 12;
1741 shader->baseShader.limits.constant_bool = 16;
1742 shader->baseShader.limits.constant_int = 16;
1743 shader->baseShader.limits.address = 1;
1744 shader->baseShader.limits.packed_output = 0;
1745 shader->baseShader.limits.sampler = 0;
1746 shader->baseShader.limits.label = 16;
1747 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1748 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1749 shader->baseShader.reg_maps.shader_version.major,
1750 shader->baseShader.reg_maps.shader_version.minor);
1751 }
1752 }
1753
1754 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1755 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1756 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1757 {
1758 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1759 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1760 unsigned int i;
1761 HRESULT hr;
1762 WORD map;
1763
1764 if (!byte_code) return WINED3DERR_INVALIDCALL;
1765
1766 shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1767 shader_init(&shader->baseShader, device, parent, parent_ops);
1768
1769 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1770 output_signature, device->d3d_vshader_constantF);
1771 if (FAILED(hr))
1772 {
1773 WARN("Failed to set function, hr %#x.\n", hr);
1774 shader_cleanup((IWineD3DBaseShader *)shader);
1775 return hr;
1776 }
1777
1778 map = shader->baseShader.reg_maps.input_registers;
1779 for (i = 0; map; map >>= 1, ++i)
1780 {
1781 if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1782
1783 shader->attributes[i].usage =
1784 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1785 shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1786 }
1787
1788 if (output_signature)
1789 {
1790 for (i = 0; i < output_signature->element_count; ++i)
1791 {
1792 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1793 reg_maps->output_registers |= 1 << e->register_idx;
1794 shader->baseShader.output_signature[e->register_idx] = *e;
1795 }
1796 }
1797
1798 vertexshader_set_limits(shader);
1799
1800 if (device->vs_selected_mode == SHADER_ARB
1801 && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1802 && shader->min_rel_offset <= shader->max_rel_offset)
1803 {
1804 if (shader->max_rel_offset - shader->min_rel_offset > 127)
1805 {
1806 FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1807 FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1808 FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1809 }
1810 else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1811 {
1812 shader->rel_offset = shader->min_rel_offset + 63;
1813 }
1814 else if (shader->max_rel_offset > 63)
1815 {
1816 shader->rel_offset = shader->min_rel_offset;
1817 }
1818 else
1819 {
1820 shader->rel_offset = 0;
1821 }
1822 }
1823
1824 shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1825 && !list_empty(&shader->baseShader.constantsF);
1826
1827 return WINED3D_OK;
1828 }
1829
1830 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1831 REFIID riid, void **object)
1832 {
1833 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1834
1835 if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1836 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1837 || IsEqualGUID(riid, &IID_IWineD3DBase)
1838 || IsEqualGUID(riid, &IID_IUnknown))
1839 {
1840 IUnknown_AddRef(iface);
1841 *object = iface;
1842 return S_OK;
1843 }
1844
1845 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1846
1847 *object = NULL;
1848 return E_NOINTERFACE;
1849 }
1850
1851 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1852 {
1853 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1854 ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1855
1856 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1857
1858 return refcount;
1859 }
1860
1861 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1862 {
1863 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1864 ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1865
1866 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1867
1868 if (!refcount)
1869 {
1870 shader_cleanup((IWineD3DBaseShader *)iface);
1871 shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1872 HeapFree(GetProcessHeap(), 0, shader);
1873 }
1874
1875 return refcount;
1876 }
1877
1878 static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1879 {
1880 TRACE("iface %p, parent %p.\n", iface, parent);
1881
1882 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1883
1884 return WINED3D_OK;
1885 }
1886
1887 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1888 {
1889 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1890
1891 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1892 }
1893
1894 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1895 {
1896 /* IUnknown methods */
1897 geometryshader_QueryInterface,
1898 geometryshader_AddRef,
1899 geometryshader_Release,
1900 /* IWineD3DBase methods */
1901 geometryshader_GetParent,
1902 /* IWineD3DBaseShader methods */
1903 geometryshader_GetFunction,
1904 };
1905
1906 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1907 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1908 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1909 {
1910 HRESULT hr;
1911
1912 shader->vtbl = &wined3d_geometryshader_vtbl;
1913 shader_init(&shader->base_shader, device, parent, parent_ops);
1914
1915 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1916 if (FAILED(hr))
1917 {
1918 WARN("Failed to set function, hr %#x.\n", hr);
1919 shader_cleanup((IWineD3DBaseShader *)shader);
1920 return hr;
1921 }
1922
1923 shader->base_shader.load_local_constsF = FALSE;
1924
1925 return WINED3D_OK;
1926 }
1927
1928 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1929 {
1930 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1931
1932 if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1933 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1934 || IsEqualGUID(riid, &IID_IWineD3DBase)
1935 || IsEqualGUID(riid, &IID_IUnknown))
1936 {
1937 IUnknown_AddRef(iface);
1938 *object = iface;
1939 return S_OK;
1940 }
1941
1942 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1943
1944 *object = NULL;
1945 return E_NOINTERFACE;
1946 }
1947
1948 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
1949 {
1950 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1951 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1952
1953 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1954
1955 return refcount;
1956 }
1957
1958 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
1959 {
1960 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1961 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1962
1963 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1964
1965 if (!refcount)
1966 {
1967 shader_cleanup((IWineD3DBaseShader *)iface);
1968 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1969 HeapFree(GetProcessHeap(), 0, shader);
1970 }
1971
1972 return refcount;
1973 }
1974
1975 static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
1976 {
1977 TRACE("iface %p, parent %p.\n", iface, parent);
1978
1979 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1980
1981 return WINED3D_OK;
1982 }
1983
1984 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
1985 {
1986 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1987
1988 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1989 }
1990
1991 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
1992 {
1993 /* IUnknown methods */
1994 pixelshader_QueryInterface,
1995 pixelshader_AddRef,
1996 pixelshader_Release,
1997 /* IWineD3DBase methods */
1998 pixelshader_GetParent,
1999 /* IWineD3DBaseShader methods */
2000 pixelshader_GetFunction
2001 };
2002
2003 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
2004 IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
2005 {
2006 IWineD3DBaseTextureImpl *texture;
2007 UINT i;
2008
2009 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2010 args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
2011 args->np2_fixup = 0;
2012
2013 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2014 {
2015 if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2016 texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
2017 if (!texture)
2018 {
2019 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2020 continue;
2021 }
2022 args->color_fixup[i] = texture->resource.format_desc->color_fixup;
2023
2024 /* Flag samplers that need NP2 texcoord fixup. */
2025 if (!texture->baseTexture.pow2Matrix_identity)
2026 {
2027 args->np2_fixup |= (1 << i);
2028 }
2029 }
2030 if (shader->baseShader.reg_maps.shader_version.major >= 3)
2031 {
2032 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2033 {
2034 args->vp_mode = pretransformed;
2035 }
2036 else if (use_vs(stateblock))
2037 {
2038 args->vp_mode = vertexshader;
2039 }
2040 else
2041 {
2042 args->vp_mode = fixedfunction;
2043 }
2044 args->fog = FOG_OFF;
2045 }
2046 else
2047 {
2048 args->vp_mode = vertexshader;
2049 if (stateblock->renderState[WINED3DRS_FOGENABLE])
2050 {
2051 switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2052 {
2053 case WINED3DFOG_NONE:
2054 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2055 || use_vs(stateblock))
2056 {
2057 args->fog = FOG_LINEAR;
2058 break;
2059 }
2060
2061 switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2062 {
2063 case WINED3DFOG_NONE: /* Fall through. */
2064 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2065 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2066 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2067 }
2068 break;
2069
2070 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2071 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2072 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2073 }
2074 }
2075 else
2076 {
2077 args->fog = FOG_OFF;
2078 }
2079 }
2080 }
2081
2082 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2083 {
2084 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2085 shader->baseShader.reg_maps.shader_version.minor);
2086
2087 shader->baseShader.limits.attributes = 0;
2088 shader->baseShader.limits.address = 0;
2089 shader->baseShader.limits.packed_output = 0;
2090
2091 switch (shader_version)
2092 {
2093 case WINED3D_SHADER_VERSION(1, 0):
2094 case WINED3D_SHADER_VERSION(1, 1):
2095 case WINED3D_SHADER_VERSION(1, 2):
2096 case WINED3D_SHADER_VERSION(1, 3):
2097 shader->baseShader.limits.temporary = 2;
2098 shader->baseShader.limits.constant_float = 8;
2099 shader->baseShader.limits.constant_int = 0;
2100 shader->baseShader.limits.constant_bool = 0;
2101 shader->baseShader.limits.texcoord = 4;
2102 shader->baseShader.limits.sampler = 4;
2103 shader->baseShader.limits.packed_input = 0;
2104 shader->baseShader.limits.label = 0;
2105 break;
2106
2107 case WINED3D_SHADER_VERSION(1, 4):
2108 shader->baseShader.limits.temporary = 6;
2109 shader->baseShader.limits.constant_float = 8;
2110 shader->baseShader.limits.constant_int = 0;
2111 shader->baseShader.limits.constant_bool = 0;
2112 shader->baseShader.limits.texcoord = 6;
2113 shader->baseShader.limits.sampler = 6;
2114 shader->baseShader.limits.packed_input = 0;
2115 shader->baseShader.limits.label = 0;
2116 break;
2117
2118 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2119 case WINED3D_SHADER_VERSION(2, 0):
2120 shader->baseShader.limits.temporary = 32;
2121 shader->baseShader.limits.constant_float = 32;
2122 shader->baseShader.limits.constant_int = 16;
2123 shader->baseShader.limits.constant_bool = 16;
2124 shader->baseShader.limits.texcoord = 8;
2125 shader->baseShader.limits.sampler = 16;
2126 shader->baseShader.limits.packed_input = 0;
2127 break;
2128
2129 case WINED3D_SHADER_VERSION(2, 1):
2130 shader->baseShader.limits.temporary = 32;
2131 shader->baseShader.limits.constant_float = 32;
2132 shader->baseShader.limits.constant_int = 16;
2133 shader->baseShader.limits.constant_bool = 16;
2134 shader->baseShader.limits.texcoord = 8;
2135 shader->baseShader.limits.sampler = 16;
2136 shader->baseShader.limits.packed_input = 0;
2137 shader->baseShader.limits.label = 16;
2138 break;
2139
2140 case WINED3D_SHADER_VERSION(4, 0):
2141 FIXME("Using 3.0 limits for 4.0 shader.\n");
2142 /* Fall through. */
2143
2144 case WINED3D_SHADER_VERSION(3, 0):
2145 shader->baseShader.limits.temporary = 32;
2146 shader->baseShader.limits.constant_float = 224;
2147 shader->baseShader.limits.constant_int = 16;
2148 shader->baseShader.limits.constant_bool = 16;
2149 shader->baseShader.limits.texcoord = 0;
2150 shader->baseShader.limits.sampler = 16;
2151 shader->baseShader.limits.packed_input = 12;
2152 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2153 break;
2154
2155 default:
2156 shader->baseShader.limits.temporary = 32;
2157 shader->baseShader.limits.constant_float = 32;
2158 shader->baseShader.limits.constant_int = 16;
2159 shader->baseShader.limits.constant_bool = 16;
2160 shader->baseShader.limits.texcoord = 8;
2161 shader->baseShader.limits.sampler = 16;
2162 shader->baseShader.limits.packed_input = 0;
2163 shader->baseShader.limits.label = 0;
2164 FIXME("Unrecognized pixel shader version %u.%u\n",
2165 shader->baseShader.reg_maps.shader_version.major,
2166 shader->baseShader.reg_maps.shader_version.minor);
2167 }
2168 }
2169
2170 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2171 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2172 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2173 {
2174 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2175 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2176 HRESULT hr;
2177
2178 if (!byte_code) return WINED3DERR_INVALIDCALL;
2179
2180 shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2181 shader_init(&shader->baseShader, device, parent, parent_ops);
2182
2183 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2184 output_signature, device->d3d_pshader_constantF);
2185 if (FAILED(hr))
2186 {
2187 WARN("Failed to set function, hr %#x.\n", hr);
2188 shader_cleanup((IWineD3DBaseShader *)shader);
2189 return hr;
2190 }
2191
2192 pixelshader_set_limits(shader);
2193
2194 for (i = 0; i < MAX_REG_INPUT; ++i)
2195 {
2196 if (shader->input_reg_used[i])
2197 {
2198 ++num_regs_used;
2199 highest_reg_used = i;
2200 }
2201 }
2202
2203 /* Don't do any register mapping magic if it is not needed, or if we can't
2204 * achieve anything anyway */
2205 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2206 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2207 {
2208 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2209 {
2210 /* This happens with relative addressing. The input mapper function
2211 * warns about this if the higher registers are declared too, so
2212 * don't write a FIXME here */
2213 WARN("More varying registers used than supported\n");
2214 }
2215
2216 for (i = 0; i < MAX_REG_INPUT; ++i)
2217 {
2218 shader->input_reg_map[i] = i;
2219 }
2220
2221 shader->declared_in_count = highest_reg_used + 1;
2222 }
2223 else
2224 {
2225 shader->declared_in_count = 0;
2226 for (i = 0; i < MAX_REG_INPUT; ++i)
2227 {
2228 if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2229 else shader->input_reg_map[i] = ~0U;
2230 }
2231 }
2232
2233 shader->baseShader.load_local_constsF = FALSE;
2234
2235 return WINED3D_OK;
2236 }
2237
2238 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2239 {
2240 WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2241 unsigned int i;
2242
2243 if (reg_maps->shader_version.major != 1) return;
2244
2245 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2246 {
2247 /* We don't sample from this sampler. */
2248 if (!sampler_type[i]) continue;
2249
2250 if (!textures[i])
2251 {
2252 WARN("No texture bound to sampler %u, using 2D.\n", i);
2253 sampler_type[i] = WINED3DSTT_2D;
2254 continue;
2255 }
2256
2257 switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2258 {
2259 case GL_TEXTURE_RECTANGLE_ARB:
2260 case GL_TEXTURE_2D:
2261 /* We have to select between texture rectangles and 2D
2262 * textures later because 2.0 and 3.0 shaders only have
2263 * WINED3DSTT_2D as well. */
2264 sampler_type[i] = WINED3DSTT_2D;
2265 break;
2266
2267 case GL_TEXTURE_3D:
2268 sampler_type[i] = WINED3DSTT_VOLUME;
2269 break;
2270
2271 case GL_TEXTURE_CUBE_MAP_ARB:
2272 sampler_type[i] = WINED3DSTT_CUBE;
2273 break;
2274
2275 default:
2276 FIXME("Unrecognized texture type %#x, using 2D.\n",
2277 IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2278 sampler_type[i] = WINED3DSTT_2D;
2279 }
2280 }
2281 }
Windows API implementation