?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Recognize the SM4 lt opcode.
[wine/multimedia.git] / dlls / wined3d / baseshader.c
blobf0fe664ae7129eb4cd87319c9d381a7104388c35
1 /*
2 * shaders implementation
3 *
4 * Copyright 2002-2003 Jason Edmeades
5 * Copyright 2002-2003 Raphael Junqueira
6 * Copyright 2004 Christian Costa
7 * Copyright 2005 Oliver Stieber
8 * Copyright 2006 Ivan Gyurdiev
9 * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
10 * Copyright 2009 Henri Verbeet for CodeWeavers
11 *
12 * This library is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public
14 * License as published by the Free Software Foundation; either
15 * version 2.1 of the License, or (at your option) any later version.
16 *
17 * This library is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
21 *
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with this library; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
25 */
26
27 #include "config.h"
28 #include <string.h>
29 #include <stdio.h>
30 #include "wined3d_private.h"
31
32 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
33 WINE_DECLARE_DEBUG_CHANNEL(d3d);
34
35 static const char *shader_opcode_names[] =
36 {
37 /* WINED3DSIH_ABS */ "abs",
38 /* WINED3DSIH_ADD */ "add",
39 /* WINED3DSIH_BEM */ "bem",
40 /* WINED3DSIH_BREAK */ "break",
41 /* WINED3DSIH_BREAKC */ "breakc",
42 /* WINED3DSIH_BREAKP */ "breakp",
43 /* WINED3DSIH_CALL */ "call",
44 /* WINED3DSIH_CALLNZ */ "callnz",
45 /* WINED3DSIH_CMP */ "cmp",
46 /* WINED3DSIH_CND */ "cnd",
47 /* WINED3DSIH_CRS */ "crs",
48 /* WINED3DSIH_DCL */ "dcl",
49 /* WINED3DSIH_DEF */ "def",
50 /* WINED3DSIH_DEFB */ "defb",
51 /* WINED3DSIH_DEFI */ "defi",
52 /* WINED3DSIH_DP2ADD */ "dp2add",
53 /* WINED3DSIH_DP3 */ "dp3",
54 /* WINED3DSIH_DP4 */ "dp4",
55 /* WINED3DSIH_DST */ "dst",
56 /* WINED3DSIH_DSX */ "dsx",
57 /* WINED3DSIH_DSY */ "dsy",
58 /* WINED3DSIH_ELSE */ "else",
59 /* WINED3DSIH_EMIT */ "emit",
60 /* WINED3DSIH_ENDIF */ "endif",
61 /* WINED3DSIH_ENDLOOP */ "endloop",
62 /* WINED3DSIH_ENDREP */ "endrep",
63 /* WINED3DSIH_EXP */ "exp",
64 /* WINED3DSIH_EXPP */ "expp",
65 /* WINED3DSIH_FRC */ "frc",
66 /* WINED3DSIH_IADD */ "iadd",
67 /* WINED3DSIH_IF */ "if",
68 /* WINED3DSIH_IFC */ "ifc",
69 /* WINED3DSIH_IGE */ "ige",
70 /* WINED3DSIH_LABEL */ "label",
71 /* WINED3DSIH_LIT */ "lit",
72 /* WINED3DSIH_LOG */ "log",
73 /* WINED3DSIH_LOGP */ "logp",
74 /* WINED3DSIH_LOOP */ "loop",
75 /* WINED3DSIH_LRP */ "lrp",
76 /* WINED3DSIH_LT */ "lt",
77 /* WINED3DSIH_M3x2 */ "m3x2",
78 /* WINED3DSIH_M3x3 */ "m3x3",
79 /* WINED3DSIH_M3x4 */ "m3x4",
80 /* WINED3DSIH_M4x3 */ "m4x3",
81 /* WINED3DSIH_M4x4 */ "m4x4",
82 /* WINED3DSIH_MAD */ "mad",
83 /* WINED3DSIH_MAX */ "max",
84 /* WINED3DSIH_MIN */ "min",
85 /* WINED3DSIH_MOV */ "mov",
86 /* WINED3DSIH_MOVA */ "mova",
87 /* WINED3DSIH_MUL */ "mul",
88 /* WINED3DSIH_NOP */ "nop",
89 /* WINED3DSIH_NRM */ "nrm",
90 /* WINED3DSIH_PHASE */ "phase",
91 /* WINED3DSIH_POW */ "pow",
92 /* WINED3DSIH_RCP */ "rcp",
93 /* WINED3DSIH_REP */ "rep",
94 /* WINED3DSIH_RET */ "ret",
95 /* WINED3DSIH_RSQ */ "rsq",
96 /* WINED3DSIH_SETP */ "setp",
97 /* WINED3DSIH_SGE */ "sge",
98 /* WINED3DSIH_SGN */ "sgn",
99 /* WINED3DSIH_SINCOS */ "sincos",
100 /* WINED3DSIH_SLT */ "slt",
101 /* WINED3DSIH_SUB */ "sub",
102 /* WINED3DSIH_TEX */ "texld",
103 /* WINED3DSIH_TEXBEM */ "texbem",
104 /* WINED3DSIH_TEXBEML */ "texbeml",
105 /* WINED3DSIH_TEXCOORD */ "texcrd",
106 /* WINED3DSIH_TEXDEPTH */ "texdepth",
107 /* WINED3DSIH_TEXDP3 */ "texdp3",
108 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
109 /* WINED3DSIH_TEXKILL */ "texkill",
110 /* WINED3DSIH_TEXLDD */ "texldd",
111 /* WINED3DSIH_TEXLDL */ "texldl",
112 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
113 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
114 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
115 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
116 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
117 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
118 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
119 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
120 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
121 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
122 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
123 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
124 };
125
126 const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
127 {
128 switch (version_token >> 16)
129 {
130 case WINED3D_SM1_VS:
131 case WINED3D_SM1_PS:
132 return &sm1_shader_frontend;
133
134 case WINED3D_SM4_PS:
135 case WINED3D_SM4_VS:
136 case WINED3D_SM4_GS:
137 return &sm4_shader_frontend;
138
139 default:
140 FIXME("Unrecognised version token %#x\n", version_token);
141 return NULL;
142 }
143 }
144
145 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
146 {
147 buffer->buffer[0] = '\0';
148 buffer->bsize = 0;
149 buffer->lineNo = 0;
150 buffer->newline = TRUE;
151 }
152
153 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
154 {
155 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
156 if (!buffer->buffer)
157 {
158 ERR("Failed to allocate shader buffer memory.\n");
159 return FALSE;
160 }
161
162 shader_buffer_clear(buffer);
163 return TRUE;
164 }
165
166 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
167 {
168 HeapFree(GetProcessHeap(), 0, buffer->buffer);
169 }
170
171 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
172 {
173 char* base = buffer->buffer + buffer->bsize;
174 int rc;
175
176 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
177
178 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
179 {
180 ERR("The buffer allocated for the shader program string "
181 "is too small at %d bytes.\n", SHADER_PGMSIZE);
182 buffer->bsize = SHADER_PGMSIZE - 1;
183 return -1;
184 }
185
186 if (buffer->newline) {
187 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
188 buffer->newline = FALSE;
189 } else {
190 TRACE("%s", base);
191 }
192
193 buffer->bsize += rc;
194 if (buffer->buffer[buffer->bsize-1] == '\n') {
195 buffer->lineNo++;
196 buffer->newline = TRUE;
197 }
198 return 0;
199 }
200
201 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
202 {
203 int ret;
204 va_list args;
205
206 va_start(args, format);
207 ret = shader_vaddline(buffer, format, args);
208 va_end(args);
209
210 return ret;
211 }
212
213 void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
214 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
215 {
216 shader->ref = 1;
217 shader->device = (IWineD3DDevice *)device;
218 shader->parent = parent;
219 shader->parent_ops = parent_ops;
220 list_init(&shader->linked_programs);
221 list_add_head(&device->shaders, &shader->shader_list_entry);
222 }
223
224 /* Convert floating point offset relative
225 * to a register file to an absolute offset for float constants */
226 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
227 {
228 switch (register_type)
229 {
230 case WINED3DSPR_CONST: return register_idx;
231 case WINED3DSPR_CONST2: return 2048 + register_idx;
232 case WINED3DSPR_CONST3: return 4096 + register_idx;
233 case WINED3DSPR_CONST4: return 6144 + register_idx;
234 default:
235 FIXME("Unsupported register type: %d\n", register_type);
236 return register_idx;
237 }
238 }
239
240 static void shader_delete_constant_list(struct list* clist) {
241
242 struct list *ptr;
243 struct local_constant* constant;
244
245 ptr = list_head(clist);
246 while (ptr) {
247 constant = LIST_ENTRY(ptr, struct local_constant, entry);
248 ptr = list_next(clist, ptr);
249 HeapFree(GetProcessHeap(), 0, constant);
250 }
251 list_init(clist);
252 }
253
254 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
255 {
256 DWORD idx, shift;
257 idx = bit >> 5;
258 shift = bit & 0x1f;
259 bitmap[idx] |= (1 << shift);
260 }
261
262 static void shader_record_register_usage(IWineD3DBaseShaderImpl *This, struct shader_reg_maps *reg_maps,
263 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
264 {
265 switch (reg->type)
266 {
267 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
268 if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
269 else reg_maps->address |= 1 << reg->idx;
270 break;
271
272 case WINED3DSPR_TEMP:
273 reg_maps->temporary |= 1 << reg->idx;
274 break;
275
276 case WINED3DSPR_INPUT:
277 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
278 {
279 if (reg->rel_addr)
280 {
281 /* If relative addressing is used, we must assume that all registers
282 * are used. Even if it is a construct like v3[aL], we can't assume
283 * that v0, v1 and v2 aren't read because aL can be negative */
284 unsigned int i;
285 for (i = 0; i < MAX_REG_INPUT; ++i)
286 {
287 ((IWineD3DPixelShaderImpl *)This)->input_reg_used[i] = TRUE;
288 }
289 }
290 else
291 {
292 ((IWineD3DPixelShaderImpl *)This)->input_reg_used[reg->idx] = TRUE;
293 }
294 }
295 else reg_maps->input_registers |= 1 << reg->idx;
296 break;
297
298 case WINED3DSPR_RASTOUT:
299 if (reg->idx == 1) reg_maps->fog = 1;
300 break;
301
302 case WINED3DSPR_MISCTYPE:
303 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
304 {
305 if (reg->idx == 0) reg_maps->vpos = 1;
306 else if (reg->idx == 1) reg_maps->usesfacing = 1;
307 }
308 break;
309
310 case WINED3DSPR_CONST:
311 if (reg->rel_addr)
312 {
313 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
314 {
315 if (reg->idx < ((IWineD3DVertexShaderImpl *)This)->min_rel_offset)
316 {
317 ((IWineD3DVertexShaderImpl *)This)->min_rel_offset = reg->idx;
318 }
319 if (reg->idx > ((IWineD3DVertexShaderImpl *)This)->max_rel_offset)
320 {
321 ((IWineD3DVertexShaderImpl *)This)->max_rel_offset = reg->idx;
322 }
323 }
324 reg_maps->usesrelconstF = TRUE;
325 }
326 else
327 {
328 set_bitmap_bit(reg_maps->constf, reg->idx);
329 }
330 break;
331
332 case WINED3DSPR_CONSTINT:
333 reg_maps->integer_constants |= (1 << reg->idx);
334 break;
335
336 case WINED3DSPR_CONSTBOOL:
337 reg_maps->boolean_constants |= (1 << reg->idx);
338 break;
339
340 case WINED3DSPR_COLOROUT:
341 reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
342 break;
343
344 default:
345 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
346 break;
347 }
348 }
349
350 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
351 {
352 switch(instr)
353 {
354 case WINED3DSIH_M4x4:
355 case WINED3DSIH_M3x4:
356 return param == 1 ? 3 : 0;
357
358 case WINED3DSIH_M4x3:
359 case WINED3DSIH_M3x3:
360 return param == 1 ? 2 : 0;
361
362 case WINED3DSIH_M3x2:
363 return param == 1 ? 1 : 0;
364
365 default:
366 return 0;
367 }
368 }
369
370 static const char *semantic_names[] =
371 {
372 /* WINED3DDECLUSAGE_POSITION */ "SV_POSITION",
373 /* WINED3DDECLUSAGE_BLENDWEIGHT */ "BLENDWEIGHT",
374 /* WINED3DDECLUSAGE_BLENDINDICES */ "BLENDINDICES",
375 /* WINED3DDECLUSAGE_NORMAL */ "NORMAL",
376 /* WINED3DDECLUSAGE_PSIZE */ "PSIZE",
377 /* WINED3DDECLUSAGE_TEXCOORD */ "TEXCOORD",
378 /* WINED3DDECLUSAGE_TANGENT */ "TANGENT",
379 /* WINED3DDECLUSAGE_BINORMAL */ "BINORMAL",
380 /* WINED3DDECLUSAGE_TESSFACTOR */ "TESSFACTOR",
381 /* WINED3DDECLUSAGE_POSITIONT */ "POSITIONT",
382 /* WINED3DDECLUSAGE_COLOR */ "COLOR",
383 /* WINED3DDECLUSAGE_FOG */ "FOG",
384 /* WINED3DDECLUSAGE_DEPTH */ "DEPTH",
385 /* WINED3DDECLUSAGE_SAMPLE */ "SAMPLE",
386 };
387
388 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
389 {
390 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
391 {
392 FIXME("Unrecognized usage %#x\n", usage);
393 return "UNRECOGNIZED";
394 }
395
396 return semantic_names[usage];
397 }
398
399 WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
400 {
401 unsigned int i;
402
403 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
404 {
405 if (!strcmp(name, semantic_names[i])) return i;
406 }
407
408 return ~0U;
409 }
410
411 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
412 {
413 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
414 }
415
416 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
417 const struct wined3d_shader_semantic *s)
418 {
419 e->semantic_name = shader_semantic_name_from_usage(s->usage);
420 e->semantic_idx = s->usage_idx;
421 e->sysval_semantic = 0;
422 e->component_type = 0;
423 e->register_idx = s->reg.reg.idx;
424 e->mask = s->reg.write_mask;
425 }
426
427 /* Note that this does not count the loop register
428 * as an address register. */
429
430 HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
431 struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
432 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
433 {
434 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
435 void *fe_data = This->baseShader.frontend_data;
436 struct wined3d_shader_version shader_version;
437 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
438 const DWORD* pToken = byte_code;
439
440 /* There are some minor differences between pixel and vertex shaders */
441
442 memset(reg_maps, 0, sizeof(*reg_maps));
443
444 /* get_registers_used is called on every compile on some 1.x shaders, which can result
445 * in stacking up a collection of local constants. Delete the old constants if existing
446 */
447 shader_delete_constant_list(&This->baseShader.constantsF);
448 shader_delete_constant_list(&This->baseShader.constantsB);
449 shader_delete_constant_list(&This->baseShader.constantsI);
450
451 fe->shader_read_header(fe_data, &pToken, &shader_version);
452 reg_maps->shader_version = shader_version;
453
454 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
455 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
456 if(!reg_maps->constf) {
457 ERR("Out of memory\n");
458 return E_OUTOFMEMORY;
459 }
460
461 while (!fe->shader_is_end(fe_data, &pToken))
462 {
463 struct wined3d_shader_instruction ins;
464 const char *comment;
465 UINT param_size;
466
467 /* Skip comments */
468 fe->shader_read_comment(&pToken, &comment);
469 if (comment) continue;
470
471 /* Fetch opcode */
472 fe->shader_read_opcode(fe_data, &pToken, &ins, &param_size);
473
474 /* Unhandled opcode, and its parameters */
475 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
476 {
477 TRACE("Skipping unrecognized instruction.\n");
478 pToken += param_size;
479 continue;
480 }
481
482 /* Handle declarations */
483 if (ins.handler_idx == WINED3DSIH_DCL)
484 {
485 struct wined3d_shader_semantic semantic;
486
487 fe->shader_read_semantic(&pToken, &semantic);
488
489 switch (semantic.reg.reg.type)
490 {
491 /* Mark input registers used. */
492 case WINED3DSPR_INPUT:
493 reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
494 shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
495 break;
496
497 /* Vshader: mark 3.0 output registers used, save token */
498 case WINED3DSPR_OUTPUT:
499 reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
500 shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
501 if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
502 break;
503
504 /* Save sampler usage token */
505 case WINED3DSPR_SAMPLER:
506 reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
507 break;
508
509 default:
510 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
511 break;
512 }
513 }
514 else if (ins.handler_idx == WINED3DSIH_DEF)
515 {
516 struct wined3d_shader_dst_param dst;
517 struct wined3d_shader_src_param rel_addr;
518
519 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
520 if (!lconst) return E_OUTOFMEMORY;
521
522 fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
523 lconst->idx = dst.reg.idx;
524
525 memcpy(lconst->value, pToken, 4 * sizeof(DWORD));
526 pToken += 4;
527
528 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
529 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
530 {
531 float *value = (float *) lconst->value;
532 if (value[0] < -1.0f) value[0] = -1.0f;
533 else if (value[0] > 1.0f) value[0] = 1.0f;
534 if (value[1] < -1.0f) value[1] = -1.0f;
535 else if (value[1] > 1.0f) value[1] = 1.0f;
536 if (value[2] < -1.0f) value[2] = -1.0f;
537 else if (value[2] > 1.0f) value[2] = 1.0f;
538 if (value[3] < -1.0f) value[3] = -1.0f;
539 else if (value[3] > 1.0f) value[3] = 1.0f;
540 }
541
542 list_add_head(&This->baseShader.constantsF, &lconst->entry);
543 }
544 else if (ins.handler_idx == WINED3DSIH_DEFI)
545 {
546 struct wined3d_shader_dst_param dst;
547 struct wined3d_shader_src_param rel_addr;
548
549 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
550 if (!lconst) return E_OUTOFMEMORY;
551
552 fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
553 lconst->idx = dst.reg.idx;
554
555 memcpy(lconst->value, pToken, 4 * sizeof(DWORD));
556 pToken += 4;
557
558 list_add_head(&This->baseShader.constantsI, &lconst->entry);
559 reg_maps->local_int_consts |= (1 << dst.reg.idx);
560 }
561 else if (ins.handler_idx == WINED3DSIH_DEFB)
562 {
563 struct wined3d_shader_dst_param dst;
564 struct wined3d_shader_src_param rel_addr;
565
566 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
567 if (!lconst) return E_OUTOFMEMORY;
568
569 fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
570 lconst->idx = dst.reg.idx;
571
572 memcpy(lconst->value, pToken, sizeof(DWORD));
573 ++pToken;
574
575 list_add_head(&This->baseShader.constantsB, &lconst->entry);
576 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
577 }
578 /* If there's a loop in the shader */
579 else if (ins.handler_idx == WINED3DSIH_LOOP
580 || ins.handler_idx == WINED3DSIH_REP)
581 {
582 struct wined3d_shader_src_param src, rel_addr;
583
584 fe->shader_read_src_param(fe_data, &pToken, &src, &rel_addr);
585
586 /* Rep and Loop always use an integer constant for the control parameters */
587 if (ins.handler_idx == WINED3DSIH_REP)
588 {
589 reg_maps->integer_constants |= 1 << src.reg.idx;
590 }
591 else
592 {
593 fe->shader_read_src_param(fe_data, &pToken, &src, &rel_addr);
594 reg_maps->integer_constants |= 1 << src.reg.idx;
595 }
596
597 cur_loop_depth++;
598 if(cur_loop_depth > max_loop_depth)
599 max_loop_depth = cur_loop_depth;
600 }
601 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
602 || ins.handler_idx == WINED3DSIH_ENDREP)
603 {
604 cur_loop_depth--;
605 }
606 /* For subroutine prototypes */
607 else if (ins.handler_idx == WINED3DSIH_LABEL)
608 {
609 struct wined3d_shader_src_param src, rel_addr;
610
611 fe->shader_read_src_param(fe_data, &pToken, &src, &rel_addr);
612 reg_maps->labels |= 1 << src.reg.idx;
613 }
614 /* Set texture, address, temporary registers */
615 else
616 {
617 int i, limit;
618 BOOL color0_mov = FALSE;
619
620 /* This will loop over all the registers and try to
621 * make a bitmask of the ones we're interested in.
622 *
623 * Relative addressing tokens are ignored, but that's
624 * okay, since we'll catch any address registers when
625 * they are initialized (required by spec) */
626
627 if (ins.dst_count)
628 {
629 struct wined3d_shader_dst_param dst_param;
630 struct wined3d_shader_src_param dst_rel_addr;
631
632 fe->shader_read_dst_param(fe_data, &pToken, &dst_param, &dst_rel_addr);
633
634 shader_record_register_usage(This, reg_maps, &dst_param.reg, shader_version.type);
635
636 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
637 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
638 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
639 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
640 && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
641 {
642 reg_maps->texcoord_mask[dst_param.reg.idx] |= dst_param.write_mask;
643 }
644
645 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
646 {
647 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)This;
648
649 if(dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
650 {
651 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
652 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
653 * the mov and perform the sRGB write correction from the source register.
654 *
655 * However, if the mov is only partial, we can't do this, and if the write
656 * comes from an instruction other than MOV it is hard to do as well. If
657 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
658
659 ps->color0_mov = FALSE;
660 if (ins.handler_idx == WINED3DSIH_MOV)
661 {
662 /* Used later when the source register is read. */
663 color0_mov = TRUE;
664 }
665 }
666 /* Also drop the MOV marker if the source register is overwritten prior to the shader
667 * end
668 */
669 else if(dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
670 {
671 ps->color0_mov = FALSE;
672 }
673 }
674
675 /* Declare 1.X samplers implicitly, based on the destination reg. number */
676 if (shader_version.major == 1
677 && (ins.handler_idx == WINED3DSIH_TEX
678 || ins.handler_idx == WINED3DSIH_TEXBEM
679 || ins.handler_idx == WINED3DSIH_TEXBEML
680 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
681 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
682 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
683 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
684 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
685 || ins.handler_idx == WINED3DSIH_TEXREG2AR
686 || ins.handler_idx == WINED3DSIH_TEXREG2GB
687 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
688 {
689 /* Fake sampler usage, only set reserved bit and ttype */
690 DWORD sampler_code = dst_param.reg.idx;
691
692 TRACE("Setting fake 2D sampler for 1.x pixelshader\n");
693 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
694
695 /* texbem is only valid with < 1.4 pixel shaders */
696 if (ins.handler_idx == WINED3DSIH_TEXBEM
697 || ins.handler_idx == WINED3DSIH_TEXBEML)
698 {
699 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
700 if (ins.handler_idx == WINED3DSIH_TEXBEML)
701 {
702 reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
703 }
704 }
705 }
706 else if (ins.handler_idx == WINED3DSIH_BEM)
707 {
708 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
709 }
710 }
711
712 if (ins.handler_idx == WINED3DSIH_NRM)
713 {
714 reg_maps->usesnrm = 1;
715 }
716 else if (ins.handler_idx == WINED3DSIH_DSY)
717 {
718 reg_maps->usesdsy = 1;
719 }
720 else if (ins.handler_idx == WINED3DSIH_DSX)
721 {
722 reg_maps->usesdsx = 1;
723 }
724 else if(ins.handler_idx == WINED3DSIH_TEXLDD)
725 {
726 reg_maps->usestexldd = 1;
727 }
728 else if(ins.handler_idx == WINED3DSIH_TEXLDL)
729 {
730 reg_maps->usestexldl = 1;
731 }
732 else if(ins.handler_idx == WINED3DSIH_MOVA)
733 {
734 reg_maps->usesmova = 1;
735 }
736 else if(ins.handler_idx == WINED3DSIH_IFC)
737 {
738 reg_maps->usesifc = 1;
739 }
740 else if(ins.handler_idx == WINED3DSIH_CALL)
741 {
742 reg_maps->usescall = 1;
743 }
744
745 limit = ins.src_count + (ins.predicate ? 1 : 0);
746 for (i = 0; i < limit; ++i)
747 {
748 struct wined3d_shader_src_param src_param, src_rel_addr;
749 unsigned int count;
750
751 fe->shader_read_src_param(fe_data, &pToken, &src_param, &src_rel_addr);
752 count = get_instr_extra_regcount(ins.handler_idx, i);
753
754 shader_record_register_usage(This, reg_maps, &src_param.reg, shader_version.type);
755 while (count)
756 {
757 ++src_param.reg.idx;
758 shader_record_register_usage(This, reg_maps, &src_param.reg, shader_version.type);
759 --count;
760 }
761
762 if(color0_mov)
763 {
764 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
765 if(src_param.reg.type == WINED3DSPR_TEMP &&
766 src_param.swizzle == WINED3DSP_NOSWIZZLE)
767 {
768 ps->color0_mov = TRUE;
769 ps->color0_reg = src_param.reg.idx;
770 }
771 }
772 }
773 }
774 }
775 reg_maps->loop_depth = max_loop_depth;
776
777 This->baseShader.functionLength = ((const char *)pToken - (const char *)byte_code);
778
779 return WINED3D_OK;
780 }
781
782 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
783 {
784 DWORD map = 1 << max;
785 map |= map - 1;
786 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
787
788 return wined3d_log2i(map);
789 }
790
791 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
792 const struct wined3d_shader_version *shader_version)
793 {
794 TRACE("dcl");
795
796 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
797 {
798 switch (semantic->sampler_type)
799 {
800 case WINED3DSTT_2D: TRACE("_2d"); break;
801 case WINED3DSTT_CUBE: TRACE("_cube"); break;
802 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
803 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
804 }
805 }
806 else
807 {
808 /* Pixel shaders 3.0 don't have usage semantics */
809 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
810 return;
811 else
812 TRACE("_");
813
814 switch (semantic->usage)
815 {
816 case WINED3DDECLUSAGE_POSITION:
817 TRACE("position%d", semantic->usage_idx);
818 break;
819 case WINED3DDECLUSAGE_BLENDINDICES:
820 TRACE("blend");
821 break;
822 case WINED3DDECLUSAGE_BLENDWEIGHT:
823 TRACE("weight");
824 break;
825 case WINED3DDECLUSAGE_NORMAL:
826 TRACE("normal%d", semantic->usage_idx);
827 break;
828 case WINED3DDECLUSAGE_PSIZE:
829 TRACE("psize");
830 break;
831 case WINED3DDECLUSAGE_COLOR:
832 if (semantic->usage_idx == 0) TRACE("color");
833 else TRACE("specular%d", (semantic->usage_idx - 1));
834 break;
835 case WINED3DDECLUSAGE_TEXCOORD:
836 TRACE("texture%d", semantic->usage_idx);
837 break;
838 case WINED3DDECLUSAGE_TANGENT:
839 TRACE("tangent");
840 break;
841 case WINED3DDECLUSAGE_BINORMAL:
842 TRACE("binormal");
843 break;
844 case WINED3DDECLUSAGE_TESSFACTOR:
845 TRACE("tessfactor");
846 break;
847 case WINED3DDECLUSAGE_POSITIONT:
848 TRACE("positionT%d", semantic->usage_idx);
849 break;
850 case WINED3DDECLUSAGE_FOG:
851 TRACE("fog");
852 break;
853 case WINED3DDECLUSAGE_DEPTH:
854 TRACE("depth");
855 break;
856 case WINED3DDECLUSAGE_SAMPLE:
857 TRACE("sample");
858 break;
859 default:
860 FIXME("unknown_semantics(0x%08x)", semantic->usage);
861 }
862 }
863 }
864
865 static void shader_dump_register(const struct wined3d_shader_register *reg,
866 const struct wined3d_shader_version *shader_version)
867 {
868 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
869 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
870 UINT offset = reg->idx;
871
872 switch (reg->type)
873 {
874 case WINED3DSPR_TEMP:
875 TRACE("r");
876 break;
877
878 case WINED3DSPR_INPUT:
879 TRACE("v");
880 break;
881
882 case WINED3DSPR_CONST:
883 case WINED3DSPR_CONST2:
884 case WINED3DSPR_CONST3:
885 case WINED3DSPR_CONST4:
886 TRACE("c");
887 offset = shader_get_float_offset(reg->type, reg->idx);
888 break;
889
890 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
891 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
892 break;
893
894 case WINED3DSPR_RASTOUT:
895 TRACE("%s", rastout_reg_names[reg->idx]);
896 break;
897
898 case WINED3DSPR_COLOROUT:
899 TRACE("oC");
900 break;
901
902 case WINED3DSPR_DEPTHOUT:
903 TRACE("oDepth");
904 break;
905
906 case WINED3DSPR_ATTROUT:
907 TRACE("oD");
908 break;
909
910 case WINED3DSPR_TEXCRDOUT:
911 /* Vertex shaders >= 3.0 use general purpose output registers
912 * (WINED3DSPR_OUTPUT), which can include an address token */
913 if (shader_version->major >= 3) TRACE("o");
914 else TRACE("oT");
915 break;
916
917 case WINED3DSPR_CONSTINT:
918 TRACE("i");
919 break;
920
921 case WINED3DSPR_CONSTBOOL:
922 TRACE("b");
923 break;
924
925 case WINED3DSPR_LABEL:
926 TRACE("l");
927 break;
928
929 case WINED3DSPR_LOOP:
930 TRACE("aL");
931 break;
932
933 case WINED3DSPR_SAMPLER:
934 TRACE("s");
935 break;
936
937 case WINED3DSPR_MISCTYPE:
938 if (reg->idx > 1) FIXME("Unhandled misctype register %d\n", reg->idx);
939 else TRACE("%s", misctype_reg_names[reg->idx]);
940 break;
941
942 case WINED3DSPR_PREDICATE:
943 TRACE("p");
944 break;
945
946 case WINED3DSPR_IMMCONST:
947 TRACE("l");
948 break;
949
950 case WINED3DSPR_CONSTBUFFER:
951 TRACE("cb");
952 break;
953
954 default:
955 TRACE("unhandled_rtype(%#x)", reg->type);
956 break;
957 }
958
959 if (reg->type == WINED3DSPR_IMMCONST)
960 {
961 TRACE("(");
962 switch (reg->immconst_type)
963 {
964 case WINED3D_IMMCONST_FLOAT:
965 TRACE("%.8e", *(const float *)reg->immconst_data);
966 break;
967
968 case WINED3D_IMMCONST_FLOAT4:
969 TRACE("%.8e, %.8e, %.8e, %.8e",
970 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
971 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
972 break;
973
974 default:
975 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
976 break;
977 }
978 TRACE(")");
979 }
980 else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
981 {
982 if (reg->array_idx != ~0U)
983 {
984 TRACE("%u[%u", offset, reg->array_idx);
985 if (reg->rel_addr)
986 {
987 TRACE(" + ");
988 shader_dump_src_param(reg->rel_addr, shader_version);
989 }
990 TRACE("]");
991 }
992 else
993 {
994 if (reg->rel_addr)
995 {
996 TRACE("[");
997 shader_dump_src_param(reg->rel_addr, shader_version);
998 TRACE(" + ");
999 }
1000 TRACE("%u", offset);
1001 if (reg->rel_addr) TRACE("]");
1002 }
1003 }
1004 }
1005
1006 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1007 const struct wined3d_shader_version *shader_version)
1008 {
1009 DWORD write_mask = param->write_mask;
1010
1011 shader_dump_register(&param->reg, shader_version);
1012
1013 if (write_mask != WINED3DSP_WRITEMASK_ALL)
1014 {
1015 static const char *write_mask_chars = "xyzw";
1016
1017 TRACE(".");
1018 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1019 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1020 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1021 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1022 }
1023 }
1024
1025 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1026 const struct wined3d_shader_version *shader_version)
1027 {
1028 DWORD src_modifier = param->modifiers;
1029 DWORD swizzle = param->swizzle;
1030
1031 if (src_modifier == WINED3DSPSM_NEG
1032 || src_modifier == WINED3DSPSM_BIASNEG
1033 || src_modifier == WINED3DSPSM_SIGNNEG
1034 || src_modifier == WINED3DSPSM_X2NEG
1035 || src_modifier == WINED3DSPSM_ABSNEG)
1036 TRACE("-");
1037 else if (src_modifier == WINED3DSPSM_COMP)
1038 TRACE("1-");
1039 else if (src_modifier == WINED3DSPSM_NOT)
1040 TRACE("!");
1041
1042 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1043 TRACE("abs(");
1044
1045 shader_dump_register(&param->reg, shader_version);
1046
1047 if (src_modifier)
1048 {
1049 switch (src_modifier)
1050 {
1051 case WINED3DSPSM_NONE: break;
1052 case WINED3DSPSM_NEG: break;
1053 case WINED3DSPSM_NOT: break;
1054 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1055 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1056 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1057 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1058 case WINED3DSPSM_COMP: break;
1059 case WINED3DSPSM_X2: TRACE("_x2"); break;
1060 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1061 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1062 case WINED3DSPSM_DW: TRACE("_dw"); break;
1063 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1064 case WINED3DSPSM_ABS: TRACE(")"); break;
1065 default:
1066 TRACE("_unknown_modifier(%#x)", src_modifier);
1067 }
1068 }
1069
1070 if (swizzle != WINED3DSP_NOSWIZZLE)
1071 {
1072 static const char *swizzle_chars = "xyzw";
1073 DWORD swizzle_x = swizzle & 0x03;
1074 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1075 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1076 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1077
1078 if (swizzle_x == swizzle_y
1079 && swizzle_x == swizzle_z
1080 && swizzle_x == swizzle_w)
1081 {
1082 TRACE(".%c", swizzle_chars[swizzle_x]);
1083 }
1084 else
1085 {
1086 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1087 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1088 }
1089 }
1090 }
1091
1092 /* Shared code in order to generate the bulk of the shader string.
1093 * NOTE: A description of how to parse tokens can be found on msdn */
1094 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1095 const shader_reg_maps *reg_maps, const DWORD *pFunction, void *backend_ctx)
1096 {
1097 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
1098 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
1099 const struct wined3d_shader_frontend *fe = This->baseShader.frontend;
1100 void *fe_data = This->baseShader.frontend_data;
1101 struct wined3d_shader_src_param src_rel_addr[4];
1102 struct wined3d_shader_src_param src_param[4];
1103 struct wined3d_shader_version shader_version;
1104 struct wined3d_shader_src_param dst_rel_addr;
1105 struct wined3d_shader_dst_param dst_param;
1106 struct wined3d_shader_instruction ins;
1107 struct wined3d_shader_context ctx;
1108 const DWORD *pToken = pFunction;
1109 DWORD i;
1110
1111 /* Initialize current parsing state */
1112 ctx.shader = iface;
1113 ctx.reg_maps = reg_maps;
1114 ctx.buffer = buffer;
1115 ctx.backend_data = backend_ctx;
1116
1117 ins.ctx = &ctx;
1118 ins.dst = &dst_param;
1119 ins.src = src_param;
1120 This->baseShader.parse_state.current_row = 0;
1121
1122 fe->shader_read_header(fe_data, &pToken, &shader_version);
1123
1124 while (!fe->shader_is_end(fe_data, &pToken))
1125 {
1126 const char *comment;
1127 UINT param_size;
1128
1129 /* Skip comment tokens */
1130 fe->shader_read_comment(&pToken, &comment);
1131 if (comment) continue;
1132
1133 /* Read opcode */
1134 fe->shader_read_opcode(fe_data, &pToken, &ins, &param_size);
1135
1136 /* Unknown opcode and its parameters */
1137 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1138 {
1139 TRACE("Skipping unrecognized instruction.\n");
1140 pToken += param_size;
1141 continue;
1142 }
1143
1144 /* Nothing to do */
1145 if (ins.handler_idx == WINED3DSIH_DCL
1146 || ins.handler_idx == WINED3DSIH_NOP
1147 || ins.handler_idx == WINED3DSIH_DEF
1148 || ins.handler_idx == WINED3DSIH_DEFI
1149 || ins.handler_idx == WINED3DSIH_DEFB
1150 || ins.handler_idx == WINED3DSIH_PHASE)
1151 {
1152 pToken += param_size;
1153 continue;
1154 }
1155
1156 /* Destination token */
1157 if (ins.dst_count) fe->shader_read_dst_param(fe_data, &pToken, &dst_param, &dst_rel_addr);
1158
1159 /* Predication token */
1160 if (ins.predicate) ins.predicate = *pToken++;
1161
1162 /* Other source tokens */
1163 for (i = 0; i < ins.src_count; ++i)
1164 {
1165 fe->shader_read_src_param(fe_data, &pToken, &src_param[i], &src_rel_addr[i]);
1166 }
1167
1168 /* Call appropriate function for output target */
1169 device->shader_backend->shader_handle_instruction(&ins);
1170 }
1171 }
1172
1173 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1174 {
1175 DWORD mmask = dst->modifiers;
1176
1177 switch (dst->shift)
1178 {
1179 case 0: break;
1180 case 13: TRACE("_d8"); break;
1181 case 14: TRACE("_d4"); break;
1182 case 15: TRACE("_d2"); break;
1183 case 1: TRACE("_x2"); break;
1184 case 2: TRACE("_x4"); break;
1185 case 3: TRACE("_x8"); break;
1186 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1187 }
1188
1189 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1190 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1191 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1192
1193 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1194 if (mmask)
1195 FIXME("_unrecognized_modifier(%#x)", mmask);
1196 }
1197
1198 void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *pFunction)
1199 {
1200 struct wined3d_shader_version shader_version;
1201 const DWORD* pToken = pFunction;
1202 const char *type_prefix;
1203 DWORD i;
1204
1205 TRACE("Parsing %p\n", pFunction);
1206
1207 fe->shader_read_header(fe_data, &pToken, &shader_version);
1208
1209 switch (shader_version.type)
1210 {
1211 case WINED3D_SHADER_TYPE_VERTEX:
1212 type_prefix = "vs";
1213 break;
1214
1215 case WINED3D_SHADER_TYPE_GEOMETRY:
1216 type_prefix = "gs";
1217 break;
1218
1219 case WINED3D_SHADER_TYPE_PIXEL:
1220 type_prefix = "ps";
1221 break;
1222
1223 default:
1224 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1225 type_prefix = "unknown";
1226 break;
1227 }
1228
1229 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1230
1231 while (!fe->shader_is_end(fe_data, &pToken))
1232 {
1233 struct wined3d_shader_instruction ins;
1234 const char *comment;
1235 UINT param_size;
1236
1237 /* comment */
1238 fe->shader_read_comment(&pToken, &comment);
1239 if (comment)
1240 {
1241 TRACE("//%s\n", comment);
1242 continue;
1243 }
1244
1245 fe->shader_read_opcode(fe_data, &pToken, &ins, &param_size);
1246 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1247 {
1248 TRACE("Skipping unrecognized instruction.\n");
1249 pToken += param_size;
1250 continue;
1251 }
1252
1253 if (ins.handler_idx == WINED3DSIH_DCL)
1254 {
1255 struct wined3d_shader_semantic semantic;
1256
1257 fe->shader_read_semantic(&pToken, &semantic);
1258
1259 shader_dump_decl_usage(&semantic, &shader_version);
1260 shader_dump_ins_modifiers(&semantic.reg);
1261 TRACE(" ");
1262 shader_dump_dst_param(&semantic.reg, &shader_version);
1263 }
1264 else if (ins.handler_idx == WINED3DSIH_DEF)
1265 {
1266 struct wined3d_shader_dst_param dst;
1267 struct wined3d_shader_src_param rel_addr;
1268
1269 fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
1270
1271 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1272 *(const float *)(pToken),
1273 *(const float *)(pToken + 1),
1274 *(const float *)(pToken + 2),
1275 *(const float *)(pToken + 3));
1276 pToken += 4;
1277 }
1278 else if (ins.handler_idx == WINED3DSIH_DEFI)
1279 {
1280 struct wined3d_shader_dst_param dst;
1281 struct wined3d_shader_src_param rel_addr;
1282
1283 fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
1284
1285 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1286 *(pToken),
1287 *(pToken + 1),
1288 *(pToken + 2),
1289 *(pToken + 3));
1290 pToken += 4;
1291 }
1292 else if (ins.handler_idx == WINED3DSIH_DEFB)
1293 {
1294 struct wined3d_shader_dst_param dst;
1295 struct wined3d_shader_src_param rel_addr;
1296
1297 fe->shader_read_dst_param(fe_data, &pToken, &dst, &rel_addr);
1298
1299 TRACE("defb b%u = %s", dst.reg.idx, *pToken ? "true" : "false");
1300 ++pToken;
1301 }
1302 else
1303 {
1304 struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1305 struct wined3d_shader_dst_param dst_param;
1306 struct wined3d_shader_src_param src_param;
1307
1308 if (ins.dst_count)
1309 {
1310 fe->shader_read_dst_param(fe_data, &pToken, &dst_param, &dst_rel_addr);
1311 }
1312
1313 /* Print out predication source token first - it follows
1314 * the destination token. */
1315 if (ins.predicate)
1316 {
1317 fe->shader_read_src_param(fe_data, &pToken, &src_param, &src_rel_addr);
1318 TRACE("(");
1319 shader_dump_src_param(&src_param, &shader_version);
1320 TRACE(") ");
1321 }
1322
1323 /* PixWin marks instructions with the coissue flag with a '+' */
1324 if (ins.coissue) TRACE("+");
1325
1326 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1327
1328 if (ins.handler_idx == WINED3DSIH_IFC
1329 || ins.handler_idx == WINED3DSIH_BREAKC)
1330 {
1331 switch (ins.flags)
1332 {
1333 case COMPARISON_GT: TRACE("_gt"); break;
1334 case COMPARISON_EQ: TRACE("_eq"); break;
1335 case COMPARISON_GE: TRACE("_ge"); break;
1336 case COMPARISON_LT: TRACE("_lt"); break;
1337 case COMPARISON_NE: TRACE("_ne"); break;
1338 case COMPARISON_LE: TRACE("_le"); break;
1339 default: TRACE("_(%u)", ins.flags);
1340 }
1341 }
1342 else if (ins.handler_idx == WINED3DSIH_TEX
1343 && shader_version.major >= 2
1344 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1345 {
1346 TRACE("p");
1347 }
1348
1349 /* We already read the destination token, print it. */
1350 if (ins.dst_count)
1351 {
1352 shader_dump_ins_modifiers(&dst_param);
1353 TRACE(" ");
1354 shader_dump_dst_param(&dst_param, &shader_version);
1355 }
1356
1357 /* Other source tokens */
1358 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1359 {
1360 fe->shader_read_src_param(fe_data, &pToken, &src_param, &src_rel_addr);
1361 TRACE(!i ? " " : ", ");
1362 shader_dump_src_param(&src_param, &shader_version);
1363 }
1364 }
1365 TRACE("\n");
1366 }
1367 }
1368
1369 void shader_cleanup(IWineD3DBaseShader *iface)
1370 {
1371 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1372
1373 ((IWineD3DDeviceImpl *)This->baseShader.device)->shader_backend->shader_destroy(iface);
1374 HeapFree(GetProcessHeap(), 0, This->baseShader.reg_maps.constf);
1375 HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1376 shader_delete_constant_list(&This->baseShader.constantsF);
1377 shader_delete_constant_list(&This->baseShader.constantsB);
1378 shader_delete_constant_list(&This->baseShader.constantsI);
1379 list_remove(&This->baseShader.shader_list_entry);
1380
1381 if (This->baseShader.frontend && This->baseShader.frontend_data)
1382 {
1383 This->baseShader.frontend->shader_free(This->baseShader.frontend_data);
1384 }
1385 }
1386
1387 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1388 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1389 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1390 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1391 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1392 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1393 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1394 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1395 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1396 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1397 static void shader_none_free(IWineD3DDevice *iface) {}
1398 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1399
1400 static void shader_none_get_caps(WINED3DDEVTYPE devtype,
1401 const struct wined3d_gl_info *gl_info, struct shader_caps *pCaps)
1402 {
1403 /* Set the shader caps to 0 for the none shader backend */
1404 pCaps->VertexShaderVersion = 0;
1405 pCaps->PixelShaderVersion = 0;
1406 pCaps->PixelShader1xMaxValue = 0.0f;
1407 }
1408
1409 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1410 {
1411 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1412 {
1413 TRACE("Checking support for fixup:\n");
1414 dump_color_fixup_desc(fixup);
1415 }
1416
1417 /* Faked to make some apps happy. */
1418 if (!is_yuv_fixup(fixup))
1419 {
1420 TRACE("[OK]\n");
1421 return TRUE;
1422 }
1423
1424 TRACE("[FAILED]\n");
1425 return FALSE;
1426 }
1427
1428 const shader_backend_t none_shader_backend = {
1429 shader_none_handle_instruction,
1430 shader_none_select,
1431 shader_none_select_depth_blt,
1432 shader_none_deselect_depth_blt,
1433 shader_none_update_float_vertex_constants,
1434 shader_none_update_float_pixel_constants,
1435 shader_none_load_constants,
1436 shader_none_load_np2fixup_constants,
1437 shader_none_destroy,
1438 shader_none_alloc,
1439 shader_none_free,
1440 shader_none_dirty_const,
1441 shader_none_get_caps,
1442 shader_none_color_fixup_supported,
1443 };
Maarten's multimedia branch