VirtualBox

source: vbox/trunk/src/VBox/Devices/Graphics/shaderlib/shader.c@ 81722

Last change on this file since 81722 was 81211, checked in by vboxsync, 5 years ago

More validation of a shader bytecode, bugref:9520
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 87.7 KB
Line 
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/*
26 * Oracle LGPL Disclaimer: For the avoidance of doubt, except that if any license choice
27 * other than GPL or LGPL is available it will apply instead, Oracle elects to use only
28 * the Lesser General Public License version 2.1 (LGPLv2) at this time for any software where
29 * a choice of LGPL license versions is made available with the language indicating
30 * that LGPLv2 or any later version may be used, or where a choice of which version
31 * of the LGPL is applied is otherwise unspecified.
32 */
33
34#include "config.h"
35#include "wine/port.h"
36
37#include <math.h>
38#include <stdio.h>
39#include <string.h>
40
41#include "wined3d_private.h"
42
43WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
44WINE_DECLARE_DEBUG_CHANNEL(d3d);
45
46static const char *shader_opcode_names[] =
47{
48 /* WINED3DSIH_ABS */ "abs",
49 /* WINED3DSIH_ADD */ "add",
50 /* WINED3DSIH_BEM */ "bem",
51 /* WINED3DSIH_BREAK */ "break",
52 /* WINED3DSIH_BREAKC */ "breakc",
53 /* WINED3DSIH_BREAKP */ "breakp",
54 /* WINED3DSIH_CALL */ "call",
55 /* WINED3DSIH_CALLNZ */ "callnz",
56 /* WINED3DSIH_CMP */ "cmp",
57 /* WINED3DSIH_CND */ "cnd",
58 /* WINED3DSIH_CRS */ "crs",
59 /* WINED3DSIH_CUT */ "cut",
60 /* WINED3DSIH_DCL */ "dcl",
61 /* WINED3DSIH_DEF */ "def",
62 /* WINED3DSIH_DEFB */ "defb",
63 /* WINED3DSIH_DEFI */ "defi",
64 /* WINED3DSIH_DP2ADD */ "dp2add",
65 /* WINED3DSIH_DP3 */ "dp3",
66 /* WINED3DSIH_DP4 */ "dp4",
67 /* WINED3DSIH_DST */ "dst",
68 /* WINED3DSIH_DSX */ "dsx",
69 /* WINED3DSIH_DSY */ "dsy",
70 /* WINED3DSIH_ELSE */ "else",
71 /* WINED3DSIH_EMIT */ "emit",
72 /* WINED3DSIH_ENDIF */ "endif",
73 /* WINED3DSIH_ENDLOOP */ "endloop",
74 /* WINED3DSIH_ENDREP */ "endrep",
75 /* WINED3DSIH_EXP */ "exp",
76 /* WINED3DSIH_EXPP */ "expp",
77 /* WINED3DSIH_FRC */ "frc",
78 /* WINED3DSIH_IADD */ "iadd",
79 /* WINED3DSIH_IF */ "if",
80 /* WINED3DSIH_IFC */ "ifc",
81 /* WINED3DSIH_IGE */ "ige",
82 /* WINED3DSIH_LABEL */ "label",
83 /* WINED3DSIH_LIT */ "lit",
84 /* WINED3DSIH_LOG */ "log",
85 /* WINED3DSIH_LOGP */ "logp",
86 /* WINED3DSIH_LOOP */ "loop",
87 /* WINED3DSIH_LRP */ "lrp",
88 /* WINED3DSIH_LT */ "lt",
89 /* WINED3DSIH_M3x2 */ "m3x2",
90 /* WINED3DSIH_M3x3 */ "m3x3",
91 /* WINED3DSIH_M3x4 */ "m3x4",
92 /* WINED3DSIH_M4x3 */ "m4x3",
93 /* WINED3DSIH_M4x4 */ "m4x4",
94 /* WINED3DSIH_MAD */ "mad",
95 /* WINED3DSIH_MAX */ "max",
96 /* WINED3DSIH_MIN */ "min",
97 /* WINED3DSIH_MOV */ "mov",
98 /* WINED3DSIH_MOVA */ "mova",
99 /* WINED3DSIH_MUL */ "mul",
100 /* WINED3DSIH_NOP */ "nop",
101 /* WINED3DSIH_NRM */ "nrm",
102 /* WINED3DSIH_PHASE */ "phase",
103 /* WINED3DSIH_POW */ "pow",
104 /* WINED3DSIH_RCP */ "rcp",
105 /* WINED3DSIH_REP */ "rep",
106 /* WINED3DSIH_RET */ "ret",
107 /* WINED3DSIH_RSQ */ "rsq",
108 /* WINED3DSIH_SETP */ "setp",
109 /* WINED3DSIH_SGE */ "sge",
110 /* WINED3DSIH_SGN */ "sgn",
111 /* WINED3DSIH_SINCOS */ "sincos",
112 /* WINED3DSIH_SLT */ "slt",
113 /* WINED3DSIH_SUB */ "sub",
114 /* WINED3DSIH_TEX */ "texld",
115 /* WINED3DSIH_TEXBEM */ "texbem",
116 /* WINED3DSIH_TEXBEML */ "texbeml",
117 /* WINED3DSIH_TEXCOORD */ "texcrd",
118 /* WINED3DSIH_TEXDEPTH */ "texdepth",
119 /* WINED3DSIH_TEXDP3 */ "texdp3",
120 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
121 /* WINED3DSIH_TEXKILL */ "texkill",
122 /* WINED3DSIH_TEXLDD */ "texldd",
123 /* WINED3DSIH_TEXLDL */ "texldl",
124 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
125 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
126 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
127 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
128 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
129 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
130 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
131 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
132 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
133 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
134 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
135 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
136};
137
138static const char *semantic_names[] =
139{
140 /* WINED3DDECLUSAGE_POSITION */ "SV_POSITION",
141 /* WINED3DDECLUSAGE_BLENDWEIGHT */ "BLENDWEIGHT",
142 /* WINED3DDECLUSAGE_BLENDINDICES */ "BLENDINDICES",
143 /* WINED3DDECLUSAGE_NORMAL */ "NORMAL",
144 /* WINED3DDECLUSAGE_PSIZE */ "PSIZE",
145 /* WINED3DDECLUSAGE_TEXCOORD */ "TEXCOORD",
146 /* WINED3DDECLUSAGE_TANGENT */ "TANGENT",
147 /* WINED3DDECLUSAGE_BINORMAL */ "BINORMAL",
148 /* WINED3DDECLUSAGE_TESSFACTOR */ "TESSFACTOR",
149 /* WINED3DDECLUSAGE_POSITIONT */ "POSITIONT",
150 /* WINED3DDECLUSAGE_COLOR */ "COLOR",
151 /* WINED3DDECLUSAGE_FOG */ "FOG",
152 /* WINED3DDECLUSAGE_DEPTH */ "DEPTH",
153 /* WINED3DDECLUSAGE_SAMPLE */ "SAMPLE",
154};
155
156static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
157{
158 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
159 {
160 FIXME("Unrecognized usage %#x.\n", usage);
161 return "UNRECOGNIZED";
162 }
163
164 return semantic_names[usage];
165}
166
167static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
168{
169 unsigned int i;
170
171 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
172 {
173 if (!strcmp(name, semantic_names[i])) return i;
174 }
175
176 return ~0U;
177}
178
179BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
180{
181 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
182}
183
184static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
185 const struct wined3d_shader_semantic *s)
186{
187 e->semantic_name = shader_semantic_name_from_usage(s->usage);
188 e->semantic_idx = s->usage_idx;
189 e->sysval_semantic = 0;
190 e->component_type = 0;
191 e->register_idx = s->reg.reg.idx;
192 e->mask = s->reg.write_mask;
193}
194
195static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
196{
197 switch (version_token >> 16)
198 {
199 case WINED3D_SM1_VS:
200 case WINED3D_SM1_PS:
201 return &sm1_shader_frontend;
202
203 case WINED3D_SM4_PS:
204 case WINED3D_SM4_VS:
205 case WINED3D_SM4_GS:
206 return &sm4_shader_frontend;
207
208 default:
209 FIXME("Unrecognised version token %#x\n", version_token);
210 return NULL;
211 }
212}
213
214void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
215{
216 buffer->buffer[0] = '\0';
217 buffer->bsize = 0;
218 buffer->lineNo = 0;
219 buffer->newline = TRUE;
220}
221
222BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
223{
224 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
225 if (!buffer->buffer)
226 {
227 ERR("Failed to allocate shader buffer memory.\n");
228 return FALSE;
229 }
230
231 shader_buffer_clear(buffer);
232 return TRUE;
233}
234
235void shader_buffer_free(struct wined3d_shader_buffer *buffer)
236{
237 HeapFree(GetProcessHeap(), 0, buffer->buffer);
238}
239
240int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
241{
242 char *base = buffer->buffer + buffer->bsize;
243 int rc;
244
245 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
246
247 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
248 {
249 ERR("The buffer allocated for the shader program string "
250 "is too small at %d bytes.\n", SHADER_PGMSIZE);
251 buffer->bsize = SHADER_PGMSIZE - 1;
252 return -1;
253 }
254
255 if (buffer->newline)
256 {
257 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
258 buffer->newline = FALSE;
259 }
260 else
261 {
262 TRACE("%s", base);
263 }
264
265 buffer->bsize += rc;
266 if (buffer->buffer[buffer->bsize-1] == '\n')
267 {
268 ++buffer->lineNo;
269 buffer->newline = TRUE;
270 }
271
272 return 0;
273}
274
275int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
276{
277 va_list args;
278 int ret;
279
280 va_start(args, format);
281 ret = shader_vaddline(buffer, format, args);
282 va_end(args);
283
284 return ret;
285}
286
287static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
288 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
289{
290 shader->ref = 1;
291 shader->device = (IWineD3DDevice *)device;
292 shader->parent = parent;
293 shader->parent_ops = parent_ops;
294 list_init(&shader->linked_programs);
295 list_add_head(&device->shaders, &shader->shader_list_entry);
296}
297
298/* Convert floating point offset relative to a register file to an absolute
299 * offset for float constants. */
300static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
301{
302 switch (register_type)
303 {
304 case WINED3DSPR_CONST: return register_idx;
305 case WINED3DSPR_CONST2: return 2048 + register_idx;
306 case WINED3DSPR_CONST3: return 4096 + register_idx;
307 case WINED3DSPR_CONST4: return 6144 + register_idx;
308 default:
309 FIXME("Unsupported register type: %u.\n", register_type);
310 return register_idx;
311 }
312}
313
314static void shader_delete_constant_list(struct list *clist)
315{
316 struct local_constant *constant;
317 struct list *ptr;
318
319 ptr = list_head(clist);
320 while (ptr)
321 {
322 constant = LIST_ENTRY(ptr, struct local_constant, entry);
323 ptr = list_next(clist, ptr);
324 HeapFree(GetProcessHeap(), 0, constant);
325 }
326 list_init(clist);
327}
328
329static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
330{
331 DWORD idx, shift;
332 idx = bit >> 5;
333 shift = bit & 0x1f;
334 bitmap[idx] |= (1 << shift);
335}
336
337static HRESULT shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct shader_reg_maps *reg_maps,
338 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
339{
340 switch (reg->type)
341 {
342 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
343 if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
344 else reg_maps->address |= 1 << reg->idx;
345 break;
346
347 case WINED3DSPR_TEMP:
348 reg_maps->temporary |= 1 << reg->idx;
349 break;
350
351 case WINED3DSPR_INPUT:
352 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
353 {
354 if (reg->rel_addr)
355 {
356 /* If relative addressing is used, we must assume that all registers
357 * are used. Even if it is a construct like v3[aL], we can't assume
358 * that v0, v1 and v2 aren't read because aL can be negative */
359 unsigned int i;
360 for (i = 0; i < MAX_REG_INPUT; ++i)
361 {
362 ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
363 }
364 }
365 else
366 {
367 unsigned int reg_idx = reg->idx;
368
369 if (reg_idx >= MAX_REG_INPUT)
370 {
371 ERR("Invalid input register index %d.\n", reg_idx);
372 return E_INVALIDARG;
373 }
374
375 ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg_idx] = TRUE;
376 }
377 }
378 else reg_maps->input_registers |= 1 << reg->idx;
379 break;
380
381 case WINED3DSPR_RASTOUT:
382 if (reg->idx == 1) reg_maps->fog = 1;
383 break;
384
385 case WINED3DSPR_MISCTYPE:
386 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
387 {
388 if (reg->idx == 0) reg_maps->vpos = 1;
389 else if (reg->idx == 1) reg_maps->usesfacing = 1;
390 }
391 break;
392
393 case WINED3DSPR_CONST:
394 if (reg->rel_addr)
395 {
396 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
397 {
398 if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
399 {
400 ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
401 }
402 if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
403 {
404 ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
405 }
406 }
407 reg_maps->usesrelconstF = TRUE;
408 }
409 else
410 {
411 set_bitmap_bit(reg_maps->constf, reg->idx);
412 }
413 break;
414
415 case WINED3DSPR_CONSTINT:
416 reg_maps->integer_constants |= (1 << reg->idx);
417 break;
418
419 case WINED3DSPR_CONSTBOOL:
420 reg_maps->boolean_constants |= (1 << reg->idx);
421 break;
422
423 case WINED3DSPR_COLOROUT:
424 reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
425 break;
426
427 default:
428 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
429 break;
430 }
431
432 return S_OK;
433}
434
435static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
436{
437 switch (instr)
438 {
439 case WINED3DSIH_M4x4:
440 case WINED3DSIH_M3x4:
441 return param == 1 ? 3 : 0;
442
443 case WINED3DSIH_M4x3:
444 case WINED3DSIH_M3x3:
445 return param == 1 ? 2 : 0;
446
447 case WINED3DSIH_M3x2:
448 return param == 1 ? 1 : 0;
449
450 default:
451 return 0;
452 }
453}
454
455/* Note that this does not count the loop register as an address register. */
456static HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
457 struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
458 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
459{
460 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
461 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
462 void *fe_data = shader->baseShader.frontend_data;
463 struct wined3d_shader_version shader_version;
464 const DWORD *ptr = byte_code;
465
466 memset(reg_maps, 0, sizeof(*reg_maps));
467
468 /* get_registers_used() is called on every compile on some 1.x shaders,
469 * which can result in stacking up a collection of local constants.
470 * Delete the old constants if existing. */
471 shader_delete_constant_list(&shader->baseShader.constantsF);
472 shader_delete_constant_list(&shader->baseShader.constantsB);
473 shader_delete_constant_list(&shader->baseShader.constantsI);
474
475 fe->shader_read_header(fe_data, &ptr, &shader_version);
476 reg_maps->shader_version = shader_version;
477
478 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
479 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
480 if (!reg_maps->constf)
481 {
482 ERR("Failed to allocate constant map memory.\n");
483 return E_OUTOFMEMORY;
484 }
485
486 while (!fe->shader_is_end(fe_data, &ptr))
487 {
488 struct wined3d_shader_instruction ins;
489 const char *comment;
490 UINT comment_size;
491 UINT param_size;
492
493 /* Skip comments. */
494 fe->shader_read_comment(&ptr, &comment, &comment_size);
495 if (comment) continue;
496
497 /* Fetch opcode. */
498 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
499
500 /* Unhandled opcode, and its parameters. */
501 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
502 {
503 TRACE("Skipping unrecognized instruction.\n");
504 ptr += param_size;
505 continue;
506 }
507
508 /* Handle declarations. */
509 if (ins.handler_idx == WINED3DSIH_DCL)
510 {
511 struct wined3d_shader_semantic semantic;
512 unsigned int reg_idx;
513
514 fe->shader_read_semantic(&ptr, &semantic);
515 reg_idx = semantic.reg.reg.idx;
516
517 switch (semantic.reg.reg.type)
518 {
519 /* Mark input registers used. */
520 case WINED3DSPR_INPUT:
521 if (reg_idx >= MAX_REG_INPUT)
522 {
523 ERR("Invalid input register index %d.\n", reg_idx);
524 return E_INVALIDARG;
525 }
526 reg_maps->input_registers |= 1 << reg_idx;
527 shader_signature_from_semantic(&input_signature[reg_idx], &semantic);
528 break;
529
530 /* Vertex shader: mark 3.0 output registers used, save token. */
531 case WINED3DSPR_OUTPUT:
532 if (reg_idx >= MAX_REG_OUTPUT)
533 {
534 ERR("Invalid output register index %d.\n", reg_idx);
535 return E_INVALIDARG;
536 }
537 reg_maps->output_registers |= 1 << reg_idx;
538 shader_signature_from_semantic(&output_signature[reg_idx], &semantic);
539 if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
540 break;
541
542 /* Save sampler usage token. */
543 case WINED3DSPR_SAMPLER:
544 if (reg_idx >= RT_ELEMENTS(reg_maps->sampler_type))
545 {
546 ERR("Invalid sampler index %d.\n", reg_idx);
547 return E_INVALIDARG;
548 }
549 reg_maps->sampler_type[reg_idx] = semantic.sampler_type;
550 break;
551
552 default:
553 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
554 break;
555 }
556 }
557 else if (ins.handler_idx == WINED3DSIH_DEF)
558 {
559 struct wined3d_shader_src_param rel_addr;
560 struct wined3d_shader_dst_param dst;
561
562 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
563 if (!lconst) return E_OUTOFMEMORY;
564
565 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
566 lconst->idx = dst.reg.idx;
567
568 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
569 ptr += 4;
570
571 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
572 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
573 {
574 float *value = (float *)lconst->value;
575 if (value[0] < -1.0f) value[0] = -1.0f;
576 else if (value[0] > 1.0f) value[0] = 1.0f;
577 if (value[1] < -1.0f) value[1] = -1.0f;
578 else if (value[1] > 1.0f) value[1] = 1.0f;
579 if (value[2] < -1.0f) value[2] = -1.0f;
580 else if (value[2] > 1.0f) value[2] = 1.0f;
581 if (value[3] < -1.0f) value[3] = -1.0f;
582 else if (value[3] > 1.0f) value[3] = 1.0f;
583 }
584
585 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
586 }
587 else if (ins.handler_idx == WINED3DSIH_DEFI)
588 {
589 struct wined3d_shader_src_param rel_addr;
590 struct wined3d_shader_dst_param dst;
591
592 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
593 if (!lconst) return E_OUTOFMEMORY;
594
595 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
596 lconst->idx = dst.reg.idx;
597
598 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
599 ptr += 4;
600
601 list_add_head(&shader->baseShader.constantsI, &lconst->entry);
602 reg_maps->local_int_consts |= (1 << dst.reg.idx);
603 }
604 else if (ins.handler_idx == WINED3DSIH_DEFB)
605 {
606 struct wined3d_shader_src_param rel_addr;
607 struct wined3d_shader_dst_param dst;
608
609 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
610 if (!lconst) return E_OUTOFMEMORY;
611
612 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
613 lconst->idx = dst.reg.idx;
614
615 memcpy(lconst->value, ptr, sizeof(DWORD));
616 ++ptr;
617
618 list_add_head(&shader->baseShader.constantsB, &lconst->entry);
619 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
620 }
621 /* If there's a loop in the shader. */
622 else if (ins.handler_idx == WINED3DSIH_LOOP
623 || ins.handler_idx == WINED3DSIH_REP)
624 {
625 struct wined3d_shader_src_param src, rel_addr;
626
627 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
628
629 /* Rep and Loop always use an integer constant for the control parameters. */
630 if (ins.handler_idx == WINED3DSIH_REP)
631 {
632 reg_maps->integer_constants |= 1 << src.reg.idx;
633 }
634 else
635 {
636 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
637 reg_maps->integer_constants |= 1 << src.reg.idx;
638 }
639
640 cur_loop_depth++;
641 if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
642 }
643 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
644 || ins.handler_idx == WINED3DSIH_ENDREP)
645 {
646 cur_loop_depth--;
647 }
648 /* For subroutine prototypes. */
649 else if (ins.handler_idx == WINED3DSIH_LABEL)
650 {
651 struct wined3d_shader_src_param src, rel_addr;
652
653 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
654 reg_maps->labels |= 1 << src.reg.idx;
655 }
656 /* Set texture, address, temporary registers. */
657 else
658 {
659 BOOL color0_mov = FALSE;
660 int i, limit;
661
662 /* This will loop over all the registers and try to
663 * make a bitmask of the ones we're interested in.
664 *
665 * Relative addressing tokens are ignored, but that's
666 * okay, since we'll catch any address registers when
667 * they are initialized (required by spec). */
668 if (ins.dst_count)
669 {
670 struct wined3d_shader_src_param dst_rel_addr;
671 struct wined3d_shader_dst_param dst_param;
672 HRESULT hr;
673
674 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
675
676 hr = shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
677 if (FAILED(hr))
678 {
679 ERR("shader_record_register_usage failed.\n");
680 return hr;
681 }
682
683 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
684 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
685 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
686 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
687 && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
688 {
689 unsigned int idx = dst_param.reg.idx;
690
691 if (idx >= MAX_REG_TEXCRD)
692 {
693 ERR("Invalid texcoord index %d.\n", idx);
694 return E_INVALIDARG;
695 }
696
697 reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
698 }
699
700 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
701 {
702 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
703
704 if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
705 {
706 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
707 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
708 * the mov and perform the sRGB write correction from the source register.
709 *
710 * However, if the mov is only partial, we can't do this, and if the write
711 * comes from an instruction other than MOV it is hard to do as well. If
712 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
713
714 ps->color0_mov = FALSE;
715 if (ins.handler_idx == WINED3DSIH_MOV)
716 {
717 /* Used later when the source register is read. */
718 color0_mov = TRUE;
719 }
720 }
721 /* Also drop the MOV marker if the source register is overwritten prior to the shader
722 * end
723 */
724 else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
725 {
726 ps->color0_mov = FALSE;
727 }
728 }
729
730 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
731 if (shader_version.major == 1
732 && (ins.handler_idx == WINED3DSIH_TEX
733 || ins.handler_idx == WINED3DSIH_TEXBEM
734 || ins.handler_idx == WINED3DSIH_TEXBEML
735 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
736 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
737 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
738 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
739 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
740 || ins.handler_idx == WINED3DSIH_TEXREG2AR
741 || ins.handler_idx == WINED3DSIH_TEXREG2GB
742 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
743 {
744 /* Fake sampler usage, only set reserved bit and type. */
745 DWORD sampler_code = dst_param.reg.idx;
746
747 if (sampler_code >= RT_ELEMENTS(reg_maps->sampler_type))
748 {
749 ERR("Invalid sampler index %d.\n", sampler_code);
750 return E_INVALIDARG;
751 }
752
753 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
754 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
755
756 /* texbem is only valid with < 1.4 pixel shaders */
757 if (ins.handler_idx == WINED3DSIH_TEXBEM
758 || ins.handler_idx == WINED3DSIH_TEXBEML)
759 {
760 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
761 if (ins.handler_idx == WINED3DSIH_TEXBEML)
762 {
763 reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
764 }
765 }
766 }
767 else if (ins.handler_idx == WINED3DSIH_BEM)
768 {
769 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
770 }
771 }
772
773 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
774 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
775 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
776 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
777 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
778 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
779 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
780 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
781
782 limit = ins.src_count + (ins.predicate ? 1 : 0);
783 for (i = 0; i < limit; ++i)
784 {
785 struct wined3d_shader_src_param src_param, src_rel_addr;
786 unsigned int count;
787
788 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
789 count = get_instr_extra_regcount(ins.handler_idx, i);
790
791 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
792 while (count)
793 {
794 ++src_param.reg.idx;
795 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
796 --count;
797 }
798
799 if (color0_mov)
800 {
801 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
802 if (src_param.reg.type == WINED3DSPR_TEMP
803 && src_param.swizzle == WINED3DSP_NOSWIZZLE)
804 {
805 ps->color0_mov = TRUE;
806 ps->color0_reg = src_param.reg.idx;
807 }
808 }
809 }
810 }
811 }
812 reg_maps->loop_depth = max_loop_depth;
813
814 shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
815
816 return WINED3D_OK;
817}
818
819unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
820{
821 DWORD map = 1 << max;
822 map |= map - 1;
823 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
824
825 return wined3d_log2i(map);
826}
827
828static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
829 const struct wined3d_shader_version *shader_version)
830{
831 TRACE("dcl");
832
833 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
834 {
835 switch (semantic->sampler_type)
836 {
837 case WINED3DSTT_2D: TRACE("_2d"); break;
838 case WINED3DSTT_CUBE: TRACE("_cube"); break;
839 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
840 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
841 }
842 }
843 else
844 {
845 /* Pixel shaders 3.0 don't have usage semantics. */
846 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
847 else TRACE("_");
848
849 switch (semantic->usage)
850 {
851 case WINED3DDECLUSAGE_POSITION:
852 TRACE("position%u", semantic->usage_idx);
853 break;
854
855 case WINED3DDECLUSAGE_BLENDINDICES:
856 TRACE("blend");
857 break;
858
859 case WINED3DDECLUSAGE_BLENDWEIGHT:
860 TRACE("weight");
861 break;
862
863 case WINED3DDECLUSAGE_NORMAL:
864 TRACE("normal%u", semantic->usage_idx);
865 break;
866
867 case WINED3DDECLUSAGE_PSIZE:
868 TRACE("psize");
869 break;
870
871 case WINED3DDECLUSAGE_COLOR:
872 if (semantic->usage_idx == 0) TRACE("color");
873 else TRACE("specular%u", (semantic->usage_idx - 1));
874 break;
875
876 case WINED3DDECLUSAGE_TEXCOORD:
877 TRACE("texture%u", semantic->usage_idx);
878 break;
879
880 case WINED3DDECLUSAGE_TANGENT:
881 TRACE("tangent");
882 break;
883
884 case WINED3DDECLUSAGE_BINORMAL:
885 TRACE("binormal");
886 break;
887
888 case WINED3DDECLUSAGE_TESSFACTOR:
889 TRACE("tessfactor");
890 break;
891
892 case WINED3DDECLUSAGE_POSITIONT:
893 TRACE("positionT%u", semantic->usage_idx);
894 break;
895
896 case WINED3DDECLUSAGE_FOG:
897 TRACE("fog");
898 break;
899
900 case WINED3DDECLUSAGE_DEPTH:
901 TRACE("depth");
902 break;
903
904 case WINED3DDECLUSAGE_SAMPLE:
905 TRACE("sample");
906 break;
907
908 default:
909 FIXME("unknown_semantics(0x%08x)", semantic->usage);
910 }
911 }
912}
913
914static void shader_dump_register(const struct wined3d_shader_register *reg,
915 const struct wined3d_shader_version *shader_version)
916{
917 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
918 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
919 UINT offset = reg->idx;
920
921 switch (reg->type)
922 {
923 case WINED3DSPR_TEMP:
924 TRACE("r");
925 break;
926
927 case WINED3DSPR_INPUT:
928 TRACE("v");
929 break;
930
931 case WINED3DSPR_CONST:
932 case WINED3DSPR_CONST2:
933 case WINED3DSPR_CONST3:
934 case WINED3DSPR_CONST4:
935 TRACE("c");
936 offset = shader_get_float_offset(reg->type, reg->idx);
937 break;
938
939 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
940 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
941 break;
942
943 case WINED3DSPR_RASTOUT:
944 TRACE("%s", rastout_reg_names[reg->idx]);
945 break;
946
947 case WINED3DSPR_COLOROUT:
948 TRACE("oC");
949 break;
950
951 case WINED3DSPR_DEPTHOUT:
952 TRACE("oDepth");
953 break;
954
955 case WINED3DSPR_ATTROUT:
956 TRACE("oD");
957 break;
958
959 case WINED3DSPR_TEXCRDOUT:
960 /* Vertex shaders >= 3.0 use general purpose output registers
961 * (WINED3DSPR_OUTPUT), which can include an address token. */
962 if (shader_version->major >= 3) TRACE("o");
963 else TRACE("oT");
964 break;
965
966 case WINED3DSPR_CONSTINT:
967 TRACE("i");
968 break;
969
970 case WINED3DSPR_CONSTBOOL:
971 TRACE("b");
972 break;
973
974 case WINED3DSPR_LABEL:
975 TRACE("l");
976 break;
977
978 case WINED3DSPR_LOOP:
979 TRACE("aL");
980 break;
981
982 case WINED3DSPR_SAMPLER:
983 TRACE("s");
984 break;
985
986 case WINED3DSPR_MISCTYPE:
987 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
988 else TRACE("%s", misctype_reg_names[reg->idx]);
989 break;
990
991 case WINED3DSPR_PREDICATE:
992 TRACE("p");
993 break;
994
995 case WINED3DSPR_IMMCONST:
996 TRACE("l");
997 break;
998
999 case WINED3DSPR_CONSTBUFFER:
1000 TRACE("cb");
1001 break;
1002
1003 default:
1004 TRACE("unhandled_rtype(%#x)", reg->type);
1005 break;
1006 }
1007
1008 if (reg->type == WINED3DSPR_IMMCONST)
1009 {
1010 TRACE("(");
1011 switch (reg->immconst_type)
1012 {
1013 case WINED3D_IMMCONST_FLOAT:
1014 TRACE("%.8e", *(const float *)reg->immconst_data);
1015 break;
1016
1017 case WINED3D_IMMCONST_FLOAT4:
1018 TRACE("%.8e, %.8e, %.8e, %.8e",
1019 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1020 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1021 break;
1022
1023 default:
1024 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1025 break;
1026 }
1027 TRACE(")");
1028 }
1029 else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
1030 {
1031 if (reg->array_idx != ~0U)
1032 {
1033 TRACE("%u[%u", offset, reg->array_idx);
1034 if (reg->rel_addr)
1035 {
1036 TRACE(" + ");
1037 shader_dump_src_param(reg->rel_addr, shader_version);
1038 }
1039 TRACE("]");
1040 }
1041 else
1042 {
1043 if (reg->rel_addr)
1044 {
1045 TRACE("[");
1046 shader_dump_src_param(reg->rel_addr, shader_version);
1047 TRACE(" + ");
1048 }
1049 TRACE("%u", offset);
1050 if (reg->rel_addr) TRACE("]");
1051 }
1052 }
1053}
1054
1055void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1056 const struct wined3d_shader_version *shader_version)
1057{
1058 DWORD write_mask = param->write_mask;
1059
1060 shader_dump_register(&param->reg, shader_version);
1061
1062 if (write_mask != WINED3DSP_WRITEMASK_ALL)
1063 {
1064 static const char *write_mask_chars = "xyzw";
1065
1066 TRACE(".");
1067 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1068 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1069 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1070 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1071 }
1072}
1073
1074void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1075 const struct wined3d_shader_version *shader_version)
1076{
1077 DWORD src_modifier = param->modifiers;
1078 DWORD swizzle = param->swizzle;
1079
1080 if (src_modifier == WINED3DSPSM_NEG
1081 || src_modifier == WINED3DSPSM_BIASNEG
1082 || src_modifier == WINED3DSPSM_SIGNNEG
1083 || src_modifier == WINED3DSPSM_X2NEG
1084 || src_modifier == WINED3DSPSM_ABSNEG)
1085 TRACE("-");
1086 else if (src_modifier == WINED3DSPSM_COMP)
1087 TRACE("1-");
1088 else if (src_modifier == WINED3DSPSM_NOT)
1089 TRACE("!");
1090
1091 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1092 TRACE("abs(");
1093
1094 shader_dump_register(&param->reg, shader_version);
1095
1096 if (src_modifier)
1097 {
1098 switch (src_modifier)
1099 {
1100 case WINED3DSPSM_NONE: break;
1101 case WINED3DSPSM_NEG: break;
1102 case WINED3DSPSM_NOT: break;
1103 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1104 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1105 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1106 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1107 case WINED3DSPSM_COMP: break;
1108 case WINED3DSPSM_X2: TRACE("_x2"); break;
1109 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1110 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1111 case WINED3DSPSM_DW: TRACE("_dw"); break;
1112 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1113 case WINED3DSPSM_ABS: TRACE(")"); break;
1114 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1115 }
1116 }
1117
1118 if (swizzle != WINED3DSP_NOSWIZZLE)
1119 {
1120 static const char *swizzle_chars = "xyzw";
1121 DWORD swizzle_x = swizzle & 0x03;
1122 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1123 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1124 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1125
1126 if (swizzle_x == swizzle_y
1127 && swizzle_x == swizzle_z
1128 && swizzle_x == swizzle_w)
1129 {
1130 TRACE(".%c", swizzle_chars[swizzle_x]);
1131 }
1132 else
1133 {
1134 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1135 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1136 }
1137 }
1138}
1139
1140/* Shared code in order to generate the bulk of the shader string.
1141 * NOTE: A description of how to parse tokens can be found on MSDN. */
1142void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1143 const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1144{
1145 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1146 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1147 const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1148 void *fe_data = shader->baseShader.frontend_data;
1149 struct wined3d_shader_src_param src_rel_addr[4];
1150 struct wined3d_shader_src_param src_param[4];
1151 struct wined3d_shader_version shader_version;
1152 struct wined3d_shader_src_param dst_rel_addr;
1153 struct wined3d_shader_dst_param dst_param;
1154 struct wined3d_shader_instruction ins;
1155 struct wined3d_shader_context ctx;
1156 const DWORD *ptr = byte_code;
1157 DWORD i;
1158
1159 /* Initialize current parsing state. */
1160 ctx.shader = iface;
1161 ctx.gl_info = &device->adapter->gl_info;
1162 ctx.reg_maps = reg_maps;
1163 ctx.buffer = buffer;
1164 ctx.backend_data = backend_ctx;
1165
1166 ins.ctx = &ctx;
1167 ins.dst = &dst_param;
1168 ins.src = src_param;
1169 shader->baseShader.parse_state.current_row = 0;
1170
1171 fe->shader_read_header(fe_data, &ptr, &shader_version);
1172
1173 while (!fe->shader_is_end(fe_data, &ptr))
1174 {
1175 const char *comment;
1176 UINT comment_size;
1177 UINT param_size;
1178
1179 /* Skip comment tokens. */
1180 fe->shader_read_comment(&ptr, &comment, &comment_size);
1181 if (comment) continue;
1182
1183 /* Read opcode. */
1184 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1185
1186 /* Unknown opcode and its parameters. */
1187 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1188 {
1189 TRACE("Skipping unrecognized instruction.\n");
1190 ptr += param_size;
1191 continue;
1192 }
1193
1194 /* Nothing to do. */
1195 if (ins.handler_idx == WINED3DSIH_DCL
1196 || ins.handler_idx == WINED3DSIH_NOP
1197 || ins.handler_idx == WINED3DSIH_DEF
1198 || ins.handler_idx == WINED3DSIH_DEFI
1199 || ins.handler_idx == WINED3DSIH_DEFB
1200 || ins.handler_idx == WINED3DSIH_PHASE)
1201 {
1202 ptr += param_size;
1203 continue;
1204 }
1205
1206 /* Destination token */
1207 if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1208
1209 /* Predication token */
1210 if (ins.predicate)
1211 {
1212#ifndef VBOX_WITH_VMSVGA
1213 FIXME("Predicates not implemented.\n");
1214#endif
1215 ins.predicate = *ptr++;
1216 }
1217
1218 /* Other source tokens */
1219 for (i = 0; i < ins.src_count; ++i)
1220 {
1221 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1222 }
1223
1224 /* Call appropriate function for output target */
1225 device->shader_backend->shader_handle_instruction(&ins);
1226 }
1227}
1228
1229static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1230{
1231 DWORD mmask = dst->modifiers;
1232
1233 switch (dst->shift)
1234 {
1235 case 0: break;
1236 case 13: TRACE("_d8"); break;
1237 case 14: TRACE("_d4"); break;
1238 case 15: TRACE("_d2"); break;
1239 case 1: TRACE("_x2"); break;
1240 case 2: TRACE("_x4"); break;
1241 case 3: TRACE("_x8"); break;
1242 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1243 }
1244
1245 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1246 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1247 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1248
1249 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1250 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1251}
1252
1253static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1254{
1255 struct wined3d_shader_version shader_version;
1256 const DWORD *ptr = byte_code;
1257 const char *type_prefix;
1258 DWORD i;
1259
1260 TRACE("Parsing %p.\n", byte_code);
1261
1262 fe->shader_read_header(fe_data, &ptr, &shader_version);
1263
1264 switch (shader_version.type)
1265 {
1266 case WINED3D_SHADER_TYPE_VERTEX:
1267 type_prefix = "vs";
1268 break;
1269
1270 case WINED3D_SHADER_TYPE_GEOMETRY:
1271 type_prefix = "gs";
1272 break;
1273
1274 case WINED3D_SHADER_TYPE_PIXEL:
1275 type_prefix = "ps";
1276 break;
1277
1278 default:
1279 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1280 type_prefix = "unknown";
1281 break;
1282 }
1283
1284 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1285
1286 while (!fe->shader_is_end(fe_data, &ptr))
1287 {
1288 struct wined3d_shader_instruction ins;
1289 const char *comment;
1290 UINT comment_size;
1291 UINT param_size;
1292
1293 /* comment */
1294 fe->shader_read_comment(&ptr, &comment, &comment_size);
1295 if (comment)
1296 {
1297 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1298 {
1299 const char *end = comment + comment_size;
1300 const char *cur = comment + 4;
1301 const char *line = cur;
1302
1303 TRACE("// TEXT\n");
1304 while (cur != end)
1305 {
1306 if (*cur == '\n')
1307 {
1308 UINT len = cur - line;
1309 if (len && *(cur - 1) == '\r') --len;
1310 TRACE("// %s\n", debugstr_an(line, len));
1311 line = ++cur;
1312 }
1313 else ++cur;
1314 }
1315 if (line != cur) TRACE("// %s\n", debugstr_an(line, cur - line));
1316 }
1317 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1318 continue;
1319 }
1320
1321 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1322 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1323 {
1324 TRACE("Skipping unrecognized instruction.\n");
1325 ptr += param_size;
1326 continue;
1327 }
1328
1329 if (ins.handler_idx == WINED3DSIH_DCL)
1330 {
1331 struct wined3d_shader_semantic semantic;
1332
1333 fe->shader_read_semantic(&ptr, &semantic);
1334
1335 shader_dump_decl_usage(&semantic, &shader_version);
1336 shader_dump_ins_modifiers(&semantic.reg);
1337 TRACE(" ");
1338 shader_dump_dst_param(&semantic.reg, &shader_version);
1339 }
1340 else if (ins.handler_idx == WINED3DSIH_DEF)
1341 {
1342 struct wined3d_shader_src_param rel_addr;
1343 struct wined3d_shader_dst_param dst;
1344
1345 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1346
1347 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1348 *(const float *)(ptr),
1349 *(const float *)(ptr + 1),
1350 *(const float *)(ptr + 2),
1351 *(const float *)(ptr + 3));
1352 ptr += 4;
1353 }
1354 else if (ins.handler_idx == WINED3DSIH_DEFI)
1355 {
1356 struct wined3d_shader_src_param rel_addr;
1357 struct wined3d_shader_dst_param dst;
1358
1359 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1360
1361 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1362 *(ptr),
1363 *(ptr + 1),
1364 *(ptr + 2),
1365 *(ptr + 3));
1366 ptr += 4;
1367 }
1368 else if (ins.handler_idx == WINED3DSIH_DEFB)
1369 {
1370 struct wined3d_shader_src_param rel_addr;
1371 struct wined3d_shader_dst_param dst;
1372
1373 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1374
1375 TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1376 ++ptr;
1377 }
1378 else
1379 {
1380 struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1381 struct wined3d_shader_dst_param dst_param;
1382 struct wined3d_shader_src_param src_param;
1383
1384 if (ins.dst_count)
1385 {
1386 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1387 }
1388
1389 /* Print out predication source token first - it follows
1390 * the destination token. */
1391 if (ins.predicate)
1392 {
1393 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1394 TRACE("(");
1395 shader_dump_src_param(&src_param, &shader_version);
1396 TRACE(") ");
1397 }
1398
1399 /* PixWin marks instructions with the coissue flag with a '+' */
1400 if (ins.coissue) TRACE("+");
1401
1402 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1403
1404 if (ins.handler_idx == WINED3DSIH_IFC
1405 || ins.handler_idx == WINED3DSIH_BREAKC)
1406 {
1407 switch (ins.flags)
1408 {
1409 case COMPARISON_GT: TRACE("_gt"); break;
1410 case COMPARISON_EQ: TRACE("_eq"); break;
1411 case COMPARISON_GE: TRACE("_ge"); break;
1412 case COMPARISON_LT: TRACE("_lt"); break;
1413 case COMPARISON_NE: TRACE("_ne"); break;
1414 case COMPARISON_LE: TRACE("_le"); break;
1415 default: TRACE("_(%u)", ins.flags);
1416 }
1417 }
1418 else if (ins.handler_idx == WINED3DSIH_TEX
1419 && shader_version.major >= 2
1420 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1421 {
1422 TRACE("p");
1423 }
1424
1425 /* We already read the destination token, print it. */
1426 if (ins.dst_count)
1427 {
1428 shader_dump_ins_modifiers(&dst_param);
1429 TRACE(" ");
1430 shader_dump_dst_param(&dst_param, &shader_version);
1431 }
1432
1433 /* Other source tokens */
1434 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1435 {
1436 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1437 TRACE(!i ? " " : ", ");
1438 shader_dump_src_param(&src_param, &shader_version);
1439 }
1440 }
1441 TRACE("\n");
1442 }
1443}
1444
1445static void shader_cleanup(IWineD3DBaseShader *iface)
1446{
1447 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1448
1449 ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1450 HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1451 HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1452 shader_delete_constant_list(&shader->baseShader.constantsF);
1453 shader_delete_constant_list(&shader->baseShader.constantsB);
1454 shader_delete_constant_list(&shader->baseShader.constantsI);
1455 list_remove(&shader->baseShader.shader_list_entry);
1456
1457 if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1458 {
1459 shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1460 }
1461}
1462
1463static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1464static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1465static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1466static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1467static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1468static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1469static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1470static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1471static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1472static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1473static void shader_none_free(IWineD3DDevice *iface) {}
1474static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1475
1476static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1477{
1478 /* Set the shader caps to 0 for the none shader backend */
1479 caps->VertexShaderVersion = 0;
1480 caps->PixelShaderVersion = 0;
1481 caps->PixelShader1xMaxValue = 0.0f;
1482}
1483
1484static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1485{
1486 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1487 {
1488 TRACE("Checking support for fixup:\n");
1489 dump_color_fixup_desc(fixup);
1490 }
1491
1492 /* Faked to make some apps happy. */
1493 if (!is_complex_fixup(fixup))
1494 {
1495 TRACE("[OK]\n");
1496 return TRUE;
1497 }
1498
1499 TRACE("[FAILED]\n");
1500 return FALSE;
1501}
1502
1503const shader_backend_t none_shader_backend = {
1504 shader_none_handle_instruction,
1505 shader_none_select,
1506 shader_none_select_depth_blt,
1507 shader_none_deselect_depth_blt,
1508 shader_none_update_float_vertex_constants,
1509 shader_none_update_float_pixel_constants,
1510 shader_none_load_constants,
1511 shader_none_load_np2fixup_constants,
1512 shader_none_destroy,
1513 shader_none_alloc,
1514 shader_none_free,
1515 shader_none_dirty_const,
1516 shader_none_get_caps,
1517 shader_none_color_fixup_supported,
1518};
1519
1520static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1521{
1522 *parent = shader->baseShader.parent;
1523 IUnknown_AddRef(*parent);
1524 TRACE("shader %p, returning %p.\n", shader, *parent);
1525}
1526
1527static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1528{
1529 if (!data)
1530 {
1531 *data_size = shader->baseShader.functionLength;
1532 return WINED3D_OK;
1533 }
1534
1535 if (*data_size < shader->baseShader.functionLength)
1536 {
1537 /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1538 * than the required size we should write the required size and
1539 * return D3DERR_MOREDATA. That's not actually true. */
1540 return WINED3DERR_INVALIDCALL;
1541 }
1542
1543 memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1544
1545 return WINED3D_OK;
1546}
1547
1548static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1549 const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1550{
1551 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1552 const struct wined3d_shader_frontend *fe;
1553 HRESULT hr;
1554
1555 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1556 shader, byte_code, output_signature, float_const_count);
1557
1558 fe = shader_select_frontend(*byte_code);
1559 if (!fe)
1560 {
1561 FIXME("Unable to find frontend for shader.\n");
1562 return WINED3DERR_INVALIDCALL;
1563 }
1564 shader->baseShader.frontend = fe;
1565 shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1566 if (!shader->baseShader.frontend_data)
1567 {
1568 FIXME("Failed to initialize frontend.\n");
1569 return WINED3DERR_INVALIDCALL;
1570 }
1571
1572 /* First pass: trace shader. */
1573 if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1574
1575 /* Initialize immediate constant lists. */
1576 list_init(&shader->baseShader.constantsF);
1577 list_init(&shader->baseShader.constantsB);
1578 list_init(&shader->baseShader.constantsI);
1579
1580 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1581 hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1582 reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1583 byte_code, float_const_count);
1584 if (FAILED(hr)) return hr;
1585
1586 shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1587 if (!shader->baseShader.function) return E_OUTOFMEMORY;
1588 memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1589
1590 return WINED3D_OK;
1591}
1592
1593static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1594{
1595 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1596
1597 if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1598 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1599 || IsEqualGUID(riid, &IID_IWineD3DBase)
1600 || IsEqualGUID(riid, &IID_IUnknown))
1601 {
1602 IUnknown_AddRef(iface);
1603 *object = iface;
1604 return S_OK;
1605 }
1606
1607 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1608
1609 *object = NULL;
1610 return E_NOINTERFACE;
1611}
1612
1613static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1614{
1615 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1616 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1617
1618 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1619
1620 return refcount;
1621}
1622
1623#ifdef VBOX_WINE_WITH_SHADER_CACHE
1624static void shader_chache_put(PVBOXEXT_HASHCACHE pCache, IWineD3DBaseShaderImpl *pShader);
1625#endif
1626
1627static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1628{
1629 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1630 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1631
1632 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1633
1634 if (!refcount)
1635 {
1636#ifdef VBOX_WINE_WITH_SHADER_CACHE
1637 shader_chache_put(&((IWineD3DDeviceImpl*)shader->baseShader.device)->vshaderCache, (IWineD3DBaseShaderImpl*)shader);
1638#else
1639 shader_cleanup((IWineD3DBaseShader *)iface);
1640#ifndef VBOX_WITH_VMSVGA
1641 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1642#endif
1643 HeapFree(GetProcessHeap(), 0, shader);
1644#endif
1645 }
1646
1647 return refcount;
1648}
1649
1650static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1651{
1652 TRACE("iface %p, parent %p.\n", iface, parent);
1653
1654 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1655
1656 return WINED3D_OK;
1657}
1658
1659static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1660{
1661 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1662
1663 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1664}
1665
1666/* Set local constants for d3d8 shaders. */
1667static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1668 UINT start_idx, const float *src_data, UINT count)
1669{
1670 IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1671 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1672 UINT i, end_idx;
1673
1674 TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1675
1676 end_idx = start_idx + count;
1677 if (end_idx > device->d3d_vshader_constantF)
1678 {
1679 WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1680 end_idx = device->d3d_vshader_constantF;
1681 }
1682
1683 for (i = start_idx; i < end_idx; ++i)
1684 {
1685 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1686 if (!lconst) return E_OUTOFMEMORY;
1687
1688 lconst->idx = i;
1689 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1690 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1691 }
1692
1693 return WINED3D_OK;
1694}
1695
1696static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1697{
1698 /* IUnknown methods */
1699 vertexshader_QueryInterface,
1700 vertexshader_AddRef,
1701 vertexshader_Release,
1702 /* IWineD3DBase methods */
1703 vertexshader_GetParent,
1704 /* IWineD3DBaseShader methods */
1705 vertexshader_GetFunction,
1706 /* IWineD3DVertexShader methods */
1707 vertexshader_SetLocalConstantsF,
1708};
1709
1710void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1711 IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1712{
1713 args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1714 args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1715 && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1716 args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1717}
1718
1719static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1720{
1721 if (usage_idx1 != usage_idx2) return FALSE;
1722 if (usage1 == usage2) return TRUE;
1723 if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1724 if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1725
1726 return FALSE;
1727}
1728
1729BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1730{
1731 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1732 WORD map = shader->baseShader.reg_maps.input_registers;
1733 unsigned int i;
1734
1735 for (i = 0; map; map >>= 1, ++i)
1736 {
1737 if (!(map & 1)) continue;
1738
1739 if (match_usage(shader->attributes[i].usage,
1740 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1741 {
1742 *regnum = i;
1743 return TRUE;
1744 }
1745 }
1746 return FALSE;
1747}
1748
1749static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1750{
1751 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1752 shader->baseShader.reg_maps.shader_version.minor);
1753 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1754
1755 shader->baseShader.limits.texcoord = 0;
1756 shader->baseShader.limits.attributes = 16;
1757 shader->baseShader.limits.packed_input = 0;
1758
1759 switch (shader_version)
1760 {
1761 case WINED3D_SHADER_VERSION(1, 0):
1762 case WINED3D_SHADER_VERSION(1, 1):
1763 shader->baseShader.limits.temporary = 12;
1764 shader->baseShader.limits.constant_bool = 0;
1765 shader->baseShader.limits.constant_int = 0;
1766 shader->baseShader.limits.address = 1;
1767 shader->baseShader.limits.packed_output = 0;
1768 shader->baseShader.limits.sampler = 0;
1769 shader->baseShader.limits.label = 0;
1770 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1771 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1772 * constants? */
1773 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1774 break;
1775
1776 case WINED3D_SHADER_VERSION(2, 0):
1777 case WINED3D_SHADER_VERSION(2, 1):
1778 shader->baseShader.limits.temporary = 12;
1779 shader->baseShader.limits.constant_bool = 16;
1780 shader->baseShader.limits.constant_int = 16;
1781 shader->baseShader.limits.address = 1;
1782 shader->baseShader.limits.packed_output = 0;
1783 shader->baseShader.limits.sampler = 0;
1784 shader->baseShader.limits.label = 16;
1785 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1786 break;
1787
1788 case WINED3D_SHADER_VERSION(4, 0):
1789 FIXME("Using 3.0 limits for 4.0 shader.\n");
1790 RT_FALL_THRU();
1791
1792 case WINED3D_SHADER_VERSION(3, 0):
1793 shader->baseShader.limits.temporary = 32;
1794 shader->baseShader.limits.constant_bool = 32;
1795 shader->baseShader.limits.constant_int = 32;
1796 shader->baseShader.limits.address = 1;
1797 shader->baseShader.limits.packed_output = 12;
1798 shader->baseShader.limits.sampler = 4;
1799 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1800 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1801 * even though they are capable of supporting much more (GL
1802 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1803 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1804 * shaders to 256. */
1805 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1806 break;
1807
1808 default:
1809 shader->baseShader.limits.temporary = 12;
1810 shader->baseShader.limits.constant_bool = 16;
1811 shader->baseShader.limits.constant_int = 16;
1812 shader->baseShader.limits.address = 1;
1813 shader->baseShader.limits.packed_output = 0;
1814 shader->baseShader.limits.sampler = 0;
1815 shader->baseShader.limits.label = 16;
1816 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1817 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1818 shader->baseShader.reg_maps.shader_version.major,
1819 shader->baseShader.reg_maps.shader_version.minor);
1820 }
1821}
1822
1823HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1824 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1825 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1826{
1827 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1828 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1829 unsigned int i;
1830 HRESULT hr;
1831 WORD map;
1832
1833 if (!byte_code) return WINED3DERR_INVALIDCALL;
1834
1835 shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1836 shader_init(&shader->baseShader, device, parent, parent_ops);
1837
1838 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1839 output_signature, device->d3d_vshader_constantF);
1840 if (FAILED(hr))
1841 {
1842 WARN("Failed to set function, hr %#x.\n", hr);
1843 shader_cleanup((IWineD3DBaseShader *)shader);
1844 return hr;
1845 }
1846
1847 map = shader->baseShader.reg_maps.input_registers;
1848 for (i = 0; map; map >>= 1, ++i)
1849 {
1850 if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1851
1852 shader->attributes[i].usage =
1853 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1854 shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1855 }
1856
1857 if (output_signature)
1858 {
1859 for (i = 0; i < output_signature->element_count; ++i)
1860 {
1861 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1862 reg_maps->output_registers |= 1 << e->register_idx;
1863 shader->baseShader.output_signature[e->register_idx] = *e;
1864 }
1865 }
1866
1867 vertexshader_set_limits(shader);
1868
1869 if (device->vs_selected_mode == SHADER_ARB
1870 && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1871 && shader->min_rel_offset <= shader->max_rel_offset)
1872 {
1873 if (shader->max_rel_offset - shader->min_rel_offset > 127)
1874 {
1875 FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1876 FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1877 FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1878 }
1879 else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1880 {
1881 shader->rel_offset = shader->min_rel_offset + 63;
1882 }
1883 else if (shader->max_rel_offset > 63)
1884 {
1885 shader->rel_offset = shader->min_rel_offset;
1886 }
1887 else
1888 {
1889 shader->rel_offset = 0;
1890 }
1891 }
1892
1893 shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1894 && !list_empty(&shader->baseShader.constantsF);
1895
1896 return WINED3D_OK;
1897}
1898
1899static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1900 REFIID riid, void **object)
1901{
1902 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1903
1904 if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1905 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1906 || IsEqualGUID(riid, &IID_IWineD3DBase)
1907 || IsEqualGUID(riid, &IID_IUnknown))
1908 {
1909 IUnknown_AddRef(iface);
1910 *object = iface;
1911 return S_OK;
1912 }
1913
1914 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1915
1916 *object = NULL;
1917 return E_NOINTERFACE;
1918}
1919
1920static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1921{
1922 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1923 ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1924
1925 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1926
1927 return refcount;
1928}
1929
1930static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1931{
1932 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1933 ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1934
1935 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1936
1937 if (!refcount)
1938 {
1939 shader_cleanup((IWineD3DBaseShader *)iface);
1940 shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1941 HeapFree(GetProcessHeap(), 0, shader);
1942 }
1943
1944 return refcount;
1945}
1946
1947static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1948{
1949 TRACE("iface %p, parent %p.\n", iface, parent);
1950
1951 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1952
1953 return WINED3D_OK;
1954}
1955
1956static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1957{
1958 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1959
1960 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1961}
1962
1963static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1964{
1965 /* IUnknown methods */
1966 geometryshader_QueryInterface,
1967 geometryshader_AddRef,
1968 geometryshader_Release,
1969 /* IWineD3DBase methods */
1970 geometryshader_GetParent,
1971 /* IWineD3DBaseShader methods */
1972 geometryshader_GetFunction,
1973};
1974
1975HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1976 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1977 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1978{
1979 HRESULT hr;
1980
1981 shader->vtbl = &wined3d_geometryshader_vtbl;
1982 shader_init(&shader->base_shader, device, parent, parent_ops);
1983
1984 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1985 if (FAILED(hr))
1986 {
1987 WARN("Failed to set function, hr %#x.\n", hr);
1988 shader_cleanup((IWineD3DBaseShader *)shader);
1989 return hr;
1990 }
1991
1992 shader->base_shader.load_local_constsF = FALSE;
1993
1994 return WINED3D_OK;
1995}
1996
1997static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1998{
1999 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
2000
2001 if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
2002 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
2003 || IsEqualGUID(riid, &IID_IWineD3DBase)
2004 || IsEqualGUID(riid, &IID_IUnknown))
2005 {
2006 IUnknown_AddRef(iface);
2007 *object = iface;
2008 return S_OK;
2009 }
2010
2011 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
2012
2013 *object = NULL;
2014 return E_NOINTERFACE;
2015}
2016
2017static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
2018{
2019 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2020 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
2021
2022 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2023
2024 return refcount;
2025}
2026
2027static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
2028{
2029 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2030 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
2031
2032 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2033
2034 if (!refcount)
2035 {
2036#ifdef VBOX_WINE_WITH_SHADER_CACHE
2037 shader_chache_put(&((IWineD3DDeviceImpl*)shader->baseShader.device)->pshaderCache, (IWineD3DBaseShaderImpl*)shader);
2038#else
2039 shader_cleanup((IWineD3DBaseShader *)iface);
2040#ifndef VBOX_WITH_VMSVGA
2041 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
2042#endif
2043 HeapFree(GetProcessHeap(), 0, shader);
2044#endif
2045 }
2046
2047 return refcount;
2048}
2049
2050static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
2051{
2052 TRACE("iface %p, parent %p.\n", iface, parent);
2053
2054 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
2055
2056 return WINED3D_OK;
2057}
2058
2059static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
2060{
2061 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
2062
2063 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
2064}
2065
2066static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2067{
2068 /* IUnknown methods */
2069 pixelshader_QueryInterface,
2070 pixelshader_AddRef,
2071 pixelshader_Release,
2072 /* IWineD3DBase methods */
2073 pixelshader_GetParent,
2074 /* IWineD3DBaseShader methods */
2075 pixelshader_GetFunction
2076};
2077
2078void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
2079 IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
2080{
2081 IWineD3DBaseTextureImpl *texture;
2082 UINT i;
2083
2084 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2085 args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
2086 args->np2_fixup = 0;
2087 args->t_mirror = 0;
2088
2089 if (shader->baseShader.reg_maps.shader_version.major==1
2090 && shader->baseShader.reg_maps.shader_version.minor<=3)
2091 {
2092 for (i=0; i<4; ++i)
2093 {
2094 DWORD flags = stateblock->textureState[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
2095 DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
2096 if (flags & WINED3DTTFF_PROJECTED)
2097 {
2098 tex_transform |= WINED3D_PSARGS_PROJECTED;
2099 }
2100 args->tex_transform |= tex_transform << (i*WINED3D_PSARGS_TEXTRANSFORM_SHIFT);
2101 }
2102 }
2103
2104 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2105 {
2106 if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2107 texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
2108 if (!texture)
2109 {
2110 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2111 continue;
2112 }
2113 args->color_fixup[i] = texture->resource.format_desc->color_fixup;
2114
2115 /* Flag samplers that need NP2 texcoord fixup. */
2116 if (!texture->baseTexture.pow2Matrix_identity)
2117 {
2118 args->np2_fixup |= (1 << i);
2119 }
2120
2121 if (texture->baseTexture.t_mirror)
2122 {
2123 args->t_mirror |= (1 << i);
2124 }
2125 }
2126 if (shader->baseShader.reg_maps.shader_version.major >= 3)
2127 {
2128 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2129 {
2130 args->vp_mode = pretransformed;
2131 }
2132 else if (use_vs(stateblock))
2133 {
2134 args->vp_mode = vertexshader;
2135 }
2136 else
2137 {
2138 args->vp_mode = fixedfunction;
2139 }
2140 args->fog = FOG_OFF;
2141 }
2142 else
2143 {
2144 args->vp_mode = vertexshader;
2145 if (stateblock->renderState[WINED3DRS_FOGENABLE])
2146 {
2147 switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2148 {
2149 case WINED3DFOG_NONE:
2150 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2151 || use_vs(stateblock))
2152 {
2153 args->fog = FOG_LINEAR;
2154 break;
2155 }
2156
2157 switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2158 {
2159 case WINED3DFOG_NONE: RT_FALL_THRU();
2160 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2161 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2162 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2163 }
2164 break;
2165
2166 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2167 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2168 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2169 }
2170 }
2171 else
2172 {
2173 args->fog = FOG_OFF;
2174 }
2175 }
2176}
2177
2178static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2179{
2180 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2181 shader->baseShader.reg_maps.shader_version.minor);
2182
2183 shader->baseShader.limits.attributes = 0;
2184 shader->baseShader.limits.address = 0;
2185 shader->baseShader.limits.packed_output = 0;
2186
2187 switch (shader_version)
2188 {
2189 case WINED3D_SHADER_VERSION(1, 0):
2190 case WINED3D_SHADER_VERSION(1, 1):
2191 case WINED3D_SHADER_VERSION(1, 2):
2192 case WINED3D_SHADER_VERSION(1, 3):
2193 shader->baseShader.limits.temporary = 2;
2194 shader->baseShader.limits.constant_float = 8;
2195 shader->baseShader.limits.constant_int = 0;
2196 shader->baseShader.limits.constant_bool = 0;
2197 shader->baseShader.limits.texcoord = 4;
2198 shader->baseShader.limits.sampler = 4;
2199 shader->baseShader.limits.packed_input = 0;
2200 shader->baseShader.limits.label = 0;
2201 break;
2202
2203 case WINED3D_SHADER_VERSION(1, 4):
2204 shader->baseShader.limits.temporary = 6;
2205 shader->baseShader.limits.constant_float = 8;
2206 shader->baseShader.limits.constant_int = 0;
2207 shader->baseShader.limits.constant_bool = 0;
2208 shader->baseShader.limits.texcoord = 6;
2209 shader->baseShader.limits.sampler = 6;
2210 shader->baseShader.limits.packed_input = 0;
2211 shader->baseShader.limits.label = 0;
2212 break;
2213
2214 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2215 case WINED3D_SHADER_VERSION(2, 0):
2216 shader->baseShader.limits.temporary = 32;
2217 shader->baseShader.limits.constant_float = 32;
2218 shader->baseShader.limits.constant_int = 16;
2219 shader->baseShader.limits.constant_bool = 16;
2220 shader->baseShader.limits.texcoord = 8;
2221 shader->baseShader.limits.sampler = 16;
2222 shader->baseShader.limits.packed_input = 0;
2223 break;
2224
2225 case WINED3D_SHADER_VERSION(2, 1):
2226 shader->baseShader.limits.temporary = 32;
2227 shader->baseShader.limits.constant_float = 32;
2228 shader->baseShader.limits.constant_int = 16;
2229 shader->baseShader.limits.constant_bool = 16;
2230 shader->baseShader.limits.texcoord = 8;
2231 shader->baseShader.limits.sampler = 16;
2232 shader->baseShader.limits.packed_input = 0;
2233 shader->baseShader.limits.label = 16;
2234 break;
2235
2236 case WINED3D_SHADER_VERSION(4, 0):
2237 FIXME("Using 3.0 limits for 4.0 shader.\n");
2238 RT_FALL_THRU();
2239
2240 case WINED3D_SHADER_VERSION(3, 0):
2241 shader->baseShader.limits.temporary = 32;
2242 shader->baseShader.limits.constant_float = 224;
2243 shader->baseShader.limits.constant_int = 16;
2244 shader->baseShader.limits.constant_bool = 16;
2245 shader->baseShader.limits.texcoord = 0;
2246 shader->baseShader.limits.sampler = 16;
2247 shader->baseShader.limits.packed_input = 12;
2248 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2249 break;
2250
2251 default:
2252 shader->baseShader.limits.temporary = 32;
2253 shader->baseShader.limits.constant_float = 32;
2254 shader->baseShader.limits.constant_int = 16;
2255 shader->baseShader.limits.constant_bool = 16;
2256 shader->baseShader.limits.texcoord = 8;
2257 shader->baseShader.limits.sampler = 16;
2258 shader->baseShader.limits.packed_input = 0;
2259 shader->baseShader.limits.label = 0;
2260 FIXME("Unrecognized pixel shader version %u.%u\n",
2261 shader->baseShader.reg_maps.shader_version.major,
2262 shader->baseShader.reg_maps.shader_version.minor);
2263 }
2264}
2265
2266HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2267 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2268 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2269{
2270 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2271 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2272 HRESULT hr;
2273
2274 if (!byte_code) return WINED3DERR_INVALIDCALL;
2275
2276 shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2277 shader_init(&shader->baseShader, device, parent, parent_ops);
2278
2279 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2280 output_signature, device->d3d_pshader_constantF);
2281 if (FAILED(hr))
2282 {
2283 WARN("Failed to set function, hr %#x.\n", hr);
2284 shader_cleanup((IWineD3DBaseShader *)shader);
2285 return hr;
2286 }
2287
2288 pixelshader_set_limits(shader);
2289
2290 for (i = 0; i < MAX_REG_INPUT; ++i)
2291 {
2292 if (shader->input_reg_used[i])
2293 {
2294 ++num_regs_used;
2295 highest_reg_used = i;
2296 }
2297 }
2298
2299 /* Don't do any register mapping magic if it is not needed, or if we can't
2300 * achieve anything anyway */
2301 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2302 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2303 {
2304 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2305 {
2306 /* This happens with relative addressing. The input mapper function
2307 * warns about this if the higher registers are declared too, so
2308 * don't write a FIXME here */
2309 WARN("More varying registers used than supported\n");
2310 }
2311
2312 for (i = 0; i < MAX_REG_INPUT; ++i)
2313 {
2314 shader->input_reg_map[i] = i;
2315 }
2316
2317 shader->declared_in_count = highest_reg_used + 1;
2318 }
2319 else
2320 {
2321 shader->declared_in_count = 0;
2322 for (i = 0; i < MAX_REG_INPUT; ++i)
2323 {
2324 if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2325 else shader->input_reg_map[i] = ~0U;
2326 }
2327 }
2328
2329 shader->baseShader.load_local_constsF = FALSE;
2330
2331 return WINED3D_OK;
2332}
2333
2334void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2335{
2336 WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2337 unsigned int i;
2338
2339 if (reg_maps->shader_version.major != 1) return;
2340
2341 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2342 {
2343 /* We don't sample from this sampler. */
2344 if (!sampler_type[i]) continue;
2345
2346 if (!textures[i])
2347 {
2348 WARN("No texture bound to sampler %u, using 2D.\n", i);
2349 sampler_type[i] = WINED3DSTT_2D;
2350 continue;
2351 }
2352
2353 switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2354 {
2355 case GL_TEXTURE_RECTANGLE_ARB:
2356 case GL_TEXTURE_2D:
2357 /* We have to select between texture rectangles and 2D
2358 * textures later because 2.0 and 3.0 shaders only have
2359 * WINED3DSTT_2D as well. */
2360 sampler_type[i] = WINED3DSTT_2D;
2361 break;
2362
2363 case GL_TEXTURE_3D:
2364 sampler_type[i] = WINED3DSTT_VOLUME;
2365 break;
2366
2367 case GL_TEXTURE_CUBE_MAP_ARB:
2368 sampler_type[i] = WINED3DSTT_CUBE;
2369 break;
2370
2371 default:
2372 FIXME("Unrecognized texture type %#x, using 2D.\n",
2373 IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2374 sampler_type[i] = WINED3DSTT_2D;
2375 }
2376 }
2377}
2378
2379#ifdef VBOX_WINE_WITH_SHADER_CACHE
2380
2381static void shader_reinit(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
2382 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2383{
2384 shader->ref = 1;
2385 shader->device = (IWineD3DDevice *)device;
2386 shader->parent = parent;
2387 shader->parent_ops = parent_ops;
2388#ifdef DEBUG
2389 {
2390 IWineD3DBaseShaderImpl *tstShader;
2391 LIST_FOR_EACH_ENTRY(tstShader, &device->shaders, IWineD3DBaseShaderImpl, baseShader.shader_list_entry) {
2392 if (&tstShader->baseShader == shader)
2393 return;
2394 }
2395 ERR("shader not in list!");
2396 }
2397#endif
2398}
2399
2400static DECLCALLBACK(uint32_t) shader_cache_hash(void *pvKey)
2401{
2402 uint32_t u32Hash, i;
2403 DWORD *function;
2404 IWineD3DBaseShaderImpl *pShader = (IWineD3DBaseShaderImpl *)pvKey;
2405 if (pShader->baseShader.u32CacheDataInited)
2406 return pShader->baseShader.u32Hash;
2407
2408 u32Hash = pShader->baseShader.functionLength;
2409 function = pShader->baseShader.function;
2410 for (i = 0; i < pShader->baseShader.functionLength / 4; ++i)
2411 {
2412 u32Hash += *function++;
2413 }
2414
2415 for (i = 0; i < pShader->baseShader.functionLength % 4; ++i)
2416 {
2417 u32Hash += *((uint8_t*)function)++;
2418 }
2419
2420 pShader->baseShader.u32Hash = u32Hash;
2421 pShader->baseShader.u32CacheDataInited = TRUE;
2422 return pShader->baseShader.u32Hash;
2423}
2424
2425static DECLCALLBACK(bool) shader_cache_equal(void *pvKey1, void *pvKey2)
2426{
2427 IWineD3DBaseShaderImpl *pShader1 = (IWineD3DBaseShaderImpl *)pvKey1;
2428 IWineD3DBaseShaderImpl *pShader2 = (IWineD3DBaseShaderImpl *)pvKey2;
2429
2430 if (pShader1 == pShader2)
2431 return TRUE;
2432
2433 if (pShader1->baseShader.functionLength != pShader2->baseShader.functionLength)
2434 return FALSE;
2435
2436 if (shader_cache_hash(pvKey1) != shader_cache_hash(pvKey2))
2437 {
2438#ifdef DEBUG_misha
2439 Assert(memcmp(pShader1->baseShader.function, pShader2->baseShader.function, pShader1->baseShader.functionLength));
2440#endif
2441 return FALSE;
2442 }
2443
2444 return !memcmp(pShader1->baseShader.function, pShader2->baseShader.function, pShader1->baseShader.functionLength);
2445}
2446
2447#define VBOX_SHADER_FROM_CACHE_ENTRY(_pEntry) RT_FROM_MEMBER(RT_FROM_MEMBER((_pEntry), IWineD3DBaseShaderClass, CacheEntry), IWineD3DBaseShaderImpl, baseShader)
2448
2449static DECLCALLBACK(void) shader_cache_cleanup_entry(void *pvKey, struct VBOXEXT_HASHCACHE_ENTRY *pEntry)
2450{
2451 IWineD3DBaseShaderImpl *pShader = VBOX_SHADER_FROM_CACHE_ENTRY(pEntry);
2452 shader_cleanup((IWineD3DBaseShader *)pShader);
2453 if (pShader->baseShader.parent)
2454 pShader->baseShader.parent_ops->wined3d_object_destroyed(pShader->baseShader.parent);
2455 HeapFree(GetProcessHeap(), 0, pShader);
2456}
2457
2458static void shader_chache_init(PVBOXEXT_HASHCACHE pCache)
2459{
2460 VBoxExtCacheInit(pCache, 1024, shader_cache_hash, shader_cache_equal, shader_cache_cleanup_entry);
2461}
2462
2463static void shader_chache_term(PVBOXEXT_HASHCACHE pCache)
2464{
2465 VBoxExtCacheTerm(pCache);
2466}
2467
2468static void shader_chache_put(PVBOXEXT_HASHCACHE pCache, IWineD3DBaseShaderImpl *pShader)
2469{
2470 if (pShader->baseShader.parent)
2471 {
2472 pShader->baseShader.parent_ops->wined3d_object_destroyed(pShader->baseShader.parent);
2473 pShader->baseShader.parent = NULL;
2474 }
2475
2476 VBoxExtCachePut(pCache, pShader, &pShader->baseShader.CacheEntry);
2477}
2478
2479static IWineD3DBaseShaderImpl* shader_chache_get(PVBOXEXT_HASHCACHE pCache, IWineD3DBaseShaderImpl *pShader)
2480{
2481 PVBOXEXT_HASHCACHE_ENTRY pEntry = VBoxExtCacheGet(pCache, pShader);
2482 IWineD3DBaseShaderImpl *pCachedShader;
2483
2484 if (!pEntry)
2485 return pShader;
2486
2487 pCachedShader = VBOX_SHADER_FROM_CACHE_ENTRY(pEntry);
2488 shader_reinit(&pCachedShader->baseShader, (IWineD3DDeviceImpl *)pShader->baseShader.device,
2489 pShader->baseShader.parent, pShader->baseShader.parent_ops);
2490 pShader->baseShader.parent = NULL;
2491 /* we can not do a IWineD3DBaseShader_Release here since this would result in putting a shader to the cache */
2492 shader_cleanup((IWineD3DBaseShader *)pShader);
2493 HeapFree(GetProcessHeap(), 0, pShader);
2494 return pCachedShader;
2495}
2496
2497IWineD3DVertexShaderImpl * vertexshader_check_cached(IWineD3DDeviceImpl *device, IWineD3DVertexShaderImpl *object)
2498{
2499 return (IWineD3DVertexShaderImpl*)shader_chache_get(&device->vshaderCache, (IWineD3DBaseShaderImpl *)object);
2500}
2501
2502IWineD3DPixelShaderImpl * pixelshader_check_cached(IWineD3DDeviceImpl *device, IWineD3DPixelShaderImpl *object)
2503{
2504 return (IWineD3DPixelShaderImpl*)shader_chache_get(&device->pshaderCache, (IWineD3DBaseShaderImpl *)object);
2505}
2506
2507void shader_chaches_init(IWineD3DDeviceImpl *device)
2508{
2509 shader_chache_init(&device->vshaderCache);
2510 shader_chache_init(&device->pshaderCache);
2511}
2512
2513void shader_chaches_term(IWineD3DDeviceImpl *device)
2514{
2515 shader_chache_term(&device->vshaderCache);
2516 shader_chache_term(&device->pshaderCache);
2517}
2518#endif
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette