?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Don't use the same va_list multiple times in shader_vaddline().
[wine/multimedia.git] / dlls / wined3d / shader.c
blob279ec423f68e6c71dbcae7f6905fbce19c42b6dd
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, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <math.h>
29 #include <stdio.h>
30 #include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_BEM */ "bem",
47 /* WINED3DSIH_BREAK */ "break",
48 /* WINED3DSIH_BREAKC */ "breakc",
49 /* WINED3DSIH_BREAKP */ "breakp",
50 /* WINED3DSIH_CALL */ "call",
51 /* WINED3DSIH_CALLNZ */ "callnz",
52 /* WINED3DSIH_CMP */ "cmp",
53 /* WINED3DSIH_CND */ "cnd",
54 /* WINED3DSIH_CRS */ "crs",
55 /* WINED3DSIH_CUT */ "cut",
56 /* WINED3DSIH_DCL */ "dcl",
57 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
58 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
59 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
60 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
61 /* WINED3DSIH_DEF */ "def",
62 /* WINED3DSIH_DEFB */ "defb",
63 /* WINED3DSIH_DEFI */ "defi",
64 /* WINED3DSIH_DIV */ "div",
65 /* WINED3DSIH_DP2 */ "dp2",
66 /* WINED3DSIH_DP2ADD */ "dp2add",
67 /* WINED3DSIH_DP3 */ "dp3",
68 /* WINED3DSIH_DP4 */ "dp4",
69 /* WINED3DSIH_DST */ "dst",
70 /* WINED3DSIH_DSX */ "dsx",
71 /* WINED3DSIH_DSY */ "dsy",
72 /* WINED3DSIH_ELSE */ "else",
73 /* WINED3DSIH_EMIT */ "emit",
74 /* WINED3DSIH_ENDIF */ "endif",
75 /* WINED3DSIH_ENDLOOP */ "endloop",
76 /* WINED3DSIH_ENDREP */ "endrep",
77 /* WINED3DSIH_EQ */ "eq",
78 /* WINED3DSIH_EXP */ "exp",
79 /* WINED3DSIH_EXPP */ "expp",
80 /* WINED3DSIH_FRC */ "frc",
81 /* WINED3DSIH_FTOI */ "ftoi",
82 /* WINED3DSIH_GE */ "ge",
83 /* WINED3DSIH_IADD */ "iadd",
84 /* WINED3DSIH_IEQ */ "ieq",
85 /* WINED3DSIH_IF */ "if",
86 /* WINED3DSIH_IFC */ "ifc",
87 /* WINED3DSIH_IGE */ "ige",
88 /* WINED3DSIH_IMUL */ "imul",
89 /* WINED3DSIH_ISHL */ "ishl",
90 /* WINED3DSIH_ITOF */ "itof",
91 /* WINED3DSIH_LABEL */ "label",
92 /* WINED3DSIH_LD */ "ld",
93 /* WINED3DSIH_LIT */ "lit",
94 /* WINED3DSIH_LOG */ "log",
95 /* WINED3DSIH_LOGP */ "logp",
96 /* WINED3DSIH_LOOP */ "loop",
97 /* WINED3DSIH_LRP */ "lrp",
98 /* WINED3DSIH_LT */ "lt",
99 /* WINED3DSIH_M3x2 */ "m3x2",
100 /* WINED3DSIH_M3x3 */ "m3x3",
101 /* WINED3DSIH_M3x4 */ "m3x4",
102 /* WINED3DSIH_M4x3 */ "m4x3",
103 /* WINED3DSIH_M4x4 */ "m4x4",
104 /* WINED3DSIH_MAD */ "mad",
105 /* WINED3DSIH_MAX */ "max",
106 /* WINED3DSIH_MIN */ "min",
107 /* WINED3DSIH_MOV */ "mov",
108 /* WINED3DSIH_MOVA */ "mova",
109 /* WINED3DSIH_MOVC */ "movc",
110 /* WINED3DSIH_MUL */ "mul",
111 /* WINED3DSIH_NE */ "ne",
112 /* WINED3DSIH_NOP */ "nop",
113 /* WINED3DSIH_NRM */ "nrm",
114 /* WINED3DSIH_OR */ "or",
115 /* WINED3DSIH_PHASE */ "phase",
116 /* WINED3DSIH_POW */ "pow",
117 /* WINED3DSIH_RCP */ "rcp",
118 /* WINED3DSIH_REP */ "rep",
119 /* WINED3DSIH_RET */ "ret",
120 /* WINED3DSIH_ROUND_NI */ "round_ni",
121 /* WINED3DSIH_RSQ */ "rsq",
122 /* WINED3DSIH_SAMPLE */ "sample",
123 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
124 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
125 /* WINED3DSIH_SETP */ "setp",
126 /* WINED3DSIH_SGE */ "sge",
127 /* WINED3DSIH_SGN */ "sgn",
128 /* WINED3DSIH_SINCOS */ "sincos",
129 /* WINED3DSIH_SLT */ "slt",
130 /* WINED3DSIH_SQRT */ "sqrt",
131 /* WINED3DSIH_SUB */ "sub",
132 /* WINED3DSIH_TEX */ "texld",
133 /* WINED3DSIH_TEXBEM */ "texbem",
134 /* WINED3DSIH_TEXBEML */ "texbeml",
135 /* WINED3DSIH_TEXCOORD */ "texcrd",
136 /* WINED3DSIH_TEXDEPTH */ "texdepth",
137 /* WINED3DSIH_TEXDP3 */ "texdp3",
138 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
139 /* WINED3DSIH_TEXKILL */ "texkill",
140 /* WINED3DSIH_TEXLDD */ "texldd",
141 /* WINED3DSIH_TEXLDL */ "texldl",
142 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
143 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
144 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
145 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
146 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
147 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
148 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
149 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
150 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
151 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
152 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
153 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
154 /* WINED3DSIH_UDIV */ "udiv",
155 /* WINED3DSIH_UGE */ "uge",
156 /* WINED3DSIH_USHR */ "ushr",
157 /* WINED3DSIH_UTOF */ "utof",
158 /* WINED3DSIH_XOR */ "xor",
159 };
160
161 static const char * const semantic_names[] =
162 {
163 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
164 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
165 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
166 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
167 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
168 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
169 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
170 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
171 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
172 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
173 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
174 /* WINED3D_DECL_USAGE_FOG */ "FOG",
175 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
176 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
177 };
178
179 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
180 {
181 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
182 {
183 FIXME("Unrecognized usage %#x.\n", usage);
184 return "UNRECOGNIZED";
185 }
186
187 return semantic_names[usage];
188 }
189
190 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
191 {
192 unsigned int i;
193
194 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
195 {
196 if (!strcmp(name, semantic_names[i])) return i;
197 }
198
199 return ~0U;
200 }
201
202 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
203 {
204 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
205 }
206
207 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
208 const struct wined3d_shader_semantic *s)
209 {
210 e->semantic_name = shader_semantic_name_from_usage(s->usage);
211 e->semantic_idx = s->usage_idx;
212 e->sysval_semantic = 0;
213 e->component_type = 0;
214 e->register_idx = s->reg.reg.idx[0].offset;
215 e->mask = s->reg.write_mask;
216 }
217
218 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
219 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
220 {
221 e->semantic_name = shader_semantic_name_from_usage(usage);
222 e->semantic_idx = usage_idx;
223 e->sysval_semantic = 0;
224 e->component_type = 0;
225 e->register_idx = reg_idx;
226 e->mask = write_mask;
227 }
228
229 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
230 {
231 switch (version_token >> 16)
232 {
233 case WINED3D_SM1_VS:
234 case WINED3D_SM1_PS:
235 return &sm1_shader_frontend;
236
237 case WINED3D_SM4_PS:
238 case WINED3D_SM4_VS:
239 case WINED3D_SM4_GS:
240 return &sm4_shader_frontend;
241
242 default:
243 FIXME("Unrecognised version token %#x\n", version_token);
244 return NULL;
245 }
246 }
247
248 void string_buffer_clear(struct wined3d_string_buffer *buffer)
249 {
250 buffer->buffer[0] = '\0';
251 buffer->content_size = 0;
252 }
253
254 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
255 {
256 buffer->buffer_size = 32;
257 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
258 {
259 ERR("Failed to allocate shader buffer memory.\n");
260 return FALSE;
261 }
262
263 string_buffer_clear(buffer);
264 return TRUE;
265 }
266
267 void string_buffer_free(struct wined3d_string_buffer *buffer)
268 {
269 HeapFree(GetProcessHeap(), 0, buffer->buffer);
270 }
271
272 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
273 {
274 char *new_buffer;
275 unsigned int new_buffer_size = buffer->buffer_size * 2;
276
277 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
278 new_buffer_size *= 2;
279 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
280 {
281 ERR("Failed to grow buffer.\n");
282 buffer->buffer[buffer->content_size] = '\0';
283 return FALSE;
284 }
285 buffer->buffer = new_buffer;
286 buffer->buffer_size = new_buffer_size;
287 return TRUE;
288 }
289
290 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
291 {
292 unsigned int rem;
293 int rc;
294
295 rem = buffer->buffer_size - buffer->content_size;
296 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
297 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
298 return rc;
299
300 buffer->content_size += rc;
301 return 0;
302 }
303
304 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
305 {
306 va_list args;
307 int ret;
308
309 for (;;)
310 {
311 va_start(args, format);
312 ret = shader_vaddline(buffer, format, args);
313 va_end(args);
314 if (!ret)
315 return ret;
316 if (!string_buffer_resize(buffer, ret))
317 return -1;
318 }
319 }
320
321 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
322 {
323 struct wined3d_string_buffer *buffer;
324
325 if (list_empty(&list->list))
326 {
327 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
328 if (!buffer || !string_buffer_init(buffer))
329 {
330 ERR("Couldn't allocate buffer for temporary string.\n");
331 if (buffer)
332 HeapFree(GetProcessHeap(), 0, buffer);
333 return NULL;
334 }
335 }
336 else
337 {
338 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
339 list_remove(&buffer->entry);
340 }
341 string_buffer_clear(buffer);
342 return buffer;
343 }
344
345 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
346 {
347 if (!buffer)
348 return 0;
349 string_buffer_clear(buffer);
350 return shader_vaddline(buffer, format, args);
351 }
352
353 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
354 {
355 va_list args;
356 int ret;
357
358 for (;;)
359 {
360 va_start(args, format);
361 ret = string_buffer_vsprintf(buffer, format, args);
362 va_end(args);
363 if (!ret)
364 return;
365 if (!string_buffer_resize(buffer, ret))
366 return;
367 }
368 }
369
370 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
371 {
372 if (!buffer)
373 return;
374 list_add_head(&list->list, &buffer->entry);
375 }
376
377 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
378 {
379 list_init(&list->list);
380 }
381
382 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
383 {
384 struct wined3d_string_buffer *buffer, *buffer_next;
385
386 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
387 {
388 string_buffer_free(buffer);
389 HeapFree(GetProcessHeap(), 0, buffer);
390 }
391 list_init(&list->list);
392 }
393
394 /* Convert floating point offset relative to a register file to an absolute
395 * offset for float constants. */
396 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
397 {
398 switch (register_type)
399 {
400 case WINED3DSPR_CONST: return register_idx;
401 case WINED3DSPR_CONST2: return 2048 + register_idx;
402 case WINED3DSPR_CONST3: return 4096 + register_idx;
403 case WINED3DSPR_CONST4: return 6144 + register_idx;
404 default:
405 FIXME("Unsupported register type: %u.\n", register_type);
406 return register_idx;
407 }
408 }
409
410 static void shader_delete_constant_list(struct list *clist)
411 {
412 struct wined3d_shader_lconst *constant;
413 struct list *ptr;
414
415 ptr = list_head(clist);
416 while (ptr)
417 {
418 constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
419 ptr = list_next(clist, ptr);
420 HeapFree(GetProcessHeap(), 0, constant);
421 }
422 list_init(clist);
423 }
424
425 static void shader_set_limits(struct wined3d_shader *shader)
426 {
427 static const struct limits_entry
428 {
429 unsigned int min_version;
430 unsigned int max_version;
431 struct wined3d_shader_limits limits;
432 }
433 vs_limits[] =
434 {
435 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
436 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
437 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
438 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
439 * even though they are capable of supporting much more (GL
440 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
441 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
442 * shaders to 256. */
443 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
444 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
445 {0}
446 },
447 gs_limits[] =
448 {
449 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
450 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
451 {0}
452 },
453 ps_limits[] =
454 {
455 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
456 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
457 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
458 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
459 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
460 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
461 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 0, 32}},
462 {0}
463 };
464 const struct limits_entry *limits_array;
465 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
466 shader->reg_maps.shader_version.minor);
467 int i = 0;
468
469 switch (shader->reg_maps.shader_version.type)
470 {
471 default:
472 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
473 /* Fall-through. */
474 case WINED3D_SHADER_TYPE_VERTEX:
475 limits_array = vs_limits;
476 break;
477 case WINED3D_SHADER_TYPE_GEOMETRY:
478 limits_array = gs_limits;
479 break;
480 case WINED3D_SHADER_TYPE_PIXEL:
481 limits_array = ps_limits;
482 break;
483 }
484
485 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
486 {
487 if (shader_version <= limits_array[i].max_version)
488 {
489 shader->limits = &limits_array[i].limits;
490 break;
491 }
492 ++i;
493 }
494 if (!shader->limits)
495 {
496 FIXME("Unexpected shader version \"%u.%u\".\n",
497 shader->reg_maps.shader_version.major,
498 shader->reg_maps.shader_version.minor);
499 shader->limits = &limits_array[max(0, i - 1)].limits;
500 }
501 }
502
503 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
504 {
505 DWORD idx, shift;
506 idx = bit >> 5;
507 shift = bit & 0x1f;
508 bitmap[idx] |= (1 << shift);
509 }
510
511 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
512 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
513 {
514 switch (reg->type)
515 {
516 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
517 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
518 reg_maps->texcoord |= 1 << reg->idx[0].offset;
519 else
520 reg_maps->address |= 1 << reg->idx[0].offset;
521 break;
522
523 case WINED3DSPR_TEMP:
524 reg_maps->temporary |= 1 << reg->idx[0].offset;
525 break;
526
527 case WINED3DSPR_INPUT:
528 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
529 {
530 if (reg->idx[0].rel_addr)
531 {
532 /* If relative addressing is used, we must assume that all registers
533 * are used. Even if it is a construct like v3[aL], we can't assume
534 * that v0, v1 and v2 aren't read because aL can be negative */
535 unsigned int i;
536 for (i = 0; i < MAX_REG_INPUT; ++i)
537 {
538 shader->u.ps.input_reg_used[i] = TRUE;
539 }
540 }
541 else
542 {
543 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
544 }
545 }
546 else
547 reg_maps->input_registers |= 1 << reg->idx[0].offset;
548 break;
549
550 case WINED3DSPR_RASTOUT:
551 if (reg->idx[0].offset == 1)
552 reg_maps->fog = 1;
553 break;
554
555 case WINED3DSPR_MISCTYPE:
556 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
557 {
558 if (!reg->idx[0].offset)
559 reg_maps->vpos = 1;
560 else if (reg->idx[0].offset == 1)
561 reg_maps->usesfacing = 1;
562 }
563 break;
564
565 case WINED3DSPR_CONST:
566 if (reg->idx[0].rel_addr)
567 {
568 if (reg->idx[0].offset < reg_maps->min_rel_offset)
569 reg_maps->min_rel_offset = reg->idx[0].offset;
570 if (reg->idx[0].offset > reg_maps->max_rel_offset)
571 reg_maps->max_rel_offset = reg->idx[0].offset;
572 reg_maps->usesrelconstF = TRUE;
573 }
574 else
575 {
576 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
577 {
578 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
579 return FALSE;
580 }
581 else
582 {
583 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
584 }
585 }
586 break;
587
588 case WINED3DSPR_CONSTINT:
589 if (reg->idx[0].offset >= shader->limits->constant_int)
590 {
591 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
592 return FALSE;
593 }
594 else
595 {
596 reg_maps->integer_constants |= (1 << reg->idx[0].offset);
597 }
598 break;
599
600 case WINED3DSPR_CONSTBOOL:
601 if (reg->idx[0].offset >= shader->limits->constant_bool)
602 {
603 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
604 return FALSE;
605 }
606 else
607 {
608 reg_maps->boolean_constants |= (1 << reg->idx[0].offset);
609 }
610 break;
611
612 case WINED3DSPR_COLOROUT:
613 reg_maps->rt_mask |= (1 << reg->idx[0].offset);
614 break;
615
616 default:
617 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
618 reg->type, reg->idx[0].offset, reg->idx[1].offset);
619 break;
620 }
621 return TRUE;
622 }
623
624 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
625 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
626 {
627 struct wined3d_shader_sampler_map_entry *entries, *entry;
628 struct wined3d_shader_sampler_map *map;
629 unsigned int i;
630
631 map = &reg_maps->sampler_map;
632 entries = map->entries;
633 for (i = 0; i < map->count; ++i)
634 {
635 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
636 return;
637 }
638
639 if (!map->size)
640 {
641 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
642 {
643 ERR("Failed to allocate sampler map entries.\n");
644 return;
645 }
646 map->size = 4;
647 map->entries = entries;
648 }
649 else if (map->count == map->size)
650 {
651 size_t new_size = map->size * 2;
652
653 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
654 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
655 {
656 ERR("Failed to resize sampler map entries.\n");
657 return;
658 }
659 map->size = new_size;
660 map->entries = entries;
661 }
662
663 entry = &entries[map->count++];
664 entry->resource_idx = resource_idx;
665 entry->sampler_idx = sampler_idx;
666 entry->bind_idx = bind_idx;
667 }
668
669 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
670 {
671 switch (instr)
672 {
673 case WINED3DSIH_M4x4:
674 case WINED3DSIH_M3x4:
675 return param == 1 ? 3 : 0;
676
677 case WINED3DSIH_M4x3:
678 case WINED3DSIH_M3x3:
679 return param == 1 ? 2 : 0;
680
681 case WINED3DSIH_M3x2:
682 return param == 1 ? 1 : 0;
683
684 default:
685 return 0;
686 }
687 }
688
689 /* Note that this does not count the loop register as an address register. */
690 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
691 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
692 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
693 {
694 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
695 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
696 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
697 void *fe_data = shader->frontend_data;
698 struct wined3d_shader_version shader_version;
699 const DWORD *ptr = byte_code;
700 unsigned int i;
701
702 memset(reg_maps, 0, sizeof(*reg_maps));
703 memset(input_signature_elements, 0, sizeof(input_signature_elements));
704 memset(output_signature_elements, 0, sizeof(output_signature_elements));
705 reg_maps->min_rel_offset = ~0U;
706
707 fe->shader_read_header(fe_data, &ptr, &shader_version);
708 reg_maps->shader_version = shader_version;
709
710 shader_set_limits(shader);
711
712 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
713 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
714 if (!reg_maps->constf)
715 {
716 ERR("Failed to allocate constant map memory.\n");
717 return E_OUTOFMEMORY;
718 }
719
720 while (!fe->shader_is_end(fe_data, &ptr))
721 {
722 struct wined3d_shader_instruction ins;
723
724 /* Fetch opcode. */
725 fe->shader_read_instruction(fe_data, &ptr, &ins);
726
727 /* Unhandled opcode, and its parameters. */
728 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
729 {
730 TRACE("Skipping unrecognized instruction.\n");
731 continue;
732 }
733
734 /* Handle declarations. */
735 if (ins.handler_idx == WINED3DSIH_DCL)
736 {
737 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
738 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
739
740 switch (semantic->reg.reg.type)
741 {
742 /* Mark input registers used. */
743 case WINED3DSPR_INPUT:
744 if (reg_idx >= MAX_REG_INPUT)
745 {
746 ERR("Invalid input register index %u.\n", reg_idx);
747 break;
748 }
749 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
750 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
751 return WINED3DERR_INVALIDCALL;
752 reg_maps->input_registers |= 1 << reg_idx;
753 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
754 break;
755
756 /* Vertex shader: mark 3.0 output registers used, save token. */
757 case WINED3DSPR_OUTPUT:
758 if (reg_idx >= MAX_REG_OUTPUT)
759 {
760 ERR("Invalid output register index %u.\n", reg_idx);
761 break;
762 }
763 reg_maps->output_registers |= 1 << reg_idx;
764 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
765 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
766 reg_maps->fog = 1;
767 break;
768
769 case WINED3DSPR_SAMPLER:
770 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
771 case WINED3DSPR_RESOURCE:
772 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
773 {
774 ERR("Invalid resource index %u.\n", reg_idx);
775 break;
776 }
777 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
778 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
779 break;
780
781 default:
782 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
783 break;
784 }
785 }
786 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
787 {
788 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
789 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
790 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
791 else
792 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
793 }
794 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
795 {
796 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
797 shader->u.gs.input_type = ins.declaration.primitive_type;
798 else
799 FIXME("Invalid instruction %#x for shader type %#x.\n",
800 ins.handler_idx, shader_version.type);
801 }
802 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
803 {
804 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
805 shader->u.gs.output_type = ins.declaration.primitive_type;
806 else
807 FIXME("Invalid instruction %#x for shader type %#x.\n",
808 ins.handler_idx, shader_version.type);
809 }
810 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
811 {
812 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
813 shader->u.gs.vertices_out = ins.declaration.count;
814 else
815 FIXME("Invalid instruction %#x for shader type %#x.\n",
816 ins.handler_idx, shader_version.type);
817 }
818 else if (ins.handler_idx == WINED3DSIH_DEF)
819 {
820 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
821 float *value;
822 if (!lconst) return E_OUTOFMEMORY;
823
824 lconst->idx = ins.dst[0].reg.idx[0].offset;
825 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
826 value = (float *)lconst->value;
827
828 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
829 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
830 {
831 if (value[0] < -1.0f) value[0] = -1.0f;
832 else if (value[0] > 1.0f) value[0] = 1.0f;
833 if (value[1] < -1.0f) value[1] = -1.0f;
834 else if (value[1] > 1.0f) value[1] = 1.0f;
835 if (value[2] < -1.0f) value[2] = -1.0f;
836 else if (value[2] > 1.0f) value[2] = 1.0f;
837 if (value[3] < -1.0f) value[3] = -1.0f;
838 else if (value[3] > 1.0f) value[3] = 1.0f;
839 }
840
841 list_add_head(&shader->constantsF, &lconst->entry);
842
843 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
844 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
845 {
846 shader->lconst_inf_or_nan = TRUE;
847 }
848 }
849 else if (ins.handler_idx == WINED3DSIH_DEFI)
850 {
851 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
852 if (!lconst) return E_OUTOFMEMORY;
853
854 lconst->idx = ins.dst[0].reg.idx[0].offset;
855 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
856
857 list_add_head(&shader->constantsI, &lconst->entry);
858 reg_maps->local_int_consts |= (1 << lconst->idx);
859 }
860 else if (ins.handler_idx == WINED3DSIH_DEFB)
861 {
862 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
863 if (!lconst) return E_OUTOFMEMORY;
864
865 lconst->idx = ins.dst[0].reg.idx[0].offset;
866 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
867
868 list_add_head(&shader->constantsB, &lconst->entry);
869 reg_maps->local_bool_consts |= (1 << lconst->idx);
870 }
871 /* For subroutine prototypes. */
872 else if (ins.handler_idx == WINED3DSIH_LABEL)
873 {
874 reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
875 }
876 /* Set texture, address, temporary registers. */
877 else
878 {
879 BOOL color0_mov = FALSE;
880 unsigned int i;
881
882 /* This will loop over all the registers and try to
883 * make a bitmask of the ones we're interested in.
884 *
885 * Relative addressing tokens are ignored, but that's
886 * okay, since we'll catch any address registers when
887 * they are initialized (required by spec). */
888 for (i = 0; i < ins.dst_count; ++i)
889 {
890 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
891 shader_version.type, constf_size))
892 return WINED3DERR_INVALIDCALL;
893
894 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
895 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
896 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
897 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
898 {
899 UINT idx = ins.dst[i].reg.idx[0].offset;
900
901 switch (ins.dst[i].reg.type)
902 {
903 case WINED3DSPR_RASTOUT:
904 switch (idx)
905 {
906 case 0: /* oPos */
907 reg_maps->output_registers |= 1 << 10;
908 shader_signature_from_usage(&output_signature_elements[10],
909 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
910 break;
911
912 case 1: /* oFog */
913 reg_maps->output_registers |= 1 << 11;
914 shader_signature_from_usage(&output_signature_elements[11],
915 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
916 break;
917
918 case 2: /* oPts */
919 reg_maps->output_registers |= 1 << 11;
920 shader_signature_from_usage(&output_signature_elements[11],
921 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
922 break;
923 }
924 break;
925
926 case WINED3DSPR_ATTROUT:
927 if (idx < 2)
928 {
929 idx += 8;
930 if (reg_maps->output_registers & (1 << idx))
931 {
932 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
933 }
934 else
935 {
936 reg_maps->output_registers |= 1 << idx;
937 shader_signature_from_usage(&output_signature_elements[idx],
938 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
939 }
940 }
941 break;
942
943 case WINED3DSPR_TEXCRDOUT:
944
945 reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
946 if (reg_maps->output_registers & (1 << idx))
947 {
948 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
949 }
950 else
951 {
952 reg_maps->output_registers |= 1 << idx;
953 shader_signature_from_usage(&output_signature_elements[idx],
954 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
955 }
956 break;
957
958 default:
959 break;
960 }
961 }
962
963 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
964 {
965 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
966 {
967 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
968 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
969 * the mov and perform the sRGB write correction from the source register.
970 *
971 * However, if the mov is only partial, we can't do this, and if the write
972 * comes from an instruction other than MOV it is hard to do as well. If
973 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
974 shader->u.ps.color0_mov = FALSE;
975 if (ins.handler_idx == WINED3DSIH_MOV
976 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
977 {
978 /* Used later when the source register is read. */
979 color0_mov = TRUE;
980 }
981 }
982 /* Also drop the MOV marker if the source register is overwritten prior to the shader
983 * end
984 */
985 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
986 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
987 {
988 shader->u.ps.color0_mov = FALSE;
989 }
990 }
991
992 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
993 if (shader_version.major == 1
994 && (ins.handler_idx == WINED3DSIH_TEX
995 || ins.handler_idx == WINED3DSIH_TEXBEM
996 || ins.handler_idx == WINED3DSIH_TEXBEML
997 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
998 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
999 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1000 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1001 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1002 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1003 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1004 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1005 {
1006 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1007
1008 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1009 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1010 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1011 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1012
1013 /* texbem is only valid with < 1.4 pixel shaders */
1014 if (ins.handler_idx == WINED3DSIH_TEXBEM
1015 || ins.handler_idx == WINED3DSIH_TEXBEML)
1016 {
1017 reg_maps->bumpmat |= 1 << reg_idx;
1018 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1019 {
1020 reg_maps->luminanceparams |= 1 << reg_idx;
1021 }
1022 }
1023 }
1024 else if (ins.handler_idx == WINED3DSIH_BEM)
1025 {
1026 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
1027 }
1028 }
1029
1030 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1031 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1032 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1033 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1034 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1035 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1036 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1037 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1038 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1039 else if (ins.handler_idx == WINED3DSIH_LOOP
1040 || ins.handler_idx == WINED3DSIH_REP)
1041 {
1042 ++cur_loop_depth;
1043 if (cur_loop_depth > max_loop_depth)
1044 max_loop_depth = cur_loop_depth;
1045 }
1046 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1047 || ins.handler_idx == WINED3DSIH_ENDREP)
1048 {
1049 --cur_loop_depth;
1050 }
1051 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1052 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1053 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1054 {
1055 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1056 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1057 }
1058
1059 if (ins.predicate)
1060 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1061 shader_version.type, constf_size))
1062 return WINED3DERR_INVALIDCALL;
1063
1064 for (i = 0; i < ins.src_count; ++i)
1065 {
1066 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1067 struct wined3d_shader_register reg = ins.src[i].reg;
1068
1069 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1070 shader_version.type, constf_size))
1071 return WINED3DERR_INVALIDCALL;
1072 while (count)
1073 {
1074 ++reg.idx[0].offset;
1075 if (!shader_record_register_usage(shader, reg_maps, &reg,
1076 shader_version.type, constf_size))
1077 return WINED3DERR_INVALIDCALL;
1078 --count;
1079 }
1080
1081 if (color0_mov)
1082 {
1083 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1084 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1085 {
1086 shader->u.ps.color0_mov = TRUE;
1087 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1088 }
1089 }
1090 }
1091 }
1092 }
1093 reg_maps->loop_depth = max_loop_depth;
1094
1095 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1096 * R0 is written to the render target. */
1097 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1098 reg_maps->rt_mask |= (1 << 0);
1099
1100 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1101
1102 if (input_signature->elements)
1103 {
1104 for (i = 0; i < input_signature->element_count; ++i)
1105 {
1106 reg_maps->input_registers |= 1 << input_signature->elements[i].register_idx;
1107 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1108 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1109 reg_maps->vpos = 1;
1110 }
1111 }
1112 else if (!input_signature->elements && reg_maps->input_registers)
1113 {
1114 unsigned int count = count_bits(reg_maps->input_registers);
1115 struct wined3d_shader_signature_element *e;
1116 unsigned int i;
1117
1118 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1119 return E_OUTOFMEMORY;
1120 input_signature->element_count = count;
1121
1122 e = input_signature->elements;
1123 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1124 {
1125 if (!(reg_maps->input_registers & (1 << i)))
1126 continue;
1127 input_signature_elements[i].register_idx = i;
1128 *e++ = input_signature_elements[i];
1129 }
1130 }
1131
1132 if (output_signature->elements)
1133 {
1134 for (i = 0; i < output_signature->element_count; ++i)
1135 {
1136 reg_maps->output_registers |= 1 << output_signature->elements[i].register_idx;
1137 }
1138 }
1139 else if (reg_maps->output_registers)
1140 {
1141 unsigned int count = count_bits(reg_maps->output_registers);
1142 struct wined3d_shader_signature_element *e;
1143
1144 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1145 return E_OUTOFMEMORY;
1146 output_signature->element_count = count;
1147
1148 e = output_signature->elements;
1149 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1150 {
1151 if (!(reg_maps->output_registers & (1 << i)))
1152 continue;
1153 *e++ = output_signature_elements[i];
1154 }
1155 }
1156
1157 return WINED3D_OK;
1158 }
1159
1160 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1161 {
1162 DWORD map = 1 << max;
1163 map |= map - 1;
1164 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1165
1166 return wined3d_log2i(map);
1167 }
1168
1169 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
1170 const struct wined3d_shader_version *shader_version)
1171 {
1172 TRACE("dcl");
1173
1174 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1175 {
1176 switch (semantic->resource_type)
1177 {
1178 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1179 TRACE("_2d");
1180 break;
1181
1182 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1183 TRACE("_volume");
1184 break;
1185
1186 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1187 TRACE("_cube");
1188 break;
1189
1190 default:
1191 TRACE("_unknown_ttype(0x%08x)", semantic->resource_type);
1192 break;
1193 }
1194 }
1195 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1196 {
1197 TRACE("_resource_");
1198 switch (semantic->resource_type)
1199 {
1200 case WINED3D_SHADER_RESOURCE_BUFFER:
1201 TRACE("buffer");
1202 break;
1203
1204 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1205 TRACE("texture1d");
1206 break;
1207
1208 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1209 TRACE("texture2d");
1210 break;
1211
1212 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1213 TRACE("texture2dms");
1214 break;
1215
1216 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1217 TRACE("texture3d");
1218 break;
1219
1220 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1221 TRACE("texturecube");
1222 break;
1223
1224 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1225 TRACE("texture1darray");
1226 break;
1227
1228 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1229 TRACE("texture2darray");
1230 break;
1231
1232 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1233 TRACE("texture2dmsarray");
1234 break;
1235
1236 default:
1237 TRACE("unknown");
1238 break;
1239 }
1240 switch (semantic->resource_data_type)
1241 {
1242 case WINED3D_DATA_FLOAT:
1243 TRACE(" (float)");
1244 break;
1245
1246 case WINED3D_DATA_INT:
1247 TRACE(" (int)");
1248 break;
1249
1250 case WINED3D_DATA_UINT:
1251 TRACE(" (uint)");
1252 break;
1253
1254 case WINED3D_DATA_UNORM:
1255 TRACE(" (unorm)");
1256 break;
1257
1258 case WINED3D_DATA_SNORM:
1259 TRACE(" (snorm)");
1260 break;
1261
1262 default:
1263 TRACE(" (unknown)");
1264 break;
1265 }
1266 }
1267 else
1268 {
1269 /* Pixel shaders 3.0 don't have usage semantics. */
1270 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
1271 else TRACE("_");
1272
1273 switch (semantic->usage)
1274 {
1275 case WINED3D_DECL_USAGE_POSITION:
1276 TRACE("position%u", semantic->usage_idx);
1277 break;
1278
1279 case WINED3D_DECL_USAGE_BLEND_INDICES:
1280 TRACE("blend");
1281 break;
1282
1283 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1284 TRACE("weight");
1285 break;
1286
1287 case WINED3D_DECL_USAGE_NORMAL:
1288 TRACE("normal%u", semantic->usage_idx);
1289 break;
1290
1291 case WINED3D_DECL_USAGE_PSIZE:
1292 TRACE("psize");
1293 break;
1294
1295 case WINED3D_DECL_USAGE_COLOR:
1296 if (!semantic->usage_idx) TRACE("color");
1297 else TRACE("specular%u", (semantic->usage_idx - 1));
1298 break;
1299
1300 case WINED3D_DECL_USAGE_TEXCOORD:
1301 TRACE("texture%u", semantic->usage_idx);
1302 break;
1303
1304 case WINED3D_DECL_USAGE_TANGENT:
1305 TRACE("tangent");
1306 break;
1307
1308 case WINED3D_DECL_USAGE_BINORMAL:
1309 TRACE("binormal");
1310 break;
1311
1312 case WINED3D_DECL_USAGE_TESS_FACTOR:
1313 TRACE("tessfactor");
1314 break;
1315
1316 case WINED3D_DECL_USAGE_POSITIONT:
1317 TRACE("positionT%u", semantic->usage_idx);
1318 break;
1319
1320 case WINED3D_DECL_USAGE_FOG:
1321 TRACE("fog");
1322 break;
1323
1324 case WINED3D_DECL_USAGE_DEPTH:
1325 TRACE("depth");
1326 break;
1327
1328 case WINED3D_DECL_USAGE_SAMPLE:
1329 TRACE("sample");
1330 break;
1331
1332 default:
1333 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1334 }
1335 }
1336 }
1337
1338 static void shader_dump_register(const struct wined3d_shader_register *reg,
1339 const struct wined3d_shader_version *shader_version)
1340 {
1341 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1342 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1343 UINT offset = reg->idx[0].offset;
1344
1345 switch (reg->type)
1346 {
1347 case WINED3DSPR_TEMP:
1348 TRACE("r");
1349 break;
1350
1351 case WINED3DSPR_INPUT:
1352 TRACE("v");
1353 break;
1354
1355 case WINED3DSPR_CONST:
1356 case WINED3DSPR_CONST2:
1357 case WINED3DSPR_CONST3:
1358 case WINED3DSPR_CONST4:
1359 TRACE("c");
1360 offset = shader_get_float_offset(reg->type, offset);
1361 break;
1362
1363 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1364 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1365 break;
1366
1367 case WINED3DSPR_RASTOUT:
1368 TRACE("%s", rastout_reg_names[offset]);
1369 break;
1370
1371 case WINED3DSPR_COLOROUT:
1372 TRACE("oC");
1373 break;
1374
1375 case WINED3DSPR_DEPTHOUT:
1376 TRACE("oDepth");
1377 break;
1378
1379 case WINED3DSPR_ATTROUT:
1380 TRACE("oD");
1381 break;
1382
1383 case WINED3DSPR_TEXCRDOUT:
1384 /* Vertex shaders >= 3.0 use general purpose output registers
1385 * (WINED3DSPR_OUTPUT), which can include an address token. */
1386 if (shader_version->major >= 3) TRACE("o");
1387 else TRACE("oT");
1388 break;
1389
1390 case WINED3DSPR_CONSTINT:
1391 TRACE("i");
1392 break;
1393
1394 case WINED3DSPR_CONSTBOOL:
1395 TRACE("b");
1396 break;
1397
1398 case WINED3DSPR_LABEL:
1399 TRACE("l");
1400 break;
1401
1402 case WINED3DSPR_LOOP:
1403 TRACE("aL");
1404 break;
1405
1406 case WINED3DSPR_SAMPLER:
1407 TRACE("s");
1408 break;
1409
1410 case WINED3DSPR_MISCTYPE:
1411 if (offset > 1)
1412 FIXME("Unhandled misctype register %u.\n", offset);
1413 else
1414 TRACE("%s", misctype_reg_names[offset]);
1415 break;
1416
1417 case WINED3DSPR_PREDICATE:
1418 TRACE("p");
1419 break;
1420
1421 case WINED3DSPR_IMMCONST:
1422 TRACE("l");
1423 break;
1424
1425 case WINED3DSPR_CONSTBUFFER:
1426 TRACE("cb");
1427 break;
1428
1429 case WINED3DSPR_PRIMID:
1430 TRACE("primID");
1431 break;
1432
1433 case WINED3DSPR_NULL:
1434 TRACE("null");
1435 break;
1436
1437 case WINED3DSPR_RESOURCE:
1438 TRACE("t");
1439 break;
1440
1441 default:
1442 TRACE("unhandled_rtype(%#x)", reg->type);
1443 break;
1444 }
1445
1446 if (reg->type == WINED3DSPR_IMMCONST)
1447 {
1448 TRACE("(");
1449 switch (reg->immconst_type)
1450 {
1451 case WINED3D_IMMCONST_SCALAR:
1452 switch (reg->data_type)
1453 {
1454 case WINED3D_DATA_FLOAT:
1455 TRACE("%.8e", *(const float *)reg->immconst_data);
1456 break;
1457 case WINED3D_DATA_INT:
1458 TRACE("%d", reg->immconst_data[0]);
1459 break;
1460 case WINED3D_DATA_RESOURCE:
1461 case WINED3D_DATA_SAMPLER:
1462 case WINED3D_DATA_UINT:
1463 TRACE("%u", reg->immconst_data[0]);
1464 break;
1465 default:
1466 TRACE("<unhandled data type %#x>", reg->data_type);
1467 break;
1468 }
1469 break;
1470
1471 case WINED3D_IMMCONST_VEC4:
1472 switch (reg->data_type)
1473 {
1474 case WINED3D_DATA_FLOAT:
1475 TRACE("%.8e, %.8e, %.8e, %.8e",
1476 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1477 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1478 break;
1479 case WINED3D_DATA_INT:
1480 TRACE("%d, %d, %d, %d",
1481 reg->immconst_data[0], reg->immconst_data[1],
1482 reg->immconst_data[2], reg->immconst_data[3]);
1483 break;
1484 case WINED3D_DATA_RESOURCE:
1485 case WINED3D_DATA_SAMPLER:
1486 case WINED3D_DATA_UINT:
1487 TRACE("%u, %u, %u, %u",
1488 reg->immconst_data[0], reg->immconst_data[1],
1489 reg->immconst_data[2], reg->immconst_data[3]);
1490 break;
1491 default:
1492 TRACE("<unhandled data type %#x>", reg->data_type);
1493 break;
1494 }
1495 break;
1496
1497 default:
1498 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1499 break;
1500 }
1501 TRACE(")");
1502 }
1503 else if (reg->type != WINED3DSPR_RASTOUT
1504 && reg->type != WINED3DSPR_MISCTYPE
1505 && reg->type != WINED3DSPR_NULL)
1506 {
1507 if (offset != ~0U)
1508 {
1509 TRACE("[");
1510 if (reg->idx[0].rel_addr)
1511 {
1512 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1513 TRACE(" + ");
1514 }
1515 TRACE("%u]", offset);
1516
1517 if (reg->idx[1].offset != ~0U)
1518 {
1519 TRACE("[");
1520 if (reg->idx[1].rel_addr)
1521 {
1522 shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1523 TRACE(" + ");
1524 }
1525 TRACE("%u]", reg->idx[1].offset);
1526 }
1527 }
1528 }
1529 }
1530
1531 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1532 const struct wined3d_shader_version *shader_version)
1533 {
1534 DWORD write_mask = param->write_mask;
1535
1536 shader_dump_register(&param->reg, shader_version);
1537
1538 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1539 {
1540 static const char write_mask_chars[] = "xyzw";
1541
1542 TRACE(".");
1543 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1544 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1545 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1546 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1547 }
1548 }
1549
1550 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1551 const struct wined3d_shader_version *shader_version)
1552 {
1553 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1554 DWORD swizzle = param->swizzle;
1555
1556 if (src_modifier == WINED3DSPSM_NEG
1557 || src_modifier == WINED3DSPSM_BIASNEG
1558 || src_modifier == WINED3DSPSM_SIGNNEG
1559 || src_modifier == WINED3DSPSM_X2NEG
1560 || src_modifier == WINED3DSPSM_ABSNEG)
1561 TRACE("-");
1562 else if (src_modifier == WINED3DSPSM_COMP)
1563 TRACE("1-");
1564 else if (src_modifier == WINED3DSPSM_NOT)
1565 TRACE("!");
1566
1567 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1568 TRACE("abs(");
1569
1570 shader_dump_register(&param->reg, shader_version);
1571
1572 if (src_modifier)
1573 {
1574 switch (src_modifier)
1575 {
1576 case WINED3DSPSM_NONE: break;
1577 case WINED3DSPSM_NEG: break;
1578 case WINED3DSPSM_NOT: break;
1579 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1580 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1581 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1582 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1583 case WINED3DSPSM_COMP: break;
1584 case WINED3DSPSM_X2: TRACE("_x2"); break;
1585 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1586 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1587 case WINED3DSPSM_DW: TRACE("_dw"); break;
1588 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1589 case WINED3DSPSM_ABS: TRACE(")"); break;
1590 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1591 }
1592 }
1593
1594 if (swizzle != WINED3DSP_NOSWIZZLE)
1595 {
1596 static const char swizzle_chars[] = "xyzw";
1597 DWORD swizzle_x = swizzle & 0x03;
1598 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1599 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1600 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1601
1602 if (swizzle_x == swizzle_y
1603 && swizzle_x == swizzle_z
1604 && swizzle_x == swizzle_w)
1605 {
1606 TRACE(".%c", swizzle_chars[swizzle_x]);
1607 }
1608 else
1609 {
1610 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1611 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1612 }
1613 }
1614 }
1615
1616 /* Shared code in order to generate the bulk of the shader string.
1617 * NOTE: A description of how to parse tokens can be found on MSDN. */
1618 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1619 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1620 {
1621 struct wined3d_device *device = shader->device;
1622 const struct wined3d_shader_frontend *fe = shader->frontend;
1623 void *fe_data = shader->frontend_data;
1624 struct wined3d_shader_version shader_version;
1625 struct wined3d_shader_loop_state loop_state;
1626 struct wined3d_shader_instruction ins;
1627 struct wined3d_shader_tex_mx tex_mx;
1628 struct wined3d_shader_context ctx;
1629 const DWORD *ptr = byte_code;
1630
1631 /* Initialize current parsing state. */
1632 tex_mx.current_row = 0;
1633 loop_state.current_depth = 0;
1634 loop_state.current_reg = 0;
1635
1636 ctx.shader = shader;
1637 ctx.gl_info = &device->adapter->gl_info;
1638 ctx.reg_maps = reg_maps;
1639 ctx.buffer = buffer;
1640 ctx.tex_mx = &tex_mx;
1641 ctx.loop_state = &loop_state;
1642 ctx.backend_data = backend_ctx;
1643 ins.ctx = &ctx;
1644
1645 fe->shader_read_header(fe_data, &ptr, &shader_version);
1646
1647 while (!fe->shader_is_end(fe_data, &ptr))
1648 {
1649 /* Read opcode. */
1650 fe->shader_read_instruction(fe_data, &ptr, &ins);
1651
1652 /* Unknown opcode and its parameters. */
1653 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1654 {
1655 TRACE("Skipping unrecognized instruction.\n");
1656 continue;
1657 }
1658
1659 if (ins.predicate)
1660 FIXME("Predicates not implemented.\n");
1661
1662 /* Call appropriate function for output target */
1663 device->shader_backend->shader_handle_instruction(&ins);
1664 }
1665 }
1666
1667 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1668 {
1669 DWORD mmask = dst->modifiers;
1670
1671 switch (dst->shift)
1672 {
1673 case 0: break;
1674 case 13: TRACE("_d8"); break;
1675 case 14: TRACE("_d4"); break;
1676 case 15: TRACE("_d2"); break;
1677 case 1: TRACE("_x2"); break;
1678 case 2: TRACE("_x4"); break;
1679 case 3: TRACE("_x8"); break;
1680 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1681 }
1682
1683 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1684 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1685 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1686
1687 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1688 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1689 }
1690
1691 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1692 {
1693 switch (primitive_type)
1694 {
1695 case WINED3D_PT_UNDEFINED:
1696 TRACE("undefined");
1697 break;
1698 case WINED3D_PT_POINTLIST:
1699 TRACE("pointlist");
1700 break;
1701 case WINED3D_PT_LINELIST:
1702 TRACE("linelist");
1703 break;
1704 case WINED3D_PT_LINESTRIP:
1705 TRACE("linestrip");
1706 break;
1707 case WINED3D_PT_TRIANGLELIST:
1708 TRACE("trianglelist");
1709 break;
1710 case WINED3D_PT_TRIANGLESTRIP:
1711 TRACE("trianglestrip");
1712 break;
1713 case WINED3D_PT_TRIANGLEFAN:
1714 TRACE("trianglefan");
1715 break;
1716 case WINED3D_PT_LINELIST_ADJ:
1717 TRACE("linelist_adj");
1718 break;
1719 case WINED3D_PT_LINESTRIP_ADJ:
1720 TRACE("linestrip_adj");
1721 break;
1722 case WINED3D_PT_TRIANGLELIST_ADJ:
1723 TRACE("trianglelist_adj");
1724 break;
1725 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1726 TRACE("trianglestrip_adj");
1727 break;
1728 default:
1729 TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1730 break;
1731 }
1732 }
1733
1734 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1735 {
1736 struct wined3d_shader_version shader_version;
1737 const DWORD *ptr = byte_code;
1738 const char *type_prefix;
1739 DWORD i;
1740
1741 TRACE("Parsing %p.\n", byte_code);
1742
1743 fe->shader_read_header(fe_data, &ptr, &shader_version);
1744
1745 switch (shader_version.type)
1746 {
1747 case WINED3D_SHADER_TYPE_VERTEX:
1748 type_prefix = "vs";
1749 break;
1750
1751 case WINED3D_SHADER_TYPE_GEOMETRY:
1752 type_prefix = "gs";
1753 break;
1754
1755 case WINED3D_SHADER_TYPE_PIXEL:
1756 type_prefix = "ps";
1757 break;
1758
1759 default:
1760 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1761 type_prefix = "unknown";
1762 break;
1763 }
1764
1765 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1766
1767 while (!fe->shader_is_end(fe_data, &ptr))
1768 {
1769 struct wined3d_shader_instruction ins;
1770
1771 fe->shader_read_instruction(fe_data, &ptr, &ins);
1772 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1773 {
1774 TRACE("Skipping unrecognized instruction.\n");
1775 continue;
1776 }
1777
1778 if (ins.handler_idx == WINED3DSIH_DCL)
1779 {
1780 shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1781 shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1782 TRACE(" ");
1783 shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1784 }
1785 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1786 {
1787 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1788 shader_dump_src_param(&ins.declaration.src, &shader_version);
1789 TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1790 }
1791 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1792 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1793 {
1794 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1795 shader_dump_primitive_type(ins.declaration.primitive_type);
1796 }
1797 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1798 {
1799 TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1800 }
1801 else if (ins.handler_idx == WINED3DSIH_DEF)
1802 {
1803 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1804 ins.dst[0].reg.idx[0].offset),
1805 *(const float *)&ins.src[0].reg.immconst_data[0],
1806 *(const float *)&ins.src[0].reg.immconst_data[1],
1807 *(const float *)&ins.src[0].reg.immconst_data[2],
1808 *(const float *)&ins.src[0].reg.immconst_data[3]);
1809 }
1810 else if (ins.handler_idx == WINED3DSIH_DEFI)
1811 {
1812 TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1813 ins.src[0].reg.immconst_data[0],
1814 ins.src[0].reg.immconst_data[1],
1815 ins.src[0].reg.immconst_data[2],
1816 ins.src[0].reg.immconst_data[3]);
1817 }
1818 else if (ins.handler_idx == WINED3DSIH_DEFB)
1819 {
1820 TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1821 }
1822 else
1823 {
1824 if (ins.predicate)
1825 {
1826 TRACE("(");
1827 shader_dump_src_param(ins.predicate, &shader_version);
1828 TRACE(") ");
1829 }
1830
1831 /* PixWin marks instructions with the coissue flag with a '+' */
1832 if (ins.coissue) TRACE("+");
1833
1834 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1835
1836 if (ins.handler_idx == WINED3DSIH_IFC
1837 || ins.handler_idx == WINED3DSIH_BREAKC)
1838 {
1839 switch (ins.flags)
1840 {
1841 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1842 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1843 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1844 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1845 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1846 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1847 default: TRACE("_(%u)", ins.flags);
1848 }
1849 }
1850 else if (ins.handler_idx == WINED3DSIH_TEX
1851 && shader_version.major >= 2
1852 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1853 {
1854 TRACE("p");
1855 }
1856
1857 for (i = 0; i < ins.dst_count; ++i)
1858 {
1859 shader_dump_ins_modifiers(&ins.dst[i]);
1860 TRACE(!i ? " " : ", ");
1861 shader_dump_dst_param(&ins.dst[i], &shader_version);
1862 }
1863
1864 /* Other source tokens */
1865 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1866 {
1867 TRACE(!i ? " " : ", ");
1868 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1869 }
1870 }
1871 TRACE("\n");
1872 }
1873 }
1874
1875 static void shader_cleanup(struct wined3d_shader *shader)
1876 {
1877 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
1878 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
1879 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
1880 shader->device->shader_backend->shader_destroy(shader);
1881 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1882 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
1883 HeapFree(GetProcessHeap(), 0, shader->function);
1884 shader_delete_constant_list(&shader->constantsF);
1885 shader_delete_constant_list(&shader->constantsB);
1886 shader_delete_constant_list(&shader->constantsI);
1887 list_remove(&shader->shader_list_entry);
1888
1889 if (shader->frontend && shader->frontend_data)
1890 shader->frontend->shader_free(shader->frontend_data);
1891 }
1892
1893 struct shader_none_priv
1894 {
1895 const struct wined3d_vertex_pipe_ops *vertex_pipe;
1896 const struct fragment_pipeline *fragment_pipe;
1897 BOOL ffp_proj_control;
1898 };
1899
1900 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1901 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1902 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
1903 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1904 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1905 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1906 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
1907 const struct wined3d_state *state) {}
1908 static void shader_none_destroy(struct wined3d_shader *shader) {}
1909 static void shader_none_free_context_data(struct wined3d_context *context) {}
1910
1911 /* Context activation is done by the caller. */
1912 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
1913 const struct wined3d_state *state)
1914 {
1915 const struct wined3d_gl_info *gl_info = context->gl_info;
1916 struct shader_none_priv *priv = shader_priv;
1917
1918 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
1919 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
1920 }
1921
1922 /* Context activation is done by the caller. */
1923 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
1924 {
1925 struct shader_none_priv *priv = shader_priv;
1926 const struct wined3d_gl_info *gl_info = context->gl_info;
1927
1928 priv->vertex_pipe->vp_enable(gl_info, FALSE);
1929 priv->fragment_pipe->enable_extension(gl_info, FALSE);
1930
1931 context->shader_update_mask = (1 << WINED3D_SHADER_TYPE_PIXEL)
1932 | (1 << WINED3D_SHADER_TYPE_VERTEX)
1933 | (1 << WINED3D_SHADER_TYPE_GEOMETRY);
1934 }
1935
1936 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
1937 const struct fragment_pipeline *fragment_pipe)
1938 {
1939 struct fragment_caps fragment_caps;
1940 void *vertex_priv, *fragment_priv;
1941 struct shader_none_priv *priv;
1942
1943 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1944 return E_OUTOFMEMORY;
1945
1946 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
1947 {
1948 ERR("Failed to initialize vertex pipe.\n");
1949 HeapFree(GetProcessHeap(), 0, priv);
1950 return E_FAIL;
1951 }
1952
1953 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1954 {
1955 ERR("Failed to initialize fragment pipe.\n");
1956 vertex_pipe->vp_free(device);
1957 HeapFree(GetProcessHeap(), 0, priv);
1958 return E_FAIL;
1959 }
1960
1961 priv->vertex_pipe = vertex_pipe;
1962 priv->fragment_pipe = fragment_pipe;
1963 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
1964 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
1965
1966 device->vertex_priv = vertex_priv;
1967 device->fragment_priv = fragment_priv;
1968 device->shader_priv = priv;
1969
1970 return WINED3D_OK;
1971 }
1972
1973 static void shader_none_free(struct wined3d_device *device)
1974 {
1975 struct shader_none_priv *priv = device->shader_priv;
1976
1977 priv->fragment_pipe->free_private(device);
1978 priv->vertex_pipe->vp_free(device);
1979 HeapFree(GetProcessHeap(), 0, priv);
1980 }
1981
1982 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
1983 {
1984 return TRUE;
1985 }
1986
1987 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1988 {
1989 /* Set the shader caps to 0 for the none shader backend */
1990 caps->vs_version = 0;
1991 caps->gs_version = 0;
1992 caps->ps_version = 0;
1993 caps->vs_uniform_count = 0;
1994 caps->ps_uniform_count = 0;
1995 caps->ps_1x_max_value = 0.0f;
1996 caps->wined3d_caps = 0;
1997 }
1998
1999 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2000 {
2001 /* We "support" every possible fixup, since we don't support any shader
2002 * model, and will never have to actually sample a texture. */
2003 return TRUE;
2004 }
2005
2006 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2007 {
2008 struct shader_none_priv *priv = shader_priv;
2009
2010 return priv->ffp_proj_control;
2011 }
2012
2013 const struct wined3d_shader_backend_ops none_shader_backend =
2014 {
2015 shader_none_handle_instruction,
2016 shader_none_select,
2017 shader_none_disable,
2018 shader_none_select_depth_blt,
2019 shader_none_deselect_depth_blt,
2020 shader_none_update_float_vertex_constants,
2021 shader_none_update_float_pixel_constants,
2022 shader_none_load_constants,
2023 shader_none_destroy,
2024 shader_none_alloc,
2025 shader_none_free,
2026 shader_none_allocate_context_data,
2027 shader_none_free_context_data,
2028 shader_none_get_caps,
2029 shader_none_color_fixup_supported,
2030 shader_none_has_ffp_proj_control,
2031 };
2032
2033 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2034 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2035 enum wined3d_shader_type type, unsigned int max_version)
2036 {
2037 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2038 const struct wined3d_shader_frontend *fe;
2039 HRESULT hr;
2040 unsigned int backend_version;
2041 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2042
2043 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
2044 shader, byte_code, output_signature, float_const_count);
2045
2046 fe = shader_select_frontend(*byte_code);
2047 if (!fe)
2048 {
2049 FIXME("Unable to find frontend for shader.\n");
2050 return WINED3DERR_INVALIDCALL;
2051 }
2052 shader->frontend = fe;
2053 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2054 if (!shader->frontend_data)
2055 {
2056 FIXME("Failed to initialize frontend.\n");
2057 return WINED3DERR_INVALIDCALL;
2058 }
2059
2060 /* First pass: trace shader. */
2061 if (TRACE_ON(d3d_shader))
2062 shader_trace_init(fe, shader->frontend_data, byte_code);
2063
2064 /* Initialize immediate constant lists. */
2065 list_init(&shader->constantsF);
2066 list_init(&shader->constantsB);
2067 list_init(&shader->constantsI);
2068 shader->lconst_inf_or_nan = FALSE;
2069
2070 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2071 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2072 &shader->output_signature, byte_code, float_const_count)))
2073 return hr;
2074
2075 if (reg_maps->shader_version.type != type)
2076 {
2077 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2078 return WINED3DERR_INVALIDCALL;
2079 }
2080 if (reg_maps->shader_version.major > max_version)
2081 {
2082 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2083 return WINED3DERR_INVALIDCALL;
2084 }
2085 switch (type)
2086 {
2087 case WINED3D_SHADER_TYPE_VERTEX:
2088 backend_version = d3d_info->limits.vs_version;
2089 break;
2090 case WINED3D_SHADER_TYPE_GEOMETRY:
2091 backend_version = d3d_info->limits.gs_version;
2092 break;
2093 case WINED3D_SHADER_TYPE_PIXEL:
2094 backend_version = d3d_info->limits.ps_version;
2095 break;
2096 default:
2097 FIXME("No backend version-checking for this shader type\n");
2098 backend_version = 0;
2099 }
2100 if (reg_maps->shader_version.major > backend_version)
2101 {
2102 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2103 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2104 return WINED3DERR_INVALIDCALL;
2105 }
2106
2107 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2108 if (!shader->function)
2109 return E_OUTOFMEMORY;
2110 memcpy(shader->function, byte_code, shader->functionLength);
2111
2112 return WINED3D_OK;
2113 }
2114
2115 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2116 {
2117 ULONG refcount = InterlockedIncrement(&shader->ref);
2118
2119 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2120
2121 return refcount;
2122 }
2123
2124 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2125 {
2126 ULONG refcount = InterlockedDecrement(&shader->ref);
2127
2128 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2129
2130 if (!refcount)
2131 {
2132 shader_cleanup(shader);
2133 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2134 HeapFree(GetProcessHeap(), 0, shader);
2135 }
2136
2137 return refcount;
2138 }
2139
2140 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2141 {
2142 TRACE("shader %p.\n", shader);
2143
2144 return shader->parent;
2145 }
2146
2147 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2148 void *byte_code, UINT *byte_code_size)
2149 {
2150 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2151
2152 if (!byte_code)
2153 {
2154 *byte_code_size = shader->functionLength;
2155 return WINED3D_OK;
2156 }
2157
2158 if (*byte_code_size < shader->functionLength)
2159 {
2160 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2161 * than the required size we should write the required size and
2162 * return D3DERR_MOREDATA. That's not actually true. */
2163 return WINED3DERR_INVALIDCALL;
2164 }
2165
2166 memcpy(byte_code, shader->function, shader->functionLength);
2167
2168 return WINED3D_OK;
2169 }
2170
2171 /* Set local constants for d3d8 shaders. */
2172 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2173 UINT start_idx, const float *src_data, UINT count)
2174 {
2175 UINT end_idx = start_idx + count;
2176 UINT i;
2177
2178 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2179
2180 if (end_idx > shader->limits->constant_float)
2181 {
2182 WARN("end_idx %u > float constants limit %u.\n",
2183 end_idx, shader->limits->constant_float);
2184 end_idx = shader->limits->constant_float;
2185 }
2186
2187 for (i = start_idx; i < end_idx; ++i)
2188 {
2189 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2190 float *value;
2191 if (!lconst)
2192 return E_OUTOFMEMORY;
2193
2194 lconst->idx = i;
2195 value = (float *)lconst->value;
2196 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2197 list_add_head(&shader->constantsF, &lconst->entry);
2198
2199 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2200 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2201 {
2202 shader->lconst_inf_or_nan = TRUE;
2203 }
2204 }
2205
2206 return WINED3D_OK;
2207 }
2208
2209 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2210 WORD swizzle_map, struct vs_compile_args *args)
2211 {
2212 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2213 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2214 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2215 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2216 args->swizzle_map = swizzle_map;
2217 }
2218
2219 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2220 {
2221 if (usage_idx1 != usage_idx2)
2222 return FALSE;
2223 if (usage1 == usage2)
2224 return TRUE;
2225 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2226 return TRUE;
2227 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2228 return TRUE;
2229
2230 return FALSE;
2231 }
2232
2233 BOOL vshader_get_input(const struct wined3d_shader *shader,
2234 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2235 {
2236 WORD map = shader->reg_maps.input_registers;
2237 unsigned int i;
2238
2239 for (i = 0; map; map >>= 1, ++i)
2240 {
2241 if (!(map & 1)) continue;
2242
2243 if (match_usage(shader->u.vs.attributes[i].usage,
2244 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2245 {
2246 *regnum = i;
2247 return TRUE;
2248 }
2249 }
2250 return FALSE;
2251 }
2252
2253 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2254 const struct wined3d_shader_signature *src, char **signature_strings)
2255 {
2256 struct wined3d_shader_signature_element *e;
2257 unsigned int i;
2258 SIZE_T len;
2259 char *ptr;
2260
2261 ptr = *signature_strings;
2262
2263 dst->element_count = src->element_count;
2264 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2265 return E_OUTOFMEMORY;
2266
2267 for (i = 0; i < src->element_count; ++i)
2268 {
2269 e = &src->elements[i];
2270 dst->elements[i] = *e;
2271
2272 len = strlen(e->semantic_name);
2273 memcpy(ptr, e->semantic_name, len + 1);
2274 dst->elements[i].semantic_name = ptr;
2275 ptr += len + 1;
2276 }
2277
2278 *signature_strings = ptr;
2279
2280 return WINED3D_OK;
2281 }
2282
2283 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2284 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2285 void *parent, const struct wined3d_parent_ops *parent_ops)
2286 {
2287 struct wined3d_shader_signature_element *e;
2288 SIZE_T total, len;
2289 unsigned int i;
2290 HRESULT hr;
2291 char *ptr;
2292
2293 if (!desc->byte_code)
2294 return WINED3DERR_INVALIDCALL;
2295
2296 shader->ref = 1;
2297 shader->device = device;
2298 shader->parent = parent;
2299 shader->parent_ops = parent_ops;
2300
2301 total = 0;
2302 if (desc->input_signature)
2303 {
2304 for (i = 0; i < desc->input_signature->element_count; ++i)
2305 {
2306 e = &desc->input_signature->elements[i];
2307 len = strlen(e->semantic_name);
2308 if (len >= ~(SIZE_T)0 - total)
2309 return E_OUTOFMEMORY;
2310
2311 total += len + 1;
2312 }
2313 }
2314 if (desc->output_signature)
2315 {
2316 for (i = 0; i < desc->output_signature->element_count; ++i)
2317 {
2318 e = &desc->output_signature->elements[i];
2319 len = strlen(e->semantic_name);
2320 if (len >= ~(SIZE_T)0 - total)
2321 return E_OUTOFMEMORY;
2322
2323 total += len + 1;
2324 }
2325 }
2326 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2327 return E_OUTOFMEMORY;
2328 ptr = shader->signature_strings;
2329
2330 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2331 desc->input_signature, &ptr)))
2332 {
2333 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2334 return hr;
2335 }
2336 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2337 desc->output_signature, &ptr)))
2338 {
2339 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2340 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2341 return hr;
2342 }
2343
2344 list_init(&shader->linked_programs);
2345 list_add_head(&device->shaders, &shader->shader_list_entry);
2346
2347 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2348 float_const_count, type, desc->max_version)))
2349 {
2350 WARN("Failed to set function, hr %#x.\n", hr);
2351 shader_cleanup(shader);
2352 }
2353
2354 return hr;
2355 }
2356
2357 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2358 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2359 {
2360 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2361 unsigned int i;
2362 HRESULT hr;
2363
2364 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2365 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2366 return hr;
2367
2368 for (i = 0; i < shader->input_signature.element_count; ++i)
2369 {
2370 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2371
2372 if (!(reg_maps->input_registers & (1 << input->register_idx)) || !input->semantic_name)
2373 continue;
2374
2375 shader->u.vs.attributes[input->register_idx].usage =
2376 shader_usage_from_semantic_name(input->semantic_name);
2377 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2378 }
2379
2380 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2381 shader->lconst_inf_or_nan;
2382
2383 return WINED3D_OK;
2384 }
2385
2386 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2387 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2388 {
2389 HRESULT hr;
2390
2391 if (FAILED(hr = shader_init(shader, device, desc, 0,
2392 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2393 return hr;
2394
2395 shader->load_local_constsF = shader->lconst_inf_or_nan;
2396
2397 return WINED3D_OK;
2398 }
2399
2400 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2401 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_gl_info *gl_info)
2402 {
2403 const struct wined3d_texture *texture;
2404 UINT i;
2405
2406 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2407 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
2408 {
2409 unsigned int rt_fmt_flags = state->fb->render_targets[0]->format_flags;
2410 if (rt_fmt_flags & WINED3DFMT_FLAG_SRGB_WRITE)
2411 {
2412 static unsigned int warned = 0;
2413
2414 args->srgb_correction = 1;
2415 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2416 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2417 "support, expect rendering artifacts.\n");
2418 }
2419 }
2420
2421 if (shader->reg_maps.shader_version.major == 1
2422 && shader->reg_maps.shader_version.minor <= 3)
2423 {
2424 for (i = 0; i < shader->limits->sampler; ++i)
2425 {
2426 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2427
2428 if (flags & WINED3D_TTFF_PROJECTED)
2429 {
2430 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2431
2432 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2433 {
2434 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2435 unsigned int j;
2436 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2437 DWORD max_valid = WINED3D_TTFF_COUNT4;
2438
2439 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2440 {
2441 struct wined3d_vertex_declaration_element *element =
2442 &state->vertex_declaration->elements[j];
2443
2444 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2445 && element->usage_idx == index)
2446 {
2447 max_valid = element->format->component_count;
2448 break;
2449 }
2450 }
2451 if (!tex_transform || tex_transform > max_valid)
2452 {
2453 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2454 tex_transform, max_valid);
2455 tex_transform = max_valid;
2456 }
2457 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2458 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2459 && tex_transform > WINED3D_TTFF_COUNT2)
2460 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2461 && tex_transform > WINED3D_TTFF_COUNT3))
2462 tex_transform |= WINED3D_PSARGS_PROJECTED;
2463 else
2464 {
2465 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2466 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2467 i, tex_transform, resource_type);
2468 }
2469 }
2470 else
2471 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2472
2473 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2474 }
2475 }
2476 }
2477 if (shader->reg_maps.shader_version.major == 1
2478 && shader->reg_maps.shader_version.minor <= 4)
2479 {
2480 for (i = 0; i < shader->limits->sampler; ++i)
2481 {
2482 const struct wined3d_texture *texture = state->textures[i];
2483
2484 if (!shader->reg_maps.resource_info[i].type)
2485 continue;
2486
2487 /* Treat unbound textures as 2D. The dummy texture will provide
2488 * the proper sample value. The tex_types bitmap defaults to
2489 * 2D because of the memset. */
2490 if (!texture)
2491 continue;
2492
2493 switch (texture->target)
2494 {
2495 /* RECT textures are distinguished from 2D textures via np2_fixup */
2496 case GL_TEXTURE_RECTANGLE_ARB:
2497 case GL_TEXTURE_2D:
2498 break;
2499
2500 case GL_TEXTURE_3D:
2501 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2502 break;
2503
2504 case GL_TEXTURE_CUBE_MAP_ARB:
2505 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2506 break;
2507 }
2508 }
2509 }
2510
2511 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2512 {
2513 if (!shader->reg_maps.resource_info[i].type)
2514 continue;
2515
2516 texture = state->textures[i];
2517 if (!texture)
2518 {
2519 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2520 continue;
2521 }
2522 args->color_fixup[i] = texture->resource.format->color_fixup;
2523
2524 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
2525 args->shadow |= 1 << i;
2526
2527 /* Flag samplers that need NP2 texcoord fixup. */
2528 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2529 args->np2_fixup |= (1 << i);
2530 }
2531 if (shader->reg_maps.shader_version.major >= 3)
2532 {
2533 if (position_transformed)
2534 args->vp_mode = pretransformed;
2535 else if (use_vs(state))
2536 args->vp_mode = vertexshader;
2537 else
2538 args->vp_mode = fixedfunction;
2539 args->fog = WINED3D_FFP_PS_FOG_OFF;
2540 }
2541 else
2542 {
2543 args->vp_mode = vertexshader;
2544 if (state->render_states[WINED3D_RS_FOGENABLE])
2545 {
2546 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2547 {
2548 case WINED3D_FOG_NONE:
2549 if (position_transformed || use_vs(state))
2550 {
2551 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2552 break;
2553 }
2554
2555 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2556 {
2557 case WINED3D_FOG_NONE: /* Fall through. */
2558 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2559 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2560 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2561 }
2562 break;
2563
2564 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2565 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2566 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2567 }
2568 }
2569 else
2570 {
2571 args->fog = WINED3D_FFP_PS_FOG_OFF;
2572 }
2573 }
2574 }
2575
2576 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2577 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2578 {
2579 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2580 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2581 HRESULT hr;
2582
2583 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
2584 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
2585 return hr;
2586
2587 for (i = 0; i < MAX_REG_INPUT; ++i)
2588 {
2589 if (shader->u.ps.input_reg_used[i])
2590 {
2591 ++num_regs_used;
2592 highest_reg_used = i;
2593 }
2594 }
2595
2596 /* Don't do any register mapping magic if it is not needed, or if we can't
2597 * achieve anything anyway */
2598 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2599 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2600 {
2601 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2602 {
2603 /* This happens with relative addressing. The input mapper function
2604 * warns about this if the higher registers are declared too, so
2605 * don't write a FIXME here */
2606 WARN("More varying registers used than supported\n");
2607 }
2608
2609 for (i = 0; i < MAX_REG_INPUT; ++i)
2610 {
2611 shader->u.ps.input_reg_map[i] = i;
2612 }
2613
2614 shader->u.ps.declared_in_count = highest_reg_used + 1;
2615 }
2616 else
2617 {
2618 shader->u.ps.declared_in_count = 0;
2619 for (i = 0; i < MAX_REG_INPUT; ++i)
2620 {
2621 if (shader->u.ps.input_reg_used[i])
2622 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2623 else shader->u.ps.input_reg_map[i] = ~0U;
2624 }
2625 }
2626
2627 shader->load_local_constsF = shader->lconst_inf_or_nan;
2628
2629 return WINED3D_OK;
2630 }
2631
2632 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
2633 {
2634 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2635 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
2636 unsigned int i;
2637
2638 if (reg_maps->shader_version.major != 1) return;
2639
2640 for (i = 0; i < shader->limits->sampler; ++i)
2641 {
2642 /* We don't sample from this sampler. */
2643 if (!resource_info[i].type)
2644 continue;
2645
2646 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
2647 {
2648 case WINED3D_SHADER_TEX_2D:
2649 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
2650 break;
2651
2652 case WINED3D_SHADER_TEX_3D:
2653 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
2654 break;
2655
2656 case WINED3D_SHADER_TEX_CUBE:
2657 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
2658 break;
2659 }
2660 }
2661 }
2662
2663 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2664 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2665 {
2666 struct wined3d_shader *object;
2667 HRESULT hr;
2668
2669 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2670 device, desc, parent, parent_ops, shader);
2671
2672 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2673 if (!object)
2674 return E_OUTOFMEMORY;
2675
2676 if (FAILED(hr = geometryshader_init(object, device, desc, parent, parent_ops)))
2677 {
2678 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2679 HeapFree(GetProcessHeap(), 0, object);
2680 return hr;
2681 }
2682
2683 TRACE("Created geometry shader %p.\n", object);
2684 *shader = object;
2685
2686 return WINED3D_OK;
2687 }
2688
2689 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2690 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2691 {
2692 struct wined3d_shader *object;
2693 HRESULT hr;
2694
2695 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2696 device, desc, parent, parent_ops, shader);
2697
2698 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2699 if (!object)
2700 return E_OUTOFMEMORY;
2701
2702 if (FAILED(hr = pixelshader_init(object, device, desc, parent, parent_ops)))
2703 {
2704 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2705 HeapFree(GetProcessHeap(), 0, object);
2706 return hr;
2707 }
2708
2709 TRACE("Created pixel shader %p.\n", object);
2710 *shader = object;
2711
2712 return WINED3D_OK;
2713 }
2714
2715 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2716 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2717 {
2718 struct wined3d_shader *object;
2719 HRESULT hr;
2720
2721 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2722 device, desc, parent, parent_ops, shader);
2723
2724 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2725 if (!object)
2726 return E_OUTOFMEMORY;
2727
2728 if (FAILED(hr = vertexshader_init(object, device, desc, parent, parent_ops)))
2729 {
2730 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2731 HeapFree(GetProcessHeap(), 0, object);
2732 return hr;
2733 }
2734
2735 TRACE("Created vertex shader %p.\n", object);
2736 *shader = object;
2737
2738 return WINED3D_OK;
2739 }
Maarten's multimedia branch