?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dbghelp: Parse the DWARF information in Mach-O modules.
[wine.git] / dlls / wined3d / shader.c
blob5b3252814ddaec25631adefebcef2aa213dc65fe
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 if (reg->idx[0].offset == 2)
554 reg_maps->point_size = 1;
555 break;
556
557 case WINED3DSPR_MISCTYPE:
558 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
559 {
560 if (!reg->idx[0].offset)
561 reg_maps->vpos = 1;
562 else if (reg->idx[0].offset == 1)
563 reg_maps->usesfacing = 1;
564 }
565 break;
566
567 case WINED3DSPR_CONST:
568 if (reg->idx[0].rel_addr)
569 {
570 if (reg->idx[0].offset < reg_maps->min_rel_offset)
571 reg_maps->min_rel_offset = reg->idx[0].offset;
572 if (reg->idx[0].offset > reg_maps->max_rel_offset)
573 reg_maps->max_rel_offset = reg->idx[0].offset;
574 reg_maps->usesrelconstF = TRUE;
575 }
576 else
577 {
578 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
579 {
580 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
581 return FALSE;
582 }
583 else
584 {
585 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
586 }
587 }
588 break;
589
590 case WINED3DSPR_CONSTINT:
591 if (reg->idx[0].offset >= shader->limits->constant_int)
592 {
593 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
594 return FALSE;
595 }
596 else
597 {
598 reg_maps->integer_constants |= (1 << reg->idx[0].offset);
599 }
600 break;
601
602 case WINED3DSPR_CONSTBOOL:
603 if (reg->idx[0].offset >= shader->limits->constant_bool)
604 {
605 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
606 return FALSE;
607 }
608 else
609 {
610 reg_maps->boolean_constants |= (1 << reg->idx[0].offset);
611 }
612 break;
613
614 case WINED3DSPR_COLOROUT:
615 reg_maps->rt_mask |= (1 << reg->idx[0].offset);
616 break;
617
618 default:
619 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
620 reg->type, reg->idx[0].offset, reg->idx[1].offset);
621 break;
622 }
623 return TRUE;
624 }
625
626 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
627 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
628 {
629 struct wined3d_shader_sampler_map_entry *entries, *entry;
630 struct wined3d_shader_sampler_map *map;
631 unsigned int i;
632
633 map = &reg_maps->sampler_map;
634 entries = map->entries;
635 for (i = 0; i < map->count; ++i)
636 {
637 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
638 return;
639 }
640
641 if (!map->size)
642 {
643 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
644 {
645 ERR("Failed to allocate sampler map entries.\n");
646 return;
647 }
648 map->size = 4;
649 map->entries = entries;
650 }
651 else if (map->count == map->size)
652 {
653 size_t new_size = map->size * 2;
654
655 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
656 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
657 {
658 ERR("Failed to resize sampler map entries.\n");
659 return;
660 }
661 map->size = new_size;
662 map->entries = entries;
663 }
664
665 entry = &entries[map->count++];
666 entry->resource_idx = resource_idx;
667 entry->sampler_idx = sampler_idx;
668 entry->bind_idx = bind_idx;
669 }
670
671 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
672 {
673 switch (instr)
674 {
675 case WINED3DSIH_M4x4:
676 case WINED3DSIH_M3x4:
677 return param == 1 ? 3 : 0;
678
679 case WINED3DSIH_M4x3:
680 case WINED3DSIH_M3x3:
681 return param == 1 ? 2 : 0;
682
683 case WINED3DSIH_M3x2:
684 return param == 1 ? 1 : 0;
685
686 default:
687 return 0;
688 }
689 }
690
691 /* Note that this does not count the loop register as an address register. */
692 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
693 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
694 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
695 {
696 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
697 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
698 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
699 void *fe_data = shader->frontend_data;
700 struct wined3d_shader_version shader_version;
701 const DWORD *ptr = byte_code;
702 unsigned int i;
703
704 memset(reg_maps, 0, sizeof(*reg_maps));
705 memset(input_signature_elements, 0, sizeof(input_signature_elements));
706 memset(output_signature_elements, 0, sizeof(output_signature_elements));
707 reg_maps->min_rel_offset = ~0U;
708
709 fe->shader_read_header(fe_data, &ptr, &shader_version);
710 reg_maps->shader_version = shader_version;
711
712 shader_set_limits(shader);
713
714 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
715 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
716 if (!reg_maps->constf)
717 {
718 ERR("Failed to allocate constant map memory.\n");
719 return E_OUTOFMEMORY;
720 }
721
722 while (!fe->shader_is_end(fe_data, &ptr))
723 {
724 struct wined3d_shader_instruction ins;
725
726 /* Fetch opcode. */
727 fe->shader_read_instruction(fe_data, &ptr, &ins);
728
729 /* Unhandled opcode, and its parameters. */
730 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
731 {
732 TRACE("Skipping unrecognized instruction.\n");
733 continue;
734 }
735
736 /* Handle declarations. */
737 if (ins.handler_idx == WINED3DSIH_DCL)
738 {
739 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
740 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
741
742 switch (semantic->reg.reg.type)
743 {
744 /* Mark input registers used. */
745 case WINED3DSPR_INPUT:
746 if (reg_idx >= MAX_REG_INPUT)
747 {
748 ERR("Invalid input register index %u.\n", reg_idx);
749 break;
750 }
751 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
752 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
753 return WINED3DERR_INVALIDCALL;
754 reg_maps->input_registers |= 1 << reg_idx;
755 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
756 break;
757
758 /* Vertex shader: mark 3.0 output registers used, save token. */
759 case WINED3DSPR_OUTPUT:
760 if (reg_idx >= MAX_REG_OUTPUT)
761 {
762 ERR("Invalid output register index %u.\n", reg_idx);
763 break;
764 }
765 reg_maps->output_registers |= 1 << reg_idx;
766 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
767 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
768 reg_maps->fog = 1;
769 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
770 reg_maps->point_size = 1;
771 break;
772
773 case WINED3DSPR_SAMPLER:
774 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
775 case WINED3DSPR_RESOURCE:
776 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
777 {
778 ERR("Invalid resource index %u.\n", reg_idx);
779 break;
780 }
781 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
782 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
783 break;
784
785 default:
786 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
787 break;
788 }
789 }
790 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
791 {
792 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
793 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
794 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
795 else
796 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
797 }
798 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
799 {
800 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
801 shader->u.gs.input_type = ins.declaration.primitive_type;
802 else
803 FIXME("Invalid instruction %#x for shader type %#x.\n",
804 ins.handler_idx, shader_version.type);
805 }
806 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
807 {
808 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
809 shader->u.gs.output_type = ins.declaration.primitive_type;
810 else
811 FIXME("Invalid instruction %#x for shader type %#x.\n",
812 ins.handler_idx, shader_version.type);
813 }
814 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
815 {
816 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
817 shader->u.gs.vertices_out = ins.declaration.count;
818 else
819 FIXME("Invalid instruction %#x for shader type %#x.\n",
820 ins.handler_idx, shader_version.type);
821 }
822 else if (ins.handler_idx == WINED3DSIH_DEF)
823 {
824 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
825 float *value;
826 if (!lconst) return E_OUTOFMEMORY;
827
828 lconst->idx = ins.dst[0].reg.idx[0].offset;
829 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
830 value = (float *)lconst->value;
831
832 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
833 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
834 {
835 if (value[0] < -1.0f) value[0] = -1.0f;
836 else if (value[0] > 1.0f) value[0] = 1.0f;
837 if (value[1] < -1.0f) value[1] = -1.0f;
838 else if (value[1] > 1.0f) value[1] = 1.0f;
839 if (value[2] < -1.0f) value[2] = -1.0f;
840 else if (value[2] > 1.0f) value[2] = 1.0f;
841 if (value[3] < -1.0f) value[3] = -1.0f;
842 else if (value[3] > 1.0f) value[3] = 1.0f;
843 }
844
845 list_add_head(&shader->constantsF, &lconst->entry);
846
847 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
848 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
849 {
850 shader->lconst_inf_or_nan = TRUE;
851 }
852 }
853 else if (ins.handler_idx == WINED3DSIH_DEFI)
854 {
855 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
856 if (!lconst) return E_OUTOFMEMORY;
857
858 lconst->idx = ins.dst[0].reg.idx[0].offset;
859 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
860
861 list_add_head(&shader->constantsI, &lconst->entry);
862 reg_maps->local_int_consts |= (1 << lconst->idx);
863 }
864 else if (ins.handler_idx == WINED3DSIH_DEFB)
865 {
866 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
867 if (!lconst) return E_OUTOFMEMORY;
868
869 lconst->idx = ins.dst[0].reg.idx[0].offset;
870 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
871
872 list_add_head(&shader->constantsB, &lconst->entry);
873 reg_maps->local_bool_consts |= (1 << lconst->idx);
874 }
875 /* For subroutine prototypes. */
876 else if (ins.handler_idx == WINED3DSIH_LABEL)
877 {
878 reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
879 }
880 /* Set texture, address, temporary registers. */
881 else
882 {
883 BOOL color0_mov = FALSE;
884 unsigned int i;
885
886 /* This will loop over all the registers and try to
887 * make a bitmask of the ones we're interested in.
888 *
889 * Relative addressing tokens are ignored, but that's
890 * okay, since we'll catch any address registers when
891 * they are initialized (required by spec). */
892 for (i = 0; i < ins.dst_count; ++i)
893 {
894 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
895 shader_version.type, constf_size))
896 return WINED3DERR_INVALIDCALL;
897
898 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
899 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
900 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
901 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
902 {
903 UINT idx = ins.dst[i].reg.idx[0].offset;
904
905 switch (ins.dst[i].reg.type)
906 {
907 case WINED3DSPR_RASTOUT:
908 switch (idx)
909 {
910 case 0: /* oPos */
911 reg_maps->output_registers |= 1 << 10;
912 shader_signature_from_usage(&output_signature_elements[10],
913 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
914 break;
915
916 case 1: /* oFog */
917 reg_maps->output_registers |= 1 << 11;
918 shader_signature_from_usage(&output_signature_elements[11],
919 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
920 break;
921
922 case 2: /* oPts */
923 reg_maps->output_registers |= 1 << 11;
924 shader_signature_from_usage(&output_signature_elements[11],
925 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
926 break;
927 }
928 break;
929
930 case WINED3DSPR_ATTROUT:
931 if (idx < 2)
932 {
933 idx += 8;
934 if (reg_maps->output_registers & (1 << idx))
935 {
936 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
937 }
938 else
939 {
940 reg_maps->output_registers |= 1 << idx;
941 shader_signature_from_usage(&output_signature_elements[idx],
942 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
943 }
944 }
945 break;
946
947 case WINED3DSPR_TEXCRDOUT:
948
949 reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
950 if (reg_maps->output_registers & (1 << idx))
951 {
952 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
953 }
954 else
955 {
956 reg_maps->output_registers |= 1 << idx;
957 shader_signature_from_usage(&output_signature_elements[idx],
958 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
959 }
960 break;
961
962 default:
963 break;
964 }
965 }
966
967 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
968 {
969 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
970 {
971 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
972 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
973 * the mov and perform the sRGB write correction from the source register.
974 *
975 * However, if the mov is only partial, we can't do this, and if the write
976 * comes from an instruction other than MOV it is hard to do as well. If
977 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
978 shader->u.ps.color0_mov = FALSE;
979 if (ins.handler_idx == WINED3DSIH_MOV
980 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
981 {
982 /* Used later when the source register is read. */
983 color0_mov = TRUE;
984 }
985 }
986 /* Also drop the MOV marker if the source register is overwritten prior to the shader
987 * end
988 */
989 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
990 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
991 {
992 shader->u.ps.color0_mov = FALSE;
993 }
994 }
995
996 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
997 if (shader_version.major == 1
998 && (ins.handler_idx == WINED3DSIH_TEX
999 || ins.handler_idx == WINED3DSIH_TEXBEM
1000 || ins.handler_idx == WINED3DSIH_TEXBEML
1001 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1002 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1003 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1004 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1005 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1006 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1007 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1008 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1009 {
1010 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1011
1012 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1013 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1014 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1015 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1016
1017 /* texbem is only valid with < 1.4 pixel shaders */
1018 if (ins.handler_idx == WINED3DSIH_TEXBEM
1019 || ins.handler_idx == WINED3DSIH_TEXBEML)
1020 {
1021 reg_maps->bumpmat |= 1 << reg_idx;
1022 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1023 {
1024 reg_maps->luminanceparams |= 1 << reg_idx;
1025 }
1026 }
1027 }
1028 else if (ins.handler_idx == WINED3DSIH_BEM)
1029 {
1030 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
1031 }
1032 }
1033
1034 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1035 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1036 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1037 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1038 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1039 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1040 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1041 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1042 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1043 else if (ins.handler_idx == WINED3DSIH_LOOP
1044 || ins.handler_idx == WINED3DSIH_REP)
1045 {
1046 ++cur_loop_depth;
1047 if (cur_loop_depth > max_loop_depth)
1048 max_loop_depth = cur_loop_depth;
1049 }
1050 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1051 || ins.handler_idx == WINED3DSIH_ENDREP)
1052 {
1053 --cur_loop_depth;
1054 }
1055 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1056 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1057 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1058 {
1059 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1060 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1061 }
1062
1063 if (ins.predicate)
1064 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1065 shader_version.type, constf_size))
1066 return WINED3DERR_INVALIDCALL;
1067
1068 for (i = 0; i < ins.src_count; ++i)
1069 {
1070 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1071 struct wined3d_shader_register reg = ins.src[i].reg;
1072
1073 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1074 shader_version.type, constf_size))
1075 return WINED3DERR_INVALIDCALL;
1076 while (count)
1077 {
1078 ++reg.idx[0].offset;
1079 if (!shader_record_register_usage(shader, reg_maps, &reg,
1080 shader_version.type, constf_size))
1081 return WINED3DERR_INVALIDCALL;
1082 --count;
1083 }
1084
1085 if (color0_mov)
1086 {
1087 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1088 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1089 {
1090 shader->u.ps.color0_mov = TRUE;
1091 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1092 }
1093 }
1094 }
1095 }
1096 }
1097 reg_maps->loop_depth = max_loop_depth;
1098
1099 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1100 * R0 is written to the render target. */
1101 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1102 reg_maps->rt_mask |= (1 << 0);
1103
1104 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1105
1106 if (input_signature->elements)
1107 {
1108 for (i = 0; i < input_signature->element_count; ++i)
1109 {
1110 reg_maps->input_registers |= 1 << input_signature->elements[i].register_idx;
1111 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1112 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1113 reg_maps->vpos = 1;
1114 }
1115 }
1116 else if (!input_signature->elements && reg_maps->input_registers)
1117 {
1118 unsigned int count = count_bits(reg_maps->input_registers);
1119 struct wined3d_shader_signature_element *e;
1120 unsigned int i;
1121
1122 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1123 return E_OUTOFMEMORY;
1124 input_signature->element_count = count;
1125
1126 e = input_signature->elements;
1127 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1128 {
1129 if (!(reg_maps->input_registers & (1 << i)))
1130 continue;
1131 input_signature_elements[i].register_idx = i;
1132 *e++ = input_signature_elements[i];
1133 }
1134 }
1135
1136 if (output_signature->elements)
1137 {
1138 for (i = 0; i < output_signature->element_count; ++i)
1139 {
1140 reg_maps->output_registers |= 1 << output_signature->elements[i].register_idx;
1141 }
1142 }
1143 else if (reg_maps->output_registers)
1144 {
1145 unsigned int count = count_bits(reg_maps->output_registers);
1146 struct wined3d_shader_signature_element *e;
1147
1148 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1149 return E_OUTOFMEMORY;
1150 output_signature->element_count = count;
1151
1152 e = output_signature->elements;
1153 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1154 {
1155 if (!(reg_maps->output_registers & (1 << i)))
1156 continue;
1157 *e++ = output_signature_elements[i];
1158 }
1159 }
1160
1161 return WINED3D_OK;
1162 }
1163
1164 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1165 {
1166 DWORD map = 1 << max;
1167 map |= map - 1;
1168 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1169
1170 return wined3d_log2i(map);
1171 }
1172
1173 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
1174 const struct wined3d_shader_version *shader_version)
1175 {
1176 TRACE("dcl");
1177
1178 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1179 {
1180 switch (semantic->resource_type)
1181 {
1182 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1183 TRACE("_2d");
1184 break;
1185
1186 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1187 TRACE("_volume");
1188 break;
1189
1190 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1191 TRACE("_cube");
1192 break;
1193
1194 default:
1195 TRACE("_unknown_ttype(0x%08x)", semantic->resource_type);
1196 break;
1197 }
1198 }
1199 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1200 {
1201 TRACE("_resource_");
1202 switch (semantic->resource_type)
1203 {
1204 case WINED3D_SHADER_RESOURCE_BUFFER:
1205 TRACE("buffer");
1206 break;
1207
1208 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1209 TRACE("texture1d");
1210 break;
1211
1212 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1213 TRACE("texture2d");
1214 break;
1215
1216 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1217 TRACE("texture2dms");
1218 break;
1219
1220 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1221 TRACE("texture3d");
1222 break;
1223
1224 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1225 TRACE("texturecube");
1226 break;
1227
1228 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1229 TRACE("texture1darray");
1230 break;
1231
1232 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1233 TRACE("texture2darray");
1234 break;
1235
1236 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1237 TRACE("texture2dmsarray");
1238 break;
1239
1240 default:
1241 TRACE("unknown");
1242 break;
1243 }
1244 switch (semantic->resource_data_type)
1245 {
1246 case WINED3D_DATA_FLOAT:
1247 TRACE(" (float)");
1248 break;
1249
1250 case WINED3D_DATA_INT:
1251 TRACE(" (int)");
1252 break;
1253
1254 case WINED3D_DATA_UINT:
1255 TRACE(" (uint)");
1256 break;
1257
1258 case WINED3D_DATA_UNORM:
1259 TRACE(" (unorm)");
1260 break;
1261
1262 case WINED3D_DATA_SNORM:
1263 TRACE(" (snorm)");
1264 break;
1265
1266 default:
1267 TRACE(" (unknown)");
1268 break;
1269 }
1270 }
1271 else
1272 {
1273 /* Pixel shaders 3.0 don't have usage semantics. */
1274 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
1275 else TRACE("_");
1276
1277 switch (semantic->usage)
1278 {
1279 case WINED3D_DECL_USAGE_POSITION:
1280 TRACE("position%u", semantic->usage_idx);
1281 break;
1282
1283 case WINED3D_DECL_USAGE_BLEND_INDICES:
1284 TRACE("blend");
1285 break;
1286
1287 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1288 TRACE("weight");
1289 break;
1290
1291 case WINED3D_DECL_USAGE_NORMAL:
1292 TRACE("normal%u", semantic->usage_idx);
1293 break;
1294
1295 case WINED3D_DECL_USAGE_PSIZE:
1296 TRACE("psize");
1297 break;
1298
1299 case WINED3D_DECL_USAGE_COLOR:
1300 if (!semantic->usage_idx) TRACE("color");
1301 else TRACE("specular%u", (semantic->usage_idx - 1));
1302 break;
1303
1304 case WINED3D_DECL_USAGE_TEXCOORD:
1305 TRACE("texture%u", semantic->usage_idx);
1306 break;
1307
1308 case WINED3D_DECL_USAGE_TANGENT:
1309 TRACE("tangent");
1310 break;
1311
1312 case WINED3D_DECL_USAGE_BINORMAL:
1313 TRACE("binormal");
1314 break;
1315
1316 case WINED3D_DECL_USAGE_TESS_FACTOR:
1317 TRACE("tessfactor");
1318 break;
1319
1320 case WINED3D_DECL_USAGE_POSITIONT:
1321 TRACE("positionT%u", semantic->usage_idx);
1322 break;
1323
1324 case WINED3D_DECL_USAGE_FOG:
1325 TRACE("fog");
1326 break;
1327
1328 case WINED3D_DECL_USAGE_DEPTH:
1329 TRACE("depth");
1330 break;
1331
1332 case WINED3D_DECL_USAGE_SAMPLE:
1333 TRACE("sample");
1334 break;
1335
1336 default:
1337 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1338 }
1339 }
1340 }
1341
1342 static void shader_dump_register(const struct wined3d_shader_register *reg,
1343 const struct wined3d_shader_version *shader_version)
1344 {
1345 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1346 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1347 UINT offset = reg->idx[0].offset;
1348
1349 switch (reg->type)
1350 {
1351 case WINED3DSPR_TEMP:
1352 TRACE("r");
1353 break;
1354
1355 case WINED3DSPR_INPUT:
1356 TRACE("v");
1357 break;
1358
1359 case WINED3DSPR_CONST:
1360 case WINED3DSPR_CONST2:
1361 case WINED3DSPR_CONST3:
1362 case WINED3DSPR_CONST4:
1363 TRACE("c");
1364 offset = shader_get_float_offset(reg->type, offset);
1365 break;
1366
1367 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1368 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1369 break;
1370
1371 case WINED3DSPR_RASTOUT:
1372 TRACE("%s", rastout_reg_names[offset]);
1373 break;
1374
1375 case WINED3DSPR_COLOROUT:
1376 TRACE("oC");
1377 break;
1378
1379 case WINED3DSPR_DEPTHOUT:
1380 TRACE("oDepth");
1381 break;
1382
1383 case WINED3DSPR_ATTROUT:
1384 TRACE("oD");
1385 break;
1386
1387 case WINED3DSPR_TEXCRDOUT:
1388 /* Vertex shaders >= 3.0 use general purpose output registers
1389 * (WINED3DSPR_OUTPUT), which can include an address token. */
1390 if (shader_version->major >= 3) TRACE("o");
1391 else TRACE("oT");
1392 break;
1393
1394 case WINED3DSPR_CONSTINT:
1395 TRACE("i");
1396 break;
1397
1398 case WINED3DSPR_CONSTBOOL:
1399 TRACE("b");
1400 break;
1401
1402 case WINED3DSPR_LABEL:
1403 TRACE("l");
1404 break;
1405
1406 case WINED3DSPR_LOOP:
1407 TRACE("aL");
1408 break;
1409
1410 case WINED3DSPR_SAMPLER:
1411 TRACE("s");
1412 break;
1413
1414 case WINED3DSPR_MISCTYPE:
1415 if (offset > 1)
1416 FIXME("Unhandled misctype register %u.\n", offset);
1417 else
1418 TRACE("%s", misctype_reg_names[offset]);
1419 break;
1420
1421 case WINED3DSPR_PREDICATE:
1422 TRACE("p");
1423 break;
1424
1425 case WINED3DSPR_IMMCONST:
1426 TRACE("l");
1427 break;
1428
1429 case WINED3DSPR_CONSTBUFFER:
1430 TRACE("cb");
1431 break;
1432
1433 case WINED3DSPR_PRIMID:
1434 TRACE("primID");
1435 break;
1436
1437 case WINED3DSPR_NULL:
1438 TRACE("null");
1439 break;
1440
1441 case WINED3DSPR_RESOURCE:
1442 TRACE("t");
1443 break;
1444
1445 default:
1446 TRACE("unhandled_rtype(%#x)", reg->type);
1447 break;
1448 }
1449
1450 if (reg->type == WINED3DSPR_IMMCONST)
1451 {
1452 TRACE("(");
1453 switch (reg->immconst_type)
1454 {
1455 case WINED3D_IMMCONST_SCALAR:
1456 switch (reg->data_type)
1457 {
1458 case WINED3D_DATA_FLOAT:
1459 TRACE("%.8e", *(const float *)reg->immconst_data);
1460 break;
1461 case WINED3D_DATA_INT:
1462 TRACE("%d", reg->immconst_data[0]);
1463 break;
1464 case WINED3D_DATA_RESOURCE:
1465 case WINED3D_DATA_SAMPLER:
1466 case WINED3D_DATA_UINT:
1467 TRACE("%u", reg->immconst_data[0]);
1468 break;
1469 default:
1470 TRACE("<unhandled data type %#x>", reg->data_type);
1471 break;
1472 }
1473 break;
1474
1475 case WINED3D_IMMCONST_VEC4:
1476 switch (reg->data_type)
1477 {
1478 case WINED3D_DATA_FLOAT:
1479 TRACE("%.8e, %.8e, %.8e, %.8e",
1480 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1481 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1482 break;
1483 case WINED3D_DATA_INT:
1484 TRACE("%d, %d, %d, %d",
1485 reg->immconst_data[0], reg->immconst_data[1],
1486 reg->immconst_data[2], reg->immconst_data[3]);
1487 break;
1488 case WINED3D_DATA_RESOURCE:
1489 case WINED3D_DATA_SAMPLER:
1490 case WINED3D_DATA_UINT:
1491 TRACE("%u, %u, %u, %u",
1492 reg->immconst_data[0], reg->immconst_data[1],
1493 reg->immconst_data[2], reg->immconst_data[3]);
1494 break;
1495 default:
1496 TRACE("<unhandled data type %#x>", reg->data_type);
1497 break;
1498 }
1499 break;
1500
1501 default:
1502 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1503 break;
1504 }
1505 TRACE(")");
1506 }
1507 else if (reg->type != WINED3DSPR_RASTOUT
1508 && reg->type != WINED3DSPR_MISCTYPE
1509 && reg->type != WINED3DSPR_NULL)
1510 {
1511 if (offset != ~0U)
1512 {
1513 TRACE("[");
1514 if (reg->idx[0].rel_addr)
1515 {
1516 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1517 TRACE(" + ");
1518 }
1519 TRACE("%u]", offset);
1520
1521 if (reg->idx[1].offset != ~0U)
1522 {
1523 TRACE("[");
1524 if (reg->idx[1].rel_addr)
1525 {
1526 shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1527 TRACE(" + ");
1528 }
1529 TRACE("%u]", reg->idx[1].offset);
1530 }
1531 }
1532 }
1533 }
1534
1535 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1536 const struct wined3d_shader_version *shader_version)
1537 {
1538 DWORD write_mask = param->write_mask;
1539
1540 shader_dump_register(&param->reg, shader_version);
1541
1542 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1543 {
1544 static const char write_mask_chars[] = "xyzw";
1545
1546 TRACE(".");
1547 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1548 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1549 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1550 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1551 }
1552 }
1553
1554 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1555 const struct wined3d_shader_version *shader_version)
1556 {
1557 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1558 DWORD swizzle = param->swizzle;
1559
1560 if (src_modifier == WINED3DSPSM_NEG
1561 || src_modifier == WINED3DSPSM_BIASNEG
1562 || src_modifier == WINED3DSPSM_SIGNNEG
1563 || src_modifier == WINED3DSPSM_X2NEG
1564 || src_modifier == WINED3DSPSM_ABSNEG)
1565 TRACE("-");
1566 else if (src_modifier == WINED3DSPSM_COMP)
1567 TRACE("1-");
1568 else if (src_modifier == WINED3DSPSM_NOT)
1569 TRACE("!");
1570
1571 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1572 TRACE("abs(");
1573
1574 shader_dump_register(&param->reg, shader_version);
1575
1576 switch (src_modifier)
1577 {
1578 case WINED3DSPSM_NONE: break;
1579 case WINED3DSPSM_NEG: break;
1580 case WINED3DSPSM_NOT: break;
1581 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1582 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1583 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1584 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1585 case WINED3DSPSM_COMP: break;
1586 case WINED3DSPSM_X2: TRACE("_x2"); break;
1587 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1588 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1589 case WINED3DSPSM_DW: TRACE("_dw"); break;
1590 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1591 case WINED3DSPSM_ABS: TRACE(")"); break;
1592 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1593 }
1594
1595 if (swizzle != WINED3DSP_NOSWIZZLE)
1596 {
1597 static const char swizzle_chars[] = "xyzw";
1598 DWORD swizzle_x = swizzle & 0x03;
1599 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1600 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1601 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1602
1603 if (swizzle_x == swizzle_y
1604 && swizzle_x == swizzle_z
1605 && swizzle_x == swizzle_w)
1606 {
1607 TRACE(".%c", swizzle_chars[swizzle_x]);
1608 }
1609 else
1610 {
1611 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1612 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1613 }
1614 }
1615 }
1616
1617 /* Shared code in order to generate the bulk of the shader string.
1618 * NOTE: A description of how to parse tokens can be found on MSDN. */
1619 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1620 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1621 {
1622 struct wined3d_device *device = shader->device;
1623 const struct wined3d_shader_frontend *fe = shader->frontend;
1624 void *fe_data = shader->frontend_data;
1625 struct wined3d_shader_version shader_version;
1626 struct wined3d_shader_loop_state loop_state;
1627 struct wined3d_shader_instruction ins;
1628 struct wined3d_shader_tex_mx tex_mx;
1629 struct wined3d_shader_context ctx;
1630 const DWORD *ptr = byte_code;
1631
1632 /* Initialize current parsing state. */
1633 tex_mx.current_row = 0;
1634 loop_state.current_depth = 0;
1635 loop_state.current_reg = 0;
1636
1637 ctx.shader = shader;
1638 ctx.gl_info = &device->adapter->gl_info;
1639 ctx.reg_maps = reg_maps;
1640 ctx.buffer = buffer;
1641 ctx.tex_mx = &tex_mx;
1642 ctx.loop_state = &loop_state;
1643 ctx.backend_data = backend_ctx;
1644 ins.ctx = &ctx;
1645
1646 fe->shader_read_header(fe_data, &ptr, &shader_version);
1647
1648 while (!fe->shader_is_end(fe_data, &ptr))
1649 {
1650 /* Read opcode. */
1651 fe->shader_read_instruction(fe_data, &ptr, &ins);
1652
1653 /* Unknown opcode and its parameters. */
1654 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1655 {
1656 TRACE("Skipping unrecognized instruction.\n");
1657 continue;
1658 }
1659
1660 if (ins.predicate)
1661 FIXME("Predicates not implemented.\n");
1662
1663 /* Call appropriate function for output target */
1664 device->shader_backend->shader_handle_instruction(&ins);
1665 }
1666 }
1667
1668 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1669 {
1670 DWORD mmask = dst->modifiers;
1671
1672 switch (dst->shift)
1673 {
1674 case 0: break;
1675 case 13: TRACE("_d8"); break;
1676 case 14: TRACE("_d4"); break;
1677 case 15: TRACE("_d2"); break;
1678 case 1: TRACE("_x2"); break;
1679 case 2: TRACE("_x4"); break;
1680 case 3: TRACE("_x8"); break;
1681 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1682 }
1683
1684 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1685 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1686 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1687
1688 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1689 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1690 }
1691
1692 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1693 {
1694 switch (primitive_type)
1695 {
1696 case WINED3D_PT_UNDEFINED:
1697 TRACE("undefined");
1698 break;
1699 case WINED3D_PT_POINTLIST:
1700 TRACE("pointlist");
1701 break;
1702 case WINED3D_PT_LINELIST:
1703 TRACE("linelist");
1704 break;
1705 case WINED3D_PT_LINESTRIP:
1706 TRACE("linestrip");
1707 break;
1708 case WINED3D_PT_TRIANGLELIST:
1709 TRACE("trianglelist");
1710 break;
1711 case WINED3D_PT_TRIANGLESTRIP:
1712 TRACE("trianglestrip");
1713 break;
1714 case WINED3D_PT_TRIANGLEFAN:
1715 TRACE("trianglefan");
1716 break;
1717 case WINED3D_PT_LINELIST_ADJ:
1718 TRACE("linelist_adj");
1719 break;
1720 case WINED3D_PT_LINESTRIP_ADJ:
1721 TRACE("linestrip_adj");
1722 break;
1723 case WINED3D_PT_TRIANGLELIST_ADJ:
1724 TRACE("trianglelist_adj");
1725 break;
1726 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1727 TRACE("trianglestrip_adj");
1728 break;
1729 default:
1730 TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1731 break;
1732 }
1733 }
1734
1735 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1736 {
1737 struct wined3d_shader_version shader_version;
1738 const DWORD *ptr = byte_code;
1739 const char *type_prefix;
1740 DWORD i;
1741
1742 TRACE("Parsing %p.\n", byte_code);
1743
1744 fe->shader_read_header(fe_data, &ptr, &shader_version);
1745
1746 switch (shader_version.type)
1747 {
1748 case WINED3D_SHADER_TYPE_VERTEX:
1749 type_prefix = "vs";
1750 break;
1751
1752 case WINED3D_SHADER_TYPE_GEOMETRY:
1753 type_prefix = "gs";
1754 break;
1755
1756 case WINED3D_SHADER_TYPE_PIXEL:
1757 type_prefix = "ps";
1758 break;
1759
1760 default:
1761 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1762 type_prefix = "unknown";
1763 break;
1764 }
1765
1766 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1767
1768 while (!fe->shader_is_end(fe_data, &ptr))
1769 {
1770 struct wined3d_shader_instruction ins;
1771
1772 fe->shader_read_instruction(fe_data, &ptr, &ins);
1773 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1774 {
1775 TRACE("Skipping unrecognized instruction.\n");
1776 continue;
1777 }
1778
1779 if (ins.handler_idx == WINED3DSIH_DCL)
1780 {
1781 shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1782 shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1783 TRACE(" ");
1784 shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1785 }
1786 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1787 {
1788 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1789 shader_dump_src_param(&ins.declaration.src, &shader_version);
1790 TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1791 }
1792 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1793 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1794 {
1795 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1796 shader_dump_primitive_type(ins.declaration.primitive_type);
1797 }
1798 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1799 {
1800 TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1801 }
1802 else if (ins.handler_idx == WINED3DSIH_DEF)
1803 {
1804 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1805 ins.dst[0].reg.idx[0].offset),
1806 *(const float *)&ins.src[0].reg.immconst_data[0],
1807 *(const float *)&ins.src[0].reg.immconst_data[1],
1808 *(const float *)&ins.src[0].reg.immconst_data[2],
1809 *(const float *)&ins.src[0].reg.immconst_data[3]);
1810 }
1811 else if (ins.handler_idx == WINED3DSIH_DEFI)
1812 {
1813 TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1814 ins.src[0].reg.immconst_data[0],
1815 ins.src[0].reg.immconst_data[1],
1816 ins.src[0].reg.immconst_data[2],
1817 ins.src[0].reg.immconst_data[3]);
1818 }
1819 else if (ins.handler_idx == WINED3DSIH_DEFB)
1820 {
1821 TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1822 }
1823 else
1824 {
1825 if (ins.predicate)
1826 {
1827 TRACE("(");
1828 shader_dump_src_param(ins.predicate, &shader_version);
1829 TRACE(") ");
1830 }
1831
1832 /* PixWin marks instructions with the coissue flag with a '+' */
1833 if (ins.coissue) TRACE("+");
1834
1835 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1836
1837 if (ins.handler_idx == WINED3DSIH_IFC
1838 || ins.handler_idx == WINED3DSIH_BREAKC)
1839 {
1840 switch (ins.flags)
1841 {
1842 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1843 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1844 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1845 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1846 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1847 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1848 default: TRACE("_(%u)", ins.flags);
1849 }
1850 }
1851 else if (ins.handler_idx == WINED3DSIH_TEX
1852 && shader_version.major >= 2
1853 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1854 {
1855 TRACE("p");
1856 }
1857
1858 for (i = 0; i < ins.dst_count; ++i)
1859 {
1860 shader_dump_ins_modifiers(&ins.dst[i]);
1861 TRACE(!i ? " " : ", ");
1862 shader_dump_dst_param(&ins.dst[i], &shader_version);
1863 }
1864
1865 /* Other source tokens */
1866 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1867 {
1868 TRACE(!i ? " " : ", ");
1869 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1870 }
1871 }
1872 TRACE("\n");
1873 }
1874 }
1875
1876 static void shader_cleanup(struct wined3d_shader *shader)
1877 {
1878 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
1879 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
1880 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
1881 shader->device->shader_backend->shader_destroy(shader);
1882 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1883 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
1884 HeapFree(GetProcessHeap(), 0, shader->function);
1885 shader_delete_constant_list(&shader->constantsF);
1886 shader_delete_constant_list(&shader->constantsB);
1887 shader_delete_constant_list(&shader->constantsI);
1888 list_remove(&shader->shader_list_entry);
1889
1890 if (shader->frontend && shader->frontend_data)
1891 shader->frontend->shader_free(shader->frontend_data);
1892 }
1893
1894 struct shader_none_priv
1895 {
1896 const struct wined3d_vertex_pipe_ops *vertex_pipe;
1897 const struct fragment_pipeline *fragment_pipe;
1898 BOOL ffp_proj_control;
1899 };
1900
1901 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1902 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1903 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
1904 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1905 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1906 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1907 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
1908 const struct wined3d_state *state) {}
1909 static void shader_none_destroy(struct wined3d_shader *shader) {}
1910 static void shader_none_free_context_data(struct wined3d_context *context) {}
1911 static void shader_none_init_context_state(struct wined3d_context *context) {}
1912
1913 /* Context activation is done by the caller. */
1914 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
1915 const struct wined3d_state *state)
1916 {
1917 const struct wined3d_gl_info *gl_info = context->gl_info;
1918 struct shader_none_priv *priv = shader_priv;
1919
1920 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
1921 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
1922 }
1923
1924 /* Context activation is done by the caller. */
1925 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
1926 {
1927 struct shader_none_priv *priv = shader_priv;
1928 const struct wined3d_gl_info *gl_info = context->gl_info;
1929
1930 priv->vertex_pipe->vp_enable(gl_info, FALSE);
1931 priv->fragment_pipe->enable_extension(gl_info, FALSE);
1932
1933 context->shader_update_mask = (1 << WINED3D_SHADER_TYPE_PIXEL)
1934 | (1 << WINED3D_SHADER_TYPE_VERTEX)
1935 | (1 << WINED3D_SHADER_TYPE_GEOMETRY);
1936 }
1937
1938 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
1939 const struct fragment_pipeline *fragment_pipe)
1940 {
1941 struct fragment_caps fragment_caps;
1942 void *vertex_priv, *fragment_priv;
1943 struct shader_none_priv *priv;
1944
1945 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1946 return E_OUTOFMEMORY;
1947
1948 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
1949 {
1950 ERR("Failed to initialize vertex pipe.\n");
1951 HeapFree(GetProcessHeap(), 0, priv);
1952 return E_FAIL;
1953 }
1954
1955 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1956 {
1957 ERR("Failed to initialize fragment pipe.\n");
1958 vertex_pipe->vp_free(device);
1959 HeapFree(GetProcessHeap(), 0, priv);
1960 return E_FAIL;
1961 }
1962
1963 priv->vertex_pipe = vertex_pipe;
1964 priv->fragment_pipe = fragment_pipe;
1965 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
1966 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
1967
1968 device->vertex_priv = vertex_priv;
1969 device->fragment_priv = fragment_priv;
1970 device->shader_priv = priv;
1971
1972 return WINED3D_OK;
1973 }
1974
1975 static void shader_none_free(struct wined3d_device *device)
1976 {
1977 struct shader_none_priv *priv = device->shader_priv;
1978
1979 priv->fragment_pipe->free_private(device);
1980 priv->vertex_pipe->vp_free(device);
1981 HeapFree(GetProcessHeap(), 0, priv);
1982 }
1983
1984 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
1985 {
1986 return TRUE;
1987 }
1988
1989 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1990 {
1991 /* Set the shader caps to 0 for the none shader backend */
1992 caps->vs_version = 0;
1993 caps->gs_version = 0;
1994 caps->ps_version = 0;
1995 caps->vs_uniform_count = 0;
1996 caps->ps_uniform_count = 0;
1997 caps->ps_1x_max_value = 0.0f;
1998 caps->wined3d_caps = 0;
1999 }
2000
2001 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2002 {
2003 /* We "support" every possible fixup, since we don't support any shader
2004 * model, and will never have to actually sample a texture. */
2005 return TRUE;
2006 }
2007
2008 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2009 {
2010 struct shader_none_priv *priv = shader_priv;
2011
2012 return priv->ffp_proj_control;
2013 }
2014
2015 const struct wined3d_shader_backend_ops none_shader_backend =
2016 {
2017 shader_none_handle_instruction,
2018 shader_none_select,
2019 shader_none_disable,
2020 shader_none_select_depth_blt,
2021 shader_none_deselect_depth_blt,
2022 shader_none_update_float_vertex_constants,
2023 shader_none_update_float_pixel_constants,
2024 shader_none_load_constants,
2025 shader_none_destroy,
2026 shader_none_alloc,
2027 shader_none_free,
2028 shader_none_allocate_context_data,
2029 shader_none_free_context_data,
2030 shader_none_init_context_state,
2031 shader_none_get_caps,
2032 shader_none_color_fixup_supported,
2033 shader_none_has_ffp_proj_control,
2034 };
2035
2036 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2037 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2038 enum wined3d_shader_type type, unsigned int max_version)
2039 {
2040 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2041 const struct wined3d_shader_frontend *fe;
2042 HRESULT hr;
2043 unsigned int backend_version;
2044 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2045
2046 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
2047 shader, byte_code, output_signature, float_const_count);
2048
2049 fe = shader_select_frontend(*byte_code);
2050 if (!fe)
2051 {
2052 FIXME("Unable to find frontend for shader.\n");
2053 return WINED3DERR_INVALIDCALL;
2054 }
2055 shader->frontend = fe;
2056 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2057 if (!shader->frontend_data)
2058 {
2059 FIXME("Failed to initialize frontend.\n");
2060 return WINED3DERR_INVALIDCALL;
2061 }
2062
2063 /* First pass: trace shader. */
2064 if (TRACE_ON(d3d_shader))
2065 shader_trace_init(fe, shader->frontend_data, byte_code);
2066
2067 /* Initialize immediate constant lists. */
2068 list_init(&shader->constantsF);
2069 list_init(&shader->constantsB);
2070 list_init(&shader->constantsI);
2071 shader->lconst_inf_or_nan = FALSE;
2072
2073 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2074 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2075 &shader->output_signature, byte_code, float_const_count)))
2076 return hr;
2077
2078 if (reg_maps->shader_version.type != type)
2079 {
2080 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2081 return WINED3DERR_INVALIDCALL;
2082 }
2083 if (reg_maps->shader_version.major > max_version)
2084 {
2085 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2086 return WINED3DERR_INVALIDCALL;
2087 }
2088 switch (type)
2089 {
2090 case WINED3D_SHADER_TYPE_VERTEX:
2091 backend_version = d3d_info->limits.vs_version;
2092 break;
2093 case WINED3D_SHADER_TYPE_GEOMETRY:
2094 backend_version = d3d_info->limits.gs_version;
2095 break;
2096 case WINED3D_SHADER_TYPE_PIXEL:
2097 backend_version = d3d_info->limits.ps_version;
2098 break;
2099 default:
2100 FIXME("No backend version-checking for this shader type\n");
2101 backend_version = 0;
2102 }
2103 if (reg_maps->shader_version.major > backend_version)
2104 {
2105 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2106 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2107 return WINED3DERR_INVALIDCALL;
2108 }
2109
2110 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2111 if (!shader->function)
2112 return E_OUTOFMEMORY;
2113 memcpy(shader->function, byte_code, shader->functionLength);
2114
2115 return WINED3D_OK;
2116 }
2117
2118 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2119 {
2120 ULONG refcount = InterlockedIncrement(&shader->ref);
2121
2122 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2123
2124 return refcount;
2125 }
2126
2127 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2128 {
2129 ULONG refcount = InterlockedDecrement(&shader->ref);
2130
2131 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2132
2133 if (!refcount)
2134 {
2135 shader_cleanup(shader);
2136 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2137 HeapFree(GetProcessHeap(), 0, shader);
2138 }
2139
2140 return refcount;
2141 }
2142
2143 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2144 {
2145 TRACE("shader %p.\n", shader);
2146
2147 return shader->parent;
2148 }
2149
2150 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2151 void *byte_code, UINT *byte_code_size)
2152 {
2153 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2154
2155 if (!byte_code)
2156 {
2157 *byte_code_size = shader->functionLength;
2158 return WINED3D_OK;
2159 }
2160
2161 if (*byte_code_size < shader->functionLength)
2162 {
2163 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2164 * than the required size we should write the required size and
2165 * return D3DERR_MOREDATA. That's not actually true. */
2166 return WINED3DERR_INVALIDCALL;
2167 }
2168
2169 memcpy(byte_code, shader->function, shader->functionLength);
2170
2171 return WINED3D_OK;
2172 }
2173
2174 /* Set local constants for d3d8 shaders. */
2175 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2176 UINT start_idx, const float *src_data, UINT count)
2177 {
2178 UINT end_idx = start_idx + count;
2179 UINT i;
2180
2181 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2182
2183 if (end_idx > shader->limits->constant_float)
2184 {
2185 WARN("end_idx %u > float constants limit %u.\n",
2186 end_idx, shader->limits->constant_float);
2187 end_idx = shader->limits->constant_float;
2188 }
2189
2190 for (i = start_idx; i < end_idx; ++i)
2191 {
2192 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2193 float *value;
2194 if (!lconst)
2195 return E_OUTOFMEMORY;
2196
2197 lconst->idx = i;
2198 value = (float *)lconst->value;
2199 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2200 list_add_head(&shader->constantsF, &lconst->entry);
2201
2202 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2203 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2204 {
2205 shader->lconst_inf_or_nan = TRUE;
2206 }
2207 }
2208
2209 return WINED3D_OK;
2210 }
2211
2212 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2213 WORD swizzle_map, struct vs_compile_args *args)
2214 {
2215 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2216 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2217 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2218 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2219 args->point_size = state->gl_primitive_type == GL_POINTS;
2220 args->per_vertex_point_size = shader->reg_maps.point_size;
2221 args->swizzle_map = swizzle_map;
2222 }
2223
2224 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2225 {
2226 if (usage_idx1 != usage_idx2)
2227 return FALSE;
2228 if (usage1 == usage2)
2229 return TRUE;
2230 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2231 return TRUE;
2232 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2233 return TRUE;
2234
2235 return FALSE;
2236 }
2237
2238 BOOL vshader_get_input(const struct wined3d_shader *shader,
2239 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2240 {
2241 WORD map = shader->reg_maps.input_registers;
2242 unsigned int i;
2243
2244 for (i = 0; map; map >>= 1, ++i)
2245 {
2246 if (!(map & 1)) continue;
2247
2248 if (match_usage(shader->u.vs.attributes[i].usage,
2249 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2250 {
2251 *regnum = i;
2252 return TRUE;
2253 }
2254 }
2255 return FALSE;
2256 }
2257
2258 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2259 const struct wined3d_shader_signature *src, char **signature_strings)
2260 {
2261 struct wined3d_shader_signature_element *e;
2262 unsigned int i;
2263 SIZE_T len;
2264 char *ptr;
2265
2266 ptr = *signature_strings;
2267
2268 dst->element_count = src->element_count;
2269 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2270 return E_OUTOFMEMORY;
2271
2272 for (i = 0; i < src->element_count; ++i)
2273 {
2274 e = &src->elements[i];
2275 dst->elements[i] = *e;
2276
2277 len = strlen(e->semantic_name);
2278 memcpy(ptr, e->semantic_name, len + 1);
2279 dst->elements[i].semantic_name = ptr;
2280 ptr += len + 1;
2281 }
2282
2283 *signature_strings = ptr;
2284
2285 return WINED3D_OK;
2286 }
2287
2288 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2289 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2290 void *parent, const struct wined3d_parent_ops *parent_ops)
2291 {
2292 struct wined3d_shader_signature_element *e;
2293 SIZE_T total, len;
2294 unsigned int i;
2295 HRESULT hr;
2296 char *ptr;
2297
2298 if (!desc->byte_code)
2299 return WINED3DERR_INVALIDCALL;
2300
2301 shader->ref = 1;
2302 shader->device = device;
2303 shader->parent = parent;
2304 shader->parent_ops = parent_ops;
2305
2306 total = 0;
2307 if (desc->input_signature)
2308 {
2309 for (i = 0; i < desc->input_signature->element_count; ++i)
2310 {
2311 e = &desc->input_signature->elements[i];
2312 len = strlen(e->semantic_name);
2313 if (len >= ~(SIZE_T)0 - total)
2314 return E_OUTOFMEMORY;
2315
2316 total += len + 1;
2317 }
2318 }
2319 if (desc->output_signature)
2320 {
2321 for (i = 0; i < desc->output_signature->element_count; ++i)
2322 {
2323 e = &desc->output_signature->elements[i];
2324 len = strlen(e->semantic_name);
2325 if (len >= ~(SIZE_T)0 - total)
2326 return E_OUTOFMEMORY;
2327
2328 total += len + 1;
2329 }
2330 }
2331 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2332 return E_OUTOFMEMORY;
2333 ptr = shader->signature_strings;
2334
2335 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2336 desc->input_signature, &ptr)))
2337 {
2338 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2339 return hr;
2340 }
2341 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2342 desc->output_signature, &ptr)))
2343 {
2344 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2345 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2346 return hr;
2347 }
2348
2349 list_init(&shader->linked_programs);
2350 list_add_head(&device->shaders, &shader->shader_list_entry);
2351
2352 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2353 float_const_count, type, desc->max_version)))
2354 {
2355 WARN("Failed to set function, hr %#x.\n", hr);
2356 shader_cleanup(shader);
2357 }
2358
2359 return hr;
2360 }
2361
2362 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2363 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2364 {
2365 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2366 unsigned int i;
2367 HRESULT hr;
2368
2369 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2370 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2371 return hr;
2372
2373 for (i = 0; i < shader->input_signature.element_count; ++i)
2374 {
2375 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2376
2377 if (!(reg_maps->input_registers & (1 << input->register_idx)) || !input->semantic_name)
2378 continue;
2379
2380 shader->u.vs.attributes[input->register_idx].usage =
2381 shader_usage_from_semantic_name(input->semantic_name);
2382 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2383 }
2384
2385 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2386 shader->lconst_inf_or_nan;
2387
2388 return WINED3D_OK;
2389 }
2390
2391 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2392 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2393 {
2394 HRESULT hr;
2395
2396 if (FAILED(hr = shader_init(shader, device, desc, 0,
2397 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2398 return hr;
2399
2400 shader->load_local_constsF = shader->lconst_inf_or_nan;
2401
2402 return WINED3D_OK;
2403 }
2404
2405 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2406 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_gl_info *gl_info)
2407 {
2408 const struct wined3d_texture *texture;
2409 UINT i;
2410
2411 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2412 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
2413 {
2414 unsigned int rt_fmt_flags = state->fb->render_targets[0]->format_flags;
2415 if (rt_fmt_flags & WINED3DFMT_FLAG_SRGB_WRITE)
2416 {
2417 static unsigned int warned = 0;
2418
2419 args->srgb_correction = 1;
2420 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2421 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2422 "support, expect rendering artifacts.\n");
2423 }
2424 }
2425
2426 if (shader->reg_maps.shader_version.major == 1
2427 && shader->reg_maps.shader_version.minor <= 3)
2428 {
2429 for (i = 0; i < shader->limits->sampler; ++i)
2430 {
2431 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2432
2433 if (flags & WINED3D_TTFF_PROJECTED)
2434 {
2435 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2436
2437 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2438 {
2439 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2440 unsigned int j;
2441 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2442 DWORD max_valid = WINED3D_TTFF_COUNT4;
2443
2444 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2445 {
2446 struct wined3d_vertex_declaration_element *element =
2447 &state->vertex_declaration->elements[j];
2448
2449 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2450 && element->usage_idx == index)
2451 {
2452 max_valid = element->format->component_count;
2453 break;
2454 }
2455 }
2456 if (!tex_transform || tex_transform > max_valid)
2457 {
2458 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2459 tex_transform, max_valid);
2460 tex_transform = max_valid;
2461 }
2462 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2463 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2464 && tex_transform > WINED3D_TTFF_COUNT2)
2465 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2466 && tex_transform > WINED3D_TTFF_COUNT3))
2467 tex_transform |= WINED3D_PSARGS_PROJECTED;
2468 else
2469 {
2470 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2471 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2472 i, tex_transform, resource_type);
2473 }
2474 }
2475 else
2476 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2477
2478 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2479 }
2480 }
2481 }
2482 if (shader->reg_maps.shader_version.major == 1
2483 && shader->reg_maps.shader_version.minor <= 4)
2484 {
2485 for (i = 0; i < shader->limits->sampler; ++i)
2486 {
2487 const struct wined3d_texture *texture = state->textures[i];
2488
2489 if (!shader->reg_maps.resource_info[i].type)
2490 continue;
2491
2492 /* Treat unbound textures as 2D. The dummy texture will provide
2493 * the proper sample value. The tex_types bitmap defaults to
2494 * 2D because of the memset. */
2495 if (!texture)
2496 continue;
2497
2498 switch (texture->target)
2499 {
2500 /* RECT textures are distinguished from 2D textures via np2_fixup */
2501 case GL_TEXTURE_RECTANGLE_ARB:
2502 case GL_TEXTURE_2D:
2503 break;
2504
2505 case GL_TEXTURE_3D:
2506 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2507 break;
2508
2509 case GL_TEXTURE_CUBE_MAP_ARB:
2510 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2511 break;
2512 }
2513 }
2514 }
2515
2516 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2517 {
2518 if (!shader->reg_maps.resource_info[i].type)
2519 continue;
2520
2521 texture = state->textures[i];
2522 if (!texture)
2523 {
2524 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2525 continue;
2526 }
2527 args->color_fixup[i] = texture->resource.format->color_fixup;
2528
2529 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
2530 args->shadow |= 1 << i;
2531
2532 /* Flag samplers that need NP2 texcoord fixup. */
2533 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2534 args->np2_fixup |= (1 << i);
2535 }
2536 if (shader->reg_maps.shader_version.major >= 3)
2537 {
2538 if (position_transformed)
2539 args->vp_mode = pretransformed;
2540 else if (use_vs(state))
2541 args->vp_mode = vertexshader;
2542 else
2543 args->vp_mode = fixedfunction;
2544 args->fog = WINED3D_FFP_PS_FOG_OFF;
2545 }
2546 else
2547 {
2548 args->vp_mode = vertexshader;
2549 if (state->render_states[WINED3D_RS_FOGENABLE])
2550 {
2551 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2552 {
2553 case WINED3D_FOG_NONE:
2554 if (position_transformed || use_vs(state))
2555 {
2556 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2557 break;
2558 }
2559
2560 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2561 {
2562 case WINED3D_FOG_NONE: /* Fall through. */
2563 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2564 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2565 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2566 }
2567 break;
2568
2569 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2570 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2571 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2572 }
2573 }
2574 else
2575 {
2576 args->fog = WINED3D_FFP_PS_FOG_OFF;
2577 }
2578 }
2579
2580 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
2581 && state->gl_primitive_type == GL_POINTS;
2582 }
2583
2584 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2585 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2586 {
2587 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2588 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2589 HRESULT hr;
2590
2591 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
2592 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
2593 return hr;
2594
2595 for (i = 0; i < MAX_REG_INPUT; ++i)
2596 {
2597 if (shader->u.ps.input_reg_used[i])
2598 {
2599 ++num_regs_used;
2600 highest_reg_used = i;
2601 }
2602 }
2603
2604 /* Don't do any register mapping magic if it is not needed, or if we can't
2605 * achieve anything anyway */
2606 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2607 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2608 {
2609 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2610 {
2611 /* This happens with relative addressing. The input mapper function
2612 * warns about this if the higher registers are declared too, so
2613 * don't write a FIXME here */
2614 WARN("More varying registers used than supported\n");
2615 }
2616
2617 for (i = 0; i < MAX_REG_INPUT; ++i)
2618 {
2619 shader->u.ps.input_reg_map[i] = i;
2620 }
2621
2622 shader->u.ps.declared_in_count = highest_reg_used + 1;
2623 }
2624 else
2625 {
2626 shader->u.ps.declared_in_count = 0;
2627 for (i = 0; i < MAX_REG_INPUT; ++i)
2628 {
2629 if (shader->u.ps.input_reg_used[i])
2630 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2631 else shader->u.ps.input_reg_map[i] = ~0U;
2632 }
2633 }
2634
2635 shader->load_local_constsF = shader->lconst_inf_or_nan;
2636
2637 return WINED3D_OK;
2638 }
2639
2640 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
2641 {
2642 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2643 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
2644 unsigned int i;
2645
2646 if (reg_maps->shader_version.major != 1) return;
2647
2648 for (i = 0; i < shader->limits->sampler; ++i)
2649 {
2650 /* We don't sample from this sampler. */
2651 if (!resource_info[i].type)
2652 continue;
2653
2654 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
2655 {
2656 case WINED3D_SHADER_TEX_2D:
2657 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
2658 break;
2659
2660 case WINED3D_SHADER_TEX_3D:
2661 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
2662 break;
2663
2664 case WINED3D_SHADER_TEX_CUBE:
2665 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
2666 break;
2667 }
2668 }
2669 }
2670
2671 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2672 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2673 {
2674 struct wined3d_shader *object;
2675 HRESULT hr;
2676
2677 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2678 device, desc, parent, parent_ops, shader);
2679
2680 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2681 if (!object)
2682 return E_OUTOFMEMORY;
2683
2684 if (FAILED(hr = geometryshader_init(object, device, desc, parent, parent_ops)))
2685 {
2686 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2687 HeapFree(GetProcessHeap(), 0, object);
2688 return hr;
2689 }
2690
2691 TRACE("Created geometry shader %p.\n", object);
2692 *shader = object;
2693
2694 return WINED3D_OK;
2695 }
2696
2697 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2698 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2699 {
2700 struct wined3d_shader *object;
2701 HRESULT hr;
2702
2703 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2704 device, desc, parent, parent_ops, shader);
2705
2706 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2707 if (!object)
2708 return E_OUTOFMEMORY;
2709
2710 if (FAILED(hr = pixelshader_init(object, device, desc, parent, parent_ops)))
2711 {
2712 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2713 HeapFree(GetProcessHeap(), 0, object);
2714 return hr;
2715 }
2716
2717 TRACE("Created pixel shader %p.\n", object);
2718 *shader = object;
2719
2720 return WINED3D_OK;
2721 }
2722
2723 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2724 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2725 {
2726 struct wined3d_shader *object;
2727 HRESULT hr;
2728
2729 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2730 device, desc, parent, parent_ops, shader);
2731
2732 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2733 if (!object)
2734 return E_OUTOFMEMORY;
2735
2736 if (FAILED(hr = vertexshader_init(object, device, desc, parent, parent_ops)))
2737 {
2738 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2739 HeapFree(GetProcessHeap(), 0, object);
2740 return hr;
2741 }
2742
2743 TRACE("Created vertex shader %p.\n", object);
2744 *shader = object;
2745
2746 return WINED3D_OK;
2747 }
Windows API implementation