?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
winedevice: Get rid of global driver_hkey variable and fix some leaks.
[wine.git] / dlls / wined3d / shader.c
blob88c064cfaafc45510215eb185d677b547151cb68
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_GLOBAL_FLAGS */ "dcl_globalFlags",
59 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
60 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
61 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
62 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
63 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
64 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
65 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
66 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
67 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
68 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
69 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
70 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
71 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
72 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
73 /* WINED3DSIH_DEF */ "def",
74 /* WINED3DSIH_DEFB */ "defb",
75 /* WINED3DSIH_DEFI */ "defi",
76 /* WINED3DSIH_DIV */ "div",
77 /* WINED3DSIH_DP2 */ "dp2",
78 /* WINED3DSIH_DP2ADD */ "dp2add",
79 /* WINED3DSIH_DP3 */ "dp3",
80 /* WINED3DSIH_DP4 */ "dp4",
81 /* WINED3DSIH_DST */ "dst",
82 /* WINED3DSIH_DSX */ "dsx",
83 /* WINED3DSIH_DSY */ "dsy",
84 /* WINED3DSIH_ELSE */ "else",
85 /* WINED3DSIH_EMIT */ "emit",
86 /* WINED3DSIH_ENDIF */ "endif",
87 /* WINED3DSIH_ENDLOOP */ "endloop",
88 /* WINED3DSIH_ENDREP */ "endrep",
89 /* WINED3DSIH_EQ */ "eq",
90 /* WINED3DSIH_EXP */ "exp",
91 /* WINED3DSIH_EXPP */ "expp",
92 /* WINED3DSIH_FRC */ "frc",
93 /* WINED3DSIH_FTOI */ "ftoi",
94 /* WINED3DSIH_FTOU */ "ftou",
95 /* WINED3DSIH_GE */ "ge",
96 /* WINED3DSIH_IADD */ "iadd",
97 /* WINED3DSIH_IEQ */ "ieq",
98 /* WINED3DSIH_IF */ "if",
99 /* WINED3DSIH_IFC */ "ifc",
100 /* WINED3DSIH_IGE */ "ige",
101 /* WINED3DSIH_ILT */ "ilt",
102 /* WINED3DSIH_IMAD */ "imad",
103 /* WINED3DSIH_IMAX */ "imax",
104 /* WINED3DSIH_IMIN */ "imin",
105 /* WINED3DSIH_IMUL */ "imul",
106 /* WINED3DSIH_INE */ "ine",
107 /* WINED3DSIH_INEG */ "ineg",
108 /* WINED3DSIH_ISHL */ "ishl",
109 /* WINED3DSIH_ITOF */ "itof",
110 /* WINED3DSIH_LABEL */ "label",
111 /* WINED3DSIH_LD */ "ld",
112 /* WINED3DSIH_LIT */ "lit",
113 /* WINED3DSIH_LOG */ "log",
114 /* WINED3DSIH_LOGP */ "logp",
115 /* WINED3DSIH_LOOP */ "loop",
116 /* WINED3DSIH_LRP */ "lrp",
117 /* WINED3DSIH_LT */ "lt",
118 /* WINED3DSIH_M3x2 */ "m3x2",
119 /* WINED3DSIH_M3x3 */ "m3x3",
120 /* WINED3DSIH_M3x4 */ "m3x4",
121 /* WINED3DSIH_M4x3 */ "m4x3",
122 /* WINED3DSIH_M4x4 */ "m4x4",
123 /* WINED3DSIH_MAD */ "mad",
124 /* WINED3DSIH_MAX */ "max",
125 /* WINED3DSIH_MIN */ "min",
126 /* WINED3DSIH_MOV */ "mov",
127 /* WINED3DSIH_MOVA */ "mova",
128 /* WINED3DSIH_MOVC */ "movc",
129 /* WINED3DSIH_MUL */ "mul",
130 /* WINED3DSIH_NE */ "ne",
131 /* WINED3DSIH_NOP */ "nop",
132 /* WINED3DSIH_NOT */ "not",
133 /* WINED3DSIH_NRM */ "nrm",
134 /* WINED3DSIH_OR */ "or",
135 /* WINED3DSIH_PHASE */ "phase",
136 /* WINED3DSIH_POW */ "pow",
137 /* WINED3DSIH_RCP */ "rcp",
138 /* WINED3DSIH_REP */ "rep",
139 /* WINED3DSIH_RESINFO */ "resinfo",
140 /* WINED3DSIH_RET */ "ret",
141 /* WINED3DSIH_ROUND_NI */ "round_ni",
142 /* WINED3DSIH_ROUND_PI */ "round_pi",
143 /* WINED3DSIH_ROUND_Z */ "round_z",
144 /* WINED3DSIH_RSQ */ "rsq",
145 /* WINED3DSIH_SAMPLE */ "sample",
146 /* WINED3DSIH_SAMPLE_B */ "sample_b",
147 /* WINED3DSIH_SAMPLE_C */ "sample_c",
148 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
149 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
150 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
151 /* WINED3DSIH_SETP */ "setp",
152 /* WINED3DSIH_SGE */ "sge",
153 /* WINED3DSIH_SGN */ "sgn",
154 /* WINED3DSIH_SINCOS */ "sincos",
155 /* WINED3DSIH_SLT */ "slt",
156 /* WINED3DSIH_SQRT */ "sqrt",
157 /* WINED3DSIH_SUB */ "sub",
158 /* WINED3DSIH_TEX */ "texld",
159 /* WINED3DSIH_TEXBEM */ "texbem",
160 /* WINED3DSIH_TEXBEML */ "texbeml",
161 /* WINED3DSIH_TEXCOORD */ "texcrd",
162 /* WINED3DSIH_TEXDEPTH */ "texdepth",
163 /* WINED3DSIH_TEXDP3 */ "texdp3",
164 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
165 /* WINED3DSIH_TEXKILL */ "texkill",
166 /* WINED3DSIH_TEXLDD */ "texldd",
167 /* WINED3DSIH_TEXLDL */ "texldl",
168 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
169 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
170 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
171 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
172 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
173 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
174 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
175 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
176 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
177 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
178 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
179 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
180 /* WINED3DSIH_UDIV */ "udiv",
181 /* WINED3DSIH_UGE */ "uge",
182 /* WINED3DSIH_USHR */ "ushr",
183 /* WINED3DSIH_UTOF */ "utof",
184 /* WINED3DSIH_XOR */ "xor",
185 };
186
187 static const char * const semantic_names[] =
188 {
189 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
190 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
191 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
192 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
193 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
194 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
195 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
196 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
197 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
198 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
199 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
200 /* WINED3D_DECL_USAGE_FOG */ "FOG",
201 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
202 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
203 };
204
205 static const struct
206 {
207 enum wined3d_sysval_semantic sysval_semantic;
208 const char *sysval_name;
209 }
210 sysval_semantic_names[] =
211 {
212 {WINED3D_SV_POSITION, "SV_Position"},
213 {WINED3D_SV_INSTANCEID, "SV_InstanceID"},
214 {WINED3D_SV_PRIMITIVEID, "SV_PrimitiveID"},
215 {WINED3D_SV_ISFRONTFACE, "SV_IsFrontFace"},
216 {WINED3D_SV_SAMPLEINDEX, "SV_SampleIndex"},
217 };
218
219 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
220 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
221
222 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
223 {
224 if (handler_idx >= sizeof(shader_opcode_names) / sizeof(*shader_opcode_names))
225 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
226
227 return shader_opcode_names[handler_idx];
228 }
229
230 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
231 {
232 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
233 {
234 FIXME("Unrecognized usage %#x.\n", usage);
235 return "UNRECOGNIZED";
236 }
237
238 return semantic_names[usage];
239 }
240
241 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
242 {
243 unsigned int i;
244
245 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
246 {
247 if (!strcmp(name, semantic_names[i])) return i;
248 }
249
250 return ~0U;
251 }
252
253 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
254 {
255 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
256 }
257
258 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
259 const struct wined3d_shader_semantic *s)
260 {
261 e->semantic_name = shader_semantic_name_from_usage(s->usage);
262 e->semantic_idx = s->usage_idx;
263 e->sysval_semantic = 0;
264 e->component_type = 0;
265 e->register_idx = s->reg.reg.idx[0].offset;
266 e->mask = s->reg.write_mask;
267 }
268
269 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
270 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
271 {
272 e->semantic_name = shader_semantic_name_from_usage(usage);
273 e->semantic_idx = usage_idx;
274 e->sysval_semantic = 0;
275 e->component_type = 0;
276 e->register_idx = reg_idx;
277 e->mask = write_mask;
278 }
279
280 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
281 {
282 switch (version_token >> 16)
283 {
284 case WINED3D_SM1_VS:
285 case WINED3D_SM1_PS:
286 return &sm1_shader_frontend;
287
288 case WINED3D_SM4_PS:
289 case WINED3D_SM4_VS:
290 case WINED3D_SM4_GS:
291 return &sm4_shader_frontend;
292
293 default:
294 FIXME("Unrecognised version token %#x\n", version_token);
295 return NULL;
296 }
297 }
298
299 void string_buffer_clear(struct wined3d_string_buffer *buffer)
300 {
301 buffer->buffer[0] = '\0';
302 buffer->content_size = 0;
303 }
304
305 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
306 {
307 buffer->buffer_size = 32;
308 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
309 {
310 ERR("Failed to allocate shader buffer memory.\n");
311 return FALSE;
312 }
313
314 string_buffer_clear(buffer);
315 return TRUE;
316 }
317
318 void string_buffer_free(struct wined3d_string_buffer *buffer)
319 {
320 HeapFree(GetProcessHeap(), 0, buffer->buffer);
321 }
322
323 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
324 {
325 char *new_buffer;
326 unsigned int new_buffer_size = buffer->buffer_size * 2;
327
328 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
329 new_buffer_size *= 2;
330 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
331 {
332 ERR("Failed to grow buffer.\n");
333 buffer->buffer[buffer->content_size] = '\0';
334 return FALSE;
335 }
336 buffer->buffer = new_buffer;
337 buffer->buffer_size = new_buffer_size;
338 return TRUE;
339 }
340
341 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
342 {
343 unsigned int rem;
344 int rc;
345
346 rem = buffer->buffer_size - buffer->content_size;
347 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
348 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
349 return rc;
350
351 buffer->content_size += rc;
352 return 0;
353 }
354
355 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
356 {
357 va_list args;
358 int ret;
359
360 for (;;)
361 {
362 va_start(args, format);
363 ret = shader_vaddline(buffer, format, args);
364 va_end(args);
365 if (!ret)
366 return ret;
367 if (!string_buffer_resize(buffer, ret))
368 return -1;
369 }
370 }
371
372 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
373 {
374 struct wined3d_string_buffer *buffer;
375
376 if (list_empty(&list->list))
377 {
378 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
379 if (!buffer || !string_buffer_init(buffer))
380 {
381 ERR("Couldn't allocate buffer for temporary string.\n");
382 if (buffer)
383 HeapFree(GetProcessHeap(), 0, buffer);
384 return NULL;
385 }
386 }
387 else
388 {
389 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
390 list_remove(&buffer->entry);
391 }
392 string_buffer_clear(buffer);
393 return buffer;
394 }
395
396 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
397 {
398 if (!buffer)
399 return 0;
400 string_buffer_clear(buffer);
401 return shader_vaddline(buffer, format, args);
402 }
403
404 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
405 {
406 va_list args;
407 int ret;
408
409 for (;;)
410 {
411 va_start(args, format);
412 ret = string_buffer_vsprintf(buffer, format, args);
413 va_end(args);
414 if (!ret)
415 return;
416 if (!string_buffer_resize(buffer, ret))
417 return;
418 }
419 }
420
421 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
422 {
423 if (!buffer)
424 return;
425 list_add_head(&list->list, &buffer->entry);
426 }
427
428 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
429 {
430 list_init(&list->list);
431 }
432
433 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
434 {
435 struct wined3d_string_buffer *buffer, *buffer_next;
436
437 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
438 {
439 string_buffer_free(buffer);
440 HeapFree(GetProcessHeap(), 0, buffer);
441 }
442 list_init(&list->list);
443 }
444
445 /* Convert floating point offset relative to a register file to an absolute
446 * offset for float constants. */
447 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
448 {
449 switch (register_type)
450 {
451 case WINED3DSPR_CONST: return register_idx;
452 case WINED3DSPR_CONST2: return 2048 + register_idx;
453 case WINED3DSPR_CONST3: return 4096 + register_idx;
454 case WINED3DSPR_CONST4: return 6144 + register_idx;
455 default:
456 FIXME("Unsupported register type: %u.\n", register_type);
457 return register_idx;
458 }
459 }
460
461 static void shader_delete_constant_list(struct list *clist)
462 {
463 struct wined3d_shader_lconst *constant, *constant_next;
464
465 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
466 HeapFree(GetProcessHeap(), 0, constant);
467 list_init(clist);
468 }
469
470 static void shader_set_limits(struct wined3d_shader *shader)
471 {
472 static const struct limits_entry
473 {
474 unsigned int min_version;
475 unsigned int max_version;
476 struct wined3d_shader_limits limits;
477 }
478 vs_limits[] =
479 {
480 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
481 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
482 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
483 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
484 * even though they are capable of supporting much more (GL
485 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
486 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
487 * shaders to 256. */
488 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
489 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
490 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
491 {0}
492 },
493 gs_limits[] =
494 {
495 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
496 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
497 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
498 {0}
499 },
500 ps_limits[] =
501 {
502 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
503 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
504 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
505 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
506 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
507 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
508 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
509 {0}
510 };
511 const struct limits_entry *limits_array;
512 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
513 shader->reg_maps.shader_version.minor);
514 int i = 0;
515
516 switch (shader->reg_maps.shader_version.type)
517 {
518 default:
519 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
520 /* Fall-through. */
521 case WINED3D_SHADER_TYPE_VERTEX:
522 limits_array = vs_limits;
523 break;
524 case WINED3D_SHADER_TYPE_GEOMETRY:
525 limits_array = gs_limits;
526 break;
527 case WINED3D_SHADER_TYPE_PIXEL:
528 limits_array = ps_limits;
529 break;
530 }
531
532 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
533 {
534 if (shader_version <= limits_array[i].max_version)
535 {
536 shader->limits = &limits_array[i].limits;
537 break;
538 }
539 ++i;
540 }
541 if (!shader->limits)
542 {
543 FIXME("Unexpected shader version \"%u.%u\".\n",
544 shader->reg_maps.shader_version.major,
545 shader->reg_maps.shader_version.minor);
546 shader->limits = &limits_array[max(0, i - 1)].limits;
547 }
548 }
549
550 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
551 {
552 DWORD idx, shift;
553 idx = bit >> 5;
554 shift = bit & 0x1f;
555 bitmap[idx] |= (1u << shift);
556 }
557
558 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
559 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
560 {
561 switch (reg->type)
562 {
563 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
564 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
565 reg_maps->texcoord |= 1u << reg->idx[0].offset;
566 else
567 reg_maps->address |= 1u << reg->idx[0].offset;
568 break;
569
570 case WINED3DSPR_TEMP:
571 reg_maps->temporary |= 1u << reg->idx[0].offset;
572 break;
573
574 case WINED3DSPR_INPUT:
575 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
576 {
577 if (reg->idx[0].rel_addr)
578 {
579 /* If relative addressing is used, we must assume that all registers
580 * are used. Even if it is a construct like v3[aL], we can't assume
581 * that v0, v1 and v2 aren't read because aL can be negative */
582 unsigned int i;
583 for (i = 0; i < MAX_REG_INPUT; ++i)
584 {
585 shader->u.ps.input_reg_used[i] = TRUE;
586 }
587 }
588 else
589 {
590 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
591 }
592 }
593 else
594 reg_maps->input_registers |= 1u << reg->idx[0].offset;
595 break;
596
597 case WINED3DSPR_RASTOUT:
598 if (reg->idx[0].offset == 1)
599 reg_maps->fog = 1;
600 if (reg->idx[0].offset == 2)
601 reg_maps->point_size = 1;
602 break;
603
604 case WINED3DSPR_MISCTYPE:
605 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
606 {
607 if (!reg->idx[0].offset)
608 reg_maps->vpos = 1;
609 else if (reg->idx[0].offset == 1)
610 reg_maps->usesfacing = 1;
611 }
612 break;
613
614 case WINED3DSPR_CONST:
615 if (reg->idx[0].rel_addr)
616 {
617 if (reg->idx[0].offset < reg_maps->min_rel_offset)
618 reg_maps->min_rel_offset = reg->idx[0].offset;
619 if (reg->idx[0].offset > reg_maps->max_rel_offset)
620 reg_maps->max_rel_offset = reg->idx[0].offset;
621 reg_maps->usesrelconstF = TRUE;
622 }
623 else
624 {
625 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
626 {
627 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
628 return FALSE;
629 }
630 else
631 {
632 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
633 }
634 }
635 break;
636
637 case WINED3DSPR_CONSTINT:
638 if (reg->idx[0].offset >= shader->limits->constant_int)
639 {
640 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
641 return FALSE;
642 }
643 else
644 {
645 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
646 }
647 break;
648
649 case WINED3DSPR_CONSTBOOL:
650 if (reg->idx[0].offset >= shader->limits->constant_bool)
651 {
652 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
653 return FALSE;
654 }
655 else
656 {
657 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
658 }
659 break;
660
661 case WINED3DSPR_COLOROUT:
662 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
663 break;
664
665 default:
666 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
667 reg->type, reg->idx[0].offset, reg->idx[1].offset);
668 break;
669 }
670 return TRUE;
671 }
672
673 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
674 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
675 {
676 struct wined3d_shader_sampler_map_entry *entries, *entry;
677 struct wined3d_shader_sampler_map *map;
678 unsigned int i;
679
680 map = &reg_maps->sampler_map;
681 entries = map->entries;
682 for (i = 0; i < map->count; ++i)
683 {
684 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
685 return;
686 }
687
688 if (!map->size)
689 {
690 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
691 {
692 ERR("Failed to allocate sampler map entries.\n");
693 return;
694 }
695 map->size = 4;
696 map->entries = entries;
697 }
698 else if (map->count == map->size)
699 {
700 size_t new_size = map->size * 2;
701
702 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
703 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
704 {
705 ERR("Failed to resize sampler map entries.\n");
706 return;
707 }
708 map->size = new_size;
709 map->entries = entries;
710 }
711
712 entry = &entries[map->count++];
713 entry->resource_idx = resource_idx;
714 entry->sampler_idx = sampler_idx;
715 entry->bind_idx = bind_idx;
716 }
717
718 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
719 {
720 switch (instr)
721 {
722 case WINED3DSIH_M4x4:
723 case WINED3DSIH_M3x4:
724 return param == 1 ? 3 : 0;
725
726 case WINED3DSIH_M4x3:
727 case WINED3DSIH_M3x3:
728 return param == 1 ? 2 : 0;
729
730 case WINED3DSIH_M3x2:
731 return param == 1 ? 1 : 0;
732
733 default:
734 return 0;
735 }
736 }
737
738 /* Note that this does not count the loop register as an address register. */
739 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
740 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
741 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
742 {
743 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
744 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
745 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
746 void *fe_data = shader->frontend_data;
747 struct wined3d_shader_version shader_version;
748 const DWORD *ptr = byte_code;
749 unsigned int i;
750
751 memset(reg_maps, 0, sizeof(*reg_maps));
752 memset(input_signature_elements, 0, sizeof(input_signature_elements));
753 memset(output_signature_elements, 0, sizeof(output_signature_elements));
754 reg_maps->min_rel_offset = ~0U;
755
756 fe->shader_read_header(fe_data, &ptr, &shader_version);
757 reg_maps->shader_version = shader_version;
758
759 shader_set_limits(shader);
760
761 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
762 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
763 if (!reg_maps->constf)
764 {
765 ERR("Failed to allocate constant map memory.\n");
766 return E_OUTOFMEMORY;
767 }
768
769 while (!fe->shader_is_end(fe_data, &ptr))
770 {
771 struct wined3d_shader_instruction ins;
772
773 /* Fetch opcode. */
774 fe->shader_read_instruction(fe_data, &ptr, &ins);
775
776 /* Unhandled opcode, and its parameters. */
777 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
778 {
779 TRACE("Skipping unrecognized instruction.\n");
780 continue;
781 }
782
783 /* Handle declarations. */
784 if (ins.handler_idx == WINED3DSIH_DCL)
785 {
786 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
787 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
788
789 switch (semantic->reg.reg.type)
790 {
791 /* Mark input registers used. */
792 case WINED3DSPR_INPUT:
793 if (reg_idx >= MAX_REG_INPUT)
794 {
795 ERR("Invalid input register index %u.\n", reg_idx);
796 break;
797 }
798 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
799 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
800 return WINED3DERR_INVALIDCALL;
801 reg_maps->input_registers |= 1u << reg_idx;
802 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
803 break;
804
805 /* Vertex shader: mark 3.0 output registers used, save token. */
806 case WINED3DSPR_OUTPUT:
807 if (reg_idx >= MAX_REG_OUTPUT)
808 {
809 ERR("Invalid output register index %u.\n", reg_idx);
810 break;
811 }
812 reg_maps->output_registers |= 1u << reg_idx;
813 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
814 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
815 reg_maps->fog = 1;
816 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
817 reg_maps->point_size = 1;
818 break;
819
820 case WINED3DSPR_SAMPLER:
821 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
822 case WINED3DSPR_RESOURCE:
823 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
824 {
825 ERR("Invalid resource index %u.\n", reg_idx);
826 break;
827 }
828 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
829 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
830 break;
831
832 default:
833 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
834 break;
835 }
836 }
837 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
838 {
839 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
840 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
841 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
842 else
843 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
844 }
845 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
846 {
847 if (reg_maps->icb)
848 FIXME("Multiple immediate constant buffers.\n");
849 reg_maps->icb = ins.declaration.icb;
850 }
851 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
852 {
853 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
854 shader->u.gs.input_type = ins.declaration.primitive_type;
855 else
856 FIXME("Invalid instruction %#x for shader type %#x.\n",
857 ins.handler_idx, shader_version.type);
858 }
859 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
860 {
861 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
862 shader->u.gs.output_type = ins.declaration.primitive_type;
863 else
864 FIXME("Invalid instruction %#x for shader type %#x.\n",
865 ins.handler_idx, shader_version.type);
866 }
867 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
868 {
869 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
870 shader->u.gs.vertices_out = ins.declaration.count;
871 else
872 FIXME("Invalid instruction %#x for shader type %#x.\n",
873 ins.handler_idx, shader_version.type);
874 }
875 else if (ins.handler_idx == WINED3DSIH_DEF)
876 {
877 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
878 float *value;
879 if (!lconst) return E_OUTOFMEMORY;
880
881 lconst->idx = ins.dst[0].reg.idx[0].offset;
882 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
883 value = (float *)lconst->value;
884
885 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
886 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
887 {
888 if (value[0] < -1.0f) value[0] = -1.0f;
889 else if (value[0] > 1.0f) value[0] = 1.0f;
890 if (value[1] < -1.0f) value[1] = -1.0f;
891 else if (value[1] > 1.0f) value[1] = 1.0f;
892 if (value[2] < -1.0f) value[2] = -1.0f;
893 else if (value[2] > 1.0f) value[2] = 1.0f;
894 if (value[3] < -1.0f) value[3] = -1.0f;
895 else if (value[3] > 1.0f) value[3] = 1.0f;
896 }
897
898 list_add_head(&shader->constantsF, &lconst->entry);
899
900 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
901 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
902 {
903 shader->lconst_inf_or_nan = TRUE;
904 }
905 }
906 else if (ins.handler_idx == WINED3DSIH_DEFI)
907 {
908 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
909 if (!lconst) return E_OUTOFMEMORY;
910
911 lconst->idx = ins.dst[0].reg.idx[0].offset;
912 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
913
914 list_add_head(&shader->constantsI, &lconst->entry);
915 reg_maps->local_int_consts |= (1u << lconst->idx);
916 }
917 else if (ins.handler_idx == WINED3DSIH_DEFB)
918 {
919 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
920 if (!lconst) return E_OUTOFMEMORY;
921
922 lconst->idx = ins.dst[0].reg.idx[0].offset;
923 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
924
925 list_add_head(&shader->constantsB, &lconst->entry);
926 reg_maps->local_bool_consts |= (1u << lconst->idx);
927 }
928 /* For subroutine prototypes. */
929 else if (ins.handler_idx == WINED3DSIH_LABEL)
930 {
931 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
932 }
933 /* Set texture, address, temporary registers. */
934 else
935 {
936 BOOL color0_mov = FALSE;
937 unsigned int i;
938
939 /* This will loop over all the registers and try to
940 * make a bitmask of the ones we're interested in.
941 *
942 * Relative addressing tokens are ignored, but that's
943 * okay, since we'll catch any address registers when
944 * they are initialized (required by spec). */
945 for (i = 0; i < ins.dst_count; ++i)
946 {
947 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
948 shader_version.type, constf_size))
949 return WINED3DERR_INVALIDCALL;
950
951 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
952 {
953 UINT idx = ins.dst[i].reg.idx[0].offset;
954
955 switch (ins.dst[i].reg.type)
956 {
957 case WINED3DSPR_RASTOUT:
958 if (shader_version.major >= 3)
959 break;
960 switch (idx)
961 {
962 case 0: /* oPos */
963 reg_maps->output_registers |= 1u << 10;
964 shader_signature_from_usage(&output_signature_elements[10],
965 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
966 break;
967
968 case 1: /* oFog */
969 reg_maps->output_registers |= 1u << 11;
970 shader_signature_from_usage(&output_signature_elements[11],
971 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
972 break;
973
974 case 2: /* oPts */
975 reg_maps->output_registers |= 1u << 11;
976 shader_signature_from_usage(&output_signature_elements[11],
977 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
978 break;
979 }
980 break;
981
982 case WINED3DSPR_ATTROUT:
983 if (shader_version.major >= 3)
984 break;
985 if (idx < 2)
986 {
987 idx += 8;
988 if (reg_maps->output_registers & (1u << idx))
989 {
990 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
991 }
992 else
993 {
994 reg_maps->output_registers |= 1u << idx;
995 shader_signature_from_usage(&output_signature_elements[idx],
996 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
997 }
998 }
999 break;
1000
1001 case WINED3DSPR_TEXCRDOUT:
1002 if (shader_version.major >= 3)
1003 {
1004 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1005 break;
1006 }
1007 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1008 if (reg_maps->output_registers & (1u << idx))
1009 {
1010 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1011 }
1012 else
1013 {
1014 reg_maps->output_registers |= 1u << idx;
1015 shader_signature_from_usage(&output_signature_elements[idx],
1016 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1017 }
1018 break;
1019
1020 default:
1021 break;
1022 }
1023 }
1024
1025 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1026 {
1027 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1028 {
1029 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1030 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1031 * the mov and perform the sRGB write correction from the source register.
1032 *
1033 * However, if the mov is only partial, we can't do this, and if the write
1034 * comes from an instruction other than MOV it is hard to do as well. If
1035 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1036 shader->u.ps.color0_mov = FALSE;
1037 if (ins.handler_idx == WINED3DSIH_MOV
1038 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1039 {
1040 /* Used later when the source register is read. */
1041 color0_mov = TRUE;
1042 }
1043 }
1044 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1045 * end
1046 */
1047 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1048 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1049 {
1050 shader->u.ps.color0_mov = FALSE;
1051 }
1052 }
1053
1054 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1055 if (shader_version.major == 1
1056 && (ins.handler_idx == WINED3DSIH_TEX
1057 || ins.handler_idx == WINED3DSIH_TEXBEM
1058 || ins.handler_idx == WINED3DSIH_TEXBEML
1059 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1060 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1061 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1062 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1063 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1064 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1065 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1066 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1067 {
1068 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1069
1070 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1071 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1072 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1073 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1074
1075 /* texbem is only valid with < 1.4 pixel shaders */
1076 if (ins.handler_idx == WINED3DSIH_TEXBEM
1077 || ins.handler_idx == WINED3DSIH_TEXBEML)
1078 {
1079 reg_maps->bumpmat |= 1u << reg_idx;
1080 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1081 {
1082 reg_maps->luminanceparams |= 1u << reg_idx;
1083 }
1084 }
1085 }
1086 else if (ins.handler_idx == WINED3DSIH_BEM)
1087 {
1088 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1089 }
1090 }
1091
1092 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1093 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1094 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1095 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1096 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1097 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1098 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1099 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1100 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1101 else if (ins.handler_idx == WINED3DSIH_LOOP
1102 || ins.handler_idx == WINED3DSIH_REP)
1103 {
1104 ++cur_loop_depth;
1105 if (cur_loop_depth > max_loop_depth)
1106 max_loop_depth = cur_loop_depth;
1107 }
1108 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1109 || ins.handler_idx == WINED3DSIH_ENDREP)
1110 {
1111 --cur_loop_depth;
1112 }
1113 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1114 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1115 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1116 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1117 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1118 {
1119 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1120 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1121 }
1122 else if (ins.handler_idx == WINED3DSIH_LD
1123 || ins.handler_idx == WINED3DSIH_RESINFO)
1124 {
1125 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1126 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1127 }
1128
1129 if (ins.predicate)
1130 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1131 shader_version.type, constf_size))
1132 return WINED3DERR_INVALIDCALL;
1133
1134 for (i = 0; i < ins.src_count; ++i)
1135 {
1136 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1137 struct wined3d_shader_register reg = ins.src[i].reg;
1138
1139 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1140 shader_version.type, constf_size))
1141 return WINED3DERR_INVALIDCALL;
1142 while (count)
1143 {
1144 ++reg.idx[0].offset;
1145 if (!shader_record_register_usage(shader, reg_maps, &reg,
1146 shader_version.type, constf_size))
1147 return WINED3DERR_INVALIDCALL;
1148 --count;
1149 }
1150
1151 if (color0_mov)
1152 {
1153 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1154 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1155 {
1156 shader->u.ps.color0_mov = TRUE;
1157 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1158 }
1159 }
1160 }
1161 }
1162 }
1163 reg_maps->loop_depth = max_loop_depth;
1164
1165 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1166 * R0 is written to the render target. */
1167 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1168 reg_maps->rt_mask |= (1u << 0);
1169
1170 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1171
1172 if (input_signature->elements)
1173 {
1174 for (i = 0; i < input_signature->element_count; ++i)
1175 {
1176 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1177 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1178 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1179 reg_maps->vpos = 1;
1180 }
1181 }
1182 else if (!input_signature->elements && reg_maps->input_registers)
1183 {
1184 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1185 struct wined3d_shader_signature_element *e;
1186 unsigned int i;
1187
1188 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1189 return E_OUTOFMEMORY;
1190 input_signature->element_count = count;
1191
1192 e = input_signature->elements;
1193 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1194 {
1195 if (!(reg_maps->input_registers & (1u << i)))
1196 continue;
1197 input_signature_elements[i].register_idx = i;
1198 *e++ = input_signature_elements[i];
1199 }
1200 }
1201
1202 if (output_signature->elements)
1203 {
1204 for (i = 0; i < output_signature->element_count; ++i)
1205 {
1206 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1207 }
1208 }
1209 else if (reg_maps->output_registers)
1210 {
1211 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1212 struct wined3d_shader_signature_element *e;
1213
1214 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1215 return E_OUTOFMEMORY;
1216 output_signature->element_count = count;
1217
1218 e = output_signature->elements;
1219 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1220 {
1221 if (!(reg_maps->output_registers & (1u << i)))
1222 continue;
1223 *e++ = output_signature_elements[i];
1224 }
1225 }
1226
1227 return WINED3D_OK;
1228 }
1229
1230 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1231 {
1232 DWORD map = 1u << max;
1233 map |= map - 1;
1234 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1235
1236 return wined3d_log2i(map);
1237 }
1238
1239 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1240 {
1241 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1242 {
1243 shader_addline(buffer, "refactoringAllowed");
1244 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1245 if (global_flags)
1246 shader_addline(buffer, " | ");
1247 }
1248
1249 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1250 {
1251 shader_addline(buffer, "enableRawAndStructuredBuffers");
1252 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1253 }
1254
1255 if (global_flags)
1256 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1257 }
1258
1259 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1260 {
1261 unsigned int i;
1262
1263 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1264 {
1265 if (sysval_semantic_names[i].sysval_semantic == semantic)
1266 {
1267 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1268 return;
1269 }
1270 }
1271
1272 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1273 }
1274
1275 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1276 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1277 {
1278 shader_addline(buffer, "dcl");
1279
1280 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1281 {
1282 switch (semantic->resource_type)
1283 {
1284 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1285 shader_addline(buffer, "_2d");
1286 break;
1287
1288 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1289 shader_addline(buffer, "_3d");
1290 break;
1291
1292 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1293 shader_addline(buffer, "_cube");
1294 break;
1295
1296 default:
1297 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1298 break;
1299 }
1300 }
1301 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1302 {
1303 shader_addline(buffer, "_resource_");
1304 switch (semantic->resource_type)
1305 {
1306 case WINED3D_SHADER_RESOURCE_BUFFER:
1307 shader_addline(buffer, "buffer");
1308 break;
1309
1310 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1311 shader_addline(buffer, "texture1d");
1312 break;
1313
1314 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1315 shader_addline(buffer, "texture2d");
1316 break;
1317
1318 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1319 shader_addline(buffer, "texture2dms");
1320 break;
1321
1322 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1323 shader_addline(buffer, "texture3d");
1324 break;
1325
1326 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1327 shader_addline(buffer, "texturecube");
1328 break;
1329
1330 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1331 shader_addline(buffer, "texture1darray");
1332 break;
1333
1334 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1335 shader_addline(buffer, "texture2darray");
1336 break;
1337
1338 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1339 shader_addline(buffer, "texture2dmsarray");
1340 break;
1341
1342 default:
1343 shader_addline(buffer, "unknown");
1344 break;
1345 }
1346 switch (semantic->resource_data_type)
1347 {
1348 case WINED3D_DATA_FLOAT:
1349 shader_addline(buffer, " (float)");
1350 break;
1351
1352 case WINED3D_DATA_INT:
1353 shader_addline(buffer, " (int)");
1354 break;
1355
1356 case WINED3D_DATA_UINT:
1357 shader_addline(buffer, " (uint)");
1358 break;
1359
1360 case WINED3D_DATA_UNORM:
1361 shader_addline(buffer, " (unorm)");
1362 break;
1363
1364 case WINED3D_DATA_SNORM:
1365 shader_addline(buffer, " (snorm)");
1366 break;
1367
1368 default:
1369 shader_addline(buffer, " (unknown)");
1370 break;
1371 }
1372 }
1373 else
1374 {
1375 /* Pixel shaders 3.0 don't have usage semantics. */
1376 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1377 return;
1378 else
1379 shader_addline(buffer, "_");
1380
1381 switch (semantic->usage)
1382 {
1383 case WINED3D_DECL_USAGE_POSITION:
1384 shader_addline(buffer, "position%u", semantic->usage_idx);
1385 break;
1386
1387 case WINED3D_DECL_USAGE_BLEND_INDICES:
1388 shader_addline(buffer, "blend");
1389 break;
1390
1391 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1392 shader_addline(buffer, "weight");
1393 break;
1394
1395 case WINED3D_DECL_USAGE_NORMAL:
1396 shader_addline(buffer, "normal%u", semantic->usage_idx);
1397 break;
1398
1399 case WINED3D_DECL_USAGE_PSIZE:
1400 shader_addline(buffer, "psize");
1401 break;
1402
1403 case WINED3D_DECL_USAGE_COLOR:
1404 if (!semantic->usage_idx)
1405 shader_addline(buffer, "color");
1406 else
1407 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1408 break;
1409
1410 case WINED3D_DECL_USAGE_TEXCOORD:
1411 shader_addline(buffer, "texture%u", semantic->usage_idx);
1412 break;
1413
1414 case WINED3D_DECL_USAGE_TANGENT:
1415 shader_addline(buffer, "tangent");
1416 break;
1417
1418 case WINED3D_DECL_USAGE_BINORMAL:
1419 shader_addline(buffer, "binormal");
1420 break;
1421
1422 case WINED3D_DECL_USAGE_TESS_FACTOR:
1423 shader_addline(buffer, "tessfactor");
1424 break;
1425
1426 case WINED3D_DECL_USAGE_POSITIONT:
1427 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1428 break;
1429
1430 case WINED3D_DECL_USAGE_FOG:
1431 shader_addline(buffer, "fog");
1432 break;
1433
1434 case WINED3D_DECL_USAGE_DEPTH:
1435 shader_addline(buffer, "depth");
1436 break;
1437
1438 case WINED3D_DECL_USAGE_SAMPLE:
1439 shader_addline(buffer, "sample");
1440 break;
1441
1442 default:
1443 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1444 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1445 }
1446 }
1447 }
1448
1449 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1450 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1451 {
1452 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1453 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1454 UINT offset = reg->idx[0].offset;
1455
1456 switch (reg->type)
1457 {
1458 case WINED3DSPR_TEMP:
1459 shader_addline(buffer, "r");
1460 break;
1461
1462 case WINED3DSPR_INPUT:
1463 shader_addline(buffer, "v");
1464 break;
1465
1466 case WINED3DSPR_CONST:
1467 case WINED3DSPR_CONST2:
1468 case WINED3DSPR_CONST3:
1469 case WINED3DSPR_CONST4:
1470 shader_addline(buffer, "c");
1471 offset = shader_get_float_offset(reg->type, offset);
1472 break;
1473
1474 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1475 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1476 break;
1477
1478 case WINED3DSPR_RASTOUT:
1479 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1480 break;
1481
1482 case WINED3DSPR_COLOROUT:
1483 shader_addline(buffer, "oC");
1484 break;
1485
1486 case WINED3DSPR_DEPTHOUT:
1487 shader_addline(buffer, "oDepth");
1488 break;
1489
1490 case WINED3DSPR_ATTROUT:
1491 shader_addline(buffer, "oD");
1492 break;
1493
1494 case WINED3DSPR_TEXCRDOUT:
1495 /* Vertex shaders >= 3.0 use general purpose output registers
1496 * (WINED3DSPR_OUTPUT), which can include an address token. */
1497 if (shader_version->major >= 3)
1498 shader_addline(buffer, "o");
1499 else
1500 shader_addline(buffer, "oT");
1501 break;
1502
1503 case WINED3DSPR_CONSTINT:
1504 shader_addline(buffer, "i");
1505 break;
1506
1507 case WINED3DSPR_CONSTBOOL:
1508 shader_addline(buffer, "b");
1509 break;
1510
1511 case WINED3DSPR_LABEL:
1512 shader_addline(buffer, "l");
1513 break;
1514
1515 case WINED3DSPR_LOOP:
1516 shader_addline(buffer, "aL");
1517 break;
1518
1519 case WINED3DSPR_SAMPLER:
1520 shader_addline(buffer, "s");
1521 break;
1522
1523 case WINED3DSPR_MISCTYPE:
1524 if (offset > 1)
1525 {
1526 FIXME("Unhandled misctype register %u.\n", offset);
1527 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1528 }
1529 else
1530 {
1531 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1532 }
1533 break;
1534
1535 case WINED3DSPR_PREDICATE:
1536 shader_addline(buffer, "p");
1537 break;
1538
1539 case WINED3DSPR_IMMCONST:
1540 shader_addline(buffer, "l");
1541 break;
1542
1543 case WINED3DSPR_CONSTBUFFER:
1544 shader_addline(buffer, "cb");
1545 break;
1546
1547 case WINED3DSPR_IMMCONSTBUFFER:
1548 shader_addline(buffer, "icb");
1549 break;
1550
1551 case WINED3DSPR_PRIMID:
1552 shader_addline(buffer, "primID");
1553 break;
1554
1555 case WINED3DSPR_NULL:
1556 shader_addline(buffer, "null");
1557 break;
1558
1559 case WINED3DSPR_RESOURCE:
1560 shader_addline(buffer, "t");
1561 break;
1562
1563 default:
1564 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1565 break;
1566 }
1567
1568 if (reg->type == WINED3DSPR_IMMCONST)
1569 {
1570 shader_addline(buffer, "(");
1571 switch (reg->immconst_type)
1572 {
1573 case WINED3D_IMMCONST_SCALAR:
1574 switch (reg->data_type)
1575 {
1576 case WINED3D_DATA_FLOAT:
1577 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1578 break;
1579 case WINED3D_DATA_INT:
1580 shader_addline(buffer, "%d", reg->immconst_data[0]);
1581 break;
1582 case WINED3D_DATA_RESOURCE:
1583 case WINED3D_DATA_SAMPLER:
1584 case WINED3D_DATA_UINT:
1585 shader_addline(buffer, "%u", reg->immconst_data[0]);
1586 break;
1587 default:
1588 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1589 break;
1590 }
1591 break;
1592
1593 case WINED3D_IMMCONST_VEC4:
1594 switch (reg->data_type)
1595 {
1596 case WINED3D_DATA_FLOAT:
1597 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1598 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1599 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1600 break;
1601 case WINED3D_DATA_INT:
1602 shader_addline(buffer, "%d, %d, %d, %d",
1603 reg->immconst_data[0], reg->immconst_data[1],
1604 reg->immconst_data[2], reg->immconst_data[3]);
1605 break;
1606 case WINED3D_DATA_RESOURCE:
1607 case WINED3D_DATA_SAMPLER:
1608 case WINED3D_DATA_UINT:
1609 shader_addline(buffer, "%u, %u, %u, %u",
1610 reg->immconst_data[0], reg->immconst_data[1],
1611 reg->immconst_data[2], reg->immconst_data[3]);
1612 break;
1613 default:
1614 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1615 break;
1616 }
1617 break;
1618
1619 default:
1620 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1621 break;
1622 }
1623 shader_addline(buffer, ")");
1624 }
1625 else if (reg->type != WINED3DSPR_RASTOUT
1626 && reg->type != WINED3DSPR_MISCTYPE
1627 && reg->type != WINED3DSPR_NULL)
1628 {
1629 if (offset != ~0u)
1630 {
1631 shader_addline(buffer, "[");
1632 if (reg->idx[0].rel_addr)
1633 {
1634 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1635 shader_addline(buffer, " + ");
1636 }
1637 shader_addline(buffer, "%u]", offset);
1638
1639 if (reg->idx[1].offset != ~0u)
1640 {
1641 shader_addline(buffer, "[");
1642 if (reg->idx[1].rel_addr)
1643 {
1644 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1645 shader_addline(buffer, " + ");
1646 }
1647 shader_addline(buffer, "%u]", reg->idx[1].offset);
1648 }
1649 }
1650 }
1651 }
1652
1653 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1654 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1655 {
1656 DWORD write_mask = param->write_mask;
1657
1658 shader_dump_register(buffer, &param->reg, shader_version);
1659
1660 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1661 {
1662 static const char write_mask_chars[] = "xyzw";
1663
1664 shader_addline(buffer, ".");
1665 if (write_mask & WINED3DSP_WRITEMASK_0)
1666 shader_addline(buffer, "%c", write_mask_chars[0]);
1667 if (write_mask & WINED3DSP_WRITEMASK_1)
1668 shader_addline(buffer, "%c", write_mask_chars[1]);
1669 if (write_mask & WINED3DSP_WRITEMASK_2)
1670 shader_addline(buffer, "%c", write_mask_chars[2]);
1671 if (write_mask & WINED3DSP_WRITEMASK_3)
1672 shader_addline(buffer, "%c", write_mask_chars[3]);
1673 }
1674 }
1675
1676 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1677 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1678 {
1679 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1680 DWORD swizzle = param->swizzle;
1681
1682 if (src_modifier == WINED3DSPSM_NEG
1683 || src_modifier == WINED3DSPSM_BIASNEG
1684 || src_modifier == WINED3DSPSM_SIGNNEG
1685 || src_modifier == WINED3DSPSM_X2NEG
1686 || src_modifier == WINED3DSPSM_ABSNEG)
1687 shader_addline(buffer, "-");
1688 else if (src_modifier == WINED3DSPSM_COMP)
1689 shader_addline(buffer, "1-");
1690 else if (src_modifier == WINED3DSPSM_NOT)
1691 shader_addline(buffer, "!");
1692
1693 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1694 shader_addline(buffer, "abs(");
1695
1696 shader_dump_register(buffer, &param->reg, shader_version);
1697
1698 switch (src_modifier)
1699 {
1700 case WINED3DSPSM_NONE: break;
1701 case WINED3DSPSM_NEG: break;
1702 case WINED3DSPSM_NOT: break;
1703 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1704 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1705 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1706 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1707 case WINED3DSPSM_COMP: break;
1708 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1709 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1710 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1711 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1712 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1713 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1714 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1715 }
1716
1717 if (swizzle != WINED3DSP_NOSWIZZLE)
1718 {
1719 static const char swizzle_chars[] = "xyzw";
1720 DWORD swizzle_x = swizzle & 0x03;
1721 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1722 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1723 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1724
1725 if (swizzle_x == swizzle_y
1726 && swizzle_x == swizzle_z
1727 && swizzle_x == swizzle_w)
1728 {
1729 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
1730 }
1731 else
1732 {
1733 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1734 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1735 }
1736 }
1737 }
1738
1739 /* Shared code in order to generate the bulk of the shader string.
1740 * NOTE: A description of how to parse tokens can be found on MSDN. */
1741 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1742 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1743 {
1744 struct wined3d_device *device = shader->device;
1745 const struct wined3d_shader_frontend *fe = shader->frontend;
1746 void *fe_data = shader->frontend_data;
1747 struct wined3d_shader_version shader_version;
1748 struct wined3d_shader_loop_state loop_state;
1749 struct wined3d_shader_instruction ins;
1750 struct wined3d_shader_tex_mx tex_mx;
1751 struct wined3d_shader_context ctx;
1752 const DWORD *ptr = byte_code;
1753
1754 /* Initialize current parsing state. */
1755 tex_mx.current_row = 0;
1756 loop_state.current_depth = 0;
1757 loop_state.current_reg = 0;
1758
1759 ctx.shader = shader;
1760 ctx.gl_info = &device->adapter->gl_info;
1761 ctx.reg_maps = reg_maps;
1762 ctx.buffer = buffer;
1763 ctx.tex_mx = &tex_mx;
1764 ctx.loop_state = &loop_state;
1765 ctx.backend_data = backend_ctx;
1766 ins.ctx = &ctx;
1767
1768 fe->shader_read_header(fe_data, &ptr, &shader_version);
1769
1770 while (!fe->shader_is_end(fe_data, &ptr))
1771 {
1772 /* Read opcode. */
1773 fe->shader_read_instruction(fe_data, &ptr, &ins);
1774
1775 /* Unknown opcode and its parameters. */
1776 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1777 {
1778 TRACE("Skipping unrecognized instruction.\n");
1779 continue;
1780 }
1781
1782 if (ins.predicate)
1783 FIXME("Predicates not implemented.\n");
1784
1785 /* Call appropriate function for output target */
1786 device->shader_backend->shader_handle_instruction(&ins);
1787 }
1788 }
1789
1790 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
1791 const struct wined3d_shader_dst_param *dst)
1792 {
1793 DWORD mmask = dst->modifiers;
1794
1795 switch (dst->shift)
1796 {
1797 case 0: break;
1798 case 13: shader_addline(buffer, "_d8"); break;
1799 case 14: shader_addline(buffer, "_d4"); break;
1800 case 15: shader_addline(buffer, "_d2"); break;
1801 case 1: shader_addline(buffer, "_x2"); break;
1802 case 2: shader_addline(buffer, "_x4"); break;
1803 case 3: shader_addline(buffer, "_x8"); break;
1804 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
1805 }
1806
1807 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
1808 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
1809 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
1810
1811 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1812 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1813 }
1814
1815 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
1816 enum wined3d_primitive_type primitive_type)
1817 {
1818 switch (primitive_type)
1819 {
1820 case WINED3D_PT_UNDEFINED:
1821 shader_addline(buffer, "undefined");
1822 break;
1823 case WINED3D_PT_POINTLIST:
1824 shader_addline(buffer, "pointlist");
1825 break;
1826 case WINED3D_PT_LINELIST:
1827 shader_addline(buffer, "linelist");
1828 break;
1829 case WINED3D_PT_LINESTRIP:
1830 shader_addline(buffer, "linestrip");
1831 break;
1832 case WINED3D_PT_TRIANGLELIST:
1833 shader_addline(buffer, "trianglelist");
1834 break;
1835 case WINED3D_PT_TRIANGLESTRIP:
1836 shader_addline(buffer, "trianglestrip");
1837 break;
1838 case WINED3D_PT_TRIANGLEFAN:
1839 shader_addline(buffer, "trianglefan");
1840 break;
1841 case WINED3D_PT_LINELIST_ADJ:
1842 shader_addline(buffer, "linelist_adj");
1843 break;
1844 case WINED3D_PT_LINESTRIP_ADJ:
1845 shader_addline(buffer, "linestrip_adj");
1846 break;
1847 case WINED3D_PT_TRIANGLELIST_ADJ:
1848 shader_addline(buffer, "trianglelist_adj");
1849 break;
1850 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1851 shader_addline(buffer, "trianglestrip_adj");
1852 break;
1853 default:
1854 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
1855 break;
1856 }
1857 }
1858
1859 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
1860 enum wined3d_shader_interpolation_mode interpolation_mode)
1861 {
1862 switch (interpolation_mode)
1863 {
1864 case WINED3DSIM_CONSTANT:
1865 shader_addline(buffer, "constant");
1866 break;
1867 case WINED3DSIM_LINEAR:
1868 shader_addline(buffer, "linear");
1869 break;
1870 case WINED3DSIM_LINEAR_CENTROID:
1871 shader_addline(buffer, "linear centroid");
1872 break;
1873 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
1874 shader_addline(buffer, "linear noperspective");
1875 break;
1876 case WINED3DSIM_LINEAR_SAMPLE:
1877 shader_addline(buffer, "linear sample");
1878 break;
1879 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
1880 shader_addline(buffer, "linear noperspective centroid");
1881 break;
1882 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
1883 shader_addline(buffer, "linear noperspective sample");
1884 break;
1885 default:
1886 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
1887 break;
1888 }
1889 }
1890
1891 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1892 {
1893 struct wined3d_shader_version shader_version;
1894 struct wined3d_string_buffer buffer;
1895 const DWORD *ptr = byte_code;
1896 const char *type_prefix;
1897 const char *p, *q;
1898 DWORD i;
1899
1900 if (!string_buffer_init(&buffer))
1901 {
1902 ERR("Failed to initialize string buffer.\n");
1903 return;
1904 }
1905
1906 TRACE("Parsing %p.\n", byte_code);
1907
1908 fe->shader_read_header(fe_data, &ptr, &shader_version);
1909
1910 switch (shader_version.type)
1911 {
1912 case WINED3D_SHADER_TYPE_VERTEX:
1913 type_prefix = "vs";
1914 break;
1915
1916 case WINED3D_SHADER_TYPE_GEOMETRY:
1917 type_prefix = "gs";
1918 break;
1919
1920 case WINED3D_SHADER_TYPE_PIXEL:
1921 type_prefix = "ps";
1922 break;
1923
1924 default:
1925 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1926 type_prefix = "unknown";
1927 break;
1928 }
1929
1930 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1931
1932 while (!fe->shader_is_end(fe_data, &ptr))
1933 {
1934 struct wined3d_shader_instruction ins;
1935
1936 fe->shader_read_instruction(fe_data, &ptr, &ins);
1937 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1938 {
1939 WARN("Skipping unrecognized instruction.\n");
1940 shader_addline(&buffer, "<unrecognized instruction>\n");
1941 continue;
1942 }
1943
1944 if (ins.handler_idx == WINED3DSIH_DCL)
1945 {
1946 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
1947 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
1948 shader_addline(&buffer, " ");
1949 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
1950 }
1951 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1952 {
1953 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1954 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
1955 shader_addline(&buffer, ", %s",
1956 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1957 }
1958 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1959 {
1960 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1961 shader_dump_global_flags(&buffer, ins.flags);
1962 }
1963 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1964 {
1965 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
1966 for (i = 0; i < ins.declaration.icb->element_count / 4; ++i)
1967 {
1968 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
1969 ins.declaration.icb->data[4 * i + 0],
1970 ins.declaration.icb->data[4 * i + 1],
1971 ins.declaration.icb->data[4 * i + 2],
1972 ins.declaration.icb->data[4 * i + 3]);
1973 }
1974 shader_addline(&buffer, "}");
1975 }
1976 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1977 {
1978 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1979 shader_dump_interpolation_mode(&buffer, ins.flags);
1980 shader_addline(&buffer, " ");
1981 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
1982 }
1983 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
1984 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
1985 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
1986 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
1987 {
1988 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1989 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
1990 shader_addline(&buffer, ", ");
1991 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
1992 }
1993 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
1994 {
1995 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1996 shader_dump_interpolation_mode(&buffer, ins.flags);
1997 shader_addline(&buffer, " ");
1998 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
1999 shader_addline(&buffer, ", ");
2000 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2001 }
2002 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2003 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2004 {
2005 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2006 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2007 }
2008 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2009 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2010 {
2011 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2012 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2013 }
2014 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2015 {
2016 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2017 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2018 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2019 shader_addline(&buffer, ", comparisonMode");
2020 }
2021 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2022 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2023 {
2024 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2025 }
2026 else if (ins.handler_idx == WINED3DSIH_DEF)
2027 {
2028 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2029 ins.dst[0].reg.idx[0].offset),
2030 *(const float *)&ins.src[0].reg.immconst_data[0],
2031 *(const float *)&ins.src[0].reg.immconst_data[1],
2032 *(const float *)&ins.src[0].reg.immconst_data[2],
2033 *(const float *)&ins.src[0].reg.immconst_data[3]);
2034 }
2035 else if (ins.handler_idx == WINED3DSIH_DEFI)
2036 {
2037 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2038 ins.src[0].reg.immconst_data[0],
2039 ins.src[0].reg.immconst_data[1],
2040 ins.src[0].reg.immconst_data[2],
2041 ins.src[0].reg.immconst_data[3]);
2042 }
2043 else if (ins.handler_idx == WINED3DSIH_DEFB)
2044 {
2045 shader_addline(&buffer, "defb b%u = %s",
2046 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2047 }
2048 else
2049 {
2050 if (ins.predicate)
2051 {
2052 shader_addline(&buffer, "(");
2053 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2054 shader_addline(&buffer, ") ");
2055 }
2056
2057 /* PixWin marks instructions with the coissue flag with a '+' */
2058 if (ins.coissue)
2059 shader_addline(&buffer, "+");
2060
2061 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2062
2063 if (ins.handler_idx == WINED3DSIH_IFC
2064 || ins.handler_idx == WINED3DSIH_BREAKC)
2065 {
2066 switch (ins.flags)
2067 {
2068 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2069 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2070 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2071 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2072 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2073 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2074 default: shader_addline(&buffer, "_(%u)", ins.flags);
2075 }
2076 }
2077 else if (ins.handler_idx == WINED3DSIH_TEX
2078 && shader_version.major >= 2
2079 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2080 {
2081 shader_addline(&buffer, "p");
2082 }
2083 else if (ins.handler_idx == WINED3DSIH_RESINFO
2084 && ins.flags)
2085 {
2086 switch (ins.flags)
2087 {
2088 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2089 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2090 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2091 }
2092 }
2093
2094 for (i = 0; i < ins.dst_count; ++i)
2095 {
2096 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2097 shader_addline(&buffer, !i ? " " : ", ");
2098 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2099 }
2100
2101 /* Other source tokens */
2102 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2103 {
2104 shader_addline(&buffer, !i ? " " : ", ");
2105 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2106 }
2107 }
2108 shader_addline(&buffer, "\n");
2109 }
2110
2111 for (p = buffer.buffer; *p; p = q)
2112 {
2113 if (!(q = strstr(p, "\n")))
2114 q = p + strlen(p);
2115 else
2116 ++q;
2117 TRACE(" %.*s", (int)(q - p), p);
2118 }
2119
2120 string_buffer_free(&buffer);
2121 }
2122
2123 static void shader_cleanup(struct wined3d_shader *shader)
2124 {
2125 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2126 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2127 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2128 shader->device->shader_backend->shader_destroy(shader);
2129 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
2130 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
2131 HeapFree(GetProcessHeap(), 0, shader->function);
2132 shader_delete_constant_list(&shader->constantsF);
2133 shader_delete_constant_list(&shader->constantsB);
2134 shader_delete_constant_list(&shader->constantsI);
2135 list_remove(&shader->shader_list_entry);
2136
2137 if (shader->frontend && shader->frontend_data)
2138 shader->frontend->shader_free(shader->frontend_data);
2139 }
2140
2141 struct shader_none_priv
2142 {
2143 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2144 const struct fragment_pipeline *fragment_pipe;
2145 BOOL ffp_proj_control;
2146 };
2147
2148 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2149 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2150 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2151 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2152 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2153 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2154 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2155 const struct wined3d_state *state) {}
2156 static void shader_none_destroy(struct wined3d_shader *shader) {}
2157 static void shader_none_free_context_data(struct wined3d_context *context) {}
2158 static void shader_none_init_context_state(struct wined3d_context *context) {}
2159
2160 /* Context activation is done by the caller. */
2161 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2162 const struct wined3d_state *state)
2163 {
2164 const struct wined3d_gl_info *gl_info = context->gl_info;
2165 struct shader_none_priv *priv = shader_priv;
2166
2167 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2168 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2169 }
2170
2171 /* Context activation is done by the caller. */
2172 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2173 {
2174 struct shader_none_priv *priv = shader_priv;
2175 const struct wined3d_gl_info *gl_info = context->gl_info;
2176
2177 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2178 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2179
2180 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2181 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2182 | (1u << WINED3D_SHADER_TYPE_GEOMETRY);
2183 }
2184
2185 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2186 const struct fragment_pipeline *fragment_pipe)
2187 {
2188 struct fragment_caps fragment_caps;
2189 void *vertex_priv, *fragment_priv;
2190 struct shader_none_priv *priv;
2191
2192 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2193 return E_OUTOFMEMORY;
2194
2195 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2196 {
2197 ERR("Failed to initialize vertex pipe.\n");
2198 HeapFree(GetProcessHeap(), 0, priv);
2199 return E_FAIL;
2200 }
2201
2202 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2203 {
2204 ERR("Failed to initialize fragment pipe.\n");
2205 vertex_pipe->vp_free(device);
2206 HeapFree(GetProcessHeap(), 0, priv);
2207 return E_FAIL;
2208 }
2209
2210 priv->vertex_pipe = vertex_pipe;
2211 priv->fragment_pipe = fragment_pipe;
2212 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2213 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2214
2215 device->vertex_priv = vertex_priv;
2216 device->fragment_priv = fragment_priv;
2217 device->shader_priv = priv;
2218
2219 return WINED3D_OK;
2220 }
2221
2222 static void shader_none_free(struct wined3d_device *device)
2223 {
2224 struct shader_none_priv *priv = device->shader_priv;
2225
2226 priv->fragment_pipe->free_private(device);
2227 priv->vertex_pipe->vp_free(device);
2228 HeapFree(GetProcessHeap(), 0, priv);
2229 }
2230
2231 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2232 {
2233 return TRUE;
2234 }
2235
2236 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2237 {
2238 /* Set the shader caps to 0 for the none shader backend */
2239 caps->vs_version = 0;
2240 caps->gs_version = 0;
2241 caps->ps_version = 0;
2242 caps->vs_uniform_count = 0;
2243 caps->ps_uniform_count = 0;
2244 caps->ps_1x_max_value = 0.0f;
2245 caps->varying_count = 0;
2246 caps->wined3d_caps = 0;
2247 }
2248
2249 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2250 {
2251 /* We "support" every possible fixup, since we don't support any shader
2252 * model, and will never have to actually sample a texture. */
2253 return TRUE;
2254 }
2255
2256 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2257 {
2258 struct shader_none_priv *priv = shader_priv;
2259
2260 return priv->ffp_proj_control;
2261 }
2262
2263 const struct wined3d_shader_backend_ops none_shader_backend =
2264 {
2265 shader_none_handle_instruction,
2266 shader_none_select,
2267 shader_none_disable,
2268 shader_none_select_depth_blt,
2269 shader_none_deselect_depth_blt,
2270 shader_none_update_float_vertex_constants,
2271 shader_none_update_float_pixel_constants,
2272 shader_none_load_constants,
2273 shader_none_destroy,
2274 shader_none_alloc,
2275 shader_none_free,
2276 shader_none_allocate_context_data,
2277 shader_none_free_context_data,
2278 shader_none_init_context_state,
2279 shader_none_get_caps,
2280 shader_none_color_fixup_supported,
2281 shader_none_has_ffp_proj_control,
2282 };
2283
2284 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2285 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2286 enum wined3d_shader_type type, unsigned int max_version)
2287 {
2288 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2289 const struct wined3d_shader_frontend *fe;
2290 HRESULT hr;
2291 unsigned int backend_version;
2292 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2293
2294 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u, type %#x, max_version %u.\n",
2295 shader, byte_code, output_signature, float_const_count, type, max_version);
2296
2297 list_init(&shader->constantsF);
2298 list_init(&shader->constantsB);
2299 list_init(&shader->constantsI);
2300 shader->lconst_inf_or_nan = FALSE;
2301
2302 fe = shader_select_frontend(*byte_code);
2303 if (!fe)
2304 {
2305 FIXME("Unable to find frontend for shader.\n");
2306 return WINED3DERR_INVALIDCALL;
2307 }
2308 shader->frontend = fe;
2309 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2310 if (!shader->frontend_data)
2311 {
2312 FIXME("Failed to initialize frontend.\n");
2313 return WINED3DERR_INVALIDCALL;
2314 }
2315
2316 /* First pass: trace shader. */
2317 if (TRACE_ON(d3d_shader))
2318 shader_trace_init(fe, shader->frontend_data, byte_code);
2319
2320 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2321 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2322 &shader->output_signature, byte_code, float_const_count)))
2323 return hr;
2324
2325 if (reg_maps->shader_version.type != type)
2326 {
2327 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2328 return WINED3DERR_INVALIDCALL;
2329 }
2330 if (reg_maps->shader_version.major > max_version)
2331 {
2332 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2333 return WINED3DERR_INVALIDCALL;
2334 }
2335 switch (type)
2336 {
2337 case WINED3D_SHADER_TYPE_VERTEX:
2338 backend_version = d3d_info->limits.vs_version;
2339 break;
2340 case WINED3D_SHADER_TYPE_GEOMETRY:
2341 backend_version = d3d_info->limits.gs_version;
2342 break;
2343 case WINED3D_SHADER_TYPE_PIXEL:
2344 backend_version = d3d_info->limits.ps_version;
2345 break;
2346 default:
2347 FIXME("No backend version-checking for this shader type\n");
2348 backend_version = 0;
2349 }
2350 if (reg_maps->shader_version.major > backend_version)
2351 {
2352 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2353 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2354 return WINED3DERR_INVALIDCALL;
2355 }
2356
2357 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2358 if (!shader->function)
2359 return E_OUTOFMEMORY;
2360 memcpy(shader->function, byte_code, shader->functionLength);
2361
2362 return WINED3D_OK;
2363 }
2364
2365 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2366 {
2367 ULONG refcount = InterlockedIncrement(&shader->ref);
2368
2369 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2370
2371 return refcount;
2372 }
2373
2374 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2375 {
2376 ULONG refcount = InterlockedDecrement(&shader->ref);
2377
2378 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2379
2380 if (!refcount)
2381 {
2382 shader_cleanup(shader);
2383 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2384 HeapFree(GetProcessHeap(), 0, shader);
2385 }
2386
2387 return refcount;
2388 }
2389
2390 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2391 {
2392 TRACE("shader %p.\n", shader);
2393
2394 return shader->parent;
2395 }
2396
2397 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2398 void *byte_code, UINT *byte_code_size)
2399 {
2400 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2401
2402 if (!byte_code)
2403 {
2404 *byte_code_size = shader->functionLength;
2405 return WINED3D_OK;
2406 }
2407
2408 if (*byte_code_size < shader->functionLength)
2409 {
2410 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2411 * than the required size we should write the required size and
2412 * return D3DERR_MOREDATA. That's not actually true. */
2413 return WINED3DERR_INVALIDCALL;
2414 }
2415
2416 memcpy(byte_code, shader->function, shader->functionLength);
2417
2418 return WINED3D_OK;
2419 }
2420
2421 /* Set local constants for d3d8 shaders. */
2422 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2423 UINT start_idx, const float *src_data, UINT count)
2424 {
2425 UINT end_idx = start_idx + count;
2426 UINT i;
2427
2428 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2429
2430 if (end_idx > shader->limits->constant_float)
2431 {
2432 WARN("end_idx %u > float constants limit %u.\n",
2433 end_idx, shader->limits->constant_float);
2434 end_idx = shader->limits->constant_float;
2435 }
2436
2437 for (i = start_idx; i < end_idx; ++i)
2438 {
2439 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2440 float *value;
2441 if (!lconst)
2442 return E_OUTOFMEMORY;
2443
2444 lconst->idx = i;
2445 value = (float *)lconst->value;
2446 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2447 list_add_head(&shader->constantsF, &lconst->entry);
2448
2449 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2450 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2451 {
2452 shader->lconst_inf_or_nan = TRUE;
2453 }
2454 }
2455
2456 return WINED3D_OK;
2457 }
2458
2459 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2460 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2461 {
2462 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2463 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2464 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2465 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2466 args->point_size = state->gl_primitive_type == GL_POINTS;
2467 args->per_vertex_point_size = shader->reg_maps.point_size;
2468 args->swizzle_map = swizzle_map;
2469 if (d3d_info->emulated_flatshading)
2470 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2471 else
2472 args->flatshading = FALSE;
2473 }
2474
2475 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2476 {
2477 if (usage_idx1 != usage_idx2)
2478 return FALSE;
2479 if (usage1 == usage2)
2480 return TRUE;
2481 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2482 return TRUE;
2483 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2484 return TRUE;
2485
2486 return FALSE;
2487 }
2488
2489 BOOL vshader_get_input(const struct wined3d_shader *shader,
2490 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2491 {
2492 WORD map = shader->reg_maps.input_registers;
2493 unsigned int i;
2494
2495 for (i = 0; map; map >>= 1, ++i)
2496 {
2497 if (!(map & 1)) continue;
2498
2499 if (match_usage(shader->u.vs.attributes[i].usage,
2500 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2501 {
2502 *regnum = i;
2503 return TRUE;
2504 }
2505 }
2506 return FALSE;
2507 }
2508
2509 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2510 const struct wined3d_shader_signature *src, char **signature_strings)
2511 {
2512 struct wined3d_shader_signature_element *e;
2513 unsigned int i;
2514 SIZE_T len;
2515 char *ptr;
2516
2517 ptr = *signature_strings;
2518
2519 dst->element_count = src->element_count;
2520 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2521 return E_OUTOFMEMORY;
2522
2523 for (i = 0; i < src->element_count; ++i)
2524 {
2525 e = &src->elements[i];
2526 dst->elements[i] = *e;
2527
2528 len = strlen(e->semantic_name);
2529 memcpy(ptr, e->semantic_name, len + 1);
2530 dst->elements[i].semantic_name = ptr;
2531 ptr += len + 1;
2532 }
2533
2534 *signature_strings = ptr;
2535
2536 return WINED3D_OK;
2537 }
2538
2539 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2540 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2541 void *parent, const struct wined3d_parent_ops *parent_ops)
2542 {
2543 struct wined3d_shader_signature_element *e;
2544 SIZE_T total, len;
2545 unsigned int i;
2546 HRESULT hr;
2547 char *ptr;
2548
2549 if (!desc->byte_code)
2550 return WINED3DERR_INVALIDCALL;
2551
2552 shader->ref = 1;
2553 shader->device = device;
2554 shader->parent = parent;
2555 shader->parent_ops = parent_ops;
2556
2557 total = 0;
2558 if (desc->input_signature)
2559 {
2560 for (i = 0; i < desc->input_signature->element_count; ++i)
2561 {
2562 e = &desc->input_signature->elements[i];
2563 len = strlen(e->semantic_name);
2564 if (len >= ~(SIZE_T)0 - total)
2565 return E_OUTOFMEMORY;
2566
2567 total += len + 1;
2568 }
2569 }
2570 if (desc->output_signature)
2571 {
2572 for (i = 0; i < desc->output_signature->element_count; ++i)
2573 {
2574 e = &desc->output_signature->elements[i];
2575 len = strlen(e->semantic_name);
2576 if (len >= ~(SIZE_T)0 - total)
2577 return E_OUTOFMEMORY;
2578
2579 total += len + 1;
2580 }
2581 }
2582 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2583 return E_OUTOFMEMORY;
2584 ptr = shader->signature_strings;
2585
2586 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2587 desc->input_signature, &ptr)))
2588 {
2589 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2590 return hr;
2591 }
2592 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2593 desc->output_signature, &ptr)))
2594 {
2595 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2596 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2597 return hr;
2598 }
2599
2600 list_init(&shader->linked_programs);
2601 list_add_head(&device->shaders, &shader->shader_list_entry);
2602
2603 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2604 float_const_count, type, desc->max_version)))
2605 {
2606 WARN("Failed to set function, hr %#x.\n", hr);
2607 shader_cleanup(shader);
2608 }
2609
2610 return hr;
2611 }
2612
2613 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2614 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2615 {
2616 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2617 unsigned int i;
2618 HRESULT hr;
2619
2620 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2621 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2622 return hr;
2623
2624 for (i = 0; i < shader->input_signature.element_count; ++i)
2625 {
2626 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2627
2628 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2629 continue;
2630
2631 shader->u.vs.attributes[input->register_idx].usage =
2632 shader_usage_from_semantic_name(input->semantic_name);
2633 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2634 }
2635
2636 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2637 shader->lconst_inf_or_nan;
2638
2639 return WINED3D_OK;
2640 }
2641
2642 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2643 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2644 {
2645 HRESULT hr;
2646
2647 if (FAILED(hr = shader_init(shader, device, desc, 0,
2648 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2649 return hr;
2650
2651 shader->load_local_constsF = shader->lconst_inf_or_nan;
2652
2653 return WINED3D_OK;
2654 }
2655
2656 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2657 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2658 {
2659 const struct wined3d_gl_info *gl_info = context->gl_info;
2660 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2661 const struct wined3d_texture *texture;
2662 UINT i;
2663
2664 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2665 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
2666 {
2667 static unsigned int warned = 0;
2668
2669 args->srgb_correction = 1;
2670 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2671 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2672 "support, expect rendering artifacts.\n");
2673 }
2674
2675 if (shader->reg_maps.shader_version.major == 1
2676 && shader->reg_maps.shader_version.minor <= 3)
2677 {
2678 for (i = 0; i < shader->limits->sampler; ++i)
2679 {
2680 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2681
2682 if (flags & WINED3D_TTFF_PROJECTED)
2683 {
2684 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2685
2686 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2687 {
2688 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2689 unsigned int j;
2690 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2691 DWORD max_valid = WINED3D_TTFF_COUNT4;
2692
2693 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2694 {
2695 struct wined3d_vertex_declaration_element *element =
2696 &state->vertex_declaration->elements[j];
2697
2698 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2699 && element->usage_idx == index)
2700 {
2701 max_valid = element->format->component_count;
2702 break;
2703 }
2704 }
2705 if (!tex_transform || tex_transform > max_valid)
2706 {
2707 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2708 tex_transform, max_valid);
2709 tex_transform = max_valid;
2710 }
2711 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2712 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2713 && tex_transform > WINED3D_TTFF_COUNT2)
2714 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2715 && tex_transform > WINED3D_TTFF_COUNT3))
2716 tex_transform |= WINED3D_PSARGS_PROJECTED;
2717 else
2718 {
2719 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2720 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2721 i, tex_transform, resource_type);
2722 }
2723 }
2724 else
2725 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2726
2727 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2728 }
2729 }
2730 }
2731 if (shader->reg_maps.shader_version.major == 1
2732 && shader->reg_maps.shader_version.minor <= 4)
2733 {
2734 for (i = 0; i < shader->limits->sampler; ++i)
2735 {
2736 const struct wined3d_texture *texture = state->textures[i];
2737
2738 if (!shader->reg_maps.resource_info[i].type)
2739 continue;
2740
2741 /* Treat unbound textures as 2D. The dummy texture will provide
2742 * the proper sample value. The tex_types bitmap defaults to
2743 * 2D because of the memset. */
2744 if (!texture)
2745 continue;
2746
2747 switch (texture->target)
2748 {
2749 /* RECT textures are distinguished from 2D textures via np2_fixup */
2750 case GL_TEXTURE_RECTANGLE_ARB:
2751 case GL_TEXTURE_2D:
2752 break;
2753
2754 case GL_TEXTURE_3D:
2755 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2756 break;
2757
2758 case GL_TEXTURE_CUBE_MAP_ARB:
2759 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2760 break;
2761 }
2762 }
2763 }
2764
2765 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2766 {
2767 if (!shader->reg_maps.resource_info[i].type)
2768 continue;
2769
2770 texture = state->textures[i];
2771 if (!texture)
2772 {
2773 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2774 continue;
2775 }
2776 args->color_fixup[i] = texture->resource.format->color_fixup;
2777
2778 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
2779 args->shadow |= 1u << i;
2780
2781 /* Flag samplers that need NP2 texcoord fixup. */
2782 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2783 args->np2_fixup |= (1u << i);
2784 }
2785 if (shader->reg_maps.shader_version.major >= 3)
2786 {
2787 if (position_transformed)
2788 args->vp_mode = pretransformed;
2789 else if (use_vs(state))
2790 args->vp_mode = vertexshader;
2791 else
2792 args->vp_mode = fixedfunction;
2793 args->fog = WINED3D_FFP_PS_FOG_OFF;
2794 }
2795 else
2796 {
2797 args->vp_mode = vertexshader;
2798 if (state->render_states[WINED3D_RS_FOGENABLE])
2799 {
2800 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2801 {
2802 case WINED3D_FOG_NONE:
2803 if (position_transformed || use_vs(state))
2804 {
2805 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2806 break;
2807 }
2808
2809 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2810 {
2811 case WINED3D_FOG_NONE: /* Fall through. */
2812 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2813 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2814 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2815 }
2816 break;
2817
2818 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2819 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2820 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2821 }
2822 }
2823 else
2824 {
2825 args->fog = WINED3D_FFP_PS_FOG_OFF;
2826 }
2827 }
2828
2829 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
2830 {
2831 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
2832
2833 args->texcoords_initialized = 0;
2834 for (i = 0; i < MAX_TEXTURES; ++i)
2835 {
2836 if (vs)
2837 {
2838 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
2839 args->texcoords_initialized |= 1u << i;
2840 }
2841 else
2842 {
2843 const struct wined3d_stream_info *si = &context->stream_info;
2844 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2845
2846 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
2847 & WINED3D_FFP_TCI_MASK
2848 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
2849 args->texcoords_initialized |= 1u << i;
2850 }
2851 }
2852 }
2853 else
2854 {
2855 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
2856 }
2857
2858 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
2859 && state->gl_primitive_type == GL_POINTS;
2860
2861 if (d3d_info->emulated_flatshading)
2862 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2863 }
2864
2865 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2866 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2867 {
2868 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2869 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2870 HRESULT hr;
2871
2872 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
2873 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
2874 return hr;
2875
2876 for (i = 0; i < MAX_REG_INPUT; ++i)
2877 {
2878 if (shader->u.ps.input_reg_used[i])
2879 {
2880 ++num_regs_used;
2881 highest_reg_used = i;
2882 }
2883 }
2884
2885 /* Don't do any register mapping magic if it is not needed, or if we can't
2886 * achieve anything anyway */
2887 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2888 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2889 {
2890 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2891 {
2892 /* This happens with relative addressing. The input mapper function
2893 * warns about this if the higher registers are declared too, so
2894 * don't write a FIXME here */
2895 WARN("More varying registers used than supported\n");
2896 }
2897
2898 for (i = 0; i < MAX_REG_INPUT; ++i)
2899 {
2900 shader->u.ps.input_reg_map[i] = i;
2901 }
2902
2903 shader->u.ps.declared_in_count = highest_reg_used + 1;
2904 }
2905 else
2906 {
2907 shader->u.ps.declared_in_count = 0;
2908 for (i = 0; i < MAX_REG_INPUT; ++i)
2909 {
2910 if (shader->u.ps.input_reg_used[i])
2911 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2912 else shader->u.ps.input_reg_map[i] = ~0U;
2913 }
2914 }
2915
2916 shader->load_local_constsF = shader->lconst_inf_or_nan;
2917
2918 return WINED3D_OK;
2919 }
2920
2921 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
2922 {
2923 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2924 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
2925 unsigned int i;
2926
2927 if (reg_maps->shader_version.major != 1) return;
2928
2929 for (i = 0; i < shader->limits->sampler; ++i)
2930 {
2931 /* We don't sample from this sampler. */
2932 if (!resource_info[i].type)
2933 continue;
2934
2935 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
2936 {
2937 case WINED3D_SHADER_TEX_2D:
2938 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
2939 break;
2940
2941 case WINED3D_SHADER_TEX_3D:
2942 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
2943 break;
2944
2945 case WINED3D_SHADER_TEX_CUBE:
2946 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
2947 break;
2948 }
2949 }
2950 }
2951
2952 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2953 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2954 {
2955 struct wined3d_shader *object;
2956 HRESULT hr;
2957
2958 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2959 device, desc, parent, parent_ops, shader);
2960
2961 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2962 if (!object)
2963 return E_OUTOFMEMORY;
2964
2965 if (FAILED(hr = geometryshader_init(object, device, desc, parent, parent_ops)))
2966 {
2967 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2968 HeapFree(GetProcessHeap(), 0, object);
2969 return hr;
2970 }
2971
2972 TRACE("Created geometry shader %p.\n", object);
2973 *shader = object;
2974
2975 return WINED3D_OK;
2976 }
2977
2978 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2979 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2980 {
2981 struct wined3d_shader *object;
2982 HRESULT hr;
2983
2984 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2985 device, desc, parent, parent_ops, shader);
2986
2987 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2988 if (!object)
2989 return E_OUTOFMEMORY;
2990
2991 if (FAILED(hr = pixelshader_init(object, device, desc, parent, parent_ops)))
2992 {
2993 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2994 HeapFree(GetProcessHeap(), 0, object);
2995 return hr;
2996 }
2997
2998 TRACE("Created pixel shader %p.\n", object);
2999 *shader = object;
3000
3001 return WINED3D_OK;
3002 }
3003
3004 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3005 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3006 {
3007 struct wined3d_shader *object;
3008 HRESULT hr;
3009
3010 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3011 device, desc, parent, parent_ops, shader);
3012
3013 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3014 if (!object)
3015 return E_OUTOFMEMORY;
3016
3017 if (FAILED(hr = vertexshader_init(object, device, desc, parent, parent_ops)))
3018 {
3019 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3020 HeapFree(GetProcessHeap(), 0, object);
3021 return hr;
3022 }
3023
3024 TRACE("Created vertex shader %p.\n", object);
3025 *shader = object;
3026
3027 return WINED3D_OK;
3028 }
Windows API implementation