?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Recognize SM 4.1 ld2dms opcode.
[wine.git] / dlls / wined3d / shader.c
blobf45fbb8b13d8c9ecf4823e22bece574ebfbd509c
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_DSX_COARSE */ "deriv_rtx_coarse",
84 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
85 /* WINED3DSIH_DSY */ "dsy",
86 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
87 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
88 /* WINED3DSIH_ELSE */ "else",
89 /* WINED3DSIH_EMIT */ "emit",
90 /* WINED3DSIH_ENDIF */ "endif",
91 /* WINED3DSIH_ENDLOOP */ "endloop",
92 /* WINED3DSIH_ENDREP */ "endrep",
93 /* WINED3DSIH_EQ */ "eq",
94 /* WINED3DSIH_EXP */ "exp",
95 /* WINED3DSIH_EXPP */ "expp",
96 /* WINED3DSIH_FRC */ "frc",
97 /* WINED3DSIH_FTOI */ "ftoi",
98 /* WINED3DSIH_FTOU */ "ftou",
99 /* WINED3DSIH_GE */ "ge",
100 /* WINED3DSIH_IADD */ "iadd",
101 /* WINED3DSIH_IEQ */ "ieq",
102 /* WINED3DSIH_IF */ "if",
103 /* WINED3DSIH_IFC */ "ifc",
104 /* WINED3DSIH_IGE */ "ige",
105 /* WINED3DSIH_ILT */ "ilt",
106 /* WINED3DSIH_IMAD */ "imad",
107 /* WINED3DSIH_IMAX */ "imax",
108 /* WINED3DSIH_IMIN */ "imin",
109 /* WINED3DSIH_IMUL */ "imul",
110 /* WINED3DSIH_INE */ "ine",
111 /* WINED3DSIH_INEG */ "ineg",
112 /* WINED3DSIH_ISHL */ "ishl",
113 /* WINED3DSIH_ITOF */ "itof",
114 /* WINED3DSIH_LABEL */ "label",
115 /* WINED3DSIH_LD */ "ld",
116 /* WINED3DSIH_LD2DMS */ "ld2dms",
117 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
118 /* WINED3DSIH_LIT */ "lit",
119 /* WINED3DSIH_LOG */ "log",
120 /* WINED3DSIH_LOGP */ "logp",
121 /* WINED3DSIH_LOOP */ "loop",
122 /* WINED3DSIH_LRP */ "lrp",
123 /* WINED3DSIH_LT */ "lt",
124 /* WINED3DSIH_M3x2 */ "m3x2",
125 /* WINED3DSIH_M3x3 */ "m3x3",
126 /* WINED3DSIH_M3x4 */ "m3x4",
127 /* WINED3DSIH_M4x3 */ "m4x3",
128 /* WINED3DSIH_M4x4 */ "m4x4",
129 /* WINED3DSIH_MAD */ "mad",
130 /* WINED3DSIH_MAX */ "max",
131 /* WINED3DSIH_MIN */ "min",
132 /* WINED3DSIH_MOV */ "mov",
133 /* WINED3DSIH_MOVA */ "mova",
134 /* WINED3DSIH_MOVC */ "movc",
135 /* WINED3DSIH_MUL */ "mul",
136 /* WINED3DSIH_NE */ "ne",
137 /* WINED3DSIH_NOP */ "nop",
138 /* WINED3DSIH_NOT */ "not",
139 /* WINED3DSIH_NRM */ "nrm",
140 /* WINED3DSIH_OR */ "or",
141 /* WINED3DSIH_PHASE */ "phase",
142 /* WINED3DSIH_POW */ "pow",
143 /* WINED3DSIH_RCP */ "rcp",
144 /* WINED3DSIH_REP */ "rep",
145 /* WINED3DSIH_RESINFO */ "resinfo",
146 /* WINED3DSIH_RET */ "ret",
147 /* WINED3DSIH_ROUND_NI */ "round_ni",
148 /* WINED3DSIH_ROUND_PI */ "round_pi",
149 /* WINED3DSIH_ROUND_Z */ "round_z",
150 /* WINED3DSIH_RSQ */ "rsq",
151 /* WINED3DSIH_SAMPLE */ "sample",
152 /* WINED3DSIH_SAMPLE_B */ "sample_b",
153 /* WINED3DSIH_SAMPLE_C */ "sample_c",
154 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
155 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
156 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
157 /* WINED3DSIH_SETP */ "setp",
158 /* WINED3DSIH_SGE */ "sge",
159 /* WINED3DSIH_SGN */ "sgn",
160 /* WINED3DSIH_SINCOS */ "sincos",
161 /* WINED3DSIH_SLT */ "slt",
162 /* WINED3DSIH_SQRT */ "sqrt",
163 /* WINED3DSIH_SUB */ "sub",
164 /* WINED3DSIH_TEX */ "texld",
165 /* WINED3DSIH_TEXBEM */ "texbem",
166 /* WINED3DSIH_TEXBEML */ "texbeml",
167 /* WINED3DSIH_TEXCOORD */ "texcrd",
168 /* WINED3DSIH_TEXDEPTH */ "texdepth",
169 /* WINED3DSIH_TEXDP3 */ "texdp3",
170 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
171 /* WINED3DSIH_TEXKILL */ "texkill",
172 /* WINED3DSIH_TEXLDD */ "texldd",
173 /* WINED3DSIH_TEXLDL */ "texldl",
174 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
175 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
176 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
177 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
178 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
179 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
180 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
181 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
182 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
183 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
184 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
185 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
186 /* WINED3DSIH_UDIV */ "udiv",
187 /* WINED3DSIH_UGE */ "uge",
188 /* WINED3DSIH_USHR */ "ushr",
189 /* WINED3DSIH_UTOF */ "utof",
190 /* WINED3DSIH_XOR */ "xor",
191 };
192
193 static const char * const semantic_names[] =
194 {
195 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
196 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
197 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
198 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
199 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
200 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
201 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
202 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
203 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
204 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
205 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
206 /* WINED3D_DECL_USAGE_FOG */ "FOG",
207 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
208 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
209 };
210
211 static const struct
212 {
213 enum wined3d_sysval_semantic sysval_semantic;
214 const char *sysval_name;
215 }
216 sysval_semantic_names[] =
217 {
218 {WINED3D_SV_POSITION, "SV_Position"},
219 {WINED3D_SV_INSTANCEID, "SV_InstanceID"},
220 {WINED3D_SV_PRIMITIVEID, "SV_PrimitiveID"},
221 {WINED3D_SV_ISFRONTFACE, "SV_IsFrontFace"},
222 {WINED3D_SV_SAMPLEINDEX, "SV_SampleIndex"},
223 };
224
225 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
226 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
227
228 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
229 {
230 if (handler_idx >= sizeof(shader_opcode_names) / sizeof(*shader_opcode_names))
231 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
232
233 return shader_opcode_names[handler_idx];
234 }
235
236 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
237 {
238 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
239 {
240 FIXME("Unrecognized usage %#x.\n", usage);
241 return "UNRECOGNIZED";
242 }
243
244 return semantic_names[usage];
245 }
246
247 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
248 {
249 unsigned int i;
250
251 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
252 {
253 if (!strcmp(name, semantic_names[i])) return i;
254 }
255
256 return ~0U;
257 }
258
259 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
260 {
261 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
262 }
263
264 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
265 const struct wined3d_shader_semantic *s)
266 {
267 e->semantic_name = shader_semantic_name_from_usage(s->usage);
268 e->semantic_idx = s->usage_idx;
269 e->sysval_semantic = 0;
270 e->component_type = 0;
271 e->register_idx = s->reg.reg.idx[0].offset;
272 e->mask = s->reg.write_mask;
273 }
274
275 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
276 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
277 {
278 e->semantic_name = shader_semantic_name_from_usage(usage);
279 e->semantic_idx = usage_idx;
280 e->sysval_semantic = 0;
281 e->component_type = 0;
282 e->register_idx = reg_idx;
283 e->mask = write_mask;
284 }
285
286 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
287 {
288 switch (version_token >> 16)
289 {
290 case WINED3D_SM1_VS:
291 case WINED3D_SM1_PS:
292 return &sm1_shader_frontend;
293
294 case WINED3D_SM4_PS:
295 case WINED3D_SM4_VS:
296 case WINED3D_SM4_GS:
297 return &sm4_shader_frontend;
298
299 default:
300 FIXME("Unrecognised version token %#x\n", version_token);
301 return NULL;
302 }
303 }
304
305 void string_buffer_clear(struct wined3d_string_buffer *buffer)
306 {
307 buffer->buffer[0] = '\0';
308 buffer->content_size = 0;
309 }
310
311 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
312 {
313 buffer->buffer_size = 32;
314 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
315 {
316 ERR("Failed to allocate shader buffer memory.\n");
317 return FALSE;
318 }
319
320 string_buffer_clear(buffer);
321 return TRUE;
322 }
323
324 void string_buffer_free(struct wined3d_string_buffer *buffer)
325 {
326 HeapFree(GetProcessHeap(), 0, buffer->buffer);
327 }
328
329 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
330 {
331 char *new_buffer;
332 unsigned int new_buffer_size = buffer->buffer_size * 2;
333
334 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
335 new_buffer_size *= 2;
336 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
337 {
338 ERR("Failed to grow buffer.\n");
339 buffer->buffer[buffer->content_size] = '\0';
340 return FALSE;
341 }
342 buffer->buffer = new_buffer;
343 buffer->buffer_size = new_buffer_size;
344 return TRUE;
345 }
346
347 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
348 {
349 unsigned int rem;
350 int rc;
351
352 rem = buffer->buffer_size - buffer->content_size;
353 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
354 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
355 return rc;
356
357 buffer->content_size += rc;
358 return 0;
359 }
360
361 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
362 {
363 va_list args;
364 int ret;
365
366 for (;;)
367 {
368 va_start(args, format);
369 ret = shader_vaddline(buffer, format, args);
370 va_end(args);
371 if (!ret)
372 return ret;
373 if (!string_buffer_resize(buffer, ret))
374 return -1;
375 }
376 }
377
378 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
379 {
380 struct wined3d_string_buffer *buffer;
381
382 if (list_empty(&list->list))
383 {
384 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
385 if (!buffer || !string_buffer_init(buffer))
386 {
387 ERR("Couldn't allocate buffer for temporary string.\n");
388 if (buffer)
389 HeapFree(GetProcessHeap(), 0, buffer);
390 return NULL;
391 }
392 }
393 else
394 {
395 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
396 list_remove(&buffer->entry);
397 }
398 string_buffer_clear(buffer);
399 return buffer;
400 }
401
402 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
403 {
404 if (!buffer)
405 return 0;
406 string_buffer_clear(buffer);
407 return shader_vaddline(buffer, format, args);
408 }
409
410 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
411 {
412 va_list args;
413 int ret;
414
415 for (;;)
416 {
417 va_start(args, format);
418 ret = string_buffer_vsprintf(buffer, format, args);
419 va_end(args);
420 if (!ret)
421 return;
422 if (!string_buffer_resize(buffer, ret))
423 return;
424 }
425 }
426
427 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
428 {
429 if (!buffer)
430 return;
431 list_add_head(&list->list, &buffer->entry);
432 }
433
434 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
435 {
436 list_init(&list->list);
437 }
438
439 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
440 {
441 struct wined3d_string_buffer *buffer, *buffer_next;
442
443 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
444 {
445 string_buffer_free(buffer);
446 HeapFree(GetProcessHeap(), 0, buffer);
447 }
448 list_init(&list->list);
449 }
450
451 /* Convert floating point offset relative to a register file to an absolute
452 * offset for float constants. */
453 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
454 {
455 switch (register_type)
456 {
457 case WINED3DSPR_CONST: return register_idx;
458 case WINED3DSPR_CONST2: return 2048 + register_idx;
459 case WINED3DSPR_CONST3: return 4096 + register_idx;
460 case WINED3DSPR_CONST4: return 6144 + register_idx;
461 default:
462 FIXME("Unsupported register type: %u.\n", register_type);
463 return register_idx;
464 }
465 }
466
467 static void shader_delete_constant_list(struct list *clist)
468 {
469 struct wined3d_shader_lconst *constant, *constant_next;
470
471 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
472 HeapFree(GetProcessHeap(), 0, constant);
473 list_init(clist);
474 }
475
476 static void shader_set_limits(struct wined3d_shader *shader)
477 {
478 static const struct limits_entry
479 {
480 unsigned int min_version;
481 unsigned int max_version;
482 struct wined3d_shader_limits limits;
483 }
484 vs_limits[] =
485 {
486 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
487 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
488 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
489 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
490 * even though they are capable of supporting much more (GL
491 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
492 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
493 * shaders to 256. */
494 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
495 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
496 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
497 {0}
498 },
499 gs_limits[] =
500 {
501 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
502 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
503 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
504 {0}
505 },
506 ps_limits[] =
507 {
508 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
509 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
510 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
511 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
512 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
513 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
514 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
515 {0}
516 };
517 const struct limits_entry *limits_array;
518 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
519 shader->reg_maps.shader_version.minor);
520 int i = 0;
521
522 switch (shader->reg_maps.shader_version.type)
523 {
524 default:
525 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
526 /* Fall-through. */
527 case WINED3D_SHADER_TYPE_VERTEX:
528 limits_array = vs_limits;
529 break;
530 case WINED3D_SHADER_TYPE_GEOMETRY:
531 limits_array = gs_limits;
532 break;
533 case WINED3D_SHADER_TYPE_PIXEL:
534 limits_array = ps_limits;
535 break;
536 }
537
538 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
539 {
540 if (shader_version <= limits_array[i].max_version)
541 {
542 shader->limits = &limits_array[i].limits;
543 break;
544 }
545 ++i;
546 }
547 if (!shader->limits)
548 {
549 FIXME("Unexpected shader version \"%u.%u\".\n",
550 shader->reg_maps.shader_version.major,
551 shader->reg_maps.shader_version.minor);
552 shader->limits = &limits_array[max(0, i - 1)].limits;
553 }
554 }
555
556 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
557 {
558 DWORD idx, shift;
559 idx = bit >> 5;
560 shift = bit & 0x1f;
561 bitmap[idx] |= (1u << shift);
562 }
563
564 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
565 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
566 {
567 switch (reg->type)
568 {
569 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
570 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
571 reg_maps->texcoord |= 1u << reg->idx[0].offset;
572 else
573 reg_maps->address |= 1u << reg->idx[0].offset;
574 break;
575
576 case WINED3DSPR_TEMP:
577 reg_maps->temporary |= 1u << reg->idx[0].offset;
578 break;
579
580 case WINED3DSPR_INPUT:
581 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
582 {
583 if (reg->idx[0].rel_addr)
584 {
585 /* If relative addressing is used, we must assume that all registers
586 * are used. Even if it is a construct like v3[aL], we can't assume
587 * that v0, v1 and v2 aren't read because aL can be negative */
588 unsigned int i;
589 for (i = 0; i < MAX_REG_INPUT; ++i)
590 {
591 shader->u.ps.input_reg_used[i] = TRUE;
592 }
593 }
594 else
595 {
596 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
597 }
598 }
599 else
600 reg_maps->input_registers |= 1u << reg->idx[0].offset;
601 break;
602
603 case WINED3DSPR_RASTOUT:
604 if (reg->idx[0].offset == 1)
605 reg_maps->fog = 1;
606 if (reg->idx[0].offset == 2)
607 reg_maps->point_size = 1;
608 break;
609
610 case WINED3DSPR_MISCTYPE:
611 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
612 {
613 if (!reg->idx[0].offset)
614 reg_maps->vpos = 1;
615 else if (reg->idx[0].offset == 1)
616 reg_maps->usesfacing = 1;
617 }
618 break;
619
620 case WINED3DSPR_CONST:
621 if (reg->idx[0].rel_addr)
622 {
623 if (reg->idx[0].offset < reg_maps->min_rel_offset)
624 reg_maps->min_rel_offset = reg->idx[0].offset;
625 if (reg->idx[0].offset > reg_maps->max_rel_offset)
626 reg_maps->max_rel_offset = reg->idx[0].offset;
627 reg_maps->usesrelconstF = TRUE;
628 }
629 else
630 {
631 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
632 {
633 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
634 return FALSE;
635 }
636 else
637 {
638 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
639 }
640 }
641 break;
642
643 case WINED3DSPR_CONSTINT:
644 if (reg->idx[0].offset >= shader->limits->constant_int)
645 {
646 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
647 return FALSE;
648 }
649 else
650 {
651 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
652 }
653 break;
654
655 case WINED3DSPR_CONSTBOOL:
656 if (reg->idx[0].offset >= shader->limits->constant_bool)
657 {
658 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
659 return FALSE;
660 }
661 else
662 {
663 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
664 }
665 break;
666
667 case WINED3DSPR_COLOROUT:
668 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
669 break;
670
671 default:
672 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
673 reg->type, reg->idx[0].offset, reg->idx[1].offset);
674 break;
675 }
676 return TRUE;
677 }
678
679 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
680 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
681 {
682 struct wined3d_shader_sampler_map_entry *entries, *entry;
683 struct wined3d_shader_sampler_map *map;
684 unsigned int i;
685
686 map = &reg_maps->sampler_map;
687 entries = map->entries;
688 for (i = 0; i < map->count; ++i)
689 {
690 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
691 return;
692 }
693
694 if (!map->size)
695 {
696 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
697 {
698 ERR("Failed to allocate sampler map entries.\n");
699 return;
700 }
701 map->size = 4;
702 map->entries = entries;
703 }
704 else if (map->count == map->size)
705 {
706 size_t new_size = map->size * 2;
707
708 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
709 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
710 {
711 ERR("Failed to resize sampler map entries.\n");
712 return;
713 }
714 map->size = new_size;
715 map->entries = entries;
716 }
717
718 entry = &entries[map->count++];
719 entry->resource_idx = resource_idx;
720 entry->sampler_idx = sampler_idx;
721 entry->bind_idx = bind_idx;
722 }
723
724 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
725 {
726 switch (instr)
727 {
728 case WINED3DSIH_M4x4:
729 case WINED3DSIH_M3x4:
730 return param == 1 ? 3 : 0;
731
732 case WINED3DSIH_M4x3:
733 case WINED3DSIH_M3x3:
734 return param == 1 ? 2 : 0;
735
736 case WINED3DSIH_M3x2:
737 return param == 1 ? 1 : 0;
738
739 default:
740 return 0;
741 }
742 }
743
744 /* Note that this does not count the loop register as an address register. */
745 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
746 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
747 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
748 {
749 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
750 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
751 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
752 void *fe_data = shader->frontend_data;
753 struct wined3d_shader_version shader_version;
754 const DWORD *ptr = byte_code;
755 unsigned int i;
756
757 memset(reg_maps, 0, sizeof(*reg_maps));
758 memset(input_signature_elements, 0, sizeof(input_signature_elements));
759 memset(output_signature_elements, 0, sizeof(output_signature_elements));
760 reg_maps->min_rel_offset = ~0U;
761
762 fe->shader_read_header(fe_data, &ptr, &shader_version);
763 reg_maps->shader_version = shader_version;
764
765 shader_set_limits(shader);
766
767 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
768 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
769 if (!reg_maps->constf)
770 {
771 ERR("Failed to allocate constant map memory.\n");
772 return E_OUTOFMEMORY;
773 }
774
775 while (!fe->shader_is_end(fe_data, &ptr))
776 {
777 struct wined3d_shader_instruction ins;
778
779 /* Fetch opcode. */
780 fe->shader_read_instruction(fe_data, &ptr, &ins);
781
782 /* Unhandled opcode, and its parameters. */
783 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
784 {
785 TRACE("Skipping unrecognized instruction.\n");
786 continue;
787 }
788
789 /* Handle declarations. */
790 if (ins.handler_idx == WINED3DSIH_DCL)
791 {
792 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
793 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
794
795 switch (semantic->reg.reg.type)
796 {
797 /* Mark input registers used. */
798 case WINED3DSPR_INPUT:
799 if (reg_idx >= MAX_REG_INPUT)
800 {
801 ERR("Invalid input register index %u.\n", reg_idx);
802 break;
803 }
804 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
805 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
806 return WINED3DERR_INVALIDCALL;
807 reg_maps->input_registers |= 1u << reg_idx;
808 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
809 break;
810
811 /* Vertex shader: mark 3.0 output registers used, save token. */
812 case WINED3DSPR_OUTPUT:
813 if (reg_idx >= MAX_REG_OUTPUT)
814 {
815 ERR("Invalid output register index %u.\n", reg_idx);
816 break;
817 }
818 reg_maps->output_registers |= 1u << reg_idx;
819 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
820 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
821 reg_maps->fog = 1;
822 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
823 reg_maps->point_size = 1;
824 break;
825
826 case WINED3DSPR_SAMPLER:
827 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
828 case WINED3DSPR_RESOURCE:
829 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
830 {
831 ERR("Invalid resource index %u.\n", reg_idx);
832 break;
833 }
834 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
835 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
836 break;
837
838 default:
839 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
840 break;
841 }
842 }
843 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
844 {
845 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
846 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
847 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
848 else
849 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
850 }
851 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
852 {
853 if (reg_maps->icb)
854 FIXME("Multiple immediate constant buffers.\n");
855 reg_maps->icb = ins.declaration.icb;
856 }
857 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
858 {
859 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
860 shader->u.gs.input_type = ins.declaration.primitive_type;
861 else
862 FIXME("Invalid instruction %#x for shader type %#x.\n",
863 ins.handler_idx, shader_version.type);
864 }
865 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
866 {
867 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
868 shader->u.gs.output_type = ins.declaration.primitive_type;
869 else
870 FIXME("Invalid instruction %#x for shader type %#x.\n",
871 ins.handler_idx, shader_version.type);
872 }
873 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
874 {
875 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
876 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
877 }
878 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
879 {
880 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
881 shader->u.gs.vertices_out = ins.declaration.count;
882 else
883 FIXME("Invalid instruction %#x for shader type %#x.\n",
884 ins.handler_idx, shader_version.type);
885 }
886 else if (ins.handler_idx == WINED3DSIH_DEF)
887 {
888 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
889 float *value;
890 if (!lconst) return E_OUTOFMEMORY;
891
892 lconst->idx = ins.dst[0].reg.idx[0].offset;
893 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
894 value = (float *)lconst->value;
895
896 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
897 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
898 {
899 if (value[0] < -1.0f) value[0] = -1.0f;
900 else if (value[0] > 1.0f) value[0] = 1.0f;
901 if (value[1] < -1.0f) value[1] = -1.0f;
902 else if (value[1] > 1.0f) value[1] = 1.0f;
903 if (value[2] < -1.0f) value[2] = -1.0f;
904 else if (value[2] > 1.0f) value[2] = 1.0f;
905 if (value[3] < -1.0f) value[3] = -1.0f;
906 else if (value[3] > 1.0f) value[3] = 1.0f;
907 }
908
909 list_add_head(&shader->constantsF, &lconst->entry);
910
911 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
912 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
913 {
914 shader->lconst_inf_or_nan = TRUE;
915 }
916 }
917 else if (ins.handler_idx == WINED3DSIH_DEFI)
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, 4 * sizeof(DWORD));
924
925 list_add_head(&shader->constantsI, &lconst->entry);
926 reg_maps->local_int_consts |= (1u << lconst->idx);
927 }
928 else if (ins.handler_idx == WINED3DSIH_DEFB)
929 {
930 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
931 if (!lconst) return E_OUTOFMEMORY;
932
933 lconst->idx = ins.dst[0].reg.idx[0].offset;
934 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
935
936 list_add_head(&shader->constantsB, &lconst->entry);
937 reg_maps->local_bool_consts |= (1u << lconst->idx);
938 }
939 /* For subroutine prototypes. */
940 else if (ins.handler_idx == WINED3DSIH_LABEL)
941 {
942 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
943 }
944 /* Set texture, address, temporary registers. */
945 else
946 {
947 BOOL color0_mov = FALSE;
948 unsigned int i;
949
950 /* This will loop over all the registers and try to
951 * make a bitmask of the ones we're interested in.
952 *
953 * Relative addressing tokens are ignored, but that's
954 * okay, since we'll catch any address registers when
955 * they are initialized (required by spec). */
956 for (i = 0; i < ins.dst_count; ++i)
957 {
958 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
959 shader_version.type, constf_size))
960 return WINED3DERR_INVALIDCALL;
961
962 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
963 {
964 UINT idx = ins.dst[i].reg.idx[0].offset;
965
966 switch (ins.dst[i].reg.type)
967 {
968 case WINED3DSPR_RASTOUT:
969 if (shader_version.major >= 3)
970 break;
971 switch (idx)
972 {
973 case 0: /* oPos */
974 reg_maps->output_registers |= 1u << 10;
975 shader_signature_from_usage(&output_signature_elements[10],
976 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
977 break;
978
979 case 1: /* oFog */
980 reg_maps->output_registers |= 1u << 11;
981 shader_signature_from_usage(&output_signature_elements[11],
982 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
983 break;
984
985 case 2: /* oPts */
986 reg_maps->output_registers |= 1u << 11;
987 shader_signature_from_usage(&output_signature_elements[11],
988 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
989 break;
990 }
991 break;
992
993 case WINED3DSPR_ATTROUT:
994 if (shader_version.major >= 3)
995 break;
996 if (idx < 2)
997 {
998 idx += 8;
999 if (reg_maps->output_registers & (1u << idx))
1000 {
1001 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1002 }
1003 else
1004 {
1005 reg_maps->output_registers |= 1u << idx;
1006 shader_signature_from_usage(&output_signature_elements[idx],
1007 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1008 }
1009 }
1010 break;
1011
1012 case WINED3DSPR_TEXCRDOUT:
1013 if (shader_version.major >= 3)
1014 {
1015 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1016 break;
1017 }
1018 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1019 if (reg_maps->output_registers & (1u << idx))
1020 {
1021 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1022 }
1023 else
1024 {
1025 reg_maps->output_registers |= 1u << idx;
1026 shader_signature_from_usage(&output_signature_elements[idx],
1027 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1028 }
1029 break;
1030
1031 default:
1032 break;
1033 }
1034 }
1035
1036 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1037 {
1038 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1039 {
1040 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1041 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1042 * the mov and perform the sRGB write correction from the source register.
1043 *
1044 * However, if the mov is only partial, we can't do this, and if the write
1045 * comes from an instruction other than MOV it is hard to do as well. If
1046 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1047 shader->u.ps.color0_mov = FALSE;
1048 if (ins.handler_idx == WINED3DSIH_MOV
1049 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1050 {
1051 /* Used later when the source register is read. */
1052 color0_mov = TRUE;
1053 }
1054 }
1055 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1056 * end
1057 */
1058 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1059 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1060 {
1061 shader->u.ps.color0_mov = FALSE;
1062 }
1063 }
1064
1065 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1066 if (shader_version.major == 1
1067 && (ins.handler_idx == WINED3DSIH_TEX
1068 || ins.handler_idx == WINED3DSIH_TEXBEM
1069 || ins.handler_idx == WINED3DSIH_TEXBEML
1070 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1071 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1072 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1073 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1074 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1075 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1076 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1077 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1078 {
1079 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1080
1081 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1082 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1083 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1084 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1085
1086 /* texbem is only valid with < 1.4 pixel shaders */
1087 if (ins.handler_idx == WINED3DSIH_TEXBEM
1088 || ins.handler_idx == WINED3DSIH_TEXBEML)
1089 {
1090 reg_maps->bumpmat |= 1u << reg_idx;
1091 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1092 {
1093 reg_maps->luminanceparams |= 1u << reg_idx;
1094 }
1095 }
1096 }
1097 else if (ins.handler_idx == WINED3DSIH_BEM)
1098 {
1099 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1100 }
1101 }
1102
1103 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1104 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1105 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1106 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1107 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1108 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1109 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1110 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1111 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1112 else if (ins.handler_idx == WINED3DSIH_LOOP
1113 || ins.handler_idx == WINED3DSIH_REP)
1114 {
1115 ++cur_loop_depth;
1116 if (cur_loop_depth > max_loop_depth)
1117 max_loop_depth = cur_loop_depth;
1118 }
1119 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1120 || ins.handler_idx == WINED3DSIH_ENDREP)
1121 {
1122 --cur_loop_depth;
1123 }
1124 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1125 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1126 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1127 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1128 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1129 {
1130 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1131 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1132 }
1133 else if (ins.handler_idx == WINED3DSIH_LD
1134 || ins.handler_idx == WINED3DSIH_RESINFO)
1135 {
1136 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1137 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1138 }
1139
1140 if (ins.predicate)
1141 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1142 shader_version.type, constf_size))
1143 return WINED3DERR_INVALIDCALL;
1144
1145 for (i = 0; i < ins.src_count; ++i)
1146 {
1147 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1148 struct wined3d_shader_register reg = ins.src[i].reg;
1149
1150 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1151 shader_version.type, constf_size))
1152 return WINED3DERR_INVALIDCALL;
1153 while (count)
1154 {
1155 ++reg.idx[0].offset;
1156 if (!shader_record_register_usage(shader, reg_maps, &reg,
1157 shader_version.type, constf_size))
1158 return WINED3DERR_INVALIDCALL;
1159 --count;
1160 }
1161
1162 if (color0_mov)
1163 {
1164 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1165 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1166 {
1167 shader->u.ps.color0_mov = TRUE;
1168 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1169 }
1170 }
1171 }
1172 }
1173 }
1174 reg_maps->loop_depth = max_loop_depth;
1175
1176 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1177 * R0 is written to the render target. */
1178 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1179 reg_maps->rt_mask |= (1u << 0);
1180
1181 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1182
1183 if (input_signature->elements)
1184 {
1185 for (i = 0; i < input_signature->element_count; ++i)
1186 {
1187 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1188 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1189 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1190 reg_maps->vpos = 1;
1191 }
1192 }
1193 else if (!input_signature->elements && reg_maps->input_registers)
1194 {
1195 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1196 struct wined3d_shader_signature_element *e;
1197 unsigned int i;
1198
1199 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1200 return E_OUTOFMEMORY;
1201 input_signature->element_count = count;
1202
1203 e = input_signature->elements;
1204 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1205 {
1206 if (!(reg_maps->input_registers & (1u << i)))
1207 continue;
1208 input_signature_elements[i].register_idx = i;
1209 *e++ = input_signature_elements[i];
1210 }
1211 }
1212
1213 if (output_signature->elements)
1214 {
1215 for (i = 0; i < output_signature->element_count; ++i)
1216 {
1217 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1218 }
1219 }
1220 else if (reg_maps->output_registers)
1221 {
1222 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1223 struct wined3d_shader_signature_element *e;
1224
1225 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1226 return E_OUTOFMEMORY;
1227 output_signature->element_count = count;
1228
1229 e = output_signature->elements;
1230 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1231 {
1232 if (!(reg_maps->output_registers & (1u << i)))
1233 continue;
1234 *e++ = output_signature_elements[i];
1235 }
1236 }
1237
1238 return WINED3D_OK;
1239 }
1240
1241 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1242 {
1243 DWORD map = 1u << max;
1244 map |= map - 1;
1245 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1246
1247 return wined3d_log2i(map);
1248 }
1249
1250 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1251 {
1252 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1253 {
1254 shader_addline(buffer, "refactoringAllowed");
1255 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1256 if (global_flags)
1257 shader_addline(buffer, " | ");
1258 }
1259
1260 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1261 {
1262 shader_addline(buffer, "enableRawAndStructuredBuffers");
1263 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1264 }
1265
1266 if (global_flags)
1267 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1268 }
1269
1270 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1271 {
1272 unsigned int i;
1273
1274 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1275 {
1276 if (sysval_semantic_names[i].sysval_semantic == semantic)
1277 {
1278 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1279 return;
1280 }
1281 }
1282
1283 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1284 }
1285
1286 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1287 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1288 {
1289 shader_addline(buffer, "dcl");
1290
1291 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1292 {
1293 switch (semantic->resource_type)
1294 {
1295 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1296 shader_addline(buffer, "_2d");
1297 break;
1298
1299 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1300 shader_addline(buffer, "_3d");
1301 break;
1302
1303 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1304 shader_addline(buffer, "_cube");
1305 break;
1306
1307 default:
1308 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1309 break;
1310 }
1311 }
1312 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1313 {
1314 shader_addline(buffer, "_resource_");
1315 switch (semantic->resource_type)
1316 {
1317 case WINED3D_SHADER_RESOURCE_BUFFER:
1318 shader_addline(buffer, "buffer");
1319 break;
1320
1321 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1322 shader_addline(buffer, "texture1d");
1323 break;
1324
1325 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1326 shader_addline(buffer, "texture2d");
1327 break;
1328
1329 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1330 shader_addline(buffer, "texture2dms");
1331 break;
1332
1333 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1334 shader_addline(buffer, "texture3d");
1335 break;
1336
1337 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1338 shader_addline(buffer, "texturecube");
1339 break;
1340
1341 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1342 shader_addline(buffer, "texture1darray");
1343 break;
1344
1345 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1346 shader_addline(buffer, "texture2darray");
1347 break;
1348
1349 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1350 shader_addline(buffer, "texture2dmsarray");
1351 break;
1352
1353 default:
1354 shader_addline(buffer, "unknown");
1355 break;
1356 }
1357 switch (semantic->resource_data_type)
1358 {
1359 case WINED3D_DATA_FLOAT:
1360 shader_addline(buffer, " (float)");
1361 break;
1362
1363 case WINED3D_DATA_INT:
1364 shader_addline(buffer, " (int)");
1365 break;
1366
1367 case WINED3D_DATA_UINT:
1368 shader_addline(buffer, " (uint)");
1369 break;
1370
1371 case WINED3D_DATA_UNORM:
1372 shader_addline(buffer, " (unorm)");
1373 break;
1374
1375 case WINED3D_DATA_SNORM:
1376 shader_addline(buffer, " (snorm)");
1377 break;
1378
1379 default:
1380 shader_addline(buffer, " (unknown)");
1381 break;
1382 }
1383 }
1384 else
1385 {
1386 /* Pixel shaders 3.0 don't have usage semantics. */
1387 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1388 return;
1389 else
1390 shader_addline(buffer, "_");
1391
1392 switch (semantic->usage)
1393 {
1394 case WINED3D_DECL_USAGE_POSITION:
1395 shader_addline(buffer, "position%u", semantic->usage_idx);
1396 break;
1397
1398 case WINED3D_DECL_USAGE_BLEND_INDICES:
1399 shader_addline(buffer, "blend");
1400 break;
1401
1402 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1403 shader_addline(buffer, "weight");
1404 break;
1405
1406 case WINED3D_DECL_USAGE_NORMAL:
1407 shader_addline(buffer, "normal%u", semantic->usage_idx);
1408 break;
1409
1410 case WINED3D_DECL_USAGE_PSIZE:
1411 shader_addline(buffer, "psize");
1412 break;
1413
1414 case WINED3D_DECL_USAGE_COLOR:
1415 if (!semantic->usage_idx)
1416 shader_addline(buffer, "color");
1417 else
1418 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1419 break;
1420
1421 case WINED3D_DECL_USAGE_TEXCOORD:
1422 shader_addline(buffer, "texture%u", semantic->usage_idx);
1423 break;
1424
1425 case WINED3D_DECL_USAGE_TANGENT:
1426 shader_addline(buffer, "tangent");
1427 break;
1428
1429 case WINED3D_DECL_USAGE_BINORMAL:
1430 shader_addline(buffer, "binormal");
1431 break;
1432
1433 case WINED3D_DECL_USAGE_TESS_FACTOR:
1434 shader_addline(buffer, "tessfactor");
1435 break;
1436
1437 case WINED3D_DECL_USAGE_POSITIONT:
1438 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1439 break;
1440
1441 case WINED3D_DECL_USAGE_FOG:
1442 shader_addline(buffer, "fog");
1443 break;
1444
1445 case WINED3D_DECL_USAGE_DEPTH:
1446 shader_addline(buffer, "depth");
1447 break;
1448
1449 case WINED3D_DECL_USAGE_SAMPLE:
1450 shader_addline(buffer, "sample");
1451 break;
1452
1453 default:
1454 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1455 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1456 }
1457 }
1458 }
1459
1460 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1461 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1462 {
1463 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1464 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1465 UINT offset = reg->idx[0].offset;
1466
1467 switch (reg->type)
1468 {
1469 case WINED3DSPR_TEMP:
1470 shader_addline(buffer, "r");
1471 break;
1472
1473 case WINED3DSPR_INPUT:
1474 shader_addline(buffer, "v");
1475 break;
1476
1477 case WINED3DSPR_CONST:
1478 case WINED3DSPR_CONST2:
1479 case WINED3DSPR_CONST3:
1480 case WINED3DSPR_CONST4:
1481 shader_addline(buffer, "c");
1482 offset = shader_get_float_offset(reg->type, offset);
1483 break;
1484
1485 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1486 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1487 break;
1488
1489 case WINED3DSPR_RASTOUT:
1490 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1491 break;
1492
1493 case WINED3DSPR_COLOROUT:
1494 shader_addline(buffer, "oC");
1495 break;
1496
1497 case WINED3DSPR_DEPTHOUT:
1498 shader_addline(buffer, "oDepth");
1499 break;
1500
1501 case WINED3DSPR_ATTROUT:
1502 shader_addline(buffer, "oD");
1503 break;
1504
1505 case WINED3DSPR_TEXCRDOUT:
1506 /* Vertex shaders >= 3.0 use general purpose output registers
1507 * (WINED3DSPR_OUTPUT), which can include an address token. */
1508 if (shader_version->major >= 3)
1509 shader_addline(buffer, "o");
1510 else
1511 shader_addline(buffer, "oT");
1512 break;
1513
1514 case WINED3DSPR_CONSTINT:
1515 shader_addline(buffer, "i");
1516 break;
1517
1518 case WINED3DSPR_CONSTBOOL:
1519 shader_addline(buffer, "b");
1520 break;
1521
1522 case WINED3DSPR_LABEL:
1523 shader_addline(buffer, "l");
1524 break;
1525
1526 case WINED3DSPR_LOOP:
1527 shader_addline(buffer, "aL");
1528 break;
1529
1530 case WINED3DSPR_SAMPLER:
1531 shader_addline(buffer, "s");
1532 break;
1533
1534 case WINED3DSPR_MISCTYPE:
1535 if (offset > 1)
1536 {
1537 FIXME("Unhandled misctype register %u.\n", offset);
1538 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1539 }
1540 else
1541 {
1542 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1543 }
1544 break;
1545
1546 case WINED3DSPR_PREDICATE:
1547 shader_addline(buffer, "p");
1548 break;
1549
1550 case WINED3DSPR_IMMCONST:
1551 shader_addline(buffer, "l");
1552 break;
1553
1554 case WINED3DSPR_CONSTBUFFER:
1555 shader_addline(buffer, "cb");
1556 break;
1557
1558 case WINED3DSPR_IMMCONSTBUFFER:
1559 shader_addline(buffer, "icb");
1560 break;
1561
1562 case WINED3DSPR_PRIMID:
1563 shader_addline(buffer, "primID");
1564 break;
1565
1566 case WINED3DSPR_NULL:
1567 shader_addline(buffer, "null");
1568 break;
1569
1570 case WINED3DSPR_RESOURCE:
1571 shader_addline(buffer, "t");
1572 break;
1573
1574 default:
1575 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1576 break;
1577 }
1578
1579 if (reg->type == WINED3DSPR_IMMCONST)
1580 {
1581 shader_addline(buffer, "(");
1582 switch (reg->immconst_type)
1583 {
1584 case WINED3D_IMMCONST_SCALAR:
1585 switch (reg->data_type)
1586 {
1587 case WINED3D_DATA_FLOAT:
1588 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1589 break;
1590 case WINED3D_DATA_INT:
1591 shader_addline(buffer, "%d", reg->immconst_data[0]);
1592 break;
1593 case WINED3D_DATA_RESOURCE:
1594 case WINED3D_DATA_SAMPLER:
1595 case WINED3D_DATA_UINT:
1596 shader_addline(buffer, "%u", reg->immconst_data[0]);
1597 break;
1598 default:
1599 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1600 break;
1601 }
1602 break;
1603
1604 case WINED3D_IMMCONST_VEC4:
1605 switch (reg->data_type)
1606 {
1607 case WINED3D_DATA_FLOAT:
1608 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1609 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1610 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1611 break;
1612 case WINED3D_DATA_INT:
1613 shader_addline(buffer, "%d, %d, %d, %d",
1614 reg->immconst_data[0], reg->immconst_data[1],
1615 reg->immconst_data[2], reg->immconst_data[3]);
1616 break;
1617 case WINED3D_DATA_RESOURCE:
1618 case WINED3D_DATA_SAMPLER:
1619 case WINED3D_DATA_UINT:
1620 shader_addline(buffer, "%u, %u, %u, %u",
1621 reg->immconst_data[0], reg->immconst_data[1],
1622 reg->immconst_data[2], reg->immconst_data[3]);
1623 break;
1624 default:
1625 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1626 break;
1627 }
1628 break;
1629
1630 default:
1631 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1632 break;
1633 }
1634 shader_addline(buffer, ")");
1635 }
1636 else if (reg->type != WINED3DSPR_RASTOUT
1637 && reg->type != WINED3DSPR_MISCTYPE
1638 && reg->type != WINED3DSPR_NULL)
1639 {
1640 if (offset != ~0u)
1641 {
1642 shader_addline(buffer, "[");
1643 if (reg->idx[0].rel_addr)
1644 {
1645 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1646 shader_addline(buffer, " + ");
1647 }
1648 shader_addline(buffer, "%u]", offset);
1649
1650 if (reg->idx[1].offset != ~0u)
1651 {
1652 shader_addline(buffer, "[");
1653 if (reg->idx[1].rel_addr)
1654 {
1655 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1656 shader_addline(buffer, " + ");
1657 }
1658 shader_addline(buffer, "%u]", reg->idx[1].offset);
1659 }
1660 }
1661 }
1662 }
1663
1664 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1665 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1666 {
1667 DWORD write_mask = param->write_mask;
1668
1669 shader_dump_register(buffer, &param->reg, shader_version);
1670
1671 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1672 {
1673 static const char write_mask_chars[] = "xyzw";
1674
1675 shader_addline(buffer, ".");
1676 if (write_mask & WINED3DSP_WRITEMASK_0)
1677 shader_addline(buffer, "%c", write_mask_chars[0]);
1678 if (write_mask & WINED3DSP_WRITEMASK_1)
1679 shader_addline(buffer, "%c", write_mask_chars[1]);
1680 if (write_mask & WINED3DSP_WRITEMASK_2)
1681 shader_addline(buffer, "%c", write_mask_chars[2]);
1682 if (write_mask & WINED3DSP_WRITEMASK_3)
1683 shader_addline(buffer, "%c", write_mask_chars[3]);
1684 }
1685 }
1686
1687 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1688 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1689 {
1690 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1691 DWORD swizzle = param->swizzle;
1692
1693 if (src_modifier == WINED3DSPSM_NEG
1694 || src_modifier == WINED3DSPSM_BIASNEG
1695 || src_modifier == WINED3DSPSM_SIGNNEG
1696 || src_modifier == WINED3DSPSM_X2NEG
1697 || src_modifier == WINED3DSPSM_ABSNEG)
1698 shader_addline(buffer, "-");
1699 else if (src_modifier == WINED3DSPSM_COMP)
1700 shader_addline(buffer, "1-");
1701 else if (src_modifier == WINED3DSPSM_NOT)
1702 shader_addline(buffer, "!");
1703
1704 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1705 shader_addline(buffer, "abs(");
1706
1707 shader_dump_register(buffer, &param->reg, shader_version);
1708
1709 switch (src_modifier)
1710 {
1711 case WINED3DSPSM_NONE: break;
1712 case WINED3DSPSM_NEG: break;
1713 case WINED3DSPSM_NOT: break;
1714 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1715 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1716 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1717 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1718 case WINED3DSPSM_COMP: break;
1719 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1720 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1721 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1722 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1723 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1724 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1725 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1726 }
1727
1728 if (swizzle != WINED3DSP_NOSWIZZLE)
1729 {
1730 static const char swizzle_chars[] = "xyzw";
1731 DWORD swizzle_x = swizzle & 0x03;
1732 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1733 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1734 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1735
1736 if (swizzle_x == swizzle_y
1737 && swizzle_x == swizzle_z
1738 && swizzle_x == swizzle_w)
1739 {
1740 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
1741 }
1742 else
1743 {
1744 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1745 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1746 }
1747 }
1748 }
1749
1750 /* Shared code in order to generate the bulk of the shader string.
1751 * NOTE: A description of how to parse tokens can be found on MSDN. */
1752 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1753 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1754 {
1755 struct wined3d_device *device = shader->device;
1756 const struct wined3d_shader_frontend *fe = shader->frontend;
1757 void *fe_data = shader->frontend_data;
1758 struct wined3d_shader_version shader_version;
1759 struct wined3d_shader_loop_state loop_state;
1760 struct wined3d_shader_instruction ins;
1761 struct wined3d_shader_tex_mx tex_mx;
1762 struct wined3d_shader_context ctx;
1763 const DWORD *ptr = byte_code;
1764
1765 /* Initialize current parsing state. */
1766 tex_mx.current_row = 0;
1767 loop_state.current_depth = 0;
1768 loop_state.current_reg = 0;
1769
1770 ctx.shader = shader;
1771 ctx.gl_info = &device->adapter->gl_info;
1772 ctx.reg_maps = reg_maps;
1773 ctx.buffer = buffer;
1774 ctx.tex_mx = &tex_mx;
1775 ctx.loop_state = &loop_state;
1776 ctx.backend_data = backend_ctx;
1777 ins.ctx = &ctx;
1778
1779 fe->shader_read_header(fe_data, &ptr, &shader_version);
1780
1781 while (!fe->shader_is_end(fe_data, &ptr))
1782 {
1783 /* Read opcode. */
1784 fe->shader_read_instruction(fe_data, &ptr, &ins);
1785
1786 /* Unknown opcode and its parameters. */
1787 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1788 {
1789 TRACE("Skipping unrecognized instruction.\n");
1790 continue;
1791 }
1792
1793 if (ins.predicate)
1794 FIXME("Predicates not implemented.\n");
1795
1796 /* Call appropriate function for output target */
1797 device->shader_backend->shader_handle_instruction(&ins);
1798 }
1799 }
1800
1801 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
1802 const struct wined3d_shader_dst_param *dst)
1803 {
1804 DWORD mmask = dst->modifiers;
1805
1806 switch (dst->shift)
1807 {
1808 case 0: break;
1809 case 13: shader_addline(buffer, "_d8"); break;
1810 case 14: shader_addline(buffer, "_d4"); break;
1811 case 15: shader_addline(buffer, "_d2"); break;
1812 case 1: shader_addline(buffer, "_x2"); break;
1813 case 2: shader_addline(buffer, "_x4"); break;
1814 case 3: shader_addline(buffer, "_x8"); break;
1815 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
1816 }
1817
1818 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
1819 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
1820 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
1821
1822 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1823 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1824 }
1825
1826 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
1827 enum wined3d_primitive_type primitive_type)
1828 {
1829 switch (primitive_type)
1830 {
1831 case WINED3D_PT_UNDEFINED:
1832 shader_addline(buffer, "undefined");
1833 break;
1834 case WINED3D_PT_POINTLIST:
1835 shader_addline(buffer, "pointlist");
1836 break;
1837 case WINED3D_PT_LINELIST:
1838 shader_addline(buffer, "linelist");
1839 break;
1840 case WINED3D_PT_LINESTRIP:
1841 shader_addline(buffer, "linestrip");
1842 break;
1843 case WINED3D_PT_TRIANGLELIST:
1844 shader_addline(buffer, "trianglelist");
1845 break;
1846 case WINED3D_PT_TRIANGLESTRIP:
1847 shader_addline(buffer, "trianglestrip");
1848 break;
1849 case WINED3D_PT_TRIANGLEFAN:
1850 shader_addline(buffer, "trianglefan");
1851 break;
1852 case WINED3D_PT_LINELIST_ADJ:
1853 shader_addline(buffer, "linelist_adj");
1854 break;
1855 case WINED3D_PT_LINESTRIP_ADJ:
1856 shader_addline(buffer, "linestrip_adj");
1857 break;
1858 case WINED3D_PT_TRIANGLELIST_ADJ:
1859 shader_addline(buffer, "trianglelist_adj");
1860 break;
1861 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1862 shader_addline(buffer, "trianglestrip_adj");
1863 break;
1864 default:
1865 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
1866 break;
1867 }
1868 }
1869
1870 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
1871 enum wined3d_shader_interpolation_mode interpolation_mode)
1872 {
1873 switch (interpolation_mode)
1874 {
1875 case WINED3DSIM_CONSTANT:
1876 shader_addline(buffer, "constant");
1877 break;
1878 case WINED3DSIM_LINEAR:
1879 shader_addline(buffer, "linear");
1880 break;
1881 case WINED3DSIM_LINEAR_CENTROID:
1882 shader_addline(buffer, "linear centroid");
1883 break;
1884 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
1885 shader_addline(buffer, "linear noperspective");
1886 break;
1887 case WINED3DSIM_LINEAR_SAMPLE:
1888 shader_addline(buffer, "linear sample");
1889 break;
1890 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
1891 shader_addline(buffer, "linear noperspective centroid");
1892 break;
1893 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
1894 shader_addline(buffer, "linear noperspective sample");
1895 break;
1896 default:
1897 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
1898 break;
1899 }
1900 }
1901
1902 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1903 {
1904 struct wined3d_shader_version shader_version;
1905 struct wined3d_string_buffer buffer;
1906 const DWORD *ptr = byte_code;
1907 const char *type_prefix;
1908 const char *p, *q;
1909 DWORD i;
1910
1911 if (!string_buffer_init(&buffer))
1912 {
1913 ERR("Failed to initialize string buffer.\n");
1914 return;
1915 }
1916
1917 TRACE("Parsing %p.\n", byte_code);
1918
1919 fe->shader_read_header(fe_data, &ptr, &shader_version);
1920
1921 switch (shader_version.type)
1922 {
1923 case WINED3D_SHADER_TYPE_VERTEX:
1924 type_prefix = "vs";
1925 break;
1926
1927 case WINED3D_SHADER_TYPE_GEOMETRY:
1928 type_prefix = "gs";
1929 break;
1930
1931 case WINED3D_SHADER_TYPE_PIXEL:
1932 type_prefix = "ps";
1933 break;
1934
1935 default:
1936 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1937 type_prefix = "unknown";
1938 break;
1939 }
1940
1941 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1942
1943 while (!fe->shader_is_end(fe_data, &ptr))
1944 {
1945 struct wined3d_shader_instruction ins;
1946
1947 fe->shader_read_instruction(fe_data, &ptr, &ins);
1948 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1949 {
1950 WARN("Skipping unrecognized instruction.\n");
1951 shader_addline(&buffer, "<unrecognized instruction>\n");
1952 continue;
1953 }
1954
1955 if (ins.handler_idx == WINED3DSIH_DCL)
1956 {
1957 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
1958 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
1959 shader_addline(&buffer, " ");
1960 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
1961 }
1962 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1963 {
1964 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1965 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
1966 shader_addline(&buffer, ", %s",
1967 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1968 }
1969 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1970 {
1971 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1972 shader_dump_global_flags(&buffer, ins.flags);
1973 }
1974 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1975 {
1976 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
1977 for (i = 0; i < ins.declaration.icb->element_count / 4; ++i)
1978 {
1979 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
1980 ins.declaration.icb->data[4 * i + 0],
1981 ins.declaration.icb->data[4 * i + 1],
1982 ins.declaration.icb->data[4 * i + 2],
1983 ins.declaration.icb->data[4 * i + 3]);
1984 }
1985 shader_addline(&buffer, "}");
1986 }
1987 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1988 {
1989 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
1990 shader_dump_interpolation_mode(&buffer, ins.flags);
1991 shader_addline(&buffer, " ");
1992 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
1993 }
1994 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
1995 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
1996 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
1997 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
1998 {
1999 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2000 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2001 shader_addline(&buffer, ", ");
2002 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2003 }
2004 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2005 {
2006 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2007 shader_dump_interpolation_mode(&buffer, ins.flags);
2008 shader_addline(&buffer, " ");
2009 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2010 shader_addline(&buffer, ", ");
2011 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2012 }
2013 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2014 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2015 {
2016 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2017 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2018 }
2019 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2020 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2021 {
2022 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2023 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2024 }
2025 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2026 {
2027 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2028 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2029 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2030 shader_addline(&buffer, ", comparisonMode");
2031 }
2032 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2033 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2034 {
2035 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2036 }
2037 else if (ins.handler_idx == WINED3DSIH_DEF)
2038 {
2039 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2040 ins.dst[0].reg.idx[0].offset),
2041 *(const float *)&ins.src[0].reg.immconst_data[0],
2042 *(const float *)&ins.src[0].reg.immconst_data[1],
2043 *(const float *)&ins.src[0].reg.immconst_data[2],
2044 *(const float *)&ins.src[0].reg.immconst_data[3]);
2045 }
2046 else if (ins.handler_idx == WINED3DSIH_DEFI)
2047 {
2048 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2049 ins.src[0].reg.immconst_data[0],
2050 ins.src[0].reg.immconst_data[1],
2051 ins.src[0].reg.immconst_data[2],
2052 ins.src[0].reg.immconst_data[3]);
2053 }
2054 else if (ins.handler_idx == WINED3DSIH_DEFB)
2055 {
2056 shader_addline(&buffer, "defb b%u = %s",
2057 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2058 }
2059 else
2060 {
2061 if (ins.predicate)
2062 {
2063 shader_addline(&buffer, "(");
2064 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2065 shader_addline(&buffer, ") ");
2066 }
2067
2068 /* PixWin marks instructions with the coissue flag with a '+' */
2069 if (ins.coissue)
2070 shader_addline(&buffer, "+");
2071
2072 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2073
2074 if (ins.handler_idx == WINED3DSIH_IFC
2075 || ins.handler_idx == WINED3DSIH_BREAKC)
2076 {
2077 switch (ins.flags)
2078 {
2079 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2080 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2081 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2082 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2083 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2084 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2085 default: shader_addline(&buffer, "_(%u)", ins.flags);
2086 }
2087 }
2088 else if (ins.handler_idx == WINED3DSIH_TEX
2089 && shader_version.major >= 2
2090 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2091 {
2092 shader_addline(&buffer, "p");
2093 }
2094 else if (ins.handler_idx == WINED3DSIH_RESINFO
2095 && ins.flags)
2096 {
2097 switch (ins.flags)
2098 {
2099 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2100 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2101 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2102 }
2103 }
2104
2105 if (wined3d_shader_instruction_has_texel_offset(&ins))
2106 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2107
2108 for (i = 0; i < ins.dst_count; ++i)
2109 {
2110 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2111 shader_addline(&buffer, !i ? " " : ", ");
2112 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2113 }
2114
2115 /* Other source tokens */
2116 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2117 {
2118 shader_addline(&buffer, !i ? " " : ", ");
2119 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2120 }
2121 }
2122 shader_addline(&buffer, "\n");
2123 }
2124
2125 for (p = buffer.buffer; *p; p = q)
2126 {
2127 if (!(q = strstr(p, "\n")))
2128 q = p + strlen(p);
2129 else
2130 ++q;
2131 TRACE(" %.*s", (int)(q - p), p);
2132 }
2133
2134 string_buffer_free(&buffer);
2135 }
2136
2137 static void shader_cleanup(struct wined3d_shader *shader)
2138 {
2139 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2140 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2141 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2142 shader->device->shader_backend->shader_destroy(shader);
2143 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
2144 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
2145 HeapFree(GetProcessHeap(), 0, shader->function);
2146 shader_delete_constant_list(&shader->constantsF);
2147 shader_delete_constant_list(&shader->constantsB);
2148 shader_delete_constant_list(&shader->constantsI);
2149 list_remove(&shader->shader_list_entry);
2150
2151 if (shader->frontend && shader->frontend_data)
2152 shader->frontend->shader_free(shader->frontend_data);
2153 }
2154
2155 struct shader_none_priv
2156 {
2157 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2158 const struct fragment_pipeline *fragment_pipe;
2159 BOOL ffp_proj_control;
2160 };
2161
2162 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2163 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2164 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2165 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2166 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2167 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2168 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2169 const struct wined3d_state *state) {}
2170 static void shader_none_destroy(struct wined3d_shader *shader) {}
2171 static void shader_none_free_context_data(struct wined3d_context *context) {}
2172 static void shader_none_init_context_state(struct wined3d_context *context) {}
2173
2174 /* Context activation is done by the caller. */
2175 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2176 const struct wined3d_state *state)
2177 {
2178 const struct wined3d_gl_info *gl_info = context->gl_info;
2179 struct shader_none_priv *priv = shader_priv;
2180
2181 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2182 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2183 }
2184
2185 /* Context activation is done by the caller. */
2186 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2187 {
2188 struct shader_none_priv *priv = shader_priv;
2189 const struct wined3d_gl_info *gl_info = context->gl_info;
2190
2191 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2192 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2193
2194 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2195 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2196 | (1u << WINED3D_SHADER_TYPE_GEOMETRY);
2197 }
2198
2199 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2200 const struct fragment_pipeline *fragment_pipe)
2201 {
2202 struct fragment_caps fragment_caps;
2203 void *vertex_priv, *fragment_priv;
2204 struct shader_none_priv *priv;
2205
2206 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2207 return E_OUTOFMEMORY;
2208
2209 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2210 {
2211 ERR("Failed to initialize vertex pipe.\n");
2212 HeapFree(GetProcessHeap(), 0, priv);
2213 return E_FAIL;
2214 }
2215
2216 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2217 {
2218 ERR("Failed to initialize fragment pipe.\n");
2219 vertex_pipe->vp_free(device);
2220 HeapFree(GetProcessHeap(), 0, priv);
2221 return E_FAIL;
2222 }
2223
2224 priv->vertex_pipe = vertex_pipe;
2225 priv->fragment_pipe = fragment_pipe;
2226 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2227 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2228
2229 device->vertex_priv = vertex_priv;
2230 device->fragment_priv = fragment_priv;
2231 device->shader_priv = priv;
2232
2233 return WINED3D_OK;
2234 }
2235
2236 static void shader_none_free(struct wined3d_device *device)
2237 {
2238 struct shader_none_priv *priv = device->shader_priv;
2239
2240 priv->fragment_pipe->free_private(device);
2241 priv->vertex_pipe->vp_free(device);
2242 HeapFree(GetProcessHeap(), 0, priv);
2243 }
2244
2245 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2246 {
2247 return TRUE;
2248 }
2249
2250 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2251 {
2252 /* Set the shader caps to 0 for the none shader backend */
2253 caps->vs_version = 0;
2254 caps->gs_version = 0;
2255 caps->ps_version = 0;
2256 caps->vs_uniform_count = 0;
2257 caps->ps_uniform_count = 0;
2258 caps->ps_1x_max_value = 0.0f;
2259 caps->varying_count = 0;
2260 caps->wined3d_caps = 0;
2261 }
2262
2263 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2264 {
2265 /* We "support" every possible fixup, since we don't support any shader
2266 * model, and will never have to actually sample a texture. */
2267 return TRUE;
2268 }
2269
2270 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2271 {
2272 struct shader_none_priv *priv = shader_priv;
2273
2274 return priv->ffp_proj_control;
2275 }
2276
2277 const struct wined3d_shader_backend_ops none_shader_backend =
2278 {
2279 shader_none_handle_instruction,
2280 shader_none_select,
2281 shader_none_disable,
2282 shader_none_select_depth_blt,
2283 shader_none_deselect_depth_blt,
2284 shader_none_update_float_vertex_constants,
2285 shader_none_update_float_pixel_constants,
2286 shader_none_load_constants,
2287 shader_none_destroy,
2288 shader_none_alloc,
2289 shader_none_free,
2290 shader_none_allocate_context_data,
2291 shader_none_free_context_data,
2292 shader_none_init_context_state,
2293 shader_none_get_caps,
2294 shader_none_color_fixup_supported,
2295 shader_none_has_ffp_proj_control,
2296 };
2297
2298 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2299 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2300 enum wined3d_shader_type type, unsigned int max_version)
2301 {
2302 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2303 const struct wined3d_shader_frontend *fe;
2304 HRESULT hr;
2305 unsigned int backend_version;
2306 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2307
2308 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u, type %#x, max_version %u.\n",
2309 shader, byte_code, output_signature, float_const_count, type, max_version);
2310
2311 list_init(&shader->constantsF);
2312 list_init(&shader->constantsB);
2313 list_init(&shader->constantsI);
2314 shader->lconst_inf_or_nan = FALSE;
2315
2316 fe = shader_select_frontend(*byte_code);
2317 if (!fe)
2318 {
2319 FIXME("Unable to find frontend for shader.\n");
2320 return WINED3DERR_INVALIDCALL;
2321 }
2322 shader->frontend = fe;
2323 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2324 if (!shader->frontend_data)
2325 {
2326 FIXME("Failed to initialize frontend.\n");
2327 return WINED3DERR_INVALIDCALL;
2328 }
2329
2330 /* First pass: trace shader. */
2331 if (TRACE_ON(d3d_shader))
2332 shader_trace_init(fe, shader->frontend_data, byte_code);
2333
2334 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2335 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2336 &shader->output_signature, byte_code, float_const_count)))
2337 return hr;
2338
2339 if (reg_maps->shader_version.type != type)
2340 {
2341 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2342 return WINED3DERR_INVALIDCALL;
2343 }
2344 if (reg_maps->shader_version.major > max_version)
2345 {
2346 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2347 return WINED3DERR_INVALIDCALL;
2348 }
2349 switch (type)
2350 {
2351 case WINED3D_SHADER_TYPE_VERTEX:
2352 backend_version = d3d_info->limits.vs_version;
2353 break;
2354 case WINED3D_SHADER_TYPE_GEOMETRY:
2355 backend_version = d3d_info->limits.gs_version;
2356 break;
2357 case WINED3D_SHADER_TYPE_PIXEL:
2358 backend_version = d3d_info->limits.ps_version;
2359 break;
2360 default:
2361 FIXME("No backend version-checking for this shader type\n");
2362 backend_version = 0;
2363 }
2364 if (reg_maps->shader_version.major > backend_version)
2365 {
2366 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2367 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2368 return WINED3DERR_INVALIDCALL;
2369 }
2370
2371 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2372 if (!shader->function)
2373 return E_OUTOFMEMORY;
2374 memcpy(shader->function, byte_code, shader->functionLength);
2375
2376 return WINED3D_OK;
2377 }
2378
2379 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2380 {
2381 ULONG refcount = InterlockedIncrement(&shader->ref);
2382
2383 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2384
2385 return refcount;
2386 }
2387
2388 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2389 {
2390 ULONG refcount = InterlockedDecrement(&shader->ref);
2391
2392 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2393
2394 if (!refcount)
2395 {
2396 shader_cleanup(shader);
2397 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2398 HeapFree(GetProcessHeap(), 0, shader);
2399 }
2400
2401 return refcount;
2402 }
2403
2404 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2405 {
2406 TRACE("shader %p.\n", shader);
2407
2408 return shader->parent;
2409 }
2410
2411 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2412 void *byte_code, UINT *byte_code_size)
2413 {
2414 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2415
2416 if (!byte_code)
2417 {
2418 *byte_code_size = shader->functionLength;
2419 return WINED3D_OK;
2420 }
2421
2422 if (*byte_code_size < shader->functionLength)
2423 {
2424 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2425 * than the required size we should write the required size and
2426 * return D3DERR_MOREDATA. That's not actually true. */
2427 return WINED3DERR_INVALIDCALL;
2428 }
2429
2430 memcpy(byte_code, shader->function, shader->functionLength);
2431
2432 return WINED3D_OK;
2433 }
2434
2435 /* Set local constants for d3d8 shaders. */
2436 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2437 UINT start_idx, const float *src_data, UINT count)
2438 {
2439 UINT end_idx = start_idx + count;
2440 UINT i;
2441
2442 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2443
2444 if (end_idx > shader->limits->constant_float)
2445 {
2446 WARN("end_idx %u > float constants limit %u.\n",
2447 end_idx, shader->limits->constant_float);
2448 end_idx = shader->limits->constant_float;
2449 }
2450
2451 for (i = start_idx; i < end_idx; ++i)
2452 {
2453 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2454 float *value;
2455 if (!lconst)
2456 return E_OUTOFMEMORY;
2457
2458 lconst->idx = i;
2459 value = (float *)lconst->value;
2460 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2461 list_add_head(&shader->constantsF, &lconst->entry);
2462
2463 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2464 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2465 {
2466 shader->lconst_inf_or_nan = TRUE;
2467 }
2468 }
2469
2470 return WINED3D_OK;
2471 }
2472
2473 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2474 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2475 {
2476 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2477 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2478 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2479 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2480 args->point_size = state->gl_primitive_type == GL_POINTS;
2481 args->per_vertex_point_size = shader->reg_maps.point_size;
2482 args->swizzle_map = swizzle_map;
2483 if (d3d_info->emulated_flatshading)
2484 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2485 else
2486 args->flatshading = FALSE;
2487 }
2488
2489 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2490 {
2491 if (usage_idx1 != usage_idx2)
2492 return FALSE;
2493 if (usage1 == usage2)
2494 return TRUE;
2495 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2496 return TRUE;
2497 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2498 return TRUE;
2499
2500 return FALSE;
2501 }
2502
2503 BOOL vshader_get_input(const struct wined3d_shader *shader,
2504 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2505 {
2506 WORD map = shader->reg_maps.input_registers;
2507 unsigned int i;
2508
2509 for (i = 0; map; map >>= 1, ++i)
2510 {
2511 if (!(map & 1)) continue;
2512
2513 if (match_usage(shader->u.vs.attributes[i].usage,
2514 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2515 {
2516 *regnum = i;
2517 return TRUE;
2518 }
2519 }
2520 return FALSE;
2521 }
2522
2523 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2524 const struct wined3d_shader_signature *src, char **signature_strings)
2525 {
2526 struct wined3d_shader_signature_element *e;
2527 unsigned int i;
2528 SIZE_T len;
2529 char *ptr;
2530
2531 ptr = *signature_strings;
2532
2533 dst->element_count = src->element_count;
2534 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2535 return E_OUTOFMEMORY;
2536
2537 for (i = 0; i < src->element_count; ++i)
2538 {
2539 e = &src->elements[i];
2540 dst->elements[i] = *e;
2541
2542 len = strlen(e->semantic_name);
2543 memcpy(ptr, e->semantic_name, len + 1);
2544 dst->elements[i].semantic_name = ptr;
2545 ptr += len + 1;
2546 }
2547
2548 *signature_strings = ptr;
2549
2550 return WINED3D_OK;
2551 }
2552
2553 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2554 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2555 void *parent, const struct wined3d_parent_ops *parent_ops)
2556 {
2557 struct wined3d_shader_signature_element *e;
2558 SIZE_T total, len;
2559 unsigned int i;
2560 HRESULT hr;
2561 char *ptr;
2562
2563 if (!desc->byte_code)
2564 return WINED3DERR_INVALIDCALL;
2565
2566 shader->ref = 1;
2567 shader->device = device;
2568 shader->parent = parent;
2569 shader->parent_ops = parent_ops;
2570
2571 total = 0;
2572 if (desc->input_signature)
2573 {
2574 for (i = 0; i < desc->input_signature->element_count; ++i)
2575 {
2576 e = &desc->input_signature->elements[i];
2577 len = strlen(e->semantic_name);
2578 if (len >= ~(SIZE_T)0 - total)
2579 return E_OUTOFMEMORY;
2580
2581 total += len + 1;
2582 }
2583 }
2584 if (desc->output_signature)
2585 {
2586 for (i = 0; i < desc->output_signature->element_count; ++i)
2587 {
2588 e = &desc->output_signature->elements[i];
2589 len = strlen(e->semantic_name);
2590 if (len >= ~(SIZE_T)0 - total)
2591 return E_OUTOFMEMORY;
2592
2593 total += len + 1;
2594 }
2595 }
2596 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2597 return E_OUTOFMEMORY;
2598 ptr = shader->signature_strings;
2599
2600 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2601 desc->input_signature, &ptr)))
2602 {
2603 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2604 return hr;
2605 }
2606 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2607 desc->output_signature, &ptr)))
2608 {
2609 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2610 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2611 return hr;
2612 }
2613
2614 list_init(&shader->linked_programs);
2615 list_add_head(&device->shaders, &shader->shader_list_entry);
2616
2617 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2618 float_const_count, type, desc->max_version)))
2619 {
2620 WARN("Failed to set function, hr %#x.\n", hr);
2621 shader_cleanup(shader);
2622 }
2623
2624 return hr;
2625 }
2626
2627 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2628 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2629 {
2630 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2631 unsigned int i;
2632 HRESULT hr;
2633
2634 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2635 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2636 return hr;
2637
2638 for (i = 0; i < shader->input_signature.element_count; ++i)
2639 {
2640 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2641
2642 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2643 continue;
2644
2645 shader->u.vs.attributes[input->register_idx].usage =
2646 shader_usage_from_semantic_name(input->semantic_name);
2647 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2648 }
2649
2650 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2651 shader->lconst_inf_or_nan;
2652
2653 return WINED3D_OK;
2654 }
2655
2656 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2657 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2658 {
2659 HRESULT hr;
2660
2661 if (FAILED(hr = shader_init(shader, device, desc, 0,
2662 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2663 return hr;
2664
2665 shader->load_local_constsF = shader->lconst_inf_or_nan;
2666
2667 return WINED3D_OK;
2668 }
2669
2670 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2671 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2672 {
2673 const struct wined3d_gl_info *gl_info = context->gl_info;
2674 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2675 const struct wined3d_texture *texture;
2676 UINT i;
2677
2678 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2679 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
2680 {
2681 static unsigned int warned = 0;
2682
2683 args->srgb_correction = 1;
2684 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2685 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2686 "support, expect rendering artifacts.\n");
2687 }
2688
2689 if (shader->reg_maps.shader_version.major == 1
2690 && shader->reg_maps.shader_version.minor <= 3)
2691 {
2692 for (i = 0; i < shader->limits->sampler; ++i)
2693 {
2694 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2695
2696 if (flags & WINED3D_TTFF_PROJECTED)
2697 {
2698 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2699
2700 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2701 {
2702 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2703 unsigned int j;
2704 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2705 DWORD max_valid = WINED3D_TTFF_COUNT4;
2706
2707 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2708 {
2709 struct wined3d_vertex_declaration_element *element =
2710 &state->vertex_declaration->elements[j];
2711
2712 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2713 && element->usage_idx == index)
2714 {
2715 max_valid = element->format->component_count;
2716 break;
2717 }
2718 }
2719 if (!tex_transform || tex_transform > max_valid)
2720 {
2721 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2722 tex_transform, max_valid);
2723 tex_transform = max_valid;
2724 }
2725 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2726 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2727 && tex_transform > WINED3D_TTFF_COUNT2)
2728 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2729 && tex_transform > WINED3D_TTFF_COUNT3))
2730 tex_transform |= WINED3D_PSARGS_PROJECTED;
2731 else
2732 {
2733 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2734 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2735 i, tex_transform, resource_type);
2736 }
2737 }
2738 else
2739 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2740
2741 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2742 }
2743 }
2744 }
2745 if (shader->reg_maps.shader_version.major == 1
2746 && shader->reg_maps.shader_version.minor <= 4)
2747 {
2748 for (i = 0; i < shader->limits->sampler; ++i)
2749 {
2750 const struct wined3d_texture *texture = state->textures[i];
2751
2752 if (!shader->reg_maps.resource_info[i].type)
2753 continue;
2754
2755 /* Treat unbound textures as 2D. The dummy texture will provide
2756 * the proper sample value. The tex_types bitmap defaults to
2757 * 2D because of the memset. */
2758 if (!texture)
2759 continue;
2760
2761 switch (texture->target)
2762 {
2763 /* RECT textures are distinguished from 2D textures via np2_fixup */
2764 case GL_TEXTURE_RECTANGLE_ARB:
2765 case GL_TEXTURE_2D:
2766 break;
2767
2768 case GL_TEXTURE_3D:
2769 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2770 break;
2771
2772 case GL_TEXTURE_CUBE_MAP_ARB:
2773 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2774 break;
2775 }
2776 }
2777 }
2778
2779 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2780 {
2781 if (!shader->reg_maps.resource_info[i].type)
2782 continue;
2783
2784 texture = state->textures[i];
2785 if (!texture)
2786 {
2787 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2788 continue;
2789 }
2790 if (can_use_texture_swizzle(gl_info, texture->resource.format))
2791 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2792 else
2793 args->color_fixup[i] = texture->resource.format->color_fixup;
2794
2795 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
2796 args->shadow |= 1u << i;
2797
2798 /* Flag samplers that need NP2 texcoord fixup. */
2799 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2800 args->np2_fixup |= (1u << i);
2801 }
2802
2803 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
2804 if (shader->reg_maps.shader_version.major >= 4)
2805 args->shadow = 0;
2806
2807 if (shader->reg_maps.shader_version.major >= 3)
2808 {
2809 if (position_transformed)
2810 args->vp_mode = pretransformed;
2811 else if (use_vs(state))
2812 args->vp_mode = vertexshader;
2813 else
2814 args->vp_mode = fixedfunction;
2815 args->fog = WINED3D_FFP_PS_FOG_OFF;
2816 }
2817 else
2818 {
2819 args->vp_mode = vertexshader;
2820 if (state->render_states[WINED3D_RS_FOGENABLE])
2821 {
2822 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2823 {
2824 case WINED3D_FOG_NONE:
2825 if (position_transformed || use_vs(state))
2826 {
2827 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2828 break;
2829 }
2830
2831 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2832 {
2833 case WINED3D_FOG_NONE: /* Fall through. */
2834 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2835 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2836 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2837 }
2838 break;
2839
2840 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2841 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2842 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2843 }
2844 }
2845 else
2846 {
2847 args->fog = WINED3D_FFP_PS_FOG_OFF;
2848 }
2849 }
2850
2851 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
2852 {
2853 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
2854
2855 args->texcoords_initialized = 0;
2856 for (i = 0; i < MAX_TEXTURES; ++i)
2857 {
2858 if (vs)
2859 {
2860 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
2861 args->texcoords_initialized |= 1u << i;
2862 }
2863 else
2864 {
2865 const struct wined3d_stream_info *si = &context->stream_info;
2866 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2867
2868 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
2869 & WINED3D_FFP_TCI_MASK
2870 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
2871 args->texcoords_initialized |= 1u << i;
2872 }
2873 }
2874 }
2875 else
2876 {
2877 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
2878 }
2879
2880 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
2881 && state->gl_primitive_type == GL_POINTS;
2882
2883 if (d3d_info->emulated_flatshading)
2884 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2885 }
2886
2887 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2888 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2889 {
2890 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2891 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2892 HRESULT hr;
2893
2894 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
2895 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
2896 return hr;
2897
2898 for (i = 0; i < MAX_REG_INPUT; ++i)
2899 {
2900 if (shader->u.ps.input_reg_used[i])
2901 {
2902 ++num_regs_used;
2903 highest_reg_used = i;
2904 }
2905 }
2906
2907 /* Don't do any register mapping magic if it is not needed, or if we can't
2908 * achieve anything anyway */
2909 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2910 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2911 {
2912 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2913 {
2914 /* This happens with relative addressing. The input mapper function
2915 * warns about this if the higher registers are declared too, so
2916 * don't write a FIXME here */
2917 WARN("More varying registers used than supported\n");
2918 }
2919
2920 for (i = 0; i < MAX_REG_INPUT; ++i)
2921 {
2922 shader->u.ps.input_reg_map[i] = i;
2923 }
2924
2925 shader->u.ps.declared_in_count = highest_reg_used + 1;
2926 }
2927 else
2928 {
2929 shader->u.ps.declared_in_count = 0;
2930 for (i = 0; i < MAX_REG_INPUT; ++i)
2931 {
2932 if (shader->u.ps.input_reg_used[i])
2933 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2934 else shader->u.ps.input_reg_map[i] = ~0U;
2935 }
2936 }
2937
2938 shader->load_local_constsF = shader->lconst_inf_or_nan;
2939
2940 return WINED3D_OK;
2941 }
2942
2943 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
2944 {
2945 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2946 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
2947 unsigned int i;
2948
2949 if (reg_maps->shader_version.major != 1) return;
2950
2951 for (i = 0; i < shader->limits->sampler; ++i)
2952 {
2953 /* We don't sample from this sampler. */
2954 if (!resource_info[i].type)
2955 continue;
2956
2957 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
2958 {
2959 case WINED3D_SHADER_TEX_2D:
2960 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
2961 break;
2962
2963 case WINED3D_SHADER_TEX_3D:
2964 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
2965 break;
2966
2967 case WINED3D_SHADER_TEX_CUBE:
2968 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
2969 break;
2970 }
2971 }
2972 }
2973
2974 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
2975 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
2976 {
2977 struct wined3d_shader *object;
2978 HRESULT hr;
2979
2980 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
2981 device, desc, parent, parent_ops, shader);
2982
2983 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2984 if (!object)
2985 return E_OUTOFMEMORY;
2986
2987 if (FAILED(hr = geometryshader_init(object, device, desc, parent, parent_ops)))
2988 {
2989 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2990 HeapFree(GetProcessHeap(), 0, object);
2991 return hr;
2992 }
2993
2994 TRACE("Created geometry shader %p.\n", object);
2995 *shader = object;
2996
2997 return WINED3D_OK;
2998 }
2999
3000 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3001 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3002 {
3003 struct wined3d_shader *object;
3004 HRESULT hr;
3005
3006 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3007 device, desc, parent, parent_ops, shader);
3008
3009 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3010 if (!object)
3011 return E_OUTOFMEMORY;
3012
3013 if (FAILED(hr = pixelshader_init(object, device, desc, parent, parent_ops)))
3014 {
3015 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3016 HeapFree(GetProcessHeap(), 0, object);
3017 return hr;
3018 }
3019
3020 TRACE("Created pixel shader %p.\n", object);
3021 *shader = object;
3022
3023 return WINED3D_OK;
3024 }
3025
3026 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3027 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3028 {
3029 struct wined3d_shader *object;
3030 HRESULT hr;
3031
3032 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3033 device, desc, parent, parent_ops, shader);
3034
3035 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3036 if (!object)
3037 return E_OUTOFMEMORY;
3038
3039 if (FAILED(hr = vertexshader_init(object, device, desc, parent, parent_ops)))
3040 {
3041 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3042 HeapFree(GetProcessHeap(), 0, object);
3043 return hr;
3044 }
3045
3046 TRACE("Created vertex shader %p.\n", object);
3047 *shader = object;
3048
3049 return WINED3D_OK;
3050 }
Windows API implementation