?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Get rid of the "reorder shaders" for SM4+.
[wine.git] / dlls / wined3d / shader.c
blobbda0b6d85a478f006ca08fbf8bb38912883244b5
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_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
60 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
61 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
62 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
63 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
64 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
65 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
66 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
67 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
68 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
69 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
70 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
71 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
72 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
73 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
74 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
75 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
76 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
77 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
78 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
79 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
80 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
81 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
82 /* WINED3DSIH_DEF */ "def",
83 /* WINED3DSIH_DEFB */ "defb",
84 /* WINED3DSIH_DEFI */ "defi",
85 /* WINED3DSIH_DIV */ "div",
86 /* WINED3DSIH_DP2 */ "dp2",
87 /* WINED3DSIH_DP2ADD */ "dp2add",
88 /* WINED3DSIH_DP3 */ "dp3",
89 /* WINED3DSIH_DP4 */ "dp4",
90 /* WINED3DSIH_DST */ "dst",
91 /* WINED3DSIH_DSX */ "dsx",
92 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
93 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
94 /* WINED3DSIH_DSY */ "dsy",
95 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
96 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
97 /* WINED3DSIH_ELSE */ "else",
98 /* WINED3DSIH_EMIT */ "emit",
99 /* WINED3DSIH_ENDIF */ "endif",
100 /* WINED3DSIH_ENDLOOP */ "endloop",
101 /* WINED3DSIH_ENDREP */ "endrep",
102 /* WINED3DSIH_EQ */ "eq",
103 /* WINED3DSIH_EXP */ "exp",
104 /* WINED3DSIH_EXPP */ "expp",
105 /* WINED3DSIH_FRC */ "frc",
106 /* WINED3DSIH_FTOI */ "ftoi",
107 /* WINED3DSIH_FTOU */ "ftou",
108 /* WINED3DSIH_GE */ "ge",
109 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
110 /* WINED3DSIH_HS_DECLS */ "hs_decls",
111 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
112 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
113 /* WINED3DSIH_IADD */ "iadd",
114 /* WINED3DSIH_IEQ */ "ieq",
115 /* WINED3DSIH_IF */ "if",
116 /* WINED3DSIH_IFC */ "ifc",
117 /* WINED3DSIH_IGE */ "ige",
118 /* WINED3DSIH_ILT */ "ilt",
119 /* WINED3DSIH_IMAD */ "imad",
120 /* WINED3DSIH_IMAX */ "imax",
121 /* WINED3DSIH_IMIN */ "imin",
122 /* WINED3DSIH_IMUL */ "imul",
123 /* WINED3DSIH_INE */ "ine",
124 /* WINED3DSIH_INEG */ "ineg",
125 /* WINED3DSIH_ISHL */ "ishl",
126 /* WINED3DSIH_ITOF */ "itof",
127 /* WINED3DSIH_LABEL */ "label",
128 /* WINED3DSIH_LD */ "ld",
129 /* WINED3DSIH_LD2DMS */ "ld2dms",
130 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
131 /* WINED3DSIH_LIT */ "lit",
132 /* WINED3DSIH_LOG */ "log",
133 /* WINED3DSIH_LOGP */ "logp",
134 /* WINED3DSIH_LOOP */ "loop",
135 /* WINED3DSIH_LRP */ "lrp",
136 /* WINED3DSIH_LT */ "lt",
137 /* WINED3DSIH_M3x2 */ "m3x2",
138 /* WINED3DSIH_M3x3 */ "m3x3",
139 /* WINED3DSIH_M3x4 */ "m3x4",
140 /* WINED3DSIH_M4x3 */ "m4x3",
141 /* WINED3DSIH_M4x4 */ "m4x4",
142 /* WINED3DSIH_MAD */ "mad",
143 /* WINED3DSIH_MAX */ "max",
144 /* WINED3DSIH_MIN */ "min",
145 /* WINED3DSIH_MOV */ "mov",
146 /* WINED3DSIH_MOVA */ "mova",
147 /* WINED3DSIH_MOVC */ "movc",
148 /* WINED3DSIH_MUL */ "mul",
149 /* WINED3DSIH_NE */ "ne",
150 /* WINED3DSIH_NOP */ "nop",
151 /* WINED3DSIH_NOT */ "not",
152 /* WINED3DSIH_NRM */ "nrm",
153 /* WINED3DSIH_OR */ "or",
154 /* WINED3DSIH_PHASE */ "phase",
155 /* WINED3DSIH_POW */ "pow",
156 /* WINED3DSIH_RCP */ "rcp",
157 /* WINED3DSIH_REP */ "rep",
158 /* WINED3DSIH_RESINFO */ "resinfo",
159 /* WINED3DSIH_RET */ "ret",
160 /* WINED3DSIH_ROUND_NE */ "round_ne",
161 /* WINED3DSIH_ROUND_NI */ "round_ni",
162 /* WINED3DSIH_ROUND_PI */ "round_pi",
163 /* WINED3DSIH_ROUND_Z */ "round_z",
164 /* WINED3DSIH_RSQ */ "rsq",
165 /* WINED3DSIH_SAMPLE */ "sample",
166 /* WINED3DSIH_SAMPLE_B */ "sample_b",
167 /* WINED3DSIH_SAMPLE_C */ "sample_c",
168 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
169 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
170 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
171 /* WINED3DSIH_SETP */ "setp",
172 /* WINED3DSIH_SGE */ "sge",
173 /* WINED3DSIH_SGN */ "sgn",
174 /* WINED3DSIH_SINCOS */ "sincos",
175 /* WINED3DSIH_SLT */ "slt",
176 /* WINED3DSIH_SQRT */ "sqrt",
177 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
178 /* WINED3DSIH_SUB */ "sub",
179 /* WINED3DSIH_TEX */ "texld",
180 /* WINED3DSIH_TEXBEM */ "texbem",
181 /* WINED3DSIH_TEXBEML */ "texbeml",
182 /* WINED3DSIH_TEXCOORD */ "texcrd",
183 /* WINED3DSIH_TEXDEPTH */ "texdepth",
184 /* WINED3DSIH_TEXDP3 */ "texdp3",
185 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
186 /* WINED3DSIH_TEXKILL */ "texkill",
187 /* WINED3DSIH_TEXLDD */ "texldd",
188 /* WINED3DSIH_TEXLDL */ "texldl",
189 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
190 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
191 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
192 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
193 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
194 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
195 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
196 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
197 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
198 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
199 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
200 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
201 /* WINED3DSIH_UDIV */ "udiv",
202 /* WINED3DSIH_UGE */ "uge",
203 /* WINED3DSIH_ULT */ "ult",
204 /* WINED3DSIH_USHR */ "ushr",
205 /* WINED3DSIH_UTOF */ "utof",
206 /* WINED3DSIH_XOR */ "xor",
207 };
208
209 static const char * const semantic_names[] =
210 {
211 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
212 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
213 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
214 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
215 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
216 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
217 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
218 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
219 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
220 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
221 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
222 /* WINED3D_DECL_USAGE_FOG */ "FOG",
223 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
224 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
225 };
226
227 static const struct
228 {
229 enum wined3d_sysval_semantic sysval_semantic;
230 const char *sysval_name;
231 }
232 sysval_semantic_names[] =
233 {
234 {WINED3D_SV_POSITION, "SV_Position"},
235 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
236 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
237 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
238 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
239 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
240 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
241 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
242 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
243 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
244 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
245 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
246 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
247 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
248 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
249 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
250 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
251 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
252 };
253
254 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
255 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
256
257 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
258 {
259 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
260 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
261
262 return shader_opcode_names[handler_idx];
263 }
264
265 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
266 {
267 if (usage >= ARRAY_SIZE(semantic_names))
268 {
269 FIXME("Unrecognized usage %#x.\n", usage);
270 return "UNRECOGNIZED";
271 }
272
273 return semantic_names[usage];
274 }
275
276 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
277 {
278 unsigned int i;
279
280 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
281 {
282 if (!strcmp(name, semantic_names[i]))
283 return i;
284 }
285
286 return ~0U;
287 }
288
289 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
290 {
291 switch (usage)
292 {
293 case WINED3D_DECL_USAGE_POSITION:
294 return WINED3D_SV_POSITION;
295 default:
296 return 0;
297 }
298 }
299
300 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
301 {
302 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
303 }
304
305 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
306 const struct wined3d_shader_semantic *s)
307 {
308 e->semantic_name = shader_semantic_name_from_usage(s->usage);
309 e->semantic_idx = s->usage_idx;
310 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
311 e->component_type = 0;
312 e->register_idx = s->reg.reg.idx[0].offset;
313 e->mask = s->reg.write_mask;
314 }
315
316 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
317 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
318 {
319 e->semantic_name = shader_semantic_name_from_usage(usage);
320 e->semantic_idx = usage_idx;
321 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
322 e->component_type = 0;
323 e->register_idx = reg_idx;
324 e->mask = write_mask;
325 }
326
327 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
328 {
329 switch (version_token >> 16)
330 {
331 case WINED3D_SM1_VS:
332 case WINED3D_SM1_PS:
333 return &sm1_shader_frontend;
334
335 case WINED3D_SM4_PS:
336 case WINED3D_SM4_VS:
337 case WINED3D_SM4_GS:
338 case WINED3D_SM5_HS:
339 case WINED3D_SM5_DS:
340 return &sm4_shader_frontend;
341
342 default:
343 FIXME("Unrecognised version token %#x.\n", version_token);
344 return NULL;
345 }
346 }
347
348 void string_buffer_clear(struct wined3d_string_buffer *buffer)
349 {
350 buffer->buffer[0] = '\0';
351 buffer->content_size = 0;
352 }
353
354 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
355 {
356 buffer->buffer_size = 32;
357 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
358 {
359 ERR("Failed to allocate shader buffer memory.\n");
360 return FALSE;
361 }
362
363 string_buffer_clear(buffer);
364 return TRUE;
365 }
366
367 void string_buffer_free(struct wined3d_string_buffer *buffer)
368 {
369 HeapFree(GetProcessHeap(), 0, buffer->buffer);
370 }
371
372 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
373 {
374 char *new_buffer;
375 unsigned int new_buffer_size = buffer->buffer_size * 2;
376
377 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
378 new_buffer_size *= 2;
379 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
380 {
381 ERR("Failed to grow buffer.\n");
382 buffer->buffer[buffer->content_size] = '\0';
383 return FALSE;
384 }
385 buffer->buffer = new_buffer;
386 buffer->buffer_size = new_buffer_size;
387 return TRUE;
388 }
389
390 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
391 {
392 unsigned int rem;
393 int rc;
394
395 rem = buffer->buffer_size - buffer->content_size;
396 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
397 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
398 return rc;
399
400 buffer->content_size += rc;
401 return 0;
402 }
403
404 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
405 {
406 va_list args;
407 int ret;
408
409 for (;;)
410 {
411 va_start(args, format);
412 ret = shader_vaddline(buffer, format, args);
413 va_end(args);
414 if (!ret)
415 return ret;
416 if (!string_buffer_resize(buffer, ret))
417 return -1;
418 }
419 }
420
421 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
422 {
423 struct wined3d_string_buffer *buffer;
424
425 if (list_empty(&list->list))
426 {
427 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
428 if (!buffer || !string_buffer_init(buffer))
429 {
430 ERR("Couldn't allocate buffer for temporary string.\n");
431 if (buffer)
432 HeapFree(GetProcessHeap(), 0, buffer);
433 return NULL;
434 }
435 }
436 else
437 {
438 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
439 list_remove(&buffer->entry);
440 }
441 string_buffer_clear(buffer);
442 return buffer;
443 }
444
445 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
446 {
447 if (!buffer)
448 return 0;
449 string_buffer_clear(buffer);
450 return shader_vaddline(buffer, format, args);
451 }
452
453 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
454 {
455 va_list args;
456 int ret;
457
458 for (;;)
459 {
460 va_start(args, format);
461 ret = string_buffer_vsprintf(buffer, format, args);
462 va_end(args);
463 if (!ret)
464 return;
465 if (!string_buffer_resize(buffer, ret))
466 return;
467 }
468 }
469
470 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
471 {
472 if (!buffer)
473 return;
474 list_add_head(&list->list, &buffer->entry);
475 }
476
477 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
478 {
479 list_init(&list->list);
480 }
481
482 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
483 {
484 struct wined3d_string_buffer *buffer, *buffer_next;
485
486 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
487 {
488 string_buffer_free(buffer);
489 HeapFree(GetProcessHeap(), 0, buffer);
490 }
491 list_init(&list->list);
492 }
493
494 /* Convert floating point offset relative to a register file to an absolute
495 * offset for float constants. */
496 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
497 {
498 switch (register_type)
499 {
500 case WINED3DSPR_CONST: return register_idx;
501 case WINED3DSPR_CONST2: return 2048 + register_idx;
502 case WINED3DSPR_CONST3: return 4096 + register_idx;
503 case WINED3DSPR_CONST4: return 6144 + register_idx;
504 default:
505 FIXME("Unsupported register type: %u.\n", register_type);
506 return register_idx;
507 }
508 }
509
510 static void shader_delete_constant_list(struct list *clist)
511 {
512 struct wined3d_shader_lconst *constant, *constant_next;
513
514 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
515 HeapFree(GetProcessHeap(), 0, constant);
516 list_init(clist);
517 }
518
519 static void shader_set_limits(struct wined3d_shader *shader)
520 {
521 static const struct limits_entry
522 {
523 unsigned int min_version;
524 unsigned int max_version;
525 struct wined3d_shader_limits limits;
526 }
527 vs_limits[] =
528 {
529 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
530 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
531 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
532 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
533 * even though they are capable of supporting much more (GL
534 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
535 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
536 * shaders to 256. */
537 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
538 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
539 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
540 {0}
541 },
542 hs_limits[] =
543 {
544 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
545 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
546 },
547 ds_limits[] =
548 {
549 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
550 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
551 },
552 gs_limits[] =
553 {
554 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
555 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
556 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
557 {0}
558 },
559 ps_limits[] =
560 {
561 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
562 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
563 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
564 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
565 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
566 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
567 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
568 {0}
569 };
570 const struct limits_entry *limits_array;
571 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
572 shader->reg_maps.shader_version.minor);
573 int i = 0;
574
575 switch (shader->reg_maps.shader_version.type)
576 {
577 default:
578 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
579 /* Fall-through. */
580 case WINED3D_SHADER_TYPE_VERTEX:
581 limits_array = vs_limits;
582 break;
583 case WINED3D_SHADER_TYPE_HULL:
584 limits_array = hs_limits;
585 break;
586 case WINED3D_SHADER_TYPE_DOMAIN:
587 limits_array = ds_limits;
588 break;
589 case WINED3D_SHADER_TYPE_GEOMETRY:
590 limits_array = gs_limits;
591 break;
592 case WINED3D_SHADER_TYPE_PIXEL:
593 limits_array = ps_limits;
594 break;
595 }
596
597 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
598 {
599 if (shader_version <= limits_array[i].max_version)
600 {
601 shader->limits = &limits_array[i].limits;
602 break;
603 }
604 ++i;
605 }
606 if (!shader->limits)
607 {
608 FIXME("Unexpected shader version \"%u.%u\".\n",
609 shader->reg_maps.shader_version.major,
610 shader->reg_maps.shader_version.minor);
611 shader->limits = &limits_array[max(0, i - 1)].limits;
612 }
613 }
614
615 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
616 {
617 DWORD idx, shift;
618 idx = bit >> 5;
619 shift = bit & 0x1f;
620 bitmap[idx] |= (1u << shift);
621 }
622
623 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
624 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
625 {
626 switch (reg->type)
627 {
628 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
629 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
630 reg_maps->texcoord |= 1u << reg->idx[0].offset;
631 else
632 reg_maps->address |= 1u << reg->idx[0].offset;
633 break;
634
635 case WINED3DSPR_TEMP:
636 reg_maps->temporary |= 1u << reg->idx[0].offset;
637 break;
638
639 case WINED3DSPR_INPUT:
640 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
641 {
642 if (reg->idx[0].rel_addr)
643 {
644 /* If relative addressing is used, we must assume that all registers
645 * are used. Even if it is a construct like v3[aL], we can't assume
646 * that v0, v1 and v2 aren't read because aL can be negative */
647 unsigned int i;
648 for (i = 0; i < MAX_REG_INPUT; ++i)
649 {
650 shader->u.ps.input_reg_used[i] = TRUE;
651 }
652 }
653 else
654 {
655 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
656 }
657 }
658 else
659 reg_maps->input_registers |= 1u << reg->idx[0].offset;
660 break;
661
662 case WINED3DSPR_RASTOUT:
663 if (reg->idx[0].offset == 1)
664 reg_maps->fog = 1;
665 if (reg->idx[0].offset == 2)
666 reg_maps->point_size = 1;
667 break;
668
669 case WINED3DSPR_MISCTYPE:
670 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
671 {
672 if (!reg->idx[0].offset)
673 reg_maps->vpos = 1;
674 else if (reg->idx[0].offset == 1)
675 reg_maps->usesfacing = 1;
676 }
677 break;
678
679 case WINED3DSPR_CONST:
680 if (reg->idx[0].rel_addr)
681 {
682 if (reg->idx[0].offset < reg_maps->min_rel_offset)
683 reg_maps->min_rel_offset = reg->idx[0].offset;
684 if (reg->idx[0].offset > reg_maps->max_rel_offset)
685 reg_maps->max_rel_offset = reg->idx[0].offset;
686 reg_maps->usesrelconstF = TRUE;
687 }
688 else
689 {
690 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
691 {
692 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
693 return FALSE;
694 }
695 else
696 {
697 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
698 }
699 }
700 break;
701
702 case WINED3DSPR_CONSTINT:
703 if (reg->idx[0].offset >= shader->limits->constant_int)
704 {
705 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
706 return FALSE;
707 }
708 else
709 {
710 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
711 }
712 break;
713
714 case WINED3DSPR_CONSTBOOL:
715 if (reg->idx[0].offset >= shader->limits->constant_bool)
716 {
717 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
718 return FALSE;
719 }
720 else
721 {
722 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
723 }
724 break;
725
726 case WINED3DSPR_COLOROUT:
727 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
728 break;
729
730 default:
731 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
732 reg->type, reg->idx[0].offset, reg->idx[1].offset);
733 break;
734 }
735 return TRUE;
736 }
737
738 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
739 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
740 {
741 struct wined3d_shader_sampler_map_entry *entries, *entry;
742 struct wined3d_shader_sampler_map *map;
743 unsigned int i;
744
745 map = &reg_maps->sampler_map;
746 entries = map->entries;
747 for (i = 0; i < map->count; ++i)
748 {
749 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
750 return;
751 }
752
753 if (!map->size)
754 {
755 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
756 {
757 ERR("Failed to allocate sampler map entries.\n");
758 return;
759 }
760 map->size = 4;
761 map->entries = entries;
762 }
763 else if (map->count == map->size)
764 {
765 size_t new_size = map->size * 2;
766
767 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
768 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
769 {
770 ERR("Failed to resize sampler map entries.\n");
771 return;
772 }
773 map->size = new_size;
774 map->entries = entries;
775 }
776
777 entry = &entries[map->count++];
778 entry->resource_idx = resource_idx;
779 entry->sampler_idx = sampler_idx;
780 entry->bind_idx = bind_idx;
781 }
782
783 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
784 {
785 switch (instr)
786 {
787 case WINED3DSIH_M4x4:
788 case WINED3DSIH_M3x4:
789 return param == 1 ? 3 : 0;
790
791 case WINED3DSIH_M4x3:
792 case WINED3DSIH_M3x3:
793 return param == 1 ? 2 : 0;
794
795 case WINED3DSIH_M3x2:
796 return param == 1 ? 1 : 0;
797
798 default:
799 return 0;
800 }
801 }
802
803 /* Note that this does not count the loop register as an address register. */
804 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
805 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
806 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
807 {
808 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
809 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
810 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
811 void *fe_data = shader->frontend_data;
812 struct wined3d_shader_version shader_version;
813 const DWORD *ptr = byte_code;
814 unsigned int i;
815
816 memset(reg_maps, 0, sizeof(*reg_maps));
817 memset(input_signature_elements, 0, sizeof(input_signature_elements));
818 memset(output_signature_elements, 0, sizeof(output_signature_elements));
819 reg_maps->min_rel_offset = ~0U;
820
821 fe->shader_read_header(fe_data, &ptr, &shader_version);
822 reg_maps->shader_version = shader_version;
823
824 shader_set_limits(shader);
825
826 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
827 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
828 if (!reg_maps->constf)
829 {
830 ERR("Failed to allocate constant map memory.\n");
831 return E_OUTOFMEMORY;
832 }
833
834 while (!fe->shader_is_end(fe_data, &ptr))
835 {
836 struct wined3d_shader_instruction ins;
837
838 /* Fetch opcode. */
839 fe->shader_read_instruction(fe_data, &ptr, &ins);
840
841 /* Unhandled opcode, and its parameters. */
842 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
843 {
844 TRACE("Skipping unrecognized instruction.\n");
845 continue;
846 }
847
848 /* Handle declarations. */
849 if (ins.handler_idx == WINED3DSIH_DCL)
850 {
851 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
852 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
853
854 switch (semantic->reg.reg.type)
855 {
856 /* Mark input registers used. */
857 case WINED3DSPR_INPUT:
858 if (reg_idx >= MAX_REG_INPUT)
859 {
860 ERR("Invalid input register index %u.\n", reg_idx);
861 break;
862 }
863 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
864 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
865 return WINED3DERR_INVALIDCALL;
866 reg_maps->input_registers |= 1u << reg_idx;
867 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
868 break;
869
870 /* Vertex shader: mark 3.0 output registers used, save token. */
871 case WINED3DSPR_OUTPUT:
872 if (reg_idx >= MAX_REG_OUTPUT)
873 {
874 ERR("Invalid output register index %u.\n", reg_idx);
875 break;
876 }
877 reg_maps->output_registers |= 1u << reg_idx;
878 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
879 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
880 reg_maps->fog = 1;
881 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
882 reg_maps->point_size = 1;
883 break;
884
885 case WINED3DSPR_SAMPLER:
886 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
887 case WINED3DSPR_RESOURCE:
888 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
889 {
890 ERR("Invalid resource index %u.\n", reg_idx);
891 break;
892 }
893 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
894 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
895 break;
896
897 default:
898 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
899 break;
900 }
901 }
902 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
903 {
904 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
905 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
906 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
907 else
908 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
909 }
910 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
911 {
912 if (reg_maps->icb)
913 FIXME("Multiple immediate constant buffers.\n");
914 reg_maps->icb = ins.declaration.icb;
915 }
916 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
917 {
918 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
919 shader->u.gs.input_type = ins.declaration.primitive_type;
920 else
921 FIXME("Invalid instruction %#x for shader type %#x.\n",
922 ins.handler_idx, shader_version.type);
923 }
924 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
925 {
926 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
927 shader->u.gs.output_type = ins.declaration.primitive_type;
928 else
929 FIXME("Invalid instruction %#x for shader type %#x.\n",
930 ins.handler_idx, shader_version.type);
931 }
932 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
933 {
934 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
935 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
936 }
937 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
938 {
939 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
940 shader->u.gs.vertices_out = ins.declaration.count;
941 else
942 FIXME("Invalid instruction %#x for shader type %#x.\n",
943 ins.handler_idx, shader_version.type);
944 }
945 else if (ins.handler_idx == WINED3DSIH_DEF)
946 {
947 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
948 float *value;
949 if (!lconst) return E_OUTOFMEMORY;
950
951 lconst->idx = ins.dst[0].reg.idx[0].offset;
952 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
953 value = (float *)lconst->value;
954
955 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
956 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
957 {
958 if (value[0] < -1.0f) value[0] = -1.0f;
959 else if (value[0] > 1.0f) value[0] = 1.0f;
960 if (value[1] < -1.0f) value[1] = -1.0f;
961 else if (value[1] > 1.0f) value[1] = 1.0f;
962 if (value[2] < -1.0f) value[2] = -1.0f;
963 else if (value[2] > 1.0f) value[2] = 1.0f;
964 if (value[3] < -1.0f) value[3] = -1.0f;
965 else if (value[3] > 1.0f) value[3] = 1.0f;
966 }
967
968 list_add_head(&shader->constantsF, &lconst->entry);
969
970 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
971 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
972 {
973 shader->lconst_inf_or_nan = TRUE;
974 }
975 }
976 else if (ins.handler_idx == WINED3DSIH_DEFI)
977 {
978 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
979 if (!lconst) return E_OUTOFMEMORY;
980
981 lconst->idx = ins.dst[0].reg.idx[0].offset;
982 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
983
984 list_add_head(&shader->constantsI, &lconst->entry);
985 reg_maps->local_int_consts |= (1u << lconst->idx);
986 }
987 else if (ins.handler_idx == WINED3DSIH_DEFB)
988 {
989 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
990 if (!lconst) return E_OUTOFMEMORY;
991
992 lconst->idx = ins.dst[0].reg.idx[0].offset;
993 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
994
995 list_add_head(&shader->constantsB, &lconst->entry);
996 reg_maps->local_bool_consts |= (1u << lconst->idx);
997 }
998 /* For subroutine prototypes. */
999 else if (ins.handler_idx == WINED3DSIH_LABEL)
1000 {
1001 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1002 }
1003 /* Set texture, address, temporary registers. */
1004 else
1005 {
1006 BOOL color0_mov = FALSE;
1007 unsigned int i;
1008
1009 /* This will loop over all the registers and try to
1010 * make a bitmask of the ones we're interested in.
1011 *
1012 * Relative addressing tokens are ignored, but that's
1013 * okay, since we'll catch any address registers when
1014 * they are initialized (required by spec). */
1015 for (i = 0; i < ins.dst_count; ++i)
1016 {
1017 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1018 shader_version.type, constf_size))
1019 return WINED3DERR_INVALIDCALL;
1020
1021 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1022 {
1023 UINT idx = ins.dst[i].reg.idx[0].offset;
1024
1025 switch (ins.dst[i].reg.type)
1026 {
1027 case WINED3DSPR_RASTOUT:
1028 if (shader_version.major >= 3)
1029 break;
1030 switch (idx)
1031 {
1032 case 0: /* oPos */
1033 reg_maps->output_registers |= 1u << 10;
1034 shader_signature_from_usage(&output_signature_elements[10],
1035 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1036 break;
1037
1038 case 1: /* oFog */
1039 reg_maps->output_registers |= 1u << 11;
1040 shader_signature_from_usage(&output_signature_elements[11],
1041 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1042 break;
1043
1044 case 2: /* oPts */
1045 reg_maps->output_registers |= 1u << 11;
1046 shader_signature_from_usage(&output_signature_elements[11],
1047 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1048 break;
1049 }
1050 break;
1051
1052 case WINED3DSPR_ATTROUT:
1053 if (shader_version.major >= 3)
1054 break;
1055 if (idx < 2)
1056 {
1057 idx += 8;
1058 if (reg_maps->output_registers & (1u << idx))
1059 {
1060 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1061 }
1062 else
1063 {
1064 reg_maps->output_registers |= 1u << idx;
1065 shader_signature_from_usage(&output_signature_elements[idx],
1066 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1067 }
1068 }
1069 break;
1070
1071 case WINED3DSPR_TEXCRDOUT:
1072 if (shader_version.major >= 3)
1073 {
1074 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1075 break;
1076 }
1077 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1078 if (reg_maps->output_registers & (1u << idx))
1079 {
1080 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1081 }
1082 else
1083 {
1084 reg_maps->output_registers |= 1u << idx;
1085 shader_signature_from_usage(&output_signature_elements[idx],
1086 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1087 }
1088 break;
1089
1090 default:
1091 break;
1092 }
1093 }
1094
1095 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1096 {
1097 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1098 {
1099 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1100 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1101 * the mov and perform the sRGB write correction from the source register.
1102 *
1103 * However, if the mov is only partial, we can't do this, and if the write
1104 * comes from an instruction other than MOV it is hard to do as well. If
1105 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1106 shader->u.ps.color0_mov = FALSE;
1107 if (ins.handler_idx == WINED3DSIH_MOV
1108 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1109 {
1110 /* Used later when the source register is read. */
1111 color0_mov = TRUE;
1112 }
1113 }
1114 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1115 * end
1116 */
1117 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1118 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1119 {
1120 shader->u.ps.color0_mov = FALSE;
1121 }
1122 }
1123
1124 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1125 if (shader_version.major == 1
1126 && (ins.handler_idx == WINED3DSIH_TEX
1127 || ins.handler_idx == WINED3DSIH_TEXBEM
1128 || ins.handler_idx == WINED3DSIH_TEXBEML
1129 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1130 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1131 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1132 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1133 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1134 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1135 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1136 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1137 {
1138 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1139
1140 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1141 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1142 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1143 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1144
1145 /* texbem is only valid with < 1.4 pixel shaders */
1146 if (ins.handler_idx == WINED3DSIH_TEXBEM
1147 || ins.handler_idx == WINED3DSIH_TEXBEML)
1148 {
1149 reg_maps->bumpmat |= 1u << reg_idx;
1150 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1151 {
1152 reg_maps->luminanceparams |= 1u << reg_idx;
1153 }
1154 }
1155 }
1156 else if (ins.handler_idx == WINED3DSIH_BEM)
1157 {
1158 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1159 }
1160 }
1161
1162 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1163 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1164 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1165 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1166 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1167 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1168 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1169 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1170 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1171 else if (ins.handler_idx == WINED3DSIH_LOOP
1172 || ins.handler_idx == WINED3DSIH_REP)
1173 {
1174 ++cur_loop_depth;
1175 if (cur_loop_depth > max_loop_depth)
1176 max_loop_depth = cur_loop_depth;
1177 }
1178 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1179 || ins.handler_idx == WINED3DSIH_ENDREP)
1180 {
1181 --cur_loop_depth;
1182 }
1183 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1184 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1185 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1186 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1187 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1188 {
1189 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1190 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1191 }
1192 else if (ins.handler_idx == WINED3DSIH_LD
1193 || ins.handler_idx == WINED3DSIH_RESINFO)
1194 {
1195 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1196 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1197 }
1198
1199 if (ins.predicate)
1200 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1201 shader_version.type, constf_size))
1202 return WINED3DERR_INVALIDCALL;
1203
1204 for (i = 0; i < ins.src_count; ++i)
1205 {
1206 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1207 struct wined3d_shader_register reg = ins.src[i].reg;
1208
1209 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1210 shader_version.type, constf_size))
1211 return WINED3DERR_INVALIDCALL;
1212 while (count)
1213 {
1214 ++reg.idx[0].offset;
1215 if (!shader_record_register_usage(shader, reg_maps, &reg,
1216 shader_version.type, constf_size))
1217 return WINED3DERR_INVALIDCALL;
1218 --count;
1219 }
1220
1221 if (color0_mov)
1222 {
1223 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1224 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1225 {
1226 shader->u.ps.color0_mov = TRUE;
1227 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1228 }
1229 }
1230 }
1231 }
1232 }
1233 reg_maps->loop_depth = max_loop_depth;
1234
1235 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1236 * R0 is written to the render target. */
1237 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1238 reg_maps->rt_mask |= (1u << 0);
1239
1240 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1241
1242 if (input_signature->elements)
1243 {
1244 for (i = 0; i < input_signature->element_count; ++i)
1245 {
1246 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1247 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1248 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1249 reg_maps->vpos = 1;
1250 }
1251 }
1252 else if (!input_signature->elements && reg_maps->input_registers)
1253 {
1254 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1255 struct wined3d_shader_signature_element *e;
1256 unsigned int i;
1257
1258 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1259 return E_OUTOFMEMORY;
1260 input_signature->element_count = count;
1261
1262 e = input_signature->elements;
1263 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1264 {
1265 if (!(reg_maps->input_registers & (1u << i)))
1266 continue;
1267 input_signature_elements[i].register_idx = i;
1268 *e++ = input_signature_elements[i];
1269 }
1270 }
1271
1272 if (output_signature->elements)
1273 {
1274 for (i = 0; i < output_signature->element_count; ++i)
1275 {
1276 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1277 }
1278 }
1279 else if (reg_maps->output_registers)
1280 {
1281 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1282 struct wined3d_shader_signature_element *e;
1283
1284 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1285 return E_OUTOFMEMORY;
1286 output_signature->element_count = count;
1287
1288 e = output_signature->elements;
1289 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1290 {
1291 if (!(reg_maps->output_registers & (1u << i)))
1292 continue;
1293 *e++ = output_signature_elements[i];
1294 }
1295 }
1296
1297 return WINED3D_OK;
1298 }
1299
1300 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1301 {
1302 DWORD map = 1u << max;
1303 map |= map - 1;
1304 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1305
1306 return wined3d_log2i(map);
1307 }
1308
1309 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1310 {
1311 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1312 {
1313 shader_addline(buffer, "refactoringAllowed");
1314 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1315 if (global_flags)
1316 shader_addline(buffer, " | ");
1317 }
1318
1319 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1320 {
1321 shader_addline(buffer, "enableRawAndStructuredBuffers");
1322 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1323 }
1324
1325 if (global_flags)
1326 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1327 }
1328
1329 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1330 enum wined3d_tessellator_domain domain)
1331 {
1332 switch (domain)
1333 {
1334 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1335 shader_addline(buffer, "line");
1336 break;
1337 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1338 shader_addline(buffer, "triangle");
1339 break;
1340 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1341 shader_addline(buffer, "quad");
1342 break;
1343 default:
1344 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1345 break;
1346 }
1347 }
1348
1349 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1350 enum wined3d_tessellator_output_primitive output_primitive)
1351 {
1352 switch (output_primitive)
1353 {
1354 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1355 shader_addline(buffer, "point");
1356 break;
1357 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1358 shader_addline(buffer, "line");
1359 break;
1360 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1361 shader_addline(buffer, "triangle_cw");
1362 break;
1363 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1364 shader_addline(buffer, "triangle_ccw");
1365 break;
1366 default:
1367 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1368 break;
1369 }
1370 }
1371
1372 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1373 enum wined3d_tessellator_partitioning partitioning)
1374 {
1375 switch (partitioning)
1376 {
1377 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1378 shader_addline(buffer, "integer");
1379 break;
1380 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1381 shader_addline(buffer, "pow2");
1382 break;
1383 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1384 shader_addline(buffer, "fractional_odd");
1385 break;
1386 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1387 shader_addline(buffer, "fractional_even");
1388 break;
1389 default:
1390 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1391 break;
1392 }
1393 }
1394
1395 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1396 {
1397 unsigned int i;
1398
1399 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1400 {
1401 if (sysval_semantic_names[i].sysval_semantic == semantic)
1402 {
1403 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1404 return;
1405 }
1406 }
1407
1408 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1409 }
1410
1411 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1412 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1413 {
1414 shader_addline(buffer, "dcl");
1415
1416 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1417 {
1418 switch (semantic->resource_type)
1419 {
1420 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1421 shader_addline(buffer, "_2d");
1422 break;
1423
1424 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1425 shader_addline(buffer, "_3d");
1426 break;
1427
1428 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1429 shader_addline(buffer, "_cube");
1430 break;
1431
1432 default:
1433 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1434 break;
1435 }
1436 }
1437 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1438 {
1439 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1440 shader_addline(buffer, "_resource_");
1441 else
1442 shader_addline(buffer, "_uav_");
1443 switch (semantic->resource_type)
1444 {
1445 case WINED3D_SHADER_RESOURCE_BUFFER:
1446 shader_addline(buffer, "buffer");
1447 break;
1448
1449 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1450 shader_addline(buffer, "texture1d");
1451 break;
1452
1453 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1454 shader_addline(buffer, "texture2d");
1455 break;
1456
1457 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1458 shader_addline(buffer, "texture2dms");
1459 break;
1460
1461 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1462 shader_addline(buffer, "texture3d");
1463 break;
1464
1465 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1466 shader_addline(buffer, "texturecube");
1467 break;
1468
1469 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1470 shader_addline(buffer, "texture1darray");
1471 break;
1472
1473 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1474 shader_addline(buffer, "texture2darray");
1475 break;
1476
1477 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1478 shader_addline(buffer, "texture2dmsarray");
1479 break;
1480
1481 default:
1482 shader_addline(buffer, "unknown");
1483 break;
1484 }
1485 switch (semantic->resource_data_type)
1486 {
1487 case WINED3D_DATA_FLOAT:
1488 shader_addline(buffer, " (float)");
1489 break;
1490
1491 case WINED3D_DATA_INT:
1492 shader_addline(buffer, " (int)");
1493 break;
1494
1495 case WINED3D_DATA_UINT:
1496 shader_addline(buffer, " (uint)");
1497 break;
1498
1499 case WINED3D_DATA_UNORM:
1500 shader_addline(buffer, " (unorm)");
1501 break;
1502
1503 case WINED3D_DATA_SNORM:
1504 shader_addline(buffer, " (snorm)");
1505 break;
1506
1507 default:
1508 shader_addline(buffer, " (unknown)");
1509 break;
1510 }
1511 }
1512 else
1513 {
1514 /* Pixel shaders 3.0 don't have usage semantics. */
1515 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1516 return;
1517 else
1518 shader_addline(buffer, "_");
1519
1520 switch (semantic->usage)
1521 {
1522 case WINED3D_DECL_USAGE_POSITION:
1523 shader_addline(buffer, "position%u", semantic->usage_idx);
1524 break;
1525
1526 case WINED3D_DECL_USAGE_BLEND_INDICES:
1527 shader_addline(buffer, "blend");
1528 break;
1529
1530 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1531 shader_addline(buffer, "weight");
1532 break;
1533
1534 case WINED3D_DECL_USAGE_NORMAL:
1535 shader_addline(buffer, "normal%u", semantic->usage_idx);
1536 break;
1537
1538 case WINED3D_DECL_USAGE_PSIZE:
1539 shader_addline(buffer, "psize");
1540 break;
1541
1542 case WINED3D_DECL_USAGE_COLOR:
1543 if (!semantic->usage_idx)
1544 shader_addline(buffer, "color");
1545 else
1546 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1547 break;
1548
1549 case WINED3D_DECL_USAGE_TEXCOORD:
1550 shader_addline(buffer, "texture%u", semantic->usage_idx);
1551 break;
1552
1553 case WINED3D_DECL_USAGE_TANGENT:
1554 shader_addline(buffer, "tangent");
1555 break;
1556
1557 case WINED3D_DECL_USAGE_BINORMAL:
1558 shader_addline(buffer, "binormal");
1559 break;
1560
1561 case WINED3D_DECL_USAGE_TESS_FACTOR:
1562 shader_addline(buffer, "tessfactor");
1563 break;
1564
1565 case WINED3D_DECL_USAGE_POSITIONT:
1566 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1567 break;
1568
1569 case WINED3D_DECL_USAGE_FOG:
1570 shader_addline(buffer, "fog");
1571 break;
1572
1573 case WINED3D_DECL_USAGE_DEPTH:
1574 shader_addline(buffer, "depth");
1575 break;
1576
1577 case WINED3D_DECL_USAGE_SAMPLE:
1578 shader_addline(buffer, "sample");
1579 break;
1580
1581 default:
1582 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1583 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1584 }
1585 }
1586 }
1587
1588 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1589 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1590 {
1591 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1592 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1593 UINT offset = reg->idx[0].offset;
1594
1595 switch (reg->type)
1596 {
1597 case WINED3DSPR_TEMP:
1598 shader_addline(buffer, "r");
1599 break;
1600
1601 case WINED3DSPR_INPUT:
1602 shader_addline(buffer, "v");
1603 break;
1604
1605 case WINED3DSPR_CONST:
1606 case WINED3DSPR_CONST2:
1607 case WINED3DSPR_CONST3:
1608 case WINED3DSPR_CONST4:
1609 shader_addline(buffer, "c");
1610 offset = shader_get_float_offset(reg->type, offset);
1611 break;
1612
1613 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1614 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1615 break;
1616
1617 case WINED3DSPR_RASTOUT:
1618 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1619 break;
1620
1621 case WINED3DSPR_COLOROUT:
1622 shader_addline(buffer, "oC");
1623 break;
1624
1625 case WINED3DSPR_DEPTHOUT:
1626 shader_addline(buffer, "oDepth");
1627 break;
1628
1629 case WINED3DSPR_ATTROUT:
1630 shader_addline(buffer, "oD");
1631 break;
1632
1633 case WINED3DSPR_TEXCRDOUT:
1634 /* Vertex shaders >= 3.0 use general purpose output registers
1635 * (WINED3DSPR_OUTPUT), which can include an address token. */
1636 if (shader_version->major >= 3)
1637 shader_addline(buffer, "o");
1638 else
1639 shader_addline(buffer, "oT");
1640 break;
1641
1642 case WINED3DSPR_CONSTINT:
1643 shader_addline(buffer, "i");
1644 break;
1645
1646 case WINED3DSPR_CONSTBOOL:
1647 shader_addline(buffer, "b");
1648 break;
1649
1650 case WINED3DSPR_LABEL:
1651 shader_addline(buffer, "l");
1652 break;
1653
1654 case WINED3DSPR_LOOP:
1655 shader_addline(buffer, "aL");
1656 break;
1657
1658 case WINED3DSPR_SAMPLER:
1659 shader_addline(buffer, "s");
1660 break;
1661
1662 case WINED3DSPR_MISCTYPE:
1663 if (offset > 1)
1664 {
1665 FIXME("Unhandled misctype register %u.\n", offset);
1666 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1667 }
1668 else
1669 {
1670 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1671 }
1672 break;
1673
1674 case WINED3DSPR_PREDICATE:
1675 shader_addline(buffer, "p");
1676 break;
1677
1678 case WINED3DSPR_IMMCONST:
1679 shader_addline(buffer, "l");
1680 break;
1681
1682 case WINED3DSPR_CONSTBUFFER:
1683 shader_addline(buffer, "cb");
1684 break;
1685
1686 case WINED3DSPR_IMMCONSTBUFFER:
1687 shader_addline(buffer, "icb");
1688 break;
1689
1690 case WINED3DSPR_PRIMID:
1691 shader_addline(buffer, "primID");
1692 break;
1693
1694 case WINED3DSPR_NULL:
1695 shader_addline(buffer, "null");
1696 break;
1697
1698 case WINED3DSPR_RESOURCE:
1699 shader_addline(buffer, "t");
1700 break;
1701
1702 case WINED3DSPR_UAV:
1703 shader_addline(buffer, "u");
1704 break;
1705
1706 case WINED3DSPR_OUTPOINTID:
1707 shader_addline(buffer, "vOutputControlPointID");
1708 break;
1709
1710 case WINED3DSPR_FORKINSTID:
1711 shader_addline(buffer, "vForkInstanceId");
1712 break;
1713
1714 case WINED3DSPR_INCONTROLPOINT:
1715 shader_addline(buffer, "vicp");
1716 break;
1717
1718 case WINED3DSPR_PATCHCONST:
1719 shader_addline(buffer, "vpc");
1720 break;
1721
1722 case WINED3DSPR_TESSCOORD:
1723 shader_addline(buffer, "vDomainLocation");
1724 break;
1725
1726 default:
1727 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1728 break;
1729 }
1730
1731 if (reg->type == WINED3DSPR_IMMCONST)
1732 {
1733 shader_addline(buffer, "(");
1734 switch (reg->immconst_type)
1735 {
1736 case WINED3D_IMMCONST_SCALAR:
1737 switch (reg->data_type)
1738 {
1739 case WINED3D_DATA_FLOAT:
1740 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1741 break;
1742 case WINED3D_DATA_INT:
1743 shader_addline(buffer, "%d", reg->immconst_data[0]);
1744 break;
1745 case WINED3D_DATA_RESOURCE:
1746 case WINED3D_DATA_SAMPLER:
1747 case WINED3D_DATA_UINT:
1748 shader_addline(buffer, "%u", reg->immconst_data[0]);
1749 break;
1750 default:
1751 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1752 break;
1753 }
1754 break;
1755
1756 case WINED3D_IMMCONST_VEC4:
1757 switch (reg->data_type)
1758 {
1759 case WINED3D_DATA_FLOAT:
1760 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1761 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1762 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1763 break;
1764 case WINED3D_DATA_INT:
1765 shader_addline(buffer, "%d, %d, %d, %d",
1766 reg->immconst_data[0], reg->immconst_data[1],
1767 reg->immconst_data[2], reg->immconst_data[3]);
1768 break;
1769 case WINED3D_DATA_RESOURCE:
1770 case WINED3D_DATA_SAMPLER:
1771 case WINED3D_DATA_UINT:
1772 shader_addline(buffer, "%u, %u, %u, %u",
1773 reg->immconst_data[0], reg->immconst_data[1],
1774 reg->immconst_data[2], reg->immconst_data[3]);
1775 break;
1776 default:
1777 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1778 break;
1779 }
1780 break;
1781
1782 default:
1783 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1784 break;
1785 }
1786 shader_addline(buffer, ")");
1787 }
1788 else if (reg->type != WINED3DSPR_RASTOUT
1789 && reg->type != WINED3DSPR_MISCTYPE
1790 && reg->type != WINED3DSPR_NULL)
1791 {
1792 if (offset != ~0u)
1793 {
1794 shader_addline(buffer, "[");
1795 if (reg->idx[0].rel_addr)
1796 {
1797 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1798 shader_addline(buffer, " + ");
1799 }
1800 shader_addline(buffer, "%u]", offset);
1801
1802 if (reg->idx[1].offset != ~0u)
1803 {
1804 shader_addline(buffer, "[");
1805 if (reg->idx[1].rel_addr)
1806 {
1807 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1808 shader_addline(buffer, " + ");
1809 }
1810 shader_addline(buffer, "%u]", reg->idx[1].offset);
1811 }
1812 }
1813 }
1814 }
1815
1816 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1817 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1818 {
1819 DWORD write_mask = param->write_mask;
1820
1821 shader_dump_register(buffer, &param->reg, shader_version);
1822
1823 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1824 {
1825 static const char write_mask_chars[] = "xyzw";
1826
1827 shader_addline(buffer, ".");
1828 if (write_mask & WINED3DSP_WRITEMASK_0)
1829 shader_addline(buffer, "%c", write_mask_chars[0]);
1830 if (write_mask & WINED3DSP_WRITEMASK_1)
1831 shader_addline(buffer, "%c", write_mask_chars[1]);
1832 if (write_mask & WINED3DSP_WRITEMASK_2)
1833 shader_addline(buffer, "%c", write_mask_chars[2]);
1834 if (write_mask & WINED3DSP_WRITEMASK_3)
1835 shader_addline(buffer, "%c", write_mask_chars[3]);
1836 }
1837 }
1838
1839 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1840 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1841 {
1842 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1843 DWORD swizzle = param->swizzle;
1844
1845 if (src_modifier == WINED3DSPSM_NEG
1846 || src_modifier == WINED3DSPSM_BIASNEG
1847 || src_modifier == WINED3DSPSM_SIGNNEG
1848 || src_modifier == WINED3DSPSM_X2NEG
1849 || src_modifier == WINED3DSPSM_ABSNEG)
1850 shader_addline(buffer, "-");
1851 else if (src_modifier == WINED3DSPSM_COMP)
1852 shader_addline(buffer, "1-");
1853 else if (src_modifier == WINED3DSPSM_NOT)
1854 shader_addline(buffer, "!");
1855
1856 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1857 shader_addline(buffer, "abs(");
1858
1859 shader_dump_register(buffer, &param->reg, shader_version);
1860
1861 switch (src_modifier)
1862 {
1863 case WINED3DSPSM_NONE: break;
1864 case WINED3DSPSM_NEG: break;
1865 case WINED3DSPSM_NOT: break;
1866 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1867 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1868 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1869 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1870 case WINED3DSPSM_COMP: break;
1871 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1872 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1873 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1874 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1875 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1876 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1877 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1878 }
1879
1880 if (swizzle != WINED3DSP_NOSWIZZLE)
1881 {
1882 static const char swizzle_chars[] = "xyzw";
1883 DWORD swizzle_x = swizzle & 0x03;
1884 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1885 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1886 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1887
1888 if (swizzle_x == swizzle_y
1889 && swizzle_x == swizzle_z
1890 && swizzle_x == swizzle_w)
1891 {
1892 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
1893 }
1894 else
1895 {
1896 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1897 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1898 }
1899 }
1900 }
1901
1902 /* Shared code in order to generate the bulk of the shader string.
1903 * NOTE: A description of how to parse tokens can be found on MSDN. */
1904 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1905 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1906 {
1907 struct wined3d_device *device = shader->device;
1908 const struct wined3d_shader_frontend *fe = shader->frontend;
1909 void *fe_data = shader->frontend_data;
1910 struct wined3d_shader_version shader_version;
1911 struct wined3d_shader_loop_state loop_state;
1912 struct wined3d_shader_instruction ins;
1913 struct wined3d_shader_tex_mx tex_mx;
1914 struct wined3d_shader_context ctx;
1915 const DWORD *ptr = byte_code;
1916
1917 /* Initialize current parsing state. */
1918 tex_mx.current_row = 0;
1919 loop_state.current_depth = 0;
1920 loop_state.current_reg = 0;
1921
1922 ctx.shader = shader;
1923 ctx.gl_info = &device->adapter->gl_info;
1924 ctx.reg_maps = reg_maps;
1925 ctx.buffer = buffer;
1926 ctx.tex_mx = &tex_mx;
1927 ctx.loop_state = &loop_state;
1928 ctx.backend_data = backend_ctx;
1929 ins.ctx = &ctx;
1930
1931 fe->shader_read_header(fe_data, &ptr, &shader_version);
1932
1933 while (!fe->shader_is_end(fe_data, &ptr))
1934 {
1935 /* Read opcode. */
1936 fe->shader_read_instruction(fe_data, &ptr, &ins);
1937
1938 /* Unknown opcode and its parameters. */
1939 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1940 {
1941 TRACE("Skipping unrecognized instruction.\n");
1942 continue;
1943 }
1944
1945 if (ins.predicate)
1946 FIXME("Predicates not implemented.\n");
1947
1948 /* Call appropriate function for output target */
1949 device->shader_backend->shader_handle_instruction(&ins);
1950 }
1951 }
1952
1953 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
1954 const struct wined3d_shader_dst_param *dst)
1955 {
1956 DWORD mmask = dst->modifiers;
1957
1958 switch (dst->shift)
1959 {
1960 case 0: break;
1961 case 13: shader_addline(buffer, "_d8"); break;
1962 case 14: shader_addline(buffer, "_d4"); break;
1963 case 15: shader_addline(buffer, "_d2"); break;
1964 case 1: shader_addline(buffer, "_x2"); break;
1965 case 2: shader_addline(buffer, "_x4"); break;
1966 case 3: shader_addline(buffer, "_x8"); break;
1967 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
1968 }
1969
1970 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
1971 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
1972 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
1973
1974 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1975 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1976 }
1977
1978 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
1979 enum wined3d_primitive_type primitive_type)
1980 {
1981 switch (primitive_type)
1982 {
1983 case WINED3D_PT_UNDEFINED:
1984 shader_addline(buffer, "undefined");
1985 break;
1986 case WINED3D_PT_POINTLIST:
1987 shader_addline(buffer, "pointlist");
1988 break;
1989 case WINED3D_PT_LINELIST:
1990 shader_addline(buffer, "linelist");
1991 break;
1992 case WINED3D_PT_LINESTRIP:
1993 shader_addline(buffer, "linestrip");
1994 break;
1995 case WINED3D_PT_TRIANGLELIST:
1996 shader_addline(buffer, "trianglelist");
1997 break;
1998 case WINED3D_PT_TRIANGLESTRIP:
1999 shader_addline(buffer, "trianglestrip");
2000 break;
2001 case WINED3D_PT_TRIANGLEFAN:
2002 shader_addline(buffer, "trianglefan");
2003 break;
2004 case WINED3D_PT_LINELIST_ADJ:
2005 shader_addline(buffer, "linelist_adj");
2006 break;
2007 case WINED3D_PT_LINESTRIP_ADJ:
2008 shader_addline(buffer, "linestrip_adj");
2009 break;
2010 case WINED3D_PT_TRIANGLELIST_ADJ:
2011 shader_addline(buffer, "trianglelist_adj");
2012 break;
2013 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2014 shader_addline(buffer, "trianglestrip_adj");
2015 break;
2016 default:
2017 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2018 break;
2019 }
2020 }
2021
2022 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2023 enum wined3d_shader_interpolation_mode interpolation_mode)
2024 {
2025 switch (interpolation_mode)
2026 {
2027 case WINED3DSIM_CONSTANT:
2028 shader_addline(buffer, "constant");
2029 break;
2030 case WINED3DSIM_LINEAR:
2031 shader_addline(buffer, "linear");
2032 break;
2033 case WINED3DSIM_LINEAR_CENTROID:
2034 shader_addline(buffer, "linear centroid");
2035 break;
2036 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2037 shader_addline(buffer, "linear noperspective");
2038 break;
2039 case WINED3DSIM_LINEAR_SAMPLE:
2040 shader_addline(buffer, "linear sample");
2041 break;
2042 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2043 shader_addline(buffer, "linear noperspective centroid");
2044 break;
2045 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2046 shader_addline(buffer, "linear noperspective sample");
2047 break;
2048 default:
2049 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2050 break;
2051 }
2052 }
2053
2054 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2055 {
2056 struct wined3d_shader_version shader_version;
2057 struct wined3d_string_buffer buffer;
2058 const DWORD *ptr = byte_code;
2059 const char *type_prefix;
2060 const char *p, *q;
2061 DWORD i;
2062
2063 if (!string_buffer_init(&buffer))
2064 {
2065 ERR("Failed to initialize string buffer.\n");
2066 return;
2067 }
2068
2069 TRACE("Parsing %p.\n", byte_code);
2070
2071 fe->shader_read_header(fe_data, &ptr, &shader_version);
2072
2073 switch (shader_version.type)
2074 {
2075 case WINED3D_SHADER_TYPE_VERTEX:
2076 type_prefix = "vs";
2077 break;
2078
2079 case WINED3D_SHADER_TYPE_HULL:
2080 type_prefix = "hs";
2081 break;
2082
2083 case WINED3D_SHADER_TYPE_DOMAIN:
2084 type_prefix = "ds";
2085 break;
2086
2087 case WINED3D_SHADER_TYPE_GEOMETRY:
2088 type_prefix = "gs";
2089 break;
2090
2091 case WINED3D_SHADER_TYPE_PIXEL:
2092 type_prefix = "ps";
2093 break;
2094
2095 default:
2096 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2097 type_prefix = "unknown";
2098 break;
2099 }
2100
2101 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2102
2103 while (!fe->shader_is_end(fe_data, &ptr))
2104 {
2105 struct wined3d_shader_instruction ins;
2106
2107 fe->shader_read_instruction(fe_data, &ptr, &ins);
2108 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2109 {
2110 WARN("Skipping unrecognized instruction.\n");
2111 shader_addline(&buffer, "<unrecognized instruction>\n");
2112 continue;
2113 }
2114
2115 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2116 {
2117 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2118 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2119 shader_addline(&buffer, " ");
2120 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2121 }
2122 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2123 {
2124 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2125 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2126 shader_addline(&buffer, ", %s",
2127 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2128 }
2129 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2130 {
2131 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2132 shader_dump_global_flags(&buffer, ins.flags);
2133 }
2134 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2135 {
2136 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2137 ins.declaration.max_tessellation_factor);
2138 }
2139 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2140 {
2141 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2142 for (i = 0; i < ins.declaration.icb->element_count / 4; ++i)
2143 {
2144 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2145 ins.declaration.icb->data[4 * i + 0],
2146 ins.declaration.icb->data[4 * i + 1],
2147 ins.declaration.icb->data[4 * i + 2],
2148 ins.declaration.icb->data[4 * i + 3]);
2149 }
2150 shader_addline(&buffer, "}");
2151 }
2152 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2153 {
2154 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2155 shader_dump_interpolation_mode(&buffer, ins.flags);
2156 shader_addline(&buffer, " ");
2157 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2158 }
2159 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2160 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2161 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2162 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2163 {
2164 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2165 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2166 shader_addline(&buffer, ", ");
2167 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2168 }
2169 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2170 {
2171 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2172 shader_dump_interpolation_mode(&buffer, ins.flags);
2173 shader_addline(&buffer, " ");
2174 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2175 shader_addline(&buffer, ", ");
2176 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2177 }
2178 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2179 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2180 {
2181 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2182 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2183 }
2184 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2185 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2186 {
2187 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2188 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2189 }
2190 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2191 {
2192 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2193 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2194 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2195 }
2196 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2197 {
2198 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2199 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2200 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2201 shader_addline(&buffer, ", comparisonMode");
2202 }
2203 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2204 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2205 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2206 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2207 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2208 {
2209 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2210 }
2211 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2212 {
2213 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2214 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2215 }
2216 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2217 {
2218 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2219 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2220 }
2221 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2222 {
2223 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2224 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2225 }
2226 else if (ins.handler_idx == WINED3DSIH_DEF)
2227 {
2228 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2229 ins.dst[0].reg.idx[0].offset),
2230 *(const float *)&ins.src[0].reg.immconst_data[0],
2231 *(const float *)&ins.src[0].reg.immconst_data[1],
2232 *(const float *)&ins.src[0].reg.immconst_data[2],
2233 *(const float *)&ins.src[0].reg.immconst_data[3]);
2234 }
2235 else if (ins.handler_idx == WINED3DSIH_DEFI)
2236 {
2237 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2238 ins.src[0].reg.immconst_data[0],
2239 ins.src[0].reg.immconst_data[1],
2240 ins.src[0].reg.immconst_data[2],
2241 ins.src[0].reg.immconst_data[3]);
2242 }
2243 else if (ins.handler_idx == WINED3DSIH_DEFB)
2244 {
2245 shader_addline(&buffer, "defb b%u = %s",
2246 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2247 }
2248 else
2249 {
2250 if (ins.predicate)
2251 {
2252 shader_addline(&buffer, "(");
2253 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2254 shader_addline(&buffer, ") ");
2255 }
2256
2257 /* PixWin marks instructions with the coissue flag with a '+' */
2258 if (ins.coissue)
2259 shader_addline(&buffer, "+");
2260
2261 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2262
2263 if (ins.handler_idx == WINED3DSIH_IFC
2264 || ins.handler_idx == WINED3DSIH_BREAKC)
2265 {
2266 switch (ins.flags)
2267 {
2268 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2269 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2270 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2271 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2272 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2273 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2274 default: shader_addline(&buffer, "_(%u)", ins.flags);
2275 }
2276 }
2277 else if (ins.handler_idx == WINED3DSIH_TEX
2278 && shader_version.major >= 2
2279 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2280 {
2281 shader_addline(&buffer, "p");
2282 }
2283 else if (ins.handler_idx == WINED3DSIH_RESINFO
2284 && ins.flags)
2285 {
2286 switch (ins.flags)
2287 {
2288 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2289 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2290 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2291 }
2292 }
2293
2294 if (wined3d_shader_instruction_has_texel_offset(&ins))
2295 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2296
2297 for (i = 0; i < ins.dst_count; ++i)
2298 {
2299 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2300 shader_addline(&buffer, !i ? " " : ", ");
2301 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2302 }
2303
2304 /* Other source tokens */
2305 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2306 {
2307 shader_addline(&buffer, !i ? " " : ", ");
2308 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2309 }
2310 }
2311 shader_addline(&buffer, "\n");
2312 }
2313
2314 for (p = buffer.buffer; *p; p = q)
2315 {
2316 if (!(q = strstr(p, "\n")))
2317 q = p + strlen(p);
2318 else
2319 ++q;
2320 TRACE(" %.*s", (int)(q - p), p);
2321 }
2322
2323 string_buffer_free(&buffer);
2324 }
2325
2326 static void shader_cleanup(struct wined3d_shader *shader)
2327 {
2328 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2329 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2330 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2331 shader->device->shader_backend->shader_destroy(shader);
2332 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
2333 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
2334 HeapFree(GetProcessHeap(), 0, shader->function);
2335 shader_delete_constant_list(&shader->constantsF);
2336 shader_delete_constant_list(&shader->constantsB);
2337 shader_delete_constant_list(&shader->constantsI);
2338 list_remove(&shader->shader_list_entry);
2339
2340 if (shader->frontend && shader->frontend_data)
2341 shader->frontend->shader_free(shader->frontend_data);
2342 }
2343
2344 struct shader_none_priv
2345 {
2346 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2347 const struct fragment_pipeline *fragment_pipe;
2348 BOOL ffp_proj_control;
2349 };
2350
2351 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2352 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2353 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2354 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2355 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2356 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2357 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2358 const struct wined3d_state *state) {}
2359 static void shader_none_destroy(struct wined3d_shader *shader) {}
2360 static void shader_none_free_context_data(struct wined3d_context *context) {}
2361 static void shader_none_init_context_state(struct wined3d_context *context) {}
2362
2363 /* Context activation is done by the caller. */
2364 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2365 const struct wined3d_state *state)
2366 {
2367 const struct wined3d_gl_info *gl_info = context->gl_info;
2368 struct shader_none_priv *priv = shader_priv;
2369
2370 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2371 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2372 }
2373
2374 /* Context activation is done by the caller. */
2375 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2376 {
2377 struct shader_none_priv *priv = shader_priv;
2378 const struct wined3d_gl_info *gl_info = context->gl_info;
2379
2380 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2381 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2382
2383 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2384 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2385 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2386 | (1u << WINED3D_SHADER_TYPE_HULL)
2387 | (1u << WINED3D_SHADER_TYPE_DOMAIN);
2388 }
2389
2390 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2391 const struct fragment_pipeline *fragment_pipe)
2392 {
2393 struct fragment_caps fragment_caps;
2394 void *vertex_priv, *fragment_priv;
2395 struct shader_none_priv *priv;
2396
2397 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2398 return E_OUTOFMEMORY;
2399
2400 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2401 {
2402 ERR("Failed to initialize vertex pipe.\n");
2403 HeapFree(GetProcessHeap(), 0, priv);
2404 return E_FAIL;
2405 }
2406
2407 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2408 {
2409 ERR("Failed to initialize fragment pipe.\n");
2410 vertex_pipe->vp_free(device);
2411 HeapFree(GetProcessHeap(), 0, priv);
2412 return E_FAIL;
2413 }
2414
2415 priv->vertex_pipe = vertex_pipe;
2416 priv->fragment_pipe = fragment_pipe;
2417 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2418 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2419
2420 device->vertex_priv = vertex_priv;
2421 device->fragment_priv = fragment_priv;
2422 device->shader_priv = priv;
2423
2424 return WINED3D_OK;
2425 }
2426
2427 static void shader_none_free(struct wined3d_device *device)
2428 {
2429 struct shader_none_priv *priv = device->shader_priv;
2430
2431 priv->fragment_pipe->free_private(device);
2432 priv->vertex_pipe->vp_free(device);
2433 HeapFree(GetProcessHeap(), 0, priv);
2434 }
2435
2436 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2437 {
2438 return TRUE;
2439 }
2440
2441 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2442 {
2443 /* Set the shader caps to 0 for the none shader backend */
2444 caps->vs_version = 0;
2445 caps->hs_version = 0;
2446 caps->ds_version = 0;
2447 caps->gs_version = 0;
2448 caps->ps_version = 0;
2449 caps->vs_uniform_count = 0;
2450 caps->ps_uniform_count = 0;
2451 caps->ps_1x_max_value = 0.0f;
2452 caps->varying_count = 0;
2453 caps->wined3d_caps = 0;
2454 }
2455
2456 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2457 {
2458 /* We "support" every possible fixup, since we don't support any shader
2459 * model, and will never have to actually sample a texture. */
2460 return TRUE;
2461 }
2462
2463 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2464 {
2465 struct shader_none_priv *priv = shader_priv;
2466
2467 return priv->ffp_proj_control;
2468 }
2469
2470 const struct wined3d_shader_backend_ops none_shader_backend =
2471 {
2472 shader_none_handle_instruction,
2473 shader_none_select,
2474 shader_none_disable,
2475 shader_none_select_depth_blt,
2476 shader_none_deselect_depth_blt,
2477 shader_none_update_float_vertex_constants,
2478 shader_none_update_float_pixel_constants,
2479 shader_none_load_constants,
2480 shader_none_destroy,
2481 shader_none_alloc,
2482 shader_none_free,
2483 shader_none_allocate_context_data,
2484 shader_none_free_context_data,
2485 shader_none_init_context_state,
2486 shader_none_get_caps,
2487 shader_none_color_fixup_supported,
2488 shader_none_has_ffp_proj_control,
2489 };
2490
2491 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2492 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2493 enum wined3d_shader_type type, unsigned int max_version)
2494 {
2495 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2496 const struct wined3d_shader_frontend *fe;
2497 HRESULT hr;
2498 unsigned int backend_version;
2499 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2500
2501 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u, type %#x, max_version %u.\n",
2502 shader, byte_code, output_signature, float_const_count, type, max_version);
2503
2504 list_init(&shader->constantsF);
2505 list_init(&shader->constantsB);
2506 list_init(&shader->constantsI);
2507 shader->lconst_inf_or_nan = FALSE;
2508
2509 fe = shader_select_frontend(*byte_code);
2510 if (!fe)
2511 {
2512 FIXME("Unable to find frontend for shader.\n");
2513 return WINED3DERR_INVALIDCALL;
2514 }
2515 shader->frontend = fe;
2516 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2517 if (!shader->frontend_data)
2518 {
2519 FIXME("Failed to initialize frontend.\n");
2520 return WINED3DERR_INVALIDCALL;
2521 }
2522
2523 /* First pass: trace shader. */
2524 if (TRACE_ON(d3d_shader))
2525 shader_trace_init(fe, shader->frontend_data, byte_code);
2526
2527 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2528 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2529 &shader->output_signature, byte_code, float_const_count)))
2530 return hr;
2531
2532 if (reg_maps->shader_version.type != type)
2533 {
2534 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2535 return WINED3DERR_INVALIDCALL;
2536 }
2537 if (reg_maps->shader_version.major > max_version)
2538 {
2539 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2540 return WINED3DERR_INVALIDCALL;
2541 }
2542 switch (type)
2543 {
2544 case WINED3D_SHADER_TYPE_VERTEX:
2545 backend_version = d3d_info->limits.vs_version;
2546 break;
2547 case WINED3D_SHADER_TYPE_HULL:
2548 backend_version = d3d_info->limits.hs_version;
2549 break;
2550 case WINED3D_SHADER_TYPE_DOMAIN:
2551 backend_version = d3d_info->limits.ds_version;
2552 break;
2553 case WINED3D_SHADER_TYPE_GEOMETRY:
2554 backend_version = d3d_info->limits.gs_version;
2555 break;
2556 case WINED3D_SHADER_TYPE_PIXEL:
2557 backend_version = d3d_info->limits.ps_version;
2558 break;
2559 default:
2560 FIXME("No backend version-checking for this shader type\n");
2561 backend_version = 0;
2562 }
2563 if (reg_maps->shader_version.major > backend_version)
2564 {
2565 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2566 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2567 return WINED3DERR_INVALIDCALL;
2568 }
2569
2570 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2571 if (!shader->function)
2572 return E_OUTOFMEMORY;
2573 memcpy(shader->function, byte_code, shader->functionLength);
2574
2575 return WINED3D_OK;
2576 }
2577
2578 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2579 {
2580 ULONG refcount = InterlockedIncrement(&shader->ref);
2581
2582 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2583
2584 return refcount;
2585 }
2586
2587 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2588 {
2589 ULONG refcount = InterlockedDecrement(&shader->ref);
2590
2591 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2592
2593 if (!refcount)
2594 {
2595 shader_cleanup(shader);
2596 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2597 HeapFree(GetProcessHeap(), 0, shader);
2598 }
2599
2600 return refcount;
2601 }
2602
2603 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2604 {
2605 TRACE("shader %p.\n", shader);
2606
2607 return shader->parent;
2608 }
2609
2610 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2611 void *byte_code, UINT *byte_code_size)
2612 {
2613 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2614
2615 if (!byte_code)
2616 {
2617 *byte_code_size = shader->functionLength;
2618 return WINED3D_OK;
2619 }
2620
2621 if (*byte_code_size < shader->functionLength)
2622 {
2623 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2624 * than the required size we should write the required size and
2625 * return D3DERR_MOREDATA. That's not actually true. */
2626 return WINED3DERR_INVALIDCALL;
2627 }
2628
2629 memcpy(byte_code, shader->function, shader->functionLength);
2630
2631 return WINED3D_OK;
2632 }
2633
2634 /* Set local constants for d3d8 shaders. */
2635 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2636 UINT start_idx, const float *src_data, UINT count)
2637 {
2638 UINT end_idx = start_idx + count;
2639 UINT i;
2640
2641 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2642
2643 if (end_idx > shader->limits->constant_float)
2644 {
2645 WARN("end_idx %u > float constants limit %u.\n",
2646 end_idx, shader->limits->constant_float);
2647 end_idx = shader->limits->constant_float;
2648 }
2649
2650 for (i = start_idx; i < end_idx; ++i)
2651 {
2652 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2653 float *value;
2654 if (!lconst)
2655 return E_OUTOFMEMORY;
2656
2657 lconst->idx = i;
2658 value = (float *)lconst->value;
2659 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2660 list_add_head(&shader->constantsF, &lconst->entry);
2661
2662 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2663 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2664 {
2665 shader->lconst_inf_or_nan = TRUE;
2666 }
2667 }
2668
2669 return WINED3D_OK;
2670 }
2671
2672 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2673 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2674 {
2675 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2676 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2677 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2678 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2679 args->point_size = state->gl_primitive_type == GL_POINTS;
2680 args->per_vertex_point_size = shader->reg_maps.point_size;
2681 if (shader->reg_maps.shader_version.major >= 4)
2682 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
2683 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
2684 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
2685 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
2686 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
2687 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
2688 else
2689 args->next_shader_input_count = 0;
2690 args->swizzle_map = swizzle_map;
2691 if (d3d_info->emulated_flatshading)
2692 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2693 else
2694 args->flatshading = 0;
2695 }
2696
2697 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2698 {
2699 if (usage_idx1 != usage_idx2)
2700 return FALSE;
2701 if (usage1 == usage2)
2702 return TRUE;
2703 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2704 return TRUE;
2705 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2706 return TRUE;
2707
2708 return FALSE;
2709 }
2710
2711 BOOL vshader_get_input(const struct wined3d_shader *shader,
2712 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2713 {
2714 WORD map = shader->reg_maps.input_registers;
2715 unsigned int i;
2716
2717 for (i = 0; map; map >>= 1, ++i)
2718 {
2719 if (!(map & 1)) continue;
2720
2721 if (match_usage(shader->u.vs.attributes[i].usage,
2722 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2723 {
2724 *regnum = i;
2725 return TRUE;
2726 }
2727 }
2728 return FALSE;
2729 }
2730
2731 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2732 const struct wined3d_shader_signature *src, char **signature_strings)
2733 {
2734 struct wined3d_shader_signature_element *e;
2735 unsigned int i;
2736 SIZE_T len;
2737 char *ptr;
2738
2739 ptr = *signature_strings;
2740
2741 dst->element_count = src->element_count;
2742 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2743 return E_OUTOFMEMORY;
2744
2745 for (i = 0; i < src->element_count; ++i)
2746 {
2747 e = &src->elements[i];
2748 dst->elements[i] = *e;
2749
2750 len = strlen(e->semantic_name);
2751 memcpy(ptr, e->semantic_name, len + 1);
2752 dst->elements[i].semantic_name = ptr;
2753 ptr += len + 1;
2754 }
2755
2756 *signature_strings = ptr;
2757
2758 return WINED3D_OK;
2759 }
2760
2761 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2762 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2763 void *parent, const struct wined3d_parent_ops *parent_ops)
2764 {
2765 struct wined3d_shader_signature_element *e;
2766 SIZE_T total, len;
2767 unsigned int i;
2768 HRESULT hr;
2769 char *ptr;
2770
2771 if (!desc->byte_code)
2772 return WINED3DERR_INVALIDCALL;
2773
2774 shader->ref = 1;
2775 shader->device = device;
2776 shader->parent = parent;
2777 shader->parent_ops = parent_ops;
2778
2779 total = 0;
2780 if (desc->input_signature)
2781 {
2782 for (i = 0; i < desc->input_signature->element_count; ++i)
2783 {
2784 e = &desc->input_signature->elements[i];
2785 len = strlen(e->semantic_name);
2786 if (len >= ~(SIZE_T)0 - total)
2787 return E_OUTOFMEMORY;
2788
2789 total += len + 1;
2790 }
2791 }
2792 if (desc->output_signature)
2793 {
2794 for (i = 0; i < desc->output_signature->element_count; ++i)
2795 {
2796 e = &desc->output_signature->elements[i];
2797 len = strlen(e->semantic_name);
2798 if (len >= ~(SIZE_T)0 - total)
2799 return E_OUTOFMEMORY;
2800
2801 total += len + 1;
2802 }
2803 }
2804 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2805 return E_OUTOFMEMORY;
2806 ptr = shader->signature_strings;
2807
2808 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2809 desc->input_signature, &ptr)))
2810 {
2811 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2812 return hr;
2813 }
2814 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2815 desc->output_signature, &ptr)))
2816 {
2817 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2818 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2819 return hr;
2820 }
2821
2822 list_init(&shader->linked_programs);
2823 list_add_head(&device->shaders, &shader->shader_list_entry);
2824
2825 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2826 float_const_count, type, desc->max_version)))
2827 {
2828 WARN("Failed to set function, hr %#x.\n", hr);
2829 shader_cleanup(shader);
2830 }
2831
2832 return hr;
2833 }
2834
2835 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2836 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2837 {
2838 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2839 unsigned int i;
2840 HRESULT hr;
2841
2842 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2843 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2844 return hr;
2845
2846 for (i = 0; i < shader->input_signature.element_count; ++i)
2847 {
2848 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2849
2850 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2851 continue;
2852
2853 shader->u.vs.attributes[input->register_idx].usage =
2854 shader_usage_from_semantic_name(input->semantic_name);
2855 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2856 }
2857
2858 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2859 shader->lconst_inf_or_nan;
2860
2861 return WINED3D_OK;
2862 }
2863
2864 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2865 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2866 {
2867 HRESULT hr;
2868
2869 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops)))
2870 return hr;
2871
2872 shader->load_local_constsF = shader->lconst_inf_or_nan;
2873
2874 return WINED3D_OK;
2875 }
2876
2877 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2878 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2879 {
2880 HRESULT hr;
2881
2882 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops)))
2883 return hr;
2884
2885 shader->load_local_constsF = shader->lconst_inf_or_nan;
2886
2887 return WINED3D_OK;
2888 }
2889
2890 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2891 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2892 {
2893 HRESULT hr;
2894
2895 if (FAILED(hr = shader_init(shader, device, desc, 0,
2896 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2897 return hr;
2898
2899 shader->load_local_constsF = shader->lconst_inf_or_nan;
2900
2901 return WINED3D_OK;
2902 }
2903
2904 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2905 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2906 {
2907 const struct wined3d_gl_info *gl_info = context->gl_info;
2908 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2909 const struct wined3d_texture *texture;
2910 UINT i;
2911
2912 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2913 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
2914 {
2915 static unsigned int warned = 0;
2916
2917 args->srgb_correction = 1;
2918 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2919 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2920 "support, expect rendering artifacts.\n");
2921 }
2922
2923 if (shader->reg_maps.shader_version.major == 1
2924 && shader->reg_maps.shader_version.minor <= 3)
2925 {
2926 for (i = 0; i < shader->limits->sampler; ++i)
2927 {
2928 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2929
2930 if (flags & WINED3D_TTFF_PROJECTED)
2931 {
2932 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2933
2934 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2935 {
2936 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2937 unsigned int j;
2938 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2939 DWORD max_valid = WINED3D_TTFF_COUNT4;
2940
2941 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2942 {
2943 struct wined3d_vertex_declaration_element *element =
2944 &state->vertex_declaration->elements[j];
2945
2946 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2947 && element->usage_idx == index)
2948 {
2949 max_valid = element->format->component_count;
2950 break;
2951 }
2952 }
2953 if (!tex_transform || tex_transform > max_valid)
2954 {
2955 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2956 tex_transform, max_valid);
2957 tex_transform = max_valid;
2958 }
2959 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2960 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2961 && tex_transform > WINED3D_TTFF_COUNT2)
2962 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2963 && tex_transform > WINED3D_TTFF_COUNT3))
2964 tex_transform |= WINED3D_PSARGS_PROJECTED;
2965 else
2966 {
2967 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2968 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2969 i, tex_transform, resource_type);
2970 }
2971 }
2972 else
2973 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2974
2975 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2976 }
2977 }
2978 }
2979 if (shader->reg_maps.shader_version.major == 1
2980 && shader->reg_maps.shader_version.minor <= 4)
2981 {
2982 for (i = 0; i < shader->limits->sampler; ++i)
2983 {
2984 const struct wined3d_texture *texture = state->textures[i];
2985
2986 if (!shader->reg_maps.resource_info[i].type)
2987 continue;
2988
2989 /* Treat unbound textures as 2D. The dummy texture will provide
2990 * the proper sample value. The tex_types bitmap defaults to
2991 * 2D because of the memset. */
2992 if (!texture)
2993 continue;
2994
2995 switch (texture->target)
2996 {
2997 /* RECT textures are distinguished from 2D textures via np2_fixup */
2998 case GL_TEXTURE_RECTANGLE_ARB:
2999 case GL_TEXTURE_2D:
3000 break;
3001
3002 case GL_TEXTURE_3D:
3003 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3004 break;
3005
3006 case GL_TEXTURE_CUBE_MAP_ARB:
3007 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3008 break;
3009 }
3010 }
3011 }
3012
3013 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3014 {
3015 if (!shader->reg_maps.resource_info[i].type)
3016 continue;
3017
3018 texture = state->textures[i];
3019 if (!texture)
3020 {
3021 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3022 continue;
3023 }
3024 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3025 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3026 else
3027 args->color_fixup[i] = texture->resource.format->color_fixup;
3028
3029 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3030 args->shadow |= 1u << i;
3031
3032 /* Flag samplers that need NP2 texcoord fixup. */
3033 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3034 args->np2_fixup |= (1u << i);
3035 }
3036
3037 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3038 if (shader->reg_maps.shader_version.major >= 4)
3039 args->shadow = 0;
3040
3041 if (shader->reg_maps.shader_version.major >= 3)
3042 {
3043 if (position_transformed)
3044 args->vp_mode = pretransformed;
3045 else if (use_vs(state))
3046 args->vp_mode = vertexshader;
3047 else
3048 args->vp_mode = fixedfunction;
3049 args->fog = WINED3D_FFP_PS_FOG_OFF;
3050 }
3051 else
3052 {
3053 args->vp_mode = vertexshader;
3054 if (state->render_states[WINED3D_RS_FOGENABLE])
3055 {
3056 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3057 {
3058 case WINED3D_FOG_NONE:
3059 if (position_transformed || use_vs(state))
3060 {
3061 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3062 break;
3063 }
3064
3065 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3066 {
3067 case WINED3D_FOG_NONE: /* Fall through. */
3068 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3069 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3070 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3071 }
3072 break;
3073
3074 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3075 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3076 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3077 }
3078 }
3079 else
3080 {
3081 args->fog = WINED3D_FFP_PS_FOG_OFF;
3082 }
3083 }
3084
3085 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3086 {
3087 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3088
3089 args->texcoords_initialized = 0;
3090 for (i = 0; i < MAX_TEXTURES; ++i)
3091 {
3092 if (vs)
3093 {
3094 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3095 args->texcoords_initialized |= 1u << i;
3096 }
3097 else
3098 {
3099 const struct wined3d_stream_info *si = &context->stream_info;
3100 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3101
3102 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3103 & WINED3D_FFP_TCI_MASK
3104 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3105 args->texcoords_initialized |= 1u << i;
3106 }
3107 }
3108 }
3109 else
3110 {
3111 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3112 }
3113
3114 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3115 && state->gl_primitive_type == GL_POINTS;
3116
3117 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3118 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3119 else
3120 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3121 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3122 : WINED3D_CMP_ALWAYS) - 1;
3123
3124 if (d3d_info->emulated_flatshading)
3125 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3126 }
3127
3128 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3129 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3130 {
3131 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3132 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3133 HRESULT hr;
3134
3135 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3136 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3137 return hr;
3138
3139 for (i = 0; i < MAX_REG_INPUT; ++i)
3140 {
3141 if (shader->u.ps.input_reg_used[i])
3142 {
3143 ++num_regs_used;
3144 highest_reg_used = i;
3145 }
3146 }
3147
3148 /* Don't do any register mapping magic if it is not needed, or if we can't
3149 * achieve anything anyway */
3150 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3151 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3152 || shader->reg_maps.shader_version.major >= 4)
3153 {
3154 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3155 {
3156 /* This happens with relative addressing. The input mapper function
3157 * warns about this if the higher registers are declared too, so
3158 * don't write a FIXME here */
3159 WARN("More varying registers used than supported\n");
3160 }
3161
3162 for (i = 0; i < MAX_REG_INPUT; ++i)
3163 {
3164 shader->u.ps.input_reg_map[i] = i;
3165 }
3166
3167 shader->u.ps.declared_in_count = highest_reg_used + 1;
3168 }
3169 else
3170 {
3171 shader->u.ps.declared_in_count = 0;
3172 for (i = 0; i < MAX_REG_INPUT; ++i)
3173 {
3174 if (shader->u.ps.input_reg_used[i])
3175 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3176 else shader->u.ps.input_reg_map[i] = ~0U;
3177 }
3178 }
3179
3180 shader->load_local_constsF = shader->lconst_inf_or_nan;
3181
3182 return WINED3D_OK;
3183 }
3184
3185 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3186 {
3187 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3188 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3189 unsigned int i;
3190
3191 if (reg_maps->shader_version.major != 1) return;
3192
3193 for (i = 0; i < shader->limits->sampler; ++i)
3194 {
3195 /* We don't sample from this sampler. */
3196 if (!resource_info[i].type)
3197 continue;
3198
3199 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3200 {
3201 case WINED3D_SHADER_TEX_2D:
3202 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3203 break;
3204
3205 case WINED3D_SHADER_TEX_3D:
3206 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3207 break;
3208
3209 case WINED3D_SHADER_TEX_CUBE:
3210 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3211 break;
3212 }
3213 }
3214 }
3215
3216 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3217 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3218 {
3219 struct wined3d_shader *object;
3220 HRESULT hr;
3221
3222 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3223 device, desc, parent, parent_ops, shader);
3224
3225 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3226 return E_OUTOFMEMORY;
3227
3228 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3229 {
3230 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3231 HeapFree(GetProcessHeap(), 0, object);
3232 return hr;
3233 }
3234
3235 TRACE("Created domain shader %p.\n", object);
3236 *shader = object;
3237
3238 return WINED3D_OK;
3239 }
3240
3241 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3242 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3243 {
3244 struct wined3d_shader *object;
3245 HRESULT hr;
3246
3247 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3248 device, desc, parent, parent_ops, shader);
3249
3250 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3251 if (!object)
3252 return E_OUTOFMEMORY;
3253
3254 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3255 {
3256 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3257 HeapFree(GetProcessHeap(), 0, object);
3258 return hr;
3259 }
3260
3261 TRACE("Created geometry shader %p.\n", object);
3262 *shader = object;
3263
3264 return WINED3D_OK;
3265 }
3266
3267 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3268 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3269 {
3270 struct wined3d_shader *object;
3271 HRESULT hr;
3272
3273 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3274 device, desc, parent, parent_ops, shader);
3275
3276 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3277 return E_OUTOFMEMORY;
3278
3279 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3280 {
3281 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3282 HeapFree(GetProcessHeap(), 0, object);
3283 return hr;
3284 }
3285
3286 TRACE("Created hull shader %p.\n", object);
3287 *shader = object;
3288
3289 return WINED3D_OK;
3290 }
3291
3292 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3293 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3294 {
3295 struct wined3d_shader *object;
3296 HRESULT hr;
3297
3298 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3299 device, desc, parent, parent_ops, shader);
3300
3301 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3302 if (!object)
3303 return E_OUTOFMEMORY;
3304
3305 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3306 {
3307 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3308 HeapFree(GetProcessHeap(), 0, object);
3309 return hr;
3310 }
3311
3312 TRACE("Created pixel shader %p.\n", object);
3313 *shader = object;
3314
3315 return WINED3D_OK;
3316 }
3317
3318 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3319 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3320 {
3321 struct wined3d_shader *object;
3322 HRESULT hr;
3323
3324 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3325 device, desc, parent, parent_ops, shader);
3326
3327 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3328 if (!object)
3329 return E_OUTOFMEMORY;
3330
3331 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3332 {
3333 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3334 HeapFree(GetProcessHeap(), 0, object);
3335 return hr;
3336 }
3337
3338 TRACE("Created vertex shader %p.\n", object);
3339 *shader = object;
3340
3341 return WINED3D_OK;
3342 }
Windows API implementation