?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
comctl32: Use SetRect() instead of open coding it.
[wine.git] / dlls / wined3d / shader.c
blob8f8b7273a8706a63e664055485bca5145c636925
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_NI */ "round_ni",
161 /* WINED3DSIH_ROUND_PI */ "round_pi",
162 /* WINED3DSIH_ROUND_Z */ "round_z",
163 /* WINED3DSIH_RSQ */ "rsq",
164 /* WINED3DSIH_SAMPLE */ "sample",
165 /* WINED3DSIH_SAMPLE_B */ "sample_b",
166 /* WINED3DSIH_SAMPLE_C */ "sample_c",
167 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
168 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
169 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
170 /* WINED3DSIH_SETP */ "setp",
171 /* WINED3DSIH_SGE */ "sge",
172 /* WINED3DSIH_SGN */ "sgn",
173 /* WINED3DSIH_SINCOS */ "sincos",
174 /* WINED3DSIH_SLT */ "slt",
175 /* WINED3DSIH_SQRT */ "sqrt",
176 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
177 /* WINED3DSIH_SUB */ "sub",
178 /* WINED3DSIH_TEX */ "texld",
179 /* WINED3DSIH_TEXBEM */ "texbem",
180 /* WINED3DSIH_TEXBEML */ "texbeml",
181 /* WINED3DSIH_TEXCOORD */ "texcrd",
182 /* WINED3DSIH_TEXDEPTH */ "texdepth",
183 /* WINED3DSIH_TEXDP3 */ "texdp3",
184 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
185 /* WINED3DSIH_TEXKILL */ "texkill",
186 /* WINED3DSIH_TEXLDD */ "texldd",
187 /* WINED3DSIH_TEXLDL */ "texldl",
188 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
189 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
190 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
191 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
192 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
193 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
194 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
195 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
196 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
197 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
198 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
199 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
200 /* WINED3DSIH_UDIV */ "udiv",
201 /* WINED3DSIH_UGE */ "uge",
202 /* WINED3DSIH_ULT */ "ult",
203 /* WINED3DSIH_USHR */ "ushr",
204 /* WINED3DSIH_UTOF */ "utof",
205 /* WINED3DSIH_XOR */ "xor",
206 };
207
208 static const char * const semantic_names[] =
209 {
210 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
211 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
212 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
213 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
214 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
215 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
216 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
217 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
218 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
219 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
220 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
221 /* WINED3D_DECL_USAGE_FOG */ "FOG",
222 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
223 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
224 };
225
226 static const struct
227 {
228 enum wined3d_sysval_semantic sysval_semantic;
229 const char *sysval_name;
230 }
231 sysval_semantic_names[] =
232 {
233 {WINED3D_SV_POSITION, "SV_Position"},
234 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
235 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
236 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
237 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
238 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
239 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
240 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
241 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
242 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
243 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
244 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
245 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
246 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
247 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
248 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
249 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
250 };
251
252 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
253 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
254
255 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
256 {
257 if (handler_idx >= sizeof(shader_opcode_names) / sizeof(*shader_opcode_names))
258 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
259
260 return shader_opcode_names[handler_idx];
261 }
262
263 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
264 {
265 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
266 {
267 FIXME("Unrecognized usage %#x.\n", usage);
268 return "UNRECOGNIZED";
269 }
270
271 return semantic_names[usage];
272 }
273
274 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
275 {
276 unsigned int i;
277
278 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
279 {
280 if (!strcmp(name, semantic_names[i])) return i;
281 }
282
283 return ~0U;
284 }
285
286 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
287 {
288 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
289 }
290
291 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
292 const struct wined3d_shader_semantic *s)
293 {
294 e->semantic_name = shader_semantic_name_from_usage(s->usage);
295 e->semantic_idx = s->usage_idx;
296 e->sysval_semantic = 0;
297 e->component_type = 0;
298 e->register_idx = s->reg.reg.idx[0].offset;
299 e->mask = s->reg.write_mask;
300 }
301
302 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
303 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
304 {
305 e->semantic_name = shader_semantic_name_from_usage(usage);
306 e->semantic_idx = usage_idx;
307 e->sysval_semantic = 0;
308 e->component_type = 0;
309 e->register_idx = reg_idx;
310 e->mask = write_mask;
311 }
312
313 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
314 {
315 switch (version_token >> 16)
316 {
317 case WINED3D_SM1_VS:
318 case WINED3D_SM1_PS:
319 return &sm1_shader_frontend;
320
321 case WINED3D_SM4_PS:
322 case WINED3D_SM4_VS:
323 case WINED3D_SM4_GS:
324 case WINED3D_SM5_HS:
325 case WINED3D_SM5_DS:
326 return &sm4_shader_frontend;
327
328 default:
329 FIXME("Unrecognised version token %#x.\n", version_token);
330 return NULL;
331 }
332 }
333
334 void string_buffer_clear(struct wined3d_string_buffer *buffer)
335 {
336 buffer->buffer[0] = '\0';
337 buffer->content_size = 0;
338 }
339
340 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
341 {
342 buffer->buffer_size = 32;
343 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
344 {
345 ERR("Failed to allocate shader buffer memory.\n");
346 return FALSE;
347 }
348
349 string_buffer_clear(buffer);
350 return TRUE;
351 }
352
353 void string_buffer_free(struct wined3d_string_buffer *buffer)
354 {
355 HeapFree(GetProcessHeap(), 0, buffer->buffer);
356 }
357
358 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
359 {
360 char *new_buffer;
361 unsigned int new_buffer_size = buffer->buffer_size * 2;
362
363 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
364 new_buffer_size *= 2;
365 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
366 {
367 ERR("Failed to grow buffer.\n");
368 buffer->buffer[buffer->content_size] = '\0';
369 return FALSE;
370 }
371 buffer->buffer = new_buffer;
372 buffer->buffer_size = new_buffer_size;
373 return TRUE;
374 }
375
376 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
377 {
378 unsigned int rem;
379 int rc;
380
381 rem = buffer->buffer_size - buffer->content_size;
382 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
383 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
384 return rc;
385
386 buffer->content_size += rc;
387 return 0;
388 }
389
390 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
391 {
392 va_list args;
393 int ret;
394
395 for (;;)
396 {
397 va_start(args, format);
398 ret = shader_vaddline(buffer, format, args);
399 va_end(args);
400 if (!ret)
401 return ret;
402 if (!string_buffer_resize(buffer, ret))
403 return -1;
404 }
405 }
406
407 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
408 {
409 struct wined3d_string_buffer *buffer;
410
411 if (list_empty(&list->list))
412 {
413 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
414 if (!buffer || !string_buffer_init(buffer))
415 {
416 ERR("Couldn't allocate buffer for temporary string.\n");
417 if (buffer)
418 HeapFree(GetProcessHeap(), 0, buffer);
419 return NULL;
420 }
421 }
422 else
423 {
424 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
425 list_remove(&buffer->entry);
426 }
427 string_buffer_clear(buffer);
428 return buffer;
429 }
430
431 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
432 {
433 if (!buffer)
434 return 0;
435 string_buffer_clear(buffer);
436 return shader_vaddline(buffer, format, args);
437 }
438
439 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
440 {
441 va_list args;
442 int ret;
443
444 for (;;)
445 {
446 va_start(args, format);
447 ret = string_buffer_vsprintf(buffer, format, args);
448 va_end(args);
449 if (!ret)
450 return;
451 if (!string_buffer_resize(buffer, ret))
452 return;
453 }
454 }
455
456 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
457 {
458 if (!buffer)
459 return;
460 list_add_head(&list->list, &buffer->entry);
461 }
462
463 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
464 {
465 list_init(&list->list);
466 }
467
468 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
469 {
470 struct wined3d_string_buffer *buffer, *buffer_next;
471
472 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
473 {
474 string_buffer_free(buffer);
475 HeapFree(GetProcessHeap(), 0, buffer);
476 }
477 list_init(&list->list);
478 }
479
480 /* Convert floating point offset relative to a register file to an absolute
481 * offset for float constants. */
482 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
483 {
484 switch (register_type)
485 {
486 case WINED3DSPR_CONST: return register_idx;
487 case WINED3DSPR_CONST2: return 2048 + register_idx;
488 case WINED3DSPR_CONST3: return 4096 + register_idx;
489 case WINED3DSPR_CONST4: return 6144 + register_idx;
490 default:
491 FIXME("Unsupported register type: %u.\n", register_type);
492 return register_idx;
493 }
494 }
495
496 static void shader_delete_constant_list(struct list *clist)
497 {
498 struct wined3d_shader_lconst *constant, *constant_next;
499
500 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
501 HeapFree(GetProcessHeap(), 0, constant);
502 list_init(clist);
503 }
504
505 static void shader_set_limits(struct wined3d_shader *shader)
506 {
507 static const struct limits_entry
508 {
509 unsigned int min_version;
510 unsigned int max_version;
511 struct wined3d_shader_limits limits;
512 }
513 vs_limits[] =
514 {
515 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
516 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
517 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
518 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
519 * even though they are capable of supporting much more (GL
520 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
521 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
522 * shaders to 256. */
523 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
524 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
525 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
526 {0}
527 },
528 hs_limits[] =
529 {
530 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
531 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
532 },
533 ds_limits[] =
534 {
535 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
536 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
537 },
538 gs_limits[] =
539 {
540 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
541 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
542 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
543 {0}
544 },
545 ps_limits[] =
546 {
547 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
548 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
549 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
550 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
551 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
552 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
553 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
554 {0}
555 };
556 const struct limits_entry *limits_array;
557 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
558 shader->reg_maps.shader_version.minor);
559 int i = 0;
560
561 switch (shader->reg_maps.shader_version.type)
562 {
563 default:
564 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
565 /* Fall-through. */
566 case WINED3D_SHADER_TYPE_VERTEX:
567 limits_array = vs_limits;
568 break;
569 case WINED3D_SHADER_TYPE_HULL:
570 limits_array = hs_limits;
571 break;
572 case WINED3D_SHADER_TYPE_DOMAIN:
573 limits_array = ds_limits;
574 break;
575 case WINED3D_SHADER_TYPE_GEOMETRY:
576 limits_array = gs_limits;
577 break;
578 case WINED3D_SHADER_TYPE_PIXEL:
579 limits_array = ps_limits;
580 break;
581 }
582
583 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
584 {
585 if (shader_version <= limits_array[i].max_version)
586 {
587 shader->limits = &limits_array[i].limits;
588 break;
589 }
590 ++i;
591 }
592 if (!shader->limits)
593 {
594 FIXME("Unexpected shader version \"%u.%u\".\n",
595 shader->reg_maps.shader_version.major,
596 shader->reg_maps.shader_version.minor);
597 shader->limits = &limits_array[max(0, i - 1)].limits;
598 }
599 }
600
601 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
602 {
603 DWORD idx, shift;
604 idx = bit >> 5;
605 shift = bit & 0x1f;
606 bitmap[idx] |= (1u << shift);
607 }
608
609 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
610 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
611 {
612 switch (reg->type)
613 {
614 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
615 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
616 reg_maps->texcoord |= 1u << reg->idx[0].offset;
617 else
618 reg_maps->address |= 1u << reg->idx[0].offset;
619 break;
620
621 case WINED3DSPR_TEMP:
622 reg_maps->temporary |= 1u << reg->idx[0].offset;
623 break;
624
625 case WINED3DSPR_INPUT:
626 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
627 {
628 if (reg->idx[0].rel_addr)
629 {
630 /* If relative addressing is used, we must assume that all registers
631 * are used. Even if it is a construct like v3[aL], we can't assume
632 * that v0, v1 and v2 aren't read because aL can be negative */
633 unsigned int i;
634 for (i = 0; i < MAX_REG_INPUT; ++i)
635 {
636 shader->u.ps.input_reg_used[i] = TRUE;
637 }
638 }
639 else
640 {
641 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
642 }
643 }
644 else
645 reg_maps->input_registers |= 1u << reg->idx[0].offset;
646 break;
647
648 case WINED3DSPR_RASTOUT:
649 if (reg->idx[0].offset == 1)
650 reg_maps->fog = 1;
651 if (reg->idx[0].offset == 2)
652 reg_maps->point_size = 1;
653 break;
654
655 case WINED3DSPR_MISCTYPE:
656 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
657 {
658 if (!reg->idx[0].offset)
659 reg_maps->vpos = 1;
660 else if (reg->idx[0].offset == 1)
661 reg_maps->usesfacing = 1;
662 }
663 break;
664
665 case WINED3DSPR_CONST:
666 if (reg->idx[0].rel_addr)
667 {
668 if (reg->idx[0].offset < reg_maps->min_rel_offset)
669 reg_maps->min_rel_offset = reg->idx[0].offset;
670 if (reg->idx[0].offset > reg_maps->max_rel_offset)
671 reg_maps->max_rel_offset = reg->idx[0].offset;
672 reg_maps->usesrelconstF = TRUE;
673 }
674 else
675 {
676 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
677 {
678 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
679 return FALSE;
680 }
681 else
682 {
683 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
684 }
685 }
686 break;
687
688 case WINED3DSPR_CONSTINT:
689 if (reg->idx[0].offset >= shader->limits->constant_int)
690 {
691 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
692 return FALSE;
693 }
694 else
695 {
696 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
697 }
698 break;
699
700 case WINED3DSPR_CONSTBOOL:
701 if (reg->idx[0].offset >= shader->limits->constant_bool)
702 {
703 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
704 return FALSE;
705 }
706 else
707 {
708 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
709 }
710 break;
711
712 case WINED3DSPR_COLOROUT:
713 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
714 break;
715
716 default:
717 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
718 reg->type, reg->idx[0].offset, reg->idx[1].offset);
719 break;
720 }
721 return TRUE;
722 }
723
724 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
725 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
726 {
727 struct wined3d_shader_sampler_map_entry *entries, *entry;
728 struct wined3d_shader_sampler_map *map;
729 unsigned int i;
730
731 map = &reg_maps->sampler_map;
732 entries = map->entries;
733 for (i = 0; i < map->count; ++i)
734 {
735 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
736 return;
737 }
738
739 if (!map->size)
740 {
741 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
742 {
743 ERR("Failed to allocate sampler map entries.\n");
744 return;
745 }
746 map->size = 4;
747 map->entries = entries;
748 }
749 else if (map->count == map->size)
750 {
751 size_t new_size = map->size * 2;
752
753 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
754 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
755 {
756 ERR("Failed to resize sampler map entries.\n");
757 return;
758 }
759 map->size = new_size;
760 map->entries = entries;
761 }
762
763 entry = &entries[map->count++];
764 entry->resource_idx = resource_idx;
765 entry->sampler_idx = sampler_idx;
766 entry->bind_idx = bind_idx;
767 }
768
769 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
770 {
771 switch (instr)
772 {
773 case WINED3DSIH_M4x4:
774 case WINED3DSIH_M3x4:
775 return param == 1 ? 3 : 0;
776
777 case WINED3DSIH_M4x3:
778 case WINED3DSIH_M3x3:
779 return param == 1 ? 2 : 0;
780
781 case WINED3DSIH_M3x2:
782 return param == 1 ? 1 : 0;
783
784 default:
785 return 0;
786 }
787 }
788
789 /* Note that this does not count the loop register as an address register. */
790 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
791 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
792 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
793 {
794 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
795 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
796 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
797 void *fe_data = shader->frontend_data;
798 struct wined3d_shader_version shader_version;
799 const DWORD *ptr = byte_code;
800 unsigned int i;
801
802 memset(reg_maps, 0, sizeof(*reg_maps));
803 memset(input_signature_elements, 0, sizeof(input_signature_elements));
804 memset(output_signature_elements, 0, sizeof(output_signature_elements));
805 reg_maps->min_rel_offset = ~0U;
806
807 fe->shader_read_header(fe_data, &ptr, &shader_version);
808 reg_maps->shader_version = shader_version;
809
810 shader_set_limits(shader);
811
812 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
813 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
814 if (!reg_maps->constf)
815 {
816 ERR("Failed to allocate constant map memory.\n");
817 return E_OUTOFMEMORY;
818 }
819
820 while (!fe->shader_is_end(fe_data, &ptr))
821 {
822 struct wined3d_shader_instruction ins;
823
824 /* Fetch opcode. */
825 fe->shader_read_instruction(fe_data, &ptr, &ins);
826
827 /* Unhandled opcode, and its parameters. */
828 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
829 {
830 TRACE("Skipping unrecognized instruction.\n");
831 continue;
832 }
833
834 /* Handle declarations. */
835 if (ins.handler_idx == WINED3DSIH_DCL)
836 {
837 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
838 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
839
840 switch (semantic->reg.reg.type)
841 {
842 /* Mark input registers used. */
843 case WINED3DSPR_INPUT:
844 if (reg_idx >= MAX_REG_INPUT)
845 {
846 ERR("Invalid input register index %u.\n", reg_idx);
847 break;
848 }
849 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
850 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
851 return WINED3DERR_INVALIDCALL;
852 reg_maps->input_registers |= 1u << reg_idx;
853 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
854 break;
855
856 /* Vertex shader: mark 3.0 output registers used, save token. */
857 case WINED3DSPR_OUTPUT:
858 if (reg_idx >= MAX_REG_OUTPUT)
859 {
860 ERR("Invalid output register index %u.\n", reg_idx);
861 break;
862 }
863 reg_maps->output_registers |= 1u << reg_idx;
864 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
865 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
866 reg_maps->fog = 1;
867 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
868 reg_maps->point_size = 1;
869 break;
870
871 case WINED3DSPR_SAMPLER:
872 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
873 case WINED3DSPR_RESOURCE:
874 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
875 {
876 ERR("Invalid resource index %u.\n", reg_idx);
877 break;
878 }
879 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
880 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
881 break;
882
883 default:
884 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
885 break;
886 }
887 }
888 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
889 {
890 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
891 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
892 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
893 else
894 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
895 }
896 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
897 {
898 if (reg_maps->icb)
899 FIXME("Multiple immediate constant buffers.\n");
900 reg_maps->icb = ins.declaration.icb;
901 }
902 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
903 {
904 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
905 shader->u.gs.input_type = ins.declaration.primitive_type;
906 else
907 FIXME("Invalid instruction %#x for shader type %#x.\n",
908 ins.handler_idx, shader_version.type);
909 }
910 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
911 {
912 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
913 shader->u.gs.output_type = ins.declaration.primitive_type;
914 else
915 FIXME("Invalid instruction %#x for shader type %#x.\n",
916 ins.handler_idx, shader_version.type);
917 }
918 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
919 {
920 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
921 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
922 }
923 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
924 {
925 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
926 shader->u.gs.vertices_out = ins.declaration.count;
927 else
928 FIXME("Invalid instruction %#x for shader type %#x.\n",
929 ins.handler_idx, shader_version.type);
930 }
931 else if (ins.handler_idx == WINED3DSIH_DEF)
932 {
933 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
934 float *value;
935 if (!lconst) return E_OUTOFMEMORY;
936
937 lconst->idx = ins.dst[0].reg.idx[0].offset;
938 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
939 value = (float *)lconst->value;
940
941 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
942 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
943 {
944 if (value[0] < -1.0f) value[0] = -1.0f;
945 else if (value[0] > 1.0f) value[0] = 1.0f;
946 if (value[1] < -1.0f) value[1] = -1.0f;
947 else if (value[1] > 1.0f) value[1] = 1.0f;
948 if (value[2] < -1.0f) value[2] = -1.0f;
949 else if (value[2] > 1.0f) value[2] = 1.0f;
950 if (value[3] < -1.0f) value[3] = -1.0f;
951 else if (value[3] > 1.0f) value[3] = 1.0f;
952 }
953
954 list_add_head(&shader->constantsF, &lconst->entry);
955
956 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
957 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
958 {
959 shader->lconst_inf_or_nan = TRUE;
960 }
961 }
962 else if (ins.handler_idx == WINED3DSIH_DEFI)
963 {
964 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
965 if (!lconst) return E_OUTOFMEMORY;
966
967 lconst->idx = ins.dst[0].reg.idx[0].offset;
968 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
969
970 list_add_head(&shader->constantsI, &lconst->entry);
971 reg_maps->local_int_consts |= (1u << lconst->idx);
972 }
973 else if (ins.handler_idx == WINED3DSIH_DEFB)
974 {
975 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
976 if (!lconst) return E_OUTOFMEMORY;
977
978 lconst->idx = ins.dst[0].reg.idx[0].offset;
979 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
980
981 list_add_head(&shader->constantsB, &lconst->entry);
982 reg_maps->local_bool_consts |= (1u << lconst->idx);
983 }
984 /* For subroutine prototypes. */
985 else if (ins.handler_idx == WINED3DSIH_LABEL)
986 {
987 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
988 }
989 /* Set texture, address, temporary registers. */
990 else
991 {
992 BOOL color0_mov = FALSE;
993 unsigned int i;
994
995 /* This will loop over all the registers and try to
996 * make a bitmask of the ones we're interested in.
997 *
998 * Relative addressing tokens are ignored, but that's
999 * okay, since we'll catch any address registers when
1000 * they are initialized (required by spec). */
1001 for (i = 0; i < ins.dst_count; ++i)
1002 {
1003 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1004 shader_version.type, constf_size))
1005 return WINED3DERR_INVALIDCALL;
1006
1007 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1008 {
1009 UINT idx = ins.dst[i].reg.idx[0].offset;
1010
1011 switch (ins.dst[i].reg.type)
1012 {
1013 case WINED3DSPR_RASTOUT:
1014 if (shader_version.major >= 3)
1015 break;
1016 switch (idx)
1017 {
1018 case 0: /* oPos */
1019 reg_maps->output_registers |= 1u << 10;
1020 shader_signature_from_usage(&output_signature_elements[10],
1021 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1022 break;
1023
1024 case 1: /* oFog */
1025 reg_maps->output_registers |= 1u << 11;
1026 shader_signature_from_usage(&output_signature_elements[11],
1027 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1028 break;
1029
1030 case 2: /* oPts */
1031 reg_maps->output_registers |= 1u << 11;
1032 shader_signature_from_usage(&output_signature_elements[11],
1033 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1034 break;
1035 }
1036 break;
1037
1038 case WINED3DSPR_ATTROUT:
1039 if (shader_version.major >= 3)
1040 break;
1041 if (idx < 2)
1042 {
1043 idx += 8;
1044 if (reg_maps->output_registers & (1u << idx))
1045 {
1046 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1047 }
1048 else
1049 {
1050 reg_maps->output_registers |= 1u << idx;
1051 shader_signature_from_usage(&output_signature_elements[idx],
1052 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1053 }
1054 }
1055 break;
1056
1057 case WINED3DSPR_TEXCRDOUT:
1058 if (shader_version.major >= 3)
1059 {
1060 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1061 break;
1062 }
1063 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1064 if (reg_maps->output_registers & (1u << idx))
1065 {
1066 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1067 }
1068 else
1069 {
1070 reg_maps->output_registers |= 1u << idx;
1071 shader_signature_from_usage(&output_signature_elements[idx],
1072 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1073 }
1074 break;
1075
1076 default:
1077 break;
1078 }
1079 }
1080
1081 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1082 {
1083 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1084 {
1085 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1086 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1087 * the mov and perform the sRGB write correction from the source register.
1088 *
1089 * However, if the mov is only partial, we can't do this, and if the write
1090 * comes from an instruction other than MOV it is hard to do as well. If
1091 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1092 shader->u.ps.color0_mov = FALSE;
1093 if (ins.handler_idx == WINED3DSIH_MOV
1094 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1095 {
1096 /* Used later when the source register is read. */
1097 color0_mov = TRUE;
1098 }
1099 }
1100 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1101 * end
1102 */
1103 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1104 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1105 {
1106 shader->u.ps.color0_mov = FALSE;
1107 }
1108 }
1109
1110 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1111 if (shader_version.major == 1
1112 && (ins.handler_idx == WINED3DSIH_TEX
1113 || ins.handler_idx == WINED3DSIH_TEXBEM
1114 || ins.handler_idx == WINED3DSIH_TEXBEML
1115 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1116 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1117 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1118 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1119 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1120 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1121 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1122 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1123 {
1124 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1125
1126 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1127 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1128 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1129 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1130
1131 /* texbem is only valid with < 1.4 pixel shaders */
1132 if (ins.handler_idx == WINED3DSIH_TEXBEM
1133 || ins.handler_idx == WINED3DSIH_TEXBEML)
1134 {
1135 reg_maps->bumpmat |= 1u << reg_idx;
1136 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1137 {
1138 reg_maps->luminanceparams |= 1u << reg_idx;
1139 }
1140 }
1141 }
1142 else if (ins.handler_idx == WINED3DSIH_BEM)
1143 {
1144 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1145 }
1146 }
1147
1148 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1149 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1150 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1151 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1152 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1153 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1154 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1155 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1156 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1157 else if (ins.handler_idx == WINED3DSIH_LOOP
1158 || ins.handler_idx == WINED3DSIH_REP)
1159 {
1160 ++cur_loop_depth;
1161 if (cur_loop_depth > max_loop_depth)
1162 max_loop_depth = cur_loop_depth;
1163 }
1164 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1165 || ins.handler_idx == WINED3DSIH_ENDREP)
1166 {
1167 --cur_loop_depth;
1168 }
1169 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1170 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1171 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1172 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1173 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1174 {
1175 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1176 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1177 }
1178 else if (ins.handler_idx == WINED3DSIH_LD
1179 || ins.handler_idx == WINED3DSIH_RESINFO)
1180 {
1181 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1182 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1183 }
1184
1185 if (ins.predicate)
1186 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1187 shader_version.type, constf_size))
1188 return WINED3DERR_INVALIDCALL;
1189
1190 for (i = 0; i < ins.src_count; ++i)
1191 {
1192 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1193 struct wined3d_shader_register reg = ins.src[i].reg;
1194
1195 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1196 shader_version.type, constf_size))
1197 return WINED3DERR_INVALIDCALL;
1198 while (count)
1199 {
1200 ++reg.idx[0].offset;
1201 if (!shader_record_register_usage(shader, reg_maps, &reg,
1202 shader_version.type, constf_size))
1203 return WINED3DERR_INVALIDCALL;
1204 --count;
1205 }
1206
1207 if (color0_mov)
1208 {
1209 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1210 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1211 {
1212 shader->u.ps.color0_mov = TRUE;
1213 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1214 }
1215 }
1216 }
1217 }
1218 }
1219 reg_maps->loop_depth = max_loop_depth;
1220
1221 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1222 * R0 is written to the render target. */
1223 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1224 reg_maps->rt_mask |= (1u << 0);
1225
1226 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1227
1228 if (input_signature->elements)
1229 {
1230 for (i = 0; i < input_signature->element_count; ++i)
1231 {
1232 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1233 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1234 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1235 reg_maps->vpos = 1;
1236 }
1237 }
1238 else if (!input_signature->elements && reg_maps->input_registers)
1239 {
1240 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1241 struct wined3d_shader_signature_element *e;
1242 unsigned int i;
1243
1244 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1245 return E_OUTOFMEMORY;
1246 input_signature->element_count = count;
1247
1248 e = input_signature->elements;
1249 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1250 {
1251 if (!(reg_maps->input_registers & (1u << i)))
1252 continue;
1253 input_signature_elements[i].register_idx = i;
1254 *e++ = input_signature_elements[i];
1255 }
1256 }
1257
1258 if (output_signature->elements)
1259 {
1260 for (i = 0; i < output_signature->element_count; ++i)
1261 {
1262 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1263 }
1264 }
1265 else if (reg_maps->output_registers)
1266 {
1267 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1268 struct wined3d_shader_signature_element *e;
1269
1270 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1271 return E_OUTOFMEMORY;
1272 output_signature->element_count = count;
1273
1274 e = output_signature->elements;
1275 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1276 {
1277 if (!(reg_maps->output_registers & (1u << i)))
1278 continue;
1279 *e++ = output_signature_elements[i];
1280 }
1281 }
1282
1283 return WINED3D_OK;
1284 }
1285
1286 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1287 {
1288 DWORD map = 1u << max;
1289 map |= map - 1;
1290 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1291
1292 return wined3d_log2i(map);
1293 }
1294
1295 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1296 {
1297 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1298 {
1299 shader_addline(buffer, "refactoringAllowed");
1300 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1301 if (global_flags)
1302 shader_addline(buffer, " | ");
1303 }
1304
1305 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1306 {
1307 shader_addline(buffer, "enableRawAndStructuredBuffers");
1308 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1309 }
1310
1311 if (global_flags)
1312 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1313 }
1314
1315 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1316 enum wined3d_tessellator_domain domain)
1317 {
1318 switch (domain)
1319 {
1320 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1321 shader_addline(buffer, "line");
1322 break;
1323 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1324 shader_addline(buffer, "triangle");
1325 break;
1326 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1327 shader_addline(buffer, "quad");
1328 break;
1329 default:
1330 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1331 break;
1332 }
1333 }
1334
1335 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1336 enum wined3d_tessellator_output_primitive output_primitive)
1337 {
1338 switch (output_primitive)
1339 {
1340 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1341 shader_addline(buffer, "point");
1342 break;
1343 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1344 shader_addline(buffer, "line");
1345 break;
1346 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1347 shader_addline(buffer, "triangle_cw");
1348 break;
1349 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1350 shader_addline(buffer, "triangle_ccw");
1351 break;
1352 default:
1353 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1354 break;
1355 }
1356 }
1357
1358 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1359 enum wined3d_tessellator_partitioning partitioning)
1360 {
1361 switch (partitioning)
1362 {
1363 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1364 shader_addline(buffer, "integer");
1365 break;
1366 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1367 shader_addline(buffer, "pow2");
1368 break;
1369 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1370 shader_addline(buffer, "fractional_odd");
1371 break;
1372 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1373 shader_addline(buffer, "fractional_even");
1374 break;
1375 default:
1376 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1377 break;
1378 }
1379 }
1380
1381 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1382 {
1383 unsigned int i;
1384
1385 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1386 {
1387 if (sysval_semantic_names[i].sysval_semantic == semantic)
1388 {
1389 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1390 return;
1391 }
1392 }
1393
1394 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1395 }
1396
1397 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1398 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1399 {
1400 shader_addline(buffer, "dcl");
1401
1402 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1403 {
1404 switch (semantic->resource_type)
1405 {
1406 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1407 shader_addline(buffer, "_2d");
1408 break;
1409
1410 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1411 shader_addline(buffer, "_3d");
1412 break;
1413
1414 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1415 shader_addline(buffer, "_cube");
1416 break;
1417
1418 default:
1419 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1420 break;
1421 }
1422 }
1423 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1424 {
1425 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1426 shader_addline(buffer, "_resource_");
1427 else
1428 shader_addline(buffer, "_uav_");
1429 switch (semantic->resource_type)
1430 {
1431 case WINED3D_SHADER_RESOURCE_BUFFER:
1432 shader_addline(buffer, "buffer");
1433 break;
1434
1435 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1436 shader_addline(buffer, "texture1d");
1437 break;
1438
1439 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1440 shader_addline(buffer, "texture2d");
1441 break;
1442
1443 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1444 shader_addline(buffer, "texture2dms");
1445 break;
1446
1447 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1448 shader_addline(buffer, "texture3d");
1449 break;
1450
1451 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1452 shader_addline(buffer, "texturecube");
1453 break;
1454
1455 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1456 shader_addline(buffer, "texture1darray");
1457 break;
1458
1459 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1460 shader_addline(buffer, "texture2darray");
1461 break;
1462
1463 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1464 shader_addline(buffer, "texture2dmsarray");
1465 break;
1466
1467 default:
1468 shader_addline(buffer, "unknown");
1469 break;
1470 }
1471 switch (semantic->resource_data_type)
1472 {
1473 case WINED3D_DATA_FLOAT:
1474 shader_addline(buffer, " (float)");
1475 break;
1476
1477 case WINED3D_DATA_INT:
1478 shader_addline(buffer, " (int)");
1479 break;
1480
1481 case WINED3D_DATA_UINT:
1482 shader_addline(buffer, " (uint)");
1483 break;
1484
1485 case WINED3D_DATA_UNORM:
1486 shader_addline(buffer, " (unorm)");
1487 break;
1488
1489 case WINED3D_DATA_SNORM:
1490 shader_addline(buffer, " (snorm)");
1491 break;
1492
1493 default:
1494 shader_addline(buffer, " (unknown)");
1495 break;
1496 }
1497 }
1498 else
1499 {
1500 /* Pixel shaders 3.0 don't have usage semantics. */
1501 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1502 return;
1503 else
1504 shader_addline(buffer, "_");
1505
1506 switch (semantic->usage)
1507 {
1508 case WINED3D_DECL_USAGE_POSITION:
1509 shader_addline(buffer, "position%u", semantic->usage_idx);
1510 break;
1511
1512 case WINED3D_DECL_USAGE_BLEND_INDICES:
1513 shader_addline(buffer, "blend");
1514 break;
1515
1516 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1517 shader_addline(buffer, "weight");
1518 break;
1519
1520 case WINED3D_DECL_USAGE_NORMAL:
1521 shader_addline(buffer, "normal%u", semantic->usage_idx);
1522 break;
1523
1524 case WINED3D_DECL_USAGE_PSIZE:
1525 shader_addline(buffer, "psize");
1526 break;
1527
1528 case WINED3D_DECL_USAGE_COLOR:
1529 if (!semantic->usage_idx)
1530 shader_addline(buffer, "color");
1531 else
1532 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1533 break;
1534
1535 case WINED3D_DECL_USAGE_TEXCOORD:
1536 shader_addline(buffer, "texture%u", semantic->usage_idx);
1537 break;
1538
1539 case WINED3D_DECL_USAGE_TANGENT:
1540 shader_addline(buffer, "tangent");
1541 break;
1542
1543 case WINED3D_DECL_USAGE_BINORMAL:
1544 shader_addline(buffer, "binormal");
1545 break;
1546
1547 case WINED3D_DECL_USAGE_TESS_FACTOR:
1548 shader_addline(buffer, "tessfactor");
1549 break;
1550
1551 case WINED3D_DECL_USAGE_POSITIONT:
1552 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1553 break;
1554
1555 case WINED3D_DECL_USAGE_FOG:
1556 shader_addline(buffer, "fog");
1557 break;
1558
1559 case WINED3D_DECL_USAGE_DEPTH:
1560 shader_addline(buffer, "depth");
1561 break;
1562
1563 case WINED3D_DECL_USAGE_SAMPLE:
1564 shader_addline(buffer, "sample");
1565 break;
1566
1567 default:
1568 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1569 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1570 }
1571 }
1572 }
1573
1574 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1575 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1576 {
1577 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1578 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1579 UINT offset = reg->idx[0].offset;
1580
1581 switch (reg->type)
1582 {
1583 case WINED3DSPR_TEMP:
1584 shader_addline(buffer, "r");
1585 break;
1586
1587 case WINED3DSPR_INPUT:
1588 shader_addline(buffer, "v");
1589 break;
1590
1591 case WINED3DSPR_CONST:
1592 case WINED3DSPR_CONST2:
1593 case WINED3DSPR_CONST3:
1594 case WINED3DSPR_CONST4:
1595 shader_addline(buffer, "c");
1596 offset = shader_get_float_offset(reg->type, offset);
1597 break;
1598
1599 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1600 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1601 break;
1602
1603 case WINED3DSPR_RASTOUT:
1604 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1605 break;
1606
1607 case WINED3DSPR_COLOROUT:
1608 shader_addline(buffer, "oC");
1609 break;
1610
1611 case WINED3DSPR_DEPTHOUT:
1612 shader_addline(buffer, "oDepth");
1613 break;
1614
1615 case WINED3DSPR_ATTROUT:
1616 shader_addline(buffer, "oD");
1617 break;
1618
1619 case WINED3DSPR_TEXCRDOUT:
1620 /* Vertex shaders >= 3.0 use general purpose output registers
1621 * (WINED3DSPR_OUTPUT), which can include an address token. */
1622 if (shader_version->major >= 3)
1623 shader_addline(buffer, "o");
1624 else
1625 shader_addline(buffer, "oT");
1626 break;
1627
1628 case WINED3DSPR_CONSTINT:
1629 shader_addline(buffer, "i");
1630 break;
1631
1632 case WINED3DSPR_CONSTBOOL:
1633 shader_addline(buffer, "b");
1634 break;
1635
1636 case WINED3DSPR_LABEL:
1637 shader_addline(buffer, "l");
1638 break;
1639
1640 case WINED3DSPR_LOOP:
1641 shader_addline(buffer, "aL");
1642 break;
1643
1644 case WINED3DSPR_SAMPLER:
1645 shader_addline(buffer, "s");
1646 break;
1647
1648 case WINED3DSPR_MISCTYPE:
1649 if (offset > 1)
1650 {
1651 FIXME("Unhandled misctype register %u.\n", offset);
1652 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1653 }
1654 else
1655 {
1656 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1657 }
1658 break;
1659
1660 case WINED3DSPR_PREDICATE:
1661 shader_addline(buffer, "p");
1662 break;
1663
1664 case WINED3DSPR_IMMCONST:
1665 shader_addline(buffer, "l");
1666 break;
1667
1668 case WINED3DSPR_CONSTBUFFER:
1669 shader_addline(buffer, "cb");
1670 break;
1671
1672 case WINED3DSPR_IMMCONSTBUFFER:
1673 shader_addline(buffer, "icb");
1674 break;
1675
1676 case WINED3DSPR_PRIMID:
1677 shader_addline(buffer, "primID");
1678 break;
1679
1680 case WINED3DSPR_NULL:
1681 shader_addline(buffer, "null");
1682 break;
1683
1684 case WINED3DSPR_RESOURCE:
1685 shader_addline(buffer, "t");
1686 break;
1687
1688 case WINED3DSPR_UAV:
1689 shader_addline(buffer, "u");
1690 break;
1691
1692 case WINED3DSPR_OUTPOINTID:
1693 shader_addline(buffer, "vOutputControlPointID");
1694 break;
1695
1696 case WINED3DSPR_FORKINSTID:
1697 shader_addline(buffer, "vForkInstanceId");
1698 break;
1699
1700 case WINED3DSPR_INCONTROLPOINT:
1701 shader_addline(buffer, "vicp");
1702 break;
1703
1704 case WINED3DSPR_PATCHCONST:
1705 shader_addline(buffer, "vpc");
1706 break;
1707
1708 case WINED3DSPR_TESSCOORD:
1709 shader_addline(buffer, "vDomainLocation");
1710 break;
1711
1712 default:
1713 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1714 break;
1715 }
1716
1717 if (reg->type == WINED3DSPR_IMMCONST)
1718 {
1719 shader_addline(buffer, "(");
1720 switch (reg->immconst_type)
1721 {
1722 case WINED3D_IMMCONST_SCALAR:
1723 switch (reg->data_type)
1724 {
1725 case WINED3D_DATA_FLOAT:
1726 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1727 break;
1728 case WINED3D_DATA_INT:
1729 shader_addline(buffer, "%d", reg->immconst_data[0]);
1730 break;
1731 case WINED3D_DATA_RESOURCE:
1732 case WINED3D_DATA_SAMPLER:
1733 case WINED3D_DATA_UINT:
1734 shader_addline(buffer, "%u", reg->immconst_data[0]);
1735 break;
1736 default:
1737 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1738 break;
1739 }
1740 break;
1741
1742 case WINED3D_IMMCONST_VEC4:
1743 switch (reg->data_type)
1744 {
1745 case WINED3D_DATA_FLOAT:
1746 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1747 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1748 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1749 break;
1750 case WINED3D_DATA_INT:
1751 shader_addline(buffer, "%d, %d, %d, %d",
1752 reg->immconst_data[0], reg->immconst_data[1],
1753 reg->immconst_data[2], reg->immconst_data[3]);
1754 break;
1755 case WINED3D_DATA_RESOURCE:
1756 case WINED3D_DATA_SAMPLER:
1757 case WINED3D_DATA_UINT:
1758 shader_addline(buffer, "%u, %u, %u, %u",
1759 reg->immconst_data[0], reg->immconst_data[1],
1760 reg->immconst_data[2], reg->immconst_data[3]);
1761 break;
1762 default:
1763 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1764 break;
1765 }
1766 break;
1767
1768 default:
1769 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1770 break;
1771 }
1772 shader_addline(buffer, ")");
1773 }
1774 else if (reg->type != WINED3DSPR_RASTOUT
1775 && reg->type != WINED3DSPR_MISCTYPE
1776 && reg->type != WINED3DSPR_NULL)
1777 {
1778 if (offset != ~0u)
1779 {
1780 shader_addline(buffer, "[");
1781 if (reg->idx[0].rel_addr)
1782 {
1783 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1784 shader_addline(buffer, " + ");
1785 }
1786 shader_addline(buffer, "%u]", offset);
1787
1788 if (reg->idx[1].offset != ~0u)
1789 {
1790 shader_addline(buffer, "[");
1791 if (reg->idx[1].rel_addr)
1792 {
1793 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1794 shader_addline(buffer, " + ");
1795 }
1796 shader_addline(buffer, "%u]", reg->idx[1].offset);
1797 }
1798 }
1799 }
1800 }
1801
1802 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1803 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1804 {
1805 DWORD write_mask = param->write_mask;
1806
1807 shader_dump_register(buffer, &param->reg, shader_version);
1808
1809 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1810 {
1811 static const char write_mask_chars[] = "xyzw";
1812
1813 shader_addline(buffer, ".");
1814 if (write_mask & WINED3DSP_WRITEMASK_0)
1815 shader_addline(buffer, "%c", write_mask_chars[0]);
1816 if (write_mask & WINED3DSP_WRITEMASK_1)
1817 shader_addline(buffer, "%c", write_mask_chars[1]);
1818 if (write_mask & WINED3DSP_WRITEMASK_2)
1819 shader_addline(buffer, "%c", write_mask_chars[2]);
1820 if (write_mask & WINED3DSP_WRITEMASK_3)
1821 shader_addline(buffer, "%c", write_mask_chars[3]);
1822 }
1823 }
1824
1825 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1826 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1827 {
1828 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1829 DWORD swizzle = param->swizzle;
1830
1831 if (src_modifier == WINED3DSPSM_NEG
1832 || src_modifier == WINED3DSPSM_BIASNEG
1833 || src_modifier == WINED3DSPSM_SIGNNEG
1834 || src_modifier == WINED3DSPSM_X2NEG
1835 || src_modifier == WINED3DSPSM_ABSNEG)
1836 shader_addline(buffer, "-");
1837 else if (src_modifier == WINED3DSPSM_COMP)
1838 shader_addline(buffer, "1-");
1839 else if (src_modifier == WINED3DSPSM_NOT)
1840 shader_addline(buffer, "!");
1841
1842 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1843 shader_addline(buffer, "abs(");
1844
1845 shader_dump_register(buffer, &param->reg, shader_version);
1846
1847 switch (src_modifier)
1848 {
1849 case WINED3DSPSM_NONE: break;
1850 case WINED3DSPSM_NEG: break;
1851 case WINED3DSPSM_NOT: break;
1852 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1853 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1854 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1855 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1856 case WINED3DSPSM_COMP: break;
1857 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1858 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1859 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1860 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1861 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1862 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1863 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1864 }
1865
1866 if (swizzle != WINED3DSP_NOSWIZZLE)
1867 {
1868 static const char swizzle_chars[] = "xyzw";
1869 DWORD swizzle_x = swizzle & 0x03;
1870 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1871 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1872 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1873
1874 if (swizzle_x == swizzle_y
1875 && swizzle_x == swizzle_z
1876 && swizzle_x == swizzle_w)
1877 {
1878 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
1879 }
1880 else
1881 {
1882 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1883 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1884 }
1885 }
1886 }
1887
1888 /* Shared code in order to generate the bulk of the shader string.
1889 * NOTE: A description of how to parse tokens can be found on MSDN. */
1890 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1891 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1892 {
1893 struct wined3d_device *device = shader->device;
1894 const struct wined3d_shader_frontend *fe = shader->frontend;
1895 void *fe_data = shader->frontend_data;
1896 struct wined3d_shader_version shader_version;
1897 struct wined3d_shader_loop_state loop_state;
1898 struct wined3d_shader_instruction ins;
1899 struct wined3d_shader_tex_mx tex_mx;
1900 struct wined3d_shader_context ctx;
1901 const DWORD *ptr = byte_code;
1902
1903 /* Initialize current parsing state. */
1904 tex_mx.current_row = 0;
1905 loop_state.current_depth = 0;
1906 loop_state.current_reg = 0;
1907
1908 ctx.shader = shader;
1909 ctx.gl_info = &device->adapter->gl_info;
1910 ctx.reg_maps = reg_maps;
1911 ctx.buffer = buffer;
1912 ctx.tex_mx = &tex_mx;
1913 ctx.loop_state = &loop_state;
1914 ctx.backend_data = backend_ctx;
1915 ins.ctx = &ctx;
1916
1917 fe->shader_read_header(fe_data, &ptr, &shader_version);
1918
1919 while (!fe->shader_is_end(fe_data, &ptr))
1920 {
1921 /* Read opcode. */
1922 fe->shader_read_instruction(fe_data, &ptr, &ins);
1923
1924 /* Unknown opcode and its parameters. */
1925 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1926 {
1927 TRACE("Skipping unrecognized instruction.\n");
1928 continue;
1929 }
1930
1931 if (ins.predicate)
1932 FIXME("Predicates not implemented.\n");
1933
1934 /* Call appropriate function for output target */
1935 device->shader_backend->shader_handle_instruction(&ins);
1936 }
1937 }
1938
1939 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
1940 const struct wined3d_shader_dst_param *dst)
1941 {
1942 DWORD mmask = dst->modifiers;
1943
1944 switch (dst->shift)
1945 {
1946 case 0: break;
1947 case 13: shader_addline(buffer, "_d8"); break;
1948 case 14: shader_addline(buffer, "_d4"); break;
1949 case 15: shader_addline(buffer, "_d2"); break;
1950 case 1: shader_addline(buffer, "_x2"); break;
1951 case 2: shader_addline(buffer, "_x4"); break;
1952 case 3: shader_addline(buffer, "_x8"); break;
1953 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
1954 }
1955
1956 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
1957 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
1958 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
1959
1960 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1961 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1962 }
1963
1964 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
1965 enum wined3d_primitive_type primitive_type)
1966 {
1967 switch (primitive_type)
1968 {
1969 case WINED3D_PT_UNDEFINED:
1970 shader_addline(buffer, "undefined");
1971 break;
1972 case WINED3D_PT_POINTLIST:
1973 shader_addline(buffer, "pointlist");
1974 break;
1975 case WINED3D_PT_LINELIST:
1976 shader_addline(buffer, "linelist");
1977 break;
1978 case WINED3D_PT_LINESTRIP:
1979 shader_addline(buffer, "linestrip");
1980 break;
1981 case WINED3D_PT_TRIANGLELIST:
1982 shader_addline(buffer, "trianglelist");
1983 break;
1984 case WINED3D_PT_TRIANGLESTRIP:
1985 shader_addline(buffer, "trianglestrip");
1986 break;
1987 case WINED3D_PT_TRIANGLEFAN:
1988 shader_addline(buffer, "trianglefan");
1989 break;
1990 case WINED3D_PT_LINELIST_ADJ:
1991 shader_addline(buffer, "linelist_adj");
1992 break;
1993 case WINED3D_PT_LINESTRIP_ADJ:
1994 shader_addline(buffer, "linestrip_adj");
1995 break;
1996 case WINED3D_PT_TRIANGLELIST_ADJ:
1997 shader_addline(buffer, "trianglelist_adj");
1998 break;
1999 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2000 shader_addline(buffer, "trianglestrip_adj");
2001 break;
2002 default:
2003 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2004 break;
2005 }
2006 }
2007
2008 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2009 enum wined3d_shader_interpolation_mode interpolation_mode)
2010 {
2011 switch (interpolation_mode)
2012 {
2013 case WINED3DSIM_CONSTANT:
2014 shader_addline(buffer, "constant");
2015 break;
2016 case WINED3DSIM_LINEAR:
2017 shader_addline(buffer, "linear");
2018 break;
2019 case WINED3DSIM_LINEAR_CENTROID:
2020 shader_addline(buffer, "linear centroid");
2021 break;
2022 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2023 shader_addline(buffer, "linear noperspective");
2024 break;
2025 case WINED3DSIM_LINEAR_SAMPLE:
2026 shader_addline(buffer, "linear sample");
2027 break;
2028 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2029 shader_addline(buffer, "linear noperspective centroid");
2030 break;
2031 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2032 shader_addline(buffer, "linear noperspective sample");
2033 break;
2034 default:
2035 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2036 break;
2037 }
2038 }
2039
2040 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2041 {
2042 struct wined3d_shader_version shader_version;
2043 struct wined3d_string_buffer buffer;
2044 const DWORD *ptr = byte_code;
2045 const char *type_prefix;
2046 const char *p, *q;
2047 DWORD i;
2048
2049 if (!string_buffer_init(&buffer))
2050 {
2051 ERR("Failed to initialize string buffer.\n");
2052 return;
2053 }
2054
2055 TRACE("Parsing %p.\n", byte_code);
2056
2057 fe->shader_read_header(fe_data, &ptr, &shader_version);
2058
2059 switch (shader_version.type)
2060 {
2061 case WINED3D_SHADER_TYPE_VERTEX:
2062 type_prefix = "vs";
2063 break;
2064
2065 case WINED3D_SHADER_TYPE_HULL:
2066 type_prefix = "hs";
2067 break;
2068
2069 case WINED3D_SHADER_TYPE_DOMAIN:
2070 type_prefix = "ds";
2071 break;
2072
2073 case WINED3D_SHADER_TYPE_GEOMETRY:
2074 type_prefix = "gs";
2075 break;
2076
2077 case WINED3D_SHADER_TYPE_PIXEL:
2078 type_prefix = "ps";
2079 break;
2080
2081 default:
2082 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2083 type_prefix = "unknown";
2084 break;
2085 }
2086
2087 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2088
2089 while (!fe->shader_is_end(fe_data, &ptr))
2090 {
2091 struct wined3d_shader_instruction ins;
2092
2093 fe->shader_read_instruction(fe_data, &ptr, &ins);
2094 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2095 {
2096 WARN("Skipping unrecognized instruction.\n");
2097 shader_addline(&buffer, "<unrecognized instruction>\n");
2098 continue;
2099 }
2100
2101 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2102 {
2103 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2104 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2105 shader_addline(&buffer, " ");
2106 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2107 }
2108 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2109 {
2110 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2111 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2112 shader_addline(&buffer, ", %s",
2113 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2114 }
2115 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2116 {
2117 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2118 shader_dump_global_flags(&buffer, ins.flags);
2119 }
2120 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2121 {
2122 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2123 ins.declaration.max_tessellation_factor);
2124 }
2125 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2126 {
2127 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2128 for (i = 0; i < ins.declaration.icb->element_count / 4; ++i)
2129 {
2130 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2131 ins.declaration.icb->data[4 * i + 0],
2132 ins.declaration.icb->data[4 * i + 1],
2133 ins.declaration.icb->data[4 * i + 2],
2134 ins.declaration.icb->data[4 * i + 3]);
2135 }
2136 shader_addline(&buffer, "}");
2137 }
2138 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2139 {
2140 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2141 shader_dump_interpolation_mode(&buffer, ins.flags);
2142 shader_addline(&buffer, " ");
2143 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2144 }
2145 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2146 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2147 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2148 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2149 {
2150 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2151 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2152 shader_addline(&buffer, ", ");
2153 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2154 }
2155 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2156 {
2157 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2158 shader_dump_interpolation_mode(&buffer, ins.flags);
2159 shader_addline(&buffer, " ");
2160 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2161 shader_addline(&buffer, ", ");
2162 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2163 }
2164 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2165 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2166 {
2167 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2168 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2169 }
2170 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2171 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2172 {
2173 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2174 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2175 }
2176 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2177 {
2178 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2179 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2180 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2181 }
2182 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2183 {
2184 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2185 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2186 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2187 shader_addline(&buffer, ", comparisonMode");
2188 }
2189 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2190 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2191 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2192 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2193 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2194 {
2195 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2196 }
2197 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2198 {
2199 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2200 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2201 }
2202 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2203 {
2204 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2205 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2206 }
2207 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2208 {
2209 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2210 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2211 }
2212 else if (ins.handler_idx == WINED3DSIH_DEF)
2213 {
2214 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2215 ins.dst[0].reg.idx[0].offset),
2216 *(const float *)&ins.src[0].reg.immconst_data[0],
2217 *(const float *)&ins.src[0].reg.immconst_data[1],
2218 *(const float *)&ins.src[0].reg.immconst_data[2],
2219 *(const float *)&ins.src[0].reg.immconst_data[3]);
2220 }
2221 else if (ins.handler_idx == WINED3DSIH_DEFI)
2222 {
2223 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2224 ins.src[0].reg.immconst_data[0],
2225 ins.src[0].reg.immconst_data[1],
2226 ins.src[0].reg.immconst_data[2],
2227 ins.src[0].reg.immconst_data[3]);
2228 }
2229 else if (ins.handler_idx == WINED3DSIH_DEFB)
2230 {
2231 shader_addline(&buffer, "defb b%u = %s",
2232 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2233 }
2234 else
2235 {
2236 if (ins.predicate)
2237 {
2238 shader_addline(&buffer, "(");
2239 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2240 shader_addline(&buffer, ") ");
2241 }
2242
2243 /* PixWin marks instructions with the coissue flag with a '+' */
2244 if (ins.coissue)
2245 shader_addline(&buffer, "+");
2246
2247 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2248
2249 if (ins.handler_idx == WINED3DSIH_IFC
2250 || ins.handler_idx == WINED3DSIH_BREAKC)
2251 {
2252 switch (ins.flags)
2253 {
2254 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2255 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2256 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2257 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2258 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2259 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2260 default: shader_addline(&buffer, "_(%u)", ins.flags);
2261 }
2262 }
2263 else if (ins.handler_idx == WINED3DSIH_TEX
2264 && shader_version.major >= 2
2265 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2266 {
2267 shader_addline(&buffer, "p");
2268 }
2269 else if (ins.handler_idx == WINED3DSIH_RESINFO
2270 && ins.flags)
2271 {
2272 switch (ins.flags)
2273 {
2274 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2275 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2276 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2277 }
2278 }
2279
2280 if (wined3d_shader_instruction_has_texel_offset(&ins))
2281 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2282
2283 for (i = 0; i < ins.dst_count; ++i)
2284 {
2285 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2286 shader_addline(&buffer, !i ? " " : ", ");
2287 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2288 }
2289
2290 /* Other source tokens */
2291 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2292 {
2293 shader_addline(&buffer, !i ? " " : ", ");
2294 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2295 }
2296 }
2297 shader_addline(&buffer, "\n");
2298 }
2299
2300 for (p = buffer.buffer; *p; p = q)
2301 {
2302 if (!(q = strstr(p, "\n")))
2303 q = p + strlen(p);
2304 else
2305 ++q;
2306 TRACE(" %.*s", (int)(q - p), p);
2307 }
2308
2309 string_buffer_free(&buffer);
2310 }
2311
2312 static void shader_cleanup(struct wined3d_shader *shader)
2313 {
2314 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2315 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2316 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2317 shader->device->shader_backend->shader_destroy(shader);
2318 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
2319 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
2320 HeapFree(GetProcessHeap(), 0, shader->function);
2321 shader_delete_constant_list(&shader->constantsF);
2322 shader_delete_constant_list(&shader->constantsB);
2323 shader_delete_constant_list(&shader->constantsI);
2324 list_remove(&shader->shader_list_entry);
2325
2326 if (shader->frontend && shader->frontend_data)
2327 shader->frontend->shader_free(shader->frontend_data);
2328 }
2329
2330 struct shader_none_priv
2331 {
2332 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2333 const struct fragment_pipeline *fragment_pipe;
2334 BOOL ffp_proj_control;
2335 };
2336
2337 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2338 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2339 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2340 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2341 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2342 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2343 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2344 const struct wined3d_state *state) {}
2345 static void shader_none_destroy(struct wined3d_shader *shader) {}
2346 static void shader_none_free_context_data(struct wined3d_context *context) {}
2347 static void shader_none_init_context_state(struct wined3d_context *context) {}
2348
2349 /* Context activation is done by the caller. */
2350 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2351 const struct wined3d_state *state)
2352 {
2353 const struct wined3d_gl_info *gl_info = context->gl_info;
2354 struct shader_none_priv *priv = shader_priv;
2355
2356 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2357 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2358 }
2359
2360 /* Context activation is done by the caller. */
2361 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2362 {
2363 struct shader_none_priv *priv = shader_priv;
2364 const struct wined3d_gl_info *gl_info = context->gl_info;
2365
2366 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2367 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2368
2369 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2370 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2371 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2372 | (1u << WINED3D_SHADER_TYPE_HULL)
2373 | (1u << WINED3D_SHADER_TYPE_DOMAIN);
2374 }
2375
2376 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2377 const struct fragment_pipeline *fragment_pipe)
2378 {
2379 struct fragment_caps fragment_caps;
2380 void *vertex_priv, *fragment_priv;
2381 struct shader_none_priv *priv;
2382
2383 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2384 return E_OUTOFMEMORY;
2385
2386 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2387 {
2388 ERR("Failed to initialize vertex pipe.\n");
2389 HeapFree(GetProcessHeap(), 0, priv);
2390 return E_FAIL;
2391 }
2392
2393 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2394 {
2395 ERR("Failed to initialize fragment pipe.\n");
2396 vertex_pipe->vp_free(device);
2397 HeapFree(GetProcessHeap(), 0, priv);
2398 return E_FAIL;
2399 }
2400
2401 priv->vertex_pipe = vertex_pipe;
2402 priv->fragment_pipe = fragment_pipe;
2403 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2404 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2405
2406 device->vertex_priv = vertex_priv;
2407 device->fragment_priv = fragment_priv;
2408 device->shader_priv = priv;
2409
2410 return WINED3D_OK;
2411 }
2412
2413 static void shader_none_free(struct wined3d_device *device)
2414 {
2415 struct shader_none_priv *priv = device->shader_priv;
2416
2417 priv->fragment_pipe->free_private(device);
2418 priv->vertex_pipe->vp_free(device);
2419 HeapFree(GetProcessHeap(), 0, priv);
2420 }
2421
2422 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2423 {
2424 return TRUE;
2425 }
2426
2427 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2428 {
2429 /* Set the shader caps to 0 for the none shader backend */
2430 caps->vs_version = 0;
2431 caps->hs_version = 0;
2432 caps->ds_version = 0;
2433 caps->gs_version = 0;
2434 caps->ps_version = 0;
2435 caps->vs_uniform_count = 0;
2436 caps->ps_uniform_count = 0;
2437 caps->ps_1x_max_value = 0.0f;
2438 caps->varying_count = 0;
2439 caps->wined3d_caps = 0;
2440 }
2441
2442 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2443 {
2444 /* We "support" every possible fixup, since we don't support any shader
2445 * model, and will never have to actually sample a texture. */
2446 return TRUE;
2447 }
2448
2449 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2450 {
2451 struct shader_none_priv *priv = shader_priv;
2452
2453 return priv->ffp_proj_control;
2454 }
2455
2456 const struct wined3d_shader_backend_ops none_shader_backend =
2457 {
2458 shader_none_handle_instruction,
2459 shader_none_select,
2460 shader_none_disable,
2461 shader_none_select_depth_blt,
2462 shader_none_deselect_depth_blt,
2463 shader_none_update_float_vertex_constants,
2464 shader_none_update_float_pixel_constants,
2465 shader_none_load_constants,
2466 shader_none_destroy,
2467 shader_none_alloc,
2468 shader_none_free,
2469 shader_none_allocate_context_data,
2470 shader_none_free_context_data,
2471 shader_none_init_context_state,
2472 shader_none_get_caps,
2473 shader_none_color_fixup_supported,
2474 shader_none_has_ffp_proj_control,
2475 };
2476
2477 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2478 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2479 enum wined3d_shader_type type, unsigned int max_version)
2480 {
2481 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2482 const struct wined3d_shader_frontend *fe;
2483 HRESULT hr;
2484 unsigned int backend_version;
2485 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2486
2487 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u, type %#x, max_version %u.\n",
2488 shader, byte_code, output_signature, float_const_count, type, max_version);
2489
2490 list_init(&shader->constantsF);
2491 list_init(&shader->constantsB);
2492 list_init(&shader->constantsI);
2493 shader->lconst_inf_or_nan = FALSE;
2494
2495 fe = shader_select_frontend(*byte_code);
2496 if (!fe)
2497 {
2498 FIXME("Unable to find frontend for shader.\n");
2499 return WINED3DERR_INVALIDCALL;
2500 }
2501 shader->frontend = fe;
2502 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2503 if (!shader->frontend_data)
2504 {
2505 FIXME("Failed to initialize frontend.\n");
2506 return WINED3DERR_INVALIDCALL;
2507 }
2508
2509 /* First pass: trace shader. */
2510 if (TRACE_ON(d3d_shader))
2511 shader_trace_init(fe, shader->frontend_data, byte_code);
2512
2513 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2514 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2515 &shader->output_signature, byte_code, float_const_count)))
2516 return hr;
2517
2518 if (reg_maps->shader_version.type != type)
2519 {
2520 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2521 return WINED3DERR_INVALIDCALL;
2522 }
2523 if (reg_maps->shader_version.major > max_version)
2524 {
2525 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2526 return WINED3DERR_INVALIDCALL;
2527 }
2528 switch (type)
2529 {
2530 case WINED3D_SHADER_TYPE_VERTEX:
2531 backend_version = d3d_info->limits.vs_version;
2532 break;
2533 case WINED3D_SHADER_TYPE_HULL:
2534 backend_version = d3d_info->limits.hs_version;
2535 break;
2536 case WINED3D_SHADER_TYPE_DOMAIN:
2537 backend_version = d3d_info->limits.ds_version;
2538 break;
2539 case WINED3D_SHADER_TYPE_GEOMETRY:
2540 backend_version = d3d_info->limits.gs_version;
2541 break;
2542 case WINED3D_SHADER_TYPE_PIXEL:
2543 backend_version = d3d_info->limits.ps_version;
2544 break;
2545 default:
2546 FIXME("No backend version-checking for this shader type\n");
2547 backend_version = 0;
2548 }
2549 if (reg_maps->shader_version.major > backend_version)
2550 {
2551 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2552 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2553 return WINED3DERR_INVALIDCALL;
2554 }
2555
2556 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2557 if (!shader->function)
2558 return E_OUTOFMEMORY;
2559 memcpy(shader->function, byte_code, shader->functionLength);
2560
2561 return WINED3D_OK;
2562 }
2563
2564 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2565 {
2566 ULONG refcount = InterlockedIncrement(&shader->ref);
2567
2568 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2569
2570 return refcount;
2571 }
2572
2573 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2574 {
2575 ULONG refcount = InterlockedDecrement(&shader->ref);
2576
2577 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2578
2579 if (!refcount)
2580 {
2581 shader_cleanup(shader);
2582 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2583 HeapFree(GetProcessHeap(), 0, shader);
2584 }
2585
2586 return refcount;
2587 }
2588
2589 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2590 {
2591 TRACE("shader %p.\n", shader);
2592
2593 return shader->parent;
2594 }
2595
2596 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2597 void *byte_code, UINT *byte_code_size)
2598 {
2599 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2600
2601 if (!byte_code)
2602 {
2603 *byte_code_size = shader->functionLength;
2604 return WINED3D_OK;
2605 }
2606
2607 if (*byte_code_size < shader->functionLength)
2608 {
2609 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2610 * than the required size we should write the required size and
2611 * return D3DERR_MOREDATA. That's not actually true. */
2612 return WINED3DERR_INVALIDCALL;
2613 }
2614
2615 memcpy(byte_code, shader->function, shader->functionLength);
2616
2617 return WINED3D_OK;
2618 }
2619
2620 /* Set local constants for d3d8 shaders. */
2621 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2622 UINT start_idx, const float *src_data, UINT count)
2623 {
2624 UINT end_idx = start_idx + count;
2625 UINT i;
2626
2627 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2628
2629 if (end_idx > shader->limits->constant_float)
2630 {
2631 WARN("end_idx %u > float constants limit %u.\n",
2632 end_idx, shader->limits->constant_float);
2633 end_idx = shader->limits->constant_float;
2634 }
2635
2636 for (i = start_idx; i < end_idx; ++i)
2637 {
2638 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2639 float *value;
2640 if (!lconst)
2641 return E_OUTOFMEMORY;
2642
2643 lconst->idx = i;
2644 value = (float *)lconst->value;
2645 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2646 list_add_head(&shader->constantsF, &lconst->entry);
2647
2648 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2649 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2650 {
2651 shader->lconst_inf_or_nan = TRUE;
2652 }
2653 }
2654
2655 return WINED3D_OK;
2656 }
2657
2658 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2659 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2660 {
2661 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2662 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2663 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2664 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2665 args->point_size = state->gl_primitive_type == GL_POINTS;
2666 args->per_vertex_point_size = shader->reg_maps.point_size;
2667 args->swizzle_map = swizzle_map;
2668 if (d3d_info->emulated_flatshading)
2669 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2670 else
2671 args->flatshading = FALSE;
2672 }
2673
2674 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2675 {
2676 if (usage_idx1 != usage_idx2)
2677 return FALSE;
2678 if (usage1 == usage2)
2679 return TRUE;
2680 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2681 return TRUE;
2682 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2683 return TRUE;
2684
2685 return FALSE;
2686 }
2687
2688 BOOL vshader_get_input(const struct wined3d_shader *shader,
2689 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2690 {
2691 WORD map = shader->reg_maps.input_registers;
2692 unsigned int i;
2693
2694 for (i = 0; map; map >>= 1, ++i)
2695 {
2696 if (!(map & 1)) continue;
2697
2698 if (match_usage(shader->u.vs.attributes[i].usage,
2699 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2700 {
2701 *regnum = i;
2702 return TRUE;
2703 }
2704 }
2705 return FALSE;
2706 }
2707
2708 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2709 const struct wined3d_shader_signature *src, char **signature_strings)
2710 {
2711 struct wined3d_shader_signature_element *e;
2712 unsigned int i;
2713 SIZE_T len;
2714 char *ptr;
2715
2716 ptr = *signature_strings;
2717
2718 dst->element_count = src->element_count;
2719 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2720 return E_OUTOFMEMORY;
2721
2722 for (i = 0; i < src->element_count; ++i)
2723 {
2724 e = &src->elements[i];
2725 dst->elements[i] = *e;
2726
2727 len = strlen(e->semantic_name);
2728 memcpy(ptr, e->semantic_name, len + 1);
2729 dst->elements[i].semantic_name = ptr;
2730 ptr += len + 1;
2731 }
2732
2733 *signature_strings = ptr;
2734
2735 return WINED3D_OK;
2736 }
2737
2738 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2739 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2740 void *parent, const struct wined3d_parent_ops *parent_ops)
2741 {
2742 struct wined3d_shader_signature_element *e;
2743 SIZE_T total, len;
2744 unsigned int i;
2745 HRESULT hr;
2746 char *ptr;
2747
2748 if (!desc->byte_code)
2749 return WINED3DERR_INVALIDCALL;
2750
2751 shader->ref = 1;
2752 shader->device = device;
2753 shader->parent = parent;
2754 shader->parent_ops = parent_ops;
2755
2756 total = 0;
2757 if (desc->input_signature)
2758 {
2759 for (i = 0; i < desc->input_signature->element_count; ++i)
2760 {
2761 e = &desc->input_signature->elements[i];
2762 len = strlen(e->semantic_name);
2763 if (len >= ~(SIZE_T)0 - total)
2764 return E_OUTOFMEMORY;
2765
2766 total += len + 1;
2767 }
2768 }
2769 if (desc->output_signature)
2770 {
2771 for (i = 0; i < desc->output_signature->element_count; ++i)
2772 {
2773 e = &desc->output_signature->elements[i];
2774 len = strlen(e->semantic_name);
2775 if (len >= ~(SIZE_T)0 - total)
2776 return E_OUTOFMEMORY;
2777
2778 total += len + 1;
2779 }
2780 }
2781 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2782 return E_OUTOFMEMORY;
2783 ptr = shader->signature_strings;
2784
2785 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2786 desc->input_signature, &ptr)))
2787 {
2788 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2789 return hr;
2790 }
2791 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2792 desc->output_signature, &ptr)))
2793 {
2794 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2795 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2796 return hr;
2797 }
2798
2799 list_init(&shader->linked_programs);
2800 list_add_head(&device->shaders, &shader->shader_list_entry);
2801
2802 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2803 float_const_count, type, desc->max_version)))
2804 {
2805 WARN("Failed to set function, hr %#x.\n", hr);
2806 shader_cleanup(shader);
2807 }
2808
2809 return hr;
2810 }
2811
2812 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2813 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2814 {
2815 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2816 unsigned int i;
2817 HRESULT hr;
2818
2819 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2820 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2821 return hr;
2822
2823 for (i = 0; i < shader->input_signature.element_count; ++i)
2824 {
2825 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2826
2827 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2828 continue;
2829
2830 shader->u.vs.attributes[input->register_idx].usage =
2831 shader_usage_from_semantic_name(input->semantic_name);
2832 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2833 }
2834
2835 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2836 shader->lconst_inf_or_nan;
2837
2838 return WINED3D_OK;
2839 }
2840
2841 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2842 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2843 {
2844 HRESULT hr;
2845
2846 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops)))
2847 return hr;
2848
2849 shader->load_local_constsF = shader->lconst_inf_or_nan;
2850
2851 return WINED3D_OK;
2852 }
2853
2854 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2855 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2856 {
2857 HRESULT hr;
2858
2859 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops)))
2860 return hr;
2861
2862 shader->load_local_constsF = shader->lconst_inf_or_nan;
2863
2864 return WINED3D_OK;
2865 }
2866
2867 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2868 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2869 {
2870 HRESULT hr;
2871
2872 if (FAILED(hr = shader_init(shader, device, desc, 0,
2873 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2874 return hr;
2875
2876 shader->load_local_constsF = shader->lconst_inf_or_nan;
2877
2878 return WINED3D_OK;
2879 }
2880
2881 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2882 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2883 {
2884 const struct wined3d_gl_info *gl_info = context->gl_info;
2885 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2886 const struct wined3d_texture *texture;
2887 UINT i;
2888
2889 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2890 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
2891 {
2892 static unsigned int warned = 0;
2893
2894 args->srgb_correction = 1;
2895 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2896 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2897 "support, expect rendering artifacts.\n");
2898 }
2899
2900 if (shader->reg_maps.shader_version.major == 1
2901 && shader->reg_maps.shader_version.minor <= 3)
2902 {
2903 for (i = 0; i < shader->limits->sampler; ++i)
2904 {
2905 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2906
2907 if (flags & WINED3D_TTFF_PROJECTED)
2908 {
2909 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2910
2911 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2912 {
2913 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2914 unsigned int j;
2915 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2916 DWORD max_valid = WINED3D_TTFF_COUNT4;
2917
2918 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2919 {
2920 struct wined3d_vertex_declaration_element *element =
2921 &state->vertex_declaration->elements[j];
2922
2923 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2924 && element->usage_idx == index)
2925 {
2926 max_valid = element->format->component_count;
2927 break;
2928 }
2929 }
2930 if (!tex_transform || tex_transform > max_valid)
2931 {
2932 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2933 tex_transform, max_valid);
2934 tex_transform = max_valid;
2935 }
2936 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2937 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2938 && tex_transform > WINED3D_TTFF_COUNT2)
2939 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2940 && tex_transform > WINED3D_TTFF_COUNT3))
2941 tex_transform |= WINED3D_PSARGS_PROJECTED;
2942 else
2943 {
2944 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2945 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2946 i, tex_transform, resource_type);
2947 }
2948 }
2949 else
2950 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2951
2952 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2953 }
2954 }
2955 }
2956 if (shader->reg_maps.shader_version.major == 1
2957 && shader->reg_maps.shader_version.minor <= 4)
2958 {
2959 for (i = 0; i < shader->limits->sampler; ++i)
2960 {
2961 const struct wined3d_texture *texture = state->textures[i];
2962
2963 if (!shader->reg_maps.resource_info[i].type)
2964 continue;
2965
2966 /* Treat unbound textures as 2D. The dummy texture will provide
2967 * the proper sample value. The tex_types bitmap defaults to
2968 * 2D because of the memset. */
2969 if (!texture)
2970 continue;
2971
2972 switch (texture->target)
2973 {
2974 /* RECT textures are distinguished from 2D textures via np2_fixup */
2975 case GL_TEXTURE_RECTANGLE_ARB:
2976 case GL_TEXTURE_2D:
2977 break;
2978
2979 case GL_TEXTURE_3D:
2980 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2981 break;
2982
2983 case GL_TEXTURE_CUBE_MAP_ARB:
2984 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2985 break;
2986 }
2987 }
2988 }
2989
2990 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2991 {
2992 if (!shader->reg_maps.resource_info[i].type)
2993 continue;
2994
2995 texture = state->textures[i];
2996 if (!texture)
2997 {
2998 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2999 continue;
3000 }
3001 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3002 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3003 else
3004 args->color_fixup[i] = texture->resource.format->color_fixup;
3005
3006 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3007 args->shadow |= 1u << i;
3008
3009 /* Flag samplers that need NP2 texcoord fixup. */
3010 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3011 args->np2_fixup |= (1u << i);
3012 }
3013
3014 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3015 if (shader->reg_maps.shader_version.major >= 4)
3016 args->shadow = 0;
3017
3018 if (shader->reg_maps.shader_version.major >= 3)
3019 {
3020 if (position_transformed)
3021 args->vp_mode = pretransformed;
3022 else if (use_vs(state))
3023 args->vp_mode = vertexshader;
3024 else
3025 args->vp_mode = fixedfunction;
3026 args->fog = WINED3D_FFP_PS_FOG_OFF;
3027 }
3028 else
3029 {
3030 args->vp_mode = vertexshader;
3031 if (state->render_states[WINED3D_RS_FOGENABLE])
3032 {
3033 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3034 {
3035 case WINED3D_FOG_NONE:
3036 if (position_transformed || use_vs(state))
3037 {
3038 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3039 break;
3040 }
3041
3042 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3043 {
3044 case WINED3D_FOG_NONE: /* Fall through. */
3045 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3046 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3047 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3048 }
3049 break;
3050
3051 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3052 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3053 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3054 }
3055 }
3056 else
3057 {
3058 args->fog = WINED3D_FFP_PS_FOG_OFF;
3059 }
3060 }
3061
3062 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3063 {
3064 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3065
3066 args->texcoords_initialized = 0;
3067 for (i = 0; i < MAX_TEXTURES; ++i)
3068 {
3069 if (vs)
3070 {
3071 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3072 args->texcoords_initialized |= 1u << i;
3073 }
3074 else
3075 {
3076 const struct wined3d_stream_info *si = &context->stream_info;
3077 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3078
3079 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3080 & WINED3D_FFP_TCI_MASK
3081 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3082 args->texcoords_initialized |= 1u << i;
3083 }
3084 }
3085 }
3086 else
3087 {
3088 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3089 }
3090
3091 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3092 && state->gl_primitive_type == GL_POINTS;
3093
3094 if (d3d_info->emulated_flatshading)
3095 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3096 }
3097
3098 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3099 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3100 {
3101 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3102 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3103 HRESULT hr;
3104
3105 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3106 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3107 return hr;
3108
3109 for (i = 0; i < MAX_REG_INPUT; ++i)
3110 {
3111 if (shader->u.ps.input_reg_used[i])
3112 {
3113 ++num_regs_used;
3114 highest_reg_used = i;
3115 }
3116 }
3117
3118 /* Don't do any register mapping magic if it is not needed, or if we can't
3119 * achieve anything anyway */
3120 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3121 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
3122 {
3123 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3124 {
3125 /* This happens with relative addressing. The input mapper function
3126 * warns about this if the higher registers are declared too, so
3127 * don't write a FIXME here */
3128 WARN("More varying registers used than supported\n");
3129 }
3130
3131 for (i = 0; i < MAX_REG_INPUT; ++i)
3132 {
3133 shader->u.ps.input_reg_map[i] = i;
3134 }
3135
3136 shader->u.ps.declared_in_count = highest_reg_used + 1;
3137 }
3138 else
3139 {
3140 shader->u.ps.declared_in_count = 0;
3141 for (i = 0; i < MAX_REG_INPUT; ++i)
3142 {
3143 if (shader->u.ps.input_reg_used[i])
3144 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3145 else shader->u.ps.input_reg_map[i] = ~0U;
3146 }
3147 }
3148
3149 shader->load_local_constsF = shader->lconst_inf_or_nan;
3150
3151 return WINED3D_OK;
3152 }
3153
3154 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3155 {
3156 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3157 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3158 unsigned int i;
3159
3160 if (reg_maps->shader_version.major != 1) return;
3161
3162 for (i = 0; i < shader->limits->sampler; ++i)
3163 {
3164 /* We don't sample from this sampler. */
3165 if (!resource_info[i].type)
3166 continue;
3167
3168 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3169 {
3170 case WINED3D_SHADER_TEX_2D:
3171 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3172 break;
3173
3174 case WINED3D_SHADER_TEX_3D:
3175 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3176 break;
3177
3178 case WINED3D_SHADER_TEX_CUBE:
3179 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3180 break;
3181 }
3182 }
3183 }
3184
3185 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3186 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3187 {
3188 struct wined3d_shader *object;
3189 HRESULT hr;
3190
3191 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3192 device, desc, parent, parent_ops, shader);
3193
3194 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3195 return E_OUTOFMEMORY;
3196
3197 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3198 {
3199 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3200 HeapFree(GetProcessHeap(), 0, object);
3201 return hr;
3202 }
3203
3204 TRACE("Created domain shader %p.\n", object);
3205 *shader = object;
3206
3207 return WINED3D_OK;
3208 }
3209
3210 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3211 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3212 {
3213 struct wined3d_shader *object;
3214 HRESULT hr;
3215
3216 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3217 device, desc, parent, parent_ops, shader);
3218
3219 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3220 if (!object)
3221 return E_OUTOFMEMORY;
3222
3223 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3224 {
3225 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3226 HeapFree(GetProcessHeap(), 0, object);
3227 return hr;
3228 }
3229
3230 TRACE("Created geometry shader %p.\n", object);
3231 *shader = object;
3232
3233 return WINED3D_OK;
3234 }
3235
3236 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3237 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3238 {
3239 struct wined3d_shader *object;
3240 HRESULT hr;
3241
3242 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3243 device, desc, parent, parent_ops, shader);
3244
3245 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3246 return E_OUTOFMEMORY;
3247
3248 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3249 {
3250 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3251 HeapFree(GetProcessHeap(), 0, object);
3252 return hr;
3253 }
3254
3255 TRACE("Created hull shader %p.\n", object);
3256 *shader = object;
3257
3258 return WINED3D_OK;
3259 }
3260
3261 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3262 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3263 {
3264 struct wined3d_shader *object;
3265 HRESULT hr;
3266
3267 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3268 device, desc, parent, parent_ops, shader);
3269
3270 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3271 if (!object)
3272 return E_OUTOFMEMORY;
3273
3274 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3275 {
3276 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3277 HeapFree(GetProcessHeap(), 0, object);
3278 return hr;
3279 }
3280
3281 TRACE("Created pixel shader %p.\n", object);
3282 *shader = object;
3283
3284 return WINED3D_OK;
3285 }
3286
3287 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3288 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3289 {
3290 struct wined3d_shader *object;
3291 HRESULT hr;
3292
3293 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3294 device, desc, parent, parent_ops, shader);
3295
3296 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3297 if (!object)
3298 return E_OUTOFMEMORY;
3299
3300 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3301 {
3302 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3303 HeapFree(GetProcessHeap(), 0, object);
3304 return hr;
3305 }
3306
3307 TRACE("Created vertex shader %p.\n", object);
3308 *shader = object;
3309
3310 return WINED3D_OK;
3311 }
Windows API implementation