?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Recognize SM5 vicp register.
[wine.git] / dlls / wined3d / shader.c
blob920abd7e25fb6184f03606f54d4d334aa5a9b5d8
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_DECLS */ "hs_decls",
110 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
111 /* WINED3DSIH_IADD */ "iadd",
112 /* WINED3DSIH_IEQ */ "ieq",
113 /* WINED3DSIH_IF */ "if",
114 /* WINED3DSIH_IFC */ "ifc",
115 /* WINED3DSIH_IGE */ "ige",
116 /* WINED3DSIH_ILT */ "ilt",
117 /* WINED3DSIH_IMAD */ "imad",
118 /* WINED3DSIH_IMAX */ "imax",
119 /* WINED3DSIH_IMIN */ "imin",
120 /* WINED3DSIH_IMUL */ "imul",
121 /* WINED3DSIH_INE */ "ine",
122 /* WINED3DSIH_INEG */ "ineg",
123 /* WINED3DSIH_ISHL */ "ishl",
124 /* WINED3DSIH_ITOF */ "itof",
125 /* WINED3DSIH_LABEL */ "label",
126 /* WINED3DSIH_LD */ "ld",
127 /* WINED3DSIH_LD2DMS */ "ld2dms",
128 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
129 /* WINED3DSIH_LIT */ "lit",
130 /* WINED3DSIH_LOG */ "log",
131 /* WINED3DSIH_LOGP */ "logp",
132 /* WINED3DSIH_LOOP */ "loop",
133 /* WINED3DSIH_LRP */ "lrp",
134 /* WINED3DSIH_LT */ "lt",
135 /* WINED3DSIH_M3x2 */ "m3x2",
136 /* WINED3DSIH_M3x3 */ "m3x3",
137 /* WINED3DSIH_M3x4 */ "m3x4",
138 /* WINED3DSIH_M4x3 */ "m4x3",
139 /* WINED3DSIH_M4x4 */ "m4x4",
140 /* WINED3DSIH_MAD */ "mad",
141 /* WINED3DSIH_MAX */ "max",
142 /* WINED3DSIH_MIN */ "min",
143 /* WINED3DSIH_MOV */ "mov",
144 /* WINED3DSIH_MOVA */ "mova",
145 /* WINED3DSIH_MOVC */ "movc",
146 /* WINED3DSIH_MUL */ "mul",
147 /* WINED3DSIH_NE */ "ne",
148 /* WINED3DSIH_NOP */ "nop",
149 /* WINED3DSIH_NOT */ "not",
150 /* WINED3DSIH_NRM */ "nrm",
151 /* WINED3DSIH_OR */ "or",
152 /* WINED3DSIH_PHASE */ "phase",
153 /* WINED3DSIH_POW */ "pow",
154 /* WINED3DSIH_RCP */ "rcp",
155 /* WINED3DSIH_REP */ "rep",
156 /* WINED3DSIH_RESINFO */ "resinfo",
157 /* WINED3DSIH_RET */ "ret",
158 /* WINED3DSIH_ROUND_NI */ "round_ni",
159 /* WINED3DSIH_ROUND_PI */ "round_pi",
160 /* WINED3DSIH_ROUND_Z */ "round_z",
161 /* WINED3DSIH_RSQ */ "rsq",
162 /* WINED3DSIH_SAMPLE */ "sample",
163 /* WINED3DSIH_SAMPLE_B */ "sample_b",
164 /* WINED3DSIH_SAMPLE_C */ "sample_c",
165 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
166 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
167 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
168 /* WINED3DSIH_SETP */ "setp",
169 /* WINED3DSIH_SGE */ "sge",
170 /* WINED3DSIH_SGN */ "sgn",
171 /* WINED3DSIH_SINCOS */ "sincos",
172 /* WINED3DSIH_SLT */ "slt",
173 /* WINED3DSIH_SQRT */ "sqrt",
174 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
175 /* WINED3DSIH_SUB */ "sub",
176 /* WINED3DSIH_TEX */ "texld",
177 /* WINED3DSIH_TEXBEM */ "texbem",
178 /* WINED3DSIH_TEXBEML */ "texbeml",
179 /* WINED3DSIH_TEXCOORD */ "texcrd",
180 /* WINED3DSIH_TEXDEPTH */ "texdepth",
181 /* WINED3DSIH_TEXDP3 */ "texdp3",
182 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
183 /* WINED3DSIH_TEXKILL */ "texkill",
184 /* WINED3DSIH_TEXLDD */ "texldd",
185 /* WINED3DSIH_TEXLDL */ "texldl",
186 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
187 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
188 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
189 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
190 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
191 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
192 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
193 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
194 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
195 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
196 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
197 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
198 /* WINED3DSIH_UDIV */ "udiv",
199 /* WINED3DSIH_UGE */ "uge",
200 /* WINED3DSIH_USHR */ "ushr",
201 /* WINED3DSIH_UTOF */ "utof",
202 /* WINED3DSIH_XOR */ "xor",
203 };
204
205 static const char * const semantic_names[] =
206 {
207 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
208 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
209 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
210 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
211 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
212 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
213 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
214 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
215 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
216 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
217 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
218 /* WINED3D_DECL_USAGE_FOG */ "FOG",
219 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
220 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
221 };
222
223 static const struct
224 {
225 enum wined3d_sysval_semantic sysval_semantic;
226 const char *sysval_name;
227 }
228 sysval_semantic_names[] =
229 {
230 {WINED3D_SV_POSITION, "SV_Position"},
231 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
232 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
233 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
234 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
235 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
236 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
237 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
238 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
239 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
240 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
241 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
242 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
243 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
244 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
245 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
246 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
247 };
248
249 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
250 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
251
252 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
253 {
254 if (handler_idx >= sizeof(shader_opcode_names) / sizeof(*shader_opcode_names))
255 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
256
257 return shader_opcode_names[handler_idx];
258 }
259
260 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
261 {
262 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
263 {
264 FIXME("Unrecognized usage %#x.\n", usage);
265 return "UNRECOGNIZED";
266 }
267
268 return semantic_names[usage];
269 }
270
271 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
272 {
273 unsigned int i;
274
275 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
276 {
277 if (!strcmp(name, semantic_names[i])) return i;
278 }
279
280 return ~0U;
281 }
282
283 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
284 {
285 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
286 }
287
288 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
289 const struct wined3d_shader_semantic *s)
290 {
291 e->semantic_name = shader_semantic_name_from_usage(s->usage);
292 e->semantic_idx = s->usage_idx;
293 e->sysval_semantic = 0;
294 e->component_type = 0;
295 e->register_idx = s->reg.reg.idx[0].offset;
296 e->mask = s->reg.write_mask;
297 }
298
299 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
300 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
301 {
302 e->semantic_name = shader_semantic_name_from_usage(usage);
303 e->semantic_idx = usage_idx;
304 e->sysval_semantic = 0;
305 e->component_type = 0;
306 e->register_idx = reg_idx;
307 e->mask = write_mask;
308 }
309
310 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
311 {
312 switch (version_token >> 16)
313 {
314 case WINED3D_SM1_VS:
315 case WINED3D_SM1_PS:
316 return &sm1_shader_frontend;
317
318 case WINED3D_SM4_PS:
319 case WINED3D_SM4_VS:
320 case WINED3D_SM4_GS:
321 case WINED3D_SM5_HS:
322 case WINED3D_SM5_DS:
323 return &sm4_shader_frontend;
324
325 default:
326 FIXME("Unrecognised version token %#x.\n", version_token);
327 return NULL;
328 }
329 }
330
331 void string_buffer_clear(struct wined3d_string_buffer *buffer)
332 {
333 buffer->buffer[0] = '\0';
334 buffer->content_size = 0;
335 }
336
337 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
338 {
339 buffer->buffer_size = 32;
340 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
341 {
342 ERR("Failed to allocate shader buffer memory.\n");
343 return FALSE;
344 }
345
346 string_buffer_clear(buffer);
347 return TRUE;
348 }
349
350 void string_buffer_free(struct wined3d_string_buffer *buffer)
351 {
352 HeapFree(GetProcessHeap(), 0, buffer->buffer);
353 }
354
355 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
356 {
357 char *new_buffer;
358 unsigned int new_buffer_size = buffer->buffer_size * 2;
359
360 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
361 new_buffer_size *= 2;
362 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
363 {
364 ERR("Failed to grow buffer.\n");
365 buffer->buffer[buffer->content_size] = '\0';
366 return FALSE;
367 }
368 buffer->buffer = new_buffer;
369 buffer->buffer_size = new_buffer_size;
370 return TRUE;
371 }
372
373 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
374 {
375 unsigned int rem;
376 int rc;
377
378 rem = buffer->buffer_size - buffer->content_size;
379 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
380 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
381 return rc;
382
383 buffer->content_size += rc;
384 return 0;
385 }
386
387 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
388 {
389 va_list args;
390 int ret;
391
392 for (;;)
393 {
394 va_start(args, format);
395 ret = shader_vaddline(buffer, format, args);
396 va_end(args);
397 if (!ret)
398 return ret;
399 if (!string_buffer_resize(buffer, ret))
400 return -1;
401 }
402 }
403
404 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
405 {
406 struct wined3d_string_buffer *buffer;
407
408 if (list_empty(&list->list))
409 {
410 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
411 if (!buffer || !string_buffer_init(buffer))
412 {
413 ERR("Couldn't allocate buffer for temporary string.\n");
414 if (buffer)
415 HeapFree(GetProcessHeap(), 0, buffer);
416 return NULL;
417 }
418 }
419 else
420 {
421 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
422 list_remove(&buffer->entry);
423 }
424 string_buffer_clear(buffer);
425 return buffer;
426 }
427
428 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
429 {
430 if (!buffer)
431 return 0;
432 string_buffer_clear(buffer);
433 return shader_vaddline(buffer, format, args);
434 }
435
436 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
437 {
438 va_list args;
439 int ret;
440
441 for (;;)
442 {
443 va_start(args, format);
444 ret = string_buffer_vsprintf(buffer, format, args);
445 va_end(args);
446 if (!ret)
447 return;
448 if (!string_buffer_resize(buffer, ret))
449 return;
450 }
451 }
452
453 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
454 {
455 if (!buffer)
456 return;
457 list_add_head(&list->list, &buffer->entry);
458 }
459
460 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
461 {
462 list_init(&list->list);
463 }
464
465 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
466 {
467 struct wined3d_string_buffer *buffer, *buffer_next;
468
469 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
470 {
471 string_buffer_free(buffer);
472 HeapFree(GetProcessHeap(), 0, buffer);
473 }
474 list_init(&list->list);
475 }
476
477 /* Convert floating point offset relative to a register file to an absolute
478 * offset for float constants. */
479 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
480 {
481 switch (register_type)
482 {
483 case WINED3DSPR_CONST: return register_idx;
484 case WINED3DSPR_CONST2: return 2048 + register_idx;
485 case WINED3DSPR_CONST3: return 4096 + register_idx;
486 case WINED3DSPR_CONST4: return 6144 + register_idx;
487 default:
488 FIXME("Unsupported register type: %u.\n", register_type);
489 return register_idx;
490 }
491 }
492
493 static void shader_delete_constant_list(struct list *clist)
494 {
495 struct wined3d_shader_lconst *constant, *constant_next;
496
497 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
498 HeapFree(GetProcessHeap(), 0, constant);
499 list_init(clist);
500 }
501
502 static void shader_set_limits(struct wined3d_shader *shader)
503 {
504 static const struct limits_entry
505 {
506 unsigned int min_version;
507 unsigned int max_version;
508 struct wined3d_shader_limits limits;
509 }
510 vs_limits[] =
511 {
512 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
513 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
514 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
515 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
516 * even though they are capable of supporting much more (GL
517 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
518 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
519 * shaders to 256. */
520 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
521 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
522 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
523 {0}
524 },
525 hs_limits[] =
526 {
527 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
528 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
529 },
530 ds_limits[] =
531 {
532 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
533 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
534 },
535 gs_limits[] =
536 {
537 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
538 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
539 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
540 {0}
541 },
542 ps_limits[] =
543 {
544 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
545 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
546 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
547 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
548 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
549 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
550 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
551 {0}
552 };
553 const struct limits_entry *limits_array;
554 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
555 shader->reg_maps.shader_version.minor);
556 int i = 0;
557
558 switch (shader->reg_maps.shader_version.type)
559 {
560 default:
561 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
562 /* Fall-through. */
563 case WINED3D_SHADER_TYPE_VERTEX:
564 limits_array = vs_limits;
565 break;
566 case WINED3D_SHADER_TYPE_HULL:
567 limits_array = hs_limits;
568 break;
569 case WINED3D_SHADER_TYPE_DOMAIN:
570 limits_array = ds_limits;
571 break;
572 case WINED3D_SHADER_TYPE_GEOMETRY:
573 limits_array = gs_limits;
574 break;
575 case WINED3D_SHADER_TYPE_PIXEL:
576 limits_array = ps_limits;
577 break;
578 }
579
580 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
581 {
582 if (shader_version <= limits_array[i].max_version)
583 {
584 shader->limits = &limits_array[i].limits;
585 break;
586 }
587 ++i;
588 }
589 if (!shader->limits)
590 {
591 FIXME("Unexpected shader version \"%u.%u\".\n",
592 shader->reg_maps.shader_version.major,
593 shader->reg_maps.shader_version.minor);
594 shader->limits = &limits_array[max(0, i - 1)].limits;
595 }
596 }
597
598 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
599 {
600 DWORD idx, shift;
601 idx = bit >> 5;
602 shift = bit & 0x1f;
603 bitmap[idx] |= (1u << shift);
604 }
605
606 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
607 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
608 {
609 switch (reg->type)
610 {
611 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
612 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
613 reg_maps->texcoord |= 1u << reg->idx[0].offset;
614 else
615 reg_maps->address |= 1u << reg->idx[0].offset;
616 break;
617
618 case WINED3DSPR_TEMP:
619 reg_maps->temporary |= 1u << reg->idx[0].offset;
620 break;
621
622 case WINED3DSPR_INPUT:
623 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
624 {
625 if (reg->idx[0].rel_addr)
626 {
627 /* If relative addressing is used, we must assume that all registers
628 * are used. Even if it is a construct like v3[aL], we can't assume
629 * that v0, v1 and v2 aren't read because aL can be negative */
630 unsigned int i;
631 for (i = 0; i < MAX_REG_INPUT; ++i)
632 {
633 shader->u.ps.input_reg_used[i] = TRUE;
634 }
635 }
636 else
637 {
638 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
639 }
640 }
641 else
642 reg_maps->input_registers |= 1u << reg->idx[0].offset;
643 break;
644
645 case WINED3DSPR_RASTOUT:
646 if (reg->idx[0].offset == 1)
647 reg_maps->fog = 1;
648 if (reg->idx[0].offset == 2)
649 reg_maps->point_size = 1;
650 break;
651
652 case WINED3DSPR_MISCTYPE:
653 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
654 {
655 if (!reg->idx[0].offset)
656 reg_maps->vpos = 1;
657 else if (reg->idx[0].offset == 1)
658 reg_maps->usesfacing = 1;
659 }
660 break;
661
662 case WINED3DSPR_CONST:
663 if (reg->idx[0].rel_addr)
664 {
665 if (reg->idx[0].offset < reg_maps->min_rel_offset)
666 reg_maps->min_rel_offset = reg->idx[0].offset;
667 if (reg->idx[0].offset > reg_maps->max_rel_offset)
668 reg_maps->max_rel_offset = reg->idx[0].offset;
669 reg_maps->usesrelconstF = TRUE;
670 }
671 else
672 {
673 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
674 {
675 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
676 return FALSE;
677 }
678 else
679 {
680 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
681 }
682 }
683 break;
684
685 case WINED3DSPR_CONSTINT:
686 if (reg->idx[0].offset >= shader->limits->constant_int)
687 {
688 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
689 return FALSE;
690 }
691 else
692 {
693 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
694 }
695 break;
696
697 case WINED3DSPR_CONSTBOOL:
698 if (reg->idx[0].offset >= shader->limits->constant_bool)
699 {
700 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
701 return FALSE;
702 }
703 else
704 {
705 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
706 }
707 break;
708
709 case WINED3DSPR_COLOROUT:
710 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
711 break;
712
713 default:
714 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
715 reg->type, reg->idx[0].offset, reg->idx[1].offset);
716 break;
717 }
718 return TRUE;
719 }
720
721 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
722 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
723 {
724 struct wined3d_shader_sampler_map_entry *entries, *entry;
725 struct wined3d_shader_sampler_map *map;
726 unsigned int i;
727
728 map = &reg_maps->sampler_map;
729 entries = map->entries;
730 for (i = 0; i < map->count; ++i)
731 {
732 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
733 return;
734 }
735
736 if (!map->size)
737 {
738 if (!(entries = HeapAlloc(GetProcessHeap(), 0, sizeof(*entries) * 4)))
739 {
740 ERR("Failed to allocate sampler map entries.\n");
741 return;
742 }
743 map->size = 4;
744 map->entries = entries;
745 }
746 else if (map->count == map->size)
747 {
748 size_t new_size = map->size * 2;
749
750 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
751 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
752 {
753 ERR("Failed to resize sampler map entries.\n");
754 return;
755 }
756 map->size = new_size;
757 map->entries = entries;
758 }
759
760 entry = &entries[map->count++];
761 entry->resource_idx = resource_idx;
762 entry->sampler_idx = sampler_idx;
763 entry->bind_idx = bind_idx;
764 }
765
766 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
767 {
768 switch (instr)
769 {
770 case WINED3DSIH_M4x4:
771 case WINED3DSIH_M3x4:
772 return param == 1 ? 3 : 0;
773
774 case WINED3DSIH_M4x3:
775 case WINED3DSIH_M3x3:
776 return param == 1 ? 2 : 0;
777
778 case WINED3DSIH_M3x2:
779 return param == 1 ? 1 : 0;
780
781 default:
782 return 0;
783 }
784 }
785
786 /* Note that this does not count the loop register as an address register. */
787 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
788 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
789 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
790 {
791 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
792 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
793 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
794 void *fe_data = shader->frontend_data;
795 struct wined3d_shader_version shader_version;
796 const DWORD *ptr = byte_code;
797 unsigned int i;
798
799 memset(reg_maps, 0, sizeof(*reg_maps));
800 memset(input_signature_elements, 0, sizeof(input_signature_elements));
801 memset(output_signature_elements, 0, sizeof(output_signature_elements));
802 reg_maps->min_rel_offset = ~0U;
803
804 fe->shader_read_header(fe_data, &ptr, &shader_version);
805 reg_maps->shader_version = shader_version;
806
807 shader_set_limits(shader);
808
809 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
810 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
811 if (!reg_maps->constf)
812 {
813 ERR("Failed to allocate constant map memory.\n");
814 return E_OUTOFMEMORY;
815 }
816
817 while (!fe->shader_is_end(fe_data, &ptr))
818 {
819 struct wined3d_shader_instruction ins;
820
821 /* Fetch opcode. */
822 fe->shader_read_instruction(fe_data, &ptr, &ins);
823
824 /* Unhandled opcode, and its parameters. */
825 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
826 {
827 TRACE("Skipping unrecognized instruction.\n");
828 continue;
829 }
830
831 /* Handle declarations. */
832 if (ins.handler_idx == WINED3DSIH_DCL)
833 {
834 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
835 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
836
837 switch (semantic->reg.reg.type)
838 {
839 /* Mark input registers used. */
840 case WINED3DSPR_INPUT:
841 if (reg_idx >= MAX_REG_INPUT)
842 {
843 ERR("Invalid input register index %u.\n", reg_idx);
844 break;
845 }
846 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
847 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
848 return WINED3DERR_INVALIDCALL;
849 reg_maps->input_registers |= 1u << reg_idx;
850 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
851 break;
852
853 /* Vertex shader: mark 3.0 output registers used, save token. */
854 case WINED3DSPR_OUTPUT:
855 if (reg_idx >= MAX_REG_OUTPUT)
856 {
857 ERR("Invalid output register index %u.\n", reg_idx);
858 break;
859 }
860 reg_maps->output_registers |= 1u << reg_idx;
861 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
862 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
863 reg_maps->fog = 1;
864 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
865 reg_maps->point_size = 1;
866 break;
867
868 case WINED3DSPR_SAMPLER:
869 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
870 case WINED3DSPR_RESOURCE:
871 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
872 {
873 ERR("Invalid resource index %u.\n", reg_idx);
874 break;
875 }
876 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
877 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
878 break;
879
880 default:
881 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
882 break;
883 }
884 }
885 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
886 {
887 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
888 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
889 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
890 else
891 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
892 }
893 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
894 {
895 if (reg_maps->icb)
896 FIXME("Multiple immediate constant buffers.\n");
897 reg_maps->icb = ins.declaration.icb;
898 }
899 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
900 {
901 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
902 shader->u.gs.input_type = ins.declaration.primitive_type;
903 else
904 FIXME("Invalid instruction %#x for shader type %#x.\n",
905 ins.handler_idx, shader_version.type);
906 }
907 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
908 {
909 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
910 shader->u.gs.output_type = ins.declaration.primitive_type;
911 else
912 FIXME("Invalid instruction %#x for shader type %#x.\n",
913 ins.handler_idx, shader_version.type);
914 }
915 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
916 {
917 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
918 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
919 }
920 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
921 {
922 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
923 shader->u.gs.vertices_out = ins.declaration.count;
924 else
925 FIXME("Invalid instruction %#x for shader type %#x.\n",
926 ins.handler_idx, shader_version.type);
927 }
928 else if (ins.handler_idx == WINED3DSIH_DEF)
929 {
930 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
931 float *value;
932 if (!lconst) return E_OUTOFMEMORY;
933
934 lconst->idx = ins.dst[0].reg.idx[0].offset;
935 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
936 value = (float *)lconst->value;
937
938 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
939 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
940 {
941 if (value[0] < -1.0f) value[0] = -1.0f;
942 else if (value[0] > 1.0f) value[0] = 1.0f;
943 if (value[1] < -1.0f) value[1] = -1.0f;
944 else if (value[1] > 1.0f) value[1] = 1.0f;
945 if (value[2] < -1.0f) value[2] = -1.0f;
946 else if (value[2] > 1.0f) value[2] = 1.0f;
947 if (value[3] < -1.0f) value[3] = -1.0f;
948 else if (value[3] > 1.0f) value[3] = 1.0f;
949 }
950
951 list_add_head(&shader->constantsF, &lconst->entry);
952
953 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
954 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
955 {
956 shader->lconst_inf_or_nan = TRUE;
957 }
958 }
959 else if (ins.handler_idx == WINED3DSIH_DEFI)
960 {
961 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
962 if (!lconst) return E_OUTOFMEMORY;
963
964 lconst->idx = ins.dst[0].reg.idx[0].offset;
965 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
966
967 list_add_head(&shader->constantsI, &lconst->entry);
968 reg_maps->local_int_consts |= (1u << lconst->idx);
969 }
970 else if (ins.handler_idx == WINED3DSIH_DEFB)
971 {
972 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
973 if (!lconst) return E_OUTOFMEMORY;
974
975 lconst->idx = ins.dst[0].reg.idx[0].offset;
976 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
977
978 list_add_head(&shader->constantsB, &lconst->entry);
979 reg_maps->local_bool_consts |= (1u << lconst->idx);
980 }
981 /* For subroutine prototypes. */
982 else if (ins.handler_idx == WINED3DSIH_LABEL)
983 {
984 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
985 }
986 /* Set texture, address, temporary registers. */
987 else
988 {
989 BOOL color0_mov = FALSE;
990 unsigned int i;
991
992 /* This will loop over all the registers and try to
993 * make a bitmask of the ones we're interested in.
994 *
995 * Relative addressing tokens are ignored, but that's
996 * okay, since we'll catch any address registers when
997 * they are initialized (required by spec). */
998 for (i = 0; i < ins.dst_count; ++i)
999 {
1000 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1001 shader_version.type, constf_size))
1002 return WINED3DERR_INVALIDCALL;
1003
1004 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1005 {
1006 UINT idx = ins.dst[i].reg.idx[0].offset;
1007
1008 switch (ins.dst[i].reg.type)
1009 {
1010 case WINED3DSPR_RASTOUT:
1011 if (shader_version.major >= 3)
1012 break;
1013 switch (idx)
1014 {
1015 case 0: /* oPos */
1016 reg_maps->output_registers |= 1u << 10;
1017 shader_signature_from_usage(&output_signature_elements[10],
1018 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1019 break;
1020
1021 case 1: /* oFog */
1022 reg_maps->output_registers |= 1u << 11;
1023 shader_signature_from_usage(&output_signature_elements[11],
1024 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1025 break;
1026
1027 case 2: /* oPts */
1028 reg_maps->output_registers |= 1u << 11;
1029 shader_signature_from_usage(&output_signature_elements[11],
1030 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1031 break;
1032 }
1033 break;
1034
1035 case WINED3DSPR_ATTROUT:
1036 if (shader_version.major >= 3)
1037 break;
1038 if (idx < 2)
1039 {
1040 idx += 8;
1041 if (reg_maps->output_registers & (1u << idx))
1042 {
1043 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1044 }
1045 else
1046 {
1047 reg_maps->output_registers |= 1u << idx;
1048 shader_signature_from_usage(&output_signature_elements[idx],
1049 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1050 }
1051 }
1052 break;
1053
1054 case WINED3DSPR_TEXCRDOUT:
1055 if (shader_version.major >= 3)
1056 {
1057 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1058 break;
1059 }
1060 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1061 if (reg_maps->output_registers & (1u << idx))
1062 {
1063 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1064 }
1065 else
1066 {
1067 reg_maps->output_registers |= 1u << idx;
1068 shader_signature_from_usage(&output_signature_elements[idx],
1069 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1070 }
1071 break;
1072
1073 default:
1074 break;
1075 }
1076 }
1077
1078 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1079 {
1080 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1081 {
1082 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1083 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1084 * the mov and perform the sRGB write correction from the source register.
1085 *
1086 * However, if the mov is only partial, we can't do this, and if the write
1087 * comes from an instruction other than MOV it is hard to do as well. If
1088 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1089 shader->u.ps.color0_mov = FALSE;
1090 if (ins.handler_idx == WINED3DSIH_MOV
1091 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1092 {
1093 /* Used later when the source register is read. */
1094 color0_mov = TRUE;
1095 }
1096 }
1097 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1098 * end
1099 */
1100 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1101 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1102 {
1103 shader->u.ps.color0_mov = FALSE;
1104 }
1105 }
1106
1107 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1108 if (shader_version.major == 1
1109 && (ins.handler_idx == WINED3DSIH_TEX
1110 || ins.handler_idx == WINED3DSIH_TEXBEM
1111 || ins.handler_idx == WINED3DSIH_TEXBEML
1112 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1113 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1114 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1115 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1116 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1117 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1118 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1119 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1120 {
1121 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1122
1123 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1124 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1125 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1126 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1127
1128 /* texbem is only valid with < 1.4 pixel shaders */
1129 if (ins.handler_idx == WINED3DSIH_TEXBEM
1130 || ins.handler_idx == WINED3DSIH_TEXBEML)
1131 {
1132 reg_maps->bumpmat |= 1u << reg_idx;
1133 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1134 {
1135 reg_maps->luminanceparams |= 1u << reg_idx;
1136 }
1137 }
1138 }
1139 else if (ins.handler_idx == WINED3DSIH_BEM)
1140 {
1141 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1142 }
1143 }
1144
1145 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1146 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
1147 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
1148 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1149 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1150 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1151 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1152 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1153 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1154 else if (ins.handler_idx == WINED3DSIH_LOOP
1155 || ins.handler_idx == WINED3DSIH_REP)
1156 {
1157 ++cur_loop_depth;
1158 if (cur_loop_depth > max_loop_depth)
1159 max_loop_depth = cur_loop_depth;
1160 }
1161 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1162 || ins.handler_idx == WINED3DSIH_ENDREP)
1163 {
1164 --cur_loop_depth;
1165 }
1166 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1167 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1168 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1169 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1170 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1171 {
1172 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1173 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1174 }
1175 else if (ins.handler_idx == WINED3DSIH_LD
1176 || ins.handler_idx == WINED3DSIH_RESINFO)
1177 {
1178 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1179 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1180 }
1181
1182 if (ins.predicate)
1183 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1184 shader_version.type, constf_size))
1185 return WINED3DERR_INVALIDCALL;
1186
1187 for (i = 0; i < ins.src_count; ++i)
1188 {
1189 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1190 struct wined3d_shader_register reg = ins.src[i].reg;
1191
1192 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1193 shader_version.type, constf_size))
1194 return WINED3DERR_INVALIDCALL;
1195 while (count)
1196 {
1197 ++reg.idx[0].offset;
1198 if (!shader_record_register_usage(shader, reg_maps, &reg,
1199 shader_version.type, constf_size))
1200 return WINED3DERR_INVALIDCALL;
1201 --count;
1202 }
1203
1204 if (color0_mov)
1205 {
1206 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1207 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1208 {
1209 shader->u.ps.color0_mov = TRUE;
1210 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1211 }
1212 }
1213 }
1214 }
1215 }
1216 reg_maps->loop_depth = max_loop_depth;
1217
1218 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1219 * R0 is written to the render target. */
1220 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1221 reg_maps->rt_mask |= (1u << 0);
1222
1223 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1224
1225 if (input_signature->elements)
1226 {
1227 for (i = 0; i < input_signature->element_count; ++i)
1228 {
1229 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1230 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL
1231 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1232 reg_maps->vpos = 1;
1233 }
1234 }
1235 else if (!input_signature->elements && reg_maps->input_registers)
1236 {
1237 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1238 struct wined3d_shader_signature_element *e;
1239 unsigned int i;
1240
1241 if (!(input_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*input_signature->elements) * count)))
1242 return E_OUTOFMEMORY;
1243 input_signature->element_count = count;
1244
1245 e = input_signature->elements;
1246 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1247 {
1248 if (!(reg_maps->input_registers & (1u << i)))
1249 continue;
1250 input_signature_elements[i].register_idx = i;
1251 *e++ = input_signature_elements[i];
1252 }
1253 }
1254
1255 if (output_signature->elements)
1256 {
1257 for (i = 0; i < output_signature->element_count; ++i)
1258 {
1259 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1260 }
1261 }
1262 else if (reg_maps->output_registers)
1263 {
1264 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1265 struct wined3d_shader_signature_element *e;
1266
1267 if (!(output_signature->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*output_signature->elements) * count)))
1268 return E_OUTOFMEMORY;
1269 output_signature->element_count = count;
1270
1271 e = output_signature->elements;
1272 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1273 {
1274 if (!(reg_maps->output_registers & (1u << i)))
1275 continue;
1276 *e++ = output_signature_elements[i];
1277 }
1278 }
1279
1280 return WINED3D_OK;
1281 }
1282
1283 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1284 {
1285 DWORD map = 1u << max;
1286 map |= map - 1;
1287 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1288
1289 return wined3d_log2i(map);
1290 }
1291
1292 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1293 {
1294 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1295 {
1296 shader_addline(buffer, "refactoringAllowed");
1297 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1298 if (global_flags)
1299 shader_addline(buffer, " | ");
1300 }
1301
1302 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1303 {
1304 shader_addline(buffer, "enableRawAndStructuredBuffers");
1305 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1306 }
1307
1308 if (global_flags)
1309 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1310 }
1311
1312 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1313 enum wined3d_tessellator_domain domain)
1314 {
1315 switch (domain)
1316 {
1317 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1318 shader_addline(buffer, "line");
1319 break;
1320 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1321 shader_addline(buffer, "triangle");
1322 break;
1323 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1324 shader_addline(buffer, "quad");
1325 break;
1326 default:
1327 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1328 break;
1329 }
1330 }
1331
1332 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1333 enum wined3d_tessellator_output_primitive output_primitive)
1334 {
1335 switch (output_primitive)
1336 {
1337 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1338 shader_addline(buffer, "point");
1339 break;
1340 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1341 shader_addline(buffer, "line");
1342 break;
1343 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1344 shader_addline(buffer, "triangle_cw");
1345 break;
1346 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1347 shader_addline(buffer, "triangle_ccw");
1348 break;
1349 default:
1350 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1351 break;
1352 }
1353 }
1354
1355 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1356 enum wined3d_tessellator_partitioning partitioning)
1357 {
1358 switch (partitioning)
1359 {
1360 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1361 shader_addline(buffer, "integer");
1362 break;
1363 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1364 shader_addline(buffer, "pow2");
1365 break;
1366 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1367 shader_addline(buffer, "fractional_odd");
1368 break;
1369 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1370 shader_addline(buffer, "fractional_even");
1371 break;
1372 default:
1373 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1374 break;
1375 }
1376 }
1377
1378 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1379 {
1380 unsigned int i;
1381
1382 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1383 {
1384 if (sysval_semantic_names[i].sysval_semantic == semantic)
1385 {
1386 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1387 return;
1388 }
1389 }
1390
1391 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1392 }
1393
1394 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1395 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1396 {
1397 shader_addline(buffer, "dcl");
1398
1399 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1400 {
1401 switch (semantic->resource_type)
1402 {
1403 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1404 shader_addline(buffer, "_2d");
1405 break;
1406
1407 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1408 shader_addline(buffer, "_3d");
1409 break;
1410
1411 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1412 shader_addline(buffer, "_cube");
1413 break;
1414
1415 default:
1416 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1417 break;
1418 }
1419 }
1420 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1421 {
1422 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1423 shader_addline(buffer, "_resource_");
1424 else
1425 shader_addline(buffer, "_uav_");
1426 switch (semantic->resource_type)
1427 {
1428 case WINED3D_SHADER_RESOURCE_BUFFER:
1429 shader_addline(buffer, "buffer");
1430 break;
1431
1432 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1433 shader_addline(buffer, "texture1d");
1434 break;
1435
1436 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1437 shader_addline(buffer, "texture2d");
1438 break;
1439
1440 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1441 shader_addline(buffer, "texture2dms");
1442 break;
1443
1444 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1445 shader_addline(buffer, "texture3d");
1446 break;
1447
1448 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1449 shader_addline(buffer, "texturecube");
1450 break;
1451
1452 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1453 shader_addline(buffer, "texture1darray");
1454 break;
1455
1456 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1457 shader_addline(buffer, "texture2darray");
1458 break;
1459
1460 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1461 shader_addline(buffer, "texture2dmsarray");
1462 break;
1463
1464 default:
1465 shader_addline(buffer, "unknown");
1466 break;
1467 }
1468 switch (semantic->resource_data_type)
1469 {
1470 case WINED3D_DATA_FLOAT:
1471 shader_addline(buffer, " (float)");
1472 break;
1473
1474 case WINED3D_DATA_INT:
1475 shader_addline(buffer, " (int)");
1476 break;
1477
1478 case WINED3D_DATA_UINT:
1479 shader_addline(buffer, " (uint)");
1480 break;
1481
1482 case WINED3D_DATA_UNORM:
1483 shader_addline(buffer, " (unorm)");
1484 break;
1485
1486 case WINED3D_DATA_SNORM:
1487 shader_addline(buffer, " (snorm)");
1488 break;
1489
1490 default:
1491 shader_addline(buffer, " (unknown)");
1492 break;
1493 }
1494 }
1495 else
1496 {
1497 /* Pixel shaders 3.0 don't have usage semantics. */
1498 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1499 return;
1500 else
1501 shader_addline(buffer, "_");
1502
1503 switch (semantic->usage)
1504 {
1505 case WINED3D_DECL_USAGE_POSITION:
1506 shader_addline(buffer, "position%u", semantic->usage_idx);
1507 break;
1508
1509 case WINED3D_DECL_USAGE_BLEND_INDICES:
1510 shader_addline(buffer, "blend");
1511 break;
1512
1513 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1514 shader_addline(buffer, "weight");
1515 break;
1516
1517 case WINED3D_DECL_USAGE_NORMAL:
1518 shader_addline(buffer, "normal%u", semantic->usage_idx);
1519 break;
1520
1521 case WINED3D_DECL_USAGE_PSIZE:
1522 shader_addline(buffer, "psize");
1523 break;
1524
1525 case WINED3D_DECL_USAGE_COLOR:
1526 if (!semantic->usage_idx)
1527 shader_addline(buffer, "color");
1528 else
1529 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1530 break;
1531
1532 case WINED3D_DECL_USAGE_TEXCOORD:
1533 shader_addline(buffer, "texture%u", semantic->usage_idx);
1534 break;
1535
1536 case WINED3D_DECL_USAGE_TANGENT:
1537 shader_addline(buffer, "tangent");
1538 break;
1539
1540 case WINED3D_DECL_USAGE_BINORMAL:
1541 shader_addline(buffer, "binormal");
1542 break;
1543
1544 case WINED3D_DECL_USAGE_TESS_FACTOR:
1545 shader_addline(buffer, "tessfactor");
1546 break;
1547
1548 case WINED3D_DECL_USAGE_POSITIONT:
1549 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1550 break;
1551
1552 case WINED3D_DECL_USAGE_FOG:
1553 shader_addline(buffer, "fog");
1554 break;
1555
1556 case WINED3D_DECL_USAGE_DEPTH:
1557 shader_addline(buffer, "depth");
1558 break;
1559
1560 case WINED3D_DECL_USAGE_SAMPLE:
1561 shader_addline(buffer, "sample");
1562 break;
1563
1564 default:
1565 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1566 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1567 }
1568 }
1569 }
1570
1571 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1572 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1573 {
1574 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1575 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1576 UINT offset = reg->idx[0].offset;
1577
1578 switch (reg->type)
1579 {
1580 case WINED3DSPR_TEMP:
1581 shader_addline(buffer, "r");
1582 break;
1583
1584 case WINED3DSPR_INPUT:
1585 shader_addline(buffer, "v");
1586 break;
1587
1588 case WINED3DSPR_CONST:
1589 case WINED3DSPR_CONST2:
1590 case WINED3DSPR_CONST3:
1591 case WINED3DSPR_CONST4:
1592 shader_addline(buffer, "c");
1593 offset = shader_get_float_offset(reg->type, offset);
1594 break;
1595
1596 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1597 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1598 break;
1599
1600 case WINED3DSPR_RASTOUT:
1601 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1602 break;
1603
1604 case WINED3DSPR_COLOROUT:
1605 shader_addline(buffer, "oC");
1606 break;
1607
1608 case WINED3DSPR_DEPTHOUT:
1609 shader_addline(buffer, "oDepth");
1610 break;
1611
1612 case WINED3DSPR_ATTROUT:
1613 shader_addline(buffer, "oD");
1614 break;
1615
1616 case WINED3DSPR_TEXCRDOUT:
1617 /* Vertex shaders >= 3.0 use general purpose output registers
1618 * (WINED3DSPR_OUTPUT), which can include an address token. */
1619 if (shader_version->major >= 3)
1620 shader_addline(buffer, "o");
1621 else
1622 shader_addline(buffer, "oT");
1623 break;
1624
1625 case WINED3DSPR_CONSTINT:
1626 shader_addline(buffer, "i");
1627 break;
1628
1629 case WINED3DSPR_CONSTBOOL:
1630 shader_addline(buffer, "b");
1631 break;
1632
1633 case WINED3DSPR_LABEL:
1634 shader_addline(buffer, "l");
1635 break;
1636
1637 case WINED3DSPR_LOOP:
1638 shader_addline(buffer, "aL");
1639 break;
1640
1641 case WINED3DSPR_SAMPLER:
1642 shader_addline(buffer, "s");
1643 break;
1644
1645 case WINED3DSPR_MISCTYPE:
1646 if (offset > 1)
1647 {
1648 FIXME("Unhandled misctype register %u.\n", offset);
1649 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1650 }
1651 else
1652 {
1653 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1654 }
1655 break;
1656
1657 case WINED3DSPR_PREDICATE:
1658 shader_addline(buffer, "p");
1659 break;
1660
1661 case WINED3DSPR_IMMCONST:
1662 shader_addline(buffer, "l");
1663 break;
1664
1665 case WINED3DSPR_CONSTBUFFER:
1666 shader_addline(buffer, "cb");
1667 break;
1668
1669 case WINED3DSPR_IMMCONSTBUFFER:
1670 shader_addline(buffer, "icb");
1671 break;
1672
1673 case WINED3DSPR_PRIMID:
1674 shader_addline(buffer, "primID");
1675 break;
1676
1677 case WINED3DSPR_NULL:
1678 shader_addline(buffer, "null");
1679 break;
1680
1681 case WINED3DSPR_RESOURCE:
1682 shader_addline(buffer, "t");
1683 break;
1684
1685 case WINED3DSPR_UAV:
1686 shader_addline(buffer, "u");
1687 break;
1688
1689 case WINED3DSPR_FORKINSTID:
1690 shader_addline(buffer, "vForkInstanceId");
1691 break;
1692
1693 case WINED3DSPR_INCONTROLPOINT:
1694 shader_addline(buffer, "vicp");
1695 break;
1696
1697 case WINED3DSPR_TESSCOORD:
1698 shader_addline(buffer, "vDomainLocation");
1699 break;
1700
1701 default:
1702 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1703 break;
1704 }
1705
1706 if (reg->type == WINED3DSPR_IMMCONST)
1707 {
1708 shader_addline(buffer, "(");
1709 switch (reg->immconst_type)
1710 {
1711 case WINED3D_IMMCONST_SCALAR:
1712 switch (reg->data_type)
1713 {
1714 case WINED3D_DATA_FLOAT:
1715 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1716 break;
1717 case WINED3D_DATA_INT:
1718 shader_addline(buffer, "%d", reg->immconst_data[0]);
1719 break;
1720 case WINED3D_DATA_RESOURCE:
1721 case WINED3D_DATA_SAMPLER:
1722 case WINED3D_DATA_UINT:
1723 shader_addline(buffer, "%u", reg->immconst_data[0]);
1724 break;
1725 default:
1726 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1727 break;
1728 }
1729 break;
1730
1731 case WINED3D_IMMCONST_VEC4:
1732 switch (reg->data_type)
1733 {
1734 case WINED3D_DATA_FLOAT:
1735 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1736 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1737 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1738 break;
1739 case WINED3D_DATA_INT:
1740 shader_addline(buffer, "%d, %d, %d, %d",
1741 reg->immconst_data[0], reg->immconst_data[1],
1742 reg->immconst_data[2], reg->immconst_data[3]);
1743 break;
1744 case WINED3D_DATA_RESOURCE:
1745 case WINED3D_DATA_SAMPLER:
1746 case WINED3D_DATA_UINT:
1747 shader_addline(buffer, "%u, %u, %u, %u",
1748 reg->immconst_data[0], reg->immconst_data[1],
1749 reg->immconst_data[2], reg->immconst_data[3]);
1750 break;
1751 default:
1752 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1753 break;
1754 }
1755 break;
1756
1757 default:
1758 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1759 break;
1760 }
1761 shader_addline(buffer, ")");
1762 }
1763 else if (reg->type != WINED3DSPR_RASTOUT
1764 && reg->type != WINED3DSPR_MISCTYPE
1765 && reg->type != WINED3DSPR_NULL)
1766 {
1767 if (offset != ~0u)
1768 {
1769 shader_addline(buffer, "[");
1770 if (reg->idx[0].rel_addr)
1771 {
1772 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1773 shader_addline(buffer, " + ");
1774 }
1775 shader_addline(buffer, "%u]", offset);
1776
1777 if (reg->idx[1].offset != ~0u)
1778 {
1779 shader_addline(buffer, "[");
1780 if (reg->idx[1].rel_addr)
1781 {
1782 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1783 shader_addline(buffer, " + ");
1784 }
1785 shader_addline(buffer, "%u]", reg->idx[1].offset);
1786 }
1787 }
1788 }
1789 }
1790
1791 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1792 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1793 {
1794 DWORD write_mask = param->write_mask;
1795
1796 shader_dump_register(buffer, &param->reg, shader_version);
1797
1798 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1799 {
1800 static const char write_mask_chars[] = "xyzw";
1801
1802 shader_addline(buffer, ".");
1803 if (write_mask & WINED3DSP_WRITEMASK_0)
1804 shader_addline(buffer, "%c", write_mask_chars[0]);
1805 if (write_mask & WINED3DSP_WRITEMASK_1)
1806 shader_addline(buffer, "%c", write_mask_chars[1]);
1807 if (write_mask & WINED3DSP_WRITEMASK_2)
1808 shader_addline(buffer, "%c", write_mask_chars[2]);
1809 if (write_mask & WINED3DSP_WRITEMASK_3)
1810 shader_addline(buffer, "%c", write_mask_chars[3]);
1811 }
1812 }
1813
1814 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1815 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1816 {
1817 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1818 DWORD swizzle = param->swizzle;
1819
1820 if (src_modifier == WINED3DSPSM_NEG
1821 || src_modifier == WINED3DSPSM_BIASNEG
1822 || src_modifier == WINED3DSPSM_SIGNNEG
1823 || src_modifier == WINED3DSPSM_X2NEG
1824 || src_modifier == WINED3DSPSM_ABSNEG)
1825 shader_addline(buffer, "-");
1826 else if (src_modifier == WINED3DSPSM_COMP)
1827 shader_addline(buffer, "1-");
1828 else if (src_modifier == WINED3DSPSM_NOT)
1829 shader_addline(buffer, "!");
1830
1831 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1832 shader_addline(buffer, "abs(");
1833
1834 shader_dump_register(buffer, &param->reg, shader_version);
1835
1836 switch (src_modifier)
1837 {
1838 case WINED3DSPSM_NONE: break;
1839 case WINED3DSPSM_NEG: break;
1840 case WINED3DSPSM_NOT: break;
1841 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1842 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1843 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1844 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1845 case WINED3DSPSM_COMP: break;
1846 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1847 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1848 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1849 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1850 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1851 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1852 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1853 }
1854
1855 if (swizzle != WINED3DSP_NOSWIZZLE)
1856 {
1857 static const char swizzle_chars[] = "xyzw";
1858 DWORD swizzle_x = swizzle & 0x03;
1859 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1860 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1861 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1862
1863 if (swizzle_x == swizzle_y
1864 && swizzle_x == swizzle_z
1865 && swizzle_x == swizzle_w)
1866 {
1867 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
1868 }
1869 else
1870 {
1871 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1872 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1873 }
1874 }
1875 }
1876
1877 /* Shared code in order to generate the bulk of the shader string.
1878 * NOTE: A description of how to parse tokens can be found on MSDN. */
1879 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1880 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1881 {
1882 struct wined3d_device *device = shader->device;
1883 const struct wined3d_shader_frontend *fe = shader->frontend;
1884 void *fe_data = shader->frontend_data;
1885 struct wined3d_shader_version shader_version;
1886 struct wined3d_shader_loop_state loop_state;
1887 struct wined3d_shader_instruction ins;
1888 struct wined3d_shader_tex_mx tex_mx;
1889 struct wined3d_shader_context ctx;
1890 const DWORD *ptr = byte_code;
1891
1892 /* Initialize current parsing state. */
1893 tex_mx.current_row = 0;
1894 loop_state.current_depth = 0;
1895 loop_state.current_reg = 0;
1896
1897 ctx.shader = shader;
1898 ctx.gl_info = &device->adapter->gl_info;
1899 ctx.reg_maps = reg_maps;
1900 ctx.buffer = buffer;
1901 ctx.tex_mx = &tex_mx;
1902 ctx.loop_state = &loop_state;
1903 ctx.backend_data = backend_ctx;
1904 ins.ctx = &ctx;
1905
1906 fe->shader_read_header(fe_data, &ptr, &shader_version);
1907
1908 while (!fe->shader_is_end(fe_data, &ptr))
1909 {
1910 /* Read opcode. */
1911 fe->shader_read_instruction(fe_data, &ptr, &ins);
1912
1913 /* Unknown opcode and its parameters. */
1914 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1915 {
1916 TRACE("Skipping unrecognized instruction.\n");
1917 continue;
1918 }
1919
1920 if (ins.predicate)
1921 FIXME("Predicates not implemented.\n");
1922
1923 /* Call appropriate function for output target */
1924 device->shader_backend->shader_handle_instruction(&ins);
1925 }
1926 }
1927
1928 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
1929 const struct wined3d_shader_dst_param *dst)
1930 {
1931 DWORD mmask = dst->modifiers;
1932
1933 switch (dst->shift)
1934 {
1935 case 0: break;
1936 case 13: shader_addline(buffer, "_d8"); break;
1937 case 14: shader_addline(buffer, "_d4"); break;
1938 case 15: shader_addline(buffer, "_d2"); break;
1939 case 1: shader_addline(buffer, "_x2"); break;
1940 case 2: shader_addline(buffer, "_x4"); break;
1941 case 3: shader_addline(buffer, "_x8"); break;
1942 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
1943 }
1944
1945 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
1946 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
1947 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
1948
1949 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1950 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1951 }
1952
1953 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
1954 enum wined3d_primitive_type primitive_type)
1955 {
1956 switch (primitive_type)
1957 {
1958 case WINED3D_PT_UNDEFINED:
1959 shader_addline(buffer, "undefined");
1960 break;
1961 case WINED3D_PT_POINTLIST:
1962 shader_addline(buffer, "pointlist");
1963 break;
1964 case WINED3D_PT_LINELIST:
1965 shader_addline(buffer, "linelist");
1966 break;
1967 case WINED3D_PT_LINESTRIP:
1968 shader_addline(buffer, "linestrip");
1969 break;
1970 case WINED3D_PT_TRIANGLELIST:
1971 shader_addline(buffer, "trianglelist");
1972 break;
1973 case WINED3D_PT_TRIANGLESTRIP:
1974 shader_addline(buffer, "trianglestrip");
1975 break;
1976 case WINED3D_PT_TRIANGLEFAN:
1977 shader_addline(buffer, "trianglefan");
1978 break;
1979 case WINED3D_PT_LINELIST_ADJ:
1980 shader_addline(buffer, "linelist_adj");
1981 break;
1982 case WINED3D_PT_LINESTRIP_ADJ:
1983 shader_addline(buffer, "linestrip_adj");
1984 break;
1985 case WINED3D_PT_TRIANGLELIST_ADJ:
1986 shader_addline(buffer, "trianglelist_adj");
1987 break;
1988 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1989 shader_addline(buffer, "trianglestrip_adj");
1990 break;
1991 default:
1992 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
1993 break;
1994 }
1995 }
1996
1997 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
1998 enum wined3d_shader_interpolation_mode interpolation_mode)
1999 {
2000 switch (interpolation_mode)
2001 {
2002 case WINED3DSIM_CONSTANT:
2003 shader_addline(buffer, "constant");
2004 break;
2005 case WINED3DSIM_LINEAR:
2006 shader_addline(buffer, "linear");
2007 break;
2008 case WINED3DSIM_LINEAR_CENTROID:
2009 shader_addline(buffer, "linear centroid");
2010 break;
2011 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2012 shader_addline(buffer, "linear noperspective");
2013 break;
2014 case WINED3DSIM_LINEAR_SAMPLE:
2015 shader_addline(buffer, "linear sample");
2016 break;
2017 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2018 shader_addline(buffer, "linear noperspective centroid");
2019 break;
2020 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2021 shader_addline(buffer, "linear noperspective sample");
2022 break;
2023 default:
2024 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2025 break;
2026 }
2027 }
2028
2029 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2030 {
2031 struct wined3d_shader_version shader_version;
2032 struct wined3d_string_buffer buffer;
2033 const DWORD *ptr = byte_code;
2034 const char *type_prefix;
2035 const char *p, *q;
2036 DWORD i;
2037
2038 if (!string_buffer_init(&buffer))
2039 {
2040 ERR("Failed to initialize string buffer.\n");
2041 return;
2042 }
2043
2044 TRACE("Parsing %p.\n", byte_code);
2045
2046 fe->shader_read_header(fe_data, &ptr, &shader_version);
2047
2048 switch (shader_version.type)
2049 {
2050 case WINED3D_SHADER_TYPE_VERTEX:
2051 type_prefix = "vs";
2052 break;
2053
2054 case WINED3D_SHADER_TYPE_HULL:
2055 type_prefix = "hs";
2056 break;
2057
2058 case WINED3D_SHADER_TYPE_DOMAIN:
2059 type_prefix = "ds";
2060 break;
2061
2062 case WINED3D_SHADER_TYPE_GEOMETRY:
2063 type_prefix = "gs";
2064 break;
2065
2066 case WINED3D_SHADER_TYPE_PIXEL:
2067 type_prefix = "ps";
2068 break;
2069
2070 default:
2071 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2072 type_prefix = "unknown";
2073 break;
2074 }
2075
2076 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2077
2078 while (!fe->shader_is_end(fe_data, &ptr))
2079 {
2080 struct wined3d_shader_instruction ins;
2081
2082 fe->shader_read_instruction(fe_data, &ptr, &ins);
2083 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2084 {
2085 WARN("Skipping unrecognized instruction.\n");
2086 shader_addline(&buffer, "<unrecognized instruction>\n");
2087 continue;
2088 }
2089
2090 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2091 {
2092 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2093 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2094 shader_addline(&buffer, " ");
2095 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2096 }
2097 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2098 {
2099 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2100 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2101 shader_addline(&buffer, ", %s",
2102 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2103 }
2104 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2105 {
2106 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2107 shader_dump_global_flags(&buffer, ins.flags);
2108 }
2109 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2110 {
2111 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2112 ins.declaration.max_tessellation_factor);
2113 }
2114 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2115 {
2116 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2117 for (i = 0; i < ins.declaration.icb->element_count / 4; ++i)
2118 {
2119 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2120 ins.declaration.icb->data[4 * i + 0],
2121 ins.declaration.icb->data[4 * i + 1],
2122 ins.declaration.icb->data[4 * i + 2],
2123 ins.declaration.icb->data[4 * i + 3]);
2124 }
2125 shader_addline(&buffer, "}");
2126 }
2127 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2128 {
2129 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2130 shader_dump_interpolation_mode(&buffer, ins.flags);
2131 shader_addline(&buffer, " ");
2132 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2133 }
2134 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2135 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2136 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2137 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2138 {
2139 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2140 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2141 shader_addline(&buffer, ", ");
2142 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2143 }
2144 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2145 {
2146 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2147 shader_dump_interpolation_mode(&buffer, ins.flags);
2148 shader_addline(&buffer, " ");
2149 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2150 shader_addline(&buffer, ", ");
2151 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2152 }
2153 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2154 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2155 {
2156 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2157 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2158 }
2159 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2160 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2161 {
2162 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2163 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2164 }
2165 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2166 {
2167 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2168 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2169 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2170 }
2171 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2172 {
2173 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2174 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2175 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2176 shader_addline(&buffer, ", comparisonMode");
2177 }
2178 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2179 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2180 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2181 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2182 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2183 {
2184 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2185 }
2186 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2187 {
2188 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2189 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2190 }
2191 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2192 {
2193 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2194 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2195 }
2196 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2197 {
2198 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2199 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2200 }
2201 else if (ins.handler_idx == WINED3DSIH_DEF)
2202 {
2203 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2204 ins.dst[0].reg.idx[0].offset),
2205 *(const float *)&ins.src[0].reg.immconst_data[0],
2206 *(const float *)&ins.src[0].reg.immconst_data[1],
2207 *(const float *)&ins.src[0].reg.immconst_data[2],
2208 *(const float *)&ins.src[0].reg.immconst_data[3]);
2209 }
2210 else if (ins.handler_idx == WINED3DSIH_DEFI)
2211 {
2212 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2213 ins.src[0].reg.immconst_data[0],
2214 ins.src[0].reg.immconst_data[1],
2215 ins.src[0].reg.immconst_data[2],
2216 ins.src[0].reg.immconst_data[3]);
2217 }
2218 else if (ins.handler_idx == WINED3DSIH_DEFB)
2219 {
2220 shader_addline(&buffer, "defb b%u = %s",
2221 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2222 }
2223 else
2224 {
2225 if (ins.predicate)
2226 {
2227 shader_addline(&buffer, "(");
2228 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2229 shader_addline(&buffer, ") ");
2230 }
2231
2232 /* PixWin marks instructions with the coissue flag with a '+' */
2233 if (ins.coissue)
2234 shader_addline(&buffer, "+");
2235
2236 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2237
2238 if (ins.handler_idx == WINED3DSIH_IFC
2239 || ins.handler_idx == WINED3DSIH_BREAKC)
2240 {
2241 switch (ins.flags)
2242 {
2243 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2244 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2245 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2246 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2247 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2248 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2249 default: shader_addline(&buffer, "_(%u)", ins.flags);
2250 }
2251 }
2252 else if (ins.handler_idx == WINED3DSIH_TEX
2253 && shader_version.major >= 2
2254 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2255 {
2256 shader_addline(&buffer, "p");
2257 }
2258 else if (ins.handler_idx == WINED3DSIH_RESINFO
2259 && ins.flags)
2260 {
2261 switch (ins.flags)
2262 {
2263 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2264 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2265 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2266 }
2267 }
2268
2269 if (wined3d_shader_instruction_has_texel_offset(&ins))
2270 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2271
2272 for (i = 0; i < ins.dst_count; ++i)
2273 {
2274 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2275 shader_addline(&buffer, !i ? " " : ", ");
2276 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2277 }
2278
2279 /* Other source tokens */
2280 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2281 {
2282 shader_addline(&buffer, !i ? " " : ", ");
2283 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2284 }
2285 }
2286 shader_addline(&buffer, "\n");
2287 }
2288
2289 for (p = buffer.buffer; *p; p = q)
2290 {
2291 if (!(q = strstr(p, "\n")))
2292 q = p + strlen(p);
2293 else
2294 ++q;
2295 TRACE(" %.*s", (int)(q - p), p);
2296 }
2297
2298 string_buffer_free(&buffer);
2299 }
2300
2301 static void shader_cleanup(struct wined3d_shader *shader)
2302 {
2303 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2304 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2305 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2306 shader->device->shader_backend->shader_destroy(shader);
2307 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
2308 HeapFree(GetProcessHeap(), 0, shader->reg_maps.sampler_map.entries);
2309 HeapFree(GetProcessHeap(), 0, shader->function);
2310 shader_delete_constant_list(&shader->constantsF);
2311 shader_delete_constant_list(&shader->constantsB);
2312 shader_delete_constant_list(&shader->constantsI);
2313 list_remove(&shader->shader_list_entry);
2314
2315 if (shader->frontend && shader->frontend_data)
2316 shader->frontend->shader_free(shader->frontend_data);
2317 }
2318
2319 struct shader_none_priv
2320 {
2321 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2322 const struct fragment_pipeline *fragment_pipe;
2323 BOOL ffp_proj_control;
2324 };
2325
2326 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2327 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2328 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2329 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2330 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2331 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2332 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2333 const struct wined3d_state *state) {}
2334 static void shader_none_destroy(struct wined3d_shader *shader) {}
2335 static void shader_none_free_context_data(struct wined3d_context *context) {}
2336 static void shader_none_init_context_state(struct wined3d_context *context) {}
2337
2338 /* Context activation is done by the caller. */
2339 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2340 const struct wined3d_state *state)
2341 {
2342 const struct wined3d_gl_info *gl_info = context->gl_info;
2343 struct shader_none_priv *priv = shader_priv;
2344
2345 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2346 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2347 }
2348
2349 /* Context activation is done by the caller. */
2350 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2351 {
2352 struct shader_none_priv *priv = shader_priv;
2353 const struct wined3d_gl_info *gl_info = context->gl_info;
2354
2355 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2356 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2357
2358 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2359 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2360 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2361 | (1u << WINED3D_SHADER_TYPE_HULL)
2362 | (1u << WINED3D_SHADER_TYPE_DOMAIN);
2363 }
2364
2365 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2366 const struct fragment_pipeline *fragment_pipe)
2367 {
2368 struct fragment_caps fragment_caps;
2369 void *vertex_priv, *fragment_priv;
2370 struct shader_none_priv *priv;
2371
2372 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2373 return E_OUTOFMEMORY;
2374
2375 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2376 {
2377 ERR("Failed to initialize vertex pipe.\n");
2378 HeapFree(GetProcessHeap(), 0, priv);
2379 return E_FAIL;
2380 }
2381
2382 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2383 {
2384 ERR("Failed to initialize fragment pipe.\n");
2385 vertex_pipe->vp_free(device);
2386 HeapFree(GetProcessHeap(), 0, priv);
2387 return E_FAIL;
2388 }
2389
2390 priv->vertex_pipe = vertex_pipe;
2391 priv->fragment_pipe = fragment_pipe;
2392 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2393 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2394
2395 device->vertex_priv = vertex_priv;
2396 device->fragment_priv = fragment_priv;
2397 device->shader_priv = priv;
2398
2399 return WINED3D_OK;
2400 }
2401
2402 static void shader_none_free(struct wined3d_device *device)
2403 {
2404 struct shader_none_priv *priv = device->shader_priv;
2405
2406 priv->fragment_pipe->free_private(device);
2407 priv->vertex_pipe->vp_free(device);
2408 HeapFree(GetProcessHeap(), 0, priv);
2409 }
2410
2411 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2412 {
2413 return TRUE;
2414 }
2415
2416 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2417 {
2418 /* Set the shader caps to 0 for the none shader backend */
2419 caps->vs_version = 0;
2420 caps->hs_version = 0;
2421 caps->ds_version = 0;
2422 caps->gs_version = 0;
2423 caps->ps_version = 0;
2424 caps->vs_uniform_count = 0;
2425 caps->ps_uniform_count = 0;
2426 caps->ps_1x_max_value = 0.0f;
2427 caps->varying_count = 0;
2428 caps->wined3d_caps = 0;
2429 }
2430
2431 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2432 {
2433 /* We "support" every possible fixup, since we don't support any shader
2434 * model, and will never have to actually sample a texture. */
2435 return TRUE;
2436 }
2437
2438 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2439 {
2440 struct shader_none_priv *priv = shader_priv;
2441
2442 return priv->ffp_proj_control;
2443 }
2444
2445 const struct wined3d_shader_backend_ops none_shader_backend =
2446 {
2447 shader_none_handle_instruction,
2448 shader_none_select,
2449 shader_none_disable,
2450 shader_none_select_depth_blt,
2451 shader_none_deselect_depth_blt,
2452 shader_none_update_float_vertex_constants,
2453 shader_none_update_float_pixel_constants,
2454 shader_none_load_constants,
2455 shader_none_destroy,
2456 shader_none_alloc,
2457 shader_none_free,
2458 shader_none_allocate_context_data,
2459 shader_none_free_context_data,
2460 shader_none_init_context_state,
2461 shader_none_get_caps,
2462 shader_none_color_fixup_supported,
2463 shader_none_has_ffp_proj_control,
2464 };
2465
2466 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2467 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
2468 enum wined3d_shader_type type, unsigned int max_version)
2469 {
2470 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2471 const struct wined3d_shader_frontend *fe;
2472 HRESULT hr;
2473 unsigned int backend_version;
2474 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2475
2476 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u, type %#x, max_version %u.\n",
2477 shader, byte_code, output_signature, float_const_count, type, max_version);
2478
2479 list_init(&shader->constantsF);
2480 list_init(&shader->constantsB);
2481 list_init(&shader->constantsI);
2482 shader->lconst_inf_or_nan = FALSE;
2483
2484 fe = shader_select_frontend(*byte_code);
2485 if (!fe)
2486 {
2487 FIXME("Unable to find frontend for shader.\n");
2488 return WINED3DERR_INVALIDCALL;
2489 }
2490 shader->frontend = fe;
2491 shader->frontend_data = fe->shader_init(byte_code, output_signature);
2492 if (!shader->frontend_data)
2493 {
2494 FIXME("Failed to initialize frontend.\n");
2495 return WINED3DERR_INVALIDCALL;
2496 }
2497
2498 /* First pass: trace shader. */
2499 if (TRACE_ON(d3d_shader))
2500 shader_trace_init(fe, shader->frontend_data, byte_code);
2501
2502 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2503 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2504 &shader->output_signature, byte_code, float_const_count)))
2505 return hr;
2506
2507 if (reg_maps->shader_version.type != type)
2508 {
2509 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2510 return WINED3DERR_INVALIDCALL;
2511 }
2512 if (reg_maps->shader_version.major > max_version)
2513 {
2514 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2515 return WINED3DERR_INVALIDCALL;
2516 }
2517 switch (type)
2518 {
2519 case WINED3D_SHADER_TYPE_VERTEX:
2520 backend_version = d3d_info->limits.vs_version;
2521 break;
2522 case WINED3D_SHADER_TYPE_HULL:
2523 backend_version = d3d_info->limits.hs_version;
2524 break;
2525 case WINED3D_SHADER_TYPE_DOMAIN:
2526 backend_version = d3d_info->limits.ds_version;
2527 break;
2528 case WINED3D_SHADER_TYPE_GEOMETRY:
2529 backend_version = d3d_info->limits.gs_version;
2530 break;
2531 case WINED3D_SHADER_TYPE_PIXEL:
2532 backend_version = d3d_info->limits.ps_version;
2533 break;
2534 default:
2535 FIXME("No backend version-checking for this shader type\n");
2536 backend_version = 0;
2537 }
2538 if (reg_maps->shader_version.major > backend_version)
2539 {
2540 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2541 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2542 return WINED3DERR_INVALIDCALL;
2543 }
2544
2545 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
2546 if (!shader->function)
2547 return E_OUTOFMEMORY;
2548 memcpy(shader->function, byte_code, shader->functionLength);
2549
2550 return WINED3D_OK;
2551 }
2552
2553 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2554 {
2555 ULONG refcount = InterlockedIncrement(&shader->ref);
2556
2557 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2558
2559 return refcount;
2560 }
2561
2562 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2563 {
2564 ULONG refcount = InterlockedDecrement(&shader->ref);
2565
2566 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2567
2568 if (!refcount)
2569 {
2570 shader_cleanup(shader);
2571 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2572 HeapFree(GetProcessHeap(), 0, shader);
2573 }
2574
2575 return refcount;
2576 }
2577
2578 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2579 {
2580 TRACE("shader %p.\n", shader);
2581
2582 return shader->parent;
2583 }
2584
2585 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2586 void *byte_code, UINT *byte_code_size)
2587 {
2588 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2589
2590 if (!byte_code)
2591 {
2592 *byte_code_size = shader->functionLength;
2593 return WINED3D_OK;
2594 }
2595
2596 if (*byte_code_size < shader->functionLength)
2597 {
2598 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2599 * than the required size we should write the required size and
2600 * return D3DERR_MOREDATA. That's not actually true. */
2601 return WINED3DERR_INVALIDCALL;
2602 }
2603
2604 memcpy(byte_code, shader->function, shader->functionLength);
2605
2606 return WINED3D_OK;
2607 }
2608
2609 /* Set local constants for d3d8 shaders. */
2610 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2611 UINT start_idx, const float *src_data, UINT count)
2612 {
2613 UINT end_idx = start_idx + count;
2614 UINT i;
2615
2616 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2617
2618 if (end_idx > shader->limits->constant_float)
2619 {
2620 WARN("end_idx %u > float constants limit %u.\n",
2621 end_idx, shader->limits->constant_float);
2622 end_idx = shader->limits->constant_float;
2623 }
2624
2625 for (i = start_idx; i < end_idx; ++i)
2626 {
2627 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2628 float *value;
2629 if (!lconst)
2630 return E_OUTOFMEMORY;
2631
2632 lconst->idx = i;
2633 value = (float *)lconst->value;
2634 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2635 list_add_head(&shader->constantsF, &lconst->entry);
2636
2637 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2638 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2639 {
2640 shader->lconst_inf_or_nan = TRUE;
2641 }
2642 }
2643
2644 return WINED3D_OK;
2645 }
2646
2647 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2648 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2649 {
2650 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2651 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2652 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2653 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2654 args->point_size = state->gl_primitive_type == GL_POINTS;
2655 args->per_vertex_point_size = shader->reg_maps.point_size;
2656 args->swizzle_map = swizzle_map;
2657 if (d3d_info->emulated_flatshading)
2658 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2659 else
2660 args->flatshading = FALSE;
2661 }
2662
2663 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2664 {
2665 if (usage_idx1 != usage_idx2)
2666 return FALSE;
2667 if (usage1 == usage2)
2668 return TRUE;
2669 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2670 return TRUE;
2671 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2672 return TRUE;
2673
2674 return FALSE;
2675 }
2676
2677 BOOL vshader_get_input(const struct wined3d_shader *shader,
2678 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2679 {
2680 WORD map = shader->reg_maps.input_registers;
2681 unsigned int i;
2682
2683 for (i = 0; map; map >>= 1, ++i)
2684 {
2685 if (!(map & 1)) continue;
2686
2687 if (match_usage(shader->u.vs.attributes[i].usage,
2688 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2689 {
2690 *regnum = i;
2691 return TRUE;
2692 }
2693 }
2694 return FALSE;
2695 }
2696
2697 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2698 const struct wined3d_shader_signature *src, char **signature_strings)
2699 {
2700 struct wined3d_shader_signature_element *e;
2701 unsigned int i;
2702 SIZE_T len;
2703 char *ptr;
2704
2705 ptr = *signature_strings;
2706
2707 dst->element_count = src->element_count;
2708 if (!(dst->elements = HeapAlloc(GetProcessHeap(), 0, sizeof(*dst->elements) * dst->element_count)))
2709 return E_OUTOFMEMORY;
2710
2711 for (i = 0; i < src->element_count; ++i)
2712 {
2713 e = &src->elements[i];
2714 dst->elements[i] = *e;
2715
2716 len = strlen(e->semantic_name);
2717 memcpy(ptr, e->semantic_name, len + 1);
2718 dst->elements[i].semantic_name = ptr;
2719 ptr += len + 1;
2720 }
2721
2722 *signature_strings = ptr;
2723
2724 return WINED3D_OK;
2725 }
2726
2727 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2728 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2729 void *parent, const struct wined3d_parent_ops *parent_ops)
2730 {
2731 struct wined3d_shader_signature_element *e;
2732 SIZE_T total, len;
2733 unsigned int i;
2734 HRESULT hr;
2735 char *ptr;
2736
2737 if (!desc->byte_code)
2738 return WINED3DERR_INVALIDCALL;
2739
2740 shader->ref = 1;
2741 shader->device = device;
2742 shader->parent = parent;
2743 shader->parent_ops = parent_ops;
2744
2745 total = 0;
2746 if (desc->input_signature)
2747 {
2748 for (i = 0; i < desc->input_signature->element_count; ++i)
2749 {
2750 e = &desc->input_signature->elements[i];
2751 len = strlen(e->semantic_name);
2752 if (len >= ~(SIZE_T)0 - total)
2753 return E_OUTOFMEMORY;
2754
2755 total += len + 1;
2756 }
2757 }
2758 if (desc->output_signature)
2759 {
2760 for (i = 0; i < desc->output_signature->element_count; ++i)
2761 {
2762 e = &desc->output_signature->elements[i];
2763 len = strlen(e->semantic_name);
2764 if (len >= ~(SIZE_T)0 - total)
2765 return E_OUTOFMEMORY;
2766
2767 total += len + 1;
2768 }
2769 }
2770 if (!(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2771 return E_OUTOFMEMORY;
2772 ptr = shader->signature_strings;
2773
2774 if (desc->input_signature && FAILED(hr = shader_signature_copy(&shader->input_signature,
2775 desc->input_signature, &ptr)))
2776 {
2777 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2778 return hr;
2779 }
2780 if (desc->output_signature && FAILED(hr = shader_signature_copy(&shader->output_signature,
2781 desc->output_signature, &ptr)))
2782 {
2783 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2784 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2785 return hr;
2786 }
2787
2788 list_init(&shader->linked_programs);
2789 list_add_head(&device->shaders, &shader->shader_list_entry);
2790
2791 if (FAILED(hr = shader_set_function(shader, desc->byte_code, desc->output_signature,
2792 float_const_count, type, desc->max_version)))
2793 {
2794 WARN("Failed to set function, hr %#x.\n", hr);
2795 shader_cleanup(shader);
2796 }
2797
2798 return hr;
2799 }
2800
2801 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2802 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2803 {
2804 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2805 unsigned int i;
2806 HRESULT hr;
2807
2808 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
2809 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
2810 return hr;
2811
2812 for (i = 0; i < shader->input_signature.element_count; ++i)
2813 {
2814 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2815
2816 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2817 continue;
2818
2819 shader->u.vs.attributes[input->register_idx].usage =
2820 shader_usage_from_semantic_name(input->semantic_name);
2821 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2822 }
2823
2824 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2825 shader->lconst_inf_or_nan;
2826
2827 return WINED3D_OK;
2828 }
2829
2830 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2831 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2832 {
2833 HRESULT hr;
2834
2835 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops)))
2836 return hr;
2837
2838 shader->load_local_constsF = shader->lconst_inf_or_nan;
2839
2840 return WINED3D_OK;
2841 }
2842
2843 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2844 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2845 {
2846 HRESULT hr;
2847
2848 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops)))
2849 return hr;
2850
2851 shader->load_local_constsF = shader->lconst_inf_or_nan;
2852
2853 return WINED3D_OK;
2854 }
2855
2856 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2857 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2858 {
2859 HRESULT hr;
2860
2861 if (FAILED(hr = shader_init(shader, device, desc, 0,
2862 WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
2863 return hr;
2864
2865 shader->load_local_constsF = shader->lconst_inf_or_nan;
2866
2867 return WINED3D_OK;
2868 }
2869
2870 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2871 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2872 {
2873 const struct wined3d_gl_info *gl_info = context->gl_info;
2874 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2875 const struct wined3d_texture *texture;
2876 UINT i;
2877
2878 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2879 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
2880 {
2881 static unsigned int warned = 0;
2882
2883 args->srgb_correction = 1;
2884 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2885 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2886 "support, expect rendering artifacts.\n");
2887 }
2888
2889 if (shader->reg_maps.shader_version.major == 1
2890 && shader->reg_maps.shader_version.minor <= 3)
2891 {
2892 for (i = 0; i < shader->limits->sampler; ++i)
2893 {
2894 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2895
2896 if (flags & WINED3D_TTFF_PROJECTED)
2897 {
2898 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2899
2900 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2901 {
2902 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2903 unsigned int j;
2904 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2905 DWORD max_valid = WINED3D_TTFF_COUNT4;
2906
2907 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2908 {
2909 struct wined3d_vertex_declaration_element *element =
2910 &state->vertex_declaration->elements[j];
2911
2912 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2913 && element->usage_idx == index)
2914 {
2915 max_valid = element->format->component_count;
2916 break;
2917 }
2918 }
2919 if (!tex_transform || tex_transform > max_valid)
2920 {
2921 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2922 tex_transform, max_valid);
2923 tex_transform = max_valid;
2924 }
2925 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2926 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2927 && tex_transform > WINED3D_TTFF_COUNT2)
2928 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2929 && tex_transform > WINED3D_TTFF_COUNT3))
2930 tex_transform |= WINED3D_PSARGS_PROJECTED;
2931 else
2932 {
2933 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2934 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2935 i, tex_transform, resource_type);
2936 }
2937 }
2938 else
2939 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2940
2941 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2942 }
2943 }
2944 }
2945 if (shader->reg_maps.shader_version.major == 1
2946 && shader->reg_maps.shader_version.minor <= 4)
2947 {
2948 for (i = 0; i < shader->limits->sampler; ++i)
2949 {
2950 const struct wined3d_texture *texture = state->textures[i];
2951
2952 if (!shader->reg_maps.resource_info[i].type)
2953 continue;
2954
2955 /* Treat unbound textures as 2D. The dummy texture will provide
2956 * the proper sample value. The tex_types bitmap defaults to
2957 * 2D because of the memset. */
2958 if (!texture)
2959 continue;
2960
2961 switch (texture->target)
2962 {
2963 /* RECT textures are distinguished from 2D textures via np2_fixup */
2964 case GL_TEXTURE_RECTANGLE_ARB:
2965 case GL_TEXTURE_2D:
2966 break;
2967
2968 case GL_TEXTURE_3D:
2969 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2970 break;
2971
2972 case GL_TEXTURE_CUBE_MAP_ARB:
2973 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2974 break;
2975 }
2976 }
2977 }
2978
2979 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2980 {
2981 if (!shader->reg_maps.resource_info[i].type)
2982 continue;
2983
2984 texture = state->textures[i];
2985 if (!texture)
2986 {
2987 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2988 continue;
2989 }
2990 if (can_use_texture_swizzle(gl_info, texture->resource.format))
2991 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2992 else
2993 args->color_fixup[i] = texture->resource.format->color_fixup;
2994
2995 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
2996 args->shadow |= 1u << i;
2997
2998 /* Flag samplers that need NP2 texcoord fixup. */
2999 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3000 args->np2_fixup |= (1u << i);
3001 }
3002
3003 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3004 if (shader->reg_maps.shader_version.major >= 4)
3005 args->shadow = 0;
3006
3007 if (shader->reg_maps.shader_version.major >= 3)
3008 {
3009 if (position_transformed)
3010 args->vp_mode = pretransformed;
3011 else if (use_vs(state))
3012 args->vp_mode = vertexshader;
3013 else
3014 args->vp_mode = fixedfunction;
3015 args->fog = WINED3D_FFP_PS_FOG_OFF;
3016 }
3017 else
3018 {
3019 args->vp_mode = vertexshader;
3020 if (state->render_states[WINED3D_RS_FOGENABLE])
3021 {
3022 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3023 {
3024 case WINED3D_FOG_NONE:
3025 if (position_transformed || use_vs(state))
3026 {
3027 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3028 break;
3029 }
3030
3031 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3032 {
3033 case WINED3D_FOG_NONE: /* Fall through. */
3034 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3035 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3036 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3037 }
3038 break;
3039
3040 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3041 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3042 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3043 }
3044 }
3045 else
3046 {
3047 args->fog = WINED3D_FFP_PS_FOG_OFF;
3048 }
3049 }
3050
3051 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3052 {
3053 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3054
3055 args->texcoords_initialized = 0;
3056 for (i = 0; i < MAX_TEXTURES; ++i)
3057 {
3058 if (vs)
3059 {
3060 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3061 args->texcoords_initialized |= 1u << i;
3062 }
3063 else
3064 {
3065 const struct wined3d_stream_info *si = &context->stream_info;
3066 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3067
3068 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3069 & WINED3D_FFP_TCI_MASK
3070 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3071 args->texcoords_initialized |= 1u << i;
3072 }
3073 }
3074 }
3075 else
3076 {
3077 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3078 }
3079
3080 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3081 && state->gl_primitive_type == GL_POINTS;
3082
3083 if (d3d_info->emulated_flatshading)
3084 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3085 }
3086
3087 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3088 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3089 {
3090 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3091 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3092 HRESULT hr;
3093
3094 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3095 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3096 return hr;
3097
3098 for (i = 0; i < MAX_REG_INPUT; ++i)
3099 {
3100 if (shader->u.ps.input_reg_used[i])
3101 {
3102 ++num_regs_used;
3103 highest_reg_used = i;
3104 }
3105 }
3106
3107 /* Don't do any register mapping magic if it is not needed, or if we can't
3108 * achieve anything anyway */
3109 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3110 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
3111 {
3112 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3113 {
3114 /* This happens with relative addressing. The input mapper function
3115 * warns about this if the higher registers are declared too, so
3116 * don't write a FIXME here */
3117 WARN("More varying registers used than supported\n");
3118 }
3119
3120 for (i = 0; i < MAX_REG_INPUT; ++i)
3121 {
3122 shader->u.ps.input_reg_map[i] = i;
3123 }
3124
3125 shader->u.ps.declared_in_count = highest_reg_used + 1;
3126 }
3127 else
3128 {
3129 shader->u.ps.declared_in_count = 0;
3130 for (i = 0; i < MAX_REG_INPUT; ++i)
3131 {
3132 if (shader->u.ps.input_reg_used[i])
3133 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3134 else shader->u.ps.input_reg_map[i] = ~0U;
3135 }
3136 }
3137
3138 shader->load_local_constsF = shader->lconst_inf_or_nan;
3139
3140 return WINED3D_OK;
3141 }
3142
3143 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3144 {
3145 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3146 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3147 unsigned int i;
3148
3149 if (reg_maps->shader_version.major != 1) return;
3150
3151 for (i = 0; i < shader->limits->sampler; ++i)
3152 {
3153 /* We don't sample from this sampler. */
3154 if (!resource_info[i].type)
3155 continue;
3156
3157 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3158 {
3159 case WINED3D_SHADER_TEX_2D:
3160 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3161 break;
3162
3163 case WINED3D_SHADER_TEX_3D:
3164 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3165 break;
3166
3167 case WINED3D_SHADER_TEX_CUBE:
3168 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3169 break;
3170 }
3171 }
3172 }
3173
3174 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3175 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3176 {
3177 struct wined3d_shader *object;
3178 HRESULT hr;
3179
3180 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3181 device, desc, parent, parent_ops, shader);
3182
3183 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3184 return E_OUTOFMEMORY;
3185
3186 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3187 {
3188 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3189 HeapFree(GetProcessHeap(), 0, object);
3190 return hr;
3191 }
3192
3193 TRACE("Created domain shader %p.\n", object);
3194 *shader = object;
3195
3196 return WINED3D_OK;
3197 }
3198
3199 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3200 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3201 {
3202 struct wined3d_shader *object;
3203 HRESULT hr;
3204
3205 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3206 device, desc, parent, parent_ops, shader);
3207
3208 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3209 if (!object)
3210 return E_OUTOFMEMORY;
3211
3212 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3213 {
3214 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3215 HeapFree(GetProcessHeap(), 0, object);
3216 return hr;
3217 }
3218
3219 TRACE("Created geometry shader %p.\n", object);
3220 *shader = object;
3221
3222 return WINED3D_OK;
3223 }
3224
3225 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3226 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3227 {
3228 struct wined3d_shader *object;
3229 HRESULT hr;
3230
3231 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3232 device, desc, parent, parent_ops, shader);
3233
3234 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3235 return E_OUTOFMEMORY;
3236
3237 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3238 {
3239 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3240 HeapFree(GetProcessHeap(), 0, object);
3241 return hr;
3242 }
3243
3244 TRACE("Created hull shader %p.\n", object);
3245 *shader = object;
3246
3247 return WINED3D_OK;
3248 }
3249
3250 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3251 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3252 {
3253 struct wined3d_shader *object;
3254 HRESULT hr;
3255
3256 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3257 device, desc, parent, parent_ops, shader);
3258
3259 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3260 if (!object)
3261 return E_OUTOFMEMORY;
3262
3263 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3264 {
3265 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3266 HeapFree(GetProcessHeap(), 0, object);
3267 return hr;
3268 }
3269
3270 TRACE("Created pixel shader %p.\n", object);
3271 *shader = object;
3272
3273 return WINED3D_OK;
3274 }
3275
3276 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3277 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3278 {
3279 struct wined3d_shader *object;
3280 HRESULT hr;
3281
3282 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3283 device, desc, parent, parent_ops, shader);
3284
3285 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3286 if (!object)
3287 return E_OUTOFMEMORY;
3288
3289 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3290 {
3291 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3292 HeapFree(GetProcessHeap(), 0, object);
3293 return hr;
3294 }
3295
3296 TRACE("Created vertex shader %p.\n", object);
3297 *shader = object;
3298
3299 return WINED3D_OK;
3300 }
Windows API implementation