projects / reactos.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Synchronize with trunk r58457.
[reactos.git] / dll / directx / wine / wined3d / shader.c
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 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 static const char * const shader_opcode_names[] =
37 {
38 /* WINED3DSIH_ABS */ "abs",
39 /* WINED3DSIH_ADD */ "add",
40 /* WINED3DSIH_AND */ "and",
41 /* WINED3DSIH_BEM */ "bem",
42 /* WINED3DSIH_BREAK */ "break",
43 /* WINED3DSIH_BREAKC */ "breakc",
44 /* WINED3DSIH_BREAKP */ "breakp",
45 /* WINED3DSIH_CALL */ "call",
46 /* WINED3DSIH_CALLNZ */ "callnz",
47 /* WINED3DSIH_CMP */ "cmp",
48 /* WINED3DSIH_CND */ "cnd",
49 /* WINED3DSIH_CRS */ "crs",
50 /* WINED3DSIH_CUT */ "cut",
51 /* WINED3DSIH_DCL */ "dcl",
52 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
53 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
54 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
55 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
56 /* WINED3DSIH_DEF */ "def",
57 /* WINED3DSIH_DEFB */ "defb",
58 /* WINED3DSIH_DEFI */ "defi",
59 /* WINED3DSIH_DIV */ "div",
60 /* WINED3DSIH_DP2ADD */ "dp2add",
61 /* WINED3DSIH_DP3 */ "dp3",
62 /* WINED3DSIH_DP4 */ "dp4",
63 /* WINED3DSIH_DST */ "dst",
64 /* WINED3DSIH_DSX */ "dsx",
65 /* WINED3DSIH_DSY */ "dsy",
66 /* WINED3DSIH_ELSE */ "else",
67 /* WINED3DSIH_EMIT */ "emit",
68 /* WINED3DSIH_ENDIF */ "endif",
69 /* WINED3DSIH_ENDLOOP */ "endloop",
70 /* WINED3DSIH_ENDREP */ "endrep",
71 /* WINED3DSIH_EQ */ "eq",
72 /* WINED3DSIH_EXP */ "exp",
73 /* WINED3DSIH_EXPP */ "expp",
74 /* WINED3DSIH_FRC */ "frc",
75 /* WINED3DSIH_FTOI */ "ftoi",
76 /* WINED3DSIH_GE */ "ge",
77 /* WINED3DSIH_IADD */ "iadd",
78 /* WINED3DSIH_IEQ */ "ieq",
79 /* WINED3DSIH_IF */ "if",
80 /* WINED3DSIH_IFC */ "ifc",
81 /* WINED3DSIH_IGE */ "ige",
82 /* WINED3DSIH_IMUL */ "imul",
83 /* WINED3DSIH_ITOF */ "itof",
84 /* WINED3DSIH_LABEL */ "label",
85 /* WINED3DSIH_LD */ "ld",
86 /* WINED3DSIH_LIT */ "lit",
87 /* WINED3DSIH_LOG */ "log",
88 /* WINED3DSIH_LOGP */ "logp",
89 /* WINED3DSIH_LOOP */ "loop",
90 /* WINED3DSIH_LRP */ "lrp",
91 /* WINED3DSIH_LT */ "lt",
92 /* WINED3DSIH_M3x2 */ "m3x2",
93 /* WINED3DSIH_M3x3 */ "m3x3",
94 /* WINED3DSIH_M3x4 */ "m3x4",
95 /* WINED3DSIH_M4x3 */ "m4x3",
96 /* WINED3DSIH_M4x4 */ "m4x4",
97 /* WINED3DSIH_MAD */ "mad",
98 /* WINED3DSIH_MAX */ "max",
99 /* WINED3DSIH_MIN */ "min",
100 /* WINED3DSIH_MOV */ "mov",
101 /* WINED3DSIH_MOVA */ "mova",
102 /* WINED3DSIH_MOVC */ "movc",
103 /* WINED3DSIH_MUL */ "mul",
104 /* WINED3DSIH_NOP */ "nop",
105 /* WINED3DSIH_NRM */ "nrm",
106 /* WINED3DSIH_PHASE */ "phase",
107 /* WINED3DSIH_POW */ "pow",
108 /* WINED3DSIH_RCP */ "rcp",
109 /* WINED3DSIH_REP */ "rep",
110 /* WINED3DSIH_RET */ "ret",
111 /* WINED3DSIH_ROUND_NI */ "round_ni",
112 /* WINED3DSIH_RSQ */ "rsq",
113 /* WINED3DSIH_SAMPLE */ "sample",
114 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
115 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
116 /* WINED3DSIH_SETP */ "setp",
117 /* WINED3DSIH_SGE */ "sge",
118 /* WINED3DSIH_SGN */ "sgn",
119 /* WINED3DSIH_SINCOS */ "sincos",
120 /* WINED3DSIH_SLT */ "slt",
121 /* WINED3DSIH_SQRT */ "sqrt",
122 /* WINED3DSIH_SUB */ "sub",
123 /* WINED3DSIH_TEX */ "texld",
124 /* WINED3DSIH_TEXBEM */ "texbem",
125 /* WINED3DSIH_TEXBEML */ "texbeml",
126 /* WINED3DSIH_TEXCOORD */ "texcrd",
127 /* WINED3DSIH_TEXDEPTH */ "texdepth",
128 /* WINED3DSIH_TEXDP3 */ "texdp3",
129 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
130 /* WINED3DSIH_TEXKILL */ "texkill",
131 /* WINED3DSIH_TEXLDD */ "texldd",
132 /* WINED3DSIH_TEXLDL */ "texldl",
133 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
134 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
135 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
136 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
137 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
138 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
139 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
140 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
141 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
142 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
143 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
144 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
145 /* WINED3DSIH_UDIV */ "udiv",
146 /* WINED3DSIH_USHR */ "ushr",
147 /* WINED3DSIH_UTOF */ "utof",
148 /* WINED3DSIH_XOR */ "xor",
149 };
150
151 static const char * const semantic_names[] =
152 {
153 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
154 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
155 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
156 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
157 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
158 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
159 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
160 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
161 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
162 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
163 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
164 /* WINED3D_DECL_USAGE_FOG */ "FOG",
165 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
166 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
167 };
168
169 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
170 {
171 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
172 {
173 FIXME("Unrecognized usage %#x.\n", usage);
174 return "UNRECOGNIZED";
175 }
176
177 return semantic_names[usage];
178 }
179
180 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
181 {
182 unsigned int i;
183
184 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
185 {
186 if (!strcmp(name, semantic_names[i])) return i;
187 }
188
189 return ~0U;
190 }
191
192 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
193 {
194 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
195 }
196
197 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
198 const struct wined3d_shader_semantic *s)
199 {
200 e->semantic_name = shader_semantic_name_from_usage(s->usage);
201 e->semantic_idx = s->usage_idx;
202 e->sysval_semantic = 0;
203 e->component_type = 0;
204 e->register_idx = s->reg.reg.idx[0].offset;
205 e->mask = s->reg.write_mask;
206 }
207
208 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
209 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
210 {
211 e->semantic_name = shader_semantic_name_from_usage(usage);
212 e->semantic_idx = usage_idx;
213 e->sysval_semantic = 0;
214 e->component_type = 0;
215 e->register_idx = reg_idx;
216 e->mask = write_mask;
217 }
218
219 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
220 {
221 switch (version_token >> 16)
222 {
223 case WINED3D_SM1_VS:
224 case WINED3D_SM1_PS:
225 return &sm1_shader_frontend;
226
227 case WINED3D_SM4_PS:
228 case WINED3D_SM4_VS:
229 case WINED3D_SM4_GS:
230 return &sm4_shader_frontend;
231
232 default:
233 FIXME("Unrecognised version token %#x\n", version_token);
234 return NULL;
235 }
236 }
237
238 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
239 {
240 buffer->buffer[0] = '\0';
241 buffer->bsize = 0;
242 buffer->lineNo = 0;
243 buffer->newline = TRUE;
244 }
245
246 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
247 {
248 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
249 if (!buffer->buffer)
250 {
251 ERR("Failed to allocate shader buffer memory.\n");
252 return FALSE;
253 }
254
255 shader_buffer_clear(buffer);
256 return TRUE;
257 }
258
259 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
260 {
261 HeapFree(GetProcessHeap(), 0, buffer->buffer);
262 }
263
264 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
265 {
266 char *base = buffer->buffer + buffer->bsize;
267 int rc;
268
269 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
270
271 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
272 {
273 ERR("The buffer allocated for the shader program string "
274 "is too small at %d bytes.\n", SHADER_PGMSIZE);
275 buffer->bsize = SHADER_PGMSIZE - 1;
276 return -1;
277 }
278
279 if (buffer->newline)
280 {
281 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
282 buffer->newline = FALSE;
283 }
284 else
285 {
286 TRACE("%s", base);
287 }
288
289 buffer->bsize += rc;
290 if (buffer->buffer[buffer->bsize-1] == '\n')
291 {
292 ++buffer->lineNo;
293 buffer->newline = TRUE;
294 }
295
296 return 0;
297 }
298
299 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
300 {
301 va_list args;
302 int ret;
303
304 va_start(args, format);
305 ret = shader_vaddline(buffer, format, args);
306 va_end(args);
307
308 return ret;
309 }
310
311 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
312 void *parent, const struct wined3d_parent_ops *parent_ops)
313 {
314 shader->ref = 1;
315 shader->device = device;
316 shader->parent = parent;
317 shader->parent_ops = parent_ops;
318 list_init(&shader->linked_programs);
319 list_add_head(&device->shaders, &shader->shader_list_entry);
320 }
321
322 /* Convert floating point offset relative to a register file to an absolute
323 * offset for float constants. */
324 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
325 {
326 switch (register_type)
327 {
328 case WINED3DSPR_CONST: return register_idx;
329 case WINED3DSPR_CONST2: return 2048 + register_idx;
330 case WINED3DSPR_CONST3: return 4096 + register_idx;
331 case WINED3DSPR_CONST4: return 6144 + register_idx;
332 default:
333 FIXME("Unsupported register type: %u.\n", register_type);
334 return register_idx;
335 }
336 }
337
338 static void shader_delete_constant_list(struct list *clist)
339 {
340 struct wined3d_shader_lconst *constant;
341 struct list *ptr;
342
343 ptr = list_head(clist);
344 while (ptr)
345 {
346 constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
347 ptr = list_next(clist, ptr);
348 HeapFree(GetProcessHeap(), 0, constant);
349 }
350 list_init(clist);
351 }
352
353 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
354 {
355 DWORD idx, shift;
356 idx = bit >> 5;
357 shift = bit & 0x1f;
358 bitmap[idx] |= (1 << shift);
359 }
360
361 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
362 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
363 {
364 switch (reg->type)
365 {
366 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
367 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
368 reg_maps->texcoord |= 1 << reg->idx[0].offset;
369 else
370 reg_maps->address |= 1 << reg->idx[0].offset;
371 break;
372
373 case WINED3DSPR_TEMP:
374 reg_maps->temporary |= 1 << reg->idx[0].offset;
375 break;
376
377 case WINED3DSPR_INPUT:
378 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
379 {
380 if (reg->idx[0].rel_addr)
381 {
382 /* If relative addressing is used, we must assume that all registers
383 * are used. Even if it is a construct like v3[aL], we can't assume
384 * that v0, v1 and v2 aren't read because aL can be negative */
385 unsigned int i;
386 for (i = 0; i < MAX_REG_INPUT; ++i)
387 {
388 shader->u.ps.input_reg_used[i] = TRUE;
389 }
390 }
391 else
392 {
393 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
394 }
395 }
396 else
397 reg_maps->input_registers |= 1 << reg->idx[0].offset;
398 break;
399
400 case WINED3DSPR_RASTOUT:
401 if (reg->idx[0].offset == 1)
402 reg_maps->fog = 1;
403 break;
404
405 case WINED3DSPR_MISCTYPE:
406 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
407 {
408 if (!reg->idx[0].offset)
409 reg_maps->vpos = 1;
410 else if (reg->idx[0].offset == 1)
411 reg_maps->usesfacing = 1;
412 }
413 break;
414
415 case WINED3DSPR_CONST:
416 if (reg->idx[0].rel_addr)
417 {
418 if (reg->idx[0].offset < reg_maps->min_rel_offset)
419 reg_maps->min_rel_offset = reg->idx[0].offset;
420 if (reg->idx[0].offset > reg_maps->max_rel_offset)
421 reg_maps->max_rel_offset = reg->idx[0].offset;
422 reg_maps->usesrelconstF = TRUE;
423 }
424 else
425 {
426 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
427 }
428 break;
429
430 case WINED3DSPR_CONSTINT:
431 reg_maps->integer_constants |= (1 << reg->idx[0].offset);
432 break;
433
434 case WINED3DSPR_CONSTBOOL:
435 reg_maps->boolean_constants |= (1 << reg->idx[0].offset);
436 break;
437
438 case WINED3DSPR_COLOROUT:
439 reg_maps->rt_mask |= (1 << reg->idx[0].offset);
440 break;
441
442 default:
443 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
444 reg->type, reg->idx[0].offset, reg->idx[1].offset);
445 break;
446 }
447 }
448
449 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
450 {
451 switch (instr)
452 {
453 case WINED3DSIH_M4x4:
454 case WINED3DSIH_M3x4:
455 return param == 1 ? 3 : 0;
456
457 case WINED3DSIH_M4x3:
458 case WINED3DSIH_M3x3:
459 return param == 1 ? 2 : 0;
460
461 case WINED3DSIH_M3x2:
462 return param == 1 ? 1 : 0;
463
464 default:
465 return 0;
466 }
467 }
468
469 /* Note that this does not count the loop register as an address register. */
470 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
471 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
472 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
473 {
474 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
475 void *fe_data = shader->frontend_data;
476 struct wined3d_shader_version shader_version;
477 const DWORD *ptr = byte_code;
478
479 memset(reg_maps, 0, sizeof(*reg_maps));
480 reg_maps->min_rel_offset = ~0U;
481
482 fe->shader_read_header(fe_data, &ptr, &shader_version);
483 reg_maps->shader_version = shader_version;
484
485 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
486 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
487 if (!reg_maps->constf)
488 {
489 ERR("Failed to allocate constant map memory.\n");
490 return E_OUTOFMEMORY;
491 }
492
493 while (!fe->shader_is_end(fe_data, &ptr))
494 {
495 struct wined3d_shader_instruction ins;
496
497 /* Fetch opcode. */
498 fe->shader_read_instruction(fe_data, &ptr, &ins);
499
500 /* Unhandled opcode, and its parameters. */
501 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
502 {
503 TRACE("Skipping unrecognized instruction.\n");
504 continue;
505 }
506
507 /* Handle declarations. */
508 if (ins.handler_idx == WINED3DSIH_DCL)
509 {
510 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
511
512 switch (semantic->reg.reg.type)
513 {
514 /* Mark input registers used. */
515 case WINED3DSPR_INPUT:
516 reg_maps->input_registers |= 1 << semantic->reg.reg.idx[0].offset;
517 shader_signature_from_semantic(&input_signature[semantic->reg.reg.idx[0].offset], semantic);
518 break;
519
520 /* Vertex shader: mark 3.0 output registers used, save token. */
521 case WINED3DSPR_OUTPUT:
522 reg_maps->output_registers |= 1 << semantic->reg.reg.idx[0].offset;
523 shader_signature_from_semantic(&output_signature[semantic->reg.reg.idx[0].offset], semantic);
524 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
525 reg_maps->fog = 1;
526 break;
527
528 /* Save sampler usage token. */
529 case WINED3DSPR_SAMPLER:
530 reg_maps->sampler_type[semantic->reg.reg.idx[0].offset] = semantic->sampler_type;
531 break;
532
533 default:
534 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
535 break;
536 }
537 }
538 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
539 {
540 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
541 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
542 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
543 else
544 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
545 }
546 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
547 {
548 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
549 shader->u.gs.input_type = ins.declaration.primitive_type;
550 else
551 FIXME("Invalid instruction %#x for shader type %#x.\n",
552 ins.handler_idx, shader_version.type);
553 }
554 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
555 {
556 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
557 shader->u.gs.output_type = ins.declaration.primitive_type;
558 else
559 FIXME("Invalid instruction %#x for shader type %#x.\n",
560 ins.handler_idx, shader_version.type);
561 }
562 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
563 {
564 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
565 shader->u.gs.vertices_out = ins.declaration.count;
566 else
567 FIXME("Invalid instruction %#x for shader type %#x.\n",
568 ins.handler_idx, shader_version.type);
569 }
570 else if (ins.handler_idx == WINED3DSIH_DEF)
571 {
572 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
573 if (!lconst) return E_OUTOFMEMORY;
574
575 lconst->idx = ins.dst[0].reg.idx[0].offset;
576 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
577
578 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
579 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
580 {
581 float *value = (float *)lconst->value;
582 if (value[0] < -1.0f) value[0] = -1.0f;
583 else if (value[0] > 1.0f) value[0] = 1.0f;
584 if (value[1] < -1.0f) value[1] = -1.0f;
585 else if (value[1] > 1.0f) value[1] = 1.0f;
586 if (value[2] < -1.0f) value[2] = -1.0f;
587 else if (value[2] > 1.0f) value[2] = 1.0f;
588 if (value[3] < -1.0f) value[3] = -1.0f;
589 else if (value[3] > 1.0f) value[3] = 1.0f;
590 }
591
592 list_add_head(&shader->constantsF, &lconst->entry);
593 }
594 else if (ins.handler_idx == WINED3DSIH_DEFI)
595 {
596 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
597 if (!lconst) return E_OUTOFMEMORY;
598
599 lconst->idx = ins.dst[0].reg.idx[0].offset;
600 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
601
602 list_add_head(&shader->constantsI, &lconst->entry);
603 reg_maps->local_int_consts |= (1 << lconst->idx);
604 }
605 else if (ins.handler_idx == WINED3DSIH_DEFB)
606 {
607 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
608 if (!lconst) return E_OUTOFMEMORY;
609
610 lconst->idx = ins.dst[0].reg.idx[0].offset;
611 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
612
613 list_add_head(&shader->constantsB, &lconst->entry);
614 reg_maps->local_bool_consts |= (1 << lconst->idx);
615 }
616 /* For subroutine prototypes. */
617 else if (ins.handler_idx == WINED3DSIH_LABEL)
618 {
619 reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
620 }
621 /* Set texture, address, temporary registers. */
622 else
623 {
624 BOOL color0_mov = FALSE;
625 unsigned int i;
626
627 /* This will loop over all the registers and try to
628 * make a bitmask of the ones we're interested in.
629 *
630 * Relative addressing tokens are ignored, but that's
631 * okay, since we'll catch any address registers when
632 * they are initialized (required by spec). */
633 for (i = 0; i < ins.dst_count; ++i)
634 {
635 shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg, shader_version.type);
636
637 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
638 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
639 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
640 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
641 {
642 UINT idx = ins.dst[i].reg.idx[0].offset;
643
644 switch (ins.dst[i].reg.type)
645 {
646 case WINED3DSPR_RASTOUT:
647 switch (idx)
648 {
649 case 0: /* oPos */
650 reg_maps->output_registers |= 1 << 10;
651 shader_signature_from_usage(&output_signature[10],
652 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
653 break;
654
655 case 1: /* oFog */
656 reg_maps->output_registers |= 1 << 11;
657 shader_signature_from_usage(&output_signature[11],
658 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
659 break;
660
661 case 2: /* oPts */
662 reg_maps->output_registers |= 1 << 11;
663 shader_signature_from_usage(&output_signature[11],
664 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
665 break;
666 }
667 break;
668
669 case WINED3DSPR_ATTROUT:
670 if (idx < 2)
671 {
672 idx += 8;
673 if (reg_maps->output_registers & (1 << idx))
674 {
675 output_signature[idx].mask |= ins.dst[i].write_mask;
676 }
677 else
678 {
679 reg_maps->output_registers |= 1 << idx;
680 shader_signature_from_usage(&output_signature[idx],
681 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
682 }
683 }
684 break;
685
686 case WINED3DSPR_TEXCRDOUT:
687
688 reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
689 if (reg_maps->output_registers & (1 << idx))
690 {
691 output_signature[idx].mask |= ins.dst[i].write_mask;
692 }
693 else
694 {
695 reg_maps->output_registers |= 1 << idx;
696 shader_signature_from_usage(&output_signature[idx],
697 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
698 }
699 break;
700
701 default:
702 break;
703 }
704 }
705
706 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
707 {
708 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
709 {
710 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
711 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
712 * the mov and perform the sRGB write correction from the source register.
713 *
714 * However, if the mov is only partial, we can't do this, and if the write
715 * comes from an instruction other than MOV it is hard to do as well. If
716 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
717 shader->u.ps.color0_mov = FALSE;
718 if (ins.handler_idx == WINED3DSIH_MOV
719 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
720 {
721 /* Used later when the source register is read. */
722 color0_mov = TRUE;
723 }
724 }
725 /* Also drop the MOV marker if the source register is overwritten prior to the shader
726 * end
727 */
728 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
729 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
730 {
731 shader->u.ps.color0_mov = FALSE;
732 }
733 }
734
735 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
736 if (shader_version.major == 1
737 && (ins.handler_idx == WINED3DSIH_TEX
738 || ins.handler_idx == WINED3DSIH_TEXBEM
739 || ins.handler_idx == WINED3DSIH_TEXBEML
740 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
741 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
742 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
743 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
744 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
745 || ins.handler_idx == WINED3DSIH_TEXREG2AR
746 || ins.handler_idx == WINED3DSIH_TEXREG2GB
747 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
748 {
749 /* Fake sampler usage, only set reserved bit and type. */
750 DWORD sampler_code = ins.dst[i].reg.idx[0].offset;
751
752 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
753 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
754
755 /* texbem is only valid with < 1.4 pixel shaders */
756 if (ins.handler_idx == WINED3DSIH_TEXBEM
757 || ins.handler_idx == WINED3DSIH_TEXBEML)
758 {
759 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
760 if (ins.handler_idx == WINED3DSIH_TEXBEML)
761 {
762 reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx[0].offset;
763 }
764 }
765 }
766 else if (ins.handler_idx == WINED3DSIH_BEM)
767 {
768 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
769 }
770 }
771
772 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
773 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
774 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
775 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
776 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
777 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
778 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
779 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
780 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
781 else if (ins.handler_idx == WINED3DSIH_LOOP
782 || ins.handler_idx == WINED3DSIH_REP)
783 {
784 ++cur_loop_depth;
785 if (cur_loop_depth > max_loop_depth)
786 max_loop_depth = cur_loop_depth;
787 }
788 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
789 || ins.handler_idx == WINED3DSIH_ENDREP)
790 --cur_loop_depth;
791
792 if (ins.predicate)
793 shader_record_register_usage(shader, reg_maps, &ins.predicate->reg, shader_version.type);
794
795 for (i = 0; i < ins.src_count; ++i)
796 {
797 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
798 struct wined3d_shader_register reg = ins.src[i].reg;
799
800 shader_record_register_usage(shader, reg_maps, &ins.src[i].reg, shader_version.type);
801 while (count)
802 {
803 ++reg.idx[0].offset;
804 shader_record_register_usage(shader, reg_maps, &reg, shader_version.type);
805 --count;
806 }
807
808 if (color0_mov)
809 {
810 if (ins.src[i].reg.type == WINED3DSPR_TEMP
811 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
812 {
813 shader->u.ps.color0_mov = TRUE;
814 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
815 }
816 }
817 }
818 }
819 }
820 reg_maps->loop_depth = max_loop_depth;
821
822 /* PS before 2.0 don't have explicit color outputs. Instead the value of
823 * R0 is written to the render target. */
824 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
825 reg_maps->rt_mask |= (1 << 0);
826
827 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
828
829 return WINED3D_OK;
830 }
831
832 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
833 {
834 DWORD map = 1 << max;
835 map |= map - 1;
836 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
837
838 return wined3d_log2i(map);
839 }
840
841 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
842 const struct wined3d_shader_version *shader_version)
843 {
844 TRACE("dcl");
845
846 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
847 {
848 switch (semantic->sampler_type)
849 {
850 case WINED3DSTT_2D: TRACE("_2d"); break;
851 case WINED3DSTT_CUBE: TRACE("_cube"); break;
852 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
853 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
854 }
855 }
856 else
857 {
858 /* Pixel shaders 3.0 don't have usage semantics. */
859 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
860 else TRACE("_");
861
862 switch (semantic->usage)
863 {
864 case WINED3D_DECL_USAGE_POSITION:
865 TRACE("position%u", semantic->usage_idx);
866 break;
867
868 case WINED3D_DECL_USAGE_BLEND_INDICES:
869 TRACE("blend");
870 break;
871
872 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
873 TRACE("weight");
874 break;
875
876 case WINED3D_DECL_USAGE_NORMAL:
877 TRACE("normal%u", semantic->usage_idx);
878 break;
879
880 case WINED3D_DECL_USAGE_PSIZE:
881 TRACE("psize");
882 break;
883
884 case WINED3D_DECL_USAGE_COLOR:
885 if (!semantic->usage_idx) TRACE("color");
886 else TRACE("specular%u", (semantic->usage_idx - 1));
887 break;
888
889 case WINED3D_DECL_USAGE_TEXCOORD:
890 TRACE("texture%u", semantic->usage_idx);
891 break;
892
893 case WINED3D_DECL_USAGE_TANGENT:
894 TRACE("tangent");
895 break;
896
897 case WINED3D_DECL_USAGE_BINORMAL:
898 TRACE("binormal");
899 break;
900
901 case WINED3D_DECL_USAGE_TESS_FACTOR:
902 TRACE("tessfactor");
903 break;
904
905 case WINED3D_DECL_USAGE_POSITIONT:
906 TRACE("positionT%u", semantic->usage_idx);
907 break;
908
909 case WINED3D_DECL_USAGE_FOG:
910 TRACE("fog");
911 break;
912
913 case WINED3D_DECL_USAGE_DEPTH:
914 TRACE("depth");
915 break;
916
917 case WINED3D_DECL_USAGE_SAMPLE:
918 TRACE("sample");
919 break;
920
921 default:
922 FIXME("unknown_semantics(0x%08x)", semantic->usage);
923 }
924 }
925 }
926
927 static void shader_dump_register(const struct wined3d_shader_register *reg,
928 const struct wined3d_shader_version *shader_version)
929 {
930 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
931 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
932 UINT offset = reg->idx[0].offset;
933
934 switch (reg->type)
935 {
936 case WINED3DSPR_TEMP:
937 TRACE("r");
938 break;
939
940 case WINED3DSPR_INPUT:
941 TRACE("v");
942 break;
943
944 case WINED3DSPR_CONST:
945 case WINED3DSPR_CONST2:
946 case WINED3DSPR_CONST3:
947 case WINED3DSPR_CONST4:
948 TRACE("c");
949 offset = shader_get_float_offset(reg->type, offset);
950 break;
951
952 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
953 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
954 break;
955
956 case WINED3DSPR_RASTOUT:
957 TRACE("%s", rastout_reg_names[offset]);
958 break;
959
960 case WINED3DSPR_COLOROUT:
961 TRACE("oC");
962 break;
963
964 case WINED3DSPR_DEPTHOUT:
965 TRACE("oDepth");
966 break;
967
968 case WINED3DSPR_ATTROUT:
969 TRACE("oD");
970 break;
971
972 case WINED3DSPR_TEXCRDOUT:
973 /* Vertex shaders >= 3.0 use general purpose output registers
974 * (WINED3DSPR_OUTPUT), which can include an address token. */
975 if (shader_version->major >= 3) TRACE("o");
976 else TRACE("oT");
977 break;
978
979 case WINED3DSPR_CONSTINT:
980 TRACE("i");
981 break;
982
983 case WINED3DSPR_CONSTBOOL:
984 TRACE("b");
985 break;
986
987 case WINED3DSPR_LABEL:
988 TRACE("l");
989 break;
990
991 case WINED3DSPR_LOOP:
992 TRACE("aL");
993 break;
994
995 case WINED3DSPR_SAMPLER:
996 TRACE("s");
997 break;
998
999 case WINED3DSPR_MISCTYPE:
1000 if (offset > 1)
1001 FIXME("Unhandled misctype register %u.\n", offset);
1002 else
1003 TRACE("%s", misctype_reg_names[offset]);
1004 break;
1005
1006 case WINED3DSPR_PREDICATE:
1007 TRACE("p");
1008 break;
1009
1010 case WINED3DSPR_IMMCONST:
1011 TRACE("l");
1012 break;
1013
1014 case WINED3DSPR_CONSTBUFFER:
1015 TRACE("cb");
1016 break;
1017
1018 case WINED3DSPR_PRIMID:
1019 TRACE("primID");
1020 break;
1021
1022 case WINED3DSPR_NULL:
1023 TRACE("null");
1024 break;
1025
1026 case WINED3DSPR_RESOURCE:
1027 TRACE("t");
1028 break;
1029
1030 default:
1031 TRACE("unhandled_rtype(%#x)", reg->type);
1032 break;
1033 }
1034
1035 if (reg->type == WINED3DSPR_IMMCONST)
1036 {
1037 TRACE("(");
1038 switch (reg->immconst_type)
1039 {
1040 case WINED3D_IMMCONST_SCALAR:
1041 switch (reg->data_type)
1042 {
1043 case WINED3D_DATA_FLOAT:
1044 TRACE("%.8e", *(const float *)reg->immconst_data);
1045 break;
1046 case WINED3D_DATA_INT:
1047 TRACE("%d", reg->immconst_data[0]);
1048 break;
1049 case WINED3D_DATA_RESOURCE:
1050 case WINED3D_DATA_SAMPLER:
1051 case WINED3D_DATA_UINT:
1052 TRACE("%u", reg->immconst_data[0]);
1053 break;
1054 default:
1055 TRACE("<unhandled data type %#x>", reg->data_type);
1056 break;
1057 }
1058 break;
1059
1060 case WINED3D_IMMCONST_VEC4:
1061 switch (reg->data_type)
1062 {
1063 case WINED3D_DATA_FLOAT:
1064 TRACE("%.8e, %.8e, %.8e, %.8e",
1065 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1066 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1067 break;
1068 case WINED3D_DATA_INT:
1069 TRACE("%d, %d, %d, %d",
1070 reg->immconst_data[0], reg->immconst_data[1],
1071 reg->immconst_data[2], reg->immconst_data[3]);
1072 break;
1073 case WINED3D_DATA_RESOURCE:
1074 case WINED3D_DATA_SAMPLER:
1075 case WINED3D_DATA_UINT:
1076 TRACE("%u, %u, %u, %u",
1077 reg->immconst_data[0], reg->immconst_data[1],
1078 reg->immconst_data[2], reg->immconst_data[3]);
1079 break;
1080 default:
1081 TRACE("<unhandled data type %#x>", reg->data_type);
1082 break;
1083 }
1084 break;
1085
1086 default:
1087 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1088 break;
1089 }
1090 TRACE(")");
1091 }
1092 else if (reg->type != WINED3DSPR_RASTOUT
1093 && reg->type != WINED3DSPR_MISCTYPE
1094 && reg->type != WINED3DSPR_NULL)
1095 {
1096 if (offset != ~0U)
1097 {
1098 TRACE("[");
1099 if (reg->idx[0].rel_addr)
1100 {
1101 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1102 TRACE(" + ");
1103 }
1104 TRACE("%u]", offset);
1105
1106 if (reg->idx[1].offset != ~0U)
1107 {
1108 TRACE("[");
1109 if (reg->idx[1].rel_addr)
1110 {
1111 shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1112 TRACE(" + ");
1113 }
1114 TRACE("%u]", reg->idx[1].offset);
1115 }
1116 }
1117 }
1118 }
1119
1120 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1121 const struct wined3d_shader_version *shader_version)
1122 {
1123 DWORD write_mask = param->write_mask;
1124
1125 shader_dump_register(&param->reg, shader_version);
1126
1127 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1128 {
1129 static const char *write_mask_chars = "xyzw";
1130
1131 TRACE(".");
1132 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1133 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1134 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1135 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1136 }
1137 }
1138
1139 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1140 const struct wined3d_shader_version *shader_version)
1141 {
1142 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1143 DWORD swizzle = param->swizzle;
1144
1145 if (src_modifier == WINED3DSPSM_NEG
1146 || src_modifier == WINED3DSPSM_BIASNEG
1147 || src_modifier == WINED3DSPSM_SIGNNEG
1148 || src_modifier == WINED3DSPSM_X2NEG
1149 || src_modifier == WINED3DSPSM_ABSNEG)
1150 TRACE("-");
1151 else if (src_modifier == WINED3DSPSM_COMP)
1152 TRACE("1-");
1153 else if (src_modifier == WINED3DSPSM_NOT)
1154 TRACE("!");
1155
1156 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1157 TRACE("abs(");
1158
1159 shader_dump_register(&param->reg, shader_version);
1160
1161 if (src_modifier)
1162 {
1163 switch (src_modifier)
1164 {
1165 case WINED3DSPSM_NONE: break;
1166 case WINED3DSPSM_NEG: break;
1167 case WINED3DSPSM_NOT: break;
1168 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1169 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1170 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1171 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1172 case WINED3DSPSM_COMP: break;
1173 case WINED3DSPSM_X2: TRACE("_x2"); break;
1174 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1175 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1176 case WINED3DSPSM_DW: TRACE("_dw"); break;
1177 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1178 case WINED3DSPSM_ABS: TRACE(")"); break;
1179 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1180 }
1181 }
1182
1183 if (swizzle != WINED3DSP_NOSWIZZLE)
1184 {
1185 static const char *swizzle_chars = "xyzw";
1186 DWORD swizzle_x = swizzle & 0x03;
1187 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1188 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1189 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1190
1191 if (swizzle_x == swizzle_y
1192 && swizzle_x == swizzle_z
1193 && swizzle_x == swizzle_w)
1194 {
1195 TRACE(".%c", swizzle_chars[swizzle_x]);
1196 }
1197 else
1198 {
1199 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1200 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1201 }
1202 }
1203 }
1204
1205 /* Shared code in order to generate the bulk of the shader string.
1206 * NOTE: A description of how to parse tokens can be found on MSDN. */
1207 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1208 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1209 {
1210 struct wined3d_device *device = shader->device;
1211 const struct wined3d_shader_frontend *fe = shader->frontend;
1212 void *fe_data = shader->frontend_data;
1213 struct wined3d_shader_version shader_version;
1214 struct wined3d_shader_loop_state loop_state;
1215 struct wined3d_shader_instruction ins;
1216 struct wined3d_shader_tex_mx tex_mx;
1217 struct wined3d_shader_context ctx;
1218 const DWORD *ptr = byte_code;
1219
1220 /* Initialize current parsing state. */
1221 tex_mx.current_row = 0;
1222 loop_state.current_depth = 0;
1223 loop_state.current_reg = 0;
1224
1225 ctx.shader = shader;
1226 ctx.gl_info = &device->adapter->gl_info;
1227 ctx.reg_maps = reg_maps;
1228 ctx.buffer = buffer;
1229 ctx.tex_mx = &tex_mx;
1230 ctx.loop_state = &loop_state;
1231 ctx.backend_data = backend_ctx;
1232 ins.ctx = &ctx;
1233
1234 fe->shader_read_header(fe_data, &ptr, &shader_version);
1235
1236 while (!fe->shader_is_end(fe_data, &ptr))
1237 {
1238 /* Read opcode. */
1239 fe->shader_read_instruction(fe_data, &ptr, &ins);
1240
1241 /* Unknown opcode and its parameters. */
1242 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1243 {
1244 TRACE("Skipping unrecognized instruction.\n");
1245 continue;
1246 }
1247
1248 if (ins.predicate)
1249 FIXME("Predicates not implemented.\n");
1250
1251 /* Call appropriate function for output target */
1252 device->shader_backend->shader_handle_instruction(&ins);
1253 }
1254 }
1255
1256 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1257 {
1258 DWORD mmask = dst->modifiers;
1259
1260 switch (dst->shift)
1261 {
1262 case 0: break;
1263 case 13: TRACE("_d8"); break;
1264 case 14: TRACE("_d4"); break;
1265 case 15: TRACE("_d2"); break;
1266 case 1: TRACE("_x2"); break;
1267 case 2: TRACE("_x4"); break;
1268 case 3: TRACE("_x8"); break;
1269 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1270 }
1271
1272 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1273 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1274 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1275
1276 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1277 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1278 }
1279
1280 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1281 {
1282 switch (primitive_type)
1283 {
1284 case WINED3D_PT_UNDEFINED:
1285 TRACE("undefined");
1286 break;
1287 case WINED3D_PT_POINTLIST:
1288 TRACE("pointlist");
1289 break;
1290 case WINED3D_PT_LINELIST:
1291 TRACE("linelist");
1292 break;
1293 case WINED3D_PT_LINESTRIP:
1294 TRACE("linestrip");
1295 break;
1296 case WINED3D_PT_TRIANGLELIST:
1297 TRACE("trianglelist");
1298 break;
1299 case WINED3D_PT_TRIANGLESTRIP:
1300 TRACE("trianglestrip");
1301 break;
1302 case WINED3D_PT_TRIANGLEFAN:
1303 TRACE("trianglefan");
1304 break;
1305 case WINED3D_PT_LINELIST_ADJ:
1306 TRACE("linelist_adj");
1307 break;
1308 case WINED3D_PT_LINESTRIP_ADJ:
1309 TRACE("linestrip_adj");
1310 break;
1311 case WINED3D_PT_TRIANGLELIST_ADJ:
1312 TRACE("trianglelist_adj");
1313 break;
1314 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1315 TRACE("trianglestrip_adj");
1316 break;
1317 default:
1318 TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1319 break;
1320 }
1321 }
1322
1323 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1324 {
1325 struct wined3d_shader_version shader_version;
1326 const DWORD *ptr = byte_code;
1327 const char *type_prefix;
1328 DWORD i;
1329
1330 TRACE("Parsing %p.\n", byte_code);
1331
1332 fe->shader_read_header(fe_data, &ptr, &shader_version);
1333
1334 switch (shader_version.type)
1335 {
1336 case WINED3D_SHADER_TYPE_VERTEX:
1337 type_prefix = "vs";
1338 break;
1339
1340 case WINED3D_SHADER_TYPE_GEOMETRY:
1341 type_prefix = "gs";
1342 break;
1343
1344 case WINED3D_SHADER_TYPE_PIXEL:
1345 type_prefix = "ps";
1346 break;
1347
1348 default:
1349 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1350 type_prefix = "unknown";
1351 break;
1352 }
1353
1354 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1355
1356 while (!fe->shader_is_end(fe_data, &ptr))
1357 {
1358 struct wined3d_shader_instruction ins;
1359
1360 fe->shader_read_instruction(fe_data, &ptr, &ins);
1361 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1362 {
1363 TRACE("Skipping unrecognized instruction.\n");
1364 continue;
1365 }
1366
1367 if (ins.handler_idx == WINED3DSIH_DCL)
1368 {
1369 shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1370 shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1371 TRACE(" ");
1372 shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1373 }
1374 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1375 {
1376 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1377 shader_dump_src_param(&ins.declaration.src, &shader_version);
1378 TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1379 }
1380 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1381 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1382 {
1383 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1384 shader_dump_primitive_type(ins.declaration.primitive_type);
1385 }
1386 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1387 {
1388 TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1389 }
1390 else if (ins.handler_idx == WINED3DSIH_DEF)
1391 {
1392 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1393 ins.dst[0].reg.idx[0].offset),
1394 *(const float *)&ins.src[0].reg.immconst_data[0],
1395 *(const float *)&ins.src[0].reg.immconst_data[1],
1396 *(const float *)&ins.src[0].reg.immconst_data[2],
1397 *(const float *)&ins.src[0].reg.immconst_data[3]);
1398 }
1399 else if (ins.handler_idx == WINED3DSIH_DEFI)
1400 {
1401 TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1402 ins.src[0].reg.immconst_data[0],
1403 ins.src[0].reg.immconst_data[1],
1404 ins.src[0].reg.immconst_data[2],
1405 ins.src[0].reg.immconst_data[3]);
1406 }
1407 else if (ins.handler_idx == WINED3DSIH_DEFB)
1408 {
1409 TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1410 }
1411 else
1412 {
1413 if (ins.predicate)
1414 {
1415 TRACE("(");
1416 shader_dump_src_param(ins.predicate, &shader_version);
1417 TRACE(") ");
1418 }
1419
1420 /* PixWin marks instructions with the coissue flag with a '+' */
1421 if (ins.coissue) TRACE("+");
1422
1423 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1424
1425 if (ins.handler_idx == WINED3DSIH_IFC
1426 || ins.handler_idx == WINED3DSIH_BREAKC)
1427 {
1428 switch (ins.flags)
1429 {
1430 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1431 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1432 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1433 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1434 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1435 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1436 default: TRACE("_(%u)", ins.flags);
1437 }
1438 }
1439 else if (ins.handler_idx == WINED3DSIH_TEX
1440 && shader_version.major >= 2
1441 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1442 {
1443 TRACE("p");
1444 }
1445
1446 for (i = 0; i < ins.dst_count; ++i)
1447 {
1448 shader_dump_ins_modifiers(&ins.dst[i]);
1449 TRACE(!i ? " " : ", ");
1450 shader_dump_dst_param(&ins.dst[i], &shader_version);
1451 }
1452
1453 /* Other source tokens */
1454 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1455 {
1456 TRACE(!i ? " " : ", ");
1457 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1458 }
1459 }
1460 TRACE("\n");
1461 }
1462 }
1463
1464 static void shader_cleanup(struct wined3d_shader *shader)
1465 {
1466 shader->device->shader_backend->shader_destroy(shader);
1467 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1468 HeapFree(GetProcessHeap(), 0, shader->function);
1469 shader_delete_constant_list(&shader->constantsF);
1470 shader_delete_constant_list(&shader->constantsB);
1471 shader_delete_constant_list(&shader->constantsI);
1472 list_remove(&shader->shader_list_entry);
1473
1474 if (shader->frontend && shader->frontend_data)
1475 shader->frontend->shader_free(shader->frontend_data);
1476 }
1477
1478 struct shader_none_priv
1479 {
1480 const struct fragment_pipeline *fragment_pipe;
1481 BOOL ffp_proj_control;
1482 };
1483
1484 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1485 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1486 enum tex_types tex_type, const SIZE *ds_mask_size) {}
1487 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1488 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1489 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1490 static void shader_none_load_constants(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1491 static void shader_none_load_np2fixup_constants(void *shader_priv,
1492 const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1493 static void shader_none_destroy(struct wined3d_shader *shader) {}
1494 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1495
1496 static void shader_none_select(const struct wined3d_context *context, enum wined3d_shader_mode vertex_mode,
1497 enum wined3d_shader_mode fragment_mode)
1498 {
1499 const struct wined3d_gl_info *gl_info = context->gl_info;
1500 struct wined3d_device *device = context->swapchain->device;
1501 struct shader_none_priv *priv = device->shader_priv;
1502
1503 priv->fragment_pipe->enable_extension(gl_info, fragment_mode == WINED3D_SHADER_MODE_FFP);
1504 }
1505
1506 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct fragment_pipeline *fragment_pipe)
1507 {
1508 struct fragment_caps fragment_caps;
1509 struct shader_none_priv *priv;
1510 void *fragment_priv;
1511
1512 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1513 return E_OUTOFMEMORY;
1514
1515 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1516 {
1517 ERR("Failed to initialize fragment pipe.\n");
1518 HeapFree(GetProcessHeap(), 0, priv);
1519 return E_FAIL;
1520 }
1521
1522 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
1523 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
1524 device->fragment_priv = fragment_priv;
1525 priv->fragment_pipe = fragment_pipe;
1526 device->shader_priv = priv;
1527
1528 return WINED3D_OK;
1529 }
1530
1531 static void shader_none_free(struct wined3d_device *device)
1532 {
1533 struct shader_none_priv *priv = device->shader_priv;
1534
1535 priv->fragment_pipe->free_private(device);
1536 HeapFree(GetProcessHeap(), 0, priv);
1537 }
1538
1539 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1540 {
1541 /* Set the shader caps to 0 for the none shader backend */
1542 caps->vs_version = 0;
1543 caps->gs_version = 0;
1544 caps->ps_version = 0;
1545 caps->vs_uniform_count = 0;
1546 caps->ps_uniform_count = 0;
1547 caps->ps_1x_max_value = 0.0f;
1548 caps->wined3d_caps = 0;
1549 }
1550
1551 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1552 {
1553 /* We "support" every possible fixup, since we don't support any shader
1554 * model, and will never have to actually sample a texture. */
1555 return TRUE;
1556 }
1557
1558 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
1559 {
1560 struct shader_none_priv *priv = shader_priv;
1561
1562 return priv->ffp_proj_control;
1563 }
1564
1565 const struct wined3d_shader_backend_ops none_shader_backend =
1566 {
1567 shader_none_handle_instruction,
1568 shader_none_select,
1569 shader_none_select_depth_blt,
1570 shader_none_deselect_depth_blt,
1571 shader_none_update_float_vertex_constants,
1572 shader_none_update_float_pixel_constants,
1573 shader_none_load_constants,
1574 shader_none_load_np2fixup_constants,
1575 shader_none_destroy,
1576 shader_none_alloc,
1577 shader_none_free,
1578 shader_none_context_destroyed,
1579 shader_none_get_caps,
1580 shader_none_color_fixup_supported,
1581 shader_none_has_ffp_proj_control,
1582 };
1583
1584 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1585 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1586 enum wined3d_shader_type type, unsigned int max_version)
1587 {
1588 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1589 const struct wined3d_shader_frontend *fe;
1590 HRESULT hr;
1591 unsigned int backend_version;
1592
1593 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1594 shader, byte_code, output_signature, float_const_count);
1595
1596 fe = shader_select_frontend(*byte_code);
1597 if (!fe)
1598 {
1599 FIXME("Unable to find frontend for shader.\n");
1600 return WINED3DERR_INVALIDCALL;
1601 }
1602 shader->frontend = fe;
1603 shader->frontend_data = fe->shader_init(byte_code, output_signature);
1604 if (!shader->frontend_data)
1605 {
1606 FIXME("Failed to initialize frontend.\n");
1607 return WINED3DERR_INVALIDCALL;
1608 }
1609
1610 /* First pass: trace shader. */
1611 if (TRACE_ON(d3d_shader))
1612 shader_trace_init(fe, shader->frontend_data, byte_code);
1613
1614 /* Initialize immediate constant lists. */
1615 list_init(&shader->constantsF);
1616 list_init(&shader->constantsB);
1617 list_init(&shader->constantsI);
1618
1619 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1620 hr = shader_get_registers_used(shader, fe,
1621 reg_maps, shader->input_signature, shader->output_signature,
1622 byte_code, float_const_count);
1623 if (FAILED(hr)) return hr;
1624
1625 if (reg_maps->shader_version.type != type)
1626 {
1627 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1628 return WINED3DERR_INVALIDCALL;
1629 }
1630 if (reg_maps->shader_version.major > max_version)
1631 {
1632 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1633 return WINED3DERR_INVALIDCALL;
1634 }
1635 switch (type)
1636 {
1637 case WINED3D_SHADER_TYPE_VERTEX:
1638 backend_version = shader->device->vs_version;
1639 break;
1640 case WINED3D_SHADER_TYPE_GEOMETRY:
1641 backend_version = shader->device->gs_version;
1642 break;
1643 case WINED3D_SHADER_TYPE_PIXEL:
1644 backend_version = shader->device->ps_version;
1645 break;
1646 default:
1647 FIXME("No backend version-checking for this shader type\n");
1648 backend_version = 0;
1649 }
1650 if (reg_maps->shader_version.major > backend_version)
1651 {
1652 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1653 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1654 return WINED3DERR_INVALIDCALL;
1655 }
1656
1657 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1658 if (!shader->function)
1659 return E_OUTOFMEMORY;
1660 memcpy(shader->function, byte_code, shader->functionLength);
1661
1662 return WINED3D_OK;
1663 }
1664
1665 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1666 {
1667 ULONG refcount = InterlockedIncrement(&shader->ref);
1668
1669 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1670
1671 return refcount;
1672 }
1673
1674 /* Do not call while under the GL lock. */
1675 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1676 {
1677 ULONG refcount = InterlockedDecrement(&shader->ref);
1678
1679 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1680
1681 if (!refcount)
1682 {
1683 shader_cleanup(shader);
1684 shader->parent_ops->wined3d_object_destroyed(shader->parent);
1685 HeapFree(GetProcessHeap(), 0, shader);
1686 }
1687
1688 return refcount;
1689 }
1690
1691 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1692 {
1693 TRACE("shader %p.\n", shader);
1694
1695 return shader->parent;
1696 }
1697
1698 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1699 void *byte_code, UINT *byte_code_size)
1700 {
1701 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1702
1703 if (!byte_code)
1704 {
1705 *byte_code_size = shader->functionLength;
1706 return WINED3D_OK;
1707 }
1708
1709 if (*byte_code_size < shader->functionLength)
1710 {
1711 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1712 * than the required size we should write the required size and
1713 * return D3DERR_MOREDATA. That's not actually true. */
1714 return WINED3DERR_INVALIDCALL;
1715 }
1716
1717 memcpy(byte_code, shader->function, shader->functionLength);
1718
1719 return WINED3D_OK;
1720 }
1721
1722 /* Set local constants for d3d8 shaders. */
1723 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1724 UINT start_idx, const float *src_data, UINT count)
1725 {
1726 UINT end_idx = start_idx + count;
1727 UINT i;
1728
1729 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1730
1731 if (end_idx > shader->limits.constant_float)
1732 {
1733 WARN("end_idx %u > float constants limit %u.\n",
1734 end_idx, shader->limits.constant_float);
1735 end_idx = shader->limits.constant_float;
1736 }
1737
1738 for (i = start_idx; i < end_idx; ++i)
1739 {
1740 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1741 if (!lconst)
1742 return E_OUTOFMEMORY;
1743
1744 lconst->idx = i;
1745 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1746 list_add_head(&shader->constantsF, &lconst->entry);
1747 }
1748
1749 return WINED3D_OK;
1750 }
1751
1752 void find_vs_compile_args(const struct wined3d_state *state,
1753 const struct wined3d_shader *shader, struct vs_compile_args *args)
1754 {
1755 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1756 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1757 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1758 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1759 args->swizzle_map = shader->device->strided_streams.swizzle_map;
1760 }
1761
1762 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1763 {
1764 if (usage_idx1 != usage_idx2)
1765 return FALSE;
1766 if (usage1 == usage2)
1767 return TRUE;
1768 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1769 return TRUE;
1770 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1771 return TRUE;
1772
1773 return FALSE;
1774 }
1775
1776 BOOL vshader_get_input(const struct wined3d_shader *shader,
1777 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1778 {
1779 WORD map = shader->reg_maps.input_registers;
1780 unsigned int i;
1781
1782 for (i = 0; map; map >>= 1, ++i)
1783 {
1784 if (!(map & 1)) continue;
1785
1786 if (match_usage(shader->u.vs.attributes[i].usage,
1787 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1788 {
1789 *regnum = i;
1790 return TRUE;
1791 }
1792 }
1793 return FALSE;
1794 }
1795
1796 static void vertexshader_set_limits(struct wined3d_shader *shader)
1797 {
1798 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1799 shader->reg_maps.shader_version.minor);
1800 struct wined3d_device *device = shader->device;
1801
1802 shader->limits.packed_input = 0;
1803
1804 switch (shader_version)
1805 {
1806 case WINED3D_SHADER_VERSION(1, 0):
1807 case WINED3D_SHADER_VERSION(1, 1):
1808 shader->limits.constant_bool = 0;
1809 shader->limits.constant_int = 0;
1810 shader->limits.packed_output = 12;
1811 shader->limits.sampler = 0;
1812 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1813 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1814 * constants? */
1815 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1816 break;
1817
1818 case WINED3D_SHADER_VERSION(2, 0):
1819 case WINED3D_SHADER_VERSION(2, 1):
1820 shader->limits.constant_bool = 16;
1821 shader->limits.constant_int = 16;
1822 shader->limits.packed_output = 12;
1823 shader->limits.sampler = 0;
1824 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1825 break;
1826
1827 case WINED3D_SHADER_VERSION(3, 0):
1828 shader->limits.constant_bool = 16;
1829 shader->limits.constant_int = 16;
1830 shader->limits.packed_output = 12;
1831 shader->limits.sampler = 4;
1832 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1833 * even though they are capable of supporting much more (GL
1834 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1835 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1836 * shaders to 256. */
1837 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1838 break;
1839
1840 case WINED3D_SHADER_VERSION(4, 0):
1841 shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1842 shader->limits.constant_int = 0;
1843 shader->limits.constant_float = 0;
1844 shader->limits.constant_bool = 0;
1845 shader->limits.packed_output = 16;
1846 shader->limits.packed_input = 0;
1847 break;
1848
1849 default:
1850 shader->limits.constant_bool = 16;
1851 shader->limits.constant_int = 16;
1852 shader->limits.packed_output = 12;
1853 shader->limits.sampler = 0;
1854 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1855 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1856 shader->reg_maps.shader_version.major,
1857 shader->reg_maps.shader_version.minor);
1858 }
1859 }
1860
1861 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1862 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1863 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1864 {
1865 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1866 unsigned int i;
1867 HRESULT hr;
1868 WORD map;
1869
1870 if (!byte_code) return WINED3DERR_INVALIDCALL;
1871
1872 shader_init(shader, device, parent, parent_ops);
1873 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1874 WINED3D_SHADER_TYPE_VERTEX, max_version);
1875 if (FAILED(hr))
1876 {
1877 WARN("Failed to set function, hr %#x.\n", hr);
1878 shader_cleanup(shader);
1879 return hr;
1880 }
1881
1882 map = reg_maps->input_registers;
1883 for (i = 0; map; map >>= 1, ++i)
1884 {
1885 if (!(map & 1) || !shader->input_signature[i].semantic_name)
1886 continue;
1887
1888 shader->u.vs.attributes[i].usage =
1889 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1890 shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1891 }
1892
1893 if (output_signature)
1894 {
1895 for (i = 0; i < output_signature->element_count; ++i)
1896 {
1897 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1898 reg_maps->output_registers |= 1 << e->register_idx;
1899 shader->output_signature[e->register_idx] = *e;
1900 }
1901 }
1902
1903 vertexshader_set_limits(shader);
1904
1905 shader->load_local_constsF = reg_maps->usesrelconstF
1906 && !list_empty(&shader->constantsF);
1907
1908 return WINED3D_OK;
1909 }
1910
1911 static void geometryshader_set_limits(struct wined3d_shader *shader)
1912 {
1913 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1914 shader->reg_maps.shader_version.minor);
1915
1916 switch (shader_version)
1917 {
1918 case WINED3D_SHADER_VERSION(4, 0):
1919 shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1920 shader->limits.constant_int = 0;
1921 shader->limits.constant_float = 0;
1922 shader->limits.constant_bool = 0;
1923 shader->limits.packed_output = 32;
1924 shader->limits.packed_input = 16;
1925 break;
1926
1927 default:
1928 memset(&shader->limits, 0, sizeof(shader->limits));
1929 FIXME("Unhandled geometry shader version \"%u.%u\".\n",
1930 shader->reg_maps.shader_version.major,
1931 shader->reg_maps.shader_version.minor);
1932 }
1933 }
1934
1935 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1936 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1937 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1938 {
1939 HRESULT hr;
1940
1941 shader_init(shader, device, parent, parent_ops);
1942 hr = shader_set_function(shader, byte_code, output_signature, 0,
1943 WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1944 if (FAILED(hr))
1945 {
1946 WARN("Failed to set function, hr %#x.\n", hr);
1947 shader_cleanup(shader);
1948 return hr;
1949 }
1950
1951 geometryshader_set_limits(shader);
1952
1953 shader->load_local_constsF = FALSE;
1954
1955 return WINED3D_OK;
1956 }
1957
1958 void find_ps_compile_args(const struct wined3d_state *state,
1959 const struct wined3d_shader *shader, struct ps_compile_args *args)
1960 {
1961 struct wined3d_device *device = shader->device;
1962 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1963 const struct wined3d_texture *texture;
1964 UINT i;
1965
1966 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1967 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1968 {
1969 const struct wined3d_surface *rt = state->fb->render_targets[0];
1970 if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
1971 {
1972 static unsigned int warned = 0;
1973
1974 args->srgb_correction = 1;
1975 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
1976 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
1977 "support, expect rendering artifacts.\n");
1978 }
1979 }
1980
1981 if (shader->reg_maps.shader_version.major == 1
1982 && shader->reg_maps.shader_version.minor <= 3)
1983 {
1984 for (i = 0; i < 4; ++i)
1985 {
1986 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
1987
1988 if (flags & WINED3D_TTFF_PROJECTED)
1989 {
1990 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
1991
1992 if (!state->vertex_shader)
1993 {
1994 unsigned int j;
1995 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
1996 DWORD max_valid = WINED3D_TTFF_COUNT4;
1997 enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
1998
1999 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2000 {
2001 struct wined3d_vertex_declaration_element *element =
2002 &state->vertex_declaration->elements[j];
2003
2004 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2005 && element->usage_idx == index)
2006 {
2007 max_valid = element->format->component_count;
2008 break;
2009 }
2010 }
2011 if (!tex_transform || tex_transform > max_valid)
2012 {
2013 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2014 tex_transform, max_valid);
2015 tex_transform = max_valid;
2016 }
2017 if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
2018 || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
2019 || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
2020 tex_transform |= WINED3D_PSARGS_PROJECTED;
2021 else
2022 {
2023 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2024 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2025 i, tex_transform, sampler_type);
2026 }
2027 }
2028 else
2029 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2030
2031 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2032 }
2033 }
2034 }
2035
2036 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2037 {
2038 if (!shader->reg_maps.sampler_type[i])
2039 continue;
2040
2041 texture = state->textures[i];
2042 if (!texture)
2043 {
2044 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2045 continue;
2046 }
2047 args->color_fixup[i] = texture->resource.format->color_fixup;
2048
2049 if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2050 args->shadow |= 1 << i;
2051
2052 /* Flag samplers that need NP2 texcoord fixup. */
2053 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2054 args->np2_fixup |= (1 << i);
2055 }
2056 if (shader->reg_maps.shader_version.major >= 3)
2057 {
2058 if (device->strided_streams.position_transformed)
2059 {
2060 args->vp_mode = pretransformed;
2061 }
2062 else if (use_vs(state))
2063 {
2064 args->vp_mode = vertexshader;
2065 }
2066 else
2067 {
2068 args->vp_mode = fixedfunction;
2069 }
2070 args->fog = FOG_OFF;
2071 }
2072 else
2073 {
2074 args->vp_mode = vertexshader;
2075 if (state->render_states[WINED3D_RS_FOGENABLE])
2076 {
2077 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2078 {
2079 case WINED3D_FOG_NONE:
2080 if (device->strided_streams.position_transformed || use_vs(state))
2081 {
2082 args->fog = FOG_LINEAR;
2083 break;
2084 }
2085
2086 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2087 {
2088 case WINED3D_FOG_NONE: /* Fall through. */
2089 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2090 case WINED3D_FOG_EXP: args->fog = FOG_EXP; break;
2091 case WINED3D_FOG_EXP2: args->fog = FOG_EXP2; break;
2092 }
2093 break;
2094
2095 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2096 case WINED3D_FOG_EXP: args->fog = FOG_EXP; break;
2097 case WINED3D_FOG_EXP2: args->fog = FOG_EXP2; break;
2098 }
2099 }
2100 else
2101 {
2102 args->fog = FOG_OFF;
2103 }
2104 }
2105 }
2106
2107 static void pixelshader_set_limits(struct wined3d_shader *shader)
2108 {
2109 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2110 shader->reg_maps.shader_version.minor);
2111
2112 shader->limits.packed_output = 0;
2113
2114 switch (shader_version)
2115 {
2116 case WINED3D_SHADER_VERSION(1, 0):
2117 case WINED3D_SHADER_VERSION(1, 1):
2118 case WINED3D_SHADER_VERSION(1, 2):
2119 case WINED3D_SHADER_VERSION(1, 3):
2120 shader->limits.constant_float = 8;
2121 shader->limits.constant_int = 0;
2122 shader->limits.constant_bool = 0;
2123 shader->limits.sampler = 4;
2124 shader->limits.packed_input = 0;
2125 break;
2126
2127 case WINED3D_SHADER_VERSION(1, 4):
2128 shader->limits.constant_float = 8;
2129 shader->limits.constant_int = 0;
2130 shader->limits.constant_bool = 0;
2131 shader->limits.sampler = 6;
2132 shader->limits.packed_input = 0;
2133 break;
2134
2135 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2136 case WINED3D_SHADER_VERSION(2, 0):
2137 shader->limits.constant_float = 32;
2138 shader->limits.constant_int = 16;
2139 shader->limits.constant_bool = 16;
2140 shader->limits.sampler = 16;
2141 shader->limits.packed_input = 0;
2142 break;
2143
2144 case WINED3D_SHADER_VERSION(2, 1):
2145 shader->limits.constant_float = 32;
2146 shader->limits.constant_int = 16;
2147 shader->limits.constant_bool = 16;
2148 shader->limits.sampler = 16;
2149 shader->limits.packed_input = 0;
2150 break;
2151
2152 case WINED3D_SHADER_VERSION(3, 0):
2153 shader->limits.constant_float = 224;
2154 shader->limits.constant_int = 16;
2155 shader->limits.constant_bool = 16;
2156 shader->limits.sampler = 16;
2157 shader->limits.packed_input = 12;
2158 break;
2159
2160 case WINED3D_SHADER_VERSION(4, 0):
2161 shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
2162 shader->limits.constant_int = 0;
2163 shader->limits.constant_float = 0;
2164 shader->limits.constant_bool = 0;
2165 shader->limits.packed_input = 32;
2166 break;
2167
2168 default:
2169 shader->limits.constant_float = 32;
2170 shader->limits.constant_int = 16;
2171 shader->limits.constant_bool = 16;
2172 shader->limits.sampler = 16;
2173 shader->limits.packed_input = 0;
2174 FIXME("Unrecognized pixel shader version %u.%u\n",
2175 shader->reg_maps.shader_version.major,
2176 shader->reg_maps.shader_version.minor);
2177 }
2178 }
2179
2180 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2181 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2182 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2183 {
2184 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2185 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2186 HRESULT hr;
2187
2188 if (!byte_code) return WINED3DERR_INVALIDCALL;
2189
2190 shader_init(shader, device, parent, parent_ops);
2191 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2192 WINED3D_SHADER_TYPE_PIXEL, max_version);
2193 if (FAILED(hr))
2194 {
2195 WARN("Failed to set function, hr %#x.\n", hr);
2196 shader_cleanup(shader);
2197 return hr;
2198 }
2199
2200 pixelshader_set_limits(shader);
2201
2202 for (i = 0; i < MAX_REG_INPUT; ++i)
2203 {
2204 if (shader->u.ps.input_reg_used[i])
2205 {
2206 ++num_regs_used;
2207 highest_reg_used = i;
2208 }
2209 }
2210
2211 /* Don't do any register mapping magic if it is not needed, or if we can't
2212 * achieve anything anyway */
2213 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2214 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2215 {
2216 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2217 {
2218 /* This happens with relative addressing. The input mapper function
2219 * warns about this if the higher registers are declared too, so
2220 * don't write a FIXME here */
2221 WARN("More varying registers used than supported\n");
2222 }
2223
2224 for (i = 0; i < MAX_REG_INPUT; ++i)
2225 {
2226 shader->u.ps.input_reg_map[i] = i;
2227 }
2228
2229 shader->u.ps.declared_in_count = highest_reg_used + 1;
2230 }
2231 else
2232 {
2233 shader->u.ps.declared_in_count = 0;
2234 for (i = 0; i < MAX_REG_INPUT; ++i)
2235 {
2236 if (shader->u.ps.input_reg_used[i])
2237 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2238 else shader->u.ps.input_reg_map[i] = ~0U;
2239 }
2240 }
2241
2242 shader->load_local_constsF = FALSE;
2243
2244 return WINED3D_OK;
2245 }
2246
2247 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2248 {
2249 enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2250 unsigned int i;
2251
2252 if (reg_maps->shader_version.major != 1) return;
2253
2254 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2255 {
2256 /* We don't sample from this sampler. */
2257 if (!sampler_type[i]) continue;
2258
2259 if (!textures[i])
2260 {
2261 WARN("No texture bound to sampler %u, using 2D.\n", i);
2262 sampler_type[i] = WINED3DSTT_2D;
2263 continue;
2264 }
2265
2266 switch (textures[i]->target)
2267 {
2268 case GL_TEXTURE_RECTANGLE_ARB:
2269 case GL_TEXTURE_2D:
2270 /* We have to select between texture rectangles and 2D
2271 * textures later because 2.0 and 3.0 shaders only have
2272 * WINED3DSTT_2D as well. */
2273 sampler_type[i] = WINED3DSTT_2D;
2274 break;
2275
2276 case GL_TEXTURE_3D:
2277 sampler_type[i] = WINED3DSTT_VOLUME;
2278 break;
2279
2280 case GL_TEXTURE_CUBE_MAP_ARB:
2281 sampler_type[i] = WINED3DSTT_CUBE;
2282 break;
2283
2284 default:
2285 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2286 sampler_type[i] = WINED3DSTT_2D;
2287 }
2288 }
2289 }
2290
2291 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2292 const struct wined3d_shader_signature *output_signature, void *parent,
2293 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2294 {
2295 struct wined3d_shader *object;
2296 HRESULT hr;
2297
2298 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2299 device, byte_code, output_signature, parent, parent_ops, shader);
2300
2301 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2302 if (!object)
2303 return E_OUTOFMEMORY;
2304
2305 hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2306 if (FAILED(hr))
2307 {
2308 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2309 HeapFree(GetProcessHeap(), 0, object);
2310 return hr;
2311 }
2312
2313 TRACE("Created geometry shader %p.\n", object);
2314 *shader = object;
2315
2316 return WINED3D_OK;
2317 }
2318
2319 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2320 const struct wined3d_shader_signature *output_signature, void *parent,
2321 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2322 {
2323 struct wined3d_shader *object;
2324 HRESULT hr;
2325
2326 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2327 device, byte_code, output_signature, parent, parent_ops, shader);
2328
2329 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2330 if (!object)
2331 return E_OUTOFMEMORY;
2332
2333 hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2334 if (FAILED(hr))
2335 {
2336 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2337 HeapFree(GetProcessHeap(), 0, object);
2338 return hr;
2339 }
2340
2341 TRACE("Created pixel shader %p.\n", object);
2342 *shader = object;
2343
2344 return WINED3D_OK;
2345 }
2346
2347 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2348 const struct wined3d_shader_signature *output_signature, void *parent,
2349 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2350 {
2351 struct wined3d_shader *object;
2352 HRESULT hr;
2353
2354 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2355 device, byte_code, output_signature, parent, parent_ops, shader);
2356
2357 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2358 if (!object)
2359 return E_OUTOFMEMORY;
2360
2361 hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2362 if (FAILED(hr))
2363 {
2364 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2365 HeapFree(GetProcessHeap(), 0, object);
2366 return hr;
2367 }
2368
2369 TRACE("Created vertex shader %p.\n", object);
2370 *shader = object;
2371
2372 return WINED3D_OK;
2373 }
A free Windows-compatible Operating System