shader.c@ 35718

Last change on this file since 35718 was 33656, checked in by vboxsync, 14 years ago
*: rebrand Sun (L)GPL disclaimers
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 80.1 KB

Line
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 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	/*
26	* Oracle LGPL Disclaimer: For the avoidance of doubt, except that if any license choice
27	* other than GPL or LGPL is available it will apply instead, Oracle elects to use only
28	* the Lesser General Public License version 2.1 (LGPLv2) at this time for any software where
29	* a choice of LGPL license versions is made available with the language indicating
30	* that LGPLv2 or any later version may be used, or where a choice of which version
31	* of the LGPL is applied is otherwise unspecified.
32	*/
33
34	#include "config.h"
35	#include "wine/port.h"
36
37	#include <math.h>
38	#include <stdio.h>
39	#include <string.h>
40
41	#include "wined3d_private.h"
42
43	WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
44	WINE_DECLARE_DEBUG_CHANNEL(d3d);
45
46	static const char *shader_opcode_names[] =
47	{
48	/* WINED3DSIH_ABS */ "abs",
49	/* WINED3DSIH_ADD */ "add",
50	/* WINED3DSIH_BEM */ "bem",
51	/* WINED3DSIH_BREAK */ "break",
52	/* WINED3DSIH_BREAKC */ "breakc",
53	/* WINED3DSIH_BREAKP */ "breakp",
54	/* WINED3DSIH_CALL */ "call",
55	/* WINED3DSIH_CALLNZ */ "callnz",
56	/* WINED3DSIH_CMP */ "cmp",
57	/* WINED3DSIH_CND */ "cnd",
58	/* WINED3DSIH_CRS */ "crs",
59	/* WINED3DSIH_CUT */ "cut",
60	/* WINED3DSIH_DCL */ "dcl",
61	/* WINED3DSIH_DEF */ "def",
62	/* WINED3DSIH_DEFB */ "defb",
63	/* WINED3DSIH_DEFI */ "defi",
64	/* WINED3DSIH_DP2ADD */ "dp2add",
65	/* WINED3DSIH_DP3 */ "dp3",
66	/* WINED3DSIH_DP4 */ "dp4",
67	/* WINED3DSIH_DST */ "dst",
68	/* WINED3DSIH_DSX */ "dsx",
69	/* WINED3DSIH_DSY */ "dsy",
70	/* WINED3DSIH_ELSE */ "else",
71	/* WINED3DSIH_EMIT */ "emit",
72	/* WINED3DSIH_ENDIF */ "endif",
73	/* WINED3DSIH_ENDLOOP */ "endloop",
74	/* WINED3DSIH_ENDREP */ "endrep",
75	/* WINED3DSIH_EXP */ "exp",
76	/* WINED3DSIH_EXPP */ "expp",
77	/* WINED3DSIH_FRC */ "frc",
78	/* WINED3DSIH_IADD */ "iadd",
79	/* WINED3DSIH_IF */ "if",
80	/* WINED3DSIH_IFC */ "ifc",
81	/* WINED3DSIH_IGE */ "ige",
82	/* WINED3DSIH_LABEL */ "label",
83	/* WINED3DSIH_LIT */ "lit",
84	/* WINED3DSIH_LOG */ "log",
85	/* WINED3DSIH_LOGP */ "logp",
86	/* WINED3DSIH_LOOP */ "loop",
87	/* WINED3DSIH_LRP */ "lrp",
88	/* WINED3DSIH_LT */ "lt",
89	/* WINED3DSIH_M3x2 */ "m3x2",
90	/* WINED3DSIH_M3x3 */ "m3x3",
91	/* WINED3DSIH_M3x4 */ "m3x4",
92	/* WINED3DSIH_M4x3 */ "m4x3",
93	/* WINED3DSIH_M4x4 */ "m4x4",
94	/* WINED3DSIH_MAD */ "mad",
95	/* WINED3DSIH_MAX */ "max",
96	/* WINED3DSIH_MIN */ "min",
97	/* WINED3DSIH_MOV */ "mov",
98	/* WINED3DSIH_MOVA */ "mova",
99	/* WINED3DSIH_MUL */ "mul",
100	/* WINED3DSIH_NOP */ "nop",
101	/* WINED3DSIH_NRM */ "nrm",
102	/* WINED3DSIH_PHASE */ "phase",
103	/* WINED3DSIH_POW */ "pow",
104	/* WINED3DSIH_RCP */ "rcp",
105	/* WINED3DSIH_REP */ "rep",
106	/* WINED3DSIH_RET */ "ret",
107	/* WINED3DSIH_RSQ */ "rsq",
108	/* WINED3DSIH_SETP */ "setp",
109	/* WINED3DSIH_SGE */ "sge",
110	/* WINED3DSIH_SGN */ "sgn",
111	/* WINED3DSIH_SINCOS */ "sincos",
112	/* WINED3DSIH_SLT */ "slt",
113	/* WINED3DSIH_SUB */ "sub",
114	/* WINED3DSIH_TEX */ "texld",
115	/* WINED3DSIH_TEXBEM */ "texbem",
116	/* WINED3DSIH_TEXBEML */ "texbeml",
117	/* WINED3DSIH_TEXCOORD */ "texcrd",
118	/* WINED3DSIH_TEXDEPTH */ "texdepth",
119	/* WINED3DSIH_TEXDP3 */ "texdp3",
120	/* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
121	/* WINED3DSIH_TEXKILL */ "texkill",
122	/* WINED3DSIH_TEXLDD */ "texldd",
123	/* WINED3DSIH_TEXLDL */ "texldl",
124	/* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
125	/* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
126	/* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
127	/* WINED3DSIH_TEXM3x3 */ "texm3x3",
128	/* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
129	/* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
130	/* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
131	/* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
132	/* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
133	/* WINED3DSIH_TEXREG2AR */ "texreg2ar",
134	/* WINED3DSIH_TEXREG2GB */ "texreg2gb",
135	/* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
136	};
137
138	static const char *semantic_names[] =
139	{
140	/* WINED3DDECLUSAGE_POSITION */ "SV_POSITION",
141	/* WINED3DDECLUSAGE_BLENDWEIGHT */ "BLENDWEIGHT",
142	/* WINED3DDECLUSAGE_BLENDINDICES */ "BLENDINDICES",
143	/* WINED3DDECLUSAGE_NORMAL */ "NORMAL",
144	/* WINED3DDECLUSAGE_PSIZE */ "PSIZE",
145	/* WINED3DDECLUSAGE_TEXCOORD */ "TEXCOORD",
146	/* WINED3DDECLUSAGE_TANGENT */ "TANGENT",
147	/* WINED3DDECLUSAGE_BINORMAL */ "BINORMAL",
148	/* WINED3DDECLUSAGE_TESSFACTOR */ "TESSFACTOR",
149	/* WINED3DDECLUSAGE_POSITIONT */ "POSITIONT",
150	/* WINED3DDECLUSAGE_COLOR */ "COLOR",
151	/* WINED3DDECLUSAGE_FOG */ "FOG",
152	/* WINED3DDECLUSAGE_DEPTH */ "DEPTH",
153	/* WINED3DDECLUSAGE_SAMPLE */ "SAMPLE",
154	};
155
156	static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
157	{
158	if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
159	{
160	FIXME("Unrecognized usage %#x.\n", usage);
161	return "UNRECOGNIZED";
162	}
163
164	return semantic_names[usage];
165	}
166
167	static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
168	{
169	unsigned int i;
170
171	for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
172	{
173	if (!strcmp(name, semantic_names[i])) return i;
174	}
175
176	return ~0U;
177	}
178
179	BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
180	{
181	return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
182	}
183
184	static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
185	const struct wined3d_shader_semantic *s)
186	{
187	e->semantic_name = shader_semantic_name_from_usage(s->usage);
188	e->semantic_idx = s->usage_idx;
189	e->sysval_semantic = 0;
190	e->component_type = 0;
191	e->register_idx = s->reg.reg.idx;
192	e->mask = s->reg.write_mask;
193	}
194
195	static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
196	{
197	switch (version_token >> 16)
198	{
199	case WINED3D_SM1_VS:
200	case WINED3D_SM1_PS:
201	return &sm1_shader_frontend;
202
203	case WINED3D_SM4_PS:
204	case WINED3D_SM4_VS:
205	case WINED3D_SM4_GS:
206	return &sm4_shader_frontend;
207
208	default:
209	FIXME("Unrecognised version token %#x\n", version_token);
210	return NULL;
211	}
212	}
213
214	void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
215	{
216	buffer->buffer[0] = '\0';
217	buffer->bsize = 0;
218	buffer->lineNo = 0;
219	buffer->newline = TRUE;
220	}
221
222	BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
223	{
224	buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
225	if (!buffer->buffer)
226	{
227	ERR("Failed to allocate shader buffer memory.\n");
228	return FALSE;
229	}
230
231	shader_buffer_clear(buffer);
232	return TRUE;
233	}
234
235	void shader_buffer_free(struct wined3d_shader_buffer *buffer)
236	{
237	HeapFree(GetProcessHeap(), 0, buffer->buffer);
238	}
239
240	int shader_vaddline(struct wined3d_shader_buffer buffer, const char format, va_list args)
241	{
242	char *base = buffer->buffer + buffer->bsize;
243	int rc;
244
245	rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
246
247	if (rc < 0 /* C89 / \|\| (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize / C99 */)
248	{
249	ERR("The buffer allocated for the shader program string "
250	"is too small at %d bytes.\n", SHADER_PGMSIZE);
251	buffer->bsize = SHADER_PGMSIZE - 1;
252	return -1;
253	}
254
255	if (buffer->newline)
256	{
257	TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
258	buffer->newline = FALSE;
259	}
260	else
261	{
262	TRACE("%s", base);
263	}
264
265	buffer->bsize += rc;
266	if (buffer->buffer[buffer->bsize-1] == '\n')
267	{
268	++buffer->lineNo;
269	buffer->newline = TRUE;
270	}
271
272	return 0;
273	}
274
275	int shader_addline(struct wined3d_shader_buffer buffer, const char format, ...)
276	{
277	va_list args;
278	int ret;
279
280	va_start(args, format);
281	ret = shader_vaddline(buffer, format, args);
282	va_end(args);
283
284	return ret;
285	}
286
287	static void shader_init(struct IWineD3DBaseShaderClass shader, IWineD3DDeviceImpl device,
288	IUnknown parent, const struct wined3d_parent_ops parent_ops)
289	{
290	shader->ref = 1;
291	shader->device = (IWineD3DDevice *)device;
292	shader->parent = parent;
293	shader->parent_ops = parent_ops;
294	list_init(&shader->linked_programs);
295	list_add_head(&device->shaders, &shader->shader_list_entry);
296	}
297
298	/* Convert floating point offset relative to a register file to an absolute
299	* offset for float constants. */
300	static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
301	{
302	switch (register_type)
303	{
304	case WINED3DSPR_CONST: return register_idx;
305	case WINED3DSPR_CONST2: return 2048 + register_idx;
306	case WINED3DSPR_CONST3: return 4096 + register_idx;
307	case WINED3DSPR_CONST4: return 6144 + register_idx;
308	default:
309	FIXME("Unsupported register type: %u.\n", register_type);
310	return register_idx;
311	}
312	}
313
314	static void shader_delete_constant_list(struct list *clist)
315	{
316	struct local_constant *constant;
317	struct list *ptr;
318
319	ptr = list_head(clist);
320	while (ptr)
321	{
322	constant = LIST_ENTRY(ptr, struct local_constant, entry);
323	ptr = list_next(clist, ptr);
324	HeapFree(GetProcessHeap(), 0, constant);
325	}
326	list_init(clist);
327	}
328
329	static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
330	{
331	DWORD idx, shift;
332	idx = bit >> 5;
333	shift = bit & 0x1f;
334	bitmap[idx] \|= (1 << shift);
335	}
336
337	static void shader_record_register_usage(IWineD3DBaseShaderImpl shader, struct shader_reg_maps reg_maps,
338	const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
339	{
340	switch (reg->type)
341	{
342	case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
343	if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord \|= 1 << reg->idx;
344	else reg_maps->address \|= 1 << reg->idx;
345	break;
346
347	case WINED3DSPR_TEMP:
348	reg_maps->temporary \|= 1 << reg->idx;
349	break;
350
351	case WINED3DSPR_INPUT:
352	if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
353	{
354	if (reg->rel_addr)
355	{
356	/* If relative addressing is used, we must assume that all registers
357	* are used. Even if it is a construct like v3[aL], we can't assume
358	* that v0, v1 and v2 aren't read because aL can be negative */
359	unsigned int i;
360	for (i = 0; i < MAX_REG_INPUT; ++i)
361	{
362	((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
363	}
364	}
365	else
366	{
367	((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg->idx] = TRUE;
368	}
369	}
370	else reg_maps->input_registers \|= 1 << reg->idx;
371	break;
372
373	case WINED3DSPR_RASTOUT:
374	if (reg->idx == 1) reg_maps->fog = 1;
375	break;
376
377	case WINED3DSPR_MISCTYPE:
378	if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
379	{
380	if (reg->idx == 0) reg_maps->vpos = 1;
381	else if (reg->idx == 1) reg_maps->usesfacing = 1;
382	}
383	break;
384
385	case WINED3DSPR_CONST:
386	if (reg->rel_addr)
387	{
388	if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
389	{
390	if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
391	{
392	((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
393	}
394	if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
395	{
396	((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
397	}
398	}
399	reg_maps->usesrelconstF = TRUE;
400	}
401	else
402	{
403	set_bitmap_bit(reg_maps->constf, reg->idx);
404	}
405	break;
406
407	case WINED3DSPR_CONSTINT:
408	reg_maps->integer_constants \|= (1 << reg->idx);
409	break;
410
411	case WINED3DSPR_CONSTBOOL:
412	reg_maps->boolean_constants \|= (1 << reg->idx);
413	break;
414
415	case WINED3DSPR_COLOROUT:
416	reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
417	break;
418
419	default:
420	TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
421	break;
422	}
423	}
424
425	static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
426	{
427	switch (instr)
428	{
429	case WINED3DSIH_M4x4:
430	case WINED3DSIH_M3x4:
431	return param == 1 ? 3 : 0;
432
433	case WINED3DSIH_M4x3:
434	case WINED3DSIH_M3x3:
435	return param == 1 ? 2 : 0;
436
437	case WINED3DSIH_M3x2:
438	return param == 1 ? 1 : 0;
439
440	default:
441	return 0;
442	}
443	}
444
445	/* Note that this does not count the loop register as an address register. */
446	static HRESULT shader_get_registers_used(IWineD3DBaseShader iface, const struct wined3d_shader_frontend fe,
447	struct shader_reg_maps reg_maps, struct wined3d_shader_signature_element input_signature,
448	struct wined3d_shader_signature_element output_signature, const DWORD byte_code, DWORD constf_size)
449	{
450	IWineD3DBaseShaderImpl shader = (IWineD3DBaseShaderImpl )iface;
451	unsigned int cur_loop_depth = 0, max_loop_depth = 0;
452	void *fe_data = shader->baseShader.frontend_data;
453	struct wined3d_shader_version shader_version;
454	const DWORD *ptr = byte_code;
455
456	memset(reg_maps, 0, sizeof(*reg_maps));
457
458	/* get_registers_used() is called on every compile on some 1.x shaders,
459	* which can result in stacking up a collection of local constants.
460	* Delete the old constants if existing. */
461	shader_delete_constant_list(&shader->baseShader.constantsF);
462	shader_delete_constant_list(&shader->baseShader.constantsB);
463	shader_delete_constant_list(&shader->baseShader.constantsI);
464
465	fe->shader_read_header(fe_data, &ptr, &shader_version);
466	reg_maps->shader_version = shader_version;
467
468	reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
469	sizeof(reg_maps->constf) ((constf_size + 31) / 32));
470	if (!reg_maps->constf)
471	{
472	ERR("Failed to allocate constant map memory.\n");
473	return E_OUTOFMEMORY;
474	}
475
476	while (!fe->shader_is_end(fe_data, &ptr))
477	{
478	struct wined3d_shader_instruction ins;
479	const char *comment;
480	UINT comment_size;
481	UINT param_size;
482
483	/* Skip comments. */
484	fe->shader_read_comment(&ptr, &comment, &comment_size);
485	if (comment) continue;
486
487	/* Fetch opcode. */
488	fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
489
490	/* Unhandled opcode, and its parameters. */
491	if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
492	{
493	TRACE("Skipping unrecognized instruction.\n");
494	ptr += param_size;
495	continue;
496	}
497
498	/* Handle declarations. */
499	if (ins.handler_idx == WINED3DSIH_DCL)
500	{
501	struct wined3d_shader_semantic semantic;
502
503	fe->shader_read_semantic(&ptr, &semantic);
504
505	switch (semantic.reg.reg.type)
506	{
507	/* Mark input registers used. */
508	case WINED3DSPR_INPUT:
509	reg_maps->input_registers \|= 1 << semantic.reg.reg.idx;
510	shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
511	break;
512
513	/* Vertex shader: mark 3.0 output registers used, save token. */
514	case WINED3DSPR_OUTPUT:
515	reg_maps->output_registers \|= 1 << semantic.reg.reg.idx;
516	shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
517	if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
518	break;
519
520	/* Save sampler usage token. */
521	case WINED3DSPR_SAMPLER:
522	reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
523	break;
524
525	default:
526	TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
527	break;
528	}
529	}
530	else if (ins.handler_idx == WINED3DSIH_DEF)
531	{
532	struct wined3d_shader_src_param rel_addr;
533	struct wined3d_shader_dst_param dst;
534
535	local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
536	if (!lconst) return E_OUTOFMEMORY;
537
538	fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
539	lconst->idx = dst.reg.idx;
540
541	memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
542	ptr += 4;
543
544	/* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
545	if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
546	{
547	float value = (float )lconst->value;
548	if (value[0] < -1.0f) value[0] = -1.0f;
549	else if (value[0] > 1.0f) value[0] = 1.0f;
550	if (value[1] < -1.0f) value[1] = -1.0f;
551	else if (value[1] > 1.0f) value[1] = 1.0f;
552	if (value[2] < -1.0f) value[2] = -1.0f;
553	else if (value[2] > 1.0f) value[2] = 1.0f;
554	if (value[3] < -1.0f) value[3] = -1.0f;
555	else if (value[3] > 1.0f) value[3] = 1.0f;
556	}
557
558	list_add_head(&shader->baseShader.constantsF, &lconst->entry);
559	}
560	else if (ins.handler_idx == WINED3DSIH_DEFI)
561	{
562	struct wined3d_shader_src_param rel_addr;
563	struct wined3d_shader_dst_param dst;
564
565	local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
566	if (!lconst) return E_OUTOFMEMORY;
567
568	fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
569	lconst->idx = dst.reg.idx;
570
571	memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
572	ptr += 4;
573
574	list_add_head(&shader->baseShader.constantsI, &lconst->entry);
575	reg_maps->local_int_consts \|= (1 << dst.reg.idx);
576	}
577	else if (ins.handler_idx == WINED3DSIH_DEFB)
578	{
579	struct wined3d_shader_src_param rel_addr;
580	struct wined3d_shader_dst_param dst;
581
582	local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
583	if (!lconst) return E_OUTOFMEMORY;
584
585	fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
586	lconst->idx = dst.reg.idx;
587
588	memcpy(lconst->value, ptr, sizeof(DWORD));
589	++ptr;
590
591	list_add_head(&shader->baseShader.constantsB, &lconst->entry);
592	reg_maps->local_bool_consts \|= (1 << dst.reg.idx);
593	}
594	/* If there's a loop in the shader. */
595	else if (ins.handler_idx == WINED3DSIH_LOOP
596	\|\| ins.handler_idx == WINED3DSIH_REP)
597	{
598	struct wined3d_shader_src_param src, rel_addr;
599
600	fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
601
602	/* Rep and Loop always use an integer constant for the control parameters. */
603	if (ins.handler_idx == WINED3DSIH_REP)
604	{
605	reg_maps->integer_constants \|= 1 << src.reg.idx;
606	}
607	else
608	{
609	fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
610	reg_maps->integer_constants \|= 1 << src.reg.idx;
611	}
612
613	cur_loop_depth++;
614	if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
615	}
616	else if (ins.handler_idx == WINED3DSIH_ENDLOOP
617	\|\| ins.handler_idx == WINED3DSIH_ENDREP)
618	{
619	cur_loop_depth--;
620	}
621	/* For subroutine prototypes. */
622	else if (ins.handler_idx == WINED3DSIH_LABEL)
623	{
624	struct wined3d_shader_src_param src, rel_addr;
625
626	fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
627	reg_maps->labels \|= 1 << src.reg.idx;
628	}
629	/* Set texture, address, temporary registers. */
630	else
631	{
632	BOOL color0_mov = FALSE;
633	int i, limit;
634
635	/* This will loop over all the registers and try to
636	* make a bitmask of the ones we're interested in.
637	*
638	* Relative addressing tokens are ignored, but that's
639	* okay, since we'll catch any address registers when
640	* they are initialized (required by spec). */
641	if (ins.dst_count)
642	{
643	struct wined3d_shader_src_param dst_rel_addr;
644	struct wined3d_shader_dst_param dst_param;
645
646	fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
647
648	shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
649
650	/* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
651	* is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
652	* shaders because TECRDOUT isn't used in them, but future register types might cause issues */
653	if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
654	&& dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
655	{
656	reg_maps->texcoord_mask[dst_param.reg.idx] \|= dst_param.write_mask;
657	}
658
659	if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
660	{
661	IWineD3DPixelShaderImpl ps = (IWineD3DPixelShaderImpl )shader;
662
663	if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
664	{
665	/* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
666	* COLOROUT 0. If we know this in advance, the ARB shader backend can skip
667	* the mov and perform the sRGB write correction from the source register.
668	*
669	* However, if the mov is only partial, we can't do this, and if the write
670	* comes from an instruction other than MOV it is hard to do as well. If
671	* COLOROUT 0 is overwritten partially later, the marker is dropped again. */
672
673	ps->color0_mov = FALSE;
674	if (ins.handler_idx == WINED3DSIH_MOV)
675	{
676	/* Used later when the source register is read. */
677	color0_mov = TRUE;
678	}
679	}
680	/* Also drop the MOV marker if the source register is overwritten prior to the shader
681	* end
682	*/
683	else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
684	{
685	ps->color0_mov = FALSE;
686	}
687	}
688
689	/* Declare 1.x samplers implicitly, based on the destination reg. number. */
690	if (shader_version.major == 1
691	&& (ins.handler_idx == WINED3DSIH_TEX
692	\|\| ins.handler_idx == WINED3DSIH_TEXBEM
693	\|\| ins.handler_idx == WINED3DSIH_TEXBEML
694	\|\| ins.handler_idx == WINED3DSIH_TEXDP3TEX
695	\|\| ins.handler_idx == WINED3DSIH_TEXM3x2TEX
696	\|\| ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
697	\|\| ins.handler_idx == WINED3DSIH_TEXM3x3TEX
698	\|\| ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
699	\|\| ins.handler_idx == WINED3DSIH_TEXREG2AR
700	\|\| ins.handler_idx == WINED3DSIH_TEXREG2GB
701	\|\| ins.handler_idx == WINED3DSIH_TEXREG2RGB))
702	{
703	/* Fake sampler usage, only set reserved bit and type. */
704	DWORD sampler_code = dst_param.reg.idx;
705
706	TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
707	reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
708
709	/* texbem is only valid with < 1.4 pixel shaders */
710	if (ins.handler_idx == WINED3DSIH_TEXBEM
711	\|\| ins.handler_idx == WINED3DSIH_TEXBEML)
712	{
713	reg_maps->bumpmat \|= 1 << dst_param.reg.idx;
714	if (ins.handler_idx == WINED3DSIH_TEXBEML)
715	{
716	reg_maps->luminanceparams \|= 1 << dst_param.reg.idx;
717	}
718	}
719	}
720	else if (ins.handler_idx == WINED3DSIH_BEM)
721	{
722	reg_maps->bumpmat \|= 1 << dst_param.reg.idx;
723	}
724	}
725
726	if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
727	else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
728	else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
729	else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
730	else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
731	else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
732	else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
733	else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
734
735	limit = ins.src_count + (ins.predicate ? 1 : 0);
736	for (i = 0; i < limit; ++i)
737	{
738	struct wined3d_shader_src_param src_param, src_rel_addr;
739	unsigned int count;
740
741	fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
742	count = get_instr_extra_regcount(ins.handler_idx, i);
743
744	shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
745	while (count)
746	{
747	++src_param.reg.idx;
748	shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
749	--count;
750	}
751
752	if (color0_mov)
753	{
754	IWineD3DPixelShaderImpl ps = (IWineD3DPixelShaderImpl )shader;
755	if (src_param.reg.type == WINED3DSPR_TEMP
756	&& src_param.swizzle == WINED3DSP_NOSWIZZLE)
757	{
758	ps->color0_mov = TRUE;
759	ps->color0_reg = src_param.reg.idx;
760	}
761	}
762	}
763	}
764	}
765	reg_maps->loop_depth = max_loop_depth;
766
767	shader->baseShader.functionLength = ((const char )ptr - (const char )byte_code);
768
769	return WINED3D_OK;
770	}
771
772	unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
773	{
774	DWORD map = 1 << max;
775	map \|= map - 1;
776	map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
777
778	return wined3d_log2i(map);
779	}
780
781	static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
782	const struct wined3d_shader_version *shader_version)
783	{
784	TRACE("dcl");
785
786	if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
787	{
788	switch (semantic->sampler_type)
789	{
790	case WINED3DSTT_2D: TRACE("_2d"); break;
791	case WINED3DSTT_CUBE: TRACE("_cube"); break;
792	case WINED3DSTT_VOLUME: TRACE("_volume"); break;
793	default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
794	}
795	}
796	else
797	{
798	/* Pixel shaders 3.0 don't have usage semantics. */
799	if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
800	else TRACE("_");
801
802	switch (semantic->usage)
803	{
804	case WINED3DDECLUSAGE_POSITION:
805	TRACE("position%u", semantic->usage_idx);
806	break;
807
808	case WINED3DDECLUSAGE_BLENDINDICES:
809	TRACE("blend");
810	break;
811
812	case WINED3DDECLUSAGE_BLENDWEIGHT:
813	TRACE("weight");
814	break;
815
816	case WINED3DDECLUSAGE_NORMAL:
817	TRACE("normal%u", semantic->usage_idx);
818	break;
819
820	case WINED3DDECLUSAGE_PSIZE:
821	TRACE("psize");
822	break;
823
824	case WINED3DDECLUSAGE_COLOR:
825	if (semantic->usage_idx == 0) TRACE("color");
826	else TRACE("specular%u", (semantic->usage_idx - 1));
827	break;
828
829	case WINED3DDECLUSAGE_TEXCOORD:
830	TRACE("texture%u", semantic->usage_idx);
831	break;
832
833	case WINED3DDECLUSAGE_TANGENT:
834	TRACE("tangent");
835	break;
836
837	case WINED3DDECLUSAGE_BINORMAL:
838	TRACE("binormal");
839	break;
840
841	case WINED3DDECLUSAGE_TESSFACTOR:
842	TRACE("tessfactor");
843	break;
844
845	case WINED3DDECLUSAGE_POSITIONT:
846	TRACE("positionT%u", semantic->usage_idx);
847	break;
848
849	case WINED3DDECLUSAGE_FOG:
850	TRACE("fog");
851	break;
852
853	case WINED3DDECLUSAGE_DEPTH:
854	TRACE("depth");
855	break;
856
857	case WINED3DDECLUSAGE_SAMPLE:
858	TRACE("sample");
859	break;
860
861	default:
862	FIXME("unknown_semantics(0x%08x)", semantic->usage);
863	}
864	}
865	}
866
867	static void shader_dump_register(const struct wined3d_shader_register *reg,
868	const struct wined3d_shader_version *shader_version)
869	{
870	static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
871	static const char * const misctype_reg_names[] = {"vPos", "vFace"};
872	UINT offset = reg->idx;
873
874	switch (reg->type)
875	{
876	case WINED3DSPR_TEMP:
877	TRACE("r");
878	break;
879
880	case WINED3DSPR_INPUT:
881	TRACE("v");
882	break;
883
884	case WINED3DSPR_CONST:
885	case WINED3DSPR_CONST2:
886	case WINED3DSPR_CONST3:
887	case WINED3DSPR_CONST4:
888	TRACE("c");
889	offset = shader_get_float_offset(reg->type, reg->idx);
890	break;
891
892	case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
893	TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
894	break;
895
896	case WINED3DSPR_RASTOUT:
897	TRACE("%s", rastout_reg_names[reg->idx]);
898	break;
899
900	case WINED3DSPR_COLOROUT:
901	TRACE("oC");
902	break;
903
904	case WINED3DSPR_DEPTHOUT:
905	TRACE("oDepth");
906	break;
907
908	case WINED3DSPR_ATTROUT:
909	TRACE("oD");
910	break;
911
912	case WINED3DSPR_TEXCRDOUT:
913	/* Vertex shaders >= 3.0 use general purpose output registers
914	* (WINED3DSPR_OUTPUT), which can include an address token. */
915	if (shader_version->major >= 3) TRACE("o");
916	else TRACE("oT");
917	break;
918
919	case WINED3DSPR_CONSTINT:
920	TRACE("i");
921	break;
922
923	case WINED3DSPR_CONSTBOOL:
924	TRACE("b");
925	break;
926
927	case WINED3DSPR_LABEL:
928	TRACE("l");
929	break;
930
931	case WINED3DSPR_LOOP:
932	TRACE("aL");
933	break;
934
935	case WINED3DSPR_SAMPLER:
936	TRACE("s");
937	break;
938
939	case WINED3DSPR_MISCTYPE:
940	if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
941	else TRACE("%s", misctype_reg_names[reg->idx]);
942	break;
943
944	case WINED3DSPR_PREDICATE:
945	TRACE("p");
946	break;
947
948	case WINED3DSPR_IMMCONST:
949	TRACE("l");
950	break;
951
952	case WINED3DSPR_CONSTBUFFER:
953	TRACE("cb");
954	break;
955
956	default:
957	TRACE("unhandled_rtype(%#x)", reg->type);
958	break;
959	}
960
961	if (reg->type == WINED3DSPR_IMMCONST)
962	{
963	TRACE("(");
964	switch (reg->immconst_type)
965	{
966	case WINED3D_IMMCONST_FLOAT:
967	TRACE("%.8e", (const float )reg->immconst_data);
968	break;
969
970	case WINED3D_IMMCONST_FLOAT4:
971	TRACE("%.8e, %.8e, %.8e, %.8e",
972	(const float )&reg->immconst_data[0], (const float )&reg->immconst_data[1],
973	(const float )&reg->immconst_data[2], (const float )&reg->immconst_data[3]);
974	break;
975
976	default:
977	TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
978	break;
979	}
980	TRACE(")");
981	}
982	else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
983	{
984	if (reg->array_idx != ~0U)
985	{
986	TRACE("%u[%u", offset, reg->array_idx);
987	if (reg->rel_addr)
988	{
989	TRACE(" + ");
990	shader_dump_src_param(reg->rel_addr, shader_version);
991	}
992	TRACE("]");
993	}
994	else
995	{
996	if (reg->rel_addr)
997	{
998	TRACE("[");
999	shader_dump_src_param(reg->rel_addr, shader_version);
1000	TRACE(" + ");
1001	}
1002	TRACE("%u", offset);
1003	if (reg->rel_addr) TRACE("]");
1004	}
1005	}
1006	}
1007
1008	void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1009	const struct wined3d_shader_version *shader_version)
1010	{
1011	DWORD write_mask = param->write_mask;
1012
1013	shader_dump_register(&param->reg, shader_version);
1014
1015	if (write_mask != WINED3DSP_WRITEMASK_ALL)
1016	{
1017	static const char *write_mask_chars = "xyzw";
1018
1019	TRACE(".");
1020	if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1021	if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1022	if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1023	if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1024	}
1025	}
1026
1027	void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1028	const struct wined3d_shader_version *shader_version)
1029	{
1030	DWORD src_modifier = param->modifiers;
1031	DWORD swizzle = param->swizzle;
1032
1033	if (src_modifier == WINED3DSPSM_NEG
1034	\|\| src_modifier == WINED3DSPSM_BIASNEG
1035	\|\| src_modifier == WINED3DSPSM_SIGNNEG
1036	\|\| src_modifier == WINED3DSPSM_X2NEG
1037	\|\| src_modifier == WINED3DSPSM_ABSNEG)
1038	TRACE("-");
1039	else if (src_modifier == WINED3DSPSM_COMP)
1040	TRACE("1-");
1041	else if (src_modifier == WINED3DSPSM_NOT)
1042	TRACE("!");
1043
1044	if (src_modifier == WINED3DSPSM_ABS \|\| src_modifier == WINED3DSPSM_ABSNEG)
1045	TRACE("abs(");
1046
1047	shader_dump_register(&param->reg, shader_version);
1048
1049	if (src_modifier)
1050	{
1051	switch (src_modifier)
1052	{
1053	case WINED3DSPSM_NONE: break;
1054	case WINED3DSPSM_NEG: break;
1055	case WINED3DSPSM_NOT: break;
1056	case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1057	case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1058	case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1059	case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1060	case WINED3DSPSM_COMP: break;
1061	case WINED3DSPSM_X2: TRACE("_x2"); break;
1062	case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1063	case WINED3DSPSM_DZ: TRACE("_dz"); break;
1064	case WINED3DSPSM_DW: TRACE("_dw"); break;
1065	case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1066	case WINED3DSPSM_ABS: TRACE(")"); break;
1067	default: TRACE("_unknown_modifier(%#x)", src_modifier);
1068	}
1069	}
1070
1071	if (swizzle != WINED3DSP_NOSWIZZLE)
1072	{
1073	static const char *swizzle_chars = "xyzw";
1074	DWORD swizzle_x = swizzle & 0x03;
1075	DWORD swizzle_y = (swizzle >> 2) & 0x03;
1076	DWORD swizzle_z = (swizzle >> 4) & 0x03;
1077	DWORD swizzle_w = (swizzle >> 6) & 0x03;
1078
1079	if (swizzle_x == swizzle_y
1080	&& swizzle_x == swizzle_z
1081	&& swizzle_x == swizzle_w)
1082	{
1083	TRACE(".%c", swizzle_chars[swizzle_x]);
1084	}
1085	else
1086	{
1087	TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1088	swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1089	}
1090	}
1091	}
1092
1093	/* Shared code in order to generate the bulk of the shader string.
1094	* NOTE: A description of how to parse tokens can be found on MSDN. */
1095	void shader_generate_main(IWineD3DBaseShader iface, struct wined3d_shader_buffer buffer,
1096	const shader_reg_maps reg_maps, const DWORD byte_code, void *backend_ctx)
1097	{
1098	IWineD3DBaseShaderImpl shader = (IWineD3DBaseShaderImpl )iface;
1099	IWineD3DDeviceImpl device = (IWineD3DDeviceImpl )shader->baseShader.device;
1100	const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1101	void *fe_data = shader->baseShader.frontend_data;
1102	struct wined3d_shader_src_param src_rel_addr[4];
1103	struct wined3d_shader_src_param src_param[4];
1104	struct wined3d_shader_version shader_version;
1105	struct wined3d_shader_src_param dst_rel_addr;
1106	struct wined3d_shader_dst_param dst_param;
1107	struct wined3d_shader_instruction ins;
1108	struct wined3d_shader_context ctx;
1109	const DWORD *ptr = byte_code;
1110	DWORD i;
1111
1112	/* Initialize current parsing state. */
1113	ctx.shader = iface;
1114	ctx.gl_info = &device->adapter->gl_info;
1115	ctx.reg_maps = reg_maps;
1116	ctx.buffer = buffer;
1117	ctx.backend_data = backend_ctx;
1118
1119	ins.ctx = &ctx;
1120	ins.dst = &dst_param;
1121	ins.src = src_param;
1122	shader->baseShader.parse_state.current_row = 0;
1123
1124	fe->shader_read_header(fe_data, &ptr, &shader_version);
1125
1126	while (!fe->shader_is_end(fe_data, &ptr))
1127	{
1128	const char *comment;
1129	UINT comment_size;
1130	UINT param_size;
1131
1132	/* Skip comment tokens. */
1133	fe->shader_read_comment(&ptr, &comment, &comment_size);
1134	if (comment) continue;
1135
1136	/* Read opcode. */
1137	fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1138
1139	/* Unknown opcode and its parameters. */
1140	if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1141	{
1142	TRACE("Skipping unrecognized instruction.\n");
1143	ptr += param_size;
1144	continue;
1145	}
1146
1147	/* Nothing to do. */
1148	if (ins.handler_idx == WINED3DSIH_DCL
1149	\|\| ins.handler_idx == WINED3DSIH_NOP
1150	\|\| ins.handler_idx == WINED3DSIH_DEF
1151	\|\| ins.handler_idx == WINED3DSIH_DEFI
1152	\|\| ins.handler_idx == WINED3DSIH_DEFB
1153	\|\| ins.handler_idx == WINED3DSIH_PHASE)
1154	{
1155	ptr += param_size;
1156	continue;
1157	}
1158
1159	/* Destination token */
1160	if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1161
1162	/* Predication token */
1163	if (ins.predicate)
1164	{
1165	FIXME("Predicates not implemented.\n");
1166	ins.predicate = *ptr++;
1167	}
1168
1169	/* Other source tokens */
1170	for (i = 0; i < ins.src_count; ++i)
1171	{
1172	fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1173	}
1174
1175	/* Call appropriate function for output target */
1176	device->shader_backend->shader_handle_instruction(&ins);
1177	}
1178	}
1179
1180	static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1181	{
1182	DWORD mmask = dst->modifiers;
1183
1184	switch (dst->shift)
1185	{
1186	case 0: break;
1187	case 13: TRACE("_d8"); break;
1188	case 14: TRACE("_d4"); break;
1189	case 15: TRACE("_d2"); break;
1190	case 1: TRACE("_x2"); break;
1191	case 2: TRACE("_x4"); break;
1192	case 3: TRACE("_x8"); break;
1193	default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1194	}
1195
1196	if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1197	if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1198	if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1199
1200	mmask &= ~(WINED3DSPDM_SATURATE \| WINED3DSPDM_PARTIALPRECISION \| WINED3DSPDM_MSAMPCENTROID);
1201	if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1202	}
1203
1204	static void shader_trace_init(const struct wined3d_shader_frontend fe, void fe_data, const DWORD *byte_code)
1205	{
1206	struct wined3d_shader_version shader_version;
1207	const DWORD *ptr = byte_code;
1208	const char *type_prefix;
1209	DWORD i;
1210
1211	TRACE("Parsing %p.\n", byte_code);
1212
1213	fe->shader_read_header(fe_data, &ptr, &shader_version);
1214
1215	switch (shader_version.type)
1216	{
1217	case WINED3D_SHADER_TYPE_VERTEX:
1218	type_prefix = "vs";
1219	break;
1220
1221	case WINED3D_SHADER_TYPE_GEOMETRY:
1222	type_prefix = "gs";
1223	break;
1224
1225	case WINED3D_SHADER_TYPE_PIXEL:
1226	type_prefix = "ps";
1227	break;
1228
1229	default:
1230	FIXME("Unhandled shader type %#x.\n", shader_version.type);
1231	type_prefix = "unknown";
1232	break;
1233	}
1234
1235	TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1236
1237	while (!fe->shader_is_end(fe_data, &ptr))
1238	{
1239	struct wined3d_shader_instruction ins;
1240	const char *comment;
1241	UINT comment_size;
1242	UINT param_size;
1243
1244	/* comment */
1245	fe->shader_read_comment(&ptr, &comment, &comment_size);
1246	if (comment)
1247	{
1248	if (comment_size > 4 && (const DWORD )comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1249	{
1250	const char *end = comment + comment_size;
1251	const char *ptr = comment + 4;
1252	const char *line = ptr;
1253
1254	TRACE("// TEXT\n");
1255	while (ptr != end)
1256	{
1257	if (*ptr == '\n')
1258	{
1259	UINT len = ptr - line;
1260	if (len && *(ptr - 1) == '\r') --len;
1261	TRACE("// %s\n", debugstr_an(line, len));
1262	line = ++ptr;
1263	}
1264	else ++ptr;
1265	}
1266	if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1267	}
1268	else TRACE("// %s\n", debugstr_an(comment, comment_size));
1269	continue;
1270	}
1271
1272	fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1273	if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1274	{
1275	TRACE("Skipping unrecognized instruction.\n");
1276	ptr += param_size;
1277	continue;
1278	}
1279
1280	if (ins.handler_idx == WINED3DSIH_DCL)
1281	{
1282	struct wined3d_shader_semantic semantic;
1283
1284	fe->shader_read_semantic(&ptr, &semantic);
1285
1286	shader_dump_decl_usage(&semantic, &shader_version);
1287	shader_dump_ins_modifiers(&semantic.reg);
1288	TRACE(" ");
1289	shader_dump_dst_param(&semantic.reg, &shader_version);
1290	}
1291	else if (ins.handler_idx == WINED3DSIH_DEF)
1292	{
1293	struct wined3d_shader_src_param rel_addr;
1294	struct wined3d_shader_dst_param dst;
1295
1296	fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1297
1298	TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1299	(const float )(ptr),
1300	(const float )(ptr + 1),
1301	(const float )(ptr + 2),
1302	(const float )(ptr + 3));
1303	ptr += 4;
1304	}
1305	else if (ins.handler_idx == WINED3DSIH_DEFI)
1306	{
1307	struct wined3d_shader_src_param rel_addr;
1308	struct wined3d_shader_dst_param dst;
1309
1310	fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1311
1312	TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1313	*(ptr),
1314	*(ptr + 1),
1315	*(ptr + 2),
1316	*(ptr + 3));
1317	ptr += 4;
1318	}
1319	else if (ins.handler_idx == WINED3DSIH_DEFB)
1320	{
1321	struct wined3d_shader_src_param rel_addr;
1322	struct wined3d_shader_dst_param dst;
1323
1324	fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1325
1326	TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1327	++ptr;
1328	}
1329	else
1330	{
1331	struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1332	struct wined3d_shader_dst_param dst_param;
1333	struct wined3d_shader_src_param src_param;
1334
1335	if (ins.dst_count)
1336	{
1337	fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1338	}
1339
1340	/* Print out predication source token first - it follows
1341	* the destination token. */
1342	if (ins.predicate)
1343	{
1344	fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1345	TRACE("(");
1346	shader_dump_src_param(&src_param, &shader_version);
1347	TRACE(") ");
1348	}
1349
1350	/* PixWin marks instructions with the coissue flag with a '+' */
1351	if (ins.coissue) TRACE("+");
1352
1353	TRACE("%s", shader_opcode_names[ins.handler_idx]);
1354
1355	if (ins.handler_idx == WINED3DSIH_IFC
1356	\|\| ins.handler_idx == WINED3DSIH_BREAKC)
1357	{
1358	switch (ins.flags)
1359	{
1360	case COMPARISON_GT: TRACE("_gt"); break;
1361	case COMPARISON_EQ: TRACE("_eq"); break;
1362	case COMPARISON_GE: TRACE("_ge"); break;
1363	case COMPARISON_LT: TRACE("_lt"); break;
1364	case COMPARISON_NE: TRACE("_ne"); break;
1365	case COMPARISON_LE: TRACE("_le"); break;
1366	default: TRACE("_(%u)", ins.flags);
1367	}
1368	}
1369	else if (ins.handler_idx == WINED3DSIH_TEX
1370	&& shader_version.major >= 2
1371	&& (ins.flags & WINED3DSI_TEXLD_PROJECT))
1372	{
1373	TRACE("p");
1374	}
1375
1376	/* We already read the destination token, print it. */
1377	if (ins.dst_count)
1378	{
1379	shader_dump_ins_modifiers(&dst_param);
1380	TRACE(" ");
1381	shader_dump_dst_param(&dst_param, &shader_version);
1382	}
1383
1384	/* Other source tokens */
1385	for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1386	{
1387	fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1388	TRACE(!i ? " " : ", ");
1389	shader_dump_src_param(&src_param, &shader_version);
1390	}
1391	}
1392	TRACE("\n");
1393	}
1394	}
1395
1396	static void shader_cleanup(IWineD3DBaseShader *iface)
1397	{
1398	IWineD3DBaseShaderImpl shader = (IWineD3DBaseShaderImpl )iface;
1399
1400	((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1401	HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1402	HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1403	shader_delete_constant_list(&shader->baseShader.constantsF);
1404	shader_delete_constant_list(&shader->baseShader.constantsB);
1405	shader_delete_constant_list(&shader->baseShader.constantsI);
1406	list_remove(&shader->baseShader.shader_list_entry);
1407
1408	if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1409	{
1410	shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1411	}
1412	}
1413
1414	static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1415	static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1416	static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1417	static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1418	static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1419	static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1420	static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1421	static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1422	static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1423	static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1424	static void shader_none_free(IWineD3DDevice *iface) {}
1425	static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1426
1427	static void shader_none_get_caps(const struct wined3d_gl_info gl_info, struct shader_caps caps)
1428	{
1429	/* Set the shader caps to 0 for the none shader backend */
1430	caps->VertexShaderVersion = 0;
1431	caps->PixelShaderVersion = 0;
1432	caps->PixelShader1xMaxValue = 0.0f;
1433	}
1434
1435	static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1436	{
1437	if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1438	{
1439	TRACE("Checking support for fixup:\n");
1440	dump_color_fixup_desc(fixup);
1441	}
1442
1443	/* Faked to make some apps happy. */
1444	if (!is_complex_fixup(fixup))
1445	{
1446	TRACE("[OK]\n");
1447	return TRUE;
1448	}
1449
1450	TRACE("[FAILED]\n");
1451	return FALSE;
1452	}
1453
1454	const shader_backend_t none_shader_backend = {
1455	shader_none_handle_instruction,
1456	shader_none_select,
1457	shader_none_select_depth_blt,
1458	shader_none_deselect_depth_blt,
1459	shader_none_update_float_vertex_constants,
1460	shader_none_update_float_pixel_constants,
1461	shader_none_load_constants,
1462	shader_none_load_np2fixup_constants,
1463	shader_none_destroy,
1464	shader_none_alloc,
1465	shader_none_free,
1466	shader_none_dirty_const,
1467	shader_none_get_caps,
1468	shader_none_color_fixup_supported,
1469	};
1470
1471	static void shader_get_parent(IWineD3DBaseShaderImpl shader, IUnknown *parent)
1472	{
1473	*parent = shader->baseShader.parent;
1474	IUnknown_AddRef(*parent);
1475	TRACE("shader %p, returning %p.\n", shader, *parent);
1476	}
1477
1478	static HRESULT shader_get_function(IWineD3DBaseShaderImpl shader, void data, UINT *data_size)
1479	{
1480	if (!data)
1481	{
1482	*data_size = shader->baseShader.functionLength;
1483	return WINED3D_OK;
1484	}
1485
1486	if (*data_size < shader->baseShader.functionLength)
1487	{
1488	/* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1489	* than the required size we should write the required size and
1490	* return D3DERR_MOREDATA. That's not actually true. */
1491	return WINED3DERR_INVALIDCALL;
1492	}
1493
1494	memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1495
1496	return WINED3D_OK;
1497	}
1498
1499	static HRESULT shader_set_function(IWineD3DBaseShaderImpl shader, const DWORD byte_code,
1500	const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1501	{
1502	struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1503	const struct wined3d_shader_frontend *fe;
1504	HRESULT hr;
1505
1506	TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1507	shader, byte_code, output_signature, float_const_count);
1508
1509	fe = shader_select_frontend(*byte_code);
1510	if (!fe)
1511	{
1512	FIXME("Unable to find frontend for shader.\n");
1513	return WINED3DERR_INVALIDCALL;
1514	}
1515	shader->baseShader.frontend = fe;
1516	shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1517	if (!shader->baseShader.frontend_data)
1518	{
1519	FIXME("Failed to initialize frontend.\n");
1520	return WINED3DERR_INVALIDCALL;
1521	}
1522
1523	/* First pass: trace shader. */
1524	if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1525
1526	/* Initialize immediate constant lists. */
1527	list_init(&shader->baseShader.constantsF);
1528	list_init(&shader->baseShader.constantsB);
1529	list_init(&shader->baseShader.constantsI);
1530
1531	/* Second pass: figure out which registers are used, what the semantics are, etc. */
1532	hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1533	reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1534	byte_code, float_const_count);
1535	if (FAILED(hr)) return hr;
1536
1537	shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1538	if (!shader->baseShader.function) return E_OUTOFMEMORY;
1539	memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1540
1541	return WINED3D_OK;
1542	}
1543
1544	static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader iface, REFIID riid, void *object)
1545	{
1546	TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1547
1548	if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1549	\|\| IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1550	\|\| IsEqualGUID(riid, &IID_IWineD3DBase)
1551	\|\| IsEqualGUID(riid, &IID_IUnknown))
1552	{
1553	IUnknown_AddRef(iface);
1554	*object = iface;
1555	return S_OK;
1556	}
1557
1558	WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1559
1560	*object = NULL;
1561	return E_NOINTERFACE;
1562	}
1563
1564	static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1565	{
1566	IWineD3DVertexShaderImpl shader = (IWineD3DVertexShaderImpl )iface;
1567	ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1568
1569	TRACE("%p increasing refcount to %u.\n", shader, refcount);
1570
1571	return refcount;
1572	}
1573
1574	static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1575	{
1576	IWineD3DVertexShaderImpl shader = (IWineD3DVertexShaderImpl )iface;
1577	ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1578
1579	TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1580
1581	if (!refcount)
1582	{
1583	shader_cleanup((IWineD3DBaseShader *)iface);
1584	shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1585	HeapFree(GetProcessHeap(), 0, shader);
1586	}
1587
1588	return refcount;
1589	}
1590
1591	static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader iface, IUnknown *parent)
1592	{
1593	TRACE("iface %p, parent %p.\n", iface, parent);
1594
1595	shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1596
1597	return WINED3D_OK;
1598	}
1599
1600	static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader iface, void data, UINT *data_size)
1601	{
1602	TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1603
1604	return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1605	}
1606
1607	/* Set local constants for d3d8 shaders. */
1608	static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1609	UINT start_idx, const float *src_data, UINT count)
1610	{
1611	IWineD3DVertexShaderImpl shader =(IWineD3DVertexShaderImpl )iface;
1612	IWineD3DDeviceImpl device = (IWineD3DDeviceImpl )shader->baseShader.device;
1613	UINT i, end_idx;
1614
1615	TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1616
1617	end_idx = start_idx + count;
1618	if (end_idx > device->d3d_vshader_constantF)
1619	{
1620	WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1621	end_idx = device->d3d_vshader_constantF;
1622	}
1623
1624	for (i = start_idx; i < end_idx; ++i)
1625	{
1626	local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1627	if (!lconst) return E_OUTOFMEMORY;
1628
1629	lconst->idx = i;
1630	memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components /, 4 sizeof(float));
1631	list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1632	}
1633
1634	return WINED3D_OK;
1635	}
1636
1637	static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1638	{
1639	/* IUnknown methods */
1640	vertexshader_QueryInterface,
1641	vertexshader_AddRef,
1642	vertexshader_Release,
1643	/* IWineD3DBase methods */
1644	vertexshader_GetParent,
1645	/* IWineD3DBaseShader methods */
1646	vertexshader_GetFunction,
1647	/* IWineD3DVertexShader methods */
1648	vertexshader_SetLocalConstantsF,
1649	};
1650
1651	void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1652	IWineD3DStateBlockImpl stateblock, struct vs_compile_args args)
1653	{
1654	args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1655	args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1656	&& stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1657	args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1658	}
1659
1660	static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1661	{
1662	if (usage_idx1 != usage_idx2) return FALSE;
1663	if (usage1 == usage2) return TRUE;
1664	if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1665	if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1666
1667	return FALSE;
1668	}
1669
1670	BOOL vshader_get_input(IWineD3DVertexShader iface, BYTE usage_req, BYTE usage_idx_req, unsigned int regnum)
1671	{
1672	IWineD3DVertexShaderImpl shader = (IWineD3DVertexShaderImpl )iface;
1673	WORD map = shader->baseShader.reg_maps.input_registers;
1674	unsigned int i;
1675
1676	for (i = 0; map; map >>= 1, ++i)
1677	{
1678	if (!(map & 1)) continue;
1679
1680	if (match_usage(shader->attributes[i].usage,
1681	shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1682	{
1683	*regnum = i;
1684	return TRUE;
1685	}
1686	}
1687	return FALSE;
1688	}
1689
1690	static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1691	{
1692	DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1693	shader->baseShader.reg_maps.shader_version.minor);
1694	IWineD3DDeviceImpl device = (IWineD3DDeviceImpl )shader->baseShader.device;
1695
1696	shader->baseShader.limits.texcoord = 0;
1697	shader->baseShader.limits.attributes = 16;
1698	shader->baseShader.limits.packed_input = 0;
1699
1700	switch (shader_version)
1701	{
1702	case WINED3D_SHADER_VERSION(1, 0):
1703	case WINED3D_SHADER_VERSION(1, 1):
1704	shader->baseShader.limits.temporary = 12;
1705	shader->baseShader.limits.constant_bool = 0;
1706	shader->baseShader.limits.constant_int = 0;
1707	shader->baseShader.limits.address = 1;
1708	shader->baseShader.limits.packed_output = 0;
1709	shader->baseShader.limits.sampler = 0;
1710	shader->baseShader.limits.label = 0;
1711	/* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1712	* a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1713	* constants? */
1714	shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1715	break;
1716
1717	case WINED3D_SHADER_VERSION(2, 0):
1718	case WINED3D_SHADER_VERSION(2, 1):
1719	shader->baseShader.limits.temporary = 12;
1720	shader->baseShader.limits.constant_bool = 16;
1721	shader->baseShader.limits.constant_int = 16;
1722	shader->baseShader.limits.address = 1;
1723	shader->baseShader.limits.packed_output = 0;
1724	shader->baseShader.limits.sampler = 0;
1725	shader->baseShader.limits.label = 16;
1726	shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1727	break;
1728
1729	case WINED3D_SHADER_VERSION(4, 0):
1730	FIXME("Using 3.0 limits for 4.0 shader.\n");
1731	/* Fall through. */
1732
1733	case WINED3D_SHADER_VERSION(3, 0):
1734	shader->baseShader.limits.temporary = 32;
1735	shader->baseShader.limits.constant_bool = 32;
1736	shader->baseShader.limits.constant_int = 32;
1737	shader->baseShader.limits.address = 1;
1738	shader->baseShader.limits.packed_output = 12;
1739	shader->baseShader.limits.sampler = 4;
1740	shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1741	/* DX10 cards on Windows advertise a d3d9 constant limit of 256
1742	* even though they are capable of supporting much more (GL
1743	* drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1744	* wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1745	* shaders to 256. */
1746	shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1747	break;
1748
1749	default:
1750	shader->baseShader.limits.temporary = 12;
1751	shader->baseShader.limits.constant_bool = 16;
1752	shader->baseShader.limits.constant_int = 16;
1753	shader->baseShader.limits.address = 1;
1754	shader->baseShader.limits.packed_output = 0;
1755	shader->baseShader.limits.sampler = 0;
1756	shader->baseShader.limits.label = 16;
1757	shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1758	FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1759	shader->baseShader.reg_maps.shader_version.major,
1760	shader->baseShader.reg_maps.shader_version.minor);
1761	}
1762	}
1763
1764	HRESULT vertexshader_init(IWineD3DVertexShaderImpl shader, IWineD3DDeviceImpl device,
1765	const DWORD byte_code, const struct wined3d_shader_signature output_signature,
1766	IUnknown parent, const struct wined3d_parent_ops parent_ops)
1767	{
1768	const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1769	struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1770	unsigned int i;
1771	HRESULT hr;
1772	WORD map;
1773
1774	if (!byte_code) return WINED3DERR_INVALIDCALL;
1775
1776	shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1777	shader_init(&shader->baseShader, device, parent, parent_ops);
1778
1779	hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1780	output_signature, device->d3d_vshader_constantF);
1781	if (FAILED(hr))
1782	{
1783	WARN("Failed to set function, hr %#x.\n", hr);
1784	shader_cleanup((IWineD3DBaseShader *)shader);
1785	return hr;
1786	}
1787
1788	map = shader->baseShader.reg_maps.input_registers;
1789	for (i = 0; map; map >>= 1, ++i)
1790	{
1791	if (!(map & 1) \|\| !shader->baseShader.input_signature[i].semantic_name) continue;
1792
1793	shader->attributes[i].usage =
1794	shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1795	shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1796	}
1797
1798	if (output_signature)
1799	{
1800	for (i = 0; i < output_signature->element_count; ++i)
1801	{
1802	struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1803	reg_maps->output_registers \|= 1 << e->register_idx;
1804	shader->baseShader.output_signature[e->register_idx] = *e;
1805	}
1806	}
1807
1808	vertexshader_set_limits(shader);
1809
1810	if (device->vs_selected_mode == SHADER_ARB
1811	&& (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1812	&& shader->min_rel_offset <= shader->max_rel_offset)
1813	{
1814	if (shader->max_rel_offset - shader->min_rel_offset > 127)
1815	{
1816	FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1817	FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1818	FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1819	}
1820	else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1821	{
1822	shader->rel_offset = shader->min_rel_offset + 63;
1823	}
1824	else if (shader->max_rel_offset > 63)
1825	{
1826	shader->rel_offset = shader->min_rel_offset;
1827	}
1828	else
1829	{
1830	shader->rel_offset = 0;
1831	}
1832	}
1833
1834	shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1835	&& !list_empty(&shader->baseShader.constantsF);
1836
1837	return WINED3D_OK;
1838	}
1839
1840	static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1841	REFIID riid, void **object)
1842	{
1843	TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1844
1845	if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1846	\|\| IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1847	\|\| IsEqualGUID(riid, &IID_IWineD3DBase)
1848	\|\| IsEqualGUID(riid, &IID_IUnknown))
1849	{
1850	IUnknown_AddRef(iface);
1851	*object = iface;
1852	return S_OK;
1853	}
1854
1855	WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1856
1857	*object = NULL;
1858	return E_NOINTERFACE;
1859	}
1860
1861	static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1862	{
1863	struct wined3d_geometryshader shader = (struct wined3d_geometryshader )iface;
1864	ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1865
1866	TRACE("%p increasing refcount to %u.\n", shader, refcount);
1867
1868	return refcount;
1869	}
1870
1871	static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1872	{
1873	struct wined3d_geometryshader shader = (struct wined3d_geometryshader )iface;
1874	ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1875
1876	TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1877
1878	if (!refcount)
1879	{
1880	shader_cleanup((IWineD3DBaseShader *)iface);
1881	shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1882	HeapFree(GetProcessHeap(), 0, shader);
1883	}
1884
1885	return refcount;
1886	}
1887
1888	static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader iface, IUnknown *parent)
1889	{
1890	TRACE("iface %p, parent %p.\n", iface, parent);
1891
1892	shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1893
1894	return WINED3D_OK;
1895	}
1896
1897	static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader iface, void data, UINT *data_size)
1898	{
1899	TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1900
1901	return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1902	}
1903
1904	static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1905	{
1906	/* IUnknown methods */
1907	geometryshader_QueryInterface,
1908	geometryshader_AddRef,
1909	geometryshader_Release,
1910	/* IWineD3DBase methods */
1911	geometryshader_GetParent,
1912	/* IWineD3DBaseShader methods */
1913	geometryshader_GetFunction,
1914	};
1915
1916	HRESULT geometryshader_init(struct wined3d_geometryshader shader, IWineD3DDeviceImpl device,
1917	const DWORD byte_code, const struct wined3d_shader_signature output_signature,
1918	IUnknown parent, const struct wined3d_parent_ops parent_ops)
1919	{
1920	HRESULT hr;
1921
1922	shader->vtbl = &wined3d_geometryshader_vtbl;
1923	shader_init(&shader->base_shader, device, parent, parent_ops);
1924
1925	hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1926	if (FAILED(hr))
1927	{
1928	WARN("Failed to set function, hr %#x.\n", hr);
1929	shader_cleanup((IWineD3DBaseShader *)shader);
1930	return hr;
1931	}
1932
1933	shader->base_shader.load_local_constsF = FALSE;
1934
1935	return WINED3D_OK;
1936	}
1937
1938	static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader iface, REFIID riid, void *object)
1939	{
1940	TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1941
1942	if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1943	\|\| IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1944	\|\| IsEqualGUID(riid, &IID_IWineD3DBase)
1945	\|\| IsEqualGUID(riid, &IID_IUnknown))
1946	{
1947	IUnknown_AddRef(iface);
1948	*object = iface;
1949	return S_OK;
1950	}
1951
1952	WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1953
1954	*object = NULL;
1955	return E_NOINTERFACE;
1956	}
1957
1958	static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
1959	{
1960	IWineD3DPixelShaderImpl shader = (IWineD3DPixelShaderImpl )iface;
1961	ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1962
1963	TRACE("%p increasing refcount to %u.\n", shader, refcount);
1964
1965	return refcount;
1966	}
1967
1968	static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
1969	{
1970	IWineD3DPixelShaderImpl shader = (IWineD3DPixelShaderImpl )iface;
1971	ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1972
1973	TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1974
1975	if (!refcount)
1976	{
1977	shader_cleanup((IWineD3DBaseShader *)iface);
1978	shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1979	HeapFree(GetProcessHeap(), 0, shader);
1980	}
1981
1982	return refcount;
1983	}
1984
1985	static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader iface, IUnknown *parent)
1986	{
1987	TRACE("iface %p, parent %p.\n", iface, parent);
1988
1989	shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1990
1991	return WINED3D_OK;
1992	}
1993
1994	static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader iface, void data, UINT *data_size)
1995	{
1996	TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1997
1998	return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1999	}
2000
2001	static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2002	{
2003	/* IUnknown methods */
2004	pixelshader_QueryInterface,
2005	pixelshader_AddRef,
2006	pixelshader_Release,
2007	/* IWineD3DBase methods */
2008	pixelshader_GetParent,
2009	/* IWineD3DBaseShader methods */
2010	pixelshader_GetFunction
2011	};
2012
2013	void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
2014	IWineD3DStateBlockImpl stateblock, struct ps_compile_args args)
2015	{
2016	IWineD3DBaseTextureImpl *texture;
2017	UINT i;
2018
2019	memset(args, 0, sizeof(args)); / FIXME: Make sure all bits are set. */
2020	args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
2021	args->np2_fixup = 0;
2022	args->t_mirror = 0;
2023
2024	for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2025	{
2026	if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2027	texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
2028	if (!texture)
2029	{
2030	args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2031	continue;
2032	}
2033	args->color_fixup[i] = texture->resource.format_desc->color_fixup;
2034
2035	/* Flag samplers that need NP2 texcoord fixup. */
2036	if (!texture->baseTexture.pow2Matrix_identity)
2037	{
2038	args->np2_fixup \|= (1 << i);
2039	}
2040
2041	if (texture->baseTexture.t_mirror)
2042	{
2043	args->t_mirror \|= (1 << i);
2044	}
2045	}
2046	if (shader->baseShader.reg_maps.shader_version.major >= 3)
2047	{
2048	if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2049	{
2050	args->vp_mode = pretransformed;
2051	}
2052	else if (use_vs(stateblock))
2053	{
2054	args->vp_mode = vertexshader;
2055	}
2056	else
2057	{
2058	args->vp_mode = fixedfunction;
2059	}
2060	args->fog = FOG_OFF;
2061	}
2062	else
2063	{
2064	args->vp_mode = vertexshader;
2065	if (stateblock->renderState[WINED3DRS_FOGENABLE])
2066	{
2067	switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2068	{
2069	case WINED3DFOG_NONE:
2070	if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2071	\|\| use_vs(stateblock))
2072	{
2073	args->fog = FOG_LINEAR;
2074	break;
2075	}
2076
2077	switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2078	{
2079	case WINED3DFOG_NONE: /* Fall through. */
2080	case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2081	case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2082	case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2083	}
2084	break;
2085
2086	case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2087	case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2088	case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2089	}
2090	}
2091	else
2092	{
2093	args->fog = FOG_OFF;
2094	}
2095	}
2096	}
2097
2098	static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2099	{
2100	DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2101	shader->baseShader.reg_maps.shader_version.minor);
2102
2103	shader->baseShader.limits.attributes = 0;
2104	shader->baseShader.limits.address = 0;
2105	shader->baseShader.limits.packed_output = 0;
2106
2107	switch (shader_version)
2108	{
2109	case WINED3D_SHADER_VERSION(1, 0):
2110	case WINED3D_SHADER_VERSION(1, 1):
2111	case WINED3D_SHADER_VERSION(1, 2):
2112	case WINED3D_SHADER_VERSION(1, 3):
2113	shader->baseShader.limits.temporary = 2;
2114	shader->baseShader.limits.constant_float = 8;
2115	shader->baseShader.limits.constant_int = 0;
2116	shader->baseShader.limits.constant_bool = 0;
2117	shader->baseShader.limits.texcoord = 4;
2118	shader->baseShader.limits.sampler = 4;
2119	shader->baseShader.limits.packed_input = 0;
2120	shader->baseShader.limits.label = 0;
2121	break;
2122
2123	case WINED3D_SHADER_VERSION(1, 4):
2124	shader->baseShader.limits.temporary = 6;
2125	shader->baseShader.limits.constant_float = 8;
2126	shader->baseShader.limits.constant_int = 0;
2127	shader->baseShader.limits.constant_bool = 0;
2128	shader->baseShader.limits.texcoord = 6;
2129	shader->baseShader.limits.sampler = 6;
2130	shader->baseShader.limits.packed_input = 0;
2131	shader->baseShader.limits.label = 0;
2132	break;
2133
2134	/* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2135	case WINED3D_SHADER_VERSION(2, 0):
2136	shader->baseShader.limits.temporary = 32;
2137	shader->baseShader.limits.constant_float = 32;
2138	shader->baseShader.limits.constant_int = 16;
2139	shader->baseShader.limits.constant_bool = 16;
2140	shader->baseShader.limits.texcoord = 8;
2141	shader->baseShader.limits.sampler = 16;
2142	shader->baseShader.limits.packed_input = 0;
2143	break;
2144
2145	case WINED3D_SHADER_VERSION(2, 1):
2146	shader->baseShader.limits.temporary = 32;
2147	shader->baseShader.limits.constant_float = 32;
2148	shader->baseShader.limits.constant_int = 16;
2149	shader->baseShader.limits.constant_bool = 16;
2150	shader->baseShader.limits.texcoord = 8;
2151	shader->baseShader.limits.sampler = 16;
2152	shader->baseShader.limits.packed_input = 0;
2153	shader->baseShader.limits.label = 16;
2154	break;
2155
2156	case WINED3D_SHADER_VERSION(4, 0):
2157	FIXME("Using 3.0 limits for 4.0 shader.\n");
2158	/* Fall through. */
2159
2160	case WINED3D_SHADER_VERSION(3, 0):
2161	shader->baseShader.limits.temporary = 32;
2162	shader->baseShader.limits.constant_float = 224;
2163	shader->baseShader.limits.constant_int = 16;
2164	shader->baseShader.limits.constant_bool = 16;
2165	shader->baseShader.limits.texcoord = 0;
2166	shader->baseShader.limits.sampler = 16;
2167	shader->baseShader.limits.packed_input = 12;
2168	shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2169	break;
2170
2171	default:
2172	shader->baseShader.limits.temporary = 32;
2173	shader->baseShader.limits.constant_float = 32;
2174	shader->baseShader.limits.constant_int = 16;
2175	shader->baseShader.limits.constant_bool = 16;
2176	shader->baseShader.limits.texcoord = 8;
2177	shader->baseShader.limits.sampler = 16;
2178	shader->baseShader.limits.packed_input = 0;
2179	shader->baseShader.limits.label = 0;
2180	FIXME("Unrecognized pixel shader version %u.%u\n",
2181	shader->baseShader.reg_maps.shader_version.major,
2182	shader->baseShader.reg_maps.shader_version.minor);
2183	}
2184	}
2185
2186	HRESULT pixelshader_init(IWineD3DPixelShaderImpl shader, IWineD3DDeviceImpl device,
2187	const DWORD byte_code, const struct wined3d_shader_signature output_signature,
2188	IUnknown parent, const struct wined3d_parent_ops parent_ops)
2189	{
2190	const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2191	unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2192	HRESULT hr;
2193
2194	if (!byte_code) return WINED3DERR_INVALIDCALL;
2195
2196	shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2197	shader_init(&shader->baseShader, device, parent, parent_ops);
2198
2199	hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2200	output_signature, device->d3d_pshader_constantF);
2201	if (FAILED(hr))
2202	{
2203	WARN("Failed to set function, hr %#x.\n", hr);
2204	shader_cleanup((IWineD3DBaseShader *)shader);
2205	return hr;
2206	}
2207
2208	pixelshader_set_limits(shader);
2209
2210	for (i = 0; i < MAX_REG_INPUT; ++i)
2211	{
2212	if (shader->input_reg_used[i])
2213	{
2214	++num_regs_used;
2215	highest_reg_used = i;
2216	}
2217	}
2218
2219	/* Don't do any register mapping magic if it is not needed, or if we can't
2220	* achieve anything anyway */
2221	if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2222	\|\| num_regs_used > (gl_info->limits.glsl_varyings / 4))
2223	{
2224	if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2225	{
2226	/* This happens with relative addressing. The input mapper function
2227	* warns about this if the higher registers are declared too, so
2228	* don't write a FIXME here */
2229	WARN("More varying registers used than supported\n");
2230	}
2231
2232	for (i = 0; i < MAX_REG_INPUT; ++i)
2233	{
2234	shader->input_reg_map[i] = i;
2235	}
2236
2237	shader->declared_in_count = highest_reg_used + 1;
2238	}
2239	else
2240	{
2241	shader->declared_in_count = 0;
2242	for (i = 0; i < MAX_REG_INPUT; ++i)
2243	{
2244	if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2245	else shader->input_reg_map[i] = ~0U;
2246	}
2247	}
2248
2249	shader->baseShader.load_local_constsF = FALSE;
2250
2251	return WINED3D_OK;
2252	}
2253
2254	void pixelshader_update_samplers(struct shader_reg_maps reg_maps, IWineD3DBaseTexture const *textures)
2255	{
2256	WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2257	unsigned int i;
2258
2259	if (reg_maps->shader_version.major != 1) return;
2260
2261	for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2262	{
2263	/* We don't sample from this sampler. */
2264	if (!sampler_type[i]) continue;
2265
2266	if (!textures[i])
2267	{
2268	WARN("No texture bound to sampler %u, using 2D.\n", i);
2269	sampler_type[i] = WINED3DSTT_2D;
2270	continue;
2271	}
2272
2273	switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2274	{
2275	case GL_TEXTURE_RECTANGLE_ARB:
2276	case GL_TEXTURE_2D:
2277	/* We have to select between texture rectangles and 2D
2278	* textures later because 2.0 and 3.0 shaders only have
2279	* WINED3DSTT_2D as well. */
2280	sampler_type[i] = WINED3DSTT_2D;
2281	break;
2282
2283	case GL_TEXTURE_3D:
2284	sampler_type[i] = WINED3DSTT_VOLUME;
2285	break;
2286
2287	case GL_TEXTURE_CUBE_MAP_ARB:
2288	sampler_type[i] = WINED3DSTT_CUBE;
2289	break;
2290
2291	default:
2292	FIXME("Unrecognized texture type %#x, using 2D.\n",
2293	IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2294	sampler_type[i] = WINED3DSTT_2D;
2295	}
2296	}
2297	}

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/Additions/WINNT/Graphics/Wine/wined3d/shader.c@ 35718

Download in other formats: