?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
quartz: Clarify debug strings.
[wine.git] / dlls / wined3d / shader.c
blob61509d97d6b1d51d2f576dcc941f278bc083559e
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <math.h>
29 #include <stdio.h>
30 #include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_BEM */ "bem",
47 /* WINED3DSIH_BFI */ "bfi",
48 /* WINED3DSIH_BFREV */ "bfrev",
49 /* WINED3DSIH_BREAK */ "break",
50 /* WINED3DSIH_BREAKC */ "breakc",
51 /* WINED3DSIH_BREAKP */ "breakp",
52 /* WINED3DSIH_BUFINFO */ "bufinfo",
53 /* WINED3DSIH_CALL */ "call",
54 /* WINED3DSIH_CALLNZ */ "callnz",
55 /* WINED3DSIH_CASE */ "case",
56 /* WINED3DSIH_CMP */ "cmp",
57 /* WINED3DSIH_CND */ "cnd",
58 /* WINED3DSIH_CONTINUE */ "continue",
59 /* WINED3DSIH_CRS */ "crs",
60 /* WINED3DSIH_CUT */ "cut",
61 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
62 /* WINED3DSIH_DCL */ "dcl",
63 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
64 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
65 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
66 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
67 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
68 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
69 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
70 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
71 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
72 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
73 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
74 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
75 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
76 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
77 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
78 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
79 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
80 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
81 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
82 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
83 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
84 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
85 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
86 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
87 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
88 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
89 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
90 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
91 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
92 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
93 /* WINED3DSIH_DEF */ "def",
94 /* WINED3DSIH_DEFAULT */ "default",
95 /* WINED3DSIH_DEFB */ "defb",
96 /* WINED3DSIH_DEFI */ "defi",
97 /* WINED3DSIH_DIV */ "div",
98 /* WINED3DSIH_DP2 */ "dp2",
99 /* WINED3DSIH_DP2ADD */ "dp2add",
100 /* WINED3DSIH_DP3 */ "dp3",
101 /* WINED3DSIH_DP4 */ "dp4",
102 /* WINED3DSIH_DST */ "dst",
103 /* WINED3DSIH_DSX */ "dsx",
104 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
105 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
106 /* WINED3DSIH_DSY */ "dsy",
107 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
108 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
109 /* WINED3DSIH_ELSE */ "else",
110 /* WINED3DSIH_EMIT */ "emit",
111 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
112 /* WINED3DSIH_ENDIF */ "endif",
113 /* WINED3DSIH_ENDLOOP */ "endloop",
114 /* WINED3DSIH_ENDREP */ "endrep",
115 /* WINED3DSIH_ENDSWITCH */ "endswitch",
116 /* WINED3DSIH_EQ */ "eq",
117 /* WINED3DSIH_EXP */ "exp",
118 /* WINED3DSIH_EXPP */ "expp",
119 /* WINED3DSIH_FRC */ "frc",
120 /* WINED3DSIH_FTOI */ "ftoi",
121 /* WINED3DSIH_FTOU */ "ftou",
122 /* WINED3DSIH_GATHER4 */ "gather4",
123 /* WINED3DSIH_GATHER4_C */ "gather4_c",
124 /* WINED3DSIH_GE */ "ge",
125 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
126 /* WINED3DSIH_HS_DECLS */ "hs_decls",
127 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
128 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
129 /* WINED3DSIH_IADD */ "iadd",
130 /* WINED3DSIH_IEQ */ "ieq",
131 /* WINED3DSIH_IF */ "if",
132 /* WINED3DSIH_IFC */ "ifc",
133 /* WINED3DSIH_IGE */ "ige",
134 /* WINED3DSIH_ILT */ "ilt",
135 /* WINED3DSIH_IMAD */ "imad",
136 /* WINED3DSIH_IMAX */ "imax",
137 /* WINED3DSIH_IMIN */ "imin",
138 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
139 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
140 /* WINED3DSIH_IMUL */ "imul",
141 /* WINED3DSIH_INE */ "ine",
142 /* WINED3DSIH_INEG */ "ineg",
143 /* WINED3DSIH_ISHL */ "ishl",
144 /* WINED3DSIH_ISHR */ "ishr",
145 /* WINED3DSIH_ITOF */ "itof",
146 /* WINED3DSIH_LABEL */ "label",
147 /* WINED3DSIH_LD */ "ld",
148 /* WINED3DSIH_LD2DMS */ "ld2dms",
149 /* WINED3DSIH_LD_RAW */ "ld_raw",
150 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
151 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
152 /* WINED3DSIH_LIT */ "lit",
153 /* WINED3DSIH_LOD */ "lod",
154 /* WINED3DSIH_LOG */ "log",
155 /* WINED3DSIH_LOGP */ "logp",
156 /* WINED3DSIH_LOOP */ "loop",
157 /* WINED3DSIH_LRP */ "lrp",
158 /* WINED3DSIH_LT */ "lt",
159 /* WINED3DSIH_M3x2 */ "m3x2",
160 /* WINED3DSIH_M3x3 */ "m3x3",
161 /* WINED3DSIH_M3x4 */ "m3x4",
162 /* WINED3DSIH_M4x3 */ "m4x3",
163 /* WINED3DSIH_M4x4 */ "m4x4",
164 /* WINED3DSIH_MAD */ "mad",
165 /* WINED3DSIH_MAX */ "max",
166 /* WINED3DSIH_MIN */ "min",
167 /* WINED3DSIH_MOV */ "mov",
168 /* WINED3DSIH_MOVA */ "mova",
169 /* WINED3DSIH_MOVC */ "movc",
170 /* WINED3DSIH_MUL */ "mul",
171 /* WINED3DSIH_NE */ "ne",
172 /* WINED3DSIH_NOP */ "nop",
173 /* WINED3DSIH_NOT */ "not",
174 /* WINED3DSIH_NRM */ "nrm",
175 /* WINED3DSIH_OR */ "or",
176 /* WINED3DSIH_PHASE */ "phase",
177 /* WINED3DSIH_POW */ "pow",
178 /* WINED3DSIH_RCP */ "rcp",
179 /* WINED3DSIH_REP */ "rep",
180 /* WINED3DSIH_RESINFO */ "resinfo",
181 /* WINED3DSIH_RET */ "ret",
182 /* WINED3DSIH_ROUND_NE */ "round_ne",
183 /* WINED3DSIH_ROUND_NI */ "round_ni",
184 /* WINED3DSIH_ROUND_PI */ "round_pi",
185 /* WINED3DSIH_ROUND_Z */ "round_z",
186 /* WINED3DSIH_RSQ */ "rsq",
187 /* WINED3DSIH_SAMPLE */ "sample",
188 /* WINED3DSIH_SAMPLE_B */ "sample_b",
189 /* WINED3DSIH_SAMPLE_C */ "sample_c",
190 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
191 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
192 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
193 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
194 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
195 /* WINED3DSIH_SETP */ "setp",
196 /* WINED3DSIH_SGE */ "sge",
197 /* WINED3DSIH_SGN */ "sgn",
198 /* WINED3DSIH_SINCOS */ "sincos",
199 /* WINED3DSIH_SLT */ "slt",
200 /* WINED3DSIH_SQRT */ "sqrt",
201 /* WINED3DSIH_STORE_RAW */ "store_raw",
202 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
203 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
204 /* WINED3DSIH_SUB */ "sub",
205 /* WINED3DSIH_SWAPC */ "swapc",
206 /* WINED3DSIH_SWITCH */ "switch",
207 /* WINED3DSIH_TEX */ "texld",
208 /* WINED3DSIH_TEXBEM */ "texbem",
209 /* WINED3DSIH_TEXBEML */ "texbeml",
210 /* WINED3DSIH_TEXCOORD */ "texcrd",
211 /* WINED3DSIH_TEXDEPTH */ "texdepth",
212 /* WINED3DSIH_TEXDP3 */ "texdp3",
213 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
214 /* WINED3DSIH_TEXKILL */ "texkill",
215 /* WINED3DSIH_TEXLDD */ "texldd",
216 /* WINED3DSIH_TEXLDL */ "texldl",
217 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
218 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
219 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
220 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
221 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
222 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
223 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
224 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
225 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
226 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
227 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
228 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
229 /* WINED3DSIH_UBFE */ "ubfe",
230 /* WINED3DSIH_UDIV */ "udiv",
231 /* WINED3DSIH_UGE */ "uge",
232 /* WINED3DSIH_ULT */ "ult",
233 /* WINED3DSIH_UMAX */ "umax",
234 /* WINED3DSIH_UMIN */ "umin",
235 /* WINED3DSIH_USHR */ "ushr",
236 /* WINED3DSIH_UTOF */ "utof",
237 /* WINED3DSIH_XOR */ "xor",
238 };
239
240 static const char * const semantic_names[] =
241 {
242 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
243 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
244 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
245 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
246 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
247 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
248 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
249 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
250 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
251 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
252 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
253 /* WINED3D_DECL_USAGE_FOG */ "FOG",
254 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
255 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
256 };
257
258 static const struct
259 {
260 enum wined3d_sysval_semantic sysval_semantic;
261 const char *sysval_name;
262 }
263 sysval_semantic_names[] =
264 {
265 {WINED3D_SV_POSITION, "SV_Position"},
266 {WINED3D_SV_CLIP_DISTANCE, "SV_ClipDistance"},
267 {WINED3D_SV_CULL_DISTANCE, "SV_CullDistance"},
268 {WINED3D_SV_RENDER_TARGET_ARRAY_INDEX, "SV_RenderTargetArrayIndex"},
269 {WINED3D_SV_VIEWPORT_ARRAY_INDEX, "SV_ViewportArrayIndex"},
270 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
271 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
272 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
273 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
274 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
275 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
276 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
277 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
278 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
279 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
280 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
281 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
282 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
283 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
284 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
285 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
286 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
287 };
288
289 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
290 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
291
292 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
293 {
294 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
295 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
296
297 return shader_opcode_names[handler_idx];
298 }
299
300 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
301 {
302 if (usage >= ARRAY_SIZE(semantic_names))
303 {
304 FIXME("Unrecognized usage %#x.\n", usage);
305 return "UNRECOGNIZED";
306 }
307
308 return semantic_names[usage];
309 }
310
311 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
312 {
313 unsigned int i;
314
315 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
316 {
317 if (!strcmp(name, semantic_names[i]))
318 return i;
319 }
320
321 return ~0U;
322 }
323
324 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
325 {
326 switch (usage)
327 {
328 case WINED3D_DECL_USAGE_POSITION:
329 return WINED3D_SV_POSITION;
330 default:
331 return 0;
332 }
333 }
334
335 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
336 {
337 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
338 }
339
340 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
341 const struct wined3d_shader_semantic *s)
342 {
343 e->semantic_name = shader_semantic_name_from_usage(s->usage);
344 e->semantic_idx = s->usage_idx;
345 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
346 e->component_type = WINED3D_TYPE_FLOAT;
347 e->register_idx = s->reg.reg.idx[0].offset;
348 e->mask = s->reg.write_mask;
349 }
350
351 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
352 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
353 {
354 e->semantic_name = shader_semantic_name_from_usage(usage);
355 e->semantic_idx = usage_idx;
356 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
357 e->component_type = WINED3D_TYPE_FLOAT;
358 e->register_idx = reg_idx;
359 e->mask = write_mask;
360 }
361
362 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
363 {
364 switch (version_token >> 16)
365 {
366 case WINED3D_SM1_VS:
367 case WINED3D_SM1_PS:
368 return &sm1_shader_frontend;
369
370 case WINED3D_SM4_PS:
371 case WINED3D_SM4_VS:
372 case WINED3D_SM4_GS:
373 case WINED3D_SM5_HS:
374 case WINED3D_SM5_DS:
375 case WINED3D_SM5_CS:
376 return &sm4_shader_frontend;
377
378 default:
379 FIXME("Unrecognised version token %#x.\n", version_token);
380 return NULL;
381 }
382 }
383
384 void string_buffer_clear(struct wined3d_string_buffer *buffer)
385 {
386 buffer->buffer[0] = '\0';
387 buffer->content_size = 0;
388 }
389
390 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
391 {
392 buffer->buffer_size = 32;
393 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
394 {
395 ERR("Failed to allocate shader buffer memory.\n");
396 return FALSE;
397 }
398
399 string_buffer_clear(buffer);
400 return TRUE;
401 }
402
403 void string_buffer_free(struct wined3d_string_buffer *buffer)
404 {
405 HeapFree(GetProcessHeap(), 0, buffer->buffer);
406 }
407
408 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
409 {
410 char *new_buffer;
411 unsigned int new_buffer_size = buffer->buffer_size * 2;
412
413 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
414 new_buffer_size *= 2;
415 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
416 {
417 ERR("Failed to grow buffer.\n");
418 buffer->buffer[buffer->content_size] = '\0';
419 return FALSE;
420 }
421 buffer->buffer = new_buffer;
422 buffer->buffer_size = new_buffer_size;
423 return TRUE;
424 }
425
426 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
427 {
428 unsigned int rem;
429 int rc;
430
431 rem = buffer->buffer_size - buffer->content_size;
432 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
433 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
434 return rc;
435
436 buffer->content_size += rc;
437 return 0;
438 }
439
440 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
441 {
442 va_list args;
443 int ret;
444
445 for (;;)
446 {
447 va_start(args, format);
448 ret = shader_vaddline(buffer, format, args);
449 va_end(args);
450 if (!ret)
451 return ret;
452 if (!string_buffer_resize(buffer, ret))
453 return -1;
454 }
455 }
456
457 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
458 {
459 struct wined3d_string_buffer *buffer;
460
461 if (list_empty(&list->list))
462 {
463 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
464 if (!buffer || !string_buffer_init(buffer))
465 {
466 ERR("Couldn't allocate buffer for temporary string.\n");
467 HeapFree(GetProcessHeap(), 0, buffer);
468 return NULL;
469 }
470 }
471 else
472 {
473 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
474 list_remove(&buffer->entry);
475 }
476 string_buffer_clear(buffer);
477 return buffer;
478 }
479
480 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
481 {
482 if (!buffer)
483 return 0;
484 string_buffer_clear(buffer);
485 return shader_vaddline(buffer, format, args);
486 }
487
488 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
489 {
490 va_list args;
491 int ret;
492
493 for (;;)
494 {
495 va_start(args, format);
496 ret = string_buffer_vsprintf(buffer, format, args);
497 va_end(args);
498 if (!ret)
499 return;
500 if (!string_buffer_resize(buffer, ret))
501 return;
502 }
503 }
504
505 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
506 {
507 if (!buffer)
508 return;
509 list_add_head(&list->list, &buffer->entry);
510 }
511
512 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
513 {
514 list_init(&list->list);
515 }
516
517 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
518 {
519 struct wined3d_string_buffer *buffer, *buffer_next;
520
521 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
522 {
523 string_buffer_free(buffer);
524 HeapFree(GetProcessHeap(), 0, buffer);
525 }
526 list_init(&list->list);
527 }
528
529 /* Convert floating point offset relative to a register file to an absolute
530 * offset for float constants. */
531 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
532 {
533 switch (register_type)
534 {
535 case WINED3DSPR_CONST: return register_idx;
536 case WINED3DSPR_CONST2: return 2048 + register_idx;
537 case WINED3DSPR_CONST3: return 4096 + register_idx;
538 case WINED3DSPR_CONST4: return 6144 + register_idx;
539 default:
540 FIXME("Unsupported register type: %u.\n", register_type);
541 return register_idx;
542 }
543 }
544
545 static void shader_delete_constant_list(struct list *clist)
546 {
547 struct wined3d_shader_lconst *constant, *constant_next;
548
549 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
550 HeapFree(GetProcessHeap(), 0, constant);
551 list_init(clist);
552 }
553
554 static void shader_set_limits(struct wined3d_shader *shader)
555 {
556 static const struct limits_entry
557 {
558 unsigned int min_version;
559 unsigned int max_version;
560 struct wined3d_shader_limits limits;
561 }
562 vs_limits[] =
563 {
564 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
565 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
566 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
567 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
568 * even though they are capable of supporting much more (GL
569 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
570 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
571 * shaders to 256. */
572 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
573 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
574 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
575 {0}
576 },
577 hs_limits[] =
578 {
579 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
580 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
581 },
582 ds_limits[] =
583 {
584 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
585 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
586 },
587 gs_limits[] =
588 {
589 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
590 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
591 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
592 {0}
593 },
594 ps_limits[] =
595 {
596 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
597 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
598 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
599 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
600 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
601 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
602 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
603 {0}
604 },
605 cs_limits[] =
606 {
607 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
608 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
609 };
610 const struct limits_entry *limits_array;
611 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
612 shader->reg_maps.shader_version.minor);
613 int i = 0;
614
615 switch (shader->reg_maps.shader_version.type)
616 {
617 default:
618 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
619 /* Fall-through. */
620 case WINED3D_SHADER_TYPE_VERTEX:
621 limits_array = vs_limits;
622 break;
623 case WINED3D_SHADER_TYPE_HULL:
624 limits_array = hs_limits;
625 break;
626 case WINED3D_SHADER_TYPE_DOMAIN:
627 limits_array = ds_limits;
628 break;
629 case WINED3D_SHADER_TYPE_GEOMETRY:
630 limits_array = gs_limits;
631 break;
632 case WINED3D_SHADER_TYPE_PIXEL:
633 limits_array = ps_limits;
634 break;
635 case WINED3D_SHADER_TYPE_COMPUTE:
636 limits_array = cs_limits;
637 break;
638 }
639
640 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
641 {
642 if (shader_version <= limits_array[i].max_version)
643 {
644 shader->limits = &limits_array[i].limits;
645 break;
646 }
647 ++i;
648 }
649 if (!shader->limits)
650 {
651 FIXME("Unexpected shader version \"%u.%u\".\n",
652 shader->reg_maps.shader_version.major,
653 shader->reg_maps.shader_version.minor);
654 shader->limits = &limits_array[max(0, i - 1)].limits;
655 }
656 }
657
658 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
659 {
660 DWORD idx, shift;
661 idx = bit >> 5;
662 shift = bit & 0x1f;
663 bitmap[idx] |= (1u << shift);
664 }
665
666 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
667 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
668 {
669 switch (reg->type)
670 {
671 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
672 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
673 reg_maps->texcoord |= 1u << reg->idx[0].offset;
674 else
675 reg_maps->address |= 1u << reg->idx[0].offset;
676 break;
677
678 case WINED3DSPR_TEMP:
679 reg_maps->temporary |= 1u << reg->idx[0].offset;
680 break;
681
682 case WINED3DSPR_INPUT:
683 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
684 {
685 if (reg->idx[0].rel_addr)
686 {
687 /* If relative addressing is used, we must assume that all registers
688 * are used. Even if it is a construct like v3[aL], we can't assume
689 * that v0, v1 and v2 aren't read because aL can be negative */
690 unsigned int i;
691 for (i = 0; i < MAX_REG_INPUT; ++i)
692 {
693 shader->u.ps.input_reg_used[i] = TRUE;
694 }
695 }
696 else
697 {
698 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
699 }
700 }
701 else
702 reg_maps->input_registers |= 1u << reg->idx[0].offset;
703 break;
704
705 case WINED3DSPR_RASTOUT:
706 if (reg->idx[0].offset == 1)
707 reg_maps->fog = 1;
708 if (reg->idx[0].offset == 2)
709 reg_maps->point_size = 1;
710 break;
711
712 case WINED3DSPR_MISCTYPE:
713 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
714 {
715 if (!reg->idx[0].offset)
716 reg_maps->vpos = 1;
717 else if (reg->idx[0].offset == 1)
718 reg_maps->usesfacing = 1;
719 }
720 break;
721
722 case WINED3DSPR_CONST:
723 if (reg->idx[0].rel_addr)
724 {
725 if (reg->idx[0].offset < reg_maps->min_rel_offset)
726 reg_maps->min_rel_offset = reg->idx[0].offset;
727 if (reg->idx[0].offset > reg_maps->max_rel_offset)
728 reg_maps->max_rel_offset = reg->idx[0].offset;
729 reg_maps->usesrelconstF = TRUE;
730 }
731 else
732 {
733 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
734 {
735 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
736 return FALSE;
737 }
738 else
739 {
740 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
741 }
742 }
743 break;
744
745 case WINED3DSPR_CONSTINT:
746 if (reg->idx[0].offset >= shader->limits->constant_int)
747 {
748 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
749 return FALSE;
750 }
751 else
752 {
753 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
754 }
755 break;
756
757 case WINED3DSPR_CONSTBOOL:
758 if (reg->idx[0].offset >= shader->limits->constant_bool)
759 {
760 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
761 return FALSE;
762 }
763 else
764 {
765 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
766 }
767 break;
768
769 case WINED3DSPR_COLOROUT:
770 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
771 break;
772
773 default:
774 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
775 reg->type, reg->idx[0].offset, reg->idx[1].offset);
776 break;
777 }
778 return TRUE;
779 }
780
781 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
782 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
783 {
784 struct wined3d_shader_sampler_map_entry *entries, *entry;
785 struct wined3d_shader_sampler_map *map;
786 unsigned int i;
787
788 map = &reg_maps->sampler_map;
789 entries = map->entries;
790 for (i = 0; i < map->count; ++i)
791 {
792 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
793 return;
794 }
795
796 if (!map->size)
797 {
798 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
799 {
800 ERR("Failed to allocate sampler map entries.\n");
801 return;
802 }
803 map->size = 4;
804 map->entries = entries;
805 }
806 else if (map->count == map->size)
807 {
808 size_t new_size = map->size * 2;
809
810 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
811 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
812 {
813 ERR("Failed to resize sampler map entries.\n");
814 return;
815 }
816 map->size = new_size;
817 map->entries = entries;
818 }
819
820 entry = &entries[map->count++];
821 entry->resource_idx = resource_idx;
822 entry->sampler_idx = sampler_idx;
823 entry->bind_idx = bind_idx;
824 }
825
826 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
827 {
828 switch (instr)
829 {
830 case WINED3DSIH_M4x4:
831 case WINED3DSIH_M3x4:
832 return param == 1 ? 3 : 0;
833
834 case WINED3DSIH_M4x3:
835 case WINED3DSIH_M3x3:
836 return param == 1 ? 2 : 0;
837
838 case WINED3DSIH_M3x2:
839 return param == 1 ? 1 : 0;
840
841 default:
842 return 0;
843 }
844 }
845
846 /* Note that this does not count the loop register as an address register. */
847 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
848 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
849 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
850 {
851 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
852 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
853 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
854 void *fe_data = shader->frontend_data;
855 struct wined3d_shader_version shader_version;
856 const DWORD *ptr = byte_code;
857 unsigned int i;
858
859 memset(reg_maps, 0, sizeof(*reg_maps));
860 memset(input_signature_elements, 0, sizeof(input_signature_elements));
861 memset(output_signature_elements, 0, sizeof(output_signature_elements));
862 reg_maps->min_rel_offset = ~0U;
863 list_init(&reg_maps->indexable_temps);
864
865 fe->shader_read_header(fe_data, &ptr, &shader_version);
866 reg_maps->shader_version = shader_version;
867
868 shader_set_limits(shader);
869
870 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
871 sizeof(*reg_maps->constf))))
872 {
873 ERR("Failed to allocate constant map memory.\n");
874 return E_OUTOFMEMORY;
875 }
876
877 while (!fe->shader_is_end(fe_data, &ptr))
878 {
879 struct wined3d_shader_instruction ins;
880
881 /* Fetch opcode. */
882 fe->shader_read_instruction(fe_data, &ptr, &ins);
883
884 /* Unhandled opcode, and its parameters. */
885 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
886 {
887 TRACE("Skipping unrecognized instruction.\n");
888 continue;
889 }
890
891 /* Handle declarations. */
892 if (ins.handler_idx == WINED3DSIH_DCL)
893 {
894 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
895 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
896
897 switch (semantic->reg.reg.type)
898 {
899 /* Mark input registers used. */
900 case WINED3DSPR_INPUT:
901 if (reg_idx >= MAX_REG_INPUT)
902 {
903 ERR("Invalid input register index %u.\n", reg_idx);
904 break;
905 }
906 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
907 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
908 return WINED3DERR_INVALIDCALL;
909 reg_maps->input_registers |= 1u << reg_idx;
910 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
911 break;
912
913 /* Vertex shader: mark 3.0 output registers used, save token. */
914 case WINED3DSPR_OUTPUT:
915 if (reg_idx >= MAX_REG_OUTPUT)
916 {
917 ERR("Invalid output register index %u.\n", reg_idx);
918 break;
919 }
920 reg_maps->output_registers |= 1u << reg_idx;
921 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
922 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
923 reg_maps->fog = 1;
924 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
925 reg_maps->point_size = 1;
926 break;
927
928 case WINED3DSPR_SAMPLER:
929 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
930 case WINED3DSPR_RESOURCE:
931 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
932 {
933 ERR("Invalid resource index %u.\n", reg_idx);
934 break;
935 }
936 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
937 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
938 break;
939
940 default:
941 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
942 break;
943 }
944 }
945 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
946 {
947 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
948 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
949 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
950 else
951 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
952 }
953 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
954 {
955 if (reg_maps->icb)
956 FIXME("Multiple immediate constant buffers.\n");
957 reg_maps->icb = ins.declaration.icb;
958 }
959 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
960 {
961 struct wined3d_shader_indexable_temp *reg;
962
963 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
964 return E_OUTOFMEMORY;
965
966 *reg = ins.declaration.indexable_temp;
967 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
968 }
969 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
970 {
971 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
972 shader->u.gs.input_type = ins.declaration.primitive_type;
973 else
974 FIXME("Invalid instruction %#x for shader type %#x.\n",
975 ins.handler_idx, shader_version.type);
976 }
977 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
978 {
979 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
980 shader->u.gs.output_type = ins.declaration.primitive_type;
981 else
982 FIXME("Invalid instruction %#x for shader type %#x.\n",
983 ins.handler_idx, shader_version.type);
984 }
985 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
986 {
987 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
988 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
989 }
990 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
991 {
992 reg_maps->temporary_count = ins.declaration.count;
993 }
994 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
995 {
996 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
997 shader->u.gs.vertices_out = ins.declaration.count;
998 else
999 FIXME("Invalid instruction %#x for shader type %#x.\n",
1000 ins.handler_idx, shader_version.type);
1001 }
1002 else if (ins.handler_idx == WINED3DSIH_DEF)
1003 {
1004 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1005 float *value;
1006 if (!lconst) return E_OUTOFMEMORY;
1007
1008 lconst->idx = ins.dst[0].reg.idx[0].offset;
1009 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
1010 value = (float *)lconst->value;
1011
1012 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1013 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1014 {
1015 if (value[0] < -1.0f) value[0] = -1.0f;
1016 else if (value[0] > 1.0f) value[0] = 1.0f;
1017 if (value[1] < -1.0f) value[1] = -1.0f;
1018 else if (value[1] > 1.0f) value[1] = 1.0f;
1019 if (value[2] < -1.0f) value[2] = -1.0f;
1020 else if (value[2] > 1.0f) value[2] = 1.0f;
1021 if (value[3] < -1.0f) value[3] = -1.0f;
1022 else if (value[3] > 1.0f) value[3] = 1.0f;
1023 }
1024
1025 list_add_head(&shader->constantsF, &lconst->entry);
1026
1027 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1028 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1029 {
1030 shader->lconst_inf_or_nan = TRUE;
1031 }
1032 }
1033 else if (ins.handler_idx == WINED3DSIH_DEFI)
1034 {
1035 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1036 if (!lconst) return E_OUTOFMEMORY;
1037
1038 lconst->idx = ins.dst[0].reg.idx[0].offset;
1039 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
1040
1041 list_add_head(&shader->constantsI, &lconst->entry);
1042 reg_maps->local_int_consts |= (1u << lconst->idx);
1043 }
1044 else if (ins.handler_idx == WINED3DSIH_DEFB)
1045 {
1046 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1047 if (!lconst) return E_OUTOFMEMORY;
1048
1049 lconst->idx = ins.dst[0].reg.idx[0].offset;
1050 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
1051
1052 list_add_head(&shader->constantsB, &lconst->entry);
1053 reg_maps->local_bool_consts |= (1u << lconst->idx);
1054 }
1055 /* For subroutine prototypes. */
1056 else if (ins.handler_idx == WINED3DSIH_LABEL)
1057 {
1058 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1059 }
1060 /* Set texture, address, temporary registers. */
1061 else
1062 {
1063 BOOL color0_mov = FALSE;
1064 unsigned int i;
1065
1066 /* This will loop over all the registers and try to
1067 * make a bitmask of the ones we're interested in.
1068 *
1069 * Relative addressing tokens are ignored, but that's
1070 * okay, since we'll catch any address registers when
1071 * they are initialized (required by spec). */
1072 for (i = 0; i < ins.dst_count; ++i)
1073 {
1074 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1075 shader_version.type, constf_size))
1076 return WINED3DERR_INVALIDCALL;
1077
1078 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1079 {
1080 UINT idx = ins.dst[i].reg.idx[0].offset;
1081
1082 switch (ins.dst[i].reg.type)
1083 {
1084 case WINED3DSPR_RASTOUT:
1085 if (shader_version.major >= 3)
1086 break;
1087 switch (idx)
1088 {
1089 case 0: /* oPos */
1090 reg_maps->output_registers |= 1u << 10;
1091 shader_signature_from_usage(&output_signature_elements[10],
1092 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1093 break;
1094
1095 case 1: /* oFog */
1096 reg_maps->output_registers |= 1u << 11;
1097 shader_signature_from_usage(&output_signature_elements[11],
1098 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1099 break;
1100
1101 case 2: /* oPts */
1102 reg_maps->output_registers |= 1u << 11;
1103 shader_signature_from_usage(&output_signature_elements[11],
1104 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1105 break;
1106 }
1107 break;
1108
1109 case WINED3DSPR_ATTROUT:
1110 if (shader_version.major >= 3)
1111 break;
1112 if (idx < 2)
1113 {
1114 idx += 8;
1115 if (reg_maps->output_registers & (1u << idx))
1116 {
1117 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1118 }
1119 else
1120 {
1121 reg_maps->output_registers |= 1u << idx;
1122 shader_signature_from_usage(&output_signature_elements[idx],
1123 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1124 }
1125 }
1126 break;
1127
1128 case WINED3DSPR_TEXCRDOUT:
1129 if (shader_version.major >= 3)
1130 {
1131 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1132 break;
1133 }
1134 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1135 if (reg_maps->output_registers & (1u << idx))
1136 {
1137 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1138 }
1139 else
1140 {
1141 reg_maps->output_registers |= 1u << idx;
1142 shader_signature_from_usage(&output_signature_elements[idx],
1143 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1144 }
1145 break;
1146
1147 default:
1148 break;
1149 }
1150 }
1151
1152 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1153 {
1154 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1155 {
1156 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1157 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1158 * the mov and perform the sRGB write correction from the source register.
1159 *
1160 * However, if the mov is only partial, we can't do this, and if the write
1161 * comes from an instruction other than MOV it is hard to do as well. If
1162 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1163 shader->u.ps.color0_mov = FALSE;
1164 if (ins.handler_idx == WINED3DSIH_MOV
1165 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1166 {
1167 /* Used later when the source register is read. */
1168 color0_mov = TRUE;
1169 }
1170 }
1171 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1172 * end
1173 */
1174 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1175 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1176 {
1177 shader->u.ps.color0_mov = FALSE;
1178 }
1179 }
1180
1181 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1182 if (shader_version.major == 1
1183 && (ins.handler_idx == WINED3DSIH_TEX
1184 || ins.handler_idx == WINED3DSIH_TEXBEM
1185 || ins.handler_idx == WINED3DSIH_TEXBEML
1186 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1187 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1188 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1189 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1190 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1191 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1192 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1193 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1194 {
1195 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1196
1197 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1198 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1199 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1200 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1201
1202 /* texbem is only valid with < 1.4 pixel shaders */
1203 if (ins.handler_idx == WINED3DSIH_TEXBEM
1204 || ins.handler_idx == WINED3DSIH_TEXBEML)
1205 {
1206 reg_maps->bumpmat |= 1u << reg_idx;
1207 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1208 {
1209 reg_maps->luminanceparams |= 1u << reg_idx;
1210 }
1211 }
1212 }
1213 else if (ins.handler_idx == WINED3DSIH_BEM)
1214 {
1215 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1216 }
1217 }
1218
1219 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
1220 else if (ins.handler_idx == WINED3DSIH_DSY
1221 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1222 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1223 {
1224 reg_maps->usesdsy = 1;
1225 }
1226 else if (ins.handler_idx == WINED3DSIH_DSX
1227 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1228 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1229 {
1230 reg_maps->usesdsx = 1;
1231 }
1232 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1233 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1234 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1235 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1236 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1237 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1238 else if (ins.handler_idx == WINED3DSIH_LOOP
1239 || ins.handler_idx == WINED3DSIH_REP)
1240 {
1241 ++cur_loop_depth;
1242 if (cur_loop_depth > max_loop_depth)
1243 max_loop_depth = cur_loop_depth;
1244 }
1245 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1246 || ins.handler_idx == WINED3DSIH_ENDREP)
1247 {
1248 --cur_loop_depth;
1249 }
1250 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1251 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1252 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1253 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1254 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1255 {
1256 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1257 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1258 }
1259 else if (ins.handler_idx == WINED3DSIH_LD
1260 || ins.handler_idx == WINED3DSIH_RESINFO)
1261 {
1262 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1263 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1264 }
1265
1266 if (ins.predicate)
1267 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1268 shader_version.type, constf_size))
1269 return WINED3DERR_INVALIDCALL;
1270
1271 for (i = 0; i < ins.src_count; ++i)
1272 {
1273 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1274 struct wined3d_shader_register reg = ins.src[i].reg;
1275
1276 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1277 shader_version.type, constf_size))
1278 return WINED3DERR_INVALIDCALL;
1279 while (count)
1280 {
1281 ++reg.idx[0].offset;
1282 if (!shader_record_register_usage(shader, reg_maps, &reg,
1283 shader_version.type, constf_size))
1284 return WINED3DERR_INVALIDCALL;
1285 --count;
1286 }
1287
1288 if (color0_mov)
1289 {
1290 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1291 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1292 {
1293 shader->u.ps.color0_mov = TRUE;
1294 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1295 }
1296 }
1297 }
1298 }
1299 }
1300 reg_maps->loop_depth = max_loop_depth;
1301
1302 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1303 * R0 is written to the render target. */
1304 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1305 reg_maps->rt_mask |= (1u << 0);
1306
1307 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1308
1309 if (input_signature->elements)
1310 {
1311 for (i = 0; i < input_signature->element_count; ++i)
1312 {
1313 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1314 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1315 {
1316 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1317 reg_maps->vpos = 1;
1318 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1319 reg_maps->usesfacing = 1;
1320 }
1321 }
1322 }
1323 else if (!input_signature->elements && reg_maps->input_registers)
1324 {
1325 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1326 struct wined3d_shader_signature_element *e;
1327 unsigned int i;
1328
1329 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1330 return E_OUTOFMEMORY;
1331 input_signature->element_count = count;
1332
1333 e = input_signature->elements;
1334 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1335 {
1336 if (!(reg_maps->input_registers & (1u << i)))
1337 continue;
1338 input_signature_elements[i].register_idx = i;
1339 *e++ = input_signature_elements[i];
1340 }
1341 }
1342
1343 if (output_signature->elements)
1344 {
1345 for (i = 0; i < output_signature->element_count; ++i)
1346 {
1347 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1348 }
1349 }
1350 else if (reg_maps->output_registers)
1351 {
1352 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1353 struct wined3d_shader_signature_element *e;
1354
1355 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1356 return E_OUTOFMEMORY;
1357 output_signature->element_count = count;
1358
1359 e = output_signature->elements;
1360 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1361 {
1362 if (!(reg_maps->output_registers & (1u << i)))
1363 continue;
1364 *e++ = output_signature_elements[i];
1365 }
1366 }
1367
1368 return WINED3D_OK;
1369 }
1370
1371 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1372 {
1373 struct wined3d_shader_indexable_temp *reg, *reg_next;
1374
1375 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1376 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1377
1378 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1379 HeapFree(GetProcessHeap(), 0, reg);
1380 list_init(&reg_maps->indexable_temps);
1381 }
1382
1383 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1384 {
1385 DWORD map = 1u << max;
1386 map |= map - 1;
1387 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1388
1389 return wined3d_log2i(map);
1390 }
1391
1392 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1393 {
1394 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1395 {
1396 shader_addline(buffer, "refactoringAllowed");
1397 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1398 if (global_flags)
1399 shader_addline(buffer, " | ");
1400 }
1401
1402 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1403 {
1404 shader_addline(buffer, "enableRawAndStructuredBuffers");
1405 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1406 }
1407
1408 if (global_flags)
1409 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1410 }
1411
1412 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1413 enum wined3d_tessellator_domain domain)
1414 {
1415 switch (domain)
1416 {
1417 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1418 shader_addline(buffer, "line");
1419 break;
1420 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1421 shader_addline(buffer, "triangle");
1422 break;
1423 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1424 shader_addline(buffer, "quad");
1425 break;
1426 default:
1427 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1428 break;
1429 }
1430 }
1431
1432 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1433 enum wined3d_tessellator_output_primitive output_primitive)
1434 {
1435 switch (output_primitive)
1436 {
1437 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1438 shader_addline(buffer, "point");
1439 break;
1440 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1441 shader_addline(buffer, "line");
1442 break;
1443 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1444 shader_addline(buffer, "triangle_cw");
1445 break;
1446 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1447 shader_addline(buffer, "triangle_ccw");
1448 break;
1449 default:
1450 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1451 break;
1452 }
1453 }
1454
1455 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1456 enum wined3d_tessellator_partitioning partitioning)
1457 {
1458 switch (partitioning)
1459 {
1460 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1461 shader_addline(buffer, "integer");
1462 break;
1463 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1464 shader_addline(buffer, "pow2");
1465 break;
1466 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1467 shader_addline(buffer, "fractional_odd");
1468 break;
1469 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1470 shader_addline(buffer, "fractional_even");
1471 break;
1472 default:
1473 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1474 break;
1475 }
1476 }
1477
1478 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1479 {
1480 unsigned int i;
1481
1482 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1483 {
1484 if (sysval_semantic_names[i].sysval_semantic == semantic)
1485 {
1486 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1487 return;
1488 }
1489 }
1490
1491 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1492 }
1493
1494 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1495 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1496 {
1497 shader_addline(buffer, "dcl");
1498
1499 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1500 {
1501 switch (semantic->resource_type)
1502 {
1503 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1504 shader_addline(buffer, "_2d");
1505 break;
1506
1507 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1508 shader_addline(buffer, "_3d");
1509 break;
1510
1511 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1512 shader_addline(buffer, "_cube");
1513 break;
1514
1515 default:
1516 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1517 break;
1518 }
1519 }
1520 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1521 {
1522 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1523 shader_addline(buffer, "_resource_");
1524 else
1525 shader_addline(buffer, "_uav_");
1526 switch (semantic->resource_type)
1527 {
1528 case WINED3D_SHADER_RESOURCE_BUFFER:
1529 shader_addline(buffer, "buffer");
1530 break;
1531
1532 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1533 shader_addline(buffer, "texture1d");
1534 break;
1535
1536 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1537 shader_addline(buffer, "texture2d");
1538 break;
1539
1540 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1541 shader_addline(buffer, "texture2dms");
1542 break;
1543
1544 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1545 shader_addline(buffer, "texture3d");
1546 break;
1547
1548 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1549 shader_addline(buffer, "texturecube");
1550 break;
1551
1552 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1553 shader_addline(buffer, "texture1darray");
1554 break;
1555
1556 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1557 shader_addline(buffer, "texture2darray");
1558 break;
1559
1560 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1561 shader_addline(buffer, "texture2dmsarray");
1562 break;
1563
1564 default:
1565 shader_addline(buffer, "unknown");
1566 break;
1567 }
1568 switch (semantic->resource_data_type)
1569 {
1570 case WINED3D_DATA_FLOAT:
1571 shader_addline(buffer, " (float)");
1572 break;
1573
1574 case WINED3D_DATA_INT:
1575 shader_addline(buffer, " (int)");
1576 break;
1577
1578 case WINED3D_DATA_UINT:
1579 shader_addline(buffer, " (uint)");
1580 break;
1581
1582 case WINED3D_DATA_UNORM:
1583 shader_addline(buffer, " (unorm)");
1584 break;
1585
1586 case WINED3D_DATA_SNORM:
1587 shader_addline(buffer, " (snorm)");
1588 break;
1589
1590 default:
1591 shader_addline(buffer, " (unknown)");
1592 break;
1593 }
1594 }
1595 else
1596 {
1597 /* Pixel shaders 3.0 don't have usage semantics. */
1598 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1599 return;
1600 else
1601 shader_addline(buffer, "_");
1602
1603 switch (semantic->usage)
1604 {
1605 case WINED3D_DECL_USAGE_POSITION:
1606 shader_addline(buffer, "position%u", semantic->usage_idx);
1607 break;
1608
1609 case WINED3D_DECL_USAGE_BLEND_INDICES:
1610 shader_addline(buffer, "blend");
1611 break;
1612
1613 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1614 shader_addline(buffer, "weight");
1615 break;
1616
1617 case WINED3D_DECL_USAGE_NORMAL:
1618 shader_addline(buffer, "normal%u", semantic->usage_idx);
1619 break;
1620
1621 case WINED3D_DECL_USAGE_PSIZE:
1622 shader_addline(buffer, "psize");
1623 break;
1624
1625 case WINED3D_DECL_USAGE_COLOR:
1626 if (!semantic->usage_idx)
1627 shader_addline(buffer, "color");
1628 else
1629 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1630 break;
1631
1632 case WINED3D_DECL_USAGE_TEXCOORD:
1633 shader_addline(buffer, "texture%u", semantic->usage_idx);
1634 break;
1635
1636 case WINED3D_DECL_USAGE_TANGENT:
1637 shader_addline(buffer, "tangent");
1638 break;
1639
1640 case WINED3D_DECL_USAGE_BINORMAL:
1641 shader_addline(buffer, "binormal");
1642 break;
1643
1644 case WINED3D_DECL_USAGE_TESS_FACTOR:
1645 shader_addline(buffer, "tessfactor");
1646 break;
1647
1648 case WINED3D_DECL_USAGE_POSITIONT:
1649 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1650 break;
1651
1652 case WINED3D_DECL_USAGE_FOG:
1653 shader_addline(buffer, "fog");
1654 break;
1655
1656 case WINED3D_DECL_USAGE_DEPTH:
1657 shader_addline(buffer, "depth");
1658 break;
1659
1660 case WINED3D_DECL_USAGE_SAMPLE:
1661 shader_addline(buffer, "sample");
1662 break;
1663
1664 default:
1665 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1666 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
1667 }
1668 }
1669 }
1670
1671 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1672 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1673 {
1674 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1675 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1676 UINT offset = reg->idx[0].offset;
1677
1678 switch (reg->type)
1679 {
1680 case WINED3DSPR_TEMP:
1681 shader_addline(buffer, "r");
1682 break;
1683
1684 case WINED3DSPR_INPUT:
1685 shader_addline(buffer, "v");
1686 break;
1687
1688 case WINED3DSPR_CONST:
1689 case WINED3DSPR_CONST2:
1690 case WINED3DSPR_CONST3:
1691 case WINED3DSPR_CONST4:
1692 shader_addline(buffer, "c");
1693 offset = shader_get_float_offset(reg->type, offset);
1694 break;
1695
1696 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1697 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1698 break;
1699
1700 case WINED3DSPR_RASTOUT:
1701 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1702 break;
1703
1704 case WINED3DSPR_COLOROUT:
1705 shader_addline(buffer, "oC");
1706 break;
1707
1708 case WINED3DSPR_DEPTHOUT:
1709 shader_addline(buffer, "oDepth");
1710 break;
1711
1712 case WINED3DSPR_ATTROUT:
1713 shader_addline(buffer, "oD");
1714 break;
1715
1716 case WINED3DSPR_TEXCRDOUT:
1717 /* Vertex shaders >= 3.0 use general purpose output registers
1718 * (WINED3DSPR_OUTPUT), which can include an address token. */
1719 if (shader_version->major >= 3)
1720 shader_addline(buffer, "o");
1721 else
1722 shader_addline(buffer, "oT");
1723 break;
1724
1725 case WINED3DSPR_CONSTINT:
1726 shader_addline(buffer, "i");
1727 break;
1728
1729 case WINED3DSPR_CONSTBOOL:
1730 shader_addline(buffer, "b");
1731 break;
1732
1733 case WINED3DSPR_LABEL:
1734 shader_addline(buffer, "l");
1735 break;
1736
1737 case WINED3DSPR_LOOP:
1738 shader_addline(buffer, "aL");
1739 break;
1740
1741 case WINED3DSPR_SAMPLER:
1742 shader_addline(buffer, "s");
1743 break;
1744
1745 case WINED3DSPR_MISCTYPE:
1746 if (offset > 1)
1747 {
1748 FIXME("Unhandled misctype register %u.\n", offset);
1749 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1750 }
1751 else
1752 {
1753 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1754 }
1755 break;
1756
1757 case WINED3DSPR_PREDICATE:
1758 shader_addline(buffer, "p");
1759 break;
1760
1761 case WINED3DSPR_IMMCONST:
1762 shader_addline(buffer, "l");
1763 break;
1764
1765 case WINED3DSPR_CONSTBUFFER:
1766 shader_addline(buffer, "cb");
1767 break;
1768
1769 case WINED3DSPR_IMMCONSTBUFFER:
1770 shader_addline(buffer, "icb");
1771 break;
1772
1773 case WINED3DSPR_PRIMID:
1774 shader_addline(buffer, "primID");
1775 break;
1776
1777 case WINED3DSPR_NULL:
1778 shader_addline(buffer, "null");
1779 break;
1780
1781 case WINED3DSPR_RESOURCE:
1782 shader_addline(buffer, "t");
1783 break;
1784
1785 case WINED3DSPR_UAV:
1786 shader_addline(buffer, "u");
1787 break;
1788
1789 case WINED3DSPR_OUTPOINTID:
1790 shader_addline(buffer, "vOutputControlPointID");
1791 break;
1792
1793 case WINED3DSPR_FORKINSTID:
1794 shader_addline(buffer, "vForkInstanceId");
1795 break;
1796
1797 case WINED3DSPR_INCONTROLPOINT:
1798 shader_addline(buffer, "vicp");
1799 break;
1800
1801 case WINED3DSPR_OUTCONTROLPOINT:
1802 shader_addline(buffer, "vocp");
1803 break;
1804
1805 case WINED3DSPR_PATCHCONST:
1806 shader_addline(buffer, "vpc");
1807 break;
1808
1809 case WINED3DSPR_TESSCOORD:
1810 shader_addline(buffer, "vDomainLocation");
1811 break;
1812
1813 case WINED3DSPR_GROUPSHAREDMEM:
1814 shader_addline(buffer, "g");
1815 break;
1816
1817 case WINED3DSPR_THREADID:
1818 shader_addline(buffer, "vThreadID");
1819 break;
1820
1821 case WINED3DSPR_THREADGROUPID:
1822 shader_addline(buffer, "vThreadGroupID");
1823 break;
1824
1825 case WINED3DSPR_LOCALTHREADID:
1826 shader_addline(buffer, "vThreadIDInGroup");
1827 break;
1828
1829 case WINED3DSPR_LOCALTHREADINDEX:
1830 shader_addline(buffer, "vThreadIDInGroupFlattened");
1831 break;
1832
1833 case WINED3DSPR_IDXTEMP:
1834 shader_addline(buffer, "x");
1835 break;
1836
1837 case WINED3DSPR_STREAM:
1838 shader_addline(buffer, "m");
1839 break;
1840
1841 default:
1842 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1843 break;
1844 }
1845
1846 if (reg->type == WINED3DSPR_IMMCONST)
1847 {
1848 shader_addline(buffer, "(");
1849 switch (reg->immconst_type)
1850 {
1851 case WINED3D_IMMCONST_SCALAR:
1852 switch (reg->data_type)
1853 {
1854 case WINED3D_DATA_FLOAT:
1855 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1856 break;
1857 case WINED3D_DATA_INT:
1858 shader_addline(buffer, "%d", reg->immconst_data[0]);
1859 break;
1860 case WINED3D_DATA_RESOURCE:
1861 case WINED3D_DATA_SAMPLER:
1862 case WINED3D_DATA_UINT:
1863 shader_addline(buffer, "%u", reg->immconst_data[0]);
1864 break;
1865 default:
1866 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1867 break;
1868 }
1869 break;
1870
1871 case WINED3D_IMMCONST_VEC4:
1872 switch (reg->data_type)
1873 {
1874 case WINED3D_DATA_FLOAT:
1875 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1876 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1877 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1878 break;
1879 case WINED3D_DATA_INT:
1880 shader_addline(buffer, "%d, %d, %d, %d",
1881 reg->immconst_data[0], reg->immconst_data[1],
1882 reg->immconst_data[2], reg->immconst_data[3]);
1883 break;
1884 case WINED3D_DATA_RESOURCE:
1885 case WINED3D_DATA_SAMPLER:
1886 case WINED3D_DATA_UINT:
1887 shader_addline(buffer, "%u, %u, %u, %u",
1888 reg->immconst_data[0], reg->immconst_data[1],
1889 reg->immconst_data[2], reg->immconst_data[3]);
1890 break;
1891 default:
1892 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1893 break;
1894 }
1895 break;
1896
1897 default:
1898 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1899 break;
1900 }
1901 shader_addline(buffer, ")");
1902 }
1903 else if (reg->type != WINED3DSPR_RASTOUT
1904 && reg->type != WINED3DSPR_MISCTYPE
1905 && reg->type != WINED3DSPR_NULL)
1906 {
1907 if (offset != ~0u)
1908 {
1909 shader_addline(buffer, "[");
1910 if (reg->idx[0].rel_addr)
1911 {
1912 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1913 shader_addline(buffer, " + ");
1914 }
1915 shader_addline(buffer, "%u]", offset);
1916
1917 if (reg->idx[1].offset != ~0u)
1918 {
1919 shader_addline(buffer, "[");
1920 if (reg->idx[1].rel_addr)
1921 {
1922 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1923 shader_addline(buffer, " + ");
1924 }
1925 shader_addline(buffer, "%u]", reg->idx[1].offset);
1926 }
1927 }
1928 }
1929 }
1930
1931 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1932 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1933 {
1934 DWORD write_mask = param->write_mask;
1935
1936 shader_dump_register(buffer, &param->reg, shader_version);
1937
1938 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1939 {
1940 static const char write_mask_chars[] = "xyzw";
1941
1942 shader_addline(buffer, ".");
1943 if (write_mask & WINED3DSP_WRITEMASK_0)
1944 shader_addline(buffer, "%c", write_mask_chars[0]);
1945 if (write_mask & WINED3DSP_WRITEMASK_1)
1946 shader_addline(buffer, "%c", write_mask_chars[1]);
1947 if (write_mask & WINED3DSP_WRITEMASK_2)
1948 shader_addline(buffer, "%c", write_mask_chars[2]);
1949 if (write_mask & WINED3DSP_WRITEMASK_3)
1950 shader_addline(buffer, "%c", write_mask_chars[3]);
1951 }
1952 }
1953
1954 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1955 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1956 {
1957 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1958 DWORD swizzle = param->swizzle;
1959
1960 if (src_modifier == WINED3DSPSM_NEG
1961 || src_modifier == WINED3DSPSM_BIASNEG
1962 || src_modifier == WINED3DSPSM_SIGNNEG
1963 || src_modifier == WINED3DSPSM_X2NEG
1964 || src_modifier == WINED3DSPSM_ABSNEG)
1965 shader_addline(buffer, "-");
1966 else if (src_modifier == WINED3DSPSM_COMP)
1967 shader_addline(buffer, "1-");
1968 else if (src_modifier == WINED3DSPSM_NOT)
1969 shader_addline(buffer, "!");
1970
1971 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1972 shader_addline(buffer, "abs(");
1973
1974 shader_dump_register(buffer, &param->reg, shader_version);
1975
1976 switch (src_modifier)
1977 {
1978 case WINED3DSPSM_NONE: break;
1979 case WINED3DSPSM_NEG: break;
1980 case WINED3DSPSM_NOT: break;
1981 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1982 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1983 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1984 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1985 case WINED3DSPSM_COMP: break;
1986 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1987 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1988 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1989 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1990 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1991 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1992 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1993 }
1994
1995 if (swizzle != WINED3DSP_NOSWIZZLE)
1996 {
1997 static const char swizzle_chars[] = "xyzw";
1998 DWORD swizzle_x = swizzle & 0x03;
1999 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2000 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2001 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2002
2003 if (swizzle_x == swizzle_y
2004 && swizzle_x == swizzle_z
2005 && swizzle_x == swizzle_w)
2006 {
2007 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2008 }
2009 else
2010 {
2011 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2012 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2013 }
2014 }
2015 }
2016
2017 /* Shared code in order to generate the bulk of the shader string.
2018 * NOTE: A description of how to parse tokens can be found on MSDN. */
2019 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2020 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
2021 {
2022 struct wined3d_device *device = shader->device;
2023 const struct wined3d_shader_frontend *fe = shader->frontend;
2024 void *fe_data = shader->frontend_data;
2025 struct wined3d_shader_version shader_version;
2026 struct wined3d_shader_loop_state loop_state;
2027 struct wined3d_shader_instruction ins;
2028 struct wined3d_shader_tex_mx tex_mx;
2029 struct wined3d_shader_context ctx;
2030 const DWORD *ptr = byte_code;
2031
2032 /* Initialize current parsing state. */
2033 tex_mx.current_row = 0;
2034 loop_state.current_depth = 0;
2035 loop_state.current_reg = 0;
2036
2037 ctx.shader = shader;
2038 ctx.gl_info = &device->adapter->gl_info;
2039 ctx.reg_maps = reg_maps;
2040 ctx.buffer = buffer;
2041 ctx.tex_mx = &tex_mx;
2042 ctx.loop_state = &loop_state;
2043 ctx.backend_data = backend_ctx;
2044 ins.ctx = &ctx;
2045
2046 fe->shader_read_header(fe_data, &ptr, &shader_version);
2047
2048 while (!fe->shader_is_end(fe_data, &ptr))
2049 {
2050 /* Read opcode. */
2051 fe->shader_read_instruction(fe_data, &ptr, &ins);
2052
2053 /* Unknown opcode and its parameters. */
2054 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2055 {
2056 TRACE("Skipping unrecognized instruction.\n");
2057 continue;
2058 }
2059
2060 if (ins.predicate)
2061 FIXME("Predicates not implemented.\n");
2062
2063 /* Call appropriate function for output target */
2064 device->shader_backend->shader_handle_instruction(&ins);
2065 }
2066 }
2067
2068 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2069 const struct wined3d_shader_dst_param *dst)
2070 {
2071 DWORD mmask = dst->modifiers;
2072
2073 switch (dst->shift)
2074 {
2075 case 0: break;
2076 case 13: shader_addline(buffer, "_d8"); break;
2077 case 14: shader_addline(buffer, "_d4"); break;
2078 case 15: shader_addline(buffer, "_d2"); break;
2079 case 1: shader_addline(buffer, "_x2"); break;
2080 case 2: shader_addline(buffer, "_x4"); break;
2081 case 3: shader_addline(buffer, "_x8"); break;
2082 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2083 }
2084
2085 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2086 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2087 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2088
2089 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2090 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2091 }
2092
2093 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2094 enum wined3d_primitive_type primitive_type)
2095 {
2096 switch (primitive_type)
2097 {
2098 case WINED3D_PT_UNDEFINED:
2099 shader_addline(buffer, "undefined");
2100 break;
2101 case WINED3D_PT_POINTLIST:
2102 shader_addline(buffer, "pointlist");
2103 break;
2104 case WINED3D_PT_LINELIST:
2105 shader_addline(buffer, "linelist");
2106 break;
2107 case WINED3D_PT_LINESTRIP:
2108 shader_addline(buffer, "linestrip");
2109 break;
2110 case WINED3D_PT_TRIANGLELIST:
2111 shader_addline(buffer, "trianglelist");
2112 break;
2113 case WINED3D_PT_TRIANGLESTRIP:
2114 shader_addline(buffer, "trianglestrip");
2115 break;
2116 case WINED3D_PT_TRIANGLEFAN:
2117 shader_addline(buffer, "trianglefan");
2118 break;
2119 case WINED3D_PT_LINELIST_ADJ:
2120 shader_addline(buffer, "linelist_adj");
2121 break;
2122 case WINED3D_PT_LINESTRIP_ADJ:
2123 shader_addline(buffer, "linestrip_adj");
2124 break;
2125 case WINED3D_PT_TRIANGLELIST_ADJ:
2126 shader_addline(buffer, "trianglelist_adj");
2127 break;
2128 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2129 shader_addline(buffer, "trianglestrip_adj");
2130 break;
2131 default:
2132 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2133 break;
2134 }
2135 }
2136
2137 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2138 enum wined3d_shader_interpolation_mode interpolation_mode)
2139 {
2140 switch (interpolation_mode)
2141 {
2142 case WINED3DSIM_CONSTANT:
2143 shader_addline(buffer, "constant");
2144 break;
2145 case WINED3DSIM_LINEAR:
2146 shader_addline(buffer, "linear");
2147 break;
2148 case WINED3DSIM_LINEAR_CENTROID:
2149 shader_addline(buffer, "linear centroid");
2150 break;
2151 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2152 shader_addline(buffer, "linear noperspective");
2153 break;
2154 case WINED3DSIM_LINEAR_SAMPLE:
2155 shader_addline(buffer, "linear sample");
2156 break;
2157 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2158 shader_addline(buffer, "linear noperspective centroid");
2159 break;
2160 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2161 shader_addline(buffer, "linear noperspective sample");
2162 break;
2163 default:
2164 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2165 break;
2166 }
2167 }
2168
2169 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2170 {
2171 struct wined3d_shader_version shader_version;
2172 struct wined3d_string_buffer buffer;
2173 const DWORD *ptr = byte_code;
2174 const char *type_prefix;
2175 const char *p, *q;
2176 DWORD i;
2177
2178 if (!string_buffer_init(&buffer))
2179 {
2180 ERR("Failed to initialize string buffer.\n");
2181 return;
2182 }
2183
2184 TRACE("Parsing %p.\n", byte_code);
2185
2186 fe->shader_read_header(fe_data, &ptr, &shader_version);
2187
2188 switch (shader_version.type)
2189 {
2190 case WINED3D_SHADER_TYPE_VERTEX:
2191 type_prefix = "vs";
2192 break;
2193
2194 case WINED3D_SHADER_TYPE_HULL:
2195 type_prefix = "hs";
2196 break;
2197
2198 case WINED3D_SHADER_TYPE_DOMAIN:
2199 type_prefix = "ds";
2200 break;
2201
2202 case WINED3D_SHADER_TYPE_GEOMETRY:
2203 type_prefix = "gs";
2204 break;
2205
2206 case WINED3D_SHADER_TYPE_PIXEL:
2207 type_prefix = "ps";
2208 break;
2209
2210 case WINED3D_SHADER_TYPE_COMPUTE:
2211 type_prefix = "cs";
2212 break;
2213
2214 default:
2215 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2216 type_prefix = "unknown";
2217 break;
2218 }
2219
2220 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2221
2222 while (!fe->shader_is_end(fe_data, &ptr))
2223 {
2224 struct wined3d_shader_instruction ins;
2225
2226 fe->shader_read_instruction(fe_data, &ptr, &ins);
2227 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2228 {
2229 WARN("Skipping unrecognized instruction.\n");
2230 shader_addline(&buffer, "<unrecognized instruction>\n");
2231 continue;
2232 }
2233
2234 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2235 {
2236 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2237 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2238 shader_addline(&buffer, " ");
2239 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2240 }
2241 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2242 {
2243 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2244 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2245 shader_addline(&buffer, ", %s",
2246 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2247 }
2248 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2249 {
2250 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2251 shader_dump_global_flags(&buffer, ins.flags);
2252 }
2253 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2254 {
2255 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2256 ins.declaration.max_tessellation_factor);
2257 }
2258 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2259 {
2260 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2261 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2262 {
2263 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2264 ins.declaration.icb->data[4 * i + 0],
2265 ins.declaration.icb->data[4 * i + 1],
2266 ins.declaration.icb->data[4 * i + 2],
2267 ins.declaration.icb->data[4 * i + 3]);
2268 }
2269 shader_addline(&buffer, "}");
2270 }
2271 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2272 {
2273 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2274 ins.declaration.indexable_temp.register_idx,
2275 ins.declaration.indexable_temp.register_size,
2276 ins.declaration.indexable_temp.component_count);
2277 }
2278 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2279 {
2280 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2281 shader_dump_interpolation_mode(&buffer, ins.flags);
2282 shader_addline(&buffer, " ");
2283 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2284 }
2285 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2286 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2287 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2288 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2289 {
2290 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2291 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2292 shader_addline(&buffer, ", ");
2293 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2294 }
2295 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2296 {
2297 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2298 shader_dump_interpolation_mode(&buffer, ins.flags);
2299 shader_addline(&buffer, " ");
2300 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2301 shader_addline(&buffer, ", ");
2302 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2303 }
2304 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2305 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2306 {
2307 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2308 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2309 }
2310 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2311 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2312 {
2313 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2314 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2315 }
2316 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2317 {
2318 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2319 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2320 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2321 }
2322 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2323 {
2324 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2325 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2326 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2327 shader_addline(&buffer, ", comparisonMode");
2328 }
2329 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2330 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2331 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2332 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2333 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2334 {
2335 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2336 }
2337 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2338 {
2339 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2340 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2341 }
2342 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2343 {
2344 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2345 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2346 }
2347 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2348 {
2349 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2350 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2351 }
2352 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2353 {
2354 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2355 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2356 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2357 }
2358 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2359 {
2360 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2361 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2362 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2363 ins.declaration.tgsm_structured.structure_count);
2364 }
2365 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2366 {
2367 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2368 ins.declaration.thread_group_size.x,
2369 ins.declaration.thread_group_size.y,
2370 ins.declaration.thread_group_size.z);
2371 }
2372 else if (ins.handler_idx == WINED3DSIH_DEF)
2373 {
2374 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2375 ins.dst[0].reg.idx[0].offset),
2376 *(const float *)&ins.src[0].reg.immconst_data[0],
2377 *(const float *)&ins.src[0].reg.immconst_data[1],
2378 *(const float *)&ins.src[0].reg.immconst_data[2],
2379 *(const float *)&ins.src[0].reg.immconst_data[3]);
2380 }
2381 else if (ins.handler_idx == WINED3DSIH_DEFI)
2382 {
2383 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2384 ins.src[0].reg.immconst_data[0],
2385 ins.src[0].reg.immconst_data[1],
2386 ins.src[0].reg.immconst_data[2],
2387 ins.src[0].reg.immconst_data[3]);
2388 }
2389 else if (ins.handler_idx == WINED3DSIH_DEFB)
2390 {
2391 shader_addline(&buffer, "defb b%u = %s",
2392 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2393 }
2394 else
2395 {
2396 if (ins.predicate)
2397 {
2398 shader_addline(&buffer, "(");
2399 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2400 shader_addline(&buffer, ") ");
2401 }
2402
2403 /* PixWin marks instructions with the coissue flag with a '+' */
2404 if (ins.coissue)
2405 shader_addline(&buffer, "+");
2406
2407 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2408
2409 if (ins.handler_idx == WINED3DSIH_BREAKP
2410 || ins.handler_idx == WINED3DSIH_IF)
2411 {
2412 switch (ins.flags)
2413 {
2414 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2415 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2416 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2417 }
2418 }
2419 else if (ins.handler_idx == WINED3DSIH_IFC
2420 || ins.handler_idx == WINED3DSIH_BREAKC)
2421 {
2422 switch (ins.flags)
2423 {
2424 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2425 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2426 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2427 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2428 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2429 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2430 default: shader_addline(&buffer, "_(%u)", ins.flags);
2431 }
2432 }
2433 else if (ins.handler_idx == WINED3DSIH_TEX
2434 && shader_version.major >= 2
2435 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2436 {
2437 shader_addline(&buffer, "p");
2438 }
2439 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2440 {
2441 switch (ins.flags)
2442 {
2443 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2444 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2445 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2446 }
2447 }
2448 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2449 {
2450 switch (ins.flags)
2451 {
2452 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2453 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2454 }
2455 }
2456
2457 if (wined3d_shader_instruction_has_texel_offset(&ins))
2458 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2459
2460 for (i = 0; i < ins.dst_count; ++i)
2461 {
2462 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2463 shader_addline(&buffer, !i ? " " : ", ");
2464 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2465 }
2466
2467 /* Other source tokens */
2468 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2469 {
2470 shader_addline(&buffer, !i ? " " : ", ");
2471 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2472 }
2473 }
2474 shader_addline(&buffer, "\n");
2475 }
2476
2477 for (p = buffer.buffer; *p; p = q)
2478 {
2479 if (!(q = strstr(p, "\n")))
2480 q = p + strlen(p);
2481 else
2482 ++q;
2483 TRACE(" %.*s", (int)(q - p), p);
2484 }
2485
2486 string_buffer_free(&buffer);
2487 }
2488
2489 static void shader_cleanup(struct wined3d_shader *shader)
2490 {
2491 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2492 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2493 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2494 shader->device->shader_backend->shader_destroy(shader);
2495 shader_cleanup_reg_maps(&shader->reg_maps);
2496 HeapFree(GetProcessHeap(), 0, shader->function);
2497 shader_delete_constant_list(&shader->constantsF);
2498 shader_delete_constant_list(&shader->constantsB);
2499 shader_delete_constant_list(&shader->constantsI);
2500 list_remove(&shader->shader_list_entry);
2501
2502 if (shader->frontend && shader->frontend_data)
2503 shader->frontend->shader_free(shader->frontend_data);
2504 }
2505
2506 struct shader_none_priv
2507 {
2508 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2509 const struct fragment_pipeline *fragment_pipe;
2510 BOOL ffp_proj_control;
2511 };
2512
2513 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2514 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2515 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2516 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2517 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2518 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2519 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2520 const struct wined3d_state *state) {}
2521 static void shader_none_destroy(struct wined3d_shader *shader) {}
2522 static void shader_none_free_context_data(struct wined3d_context *context) {}
2523 static void shader_none_init_context_state(struct wined3d_context *context) {}
2524
2525 /* Context activation is done by the caller. */
2526 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2527 const struct wined3d_state *state)
2528 {
2529 const struct wined3d_gl_info *gl_info = context->gl_info;
2530 struct shader_none_priv *priv = shader_priv;
2531
2532 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2533 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2534 }
2535
2536 /* Context activation is done by the caller. */
2537 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2538 {
2539 struct shader_none_priv *priv = shader_priv;
2540 const struct wined3d_gl_info *gl_info = context->gl_info;
2541
2542 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2543 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2544
2545 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2546 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2547 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2548 | (1u << WINED3D_SHADER_TYPE_HULL)
2549 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2550 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2551 }
2552
2553 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2554 const struct fragment_pipeline *fragment_pipe)
2555 {
2556 struct fragment_caps fragment_caps;
2557 void *vertex_priv, *fragment_priv;
2558 struct shader_none_priv *priv;
2559
2560 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2561 return E_OUTOFMEMORY;
2562
2563 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2564 {
2565 ERR("Failed to initialize vertex pipe.\n");
2566 HeapFree(GetProcessHeap(), 0, priv);
2567 return E_FAIL;
2568 }
2569
2570 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2571 {
2572 ERR("Failed to initialize fragment pipe.\n");
2573 vertex_pipe->vp_free(device);
2574 HeapFree(GetProcessHeap(), 0, priv);
2575 return E_FAIL;
2576 }
2577
2578 priv->vertex_pipe = vertex_pipe;
2579 priv->fragment_pipe = fragment_pipe;
2580 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2581 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2582
2583 device->vertex_priv = vertex_priv;
2584 device->fragment_priv = fragment_priv;
2585 device->shader_priv = priv;
2586
2587 return WINED3D_OK;
2588 }
2589
2590 static void shader_none_free(struct wined3d_device *device)
2591 {
2592 struct shader_none_priv *priv = device->shader_priv;
2593
2594 priv->fragment_pipe->free_private(device);
2595 priv->vertex_pipe->vp_free(device);
2596 HeapFree(GetProcessHeap(), 0, priv);
2597 }
2598
2599 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2600 {
2601 return TRUE;
2602 }
2603
2604 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2605 {
2606 /* Set the shader caps to 0 for the none shader backend */
2607 caps->vs_version = 0;
2608 caps->hs_version = 0;
2609 caps->ds_version = 0;
2610 caps->gs_version = 0;
2611 caps->ps_version = 0;
2612 caps->cs_version = 0;
2613 caps->vs_uniform_count = 0;
2614 caps->ps_uniform_count = 0;
2615 caps->ps_1x_max_value = 0.0f;
2616 caps->varying_count = 0;
2617 caps->wined3d_caps = 0;
2618 }
2619
2620 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2621 {
2622 /* We "support" every possible fixup, since we don't support any shader
2623 * model, and will never have to actually sample a texture. */
2624 return TRUE;
2625 }
2626
2627 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2628 {
2629 struct shader_none_priv *priv = shader_priv;
2630
2631 return priv->ffp_proj_control;
2632 }
2633
2634 const struct wined3d_shader_backend_ops none_shader_backend =
2635 {
2636 shader_none_handle_instruction,
2637 shader_none_select,
2638 shader_none_disable,
2639 shader_none_select_depth_blt,
2640 shader_none_deselect_depth_blt,
2641 shader_none_update_float_vertex_constants,
2642 shader_none_update_float_pixel_constants,
2643 shader_none_load_constants,
2644 shader_none_destroy,
2645 shader_none_alloc,
2646 shader_none_free,
2647 shader_none_allocate_context_data,
2648 shader_none_free_context_data,
2649 shader_none_init_context_state,
2650 shader_none_get_caps,
2651 shader_none_color_fixup_supported,
2652 shader_none_has_ffp_proj_control,
2653 };
2654
2655 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2656 DWORD float_const_count, enum wined3d_shader_type type, unsigned int max_version)
2657 {
2658 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2659 const struct wined3d_shader_frontend *fe;
2660 HRESULT hr;
2661 unsigned int backend_version;
2662 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2663
2664 TRACE("shader %p, byte_code %p, float_const_count %u, type %#x, max_version %u.\n",
2665 shader, byte_code, float_const_count, type, max_version);
2666
2667 list_init(&shader->constantsF);
2668 list_init(&shader->constantsB);
2669 list_init(&shader->constantsI);
2670 shader->lconst_inf_or_nan = FALSE;
2671 list_init(&reg_maps->indexable_temps);
2672
2673 fe = shader_select_frontend(*byte_code);
2674 if (!fe)
2675 {
2676 FIXME("Unable to find frontend for shader.\n");
2677 return WINED3DERR_INVALIDCALL;
2678 }
2679 shader->frontend = fe;
2680 shader->frontend_data = fe->shader_init(byte_code, &shader->output_signature);
2681 if (!shader->frontend_data)
2682 {
2683 FIXME("Failed to initialize frontend.\n");
2684 return WINED3DERR_INVALIDCALL;
2685 }
2686
2687 /* First pass: trace shader. */
2688 if (TRACE_ON(d3d_shader))
2689 shader_trace_init(fe, shader->frontend_data, byte_code);
2690
2691 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2692 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2693 &shader->output_signature, byte_code, float_const_count)))
2694 return hr;
2695
2696 if (reg_maps->shader_version.type != type)
2697 {
2698 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2699 return WINED3DERR_INVALIDCALL;
2700 }
2701 if (reg_maps->shader_version.major > max_version)
2702 {
2703 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2704 return WINED3DERR_INVALIDCALL;
2705 }
2706 switch (type)
2707 {
2708 case WINED3D_SHADER_TYPE_VERTEX:
2709 backend_version = d3d_info->limits.vs_version;
2710 break;
2711 case WINED3D_SHADER_TYPE_HULL:
2712 backend_version = d3d_info->limits.hs_version;
2713 break;
2714 case WINED3D_SHADER_TYPE_DOMAIN:
2715 backend_version = d3d_info->limits.ds_version;
2716 break;
2717 case WINED3D_SHADER_TYPE_GEOMETRY:
2718 backend_version = d3d_info->limits.gs_version;
2719 break;
2720 case WINED3D_SHADER_TYPE_PIXEL:
2721 backend_version = d3d_info->limits.ps_version;
2722 break;
2723 case WINED3D_SHADER_TYPE_COMPUTE:
2724 backend_version = d3d_info->limits.cs_version;
2725 break;
2726 default:
2727 FIXME("No backend version-checking for this shader type.\n");
2728 backend_version = 0;
2729 }
2730 if (reg_maps->shader_version.major > backend_version)
2731 {
2732 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2733 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2734 return WINED3DERR_INVALIDCALL;
2735 }
2736
2737 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength)))
2738 return E_OUTOFMEMORY;
2739 memcpy(shader->function, byte_code, shader->functionLength);
2740
2741 return WINED3D_OK;
2742 }
2743
2744 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2745 {
2746 ULONG refcount = InterlockedIncrement(&shader->ref);
2747
2748 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2749
2750 return refcount;
2751 }
2752
2753 static void wined3d_shader_destroy_object(void *object)
2754 {
2755 shader_cleanup(object);
2756 HeapFree(GetProcessHeap(), 0, object);
2757 }
2758
2759 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2760 {
2761 ULONG refcount = InterlockedDecrement(&shader->ref);
2762
2763 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2764
2765 if (!refcount)
2766 {
2767 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2768 wined3d_cs_emit_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2769 }
2770
2771 return refcount;
2772 }
2773
2774 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2775 {
2776 TRACE("shader %p.\n", shader);
2777
2778 return shader->parent;
2779 }
2780
2781 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2782 void *byte_code, UINT *byte_code_size)
2783 {
2784 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2785
2786 if (!byte_code)
2787 {
2788 *byte_code_size = shader->functionLength;
2789 return WINED3D_OK;
2790 }
2791
2792 if (*byte_code_size < shader->functionLength)
2793 {
2794 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2795 * than the required size we should write the required size and
2796 * return D3DERR_MOREDATA. That's not actually true. */
2797 return WINED3DERR_INVALIDCALL;
2798 }
2799
2800 memcpy(byte_code, shader->function, shader->functionLength);
2801
2802 return WINED3D_OK;
2803 }
2804
2805 /* Set local constants for d3d8 shaders. */
2806 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2807 UINT start_idx, const float *src_data, UINT count)
2808 {
2809 UINT end_idx = start_idx + count;
2810 UINT i;
2811
2812 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2813
2814 if (end_idx > shader->limits->constant_float)
2815 {
2816 WARN("end_idx %u > float constants limit %u.\n",
2817 end_idx, shader->limits->constant_float);
2818 end_idx = shader->limits->constant_float;
2819 }
2820
2821 for (i = start_idx; i < end_idx; ++i)
2822 {
2823 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2824 float *value;
2825 if (!lconst)
2826 return E_OUTOFMEMORY;
2827
2828 lconst->idx = i;
2829 value = (float *)lconst->value;
2830 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2831 list_add_head(&shader->constantsF, &lconst->entry);
2832
2833 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2834 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2835 {
2836 shader->lconst_inf_or_nan = TRUE;
2837 }
2838 }
2839
2840 return WINED3D_OK;
2841 }
2842
2843 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2844 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2845 {
2846 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2847 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2848 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2849 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2850 args->point_size = state->gl_primitive_type == GL_POINTS;
2851 args->per_vertex_point_size = shader->reg_maps.point_size;
2852 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
2853 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
2854 : WINED3D_SHADER_TYPE_PIXEL;
2855 if (shader->reg_maps.shader_version.major >= 4)
2856 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
2857 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
2858 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
2859 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
2860 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
2861 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
2862 else
2863 args->next_shader_input_count = 0;
2864 args->swizzle_map = swizzle_map;
2865 if (d3d_info->emulated_flatshading)
2866 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2867 else
2868 args->flatshading = 0;
2869 }
2870
2871 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2872 {
2873 if (usage_idx1 != usage_idx2)
2874 return FALSE;
2875 if (usage1 == usage2)
2876 return TRUE;
2877 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2878 return TRUE;
2879 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2880 return TRUE;
2881
2882 return FALSE;
2883 }
2884
2885 BOOL vshader_get_input(const struct wined3d_shader *shader,
2886 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2887 {
2888 WORD map = shader->reg_maps.input_registers;
2889 unsigned int i;
2890
2891 for (i = 0; map; map >>= 1, ++i)
2892 {
2893 if (!(map & 1)) continue;
2894
2895 if (match_usage(shader->u.vs.attributes[i].usage,
2896 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2897 {
2898 *regnum = i;
2899 return TRUE;
2900 }
2901 }
2902 return FALSE;
2903 }
2904
2905 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2906 const struct wined3d_shader_signature *src, char **signature_strings)
2907 {
2908 struct wined3d_shader_signature_element *e;
2909 unsigned int i;
2910 SIZE_T len;
2911 char *ptr;
2912
2913 if (!src->element_count)
2914 return WINED3D_OK;
2915
2916 ptr = *signature_strings;
2917
2918 dst->element_count = src->element_count;
2919 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
2920 return E_OUTOFMEMORY;
2921
2922 for (i = 0; i < src->element_count; ++i)
2923 {
2924 e = &src->elements[i];
2925 dst->elements[i] = *e;
2926
2927 len = strlen(e->semantic_name);
2928 memcpy(ptr, e->semantic_name, len + 1);
2929 dst->elements[i].semantic_name = ptr;
2930 ptr += len + 1;
2931 }
2932
2933 *signature_strings = ptr;
2934
2935 return WINED3D_OK;
2936 }
2937
2938 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2939 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2940 void *parent, const struct wined3d_parent_ops *parent_ops)
2941 {
2942 struct wined3d_shader_signature_element *e;
2943 SIZE_T total, len;
2944 unsigned int i;
2945 HRESULT hr;
2946 char *ptr;
2947
2948 if (!desc->byte_code)
2949 return WINED3DERR_INVALIDCALL;
2950
2951 shader->ref = 1;
2952 shader->device = device;
2953 shader->parent = parent;
2954 shader->parent_ops = parent_ops;
2955
2956 total = 0;
2957 for (i = 0; i < desc->input_signature.element_count; ++i)
2958 {
2959 e = &desc->input_signature.elements[i];
2960 len = strlen(e->semantic_name);
2961 if (len >= ~(SIZE_T)0 - total)
2962 return E_OUTOFMEMORY;
2963
2964 total += len + 1;
2965 }
2966 for (i = 0; i < desc->output_signature.element_count; ++i)
2967 {
2968 e = &desc->output_signature.elements[i];
2969 len = strlen(e->semantic_name);
2970 if (len >= ~(SIZE_T)0 - total)
2971 return E_OUTOFMEMORY;
2972
2973 total += len + 1;
2974 }
2975 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2976 return E_OUTOFMEMORY;
2977 ptr = shader->signature_strings;
2978
2979 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
2980 {
2981 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2982 return hr;
2983 }
2984 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
2985 {
2986 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2987 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2988 return hr;
2989 }
2990
2991 list_init(&shader->linked_programs);
2992 list_add_head(&device->shaders, &shader->shader_list_entry);
2993
2994 if (FAILED(hr = shader_set_function(shader, desc->byte_code,
2995 float_const_count, type, desc->max_version)))
2996 {
2997 WARN("Failed to set function, hr %#x.\n", hr);
2998 shader_cleanup(shader);
2999 }
3000
3001 shader->load_local_constsF = shader->lconst_inf_or_nan;
3002
3003 return hr;
3004 }
3005
3006 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3007 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3008 {
3009 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3010 unsigned int i;
3011 HRESULT hr;
3012
3013 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3014 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3015 return hr;
3016
3017 for (i = 0; i < shader->input_signature.element_count; ++i)
3018 {
3019 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3020
3021 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3022 continue;
3023
3024 shader->u.vs.attributes[input->register_idx].usage =
3025 shader_usage_from_semantic_name(input->semantic_name);
3026 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3027 }
3028
3029 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3030 shader->load_local_constsF = TRUE;
3031
3032 return WINED3D_OK;
3033 }
3034
3035 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3036 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3037 {
3038 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3039 }
3040
3041 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3042 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3043 {
3044 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3045 }
3046
3047 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3048 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3049 {
3050 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3051 }
3052
3053 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3054 struct gs_compile_args *args)
3055 {
3056 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3057 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3058 }
3059
3060 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3061 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3062 {
3063 const struct wined3d_gl_info *gl_info = context->gl_info;
3064 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3065 const struct wined3d_texture *texture;
3066 UINT i;
3067
3068 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3069 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3070 {
3071 static unsigned int warned = 0;
3072
3073 args->srgb_correction = 1;
3074 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3075 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3076 "support, expect rendering artifacts.\n");
3077 }
3078
3079 if (shader->reg_maps.shader_version.major == 1
3080 && shader->reg_maps.shader_version.minor <= 3)
3081 {
3082 for (i = 0; i < shader->limits->sampler; ++i)
3083 {
3084 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3085
3086 if (flags & WINED3D_TTFF_PROJECTED)
3087 {
3088 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3089
3090 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3091 {
3092 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3093 unsigned int j;
3094 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3095 DWORD max_valid = WINED3D_TTFF_COUNT4;
3096
3097 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3098 {
3099 struct wined3d_vertex_declaration_element *element =
3100 &state->vertex_declaration->elements[j];
3101
3102 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3103 && element->usage_idx == index)
3104 {
3105 max_valid = element->format->component_count;
3106 break;
3107 }
3108 }
3109 if (!tex_transform || tex_transform > max_valid)
3110 {
3111 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3112 tex_transform, max_valid);
3113 tex_transform = max_valid;
3114 }
3115 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3116 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3117 && tex_transform > WINED3D_TTFF_COUNT2)
3118 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3119 && tex_transform > WINED3D_TTFF_COUNT3))
3120 tex_transform |= WINED3D_PSARGS_PROJECTED;
3121 else
3122 {
3123 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3124 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3125 i, tex_transform, resource_type);
3126 }
3127 }
3128 else
3129 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3130
3131 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3132 }
3133 }
3134 }
3135 if (shader->reg_maps.shader_version.major == 1
3136 && shader->reg_maps.shader_version.minor <= 4)
3137 {
3138 for (i = 0; i < shader->limits->sampler; ++i)
3139 {
3140 const struct wined3d_texture *texture = state->textures[i];
3141
3142 if (!shader->reg_maps.resource_info[i].type)
3143 continue;
3144
3145 /* Treat unbound textures as 2D. The dummy texture will provide
3146 * the proper sample value. The tex_types bitmap defaults to
3147 * 2D because of the memset. */
3148 if (!texture)
3149 continue;
3150
3151 switch (texture->target)
3152 {
3153 /* RECT textures are distinguished from 2D textures via np2_fixup */
3154 case GL_TEXTURE_RECTANGLE_ARB:
3155 case GL_TEXTURE_2D:
3156 break;
3157
3158 case GL_TEXTURE_3D:
3159 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3160 break;
3161
3162 case GL_TEXTURE_CUBE_MAP_ARB:
3163 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3164 break;
3165 }
3166 }
3167 }
3168
3169 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3170 {
3171 if (!shader->reg_maps.resource_info[i].type)
3172 continue;
3173
3174 texture = state->textures[i];
3175 if (!texture)
3176 {
3177 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3178 continue;
3179 }
3180 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3181 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3182 else
3183 args->color_fixup[i] = texture->resource.format->color_fixup;
3184
3185 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3186 args->shadow |= 1u << i;
3187
3188 /* Flag samplers that need NP2 texcoord fixup. */
3189 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3190 args->np2_fixup |= (1u << i);
3191 }
3192
3193 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3194 if (shader->reg_maps.shader_version.major >= 4)
3195 args->shadow = 0;
3196
3197 if (shader->reg_maps.shader_version.major >= 3)
3198 {
3199 if (position_transformed)
3200 args->vp_mode = pretransformed;
3201 else if (use_vs(state))
3202 args->vp_mode = vertexshader;
3203 else
3204 args->vp_mode = fixedfunction;
3205 args->fog = WINED3D_FFP_PS_FOG_OFF;
3206 }
3207 else
3208 {
3209 args->vp_mode = vertexshader;
3210 if (state->render_states[WINED3D_RS_FOGENABLE])
3211 {
3212 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3213 {
3214 case WINED3D_FOG_NONE:
3215 if (position_transformed || use_vs(state))
3216 {
3217 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3218 break;
3219 }
3220
3221 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3222 {
3223 case WINED3D_FOG_NONE: /* Fall through. */
3224 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3225 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3226 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3227 }
3228 break;
3229
3230 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3231 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3232 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3233 }
3234 }
3235 else
3236 {
3237 args->fog = WINED3D_FFP_PS_FOG_OFF;
3238 }
3239 }
3240
3241 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3242 {
3243 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3244
3245 args->texcoords_initialized = 0;
3246 for (i = 0; i < MAX_TEXTURES; ++i)
3247 {
3248 if (vs)
3249 {
3250 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3251 args->texcoords_initialized |= 1u << i;
3252 }
3253 else
3254 {
3255 const struct wined3d_stream_info *si = &context->stream_info;
3256 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3257
3258 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3259 & WINED3D_FFP_TCI_MASK
3260 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3261 args->texcoords_initialized |= 1u << i;
3262 }
3263 }
3264 }
3265 else
3266 {
3267 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3268 }
3269
3270 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3271 && state->gl_primitive_type == GL_POINTS;
3272
3273 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3274 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3275 else
3276 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3277 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3278 : WINED3D_CMP_ALWAYS) - 1;
3279
3280 if (d3d_info->emulated_flatshading)
3281 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3282
3283 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3284 ? context->render_offscreen : 0;
3285 }
3286
3287 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3288 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3289 {
3290 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3291 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3292 HRESULT hr;
3293
3294 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3295 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3296 return hr;
3297
3298 for (i = 0; i < MAX_REG_INPUT; ++i)
3299 {
3300 if (shader->u.ps.input_reg_used[i])
3301 {
3302 ++num_regs_used;
3303 highest_reg_used = i;
3304 }
3305 }
3306
3307 /* Don't do any register mapping magic if it is not needed, or if we can't
3308 * achieve anything anyway */
3309 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3310 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3311 || shader->reg_maps.shader_version.major >= 4)
3312 {
3313 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3314 {
3315 /* This happens with relative addressing. The input mapper function
3316 * warns about this if the higher registers are declared too, so
3317 * don't write a FIXME here */
3318 WARN("More varying registers used than supported\n");
3319 }
3320
3321 for (i = 0; i < MAX_REG_INPUT; ++i)
3322 {
3323 shader->u.ps.input_reg_map[i] = i;
3324 }
3325
3326 shader->u.ps.declared_in_count = highest_reg_used + 1;
3327 }
3328 else
3329 {
3330 shader->u.ps.declared_in_count = 0;
3331 for (i = 0; i < MAX_REG_INPUT; ++i)
3332 {
3333 if (shader->u.ps.input_reg_used[i])
3334 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3335 else shader->u.ps.input_reg_map[i] = ~0U;
3336 }
3337 }
3338
3339 return WINED3D_OK;
3340 }
3341
3342 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3343 {
3344 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3345 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3346 unsigned int i;
3347
3348 if (reg_maps->shader_version.major != 1) return;
3349
3350 for (i = 0; i < shader->limits->sampler; ++i)
3351 {
3352 /* We don't sample from this sampler. */
3353 if (!resource_info[i].type)
3354 continue;
3355
3356 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3357 {
3358 case WINED3D_SHADER_TEX_2D:
3359 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3360 break;
3361
3362 case WINED3D_SHADER_TEX_3D:
3363 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3364 break;
3365
3366 case WINED3D_SHADER_TEX_CUBE:
3367 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3368 break;
3369 }
3370 }
3371 }
3372
3373 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3374 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3375 {
3376 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3377 }
3378
3379 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3380 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3381 {
3382 struct wined3d_shader *object;
3383 HRESULT hr;
3384
3385 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3386 device, desc, parent, parent_ops, shader);
3387
3388 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3389 return E_OUTOFMEMORY;
3390
3391 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3392 {
3393 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3394 HeapFree(GetProcessHeap(), 0, object);
3395 return hr;
3396 }
3397
3398 TRACE("Created compute shader %p.\n", object);
3399 *shader = object;
3400
3401 return WINED3D_OK;
3402 }
3403
3404 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3405 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3406 {
3407 struct wined3d_shader *object;
3408 HRESULT hr;
3409
3410 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3411 device, desc, parent, parent_ops, shader);
3412
3413 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3414 return E_OUTOFMEMORY;
3415
3416 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3417 {
3418 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3419 HeapFree(GetProcessHeap(), 0, object);
3420 return hr;
3421 }
3422
3423 TRACE("Created domain shader %p.\n", object);
3424 *shader = object;
3425
3426 return WINED3D_OK;
3427 }
3428
3429 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3430 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3431 {
3432 struct wined3d_shader *object;
3433 HRESULT hr;
3434
3435 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3436 device, desc, parent, parent_ops, shader);
3437
3438 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3439 if (!object)
3440 return E_OUTOFMEMORY;
3441
3442 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3443 {
3444 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3445 HeapFree(GetProcessHeap(), 0, object);
3446 return hr;
3447 }
3448
3449 TRACE("Created geometry shader %p.\n", object);
3450 *shader = object;
3451
3452 return WINED3D_OK;
3453 }
3454
3455 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3456 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3457 {
3458 struct wined3d_shader *object;
3459 HRESULT hr;
3460
3461 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3462 device, desc, parent, parent_ops, shader);
3463
3464 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3465 return E_OUTOFMEMORY;
3466
3467 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3468 {
3469 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3470 HeapFree(GetProcessHeap(), 0, object);
3471 return hr;
3472 }
3473
3474 TRACE("Created hull shader %p.\n", object);
3475 *shader = object;
3476
3477 return WINED3D_OK;
3478 }
3479
3480 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3481 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3482 {
3483 struct wined3d_shader *object;
3484 HRESULT hr;
3485
3486 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3487 device, desc, parent, parent_ops, shader);
3488
3489 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3490 if (!object)
3491 return E_OUTOFMEMORY;
3492
3493 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3494 {
3495 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3496 HeapFree(GetProcessHeap(), 0, object);
3497 return hr;
3498 }
3499
3500 TRACE("Created pixel shader %p.\n", object);
3501 *shader = object;
3502
3503 return WINED3D_OK;
3504 }
3505
3506 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3507 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3508 {
3509 struct wined3d_shader *object;
3510 HRESULT hr;
3511
3512 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3513 device, desc, parent, parent_ops, shader);
3514
3515 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3516 if (!object)
3517 return E_OUTOFMEMORY;
3518
3519 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3520 {
3521 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3522 HeapFree(GetProcessHeap(), 0, object);
3523 return hr;
3524 }
3525
3526 TRACE("Created vertex shader %p.\n", object);
3527 *shader = object;
3528
3529 return WINED3D_OK;
3530 }
Windows API implementation