?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Handle SM5 derivative instructions in shader_get_registers_used().
[wine.git] / dlls / wined3d / shader.c
blobf1940c6a618489be09493fa8f45f8bc1f2af3654
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, 12}},
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 && input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1316 reg_maps->vpos = 1;
1317 }
1318 }
1319 else if (!input_signature->elements && reg_maps->input_registers)
1320 {
1321 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1322 struct wined3d_shader_signature_element *e;
1323 unsigned int i;
1324
1325 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1326 return E_OUTOFMEMORY;
1327 input_signature->element_count = count;
1328
1329 e = input_signature->elements;
1330 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1331 {
1332 if (!(reg_maps->input_registers & (1u << i)))
1333 continue;
1334 input_signature_elements[i].register_idx = i;
1335 *e++ = input_signature_elements[i];
1336 }
1337 }
1338
1339 if (output_signature->elements)
1340 {
1341 for (i = 0; i < output_signature->element_count; ++i)
1342 {
1343 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1344 }
1345 }
1346 else if (reg_maps->output_registers)
1347 {
1348 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1349 struct wined3d_shader_signature_element *e;
1350
1351 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1352 return E_OUTOFMEMORY;
1353 output_signature->element_count = count;
1354
1355 e = output_signature->elements;
1356 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1357 {
1358 if (!(reg_maps->output_registers & (1u << i)))
1359 continue;
1360 *e++ = output_signature_elements[i];
1361 }
1362 }
1363
1364 return WINED3D_OK;
1365 }
1366
1367 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1368 {
1369 struct wined3d_shader_indexable_temp *reg, *reg_next;
1370
1371 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1372 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1373
1374 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1375 HeapFree(GetProcessHeap(), 0, reg);
1376 list_init(&reg_maps->indexable_temps);
1377 }
1378
1379 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1380 {
1381 DWORD map = 1u << max;
1382 map |= map - 1;
1383 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1384
1385 return wined3d_log2i(map);
1386 }
1387
1388 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1389 {
1390 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1391 {
1392 shader_addline(buffer, "refactoringAllowed");
1393 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1394 if (global_flags)
1395 shader_addline(buffer, " | ");
1396 }
1397
1398 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1399 {
1400 shader_addline(buffer, "enableRawAndStructuredBuffers");
1401 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1402 }
1403
1404 if (global_flags)
1405 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1406 }
1407
1408 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1409 enum wined3d_tessellator_domain domain)
1410 {
1411 switch (domain)
1412 {
1413 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1414 shader_addline(buffer, "line");
1415 break;
1416 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1417 shader_addline(buffer, "triangle");
1418 break;
1419 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1420 shader_addline(buffer, "quad");
1421 break;
1422 default:
1423 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1424 break;
1425 }
1426 }
1427
1428 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1429 enum wined3d_tessellator_output_primitive output_primitive)
1430 {
1431 switch (output_primitive)
1432 {
1433 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1434 shader_addline(buffer, "point");
1435 break;
1436 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1437 shader_addline(buffer, "line");
1438 break;
1439 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1440 shader_addline(buffer, "triangle_cw");
1441 break;
1442 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1443 shader_addline(buffer, "triangle_ccw");
1444 break;
1445 default:
1446 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1447 break;
1448 }
1449 }
1450
1451 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1452 enum wined3d_tessellator_partitioning partitioning)
1453 {
1454 switch (partitioning)
1455 {
1456 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1457 shader_addline(buffer, "integer");
1458 break;
1459 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1460 shader_addline(buffer, "pow2");
1461 break;
1462 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1463 shader_addline(buffer, "fractional_odd");
1464 break;
1465 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1466 shader_addline(buffer, "fractional_even");
1467 break;
1468 default:
1469 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1470 break;
1471 }
1472 }
1473
1474 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1475 {
1476 unsigned int i;
1477
1478 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1479 {
1480 if (sysval_semantic_names[i].sysval_semantic == semantic)
1481 {
1482 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1483 return;
1484 }
1485 }
1486
1487 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1488 }
1489
1490 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1491 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1492 {
1493 shader_addline(buffer, "dcl");
1494
1495 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1496 {
1497 switch (semantic->resource_type)
1498 {
1499 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1500 shader_addline(buffer, "_2d");
1501 break;
1502
1503 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1504 shader_addline(buffer, "_3d");
1505 break;
1506
1507 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1508 shader_addline(buffer, "_cube");
1509 break;
1510
1511 default:
1512 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1513 break;
1514 }
1515 }
1516 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1517 {
1518 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1519 shader_addline(buffer, "_resource_");
1520 else
1521 shader_addline(buffer, "_uav_");
1522 switch (semantic->resource_type)
1523 {
1524 case WINED3D_SHADER_RESOURCE_BUFFER:
1525 shader_addline(buffer, "buffer");
1526 break;
1527
1528 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1529 shader_addline(buffer, "texture1d");
1530 break;
1531
1532 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1533 shader_addline(buffer, "texture2d");
1534 break;
1535
1536 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1537 shader_addline(buffer, "texture2dms");
1538 break;
1539
1540 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1541 shader_addline(buffer, "texture3d");
1542 break;
1543
1544 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1545 shader_addline(buffer, "texturecube");
1546 break;
1547
1548 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1549 shader_addline(buffer, "texture1darray");
1550 break;
1551
1552 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1553 shader_addline(buffer, "texture2darray");
1554 break;
1555
1556 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1557 shader_addline(buffer, "texture2dmsarray");
1558 break;
1559
1560 default:
1561 shader_addline(buffer, "unknown");
1562 break;
1563 }
1564 switch (semantic->resource_data_type)
1565 {
1566 case WINED3D_DATA_FLOAT:
1567 shader_addline(buffer, " (float)");
1568 break;
1569
1570 case WINED3D_DATA_INT:
1571 shader_addline(buffer, " (int)");
1572 break;
1573
1574 case WINED3D_DATA_UINT:
1575 shader_addline(buffer, " (uint)");
1576 break;
1577
1578 case WINED3D_DATA_UNORM:
1579 shader_addline(buffer, " (unorm)");
1580 break;
1581
1582 case WINED3D_DATA_SNORM:
1583 shader_addline(buffer, " (snorm)");
1584 break;
1585
1586 default:
1587 shader_addline(buffer, " (unknown)");
1588 break;
1589 }
1590 }
1591 else
1592 {
1593 /* Pixel shaders 3.0 don't have usage semantics. */
1594 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1595 return;
1596 else
1597 shader_addline(buffer, "_");
1598
1599 switch (semantic->usage)
1600 {
1601 case WINED3D_DECL_USAGE_POSITION:
1602 shader_addline(buffer, "position%u", semantic->usage_idx);
1603 break;
1604
1605 case WINED3D_DECL_USAGE_BLEND_INDICES:
1606 shader_addline(buffer, "blend");
1607 break;
1608
1609 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1610 shader_addline(buffer, "weight");
1611 break;
1612
1613 case WINED3D_DECL_USAGE_NORMAL:
1614 shader_addline(buffer, "normal%u", semantic->usage_idx);
1615 break;
1616
1617 case WINED3D_DECL_USAGE_PSIZE:
1618 shader_addline(buffer, "psize");
1619 break;
1620
1621 case WINED3D_DECL_USAGE_COLOR:
1622 if (!semantic->usage_idx)
1623 shader_addline(buffer, "color");
1624 else
1625 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1626 break;
1627
1628 case WINED3D_DECL_USAGE_TEXCOORD:
1629 shader_addline(buffer, "texture%u", semantic->usage_idx);
1630 break;
1631
1632 case WINED3D_DECL_USAGE_TANGENT:
1633 shader_addline(buffer, "tangent");
1634 break;
1635
1636 case WINED3D_DECL_USAGE_BINORMAL:
1637 shader_addline(buffer, "binormal");
1638 break;
1639
1640 case WINED3D_DECL_USAGE_TESS_FACTOR:
1641 shader_addline(buffer, "tessfactor");
1642 break;
1643
1644 case WINED3D_DECL_USAGE_POSITIONT:
1645 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1646 break;
1647
1648 case WINED3D_DECL_USAGE_FOG:
1649 shader_addline(buffer, "fog");
1650 break;
1651
1652 case WINED3D_DECL_USAGE_DEPTH:
1653 shader_addline(buffer, "depth");
1654 break;
1655
1656 case WINED3D_DECL_USAGE_SAMPLE:
1657 shader_addline(buffer, "sample");
1658 break;
1659
1660 default:
1661 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1662 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1663 }
1664 }
1665 }
1666
1667 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1668 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1669 {
1670 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1671 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1672 UINT offset = reg->idx[0].offset;
1673
1674 switch (reg->type)
1675 {
1676 case WINED3DSPR_TEMP:
1677 shader_addline(buffer, "r");
1678 break;
1679
1680 case WINED3DSPR_INPUT:
1681 shader_addline(buffer, "v");
1682 break;
1683
1684 case WINED3DSPR_CONST:
1685 case WINED3DSPR_CONST2:
1686 case WINED3DSPR_CONST3:
1687 case WINED3DSPR_CONST4:
1688 shader_addline(buffer, "c");
1689 offset = shader_get_float_offset(reg->type, offset);
1690 break;
1691
1692 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1693 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1694 break;
1695
1696 case WINED3DSPR_RASTOUT:
1697 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1698 break;
1699
1700 case WINED3DSPR_COLOROUT:
1701 shader_addline(buffer, "oC");
1702 break;
1703
1704 case WINED3DSPR_DEPTHOUT:
1705 shader_addline(buffer, "oDepth");
1706 break;
1707
1708 case WINED3DSPR_ATTROUT:
1709 shader_addline(buffer, "oD");
1710 break;
1711
1712 case WINED3DSPR_TEXCRDOUT:
1713 /* Vertex shaders >= 3.0 use general purpose output registers
1714 * (WINED3DSPR_OUTPUT), which can include an address token. */
1715 if (shader_version->major >= 3)
1716 shader_addline(buffer, "o");
1717 else
1718 shader_addline(buffer, "oT");
1719 break;
1720
1721 case WINED3DSPR_CONSTINT:
1722 shader_addline(buffer, "i");
1723 break;
1724
1725 case WINED3DSPR_CONSTBOOL:
1726 shader_addline(buffer, "b");
1727 break;
1728
1729 case WINED3DSPR_LABEL:
1730 shader_addline(buffer, "l");
1731 break;
1732
1733 case WINED3DSPR_LOOP:
1734 shader_addline(buffer, "aL");
1735 break;
1736
1737 case WINED3DSPR_SAMPLER:
1738 shader_addline(buffer, "s");
1739 break;
1740
1741 case WINED3DSPR_MISCTYPE:
1742 if (offset > 1)
1743 {
1744 FIXME("Unhandled misctype register %u.\n", offset);
1745 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1746 }
1747 else
1748 {
1749 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1750 }
1751 break;
1752
1753 case WINED3DSPR_PREDICATE:
1754 shader_addline(buffer, "p");
1755 break;
1756
1757 case WINED3DSPR_IMMCONST:
1758 shader_addline(buffer, "l");
1759 break;
1760
1761 case WINED3DSPR_CONSTBUFFER:
1762 shader_addline(buffer, "cb");
1763 break;
1764
1765 case WINED3DSPR_IMMCONSTBUFFER:
1766 shader_addline(buffer, "icb");
1767 break;
1768
1769 case WINED3DSPR_PRIMID:
1770 shader_addline(buffer, "primID");
1771 break;
1772
1773 case WINED3DSPR_NULL:
1774 shader_addline(buffer, "null");
1775 break;
1776
1777 case WINED3DSPR_RESOURCE:
1778 shader_addline(buffer, "t");
1779 break;
1780
1781 case WINED3DSPR_UAV:
1782 shader_addline(buffer, "u");
1783 break;
1784
1785 case WINED3DSPR_OUTPOINTID:
1786 shader_addline(buffer, "vOutputControlPointID");
1787 break;
1788
1789 case WINED3DSPR_FORKINSTID:
1790 shader_addline(buffer, "vForkInstanceId");
1791 break;
1792
1793 case WINED3DSPR_INCONTROLPOINT:
1794 shader_addline(buffer, "vicp");
1795 break;
1796
1797 case WINED3DSPR_OUTCONTROLPOINT:
1798 shader_addline(buffer, "vocp");
1799 break;
1800
1801 case WINED3DSPR_PATCHCONST:
1802 shader_addline(buffer, "vpc");
1803 break;
1804
1805 case WINED3DSPR_TESSCOORD:
1806 shader_addline(buffer, "vDomainLocation");
1807 break;
1808
1809 case WINED3DSPR_GROUPSHAREDMEM:
1810 shader_addline(buffer, "g");
1811 break;
1812
1813 case WINED3DSPR_THREADID:
1814 shader_addline(buffer, "vThreadID");
1815 break;
1816
1817 case WINED3DSPR_THREADGROUPID:
1818 shader_addline(buffer, "vThreadGroupID");
1819 break;
1820
1821 case WINED3DSPR_LOCALTHREADID:
1822 shader_addline(buffer, "vThreadIDInGroup");
1823 break;
1824
1825 case WINED3DSPR_LOCALTHREADINDEX:
1826 shader_addline(buffer, "vThreadIDInGroupFlattened");
1827 break;
1828
1829 case WINED3DSPR_IDXTEMP:
1830 shader_addline(buffer, "x");
1831 break;
1832
1833 case WINED3DSPR_STREAM:
1834 shader_addline(buffer, "m");
1835 break;
1836
1837 default:
1838 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1839 break;
1840 }
1841
1842 if (reg->type == WINED3DSPR_IMMCONST)
1843 {
1844 shader_addline(buffer, "(");
1845 switch (reg->immconst_type)
1846 {
1847 case WINED3D_IMMCONST_SCALAR:
1848 switch (reg->data_type)
1849 {
1850 case WINED3D_DATA_FLOAT:
1851 shader_addline(buffer, "%.8e", *(const float *)reg->immconst_data);
1852 break;
1853 case WINED3D_DATA_INT:
1854 shader_addline(buffer, "%d", reg->immconst_data[0]);
1855 break;
1856 case WINED3D_DATA_RESOURCE:
1857 case WINED3D_DATA_SAMPLER:
1858 case WINED3D_DATA_UINT:
1859 shader_addline(buffer, "%u", reg->immconst_data[0]);
1860 break;
1861 default:
1862 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1863 break;
1864 }
1865 break;
1866
1867 case WINED3D_IMMCONST_VEC4:
1868 switch (reg->data_type)
1869 {
1870 case WINED3D_DATA_FLOAT:
1871 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1872 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1873 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1874 break;
1875 case WINED3D_DATA_INT:
1876 shader_addline(buffer, "%d, %d, %d, %d",
1877 reg->immconst_data[0], reg->immconst_data[1],
1878 reg->immconst_data[2], reg->immconst_data[3]);
1879 break;
1880 case WINED3D_DATA_RESOURCE:
1881 case WINED3D_DATA_SAMPLER:
1882 case WINED3D_DATA_UINT:
1883 shader_addline(buffer, "%u, %u, %u, %u",
1884 reg->immconst_data[0], reg->immconst_data[1],
1885 reg->immconst_data[2], reg->immconst_data[3]);
1886 break;
1887 default:
1888 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1889 break;
1890 }
1891 break;
1892
1893 default:
1894 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
1895 break;
1896 }
1897 shader_addline(buffer, ")");
1898 }
1899 else if (reg->type != WINED3DSPR_RASTOUT
1900 && reg->type != WINED3DSPR_MISCTYPE
1901 && reg->type != WINED3DSPR_NULL)
1902 {
1903 if (offset != ~0u)
1904 {
1905 shader_addline(buffer, "[");
1906 if (reg->idx[0].rel_addr)
1907 {
1908 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
1909 shader_addline(buffer, " + ");
1910 }
1911 shader_addline(buffer, "%u]", offset);
1912
1913 if (reg->idx[1].offset != ~0u)
1914 {
1915 shader_addline(buffer, "[");
1916 if (reg->idx[1].rel_addr)
1917 {
1918 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
1919 shader_addline(buffer, " + ");
1920 }
1921 shader_addline(buffer, "%u]", reg->idx[1].offset);
1922 }
1923 }
1924 }
1925 }
1926
1927 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
1928 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
1929 {
1930 DWORD write_mask = param->write_mask;
1931
1932 shader_dump_register(buffer, &param->reg, shader_version);
1933
1934 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1935 {
1936 static const char write_mask_chars[] = "xyzw";
1937
1938 shader_addline(buffer, ".");
1939 if (write_mask & WINED3DSP_WRITEMASK_0)
1940 shader_addline(buffer, "%c", write_mask_chars[0]);
1941 if (write_mask & WINED3DSP_WRITEMASK_1)
1942 shader_addline(buffer, "%c", write_mask_chars[1]);
1943 if (write_mask & WINED3DSP_WRITEMASK_2)
1944 shader_addline(buffer, "%c", write_mask_chars[2]);
1945 if (write_mask & WINED3DSP_WRITEMASK_3)
1946 shader_addline(buffer, "%c", write_mask_chars[3]);
1947 }
1948 }
1949
1950 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
1951 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
1952 {
1953 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1954 DWORD swizzle = param->swizzle;
1955
1956 if (src_modifier == WINED3DSPSM_NEG
1957 || src_modifier == WINED3DSPSM_BIASNEG
1958 || src_modifier == WINED3DSPSM_SIGNNEG
1959 || src_modifier == WINED3DSPSM_X2NEG
1960 || src_modifier == WINED3DSPSM_ABSNEG)
1961 shader_addline(buffer, "-");
1962 else if (src_modifier == WINED3DSPSM_COMP)
1963 shader_addline(buffer, "1-");
1964 else if (src_modifier == WINED3DSPSM_NOT)
1965 shader_addline(buffer, "!");
1966
1967 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1968 shader_addline(buffer, "abs(");
1969
1970 shader_dump_register(buffer, &param->reg, shader_version);
1971
1972 switch (src_modifier)
1973 {
1974 case WINED3DSPSM_NONE: break;
1975 case WINED3DSPSM_NEG: break;
1976 case WINED3DSPSM_NOT: break;
1977 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
1978 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
1979 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
1980 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
1981 case WINED3DSPSM_COMP: break;
1982 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
1983 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
1984 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
1985 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
1986 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
1987 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
1988 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
1989 }
1990
1991 if (swizzle != WINED3DSP_NOSWIZZLE)
1992 {
1993 static const char swizzle_chars[] = "xyzw";
1994 DWORD swizzle_x = swizzle & 0x03;
1995 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1996 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1997 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1998
1999 if (swizzle_x == swizzle_y
2000 && swizzle_x == swizzle_z
2001 && swizzle_x == swizzle_w)
2002 {
2003 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2004 }
2005 else
2006 {
2007 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2008 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2009 }
2010 }
2011 }
2012
2013 /* Shared code in order to generate the bulk of the shader string.
2014 * NOTE: A description of how to parse tokens can be found on MSDN. */
2015 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2016 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
2017 {
2018 struct wined3d_device *device = shader->device;
2019 const struct wined3d_shader_frontend *fe = shader->frontend;
2020 void *fe_data = shader->frontend_data;
2021 struct wined3d_shader_version shader_version;
2022 struct wined3d_shader_loop_state loop_state;
2023 struct wined3d_shader_instruction ins;
2024 struct wined3d_shader_tex_mx tex_mx;
2025 struct wined3d_shader_context ctx;
2026 const DWORD *ptr = byte_code;
2027
2028 /* Initialize current parsing state. */
2029 tex_mx.current_row = 0;
2030 loop_state.current_depth = 0;
2031 loop_state.current_reg = 0;
2032
2033 ctx.shader = shader;
2034 ctx.gl_info = &device->adapter->gl_info;
2035 ctx.reg_maps = reg_maps;
2036 ctx.buffer = buffer;
2037 ctx.tex_mx = &tex_mx;
2038 ctx.loop_state = &loop_state;
2039 ctx.backend_data = backend_ctx;
2040 ins.ctx = &ctx;
2041
2042 fe->shader_read_header(fe_data, &ptr, &shader_version);
2043
2044 while (!fe->shader_is_end(fe_data, &ptr))
2045 {
2046 /* Read opcode. */
2047 fe->shader_read_instruction(fe_data, &ptr, &ins);
2048
2049 /* Unknown opcode and its parameters. */
2050 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2051 {
2052 TRACE("Skipping unrecognized instruction.\n");
2053 continue;
2054 }
2055
2056 if (ins.predicate)
2057 FIXME("Predicates not implemented.\n");
2058
2059 /* Call appropriate function for output target */
2060 device->shader_backend->shader_handle_instruction(&ins);
2061 }
2062 }
2063
2064 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2065 const struct wined3d_shader_dst_param *dst)
2066 {
2067 DWORD mmask = dst->modifiers;
2068
2069 switch (dst->shift)
2070 {
2071 case 0: break;
2072 case 13: shader_addline(buffer, "_d8"); break;
2073 case 14: shader_addline(buffer, "_d4"); break;
2074 case 15: shader_addline(buffer, "_d2"); break;
2075 case 1: shader_addline(buffer, "_x2"); break;
2076 case 2: shader_addline(buffer, "_x4"); break;
2077 case 3: shader_addline(buffer, "_x8"); break;
2078 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2079 }
2080
2081 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2082 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2083 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2084
2085 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2086 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
2087 }
2088
2089 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2090 enum wined3d_primitive_type primitive_type)
2091 {
2092 switch (primitive_type)
2093 {
2094 case WINED3D_PT_UNDEFINED:
2095 shader_addline(buffer, "undefined");
2096 break;
2097 case WINED3D_PT_POINTLIST:
2098 shader_addline(buffer, "pointlist");
2099 break;
2100 case WINED3D_PT_LINELIST:
2101 shader_addline(buffer, "linelist");
2102 break;
2103 case WINED3D_PT_LINESTRIP:
2104 shader_addline(buffer, "linestrip");
2105 break;
2106 case WINED3D_PT_TRIANGLELIST:
2107 shader_addline(buffer, "trianglelist");
2108 break;
2109 case WINED3D_PT_TRIANGLESTRIP:
2110 shader_addline(buffer, "trianglestrip");
2111 break;
2112 case WINED3D_PT_TRIANGLEFAN:
2113 shader_addline(buffer, "trianglefan");
2114 break;
2115 case WINED3D_PT_LINELIST_ADJ:
2116 shader_addline(buffer, "linelist_adj");
2117 break;
2118 case WINED3D_PT_LINESTRIP_ADJ:
2119 shader_addline(buffer, "linestrip_adj");
2120 break;
2121 case WINED3D_PT_TRIANGLELIST_ADJ:
2122 shader_addline(buffer, "trianglelist_adj");
2123 break;
2124 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2125 shader_addline(buffer, "trianglestrip_adj");
2126 break;
2127 default:
2128 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2129 break;
2130 }
2131 }
2132
2133 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2134 enum wined3d_shader_interpolation_mode interpolation_mode)
2135 {
2136 switch (interpolation_mode)
2137 {
2138 case WINED3DSIM_CONSTANT:
2139 shader_addline(buffer, "constant");
2140 break;
2141 case WINED3DSIM_LINEAR:
2142 shader_addline(buffer, "linear");
2143 break;
2144 case WINED3DSIM_LINEAR_CENTROID:
2145 shader_addline(buffer, "linear centroid");
2146 break;
2147 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2148 shader_addline(buffer, "linear noperspective");
2149 break;
2150 case WINED3DSIM_LINEAR_SAMPLE:
2151 shader_addline(buffer, "linear sample");
2152 break;
2153 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2154 shader_addline(buffer, "linear noperspective centroid");
2155 break;
2156 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2157 shader_addline(buffer, "linear noperspective sample");
2158 break;
2159 default:
2160 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2161 break;
2162 }
2163 }
2164
2165 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2166 {
2167 struct wined3d_shader_version shader_version;
2168 struct wined3d_string_buffer buffer;
2169 const DWORD *ptr = byte_code;
2170 const char *type_prefix;
2171 const char *p, *q;
2172 DWORD i;
2173
2174 if (!string_buffer_init(&buffer))
2175 {
2176 ERR("Failed to initialize string buffer.\n");
2177 return;
2178 }
2179
2180 TRACE("Parsing %p.\n", byte_code);
2181
2182 fe->shader_read_header(fe_data, &ptr, &shader_version);
2183
2184 switch (shader_version.type)
2185 {
2186 case WINED3D_SHADER_TYPE_VERTEX:
2187 type_prefix = "vs";
2188 break;
2189
2190 case WINED3D_SHADER_TYPE_HULL:
2191 type_prefix = "hs";
2192 break;
2193
2194 case WINED3D_SHADER_TYPE_DOMAIN:
2195 type_prefix = "ds";
2196 break;
2197
2198 case WINED3D_SHADER_TYPE_GEOMETRY:
2199 type_prefix = "gs";
2200 break;
2201
2202 case WINED3D_SHADER_TYPE_PIXEL:
2203 type_prefix = "ps";
2204 break;
2205
2206 case WINED3D_SHADER_TYPE_COMPUTE:
2207 type_prefix = "cs";
2208 break;
2209
2210 default:
2211 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2212 type_prefix = "unknown";
2213 break;
2214 }
2215
2216 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2217
2218 while (!fe->shader_is_end(fe_data, &ptr))
2219 {
2220 struct wined3d_shader_instruction ins;
2221
2222 fe->shader_read_instruction(fe_data, &ptr, &ins);
2223 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2224 {
2225 WARN("Skipping unrecognized instruction.\n");
2226 shader_addline(&buffer, "<unrecognized instruction>\n");
2227 continue;
2228 }
2229
2230 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2231 {
2232 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2233 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2234 shader_addline(&buffer, " ");
2235 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2236 }
2237 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2238 {
2239 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2240 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2241 shader_addline(&buffer, ", %s",
2242 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2243 }
2244 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2245 {
2246 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2247 shader_dump_global_flags(&buffer, ins.flags);
2248 }
2249 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2250 {
2251 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2252 ins.declaration.max_tessellation_factor);
2253 }
2254 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2255 {
2256 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2257 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2258 {
2259 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2260 ins.declaration.icb->data[4 * i + 0],
2261 ins.declaration.icb->data[4 * i + 1],
2262 ins.declaration.icb->data[4 * i + 2],
2263 ins.declaration.icb->data[4 * i + 3]);
2264 }
2265 shader_addline(&buffer, "}");
2266 }
2267 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2268 {
2269 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2270 ins.declaration.indexable_temp.register_idx,
2271 ins.declaration.indexable_temp.register_size,
2272 ins.declaration.indexable_temp.component_count);
2273 }
2274 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2275 {
2276 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2277 shader_dump_interpolation_mode(&buffer, ins.flags);
2278 shader_addline(&buffer, " ");
2279 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2280 }
2281 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2282 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2283 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2284 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2285 {
2286 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2287 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2288 shader_addline(&buffer, ", ");
2289 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2290 }
2291 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2292 {
2293 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2294 shader_dump_interpolation_mode(&buffer, ins.flags);
2295 shader_addline(&buffer, " ");
2296 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2297 shader_addline(&buffer, ", ");
2298 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2299 }
2300 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2301 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2302 {
2303 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2304 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2305 }
2306 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2307 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2308 {
2309 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2310 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2311 }
2312 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2313 {
2314 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2315 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2316 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2317 }
2318 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2319 {
2320 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2321 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2322 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2323 shader_addline(&buffer, ", comparisonMode");
2324 }
2325 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2326 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2327 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2328 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2329 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2330 {
2331 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2332 }
2333 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2334 {
2335 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2336 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2337 }
2338 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2339 {
2340 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2341 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2342 }
2343 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2344 {
2345 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2346 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2347 }
2348 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2349 {
2350 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2351 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2352 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2353 }
2354 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2355 {
2356 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2357 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2358 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2359 ins.declaration.tgsm_structured.structure_count);
2360 }
2361 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2362 {
2363 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2364 ins.declaration.thread_group_size.x,
2365 ins.declaration.thread_group_size.y,
2366 ins.declaration.thread_group_size.z);
2367 }
2368 else if (ins.handler_idx == WINED3DSIH_DEF)
2369 {
2370 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2371 ins.dst[0].reg.idx[0].offset),
2372 *(const float *)&ins.src[0].reg.immconst_data[0],
2373 *(const float *)&ins.src[0].reg.immconst_data[1],
2374 *(const float *)&ins.src[0].reg.immconst_data[2],
2375 *(const float *)&ins.src[0].reg.immconst_data[3]);
2376 }
2377 else if (ins.handler_idx == WINED3DSIH_DEFI)
2378 {
2379 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2380 ins.src[0].reg.immconst_data[0],
2381 ins.src[0].reg.immconst_data[1],
2382 ins.src[0].reg.immconst_data[2],
2383 ins.src[0].reg.immconst_data[3]);
2384 }
2385 else if (ins.handler_idx == WINED3DSIH_DEFB)
2386 {
2387 shader_addline(&buffer, "defb b%u = %s",
2388 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
2389 }
2390 else
2391 {
2392 if (ins.predicate)
2393 {
2394 shader_addline(&buffer, "(");
2395 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2396 shader_addline(&buffer, ") ");
2397 }
2398
2399 /* PixWin marks instructions with the coissue flag with a '+' */
2400 if (ins.coissue)
2401 shader_addline(&buffer, "+");
2402
2403 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2404
2405 if (ins.handler_idx == WINED3DSIH_BREAKP
2406 || ins.handler_idx == WINED3DSIH_IF)
2407 {
2408 switch (ins.flags)
2409 {
2410 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2411 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2412 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2413 }
2414 }
2415 else if (ins.handler_idx == WINED3DSIH_IFC
2416 || ins.handler_idx == WINED3DSIH_BREAKC)
2417 {
2418 switch (ins.flags)
2419 {
2420 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2421 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2422 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2423 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2424 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2425 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2426 default: shader_addline(&buffer, "_(%u)", ins.flags);
2427 }
2428 }
2429 else if (ins.handler_idx == WINED3DSIH_TEX
2430 && shader_version.major >= 2
2431 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2432 {
2433 shader_addline(&buffer, "p");
2434 }
2435 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2436 {
2437 switch (ins.flags)
2438 {
2439 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2440 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2441 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2442 }
2443 }
2444 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2445 {
2446 switch (ins.flags)
2447 {
2448 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2449 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2450 }
2451 }
2452
2453 if (wined3d_shader_instruction_has_texel_offset(&ins))
2454 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2455
2456 for (i = 0; i < ins.dst_count; ++i)
2457 {
2458 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2459 shader_addline(&buffer, !i ? " " : ", ");
2460 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2461 }
2462
2463 /* Other source tokens */
2464 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2465 {
2466 shader_addline(&buffer, !i ? " " : ", ");
2467 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2468 }
2469 }
2470 shader_addline(&buffer, "\n");
2471 }
2472
2473 for (p = buffer.buffer; *p; p = q)
2474 {
2475 if (!(q = strstr(p, "\n")))
2476 q = p + strlen(p);
2477 else
2478 ++q;
2479 TRACE(" %.*s", (int)(q - p), p);
2480 }
2481
2482 string_buffer_free(&buffer);
2483 }
2484
2485 static void shader_cleanup(struct wined3d_shader *shader)
2486 {
2487 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2488 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2489 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2490 shader->device->shader_backend->shader_destroy(shader);
2491 shader_cleanup_reg_maps(&shader->reg_maps);
2492 HeapFree(GetProcessHeap(), 0, shader->function);
2493 shader_delete_constant_list(&shader->constantsF);
2494 shader_delete_constant_list(&shader->constantsB);
2495 shader_delete_constant_list(&shader->constantsI);
2496 list_remove(&shader->shader_list_entry);
2497
2498 if (shader->frontend && shader->frontend_data)
2499 shader->frontend->shader_free(shader->frontend_data);
2500 }
2501
2502 struct shader_none_priv
2503 {
2504 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2505 const struct fragment_pipeline *fragment_pipe;
2506 BOOL ffp_proj_control;
2507 };
2508
2509 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2510 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2511 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2512 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2513 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2514 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2515 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2516 const struct wined3d_state *state) {}
2517 static void shader_none_destroy(struct wined3d_shader *shader) {}
2518 static void shader_none_free_context_data(struct wined3d_context *context) {}
2519 static void shader_none_init_context_state(struct wined3d_context *context) {}
2520
2521 /* Context activation is done by the caller. */
2522 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2523 const struct wined3d_state *state)
2524 {
2525 const struct wined3d_gl_info *gl_info = context->gl_info;
2526 struct shader_none_priv *priv = shader_priv;
2527
2528 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2529 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2530 }
2531
2532 /* Context activation is done by the caller. */
2533 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2534 {
2535 struct shader_none_priv *priv = shader_priv;
2536 const struct wined3d_gl_info *gl_info = context->gl_info;
2537
2538 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2539 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2540
2541 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2542 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2543 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2544 | (1u << WINED3D_SHADER_TYPE_HULL)
2545 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2546 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2547 }
2548
2549 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2550 const struct fragment_pipeline *fragment_pipe)
2551 {
2552 struct fragment_caps fragment_caps;
2553 void *vertex_priv, *fragment_priv;
2554 struct shader_none_priv *priv;
2555
2556 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2557 return E_OUTOFMEMORY;
2558
2559 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2560 {
2561 ERR("Failed to initialize vertex pipe.\n");
2562 HeapFree(GetProcessHeap(), 0, priv);
2563 return E_FAIL;
2564 }
2565
2566 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2567 {
2568 ERR("Failed to initialize fragment pipe.\n");
2569 vertex_pipe->vp_free(device);
2570 HeapFree(GetProcessHeap(), 0, priv);
2571 return E_FAIL;
2572 }
2573
2574 priv->vertex_pipe = vertex_pipe;
2575 priv->fragment_pipe = fragment_pipe;
2576 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2577 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2578
2579 device->vertex_priv = vertex_priv;
2580 device->fragment_priv = fragment_priv;
2581 device->shader_priv = priv;
2582
2583 return WINED3D_OK;
2584 }
2585
2586 static void shader_none_free(struct wined3d_device *device)
2587 {
2588 struct shader_none_priv *priv = device->shader_priv;
2589
2590 priv->fragment_pipe->free_private(device);
2591 priv->vertex_pipe->vp_free(device);
2592 HeapFree(GetProcessHeap(), 0, priv);
2593 }
2594
2595 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2596 {
2597 return TRUE;
2598 }
2599
2600 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2601 {
2602 /* Set the shader caps to 0 for the none shader backend */
2603 caps->vs_version = 0;
2604 caps->hs_version = 0;
2605 caps->ds_version = 0;
2606 caps->gs_version = 0;
2607 caps->ps_version = 0;
2608 caps->cs_version = 0;
2609 caps->vs_uniform_count = 0;
2610 caps->ps_uniform_count = 0;
2611 caps->ps_1x_max_value = 0.0f;
2612 caps->varying_count = 0;
2613 caps->wined3d_caps = 0;
2614 }
2615
2616 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2617 {
2618 /* We "support" every possible fixup, since we don't support any shader
2619 * model, and will never have to actually sample a texture. */
2620 return TRUE;
2621 }
2622
2623 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2624 {
2625 struct shader_none_priv *priv = shader_priv;
2626
2627 return priv->ffp_proj_control;
2628 }
2629
2630 const struct wined3d_shader_backend_ops none_shader_backend =
2631 {
2632 shader_none_handle_instruction,
2633 shader_none_select,
2634 shader_none_disable,
2635 shader_none_select_depth_blt,
2636 shader_none_deselect_depth_blt,
2637 shader_none_update_float_vertex_constants,
2638 shader_none_update_float_pixel_constants,
2639 shader_none_load_constants,
2640 shader_none_destroy,
2641 shader_none_alloc,
2642 shader_none_free,
2643 shader_none_allocate_context_data,
2644 shader_none_free_context_data,
2645 shader_none_init_context_state,
2646 shader_none_get_caps,
2647 shader_none_color_fixup_supported,
2648 shader_none_has_ffp_proj_control,
2649 };
2650
2651 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2652 DWORD float_const_count, enum wined3d_shader_type type, unsigned int max_version)
2653 {
2654 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2655 const struct wined3d_shader_frontend *fe;
2656 HRESULT hr;
2657 unsigned int backend_version;
2658 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2659
2660 TRACE("shader %p, byte_code %p, float_const_count %u, type %#x, max_version %u.\n",
2661 shader, byte_code, float_const_count, type, max_version);
2662
2663 list_init(&shader->constantsF);
2664 list_init(&shader->constantsB);
2665 list_init(&shader->constantsI);
2666 shader->lconst_inf_or_nan = FALSE;
2667 list_init(&reg_maps->indexable_temps);
2668
2669 fe = shader_select_frontend(*byte_code);
2670 if (!fe)
2671 {
2672 FIXME("Unable to find frontend for shader.\n");
2673 return WINED3DERR_INVALIDCALL;
2674 }
2675 shader->frontend = fe;
2676 shader->frontend_data = fe->shader_init(byte_code, &shader->output_signature);
2677 if (!shader->frontend_data)
2678 {
2679 FIXME("Failed to initialize frontend.\n");
2680 return WINED3DERR_INVALIDCALL;
2681 }
2682
2683 /* First pass: trace shader. */
2684 if (TRACE_ON(d3d_shader))
2685 shader_trace_init(fe, shader->frontend_data, byte_code);
2686
2687 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2688 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2689 &shader->output_signature, byte_code, float_const_count)))
2690 return hr;
2691
2692 if (reg_maps->shader_version.type != type)
2693 {
2694 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2695 return WINED3DERR_INVALIDCALL;
2696 }
2697 if (reg_maps->shader_version.major > max_version)
2698 {
2699 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2700 return WINED3DERR_INVALIDCALL;
2701 }
2702 switch (type)
2703 {
2704 case WINED3D_SHADER_TYPE_VERTEX:
2705 backend_version = d3d_info->limits.vs_version;
2706 break;
2707 case WINED3D_SHADER_TYPE_HULL:
2708 backend_version = d3d_info->limits.hs_version;
2709 break;
2710 case WINED3D_SHADER_TYPE_DOMAIN:
2711 backend_version = d3d_info->limits.ds_version;
2712 break;
2713 case WINED3D_SHADER_TYPE_GEOMETRY:
2714 backend_version = d3d_info->limits.gs_version;
2715 break;
2716 case WINED3D_SHADER_TYPE_PIXEL:
2717 backend_version = d3d_info->limits.ps_version;
2718 break;
2719 case WINED3D_SHADER_TYPE_COMPUTE:
2720 backend_version = d3d_info->limits.cs_version;
2721 break;
2722 default:
2723 FIXME("No backend version-checking for this shader type.\n");
2724 backend_version = 0;
2725 }
2726 if (reg_maps->shader_version.major > backend_version)
2727 {
2728 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2729 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2730 return WINED3DERR_INVALIDCALL;
2731 }
2732
2733 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength)))
2734 return E_OUTOFMEMORY;
2735 memcpy(shader->function, byte_code, shader->functionLength);
2736
2737 return WINED3D_OK;
2738 }
2739
2740 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2741 {
2742 ULONG refcount = InterlockedIncrement(&shader->ref);
2743
2744 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2745
2746 return refcount;
2747 }
2748
2749 static void wined3d_shader_destroy_object(void *object)
2750 {
2751 shader_cleanup(object);
2752 HeapFree(GetProcessHeap(), 0, object);
2753 }
2754
2755 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2756 {
2757 ULONG refcount = InterlockedDecrement(&shader->ref);
2758
2759 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2760
2761 if (!refcount)
2762 {
2763 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2764 wined3d_cs_emit_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2765 }
2766
2767 return refcount;
2768 }
2769
2770 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2771 {
2772 TRACE("shader %p.\n", shader);
2773
2774 return shader->parent;
2775 }
2776
2777 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2778 void *byte_code, UINT *byte_code_size)
2779 {
2780 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2781
2782 if (!byte_code)
2783 {
2784 *byte_code_size = shader->functionLength;
2785 return WINED3D_OK;
2786 }
2787
2788 if (*byte_code_size < shader->functionLength)
2789 {
2790 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2791 * than the required size we should write the required size and
2792 * return D3DERR_MOREDATA. That's not actually true. */
2793 return WINED3DERR_INVALIDCALL;
2794 }
2795
2796 memcpy(byte_code, shader->function, shader->functionLength);
2797
2798 return WINED3D_OK;
2799 }
2800
2801 /* Set local constants for d3d8 shaders. */
2802 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2803 UINT start_idx, const float *src_data, UINT count)
2804 {
2805 UINT end_idx = start_idx + count;
2806 UINT i;
2807
2808 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2809
2810 if (end_idx > shader->limits->constant_float)
2811 {
2812 WARN("end_idx %u > float constants limit %u.\n",
2813 end_idx, shader->limits->constant_float);
2814 end_idx = shader->limits->constant_float;
2815 }
2816
2817 for (i = start_idx; i < end_idx; ++i)
2818 {
2819 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2820 float *value;
2821 if (!lconst)
2822 return E_OUTOFMEMORY;
2823
2824 lconst->idx = i;
2825 value = (float *)lconst->value;
2826 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2827 list_add_head(&shader->constantsF, &lconst->entry);
2828
2829 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2830 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2831 {
2832 shader->lconst_inf_or_nan = TRUE;
2833 }
2834 }
2835
2836 return WINED3D_OK;
2837 }
2838
2839 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2840 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2841 {
2842 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2843 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2844 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2845 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2846 args->point_size = state->gl_primitive_type == GL_POINTS;
2847 args->per_vertex_point_size = shader->reg_maps.point_size;
2848 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
2849 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
2850 : WINED3D_SHADER_TYPE_PIXEL;
2851 if (shader->reg_maps.shader_version.major >= 4)
2852 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
2853 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
2854 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
2855 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
2856 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
2857 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
2858 else
2859 args->next_shader_input_count = 0;
2860 args->swizzle_map = swizzle_map;
2861 if (d3d_info->emulated_flatshading)
2862 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2863 else
2864 args->flatshading = 0;
2865 }
2866
2867 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2868 {
2869 if (usage_idx1 != usage_idx2)
2870 return FALSE;
2871 if (usage1 == usage2)
2872 return TRUE;
2873 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2874 return TRUE;
2875 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2876 return TRUE;
2877
2878 return FALSE;
2879 }
2880
2881 BOOL vshader_get_input(const struct wined3d_shader *shader,
2882 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
2883 {
2884 WORD map = shader->reg_maps.input_registers;
2885 unsigned int i;
2886
2887 for (i = 0; map; map >>= 1, ++i)
2888 {
2889 if (!(map & 1)) continue;
2890
2891 if (match_usage(shader->u.vs.attributes[i].usage,
2892 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2893 {
2894 *regnum = i;
2895 return TRUE;
2896 }
2897 }
2898 return FALSE;
2899 }
2900
2901 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
2902 const struct wined3d_shader_signature *src, char **signature_strings)
2903 {
2904 struct wined3d_shader_signature_element *e;
2905 unsigned int i;
2906 SIZE_T len;
2907 char *ptr;
2908
2909 if (!src->element_count)
2910 return WINED3D_OK;
2911
2912 ptr = *signature_strings;
2913
2914 dst->element_count = src->element_count;
2915 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
2916 return E_OUTOFMEMORY;
2917
2918 for (i = 0; i < src->element_count; ++i)
2919 {
2920 e = &src->elements[i];
2921 dst->elements[i] = *e;
2922
2923 len = strlen(e->semantic_name);
2924 memcpy(ptr, e->semantic_name, len + 1);
2925 dst->elements[i].semantic_name = ptr;
2926 ptr += len + 1;
2927 }
2928
2929 *signature_strings = ptr;
2930
2931 return WINED3D_OK;
2932 }
2933
2934 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2935 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
2936 void *parent, const struct wined3d_parent_ops *parent_ops)
2937 {
2938 struct wined3d_shader_signature_element *e;
2939 SIZE_T total, len;
2940 unsigned int i;
2941 HRESULT hr;
2942 char *ptr;
2943
2944 if (!desc->byte_code)
2945 return WINED3DERR_INVALIDCALL;
2946
2947 shader->ref = 1;
2948 shader->device = device;
2949 shader->parent = parent;
2950 shader->parent_ops = parent_ops;
2951
2952 total = 0;
2953 for (i = 0; i < desc->input_signature.element_count; ++i)
2954 {
2955 e = &desc->input_signature.elements[i];
2956 len = strlen(e->semantic_name);
2957 if (len >= ~(SIZE_T)0 - total)
2958 return E_OUTOFMEMORY;
2959
2960 total += len + 1;
2961 }
2962 for (i = 0; i < desc->output_signature.element_count; ++i)
2963 {
2964 e = &desc->output_signature.elements[i];
2965 len = strlen(e->semantic_name);
2966 if (len >= ~(SIZE_T)0 - total)
2967 return E_OUTOFMEMORY;
2968
2969 total += len + 1;
2970 }
2971 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
2972 return E_OUTOFMEMORY;
2973 ptr = shader->signature_strings;
2974
2975 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
2976 {
2977 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2978 return hr;
2979 }
2980 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
2981 {
2982 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2983 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2984 return hr;
2985 }
2986
2987 list_init(&shader->linked_programs);
2988 list_add_head(&device->shaders, &shader->shader_list_entry);
2989
2990 if (FAILED(hr = shader_set_function(shader, desc->byte_code,
2991 float_const_count, type, desc->max_version)))
2992 {
2993 WARN("Failed to set function, hr %#x.\n", hr);
2994 shader_cleanup(shader);
2995 }
2996
2997 shader->load_local_constsF = shader->lconst_inf_or_nan;
2998
2999 return hr;
3000 }
3001
3002 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3003 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3004 {
3005 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3006 unsigned int i;
3007 HRESULT hr;
3008
3009 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3010 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3011 return hr;
3012
3013 for (i = 0; i < shader->input_signature.element_count; ++i)
3014 {
3015 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3016
3017 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3018 continue;
3019
3020 shader->u.vs.attributes[input->register_idx].usage =
3021 shader_usage_from_semantic_name(input->semantic_name);
3022 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3023 }
3024
3025 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3026 shader->load_local_constsF = TRUE;
3027
3028 return WINED3D_OK;
3029 }
3030
3031 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3032 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3033 {
3034 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3035 }
3036
3037 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3038 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3039 {
3040 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3041 }
3042
3043 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3044 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3045 {
3046 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3047 }
3048
3049 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3050 struct gs_compile_args *args)
3051 {
3052 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3053 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3054 }
3055
3056 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3057 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3058 {
3059 const struct wined3d_gl_info *gl_info = context->gl_info;
3060 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3061 const struct wined3d_texture *texture;
3062 UINT i;
3063
3064 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3065 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3066 {
3067 static unsigned int warned = 0;
3068
3069 args->srgb_correction = 1;
3070 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3071 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3072 "support, expect rendering artifacts.\n");
3073 }
3074
3075 if (shader->reg_maps.shader_version.major == 1
3076 && shader->reg_maps.shader_version.minor <= 3)
3077 {
3078 for (i = 0; i < shader->limits->sampler; ++i)
3079 {
3080 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3081
3082 if (flags & WINED3D_TTFF_PROJECTED)
3083 {
3084 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3085
3086 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3087 {
3088 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3089 unsigned int j;
3090 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3091 DWORD max_valid = WINED3D_TTFF_COUNT4;
3092
3093 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3094 {
3095 struct wined3d_vertex_declaration_element *element =
3096 &state->vertex_declaration->elements[j];
3097
3098 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3099 && element->usage_idx == index)
3100 {
3101 max_valid = element->format->component_count;
3102 break;
3103 }
3104 }
3105 if (!tex_transform || tex_transform > max_valid)
3106 {
3107 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3108 tex_transform, max_valid);
3109 tex_transform = max_valid;
3110 }
3111 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3112 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3113 && tex_transform > WINED3D_TTFF_COUNT2)
3114 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3115 && tex_transform > WINED3D_TTFF_COUNT3))
3116 tex_transform |= WINED3D_PSARGS_PROJECTED;
3117 else
3118 {
3119 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3120 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3121 i, tex_transform, resource_type);
3122 }
3123 }
3124 else
3125 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3126
3127 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3128 }
3129 }
3130 }
3131 if (shader->reg_maps.shader_version.major == 1
3132 && shader->reg_maps.shader_version.minor <= 4)
3133 {
3134 for (i = 0; i < shader->limits->sampler; ++i)
3135 {
3136 const struct wined3d_texture *texture = state->textures[i];
3137
3138 if (!shader->reg_maps.resource_info[i].type)
3139 continue;
3140
3141 /* Treat unbound textures as 2D. The dummy texture will provide
3142 * the proper sample value. The tex_types bitmap defaults to
3143 * 2D because of the memset. */
3144 if (!texture)
3145 continue;
3146
3147 switch (texture->target)
3148 {
3149 /* RECT textures are distinguished from 2D textures via np2_fixup */
3150 case GL_TEXTURE_RECTANGLE_ARB:
3151 case GL_TEXTURE_2D:
3152 break;
3153
3154 case GL_TEXTURE_3D:
3155 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3156 break;
3157
3158 case GL_TEXTURE_CUBE_MAP_ARB:
3159 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3160 break;
3161 }
3162 }
3163 }
3164
3165 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3166 {
3167 if (!shader->reg_maps.resource_info[i].type)
3168 continue;
3169
3170 texture = state->textures[i];
3171 if (!texture)
3172 {
3173 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3174 continue;
3175 }
3176 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3177 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3178 else
3179 args->color_fixup[i] = texture->resource.format->color_fixup;
3180
3181 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3182 args->shadow |= 1u << i;
3183
3184 /* Flag samplers that need NP2 texcoord fixup. */
3185 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3186 args->np2_fixup |= (1u << i);
3187 }
3188
3189 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3190 if (shader->reg_maps.shader_version.major >= 4)
3191 args->shadow = 0;
3192
3193 if (shader->reg_maps.shader_version.major >= 3)
3194 {
3195 if (position_transformed)
3196 args->vp_mode = pretransformed;
3197 else if (use_vs(state))
3198 args->vp_mode = vertexshader;
3199 else
3200 args->vp_mode = fixedfunction;
3201 args->fog = WINED3D_FFP_PS_FOG_OFF;
3202 }
3203 else
3204 {
3205 args->vp_mode = vertexshader;
3206 if (state->render_states[WINED3D_RS_FOGENABLE])
3207 {
3208 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3209 {
3210 case WINED3D_FOG_NONE:
3211 if (position_transformed || use_vs(state))
3212 {
3213 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3214 break;
3215 }
3216
3217 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3218 {
3219 case WINED3D_FOG_NONE: /* Fall through. */
3220 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3221 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3222 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3223 }
3224 break;
3225
3226 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3227 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3228 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3229 }
3230 }
3231 else
3232 {
3233 args->fog = WINED3D_FFP_PS_FOG_OFF;
3234 }
3235 }
3236
3237 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3238 {
3239 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3240
3241 args->texcoords_initialized = 0;
3242 for (i = 0; i < MAX_TEXTURES; ++i)
3243 {
3244 if (vs)
3245 {
3246 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3247 args->texcoords_initialized |= 1u << i;
3248 }
3249 else
3250 {
3251 const struct wined3d_stream_info *si = &context->stream_info;
3252 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3253
3254 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3255 & WINED3D_FFP_TCI_MASK
3256 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3257 args->texcoords_initialized |= 1u << i;
3258 }
3259 }
3260 }
3261 else
3262 {
3263 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3264 }
3265
3266 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3267 && state->gl_primitive_type == GL_POINTS;
3268
3269 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3270 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3271 else
3272 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3273 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3274 : WINED3D_CMP_ALWAYS) - 1;
3275
3276 if (d3d_info->emulated_flatshading)
3277 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3278 }
3279
3280 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3281 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3282 {
3283 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3284 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3285 HRESULT hr;
3286
3287 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3288 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3289 return hr;
3290
3291 for (i = 0; i < MAX_REG_INPUT; ++i)
3292 {
3293 if (shader->u.ps.input_reg_used[i])
3294 {
3295 ++num_regs_used;
3296 highest_reg_used = i;
3297 }
3298 }
3299
3300 /* Don't do any register mapping magic if it is not needed, or if we can't
3301 * achieve anything anyway */
3302 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3303 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3304 || shader->reg_maps.shader_version.major >= 4)
3305 {
3306 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3307 {
3308 /* This happens with relative addressing. The input mapper function
3309 * warns about this if the higher registers are declared too, so
3310 * don't write a FIXME here */
3311 WARN("More varying registers used than supported\n");
3312 }
3313
3314 for (i = 0; i < MAX_REG_INPUT; ++i)
3315 {
3316 shader->u.ps.input_reg_map[i] = i;
3317 }
3318
3319 shader->u.ps.declared_in_count = highest_reg_used + 1;
3320 }
3321 else
3322 {
3323 shader->u.ps.declared_in_count = 0;
3324 for (i = 0; i < MAX_REG_INPUT; ++i)
3325 {
3326 if (shader->u.ps.input_reg_used[i])
3327 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3328 else shader->u.ps.input_reg_map[i] = ~0U;
3329 }
3330 }
3331
3332 return WINED3D_OK;
3333 }
3334
3335 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3336 {
3337 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3338 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3339 unsigned int i;
3340
3341 if (reg_maps->shader_version.major != 1) return;
3342
3343 for (i = 0; i < shader->limits->sampler; ++i)
3344 {
3345 /* We don't sample from this sampler. */
3346 if (!resource_info[i].type)
3347 continue;
3348
3349 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3350 {
3351 case WINED3D_SHADER_TEX_2D:
3352 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3353 break;
3354
3355 case WINED3D_SHADER_TEX_3D:
3356 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3357 break;
3358
3359 case WINED3D_SHADER_TEX_CUBE:
3360 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3361 break;
3362 }
3363 }
3364 }
3365
3366 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3367 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3368 {
3369 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3370 }
3371
3372 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3373 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3374 {
3375 struct wined3d_shader *object;
3376 HRESULT hr;
3377
3378 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3379 device, desc, parent, parent_ops, shader);
3380
3381 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3382 return E_OUTOFMEMORY;
3383
3384 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3385 {
3386 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3387 HeapFree(GetProcessHeap(), 0, object);
3388 return hr;
3389 }
3390
3391 TRACE("Created compute shader %p.\n", object);
3392 *shader = object;
3393
3394 return WINED3D_OK;
3395 }
3396
3397 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3398 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3399 {
3400 struct wined3d_shader *object;
3401 HRESULT hr;
3402
3403 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3404 device, desc, parent, parent_ops, shader);
3405
3406 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3407 return E_OUTOFMEMORY;
3408
3409 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3410 {
3411 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3412 HeapFree(GetProcessHeap(), 0, object);
3413 return hr;
3414 }
3415
3416 TRACE("Created domain shader %p.\n", object);
3417 *shader = object;
3418
3419 return WINED3D_OK;
3420 }
3421
3422 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3423 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3424 {
3425 struct wined3d_shader *object;
3426 HRESULT hr;
3427
3428 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3429 device, desc, parent, parent_ops, shader);
3430
3431 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3432 if (!object)
3433 return E_OUTOFMEMORY;
3434
3435 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3436 {
3437 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3438 HeapFree(GetProcessHeap(), 0, object);
3439 return hr;
3440 }
3441
3442 TRACE("Created geometry shader %p.\n", object);
3443 *shader = object;
3444
3445 return WINED3D_OK;
3446 }
3447
3448 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3449 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3450 {
3451 struct wined3d_shader *object;
3452 HRESULT hr;
3453
3454 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3455 device, desc, parent, parent_ops, shader);
3456
3457 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3458 return E_OUTOFMEMORY;
3459
3460 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3461 {
3462 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3463 HeapFree(GetProcessHeap(), 0, object);
3464 return hr;
3465 }
3466
3467 TRACE("Created hull shader %p.\n", object);
3468 *shader = object;
3469
3470 return WINED3D_OK;
3471 }
3472
3473 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3474 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3475 {
3476 struct wined3d_shader *object;
3477 HRESULT hr;
3478
3479 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3480 device, desc, parent, parent_ops, shader);
3481
3482 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3483 if (!object)
3484 return E_OUTOFMEMORY;
3485
3486 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3487 {
3488 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3489 HeapFree(GetProcessHeap(), 0, object);
3490 return hr;
3491 }
3492
3493 TRACE("Created pixel shader %p.\n", object);
3494 *shader = object;
3495
3496 return WINED3D_OK;
3497 }
3498
3499 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3500 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3501 {
3502 struct wined3d_shader *object;
3503 HRESULT hr;
3504
3505 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3506 device, desc, parent, parent_ops, shader);
3507
3508 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3509 if (!object)
3510 return E_OUTOFMEMORY;
3511
3512 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3513 {
3514 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3515 HeapFree(GetProcessHeap(), 0, object);
3516 return hr;
3517 }
3518
3519 TRACE("Created vertex shader %p.\n", object);
3520 *shader = object;
3521
3522 return WINED3D_OK;
3523 }
Windows API implementation