?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Get rid of surface_load_ds_location().
[wine.git] / dlls / wined3d / shader.c
blob4ca82c2ed3cea5be45e2e81349b98a83d9910111
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_ATOMIC_AND */ "atomic_and",
47 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
48 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
49 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
50 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
51 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
52 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
53 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
54 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
55 /* WINED3DSIH_BEM */ "bem",
56 /* WINED3DSIH_BFI */ "bfi",
57 /* WINED3DSIH_BFREV */ "bfrev",
58 /* WINED3DSIH_BREAK */ "break",
59 /* WINED3DSIH_BREAKC */ "breakc",
60 /* WINED3DSIH_BREAKP */ "breakp",
61 /* WINED3DSIH_BUFINFO */ "bufinfo",
62 /* WINED3DSIH_CALL */ "call",
63 /* WINED3DSIH_CALLNZ */ "callnz",
64 /* WINED3DSIH_CASE */ "case",
65 /* WINED3DSIH_CMP */ "cmp",
66 /* WINED3DSIH_CND */ "cnd",
67 /* WINED3DSIH_CONTINUE */ "continue",
68 /* WINED3DSIH_CRS */ "crs",
69 /* WINED3DSIH_CUT */ "cut",
70 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
71 /* WINED3DSIH_DCL */ "dcl",
72 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
73 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
74 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
75 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
76 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
77 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
78 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
79 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
80 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
81 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
82 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
83 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
84 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
85 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
86 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
87 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
88 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
89 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
90 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
91 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
92 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
93 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
94 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
95 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
96 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
97 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
98 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
99 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
100 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
101 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
102 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
103 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
104 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
105 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
106 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
107 /* WINED3DSIH_DEF */ "def",
108 /* WINED3DSIH_DEFAULT */ "default",
109 /* WINED3DSIH_DEFB */ "defb",
110 /* WINED3DSIH_DEFI */ "defi",
111 /* WINED3DSIH_DIV */ "div",
112 /* WINED3DSIH_DP2 */ "dp2",
113 /* WINED3DSIH_DP2ADD */ "dp2add",
114 /* WINED3DSIH_DP3 */ "dp3",
115 /* WINED3DSIH_DP4 */ "dp4",
116 /* WINED3DSIH_DST */ "dst",
117 /* WINED3DSIH_DSX */ "dsx",
118 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
119 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
120 /* WINED3DSIH_DSY */ "dsy",
121 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
122 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
123 /* WINED3DSIH_ELSE */ "else",
124 /* WINED3DSIH_EMIT */ "emit",
125 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
126 /* WINED3DSIH_ENDIF */ "endif",
127 /* WINED3DSIH_ENDLOOP */ "endloop",
128 /* WINED3DSIH_ENDREP */ "endrep",
129 /* WINED3DSIH_ENDSWITCH */ "endswitch",
130 /* WINED3DSIH_EQ */ "eq",
131 /* WINED3DSIH_EXP */ "exp",
132 /* WINED3DSIH_EXPP */ "expp",
133 /* WINED3DSIH_FCALL */ "fcall",
134 /* WINED3DSIH_FRC */ "frc",
135 /* WINED3DSIH_FTOI */ "ftoi",
136 /* WINED3DSIH_FTOU */ "ftou",
137 /* WINED3DSIH_GATHER4 */ "gather4",
138 /* WINED3DSIH_GATHER4_C */ "gather4_c",
139 /* WINED3DSIH_GE */ "ge",
140 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
141 /* WINED3DSIH_HS_DECLS */ "hs_decls",
142 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
143 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
144 /* WINED3DSIH_IADD */ "iadd",
145 /* WINED3DSIH_IEQ */ "ieq",
146 /* WINED3DSIH_IF */ "if",
147 /* WINED3DSIH_IFC */ "ifc",
148 /* WINED3DSIH_IGE */ "ige",
149 /* WINED3DSIH_ILT */ "ilt",
150 /* WINED3DSIH_IMAD */ "imad",
151 /* WINED3DSIH_IMAX */ "imax",
152 /* WINED3DSIH_IMIN */ "imin",
153 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
154 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
155 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
156 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
157 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
158 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
159 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
160 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
161 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
162 /* WINED3DSIH_IMUL */ "imul",
163 /* WINED3DSIH_INE */ "ine",
164 /* WINED3DSIH_INEG */ "ineg",
165 /* WINED3DSIH_ISHL */ "ishl",
166 /* WINED3DSIH_ISHR */ "ishr",
167 /* WINED3DSIH_ITOF */ "itof",
168 /* WINED3DSIH_LABEL */ "label",
169 /* WINED3DSIH_LD */ "ld",
170 /* WINED3DSIH_LD2DMS */ "ld2dms",
171 /* WINED3DSIH_LD_RAW */ "ld_raw",
172 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
173 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
174 /* WINED3DSIH_LIT */ "lit",
175 /* WINED3DSIH_LOD */ "lod",
176 /* WINED3DSIH_LOG */ "log",
177 /* WINED3DSIH_LOGP */ "logp",
178 /* WINED3DSIH_LOOP */ "loop",
179 /* WINED3DSIH_LRP */ "lrp",
180 /* WINED3DSIH_LT */ "lt",
181 /* WINED3DSIH_M3x2 */ "m3x2",
182 /* WINED3DSIH_M3x3 */ "m3x3",
183 /* WINED3DSIH_M3x4 */ "m3x4",
184 /* WINED3DSIH_M4x3 */ "m4x3",
185 /* WINED3DSIH_M4x4 */ "m4x4",
186 /* WINED3DSIH_MAD */ "mad",
187 /* WINED3DSIH_MAX */ "max",
188 /* WINED3DSIH_MIN */ "min",
189 /* WINED3DSIH_MOV */ "mov",
190 /* WINED3DSIH_MOVA */ "mova",
191 /* WINED3DSIH_MOVC */ "movc",
192 /* WINED3DSIH_MUL */ "mul",
193 /* WINED3DSIH_NE */ "ne",
194 /* WINED3DSIH_NOP */ "nop",
195 /* WINED3DSIH_NOT */ "not",
196 /* WINED3DSIH_NRM */ "nrm",
197 /* WINED3DSIH_OR */ "or",
198 /* WINED3DSIH_PHASE */ "phase",
199 /* WINED3DSIH_POW */ "pow",
200 /* WINED3DSIH_RCP */ "rcp",
201 /* WINED3DSIH_REP */ "rep",
202 /* WINED3DSIH_RESINFO */ "resinfo",
203 /* WINED3DSIH_RET */ "ret",
204 /* WINED3DSIH_ROUND_NE */ "round_ne",
205 /* WINED3DSIH_ROUND_NI */ "round_ni",
206 /* WINED3DSIH_ROUND_PI */ "round_pi",
207 /* WINED3DSIH_ROUND_Z */ "round_z",
208 /* WINED3DSIH_RSQ */ "rsq",
209 /* WINED3DSIH_SAMPLE */ "sample",
210 /* WINED3DSIH_SAMPLE_B */ "sample_b",
211 /* WINED3DSIH_SAMPLE_C */ "sample_c",
212 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
213 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
214 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
215 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
216 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
217 /* WINED3DSIH_SETP */ "setp",
218 /* WINED3DSIH_SGE */ "sge",
219 /* WINED3DSIH_SGN */ "sgn",
220 /* WINED3DSIH_SINCOS */ "sincos",
221 /* WINED3DSIH_SLT */ "slt",
222 /* WINED3DSIH_SQRT */ "sqrt",
223 /* WINED3DSIH_STORE_RAW */ "store_raw",
224 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
225 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
226 /* WINED3DSIH_SUB */ "sub",
227 /* WINED3DSIH_SWAPC */ "swapc",
228 /* WINED3DSIH_SWITCH */ "switch",
229 /* WINED3DSIH_SYNC */ "sync",
230 /* WINED3DSIH_TEX */ "texld",
231 /* WINED3DSIH_TEXBEM */ "texbem",
232 /* WINED3DSIH_TEXBEML */ "texbeml",
233 /* WINED3DSIH_TEXCOORD */ "texcrd",
234 /* WINED3DSIH_TEXDEPTH */ "texdepth",
235 /* WINED3DSIH_TEXDP3 */ "texdp3",
236 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
237 /* WINED3DSIH_TEXKILL */ "texkill",
238 /* WINED3DSIH_TEXLDD */ "texldd",
239 /* WINED3DSIH_TEXLDL */ "texldl",
240 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
241 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
242 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
243 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
244 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
245 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
246 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
247 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
248 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
249 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
250 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
251 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
252 /* WINED3DSIH_UBFE */ "ubfe",
253 /* WINED3DSIH_UDIV */ "udiv",
254 /* WINED3DSIH_UGE */ "uge",
255 /* WINED3DSIH_ULT */ "ult",
256 /* WINED3DSIH_UMAX */ "umax",
257 /* WINED3DSIH_UMIN */ "umin",
258 /* WINED3DSIH_USHR */ "ushr",
259 /* WINED3DSIH_UTOF */ "utof",
260 /* WINED3DSIH_XOR */ "xor",
261 };
262
263 static const char * const semantic_names[] =
264 {
265 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
266 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
267 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
268 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
269 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
270 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
271 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
272 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
273 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
274 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
275 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
276 /* WINED3D_DECL_USAGE_FOG */ "FOG",
277 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
278 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
279 };
280
281 static const struct
282 {
283 enum wined3d_sysval_semantic sysval_semantic;
284 const char *sysval_name;
285 }
286 sysval_semantic_names[] =
287 {
288 {WINED3D_SV_POSITION, "SV_Position"},
289 {WINED3D_SV_CLIP_DISTANCE, "SV_ClipDistance"},
290 {WINED3D_SV_CULL_DISTANCE, "SV_CullDistance"},
291 {WINED3D_SV_RENDER_TARGET_ARRAY_INDEX, "SV_RenderTargetArrayIndex"},
292 {WINED3D_SV_VIEWPORT_ARRAY_INDEX, "SV_ViewportArrayIndex"},
293 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
294 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
295 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
296 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
297 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
298 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
299 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
300 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
301 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
302 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
303 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
304 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
305 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
306 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
307 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
308 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
309 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
310 };
311
312 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
313 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
314
315 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
316 {
317 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
318 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
319
320 return shader_opcode_names[handler_idx];
321 }
322
323 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
324 {
325 if (usage >= ARRAY_SIZE(semantic_names))
326 {
327 FIXME("Unrecognized usage %#x.\n", usage);
328 return "UNRECOGNIZED";
329 }
330
331 return semantic_names[usage];
332 }
333
334 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
335 {
336 unsigned int i;
337
338 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
339 {
340 if (!strcmp(name, semantic_names[i]))
341 return i;
342 }
343
344 return ~0U;
345 }
346
347 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
348 {
349 switch (usage)
350 {
351 case WINED3D_DECL_USAGE_POSITION:
352 return WINED3D_SV_POSITION;
353 default:
354 return 0;
355 }
356 }
357
358 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
359 {
360 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
361 }
362
363 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
364 const struct wined3d_shader_semantic *s)
365 {
366 e->semantic_name = shader_semantic_name_from_usage(s->usage);
367 e->semantic_idx = s->usage_idx;
368 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
369 e->component_type = WINED3D_TYPE_FLOAT;
370 e->register_idx = s->reg.reg.idx[0].offset;
371 e->mask = s->reg.write_mask;
372 }
373
374 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
375 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
376 {
377 e->semantic_name = shader_semantic_name_from_usage(usage);
378 e->semantic_idx = usage_idx;
379 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
380 e->component_type = WINED3D_TYPE_FLOAT;
381 e->register_idx = reg_idx;
382 e->mask = write_mask;
383 }
384
385 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
386 {
387 switch (version_token >> 16)
388 {
389 case WINED3D_SM1_VS:
390 case WINED3D_SM1_PS:
391 return &sm1_shader_frontend;
392
393 case WINED3D_SM4_PS:
394 case WINED3D_SM4_VS:
395 case WINED3D_SM4_GS:
396 case WINED3D_SM5_HS:
397 case WINED3D_SM5_DS:
398 case WINED3D_SM5_CS:
399 return &sm4_shader_frontend;
400
401 default:
402 FIXME("Unrecognised version token %#x.\n", version_token);
403 return NULL;
404 }
405 }
406
407 void string_buffer_clear(struct wined3d_string_buffer *buffer)
408 {
409 buffer->buffer[0] = '\0';
410 buffer->content_size = 0;
411 }
412
413 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
414 {
415 buffer->buffer_size = 32;
416 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
417 {
418 ERR("Failed to allocate shader buffer memory.\n");
419 return FALSE;
420 }
421
422 string_buffer_clear(buffer);
423 return TRUE;
424 }
425
426 void string_buffer_free(struct wined3d_string_buffer *buffer)
427 {
428 HeapFree(GetProcessHeap(), 0, buffer->buffer);
429 }
430
431 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
432 {
433 char *new_buffer;
434 unsigned int new_buffer_size = buffer->buffer_size * 2;
435
436 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
437 new_buffer_size *= 2;
438 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
439 {
440 ERR("Failed to grow buffer.\n");
441 buffer->buffer[buffer->content_size] = '\0';
442 return FALSE;
443 }
444 buffer->buffer = new_buffer;
445 buffer->buffer_size = new_buffer_size;
446 return TRUE;
447 }
448
449 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
450 {
451 unsigned int rem;
452 int rc;
453
454 rem = buffer->buffer_size - buffer->content_size;
455 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
456 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
457 return rc;
458
459 buffer->content_size += rc;
460 return 0;
461 }
462
463 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
464 {
465 va_list args;
466 int ret;
467
468 for (;;)
469 {
470 va_start(args, format);
471 ret = shader_vaddline(buffer, format, args);
472 va_end(args);
473 if (!ret)
474 return ret;
475 if (!string_buffer_resize(buffer, ret))
476 return -1;
477 }
478 }
479
480 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
481 {
482 struct wined3d_string_buffer *buffer;
483
484 if (list_empty(&list->list))
485 {
486 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
487 if (!buffer || !string_buffer_init(buffer))
488 {
489 ERR("Couldn't allocate buffer for temporary string.\n");
490 HeapFree(GetProcessHeap(), 0, buffer);
491 return NULL;
492 }
493 }
494 else
495 {
496 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
497 list_remove(&buffer->entry);
498 }
499 string_buffer_clear(buffer);
500 return buffer;
501 }
502
503 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
504 {
505 if (!buffer)
506 return 0;
507 string_buffer_clear(buffer);
508 return shader_vaddline(buffer, format, args);
509 }
510
511 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
512 {
513 va_list args;
514 int ret;
515
516 for (;;)
517 {
518 va_start(args, format);
519 ret = string_buffer_vsprintf(buffer, format, args);
520 va_end(args);
521 if (!ret)
522 return;
523 if (!string_buffer_resize(buffer, ret))
524 return;
525 }
526 }
527
528 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
529 {
530 if (!buffer)
531 return;
532 list_add_head(&list->list, &buffer->entry);
533 }
534
535 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
536 {
537 list_init(&list->list);
538 }
539
540 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
541 {
542 struct wined3d_string_buffer *buffer, *buffer_next;
543
544 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
545 {
546 string_buffer_free(buffer);
547 HeapFree(GetProcessHeap(), 0, buffer);
548 }
549 list_init(&list->list);
550 }
551
552 /* Convert floating point offset relative to a register file to an absolute
553 * offset for float constants. */
554 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
555 {
556 switch (register_type)
557 {
558 case WINED3DSPR_CONST: return register_idx;
559 case WINED3DSPR_CONST2: return 2048 + register_idx;
560 case WINED3DSPR_CONST3: return 4096 + register_idx;
561 case WINED3DSPR_CONST4: return 6144 + register_idx;
562 default:
563 FIXME("Unsupported register type: %u.\n", register_type);
564 return register_idx;
565 }
566 }
567
568 static void shader_delete_constant_list(struct list *clist)
569 {
570 struct wined3d_shader_lconst *constant, *constant_next;
571
572 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
573 HeapFree(GetProcessHeap(), 0, constant);
574 list_init(clist);
575 }
576
577 static void shader_set_limits(struct wined3d_shader *shader)
578 {
579 static const struct limits_entry
580 {
581 unsigned int min_version;
582 unsigned int max_version;
583 struct wined3d_shader_limits limits;
584 }
585 vs_limits[] =
586 {
587 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
588 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
589 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
590 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
591 * even though they are capable of supporting much more (GL
592 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
593 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
594 * shaders to 256. */
595 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
596 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
597 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
598 {0}
599 },
600 hs_limits[] =
601 {
602 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
603 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
604 },
605 ds_limits[] =
606 {
607 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
608 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
609 },
610 gs_limits[] =
611 {
612 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
613 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
614 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
615 {0}
616 },
617 ps_limits[] =
618 {
619 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
620 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
621 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
622 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
623 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
624 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
625 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
626 {0}
627 },
628 cs_limits[] =
629 {
630 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
631 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
632 };
633 const struct limits_entry *limits_array;
634 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
635 shader->reg_maps.shader_version.minor);
636 int i = 0;
637
638 switch (shader->reg_maps.shader_version.type)
639 {
640 default:
641 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
642 /* Fall-through. */
643 case WINED3D_SHADER_TYPE_VERTEX:
644 limits_array = vs_limits;
645 break;
646 case WINED3D_SHADER_TYPE_HULL:
647 limits_array = hs_limits;
648 break;
649 case WINED3D_SHADER_TYPE_DOMAIN:
650 limits_array = ds_limits;
651 break;
652 case WINED3D_SHADER_TYPE_GEOMETRY:
653 limits_array = gs_limits;
654 break;
655 case WINED3D_SHADER_TYPE_PIXEL:
656 limits_array = ps_limits;
657 break;
658 case WINED3D_SHADER_TYPE_COMPUTE:
659 limits_array = cs_limits;
660 break;
661 }
662
663 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
664 {
665 if (shader_version <= limits_array[i].max_version)
666 {
667 shader->limits = &limits_array[i].limits;
668 break;
669 }
670 ++i;
671 }
672 if (!shader->limits)
673 {
674 FIXME("Unexpected shader version \"%u.%u\".\n",
675 shader->reg_maps.shader_version.major,
676 shader->reg_maps.shader_version.minor);
677 shader->limits = &limits_array[max(0, i - 1)].limits;
678 }
679 }
680
681 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
682 {
683 DWORD idx, shift;
684 idx = bit >> 5;
685 shift = bit & 0x1f;
686 bitmap[idx] |= (1u << shift);
687 }
688
689 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
690 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
691 {
692 switch (reg->type)
693 {
694 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
695 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
696 reg_maps->texcoord |= 1u << reg->idx[0].offset;
697 else
698 reg_maps->address |= 1u << reg->idx[0].offset;
699 break;
700
701 case WINED3DSPR_TEMP:
702 reg_maps->temporary |= 1u << reg->idx[0].offset;
703 break;
704
705 case WINED3DSPR_INPUT:
706 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
707 {
708 if (reg->idx[0].rel_addr)
709 {
710 /* If relative addressing is used, we must assume that all registers
711 * are used. Even if it is a construct like v3[aL], we can't assume
712 * that v0, v1 and v2 aren't read because aL can be negative */
713 unsigned int i;
714 for (i = 0; i < MAX_REG_INPUT; ++i)
715 {
716 shader->u.ps.input_reg_used[i] = TRUE;
717 }
718 }
719 else
720 {
721 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
722 }
723 }
724 else
725 reg_maps->input_registers |= 1u << reg->idx[0].offset;
726 break;
727
728 case WINED3DSPR_RASTOUT:
729 if (reg->idx[0].offset == 1)
730 reg_maps->fog = 1;
731 if (reg->idx[0].offset == 2)
732 reg_maps->point_size = 1;
733 break;
734
735 case WINED3DSPR_MISCTYPE:
736 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
737 {
738 if (!reg->idx[0].offset)
739 reg_maps->vpos = 1;
740 else if (reg->idx[0].offset == 1)
741 reg_maps->usesfacing = 1;
742 }
743 break;
744
745 case WINED3DSPR_CONST:
746 if (reg->idx[0].rel_addr)
747 {
748 if (reg->idx[0].offset < reg_maps->min_rel_offset)
749 reg_maps->min_rel_offset = reg->idx[0].offset;
750 if (reg->idx[0].offset > reg_maps->max_rel_offset)
751 reg_maps->max_rel_offset = reg->idx[0].offset;
752 reg_maps->usesrelconstF = TRUE;
753 }
754 else
755 {
756 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
757 {
758 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
759 return FALSE;
760 }
761 else
762 {
763 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
764 }
765 }
766 break;
767
768 case WINED3DSPR_CONSTINT:
769 if (reg->idx[0].offset >= shader->limits->constant_int)
770 {
771 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
772 return FALSE;
773 }
774 else
775 {
776 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
777 }
778 break;
779
780 case WINED3DSPR_CONSTBOOL:
781 if (reg->idx[0].offset >= shader->limits->constant_bool)
782 {
783 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
784 return FALSE;
785 }
786 else
787 {
788 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
789 }
790 break;
791
792 case WINED3DSPR_COLOROUT:
793 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
794 break;
795
796 default:
797 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
798 reg->type, reg->idx[0].offset, reg->idx[1].offset);
799 break;
800 }
801 return TRUE;
802 }
803
804 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
805 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
806 {
807 struct wined3d_shader_sampler_map_entry *entries, *entry;
808 struct wined3d_shader_sampler_map *map;
809 unsigned int i;
810
811 map = &reg_maps->sampler_map;
812 entries = map->entries;
813 for (i = 0; i < map->count; ++i)
814 {
815 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
816 return;
817 }
818
819 if (!map->size)
820 {
821 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
822 {
823 ERR("Failed to allocate sampler map entries.\n");
824 return;
825 }
826 map->size = 4;
827 map->entries = entries;
828 }
829 else if (map->count == map->size)
830 {
831 size_t new_size = map->size * 2;
832
833 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
834 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
835 {
836 ERR("Failed to resize sampler map entries.\n");
837 return;
838 }
839 map->size = new_size;
840 map->entries = entries;
841 }
842
843 entry = &entries[map->count++];
844 entry->resource_idx = resource_idx;
845 entry->sampler_idx = sampler_idx;
846 entry->bind_idx = bind_idx;
847 }
848
849 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
850 {
851 switch (instr)
852 {
853 case WINED3DSIH_M4x4:
854 case WINED3DSIH_M3x4:
855 return param == 1 ? 3 : 0;
856
857 case WINED3DSIH_M4x3:
858 case WINED3DSIH_M3x3:
859 return param == 1 ? 2 : 0;
860
861 case WINED3DSIH_M3x2:
862 return param == 1 ? 1 : 0;
863
864 default:
865 return 0;
866 }
867 }
868
869 /* Note that this does not count the loop register as an address register. */
870 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
871 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
872 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
873 {
874 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
875 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
876 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
877 void *fe_data = shader->frontend_data;
878 struct wined3d_shader_version shader_version;
879 const DWORD *ptr = byte_code;
880 unsigned int i;
881
882 memset(reg_maps, 0, sizeof(*reg_maps));
883 memset(input_signature_elements, 0, sizeof(input_signature_elements));
884 memset(output_signature_elements, 0, sizeof(output_signature_elements));
885 reg_maps->min_rel_offset = ~0U;
886 list_init(&reg_maps->indexable_temps);
887
888 fe->shader_read_header(fe_data, &ptr, &shader_version);
889 reg_maps->shader_version = shader_version;
890
891 shader_set_limits(shader);
892
893 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
894 sizeof(*reg_maps->constf))))
895 {
896 ERR("Failed to allocate constant map memory.\n");
897 return E_OUTOFMEMORY;
898 }
899
900 while (!fe->shader_is_end(fe_data, &ptr))
901 {
902 struct wined3d_shader_instruction ins;
903
904 /* Fetch opcode. */
905 fe->shader_read_instruction(fe_data, &ptr, &ins);
906
907 /* Unhandled opcode, and its parameters. */
908 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
909 {
910 TRACE("Skipping unrecognized instruction.\n");
911 continue;
912 }
913
914 /* Handle declarations. */
915 if (ins.handler_idx == WINED3DSIH_DCL
916 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
917 {
918 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
919 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
920
921 switch (semantic->reg.reg.type)
922 {
923 /* Mark input registers used. */
924 case WINED3DSPR_INPUT:
925 if (reg_idx >= MAX_REG_INPUT)
926 {
927 ERR("Invalid input register index %u.\n", reg_idx);
928 break;
929 }
930 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
931 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
932 return WINED3DERR_INVALIDCALL;
933 reg_maps->input_registers |= 1u << reg_idx;
934 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
935 break;
936
937 /* Vertex shader: mark 3.0 output registers used, save token. */
938 case WINED3DSPR_OUTPUT:
939 if (reg_idx >= MAX_REG_OUTPUT)
940 {
941 ERR("Invalid output register index %u.\n", reg_idx);
942 break;
943 }
944 reg_maps->output_registers |= 1u << reg_idx;
945 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
946 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
947 reg_maps->fog = 1;
948 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
949 reg_maps->point_size = 1;
950 break;
951
952 case WINED3DSPR_SAMPLER:
953 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
954 case WINED3DSPR_RESOURCE:
955 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
956 {
957 ERR("Invalid resource index %u.\n", reg_idx);
958 break;
959 }
960 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
961 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
962 break;
963
964 case WINED3DSPR_UAV:
965 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
966 {
967 ERR("Invalid UAV resource index %u.\n", reg_idx);
968 break;
969 }
970 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
971 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
972 break;
973
974 default:
975 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
976 break;
977 }
978 }
979 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
980 {
981 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
982 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
983 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
984 else
985 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
986 }
987 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
988 {
989 if (reg_maps->icb)
990 FIXME("Multiple immediate constant buffers.\n");
991 reg_maps->icb = ins.declaration.icb;
992 }
993 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
994 {
995 struct wined3d_shader_indexable_temp *reg;
996
997 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
998 return E_OUTOFMEMORY;
999
1000 *reg = ins.declaration.indexable_temp;
1001 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1002 }
1003 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1004 {
1005 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1006 shader->u.gs.input_type = ins.declaration.primitive_type;
1007 else
1008 FIXME("Invalid instruction %#x for shader type %#x.\n",
1009 ins.handler_idx, shader_version.type);
1010 }
1011 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1012 {
1013 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1014 shader->u.gs.output_type = ins.declaration.primitive_type;
1015 else
1016 FIXME("Invalid instruction %#x for shader type %#x.\n",
1017 ins.handler_idx, shader_version.type);
1018 }
1019 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1020 {
1021 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1022 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1023 }
1024 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1025 {
1026 reg_maps->temporary_count = ins.declaration.count;
1027 }
1028 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1029 {
1030 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1031 shader->u.gs.vertices_out = ins.declaration.count;
1032 else
1033 FIXME("Invalid instruction %#x for shader type %#x.\n",
1034 ins.handler_idx, shader_version.type);
1035 }
1036 else if (ins.handler_idx == WINED3DSIH_DEF)
1037 {
1038 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1039 float *value;
1040 if (!lconst) return E_OUTOFMEMORY;
1041
1042 lconst->idx = ins.dst[0].reg.idx[0].offset;
1043 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1044 value = (float *)lconst->value;
1045
1046 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1047 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1048 {
1049 if (value[0] < -1.0f) value[0] = -1.0f;
1050 else if (value[0] > 1.0f) value[0] = 1.0f;
1051 if (value[1] < -1.0f) value[1] = -1.0f;
1052 else if (value[1] > 1.0f) value[1] = 1.0f;
1053 if (value[2] < -1.0f) value[2] = -1.0f;
1054 else if (value[2] > 1.0f) value[2] = 1.0f;
1055 if (value[3] < -1.0f) value[3] = -1.0f;
1056 else if (value[3] > 1.0f) value[3] = 1.0f;
1057 }
1058
1059 list_add_head(&shader->constantsF, &lconst->entry);
1060
1061 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1062 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1063 {
1064 shader->lconst_inf_or_nan = TRUE;
1065 }
1066 }
1067 else if (ins.handler_idx == WINED3DSIH_DEFI)
1068 {
1069 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1070 if (!lconst) return E_OUTOFMEMORY;
1071
1072 lconst->idx = ins.dst[0].reg.idx[0].offset;
1073 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1074
1075 list_add_head(&shader->constantsI, &lconst->entry);
1076 reg_maps->local_int_consts |= (1u << lconst->idx);
1077 }
1078 else if (ins.handler_idx == WINED3DSIH_DEFB)
1079 {
1080 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1081 if (!lconst) return E_OUTOFMEMORY;
1082
1083 lconst->idx = ins.dst[0].reg.idx[0].offset;
1084 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1085
1086 list_add_head(&shader->constantsB, &lconst->entry);
1087 reg_maps->local_bool_consts |= (1u << lconst->idx);
1088 }
1089 /* For subroutine prototypes. */
1090 else if (ins.handler_idx == WINED3DSIH_LABEL)
1091 {
1092 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1093 }
1094 /* Set texture, address, temporary registers. */
1095 else
1096 {
1097 BOOL color0_mov = FALSE;
1098 unsigned int i;
1099
1100 /* This will loop over all the registers and try to
1101 * make a bitmask of the ones we're interested in.
1102 *
1103 * Relative addressing tokens are ignored, but that's
1104 * okay, since we'll catch any address registers when
1105 * they are initialized (required by spec). */
1106 for (i = 0; i < ins.dst_count; ++i)
1107 {
1108 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1109 shader_version.type, constf_size))
1110 return WINED3DERR_INVALIDCALL;
1111
1112 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1113 {
1114 UINT idx = ins.dst[i].reg.idx[0].offset;
1115
1116 switch (ins.dst[i].reg.type)
1117 {
1118 case WINED3DSPR_RASTOUT:
1119 if (shader_version.major >= 3)
1120 break;
1121 switch (idx)
1122 {
1123 case 0: /* oPos */
1124 reg_maps->output_registers |= 1u << 10;
1125 shader_signature_from_usage(&output_signature_elements[10],
1126 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1127 break;
1128
1129 case 1: /* oFog */
1130 reg_maps->output_registers |= 1u << 11;
1131 shader_signature_from_usage(&output_signature_elements[11],
1132 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1133 break;
1134
1135 case 2: /* oPts */
1136 reg_maps->output_registers |= 1u << 11;
1137 shader_signature_from_usage(&output_signature_elements[11],
1138 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1139 break;
1140 }
1141 break;
1142
1143 case WINED3DSPR_ATTROUT:
1144 if (shader_version.major >= 3)
1145 break;
1146 if (idx < 2)
1147 {
1148 idx += 8;
1149 if (reg_maps->output_registers & (1u << idx))
1150 {
1151 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1152 }
1153 else
1154 {
1155 reg_maps->output_registers |= 1u << idx;
1156 shader_signature_from_usage(&output_signature_elements[idx],
1157 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1158 }
1159 }
1160 break;
1161
1162 case WINED3DSPR_TEXCRDOUT:
1163 if (shader_version.major >= 3)
1164 {
1165 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1166 break;
1167 }
1168 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1169 if (reg_maps->output_registers & (1u << idx))
1170 {
1171 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1172 }
1173 else
1174 {
1175 reg_maps->output_registers |= 1u << idx;
1176 shader_signature_from_usage(&output_signature_elements[idx],
1177 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1178 }
1179 break;
1180
1181 default:
1182 break;
1183 }
1184 }
1185
1186 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1187 {
1188 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1189 {
1190 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1191 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1192 * the mov and perform the sRGB write correction from the source register.
1193 *
1194 * However, if the mov is only partial, we can't do this, and if the write
1195 * comes from an instruction other than MOV it is hard to do as well. If
1196 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1197 shader->u.ps.color0_mov = FALSE;
1198 if (ins.handler_idx == WINED3DSIH_MOV
1199 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1200 {
1201 /* Used later when the source register is read. */
1202 color0_mov = TRUE;
1203 }
1204 }
1205 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1206 * end
1207 */
1208 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1209 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1210 {
1211 shader->u.ps.color0_mov = FALSE;
1212 }
1213 }
1214
1215 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1216 if (shader_version.major == 1
1217 && (ins.handler_idx == WINED3DSIH_TEX
1218 || ins.handler_idx == WINED3DSIH_TEXBEM
1219 || ins.handler_idx == WINED3DSIH_TEXBEML
1220 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1221 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1222 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1223 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1224 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1225 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1226 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1227 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1228 {
1229 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1230
1231 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1232 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1233 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1234 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1235
1236 /* texbem is only valid with < 1.4 pixel shaders */
1237 if (ins.handler_idx == WINED3DSIH_TEXBEM
1238 || ins.handler_idx == WINED3DSIH_TEXBEML)
1239 {
1240 reg_maps->bumpmat |= 1u << reg_idx;
1241 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1242 {
1243 reg_maps->luminanceparams |= 1u << reg_idx;
1244 }
1245 }
1246 }
1247 else if (ins.handler_idx == WINED3DSIH_BEM)
1248 {
1249 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1250 }
1251 }
1252
1253 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD
1254 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED)
1255 {
1256 unsigned int reg_idx;
1257 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD)
1258 reg_idx = ins.dst[0].reg.idx[0].offset;
1259 else
1260 reg_idx = ins.src[1].reg.idx[0].offset;
1261 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1262 {
1263 ERR("Invalid UAV index %u.\n", reg_idx);
1264 break;
1265 }
1266 reg_maps->uav_read_mask |= (1u << reg_idx);
1267 }
1268 else if (ins.handler_idx == WINED3DSIH_NRM)
1269 {
1270 reg_maps->usesnrm = 1;
1271 }
1272 else if (ins.handler_idx == WINED3DSIH_DSY
1273 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1274 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1275 {
1276 reg_maps->usesdsy = 1;
1277 }
1278 else if (ins.handler_idx == WINED3DSIH_DSX
1279 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1280 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1281 {
1282 reg_maps->usesdsx = 1;
1283 }
1284 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1285 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1286 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1287 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1288 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1289 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1290 else if (ins.handler_idx == WINED3DSIH_LOOP
1291 || ins.handler_idx == WINED3DSIH_REP)
1292 {
1293 ++cur_loop_depth;
1294 if (cur_loop_depth > max_loop_depth)
1295 max_loop_depth = cur_loop_depth;
1296 }
1297 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1298 || ins.handler_idx == WINED3DSIH_ENDREP)
1299 {
1300 --cur_loop_depth;
1301 }
1302 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1303 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1304 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1305 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1306 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1307 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1308 {
1309 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1310 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1311 }
1312 else if (ins.handler_idx == WINED3DSIH_LD
1313 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1314 {
1315 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1316 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1317 }
1318
1319 if (ins.predicate)
1320 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1321 shader_version.type, constf_size))
1322 return WINED3DERR_INVALIDCALL;
1323
1324 for (i = 0; i < ins.src_count; ++i)
1325 {
1326 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1327 struct wined3d_shader_register reg = ins.src[i].reg;
1328
1329 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1330 shader_version.type, constf_size))
1331 return WINED3DERR_INVALIDCALL;
1332 while (count)
1333 {
1334 ++reg.idx[0].offset;
1335 if (!shader_record_register_usage(shader, reg_maps, &reg,
1336 shader_version.type, constf_size))
1337 return WINED3DERR_INVALIDCALL;
1338 --count;
1339 }
1340
1341 if (color0_mov)
1342 {
1343 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1344 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1345 {
1346 shader->u.ps.color0_mov = TRUE;
1347 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1348 }
1349 }
1350 }
1351 }
1352 }
1353 reg_maps->loop_depth = max_loop_depth;
1354
1355 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1356 * R0 is written to the render target. */
1357 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1358 reg_maps->rt_mask |= (1u << 0);
1359
1360 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1361
1362 if (input_signature->elements)
1363 {
1364 for (i = 0; i < input_signature->element_count; ++i)
1365 {
1366 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1367 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1368 {
1369 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1370 reg_maps->vpos = 1;
1371 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1372 reg_maps->usesfacing = 1;
1373 }
1374 }
1375 }
1376 else if (!input_signature->elements && reg_maps->input_registers)
1377 {
1378 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1379 struct wined3d_shader_signature_element *e;
1380 unsigned int i;
1381
1382 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1383 return E_OUTOFMEMORY;
1384 input_signature->element_count = count;
1385
1386 e = input_signature->elements;
1387 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1388 {
1389 if (!(reg_maps->input_registers & (1u << i)))
1390 continue;
1391 input_signature_elements[i].register_idx = i;
1392 *e++ = input_signature_elements[i];
1393 }
1394 }
1395
1396 if (output_signature->elements)
1397 {
1398 for (i = 0; i < output_signature->element_count; ++i)
1399 {
1400 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1401 }
1402 }
1403 else if (reg_maps->output_registers)
1404 {
1405 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1406 struct wined3d_shader_signature_element *e;
1407
1408 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1409 return E_OUTOFMEMORY;
1410 output_signature->element_count = count;
1411
1412 e = output_signature->elements;
1413 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1414 {
1415 if (!(reg_maps->output_registers & (1u << i)))
1416 continue;
1417 *e++ = output_signature_elements[i];
1418 }
1419 }
1420
1421 return WINED3D_OK;
1422 }
1423
1424 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1425 {
1426 struct wined3d_shader_indexable_temp *reg, *reg_next;
1427
1428 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1429 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1430
1431 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1432 HeapFree(GetProcessHeap(), 0, reg);
1433 list_init(&reg_maps->indexable_temps);
1434 }
1435
1436 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1437 {
1438 DWORD map = 1u << max;
1439 map |= map - 1;
1440 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1441
1442 return wined3d_log2i(map);
1443 }
1444
1445 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1446 {
1447 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1448 {
1449 shader_addline(buffer, "refactoringAllowed");
1450 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1451 if (global_flags)
1452 shader_addline(buffer, " | ");
1453 }
1454
1455 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1456 {
1457 shader_addline(buffer, "enableRawAndStructuredBuffers");
1458 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1459 }
1460
1461 if (global_flags)
1462 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1463 }
1464
1465 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1466 {
1467 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1468 {
1469 shader_addline(buffer, "_g");
1470 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1471 }
1472 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1473 {
1474 shader_addline(buffer, "_t");
1475 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1476 }
1477
1478 if (sync_flags)
1479 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1480 }
1481
1482 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1483 {
1484 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1485 {
1486 shader_addline(buffer, "_glc");
1487 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1488 }
1489 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1490 {
1491 shader_addline(buffer, "_opc");
1492 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1493 }
1494
1495 if (uav_flags)
1496 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1497 }
1498
1499 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1500 enum wined3d_tessellator_domain domain)
1501 {
1502 switch (domain)
1503 {
1504 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1505 shader_addline(buffer, "line");
1506 break;
1507 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1508 shader_addline(buffer, "triangle");
1509 break;
1510 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1511 shader_addline(buffer, "quad");
1512 break;
1513 default:
1514 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1515 break;
1516 }
1517 }
1518
1519 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1520 enum wined3d_tessellator_output_primitive output_primitive)
1521 {
1522 switch (output_primitive)
1523 {
1524 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1525 shader_addline(buffer, "point");
1526 break;
1527 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1528 shader_addline(buffer, "line");
1529 break;
1530 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1531 shader_addline(buffer, "triangle_cw");
1532 break;
1533 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1534 shader_addline(buffer, "triangle_ccw");
1535 break;
1536 default:
1537 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1538 break;
1539 }
1540 }
1541
1542 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1543 enum wined3d_tessellator_partitioning partitioning)
1544 {
1545 switch (partitioning)
1546 {
1547 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1548 shader_addline(buffer, "integer");
1549 break;
1550 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1551 shader_addline(buffer, "pow2");
1552 break;
1553 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1554 shader_addline(buffer, "fractional_odd");
1555 break;
1556 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1557 shader_addline(buffer, "fractional_even");
1558 break;
1559 default:
1560 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1561 break;
1562 }
1563 }
1564
1565 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1566 {
1567 unsigned int i;
1568
1569 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1570 {
1571 if (sysval_semantic_names[i].sysval_semantic == semantic)
1572 {
1573 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1574 return;
1575 }
1576 }
1577
1578 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1579 }
1580
1581 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1582 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1583 {
1584 shader_addline(buffer, "dcl");
1585
1586 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1587 {
1588 switch (semantic->resource_type)
1589 {
1590 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1591 shader_addline(buffer, "_2d");
1592 break;
1593
1594 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1595 shader_addline(buffer, "_3d");
1596 break;
1597
1598 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1599 shader_addline(buffer, "_cube");
1600 break;
1601
1602 default:
1603 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1604 break;
1605 }
1606 }
1607 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1608 {
1609 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1610 shader_addline(buffer, "_resource_");
1611 else
1612 shader_addline(buffer, "_uav_");
1613 switch (semantic->resource_type)
1614 {
1615 case WINED3D_SHADER_RESOURCE_BUFFER:
1616 shader_addline(buffer, "buffer");
1617 break;
1618
1619 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1620 shader_addline(buffer, "texture1d");
1621 break;
1622
1623 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1624 shader_addline(buffer, "texture2d");
1625 break;
1626
1627 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1628 shader_addline(buffer, "texture2dms");
1629 break;
1630
1631 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1632 shader_addline(buffer, "texture3d");
1633 break;
1634
1635 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1636 shader_addline(buffer, "texturecube");
1637 break;
1638
1639 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1640 shader_addline(buffer, "texture1darray");
1641 break;
1642
1643 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1644 shader_addline(buffer, "texture2darray");
1645 break;
1646
1647 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1648 shader_addline(buffer, "texture2dmsarray");
1649 break;
1650
1651 default:
1652 shader_addline(buffer, "unknown");
1653 break;
1654 }
1655 switch (semantic->resource_data_type)
1656 {
1657 case WINED3D_DATA_FLOAT:
1658 shader_addline(buffer, " (float)");
1659 break;
1660
1661 case WINED3D_DATA_INT:
1662 shader_addline(buffer, " (int)");
1663 break;
1664
1665 case WINED3D_DATA_UINT:
1666 shader_addline(buffer, " (uint)");
1667 break;
1668
1669 case WINED3D_DATA_UNORM:
1670 shader_addline(buffer, " (unorm)");
1671 break;
1672
1673 case WINED3D_DATA_SNORM:
1674 shader_addline(buffer, " (snorm)");
1675 break;
1676
1677 default:
1678 shader_addline(buffer, " (unknown)");
1679 break;
1680 }
1681 }
1682 else
1683 {
1684 /* Pixel shaders 3.0 don't have usage semantics. */
1685 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1686 return;
1687 else
1688 shader_addline(buffer, "_");
1689
1690 switch (semantic->usage)
1691 {
1692 case WINED3D_DECL_USAGE_POSITION:
1693 shader_addline(buffer, "position%u", semantic->usage_idx);
1694 break;
1695
1696 case WINED3D_DECL_USAGE_BLEND_INDICES:
1697 shader_addline(buffer, "blend");
1698 break;
1699
1700 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1701 shader_addline(buffer, "weight");
1702 break;
1703
1704 case WINED3D_DECL_USAGE_NORMAL:
1705 shader_addline(buffer, "normal%u", semantic->usage_idx);
1706 break;
1707
1708 case WINED3D_DECL_USAGE_PSIZE:
1709 shader_addline(buffer, "psize");
1710 break;
1711
1712 case WINED3D_DECL_USAGE_COLOR:
1713 if (!semantic->usage_idx)
1714 shader_addline(buffer, "color");
1715 else
1716 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1717 break;
1718
1719 case WINED3D_DECL_USAGE_TEXCOORD:
1720 shader_addline(buffer, "texture%u", semantic->usage_idx);
1721 break;
1722
1723 case WINED3D_DECL_USAGE_TANGENT:
1724 shader_addline(buffer, "tangent");
1725 break;
1726
1727 case WINED3D_DECL_USAGE_BINORMAL:
1728 shader_addline(buffer, "binormal");
1729 break;
1730
1731 case WINED3D_DECL_USAGE_TESS_FACTOR:
1732 shader_addline(buffer, "tessfactor");
1733 break;
1734
1735 case WINED3D_DECL_USAGE_POSITIONT:
1736 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1737 break;
1738
1739 case WINED3D_DECL_USAGE_FOG:
1740 shader_addline(buffer, "fog");
1741 break;
1742
1743 case WINED3D_DECL_USAGE_DEPTH:
1744 shader_addline(buffer, "depth");
1745 break;
1746
1747 case WINED3D_DECL_USAGE_SAMPLE:
1748 shader_addline(buffer, "sample");
1749 break;
1750
1751 default:
1752 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1753 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
1754 }
1755 }
1756 }
1757
1758 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1759 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1760 {
1761 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1762 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1763 UINT offset = reg->idx[0].offset;
1764
1765 switch (reg->type)
1766 {
1767 case WINED3DSPR_TEMP:
1768 shader_addline(buffer, "r");
1769 break;
1770
1771 case WINED3DSPR_INPUT:
1772 shader_addline(buffer, "v");
1773 break;
1774
1775 case WINED3DSPR_CONST:
1776 case WINED3DSPR_CONST2:
1777 case WINED3DSPR_CONST3:
1778 case WINED3DSPR_CONST4:
1779 shader_addline(buffer, "c");
1780 offset = shader_get_float_offset(reg->type, offset);
1781 break;
1782
1783 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1784 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1785 break;
1786
1787 case WINED3DSPR_RASTOUT:
1788 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1789 break;
1790
1791 case WINED3DSPR_COLOROUT:
1792 shader_addline(buffer, "oC");
1793 break;
1794
1795 case WINED3DSPR_DEPTHOUT:
1796 shader_addline(buffer, "oDepth");
1797 break;
1798
1799 case WINED3DSPR_ATTROUT:
1800 shader_addline(buffer, "oD");
1801 break;
1802
1803 case WINED3DSPR_TEXCRDOUT:
1804 /* Vertex shaders >= 3.0 use general purpose output registers
1805 * (WINED3DSPR_OUTPUT), which can include an address token. */
1806 if (shader_version->major >= 3)
1807 shader_addline(buffer, "o");
1808 else
1809 shader_addline(buffer, "oT");
1810 break;
1811
1812 case WINED3DSPR_CONSTINT:
1813 shader_addline(buffer, "i");
1814 break;
1815
1816 case WINED3DSPR_CONSTBOOL:
1817 shader_addline(buffer, "b");
1818 break;
1819
1820 case WINED3DSPR_LABEL:
1821 shader_addline(buffer, "l");
1822 break;
1823
1824 case WINED3DSPR_LOOP:
1825 shader_addline(buffer, "aL");
1826 break;
1827
1828 case WINED3DSPR_SAMPLER:
1829 shader_addline(buffer, "s");
1830 break;
1831
1832 case WINED3DSPR_MISCTYPE:
1833 if (offset > 1)
1834 {
1835 FIXME("Unhandled misctype register %u.\n", offset);
1836 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1837 }
1838 else
1839 {
1840 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1841 }
1842 break;
1843
1844 case WINED3DSPR_PREDICATE:
1845 shader_addline(buffer, "p");
1846 break;
1847
1848 case WINED3DSPR_IMMCONST:
1849 shader_addline(buffer, "l");
1850 break;
1851
1852 case WINED3DSPR_CONSTBUFFER:
1853 shader_addline(buffer, "cb");
1854 break;
1855
1856 case WINED3DSPR_IMMCONSTBUFFER:
1857 shader_addline(buffer, "icb");
1858 break;
1859
1860 case WINED3DSPR_PRIMID:
1861 shader_addline(buffer, "primID");
1862 break;
1863
1864 case WINED3DSPR_NULL:
1865 shader_addline(buffer, "null");
1866 break;
1867
1868 case WINED3DSPR_RESOURCE:
1869 shader_addline(buffer, "t");
1870 break;
1871
1872 case WINED3DSPR_UAV:
1873 shader_addline(buffer, "u");
1874 break;
1875
1876 case WINED3DSPR_OUTPOINTID:
1877 shader_addline(buffer, "vOutputControlPointID");
1878 break;
1879
1880 case WINED3DSPR_FORKINSTID:
1881 shader_addline(buffer, "vForkInstanceId");
1882 break;
1883
1884 case WINED3DSPR_INCONTROLPOINT:
1885 shader_addline(buffer, "vicp");
1886 break;
1887
1888 case WINED3DSPR_OUTCONTROLPOINT:
1889 shader_addline(buffer, "vocp");
1890 break;
1891
1892 case WINED3DSPR_PATCHCONST:
1893 shader_addline(buffer, "vpc");
1894 break;
1895
1896 case WINED3DSPR_TESSCOORD:
1897 shader_addline(buffer, "vDomainLocation");
1898 break;
1899
1900 case WINED3DSPR_GROUPSHAREDMEM:
1901 shader_addline(buffer, "g");
1902 break;
1903
1904 case WINED3DSPR_THREADID:
1905 shader_addline(buffer, "vThreadID");
1906 break;
1907
1908 case WINED3DSPR_THREADGROUPID:
1909 shader_addline(buffer, "vThreadGroupID");
1910 break;
1911
1912 case WINED3DSPR_LOCALTHREADID:
1913 shader_addline(buffer, "vThreadIDInGroup");
1914 break;
1915
1916 case WINED3DSPR_LOCALTHREADINDEX:
1917 shader_addline(buffer, "vThreadIDInGroupFlattened");
1918 break;
1919
1920 case WINED3DSPR_IDXTEMP:
1921 shader_addline(buffer, "x");
1922 break;
1923
1924 case WINED3DSPR_STREAM:
1925 shader_addline(buffer, "m");
1926 break;
1927
1928 case WINED3DSPR_FUNCTIONBODY:
1929 shader_addline(buffer, "fb");
1930 break;
1931
1932 case WINED3DSPR_FUNCTIONPOINTER:
1933 shader_addline(buffer, "fp");
1934 break;
1935
1936 case WINED3DSPR_COVERAGE:
1937 shader_addline(buffer, "vCoverage");
1938 break;
1939
1940 case WINED3DSPR_SAMPLEMASK:
1941 shader_addline(buffer, "oMask");
1942 break;
1943
1944 default:
1945 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1946 break;
1947 }
1948
1949 if (reg->type == WINED3DSPR_IMMCONST)
1950 {
1951 shader_addline(buffer, "(");
1952 switch (reg->immconst_type)
1953 {
1954 case WINED3D_IMMCONST_SCALAR:
1955 switch (reg->data_type)
1956 {
1957 case WINED3D_DATA_FLOAT:
1958 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
1959 break;
1960 case WINED3D_DATA_INT:
1961 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
1962 break;
1963 case WINED3D_DATA_RESOURCE:
1964 case WINED3D_DATA_SAMPLER:
1965 case WINED3D_DATA_UINT:
1966 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
1967 break;
1968 default:
1969 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1970 break;
1971 }
1972 break;
1973
1974 case WINED3D_IMMCONST_VEC4:
1975 switch (reg->data_type)
1976 {
1977 case WINED3D_DATA_FLOAT:
1978 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1979 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
1980 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
1981 break;
1982 case WINED3D_DATA_INT:
1983 shader_addline(buffer, "%d, %d, %d, %d",
1984 reg->u.immconst_data[0], reg->u.immconst_data[1],
1985 reg->u.immconst_data[2], reg->u.immconst_data[3]);
1986 break;
1987 case WINED3D_DATA_RESOURCE:
1988 case WINED3D_DATA_SAMPLER:
1989 case WINED3D_DATA_UINT:
1990 shader_addline(buffer, "%u, %u, %u, %u",
1991 reg->u.immconst_data[0], reg->u.immconst_data[1],
1992 reg->u.immconst_data[2], reg->u.immconst_data[3]);
1993 break;
1994 default:
1995 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1996 break;
1997 }
1998 break;
1999
2000 default:
2001 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2002 break;
2003 }
2004 shader_addline(buffer, ")");
2005 }
2006 else if (reg->type != WINED3DSPR_RASTOUT
2007 && reg->type != WINED3DSPR_MISCTYPE
2008 && reg->type != WINED3DSPR_NULL)
2009 {
2010 if (offset != ~0u)
2011 {
2012 shader_addline(buffer, "[");
2013 if (reg->idx[0].rel_addr)
2014 {
2015 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2016 shader_addline(buffer, " + ");
2017 }
2018 shader_addline(buffer, "%u]", offset);
2019
2020 if (reg->idx[1].offset != ~0u)
2021 {
2022 shader_addline(buffer, "[");
2023 if (reg->idx[1].rel_addr)
2024 {
2025 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2026 shader_addline(buffer, " + ");
2027 }
2028 shader_addline(buffer, "%u]", reg->idx[1].offset);
2029 }
2030 }
2031
2032 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2033 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2034 }
2035 }
2036
2037 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2038 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2039 {
2040 DWORD write_mask = param->write_mask;
2041
2042 shader_dump_register(buffer, &param->reg, shader_version);
2043
2044 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2045 {
2046 static const char write_mask_chars[] = "xyzw";
2047
2048 shader_addline(buffer, ".");
2049 if (write_mask & WINED3DSP_WRITEMASK_0)
2050 shader_addline(buffer, "%c", write_mask_chars[0]);
2051 if (write_mask & WINED3DSP_WRITEMASK_1)
2052 shader_addline(buffer, "%c", write_mask_chars[1]);
2053 if (write_mask & WINED3DSP_WRITEMASK_2)
2054 shader_addline(buffer, "%c", write_mask_chars[2]);
2055 if (write_mask & WINED3DSP_WRITEMASK_3)
2056 shader_addline(buffer, "%c", write_mask_chars[3]);
2057 }
2058 }
2059
2060 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2061 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2062 {
2063 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2064 DWORD swizzle = param->swizzle;
2065
2066 if (src_modifier == WINED3DSPSM_NEG
2067 || src_modifier == WINED3DSPSM_BIASNEG
2068 || src_modifier == WINED3DSPSM_SIGNNEG
2069 || src_modifier == WINED3DSPSM_X2NEG
2070 || src_modifier == WINED3DSPSM_ABSNEG)
2071 shader_addline(buffer, "-");
2072 else if (src_modifier == WINED3DSPSM_COMP)
2073 shader_addline(buffer, "1-");
2074 else if (src_modifier == WINED3DSPSM_NOT)
2075 shader_addline(buffer, "!");
2076
2077 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2078 shader_addline(buffer, "abs(");
2079
2080 shader_dump_register(buffer, &param->reg, shader_version);
2081
2082 switch (src_modifier)
2083 {
2084 case WINED3DSPSM_NONE: break;
2085 case WINED3DSPSM_NEG: break;
2086 case WINED3DSPSM_NOT: break;
2087 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2088 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2089 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2090 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2091 case WINED3DSPSM_COMP: break;
2092 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2093 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2094 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2095 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2096 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2097 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2098 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2099 }
2100
2101 if (swizzle != WINED3DSP_NOSWIZZLE)
2102 {
2103 static const char swizzle_chars[] = "xyzw";
2104 DWORD swizzle_x = swizzle & 0x03;
2105 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2106 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2107 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2108
2109 if (swizzle_x == swizzle_y
2110 && swizzle_x == swizzle_z
2111 && swizzle_x == swizzle_w)
2112 {
2113 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2114 }
2115 else
2116 {
2117 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2118 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2119 }
2120 }
2121 }
2122
2123 /* Shared code in order to generate the bulk of the shader string.
2124 * NOTE: A description of how to parse tokens can be found on MSDN. */
2125 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2126 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
2127 {
2128 struct wined3d_device *device = shader->device;
2129 const struct wined3d_shader_frontend *fe = shader->frontend;
2130 void *fe_data = shader->frontend_data;
2131 struct wined3d_shader_version shader_version;
2132 struct wined3d_shader_parser_state state;
2133 struct wined3d_shader_instruction ins;
2134 struct wined3d_shader_tex_mx tex_mx;
2135 struct wined3d_shader_context ctx;
2136 const DWORD *ptr = byte_code;
2137
2138 /* Initialize current parsing state. */
2139 tex_mx.current_row = 0;
2140 state.current_loop_depth = 0;
2141 state.current_loop_reg = 0;
2142 state.in_subroutine = FALSE;
2143
2144 ctx.shader = shader;
2145 ctx.gl_info = &device->adapter->gl_info;
2146 ctx.reg_maps = reg_maps;
2147 ctx.buffer = buffer;
2148 ctx.tex_mx = &tex_mx;
2149 ctx.state = &state;
2150 ctx.backend_data = backend_ctx;
2151 ins.ctx = &ctx;
2152
2153 fe->shader_read_header(fe_data, &ptr, &shader_version);
2154
2155 while (!fe->shader_is_end(fe_data, &ptr))
2156 {
2157 /* Read opcode. */
2158 fe->shader_read_instruction(fe_data, &ptr, &ins);
2159
2160 /* Unknown opcode and its parameters. */
2161 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2162 {
2163 TRACE("Skipping unrecognized instruction.\n");
2164 continue;
2165 }
2166
2167 if (ins.predicate)
2168 FIXME("Predicates not implemented.\n");
2169
2170 /* Call appropriate function for output target */
2171 device->shader_backend->shader_handle_instruction(&ins);
2172 }
2173 }
2174
2175 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2176 const struct wined3d_shader_dst_param *dst)
2177 {
2178 DWORD mmask = dst->modifiers;
2179
2180 switch (dst->shift)
2181 {
2182 case 0: break;
2183 case 13: shader_addline(buffer, "_d8"); break;
2184 case 14: shader_addline(buffer, "_d4"); break;
2185 case 15: shader_addline(buffer, "_d2"); break;
2186 case 1: shader_addline(buffer, "_x2"); break;
2187 case 2: shader_addline(buffer, "_x4"); break;
2188 case 3: shader_addline(buffer, "_x8"); break;
2189 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2190 }
2191
2192 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2193 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2194 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2195
2196 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2197 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2198 }
2199
2200 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2201 enum wined3d_primitive_type primitive_type)
2202 {
2203 switch (primitive_type)
2204 {
2205 case WINED3D_PT_UNDEFINED:
2206 shader_addline(buffer, "undefined");
2207 break;
2208 case WINED3D_PT_POINTLIST:
2209 shader_addline(buffer, "pointlist");
2210 break;
2211 case WINED3D_PT_LINELIST:
2212 shader_addline(buffer, "linelist");
2213 break;
2214 case WINED3D_PT_LINESTRIP:
2215 shader_addline(buffer, "linestrip");
2216 break;
2217 case WINED3D_PT_TRIANGLELIST:
2218 shader_addline(buffer, "trianglelist");
2219 break;
2220 case WINED3D_PT_TRIANGLESTRIP:
2221 shader_addline(buffer, "trianglestrip");
2222 break;
2223 case WINED3D_PT_TRIANGLEFAN:
2224 shader_addline(buffer, "trianglefan");
2225 break;
2226 case WINED3D_PT_LINELIST_ADJ:
2227 shader_addline(buffer, "linelist_adj");
2228 break;
2229 case WINED3D_PT_LINESTRIP_ADJ:
2230 shader_addline(buffer, "linestrip_adj");
2231 break;
2232 case WINED3D_PT_TRIANGLELIST_ADJ:
2233 shader_addline(buffer, "trianglelist_adj");
2234 break;
2235 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2236 shader_addline(buffer, "trianglestrip_adj");
2237 break;
2238 default:
2239 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2240 break;
2241 }
2242 }
2243
2244 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2245 enum wined3d_shader_interpolation_mode interpolation_mode)
2246 {
2247 switch (interpolation_mode)
2248 {
2249 case WINED3DSIM_CONSTANT:
2250 shader_addline(buffer, "constant");
2251 break;
2252 case WINED3DSIM_LINEAR:
2253 shader_addline(buffer, "linear");
2254 break;
2255 case WINED3DSIM_LINEAR_CENTROID:
2256 shader_addline(buffer, "linear centroid");
2257 break;
2258 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2259 shader_addline(buffer, "linear noperspective");
2260 break;
2261 case WINED3DSIM_LINEAR_SAMPLE:
2262 shader_addline(buffer, "linear sample");
2263 break;
2264 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2265 shader_addline(buffer, "linear noperspective centroid");
2266 break;
2267 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2268 shader_addline(buffer, "linear noperspective sample");
2269 break;
2270 default:
2271 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2272 break;
2273 }
2274 }
2275
2276 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2277 {
2278 struct wined3d_shader_version shader_version;
2279 struct wined3d_string_buffer buffer;
2280 const DWORD *ptr = byte_code;
2281 const char *type_prefix;
2282 const char *p, *q;
2283 DWORD i;
2284
2285 if (!string_buffer_init(&buffer))
2286 {
2287 ERR("Failed to initialize string buffer.\n");
2288 return;
2289 }
2290
2291 TRACE("Parsing %p.\n", byte_code);
2292
2293 fe->shader_read_header(fe_data, &ptr, &shader_version);
2294
2295 switch (shader_version.type)
2296 {
2297 case WINED3D_SHADER_TYPE_VERTEX:
2298 type_prefix = "vs";
2299 break;
2300
2301 case WINED3D_SHADER_TYPE_HULL:
2302 type_prefix = "hs";
2303 break;
2304
2305 case WINED3D_SHADER_TYPE_DOMAIN:
2306 type_prefix = "ds";
2307 break;
2308
2309 case WINED3D_SHADER_TYPE_GEOMETRY:
2310 type_prefix = "gs";
2311 break;
2312
2313 case WINED3D_SHADER_TYPE_PIXEL:
2314 type_prefix = "ps";
2315 break;
2316
2317 case WINED3D_SHADER_TYPE_COMPUTE:
2318 type_prefix = "cs";
2319 break;
2320
2321 default:
2322 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2323 type_prefix = "unknown";
2324 break;
2325 }
2326
2327 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2328
2329 while (!fe->shader_is_end(fe_data, &ptr))
2330 {
2331 struct wined3d_shader_instruction ins;
2332
2333 fe->shader_read_instruction(fe_data, &ptr, &ins);
2334 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2335 {
2336 WARN("Skipping unrecognized instruction.\n");
2337 shader_addline(&buffer, "<unrecognized instruction>\n");
2338 continue;
2339 }
2340
2341 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2342 {
2343 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2344 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2345 shader_addline(&buffer, " ");
2346 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2347 }
2348 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2349 {
2350 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2351 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2352 shader_addline(&buffer, ", %s",
2353 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2354 }
2355 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2356 {
2357 shader_addline(&buffer, "%s fb%u",
2358 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2359 }
2360 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2361 {
2362 shader_addline(&buffer, "%s ft%u = {...}",
2363 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2364 }
2365 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2366 {
2367 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2368 shader_dump_global_flags(&buffer, ins.flags);
2369 }
2370 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2371 {
2372 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2373 ins.declaration.max_tessellation_factor);
2374 }
2375 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2376 {
2377 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2378 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2379 {
2380 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2381 ins.declaration.icb->data[4 * i + 0],
2382 ins.declaration.icb->data[4 * i + 1],
2383 ins.declaration.icb->data[4 * i + 2],
2384 ins.declaration.icb->data[4 * i + 3]);
2385 }
2386 shader_addline(&buffer, "}");
2387 }
2388 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2389 {
2390 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2391 ins.declaration.indexable_temp.register_idx,
2392 ins.declaration.indexable_temp.register_size,
2393 ins.declaration.indexable_temp.component_count);
2394 }
2395 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2396 {
2397 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2398 shader_dump_interpolation_mode(&buffer, ins.flags);
2399 shader_addline(&buffer, " ");
2400 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2401 }
2402 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2403 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2404 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2405 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2406 {
2407 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2408 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2409 shader_addline(&buffer, ", ");
2410 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2411 }
2412 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2413 {
2414 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2415 shader_dump_interpolation_mode(&buffer, ins.flags);
2416 shader_addline(&buffer, " ");
2417 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2418 shader_addline(&buffer, ", ");
2419 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2420 }
2421 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2422 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2423 {
2424 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2425 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2426 }
2427 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2428 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2429 {
2430 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2431 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2432 }
2433 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2434 {
2435 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2436 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2437 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2438 }
2439 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2440 {
2441 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2442 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2443 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2444 }
2445 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2446 {
2447 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2448 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2449 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2450 shader_addline(&buffer, ", comparisonMode");
2451 }
2452 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2453 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2454 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2455 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2456 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2457 {
2458 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2459 }
2460 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2461 {
2462 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2463 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2464 }
2465 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2466 {
2467 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2468 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2469 }
2470 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2471 {
2472 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2473 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2474 }
2475 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2476 {
2477 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2478 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2479 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2480 }
2481 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2482 {
2483 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2484 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2485 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2486 ins.declaration.tgsm_structured.structure_count);
2487 }
2488 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2489 {
2490 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2491 ins.declaration.thread_group_size.x,
2492 ins.declaration.thread_group_size.y,
2493 ins.declaration.thread_group_size.z);
2494 }
2495 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2496 {
2497 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2498 shader_dump_uav_flags(&buffer, ins.flags);
2499 shader_addline(&buffer, " ");
2500 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2501 }
2502 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2503 {
2504 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2505 shader_dump_uav_flags(&buffer, ins.flags);
2506 shader_addline(&buffer, " ");
2507 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2508 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2509 }
2510 else if (ins.handler_idx == WINED3DSIH_DEF)
2511 {
2512 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2513 ins.dst[0].reg.idx[0].offset),
2514 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2515 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2516 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2517 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2518 }
2519 else if (ins.handler_idx == WINED3DSIH_DEFI)
2520 {
2521 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2522 ins.src[0].reg.u.immconst_data[0],
2523 ins.src[0].reg.u.immconst_data[1],
2524 ins.src[0].reg.u.immconst_data[2],
2525 ins.src[0].reg.u.immconst_data[3]);
2526 }
2527 else if (ins.handler_idx == WINED3DSIH_DEFB)
2528 {
2529 shader_addline(&buffer, "defb b%u = %s",
2530 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2531 }
2532 else
2533 {
2534 if (ins.predicate)
2535 {
2536 shader_addline(&buffer, "(");
2537 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2538 shader_addline(&buffer, ") ");
2539 }
2540
2541 /* PixWin marks instructions with the coissue flag with a '+' */
2542 if (ins.coissue)
2543 shader_addline(&buffer, "+");
2544
2545 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2546
2547 if (ins.handler_idx == WINED3DSIH_BREAKP
2548 || ins.handler_idx == WINED3DSIH_IF)
2549 {
2550 switch (ins.flags)
2551 {
2552 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2553 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2554 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2555 }
2556 }
2557 else if (ins.handler_idx == WINED3DSIH_IFC
2558 || ins.handler_idx == WINED3DSIH_BREAKC)
2559 {
2560 switch (ins.flags)
2561 {
2562 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2563 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2564 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2565 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2566 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2567 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2568 default: shader_addline(&buffer, "_(%u)", ins.flags);
2569 }
2570 }
2571 else if (ins.handler_idx == WINED3DSIH_TEX
2572 && shader_version.major >= 2
2573 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2574 {
2575 shader_addline(&buffer, "p");
2576 }
2577 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2578 {
2579 switch (ins.flags)
2580 {
2581 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2582 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2583 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2584 }
2585 }
2586 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2587 {
2588 switch (ins.flags)
2589 {
2590 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2591 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2592 }
2593 }
2594 else if (ins.handler_idx == WINED3DSIH_SYNC)
2595 {
2596 shader_dump_sync_flags(&buffer, ins.flags);
2597 }
2598
2599 if (wined3d_shader_instruction_has_texel_offset(&ins))
2600 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2601
2602 for (i = 0; i < ins.dst_count; ++i)
2603 {
2604 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2605 shader_addline(&buffer, !i ? " " : ", ");
2606 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2607 }
2608
2609 /* Other source tokens */
2610 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2611 {
2612 shader_addline(&buffer, !i ? " " : ", ");
2613 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2614 }
2615 }
2616 shader_addline(&buffer, "\n");
2617 }
2618
2619 for (p = buffer.buffer; *p; p = q)
2620 {
2621 if (!(q = strstr(p, "\n")))
2622 q = p + strlen(p);
2623 else
2624 ++q;
2625 TRACE(" %.*s", (int)(q - p), p);
2626 }
2627
2628 string_buffer_free(&buffer);
2629 }
2630
2631 static void shader_cleanup(struct wined3d_shader *shader)
2632 {
2633 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2634 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2635 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2636 shader->device->shader_backend->shader_destroy(shader);
2637 shader_cleanup_reg_maps(&shader->reg_maps);
2638 HeapFree(GetProcessHeap(), 0, shader->function);
2639 shader_delete_constant_list(&shader->constantsF);
2640 shader_delete_constant_list(&shader->constantsB);
2641 shader_delete_constant_list(&shader->constantsI);
2642 list_remove(&shader->shader_list_entry);
2643
2644 if (shader->frontend && shader->frontend_data)
2645 shader->frontend->shader_free(shader->frontend_data);
2646 }
2647
2648 struct shader_none_priv
2649 {
2650 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2651 const struct fragment_pipeline *fragment_pipe;
2652 BOOL ffp_proj_control;
2653 };
2654
2655 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2656 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2657 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2658 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2659 const struct wined3d_state *state) {}
2660 static void shader_none_destroy(struct wined3d_shader *shader) {}
2661 static void shader_none_free_context_data(struct wined3d_context *context) {}
2662 static void shader_none_init_context_state(struct wined3d_context *context) {}
2663
2664 /* Context activation is done by the caller. */
2665 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2666 const struct wined3d_state *state)
2667 {
2668 const struct wined3d_gl_info *gl_info = context->gl_info;
2669 struct shader_none_priv *priv = shader_priv;
2670
2671 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2672 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2673 }
2674
2675 /* Context activation is done by the caller. */
2676 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2677 {
2678 struct shader_none_priv *priv = shader_priv;
2679 const struct wined3d_gl_info *gl_info = context->gl_info;
2680
2681 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2682 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2683
2684 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2685 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2686 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2687 | (1u << WINED3D_SHADER_TYPE_HULL)
2688 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2689 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2690 }
2691
2692 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2693 const struct fragment_pipeline *fragment_pipe)
2694 {
2695 struct fragment_caps fragment_caps;
2696 void *vertex_priv, *fragment_priv;
2697 struct shader_none_priv *priv;
2698
2699 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2700 return E_OUTOFMEMORY;
2701
2702 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2703 {
2704 ERR("Failed to initialize vertex pipe.\n");
2705 HeapFree(GetProcessHeap(), 0, priv);
2706 return E_FAIL;
2707 }
2708
2709 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2710 {
2711 ERR("Failed to initialize fragment pipe.\n");
2712 vertex_pipe->vp_free(device);
2713 HeapFree(GetProcessHeap(), 0, priv);
2714 return E_FAIL;
2715 }
2716
2717 priv->vertex_pipe = vertex_pipe;
2718 priv->fragment_pipe = fragment_pipe;
2719 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2720 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2721
2722 device->vertex_priv = vertex_priv;
2723 device->fragment_priv = fragment_priv;
2724 device->shader_priv = priv;
2725
2726 return WINED3D_OK;
2727 }
2728
2729 static void shader_none_free(struct wined3d_device *device)
2730 {
2731 struct shader_none_priv *priv = device->shader_priv;
2732
2733 priv->fragment_pipe->free_private(device);
2734 priv->vertex_pipe->vp_free(device);
2735 HeapFree(GetProcessHeap(), 0, priv);
2736 }
2737
2738 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2739 {
2740 return TRUE;
2741 }
2742
2743 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2744 {
2745 /* Set the shader caps to 0 for the none shader backend */
2746 caps->vs_version = 0;
2747 caps->hs_version = 0;
2748 caps->ds_version = 0;
2749 caps->gs_version = 0;
2750 caps->ps_version = 0;
2751 caps->cs_version = 0;
2752 caps->vs_uniform_count = 0;
2753 caps->ps_uniform_count = 0;
2754 caps->ps_1x_max_value = 0.0f;
2755 caps->varying_count = 0;
2756 caps->wined3d_caps = 0;
2757 }
2758
2759 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2760 {
2761 /* We "support" every possible fixup, since we don't support any shader
2762 * model, and will never have to actually sample a texture. */
2763 return TRUE;
2764 }
2765
2766 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2767 {
2768 struct shader_none_priv *priv = shader_priv;
2769
2770 return priv->ffp_proj_control;
2771 }
2772
2773 const struct wined3d_shader_backend_ops none_shader_backend =
2774 {
2775 shader_none_handle_instruction,
2776 shader_none_select,
2777 shader_none_disable,
2778 shader_none_update_float_vertex_constants,
2779 shader_none_update_float_pixel_constants,
2780 shader_none_load_constants,
2781 shader_none_destroy,
2782 shader_none_alloc,
2783 shader_none_free,
2784 shader_none_allocate_context_data,
2785 shader_none_free_context_data,
2786 shader_none_init_context_state,
2787 shader_none_get_caps,
2788 shader_none_color_fixup_supported,
2789 shader_none_has_ffp_proj_control,
2790 };
2791
2792 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2793 DWORD float_const_count, enum wined3d_shader_type type, unsigned int max_version)
2794 {
2795 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2796 const struct wined3d_shader_frontend *fe;
2797 HRESULT hr;
2798 unsigned int backend_version;
2799 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2800
2801 TRACE("shader %p, byte_code %p, float_const_count %u, type %#x, max_version %u.\n",
2802 shader, byte_code, float_const_count, type, max_version);
2803
2804 list_init(&shader->constantsF);
2805 list_init(&shader->constantsB);
2806 list_init(&shader->constantsI);
2807 shader->lconst_inf_or_nan = FALSE;
2808 list_init(&reg_maps->indexable_temps);
2809
2810 fe = shader_select_frontend(*byte_code);
2811 if (!fe)
2812 {
2813 FIXME("Unable to find frontend for shader.\n");
2814 return WINED3DERR_INVALIDCALL;
2815 }
2816 shader->frontend = fe;
2817 shader->frontend_data = fe->shader_init(byte_code, &shader->output_signature);
2818 if (!shader->frontend_data)
2819 {
2820 FIXME("Failed to initialize frontend.\n");
2821 return WINED3DERR_INVALIDCALL;
2822 }
2823
2824 /* First pass: trace shader. */
2825 if (TRACE_ON(d3d_shader))
2826 shader_trace_init(fe, shader->frontend_data, byte_code);
2827
2828 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2829 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2830 &shader->output_signature, byte_code, float_const_count)))
2831 return hr;
2832
2833 if (reg_maps->shader_version.type != type)
2834 {
2835 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2836 return WINED3DERR_INVALIDCALL;
2837 }
2838 if (reg_maps->shader_version.major > max_version)
2839 {
2840 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2841 return WINED3DERR_INVALIDCALL;
2842 }
2843 switch (type)
2844 {
2845 case WINED3D_SHADER_TYPE_VERTEX:
2846 backend_version = d3d_info->limits.vs_version;
2847 break;
2848 case WINED3D_SHADER_TYPE_HULL:
2849 backend_version = d3d_info->limits.hs_version;
2850 break;
2851 case WINED3D_SHADER_TYPE_DOMAIN:
2852 backend_version = d3d_info->limits.ds_version;
2853 break;
2854 case WINED3D_SHADER_TYPE_GEOMETRY:
2855 backend_version = d3d_info->limits.gs_version;
2856 break;
2857 case WINED3D_SHADER_TYPE_PIXEL:
2858 backend_version = d3d_info->limits.ps_version;
2859 break;
2860 case WINED3D_SHADER_TYPE_COMPUTE:
2861 backend_version = d3d_info->limits.cs_version;
2862 break;
2863 default:
2864 FIXME("No backend version-checking for this shader type.\n");
2865 backend_version = 0;
2866 }
2867 if (reg_maps->shader_version.major > backend_version)
2868 {
2869 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2870 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2871 return WINED3DERR_INVALIDCALL;
2872 }
2873
2874 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength)))
2875 return E_OUTOFMEMORY;
2876 memcpy(shader->function, byte_code, shader->functionLength);
2877
2878 return WINED3D_OK;
2879 }
2880
2881 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2882 {
2883 ULONG refcount = InterlockedIncrement(&shader->ref);
2884
2885 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2886
2887 return refcount;
2888 }
2889
2890 static void wined3d_shader_destroy_object(void *object)
2891 {
2892 shader_cleanup(object);
2893 HeapFree(GetProcessHeap(), 0, object);
2894 }
2895
2896 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2897 {
2898 ULONG refcount = InterlockedDecrement(&shader->ref);
2899
2900 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2901
2902 if (!refcount)
2903 {
2904 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2905 wined3d_cs_emit_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2906 }
2907
2908 return refcount;
2909 }
2910
2911 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2912 {
2913 TRACE("shader %p.\n", shader);
2914
2915 return shader->parent;
2916 }
2917
2918 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2919 void *byte_code, UINT *byte_code_size)
2920 {
2921 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2922
2923 if (!byte_code)
2924 {
2925 *byte_code_size = shader->functionLength;
2926 return WINED3D_OK;
2927 }
2928
2929 if (*byte_code_size < shader->functionLength)
2930 {
2931 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2932 * than the required size we should write the required size and
2933 * return D3DERR_MOREDATA. That's not actually true. */
2934 return WINED3DERR_INVALIDCALL;
2935 }
2936
2937 memcpy(byte_code, shader->function, shader->functionLength);
2938
2939 return WINED3D_OK;
2940 }
2941
2942 /* Set local constants for d3d8 shaders. */
2943 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2944 UINT start_idx, const float *src_data, UINT count)
2945 {
2946 UINT end_idx = start_idx + count;
2947 UINT i;
2948
2949 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2950
2951 if (end_idx > shader->limits->constant_float)
2952 {
2953 WARN("end_idx %u > float constants limit %u.\n",
2954 end_idx, shader->limits->constant_float);
2955 end_idx = shader->limits->constant_float;
2956 }
2957
2958 for (i = start_idx; i < end_idx; ++i)
2959 {
2960 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2961 float *value;
2962 if (!lconst)
2963 return E_OUTOFMEMORY;
2964
2965 lconst->idx = i;
2966 value = (float *)lconst->value;
2967 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2968 list_add_head(&shader->constantsF, &lconst->entry);
2969
2970 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2971 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2972 {
2973 shader->lconst_inf_or_nan = TRUE;
2974 }
2975 }
2976
2977 return WINED3D_OK;
2978 }
2979
2980 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2981 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2982 {
2983 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2984 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2985 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2986 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2987 args->point_size = state->gl_primitive_type == GL_POINTS;
2988 args->per_vertex_point_size = shader->reg_maps.point_size;
2989 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
2990 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
2991 : WINED3D_SHADER_TYPE_PIXEL;
2992 if (shader->reg_maps.shader_version.major >= 4)
2993 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
2994 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
2995 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
2996 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
2997 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
2998 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
2999 else
3000 args->next_shader_input_count = 0;
3001 args->swizzle_map = swizzle_map;
3002 if (d3d_info->emulated_flatshading)
3003 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3004 else
3005 args->flatshading = 0;
3006 }
3007
3008 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3009 {
3010 if (usage_idx1 != usage_idx2)
3011 return FALSE;
3012 if (usage1 == usage2)
3013 return TRUE;
3014 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3015 return TRUE;
3016 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3017 return TRUE;
3018
3019 return FALSE;
3020 }
3021
3022 BOOL vshader_get_input(const struct wined3d_shader *shader,
3023 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3024 {
3025 WORD map = shader->reg_maps.input_registers;
3026 unsigned int i;
3027
3028 for (i = 0; map; map >>= 1, ++i)
3029 {
3030 if (!(map & 1)) continue;
3031
3032 if (match_usage(shader->u.vs.attributes[i].usage,
3033 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3034 {
3035 *regnum = i;
3036 return TRUE;
3037 }
3038 }
3039 return FALSE;
3040 }
3041
3042 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3043 const struct wined3d_shader_signature *src, char **signature_strings)
3044 {
3045 struct wined3d_shader_signature_element *e;
3046 unsigned int i;
3047 SIZE_T len;
3048 char *ptr;
3049
3050 if (!src->element_count)
3051 return WINED3D_OK;
3052
3053 ptr = *signature_strings;
3054
3055 dst->element_count = src->element_count;
3056 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3057 return E_OUTOFMEMORY;
3058
3059 for (i = 0; i < src->element_count; ++i)
3060 {
3061 e = &src->elements[i];
3062 dst->elements[i] = *e;
3063
3064 len = strlen(e->semantic_name);
3065 memcpy(ptr, e->semantic_name, len + 1);
3066 dst->elements[i].semantic_name = ptr;
3067 ptr += len + 1;
3068 }
3069
3070 *signature_strings = ptr;
3071
3072 return WINED3D_OK;
3073 }
3074
3075 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3076 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3077 void *parent, const struct wined3d_parent_ops *parent_ops)
3078 {
3079 struct wined3d_shader_signature_element *e;
3080 SIZE_T total, len;
3081 unsigned int i;
3082 HRESULT hr;
3083 char *ptr;
3084
3085 if (!desc->byte_code)
3086 return WINED3DERR_INVALIDCALL;
3087
3088 shader->ref = 1;
3089 shader->device = device;
3090 shader->parent = parent;
3091 shader->parent_ops = parent_ops;
3092
3093 total = 0;
3094 for (i = 0; i < desc->input_signature.element_count; ++i)
3095 {
3096 e = &desc->input_signature.elements[i];
3097 len = strlen(e->semantic_name);
3098 if (len >= ~(SIZE_T)0 - total)
3099 return E_OUTOFMEMORY;
3100
3101 total += len + 1;
3102 }
3103 for (i = 0; i < desc->output_signature.element_count; ++i)
3104 {
3105 e = &desc->output_signature.elements[i];
3106 len = strlen(e->semantic_name);
3107 if (len >= ~(SIZE_T)0 - total)
3108 return E_OUTOFMEMORY;
3109
3110 total += len + 1;
3111 }
3112 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3113 return E_OUTOFMEMORY;
3114 ptr = shader->signature_strings;
3115
3116 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3117 {
3118 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3119 return hr;
3120 }
3121 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3122 {
3123 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3124 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3125 return hr;
3126 }
3127
3128 list_init(&shader->linked_programs);
3129 list_add_head(&device->shaders, &shader->shader_list_entry);
3130
3131 if (FAILED(hr = shader_set_function(shader, desc->byte_code,
3132 float_const_count, type, desc->max_version)))
3133 {
3134 WARN("Failed to set function, hr %#x.\n", hr);
3135 shader_cleanup(shader);
3136 }
3137
3138 shader->load_local_constsF = shader->lconst_inf_or_nan;
3139
3140 return hr;
3141 }
3142
3143 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3144 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3145 {
3146 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3147 unsigned int i;
3148 HRESULT hr;
3149
3150 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3151 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3152 return hr;
3153
3154 for (i = 0; i < shader->input_signature.element_count; ++i)
3155 {
3156 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3157
3158 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3159 continue;
3160
3161 shader->u.vs.attributes[input->register_idx].usage =
3162 shader_usage_from_semantic_name(input->semantic_name);
3163 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3164 }
3165
3166 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3167 shader->load_local_constsF = TRUE;
3168
3169 return WINED3D_OK;
3170 }
3171
3172 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3173 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3174 {
3175 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3176 }
3177
3178 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3179 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3180 {
3181 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3182 }
3183
3184 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3185 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3186 {
3187 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3188 }
3189
3190 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3191 struct gs_compile_args *args)
3192 {
3193 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3194 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3195 }
3196
3197 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3198 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3199 {
3200 const struct wined3d_gl_info *gl_info = context->gl_info;
3201 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3202 const struct wined3d_texture *texture;
3203 UINT i;
3204
3205 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3206 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3207 {
3208 static unsigned int warned = 0;
3209
3210 args->srgb_correction = 1;
3211 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3212 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3213 "support, expect rendering artifacts.\n");
3214 }
3215
3216 if (shader->reg_maps.shader_version.major == 1
3217 && shader->reg_maps.shader_version.minor <= 3)
3218 {
3219 for (i = 0; i < shader->limits->sampler; ++i)
3220 {
3221 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3222
3223 if (flags & WINED3D_TTFF_PROJECTED)
3224 {
3225 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3226
3227 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3228 {
3229 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3230 unsigned int j;
3231 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3232 DWORD max_valid = WINED3D_TTFF_COUNT4;
3233
3234 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3235 {
3236 struct wined3d_vertex_declaration_element *element =
3237 &state->vertex_declaration->elements[j];
3238
3239 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3240 && element->usage_idx == index)
3241 {
3242 max_valid = element->format->component_count;
3243 break;
3244 }
3245 }
3246 if (!tex_transform || tex_transform > max_valid)
3247 {
3248 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3249 tex_transform, max_valid);
3250 tex_transform = max_valid;
3251 }
3252 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3253 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3254 && tex_transform > WINED3D_TTFF_COUNT2)
3255 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3256 && tex_transform > WINED3D_TTFF_COUNT3))
3257 tex_transform |= WINED3D_PSARGS_PROJECTED;
3258 else
3259 {
3260 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3261 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3262 i, tex_transform, resource_type);
3263 }
3264 }
3265 else
3266 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3267
3268 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3269 }
3270 }
3271 }
3272 if (shader->reg_maps.shader_version.major == 1
3273 && shader->reg_maps.shader_version.minor <= 4)
3274 {
3275 for (i = 0; i < shader->limits->sampler; ++i)
3276 {
3277 const struct wined3d_texture *texture = state->textures[i];
3278
3279 if (!shader->reg_maps.resource_info[i].type)
3280 continue;
3281
3282 /* Treat unbound textures as 2D. The dummy texture will provide
3283 * the proper sample value. The tex_types bitmap defaults to
3284 * 2D because of the memset. */
3285 if (!texture)
3286 continue;
3287
3288 switch (texture->target)
3289 {
3290 /* RECT textures are distinguished from 2D textures via np2_fixup */
3291 case GL_TEXTURE_RECTANGLE_ARB:
3292 case GL_TEXTURE_2D:
3293 break;
3294
3295 case GL_TEXTURE_3D:
3296 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3297 break;
3298
3299 case GL_TEXTURE_CUBE_MAP_ARB:
3300 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3301 break;
3302 }
3303 }
3304 }
3305
3306 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3307 {
3308 if (!shader->reg_maps.resource_info[i].type)
3309 continue;
3310
3311 texture = state->textures[i];
3312 if (!texture)
3313 {
3314 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3315 continue;
3316 }
3317 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3318 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3319 else
3320 args->color_fixup[i] = texture->resource.format->color_fixup;
3321
3322 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3323 args->shadow |= 1u << i;
3324
3325 /* Flag samplers that need NP2 texcoord fixup. */
3326 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3327 args->np2_fixup |= (1u << i);
3328 }
3329
3330 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3331 if (shader->reg_maps.shader_version.major >= 4)
3332 args->shadow = 0;
3333
3334 if (shader->reg_maps.shader_version.major >= 3)
3335 {
3336 if (position_transformed)
3337 args->vp_mode = pretransformed;
3338 else if (use_vs(state))
3339 args->vp_mode = vertexshader;
3340 else
3341 args->vp_mode = fixedfunction;
3342 args->fog = WINED3D_FFP_PS_FOG_OFF;
3343 }
3344 else
3345 {
3346 args->vp_mode = vertexshader;
3347 if (state->render_states[WINED3D_RS_FOGENABLE])
3348 {
3349 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3350 {
3351 case WINED3D_FOG_NONE:
3352 if (position_transformed || use_vs(state))
3353 {
3354 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3355 break;
3356 }
3357
3358 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3359 {
3360 case WINED3D_FOG_NONE: /* Fall through. */
3361 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3362 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3363 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3364 }
3365 break;
3366
3367 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3368 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3369 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3370 }
3371 }
3372 else
3373 {
3374 args->fog = WINED3D_FFP_PS_FOG_OFF;
3375 }
3376 }
3377
3378 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3379 {
3380 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3381
3382 args->texcoords_initialized = 0;
3383 for (i = 0; i < MAX_TEXTURES; ++i)
3384 {
3385 if (vs)
3386 {
3387 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3388 args->texcoords_initialized |= 1u << i;
3389 }
3390 else
3391 {
3392 const struct wined3d_stream_info *si = &context->stream_info;
3393 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3394
3395 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3396 & WINED3D_FFP_TCI_MASK
3397 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3398 args->texcoords_initialized |= 1u << i;
3399 }
3400 }
3401 }
3402 else
3403 {
3404 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3405 }
3406
3407 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3408 && state->gl_primitive_type == GL_POINTS;
3409
3410 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3411 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3412 else
3413 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3414 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3415 : WINED3D_CMP_ALWAYS) - 1;
3416
3417 if (d3d_info->emulated_flatshading)
3418 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3419
3420 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3421 ? context->render_offscreen : 0;
3422 }
3423
3424 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3425 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3426 {
3427 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3428 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3429 HRESULT hr;
3430
3431 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3432 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3433 return hr;
3434
3435 for (i = 0; i < MAX_REG_INPUT; ++i)
3436 {
3437 if (shader->u.ps.input_reg_used[i])
3438 {
3439 ++num_regs_used;
3440 highest_reg_used = i;
3441 }
3442 }
3443
3444 /* Don't do any register mapping magic if it is not needed, or if we can't
3445 * achieve anything anyway */
3446 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3447 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3448 || shader->reg_maps.shader_version.major >= 4)
3449 {
3450 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3451 {
3452 /* This happens with relative addressing. The input mapper function
3453 * warns about this if the higher registers are declared too, so
3454 * don't write a FIXME here */
3455 WARN("More varying registers used than supported\n");
3456 }
3457
3458 for (i = 0; i < MAX_REG_INPUT; ++i)
3459 {
3460 shader->u.ps.input_reg_map[i] = i;
3461 }
3462
3463 shader->u.ps.declared_in_count = highest_reg_used + 1;
3464 }
3465 else
3466 {
3467 shader->u.ps.declared_in_count = 0;
3468 for (i = 0; i < MAX_REG_INPUT; ++i)
3469 {
3470 if (shader->u.ps.input_reg_used[i])
3471 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3472 else shader->u.ps.input_reg_map[i] = ~0U;
3473 }
3474 }
3475
3476 return WINED3D_OK;
3477 }
3478
3479 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3480 {
3481 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3482 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3483 unsigned int i;
3484
3485 if (reg_maps->shader_version.major != 1) return;
3486
3487 for (i = 0; i < shader->limits->sampler; ++i)
3488 {
3489 /* We don't sample from this sampler. */
3490 if (!resource_info[i].type)
3491 continue;
3492
3493 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3494 {
3495 case WINED3D_SHADER_TEX_2D:
3496 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3497 break;
3498
3499 case WINED3D_SHADER_TEX_3D:
3500 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3501 break;
3502
3503 case WINED3D_SHADER_TEX_CUBE:
3504 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3505 break;
3506 }
3507 }
3508 }
3509
3510 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3511 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3512 {
3513 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3514 }
3515
3516 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3517 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3518 {
3519 struct wined3d_shader *object;
3520 HRESULT hr;
3521
3522 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3523 device, desc, parent, parent_ops, shader);
3524
3525 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3526 return E_OUTOFMEMORY;
3527
3528 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3529 {
3530 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3531 HeapFree(GetProcessHeap(), 0, object);
3532 return hr;
3533 }
3534
3535 TRACE("Created compute shader %p.\n", object);
3536 *shader = object;
3537
3538 return WINED3D_OK;
3539 }
3540
3541 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3542 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3543 {
3544 struct wined3d_shader *object;
3545 HRESULT hr;
3546
3547 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3548 device, desc, parent, parent_ops, shader);
3549
3550 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3551 return E_OUTOFMEMORY;
3552
3553 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3554 {
3555 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3556 HeapFree(GetProcessHeap(), 0, object);
3557 return hr;
3558 }
3559
3560 TRACE("Created domain shader %p.\n", object);
3561 *shader = object;
3562
3563 return WINED3D_OK;
3564 }
3565
3566 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3567 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3568 {
3569 struct wined3d_shader *object;
3570 HRESULT hr;
3571
3572 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3573 device, desc, parent, parent_ops, shader);
3574
3575 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3576 if (!object)
3577 return E_OUTOFMEMORY;
3578
3579 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3580 {
3581 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3582 HeapFree(GetProcessHeap(), 0, object);
3583 return hr;
3584 }
3585
3586 TRACE("Created geometry shader %p.\n", object);
3587 *shader = object;
3588
3589 return WINED3D_OK;
3590 }
3591
3592 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3593 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3594 {
3595 struct wined3d_shader *object;
3596 HRESULT hr;
3597
3598 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3599 device, desc, parent, parent_ops, shader);
3600
3601 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3602 return E_OUTOFMEMORY;
3603
3604 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3605 {
3606 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3607 HeapFree(GetProcessHeap(), 0, object);
3608 return hr;
3609 }
3610
3611 TRACE("Created hull shader %p.\n", object);
3612 *shader = object;
3613
3614 return WINED3D_OK;
3615 }
3616
3617 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3618 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3619 {
3620 struct wined3d_shader *object;
3621 HRESULT hr;
3622
3623 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3624 device, desc, parent, parent_ops, shader);
3625
3626 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3627 if (!object)
3628 return E_OUTOFMEMORY;
3629
3630 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3631 {
3632 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3633 HeapFree(GetProcessHeap(), 0, object);
3634 return hr;
3635 }
3636
3637 TRACE("Created pixel shader %p.\n", object);
3638 *shader = object;
3639
3640 return WINED3D_OK;
3641 }
3642
3643 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3644 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3645 {
3646 struct wined3d_shader *object;
3647 HRESULT hr;
3648
3649 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3650 device, desc, parent, parent_ops, shader);
3651
3652 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3653 if (!object)
3654 return E_OUTOFMEMORY;
3655
3656 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3657 {
3658 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3659 HeapFree(GetProcessHeap(), 0, object);
3660 return hr;
3661 }
3662
3663 TRACE("Created vertex shader %p.\n", object);
3664 *shader = object;
3665
3666 return WINED3D_OK;
3667 }
Windows API implementation