?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
user32: Move *RegisterDeviceNotification*() to input.c.
[wine.git] / dlls / wined3d / shader.c
blobf4c4eb07d88a3766412c8bf0437cec29d8f51d47
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 <stdio.h>
29 #include <string.h>
30
31 #include "wined3d_private.h"
32
33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
34
35 const struct wined3d_vec4 wined3d_srgb_const[] =
36 {
37 /* pow, mul_high, sub_high, mul_low */
38 {0.41666f, 1.055f, 0.055f, 12.92f},
39 /* cmp */
40 {0.0031308f, 0.0f, 0.0f, 0.0f},
41 };
42
43 static const char * const shader_opcode_names[] =
44 {
45 /* WINED3DSIH_ABS */ "abs",
46 /* WINED3DSIH_ADD */ "add",
47 /* WINED3DSIH_AND */ "and",
48 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
49 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
50 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
51 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
52 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
53 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
54 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
55 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
56 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
57 /* WINED3DSIH_BEM */ "bem",
58 /* WINED3DSIH_BFI */ "bfi",
59 /* WINED3DSIH_BFREV */ "bfrev",
60 /* WINED3DSIH_BREAK */ "break",
61 /* WINED3DSIH_BREAKC */ "breakc",
62 /* WINED3DSIH_BREAKP */ "breakp",
63 /* WINED3DSIH_BUFINFO */ "bufinfo",
64 /* WINED3DSIH_CALL */ "call",
65 /* WINED3DSIH_CALLNZ */ "callnz",
66 /* WINED3DSIH_CASE */ "case",
67 /* WINED3DSIH_CMP */ "cmp",
68 /* WINED3DSIH_CND */ "cnd",
69 /* WINED3DSIH_CONTINUE */ "continue",
70 /* WINED3DSIH_CONTINUEP */ "continuec",
71 /* WINED3DSIH_COUNTBITS */ "countbits",
72 /* WINED3DSIH_CRS */ "crs",
73 /* WINED3DSIH_CUT */ "cut",
74 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
75 /* WINED3DSIH_DCL */ "dcl",
76 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
77 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
78 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
79 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
80 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
81 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
82 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
83 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
84 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
85 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
86 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
87 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
88 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
89 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
90 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
91 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
92 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
93 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
94 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
95 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
96 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
97 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
98 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
99 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
100 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
101 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
102 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
103 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
104 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
105 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
106 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
107 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
108 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
109 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
110 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
111 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
112 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
113 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
114 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
115 /* WINED3DSIH_DEF */ "def",
116 /* WINED3DSIH_DEFAULT */ "default",
117 /* WINED3DSIH_DEFB */ "defb",
118 /* WINED3DSIH_DEFI */ "defi",
119 /* WINED3DSIH_DIV */ "div",
120 /* WINED3DSIH_DP2 */ "dp2",
121 /* WINED3DSIH_DP2ADD */ "dp2add",
122 /* WINED3DSIH_DP3 */ "dp3",
123 /* WINED3DSIH_DP4 */ "dp4",
124 /* WINED3DSIH_DST */ "dst",
125 /* WINED3DSIH_DSX */ "dsx",
126 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
127 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
128 /* WINED3DSIH_DSY */ "dsy",
129 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
130 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
131 /* WINED3DSIH_ELSE */ "else",
132 /* WINED3DSIH_EMIT */ "emit",
133 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
134 /* WINED3DSIH_ENDIF */ "endif",
135 /* WINED3DSIH_ENDLOOP */ "endloop",
136 /* WINED3DSIH_ENDREP */ "endrep",
137 /* WINED3DSIH_ENDSWITCH */ "endswitch",
138 /* WINED3DSIH_EQ */ "eq",
139 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
140 /* WINED3DSIH_EXP */ "exp",
141 /* WINED3DSIH_EXPP */ "expp",
142 /* WINED3DSIH_F16TOF32 */ "f16tof32",
143 /* WINED3DSIH_F32TOF16 */ "f32tof16",
144 /* WINED3DSIH_FCALL */ "fcall",
145 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
146 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
147 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
148 /* WINED3DSIH_FRC */ "frc",
149 /* WINED3DSIH_FTOI */ "ftoi",
150 /* WINED3DSIH_FTOU */ "ftou",
151 /* WINED3DSIH_GATHER4 */ "gather4",
152 /* WINED3DSIH_GATHER4_C */ "gather4_c",
153 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
154 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
155 /* WINED3DSIH_GE */ "ge",
156 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
157 /* WINED3DSIH_HS_DECLS */ "hs_decls",
158 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
159 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
160 /* WINED3DSIH_IADD */ "iadd",
161 /* WINED3DSIH_IBFE */ "ibfe",
162 /* WINED3DSIH_IEQ */ "ieq",
163 /* WINED3DSIH_IF */ "if",
164 /* WINED3DSIH_IFC */ "ifc",
165 /* WINED3DSIH_IGE */ "ige",
166 /* WINED3DSIH_ILT */ "ilt",
167 /* WINED3DSIH_IMAD */ "imad",
168 /* WINED3DSIH_IMAX */ "imax",
169 /* WINED3DSIH_IMIN */ "imin",
170 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
171 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
172 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
173 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
174 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
175 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
176 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
177 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
178 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
179 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
180 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
181 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
182 /* WINED3DSIH_IMUL */ "imul",
183 /* WINED3DSIH_INE */ "ine",
184 /* WINED3DSIH_INEG */ "ineg",
185 /* WINED3DSIH_ISHL */ "ishl",
186 /* WINED3DSIH_ISHR */ "ishr",
187 /* WINED3DSIH_ITOF */ "itof",
188 /* WINED3DSIH_LABEL */ "label",
189 /* WINED3DSIH_LD */ "ld",
190 /* WINED3DSIH_LD2DMS */ "ld2dms",
191 /* WINED3DSIH_LD_RAW */ "ld_raw",
192 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
193 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
194 /* WINED3DSIH_LIT */ "lit",
195 /* WINED3DSIH_LOD */ "lod",
196 /* WINED3DSIH_LOG */ "log",
197 /* WINED3DSIH_LOGP */ "logp",
198 /* WINED3DSIH_LOOP */ "loop",
199 /* WINED3DSIH_LRP */ "lrp",
200 /* WINED3DSIH_LT */ "lt",
201 /* WINED3DSIH_M3x2 */ "m3x2",
202 /* WINED3DSIH_M3x3 */ "m3x3",
203 /* WINED3DSIH_M3x4 */ "m3x4",
204 /* WINED3DSIH_M4x3 */ "m4x3",
205 /* WINED3DSIH_M4x4 */ "m4x4",
206 /* WINED3DSIH_MAD */ "mad",
207 /* WINED3DSIH_MAX */ "max",
208 /* WINED3DSIH_MIN */ "min",
209 /* WINED3DSIH_MOV */ "mov",
210 /* WINED3DSIH_MOVA */ "mova",
211 /* WINED3DSIH_MOVC */ "movc",
212 /* WINED3DSIH_MUL */ "mul",
213 /* WINED3DSIH_NE */ "ne",
214 /* WINED3DSIH_NOP */ "nop",
215 /* WINED3DSIH_NOT */ "not",
216 /* WINED3DSIH_NRM */ "nrm",
217 /* WINED3DSIH_OR */ "or",
218 /* WINED3DSIH_PHASE */ "phase",
219 /* WINED3DSIH_POW */ "pow",
220 /* WINED3DSIH_RCP */ "rcp",
221 /* WINED3DSIH_REP */ "rep",
222 /* WINED3DSIH_RESINFO */ "resinfo",
223 /* WINED3DSIH_RET */ "ret",
224 /* WINED3DSIH_RETP */ "retp",
225 /* WINED3DSIH_ROUND_NE */ "round_ne",
226 /* WINED3DSIH_ROUND_NI */ "round_ni",
227 /* WINED3DSIH_ROUND_PI */ "round_pi",
228 /* WINED3DSIH_ROUND_Z */ "round_z",
229 /* WINED3DSIH_RSQ */ "rsq",
230 /* WINED3DSIH_SAMPLE */ "sample",
231 /* WINED3DSIH_SAMPLE_B */ "sample_b",
232 /* WINED3DSIH_SAMPLE_C */ "sample_c",
233 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
234 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
235 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
236 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
237 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
238 /* WINED3DSIH_SETP */ "setp",
239 /* WINED3DSIH_SGE */ "sge",
240 /* WINED3DSIH_SGN */ "sgn",
241 /* WINED3DSIH_SINCOS */ "sincos",
242 /* WINED3DSIH_SLT */ "slt",
243 /* WINED3DSIH_SQRT */ "sqrt",
244 /* WINED3DSIH_STORE_RAW */ "store_raw",
245 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
246 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
247 /* WINED3DSIH_SUB */ "sub",
248 /* WINED3DSIH_SWAPC */ "swapc",
249 /* WINED3DSIH_SWITCH */ "switch",
250 /* WINED3DSIH_SYNC */ "sync",
251 /* WINED3DSIH_TEX */ "texld",
252 /* WINED3DSIH_TEXBEM */ "texbem",
253 /* WINED3DSIH_TEXBEML */ "texbeml",
254 /* WINED3DSIH_TEXCOORD */ "texcrd",
255 /* WINED3DSIH_TEXDEPTH */ "texdepth",
256 /* WINED3DSIH_TEXDP3 */ "texdp3",
257 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
258 /* WINED3DSIH_TEXKILL */ "texkill",
259 /* WINED3DSIH_TEXLDD */ "texldd",
260 /* WINED3DSIH_TEXLDL */ "texldl",
261 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
262 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
263 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
264 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
265 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
266 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
267 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
268 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
269 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
270 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
271 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
272 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
273 /* WINED3DSIH_UBFE */ "ubfe",
274 /* WINED3DSIH_UDIV */ "udiv",
275 /* WINED3DSIH_UGE */ "uge",
276 /* WINED3DSIH_ULT */ "ult",
277 /* WINED3DSIH_UMAX */ "umax",
278 /* WINED3DSIH_UMIN */ "umin",
279 /* WINED3DSIH_UMUL */ "umul",
280 /* WINED3DSIH_USHR */ "ushr",
281 /* WINED3DSIH_UTOF */ "utof",
282 /* WINED3DSIH_XOR */ "xor",
283 };
284
285 static const char * const semantic_names[] =
286 {
287 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
288 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
289 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
290 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
291 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
292 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
293 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
294 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
295 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
296 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
297 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
298 /* WINED3D_DECL_USAGE_FOG */ "FOG",
299 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
300 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
301 };
302
303 static const struct
304 {
305 enum wined3d_shader_input_sysval_semantic sysval_semantic;
306 const char *sysval_name;
307 }
308 shader_input_sysval_semantic_names[] =
309 {
310 {WINED3D_SIV_POSITION, "position"},
311 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
312 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
313 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
314 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
315 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
316 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
317 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
318 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
319 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
320 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
322 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
323 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
324 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
325 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
326 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
327 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
328 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
329 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
330 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
331 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
332 };
333
334 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
335 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
336
337 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
338 {
339 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
340 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
341
342 return shader_opcode_names[handler_idx];
343 }
344
345 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
346 {
347 if (usage >= ARRAY_SIZE(semantic_names))
348 {
349 FIXME("Unrecognized usage %#x.\n", usage);
350 return "UNRECOGNIZED";
351 }
352
353 return semantic_names[usage];
354 }
355
356 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
357 {
358 unsigned int i;
359
360 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
361 {
362 if (!strcmp(name, semantic_names[i]))
363 return i;
364 }
365
366 return ~0U;
367 }
368
369 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
370 {
371 switch (usage)
372 {
373 case WINED3D_DECL_USAGE_POSITION:
374 return WINED3D_SV_POSITION;
375 default:
376 return 0;
377 }
378 }
379
380 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
381 {
382 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
383 }
384
385 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
386 const struct wined3d_shader_semantic *s)
387 {
388 e->semantic_name = shader_semantic_name_from_usage(s->usage);
389 e->semantic_idx = s->usage_idx;
390 e->stream_idx = 0;
391 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
392 e->component_type = WINED3D_TYPE_FLOAT;
393 e->register_idx = s->reg.reg.idx[0].offset;
394 e->mask = s->reg.write_mask;
395 }
396
397 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
398 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
399 {
400 e->semantic_name = shader_semantic_name_from_usage(usage);
401 e->semantic_idx = usage_idx;
402 e->stream_idx = 0;
403 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
404 e->component_type = WINED3D_TYPE_FLOAT;
405 e->register_idx = reg_idx;
406 e->mask = write_mask;
407 }
408
409 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
410 {
411 switch (format)
412 {
413 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
414 return &sm1_shader_frontend;
415
416 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
417 return &sm4_shader_frontend;
418
419 default:
420 WARN("Invalid byte code format %#x specified.\n", format);
421 return NULL;
422 }
423 }
424
425 void string_buffer_clear(struct wined3d_string_buffer *buffer)
426 {
427 buffer->buffer[0] = '\0';
428 buffer->content_size = 0;
429 }
430
431 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
432 {
433 buffer->buffer_size = 32;
434 if (!(buffer->buffer = heap_alloc(buffer->buffer_size)))
435 {
436 ERR("Failed to allocate shader buffer memory.\n");
437 return FALSE;
438 }
439
440 string_buffer_clear(buffer);
441 return TRUE;
442 }
443
444 void string_buffer_free(struct wined3d_string_buffer *buffer)
445 {
446 heap_free(buffer->buffer);
447 }
448
449 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
450 {
451 char *new_buffer;
452 unsigned int new_buffer_size = buffer->buffer_size * 2;
453
454 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
455 new_buffer_size *= 2;
456 if (!(new_buffer = heap_realloc(buffer->buffer, new_buffer_size)))
457 {
458 ERR("Failed to grow buffer.\n");
459 buffer->buffer[buffer->content_size] = '\0';
460 return FALSE;
461 }
462 buffer->buffer = new_buffer;
463 buffer->buffer_size = new_buffer_size;
464 return TRUE;
465 }
466
467 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
468 {
469 unsigned int rem;
470 int rc;
471
472 rem = buffer->buffer_size - buffer->content_size;
473 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
474 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
475 return rc;
476
477 buffer->content_size += rc;
478 return 0;
479 }
480
481 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
482 {
483 va_list args;
484 int ret;
485
486 for (;;)
487 {
488 va_start(args, format);
489 ret = shader_vaddline(buffer, format, args);
490 va_end(args);
491 if (!ret)
492 return ret;
493 if (!string_buffer_resize(buffer, ret))
494 return -1;
495 }
496 }
497
498 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
499 {
500 struct wined3d_string_buffer *buffer;
501
502 if (list_empty(&list->list))
503 {
504 buffer = heap_alloc(sizeof(*buffer));
505 if (!buffer || !string_buffer_init(buffer))
506 {
507 ERR("Couldn't allocate buffer for temporary string.\n");
508 heap_free(buffer);
509 return NULL;
510 }
511 }
512 else
513 {
514 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
515 list_remove(&buffer->entry);
516 }
517 string_buffer_clear(buffer);
518 return buffer;
519 }
520
521 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
522 {
523 if (!buffer)
524 return 0;
525 string_buffer_clear(buffer);
526 return shader_vaddline(buffer, format, args);
527 }
528
529 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
530 {
531 va_list args;
532 int ret;
533
534 for (;;)
535 {
536 va_start(args, format);
537 ret = string_buffer_vsprintf(buffer, format, args);
538 va_end(args);
539 if (!ret)
540 return;
541 if (!string_buffer_resize(buffer, ret))
542 return;
543 }
544 }
545
546 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
547 {
548 if (!buffer)
549 return;
550 list_add_head(&list->list, &buffer->entry);
551 }
552
553 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
554 {
555 list_init(&list->list);
556 }
557
558 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
559 {
560 struct wined3d_string_buffer *buffer, *buffer_next;
561
562 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
563 {
564 string_buffer_free(buffer);
565 heap_free(buffer);
566 }
567 list_init(&list->list);
568 }
569
570 /* Convert floating point offset relative to a register file to an absolute
571 * offset for float constants. */
572 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
573 {
574 switch (register_type)
575 {
576 case WINED3DSPR_CONST: return register_idx;
577 case WINED3DSPR_CONST2: return 2048 + register_idx;
578 case WINED3DSPR_CONST3: return 4096 + register_idx;
579 case WINED3DSPR_CONST4: return 6144 + register_idx;
580 default:
581 FIXME("Unsupported register type: %u.\n", register_type);
582 return register_idx;
583 }
584 }
585
586 static void shader_delete_constant_list(struct list *clist)
587 {
588 struct wined3d_shader_lconst *constant, *constant_next;
589
590 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
591 heap_free(constant);
592 list_init(clist);
593 }
594
595 static void shader_set_limits(struct wined3d_shader *shader)
596 {
597 static const struct limits_entry
598 {
599 unsigned int min_version;
600 unsigned int max_version;
601 struct wined3d_shader_limits limits;
602 }
603 vs_limits[] =
604 {
605 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
606 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
607 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
608 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
609 * even though they are capable of supporting much more (GL
610 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
611 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
612 * shaders to 256. */
613 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
614 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
615 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
616 {0}
617 },
618 hs_limits[] =
619 {
620 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
621 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
622 },
623 ds_limits[] =
624 {
625 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
626 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
627 },
628 gs_limits[] =
629 {
630 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
631 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
632 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
633 {0}
634 },
635 ps_limits[] =
636 {
637 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
638 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
639 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
640 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
641 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
642 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
643 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
644 {0}
645 },
646 cs_limits[] =
647 {
648 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
649 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
650 };
651 const struct limits_entry *limits_array;
652 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
653 shader->reg_maps.shader_version.minor);
654 int i = 0;
655
656 switch (shader->reg_maps.shader_version.type)
657 {
658 default:
659 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
660 /* Fall-through. */
661 case WINED3D_SHADER_TYPE_VERTEX:
662 limits_array = vs_limits;
663 break;
664 case WINED3D_SHADER_TYPE_HULL:
665 limits_array = hs_limits;
666 break;
667 case WINED3D_SHADER_TYPE_DOMAIN:
668 limits_array = ds_limits;
669 break;
670 case WINED3D_SHADER_TYPE_GEOMETRY:
671 limits_array = gs_limits;
672 break;
673 case WINED3D_SHADER_TYPE_PIXEL:
674 limits_array = ps_limits;
675 break;
676 case WINED3D_SHADER_TYPE_COMPUTE:
677 limits_array = cs_limits;
678 break;
679 }
680
681 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
682 {
683 if (shader_version <= limits_array[i].max_version)
684 {
685 shader->limits = &limits_array[i].limits;
686 break;
687 }
688 ++i;
689 }
690 if (!shader->limits)
691 {
692 FIXME("Unexpected shader version \"%u.%u\".\n",
693 shader->reg_maps.shader_version.major,
694 shader->reg_maps.shader_version.minor);
695 shader->limits = &limits_array[max(0, i - 1)].limits;
696 }
697 }
698
699 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
700 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
701 {
702 switch (reg->type)
703 {
704 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
705 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
706 reg_maps->texcoord |= 1u << reg->idx[0].offset;
707 else
708 reg_maps->address |= 1u << reg->idx[0].offset;
709 break;
710
711 case WINED3DSPR_TEMP:
712 reg_maps->temporary |= 1u << reg->idx[0].offset;
713 break;
714
715 case WINED3DSPR_INPUT:
716 if (reg->idx[0].rel_addr)
717 reg_maps->input_rel_addressing = 1;
718 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
719 {
720 /* If relative addressing is used, we must assume that all
721 * registers are used. Even if it is a construct like v3[aL],
722 * we can't assume that v0, v1 and v2 aren't read because aL
723 * can be negative. */
724 if (reg->idx[0].rel_addr)
725 shader->u.ps.input_reg_used = ~0u;
726 else
727 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
728 }
729 else
730 {
731 reg_maps->input_registers |= 1u << reg->idx[0].offset;
732 }
733 break;
734
735 case WINED3DSPR_RASTOUT:
736 if (reg->idx[0].offset == 1)
737 reg_maps->fog = 1;
738 if (reg->idx[0].offset == 2)
739 reg_maps->point_size = 1;
740 break;
741
742 case WINED3DSPR_MISCTYPE:
743 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
744 {
745 if (!reg->idx[0].offset)
746 reg_maps->vpos = 1;
747 else if (reg->idx[0].offset == 1)
748 reg_maps->usesfacing = 1;
749 }
750 break;
751
752 case WINED3DSPR_CONST:
753 if (reg->idx[0].rel_addr)
754 {
755 if (reg->idx[0].offset < reg_maps->min_rel_offset)
756 reg_maps->min_rel_offset = reg->idx[0].offset;
757 if (reg->idx[0].offset > reg_maps->max_rel_offset)
758 reg_maps->max_rel_offset = reg->idx[0].offset;
759 reg_maps->usesrelconstF = TRUE;
760 }
761 else
762 {
763 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
764 {
765 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
766 return FALSE;
767 }
768 else
769 {
770 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
771 }
772 }
773 break;
774
775 case WINED3DSPR_CONSTINT:
776 if (reg->idx[0].offset >= shader->limits->constant_int)
777 {
778 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
779 return FALSE;
780 }
781 else
782 {
783 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
784 }
785 break;
786
787 case WINED3DSPR_CONSTBOOL:
788 if (reg->idx[0].offset >= shader->limits->constant_bool)
789 {
790 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
791 return FALSE;
792 }
793 else
794 {
795 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
796 }
797 break;
798
799 case WINED3DSPR_COLOROUT:
800 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
801 break;
802
803 case WINED3DSPR_OUTCONTROLPOINT:
804 reg_maps->vocp = 1;
805 break;
806
807 case WINED3DSPR_SAMPLEMASK:
808 reg_maps->sample_mask = 1;
809 break;
810
811 default:
812 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
813 reg->type, reg->idx[0].offset, reg->idx[1].offset);
814 break;
815 }
816 return TRUE;
817 }
818
819 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
820 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
821 {
822 struct wined3d_shader_sampler_map_entry *entries, *entry;
823 struct wined3d_shader_sampler_map *map;
824 unsigned int i;
825
826 map = &reg_maps->sampler_map;
827 entries = map->entries;
828 for (i = 0; i < map->count; ++i)
829 {
830 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
831 return;
832 }
833
834 if (!map->size)
835 {
836 if (!(entries = heap_calloc(4, sizeof(*entries))))
837 {
838 ERR("Failed to allocate sampler map entries.\n");
839 return;
840 }
841 map->size = 4;
842 map->entries = entries;
843 }
844 else if (map->count == map->size)
845 {
846 size_t new_size = map->size * 2;
847
848 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
849 || !(entries = heap_realloc(entries, sizeof(*entries) * new_size)))
850 {
851 ERR("Failed to resize sampler map entries.\n");
852 return;
853 }
854 map->size = new_size;
855 map->entries = entries;
856 }
857
858 entry = &entries[map->count++];
859 entry->resource_idx = resource_idx;
860 entry->sampler_idx = sampler_idx;
861 entry->bind_idx = bind_idx;
862 }
863
864 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
865 {
866 switch (instr)
867 {
868 case WINED3DSIH_M4x4:
869 case WINED3DSIH_M3x4:
870 return param == 1 ? 3 : 0;
871
872 case WINED3DSIH_M4x3:
873 case WINED3DSIH_M3x3:
874 return param == 1 ? 2 : 0;
875
876 case WINED3DSIH_M3x2:
877 return param == 1 ? 1 : 0;
878
879 default:
880 return 0;
881 }
882 }
883
884 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
885 unsigned int register_idx, unsigned int size, unsigned int stride)
886 {
887 struct wined3d_shader_tgsm *tgsm;
888
889 if (register_idx >= MAX_TGSM_REGISTERS)
890 {
891 ERR("Invalid TGSM register index %u.\n", register_idx);
892 return S_OK;
893 }
894 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
895 {
896 FIXME("TGSM declarations are allowed only in compute shaders.\n");
897 return S_OK;
898 }
899
900 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
901 register_idx + 1, sizeof(*reg_maps->tgsm)))
902 return E_OUTOFMEMORY;
903
904 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
905 tgsm = &reg_maps->tgsm[register_idx];
906 tgsm->size = size;
907 tgsm->stride = stride;
908 return S_OK;
909 }
910
911 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
912 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
913 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
914 {
915 struct wined3d_shader_phase *phase;
916
917 if ((phase = *current_phase))
918 {
919 phase->end = previous_instruction_ptr;
920 *current_phase = NULL;
921 }
922
923 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
924 {
925 ERR("Unexpected shader type %s.\n", debug_shader_type(shader->reg_maps.shader_version.type));
926 return E_FAIL;
927 }
928
929 switch (ins->handler_idx)
930 {
931 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
932 if (shader->u.hs.phases.control_point)
933 {
934 FIXME("Multiple control point phases.\n");
935 heap_free(shader->u.hs.phases.control_point);
936 }
937 if (!(shader->u.hs.phases.control_point = heap_alloc_zero(sizeof(*shader->u.hs.phases.control_point))))
938 return E_OUTOFMEMORY;
939 phase = shader->u.hs.phases.control_point;
940 break;
941 case WINED3DSIH_HS_FORK_PHASE:
942 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
943 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
944 sizeof(*shader->u.hs.phases.fork)))
945 return E_OUTOFMEMORY;
946 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
947 break;
948 case WINED3DSIH_HS_JOIN_PHASE:
949 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
950 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
951 sizeof(*shader->u.hs.phases.join)))
952 return E_OUTOFMEMORY;
953 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
954 break;
955 default:
956 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
957 return E_FAIL;
958 }
959
960 phase->start = current_instruction_ptr;
961 *current_phase = phase;
962
963 return WINED3D_OK;
964 }
965
966 static HRESULT shader_calculate_clip_or_cull_distance_mask(
967 const struct wined3d_shader_signature_element *e, unsigned int *mask)
968 {
969 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
970 * the only allowed semantic indices.
971 */
972 if (e->semantic_idx >= WINED3D_MAX_CLIP_DISTANCES / 4)
973 {
974 *mask = 0;
975 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
976 return WINED3DERR_INVALIDCALL;
977 }
978
979 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
980 return WINED3D_OK;
981 }
982
983 static void wined3d_insert_interpolation_mode(DWORD *packed_interpolation_mode,
984 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
985 {
986 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
987 FIXME("Unexpected interpolation mode %#x.\n", mode);
988
989 wined3d_insert_bits(packed_interpolation_mode,
990 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
991 }
992
993 static HRESULT shader_scan_output_signature(struct wined3d_shader *shader)
994 {
995 const struct wined3d_shader_signature *output_signature = &shader->output_signature;
996 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
997 unsigned int i;
998 HRESULT hr;
999
1000 for (i = 0; i < output_signature->element_count; ++i)
1001 {
1002 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1003 unsigned int mask;
1004
1005 reg_maps->output_registers |= 1u << e->register_idx;
1006 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
1007 {
1008 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1009 return hr;
1010 reg_maps->clip_distance_mask |= mask;
1011 }
1012 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
1013 {
1014 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1015 return hr;
1016 reg_maps->cull_distance_mask |= mask;
1017 }
1018 else if (e->sysval_semantic == WINED3D_SV_VIEWPORT_ARRAY_INDEX)
1019 {
1020 reg_maps->viewport_array = 1;
1021 }
1022 }
1023
1024 return WINED3D_OK;
1025 }
1026
1027 /* Note that this does not count the loop register as an address register. */
1028 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, DWORD constf_size)
1029 {
1030 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
1031 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
1032 struct wined3d_shader_signature *output_signature = &shader->output_signature;
1033 struct wined3d_shader_signature *input_signature = &shader->input_signature;
1034 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1035 const struct wined3d_shader_frontend *fe = shader->frontend;
1036 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
1037 struct wined3d_shader_version shader_version;
1038 struct wined3d_shader_phase *phase = NULL;
1039 const DWORD *ptr, *prev_ins, *current_ins;
1040 void *fe_data = shader->frontend_data;
1041 unsigned int i;
1042 HRESULT hr;
1043
1044 memset(reg_maps, 0, sizeof(*reg_maps));
1045 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1046 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1047 reg_maps->min_rel_offset = ~0U;
1048 list_init(&reg_maps->indexable_temps);
1049
1050 fe->shader_read_header(fe_data, &ptr, &shader_version);
1051 prev_ins = current_ins = ptr;
1052 reg_maps->shader_version = shader_version;
1053
1054 shader_set_limits(shader);
1055
1056 if (!(reg_maps->constf = heap_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1057 sizeof(*reg_maps->constf))))
1058 {
1059 ERR("Failed to allocate constant map memory.\n");
1060 return E_OUTOFMEMORY;
1061 }
1062
1063 while (!fe->shader_is_end(fe_data, &ptr))
1064 {
1065 struct wined3d_shader_instruction ins;
1066
1067 current_ins = ptr;
1068 /* Fetch opcode. */
1069 fe->shader_read_instruction(fe_data, &ptr, &ins);
1070
1071 /* Unhandled opcode, and its parameters. */
1072 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1073 {
1074 WARN("Encountered unrecognised or invalid instruction.\n");
1075 return WINED3DERR_INVALIDCALL;
1076 }
1077
1078 /* Handle declarations. */
1079 if (ins.handler_idx == WINED3DSIH_DCL
1080 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1081 {
1082 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1083 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1084
1085 switch (semantic->reg.reg.type)
1086 {
1087 /* Mark input registers used. */
1088 case WINED3DSPR_INPUT:
1089 if (reg_idx >= MAX_REG_INPUT)
1090 {
1091 ERR("Invalid input register index %u.\n", reg_idx);
1092 break;
1093 }
1094 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1095 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1096 return WINED3DERR_INVALIDCALL;
1097 reg_maps->input_registers |= 1u << reg_idx;
1098 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1099 break;
1100
1101 /* Vertex shader: mark 3.0 output registers used, save token. */
1102 case WINED3DSPR_OUTPUT:
1103 if (reg_idx >= MAX_REG_OUTPUT)
1104 {
1105 ERR("Invalid output register index %u.\n", reg_idx);
1106 break;
1107 }
1108 reg_maps->output_registers |= 1u << reg_idx;
1109 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1110 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1111 reg_maps->fog = 1;
1112 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1113 reg_maps->point_size = 1;
1114 break;
1115
1116 case WINED3DSPR_SAMPLER:
1117 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1118 case WINED3DSPR_RESOURCE:
1119 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1120 {
1121 ERR("Invalid resource index %u.\n", reg_idx);
1122 break;
1123 }
1124 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1125 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1126 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1127 break;
1128
1129 case WINED3DSPR_UAV:
1130 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1131 {
1132 ERR("Invalid UAV resource index %u.\n", reg_idx);
1133 break;
1134 }
1135 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1136 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1137 if (ins.flags)
1138 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1139 break;
1140
1141 default:
1142 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1143 break;
1144 }
1145 }
1146 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1147 {
1148 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1149 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1150 {
1151 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1152 }
1153 else
1154 {
1155 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1156 wined3d_bitmap_set(&reg_maps->cb_map, reg->idx[0].offset);
1157 }
1158 }
1159 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1160 {
1161 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1162 {
1163 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1164 shader->u.ps.force_early_depth_stencil = TRUE;
1165 else
1166 FIXME("Invalid instruction %#x for shader type %#x.\n",
1167 ins.handler_idx, shader_version.type);
1168 }
1169 else
1170 {
1171 WARN("Ignoring global flags %#x.\n", ins.flags);
1172 }
1173 }
1174 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1175 {
1176 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1177 shader->u.gs.instance_count = ins.declaration.count;
1178 else
1179 FIXME("Invalid instruction %#x for shader type %#x.\n",
1180 ins.handler_idx, shader_version.type);
1181 }
1182 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1183 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1184 {
1185 if (phase)
1186 phase->instance_count = ins.declaration.count;
1187 else
1188 FIXME("Instruction %s outside of shader phase.\n",
1189 debug_d3dshaderinstructionhandler(ins.handler_idx));
1190 }
1191 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1192 {
1193 if (reg_maps->icb)
1194 FIXME("Multiple immediate constant buffers.\n");
1195 reg_maps->icb = ins.declaration.icb;
1196 }
1197 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1198 {
1199 if (phase)
1200 {
1201 FIXME("Indexable temporary registers not supported.\n");
1202 }
1203 else
1204 {
1205 struct wined3d_shader_indexable_temp *reg;
1206
1207 if (!(reg = heap_alloc(sizeof(*reg))))
1208 return E_OUTOFMEMORY;
1209
1210 *reg = ins.declaration.indexable_temp;
1211 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1212 }
1213 }
1214 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1215 {
1216 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1217 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1218 else
1219 FIXME("Invalid instruction %#x for shader type %#x.\n",
1220 ins.handler_idx, shader_version.type);
1221 }
1222 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1223 {
1224 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1225 if (reg_idx >= MAX_REG_INPUT)
1226 {
1227 ERR("Invalid register index %u.\n", reg_idx);
1228 break;
1229 }
1230 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1231 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1232 else
1233 FIXME("Invalid instruction %#x for shader type %#x.\n",
1234 ins.handler_idx, shader_version.type);
1235 }
1236 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1237 {
1238 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1239 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1240 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1241 {
1242 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1243 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1244 else
1245 FIXME("Invalid instruction %#x for shader type %#x.\n",
1246 ins.handler_idx, shader_version.type);
1247 }
1248 }
1249 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1250 {
1251 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1252 shader->u.hs.output_vertex_count = ins.declaration.count;
1253 else
1254 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1255 }
1256 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1257 {
1258 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1259 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1260 else
1261 FIXME("Invalid instruction %#x for shader type %#x.\n",
1262 ins.handler_idx, shader_version.type);
1263 }
1264 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1265 {
1266 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1267 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1268 {
1269 ERR("Invalid resource index %u.\n", reg_idx);
1270 break;
1271 }
1272 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1273 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1274 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1275 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1276 }
1277 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1278 {
1279 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1280 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1281 {
1282 ERR("Invalid resource index %u.\n", reg_idx);
1283 break;
1284 }
1285 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1286 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1287 reg_maps->resource_info[reg_idx].flags = 0;
1288 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1289 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1290 }
1291 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1292 {
1293 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1294 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1295 }
1296 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1297 {
1298 if (phase)
1299 phase->temporary_count = ins.declaration.count;
1300 else
1301 reg_maps->temporary_count = ins.declaration.count;
1302 }
1303 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1304 {
1305 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1306 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1307 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1308 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1309 }
1310 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1311 {
1312 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1313 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1314 else
1315 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1316 }
1317 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1318 {
1319 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1320 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1321 else
1322 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1323 }
1324 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1325 {
1326 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1327 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1328 return hr;
1329 }
1330 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1331 {
1332 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1333 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1334 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1335 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1336 return hr;
1337 }
1338 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1339 {
1340 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1341 {
1342 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1343 }
1344 else
1345 {
1346 FIXME("Invalid instruction %#x for shader type %#x.\n",
1347 ins.handler_idx, shader_version.type);
1348 }
1349 }
1350 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1351 {
1352 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1353 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1354 {
1355 ERR("Invalid UAV resource index %u.\n", reg_idx);
1356 break;
1357 }
1358 if (ins.flags)
1359 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1360 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1361 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1362 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1363 }
1364 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1365 {
1366 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1367 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1368 {
1369 ERR("Invalid UAV resource index %u.\n", reg_idx);
1370 break;
1371 }
1372 if (ins.flags)
1373 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1374 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1375 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1376 reg_maps->uav_resource_info[reg_idx].flags = 0;
1377 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1378 }
1379 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1380 {
1381 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1382 shader->u.gs.vertices_out = ins.declaration.count;
1383 else
1384 FIXME("Invalid instruction %#x for shader type %#x.\n",
1385 ins.handler_idx, shader_version.type);
1386 }
1387 else if (ins.handler_idx == WINED3DSIH_DEF)
1388 {
1389 struct wined3d_shader_lconst *lconst;
1390 float *value;
1391
1392 if (!(lconst = heap_alloc(sizeof(*lconst))))
1393 return E_OUTOFMEMORY;
1394
1395 lconst->idx = ins.dst[0].reg.idx[0].offset;
1396 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1397 value = (float *)lconst->value;
1398
1399 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1400 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1401 {
1402 if (value[0] < -1.0f) value[0] = -1.0f;
1403 else if (value[0] > 1.0f) value[0] = 1.0f;
1404 if (value[1] < -1.0f) value[1] = -1.0f;
1405 else if (value[1] > 1.0f) value[1] = 1.0f;
1406 if (value[2] < -1.0f) value[2] = -1.0f;
1407 else if (value[2] > 1.0f) value[2] = 1.0f;
1408 if (value[3] < -1.0f) value[3] = -1.0f;
1409 else if (value[3] > 1.0f) value[3] = 1.0f;
1410 }
1411
1412 list_add_head(&shader->constantsF, &lconst->entry);
1413
1414 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1415 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1416 {
1417 shader->lconst_inf_or_nan = TRUE;
1418 }
1419 }
1420 else if (ins.handler_idx == WINED3DSIH_DEFI)
1421 {
1422 struct wined3d_shader_lconst *lconst;
1423
1424 if (!(lconst = heap_alloc(sizeof(*lconst))))
1425 return E_OUTOFMEMORY;
1426
1427 lconst->idx = ins.dst[0].reg.idx[0].offset;
1428 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1429
1430 list_add_head(&shader->constantsI, &lconst->entry);
1431 reg_maps->local_int_consts |= (1u << lconst->idx);
1432 }
1433 else if (ins.handler_idx == WINED3DSIH_DEFB)
1434 {
1435 struct wined3d_shader_lconst *lconst;
1436
1437 if (!(lconst = heap_alloc(sizeof(*lconst))))
1438 return E_OUTOFMEMORY;
1439
1440 lconst->idx = ins.dst[0].reg.idx[0].offset;
1441 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1442
1443 list_add_head(&shader->constantsB, &lconst->entry);
1444 reg_maps->local_bool_consts |= (1u << lconst->idx);
1445 }
1446 /* Handle shader phases. */
1447 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1448 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1449 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1450 {
1451 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1452 return hr;
1453 }
1454 /* For subroutine prototypes. */
1455 else if (ins.handler_idx == WINED3DSIH_LABEL)
1456 {
1457 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1458 }
1459 /* Set texture, address, temporary registers. */
1460 else
1461 {
1462 BOOL color0_mov = FALSE;
1463 unsigned int i;
1464
1465 /* This will loop over all the registers and try to
1466 * make a bitmask of the ones we're interested in.
1467 *
1468 * Relative addressing tokens are ignored, but that's
1469 * okay, since we'll catch any address registers when
1470 * they are initialized (required by spec). */
1471 for (i = 0; i < ins.dst_count; ++i)
1472 {
1473 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1474 shader_version.type, constf_size))
1475 return WINED3DERR_INVALIDCALL;
1476
1477 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1478 {
1479 UINT idx = ins.dst[i].reg.idx[0].offset;
1480
1481 switch (ins.dst[i].reg.type)
1482 {
1483 case WINED3DSPR_RASTOUT:
1484 if (shader_version.major >= 3)
1485 break;
1486 switch (idx)
1487 {
1488 case 0: /* oPos */
1489 reg_maps->output_registers |= 1u << 10;
1490 shader_signature_from_usage(&output_signature_elements[10],
1491 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1492 break;
1493
1494 case 1: /* oFog */
1495 reg_maps->output_registers |= 1u << 11;
1496 shader_signature_from_usage(&output_signature_elements[11],
1497 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1498 break;
1499
1500 case 2: /* oPts */
1501 reg_maps->output_registers |= 1u << 11;
1502 shader_signature_from_usage(&output_signature_elements[11],
1503 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1504 break;
1505 }
1506 break;
1507
1508 case WINED3DSPR_ATTROUT:
1509 if (shader_version.major >= 3)
1510 break;
1511 if (idx < 2)
1512 {
1513 idx += 8;
1514 if (reg_maps->output_registers & (1u << idx))
1515 {
1516 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1517 }
1518 else
1519 {
1520 reg_maps->output_registers |= 1u << idx;
1521 shader_signature_from_usage(&output_signature_elements[idx],
1522 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1523 }
1524 }
1525 break;
1526
1527 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1528 if (shader_version.major >= 3)
1529 {
1530 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1531 {
1532 WARN("Invalid output register index %u.\n", idx);
1533 break;
1534 }
1535 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1536 break;
1537 }
1538 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1539 {
1540 WARN("Invalid texcoord index %u.\n", idx);
1541 break;
1542 }
1543 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1544 if (reg_maps->output_registers & (1u << idx))
1545 {
1546 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1547 }
1548 else
1549 {
1550 reg_maps->output_registers |= 1u << idx;
1551 shader_signature_from_usage(&output_signature_elements[idx],
1552 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1553 }
1554 break;
1555
1556 default:
1557 break;
1558 }
1559 }
1560
1561 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1562 {
1563 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1564 {
1565 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1566 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1567 * the mov and perform the sRGB write correction from the source register.
1568 *
1569 * However, if the mov is only partial, we can't do this, and if the write
1570 * comes from an instruction other than MOV it is hard to do as well. If
1571 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1572 shader->u.ps.color0_mov = FALSE;
1573 if (ins.handler_idx == WINED3DSIH_MOV
1574 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1575 {
1576 /* Used later when the source register is read. */
1577 color0_mov = TRUE;
1578 }
1579 }
1580 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1581 * end
1582 */
1583 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1584 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1585 {
1586 shader->u.ps.color0_mov = FALSE;
1587 }
1588 }
1589
1590 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1591 if (shader_version.major == 1
1592 && (ins.handler_idx == WINED3DSIH_TEX
1593 || ins.handler_idx == WINED3DSIH_TEXBEM
1594 || ins.handler_idx == WINED3DSIH_TEXBEML
1595 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1596 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1597 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1598 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1599 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1600 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1601 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1602 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1603 {
1604 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1605
1606 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1607 {
1608 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1609 continue;
1610 }
1611
1612 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1613 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1614 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1615 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1616 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1617
1618 /* texbem is only valid with < 1.4 pixel shaders */
1619 if (ins.handler_idx == WINED3DSIH_TEXBEM
1620 || ins.handler_idx == WINED3DSIH_TEXBEML)
1621 {
1622 reg_maps->bumpmat |= 1u << reg_idx;
1623 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1624 {
1625 reg_maps->luminanceparams |= 1u << reg_idx;
1626 }
1627 }
1628 }
1629 else if (ins.handler_idx == WINED3DSIH_BEM)
1630 {
1631 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1632 }
1633 }
1634
1635 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1636 {
1637 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1638 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1639 {
1640 ERR("Invalid UAV index %u.\n", reg_idx);
1641 break;
1642 }
1643 reg_maps->uav_counter_mask |= (1u << reg_idx);
1644 }
1645 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1646 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1647 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1648 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1649 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1650 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1651 {
1652 unsigned int reg_idx;
1653 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1654 reg_idx = ins.src[1].reg.idx[0].offset;
1655 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1656 reg_idx = ins.src[2].reg.idx[0].offset;
1657 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1658 reg_idx = ins.dst[0].reg.idx[0].offset;
1659 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1660 reg_idx = ins.src[0].reg.idx[0].offset;
1661 else
1662 reg_idx = ins.dst[1].reg.idx[0].offset;
1663 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1664 {
1665 ERR("Invalid UAV index %u.\n", reg_idx);
1666 break;
1667 }
1668 reg_maps->uav_read_mask |= (1u << reg_idx);
1669 }
1670 else if (ins.handler_idx == WINED3DSIH_NRM)
1671 {
1672 reg_maps->usesnrm = 1;
1673 }
1674 else if (ins.handler_idx == WINED3DSIH_DSY
1675 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1676 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1677 {
1678 reg_maps->usesdsy = 1;
1679 }
1680 else if (ins.handler_idx == WINED3DSIH_DSX
1681 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1682 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1683 {
1684 reg_maps->usesdsx = 1;
1685 }
1686 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1687 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1688 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1689 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1690 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1691 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1692 else if (ins.handler_idx == WINED3DSIH_LOOP
1693 || ins.handler_idx == WINED3DSIH_REP)
1694 {
1695 ++cur_loop_depth;
1696 if (cur_loop_depth > max_loop_depth)
1697 max_loop_depth = cur_loop_depth;
1698 }
1699 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1700 || ins.handler_idx == WINED3DSIH_ENDREP)
1701 {
1702 --cur_loop_depth;
1703 }
1704 else if (ins.handler_idx == WINED3DSIH_GATHER4
1705 || ins.handler_idx == WINED3DSIH_GATHER4_C
1706 || ins.handler_idx == WINED3DSIH_SAMPLE
1707 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1708 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1709 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1710 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1711 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1712 {
1713 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1714 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1715 }
1716 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1717 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1718 {
1719 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1720 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1721 }
1722 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1723 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1724 {
1725 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1726 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1727 }
1728 else if (ins.handler_idx == WINED3DSIH_LD
1729 || ins.handler_idx == WINED3DSIH_LD2DMS
1730 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1731 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1732 {
1733 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1734 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1735 }
1736 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1737 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1738 {
1739 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1740 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1741 }
1742
1743 if (ins.predicate)
1744 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1745 shader_version.type, constf_size))
1746 return WINED3DERR_INVALIDCALL;
1747
1748 for (i = 0; i < ins.src_count; ++i)
1749 {
1750 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1751 struct wined3d_shader_register reg = ins.src[i].reg;
1752
1753 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1754 shader_version.type, constf_size))
1755 return WINED3DERR_INVALIDCALL;
1756 while (count)
1757 {
1758 ++reg.idx[0].offset;
1759 if (!shader_record_register_usage(shader, reg_maps, &reg,
1760 shader_version.type, constf_size))
1761 return WINED3DERR_INVALIDCALL;
1762 --count;
1763 }
1764
1765 if (color0_mov)
1766 {
1767 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1768 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1769 {
1770 shader->u.ps.color0_mov = TRUE;
1771 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1772 }
1773 }
1774 }
1775 }
1776
1777 prev_ins = current_ins;
1778 }
1779 reg_maps->loop_depth = max_loop_depth;
1780
1781 if (phase)
1782 {
1783 phase->end = prev_ins;
1784 phase = NULL;
1785 }
1786
1787 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1788 * R0 is written to the render target. */
1789 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1790 reg_maps->rt_mask |= (1u << 0);
1791
1792 if (input_signature->elements)
1793 {
1794 for (i = 0; i < input_signature->element_count; ++i)
1795 {
1796 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1797 {
1798 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1799 {
1800 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1801 return WINED3DERR_INVALIDCALL;
1802 }
1803 }
1804 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1805 {
1806 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1807 reg_maps->vpos = 1;
1808 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1809 reg_maps->usesfacing = 1;
1810 }
1811 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1812 }
1813 }
1814 else if (!input_signature->elements && reg_maps->input_registers)
1815 {
1816 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1817 struct wined3d_shader_signature_element *e;
1818 unsigned int i;
1819
1820 if (!(input_signature->elements = heap_calloc(count, sizeof(*input_signature->elements))))
1821 return E_OUTOFMEMORY;
1822 input_signature->element_count = count;
1823
1824 e = input_signature->elements;
1825 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1826 {
1827 if (!(reg_maps->input_registers & (1u << i)))
1828 continue;
1829 input_signature_elements[i].register_idx = i;
1830 *e++ = input_signature_elements[i];
1831 }
1832 }
1833
1834 if (output_signature->elements)
1835 {
1836 if (FAILED(hr = shader_scan_output_signature(shader)))
1837 return hr;
1838 }
1839 else if (reg_maps->output_registers)
1840 {
1841 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1842 struct wined3d_shader_signature_element *e;
1843
1844 if (!(output_signature->elements = heap_calloc(count, sizeof(*output_signature->elements))))
1845 return E_OUTOFMEMORY;
1846 output_signature->element_count = count;
1847
1848 e = output_signature->elements;
1849 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1850 {
1851 if (!(reg_maps->output_registers & (1u << i)))
1852 continue;
1853 *e++ = output_signature_elements[i];
1854 }
1855 }
1856
1857 return WINED3D_OK;
1858 }
1859
1860 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1861 {
1862 struct wined3d_shader_indexable_temp *reg, *reg_next;
1863
1864 heap_free(reg_maps->constf);
1865 heap_free(reg_maps->sampler_map.entries);
1866
1867 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1868 heap_free(reg);
1869 list_init(&reg_maps->indexable_temps);
1870
1871 heap_free(reg_maps->tgsm);
1872 }
1873
1874 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1875 {
1876 DWORD map = 1u << max;
1877 map |= map - 1;
1878 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1879
1880 return wined3d_log2i(map);
1881 }
1882
1883 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1884 {
1885 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1886 {
1887 shader_addline(buffer, "refactoringAllowed");
1888 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1889 if (global_flags)
1890 shader_addline(buffer, " | ");
1891 }
1892
1893 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1894 {
1895 shader_addline(buffer, "forceEarlyDepthStencil");
1896 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1897 if (global_flags)
1898 shader_addline(buffer, " | ");
1899 }
1900
1901 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1902 {
1903 shader_addline(buffer, "enableRawAndStructuredBuffers");
1904 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1905 }
1906
1907 if (global_flags)
1908 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1909 }
1910
1911 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1912 {
1913 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1914 {
1915 shader_addline(buffer, "_g");
1916 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1917 }
1918 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1919 {
1920 shader_addline(buffer, "_t");
1921 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1922 }
1923
1924 if (sync_flags)
1925 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1926 }
1927
1928 static void shader_dump_precise_flags(struct wined3d_string_buffer *buffer, DWORD flags)
1929 {
1930 if (!(flags & WINED3DSI_PRECISE_XYZW))
1931 return;
1932
1933 shader_addline(buffer, " [precise");
1934 if (flags != WINED3DSI_PRECISE_XYZW)
1935 {
1936 shader_addline(buffer, "(%s%s%s%s)",
1937 flags & WINED3DSI_PRECISE_X ? "x" : "",
1938 flags & WINED3DSI_PRECISE_Y ? "y" : "",
1939 flags & WINED3DSI_PRECISE_Z ? "z" : "",
1940 flags & WINED3DSI_PRECISE_W ? "w" : "");
1941 }
1942 shader_addline(buffer, "]");
1943 }
1944
1945 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1946 {
1947 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1948 {
1949 shader_addline(buffer, "_glc");
1950 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1951 }
1952 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1953 {
1954 shader_addline(buffer, "_opc");
1955 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1956 }
1957
1958 if (uav_flags)
1959 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1960 }
1961
1962 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1963 enum wined3d_tessellator_domain domain)
1964 {
1965 switch (domain)
1966 {
1967 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1968 shader_addline(buffer, "line");
1969 break;
1970 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1971 shader_addline(buffer, "triangle");
1972 break;
1973 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1974 shader_addline(buffer, "quad");
1975 break;
1976 default:
1977 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1978 break;
1979 }
1980 }
1981
1982 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1983 enum wined3d_tessellator_output_primitive output_primitive)
1984 {
1985 switch (output_primitive)
1986 {
1987 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1988 shader_addline(buffer, "point");
1989 break;
1990 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1991 shader_addline(buffer, "line");
1992 break;
1993 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1994 shader_addline(buffer, "triangle_cw");
1995 break;
1996 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1997 shader_addline(buffer, "triangle_ccw");
1998 break;
1999 default:
2000 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
2001 break;
2002 }
2003 }
2004
2005 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
2006 enum wined3d_tessellator_partitioning partitioning)
2007 {
2008 switch (partitioning)
2009 {
2010 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
2011 shader_addline(buffer, "integer");
2012 break;
2013 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
2014 shader_addline(buffer, "pow2");
2015 break;
2016 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
2017 shader_addline(buffer, "fractional_odd");
2018 break;
2019 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
2020 shader_addline(buffer, "fractional_even");
2021 break;
2022 default:
2023 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
2024 break;
2025 }
2026 }
2027
2028 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
2029 enum wined3d_shader_input_sysval_semantic semantic)
2030 {
2031 unsigned int i;
2032
2033 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
2034 {
2035 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
2036 {
2037 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
2038 return;
2039 }
2040 }
2041
2042 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
2043 }
2044
2045 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
2046 const struct wined3d_shader_semantic *semantic, unsigned int flags,
2047 const struct wined3d_shader_version *shader_version)
2048 {
2049 shader_addline(buffer, "dcl");
2050
2051 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
2052 {
2053 switch (semantic->resource_type)
2054 {
2055 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2056 shader_addline(buffer, "_2d");
2057 break;
2058
2059 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2060 shader_addline(buffer, "_3d");
2061 break;
2062
2063 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2064 shader_addline(buffer, "_cube");
2065 break;
2066
2067 default:
2068 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2069 break;
2070 }
2071 }
2072 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2073 {
2074 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2075 shader_addline(buffer, "_resource_");
2076 else
2077 shader_addline(buffer, "_uav_");
2078 switch (semantic->resource_type)
2079 {
2080 case WINED3D_SHADER_RESOURCE_BUFFER:
2081 shader_addline(buffer, "buffer");
2082 break;
2083
2084 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
2085 shader_addline(buffer, "texture1d");
2086 break;
2087
2088 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2089 shader_addline(buffer, "texture2d");
2090 break;
2091
2092 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
2093 shader_addline(buffer, "texture2dms");
2094 break;
2095
2096 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2097 shader_addline(buffer, "texture3d");
2098 break;
2099
2100 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2101 shader_addline(buffer, "texturecube");
2102 break;
2103
2104 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
2105 shader_addline(buffer, "texture1darray");
2106 break;
2107
2108 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
2109 shader_addline(buffer, "texture2darray");
2110 break;
2111
2112 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
2113 shader_addline(buffer, "texture2dmsarray");
2114 break;
2115
2116 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
2117 shader_addline(buffer, "texturecubearray");
2118 break;
2119
2120 default:
2121 shader_addline(buffer, "unknown");
2122 break;
2123 }
2124 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2125 shader_dump_uav_flags(buffer, flags);
2126 switch (semantic->resource_data_type)
2127 {
2128 case WINED3D_DATA_FLOAT:
2129 shader_addline(buffer, " (float)");
2130 break;
2131
2132 case WINED3D_DATA_INT:
2133 shader_addline(buffer, " (int)");
2134 break;
2135
2136 case WINED3D_DATA_UINT:
2137 shader_addline(buffer, " (uint)");
2138 break;
2139
2140 case WINED3D_DATA_UNORM:
2141 shader_addline(buffer, " (unorm)");
2142 break;
2143
2144 case WINED3D_DATA_SNORM:
2145 shader_addline(buffer, " (snorm)");
2146 break;
2147
2148 default:
2149 shader_addline(buffer, " (unknown)");
2150 break;
2151 }
2152 }
2153 else
2154 {
2155 /* Pixel shaders 3.0 don't have usage semantics. */
2156 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2157 return;
2158 else
2159 shader_addline(buffer, "_");
2160
2161 switch (semantic->usage)
2162 {
2163 case WINED3D_DECL_USAGE_POSITION:
2164 shader_addline(buffer, "position%u", semantic->usage_idx);
2165 break;
2166
2167 case WINED3D_DECL_USAGE_BLEND_INDICES:
2168 shader_addline(buffer, "blend");
2169 break;
2170
2171 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2172 shader_addline(buffer, "weight");
2173 break;
2174
2175 case WINED3D_DECL_USAGE_NORMAL:
2176 shader_addline(buffer, "normal%u", semantic->usage_idx);
2177 break;
2178
2179 case WINED3D_DECL_USAGE_PSIZE:
2180 shader_addline(buffer, "psize");
2181 break;
2182
2183 case WINED3D_DECL_USAGE_COLOR:
2184 if (!semantic->usage_idx)
2185 shader_addline(buffer, "color");
2186 else
2187 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2188 break;
2189
2190 case WINED3D_DECL_USAGE_TEXCOORD:
2191 shader_addline(buffer, "texture%u", semantic->usage_idx);
2192 break;
2193
2194 case WINED3D_DECL_USAGE_TANGENT:
2195 shader_addline(buffer, "tangent");
2196 break;
2197
2198 case WINED3D_DECL_USAGE_BINORMAL:
2199 shader_addline(buffer, "binormal");
2200 break;
2201
2202 case WINED3D_DECL_USAGE_TESS_FACTOR:
2203 shader_addline(buffer, "tessfactor");
2204 break;
2205
2206 case WINED3D_DECL_USAGE_POSITIONT:
2207 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2208 break;
2209
2210 case WINED3D_DECL_USAGE_FOG:
2211 shader_addline(buffer, "fog");
2212 break;
2213
2214 case WINED3D_DECL_USAGE_DEPTH:
2215 shader_addline(buffer, "depth");
2216 break;
2217
2218 case WINED3D_DECL_USAGE_SAMPLE:
2219 shader_addline(buffer, "sample");
2220 break;
2221
2222 default:
2223 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2224 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2225 }
2226 }
2227 }
2228
2229 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2230 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2231 {
2232 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2233 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2234 UINT offset = reg->idx[0].offset;
2235
2236 switch (reg->type)
2237 {
2238 case WINED3DSPR_TEMP:
2239 shader_addline(buffer, "r");
2240 break;
2241
2242 case WINED3DSPR_INPUT:
2243 shader_addline(buffer, "v");
2244 break;
2245
2246 case WINED3DSPR_CONST:
2247 case WINED3DSPR_CONST2:
2248 case WINED3DSPR_CONST3:
2249 case WINED3DSPR_CONST4:
2250 shader_addline(buffer, "c");
2251 offset = shader_get_float_offset(reg->type, offset);
2252 break;
2253
2254 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2255 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2256 break;
2257
2258 case WINED3DSPR_RASTOUT:
2259 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2260 break;
2261
2262 case WINED3DSPR_COLOROUT:
2263 shader_addline(buffer, "oC");
2264 break;
2265
2266 case WINED3DSPR_DEPTHOUT:
2267 shader_addline(buffer, "oDepth");
2268 break;
2269
2270 case WINED3DSPR_DEPTHOUTGE:
2271 shader_addline(buffer, "oDepthGE");
2272 break;
2273
2274 case WINED3DSPR_DEPTHOUTLE:
2275 shader_addline(buffer, "oDepthLE");
2276 break;
2277
2278 case WINED3DSPR_ATTROUT:
2279 shader_addline(buffer, "oD");
2280 break;
2281
2282 case WINED3DSPR_TEXCRDOUT:
2283 /* Vertex shaders >= 3.0 use general purpose output registers
2284 * (WINED3DSPR_OUTPUT), which can include an address token. */
2285 if (shader_version->major >= 3)
2286 shader_addline(buffer, "o");
2287 else
2288 shader_addline(buffer, "oT");
2289 break;
2290
2291 case WINED3DSPR_CONSTINT:
2292 shader_addline(buffer, "i");
2293 break;
2294
2295 case WINED3DSPR_CONSTBOOL:
2296 shader_addline(buffer, "b");
2297 break;
2298
2299 case WINED3DSPR_LABEL:
2300 shader_addline(buffer, "l");
2301 break;
2302
2303 case WINED3DSPR_LOOP:
2304 shader_addline(buffer, "aL");
2305 break;
2306
2307 case WINED3DSPR_SAMPLER:
2308 shader_addline(buffer, "s");
2309 break;
2310
2311 case WINED3DSPR_MISCTYPE:
2312 if (offset > 1)
2313 {
2314 FIXME("Unhandled misctype register %u.\n", offset);
2315 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2316 }
2317 else
2318 {
2319 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2320 }
2321 break;
2322
2323 case WINED3DSPR_PREDICATE:
2324 shader_addline(buffer, "p");
2325 break;
2326
2327 case WINED3DSPR_IMMCONST:
2328 shader_addline(buffer, "l");
2329 break;
2330
2331 case WINED3DSPR_CONSTBUFFER:
2332 shader_addline(buffer, "cb");
2333 break;
2334
2335 case WINED3DSPR_IMMCONSTBUFFER:
2336 shader_addline(buffer, "icb");
2337 break;
2338
2339 case WINED3DSPR_PRIMID:
2340 shader_addline(buffer, "primID");
2341 break;
2342
2343 case WINED3DSPR_NULL:
2344 shader_addline(buffer, "null");
2345 break;
2346
2347 case WINED3DSPR_RASTERIZER:
2348 shader_addline(buffer, "rasterizer");
2349 break;
2350
2351 case WINED3DSPR_RESOURCE:
2352 shader_addline(buffer, "t");
2353 break;
2354
2355 case WINED3DSPR_UAV:
2356 shader_addline(buffer, "u");
2357 break;
2358
2359 case WINED3DSPR_OUTPOINTID:
2360 shader_addline(buffer, "vOutputControlPointID");
2361 break;
2362
2363 case WINED3DSPR_FORKINSTID:
2364 shader_addline(buffer, "vForkInstanceId");
2365 break;
2366
2367 case WINED3DSPR_JOININSTID:
2368 shader_addline(buffer, "vJoinInstanceId");
2369 break;
2370
2371 case WINED3DSPR_INCONTROLPOINT:
2372 shader_addline(buffer, "vicp");
2373 break;
2374
2375 case WINED3DSPR_OUTCONTROLPOINT:
2376 shader_addline(buffer, "vocp");
2377 break;
2378
2379 case WINED3DSPR_PATCHCONST:
2380 shader_addline(buffer, "vpc");
2381 break;
2382
2383 case WINED3DSPR_TESSCOORD:
2384 shader_addline(buffer, "vDomainLocation");
2385 break;
2386
2387 case WINED3DSPR_GROUPSHAREDMEM:
2388 shader_addline(buffer, "g");
2389 break;
2390
2391 case WINED3DSPR_THREADID:
2392 shader_addline(buffer, "vThreadID");
2393 break;
2394
2395 case WINED3DSPR_THREADGROUPID:
2396 shader_addline(buffer, "vThreadGroupID");
2397 break;
2398
2399 case WINED3DSPR_LOCALTHREADID:
2400 shader_addline(buffer, "vThreadIDInGroup");
2401 break;
2402
2403 case WINED3DSPR_LOCALTHREADINDEX:
2404 shader_addline(buffer, "vThreadIDInGroupFlattened");
2405 break;
2406
2407 case WINED3DSPR_IDXTEMP:
2408 shader_addline(buffer, "x");
2409 break;
2410
2411 case WINED3DSPR_STREAM:
2412 shader_addline(buffer, "m");
2413 break;
2414
2415 case WINED3DSPR_FUNCTIONBODY:
2416 shader_addline(buffer, "fb");
2417 break;
2418
2419 case WINED3DSPR_FUNCTIONPOINTER:
2420 shader_addline(buffer, "fp");
2421 break;
2422
2423 case WINED3DSPR_COVERAGE:
2424 shader_addline(buffer, "vCoverage");
2425 break;
2426
2427 case WINED3DSPR_SAMPLEMASK:
2428 shader_addline(buffer, "oMask");
2429 break;
2430
2431 case WINED3DSPR_GSINSTID:
2432 shader_addline(buffer, "vGSInstanceID");
2433 break;
2434
2435 default:
2436 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2437 break;
2438 }
2439
2440 if (reg->type == WINED3DSPR_IMMCONST)
2441 {
2442 shader_addline(buffer, "(");
2443 switch (reg->immconst_type)
2444 {
2445 case WINED3D_IMMCONST_SCALAR:
2446 switch (reg->data_type)
2447 {
2448 case WINED3D_DATA_FLOAT:
2449 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2450 break;
2451 case WINED3D_DATA_INT:
2452 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2453 break;
2454 case WINED3D_DATA_RESOURCE:
2455 case WINED3D_DATA_SAMPLER:
2456 case WINED3D_DATA_UINT:
2457 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2458 break;
2459 default:
2460 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2461 break;
2462 }
2463 break;
2464
2465 case WINED3D_IMMCONST_VEC4:
2466 switch (reg->data_type)
2467 {
2468 case WINED3D_DATA_FLOAT:
2469 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2470 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2471 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2472 break;
2473 case WINED3D_DATA_INT:
2474 shader_addline(buffer, "%d, %d, %d, %d",
2475 reg->u.immconst_data[0], reg->u.immconst_data[1],
2476 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2477 break;
2478 case WINED3D_DATA_RESOURCE:
2479 case WINED3D_DATA_SAMPLER:
2480 case WINED3D_DATA_UINT:
2481 shader_addline(buffer, "%u, %u, %u, %u",
2482 reg->u.immconst_data[0], reg->u.immconst_data[1],
2483 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2484 break;
2485 default:
2486 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2487 break;
2488 }
2489 break;
2490
2491 default:
2492 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2493 break;
2494 }
2495 shader_addline(buffer, ")");
2496 }
2497 else if (reg->type != WINED3DSPR_RASTOUT
2498 && reg->type != WINED3DSPR_MISCTYPE
2499 && reg->type != WINED3DSPR_NULL)
2500 {
2501 if (offset != ~0u)
2502 {
2503 shader_addline(buffer, "[");
2504 if (reg->idx[0].rel_addr)
2505 {
2506 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2507 shader_addline(buffer, " + ");
2508 }
2509 shader_addline(buffer, "%u]", offset);
2510
2511 if (reg->idx[1].offset != ~0u)
2512 {
2513 shader_addline(buffer, "[");
2514 if (reg->idx[1].rel_addr)
2515 {
2516 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2517 shader_addline(buffer, " + ");
2518 }
2519 shader_addline(buffer, "%u]", reg->idx[1].offset);
2520 }
2521 }
2522
2523 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2524 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2525 }
2526 }
2527
2528 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2529 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2530 {
2531 DWORD write_mask = param->write_mask;
2532
2533 shader_dump_register(buffer, &param->reg, shader_version);
2534
2535 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2536 {
2537 static const char write_mask_chars[] = "xyzw";
2538
2539 shader_addline(buffer, ".");
2540 if (write_mask & WINED3DSP_WRITEMASK_0)
2541 shader_addline(buffer, "%c", write_mask_chars[0]);
2542 if (write_mask & WINED3DSP_WRITEMASK_1)
2543 shader_addline(buffer, "%c", write_mask_chars[1]);
2544 if (write_mask & WINED3DSP_WRITEMASK_2)
2545 shader_addline(buffer, "%c", write_mask_chars[2]);
2546 if (write_mask & WINED3DSP_WRITEMASK_3)
2547 shader_addline(buffer, "%c", write_mask_chars[3]);
2548 }
2549 }
2550
2551 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2552 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2553 {
2554 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2555 DWORD swizzle = param->swizzle;
2556
2557 if (src_modifier == WINED3DSPSM_NEG
2558 || src_modifier == WINED3DSPSM_BIASNEG
2559 || src_modifier == WINED3DSPSM_SIGNNEG
2560 || src_modifier == WINED3DSPSM_X2NEG
2561 || src_modifier == WINED3DSPSM_ABSNEG)
2562 shader_addline(buffer, "-");
2563 else if (src_modifier == WINED3DSPSM_COMP)
2564 shader_addline(buffer, "1-");
2565 else if (src_modifier == WINED3DSPSM_NOT)
2566 shader_addline(buffer, "!");
2567
2568 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2569 shader_addline(buffer, "abs(");
2570
2571 shader_dump_register(buffer, &param->reg, shader_version);
2572
2573 switch (src_modifier)
2574 {
2575 case WINED3DSPSM_NONE: break;
2576 case WINED3DSPSM_NEG: break;
2577 case WINED3DSPSM_NOT: break;
2578 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2579 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2580 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2581 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2582 case WINED3DSPSM_COMP: break;
2583 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2584 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2585 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2586 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2587 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2588 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2589 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2590 }
2591
2592 if (swizzle != WINED3DSP_NOSWIZZLE)
2593 {
2594 static const char swizzle_chars[] = "xyzw";
2595 DWORD swizzle_x = swizzle & 0x03;
2596 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2597 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2598 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2599
2600 if (swizzle_x == swizzle_y
2601 && swizzle_x == swizzle_z
2602 && swizzle_x == swizzle_w)
2603 {
2604 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2605 }
2606 else
2607 {
2608 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2609 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2610 }
2611 }
2612 }
2613
2614 /* Shared code in order to generate the bulk of the shader string. */
2615 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2616 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2617 const DWORD *start, const DWORD *end)
2618 {
2619 struct wined3d_device *device = shader->device;
2620 const struct wined3d_shader_frontend *fe = shader->frontend;
2621 void *fe_data = shader->frontend_data;
2622 struct wined3d_shader_version shader_version;
2623 struct wined3d_shader_parser_state state;
2624 struct wined3d_shader_instruction ins;
2625 struct wined3d_shader_tex_mx tex_mx;
2626 struct wined3d_shader_context ctx;
2627 const DWORD *ptr;
2628
2629 /* Initialize current parsing state. */
2630 tex_mx.current_row = 0;
2631 state.current_loop_depth = 0;
2632 state.current_loop_reg = 0;
2633 state.in_subroutine = FALSE;
2634
2635 ctx.shader = shader;
2636 ctx.reg_maps = reg_maps;
2637 ctx.buffer = buffer;
2638 ctx.tex_mx = &tex_mx;
2639 ctx.state = &state;
2640 ctx.backend_data = backend_ctx;
2641 ins.ctx = &ctx;
2642
2643 fe->shader_read_header(fe_data, &ptr, &shader_version);
2644 if (start)
2645 ptr = start;
2646
2647 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2648 {
2649 /* Read opcode. */
2650 fe->shader_read_instruction(fe_data, &ptr, &ins);
2651
2652 /* Unknown opcode and its parameters. */
2653 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2654 {
2655 WARN("Encountered unrecognised or invalid instruction.\n");
2656 return WINED3DERR_INVALIDCALL;
2657 }
2658
2659 if (ins.predicate)
2660 FIXME("Predicates not implemented.\n");
2661
2662 /* Call appropriate function for output target */
2663 device->shader_backend->shader_handle_instruction(&ins);
2664 }
2665
2666 return WINED3D_OK;
2667 }
2668
2669 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2670 const struct wined3d_shader_dst_param *dst)
2671 {
2672 DWORD mmask = dst->modifiers;
2673
2674 switch (dst->shift)
2675 {
2676 case 0: break;
2677 case 13: shader_addline(buffer, "_d8"); break;
2678 case 14: shader_addline(buffer, "_d4"); break;
2679 case 15: shader_addline(buffer, "_d2"); break;
2680 case 1: shader_addline(buffer, "_x2"); break;
2681 case 2: shader_addline(buffer, "_x4"); break;
2682 case 3: shader_addline(buffer, "_x8"); break;
2683 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2684 }
2685
2686 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2687 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2688 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2689
2690 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2691 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2692 }
2693
2694 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2695 const struct wined3d_shader_primitive_type *primitive_type)
2696 {
2697 switch (primitive_type->type)
2698 {
2699 case WINED3D_PT_UNDEFINED:
2700 shader_addline(buffer, "undefined");
2701 break;
2702 case WINED3D_PT_POINTLIST:
2703 shader_addline(buffer, "pointlist");
2704 break;
2705 case WINED3D_PT_LINELIST:
2706 shader_addline(buffer, "linelist");
2707 break;
2708 case WINED3D_PT_LINESTRIP:
2709 shader_addline(buffer, "linestrip");
2710 break;
2711 case WINED3D_PT_TRIANGLELIST:
2712 shader_addline(buffer, "trianglelist");
2713 break;
2714 case WINED3D_PT_TRIANGLESTRIP:
2715 shader_addline(buffer, "trianglestrip");
2716 break;
2717 case WINED3D_PT_TRIANGLEFAN:
2718 shader_addline(buffer, "trianglefan");
2719 break;
2720 case WINED3D_PT_LINELIST_ADJ:
2721 shader_addline(buffer, "linelist_adj");
2722 break;
2723 case WINED3D_PT_LINESTRIP_ADJ:
2724 shader_addline(buffer, "linestrip_adj");
2725 break;
2726 case WINED3D_PT_TRIANGLELIST_ADJ:
2727 shader_addline(buffer, "trianglelist_adj");
2728 break;
2729 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2730 shader_addline(buffer, "trianglestrip_adj");
2731 break;
2732 case WINED3D_PT_PATCH:
2733 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2734 break;
2735 default:
2736 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2737 break;
2738 }
2739 }
2740
2741 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2742 enum wined3d_shader_interpolation_mode interpolation_mode)
2743 {
2744 switch (interpolation_mode)
2745 {
2746 case WINED3DSIM_CONSTANT:
2747 shader_addline(buffer, "constant");
2748 break;
2749 case WINED3DSIM_LINEAR:
2750 shader_addline(buffer, "linear");
2751 break;
2752 case WINED3DSIM_LINEAR_CENTROID:
2753 shader_addline(buffer, "linear centroid");
2754 break;
2755 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2756 shader_addline(buffer, "linear noperspective");
2757 break;
2758 case WINED3DSIM_LINEAR_SAMPLE:
2759 shader_addline(buffer, "linear sample");
2760 break;
2761 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2762 shader_addline(buffer, "linear noperspective centroid");
2763 break;
2764 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2765 shader_addline(buffer, "linear noperspective sample");
2766 break;
2767 default:
2768 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2769 break;
2770 }
2771 }
2772
2773 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2774 {
2775 struct wined3d_shader_version shader_version;
2776 struct wined3d_string_buffer buffer;
2777 const char *type_prefix;
2778 const char *p, *q;
2779 const DWORD *ptr;
2780 DWORD i;
2781
2782 if (!string_buffer_init(&buffer))
2783 {
2784 ERR("Failed to initialize string buffer.\n");
2785 return;
2786 }
2787
2788 fe->shader_read_header(fe_data, &ptr, &shader_version);
2789
2790 TRACE("Parsing %p.\n", ptr);
2791
2792 switch (shader_version.type)
2793 {
2794 case WINED3D_SHADER_TYPE_VERTEX:
2795 type_prefix = "vs";
2796 break;
2797
2798 case WINED3D_SHADER_TYPE_HULL:
2799 type_prefix = "hs";
2800 break;
2801
2802 case WINED3D_SHADER_TYPE_DOMAIN:
2803 type_prefix = "ds";
2804 break;
2805
2806 case WINED3D_SHADER_TYPE_GEOMETRY:
2807 type_prefix = "gs";
2808 break;
2809
2810 case WINED3D_SHADER_TYPE_PIXEL:
2811 type_prefix = "ps";
2812 break;
2813
2814 case WINED3D_SHADER_TYPE_COMPUTE:
2815 type_prefix = "cs";
2816 break;
2817
2818 default:
2819 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2820 type_prefix = "unknown";
2821 break;
2822 }
2823
2824 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2825
2826 while (!fe->shader_is_end(fe_data, &ptr))
2827 {
2828 struct wined3d_shader_instruction ins;
2829
2830 fe->shader_read_instruction(fe_data, &ptr, &ins);
2831 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2832 {
2833 WARN("Skipping unrecognized instruction.\n");
2834 shader_addline(&buffer, "<unrecognized instruction>\n");
2835 continue;
2836 }
2837
2838 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2839 {
2840 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2841 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2842 shader_addline(&buffer, " ");
2843 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2844 }
2845 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2846 {
2847 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2848 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2849 shader_addline(&buffer, ", %s",
2850 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2851 }
2852 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2853 {
2854 shader_addline(&buffer, "%s fb%u",
2855 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2856 }
2857 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2858 {
2859 shader_addline(&buffer, "%s ft%u = {...}",
2860 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2861 }
2862 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2863 {
2864 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2865 shader_dump_global_flags(&buffer, ins.flags);
2866 }
2867 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2868 {
2869 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2870 ins.declaration.max_tessellation_factor);
2871 }
2872 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2873 {
2874 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2875 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2876 {
2877 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2878 ins.declaration.icb->data[4 * i + 0],
2879 ins.declaration.icb->data[4 * i + 1],
2880 ins.declaration.icb->data[4 * i + 2],
2881 ins.declaration.icb->data[4 * i + 3]);
2882 }
2883 shader_addline(&buffer, "}");
2884 }
2885 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2886 {
2887 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2888 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2889 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2890 }
2891 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2892 {
2893 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2894 ins.declaration.indexable_temp.register_idx,
2895 ins.declaration.indexable_temp.register_size,
2896 ins.declaration.indexable_temp.component_count);
2897 }
2898 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2899 {
2900 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2901 shader_dump_interpolation_mode(&buffer, ins.flags);
2902 shader_addline(&buffer, " ");
2903 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2904 }
2905 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2906 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2907 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2908 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2909 {
2910 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2911 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2912 shader_addline(&buffer, ", ");
2913 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2914 }
2915 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2916 {
2917 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2918 shader_dump_interpolation_mode(&buffer, ins.flags);
2919 shader_addline(&buffer, " ");
2920 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2921 shader_addline(&buffer, ", ");
2922 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2923 }
2924 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2925 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2926 {
2927 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2928 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2929 }
2930 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2931 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2932 {
2933 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2934 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2935 }
2936 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2937 {
2938 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2939 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2940 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2941 }
2942 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2943 {
2944 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2945 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2946 }
2947 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2948 {
2949 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2950 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2951 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2952 }
2953 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2954 {
2955 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2956 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2957 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2958 shader_addline(&buffer, ", comparisonMode");
2959 }
2960 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2961 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2962 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2963 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2964 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2965 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2966 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2967 {
2968 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2969 }
2970 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2971 {
2972 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2973 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2974 }
2975 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2976 {
2977 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2978 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2979 }
2980 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2981 {
2982 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2983 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2984 }
2985 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2986 {
2987 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2988 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2989 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2990 }
2991 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2992 {
2993 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2994 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2995 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2996 ins.declaration.tgsm_structured.structure_count);
2997 }
2998 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2999 {
3000 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
3001 ins.declaration.thread_group_size.x,
3002 ins.declaration.thread_group_size.y,
3003 ins.declaration.thread_group_size.z);
3004 }
3005 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
3006 {
3007 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3008 shader_dump_uav_flags(&buffer, ins.flags);
3009 shader_addline(&buffer, " ");
3010 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
3011 }
3012 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
3013 {
3014 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3015 shader_dump_uav_flags(&buffer, ins.flags);
3016 shader_addline(&buffer, " ");
3017 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
3018 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
3019 }
3020 else if (ins.handler_idx == WINED3DSIH_DEF)
3021 {
3022 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
3023 ins.dst[0].reg.idx[0].offset),
3024 *(const float *)&ins.src[0].reg.u.immconst_data[0],
3025 *(const float *)&ins.src[0].reg.u.immconst_data[1],
3026 *(const float *)&ins.src[0].reg.u.immconst_data[2],
3027 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
3028 }
3029 else if (ins.handler_idx == WINED3DSIH_DEFI)
3030 {
3031 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
3032 ins.src[0].reg.u.immconst_data[0],
3033 ins.src[0].reg.u.immconst_data[1],
3034 ins.src[0].reg.u.immconst_data[2],
3035 ins.src[0].reg.u.immconst_data[3]);
3036 }
3037 else if (ins.handler_idx == WINED3DSIH_DEFB)
3038 {
3039 shader_addline(&buffer, "defb b%u = %s",
3040 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
3041 }
3042 else
3043 {
3044 if (ins.predicate)
3045 {
3046 shader_addline(&buffer, "(");
3047 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
3048 shader_addline(&buffer, ") ");
3049 }
3050
3051 /* PixWin marks instructions with the coissue flag with a '+' */
3052 if (ins.coissue)
3053 shader_addline(&buffer, "+");
3054
3055 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3056
3057 if (ins.handler_idx == WINED3DSIH_BREAKP
3058 || ins.handler_idx == WINED3DSIH_CONTINUEP
3059 || ins.handler_idx == WINED3DSIH_IF
3060 || ins.handler_idx == WINED3DSIH_RETP
3061 || ins.handler_idx == WINED3DSIH_TEXKILL)
3062 {
3063 switch (ins.flags)
3064 {
3065 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3066 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3067 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3068 }
3069 }
3070 else if (ins.handler_idx == WINED3DSIH_IFC
3071 || ins.handler_idx == WINED3DSIH_BREAKC)
3072 {
3073 switch (ins.flags)
3074 {
3075 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3076 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3077 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3078 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3079 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3080 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3081 default: shader_addline(&buffer, "_(%u)", ins.flags);
3082 }
3083 }
3084 else if (ins.handler_idx == WINED3DSIH_TEX
3085 && shader_version.major >= 2
3086 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3087 {
3088 shader_addline(&buffer, "p");
3089 }
3090 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3091 {
3092 switch (ins.flags)
3093 {
3094 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3095 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3096 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3097 }
3098 }
3099 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3100 {
3101 switch (ins.flags)
3102 {
3103 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3104 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3105 }
3106 }
3107 else if (ins.handler_idx == WINED3DSIH_SYNC)
3108 {
3109 shader_dump_sync_flags(&buffer, ins.flags);
3110 }
3111 else
3112 {
3113 shader_dump_precise_flags(&buffer, ins.flags);
3114 }
3115
3116 if (wined3d_shader_instruction_has_texel_offset(&ins))
3117 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3118
3119 for (i = 0; i < ins.dst_count; ++i)
3120 {
3121 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
3122 shader_addline(&buffer, !i ? " " : ", ");
3123 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
3124 }
3125
3126 /* Other source tokens */
3127 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
3128 {
3129 shader_addline(&buffer, !i ? " " : ", ");
3130 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
3131 }
3132 }
3133 shader_addline(&buffer, "\n");
3134 }
3135
3136 for (p = buffer.buffer; *p; p = q)
3137 {
3138 if (!(q = strstr(p, "\n")))
3139 q = p + strlen(p);
3140 else
3141 ++q;
3142 TRACE(" %.*s", (int)(q - p), p);
3143 }
3144
3145 string_buffer_free(&buffer);
3146 }
3147
3148 static void shader_cleanup(struct wined3d_shader *shader)
3149 {
3150 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3151 {
3152 heap_free(shader->u.hs.phases.control_point);
3153 heap_free(shader->u.hs.phases.fork);
3154 heap_free(shader->u.hs.phases.join);
3155 }
3156 else if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
3157 {
3158 heap_free((void *)shader->u.gs.so_desc.elements);
3159 }
3160
3161 heap_free(shader->patch_constant_signature.elements);
3162 heap_free(shader->output_signature.elements);
3163 heap_free(shader->input_signature.elements);
3164 shader->device->shader_backend->shader_destroy(shader);
3165 shader_cleanup_reg_maps(&shader->reg_maps);
3166 heap_free(shader->byte_code);
3167 shader_delete_constant_list(&shader->constantsF);
3168 shader_delete_constant_list(&shader->constantsB);
3169 shader_delete_constant_list(&shader->constantsI);
3170 list_remove(&shader->shader_list_entry);
3171
3172 if (shader->frontend && shader->frontend_data)
3173 shader->frontend->shader_free(shader->frontend_data);
3174 }
3175
3176 struct shader_none_priv
3177 {
3178 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3179 const struct wined3d_fragment_pipe_ops *fragment_pipe;
3180 BOOL ffp_proj_control;
3181 };
3182
3183 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3184 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3185 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3186 const struct wined3d_state *state) {}
3187 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3188 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3189 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3190 const struct wined3d_state *state) {}
3191 static void shader_none_destroy(struct wined3d_shader *shader) {}
3192 static void shader_none_free_context_data(struct wined3d_context *context) {}
3193 static void shader_none_init_context_state(struct wined3d_context *context) {}
3194
3195 /* Context activation is done by the caller. */
3196 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3197 const struct wined3d_state *state)
3198 {
3199 struct shader_none_priv *priv = shader_priv;
3200
3201 priv->vertex_pipe->vp_enable(context, !use_vs(state));
3202 priv->fragment_pipe->fp_enable(context, !use_ps(state));
3203 }
3204
3205 /* Context activation is done by the caller. */
3206 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3207 {
3208 struct shader_none_priv *priv = shader_priv;
3209
3210 priv->vertex_pipe->vp_enable(context, FALSE);
3211 priv->fragment_pipe->fp_enable(context, FALSE);
3212
3213 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3214 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3215 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3216 | (1u << WINED3D_SHADER_TYPE_HULL)
3217 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3218 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3219 }
3220
3221 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3222 const struct wined3d_fragment_pipe_ops *fragment_pipe)
3223 {
3224 struct fragment_caps fragment_caps;
3225 void *vertex_priv, *fragment_priv;
3226 struct shader_none_priv *priv;
3227
3228 if (!(priv = heap_alloc(sizeof(*priv))))
3229 return E_OUTOFMEMORY;
3230
3231 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3232 {
3233 ERR("Failed to initialize vertex pipe.\n");
3234 heap_free(priv);
3235 return E_FAIL;
3236 }
3237
3238 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3239 {
3240 ERR("Failed to initialize fragment pipe.\n");
3241 vertex_pipe->vp_free(device, NULL);
3242 heap_free(priv);
3243 return E_FAIL;
3244 }
3245
3246 priv->vertex_pipe = vertex_pipe;
3247 priv->fragment_pipe = fragment_pipe;
3248 fragment_pipe->get_caps(device->adapter, &fragment_caps);
3249 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3250
3251 device->vertex_priv = vertex_priv;
3252 device->fragment_priv = fragment_priv;
3253 device->shader_priv = priv;
3254
3255 return WINED3D_OK;
3256 }
3257
3258 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
3259 {
3260 struct shader_none_priv *priv = device->shader_priv;
3261
3262 priv->fragment_pipe->free_private(device, context);
3263 priv->vertex_pipe->vp_free(device, context);
3264 heap_free(priv);
3265 }
3266
3267 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3268 {
3269 return TRUE;
3270 }
3271
3272 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
3273 {
3274 /* Set the shader caps to 0 for the none shader backend */
3275 memset(caps, 0, sizeof(*caps));
3276 }
3277
3278 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3279 {
3280 /* We "support" every possible fixup, since we don't support any shader
3281 * model, and will never have to actually sample a texture. */
3282 return TRUE;
3283 }
3284
3285 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3286 {
3287 struct shader_none_priv *priv = shader_priv;
3288
3289 return priv->ffp_proj_control;
3290 }
3291
3292 const struct wined3d_shader_backend_ops none_shader_backend =
3293 {
3294 shader_none_handle_instruction,
3295 shader_none_precompile,
3296 shader_none_select,
3297 shader_none_select_compute,
3298 shader_none_disable,
3299 shader_none_update_float_vertex_constants,
3300 shader_none_update_float_pixel_constants,
3301 shader_none_load_constants,
3302 shader_none_destroy,
3303 shader_none_alloc,
3304 shader_none_free,
3305 shader_none_allocate_context_data,
3306 shader_none_free_context_data,
3307 shader_none_init_context_state,
3308 shader_none_get_caps,
3309 shader_none_color_fixup_supported,
3310 shader_none_has_ffp_proj_control,
3311 };
3312
3313 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
3314 {
3315 switch (level)
3316 {
3317 case WINED3D_FEATURE_LEVEL_11_1:
3318 case WINED3D_FEATURE_LEVEL_11:
3319 return 5;
3320 case WINED3D_FEATURE_LEVEL_10_1:
3321 case WINED3D_FEATURE_LEVEL_10:
3322 return 4;
3323 case WINED3D_FEATURE_LEVEL_9_3:
3324 return 3;
3325 case WINED3D_FEATURE_LEVEL_9_2:
3326 case WINED3D_FEATURE_LEVEL_9_1:
3327 return 2;
3328 default:
3329 return 1;
3330 }
3331 }
3332
3333 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
3334 enum wined3d_shader_type type, unsigned int float_const_count)
3335 {
3336 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3337 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3338 const struct wined3d_shader_version *version = &reg_maps->shader_version;
3339 const struct wined3d_shader_frontend *fe;
3340 unsigned int backend_version;
3341 HRESULT hr;
3342
3343 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
3344 shader, device, debug_shader_type(type), float_const_count);
3345
3346 fe = shader->frontend;
3347 if (!(shader->frontend_data = fe->shader_init(shader->function,
3348 shader->functionLength, &shader->output_signature)))
3349 {
3350 FIXME("Failed to initialize frontend.\n");
3351 return WINED3DERR_INVALIDCALL;
3352 }
3353
3354 /* First pass: trace shader. */
3355 if (TRACE_ON(d3d_shader))
3356 shader_trace_init(fe, shader->frontend_data);
3357
3358 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3359 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
3360 return hr;
3361
3362 if (version->type != type)
3363 {
3364 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
3365 return WINED3DERR_INVALIDCALL;
3366 }
3367 if (version->major > shader_max_version_from_feature_level(device->feature_level))
3368 {
3369 WARN("Shader version %u not supported by this device.\n", version->major);
3370 return WINED3DERR_INVALIDCALL;
3371 }
3372 switch (type)
3373 {
3374 case WINED3D_SHADER_TYPE_VERTEX:
3375 backend_version = d3d_info->limits.vs_version;
3376 break;
3377 case WINED3D_SHADER_TYPE_HULL:
3378 backend_version = d3d_info->limits.hs_version;
3379 break;
3380 case WINED3D_SHADER_TYPE_DOMAIN:
3381 backend_version = d3d_info->limits.ds_version;
3382 break;
3383 case WINED3D_SHADER_TYPE_GEOMETRY:
3384 backend_version = d3d_info->limits.gs_version;
3385 break;
3386 case WINED3D_SHADER_TYPE_PIXEL:
3387 backend_version = d3d_info->limits.ps_version;
3388 break;
3389 case WINED3D_SHADER_TYPE_COMPUTE:
3390 backend_version = d3d_info->limits.cs_version;
3391 break;
3392 default:
3393 FIXME("No backend version-checking for this shader type.\n");
3394 backend_version = 0;
3395 }
3396 if (version->major > backend_version)
3397 {
3398 WARN("Shader version %u.%u not supported by the current shader backend.\n",
3399 version->major, version->minor);
3400 return WINED3DERR_INVALIDCALL;
3401 }
3402
3403 shader->load_local_constsF = shader->lconst_inf_or_nan;
3404
3405 return WINED3D_OK;
3406 }
3407
3408 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3409 {
3410 ULONG refcount = InterlockedIncrement(&shader->ref);
3411
3412 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3413
3414 return refcount;
3415 }
3416
3417 static void wined3d_shader_init_object(void *object)
3418 {
3419 struct wined3d_shader *shader = object;
3420 struct wined3d_device *device = shader->device;
3421
3422 list_add_head(&device->shaders, &shader->shader_list_entry);
3423
3424 device->shader_backend->shader_precompile(device->shader_priv, shader);
3425 }
3426
3427 static void wined3d_shader_destroy_object(void *object)
3428 {
3429 shader_cleanup(object);
3430 heap_free(object);
3431 }
3432
3433 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3434 {
3435 ULONG refcount = InterlockedDecrement(&shader->ref);
3436
3437 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3438
3439 if (!refcount)
3440 {
3441 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3442 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3443 }
3444
3445 return refcount;
3446 }
3447
3448 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3449 {
3450 TRACE("shader %p.\n", shader);
3451
3452 return shader->parent;
3453 }
3454
3455 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3456 void *byte_code, UINT *byte_code_size)
3457 {
3458 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3459
3460 if (!byte_code)
3461 {
3462 *byte_code_size = shader->byte_code_size;
3463 return WINED3D_OK;
3464 }
3465
3466 if (*byte_code_size < shader->byte_code_size)
3467 {
3468 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3469 * than the required size we should write the required size and
3470 * return D3DERR_MOREDATA. That's not actually true. */
3471 return WINED3DERR_INVALIDCALL;
3472 }
3473
3474 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
3475
3476 return WINED3D_OK;
3477 }
3478
3479 /* Set local constants for d3d8 shaders. */
3480 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3481 UINT start_idx, const float *src_data, UINT count)
3482 {
3483 UINT end_idx = start_idx + count;
3484 UINT i;
3485
3486 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3487
3488 if (end_idx > shader->limits->constant_float)
3489 {
3490 WARN("end_idx %u > float constants limit %u.\n",
3491 end_idx, shader->limits->constant_float);
3492 end_idx = shader->limits->constant_float;
3493 }
3494
3495 for (i = start_idx; i < end_idx; ++i)
3496 {
3497 struct wined3d_shader_lconst *lconst;
3498 float *value;
3499
3500 if (!(lconst = heap_alloc(sizeof(*lconst))))
3501 return E_OUTOFMEMORY;
3502
3503 lconst->idx = i;
3504 value = (float *)lconst->value;
3505 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3506 list_add_head(&shader->constantsF, &lconst->entry);
3507
3508 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3509 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3510 {
3511 shader->lconst_inf_or_nan = TRUE;
3512 }
3513 }
3514
3515 return WINED3D_OK;
3516 }
3517
3518 static void init_interpolation_compile_args(DWORD *interpolation_args,
3519 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
3520 {
3521 if (!d3d_info->shader_output_interpolation || !pixel_shader
3522 || pixel_shader->reg_maps.shader_version.major < 4)
3523 {
3524 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3525 return;
3526 }
3527
3528 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3529 sizeof(pixel_shader->u.ps.interpolation_mode));
3530 }
3531
3532 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3533 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_context *context)
3534 {
3535 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3536 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3537 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3538 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3539
3540 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3541 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3542 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3543 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3544 args->point_size = state->primitive_type == WINED3D_PT_POINTLIST;
3545 args->per_vertex_point_size = shader->reg_maps.point_size;
3546 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
3547 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3548 if (shader->reg_maps.shader_version.major >= 4)
3549 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3550 : geometry_shader ? geometry_shader->limits->packed_input
3551 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3552 else
3553 args->next_shader_input_count = 0;
3554 args->swizzle_map = swizzle_map;
3555 if (d3d_info->emulated_flatshading)
3556 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3557 else
3558 args->flatshading = 0;
3559
3560 init_interpolation_compile_args(args->interpolation_mode,
3561 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
3562 }
3563
3564 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3565 {
3566 if (usage_idx1 != usage_idx2)
3567 return FALSE;
3568 if (usage1 == usage2)
3569 return TRUE;
3570 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3571 return TRUE;
3572 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3573 return TRUE;
3574
3575 return FALSE;
3576 }
3577
3578 BOOL vshader_get_input(const struct wined3d_shader *shader,
3579 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3580 {
3581 WORD map = shader->reg_maps.input_registers;
3582 unsigned int i;
3583
3584 for (i = 0; map; map >>= 1, ++i)
3585 {
3586 if (!(map & 1)) continue;
3587
3588 if (match_usage(shader->u.vs.attributes[i].usage,
3589 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3590 {
3591 *regnum = i;
3592 return TRUE;
3593 }
3594 }
3595 return FALSE;
3596 }
3597
3598 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3599 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3600 {
3601 HRESULT hr;
3602
3603 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
3604
3605 if (!desc->byte_code)
3606 return WINED3DERR_INVALIDCALL;
3607
3608 shader->ref = 1;
3609 shader->device = device;
3610 shader->parent = parent;
3611 shader->parent_ops = parent_ops;
3612
3613 list_init(&shader->linked_programs);
3614 list_init(&shader->constantsF);
3615 list_init(&shader->constantsB);
3616 list_init(&shader->constantsI);
3617 shader->lconst_inf_or_nan = FALSE;
3618 list_init(&shader->reg_maps.indexable_temps);
3619 list_init(&shader->shader_list_entry);
3620
3621 if (desc->byte_code_size == ~(size_t)0)
3622 {
3623 struct wined3d_shader_version shader_version;
3624 const struct wined3d_shader_frontend *fe;
3625 struct wined3d_shader_instruction ins;
3626 const DWORD *ptr;
3627 void *fe_data;
3628
3629 if (!(shader->frontend = shader_select_frontend(WINED3D_SHADER_BYTE_CODE_FORMAT_SM1)))
3630 {
3631 FIXME("Unable to find frontend for shader.\n");
3632 hr = WINED3DERR_INVALIDCALL;
3633 goto fail;
3634 }
3635
3636 fe = shader->frontend;
3637 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
3638 {
3639 WARN("Failed to initialise frontend data.\n");
3640 hr = WINED3DERR_INVALIDCALL;
3641 goto fail;
3642 }
3643
3644 fe->shader_read_header(fe_data, &ptr, &shader_version);
3645 while (!fe->shader_is_end(fe_data, &ptr))
3646 fe->shader_read_instruction(fe_data, &ptr, &ins);
3647
3648 fe->shader_free(fe_data);
3649
3650 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3651
3652 if (!(shader->byte_code = heap_alloc(shader->byte_code_size)))
3653 {
3654 hr = E_OUTOFMEMORY;
3655 goto fail;
3656 }
3657 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
3658
3659 shader->function = shader->byte_code;
3660 shader->functionLength = shader->byte_code_size;
3661 }
3662 else
3663 {
3664 enum wined3d_shader_byte_code_format format;
3665 unsigned int max_version;
3666
3667 if (!(shader->byte_code = heap_alloc(desc->byte_code_size)))
3668 {
3669 hr = E_OUTOFMEMORY;
3670 goto fail;
3671 }
3672 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
3673 shader->byte_code_size = desc->byte_code_size;
3674
3675 max_version = shader_max_version_from_feature_level(device->feature_level);
3676 if (FAILED(hr = shader_extract_from_dxbc(shader, max_version, &format)))
3677 goto fail;
3678
3679 if (!(shader->frontend = shader_select_frontend(format)))
3680 {
3681 FIXME("Unable to find frontend for shader.\n");
3682 hr = WINED3DERR_INVALIDCALL;
3683 goto fail;
3684 }
3685 }
3686
3687 return WINED3D_OK;
3688
3689 fail:
3690 shader_cleanup(shader);
3691 return hr;
3692 }
3693
3694 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3695 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3696 {
3697 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3698 unsigned int i;
3699 HRESULT hr;
3700
3701 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3702 return hr;
3703
3704 if (FAILED(hr = shader_set_function(shader, device,
3705 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
3706 {
3707 shader_cleanup(shader);
3708 return hr;
3709 }
3710
3711 for (i = 0; i < shader->input_signature.element_count; ++i)
3712 {
3713 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3714
3715 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3716 continue;
3717
3718 shader->u.vs.attributes[input->register_idx].usage =
3719 shader_usage_from_semantic_name(input->semantic_name);
3720 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3721 }
3722
3723 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3724 shader->load_local_constsF = TRUE;
3725
3726 return WINED3D_OK;
3727 }
3728
3729 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
3730 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
3731 {
3732 struct wined3d_shader_signature_element *e = s->elements;
3733 unsigned int i;
3734
3735 for (i = 0; i < s->element_count; ++i)
3736 {
3737 if (e[i].stream_idx == stream_idx
3738 && !_strnicmp(e[i].semantic_name, semantic_name, -1)
3739 && e[i].semantic_idx == semantic_idx)
3740 return &e[i];
3741 }
3742
3743 return NULL;
3744 }
3745
3746 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
3747 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
3748 {
3749 const struct wined3d_shader_signature_element *output;
3750 unsigned int idx;
3751
3752 if (!(output = shader_find_signature_element(&shader->output_signature,
3753 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
3754 return FALSE;
3755
3756 for (idx = 0; idx < 4; ++idx)
3757 {
3758 if (output->mask & (1u << idx))
3759 break;
3760 }
3761 idx += so_element->component_idx;
3762
3763 *register_idx = output->register_idx;
3764 *component_idx = idx;
3765 return TRUE;
3766 }
3767
3768 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
3769 const struct wined3d_stream_output_desc *so_desc)
3770 {
3771 const struct wined3d_shader_frontend *fe = shader->frontend;
3772 const struct wined3d_shader_signature_element *output;
3773 unsigned int i, component_idx, register_idx, mask;
3774 struct wined3d_stream_output_element *elements;
3775 struct wined3d_shader_version shader_version;
3776 const DWORD *ptr;
3777 void *fe_data;
3778 HRESULT hr;
3779
3780 if (!so_desc)
3781 return WINED3D_OK;
3782
3783 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
3784 {
3785 WARN("Failed to initialise frontend data.\n");
3786 return WINED3DERR_INVALIDCALL;
3787 }
3788 fe->shader_read_header(fe_data, &ptr, &shader_version);
3789 fe->shader_free(fe_data);
3790
3791 switch (shader_version.type)
3792 {
3793 case WINED3D_SHADER_TYPE_VERTEX:
3794 case WINED3D_SHADER_TYPE_DOMAIN:
3795 shader->function = NULL;
3796 shader->functionLength = 0;
3797 break;
3798 case WINED3D_SHADER_TYPE_GEOMETRY:
3799 break;
3800 default:
3801 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
3802 return E_INVALIDARG;
3803 }
3804
3805 if (!shader->function)
3806 {
3807 shader->reg_maps.shader_version = shader_version;
3808 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
3809 shader_set_limits(shader);
3810 if (FAILED(hr = shader_scan_output_signature(shader)))
3811 return hr;
3812 }
3813
3814 if (!(elements = heap_calloc(so_desc->element_count, sizeof(*elements))))
3815 return E_OUTOFMEMORY;
3816
3817 shader->u.gs.so_desc = *so_desc;
3818 shader->u.gs.so_desc.elements = elements;
3819 memcpy(elements, so_desc->elements, so_desc->element_count * sizeof(*elements));
3820
3821 for (i = 0; i < so_desc->element_count; ++i)
3822 {
3823 struct wined3d_stream_output_element *e = &elements[i];
3824
3825 if (!e->semantic_name)
3826 continue;
3827 if (!(output = shader_find_signature_element(&shader->output_signature,
3828 e->stream_idx, e->semantic_name, e->semantic_idx))
3829 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
3830 {
3831 WARN("Failed to find output signature element for stream output entry.\n");
3832 return E_INVALIDARG;
3833 }
3834
3835 e->semantic_name = output->semantic_name;
3836
3837 mask = ((1u << e->component_count) - 1) << component_idx;
3838 if ((output->mask & 0xff & mask) != mask)
3839 {
3840 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
3841 component_idx, e->component_count, mask, output->mask & 0xff);
3842 return E_INVALIDARG;
3843 }
3844 }
3845
3846 return WINED3D_OK;
3847 }
3848
3849 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3850 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3851 void *parent, const struct wined3d_parent_ops *parent_ops)
3852 {
3853 HRESULT hr;
3854
3855 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3856 return hr;
3857
3858 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
3859 goto fail;
3860
3861 if (shader->function
3862 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
3863 goto fail;
3864
3865 return WINED3D_OK;
3866
3867 fail:
3868 shader_cleanup(shader);
3869 return hr;
3870 }
3871
3872 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3873 struct ds_compile_args *args, const struct wined3d_context *context)
3874 {
3875 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3876 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3877 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3878
3879 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3880 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3881
3882 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3883 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3884 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3885
3886 args->render_offscreen = context->render_offscreen;
3887
3888 init_interpolation_compile_args(args->interpolation_mode,
3889 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
3890
3891 args->padding = 0;
3892 }
3893
3894 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3895 struct gs_compile_args *args, const struct wined3d_context *context)
3896 {
3897 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3898
3899 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3900
3901 if (!(args->primitive_type = shader->u.gs.input_type))
3902 args->primitive_type = state->primitive_type;
3903
3904 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
3905 }
3906
3907 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3908 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3909 {
3910 const struct wined3d_gl_info *gl_info = &context->device->adapter->gl_info;
3911 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3912 struct wined3d_texture *texture;
3913 unsigned int i;
3914
3915 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3916 if (!d3d_info->srgb_write_control && needs_srgb_write(d3d_info, state, &state->fb))
3917 {
3918 static unsigned int warned = 0;
3919
3920 args->srgb_correction = 1;
3921 if (state->blend_state && state->blend_state->desc.rt[0].enable && !warned++)
3922 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3923 "support, expect rendering artifacts.\n");
3924 }
3925
3926 if (shader->reg_maps.shader_version.major == 1
3927 && shader->reg_maps.shader_version.minor <= 3)
3928 {
3929 for (i = 0; i < shader->limits->sampler; ++i)
3930 {
3931 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3932
3933 if (flags & WINED3D_TTFF_PROJECTED)
3934 {
3935 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3936
3937 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3938 {
3939 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3940 unsigned int j;
3941 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3942 DWORD max_valid = WINED3D_TTFF_COUNT4;
3943
3944 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3945 {
3946 struct wined3d_vertex_declaration_element *element =
3947 &state->vertex_declaration->elements[j];
3948
3949 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3950 && element->usage_idx == index)
3951 {
3952 max_valid = element->format->component_count;
3953 break;
3954 }
3955 }
3956 if (!tex_transform || tex_transform > max_valid)
3957 {
3958 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3959 tex_transform, max_valid);
3960 tex_transform = max_valid;
3961 }
3962 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3963 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3964 && tex_transform > WINED3D_TTFF_COUNT2)
3965 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3966 && tex_transform > WINED3D_TTFF_COUNT3))
3967 tex_transform |= WINED3D_PSARGS_PROJECTED;
3968 else
3969 {
3970 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3971 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3972 i, tex_transform, resource_type);
3973 }
3974 }
3975 else
3976 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3977
3978 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3979 }
3980 }
3981 }
3982 if (shader->reg_maps.shader_version.major == 1
3983 && shader->reg_maps.shader_version.minor <= 4)
3984 {
3985 for (i = 0; i < shader->limits->sampler; ++i)
3986 {
3987 if (!shader->reg_maps.resource_info[i].type)
3988 continue;
3989
3990 /* Treat unbound textures as 2D. The dummy texture will provide
3991 * the proper sample value. The tex_types bitmap defaults to
3992 * 2D because of the memset. */
3993 if (!(texture = state->textures[i]))
3994 continue;
3995
3996 switch (wined3d_texture_gl(texture)->target)
3997 {
3998 /* RECT textures are distinguished from 2D textures via np2_fixup */
3999 default:
4000 break;
4001
4002 case GL_TEXTURE_3D:
4003 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4004 break;
4005
4006 case GL_TEXTURE_CUBE_MAP_ARB:
4007 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4008 break;
4009 }
4010 }
4011 }
4012 else if (shader->reg_maps.shader_version.major <= 3)
4013 {
4014 for (i = 0; i < shader->limits->sampler; ++i)
4015 {
4016 enum wined3d_shader_resource_type resource_type;
4017 enum wined3d_shader_tex_types tex_type;
4018
4019 if (!(resource_type = shader->reg_maps.resource_info[i].type))
4020 continue;
4021
4022 switch (resource_type)
4023 {
4024 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
4025 tex_type = WINED3D_SHADER_TEX_3D;
4026 break;
4027 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
4028 tex_type = WINED3D_SHADER_TEX_CUBE;
4029 break;
4030 default:
4031 tex_type = WINED3D_SHADER_TEX_2D;
4032 break;
4033 }
4034
4035 if ((texture = state->textures[i]))
4036 {
4037 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
4038 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4039 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
4040 tex_type = WINED3D_SHADER_TEX_2D;
4041 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
4042 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
4043 tex_type = WINED3D_SHADER_TEX_3D;
4044 }
4045 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4046 }
4047 }
4048
4049 if (shader->reg_maps.shader_version.major >= 4)
4050 {
4051 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
4052 args->shadow = 0;
4053 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4054 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4055 args->np2_fixup = 0;
4056 }
4057 else
4058 {
4059 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4060 {
4061 if (!shader->reg_maps.resource_info[i].type)
4062 continue;
4063
4064 texture = state->textures[i];
4065 if (!texture)
4066 {
4067 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4068 continue;
4069 }
4070 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
4071 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4072 else
4073 args->color_fixup[i] = texture->resource.format->color_fixup;
4074
4075 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
4076 args->shadow |= 1u << i;
4077
4078 /* Flag samplers that need NP2 texcoord fixup. */
4079 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
4080 args->np2_fixup |= (1u << i);
4081 }
4082 }
4083
4084 if (shader->reg_maps.shader_version.major >= 3)
4085 {
4086 if (position_transformed)
4087 args->vp_mode = WINED3D_VP_MODE_NONE;
4088 else if (use_vs(state))
4089 args->vp_mode = WINED3D_VP_MODE_SHADER;
4090 else
4091 args->vp_mode = WINED3D_VP_MODE_FF;
4092 args->fog = WINED3D_FFP_PS_FOG_OFF;
4093 }
4094 else
4095 {
4096 args->vp_mode = WINED3D_VP_MODE_SHADER;
4097 if (state->render_states[WINED3D_RS_FOGENABLE])
4098 {
4099 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
4100 {
4101 case WINED3D_FOG_NONE:
4102 if (position_transformed || use_vs(state))
4103 {
4104 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
4105 break;
4106 }
4107
4108 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
4109 {
4110 case WINED3D_FOG_NONE: /* Fall through. */
4111 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4112 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4113 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4114 }
4115 break;
4116
4117 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4118 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4119 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4120 }
4121 }
4122 else
4123 {
4124 args->fog = WINED3D_FFP_PS_FOG_OFF;
4125 }
4126 }
4127
4128 if (!d3d_info->full_ffp_varyings)
4129 {
4130 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
4131
4132 args->texcoords_initialized = 0;
4133 for (i = 0; i < WINED3D_MAX_TEXTURES; ++i)
4134 {
4135 if (vs)
4136 {
4137 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
4138 args->texcoords_initialized |= 1u << i;
4139 }
4140 else
4141 {
4142 const struct wined3d_stream_info *si = &context->stream_info;
4143 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
4144
4145 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
4146 & WINED3D_FFP_TCI_MASK
4147 || (coord_idx < WINED3D_MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
4148 args->texcoords_initialized |= 1u << i;
4149 }
4150 }
4151 }
4152 else
4153 {
4154 args->texcoords_initialized = (1u << WINED3D_MAX_TEXTURES) - 1;
4155 }
4156
4157 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
4158 && state->primitive_type == WINED3D_PT_POINTLIST;
4159
4160 if (d3d_info->ffp_alpha_test)
4161 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
4162 else
4163 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4164 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4165 : WINED3D_CMP_ALWAYS) - 1;
4166
4167 if (d3d_info->emulated_flatshading)
4168 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4169
4170 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
4171 ? context->render_offscreen : 0;
4172
4173 for (i = 0; i < ARRAY_SIZE(state->fb.render_targets); ++i)
4174 {
4175 struct wined3d_rendertarget_view *rtv = state->fb.render_targets[i];
4176 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
4177 args->rt_alpha_swizzle |= 1u << i;
4178 }
4179
4180 args->dual_source_blend = state->blend_state && state->blend_state->dual_source;
4181 }
4182
4183 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4184 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4185 {
4186 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4187 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4188 HRESULT hr;
4189
4190 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
4191 return hr;
4192
4193 if (FAILED(hr = shader_set_function(shader, device,
4194 WINED3D_SHADER_TYPE_PIXEL, device->adapter->d3d_info.limits.ps_uniform_count)))
4195 {
4196 shader_cleanup(shader);
4197 return hr;
4198 }
4199
4200 for (i = 0; i < MAX_REG_INPUT; ++i)
4201 {
4202 if (shader->u.ps.input_reg_used & (1u << i))
4203 {
4204 ++num_regs_used;
4205 highest_reg_used = i;
4206 }
4207 }
4208
4209 /* Don't do any register mapping magic if it is not needed, or if we can't
4210 * achieve anything anyway */
4211 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4212 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4213 || shader->reg_maps.shader_version.major >= 4)
4214 {
4215 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4216 {
4217 /* This happens with relative addressing. The input mapper function
4218 * warns about this if the higher registers are declared too, so
4219 * don't write a FIXME here */
4220 WARN("More varying registers used than supported\n");
4221 }
4222
4223 for (i = 0; i < MAX_REG_INPUT; ++i)
4224 {
4225 shader->u.ps.input_reg_map[i] = i;
4226 }
4227
4228 shader->u.ps.declared_in_count = highest_reg_used + 1;
4229 }
4230 else
4231 {
4232 shader->u.ps.declared_in_count = 0;
4233 for (i = 0; i < MAX_REG_INPUT; ++i)
4234 {
4235 if (shader->u.ps.input_reg_used & (1u << i))
4236 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4237 else shader->u.ps.input_reg_map[i] = ~0U;
4238 }
4239 }
4240
4241 return WINED3D_OK;
4242 }
4243
4244 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
4245 unsigned int resource_idx, DWORD tex_types)
4246 {
4247 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
4248 {
4249 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
4250 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
4251 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
4252 };
4253
4254 unsigned int idx;
4255
4256 if (reg_maps->shader_version.major > 3)
4257 return reg_maps->resource_info[resource_idx].type;
4258
4259 if (!reg_maps->resource_info[resource_idx].type)
4260 return 0;
4261
4262 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
4263 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
4264 return shader_resource_type_from_shader_tex_types[idx];
4265 }
4266
4267 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4268 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4269 {
4270 struct wined3d_shader *object;
4271 HRESULT hr;
4272
4273 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4274 device, desc, parent, parent_ops, shader);
4275
4276 if (!(object = heap_alloc_zero(sizeof(*object))))
4277 return E_OUTOFMEMORY;
4278
4279 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4280 {
4281 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4282 heap_free(object);
4283 return hr;
4284 }
4285
4286 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
4287 {
4288 shader_cleanup(object);
4289 heap_free(object);
4290 return hr;
4291 }
4292
4293 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4294
4295 TRACE("Created compute shader %p.\n", object);
4296 *shader = object;
4297
4298 return WINED3D_OK;
4299 }
4300
4301 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4302 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4303 {
4304 struct wined3d_shader *object;
4305 HRESULT hr;
4306
4307 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4308 device, desc, parent, parent_ops, shader);
4309
4310 if (!(object = heap_alloc_zero(sizeof(*object))))
4311 return E_OUTOFMEMORY;
4312
4313 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4314 {
4315 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4316 heap_free(object);
4317 return hr;
4318 }
4319
4320 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
4321 {
4322 shader_cleanup(object);
4323 heap_free(object);
4324 return hr;
4325 }
4326
4327 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4328
4329 TRACE("Created domain shader %p.\n", object);
4330 *shader = object;
4331
4332 return WINED3D_OK;
4333 }
4334
4335 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4336 const struct wined3d_stream_output_desc *so_desc, void *parent,
4337 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4338 {
4339 struct wined3d_shader *object;
4340 HRESULT hr;
4341
4342 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4343 device, desc, so_desc, parent, parent_ops, shader);
4344
4345 if (!(object = heap_alloc_zero(sizeof(*object))))
4346 return E_OUTOFMEMORY;
4347
4348 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4349 {
4350 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4351 heap_free(object);
4352 return hr;
4353 }
4354
4355 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4356
4357 TRACE("Created geometry shader %p.\n", object);
4358 *shader = object;
4359
4360 return WINED3D_OK;
4361 }
4362
4363 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4364 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4365 {
4366 struct wined3d_shader *object;
4367 HRESULT hr;
4368
4369 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4370 device, desc, parent, parent_ops, shader);
4371
4372 if (!(object = heap_alloc_zero(sizeof(*object))))
4373 return E_OUTOFMEMORY;
4374
4375 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4376 {
4377 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4378 heap_free(object);
4379 return hr;
4380 }
4381
4382 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
4383 {
4384 shader_cleanup(object);
4385 heap_free(object);
4386 return hr;
4387 }
4388
4389 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4390
4391 TRACE("Created hull shader %p.\n", object);
4392 *shader = object;
4393
4394 return WINED3D_OK;
4395 }
4396
4397 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4398 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4399 {
4400 struct wined3d_shader *object;
4401 HRESULT hr;
4402
4403 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4404 device, desc, parent, parent_ops, shader);
4405
4406 if (!(object = heap_alloc_zero(sizeof(*object))))
4407 return E_OUTOFMEMORY;
4408
4409 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4410 {
4411 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4412 heap_free(object);
4413 return hr;
4414 }
4415
4416 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4417
4418 TRACE("Created pixel shader %p.\n", object);
4419 *shader = object;
4420
4421 return WINED3D_OK;
4422 }
4423
4424 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4425 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4426 {
4427 struct wined3d_shader *object;
4428 HRESULT hr;
4429
4430 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4431 device, desc, parent, parent_ops, shader);
4432
4433 if (!(object = heap_alloc_zero(sizeof(*object))))
4434 return E_OUTOFMEMORY;
4435
4436 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4437 {
4438 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4439 heap_free(object);
4440 return hr;
4441 }
4442
4443 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4444
4445 TRACE("Created vertex shader %p.\n", object);
4446 *shader = object;
4447
4448 return WINED3D_OK;
4449 }
Windows API implementation