?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Move the blend enable state to wined3d_blend_state.
[wine.git] / dlls / wined3d / shader.c
blob13e55cff8aa2a5d88f5e33b6458e3cd56c7e898c
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 break;
1127
1128 case WINED3DSPR_UAV:
1129 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1130 {
1131 ERR("Invalid UAV resource index %u.\n", reg_idx);
1132 break;
1133 }
1134 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1135 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1136 if (ins.flags)
1137 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1138 break;
1139
1140 default:
1141 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1142 break;
1143 }
1144 }
1145 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1146 {
1147 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1148 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1149 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1150 else
1151 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1152 }
1153 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1154 {
1155 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1156 {
1157 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1158 shader->u.ps.force_early_depth_stencil = TRUE;
1159 else
1160 FIXME("Invalid instruction %#x for shader type %#x.\n",
1161 ins.handler_idx, shader_version.type);
1162 }
1163 else
1164 {
1165 WARN("Ignoring global flags %#x.\n", ins.flags);
1166 }
1167 }
1168 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1169 {
1170 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1171 shader->u.gs.instance_count = ins.declaration.count;
1172 else
1173 FIXME("Invalid instruction %#x for shader type %#x.\n",
1174 ins.handler_idx, shader_version.type);
1175 }
1176 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1177 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1178 {
1179 if (phase)
1180 phase->instance_count = ins.declaration.count;
1181 else
1182 FIXME("Instruction %s outside of shader phase.\n",
1183 debug_d3dshaderinstructionhandler(ins.handler_idx));
1184 }
1185 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1186 {
1187 if (reg_maps->icb)
1188 FIXME("Multiple immediate constant buffers.\n");
1189 reg_maps->icb = ins.declaration.icb;
1190 }
1191 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1192 {
1193 if (phase)
1194 {
1195 FIXME("Indexable temporary registers not supported.\n");
1196 }
1197 else
1198 {
1199 struct wined3d_shader_indexable_temp *reg;
1200
1201 if (!(reg = heap_alloc(sizeof(*reg))))
1202 return E_OUTOFMEMORY;
1203
1204 *reg = ins.declaration.indexable_temp;
1205 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1206 }
1207 }
1208 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1209 {
1210 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1211 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1212 else
1213 FIXME("Invalid instruction %#x for shader type %#x.\n",
1214 ins.handler_idx, shader_version.type);
1215 }
1216 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1217 {
1218 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1219 if (reg_idx >= MAX_REG_INPUT)
1220 {
1221 ERR("Invalid register index %u.\n", reg_idx);
1222 break;
1223 }
1224 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1225 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1226 else
1227 FIXME("Invalid instruction %#x for shader type %#x.\n",
1228 ins.handler_idx, shader_version.type);
1229 }
1230 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1231 {
1232 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1233 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1234 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1235 {
1236 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1237 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1238 else
1239 FIXME("Invalid instruction %#x for shader type %#x.\n",
1240 ins.handler_idx, shader_version.type);
1241 }
1242 }
1243 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1244 {
1245 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1246 shader->u.hs.output_vertex_count = ins.declaration.count;
1247 else
1248 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1249 }
1250 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1251 {
1252 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1253 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1254 else
1255 FIXME("Invalid instruction %#x for shader type %#x.\n",
1256 ins.handler_idx, shader_version.type);
1257 }
1258 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1259 {
1260 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1261 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1262 {
1263 ERR("Invalid resource index %u.\n", reg_idx);
1264 break;
1265 }
1266 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1267 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1268 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1269 }
1270 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1271 {
1272 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1273 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1274 {
1275 ERR("Invalid resource index %u.\n", reg_idx);
1276 break;
1277 }
1278 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1279 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1280 reg_maps->resource_info[reg_idx].flags = 0;
1281 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1282 }
1283 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1284 {
1285 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1286 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1287 }
1288 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1289 {
1290 if (phase)
1291 phase->temporary_count = ins.declaration.count;
1292 else
1293 reg_maps->temporary_count = ins.declaration.count;
1294 }
1295 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1296 {
1297 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1298 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1299 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1300 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1301 }
1302 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1303 {
1304 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1305 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1306 else
1307 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1308 }
1309 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1310 {
1311 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1312 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1313 else
1314 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1315 }
1316 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1317 {
1318 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1319 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1320 return hr;
1321 }
1322 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1323 {
1324 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1325 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1326 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1327 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1328 return hr;
1329 }
1330 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1331 {
1332 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1333 {
1334 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1335 }
1336 else
1337 {
1338 FIXME("Invalid instruction %#x for shader type %#x.\n",
1339 ins.handler_idx, shader_version.type);
1340 }
1341 }
1342 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1343 {
1344 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1345 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1346 {
1347 ERR("Invalid UAV resource index %u.\n", reg_idx);
1348 break;
1349 }
1350 if (ins.flags)
1351 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1352 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1353 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1354 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1355 }
1356 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1357 {
1358 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1359 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1360 {
1361 ERR("Invalid UAV resource index %u.\n", reg_idx);
1362 break;
1363 }
1364 if (ins.flags)
1365 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1366 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1367 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1368 reg_maps->uav_resource_info[reg_idx].flags = 0;
1369 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1370 }
1371 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1372 {
1373 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1374 shader->u.gs.vertices_out = ins.declaration.count;
1375 else
1376 FIXME("Invalid instruction %#x for shader type %#x.\n",
1377 ins.handler_idx, shader_version.type);
1378 }
1379 else if (ins.handler_idx == WINED3DSIH_DEF)
1380 {
1381 struct wined3d_shader_lconst *lconst;
1382 float *value;
1383
1384 if (!(lconst = heap_alloc(sizeof(*lconst))))
1385 return E_OUTOFMEMORY;
1386
1387 lconst->idx = ins.dst[0].reg.idx[0].offset;
1388 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1389 value = (float *)lconst->value;
1390
1391 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1392 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1393 {
1394 if (value[0] < -1.0f) value[0] = -1.0f;
1395 else if (value[0] > 1.0f) value[0] = 1.0f;
1396 if (value[1] < -1.0f) value[1] = -1.0f;
1397 else if (value[1] > 1.0f) value[1] = 1.0f;
1398 if (value[2] < -1.0f) value[2] = -1.0f;
1399 else if (value[2] > 1.0f) value[2] = 1.0f;
1400 if (value[3] < -1.0f) value[3] = -1.0f;
1401 else if (value[3] > 1.0f) value[3] = 1.0f;
1402 }
1403
1404 list_add_head(&shader->constantsF, &lconst->entry);
1405
1406 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1407 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1408 {
1409 shader->lconst_inf_or_nan = TRUE;
1410 }
1411 }
1412 else if (ins.handler_idx == WINED3DSIH_DEFI)
1413 {
1414 struct wined3d_shader_lconst *lconst;
1415
1416 if (!(lconst = heap_alloc(sizeof(*lconst))))
1417 return E_OUTOFMEMORY;
1418
1419 lconst->idx = ins.dst[0].reg.idx[0].offset;
1420 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1421
1422 list_add_head(&shader->constantsI, &lconst->entry);
1423 reg_maps->local_int_consts |= (1u << lconst->idx);
1424 }
1425 else if (ins.handler_idx == WINED3DSIH_DEFB)
1426 {
1427 struct wined3d_shader_lconst *lconst;
1428
1429 if (!(lconst = heap_alloc(sizeof(*lconst))))
1430 return E_OUTOFMEMORY;
1431
1432 lconst->idx = ins.dst[0].reg.idx[0].offset;
1433 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1434
1435 list_add_head(&shader->constantsB, &lconst->entry);
1436 reg_maps->local_bool_consts |= (1u << lconst->idx);
1437 }
1438 /* Handle shader phases. */
1439 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1440 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1441 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1442 {
1443 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1444 return hr;
1445 }
1446 /* For subroutine prototypes. */
1447 else if (ins.handler_idx == WINED3DSIH_LABEL)
1448 {
1449 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1450 }
1451 /* Set texture, address, temporary registers. */
1452 else
1453 {
1454 BOOL color0_mov = FALSE;
1455 unsigned int i;
1456
1457 /* This will loop over all the registers and try to
1458 * make a bitmask of the ones we're interested in.
1459 *
1460 * Relative addressing tokens are ignored, but that's
1461 * okay, since we'll catch any address registers when
1462 * they are initialized (required by spec). */
1463 for (i = 0; i < ins.dst_count; ++i)
1464 {
1465 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1466 shader_version.type, constf_size))
1467 return WINED3DERR_INVALIDCALL;
1468
1469 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1470 {
1471 UINT idx = ins.dst[i].reg.idx[0].offset;
1472
1473 switch (ins.dst[i].reg.type)
1474 {
1475 case WINED3DSPR_RASTOUT:
1476 if (shader_version.major >= 3)
1477 break;
1478 switch (idx)
1479 {
1480 case 0: /* oPos */
1481 reg_maps->output_registers |= 1u << 10;
1482 shader_signature_from_usage(&output_signature_elements[10],
1483 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1484 break;
1485
1486 case 1: /* oFog */
1487 reg_maps->output_registers |= 1u << 11;
1488 shader_signature_from_usage(&output_signature_elements[11],
1489 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1490 break;
1491
1492 case 2: /* oPts */
1493 reg_maps->output_registers |= 1u << 11;
1494 shader_signature_from_usage(&output_signature_elements[11],
1495 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1496 break;
1497 }
1498 break;
1499
1500 case WINED3DSPR_ATTROUT:
1501 if (shader_version.major >= 3)
1502 break;
1503 if (idx < 2)
1504 {
1505 idx += 8;
1506 if (reg_maps->output_registers & (1u << idx))
1507 {
1508 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1509 }
1510 else
1511 {
1512 reg_maps->output_registers |= 1u << idx;
1513 shader_signature_from_usage(&output_signature_elements[idx],
1514 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1515 }
1516 }
1517 break;
1518
1519 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1520 if (shader_version.major >= 3)
1521 {
1522 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1523 {
1524 WARN("Invalid output register index %u.\n", idx);
1525 break;
1526 }
1527 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1528 break;
1529 }
1530 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1531 {
1532 WARN("Invalid texcoord index %u.\n", idx);
1533 break;
1534 }
1535 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1536 if (reg_maps->output_registers & (1u << idx))
1537 {
1538 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1539 }
1540 else
1541 {
1542 reg_maps->output_registers |= 1u << idx;
1543 shader_signature_from_usage(&output_signature_elements[idx],
1544 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1545 }
1546 break;
1547
1548 default:
1549 break;
1550 }
1551 }
1552
1553 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1554 {
1555 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1556 {
1557 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1558 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1559 * the mov and perform the sRGB write correction from the source register.
1560 *
1561 * However, if the mov is only partial, we can't do this, and if the write
1562 * comes from an instruction other than MOV it is hard to do as well. If
1563 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1564 shader->u.ps.color0_mov = FALSE;
1565 if (ins.handler_idx == WINED3DSIH_MOV
1566 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1567 {
1568 /* Used later when the source register is read. */
1569 color0_mov = TRUE;
1570 }
1571 }
1572 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1573 * end
1574 */
1575 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1576 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1577 {
1578 shader->u.ps.color0_mov = FALSE;
1579 }
1580 }
1581
1582 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1583 if (shader_version.major == 1
1584 && (ins.handler_idx == WINED3DSIH_TEX
1585 || ins.handler_idx == WINED3DSIH_TEXBEM
1586 || ins.handler_idx == WINED3DSIH_TEXBEML
1587 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1588 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1589 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1590 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1591 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1592 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1593 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1594 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1595 {
1596 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1597
1598 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1599 {
1600 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1601 continue;
1602 }
1603
1604 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1605 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1606 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1607 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1608
1609 /* texbem is only valid with < 1.4 pixel shaders */
1610 if (ins.handler_idx == WINED3DSIH_TEXBEM
1611 || ins.handler_idx == WINED3DSIH_TEXBEML)
1612 {
1613 reg_maps->bumpmat |= 1u << reg_idx;
1614 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1615 {
1616 reg_maps->luminanceparams |= 1u << reg_idx;
1617 }
1618 }
1619 }
1620 else if (ins.handler_idx == WINED3DSIH_BEM)
1621 {
1622 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1623 }
1624 }
1625
1626 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1627 {
1628 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1629 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1630 {
1631 ERR("Invalid UAV index %u.\n", reg_idx);
1632 break;
1633 }
1634 reg_maps->uav_counter_mask |= (1u << reg_idx);
1635 }
1636 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1637 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1638 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1639 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1640 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1641 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1642 {
1643 unsigned int reg_idx;
1644 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1645 reg_idx = ins.src[1].reg.idx[0].offset;
1646 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1647 reg_idx = ins.src[2].reg.idx[0].offset;
1648 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1649 reg_idx = ins.dst[0].reg.idx[0].offset;
1650 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1651 reg_idx = ins.src[0].reg.idx[0].offset;
1652 else
1653 reg_idx = ins.dst[1].reg.idx[0].offset;
1654 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1655 {
1656 ERR("Invalid UAV index %u.\n", reg_idx);
1657 break;
1658 }
1659 reg_maps->uav_read_mask |= (1u << reg_idx);
1660 }
1661 else if (ins.handler_idx == WINED3DSIH_NRM)
1662 {
1663 reg_maps->usesnrm = 1;
1664 }
1665 else if (ins.handler_idx == WINED3DSIH_DSY
1666 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1667 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1668 {
1669 reg_maps->usesdsy = 1;
1670 }
1671 else if (ins.handler_idx == WINED3DSIH_DSX
1672 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1673 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1674 {
1675 reg_maps->usesdsx = 1;
1676 }
1677 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1678 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1679 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1680 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1681 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1682 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1683 else if (ins.handler_idx == WINED3DSIH_LOOP
1684 || ins.handler_idx == WINED3DSIH_REP)
1685 {
1686 ++cur_loop_depth;
1687 if (cur_loop_depth > max_loop_depth)
1688 max_loop_depth = cur_loop_depth;
1689 }
1690 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1691 || ins.handler_idx == WINED3DSIH_ENDREP)
1692 {
1693 --cur_loop_depth;
1694 }
1695 else if (ins.handler_idx == WINED3DSIH_GATHER4
1696 || ins.handler_idx == WINED3DSIH_GATHER4_C
1697 || ins.handler_idx == WINED3DSIH_SAMPLE
1698 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1699 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1700 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1701 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1702 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1703 {
1704 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1705 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1706 }
1707 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1708 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1709 {
1710 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1711 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1712 }
1713 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1714 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1715 {
1716 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1717 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1718 }
1719 else if (ins.handler_idx == WINED3DSIH_LD
1720 || ins.handler_idx == WINED3DSIH_LD2DMS
1721 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1722 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1723 {
1724 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1725 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1726 }
1727 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1728 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1729 {
1730 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1731 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1732 }
1733
1734 if (ins.predicate)
1735 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1736 shader_version.type, constf_size))
1737 return WINED3DERR_INVALIDCALL;
1738
1739 for (i = 0; i < ins.src_count; ++i)
1740 {
1741 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1742 struct wined3d_shader_register reg = ins.src[i].reg;
1743
1744 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1745 shader_version.type, constf_size))
1746 return WINED3DERR_INVALIDCALL;
1747 while (count)
1748 {
1749 ++reg.idx[0].offset;
1750 if (!shader_record_register_usage(shader, reg_maps, &reg,
1751 shader_version.type, constf_size))
1752 return WINED3DERR_INVALIDCALL;
1753 --count;
1754 }
1755
1756 if (color0_mov)
1757 {
1758 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1759 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1760 {
1761 shader->u.ps.color0_mov = TRUE;
1762 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1763 }
1764 }
1765 }
1766 }
1767
1768 prev_ins = current_ins;
1769 }
1770 reg_maps->loop_depth = max_loop_depth;
1771
1772 if (phase)
1773 {
1774 phase->end = prev_ins;
1775 phase = NULL;
1776 }
1777
1778 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1779 * R0 is written to the render target. */
1780 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1781 reg_maps->rt_mask |= (1u << 0);
1782
1783 if (input_signature->elements)
1784 {
1785 for (i = 0; i < input_signature->element_count; ++i)
1786 {
1787 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1788 {
1789 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1790 {
1791 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1792 return WINED3DERR_INVALIDCALL;
1793 }
1794 }
1795 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1796 {
1797 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1798 reg_maps->vpos = 1;
1799 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1800 reg_maps->usesfacing = 1;
1801 }
1802 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1803 }
1804 }
1805 else if (!input_signature->elements && reg_maps->input_registers)
1806 {
1807 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1808 struct wined3d_shader_signature_element *e;
1809 unsigned int i;
1810
1811 if (!(input_signature->elements = heap_calloc(count, sizeof(*input_signature->elements))))
1812 return E_OUTOFMEMORY;
1813 input_signature->element_count = count;
1814
1815 e = input_signature->elements;
1816 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1817 {
1818 if (!(reg_maps->input_registers & (1u << i)))
1819 continue;
1820 input_signature_elements[i].register_idx = i;
1821 *e++ = input_signature_elements[i];
1822 }
1823 }
1824
1825 if (output_signature->elements)
1826 {
1827 if (FAILED(hr = shader_scan_output_signature(shader)))
1828 return hr;
1829 }
1830 else if (reg_maps->output_registers)
1831 {
1832 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1833 struct wined3d_shader_signature_element *e;
1834
1835 if (!(output_signature->elements = heap_calloc(count, sizeof(*output_signature->elements))))
1836 return E_OUTOFMEMORY;
1837 output_signature->element_count = count;
1838
1839 e = output_signature->elements;
1840 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1841 {
1842 if (!(reg_maps->output_registers & (1u << i)))
1843 continue;
1844 *e++ = output_signature_elements[i];
1845 }
1846 }
1847
1848 return WINED3D_OK;
1849 }
1850
1851 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1852 {
1853 struct wined3d_shader_indexable_temp *reg, *reg_next;
1854
1855 heap_free(reg_maps->constf);
1856 heap_free(reg_maps->sampler_map.entries);
1857
1858 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1859 heap_free(reg);
1860 list_init(&reg_maps->indexable_temps);
1861
1862 heap_free(reg_maps->tgsm);
1863 }
1864
1865 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1866 {
1867 DWORD map = 1u << max;
1868 map |= map - 1;
1869 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1870
1871 return wined3d_log2i(map);
1872 }
1873
1874 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1875 {
1876 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1877 {
1878 shader_addline(buffer, "refactoringAllowed");
1879 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1880 if (global_flags)
1881 shader_addline(buffer, " | ");
1882 }
1883
1884 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1885 {
1886 shader_addline(buffer, "forceEarlyDepthStencil");
1887 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1888 if (global_flags)
1889 shader_addline(buffer, " | ");
1890 }
1891
1892 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1893 {
1894 shader_addline(buffer, "enableRawAndStructuredBuffers");
1895 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1896 }
1897
1898 if (global_flags)
1899 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1900 }
1901
1902 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1903 {
1904 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1905 {
1906 shader_addline(buffer, "_g");
1907 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1908 }
1909 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1910 {
1911 shader_addline(buffer, "_t");
1912 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1913 }
1914
1915 if (sync_flags)
1916 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1917 }
1918
1919 static void shader_dump_precise_flags(struct wined3d_string_buffer *buffer, DWORD flags)
1920 {
1921 if (!(flags & WINED3DSI_PRECISE_XYZW))
1922 return;
1923
1924 shader_addline(buffer, " [precise");
1925 if (flags != WINED3DSI_PRECISE_XYZW)
1926 {
1927 shader_addline(buffer, "(%s%s%s%s)",
1928 flags & WINED3DSI_PRECISE_X ? "x" : "",
1929 flags & WINED3DSI_PRECISE_Y ? "y" : "",
1930 flags & WINED3DSI_PRECISE_Z ? "z" : "",
1931 flags & WINED3DSI_PRECISE_W ? "w" : "");
1932 }
1933 shader_addline(buffer, "]");
1934 }
1935
1936 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1937 {
1938 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1939 {
1940 shader_addline(buffer, "_glc");
1941 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1942 }
1943 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1944 {
1945 shader_addline(buffer, "_opc");
1946 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1947 }
1948
1949 if (uav_flags)
1950 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1951 }
1952
1953 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1954 enum wined3d_tessellator_domain domain)
1955 {
1956 switch (domain)
1957 {
1958 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1959 shader_addline(buffer, "line");
1960 break;
1961 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1962 shader_addline(buffer, "triangle");
1963 break;
1964 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1965 shader_addline(buffer, "quad");
1966 break;
1967 default:
1968 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1969 break;
1970 }
1971 }
1972
1973 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1974 enum wined3d_tessellator_output_primitive output_primitive)
1975 {
1976 switch (output_primitive)
1977 {
1978 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1979 shader_addline(buffer, "point");
1980 break;
1981 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1982 shader_addline(buffer, "line");
1983 break;
1984 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1985 shader_addline(buffer, "triangle_cw");
1986 break;
1987 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1988 shader_addline(buffer, "triangle_ccw");
1989 break;
1990 default:
1991 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1992 break;
1993 }
1994 }
1995
1996 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1997 enum wined3d_tessellator_partitioning partitioning)
1998 {
1999 switch (partitioning)
2000 {
2001 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
2002 shader_addline(buffer, "integer");
2003 break;
2004 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
2005 shader_addline(buffer, "pow2");
2006 break;
2007 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
2008 shader_addline(buffer, "fractional_odd");
2009 break;
2010 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
2011 shader_addline(buffer, "fractional_even");
2012 break;
2013 default:
2014 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
2015 break;
2016 }
2017 }
2018
2019 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
2020 enum wined3d_shader_input_sysval_semantic semantic)
2021 {
2022 unsigned int i;
2023
2024 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
2025 {
2026 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
2027 {
2028 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
2029 return;
2030 }
2031 }
2032
2033 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
2034 }
2035
2036 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
2037 const struct wined3d_shader_semantic *semantic, unsigned int flags,
2038 const struct wined3d_shader_version *shader_version)
2039 {
2040 shader_addline(buffer, "dcl");
2041
2042 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
2043 {
2044 switch (semantic->resource_type)
2045 {
2046 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2047 shader_addline(buffer, "_2d");
2048 break;
2049
2050 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2051 shader_addline(buffer, "_3d");
2052 break;
2053
2054 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2055 shader_addline(buffer, "_cube");
2056 break;
2057
2058 default:
2059 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2060 break;
2061 }
2062 }
2063 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2064 {
2065 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2066 shader_addline(buffer, "_resource_");
2067 else
2068 shader_addline(buffer, "_uav_");
2069 switch (semantic->resource_type)
2070 {
2071 case WINED3D_SHADER_RESOURCE_BUFFER:
2072 shader_addline(buffer, "buffer");
2073 break;
2074
2075 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
2076 shader_addline(buffer, "texture1d");
2077 break;
2078
2079 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2080 shader_addline(buffer, "texture2d");
2081 break;
2082
2083 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
2084 shader_addline(buffer, "texture2dms");
2085 break;
2086
2087 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2088 shader_addline(buffer, "texture3d");
2089 break;
2090
2091 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2092 shader_addline(buffer, "texturecube");
2093 break;
2094
2095 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
2096 shader_addline(buffer, "texture1darray");
2097 break;
2098
2099 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
2100 shader_addline(buffer, "texture2darray");
2101 break;
2102
2103 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
2104 shader_addline(buffer, "texture2dmsarray");
2105 break;
2106
2107 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
2108 shader_addline(buffer, "texturecubearray");
2109 break;
2110
2111 default:
2112 shader_addline(buffer, "unknown");
2113 break;
2114 }
2115 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2116 shader_dump_uav_flags(buffer, flags);
2117 switch (semantic->resource_data_type)
2118 {
2119 case WINED3D_DATA_FLOAT:
2120 shader_addline(buffer, " (float)");
2121 break;
2122
2123 case WINED3D_DATA_INT:
2124 shader_addline(buffer, " (int)");
2125 break;
2126
2127 case WINED3D_DATA_UINT:
2128 shader_addline(buffer, " (uint)");
2129 break;
2130
2131 case WINED3D_DATA_UNORM:
2132 shader_addline(buffer, " (unorm)");
2133 break;
2134
2135 case WINED3D_DATA_SNORM:
2136 shader_addline(buffer, " (snorm)");
2137 break;
2138
2139 default:
2140 shader_addline(buffer, " (unknown)");
2141 break;
2142 }
2143 }
2144 else
2145 {
2146 /* Pixel shaders 3.0 don't have usage semantics. */
2147 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2148 return;
2149 else
2150 shader_addline(buffer, "_");
2151
2152 switch (semantic->usage)
2153 {
2154 case WINED3D_DECL_USAGE_POSITION:
2155 shader_addline(buffer, "position%u", semantic->usage_idx);
2156 break;
2157
2158 case WINED3D_DECL_USAGE_BLEND_INDICES:
2159 shader_addline(buffer, "blend");
2160 break;
2161
2162 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2163 shader_addline(buffer, "weight");
2164 break;
2165
2166 case WINED3D_DECL_USAGE_NORMAL:
2167 shader_addline(buffer, "normal%u", semantic->usage_idx);
2168 break;
2169
2170 case WINED3D_DECL_USAGE_PSIZE:
2171 shader_addline(buffer, "psize");
2172 break;
2173
2174 case WINED3D_DECL_USAGE_COLOR:
2175 if (!semantic->usage_idx)
2176 shader_addline(buffer, "color");
2177 else
2178 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2179 break;
2180
2181 case WINED3D_DECL_USAGE_TEXCOORD:
2182 shader_addline(buffer, "texture%u", semantic->usage_idx);
2183 break;
2184
2185 case WINED3D_DECL_USAGE_TANGENT:
2186 shader_addline(buffer, "tangent");
2187 break;
2188
2189 case WINED3D_DECL_USAGE_BINORMAL:
2190 shader_addline(buffer, "binormal");
2191 break;
2192
2193 case WINED3D_DECL_USAGE_TESS_FACTOR:
2194 shader_addline(buffer, "tessfactor");
2195 break;
2196
2197 case WINED3D_DECL_USAGE_POSITIONT:
2198 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2199 break;
2200
2201 case WINED3D_DECL_USAGE_FOG:
2202 shader_addline(buffer, "fog");
2203 break;
2204
2205 case WINED3D_DECL_USAGE_DEPTH:
2206 shader_addline(buffer, "depth");
2207 break;
2208
2209 case WINED3D_DECL_USAGE_SAMPLE:
2210 shader_addline(buffer, "sample");
2211 break;
2212
2213 default:
2214 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2215 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2216 }
2217 }
2218 }
2219
2220 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2221 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2222 {
2223 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2224 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2225 UINT offset = reg->idx[0].offset;
2226
2227 switch (reg->type)
2228 {
2229 case WINED3DSPR_TEMP:
2230 shader_addline(buffer, "r");
2231 break;
2232
2233 case WINED3DSPR_INPUT:
2234 shader_addline(buffer, "v");
2235 break;
2236
2237 case WINED3DSPR_CONST:
2238 case WINED3DSPR_CONST2:
2239 case WINED3DSPR_CONST3:
2240 case WINED3DSPR_CONST4:
2241 shader_addline(buffer, "c");
2242 offset = shader_get_float_offset(reg->type, offset);
2243 break;
2244
2245 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2246 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2247 break;
2248
2249 case WINED3DSPR_RASTOUT:
2250 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2251 break;
2252
2253 case WINED3DSPR_COLOROUT:
2254 shader_addline(buffer, "oC");
2255 break;
2256
2257 case WINED3DSPR_DEPTHOUT:
2258 shader_addline(buffer, "oDepth");
2259 break;
2260
2261 case WINED3DSPR_DEPTHOUTGE:
2262 shader_addline(buffer, "oDepthGE");
2263 break;
2264
2265 case WINED3DSPR_DEPTHOUTLE:
2266 shader_addline(buffer, "oDepthLE");
2267 break;
2268
2269 case WINED3DSPR_ATTROUT:
2270 shader_addline(buffer, "oD");
2271 break;
2272
2273 case WINED3DSPR_TEXCRDOUT:
2274 /* Vertex shaders >= 3.0 use general purpose output registers
2275 * (WINED3DSPR_OUTPUT), which can include an address token. */
2276 if (shader_version->major >= 3)
2277 shader_addline(buffer, "o");
2278 else
2279 shader_addline(buffer, "oT");
2280 break;
2281
2282 case WINED3DSPR_CONSTINT:
2283 shader_addline(buffer, "i");
2284 break;
2285
2286 case WINED3DSPR_CONSTBOOL:
2287 shader_addline(buffer, "b");
2288 break;
2289
2290 case WINED3DSPR_LABEL:
2291 shader_addline(buffer, "l");
2292 break;
2293
2294 case WINED3DSPR_LOOP:
2295 shader_addline(buffer, "aL");
2296 break;
2297
2298 case WINED3DSPR_SAMPLER:
2299 shader_addline(buffer, "s");
2300 break;
2301
2302 case WINED3DSPR_MISCTYPE:
2303 if (offset > 1)
2304 {
2305 FIXME("Unhandled misctype register %u.\n", offset);
2306 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2307 }
2308 else
2309 {
2310 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2311 }
2312 break;
2313
2314 case WINED3DSPR_PREDICATE:
2315 shader_addline(buffer, "p");
2316 break;
2317
2318 case WINED3DSPR_IMMCONST:
2319 shader_addline(buffer, "l");
2320 break;
2321
2322 case WINED3DSPR_CONSTBUFFER:
2323 shader_addline(buffer, "cb");
2324 break;
2325
2326 case WINED3DSPR_IMMCONSTBUFFER:
2327 shader_addline(buffer, "icb");
2328 break;
2329
2330 case WINED3DSPR_PRIMID:
2331 shader_addline(buffer, "primID");
2332 break;
2333
2334 case WINED3DSPR_NULL:
2335 shader_addline(buffer, "null");
2336 break;
2337
2338 case WINED3DSPR_RASTERIZER:
2339 shader_addline(buffer, "rasterizer");
2340 break;
2341
2342 case WINED3DSPR_RESOURCE:
2343 shader_addline(buffer, "t");
2344 break;
2345
2346 case WINED3DSPR_UAV:
2347 shader_addline(buffer, "u");
2348 break;
2349
2350 case WINED3DSPR_OUTPOINTID:
2351 shader_addline(buffer, "vOutputControlPointID");
2352 break;
2353
2354 case WINED3DSPR_FORKINSTID:
2355 shader_addline(buffer, "vForkInstanceId");
2356 break;
2357
2358 case WINED3DSPR_JOININSTID:
2359 shader_addline(buffer, "vJoinInstanceId");
2360 break;
2361
2362 case WINED3DSPR_INCONTROLPOINT:
2363 shader_addline(buffer, "vicp");
2364 break;
2365
2366 case WINED3DSPR_OUTCONTROLPOINT:
2367 shader_addline(buffer, "vocp");
2368 break;
2369
2370 case WINED3DSPR_PATCHCONST:
2371 shader_addline(buffer, "vpc");
2372 break;
2373
2374 case WINED3DSPR_TESSCOORD:
2375 shader_addline(buffer, "vDomainLocation");
2376 break;
2377
2378 case WINED3DSPR_GROUPSHAREDMEM:
2379 shader_addline(buffer, "g");
2380 break;
2381
2382 case WINED3DSPR_THREADID:
2383 shader_addline(buffer, "vThreadID");
2384 break;
2385
2386 case WINED3DSPR_THREADGROUPID:
2387 shader_addline(buffer, "vThreadGroupID");
2388 break;
2389
2390 case WINED3DSPR_LOCALTHREADID:
2391 shader_addline(buffer, "vThreadIDInGroup");
2392 break;
2393
2394 case WINED3DSPR_LOCALTHREADINDEX:
2395 shader_addline(buffer, "vThreadIDInGroupFlattened");
2396 break;
2397
2398 case WINED3DSPR_IDXTEMP:
2399 shader_addline(buffer, "x");
2400 break;
2401
2402 case WINED3DSPR_STREAM:
2403 shader_addline(buffer, "m");
2404 break;
2405
2406 case WINED3DSPR_FUNCTIONBODY:
2407 shader_addline(buffer, "fb");
2408 break;
2409
2410 case WINED3DSPR_FUNCTIONPOINTER:
2411 shader_addline(buffer, "fp");
2412 break;
2413
2414 case WINED3DSPR_COVERAGE:
2415 shader_addline(buffer, "vCoverage");
2416 break;
2417
2418 case WINED3DSPR_SAMPLEMASK:
2419 shader_addline(buffer, "oMask");
2420 break;
2421
2422 case WINED3DSPR_GSINSTID:
2423 shader_addline(buffer, "vGSInstanceID");
2424 break;
2425
2426 default:
2427 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2428 break;
2429 }
2430
2431 if (reg->type == WINED3DSPR_IMMCONST)
2432 {
2433 shader_addline(buffer, "(");
2434 switch (reg->immconst_type)
2435 {
2436 case WINED3D_IMMCONST_SCALAR:
2437 switch (reg->data_type)
2438 {
2439 case WINED3D_DATA_FLOAT:
2440 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2441 break;
2442 case WINED3D_DATA_INT:
2443 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2444 break;
2445 case WINED3D_DATA_RESOURCE:
2446 case WINED3D_DATA_SAMPLER:
2447 case WINED3D_DATA_UINT:
2448 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2449 break;
2450 default:
2451 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2452 break;
2453 }
2454 break;
2455
2456 case WINED3D_IMMCONST_VEC4:
2457 switch (reg->data_type)
2458 {
2459 case WINED3D_DATA_FLOAT:
2460 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2461 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2462 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2463 break;
2464 case WINED3D_DATA_INT:
2465 shader_addline(buffer, "%d, %d, %d, %d",
2466 reg->u.immconst_data[0], reg->u.immconst_data[1],
2467 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2468 break;
2469 case WINED3D_DATA_RESOURCE:
2470 case WINED3D_DATA_SAMPLER:
2471 case WINED3D_DATA_UINT:
2472 shader_addline(buffer, "%u, %u, %u, %u",
2473 reg->u.immconst_data[0], reg->u.immconst_data[1],
2474 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2475 break;
2476 default:
2477 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2478 break;
2479 }
2480 break;
2481
2482 default:
2483 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2484 break;
2485 }
2486 shader_addline(buffer, ")");
2487 }
2488 else if (reg->type != WINED3DSPR_RASTOUT
2489 && reg->type != WINED3DSPR_MISCTYPE
2490 && reg->type != WINED3DSPR_NULL)
2491 {
2492 if (offset != ~0u)
2493 {
2494 shader_addline(buffer, "[");
2495 if (reg->idx[0].rel_addr)
2496 {
2497 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2498 shader_addline(buffer, " + ");
2499 }
2500 shader_addline(buffer, "%u]", offset);
2501
2502 if (reg->idx[1].offset != ~0u)
2503 {
2504 shader_addline(buffer, "[");
2505 if (reg->idx[1].rel_addr)
2506 {
2507 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2508 shader_addline(buffer, " + ");
2509 }
2510 shader_addline(buffer, "%u]", reg->idx[1].offset);
2511 }
2512 }
2513
2514 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2515 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2516 }
2517 }
2518
2519 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2520 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2521 {
2522 DWORD write_mask = param->write_mask;
2523
2524 shader_dump_register(buffer, &param->reg, shader_version);
2525
2526 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2527 {
2528 static const char write_mask_chars[] = "xyzw";
2529
2530 shader_addline(buffer, ".");
2531 if (write_mask & WINED3DSP_WRITEMASK_0)
2532 shader_addline(buffer, "%c", write_mask_chars[0]);
2533 if (write_mask & WINED3DSP_WRITEMASK_1)
2534 shader_addline(buffer, "%c", write_mask_chars[1]);
2535 if (write_mask & WINED3DSP_WRITEMASK_2)
2536 shader_addline(buffer, "%c", write_mask_chars[2]);
2537 if (write_mask & WINED3DSP_WRITEMASK_3)
2538 shader_addline(buffer, "%c", write_mask_chars[3]);
2539 }
2540 }
2541
2542 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2543 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2544 {
2545 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2546 DWORD swizzle = param->swizzle;
2547
2548 if (src_modifier == WINED3DSPSM_NEG
2549 || src_modifier == WINED3DSPSM_BIASNEG
2550 || src_modifier == WINED3DSPSM_SIGNNEG
2551 || src_modifier == WINED3DSPSM_X2NEG
2552 || src_modifier == WINED3DSPSM_ABSNEG)
2553 shader_addline(buffer, "-");
2554 else if (src_modifier == WINED3DSPSM_COMP)
2555 shader_addline(buffer, "1-");
2556 else if (src_modifier == WINED3DSPSM_NOT)
2557 shader_addline(buffer, "!");
2558
2559 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2560 shader_addline(buffer, "abs(");
2561
2562 shader_dump_register(buffer, &param->reg, shader_version);
2563
2564 switch (src_modifier)
2565 {
2566 case WINED3DSPSM_NONE: break;
2567 case WINED3DSPSM_NEG: break;
2568 case WINED3DSPSM_NOT: break;
2569 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2570 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2571 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2572 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2573 case WINED3DSPSM_COMP: break;
2574 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2575 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2576 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2577 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2578 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2579 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2580 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2581 }
2582
2583 if (swizzle != WINED3DSP_NOSWIZZLE)
2584 {
2585 static const char swizzle_chars[] = "xyzw";
2586 DWORD swizzle_x = swizzle & 0x03;
2587 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2588 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2589 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2590
2591 if (swizzle_x == swizzle_y
2592 && swizzle_x == swizzle_z
2593 && swizzle_x == swizzle_w)
2594 {
2595 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2596 }
2597 else
2598 {
2599 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2600 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2601 }
2602 }
2603 }
2604
2605 /* Shared code in order to generate the bulk of the shader string. */
2606 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2607 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2608 const DWORD *start, const DWORD *end)
2609 {
2610 struct wined3d_device *device = shader->device;
2611 const struct wined3d_shader_frontend *fe = shader->frontend;
2612 void *fe_data = shader->frontend_data;
2613 struct wined3d_shader_version shader_version;
2614 struct wined3d_shader_parser_state state;
2615 struct wined3d_shader_instruction ins;
2616 struct wined3d_shader_tex_mx tex_mx;
2617 struct wined3d_shader_context ctx;
2618 const DWORD *ptr;
2619
2620 /* Initialize current parsing state. */
2621 tex_mx.current_row = 0;
2622 state.current_loop_depth = 0;
2623 state.current_loop_reg = 0;
2624 state.in_subroutine = FALSE;
2625
2626 ctx.shader = shader;
2627 ctx.reg_maps = reg_maps;
2628 ctx.buffer = buffer;
2629 ctx.tex_mx = &tex_mx;
2630 ctx.state = &state;
2631 ctx.backend_data = backend_ctx;
2632 ins.ctx = &ctx;
2633
2634 fe->shader_read_header(fe_data, &ptr, &shader_version);
2635 if (start)
2636 ptr = start;
2637
2638 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2639 {
2640 /* Read opcode. */
2641 fe->shader_read_instruction(fe_data, &ptr, &ins);
2642
2643 /* Unknown opcode and its parameters. */
2644 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2645 {
2646 WARN("Encountered unrecognised or invalid instruction.\n");
2647 return WINED3DERR_INVALIDCALL;
2648 }
2649
2650 if (ins.predicate)
2651 FIXME("Predicates not implemented.\n");
2652
2653 /* Call appropriate function for output target */
2654 device->shader_backend->shader_handle_instruction(&ins);
2655 }
2656
2657 return WINED3D_OK;
2658 }
2659
2660 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2661 const struct wined3d_shader_dst_param *dst)
2662 {
2663 DWORD mmask = dst->modifiers;
2664
2665 switch (dst->shift)
2666 {
2667 case 0: break;
2668 case 13: shader_addline(buffer, "_d8"); break;
2669 case 14: shader_addline(buffer, "_d4"); break;
2670 case 15: shader_addline(buffer, "_d2"); break;
2671 case 1: shader_addline(buffer, "_x2"); break;
2672 case 2: shader_addline(buffer, "_x4"); break;
2673 case 3: shader_addline(buffer, "_x8"); break;
2674 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2675 }
2676
2677 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2678 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2679 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2680
2681 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2682 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2683 }
2684
2685 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2686 const struct wined3d_shader_primitive_type *primitive_type)
2687 {
2688 switch (primitive_type->type)
2689 {
2690 case WINED3D_PT_UNDEFINED:
2691 shader_addline(buffer, "undefined");
2692 break;
2693 case WINED3D_PT_POINTLIST:
2694 shader_addline(buffer, "pointlist");
2695 break;
2696 case WINED3D_PT_LINELIST:
2697 shader_addline(buffer, "linelist");
2698 break;
2699 case WINED3D_PT_LINESTRIP:
2700 shader_addline(buffer, "linestrip");
2701 break;
2702 case WINED3D_PT_TRIANGLELIST:
2703 shader_addline(buffer, "trianglelist");
2704 break;
2705 case WINED3D_PT_TRIANGLESTRIP:
2706 shader_addline(buffer, "trianglestrip");
2707 break;
2708 case WINED3D_PT_TRIANGLEFAN:
2709 shader_addline(buffer, "trianglefan");
2710 break;
2711 case WINED3D_PT_LINELIST_ADJ:
2712 shader_addline(buffer, "linelist_adj");
2713 break;
2714 case WINED3D_PT_LINESTRIP_ADJ:
2715 shader_addline(buffer, "linestrip_adj");
2716 break;
2717 case WINED3D_PT_TRIANGLELIST_ADJ:
2718 shader_addline(buffer, "trianglelist_adj");
2719 break;
2720 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2721 shader_addline(buffer, "trianglestrip_adj");
2722 break;
2723 case WINED3D_PT_PATCH:
2724 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2725 break;
2726 default:
2727 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2728 break;
2729 }
2730 }
2731
2732 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2733 enum wined3d_shader_interpolation_mode interpolation_mode)
2734 {
2735 switch (interpolation_mode)
2736 {
2737 case WINED3DSIM_CONSTANT:
2738 shader_addline(buffer, "constant");
2739 break;
2740 case WINED3DSIM_LINEAR:
2741 shader_addline(buffer, "linear");
2742 break;
2743 case WINED3DSIM_LINEAR_CENTROID:
2744 shader_addline(buffer, "linear centroid");
2745 break;
2746 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2747 shader_addline(buffer, "linear noperspective");
2748 break;
2749 case WINED3DSIM_LINEAR_SAMPLE:
2750 shader_addline(buffer, "linear sample");
2751 break;
2752 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2753 shader_addline(buffer, "linear noperspective centroid");
2754 break;
2755 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2756 shader_addline(buffer, "linear noperspective sample");
2757 break;
2758 default:
2759 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2760 break;
2761 }
2762 }
2763
2764 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2765 {
2766 struct wined3d_shader_version shader_version;
2767 struct wined3d_string_buffer buffer;
2768 const char *type_prefix;
2769 const char *p, *q;
2770 const DWORD *ptr;
2771 DWORD i;
2772
2773 if (!string_buffer_init(&buffer))
2774 {
2775 ERR("Failed to initialize string buffer.\n");
2776 return;
2777 }
2778
2779 fe->shader_read_header(fe_data, &ptr, &shader_version);
2780
2781 TRACE("Parsing %p.\n", ptr);
2782
2783 switch (shader_version.type)
2784 {
2785 case WINED3D_SHADER_TYPE_VERTEX:
2786 type_prefix = "vs";
2787 break;
2788
2789 case WINED3D_SHADER_TYPE_HULL:
2790 type_prefix = "hs";
2791 break;
2792
2793 case WINED3D_SHADER_TYPE_DOMAIN:
2794 type_prefix = "ds";
2795 break;
2796
2797 case WINED3D_SHADER_TYPE_GEOMETRY:
2798 type_prefix = "gs";
2799 break;
2800
2801 case WINED3D_SHADER_TYPE_PIXEL:
2802 type_prefix = "ps";
2803 break;
2804
2805 case WINED3D_SHADER_TYPE_COMPUTE:
2806 type_prefix = "cs";
2807 break;
2808
2809 default:
2810 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2811 type_prefix = "unknown";
2812 break;
2813 }
2814
2815 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2816
2817 while (!fe->shader_is_end(fe_data, &ptr))
2818 {
2819 struct wined3d_shader_instruction ins;
2820
2821 fe->shader_read_instruction(fe_data, &ptr, &ins);
2822 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2823 {
2824 WARN("Skipping unrecognized instruction.\n");
2825 shader_addline(&buffer, "<unrecognized instruction>\n");
2826 continue;
2827 }
2828
2829 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2830 {
2831 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2832 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2833 shader_addline(&buffer, " ");
2834 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2835 }
2836 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2837 {
2838 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2839 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2840 shader_addline(&buffer, ", %s",
2841 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2842 }
2843 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2844 {
2845 shader_addline(&buffer, "%s fb%u",
2846 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2847 }
2848 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2849 {
2850 shader_addline(&buffer, "%s ft%u = {...}",
2851 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2852 }
2853 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2854 {
2855 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2856 shader_dump_global_flags(&buffer, ins.flags);
2857 }
2858 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2859 {
2860 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2861 ins.declaration.max_tessellation_factor);
2862 }
2863 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2864 {
2865 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2866 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2867 {
2868 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2869 ins.declaration.icb->data[4 * i + 0],
2870 ins.declaration.icb->data[4 * i + 1],
2871 ins.declaration.icb->data[4 * i + 2],
2872 ins.declaration.icb->data[4 * i + 3]);
2873 }
2874 shader_addline(&buffer, "}");
2875 }
2876 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2877 {
2878 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2879 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2880 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2881 }
2882 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2883 {
2884 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2885 ins.declaration.indexable_temp.register_idx,
2886 ins.declaration.indexable_temp.register_size,
2887 ins.declaration.indexable_temp.component_count);
2888 }
2889 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2890 {
2891 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2892 shader_dump_interpolation_mode(&buffer, ins.flags);
2893 shader_addline(&buffer, " ");
2894 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2895 }
2896 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2897 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2898 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2899 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2900 {
2901 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2902 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2903 shader_addline(&buffer, ", ");
2904 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2905 }
2906 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2907 {
2908 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2909 shader_dump_interpolation_mode(&buffer, ins.flags);
2910 shader_addline(&buffer, " ");
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
2916 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2917 {
2918 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2919 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2920 }
2921 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2922 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2923 {
2924 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2925 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2926 }
2927 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2928 {
2929 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2930 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2931 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2932 }
2933 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2934 {
2935 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2936 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2937 }
2938 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2939 {
2940 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2941 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2942 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2943 }
2944 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2945 {
2946 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2947 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2948 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2949 shader_addline(&buffer, ", comparisonMode");
2950 }
2951 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2952 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2953 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2954 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2955 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2956 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2957 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2958 {
2959 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2960 }
2961 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2962 {
2963 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2964 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2965 }
2966 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2967 {
2968 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2969 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2970 }
2971 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2972 {
2973 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2974 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2975 }
2976 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2977 {
2978 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2979 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2980 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2981 }
2982 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2983 {
2984 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2985 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2986 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2987 ins.declaration.tgsm_structured.structure_count);
2988 }
2989 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2990 {
2991 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2992 ins.declaration.thread_group_size.x,
2993 ins.declaration.thread_group_size.y,
2994 ins.declaration.thread_group_size.z);
2995 }
2996 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2997 {
2998 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2999 shader_dump_uav_flags(&buffer, ins.flags);
3000 shader_addline(&buffer, " ");
3001 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
3002 }
3003 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
3004 {
3005 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3006 shader_dump_uav_flags(&buffer, ins.flags);
3007 shader_addline(&buffer, " ");
3008 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
3009 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
3010 }
3011 else if (ins.handler_idx == WINED3DSIH_DEF)
3012 {
3013 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
3014 ins.dst[0].reg.idx[0].offset),
3015 *(const float *)&ins.src[0].reg.u.immconst_data[0],
3016 *(const float *)&ins.src[0].reg.u.immconst_data[1],
3017 *(const float *)&ins.src[0].reg.u.immconst_data[2],
3018 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
3019 }
3020 else if (ins.handler_idx == WINED3DSIH_DEFI)
3021 {
3022 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
3023 ins.src[0].reg.u.immconst_data[0],
3024 ins.src[0].reg.u.immconst_data[1],
3025 ins.src[0].reg.u.immconst_data[2],
3026 ins.src[0].reg.u.immconst_data[3]);
3027 }
3028 else if (ins.handler_idx == WINED3DSIH_DEFB)
3029 {
3030 shader_addline(&buffer, "defb b%u = %s",
3031 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
3032 }
3033 else
3034 {
3035 if (ins.predicate)
3036 {
3037 shader_addline(&buffer, "(");
3038 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
3039 shader_addline(&buffer, ") ");
3040 }
3041
3042 /* PixWin marks instructions with the coissue flag with a '+' */
3043 if (ins.coissue)
3044 shader_addline(&buffer, "+");
3045
3046 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3047
3048 if (ins.handler_idx == WINED3DSIH_BREAKP
3049 || ins.handler_idx == WINED3DSIH_CONTINUEP
3050 || ins.handler_idx == WINED3DSIH_IF
3051 || ins.handler_idx == WINED3DSIH_RETP
3052 || ins.handler_idx == WINED3DSIH_TEXKILL)
3053 {
3054 switch (ins.flags)
3055 {
3056 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3057 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3058 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3059 }
3060 }
3061 else if (ins.handler_idx == WINED3DSIH_IFC
3062 || ins.handler_idx == WINED3DSIH_BREAKC)
3063 {
3064 switch (ins.flags)
3065 {
3066 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3067 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3068 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3069 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3070 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3071 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3072 default: shader_addline(&buffer, "_(%u)", ins.flags);
3073 }
3074 }
3075 else if (ins.handler_idx == WINED3DSIH_TEX
3076 && shader_version.major >= 2
3077 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3078 {
3079 shader_addline(&buffer, "p");
3080 }
3081 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3082 {
3083 switch (ins.flags)
3084 {
3085 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3086 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3087 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3088 }
3089 }
3090 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3091 {
3092 switch (ins.flags)
3093 {
3094 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3095 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3096 }
3097 }
3098 else if (ins.handler_idx == WINED3DSIH_SYNC)
3099 {
3100 shader_dump_sync_flags(&buffer, ins.flags);
3101 }
3102 else
3103 {
3104 shader_dump_precise_flags(&buffer, ins.flags);
3105 }
3106
3107 if (wined3d_shader_instruction_has_texel_offset(&ins))
3108 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3109
3110 for (i = 0; i < ins.dst_count; ++i)
3111 {
3112 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
3113 shader_addline(&buffer, !i ? " " : ", ");
3114 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
3115 }
3116
3117 /* Other source tokens */
3118 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
3119 {
3120 shader_addline(&buffer, !i ? " " : ", ");
3121 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
3122 }
3123 }
3124 shader_addline(&buffer, "\n");
3125 }
3126
3127 for (p = buffer.buffer; *p; p = q)
3128 {
3129 if (!(q = strstr(p, "\n")))
3130 q = p + strlen(p);
3131 else
3132 ++q;
3133 TRACE(" %.*s", (int)(q - p), p);
3134 }
3135
3136 string_buffer_free(&buffer);
3137 }
3138
3139 static void shader_cleanup(struct wined3d_shader *shader)
3140 {
3141 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3142 {
3143 heap_free(shader->u.hs.phases.control_point);
3144 heap_free(shader->u.hs.phases.fork);
3145 heap_free(shader->u.hs.phases.join);
3146 }
3147 else if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
3148 {
3149 heap_free((void *)shader->u.gs.so_desc.elements);
3150 }
3151
3152 heap_free(shader->patch_constant_signature.elements);
3153 heap_free(shader->output_signature.elements);
3154 heap_free(shader->input_signature.elements);
3155 shader->device->shader_backend->shader_destroy(shader);
3156 shader_cleanup_reg_maps(&shader->reg_maps);
3157 heap_free(shader->byte_code);
3158 shader_delete_constant_list(&shader->constantsF);
3159 shader_delete_constant_list(&shader->constantsB);
3160 shader_delete_constant_list(&shader->constantsI);
3161 list_remove(&shader->shader_list_entry);
3162
3163 if (shader->frontend && shader->frontend_data)
3164 shader->frontend->shader_free(shader->frontend_data);
3165 }
3166
3167 struct shader_none_priv
3168 {
3169 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3170 const struct wined3d_fragment_pipe_ops *fragment_pipe;
3171 BOOL ffp_proj_control;
3172 };
3173
3174 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3175 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3176 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3177 const struct wined3d_state *state) {}
3178 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3179 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3180 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3181 const struct wined3d_state *state) {}
3182 static void shader_none_destroy(struct wined3d_shader *shader) {}
3183 static void shader_none_free_context_data(struct wined3d_context *context) {}
3184 static void shader_none_init_context_state(struct wined3d_context *context) {}
3185
3186 /* Context activation is done by the caller. */
3187 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3188 const struct wined3d_state *state)
3189 {
3190 struct shader_none_priv *priv = shader_priv;
3191
3192 priv->vertex_pipe->vp_enable(context, !use_vs(state));
3193 priv->fragment_pipe->fp_enable(context, !use_ps(state));
3194 }
3195
3196 /* Context activation is done by the caller. */
3197 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3198 {
3199 struct shader_none_priv *priv = shader_priv;
3200
3201 priv->vertex_pipe->vp_enable(context, FALSE);
3202 priv->fragment_pipe->fp_enable(context, FALSE);
3203
3204 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3205 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3206 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3207 | (1u << WINED3D_SHADER_TYPE_HULL)
3208 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3209 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3210 }
3211
3212 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3213 const struct wined3d_fragment_pipe_ops *fragment_pipe)
3214 {
3215 struct fragment_caps fragment_caps;
3216 void *vertex_priv, *fragment_priv;
3217 struct shader_none_priv *priv;
3218
3219 if (!(priv = heap_alloc(sizeof(*priv))))
3220 return E_OUTOFMEMORY;
3221
3222 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3223 {
3224 ERR("Failed to initialize vertex pipe.\n");
3225 heap_free(priv);
3226 return E_FAIL;
3227 }
3228
3229 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3230 {
3231 ERR("Failed to initialize fragment pipe.\n");
3232 vertex_pipe->vp_free(device, NULL);
3233 heap_free(priv);
3234 return E_FAIL;
3235 }
3236
3237 priv->vertex_pipe = vertex_pipe;
3238 priv->fragment_pipe = fragment_pipe;
3239 fragment_pipe->get_caps(device->adapter, &fragment_caps);
3240 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3241
3242 device->vertex_priv = vertex_priv;
3243 device->fragment_priv = fragment_priv;
3244 device->shader_priv = priv;
3245
3246 return WINED3D_OK;
3247 }
3248
3249 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
3250 {
3251 struct shader_none_priv *priv = device->shader_priv;
3252
3253 priv->fragment_pipe->free_private(device, context);
3254 priv->vertex_pipe->vp_free(device, context);
3255 heap_free(priv);
3256 }
3257
3258 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3259 {
3260 return TRUE;
3261 }
3262
3263 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
3264 {
3265 /* Set the shader caps to 0 for the none shader backend */
3266 memset(caps, 0, sizeof(*caps));
3267 }
3268
3269 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3270 {
3271 /* We "support" every possible fixup, since we don't support any shader
3272 * model, and will never have to actually sample a texture. */
3273 return TRUE;
3274 }
3275
3276 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3277 {
3278 struct shader_none_priv *priv = shader_priv;
3279
3280 return priv->ffp_proj_control;
3281 }
3282
3283 const struct wined3d_shader_backend_ops none_shader_backend =
3284 {
3285 shader_none_handle_instruction,
3286 shader_none_precompile,
3287 shader_none_select,
3288 shader_none_select_compute,
3289 shader_none_disable,
3290 shader_none_update_float_vertex_constants,
3291 shader_none_update_float_pixel_constants,
3292 shader_none_load_constants,
3293 shader_none_destroy,
3294 shader_none_alloc,
3295 shader_none_free,
3296 shader_none_allocate_context_data,
3297 shader_none_free_context_data,
3298 shader_none_init_context_state,
3299 shader_none_get_caps,
3300 shader_none_color_fixup_supported,
3301 shader_none_has_ffp_proj_control,
3302 };
3303
3304 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
3305 {
3306 switch (level)
3307 {
3308 case WINED3D_FEATURE_LEVEL_11_1:
3309 case WINED3D_FEATURE_LEVEL_11:
3310 return 5;
3311 case WINED3D_FEATURE_LEVEL_10_1:
3312 case WINED3D_FEATURE_LEVEL_10:
3313 return 4;
3314 case WINED3D_FEATURE_LEVEL_9_3:
3315 return 3;
3316 case WINED3D_FEATURE_LEVEL_9_2:
3317 case WINED3D_FEATURE_LEVEL_9_1:
3318 return 2;
3319 default:
3320 return 1;
3321 }
3322 }
3323
3324 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
3325 enum wined3d_shader_type type, unsigned int float_const_count)
3326 {
3327 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3328 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3329 const struct wined3d_shader_version *version = &reg_maps->shader_version;
3330 const struct wined3d_shader_frontend *fe;
3331 unsigned int backend_version;
3332 HRESULT hr;
3333
3334 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
3335 shader, device, debug_shader_type(type), float_const_count);
3336
3337 fe = shader->frontend;
3338 if (!(shader->frontend_data = fe->shader_init(shader->function,
3339 shader->functionLength, &shader->output_signature)))
3340 {
3341 FIXME("Failed to initialize frontend.\n");
3342 return WINED3DERR_INVALIDCALL;
3343 }
3344
3345 /* First pass: trace shader. */
3346 if (TRACE_ON(d3d_shader))
3347 shader_trace_init(fe, shader->frontend_data);
3348
3349 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3350 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
3351 return hr;
3352
3353 if (version->type != type)
3354 {
3355 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
3356 return WINED3DERR_INVALIDCALL;
3357 }
3358 if (version->major > shader_max_version_from_feature_level(device->feature_level))
3359 {
3360 WARN("Shader version %u not supported by this device.\n", version->major);
3361 return WINED3DERR_INVALIDCALL;
3362 }
3363 switch (type)
3364 {
3365 case WINED3D_SHADER_TYPE_VERTEX:
3366 backend_version = d3d_info->limits.vs_version;
3367 break;
3368 case WINED3D_SHADER_TYPE_HULL:
3369 backend_version = d3d_info->limits.hs_version;
3370 break;
3371 case WINED3D_SHADER_TYPE_DOMAIN:
3372 backend_version = d3d_info->limits.ds_version;
3373 break;
3374 case WINED3D_SHADER_TYPE_GEOMETRY:
3375 backend_version = d3d_info->limits.gs_version;
3376 break;
3377 case WINED3D_SHADER_TYPE_PIXEL:
3378 backend_version = d3d_info->limits.ps_version;
3379 break;
3380 case WINED3D_SHADER_TYPE_COMPUTE:
3381 backend_version = d3d_info->limits.cs_version;
3382 break;
3383 default:
3384 FIXME("No backend version-checking for this shader type.\n");
3385 backend_version = 0;
3386 }
3387 if (version->major > backend_version)
3388 {
3389 WARN("Shader version %u.%u not supported by the current shader backend.\n",
3390 version->major, version->minor);
3391 return WINED3DERR_INVALIDCALL;
3392 }
3393
3394 shader->load_local_constsF = shader->lconst_inf_or_nan;
3395
3396 return WINED3D_OK;
3397 }
3398
3399 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3400 {
3401 ULONG refcount = InterlockedIncrement(&shader->ref);
3402
3403 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3404
3405 return refcount;
3406 }
3407
3408 static void wined3d_shader_init_object(void *object)
3409 {
3410 struct wined3d_shader *shader = object;
3411 struct wined3d_device *device = shader->device;
3412
3413 list_add_head(&device->shaders, &shader->shader_list_entry);
3414
3415 device->shader_backend->shader_precompile(device->shader_priv, shader);
3416 }
3417
3418 static void wined3d_shader_destroy_object(void *object)
3419 {
3420 shader_cleanup(object);
3421 heap_free(object);
3422 }
3423
3424 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3425 {
3426 ULONG refcount = InterlockedDecrement(&shader->ref);
3427
3428 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3429
3430 if (!refcount)
3431 {
3432 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3433 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3434 }
3435
3436 return refcount;
3437 }
3438
3439 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3440 {
3441 TRACE("shader %p.\n", shader);
3442
3443 return shader->parent;
3444 }
3445
3446 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3447 void *byte_code, UINT *byte_code_size)
3448 {
3449 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3450
3451 if (!byte_code)
3452 {
3453 *byte_code_size = shader->byte_code_size;
3454 return WINED3D_OK;
3455 }
3456
3457 if (*byte_code_size < shader->byte_code_size)
3458 {
3459 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3460 * than the required size we should write the required size and
3461 * return D3DERR_MOREDATA. That's not actually true. */
3462 return WINED3DERR_INVALIDCALL;
3463 }
3464
3465 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
3466
3467 return WINED3D_OK;
3468 }
3469
3470 /* Set local constants for d3d8 shaders. */
3471 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3472 UINT start_idx, const float *src_data, UINT count)
3473 {
3474 UINT end_idx = start_idx + count;
3475 UINT i;
3476
3477 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3478
3479 if (end_idx > shader->limits->constant_float)
3480 {
3481 WARN("end_idx %u > float constants limit %u.\n",
3482 end_idx, shader->limits->constant_float);
3483 end_idx = shader->limits->constant_float;
3484 }
3485
3486 for (i = start_idx; i < end_idx; ++i)
3487 {
3488 struct wined3d_shader_lconst *lconst;
3489 float *value;
3490
3491 if (!(lconst = heap_alloc(sizeof(*lconst))))
3492 return E_OUTOFMEMORY;
3493
3494 lconst->idx = i;
3495 value = (float *)lconst->value;
3496 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3497 list_add_head(&shader->constantsF, &lconst->entry);
3498
3499 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3500 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3501 {
3502 shader->lconst_inf_or_nan = TRUE;
3503 }
3504 }
3505
3506 return WINED3D_OK;
3507 }
3508
3509 static void init_interpolation_compile_args(DWORD *interpolation_args,
3510 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
3511 {
3512 if (!d3d_info->shader_output_interpolation || !pixel_shader
3513 || pixel_shader->reg_maps.shader_version.major < 4)
3514 {
3515 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3516 return;
3517 }
3518
3519 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3520 sizeof(pixel_shader->u.ps.interpolation_mode));
3521 }
3522
3523 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3524 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_context *context)
3525 {
3526 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3527 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3528 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3529 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3530
3531 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3532 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3533 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3534 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3535 args->point_size = state->gl_primitive_type == GL_POINTS;
3536 args->per_vertex_point_size = shader->reg_maps.point_size;
3537 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
3538 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3539 if (shader->reg_maps.shader_version.major >= 4)
3540 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3541 : geometry_shader ? geometry_shader->limits->packed_input
3542 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3543 else
3544 args->next_shader_input_count = 0;
3545 args->swizzle_map = swizzle_map;
3546 if (d3d_info->emulated_flatshading)
3547 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3548 else
3549 args->flatshading = 0;
3550
3551 init_interpolation_compile_args(args->interpolation_mode,
3552 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
3553 }
3554
3555 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3556 {
3557 if (usage_idx1 != usage_idx2)
3558 return FALSE;
3559 if (usage1 == usage2)
3560 return TRUE;
3561 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3562 return TRUE;
3563 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3564 return TRUE;
3565
3566 return FALSE;
3567 }
3568
3569 BOOL vshader_get_input(const struct wined3d_shader *shader,
3570 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3571 {
3572 WORD map = shader->reg_maps.input_registers;
3573 unsigned int i;
3574
3575 for (i = 0; map; map >>= 1, ++i)
3576 {
3577 if (!(map & 1)) continue;
3578
3579 if (match_usage(shader->u.vs.attributes[i].usage,
3580 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3581 {
3582 *regnum = i;
3583 return TRUE;
3584 }
3585 }
3586 return FALSE;
3587 }
3588
3589 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3590 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3591 {
3592 HRESULT hr;
3593
3594 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
3595
3596 if (!desc->byte_code)
3597 return WINED3DERR_INVALIDCALL;
3598
3599 shader->ref = 1;
3600 shader->device = device;
3601 shader->parent = parent;
3602 shader->parent_ops = parent_ops;
3603
3604 list_init(&shader->linked_programs);
3605 list_init(&shader->constantsF);
3606 list_init(&shader->constantsB);
3607 list_init(&shader->constantsI);
3608 shader->lconst_inf_or_nan = FALSE;
3609 list_init(&shader->reg_maps.indexable_temps);
3610 list_init(&shader->shader_list_entry);
3611
3612 if (desc->byte_code_size == ~(size_t)0)
3613 {
3614 struct wined3d_shader_version shader_version;
3615 const struct wined3d_shader_frontend *fe;
3616 struct wined3d_shader_instruction ins;
3617 const DWORD *ptr;
3618 void *fe_data;
3619
3620 if (!(shader->frontend = shader_select_frontend(WINED3D_SHADER_BYTE_CODE_FORMAT_SM1)))
3621 {
3622 FIXME("Unable to find frontend for shader.\n");
3623 hr = WINED3DERR_INVALIDCALL;
3624 goto fail;
3625 }
3626
3627 fe = shader->frontend;
3628 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
3629 {
3630 WARN("Failed to initialise frontend data.\n");
3631 hr = WINED3DERR_INVALIDCALL;
3632 goto fail;
3633 }
3634
3635 fe->shader_read_header(fe_data, &ptr, &shader_version);
3636 while (!fe->shader_is_end(fe_data, &ptr))
3637 fe->shader_read_instruction(fe_data, &ptr, &ins);
3638
3639 fe->shader_free(fe_data);
3640
3641 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3642
3643 if (!(shader->byte_code = heap_alloc(shader->byte_code_size)))
3644 {
3645 hr = E_OUTOFMEMORY;
3646 goto fail;
3647 }
3648 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
3649
3650 shader->function = shader->byte_code;
3651 shader->functionLength = shader->byte_code_size;
3652 }
3653 else
3654 {
3655 enum wined3d_shader_byte_code_format format;
3656 unsigned int max_version;
3657
3658 if (!(shader->byte_code = heap_alloc(desc->byte_code_size)))
3659 {
3660 hr = E_OUTOFMEMORY;
3661 goto fail;
3662 }
3663 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
3664 shader->byte_code_size = desc->byte_code_size;
3665
3666 max_version = shader_max_version_from_feature_level(device->feature_level);
3667 if (FAILED(hr = shader_extract_from_dxbc(shader, max_version, &format)))
3668 goto fail;
3669
3670 if (!(shader->frontend = shader_select_frontend(format)))
3671 {
3672 FIXME("Unable to find frontend for shader.\n");
3673 hr = WINED3DERR_INVALIDCALL;
3674 goto fail;
3675 }
3676 }
3677
3678 return WINED3D_OK;
3679
3680 fail:
3681 shader_cleanup(shader);
3682 return hr;
3683 }
3684
3685 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3686 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3687 {
3688 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3689 unsigned int i;
3690 HRESULT hr;
3691
3692 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3693 return hr;
3694
3695 if (FAILED(hr = shader_set_function(shader, device,
3696 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
3697 {
3698 shader_cleanup(shader);
3699 return hr;
3700 }
3701
3702 for (i = 0; i < shader->input_signature.element_count; ++i)
3703 {
3704 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3705
3706 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3707 continue;
3708
3709 shader->u.vs.attributes[input->register_idx].usage =
3710 shader_usage_from_semantic_name(input->semantic_name);
3711 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3712 }
3713
3714 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3715 shader->load_local_constsF = TRUE;
3716
3717 return WINED3D_OK;
3718 }
3719
3720 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
3721 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
3722 {
3723 struct wined3d_shader_signature_element *e = s->elements;
3724 unsigned int i;
3725
3726 for (i = 0; i < s->element_count; ++i)
3727 {
3728 if (e[i].stream_idx == stream_idx
3729 && !_strnicmp(e[i].semantic_name, semantic_name, -1)
3730 && e[i].semantic_idx == semantic_idx)
3731 return &e[i];
3732 }
3733
3734 return NULL;
3735 }
3736
3737 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
3738 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
3739 {
3740 const struct wined3d_shader_signature_element *output;
3741 unsigned int idx;
3742
3743 if (!(output = shader_find_signature_element(&shader->output_signature,
3744 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
3745 return FALSE;
3746
3747 for (idx = 0; idx < 4; ++idx)
3748 {
3749 if (output->mask & (1u << idx))
3750 break;
3751 }
3752 idx += so_element->component_idx;
3753
3754 *register_idx = output->register_idx;
3755 *component_idx = idx;
3756 return TRUE;
3757 }
3758
3759 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
3760 const struct wined3d_stream_output_desc *so_desc)
3761 {
3762 const struct wined3d_shader_frontend *fe = shader->frontend;
3763 const struct wined3d_shader_signature_element *output;
3764 unsigned int i, component_idx, register_idx, mask;
3765 struct wined3d_stream_output_element *elements;
3766 struct wined3d_shader_version shader_version;
3767 const DWORD *ptr;
3768 void *fe_data;
3769 HRESULT hr;
3770
3771 if (!so_desc)
3772 return WINED3D_OK;
3773
3774 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
3775 {
3776 WARN("Failed to initialise frontend data.\n");
3777 return WINED3DERR_INVALIDCALL;
3778 }
3779 fe->shader_read_header(fe_data, &ptr, &shader_version);
3780 fe->shader_free(fe_data);
3781
3782 switch (shader_version.type)
3783 {
3784 case WINED3D_SHADER_TYPE_VERTEX:
3785 case WINED3D_SHADER_TYPE_DOMAIN:
3786 shader->function = NULL;
3787 shader->functionLength = 0;
3788 break;
3789 case WINED3D_SHADER_TYPE_GEOMETRY:
3790 break;
3791 default:
3792 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
3793 return E_INVALIDARG;
3794 }
3795
3796 if (!shader->function)
3797 {
3798 shader->reg_maps.shader_version = shader_version;
3799 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
3800 shader_set_limits(shader);
3801 if (FAILED(hr = shader_scan_output_signature(shader)))
3802 return hr;
3803 }
3804
3805 if (!(elements = heap_calloc(so_desc->element_count, sizeof(*elements))))
3806 return E_OUTOFMEMORY;
3807
3808 shader->u.gs.so_desc = *so_desc;
3809 shader->u.gs.so_desc.elements = elements;
3810 memcpy(elements, so_desc->elements, so_desc->element_count * sizeof(*elements));
3811
3812 for (i = 0; i < so_desc->element_count; ++i)
3813 {
3814 struct wined3d_stream_output_element *e = &elements[i];
3815
3816 if (!e->semantic_name)
3817 continue;
3818 if (!(output = shader_find_signature_element(&shader->output_signature,
3819 e->stream_idx, e->semantic_name, e->semantic_idx))
3820 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
3821 {
3822 WARN("Failed to find output signature element for stream output entry.\n");
3823 return E_INVALIDARG;
3824 }
3825
3826 e->semantic_name = output->semantic_name;
3827
3828 mask = ((1u << e->component_count) - 1) << component_idx;
3829 if ((output->mask & 0xff & mask) != mask)
3830 {
3831 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
3832 component_idx, e->component_count, mask, output->mask & 0xff);
3833 return E_INVALIDARG;
3834 }
3835 }
3836
3837 return WINED3D_OK;
3838 }
3839
3840 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3841 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3842 void *parent, const struct wined3d_parent_ops *parent_ops)
3843 {
3844 HRESULT hr;
3845
3846 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3847 return hr;
3848
3849 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
3850 goto fail;
3851
3852 if (shader->function
3853 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
3854 goto fail;
3855
3856 return WINED3D_OK;
3857
3858 fail:
3859 shader_cleanup(shader);
3860 return hr;
3861 }
3862
3863 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3864 struct ds_compile_args *args, const struct wined3d_context *context)
3865 {
3866 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3867 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3868 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3869
3870 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3871 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3872
3873 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3874 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3875 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3876
3877 args->render_offscreen = context->render_offscreen;
3878
3879 init_interpolation_compile_args(args->interpolation_mode,
3880 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
3881
3882 args->padding = 0;
3883 }
3884
3885 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3886 struct gs_compile_args *args, const struct wined3d_context *context)
3887 {
3888 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3889
3890 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3891
3892 if (!(args->primitive_type = shader->u.gs.input_type))
3893 args->primitive_type = d3d_primitive_type_from_gl(state->gl_primitive_type);
3894
3895 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
3896 }
3897
3898 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3899 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3900 {
3901 const struct wined3d_gl_info *gl_info = &context->device->adapter->gl_info;
3902 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3903 struct wined3d_texture *texture;
3904 unsigned int i;
3905
3906 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3907 if (!d3d_info->srgb_write_control && needs_srgb_write(context, state, state->fb))
3908 {
3909 static unsigned int warned = 0;
3910
3911 args->srgb_correction = 1;
3912 if (state->blend_state && state->blend_state->desc.enable && !warned++)
3913 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3914 "support, expect rendering artifacts.\n");
3915 }
3916
3917 if (shader->reg_maps.shader_version.major == 1
3918 && shader->reg_maps.shader_version.minor <= 3)
3919 {
3920 for (i = 0; i < shader->limits->sampler; ++i)
3921 {
3922 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3923
3924 if (flags & WINED3D_TTFF_PROJECTED)
3925 {
3926 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3927
3928 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3929 {
3930 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3931 unsigned int j;
3932 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3933 DWORD max_valid = WINED3D_TTFF_COUNT4;
3934
3935 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3936 {
3937 struct wined3d_vertex_declaration_element *element =
3938 &state->vertex_declaration->elements[j];
3939
3940 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3941 && element->usage_idx == index)
3942 {
3943 max_valid = element->format->component_count;
3944 break;
3945 }
3946 }
3947 if (!tex_transform || tex_transform > max_valid)
3948 {
3949 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3950 tex_transform, max_valid);
3951 tex_transform = max_valid;
3952 }
3953 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3954 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3955 && tex_transform > WINED3D_TTFF_COUNT2)
3956 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3957 && tex_transform > WINED3D_TTFF_COUNT3))
3958 tex_transform |= WINED3D_PSARGS_PROJECTED;
3959 else
3960 {
3961 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3962 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3963 i, tex_transform, resource_type);
3964 }
3965 }
3966 else
3967 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3968
3969 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3970 }
3971 }
3972 }
3973 if (shader->reg_maps.shader_version.major == 1
3974 && shader->reg_maps.shader_version.minor <= 4)
3975 {
3976 for (i = 0; i < shader->limits->sampler; ++i)
3977 {
3978 if (!shader->reg_maps.resource_info[i].type)
3979 continue;
3980
3981 /* Treat unbound textures as 2D. The dummy texture will provide
3982 * the proper sample value. The tex_types bitmap defaults to
3983 * 2D because of the memset. */
3984 if (!(texture = state->textures[i]))
3985 continue;
3986
3987 switch (wined3d_texture_gl(texture)->target)
3988 {
3989 /* RECT textures are distinguished from 2D textures via np2_fixup */
3990 default:
3991 break;
3992
3993 case GL_TEXTURE_3D:
3994 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3995 break;
3996
3997 case GL_TEXTURE_CUBE_MAP_ARB:
3998 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3999 break;
4000 }
4001 }
4002 }
4003 else if (shader->reg_maps.shader_version.major <= 3)
4004 {
4005 for (i = 0; i < shader->limits->sampler; ++i)
4006 {
4007 enum wined3d_shader_resource_type resource_type;
4008 enum wined3d_shader_tex_types tex_type;
4009
4010 if (!(resource_type = shader->reg_maps.resource_info[i].type))
4011 continue;
4012
4013 switch (resource_type)
4014 {
4015 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
4016 tex_type = WINED3D_SHADER_TEX_3D;
4017 break;
4018 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
4019 tex_type = WINED3D_SHADER_TEX_CUBE;
4020 break;
4021 default:
4022 tex_type = WINED3D_SHADER_TEX_2D;
4023 break;
4024 }
4025
4026 if ((texture = state->textures[i]))
4027 {
4028 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
4029 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4030 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
4031 tex_type = WINED3D_SHADER_TEX_2D;
4032 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
4033 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
4034 tex_type = WINED3D_SHADER_TEX_3D;
4035 }
4036 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4037 }
4038 }
4039
4040 if (shader->reg_maps.shader_version.major >= 4)
4041 {
4042 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
4043 args->shadow = 0;
4044 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4045 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4046 args->np2_fixup = 0;
4047 }
4048 else
4049 {
4050 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4051 {
4052 if (!shader->reg_maps.resource_info[i].type)
4053 continue;
4054
4055 texture = state->textures[i];
4056 if (!texture)
4057 {
4058 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4059 continue;
4060 }
4061 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
4062 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4063 else
4064 args->color_fixup[i] = texture->resource.format->color_fixup;
4065
4066 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
4067 args->shadow |= 1u << i;
4068
4069 /* Flag samplers that need NP2 texcoord fixup. */
4070 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
4071 args->np2_fixup |= (1u << i);
4072 }
4073 }
4074
4075 if (shader->reg_maps.shader_version.major >= 3)
4076 {
4077 if (position_transformed)
4078 args->vp_mode = WINED3D_VP_MODE_NONE;
4079 else if (use_vs(state))
4080 args->vp_mode = WINED3D_VP_MODE_SHADER;
4081 else
4082 args->vp_mode = WINED3D_VP_MODE_FF;
4083 args->fog = WINED3D_FFP_PS_FOG_OFF;
4084 }
4085 else
4086 {
4087 args->vp_mode = WINED3D_VP_MODE_SHADER;
4088 if (state->render_states[WINED3D_RS_FOGENABLE])
4089 {
4090 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
4091 {
4092 case WINED3D_FOG_NONE:
4093 if (position_transformed || use_vs(state))
4094 {
4095 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
4096 break;
4097 }
4098
4099 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
4100 {
4101 case WINED3D_FOG_NONE: /* Fall through. */
4102 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4103 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4104 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4105 }
4106 break;
4107
4108 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4109 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4110 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4111 }
4112 }
4113 else
4114 {
4115 args->fog = WINED3D_FFP_PS_FOG_OFF;
4116 }
4117 }
4118
4119 if (!d3d_info->full_ffp_varyings)
4120 {
4121 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
4122
4123 args->texcoords_initialized = 0;
4124 for (i = 0; i < WINED3D_MAX_TEXTURES; ++i)
4125 {
4126 if (vs)
4127 {
4128 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
4129 args->texcoords_initialized |= 1u << i;
4130 }
4131 else
4132 {
4133 const struct wined3d_stream_info *si = &context->stream_info;
4134 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
4135
4136 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
4137 & WINED3D_FFP_TCI_MASK
4138 || (coord_idx < WINED3D_MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
4139 args->texcoords_initialized |= 1u << i;
4140 }
4141 }
4142 }
4143 else
4144 {
4145 args->texcoords_initialized = (1u << WINED3D_MAX_TEXTURES) - 1;
4146 }
4147
4148 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
4149 && state->gl_primitive_type == GL_POINTS;
4150
4151 if (d3d_info->ffp_alpha_test)
4152 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
4153 else
4154 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4155 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4156 : WINED3D_CMP_ALWAYS) - 1;
4157
4158 if (d3d_info->emulated_flatshading)
4159 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4160
4161 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
4162 ? context->render_offscreen : 0;
4163
4164 for (i = 0; i < ARRAY_SIZE(state->fb->render_targets); ++i)
4165 {
4166 struct wined3d_rendertarget_view *rtv = state->fb->render_targets[i];
4167 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
4168 args->rt_alpha_swizzle |= 1u << i;
4169 }
4170 }
4171
4172 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4173 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4174 {
4175 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4176 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4177 HRESULT hr;
4178
4179 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
4180 return hr;
4181
4182 if (FAILED(hr = shader_set_function(shader, device,
4183 WINED3D_SHADER_TYPE_PIXEL, device->adapter->d3d_info.limits.ps_uniform_count)))
4184 {
4185 shader_cleanup(shader);
4186 return hr;
4187 }
4188
4189 for (i = 0; i < MAX_REG_INPUT; ++i)
4190 {
4191 if (shader->u.ps.input_reg_used & (1u << i))
4192 {
4193 ++num_regs_used;
4194 highest_reg_used = i;
4195 }
4196 }
4197
4198 /* Don't do any register mapping magic if it is not needed, or if we can't
4199 * achieve anything anyway */
4200 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4201 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4202 || shader->reg_maps.shader_version.major >= 4)
4203 {
4204 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4205 {
4206 /* This happens with relative addressing. The input mapper function
4207 * warns about this if the higher registers are declared too, so
4208 * don't write a FIXME here */
4209 WARN("More varying registers used than supported\n");
4210 }
4211
4212 for (i = 0; i < MAX_REG_INPUT; ++i)
4213 {
4214 shader->u.ps.input_reg_map[i] = i;
4215 }
4216
4217 shader->u.ps.declared_in_count = highest_reg_used + 1;
4218 }
4219 else
4220 {
4221 shader->u.ps.declared_in_count = 0;
4222 for (i = 0; i < MAX_REG_INPUT; ++i)
4223 {
4224 if (shader->u.ps.input_reg_used & (1u << i))
4225 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4226 else shader->u.ps.input_reg_map[i] = ~0U;
4227 }
4228 }
4229
4230 return WINED3D_OK;
4231 }
4232
4233 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
4234 unsigned int resource_idx, DWORD tex_types)
4235 {
4236 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
4237 {
4238 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
4239 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
4240 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
4241 };
4242
4243 unsigned int idx;
4244
4245 if (reg_maps->shader_version.major > 3)
4246 return reg_maps->resource_info[resource_idx].type;
4247
4248 if (!reg_maps->resource_info[resource_idx].type)
4249 return 0;
4250
4251 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
4252 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
4253 return shader_resource_type_from_shader_tex_types[idx];
4254 }
4255
4256 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4257 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4258 {
4259 struct wined3d_shader *object;
4260 HRESULT hr;
4261
4262 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4263 device, desc, parent, parent_ops, shader);
4264
4265 if (!(object = heap_alloc_zero(sizeof(*object))))
4266 return E_OUTOFMEMORY;
4267
4268 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4269 {
4270 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4271 heap_free(object);
4272 return hr;
4273 }
4274
4275 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
4276 {
4277 shader_cleanup(object);
4278 heap_free(object);
4279 return hr;
4280 }
4281
4282 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4283
4284 TRACE("Created compute shader %p.\n", object);
4285 *shader = object;
4286
4287 return WINED3D_OK;
4288 }
4289
4290 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4291 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4292 {
4293 struct wined3d_shader *object;
4294 HRESULT hr;
4295
4296 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4297 device, desc, parent, parent_ops, shader);
4298
4299 if (!(object = heap_alloc_zero(sizeof(*object))))
4300 return E_OUTOFMEMORY;
4301
4302 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4303 {
4304 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4305 heap_free(object);
4306 return hr;
4307 }
4308
4309 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
4310 {
4311 shader_cleanup(object);
4312 heap_free(object);
4313 return hr;
4314 }
4315
4316 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4317
4318 TRACE("Created domain shader %p.\n", object);
4319 *shader = object;
4320
4321 return WINED3D_OK;
4322 }
4323
4324 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4325 const struct wined3d_stream_output_desc *so_desc, void *parent,
4326 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4327 {
4328 struct wined3d_shader *object;
4329 HRESULT hr;
4330
4331 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4332 device, desc, so_desc, parent, parent_ops, shader);
4333
4334 if (!(object = heap_alloc_zero(sizeof(*object))))
4335 return E_OUTOFMEMORY;
4336
4337 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4338 {
4339 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4340 heap_free(object);
4341 return hr;
4342 }
4343
4344 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4345
4346 TRACE("Created geometry shader %p.\n", object);
4347 *shader = object;
4348
4349 return WINED3D_OK;
4350 }
4351
4352 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4353 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4354 {
4355 struct wined3d_shader *object;
4356 HRESULT hr;
4357
4358 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4359 device, desc, parent, parent_ops, shader);
4360
4361 if (!(object = heap_alloc_zero(sizeof(*object))))
4362 return E_OUTOFMEMORY;
4363
4364 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4365 {
4366 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4367 heap_free(object);
4368 return hr;
4369 }
4370
4371 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
4372 {
4373 shader_cleanup(object);
4374 heap_free(object);
4375 return hr;
4376 }
4377
4378 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4379
4380 TRACE("Created hull shader %p.\n", object);
4381 *shader = object;
4382
4383 return WINED3D_OK;
4384 }
4385
4386 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4387 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4388 {
4389 struct wined3d_shader *object;
4390 HRESULT hr;
4391
4392 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4393 device, desc, parent, parent_ops, shader);
4394
4395 if (!(object = heap_alloc_zero(sizeof(*object))))
4396 return E_OUTOFMEMORY;
4397
4398 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4399 {
4400 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4401 heap_free(object);
4402 return hr;
4403 }
4404
4405 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4406
4407 TRACE("Created pixel shader %p.\n", object);
4408 *shader = object;
4409
4410 return WINED3D_OK;
4411 }
4412
4413 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4414 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4415 {
4416 struct wined3d_shader *object;
4417 HRESULT hr;
4418
4419 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4420 device, desc, parent, parent_ops, shader);
4421
4422 if (!(object = heap_alloc_zero(sizeof(*object))))
4423 return E_OUTOFMEMORY;
4424
4425 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4426 {
4427 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4428 heap_free(object);
4429 return hr;
4430 }
4431
4432 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4433
4434 TRACE("Created vertex shader %p.\n", object);
4435 *shader = object;
4436
4437 return WINED3D_OK;
4438 }
Windows API implementation