?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Avoid calling _strnicmp() with limit -1.
[wine.git] / dlls / wined3d / shader.c
bloba296e34bc9fd671175d3534f7c01f272877680c6
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 <stdio.h>
26 #include <string.h>
27
28 #include "wined3d_private.h"
29
30 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
31
32 const struct wined3d_vec4 wined3d_srgb_const[] =
33 {
34 /* pow, mul_high, sub_high, mul_low */
35 {0.41666f, 1.055f, 0.055f, 12.92f},
36 /* cmp */
37 {0.0031308f, 0.0f, 0.0f, 0.0f},
38 };
39
40 static const char * const shader_opcode_names[] =
41 {
42 /* WINED3DSIH_ABS */ "abs",
43 /* WINED3DSIH_ADD */ "add",
44 /* WINED3DSIH_AND */ "and",
45 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
46 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
47 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
48 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
49 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
50 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
51 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
52 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
53 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
54 /* WINED3DSIH_BEM */ "bem",
55 /* WINED3DSIH_BFI */ "bfi",
56 /* WINED3DSIH_BFREV */ "bfrev",
57 /* WINED3DSIH_BREAK */ "break",
58 /* WINED3DSIH_BREAKC */ "breakc",
59 /* WINED3DSIH_BREAKP */ "breakp",
60 /* WINED3DSIH_BUFINFO */ "bufinfo",
61 /* WINED3DSIH_CALL */ "call",
62 /* WINED3DSIH_CALLNZ */ "callnz",
63 /* WINED3DSIH_CASE */ "case",
64 /* WINED3DSIH_CMP */ "cmp",
65 /* WINED3DSIH_CND */ "cnd",
66 /* WINED3DSIH_CONTINUE */ "continue",
67 /* WINED3DSIH_CONTINUEP */ "continuec",
68 /* WINED3DSIH_COUNTBITS */ "countbits",
69 /* WINED3DSIH_CRS */ "crs",
70 /* WINED3DSIH_CUT */ "cut",
71 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
72 /* WINED3DSIH_DCL */ "dcl",
73 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
74 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
75 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
76 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
77 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
78 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
79 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
81 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
82 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
83 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
84 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
85 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
86 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
87 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
88 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
89 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
90 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
91 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
92 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
93 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
94 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
95 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
96 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
97 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
98 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
99 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
100 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
101 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
102 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
103 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
104 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
105 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
106 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
107 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
108 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
109 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
110 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
111 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
112 /* WINED3DSIH_DEF */ "def",
113 /* WINED3DSIH_DEFAULT */ "default",
114 /* WINED3DSIH_DEFB */ "defb",
115 /* WINED3DSIH_DEFI */ "defi",
116 /* WINED3DSIH_DIV */ "div",
117 /* WINED3DSIH_DP2 */ "dp2",
118 /* WINED3DSIH_DP2ADD */ "dp2add",
119 /* WINED3DSIH_DP3 */ "dp3",
120 /* WINED3DSIH_DP4 */ "dp4",
121 /* WINED3DSIH_DST */ "dst",
122 /* WINED3DSIH_DSX */ "dsx",
123 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
124 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
125 /* WINED3DSIH_DSY */ "dsy",
126 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
127 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
128 /* WINED3DSIH_ELSE */ "else",
129 /* WINED3DSIH_EMIT */ "emit",
130 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
131 /* WINED3DSIH_ENDIF */ "endif",
132 /* WINED3DSIH_ENDLOOP */ "endloop",
133 /* WINED3DSIH_ENDREP */ "endrep",
134 /* WINED3DSIH_ENDSWITCH */ "endswitch",
135 /* WINED3DSIH_EQ */ "eq",
136 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
137 /* WINED3DSIH_EXP */ "exp",
138 /* WINED3DSIH_EXPP */ "expp",
139 /* WINED3DSIH_F16TOF32 */ "f16tof32",
140 /* WINED3DSIH_F32TOF16 */ "f32tof16",
141 /* WINED3DSIH_FCALL */ "fcall",
142 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
143 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
144 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
145 /* WINED3DSIH_FRC */ "frc",
146 /* WINED3DSIH_FTOI */ "ftoi",
147 /* WINED3DSIH_FTOU */ "ftou",
148 /* WINED3DSIH_GATHER4 */ "gather4",
149 /* WINED3DSIH_GATHER4_C */ "gather4_c",
150 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
151 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
152 /* WINED3DSIH_GE */ "ge",
153 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
154 /* WINED3DSIH_HS_DECLS */ "hs_decls",
155 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
156 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
157 /* WINED3DSIH_IADD */ "iadd",
158 /* WINED3DSIH_IBFE */ "ibfe",
159 /* WINED3DSIH_IEQ */ "ieq",
160 /* WINED3DSIH_IF */ "if",
161 /* WINED3DSIH_IFC */ "ifc",
162 /* WINED3DSIH_IGE */ "ige",
163 /* WINED3DSIH_ILT */ "ilt",
164 /* WINED3DSIH_IMAD */ "imad",
165 /* WINED3DSIH_IMAX */ "imax",
166 /* WINED3DSIH_IMIN */ "imin",
167 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
168 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
169 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
170 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
171 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
172 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
173 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
174 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
175 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
176 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
177 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
178 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
179 /* WINED3DSIH_IMUL */ "imul",
180 /* WINED3DSIH_INE */ "ine",
181 /* WINED3DSIH_INEG */ "ineg",
182 /* WINED3DSIH_ISHL */ "ishl",
183 /* WINED3DSIH_ISHR */ "ishr",
184 /* WINED3DSIH_ITOF */ "itof",
185 /* WINED3DSIH_LABEL */ "label",
186 /* WINED3DSIH_LD */ "ld",
187 /* WINED3DSIH_LD2DMS */ "ld2dms",
188 /* WINED3DSIH_LD_RAW */ "ld_raw",
189 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
190 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
191 /* WINED3DSIH_LIT */ "lit",
192 /* WINED3DSIH_LOD */ "lod",
193 /* WINED3DSIH_LOG */ "log",
194 /* WINED3DSIH_LOGP */ "logp",
195 /* WINED3DSIH_LOOP */ "loop",
196 /* WINED3DSIH_LRP */ "lrp",
197 /* WINED3DSIH_LT */ "lt",
198 /* WINED3DSIH_M3x2 */ "m3x2",
199 /* WINED3DSIH_M3x3 */ "m3x3",
200 /* WINED3DSIH_M3x4 */ "m3x4",
201 /* WINED3DSIH_M4x3 */ "m4x3",
202 /* WINED3DSIH_M4x4 */ "m4x4",
203 /* WINED3DSIH_MAD */ "mad",
204 /* WINED3DSIH_MAX */ "max",
205 /* WINED3DSIH_MIN */ "min",
206 /* WINED3DSIH_MOV */ "mov",
207 /* WINED3DSIH_MOVA */ "mova",
208 /* WINED3DSIH_MOVC */ "movc",
209 /* WINED3DSIH_MUL */ "mul",
210 /* WINED3DSIH_NE */ "ne",
211 /* WINED3DSIH_NOP */ "nop",
212 /* WINED3DSIH_NOT */ "not",
213 /* WINED3DSIH_NRM */ "nrm",
214 /* WINED3DSIH_OR */ "or",
215 /* WINED3DSIH_PHASE */ "phase",
216 /* WINED3DSIH_POW */ "pow",
217 /* WINED3DSIH_RCP */ "rcp",
218 /* WINED3DSIH_REP */ "rep",
219 /* WINED3DSIH_RESINFO */ "resinfo",
220 /* WINED3DSIH_RET */ "ret",
221 /* WINED3DSIH_RETP */ "retp",
222 /* WINED3DSIH_ROUND_NE */ "round_ne",
223 /* WINED3DSIH_ROUND_NI */ "round_ni",
224 /* WINED3DSIH_ROUND_PI */ "round_pi",
225 /* WINED3DSIH_ROUND_Z */ "round_z",
226 /* WINED3DSIH_RSQ */ "rsq",
227 /* WINED3DSIH_SAMPLE */ "sample",
228 /* WINED3DSIH_SAMPLE_B */ "sample_b",
229 /* WINED3DSIH_SAMPLE_C */ "sample_c",
230 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
231 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
232 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
233 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
234 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
235 /* WINED3DSIH_SETP */ "setp",
236 /* WINED3DSIH_SGE */ "sge",
237 /* WINED3DSIH_SGN */ "sgn",
238 /* WINED3DSIH_SINCOS */ "sincos",
239 /* WINED3DSIH_SLT */ "slt",
240 /* WINED3DSIH_SQRT */ "sqrt",
241 /* WINED3DSIH_STORE_RAW */ "store_raw",
242 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
243 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
244 /* WINED3DSIH_SUB */ "sub",
245 /* WINED3DSIH_SWAPC */ "swapc",
246 /* WINED3DSIH_SWITCH */ "switch",
247 /* WINED3DSIH_SYNC */ "sync",
248 /* WINED3DSIH_TEX */ "texld",
249 /* WINED3DSIH_TEXBEM */ "texbem",
250 /* WINED3DSIH_TEXBEML */ "texbeml",
251 /* WINED3DSIH_TEXCOORD */ "texcrd",
252 /* WINED3DSIH_TEXDEPTH */ "texdepth",
253 /* WINED3DSIH_TEXDP3 */ "texdp3",
254 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
255 /* WINED3DSIH_TEXKILL */ "texkill",
256 /* WINED3DSIH_TEXLDD */ "texldd",
257 /* WINED3DSIH_TEXLDL */ "texldl",
258 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
259 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
260 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
261 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
262 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
263 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
264 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
265 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
266 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
267 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
268 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
269 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
270 /* WINED3DSIH_UBFE */ "ubfe",
271 /* WINED3DSIH_UDIV */ "udiv",
272 /* WINED3DSIH_UGE */ "uge",
273 /* WINED3DSIH_ULT */ "ult",
274 /* WINED3DSIH_UMAX */ "umax",
275 /* WINED3DSIH_UMIN */ "umin",
276 /* WINED3DSIH_UMUL */ "umul",
277 /* WINED3DSIH_USHR */ "ushr",
278 /* WINED3DSIH_UTOF */ "utof",
279 /* WINED3DSIH_XOR */ "xor",
280 };
281
282 static const char * const semantic_names[] =
283 {
284 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
285 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
286 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
287 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
288 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
289 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
290 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
291 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
292 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
293 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
294 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
295 /* WINED3D_DECL_USAGE_FOG */ "FOG",
296 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
297 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
298 };
299
300 static const struct
301 {
302 enum wined3d_shader_input_sysval_semantic sysval_semantic;
303 const char *sysval_name;
304 }
305 shader_input_sysval_semantic_names[] =
306 {
307 {WINED3D_SIV_POSITION, "position"},
308 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
309 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
310 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
311 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
312 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
313 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
314 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
315 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
316 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
317 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
318 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
319 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
320 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
322 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
323 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
324 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
325 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
326 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
327 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
328 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
329 };
330
331 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
332 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
333
334 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
335 {
336 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
337 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
338
339 return shader_opcode_names[handler_idx];
340 }
341
342 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
343 {
344 if (usage >= ARRAY_SIZE(semantic_names))
345 {
346 FIXME("Unrecognized usage %#x.\n", usage);
347 return "UNRECOGNIZED";
348 }
349
350 return semantic_names[usage];
351 }
352
353 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
354 {
355 unsigned int i;
356
357 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
358 {
359 if (!strcmp(name, semantic_names[i]))
360 return i;
361 }
362
363 return ~0U;
364 }
365
366 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
367 {
368 switch (usage)
369 {
370 case WINED3D_DECL_USAGE_POSITION:
371 return WINED3D_SV_POSITION;
372 default:
373 return 0;
374 }
375 }
376
377 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
378 {
379 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
380 }
381
382 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
383 const struct wined3d_shader_semantic *s)
384 {
385 e->semantic_name = shader_semantic_name_from_usage(s->usage);
386 e->semantic_idx = s->usage_idx;
387 e->stream_idx = 0;
388 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
389 e->component_type = WINED3D_TYPE_FLOAT;
390 e->register_idx = s->reg.reg.idx[0].offset;
391 e->mask = s->reg.write_mask;
392 }
393
394 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
395 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
396 {
397 e->semantic_name = shader_semantic_name_from_usage(usage);
398 e->semantic_idx = usage_idx;
399 e->stream_idx = 0;
400 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
401 e->component_type = WINED3D_TYPE_FLOAT;
402 e->register_idx = reg_idx;
403 e->mask = write_mask;
404 }
405
406 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
407 {
408 switch (format)
409 {
410 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
411 return &sm1_shader_frontend;
412
413 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
414 return &sm4_shader_frontend;
415
416 default:
417 WARN("Invalid byte code format %#x specified.\n", format);
418 return NULL;
419 }
420 }
421
422 void string_buffer_clear(struct wined3d_string_buffer *buffer)
423 {
424 buffer->buffer[0] = '\0';
425 buffer->content_size = 0;
426 }
427
428 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
429 {
430 buffer->buffer_size = 32;
431 if (!(buffer->buffer = heap_alloc(buffer->buffer_size)))
432 {
433 ERR("Failed to allocate shader buffer memory.\n");
434 return FALSE;
435 }
436
437 string_buffer_clear(buffer);
438 return TRUE;
439 }
440
441 void string_buffer_free(struct wined3d_string_buffer *buffer)
442 {
443 heap_free(buffer->buffer);
444 }
445
446 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
447 {
448 char *new_buffer;
449 unsigned int new_buffer_size = buffer->buffer_size * 2;
450
451 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
452 new_buffer_size *= 2;
453 if (!(new_buffer = heap_realloc(buffer->buffer, new_buffer_size)))
454 {
455 ERR("Failed to grow buffer.\n");
456 buffer->buffer[buffer->content_size] = '\0';
457 return FALSE;
458 }
459 buffer->buffer = new_buffer;
460 buffer->buffer_size = new_buffer_size;
461 return TRUE;
462 }
463
464 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
465 {
466 unsigned int rem;
467 int rc;
468
469 rem = buffer->buffer_size - buffer->content_size;
470 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
471 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
472 return rc;
473
474 buffer->content_size += rc;
475 return 0;
476 }
477
478 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
479 {
480 va_list args;
481 int ret;
482
483 for (;;)
484 {
485 va_start(args, format);
486 ret = shader_vaddline(buffer, format, args);
487 va_end(args);
488 if (!ret)
489 return ret;
490 if (!string_buffer_resize(buffer, ret))
491 return -1;
492 }
493 }
494
495 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
496 {
497 struct wined3d_string_buffer *buffer;
498
499 if (list_empty(&list->list))
500 {
501 buffer = heap_alloc(sizeof(*buffer));
502 if (!buffer || !string_buffer_init(buffer))
503 {
504 ERR("Couldn't allocate buffer for temporary string.\n");
505 heap_free(buffer);
506 return NULL;
507 }
508 }
509 else
510 {
511 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
512 list_remove(&buffer->entry);
513 }
514 string_buffer_clear(buffer);
515 return buffer;
516 }
517
518 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
519 {
520 if (!buffer)
521 return 0;
522 string_buffer_clear(buffer);
523 return shader_vaddline(buffer, format, args);
524 }
525
526 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
527 {
528 va_list args;
529 int ret;
530
531 for (;;)
532 {
533 va_start(args, format);
534 ret = string_buffer_vsprintf(buffer, format, args);
535 va_end(args);
536 if (!ret)
537 return;
538 if (!string_buffer_resize(buffer, ret))
539 return;
540 }
541 }
542
543 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
544 {
545 if (!buffer)
546 return;
547 list_add_head(&list->list, &buffer->entry);
548 }
549
550 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
551 {
552 list_init(&list->list);
553 }
554
555 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
556 {
557 struct wined3d_string_buffer *buffer, *buffer_next;
558
559 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
560 {
561 string_buffer_free(buffer);
562 heap_free(buffer);
563 }
564 list_init(&list->list);
565 }
566
567 /* Convert floating point offset relative to a register file to an absolute
568 * offset for float constants. */
569 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
570 {
571 switch (register_type)
572 {
573 case WINED3DSPR_CONST: return register_idx;
574 case WINED3DSPR_CONST2: return 2048 + register_idx;
575 case WINED3DSPR_CONST3: return 4096 + register_idx;
576 case WINED3DSPR_CONST4: return 6144 + register_idx;
577 default:
578 FIXME("Unsupported register type: %u.\n", register_type);
579 return register_idx;
580 }
581 }
582
583 static void shader_delete_constant_list(struct list *clist)
584 {
585 struct wined3d_shader_lconst *constant, *constant_next;
586
587 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
588 heap_free(constant);
589 list_init(clist);
590 }
591
592 static void shader_set_limits(struct wined3d_shader *shader)
593 {
594 static const struct limits_entry
595 {
596 unsigned int min_version;
597 unsigned int max_version;
598 struct wined3d_shader_limits limits;
599 }
600 vs_limits[] =
601 {
602 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
603 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
604 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
605 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
606 * even though they are capable of supporting much more (GL
607 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
608 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
609 * shaders to 256. */
610 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
611 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
612 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
613 {0}
614 },
615 hs_limits[] =
616 {
617 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
618 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
619 },
620 ds_limits[] =
621 {
622 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
623 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
624 },
625 gs_limits[] =
626 {
627 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
628 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
629 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
630 {0}
631 },
632 ps_limits[] =
633 {
634 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
635 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
636 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
637 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
638 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
639 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
640 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
641 {0}
642 },
643 cs_limits[] =
644 {
645 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
646 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
647 };
648 const struct limits_entry *limits_array;
649 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
650 shader->reg_maps.shader_version.minor);
651 int i = 0;
652
653 switch (shader->reg_maps.shader_version.type)
654 {
655 default:
656 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
657 /* Fall-through. */
658 case WINED3D_SHADER_TYPE_VERTEX:
659 limits_array = vs_limits;
660 break;
661 case WINED3D_SHADER_TYPE_HULL:
662 limits_array = hs_limits;
663 break;
664 case WINED3D_SHADER_TYPE_DOMAIN:
665 limits_array = ds_limits;
666 break;
667 case WINED3D_SHADER_TYPE_GEOMETRY:
668 limits_array = gs_limits;
669 break;
670 case WINED3D_SHADER_TYPE_PIXEL:
671 limits_array = ps_limits;
672 break;
673 case WINED3D_SHADER_TYPE_COMPUTE:
674 limits_array = cs_limits;
675 break;
676 }
677
678 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
679 {
680 if (shader_version <= limits_array[i].max_version)
681 {
682 shader->limits = &limits_array[i].limits;
683 break;
684 }
685 ++i;
686 }
687 if (!shader->limits)
688 {
689 FIXME("Unexpected shader version \"%u.%u\".\n",
690 shader->reg_maps.shader_version.major,
691 shader->reg_maps.shader_version.minor);
692 shader->limits = &limits_array[max(0, i - 1)].limits;
693 }
694 }
695
696 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
697 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
698 {
699 switch (reg->type)
700 {
701 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
702 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
703 reg_maps->texcoord |= 1u << reg->idx[0].offset;
704 else
705 reg_maps->address |= 1u << reg->idx[0].offset;
706 break;
707
708 case WINED3DSPR_TEMP:
709 reg_maps->temporary |= 1u << reg->idx[0].offset;
710 break;
711
712 case WINED3DSPR_INPUT:
713 if (reg->idx[0].rel_addr)
714 reg_maps->input_rel_addressing = 1;
715 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
716 {
717 /* If relative addressing is used, we must assume that all
718 * registers are used. Even if it is a construct like v3[aL],
719 * we can't assume that v0, v1 and v2 aren't read because aL
720 * can be negative. */
721 if (reg->idx[0].rel_addr)
722 shader->u.ps.input_reg_used = ~0u;
723 else
724 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
725 }
726 else
727 {
728 reg_maps->input_registers |= 1u << reg->idx[0].offset;
729 }
730 break;
731
732 case WINED3DSPR_RASTOUT:
733 if (reg->idx[0].offset == 1)
734 reg_maps->fog = 1;
735 if (reg->idx[0].offset == 2)
736 reg_maps->point_size = 1;
737 break;
738
739 case WINED3DSPR_MISCTYPE:
740 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
741 {
742 if (!reg->idx[0].offset)
743 reg_maps->vpos = 1;
744 else if (reg->idx[0].offset == 1)
745 reg_maps->usesfacing = 1;
746 }
747 break;
748
749 case WINED3DSPR_CONST:
750 if (reg->idx[0].rel_addr)
751 {
752 if (reg->idx[0].offset < reg_maps->min_rel_offset)
753 reg_maps->min_rel_offset = reg->idx[0].offset;
754 if (reg->idx[0].offset > reg_maps->max_rel_offset)
755 reg_maps->max_rel_offset = reg->idx[0].offset;
756 reg_maps->usesrelconstF = TRUE;
757 }
758 else
759 {
760 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
761 {
762 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
763 return FALSE;
764 }
765 else
766 {
767 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
768 }
769 }
770 break;
771
772 case WINED3DSPR_CONSTINT:
773 if (reg->idx[0].offset >= shader->limits->constant_int)
774 {
775 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
776 return FALSE;
777 }
778 else
779 {
780 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
781 }
782 break;
783
784 case WINED3DSPR_CONSTBOOL:
785 if (reg->idx[0].offset >= shader->limits->constant_bool)
786 {
787 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
788 return FALSE;
789 }
790 else
791 {
792 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
793 }
794 break;
795
796 case WINED3DSPR_COLOROUT:
797 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
798 break;
799
800 case WINED3DSPR_OUTCONTROLPOINT:
801 reg_maps->vocp = 1;
802 break;
803
804 case WINED3DSPR_SAMPLEMASK:
805 reg_maps->sample_mask = 1;
806 break;
807
808 default:
809 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
810 reg->type, reg->idx[0].offset, reg->idx[1].offset);
811 break;
812 }
813 return TRUE;
814 }
815
816 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
817 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
818 {
819 struct wined3d_shader_sampler_map_entry *entries, *entry;
820 struct wined3d_shader_sampler_map *map;
821 unsigned int i;
822
823 map = &reg_maps->sampler_map;
824 entries = map->entries;
825 for (i = 0; i < map->count; ++i)
826 {
827 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
828 return;
829 }
830
831 if (!map->size)
832 {
833 if (!(entries = heap_calloc(4, sizeof(*entries))))
834 {
835 ERR("Failed to allocate sampler map entries.\n");
836 return;
837 }
838 map->size = 4;
839 map->entries = entries;
840 }
841 else if (map->count == map->size)
842 {
843 size_t new_size = map->size * 2;
844
845 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
846 || !(entries = heap_realloc(entries, sizeof(*entries) * new_size)))
847 {
848 ERR("Failed to resize sampler map entries.\n");
849 return;
850 }
851 map->size = new_size;
852 map->entries = entries;
853 }
854
855 entry = &entries[map->count++];
856 entry->resource_idx = resource_idx;
857 entry->sampler_idx = sampler_idx;
858 entry->bind_idx = bind_idx;
859 }
860
861 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
862 {
863 switch (instr)
864 {
865 case WINED3DSIH_M4x4:
866 case WINED3DSIH_M3x4:
867 return param == 1 ? 3 : 0;
868
869 case WINED3DSIH_M4x3:
870 case WINED3DSIH_M3x3:
871 return param == 1 ? 2 : 0;
872
873 case WINED3DSIH_M3x2:
874 return param == 1 ? 1 : 0;
875
876 default:
877 return 0;
878 }
879 }
880
881 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
882 unsigned int register_idx, unsigned int size, unsigned int stride)
883 {
884 struct wined3d_shader_tgsm *tgsm;
885
886 if (register_idx >= MAX_TGSM_REGISTERS)
887 {
888 ERR("Invalid TGSM register index %u.\n", register_idx);
889 return S_OK;
890 }
891 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
892 {
893 FIXME("TGSM declarations are allowed only in compute shaders.\n");
894 return S_OK;
895 }
896
897 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
898 register_idx + 1, sizeof(*reg_maps->tgsm)))
899 return E_OUTOFMEMORY;
900
901 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
902 tgsm = &reg_maps->tgsm[register_idx];
903 tgsm->size = size;
904 tgsm->stride = stride;
905 return S_OK;
906 }
907
908 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
909 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
910 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
911 {
912 struct wined3d_shader_phase *phase;
913
914 if ((phase = *current_phase))
915 {
916 phase->end = previous_instruction_ptr;
917 *current_phase = NULL;
918 }
919
920 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
921 {
922 ERR("Unexpected shader type %s.\n", debug_shader_type(shader->reg_maps.shader_version.type));
923 return E_FAIL;
924 }
925
926 switch (ins->handler_idx)
927 {
928 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
929 if (shader->u.hs.phases.control_point)
930 {
931 FIXME("Multiple control point phases.\n");
932 heap_free(shader->u.hs.phases.control_point);
933 }
934 if (!(shader->u.hs.phases.control_point = heap_alloc_zero(sizeof(*shader->u.hs.phases.control_point))))
935 return E_OUTOFMEMORY;
936 phase = shader->u.hs.phases.control_point;
937 break;
938 case WINED3DSIH_HS_FORK_PHASE:
939 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
940 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
941 sizeof(*shader->u.hs.phases.fork)))
942 return E_OUTOFMEMORY;
943 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
944 break;
945 case WINED3DSIH_HS_JOIN_PHASE:
946 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
947 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
948 sizeof(*shader->u.hs.phases.join)))
949 return E_OUTOFMEMORY;
950 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
951 break;
952 default:
953 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
954 return E_FAIL;
955 }
956
957 phase->start = current_instruction_ptr;
958 *current_phase = phase;
959
960 return WINED3D_OK;
961 }
962
963 static HRESULT shader_calculate_clip_or_cull_distance_mask(
964 const struct wined3d_shader_signature_element *e, unsigned int *mask)
965 {
966 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
967 * the only allowed semantic indices.
968 */
969 if (e->semantic_idx >= WINED3D_MAX_CLIP_DISTANCES / 4)
970 {
971 *mask = 0;
972 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
973 return WINED3DERR_INVALIDCALL;
974 }
975
976 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
977 return WINED3D_OK;
978 }
979
980 static void wined3d_insert_interpolation_mode(DWORD *packed_interpolation_mode,
981 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
982 {
983 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
984 FIXME("Unexpected interpolation mode %#x.\n", mode);
985
986 wined3d_insert_bits(packed_interpolation_mode,
987 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
988 }
989
990 static HRESULT shader_scan_output_signature(struct wined3d_shader *shader)
991 {
992 const struct wined3d_shader_signature *output_signature = &shader->output_signature;
993 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
994 unsigned int i;
995 HRESULT hr;
996
997 for (i = 0; i < output_signature->element_count; ++i)
998 {
999 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1000 unsigned int mask;
1001
1002 reg_maps->output_registers |= 1u << e->register_idx;
1003 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
1004 {
1005 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1006 return hr;
1007 reg_maps->clip_distance_mask |= mask;
1008 }
1009 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
1010 {
1011 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1012 return hr;
1013 reg_maps->cull_distance_mask |= mask;
1014 }
1015 else if (e->sysval_semantic == WINED3D_SV_VIEWPORT_ARRAY_INDEX)
1016 {
1017 reg_maps->viewport_array = 1;
1018 }
1019 }
1020
1021 return WINED3D_OK;
1022 }
1023
1024 /* Note that this does not count the loop register as an address register. */
1025 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, DWORD constf_size)
1026 {
1027 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
1028 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
1029 struct wined3d_shader_signature *output_signature = &shader->output_signature;
1030 struct wined3d_shader_signature *input_signature = &shader->input_signature;
1031 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1032 const struct wined3d_shader_frontend *fe = shader->frontend;
1033 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
1034 struct wined3d_shader_version shader_version;
1035 struct wined3d_shader_phase *phase = NULL;
1036 const DWORD *ptr, *prev_ins, *current_ins;
1037 void *fe_data = shader->frontend_data;
1038 unsigned int i;
1039 HRESULT hr;
1040
1041 memset(reg_maps, 0, sizeof(*reg_maps));
1042 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1043 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1044 reg_maps->min_rel_offset = ~0U;
1045 list_init(&reg_maps->indexable_temps);
1046
1047 fe->shader_read_header(fe_data, &ptr, &shader_version);
1048 prev_ins = current_ins = ptr;
1049 reg_maps->shader_version = shader_version;
1050
1051 shader_set_limits(shader);
1052
1053 if (!(reg_maps->constf = heap_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1054 sizeof(*reg_maps->constf))))
1055 {
1056 ERR("Failed to allocate constant map memory.\n");
1057 return E_OUTOFMEMORY;
1058 }
1059
1060 while (!fe->shader_is_end(fe_data, &ptr))
1061 {
1062 struct wined3d_shader_instruction ins;
1063
1064 current_ins = ptr;
1065 /* Fetch opcode. */
1066 fe->shader_read_instruction(fe_data, &ptr, &ins);
1067
1068 /* Unhandled opcode, and its parameters. */
1069 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1070 {
1071 WARN("Encountered unrecognised or invalid instruction.\n");
1072 return WINED3DERR_INVALIDCALL;
1073 }
1074
1075 /* Handle declarations. */
1076 if (ins.handler_idx == WINED3DSIH_DCL
1077 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1078 {
1079 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1080 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1081
1082 switch (semantic->reg.reg.type)
1083 {
1084 /* Mark input registers used. */
1085 case WINED3DSPR_INPUT:
1086 if (reg_idx >= MAX_REG_INPUT)
1087 {
1088 ERR("Invalid input register index %u.\n", reg_idx);
1089 break;
1090 }
1091 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1092 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1093 return WINED3DERR_INVALIDCALL;
1094 reg_maps->input_registers |= 1u << reg_idx;
1095 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1096 break;
1097
1098 /* Vertex shader: mark 3.0 output registers used, save token. */
1099 case WINED3DSPR_OUTPUT:
1100 if (reg_idx >= MAX_REG_OUTPUT)
1101 {
1102 ERR("Invalid output register index %u.\n", reg_idx);
1103 break;
1104 }
1105 reg_maps->output_registers |= 1u << reg_idx;
1106 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1107 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1108 reg_maps->fog = 1;
1109 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1110 reg_maps->point_size = 1;
1111 break;
1112
1113 case WINED3DSPR_SAMPLER:
1114 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1115 case WINED3DSPR_RESOURCE:
1116 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1117 {
1118 ERR("Invalid resource index %u.\n", reg_idx);
1119 break;
1120 }
1121 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1122 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1123 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1124 break;
1125
1126 case WINED3DSPR_UAV:
1127 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1128 {
1129 ERR("Invalid UAV resource index %u.\n", reg_idx);
1130 break;
1131 }
1132 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1133 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1134 if (ins.flags)
1135 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1136 break;
1137
1138 default:
1139 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1140 break;
1141 }
1142 }
1143 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1144 {
1145 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1146 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1147 {
1148 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1149 }
1150 else
1151 {
1152 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1153 wined3d_bitmap_set(&reg_maps->cb_map, reg->idx[0].offset);
1154 }
1155 }
1156 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1157 {
1158 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1159 {
1160 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1161 shader->u.ps.force_early_depth_stencil = TRUE;
1162 else
1163 FIXME("Invalid instruction %#x for shader type %#x.\n",
1164 ins.handler_idx, shader_version.type);
1165 }
1166 else
1167 {
1168 WARN("Ignoring global flags %#x.\n", ins.flags);
1169 }
1170 }
1171 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1172 {
1173 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1174 shader->u.gs.instance_count = ins.declaration.count;
1175 else
1176 FIXME("Invalid instruction %#x for shader type %#x.\n",
1177 ins.handler_idx, shader_version.type);
1178 }
1179 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1180 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1181 {
1182 if (phase)
1183 phase->instance_count = ins.declaration.count;
1184 else
1185 FIXME("Instruction %s outside of shader phase.\n",
1186 debug_d3dshaderinstructionhandler(ins.handler_idx));
1187 }
1188 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1189 {
1190 if (reg_maps->icb)
1191 FIXME("Multiple immediate constant buffers.\n");
1192 reg_maps->icb = ins.declaration.icb;
1193 }
1194 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1195 {
1196 if (phase)
1197 {
1198 FIXME("Indexable temporary registers not supported.\n");
1199 }
1200 else
1201 {
1202 struct wined3d_shader_indexable_temp *reg;
1203
1204 if (!(reg = heap_alloc(sizeof(*reg))))
1205 return E_OUTOFMEMORY;
1206
1207 *reg = ins.declaration.indexable_temp;
1208 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1209 }
1210 }
1211 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1212 {
1213 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1214 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1215 else
1216 FIXME("Invalid instruction %#x for shader type %#x.\n",
1217 ins.handler_idx, shader_version.type);
1218 }
1219 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1220 {
1221 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1222 if (reg_idx >= MAX_REG_INPUT)
1223 {
1224 ERR("Invalid register index %u.\n", reg_idx);
1225 break;
1226 }
1227 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1228 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1229 else
1230 FIXME("Invalid instruction %#x for shader type %#x.\n",
1231 ins.handler_idx, shader_version.type);
1232 }
1233 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1234 {
1235 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1236 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1237 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1238 {
1239 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1240 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1241 else
1242 FIXME("Invalid instruction %#x for shader type %#x.\n",
1243 ins.handler_idx, shader_version.type);
1244 }
1245 }
1246 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1247 {
1248 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1249 shader->u.hs.output_vertex_count = ins.declaration.count;
1250 else
1251 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1252 }
1253 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1254 {
1255 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1256 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1257 else
1258 FIXME("Invalid instruction %#x for shader type %#x.\n",
1259 ins.handler_idx, shader_version.type);
1260 }
1261 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1262 {
1263 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1264 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1265 {
1266 ERR("Invalid resource index %u.\n", reg_idx);
1267 break;
1268 }
1269 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1270 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1271 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1272 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1273 }
1274 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1275 {
1276 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1277 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1278 {
1279 ERR("Invalid resource index %u.\n", reg_idx);
1280 break;
1281 }
1282 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1283 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1284 reg_maps->resource_info[reg_idx].flags = 0;
1285 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1286 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1287 }
1288 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1289 {
1290 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1291 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1292 }
1293 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1294 {
1295 if (phase)
1296 phase->temporary_count = ins.declaration.count;
1297 else
1298 reg_maps->temporary_count = ins.declaration.count;
1299 }
1300 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1301 {
1302 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1303 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1304 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1305 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1306 }
1307 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1308 {
1309 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1310 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1311 else
1312 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1313 }
1314 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1315 {
1316 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1317 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1318 else
1319 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1320 }
1321 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1322 {
1323 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1324 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1325 return hr;
1326 }
1327 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1328 {
1329 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1330 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1331 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1332 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1333 return hr;
1334 }
1335 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1336 {
1337 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1338 {
1339 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1340 }
1341 else
1342 {
1343 FIXME("Invalid instruction %#x for shader type %#x.\n",
1344 ins.handler_idx, shader_version.type);
1345 }
1346 }
1347 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1348 {
1349 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1350 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1351 {
1352 ERR("Invalid UAV resource index %u.\n", reg_idx);
1353 break;
1354 }
1355 if (ins.flags)
1356 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1357 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1358 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1359 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1360 }
1361 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1362 {
1363 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1364 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1365 {
1366 ERR("Invalid UAV resource index %u.\n", reg_idx);
1367 break;
1368 }
1369 if (ins.flags)
1370 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1371 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1372 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1373 reg_maps->uav_resource_info[reg_idx].flags = 0;
1374 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1375 }
1376 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1377 {
1378 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1379 shader->u.gs.vertices_out = ins.declaration.count;
1380 else
1381 FIXME("Invalid instruction %#x for shader type %#x.\n",
1382 ins.handler_idx, shader_version.type);
1383 }
1384 else if (ins.handler_idx == WINED3DSIH_DEF)
1385 {
1386 struct wined3d_shader_lconst *lconst;
1387 float *value;
1388
1389 if (!(lconst = heap_alloc(sizeof(*lconst))))
1390 return E_OUTOFMEMORY;
1391
1392 lconst->idx = ins.dst[0].reg.idx[0].offset;
1393 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1394 value = (float *)lconst->value;
1395
1396 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1397 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1398 {
1399 if (value[0] < -1.0f) value[0] = -1.0f;
1400 else if (value[0] > 1.0f) value[0] = 1.0f;
1401 if (value[1] < -1.0f) value[1] = -1.0f;
1402 else if (value[1] > 1.0f) value[1] = 1.0f;
1403 if (value[2] < -1.0f) value[2] = -1.0f;
1404 else if (value[2] > 1.0f) value[2] = 1.0f;
1405 if (value[3] < -1.0f) value[3] = -1.0f;
1406 else if (value[3] > 1.0f) value[3] = 1.0f;
1407 }
1408
1409 list_add_head(&shader->constantsF, &lconst->entry);
1410
1411 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1412 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1413 {
1414 shader->lconst_inf_or_nan = TRUE;
1415 }
1416 }
1417 else if (ins.handler_idx == WINED3DSIH_DEFI)
1418 {
1419 struct wined3d_shader_lconst *lconst;
1420
1421 if (!(lconst = heap_alloc(sizeof(*lconst))))
1422 return E_OUTOFMEMORY;
1423
1424 lconst->idx = ins.dst[0].reg.idx[0].offset;
1425 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1426
1427 list_add_head(&shader->constantsI, &lconst->entry);
1428 reg_maps->local_int_consts |= (1u << lconst->idx);
1429 }
1430 else if (ins.handler_idx == WINED3DSIH_DEFB)
1431 {
1432 struct wined3d_shader_lconst *lconst;
1433
1434 if (!(lconst = heap_alloc(sizeof(*lconst))))
1435 return E_OUTOFMEMORY;
1436
1437 lconst->idx = ins.dst[0].reg.idx[0].offset;
1438 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1439
1440 list_add_head(&shader->constantsB, &lconst->entry);
1441 reg_maps->local_bool_consts |= (1u << lconst->idx);
1442 }
1443 /* Handle shader phases. */
1444 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1445 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1446 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1447 {
1448 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1449 return hr;
1450 }
1451 /* For subroutine prototypes. */
1452 else if (ins.handler_idx == WINED3DSIH_LABEL)
1453 {
1454 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1455 }
1456 /* Set texture, address, temporary registers. */
1457 else
1458 {
1459 BOOL color0_mov = FALSE;
1460 unsigned int i;
1461
1462 /* This will loop over all the registers and try to
1463 * make a bitmask of the ones we're interested in.
1464 *
1465 * Relative addressing tokens are ignored, but that's
1466 * okay, since we'll catch any address registers when
1467 * they are initialized (required by spec). */
1468 for (i = 0; i < ins.dst_count; ++i)
1469 {
1470 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1471 shader_version.type, constf_size))
1472 return WINED3DERR_INVALIDCALL;
1473
1474 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1475 {
1476 UINT idx = ins.dst[i].reg.idx[0].offset;
1477
1478 switch (ins.dst[i].reg.type)
1479 {
1480 case WINED3DSPR_RASTOUT:
1481 if (shader_version.major >= 3)
1482 break;
1483 switch (idx)
1484 {
1485 case 0: /* oPos */
1486 reg_maps->output_registers |= 1u << 10;
1487 shader_signature_from_usage(&output_signature_elements[10],
1488 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1489 break;
1490
1491 case 1: /* oFog */
1492 reg_maps->output_registers |= 1u << 11;
1493 shader_signature_from_usage(&output_signature_elements[11],
1494 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1495 break;
1496
1497 case 2: /* oPts */
1498 reg_maps->output_registers |= 1u << 11;
1499 shader_signature_from_usage(&output_signature_elements[11],
1500 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1501 break;
1502 }
1503 break;
1504
1505 case WINED3DSPR_ATTROUT:
1506 if (shader_version.major >= 3)
1507 break;
1508 if (idx < 2)
1509 {
1510 idx += 8;
1511 if (reg_maps->output_registers & (1u << idx))
1512 {
1513 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1514 }
1515 else
1516 {
1517 reg_maps->output_registers |= 1u << idx;
1518 shader_signature_from_usage(&output_signature_elements[idx],
1519 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1520 }
1521 }
1522 break;
1523
1524 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1525 if (shader_version.major >= 3)
1526 {
1527 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1528 {
1529 WARN("Invalid output register index %u.\n", idx);
1530 break;
1531 }
1532 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1533 break;
1534 }
1535 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1536 {
1537 WARN("Invalid texcoord index %u.\n", idx);
1538 break;
1539 }
1540 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1541 if (reg_maps->output_registers & (1u << idx))
1542 {
1543 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1544 }
1545 else
1546 {
1547 reg_maps->output_registers |= 1u << idx;
1548 shader_signature_from_usage(&output_signature_elements[idx],
1549 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1550 }
1551 break;
1552
1553 default:
1554 break;
1555 }
1556 }
1557
1558 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1559 {
1560 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1561 {
1562 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1563 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1564 * the mov and perform the sRGB write correction from the source register.
1565 *
1566 * However, if the mov is only partial, we can't do this, and if the write
1567 * comes from an instruction other than MOV it is hard to do as well. If
1568 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1569 shader->u.ps.color0_mov = FALSE;
1570 if (ins.handler_idx == WINED3DSIH_MOV
1571 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1572 {
1573 /* Used later when the source register is read. */
1574 color0_mov = TRUE;
1575 }
1576 }
1577 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1578 * end
1579 */
1580 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1581 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1582 {
1583 shader->u.ps.color0_mov = FALSE;
1584 }
1585 }
1586
1587 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1588 if (shader_version.major == 1
1589 && (ins.handler_idx == WINED3DSIH_TEX
1590 || ins.handler_idx == WINED3DSIH_TEXBEM
1591 || ins.handler_idx == WINED3DSIH_TEXBEML
1592 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1593 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1594 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1595 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1596 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1597 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1598 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1599 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1600 {
1601 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1602
1603 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1604 {
1605 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1606 continue;
1607 }
1608
1609 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1610 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1611 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1612 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1613 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1614
1615 /* texbem is only valid with < 1.4 pixel shaders */
1616 if (ins.handler_idx == WINED3DSIH_TEXBEM
1617 || ins.handler_idx == WINED3DSIH_TEXBEML)
1618 {
1619 reg_maps->bumpmat |= 1u << reg_idx;
1620 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1621 {
1622 reg_maps->luminanceparams |= 1u << reg_idx;
1623 }
1624 }
1625 }
1626 else if (ins.handler_idx == WINED3DSIH_BEM)
1627 {
1628 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1629 }
1630 }
1631
1632 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1633 {
1634 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1635 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1636 {
1637 ERR("Invalid UAV index %u.\n", reg_idx);
1638 break;
1639 }
1640 reg_maps->uav_counter_mask |= (1u << reg_idx);
1641 }
1642 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1643 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1644 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1645 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1646 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1647 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1648 {
1649 const struct wined3d_shader_register *reg;
1650
1651 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1652 reg = &ins.src[1].reg;
1653 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1654 reg = &ins.src[2].reg;
1655 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1656 reg = &ins.dst[0].reg;
1657 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1658 reg = &ins.src[0].reg;
1659 else
1660 reg = &ins.dst[1].reg;
1661
1662 if (reg->type == WINED3DSPR_UAV)
1663 {
1664 if (reg->idx[0].offset >= MAX_UNORDERED_ACCESS_VIEWS)
1665 {
1666 ERR("Invalid UAV index %u.\n", reg->idx[0].offset);
1667 break;
1668 }
1669 reg_maps->uav_read_mask |= (1u << reg->idx[0].offset);
1670 }
1671 }
1672 else if (ins.handler_idx == WINED3DSIH_NRM)
1673 {
1674 reg_maps->usesnrm = 1;
1675 }
1676 else if (ins.handler_idx == WINED3DSIH_DSY
1677 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1678 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1679 {
1680 reg_maps->usesdsy = 1;
1681 }
1682 else if (ins.handler_idx == WINED3DSIH_DSX
1683 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1684 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1685 {
1686 reg_maps->usesdsx = 1;
1687 }
1688 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1689 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1690 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1691 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1692 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1693 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1694 else if (ins.handler_idx == WINED3DSIH_LOOP
1695 || ins.handler_idx == WINED3DSIH_REP)
1696 {
1697 ++cur_loop_depth;
1698 if (cur_loop_depth > max_loop_depth)
1699 max_loop_depth = cur_loop_depth;
1700 }
1701 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1702 || ins.handler_idx == WINED3DSIH_ENDREP)
1703 {
1704 --cur_loop_depth;
1705 }
1706 else if (ins.handler_idx == WINED3DSIH_GATHER4
1707 || ins.handler_idx == WINED3DSIH_GATHER4_C
1708 || ins.handler_idx == WINED3DSIH_SAMPLE
1709 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1710 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1711 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1712 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1713 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1714 {
1715 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1716 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1717 }
1718 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1719 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1720 {
1721 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1722 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1723 }
1724 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1725 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1726 {
1727 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1728 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1729 }
1730 else if (ins.handler_idx == WINED3DSIH_LD
1731 || ins.handler_idx == WINED3DSIH_LD2DMS
1732 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1733 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1734 {
1735 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1736 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1737 }
1738 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1739 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1740 {
1741 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1742 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1743 }
1744
1745 if (ins.predicate)
1746 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1747 shader_version.type, constf_size))
1748 return WINED3DERR_INVALIDCALL;
1749
1750 for (i = 0; i < ins.src_count; ++i)
1751 {
1752 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1753 struct wined3d_shader_register reg = ins.src[i].reg;
1754
1755 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1756 shader_version.type, constf_size))
1757 return WINED3DERR_INVALIDCALL;
1758 while (count)
1759 {
1760 ++reg.idx[0].offset;
1761 if (!shader_record_register_usage(shader, reg_maps, &reg,
1762 shader_version.type, constf_size))
1763 return WINED3DERR_INVALIDCALL;
1764 --count;
1765 }
1766
1767 if (color0_mov)
1768 {
1769 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1770 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1771 {
1772 shader->u.ps.color0_mov = TRUE;
1773 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1774 }
1775 }
1776 }
1777 }
1778
1779 prev_ins = current_ins;
1780 }
1781 reg_maps->loop_depth = max_loop_depth;
1782
1783 if (phase)
1784 {
1785 phase->end = prev_ins;
1786 phase = NULL;
1787 }
1788
1789 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1790 * R0 is written to the render target. */
1791 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1792 reg_maps->rt_mask |= (1u << 0);
1793
1794 if (input_signature->elements)
1795 {
1796 for (i = 0; i < input_signature->element_count; ++i)
1797 {
1798 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1799 {
1800 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1801 {
1802 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1803 return WINED3DERR_INVALIDCALL;
1804 }
1805 }
1806 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1807 {
1808 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1809 reg_maps->vpos = 1;
1810 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1811 reg_maps->usesfacing = 1;
1812 }
1813 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1814 }
1815 }
1816 else if (!input_signature->elements && reg_maps->input_registers)
1817 {
1818 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1819 struct wined3d_shader_signature_element *e;
1820 unsigned int i;
1821
1822 if (!(input_signature->elements = heap_calloc(count, sizeof(*input_signature->elements))))
1823 return E_OUTOFMEMORY;
1824 input_signature->element_count = count;
1825
1826 e = input_signature->elements;
1827 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1828 {
1829 if (!(reg_maps->input_registers & (1u << i)))
1830 continue;
1831 input_signature_elements[i].register_idx = i;
1832 *e++ = input_signature_elements[i];
1833 }
1834 }
1835
1836 if (output_signature->elements)
1837 {
1838 if (FAILED(hr = shader_scan_output_signature(shader)))
1839 return hr;
1840 }
1841 else if (reg_maps->output_registers)
1842 {
1843 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1844 struct wined3d_shader_signature_element *e;
1845
1846 if (!(output_signature->elements = heap_calloc(count, sizeof(*output_signature->elements))))
1847 return E_OUTOFMEMORY;
1848 output_signature->element_count = count;
1849
1850 e = output_signature->elements;
1851 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1852 {
1853 if (!(reg_maps->output_registers & (1u << i)))
1854 continue;
1855 *e++ = output_signature_elements[i];
1856 }
1857 }
1858
1859 return WINED3D_OK;
1860 }
1861
1862 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1863 {
1864 struct wined3d_shader_indexable_temp *reg, *reg_next;
1865
1866 heap_free(reg_maps->constf);
1867 heap_free(reg_maps->sampler_map.entries);
1868
1869 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1870 heap_free(reg);
1871 list_init(&reg_maps->indexable_temps);
1872
1873 heap_free(reg_maps->tgsm);
1874 }
1875
1876 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1877 {
1878 DWORD map = 1u << max;
1879 map |= map - 1;
1880 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1881
1882 return wined3d_log2i(map);
1883 }
1884
1885 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1886 {
1887 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1888 {
1889 shader_addline(buffer, "refactoringAllowed");
1890 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1891 if (global_flags)
1892 shader_addline(buffer, " | ");
1893 }
1894
1895 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1896 {
1897 shader_addline(buffer, "forceEarlyDepthStencil");
1898 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1899 if (global_flags)
1900 shader_addline(buffer, " | ");
1901 }
1902
1903 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1904 {
1905 shader_addline(buffer, "enableRawAndStructuredBuffers");
1906 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1907 }
1908
1909 if (global_flags)
1910 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1911 }
1912
1913 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1914 {
1915 if (sync_flags & WINED3DSSF_GLOBAL_UAV)
1916 {
1917 shader_addline(buffer, "_uglobal");
1918 sync_flags &= ~WINED3DSSF_GLOBAL_UAV;
1919 }
1920 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1921 {
1922 shader_addline(buffer, "_g");
1923 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1924 }
1925 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1926 {
1927 shader_addline(buffer, "_t");
1928 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1929 }
1930
1931 if (sync_flags)
1932 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1933 }
1934
1935 static void shader_dump_precise_flags(struct wined3d_string_buffer *buffer, DWORD flags)
1936 {
1937 if (!(flags & WINED3DSI_PRECISE_XYZW))
1938 return;
1939
1940 shader_addline(buffer, " [precise");
1941 if (flags != WINED3DSI_PRECISE_XYZW)
1942 {
1943 shader_addline(buffer, "(%s%s%s%s)",
1944 flags & WINED3DSI_PRECISE_X ? "x" : "",
1945 flags & WINED3DSI_PRECISE_Y ? "y" : "",
1946 flags & WINED3DSI_PRECISE_Z ? "z" : "",
1947 flags & WINED3DSI_PRECISE_W ? "w" : "");
1948 }
1949 shader_addline(buffer, "]");
1950 }
1951
1952 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1953 {
1954 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1955 {
1956 shader_addline(buffer, "_glc");
1957 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1958 }
1959 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1960 {
1961 shader_addline(buffer, "_opc");
1962 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1963 }
1964
1965 if (uav_flags)
1966 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1967 }
1968
1969 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1970 enum wined3d_tessellator_domain domain)
1971 {
1972 switch (domain)
1973 {
1974 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1975 shader_addline(buffer, "line");
1976 break;
1977 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1978 shader_addline(buffer, "triangle");
1979 break;
1980 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1981 shader_addline(buffer, "quad");
1982 break;
1983 default:
1984 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1985 break;
1986 }
1987 }
1988
1989 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1990 enum wined3d_tessellator_output_primitive output_primitive)
1991 {
1992 switch (output_primitive)
1993 {
1994 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1995 shader_addline(buffer, "point");
1996 break;
1997 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1998 shader_addline(buffer, "line");
1999 break;
2000 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
2001 shader_addline(buffer, "triangle_cw");
2002 break;
2003 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
2004 shader_addline(buffer, "triangle_ccw");
2005 break;
2006 default:
2007 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
2008 break;
2009 }
2010 }
2011
2012 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
2013 enum wined3d_tessellator_partitioning partitioning)
2014 {
2015 switch (partitioning)
2016 {
2017 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
2018 shader_addline(buffer, "integer");
2019 break;
2020 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
2021 shader_addline(buffer, "pow2");
2022 break;
2023 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
2024 shader_addline(buffer, "fractional_odd");
2025 break;
2026 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
2027 shader_addline(buffer, "fractional_even");
2028 break;
2029 default:
2030 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
2031 break;
2032 }
2033 }
2034
2035 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
2036 enum wined3d_shader_input_sysval_semantic semantic)
2037 {
2038 unsigned int i;
2039
2040 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
2041 {
2042 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
2043 {
2044 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
2045 return;
2046 }
2047 }
2048
2049 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
2050 }
2051
2052 static void shader_dump_resource_type(struct wined3d_string_buffer *buffer, enum wined3d_shader_resource_type type)
2053 {
2054 static const char *const resource_type_names[] =
2055 {
2056 /* WINED3D_SHADER_RESOURCE_NONE */ "none",
2057 /* WINED3D_SHADER_RESOURCE_BUFFER */ "buffer",
2058 /* WINED3D_SHADER_RESOURCE_TEXTURE_1D */ "texture1d",
2059 /* WINED3D_SHADER_RESOURCE_TEXTURE_2D */ "texture2d",
2060 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DMS */ "texture2dms",
2061 /* WINED3D_SHADER_RESOURCE_TEXTURE_3D */ "texture3d",
2062 /* WINED3D_SHADER_RESOURCE_TEXTURE_CUBE */ "texturecube",
2063 /* WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY */ "texture1darray",
2064 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY */ "texture2darray",
2065 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY */ "texture2dmsarray",
2066 /* WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY */ "texturecubearray",
2067 };
2068
2069 if (type < ARRAY_SIZE(resource_type_names))
2070 shader_addline(buffer, "%s", resource_type_names[type]);
2071 else
2072 shader_addline(buffer, "unknown");
2073 }
2074
2075 static void shader_dump_data_type(struct wined3d_string_buffer *buffer, enum wined3d_data_type type)
2076 {
2077 static const char *const data_type_names[] =
2078 {
2079 /* WINED3D_DATA_FLOAT */ "(float)",
2080 /* WINED3D_DATA_INT */ "(int)",
2081 /* WINED3D_DATA_RESOURCE */ "(resource)",
2082 /* WINED3D_DATA_SAMPLER */ "(sampler)",
2083 /* WINED3D_DATA_UAV */ "(uav)",
2084 /* WINED3D_DATA_UINT */ "(uint)",
2085 /* WINED3D_DATA_UNORM */ "(unorm)",
2086 /* WINED3D_DATA_SNORM */ "(snorm)",
2087 /* WINED3D_DATA_OPAQUE */ "(opaque)",
2088 };
2089
2090 if (type < ARRAY_SIZE(data_type_names))
2091 shader_addline(buffer, "%s", data_type_names[type]);
2092 else
2093 shader_addline(buffer, "(unknown)");
2094 }
2095
2096 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
2097 const struct wined3d_shader_semantic *semantic, unsigned int flags,
2098 const struct wined3d_shader_version *shader_version)
2099 {
2100 shader_addline(buffer, "dcl");
2101
2102 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
2103 {
2104 switch (semantic->resource_type)
2105 {
2106 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2107 shader_addline(buffer, "_2d");
2108 break;
2109
2110 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2111 shader_addline(buffer, "_3d");
2112 break;
2113
2114 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2115 shader_addline(buffer, "_cube");
2116 break;
2117
2118 default:
2119 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2120 break;
2121 }
2122 }
2123 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2124 {
2125 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2126 shader_addline(buffer, "_resource_");
2127 else
2128 shader_addline(buffer, "_uav_");
2129 shader_dump_resource_type(buffer, semantic->resource_type);
2130 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2131 shader_dump_uav_flags(buffer, flags);
2132 shader_dump_data_type(buffer, semantic->resource_data_type);
2133 }
2134 else
2135 {
2136 /* Pixel shaders 3.0 don't have usage semantics. */
2137 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2138 return;
2139 else
2140 shader_addline(buffer, "_");
2141
2142 switch (semantic->usage)
2143 {
2144 case WINED3D_DECL_USAGE_POSITION:
2145 shader_addline(buffer, "position%u", semantic->usage_idx);
2146 break;
2147
2148 case WINED3D_DECL_USAGE_BLEND_INDICES:
2149 shader_addline(buffer, "blend");
2150 break;
2151
2152 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2153 shader_addline(buffer, "weight");
2154 break;
2155
2156 case WINED3D_DECL_USAGE_NORMAL:
2157 shader_addline(buffer, "normal%u", semantic->usage_idx);
2158 break;
2159
2160 case WINED3D_DECL_USAGE_PSIZE:
2161 shader_addline(buffer, "psize");
2162 break;
2163
2164 case WINED3D_DECL_USAGE_COLOR:
2165 if (!semantic->usage_idx)
2166 shader_addline(buffer, "color");
2167 else
2168 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2169 break;
2170
2171 case WINED3D_DECL_USAGE_TEXCOORD:
2172 shader_addline(buffer, "texture%u", semantic->usage_idx);
2173 break;
2174
2175 case WINED3D_DECL_USAGE_TANGENT:
2176 shader_addline(buffer, "tangent");
2177 break;
2178
2179 case WINED3D_DECL_USAGE_BINORMAL:
2180 shader_addline(buffer, "binormal");
2181 break;
2182
2183 case WINED3D_DECL_USAGE_TESS_FACTOR:
2184 shader_addline(buffer, "tessfactor");
2185 break;
2186
2187 case WINED3D_DECL_USAGE_POSITIONT:
2188 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2189 break;
2190
2191 case WINED3D_DECL_USAGE_FOG:
2192 shader_addline(buffer, "fog");
2193 break;
2194
2195 case WINED3D_DECL_USAGE_DEPTH:
2196 shader_addline(buffer, "depth");
2197 break;
2198
2199 case WINED3D_DECL_USAGE_SAMPLE:
2200 shader_addline(buffer, "sample");
2201 break;
2202
2203 default:
2204 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2205 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2206 }
2207 }
2208 }
2209
2210 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2211 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2212 {
2213 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2214 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2215 UINT offset = reg->idx[0].offset;
2216
2217 switch (reg->type)
2218 {
2219 case WINED3DSPR_TEMP:
2220 shader_addline(buffer, "r");
2221 break;
2222
2223 case WINED3DSPR_INPUT:
2224 shader_addline(buffer, "v");
2225 break;
2226
2227 case WINED3DSPR_CONST:
2228 case WINED3DSPR_CONST2:
2229 case WINED3DSPR_CONST3:
2230 case WINED3DSPR_CONST4:
2231 shader_addline(buffer, "c");
2232 offset = shader_get_float_offset(reg->type, offset);
2233 break;
2234
2235 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2236 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2237 break;
2238
2239 case WINED3DSPR_RASTOUT:
2240 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2241 break;
2242
2243 case WINED3DSPR_COLOROUT:
2244 shader_addline(buffer, "oC");
2245 break;
2246
2247 case WINED3DSPR_DEPTHOUT:
2248 shader_addline(buffer, "oDepth");
2249 break;
2250
2251 case WINED3DSPR_DEPTHOUTGE:
2252 shader_addline(buffer, "oDepthGE");
2253 break;
2254
2255 case WINED3DSPR_DEPTHOUTLE:
2256 shader_addline(buffer, "oDepthLE");
2257 break;
2258
2259 case WINED3DSPR_ATTROUT:
2260 shader_addline(buffer, "oD");
2261 break;
2262
2263 case WINED3DSPR_TEXCRDOUT:
2264 /* Vertex shaders >= 3.0 use general purpose output registers
2265 * (WINED3DSPR_OUTPUT), which can include an address token. */
2266 if (shader_version->major >= 3)
2267 shader_addline(buffer, "o");
2268 else
2269 shader_addline(buffer, "oT");
2270 break;
2271
2272 case WINED3DSPR_CONSTINT:
2273 shader_addline(buffer, "i");
2274 break;
2275
2276 case WINED3DSPR_CONSTBOOL:
2277 shader_addline(buffer, "b");
2278 break;
2279
2280 case WINED3DSPR_LABEL:
2281 shader_addline(buffer, "l");
2282 break;
2283
2284 case WINED3DSPR_LOOP:
2285 shader_addline(buffer, "aL");
2286 break;
2287
2288 case WINED3DSPR_SAMPLER:
2289 shader_addline(buffer, "s");
2290 break;
2291
2292 case WINED3DSPR_MISCTYPE:
2293 if (offset > 1)
2294 {
2295 FIXME("Unhandled misctype register %u.\n", offset);
2296 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2297 }
2298 else
2299 {
2300 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2301 }
2302 break;
2303
2304 case WINED3DSPR_PREDICATE:
2305 shader_addline(buffer, "p");
2306 break;
2307
2308 case WINED3DSPR_IMMCONST:
2309 shader_addline(buffer, "l");
2310 break;
2311
2312 case WINED3DSPR_CONSTBUFFER:
2313 shader_addline(buffer, "cb");
2314 break;
2315
2316 case WINED3DSPR_IMMCONSTBUFFER:
2317 shader_addline(buffer, "icb");
2318 break;
2319
2320 case WINED3DSPR_PRIMID:
2321 shader_addline(buffer, "primID");
2322 break;
2323
2324 case WINED3DSPR_NULL:
2325 shader_addline(buffer, "null");
2326 break;
2327
2328 case WINED3DSPR_RASTERIZER:
2329 shader_addline(buffer, "rasterizer");
2330 break;
2331
2332 case WINED3DSPR_RESOURCE:
2333 shader_addline(buffer, "t");
2334 break;
2335
2336 case WINED3DSPR_UAV:
2337 shader_addline(buffer, "u");
2338 break;
2339
2340 case WINED3DSPR_OUTPOINTID:
2341 shader_addline(buffer, "vOutputControlPointID");
2342 break;
2343
2344 case WINED3DSPR_FORKINSTID:
2345 shader_addline(buffer, "vForkInstanceId");
2346 break;
2347
2348 case WINED3DSPR_JOININSTID:
2349 shader_addline(buffer, "vJoinInstanceId");
2350 break;
2351
2352 case WINED3DSPR_INCONTROLPOINT:
2353 shader_addline(buffer, "vicp");
2354 break;
2355
2356 case WINED3DSPR_OUTCONTROLPOINT:
2357 shader_addline(buffer, "vocp");
2358 break;
2359
2360 case WINED3DSPR_PATCHCONST:
2361 shader_addline(buffer, "vpc");
2362 break;
2363
2364 case WINED3DSPR_TESSCOORD:
2365 shader_addline(buffer, "vDomainLocation");
2366 break;
2367
2368 case WINED3DSPR_GROUPSHAREDMEM:
2369 shader_addline(buffer, "g");
2370 break;
2371
2372 case WINED3DSPR_THREADID:
2373 shader_addline(buffer, "vThreadID");
2374 break;
2375
2376 case WINED3DSPR_THREADGROUPID:
2377 shader_addline(buffer, "vThreadGroupID");
2378 break;
2379
2380 case WINED3DSPR_LOCALTHREADID:
2381 shader_addline(buffer, "vThreadIDInGroup");
2382 break;
2383
2384 case WINED3DSPR_LOCALTHREADINDEX:
2385 shader_addline(buffer, "vThreadIDInGroupFlattened");
2386 break;
2387
2388 case WINED3DSPR_IDXTEMP:
2389 shader_addline(buffer, "x");
2390 break;
2391
2392 case WINED3DSPR_STREAM:
2393 shader_addline(buffer, "m");
2394 break;
2395
2396 case WINED3DSPR_FUNCTIONBODY:
2397 shader_addline(buffer, "fb");
2398 break;
2399
2400 case WINED3DSPR_FUNCTIONPOINTER:
2401 shader_addline(buffer, "fp");
2402 break;
2403
2404 case WINED3DSPR_COVERAGE:
2405 shader_addline(buffer, "vCoverage");
2406 break;
2407
2408 case WINED3DSPR_SAMPLEMASK:
2409 shader_addline(buffer, "oMask");
2410 break;
2411
2412 case WINED3DSPR_GSINSTID:
2413 shader_addline(buffer, "vGSInstanceID");
2414 break;
2415
2416 default:
2417 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2418 break;
2419 }
2420
2421 if (reg->type == WINED3DSPR_IMMCONST)
2422 {
2423 shader_addline(buffer, "(");
2424 switch (reg->immconst_type)
2425 {
2426 case WINED3D_IMMCONST_SCALAR:
2427 switch (reg->data_type)
2428 {
2429 case WINED3D_DATA_FLOAT:
2430 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2431 break;
2432 case WINED3D_DATA_INT:
2433 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2434 break;
2435 case WINED3D_DATA_RESOURCE:
2436 case WINED3D_DATA_SAMPLER:
2437 case WINED3D_DATA_UINT:
2438 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2439 break;
2440 default:
2441 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2442 break;
2443 }
2444 break;
2445
2446 case WINED3D_IMMCONST_VEC4:
2447 switch (reg->data_type)
2448 {
2449 case WINED3D_DATA_FLOAT:
2450 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2451 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2452 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2453 break;
2454 case WINED3D_DATA_INT:
2455 shader_addline(buffer, "%d, %d, %d, %d",
2456 reg->u.immconst_data[0], reg->u.immconst_data[1],
2457 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2458 break;
2459 case WINED3D_DATA_RESOURCE:
2460 case WINED3D_DATA_SAMPLER:
2461 case WINED3D_DATA_UINT:
2462 shader_addline(buffer, "%u, %u, %u, %u",
2463 reg->u.immconst_data[0], reg->u.immconst_data[1],
2464 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2465 break;
2466 default:
2467 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2468 break;
2469 }
2470 break;
2471
2472 default:
2473 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2474 break;
2475 }
2476 shader_addline(buffer, ")");
2477 }
2478 else if (reg->type != WINED3DSPR_RASTOUT
2479 && reg->type != WINED3DSPR_MISCTYPE
2480 && reg->type != WINED3DSPR_NULL)
2481 {
2482 if (offset != ~0u)
2483 {
2484 shader_addline(buffer, "[");
2485 if (reg->idx[0].rel_addr)
2486 {
2487 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2488 shader_addline(buffer, " + ");
2489 }
2490 shader_addline(buffer, "%u]", offset);
2491
2492 if (reg->idx[1].offset != ~0u)
2493 {
2494 shader_addline(buffer, "[");
2495 if (reg->idx[1].rel_addr)
2496 {
2497 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2498 shader_addline(buffer, " + ");
2499 }
2500 shader_addline(buffer, "%u]", reg->idx[1].offset);
2501 }
2502 }
2503
2504 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2505 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2506 }
2507 }
2508
2509 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2510 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2511 {
2512 DWORD write_mask = param->write_mask;
2513
2514 shader_dump_register(buffer, &param->reg, shader_version);
2515
2516 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2517 {
2518 static const char write_mask_chars[] = "xyzw";
2519
2520 shader_addline(buffer, ".");
2521 if (write_mask & WINED3DSP_WRITEMASK_0)
2522 shader_addline(buffer, "%c", write_mask_chars[0]);
2523 if (write_mask & WINED3DSP_WRITEMASK_1)
2524 shader_addline(buffer, "%c", write_mask_chars[1]);
2525 if (write_mask & WINED3DSP_WRITEMASK_2)
2526 shader_addline(buffer, "%c", write_mask_chars[2]);
2527 if (write_mask & WINED3DSP_WRITEMASK_3)
2528 shader_addline(buffer, "%c", write_mask_chars[3]);
2529 }
2530 }
2531
2532 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2533 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2534 {
2535 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2536 DWORD swizzle = param->swizzle;
2537
2538 if (src_modifier == WINED3DSPSM_NEG
2539 || src_modifier == WINED3DSPSM_BIASNEG
2540 || src_modifier == WINED3DSPSM_SIGNNEG
2541 || src_modifier == WINED3DSPSM_X2NEG
2542 || src_modifier == WINED3DSPSM_ABSNEG)
2543 shader_addline(buffer, "-");
2544 else if (src_modifier == WINED3DSPSM_COMP)
2545 shader_addline(buffer, "1-");
2546 else if (src_modifier == WINED3DSPSM_NOT)
2547 shader_addline(buffer, "!");
2548
2549 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2550 shader_addline(buffer, "abs(");
2551
2552 shader_dump_register(buffer, &param->reg, shader_version);
2553
2554 switch (src_modifier)
2555 {
2556 case WINED3DSPSM_NONE: break;
2557 case WINED3DSPSM_NEG: break;
2558 case WINED3DSPSM_NOT: break;
2559 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2560 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2561 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2562 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2563 case WINED3DSPSM_COMP: break;
2564 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2565 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2566 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2567 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2568 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2569 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2570 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2571 }
2572
2573 if (swizzle != WINED3DSP_NOSWIZZLE)
2574 {
2575 static const char swizzle_chars[] = "xyzw";
2576 DWORD swizzle_x = swizzle & 0x03;
2577 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2578 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2579 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2580
2581 if (swizzle_x == swizzle_y
2582 && swizzle_x == swizzle_z
2583 && swizzle_x == swizzle_w)
2584 {
2585 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2586 }
2587 else
2588 {
2589 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2590 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2591 }
2592 }
2593 }
2594
2595 /* Shared code in order to generate the bulk of the shader string. */
2596 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2597 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2598 const DWORD *start, const DWORD *end)
2599 {
2600 struct wined3d_device *device = shader->device;
2601 const struct wined3d_shader_frontend *fe = shader->frontend;
2602 void *fe_data = shader->frontend_data;
2603 struct wined3d_shader_version shader_version;
2604 struct wined3d_shader_parser_state state;
2605 struct wined3d_shader_instruction ins;
2606 struct wined3d_shader_tex_mx tex_mx;
2607 struct wined3d_shader_context ctx;
2608 const DWORD *ptr;
2609
2610 /* Initialize current parsing state. */
2611 tex_mx.current_row = 0;
2612 state.current_loop_depth = 0;
2613 state.current_loop_reg = 0;
2614 state.in_subroutine = FALSE;
2615
2616 ctx.shader = shader;
2617 ctx.reg_maps = reg_maps;
2618 ctx.buffer = buffer;
2619 ctx.tex_mx = &tex_mx;
2620 ctx.state = &state;
2621 ctx.backend_data = backend_ctx;
2622 ins.ctx = &ctx;
2623
2624 fe->shader_read_header(fe_data, &ptr, &shader_version);
2625 if (start)
2626 ptr = start;
2627
2628 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2629 {
2630 /* Read opcode. */
2631 fe->shader_read_instruction(fe_data, &ptr, &ins);
2632
2633 /* Unknown opcode and its parameters. */
2634 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2635 {
2636 WARN("Encountered unrecognised or invalid instruction.\n");
2637 return WINED3DERR_INVALIDCALL;
2638 }
2639
2640 if (ins.predicate)
2641 FIXME("Predicates not implemented.\n");
2642
2643 /* Call appropriate function for output target */
2644 device->shader_backend->shader_handle_instruction(&ins);
2645 }
2646
2647 return WINED3D_OK;
2648 }
2649
2650 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2651 const struct wined3d_shader_dst_param *dst)
2652 {
2653 DWORD mmask = dst->modifiers;
2654
2655 switch (dst->shift)
2656 {
2657 case 0: break;
2658 case 13: shader_addline(buffer, "_d8"); break;
2659 case 14: shader_addline(buffer, "_d4"); break;
2660 case 15: shader_addline(buffer, "_d2"); break;
2661 case 1: shader_addline(buffer, "_x2"); break;
2662 case 2: shader_addline(buffer, "_x4"); break;
2663 case 3: shader_addline(buffer, "_x8"); break;
2664 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2665 }
2666
2667 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2668 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2669 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2670
2671 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2672 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2673 }
2674
2675 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2676 const struct wined3d_shader_primitive_type *primitive_type)
2677 {
2678 switch (primitive_type->type)
2679 {
2680 case WINED3D_PT_UNDEFINED:
2681 shader_addline(buffer, "undefined");
2682 break;
2683 case WINED3D_PT_POINTLIST:
2684 shader_addline(buffer, "pointlist");
2685 break;
2686 case WINED3D_PT_LINELIST:
2687 shader_addline(buffer, "linelist");
2688 break;
2689 case WINED3D_PT_LINESTRIP:
2690 shader_addline(buffer, "linestrip");
2691 break;
2692 case WINED3D_PT_TRIANGLELIST:
2693 shader_addline(buffer, "trianglelist");
2694 break;
2695 case WINED3D_PT_TRIANGLESTRIP:
2696 shader_addline(buffer, "trianglestrip");
2697 break;
2698 case WINED3D_PT_TRIANGLEFAN:
2699 shader_addline(buffer, "trianglefan");
2700 break;
2701 case WINED3D_PT_LINELIST_ADJ:
2702 shader_addline(buffer, "linelist_adj");
2703 break;
2704 case WINED3D_PT_LINESTRIP_ADJ:
2705 shader_addline(buffer, "linestrip_adj");
2706 break;
2707 case WINED3D_PT_TRIANGLELIST_ADJ:
2708 shader_addline(buffer, "trianglelist_adj");
2709 break;
2710 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2711 shader_addline(buffer, "trianglestrip_adj");
2712 break;
2713 case WINED3D_PT_PATCH:
2714 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2715 break;
2716 default:
2717 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2718 break;
2719 }
2720 }
2721
2722 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2723 enum wined3d_shader_interpolation_mode interpolation_mode)
2724 {
2725 switch (interpolation_mode)
2726 {
2727 case WINED3DSIM_CONSTANT:
2728 shader_addline(buffer, "constant");
2729 break;
2730 case WINED3DSIM_LINEAR:
2731 shader_addline(buffer, "linear");
2732 break;
2733 case WINED3DSIM_LINEAR_CENTROID:
2734 shader_addline(buffer, "linear centroid");
2735 break;
2736 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2737 shader_addline(buffer, "linear noperspective");
2738 break;
2739 case WINED3DSIM_LINEAR_SAMPLE:
2740 shader_addline(buffer, "linear sample");
2741 break;
2742 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2743 shader_addline(buffer, "linear noperspective centroid");
2744 break;
2745 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2746 shader_addline(buffer, "linear noperspective sample");
2747 break;
2748 default:
2749 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2750 break;
2751 }
2752 }
2753
2754 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2755 {
2756 struct wined3d_shader_version shader_version;
2757 struct wined3d_string_buffer buffer;
2758 const char *type_prefix;
2759 const char *p, *q;
2760 const DWORD *ptr;
2761 DWORD i;
2762
2763 if (!string_buffer_init(&buffer))
2764 {
2765 ERR("Failed to initialize string buffer.\n");
2766 return;
2767 }
2768
2769 fe->shader_read_header(fe_data, &ptr, &shader_version);
2770
2771 TRACE("Parsing %p.\n", ptr);
2772
2773 switch (shader_version.type)
2774 {
2775 case WINED3D_SHADER_TYPE_VERTEX:
2776 type_prefix = "vs";
2777 break;
2778
2779 case WINED3D_SHADER_TYPE_HULL:
2780 type_prefix = "hs";
2781 break;
2782
2783 case WINED3D_SHADER_TYPE_DOMAIN:
2784 type_prefix = "ds";
2785 break;
2786
2787 case WINED3D_SHADER_TYPE_GEOMETRY:
2788 type_prefix = "gs";
2789 break;
2790
2791 case WINED3D_SHADER_TYPE_PIXEL:
2792 type_prefix = "ps";
2793 break;
2794
2795 case WINED3D_SHADER_TYPE_COMPUTE:
2796 type_prefix = "cs";
2797 break;
2798
2799 default:
2800 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2801 type_prefix = "unknown";
2802 break;
2803 }
2804
2805 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2806
2807 while (!fe->shader_is_end(fe_data, &ptr))
2808 {
2809 struct wined3d_shader_instruction ins;
2810
2811 fe->shader_read_instruction(fe_data, &ptr, &ins);
2812 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2813 {
2814 WARN("Skipping unrecognized instruction.\n");
2815 shader_addline(&buffer, "<unrecognized instruction>\n");
2816 continue;
2817 }
2818
2819 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2820 {
2821 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2822 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2823 shader_addline(&buffer, " ");
2824 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2825 }
2826 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2827 {
2828 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2829 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2830 shader_addline(&buffer, ", %s",
2831 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2832 }
2833 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2834 {
2835 shader_addline(&buffer, "%s fb%u",
2836 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2837 }
2838 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2839 {
2840 shader_addline(&buffer, "%s ft%u = {...}",
2841 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2842 }
2843 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2844 {
2845 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2846 shader_dump_global_flags(&buffer, ins.flags);
2847 }
2848 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2849 {
2850 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2851 ins.declaration.max_tessellation_factor);
2852 }
2853 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2854 {
2855 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2856 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2857 {
2858 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2859 ins.declaration.icb->data[4 * i + 0],
2860 ins.declaration.icb->data[4 * i + 1],
2861 ins.declaration.icb->data[4 * i + 2],
2862 ins.declaration.icb->data[4 * i + 3]);
2863 }
2864 shader_addline(&buffer, "}");
2865 }
2866 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2867 {
2868 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2869 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2870 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2871 }
2872 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2873 {
2874 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2875 ins.declaration.indexable_temp.register_idx,
2876 ins.declaration.indexable_temp.register_size,
2877 ins.declaration.indexable_temp.component_count);
2878 }
2879 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2880 {
2881 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2882 shader_dump_interpolation_mode(&buffer, ins.flags);
2883 shader_addline(&buffer, " ");
2884 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2885 }
2886 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2887 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2888 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2889 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2890 {
2891 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2892 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2893 shader_addline(&buffer, ", ");
2894 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2895 }
2896 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2897 {
2898 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2899 shader_dump_interpolation_mode(&buffer, ins.flags);
2900 shader_addline(&buffer, " ");
2901 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2902 shader_addline(&buffer, ", ");
2903 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2904 }
2905 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2906 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2907 {
2908 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2909 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2910 }
2911 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2912 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2913 {
2914 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2915 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2916 }
2917 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2918 {
2919 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2920 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2921 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2922 }
2923 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2924 {
2925 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2926 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2927 }
2928 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2929 {
2930 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2931 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2932 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2933 }
2934 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2935 {
2936 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2937 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2938 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2939 shader_addline(&buffer, ", comparisonMode");
2940 }
2941 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2942 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2943 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2944 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2945 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2946 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2947 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2948 {
2949 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2950 }
2951 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2952 {
2953 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2954 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2955 }
2956 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2957 {
2958 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2959 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2960 }
2961 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2962 {
2963 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2964 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2965 }
2966 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2967 {
2968 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2969 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2970 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2971 }
2972 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2973 {
2974 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2975 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2976 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2977 ins.declaration.tgsm_structured.structure_count);
2978 }
2979 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2980 {
2981 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2982 ins.declaration.thread_group_size.x,
2983 ins.declaration.thread_group_size.y,
2984 ins.declaration.thread_group_size.z);
2985 }
2986 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2987 {
2988 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2989 shader_dump_uav_flags(&buffer, ins.flags);
2990 shader_addline(&buffer, " ");
2991 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2992 }
2993 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2994 {
2995 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2996 shader_dump_uav_flags(&buffer, ins.flags);
2997 shader_addline(&buffer, " ");
2998 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2999 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
3000 }
3001 else if (ins.handler_idx == WINED3DSIH_DEF)
3002 {
3003 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
3004 ins.dst[0].reg.idx[0].offset),
3005 *(const float *)&ins.src[0].reg.u.immconst_data[0],
3006 *(const float *)&ins.src[0].reg.u.immconst_data[1],
3007 *(const float *)&ins.src[0].reg.u.immconst_data[2],
3008 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
3009 }
3010 else if (ins.handler_idx == WINED3DSIH_DEFI)
3011 {
3012 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
3013 ins.src[0].reg.u.immconst_data[0],
3014 ins.src[0].reg.u.immconst_data[1],
3015 ins.src[0].reg.u.immconst_data[2],
3016 ins.src[0].reg.u.immconst_data[3]);
3017 }
3018 else if (ins.handler_idx == WINED3DSIH_DEFB)
3019 {
3020 shader_addline(&buffer, "defb b%u = %s",
3021 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
3022 }
3023 else
3024 {
3025 if (ins.predicate)
3026 {
3027 shader_addline(&buffer, "(");
3028 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
3029 shader_addline(&buffer, ") ");
3030 }
3031
3032 /* PixWin marks instructions with the coissue flag with a '+' */
3033 if (ins.coissue)
3034 shader_addline(&buffer, "+");
3035
3036 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3037
3038 if (ins.handler_idx == WINED3DSIH_BREAKP
3039 || ins.handler_idx == WINED3DSIH_CONTINUEP
3040 || ins.handler_idx == WINED3DSIH_IF
3041 || ins.handler_idx == WINED3DSIH_RETP
3042 || ins.handler_idx == WINED3DSIH_TEXKILL)
3043 {
3044 switch (ins.flags)
3045 {
3046 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3047 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3048 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3049 }
3050 }
3051 else if (ins.handler_idx == WINED3DSIH_IFC
3052 || ins.handler_idx == WINED3DSIH_BREAKC)
3053 {
3054 switch (ins.flags)
3055 {
3056 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3057 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3058 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3059 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3060 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3061 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3062 default: shader_addline(&buffer, "_(%u)", ins.flags);
3063 }
3064 }
3065 else if (ins.handler_idx == WINED3DSIH_TEX
3066 && shader_version.major >= 2
3067 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3068 {
3069 shader_addline(&buffer, "p");
3070 }
3071 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3072 {
3073 switch (ins.flags)
3074 {
3075 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3076 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3077 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3078 }
3079 }
3080 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3081 {
3082 switch (ins.flags)
3083 {
3084 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3085 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3086 }
3087 }
3088 else if (ins.handler_idx == WINED3DSIH_SYNC)
3089 {
3090 shader_dump_sync_flags(&buffer, ins.flags);
3091 }
3092 else
3093 {
3094 shader_dump_precise_flags(&buffer, ins.flags);
3095 }
3096
3097 if (wined3d_shader_instruction_has_texel_offset(&ins))
3098 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3099
3100 if (ins.resource_type != WINED3D_SHADER_RESOURCE_NONE)
3101 {
3102 shader_addline(&buffer, "(");
3103 shader_dump_resource_type(&buffer, ins.resource_type);
3104 shader_addline(&buffer, ")");
3105 }
3106
3107 if (ins.resource_data_type != WINED3D_DATA_FLOAT)
3108 shader_dump_data_type(&buffer, ins.resource_data_type);
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
3148 heap_free(shader->patch_constant_signature.elements);
3149 heap_free(shader->output_signature.elements);
3150 heap_free(shader->input_signature.elements);
3151 shader->device->shader_backend->shader_destroy(shader);
3152 shader_cleanup_reg_maps(&shader->reg_maps);
3153 heap_free(shader->byte_code);
3154 shader_delete_constant_list(&shader->constantsF);
3155 shader_delete_constant_list(&shader->constantsB);
3156 shader_delete_constant_list(&shader->constantsI);
3157 list_remove(&shader->shader_list_entry);
3158
3159 if (shader->frontend && shader->frontend_data)
3160 shader->frontend->shader_free(shader->frontend_data);
3161 }
3162
3163 struct shader_none_priv
3164 {
3165 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3166 const struct wined3d_fragment_pipe_ops *fragment_pipe;
3167 BOOL ffp_proj_control;
3168 };
3169
3170 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3171 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3172 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3173 const struct wined3d_state *state) {}
3174 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3175 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3176 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3177 const struct wined3d_state *state) {}
3178 static void shader_none_destroy(struct wined3d_shader *shader) {}
3179 static void shader_none_free_context_data(struct wined3d_context *context) {}
3180 static void shader_none_init_context_state(struct wined3d_context *context) {}
3181
3182 /* Context activation is done by the caller. */
3183 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3184 const struct wined3d_state *state)
3185 {
3186 struct shader_none_priv *priv = shader_priv;
3187
3188 priv->vertex_pipe->vp_enable(context, !use_vs(state));
3189 priv->fragment_pipe->fp_enable(context, !use_ps(state));
3190 }
3191
3192 /* Context activation is done by the caller. */
3193 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3194 {
3195 struct shader_none_priv *priv = shader_priv;
3196
3197 priv->vertex_pipe->vp_enable(context, FALSE);
3198 priv->fragment_pipe->fp_enable(context, FALSE);
3199
3200 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3201 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3202 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3203 | (1u << WINED3D_SHADER_TYPE_HULL)
3204 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3205 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3206 }
3207
3208 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3209 const struct wined3d_fragment_pipe_ops *fragment_pipe)
3210 {
3211 struct fragment_caps fragment_caps;
3212 void *vertex_priv, *fragment_priv;
3213 struct shader_none_priv *priv;
3214
3215 if (!(priv = heap_alloc(sizeof(*priv))))
3216 return E_OUTOFMEMORY;
3217
3218 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3219 {
3220 ERR("Failed to initialize vertex pipe.\n");
3221 heap_free(priv);
3222 return E_FAIL;
3223 }
3224
3225 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3226 {
3227 ERR("Failed to initialize fragment pipe.\n");
3228 vertex_pipe->vp_free(device, NULL);
3229 heap_free(priv);
3230 return E_FAIL;
3231 }
3232
3233 priv->vertex_pipe = vertex_pipe;
3234 priv->fragment_pipe = fragment_pipe;
3235 fragment_pipe->get_caps(device->adapter, &fragment_caps);
3236 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3237
3238 device->vertex_priv = vertex_priv;
3239 device->fragment_priv = fragment_priv;
3240 device->shader_priv = priv;
3241
3242 return WINED3D_OK;
3243 }
3244
3245 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
3246 {
3247 struct shader_none_priv *priv = device->shader_priv;
3248
3249 priv->fragment_pipe->free_private(device, context);
3250 priv->vertex_pipe->vp_free(device, context);
3251 heap_free(priv);
3252 }
3253
3254 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3255 {
3256 return TRUE;
3257 }
3258
3259 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
3260 {
3261 /* Set the shader caps to 0 for the none shader backend */
3262 memset(caps, 0, sizeof(*caps));
3263 }
3264
3265 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3266 {
3267 /* We "support" every possible fixup, since we don't support any shader
3268 * model, and will never have to actually sample a texture. */
3269 return TRUE;
3270 }
3271
3272 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3273 {
3274 struct shader_none_priv *priv = shader_priv;
3275
3276 return priv->ffp_proj_control;
3277 }
3278
3279 static uint64_t shader_none_shader_compile(struct wined3d_context *context, const struct wined3d_shader_desc *shader_desc,
3280 enum wined3d_shader_type shader_type)
3281 {
3282 return 0;
3283 }
3284
3285 const struct wined3d_shader_backend_ops none_shader_backend =
3286 {
3287 shader_none_handle_instruction,
3288 shader_none_precompile,
3289 shader_none_select,
3290 shader_none_select_compute,
3291 shader_none_disable,
3292 shader_none_update_float_vertex_constants,
3293 shader_none_update_float_pixel_constants,
3294 shader_none_load_constants,
3295 shader_none_destroy,
3296 shader_none_alloc,
3297 shader_none_free,
3298 shader_none_allocate_context_data,
3299 shader_none_free_context_data,
3300 shader_none_init_context_state,
3301 shader_none_get_caps,
3302 shader_none_color_fixup_supported,
3303 shader_none_has_ffp_proj_control,
3304 shader_none_shader_compile,
3305 };
3306
3307 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
3308 {
3309 switch (level)
3310 {
3311 case WINED3D_FEATURE_LEVEL_11_1:
3312 case WINED3D_FEATURE_LEVEL_11:
3313 return 5;
3314 case WINED3D_FEATURE_LEVEL_10_1:
3315 case WINED3D_FEATURE_LEVEL_10:
3316 return 4;
3317 case WINED3D_FEATURE_LEVEL_9_3:
3318 return 3;
3319 case WINED3D_FEATURE_LEVEL_9_2:
3320 case WINED3D_FEATURE_LEVEL_9_1:
3321 return 2;
3322 default:
3323 return 1;
3324 }
3325 }
3326
3327 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
3328 enum wined3d_shader_type type, unsigned int float_const_count)
3329 {
3330 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3331 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3332 const struct wined3d_shader_version *version = &reg_maps->shader_version;
3333 const struct wined3d_shader_frontend *fe;
3334 unsigned int backend_version;
3335 HRESULT hr;
3336
3337 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
3338 shader, device, debug_shader_type(type), float_const_count);
3339
3340 fe = shader->frontend;
3341 if (!(shader->frontend_data = fe->shader_init(shader->function,
3342 shader->functionLength, &shader->output_signature)))
3343 {
3344 FIXME("Failed to initialize frontend.\n");
3345 return WINED3DERR_INVALIDCALL;
3346 }
3347
3348 /* First pass: trace shader. */
3349 if (TRACE_ON(d3d_shader))
3350 shader_trace_init(fe, shader->frontend_data);
3351
3352 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3353 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
3354 return hr;
3355
3356 if (version->type != type)
3357 {
3358 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
3359 return WINED3DERR_INVALIDCALL;
3360 }
3361 if (version->major > shader_max_version_from_feature_level(device->cs->c.state->feature_level))
3362 {
3363 WARN("Shader version %u not supported by this device.\n", version->major);
3364 return WINED3DERR_INVALIDCALL;
3365 }
3366 switch (type)
3367 {
3368 case WINED3D_SHADER_TYPE_VERTEX:
3369 backend_version = d3d_info->limits.vs_version;
3370 break;
3371 case WINED3D_SHADER_TYPE_HULL:
3372 backend_version = d3d_info->limits.hs_version;
3373 break;
3374 case WINED3D_SHADER_TYPE_DOMAIN:
3375 backend_version = d3d_info->limits.ds_version;
3376 break;
3377 case WINED3D_SHADER_TYPE_GEOMETRY:
3378 backend_version = d3d_info->limits.gs_version;
3379 break;
3380 case WINED3D_SHADER_TYPE_PIXEL:
3381 backend_version = d3d_info->limits.ps_version;
3382 break;
3383 case WINED3D_SHADER_TYPE_COMPUTE:
3384 backend_version = d3d_info->limits.cs_version;
3385 break;
3386 default:
3387 FIXME("No backend version-checking for this shader type.\n");
3388 backend_version = 0;
3389 }
3390 if (version->major > backend_version)
3391 {
3392 WARN("Shader version %u.%u not supported by the current shader backend.\n",
3393 version->major, version->minor);
3394 return WINED3DERR_INVALIDCALL;
3395 }
3396
3397 shader->load_local_constsF = shader->lconst_inf_or_nan;
3398
3399 return WINED3D_OK;
3400 }
3401
3402 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3403 {
3404 ULONG refcount = InterlockedIncrement(&shader->ref);
3405
3406 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3407
3408 return refcount;
3409 }
3410
3411 static void wined3d_shader_init_object(void *object)
3412 {
3413 struct wined3d_shader *shader = object;
3414 struct wined3d_device *device = shader->device;
3415
3416 TRACE("shader %p.\n", shader);
3417
3418 list_add_head(&device->shaders, &shader->shader_list_entry);
3419
3420 device->shader_backend->shader_precompile(device->shader_priv, shader);
3421 }
3422
3423 static void wined3d_shader_destroy_object(void *object)
3424 {
3425 TRACE("object %p.\n", object);
3426
3427 shader_cleanup(object);
3428 heap_free(object);
3429 }
3430
3431 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3432 {
3433 ULONG refcount = InterlockedDecrement(&shader->ref);
3434
3435 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3436
3437 if (!refcount)
3438 {
3439 wined3d_mutex_lock();
3440 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3441 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3442 wined3d_mutex_unlock();
3443 }
3444
3445 return refcount;
3446 }
3447
3448 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3449 {
3450 TRACE("shader %p.\n", shader);
3451
3452 return shader->parent;
3453 }
3454
3455 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3456 void *byte_code, UINT *byte_code_size)
3457 {
3458 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3459
3460 if (!byte_code)
3461 {
3462 *byte_code_size = shader->byte_code_size;
3463 return WINED3D_OK;
3464 }
3465
3466 if (*byte_code_size < shader->byte_code_size)
3467 {
3468 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3469 * than the required size we should write the required size and
3470 * return D3DERR_MOREDATA. That's not actually true. */
3471 return WINED3DERR_INVALIDCALL;
3472 }
3473
3474 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
3475
3476 return WINED3D_OK;
3477 }
3478
3479 /* Set local constants for d3d8 shaders. */
3480 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3481 UINT start_idx, const float *src_data, UINT count)
3482 {
3483 UINT end_idx = start_idx + count;
3484 UINT i;
3485
3486 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3487
3488 if (end_idx > shader->limits->constant_float)
3489 {
3490 WARN("end_idx %u > float constants limit %u.\n",
3491 end_idx, shader->limits->constant_float);
3492 end_idx = shader->limits->constant_float;
3493 }
3494
3495 for (i = start_idx; i < end_idx; ++i)
3496 {
3497 struct wined3d_shader_lconst *lconst;
3498 float *value;
3499
3500 if (!(lconst = heap_alloc(sizeof(*lconst))))
3501 return E_OUTOFMEMORY;
3502
3503 lconst->idx = i;
3504 value = (float *)lconst->value;
3505 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3506 list_add_head(&shader->constantsF, &lconst->entry);
3507
3508 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3509 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3510 {
3511 shader->lconst_inf_or_nan = TRUE;
3512 }
3513 }
3514
3515 return WINED3D_OK;
3516 }
3517
3518 static void init_interpolation_compile_args(DWORD *interpolation_args,
3519 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
3520 {
3521 if (!d3d_info->shader_output_interpolation || !pixel_shader
3522 || pixel_shader->reg_maps.shader_version.major < 4)
3523 {
3524 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3525 return;
3526 }
3527
3528 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3529 sizeof(pixel_shader->u.ps.interpolation_mode));
3530 }
3531
3532 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3533 struct vs_compile_args *args, const struct wined3d_context *context)
3534 {
3535 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3536 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3537 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3538 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3539 WORD swizzle_map = context->stream_info.swizzle_map;
3540
3541 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3542 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3543 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3544 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3545 args->point_size = state->primitive_type == WINED3D_PT_POINTLIST;
3546 args->per_vertex_point_size = shader->reg_maps.point_size;
3547 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
3548 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3549 if (shader->reg_maps.shader_version.major >= 4)
3550 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3551 : geometry_shader ? geometry_shader->limits->packed_input
3552 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3553 else
3554 args->next_shader_input_count = 0;
3555 args->swizzle_map = swizzle_map;
3556 if (d3d_info->emulated_flatshading)
3557 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3558 else
3559 args->flatshading = 0;
3560
3561 init_interpolation_compile_args(args->interpolation_mode,
3562 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
3563 }
3564
3565 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3566 {
3567 if (usage_idx1 != usage_idx2)
3568 return FALSE;
3569 if (usage1 == usage2)
3570 return TRUE;
3571 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3572 return TRUE;
3573 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3574 return TRUE;
3575
3576 return FALSE;
3577 }
3578
3579 bool vshader_get_input(const struct wined3d_shader *shader,
3580 uint8_t usage_req, uint8_t usage_idx_req, unsigned int *regnum)
3581 {
3582 uint32_t map = shader->reg_maps.input_registers & 0xffff;
3583 unsigned int i;
3584
3585 while (map)
3586 {
3587 i = wined3d_bit_scan(&map);
3588 if (match_usage(shader->u.vs.attributes[i].usage,
3589 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3590 {
3591 *regnum = i;
3592 return true;
3593 }
3594 }
3595
3596 return false;
3597 }
3598
3599 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3600 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3601 {
3602 HRESULT hr;
3603
3604 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
3605
3606 if (!desc->byte_code)
3607 return WINED3DERR_INVALIDCALL;
3608
3609 shader->ref = 1;
3610 shader->device = device;
3611 shader->parent = parent;
3612 shader->parent_ops = parent_ops;
3613
3614 list_init(&shader->linked_programs);
3615 list_init(&shader->constantsF);
3616 list_init(&shader->constantsB);
3617 list_init(&shader->constantsI);
3618 shader->lconst_inf_or_nan = FALSE;
3619 list_init(&shader->reg_maps.indexable_temps);
3620 list_init(&shader->shader_list_entry);
3621
3622 if (desc->byte_code_size == ~(size_t)0)
3623 {
3624 struct wined3d_shader_version shader_version;
3625 const struct wined3d_shader_frontend *fe;
3626 struct wined3d_shader_instruction ins;
3627 const DWORD *ptr;
3628 void *fe_data;
3629
3630 if (!(shader->frontend = shader_select_frontend(WINED3D_SHADER_BYTE_CODE_FORMAT_SM1)))
3631 {
3632 FIXME("Unable to find frontend for shader.\n");
3633 hr = WINED3DERR_INVALIDCALL;
3634 goto fail;
3635 }
3636
3637 fe = shader->frontend;
3638 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
3639 {
3640 WARN("Failed to initialise frontend data.\n");
3641 hr = WINED3DERR_INVALIDCALL;
3642 goto fail;
3643 }
3644
3645 fe->shader_read_header(fe_data, &ptr, &shader_version);
3646 while (!fe->shader_is_end(fe_data, &ptr))
3647 fe->shader_read_instruction(fe_data, &ptr, &ins);
3648
3649 fe->shader_free(fe_data);
3650
3651 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3652
3653 if (!(shader->byte_code = heap_alloc(shader->byte_code_size)))
3654 {
3655 hr = E_OUTOFMEMORY;
3656 goto fail;
3657 }
3658 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
3659
3660 shader->function = shader->byte_code;
3661 shader->functionLength = shader->byte_code_size;
3662 }
3663 else
3664 {
3665 enum wined3d_shader_byte_code_format format;
3666 unsigned int max_version;
3667
3668 if (!(shader->byte_code = heap_alloc(desc->byte_code_size)))
3669 {
3670 hr = E_OUTOFMEMORY;
3671 goto fail;
3672 }
3673 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
3674 shader->byte_code_size = desc->byte_code_size;
3675
3676 max_version = shader_max_version_from_feature_level(device->cs->c.state->feature_level);
3677 if (FAILED(hr = shader_extract_from_dxbc(shader, max_version, &format)))
3678 goto fail;
3679
3680 if (!(shader->frontend = shader_select_frontend(format)))
3681 {
3682 FIXME("Unable to find frontend for shader.\n");
3683 hr = WINED3DERR_INVALIDCALL;
3684 goto fail;
3685 }
3686 }
3687
3688 return WINED3D_OK;
3689
3690 fail:
3691 shader_cleanup(shader);
3692 return hr;
3693 }
3694
3695 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3696 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3697 {
3698 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3699 unsigned int i;
3700 HRESULT hr;
3701
3702 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3703 return hr;
3704
3705 if (FAILED(hr = shader_set_function(shader, device,
3706 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
3707 {
3708 shader_cleanup(shader);
3709 return hr;
3710 }
3711
3712 for (i = 0; i < shader->input_signature.element_count; ++i)
3713 {
3714 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3715
3716 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3717 continue;
3718
3719 shader->u.vs.attributes[input->register_idx].usage =
3720 shader_usage_from_semantic_name(input->semantic_name);
3721 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3722 }
3723
3724 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3725 shader->load_local_constsF = TRUE;
3726
3727 return WINED3D_OK;
3728 }
3729
3730 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
3731 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
3732 {
3733 struct wined3d_shader_signature_element *e = s->elements;
3734 unsigned int i;
3735
3736 for (i = 0; i < s->element_count; ++i)
3737 {
3738 if (e[i].stream_idx == stream_idx
3739 && !stricmp(e[i].semantic_name, semantic_name)
3740 && e[i].semantic_idx == semantic_idx)
3741 return &e[i];
3742 }
3743
3744 return NULL;
3745 }
3746
3747 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
3748 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
3749 {
3750 const struct wined3d_shader_signature_element *output;
3751 unsigned int idx;
3752
3753 if (!(output = shader_find_signature_element(&shader->output_signature,
3754 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
3755 return FALSE;
3756
3757 for (idx = 0; idx < 4; ++idx)
3758 {
3759 if (output->mask & (1u << idx))
3760 break;
3761 }
3762 idx += so_element->component_idx;
3763
3764 *register_idx = output->register_idx;
3765 *component_idx = idx;
3766 return TRUE;
3767 }
3768
3769 static HRESULT geometry_shader_init_so_desc(struct wined3d_geometry_shader *gs, struct wined3d_device *device,
3770 const struct wined3d_stream_output_desc *so_desc)
3771 {
3772 struct wined3d_so_desc_entry *s;
3773 struct wine_rb_entry *entry;
3774 unsigned int i;
3775 size_t size;
3776 char *name;
3777
3778 if ((entry = wine_rb_get(&device->so_descs, so_desc)))
3779 {
3780 gs->so_desc = &WINE_RB_ENTRY_VALUE(entry, struct wined3d_so_desc_entry, entry)->desc;
3781 return WINED3D_OK;
3782 }
3783
3784 size = FIELD_OFFSET(struct wined3d_so_desc_entry, elements[so_desc->element_count]);
3785 for (i = 0; i < so_desc->element_count; ++i)
3786 {
3787 const char *n = so_desc->elements[i].semantic_name;
3788
3789 if (n)
3790 size += strlen(n) + 1;
3791 }
3792 if (!(s = heap_alloc(size)))
3793 return E_OUTOFMEMORY;
3794
3795 s->desc = *so_desc;
3796
3797 memcpy(s->elements, so_desc->elements, so_desc->element_count * sizeof(*s->elements));
3798 s->desc.elements = s->elements;
3799
3800 name = (char *)&s->elements[s->desc.element_count];
3801 for (i = 0; i < so_desc->element_count; ++i)
3802 {
3803 struct wined3d_stream_output_element *e = &s->elements[i];
3804
3805 if (!e->semantic_name)
3806 continue;
3807
3808 size = strlen(e->semantic_name) + 1;
3809 memcpy(name, e->semantic_name, size);
3810 e->semantic_name = name;
3811 name += size;
3812 }
3813
3814 if (wine_rb_put(&device->so_descs, &s->desc, &s->entry) == -1)
3815 {
3816 heap_free(s);
3817 return E_FAIL;
3818 }
3819 gs->so_desc = &s->desc;
3820
3821 return WINED3D_OK;
3822 }
3823
3824 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
3825 const struct wined3d_stream_output_desc *so_desc)
3826 {
3827 const struct wined3d_shader_frontend *fe = shader->frontend;
3828 const struct wined3d_shader_signature_element *output;
3829 unsigned int i, component_idx, register_idx, mask;
3830 struct wined3d_shader_version shader_version;
3831 const DWORD *ptr;
3832 void *fe_data;
3833 HRESULT hr;
3834
3835 if (!so_desc)
3836 return WINED3D_OK;
3837
3838 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
3839 {
3840 WARN("Failed to initialise frontend data.\n");
3841 return WINED3DERR_INVALIDCALL;
3842 }
3843 fe->shader_read_header(fe_data, &ptr, &shader_version);
3844 fe->shader_free(fe_data);
3845
3846 switch (shader_version.type)
3847 {
3848 case WINED3D_SHADER_TYPE_VERTEX:
3849 case WINED3D_SHADER_TYPE_DOMAIN:
3850 shader->function = NULL;
3851 shader->functionLength = 0;
3852 break;
3853 case WINED3D_SHADER_TYPE_GEOMETRY:
3854 break;
3855 default:
3856 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
3857 return E_INVALIDARG;
3858 }
3859
3860 if (!shader->function)
3861 {
3862 shader->reg_maps.shader_version = shader_version;
3863 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
3864 shader_set_limits(shader);
3865 if (FAILED(hr = shader_scan_output_signature(shader)))
3866 return hr;
3867 }
3868
3869 for (i = 0; i < so_desc->element_count; ++i)
3870 {
3871 const struct wined3d_stream_output_element *e = &so_desc->elements[i];
3872
3873 if (!e->semantic_name)
3874 continue;
3875 if (!(output = shader_find_signature_element(&shader->output_signature,
3876 e->stream_idx, e->semantic_name, e->semantic_idx))
3877 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
3878 {
3879 WARN("Failed to find output signature element for stream output entry.\n");
3880 return E_INVALIDARG;
3881 }
3882
3883 mask = wined3d_mask_from_size(e->component_count) << component_idx;
3884 if ((output->mask & 0xff & mask) != mask)
3885 {
3886 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
3887 component_idx, e->component_count, mask, output->mask & 0xff);
3888 return E_INVALIDARG;
3889 }
3890 }
3891
3892 if (FAILED(hr = geometry_shader_init_so_desc(&shader->u.gs, shader->device, so_desc)))
3893 {
3894 WARN("Failed to initialise stream output description, hr %#x.\n", hr);
3895 return hr;
3896 }
3897
3898 return WINED3D_OK;
3899 }
3900
3901 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3902 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3903 void *parent, const struct wined3d_parent_ops *parent_ops)
3904 {
3905 HRESULT hr;
3906
3907 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3908 return hr;
3909
3910 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
3911 goto fail;
3912
3913 if (shader->function
3914 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
3915 goto fail;
3916
3917 return WINED3D_OK;
3918
3919 fail:
3920 shader_cleanup(shader);
3921 return hr;
3922 }
3923
3924 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3925 struct ds_compile_args *args, const struct wined3d_context *context)
3926 {
3927 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3928 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3929 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3930
3931 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3932 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3933
3934 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3935 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3936 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3937
3938 args->render_offscreen = context->render_offscreen;
3939
3940 init_interpolation_compile_args(args->interpolation_mode,
3941 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
3942
3943 args->padding = 0;
3944 }
3945
3946 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3947 struct gs_compile_args *args, const struct wined3d_context *context)
3948 {
3949 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3950
3951 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3952
3953 if (!(args->primitive_type = shader->u.gs.input_type))
3954 args->primitive_type = state->primitive_type;
3955
3956 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
3957 }
3958
3959 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3960 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3961 {
3962 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3963 struct wined3d_texture *texture;
3964 unsigned int i;
3965
3966 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3967 if (!d3d_info->srgb_write_control && needs_srgb_write(d3d_info, state, &state->fb))
3968 {
3969 static unsigned int warned = 0;
3970
3971 args->srgb_correction = 1;
3972 if (state->blend_state && state->blend_state->desc.rt[0].enable && !warned++)
3973 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3974 "support, expect rendering artifacts.\n");
3975 }
3976
3977 if (shader->reg_maps.shader_version.major == 1
3978 && shader->reg_maps.shader_version.minor <= 3)
3979 {
3980 for (i = 0; i < shader->limits->sampler; ++i)
3981 {
3982 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3983
3984 if (flags & WINED3D_TTFF_PROJECTED)
3985 {
3986 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3987
3988 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3989 {
3990 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3991 unsigned int j;
3992 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3993 DWORD max_valid = WINED3D_TTFF_COUNT4;
3994
3995 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3996 {
3997 struct wined3d_vertex_declaration_element *element =
3998 &state->vertex_declaration->elements[j];
3999
4000 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
4001 && element->usage_idx == index)
4002 {
4003 max_valid = element->format->component_count;
4004 break;
4005 }
4006 }
4007 if (!tex_transform || tex_transform > max_valid)
4008 {
4009 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
4010 tex_transform, max_valid);
4011 tex_transform = max_valid;
4012 }
4013 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
4014 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
4015 && tex_transform > WINED3D_TTFF_COUNT2)
4016 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4017 && tex_transform > WINED3D_TTFF_COUNT3))
4018 tex_transform |= WINED3D_PSARGS_PROJECTED;
4019 else
4020 {
4021 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
4022 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
4023 i, tex_transform, resource_type);
4024 }
4025 }
4026 else
4027 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
4028
4029 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
4030 }
4031 }
4032 }
4033 if (shader->reg_maps.shader_version.major == 1
4034 && shader->reg_maps.shader_version.minor <= 4)
4035 {
4036 for (i = 0; i < shader->limits->sampler; ++i)
4037 {
4038 if (!shader->reg_maps.resource_info[i].type)
4039 continue;
4040
4041 /* Treat unbound textures as 2D. The dummy texture will provide
4042 * the proper sample value. The tex_types bitmap defaults to
4043 * 2D because of the memset. */
4044 if (!(texture = state->textures[i]))
4045 continue;
4046
4047 switch (wined3d_texture_gl(texture)->target)
4048 {
4049 /* RECT textures are distinguished from 2D textures via np2_fixup */
4050 default:
4051 break;
4052
4053 case GL_TEXTURE_3D:
4054 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4055 break;
4056
4057 case GL_TEXTURE_CUBE_MAP_ARB:
4058 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4059 break;
4060 }
4061 }
4062 }
4063 else if (shader->reg_maps.shader_version.major <= 3)
4064 {
4065 for (i = 0; i < shader->limits->sampler; ++i)
4066 {
4067 enum wined3d_shader_resource_type resource_type;
4068 enum wined3d_shader_tex_types tex_type;
4069
4070 if (!(resource_type = shader->reg_maps.resource_info[i].type))
4071 continue;
4072
4073 switch (resource_type)
4074 {
4075 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
4076 tex_type = WINED3D_SHADER_TEX_3D;
4077 break;
4078 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
4079 tex_type = WINED3D_SHADER_TEX_CUBE;
4080 break;
4081 default:
4082 tex_type = WINED3D_SHADER_TEX_2D;
4083 break;
4084 }
4085
4086 if ((texture = state->textures[i]))
4087 {
4088 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
4089 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4090 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
4091 tex_type = WINED3D_SHADER_TEX_2D;
4092 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
4093 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
4094 tex_type = WINED3D_SHADER_TEX_3D;
4095 }
4096 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4097 }
4098 }
4099
4100 if (shader->reg_maps.shader_version.major >= 4)
4101 {
4102 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
4103 args->shadow = 0;
4104 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4105 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4106 args->np2_fixup = 0;
4107 }
4108 else
4109 {
4110 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4111 {
4112 if (!shader->reg_maps.resource_info[i].type)
4113 continue;
4114
4115 texture = state->textures[i];
4116 if (!texture)
4117 {
4118 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4119 continue;
4120 }
4121 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
4122 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4123 else
4124 args->color_fixup[i] = texture->resource.format->color_fixup;
4125
4126 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
4127 args->shadow |= 1u << i;
4128
4129 /* Flag samplers that need NP2 texcoord fixup. */
4130 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
4131 args->np2_fixup |= (1u << i);
4132 }
4133 }
4134
4135 if (shader->reg_maps.shader_version.major >= 3)
4136 {
4137 if (position_transformed)
4138 args->vp_mode = WINED3D_VP_MODE_NONE;
4139 else if (use_vs(state))
4140 args->vp_mode = WINED3D_VP_MODE_SHADER;
4141 else
4142 args->vp_mode = WINED3D_VP_MODE_FF;
4143 args->fog = WINED3D_FFP_PS_FOG_OFF;
4144 }
4145 else
4146 {
4147 args->vp_mode = WINED3D_VP_MODE_SHADER;
4148 if (state->render_states[WINED3D_RS_FOGENABLE])
4149 {
4150 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
4151 {
4152 case WINED3D_FOG_NONE:
4153 if (position_transformed || use_vs(state))
4154 {
4155 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
4156 break;
4157 }
4158
4159 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
4160 {
4161 case WINED3D_FOG_NONE: /* Fall through. */
4162 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4163 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4164 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4165 }
4166 break;
4167
4168 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4169 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4170 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4171 }
4172 }
4173 else
4174 {
4175 args->fog = WINED3D_FFP_PS_FOG_OFF;
4176 }
4177 }
4178
4179 if (!d3d_info->full_ffp_varyings)
4180 {
4181 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
4182
4183 args->texcoords_initialized = 0;
4184 for (i = 0; i < WINED3D_MAX_TEXTURES; ++i)
4185 {
4186 if (vs)
4187 {
4188 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
4189 args->texcoords_initialized |= 1u << i;
4190 }
4191 else
4192 {
4193 const struct wined3d_stream_info *si = &context->stream_info;
4194 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
4195
4196 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
4197 & WINED3D_FFP_TCI_MASK
4198 || (coord_idx < WINED3D_MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
4199 args->texcoords_initialized |= 1u << i;
4200 }
4201 }
4202 }
4203 else
4204 {
4205 args->texcoords_initialized = wined3d_mask_from_size(WINED3D_MAX_TEXTURES);
4206 }
4207
4208 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
4209 && state->primitive_type == WINED3D_PT_POINTLIST;
4210
4211 if (d3d_info->ffp_alpha_test)
4212 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
4213 else
4214 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4215 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4216 : WINED3D_CMP_ALWAYS) - 1;
4217
4218 if (d3d_info->emulated_flatshading)
4219 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4220
4221 args->y_correction = (shader->reg_maps.vpos && d3d_info->frag_coord_correction)
4222 || (shader->reg_maps.usesdsy && wined3d_settings.offscreen_rendering_mode != ORM_FBO)
4223 ? !context->render_offscreen : 0;
4224
4225 for (i = 0; i < ARRAY_SIZE(state->fb.render_targets); ++i)
4226 {
4227 struct wined3d_rendertarget_view *rtv = state->fb.render_targets[i];
4228 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
4229 args->rt_alpha_swizzle |= 1u << i;
4230 }
4231
4232 args->dual_source_blend = state->blend_state && state->blend_state->dual_source;
4233 }
4234
4235 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4236 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4237 {
4238 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4239 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4240 HRESULT hr;
4241
4242 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
4243 return hr;
4244
4245 if (FAILED(hr = shader_set_function(shader, device,
4246 WINED3D_SHADER_TYPE_PIXEL, device->adapter->d3d_info.limits.ps_uniform_count)))
4247 {
4248 shader_cleanup(shader);
4249 return hr;
4250 }
4251
4252 for (i = 0; i < MAX_REG_INPUT; ++i)
4253 {
4254 if (shader->u.ps.input_reg_used & (1u << i))
4255 {
4256 ++num_regs_used;
4257 highest_reg_used = i;
4258 }
4259 }
4260
4261 /* Don't do any register mapping magic if it is not needed, or if we can't
4262 * achieve anything anyway */
4263 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4264 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4265 || shader->reg_maps.shader_version.major >= 4)
4266 {
4267 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4268 {
4269 /* This happens with relative addressing. The input mapper function
4270 * warns about this if the higher registers are declared too, so
4271 * don't write a FIXME here */
4272 WARN("More varying registers used than supported\n");
4273 }
4274
4275 for (i = 0; i < MAX_REG_INPUT; ++i)
4276 {
4277 shader->u.ps.input_reg_map[i] = i;
4278 }
4279
4280 shader->u.ps.declared_in_count = highest_reg_used + 1;
4281 }
4282 else
4283 {
4284 shader->u.ps.declared_in_count = 0;
4285 for (i = 0; i < MAX_REG_INPUT; ++i)
4286 {
4287 if (shader->u.ps.input_reg_used & (1u << i))
4288 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4289 else shader->u.ps.input_reg_map[i] = ~0U;
4290 }
4291 }
4292
4293 return WINED3D_OK;
4294 }
4295
4296 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
4297 unsigned int resource_idx, DWORD tex_types)
4298 {
4299 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
4300 {
4301 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
4302 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
4303 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
4304 };
4305
4306 unsigned int idx;
4307
4308 if (reg_maps->shader_version.major > 3)
4309 return reg_maps->resource_info[resource_idx].type;
4310
4311 if (!reg_maps->resource_info[resource_idx].type)
4312 return 0;
4313
4314 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
4315 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
4316 return shader_resource_type_from_shader_tex_types[idx];
4317 }
4318
4319 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4320 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4321 {
4322 struct wined3d_shader *object;
4323 HRESULT hr;
4324
4325 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4326 device, desc, parent, parent_ops, shader);
4327
4328 if (!(object = heap_alloc_zero(sizeof(*object))))
4329 return E_OUTOFMEMORY;
4330
4331 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4332 {
4333 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4334 heap_free(object);
4335 return hr;
4336 }
4337
4338 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
4339 {
4340 shader_cleanup(object);
4341 heap_free(object);
4342 return hr;
4343 }
4344
4345 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4346
4347 TRACE("Created compute shader %p.\n", object);
4348 *shader = object;
4349
4350 return WINED3D_OK;
4351 }
4352
4353 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4354 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4355 {
4356 struct wined3d_shader *object;
4357 HRESULT hr;
4358
4359 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4360 device, desc, parent, parent_ops, shader);
4361
4362 if (!(object = heap_alloc_zero(sizeof(*object))))
4363 return E_OUTOFMEMORY;
4364
4365 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4366 {
4367 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4368 heap_free(object);
4369 return hr;
4370 }
4371
4372 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
4373 {
4374 shader_cleanup(object);
4375 heap_free(object);
4376 return hr;
4377 }
4378
4379 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4380
4381 TRACE("Created domain shader %p.\n", object);
4382 *shader = object;
4383
4384 return WINED3D_OK;
4385 }
4386
4387 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4388 const struct wined3d_stream_output_desc *so_desc, void *parent,
4389 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4390 {
4391 struct wined3d_shader *object;
4392 HRESULT hr;
4393
4394 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4395 device, desc, so_desc, parent, parent_ops, shader);
4396
4397 if (!(object = heap_alloc_zero(sizeof(*object))))
4398 return E_OUTOFMEMORY;
4399
4400 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4401 {
4402 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4403 heap_free(object);
4404 return hr;
4405 }
4406
4407 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4408
4409 TRACE("Created geometry shader %p.\n", object);
4410 *shader = object;
4411
4412 return WINED3D_OK;
4413 }
4414
4415 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4416 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4417 {
4418 struct wined3d_shader *object;
4419 HRESULT hr;
4420
4421 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4422 device, desc, parent, parent_ops, shader);
4423
4424 if (!(object = heap_alloc_zero(sizeof(*object))))
4425 return E_OUTOFMEMORY;
4426
4427 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4428 {
4429 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4430 heap_free(object);
4431 return hr;
4432 }
4433
4434 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
4435 {
4436 shader_cleanup(object);
4437 heap_free(object);
4438 return hr;
4439 }
4440
4441 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4442
4443 TRACE("Created hull shader %p.\n", object);
4444 *shader = object;
4445
4446 return WINED3D_OK;
4447 }
4448
4449 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4450 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4451 {
4452 struct wined3d_shader *object;
4453 HRESULT hr;
4454
4455 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4456 device, desc, parent, parent_ops, shader);
4457
4458 if (!(object = heap_alloc_zero(sizeof(*object))))
4459 return E_OUTOFMEMORY;
4460
4461 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4462 {
4463 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4464 heap_free(object);
4465 return hr;
4466 }
4467
4468 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4469
4470 TRACE("Created pixel shader %p.\n", object);
4471 *shader = object;
4472
4473 return WINED3D_OK;
4474 }
4475
4476 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4477 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4478 {
4479 struct wined3d_shader *object;
4480 HRESULT hr;
4481
4482 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4483 device, desc, parent, parent_ops, shader);
4484
4485 if (!(object = heap_alloc_zero(sizeof(*object))))
4486 return E_OUTOFMEMORY;
4487
4488 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4489 {
4490 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4491 heap_free(object);
4492 return hr;
4493 }
4494
4495 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4496
4497 TRACE("Created vertex shader %p.\n", object);
4498 *shader = object;
4499
4500 return WINED3D_OK;
4501 }
Windows API implementation