?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Use wined3d_mask_from_size() in shader_glsl_gather4().
[wine.git] / dlls / wined3d / shader.c
blob43c6bb402c01a9968e6ae52e42399780c643a77f
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26
27 #include <stdio.h>
28 #include <string.h>
29
30 #include "wined3d_private.h"
31
32 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
33
34 const struct wined3d_vec4 wined3d_srgb_const[] =
35 {
36 /* pow, mul_high, sub_high, mul_low */
37 {0.41666f, 1.055f, 0.055f, 12.92f},
38 /* cmp */
39 {0.0031308f, 0.0f, 0.0f, 0.0f},
40 };
41
42 static const char * const shader_opcode_names[] =
43 {
44 /* WINED3DSIH_ABS */ "abs",
45 /* WINED3DSIH_ADD */ "add",
46 /* WINED3DSIH_AND */ "and",
47 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
48 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
49 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
50 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
51 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
52 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
53 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
54 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
55 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
56 /* WINED3DSIH_BEM */ "bem",
57 /* WINED3DSIH_BFI */ "bfi",
58 /* WINED3DSIH_BFREV */ "bfrev",
59 /* WINED3DSIH_BREAK */ "break",
60 /* WINED3DSIH_BREAKC */ "breakc",
61 /* WINED3DSIH_BREAKP */ "breakp",
62 /* WINED3DSIH_BUFINFO */ "bufinfo",
63 /* WINED3DSIH_CALL */ "call",
64 /* WINED3DSIH_CALLNZ */ "callnz",
65 /* WINED3DSIH_CASE */ "case",
66 /* WINED3DSIH_CMP */ "cmp",
67 /* WINED3DSIH_CND */ "cnd",
68 /* WINED3DSIH_CONTINUE */ "continue",
69 /* WINED3DSIH_CONTINUEP */ "continuec",
70 /* WINED3DSIH_COUNTBITS */ "countbits",
71 /* WINED3DSIH_CRS */ "crs",
72 /* WINED3DSIH_CUT */ "cut",
73 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
74 /* WINED3DSIH_DCL */ "dcl",
75 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
76 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
77 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
78 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
79 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
80 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
81 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
82 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
83 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
84 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
85 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
86 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
87 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
88 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
89 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
90 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
91 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
92 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
93 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
94 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
95 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
96 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
97 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
98 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
99 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
100 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
101 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
102 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
103 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
104 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
105 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
106 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
107 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
108 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
109 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
110 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
111 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
112 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
113 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
114 /* WINED3DSIH_DEF */ "def",
115 /* WINED3DSIH_DEFAULT */ "default",
116 /* WINED3DSIH_DEFB */ "defb",
117 /* WINED3DSIH_DEFI */ "defi",
118 /* WINED3DSIH_DIV */ "div",
119 /* WINED3DSIH_DP2 */ "dp2",
120 /* WINED3DSIH_DP2ADD */ "dp2add",
121 /* WINED3DSIH_DP3 */ "dp3",
122 /* WINED3DSIH_DP4 */ "dp4",
123 /* WINED3DSIH_DST */ "dst",
124 /* WINED3DSIH_DSX */ "dsx",
125 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
126 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
127 /* WINED3DSIH_DSY */ "dsy",
128 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
129 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
130 /* WINED3DSIH_ELSE */ "else",
131 /* WINED3DSIH_EMIT */ "emit",
132 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
133 /* WINED3DSIH_ENDIF */ "endif",
134 /* WINED3DSIH_ENDLOOP */ "endloop",
135 /* WINED3DSIH_ENDREP */ "endrep",
136 /* WINED3DSIH_ENDSWITCH */ "endswitch",
137 /* WINED3DSIH_EQ */ "eq",
138 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
139 /* WINED3DSIH_EXP */ "exp",
140 /* WINED3DSIH_EXPP */ "expp",
141 /* WINED3DSIH_F16TOF32 */ "f16tof32",
142 /* WINED3DSIH_F32TOF16 */ "f32tof16",
143 /* WINED3DSIH_FCALL */ "fcall",
144 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
145 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
146 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
147 /* WINED3DSIH_FRC */ "frc",
148 /* WINED3DSIH_FTOI */ "ftoi",
149 /* WINED3DSIH_FTOU */ "ftou",
150 /* WINED3DSIH_GATHER4 */ "gather4",
151 /* WINED3DSIH_GATHER4_C */ "gather4_c",
152 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
153 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
154 /* WINED3DSIH_GE */ "ge",
155 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
156 /* WINED3DSIH_HS_DECLS */ "hs_decls",
157 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
158 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
159 /* WINED3DSIH_IADD */ "iadd",
160 /* WINED3DSIH_IBFE */ "ibfe",
161 /* WINED3DSIH_IEQ */ "ieq",
162 /* WINED3DSIH_IF */ "if",
163 /* WINED3DSIH_IFC */ "ifc",
164 /* WINED3DSIH_IGE */ "ige",
165 /* WINED3DSIH_ILT */ "ilt",
166 /* WINED3DSIH_IMAD */ "imad",
167 /* WINED3DSIH_IMAX */ "imax",
168 /* WINED3DSIH_IMIN */ "imin",
169 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
170 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
171 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
172 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
173 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
174 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
175 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
176 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
177 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
178 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
179 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
180 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
181 /* WINED3DSIH_IMUL */ "imul",
182 /* WINED3DSIH_INE */ "ine",
183 /* WINED3DSIH_INEG */ "ineg",
184 /* WINED3DSIH_ISHL */ "ishl",
185 /* WINED3DSIH_ISHR */ "ishr",
186 /* WINED3DSIH_ITOF */ "itof",
187 /* WINED3DSIH_LABEL */ "label",
188 /* WINED3DSIH_LD */ "ld",
189 /* WINED3DSIH_LD2DMS */ "ld2dms",
190 /* WINED3DSIH_LD_RAW */ "ld_raw",
191 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
192 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
193 /* WINED3DSIH_LIT */ "lit",
194 /* WINED3DSIH_LOD */ "lod",
195 /* WINED3DSIH_LOG */ "log",
196 /* WINED3DSIH_LOGP */ "logp",
197 /* WINED3DSIH_LOOP */ "loop",
198 /* WINED3DSIH_LRP */ "lrp",
199 /* WINED3DSIH_LT */ "lt",
200 /* WINED3DSIH_M3x2 */ "m3x2",
201 /* WINED3DSIH_M3x3 */ "m3x3",
202 /* WINED3DSIH_M3x4 */ "m3x4",
203 /* WINED3DSIH_M4x3 */ "m4x3",
204 /* WINED3DSIH_M4x4 */ "m4x4",
205 /* WINED3DSIH_MAD */ "mad",
206 /* WINED3DSIH_MAX */ "max",
207 /* WINED3DSIH_MIN */ "min",
208 /* WINED3DSIH_MOV */ "mov",
209 /* WINED3DSIH_MOVA */ "mova",
210 /* WINED3DSIH_MOVC */ "movc",
211 /* WINED3DSIH_MUL */ "mul",
212 /* WINED3DSIH_NE */ "ne",
213 /* WINED3DSIH_NOP */ "nop",
214 /* WINED3DSIH_NOT */ "not",
215 /* WINED3DSIH_NRM */ "nrm",
216 /* WINED3DSIH_OR */ "or",
217 /* WINED3DSIH_PHASE */ "phase",
218 /* WINED3DSIH_POW */ "pow",
219 /* WINED3DSIH_RCP */ "rcp",
220 /* WINED3DSIH_REP */ "rep",
221 /* WINED3DSIH_RESINFO */ "resinfo",
222 /* WINED3DSIH_RET */ "ret",
223 /* WINED3DSIH_RETP */ "retp",
224 /* WINED3DSIH_ROUND_NE */ "round_ne",
225 /* WINED3DSIH_ROUND_NI */ "round_ni",
226 /* WINED3DSIH_ROUND_PI */ "round_pi",
227 /* WINED3DSIH_ROUND_Z */ "round_z",
228 /* WINED3DSIH_RSQ */ "rsq",
229 /* WINED3DSIH_SAMPLE */ "sample",
230 /* WINED3DSIH_SAMPLE_B */ "sample_b",
231 /* WINED3DSIH_SAMPLE_C */ "sample_c",
232 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
233 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
234 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
235 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
236 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
237 /* WINED3DSIH_SETP */ "setp",
238 /* WINED3DSIH_SGE */ "sge",
239 /* WINED3DSIH_SGN */ "sgn",
240 /* WINED3DSIH_SINCOS */ "sincos",
241 /* WINED3DSIH_SLT */ "slt",
242 /* WINED3DSIH_SQRT */ "sqrt",
243 /* WINED3DSIH_STORE_RAW */ "store_raw",
244 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
245 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
246 /* WINED3DSIH_SUB */ "sub",
247 /* WINED3DSIH_SWAPC */ "swapc",
248 /* WINED3DSIH_SWITCH */ "switch",
249 /* WINED3DSIH_SYNC */ "sync",
250 /* WINED3DSIH_TEX */ "texld",
251 /* WINED3DSIH_TEXBEM */ "texbem",
252 /* WINED3DSIH_TEXBEML */ "texbeml",
253 /* WINED3DSIH_TEXCOORD */ "texcrd",
254 /* WINED3DSIH_TEXDEPTH */ "texdepth",
255 /* WINED3DSIH_TEXDP3 */ "texdp3",
256 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
257 /* WINED3DSIH_TEXKILL */ "texkill",
258 /* WINED3DSIH_TEXLDD */ "texldd",
259 /* WINED3DSIH_TEXLDL */ "texldl",
260 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
261 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
262 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
263 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
264 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
265 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
266 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
267 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
268 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
269 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
270 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
271 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
272 /* WINED3DSIH_UBFE */ "ubfe",
273 /* WINED3DSIH_UDIV */ "udiv",
274 /* WINED3DSIH_UGE */ "uge",
275 /* WINED3DSIH_ULT */ "ult",
276 /* WINED3DSIH_UMAX */ "umax",
277 /* WINED3DSIH_UMIN */ "umin",
278 /* WINED3DSIH_UMUL */ "umul",
279 /* WINED3DSIH_USHR */ "ushr",
280 /* WINED3DSIH_UTOF */ "utof",
281 /* WINED3DSIH_XOR */ "xor",
282 };
283
284 static const char * const semantic_names[] =
285 {
286 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
287 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
288 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
289 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
290 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
291 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
292 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
293 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
294 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
295 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
296 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
297 /* WINED3D_DECL_USAGE_FOG */ "FOG",
298 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
299 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
300 };
301
302 static const struct
303 {
304 enum wined3d_shader_input_sysval_semantic sysval_semantic;
305 const char *sysval_name;
306 }
307 shader_input_sysval_semantic_names[] =
308 {
309 {WINED3D_SIV_POSITION, "position"},
310 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
311 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
312 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
313 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
314 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
315 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
316 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
317 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
318 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
319 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
320 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
322 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
323 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
324 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
325 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
326 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
327 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
328 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
329 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
330 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
331 };
332
333 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
334 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
335
336 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
337 {
338 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
339 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
340
341 return shader_opcode_names[handler_idx];
342 }
343
344 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
345 {
346 if (usage >= ARRAY_SIZE(semantic_names))
347 {
348 FIXME("Unrecognized usage %#x.\n", usage);
349 return "UNRECOGNIZED";
350 }
351
352 return semantic_names[usage];
353 }
354
355 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
356 {
357 unsigned int i;
358
359 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
360 {
361 if (!strcmp(name, semantic_names[i]))
362 return i;
363 }
364
365 return ~0U;
366 }
367
368 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
369 {
370 switch (usage)
371 {
372 case WINED3D_DECL_USAGE_POSITION:
373 return WINED3D_SV_POSITION;
374 default:
375 return 0;
376 }
377 }
378
379 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
380 {
381 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
382 }
383
384 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
385 const struct wined3d_shader_semantic *s)
386 {
387 e->semantic_name = shader_semantic_name_from_usage(s->usage);
388 e->semantic_idx = s->usage_idx;
389 e->stream_idx = 0;
390 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
391 e->component_type = WINED3D_TYPE_FLOAT;
392 e->register_idx = s->reg.reg.idx[0].offset;
393 e->mask = s->reg.write_mask;
394 }
395
396 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
397 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
398 {
399 e->semantic_name = shader_semantic_name_from_usage(usage);
400 e->semantic_idx = usage_idx;
401 e->stream_idx = 0;
402 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
403 e->component_type = WINED3D_TYPE_FLOAT;
404 e->register_idx = reg_idx;
405 e->mask = write_mask;
406 }
407
408 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
409 {
410 switch (format)
411 {
412 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
413 return &sm1_shader_frontend;
414
415 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
416 return &sm4_shader_frontend;
417
418 default:
419 WARN("Invalid byte code format %#x specified.\n", format);
420 return NULL;
421 }
422 }
423
424 void string_buffer_clear(struct wined3d_string_buffer *buffer)
425 {
426 buffer->buffer[0] = '\0';
427 buffer->content_size = 0;
428 }
429
430 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
431 {
432 buffer->buffer_size = 32;
433 if (!(buffer->buffer = heap_alloc(buffer->buffer_size)))
434 {
435 ERR("Failed to allocate shader buffer memory.\n");
436 return FALSE;
437 }
438
439 string_buffer_clear(buffer);
440 return TRUE;
441 }
442
443 void string_buffer_free(struct wined3d_string_buffer *buffer)
444 {
445 heap_free(buffer->buffer);
446 }
447
448 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
449 {
450 char *new_buffer;
451 unsigned int new_buffer_size = buffer->buffer_size * 2;
452
453 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
454 new_buffer_size *= 2;
455 if (!(new_buffer = heap_realloc(buffer->buffer, new_buffer_size)))
456 {
457 ERR("Failed to grow buffer.\n");
458 buffer->buffer[buffer->content_size] = '\0';
459 return FALSE;
460 }
461 buffer->buffer = new_buffer;
462 buffer->buffer_size = new_buffer_size;
463 return TRUE;
464 }
465
466 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
467 {
468 unsigned int rem;
469 int rc;
470
471 rem = buffer->buffer_size - buffer->content_size;
472 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
473 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
474 return rc;
475
476 buffer->content_size += rc;
477 return 0;
478 }
479
480 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
481 {
482 va_list args;
483 int ret;
484
485 for (;;)
486 {
487 va_start(args, format);
488 ret = shader_vaddline(buffer, format, args);
489 va_end(args);
490 if (!ret)
491 return ret;
492 if (!string_buffer_resize(buffer, ret))
493 return -1;
494 }
495 }
496
497 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
498 {
499 struct wined3d_string_buffer *buffer;
500
501 if (list_empty(&list->list))
502 {
503 buffer = heap_alloc(sizeof(*buffer));
504 if (!buffer || !string_buffer_init(buffer))
505 {
506 ERR("Couldn't allocate buffer for temporary string.\n");
507 heap_free(buffer);
508 return NULL;
509 }
510 }
511 else
512 {
513 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
514 list_remove(&buffer->entry);
515 }
516 string_buffer_clear(buffer);
517 return buffer;
518 }
519
520 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
521 {
522 if (!buffer)
523 return 0;
524 string_buffer_clear(buffer);
525 return shader_vaddline(buffer, format, args);
526 }
527
528 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
529 {
530 va_list args;
531 int ret;
532
533 for (;;)
534 {
535 va_start(args, format);
536 ret = string_buffer_vsprintf(buffer, format, args);
537 va_end(args);
538 if (!ret)
539 return;
540 if (!string_buffer_resize(buffer, ret))
541 return;
542 }
543 }
544
545 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
546 {
547 if (!buffer)
548 return;
549 list_add_head(&list->list, &buffer->entry);
550 }
551
552 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
553 {
554 list_init(&list->list);
555 }
556
557 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
558 {
559 struct wined3d_string_buffer *buffer, *buffer_next;
560
561 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
562 {
563 string_buffer_free(buffer);
564 heap_free(buffer);
565 }
566 list_init(&list->list);
567 }
568
569 /* Convert floating point offset relative to a register file to an absolute
570 * offset for float constants. */
571 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
572 {
573 switch (register_type)
574 {
575 case WINED3DSPR_CONST: return register_idx;
576 case WINED3DSPR_CONST2: return 2048 + register_idx;
577 case WINED3DSPR_CONST3: return 4096 + register_idx;
578 case WINED3DSPR_CONST4: return 6144 + register_idx;
579 default:
580 FIXME("Unsupported register type: %u.\n", register_type);
581 return register_idx;
582 }
583 }
584
585 static void shader_delete_constant_list(struct list *clist)
586 {
587 struct wined3d_shader_lconst *constant, *constant_next;
588
589 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
590 heap_free(constant);
591 list_init(clist);
592 }
593
594 static void shader_set_limits(struct wined3d_shader *shader)
595 {
596 static const struct limits_entry
597 {
598 unsigned int min_version;
599 unsigned int max_version;
600 struct wined3d_shader_limits limits;
601 }
602 vs_limits[] =
603 {
604 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
605 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
606 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
607 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
608 * even though they are capable of supporting much more (GL
609 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
610 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
611 * shaders to 256. */
612 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
613 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
614 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
615 {0}
616 },
617 hs_limits[] =
618 {
619 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
620 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
621 },
622 ds_limits[] =
623 {
624 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
625 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
626 },
627 gs_limits[] =
628 {
629 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
630 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
631 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
632 {0}
633 },
634 ps_limits[] =
635 {
636 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
637 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
638 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
639 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
640 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
641 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
642 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
643 {0}
644 },
645 cs_limits[] =
646 {
647 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
648 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
649 };
650 const struct limits_entry *limits_array;
651 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
652 shader->reg_maps.shader_version.minor);
653 int i = 0;
654
655 switch (shader->reg_maps.shader_version.type)
656 {
657 default:
658 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
659 /* Fall-through. */
660 case WINED3D_SHADER_TYPE_VERTEX:
661 limits_array = vs_limits;
662 break;
663 case WINED3D_SHADER_TYPE_HULL:
664 limits_array = hs_limits;
665 break;
666 case WINED3D_SHADER_TYPE_DOMAIN:
667 limits_array = ds_limits;
668 break;
669 case WINED3D_SHADER_TYPE_GEOMETRY:
670 limits_array = gs_limits;
671 break;
672 case WINED3D_SHADER_TYPE_PIXEL:
673 limits_array = ps_limits;
674 break;
675 case WINED3D_SHADER_TYPE_COMPUTE:
676 limits_array = cs_limits;
677 break;
678 }
679
680 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
681 {
682 if (shader_version <= limits_array[i].max_version)
683 {
684 shader->limits = &limits_array[i].limits;
685 break;
686 }
687 ++i;
688 }
689 if (!shader->limits)
690 {
691 FIXME("Unexpected shader version \"%u.%u\".\n",
692 shader->reg_maps.shader_version.major,
693 shader->reg_maps.shader_version.minor);
694 shader->limits = &limits_array[max(0, i - 1)].limits;
695 }
696 }
697
698 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
699 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
700 {
701 switch (reg->type)
702 {
703 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
704 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
705 reg_maps->texcoord |= 1u << reg->idx[0].offset;
706 else
707 reg_maps->address |= 1u << reg->idx[0].offset;
708 break;
709
710 case WINED3DSPR_TEMP:
711 reg_maps->temporary |= 1u << reg->idx[0].offset;
712 break;
713
714 case WINED3DSPR_INPUT:
715 if (reg->idx[0].rel_addr)
716 reg_maps->input_rel_addressing = 1;
717 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
718 {
719 /* If relative addressing is used, we must assume that all
720 * registers are used. Even if it is a construct like v3[aL],
721 * we can't assume that v0, v1 and v2 aren't read because aL
722 * can be negative. */
723 if (reg->idx[0].rel_addr)
724 shader->u.ps.input_reg_used = ~0u;
725 else
726 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
727 }
728 else
729 {
730 reg_maps->input_registers |= 1u << reg->idx[0].offset;
731 }
732 break;
733
734 case WINED3DSPR_RASTOUT:
735 if (reg->idx[0].offset == 1)
736 reg_maps->fog = 1;
737 if (reg->idx[0].offset == 2)
738 reg_maps->point_size = 1;
739 break;
740
741 case WINED3DSPR_MISCTYPE:
742 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
743 {
744 if (!reg->idx[0].offset)
745 reg_maps->vpos = 1;
746 else if (reg->idx[0].offset == 1)
747 reg_maps->usesfacing = 1;
748 }
749 break;
750
751 case WINED3DSPR_CONST:
752 if (reg->idx[0].rel_addr)
753 {
754 if (reg->idx[0].offset < reg_maps->min_rel_offset)
755 reg_maps->min_rel_offset = reg->idx[0].offset;
756 if (reg->idx[0].offset > reg_maps->max_rel_offset)
757 reg_maps->max_rel_offset = reg->idx[0].offset;
758 reg_maps->usesrelconstF = TRUE;
759 }
760 else
761 {
762 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
763 {
764 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
765 return FALSE;
766 }
767 else
768 {
769 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
770 }
771 }
772 break;
773
774 case WINED3DSPR_CONSTINT:
775 if (reg->idx[0].offset >= shader->limits->constant_int)
776 {
777 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
778 return FALSE;
779 }
780 else
781 {
782 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
783 }
784 break;
785
786 case WINED3DSPR_CONSTBOOL:
787 if (reg->idx[0].offset >= shader->limits->constant_bool)
788 {
789 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
790 return FALSE;
791 }
792 else
793 {
794 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
795 }
796 break;
797
798 case WINED3DSPR_COLOROUT:
799 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
800 break;
801
802 case WINED3DSPR_OUTCONTROLPOINT:
803 reg_maps->vocp = 1;
804 break;
805
806 case WINED3DSPR_SAMPLEMASK:
807 reg_maps->sample_mask = 1;
808 break;
809
810 default:
811 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
812 reg->type, reg->idx[0].offset, reg->idx[1].offset);
813 break;
814 }
815 return TRUE;
816 }
817
818 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
819 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
820 {
821 struct wined3d_shader_sampler_map_entry *entries, *entry;
822 struct wined3d_shader_sampler_map *map;
823 unsigned int i;
824
825 map = &reg_maps->sampler_map;
826 entries = map->entries;
827 for (i = 0; i < map->count; ++i)
828 {
829 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
830 return;
831 }
832
833 if (!map->size)
834 {
835 if (!(entries = heap_calloc(4, sizeof(*entries))))
836 {
837 ERR("Failed to allocate sampler map entries.\n");
838 return;
839 }
840 map->size = 4;
841 map->entries = entries;
842 }
843 else if (map->count == map->size)
844 {
845 size_t new_size = map->size * 2;
846
847 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
848 || !(entries = heap_realloc(entries, sizeof(*entries) * new_size)))
849 {
850 ERR("Failed to resize sampler map entries.\n");
851 return;
852 }
853 map->size = new_size;
854 map->entries = entries;
855 }
856
857 entry = &entries[map->count++];
858 entry->resource_idx = resource_idx;
859 entry->sampler_idx = sampler_idx;
860 entry->bind_idx = bind_idx;
861 }
862
863 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
864 {
865 switch (instr)
866 {
867 case WINED3DSIH_M4x4:
868 case WINED3DSIH_M3x4:
869 return param == 1 ? 3 : 0;
870
871 case WINED3DSIH_M4x3:
872 case WINED3DSIH_M3x3:
873 return param == 1 ? 2 : 0;
874
875 case WINED3DSIH_M3x2:
876 return param == 1 ? 1 : 0;
877
878 default:
879 return 0;
880 }
881 }
882
883 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
884 unsigned int register_idx, unsigned int size, unsigned int stride)
885 {
886 struct wined3d_shader_tgsm *tgsm;
887
888 if (register_idx >= MAX_TGSM_REGISTERS)
889 {
890 ERR("Invalid TGSM register index %u.\n", register_idx);
891 return S_OK;
892 }
893 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
894 {
895 FIXME("TGSM declarations are allowed only in compute shaders.\n");
896 return S_OK;
897 }
898
899 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
900 register_idx + 1, sizeof(*reg_maps->tgsm)))
901 return E_OUTOFMEMORY;
902
903 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
904 tgsm = &reg_maps->tgsm[register_idx];
905 tgsm->size = size;
906 tgsm->stride = stride;
907 return S_OK;
908 }
909
910 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
911 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
912 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
913 {
914 struct wined3d_shader_phase *phase;
915
916 if ((phase = *current_phase))
917 {
918 phase->end = previous_instruction_ptr;
919 *current_phase = NULL;
920 }
921
922 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
923 {
924 ERR("Unexpected shader type %s.\n", debug_shader_type(shader->reg_maps.shader_version.type));
925 return E_FAIL;
926 }
927
928 switch (ins->handler_idx)
929 {
930 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
931 if (shader->u.hs.phases.control_point)
932 {
933 FIXME("Multiple control point phases.\n");
934 heap_free(shader->u.hs.phases.control_point);
935 }
936 if (!(shader->u.hs.phases.control_point = heap_alloc_zero(sizeof(*shader->u.hs.phases.control_point))))
937 return E_OUTOFMEMORY;
938 phase = shader->u.hs.phases.control_point;
939 break;
940 case WINED3DSIH_HS_FORK_PHASE:
941 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
942 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
943 sizeof(*shader->u.hs.phases.fork)))
944 return E_OUTOFMEMORY;
945 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
946 break;
947 case WINED3DSIH_HS_JOIN_PHASE:
948 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
949 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
950 sizeof(*shader->u.hs.phases.join)))
951 return E_OUTOFMEMORY;
952 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
953 break;
954 default:
955 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
956 return E_FAIL;
957 }
958
959 phase->start = current_instruction_ptr;
960 *current_phase = phase;
961
962 return WINED3D_OK;
963 }
964
965 static HRESULT shader_calculate_clip_or_cull_distance_mask(
966 const struct wined3d_shader_signature_element *e, unsigned int *mask)
967 {
968 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
969 * the only allowed semantic indices.
970 */
971 if (e->semantic_idx >= WINED3D_MAX_CLIP_DISTANCES / 4)
972 {
973 *mask = 0;
974 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
975 return WINED3DERR_INVALIDCALL;
976 }
977
978 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
979 return WINED3D_OK;
980 }
981
982 static void wined3d_insert_interpolation_mode(DWORD *packed_interpolation_mode,
983 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
984 {
985 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
986 FIXME("Unexpected interpolation mode %#x.\n", mode);
987
988 wined3d_insert_bits(packed_interpolation_mode,
989 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
990 }
991
992 static HRESULT shader_scan_output_signature(struct wined3d_shader *shader)
993 {
994 const struct wined3d_shader_signature *output_signature = &shader->output_signature;
995 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
996 unsigned int i;
997 HRESULT hr;
998
999 for (i = 0; i < output_signature->element_count; ++i)
1000 {
1001 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1002 unsigned int mask;
1003
1004 reg_maps->output_registers |= 1u << e->register_idx;
1005 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
1006 {
1007 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1008 return hr;
1009 reg_maps->clip_distance_mask |= mask;
1010 }
1011 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
1012 {
1013 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1014 return hr;
1015 reg_maps->cull_distance_mask |= mask;
1016 }
1017 else if (e->sysval_semantic == WINED3D_SV_VIEWPORT_ARRAY_INDEX)
1018 {
1019 reg_maps->viewport_array = 1;
1020 }
1021 }
1022
1023 return WINED3D_OK;
1024 }
1025
1026 /* Note that this does not count the loop register as an address register. */
1027 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, DWORD constf_size)
1028 {
1029 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
1030 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
1031 struct wined3d_shader_signature *output_signature = &shader->output_signature;
1032 struct wined3d_shader_signature *input_signature = &shader->input_signature;
1033 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1034 const struct wined3d_shader_frontend *fe = shader->frontend;
1035 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
1036 struct wined3d_shader_version shader_version;
1037 struct wined3d_shader_phase *phase = NULL;
1038 const DWORD *ptr, *prev_ins, *current_ins;
1039 void *fe_data = shader->frontend_data;
1040 unsigned int i;
1041 HRESULT hr;
1042
1043 memset(reg_maps, 0, sizeof(*reg_maps));
1044 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1045 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1046 reg_maps->min_rel_offset = ~0U;
1047 list_init(&reg_maps->indexable_temps);
1048
1049 fe->shader_read_header(fe_data, &ptr, &shader_version);
1050 prev_ins = current_ins = ptr;
1051 reg_maps->shader_version = shader_version;
1052
1053 shader_set_limits(shader);
1054
1055 if (!(reg_maps->constf = heap_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1056 sizeof(*reg_maps->constf))))
1057 {
1058 ERR("Failed to allocate constant map memory.\n");
1059 return E_OUTOFMEMORY;
1060 }
1061
1062 while (!fe->shader_is_end(fe_data, &ptr))
1063 {
1064 struct wined3d_shader_instruction ins;
1065
1066 current_ins = ptr;
1067 /* Fetch opcode. */
1068 fe->shader_read_instruction(fe_data, &ptr, &ins);
1069
1070 /* Unhandled opcode, and its parameters. */
1071 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1072 {
1073 WARN("Encountered unrecognised or invalid instruction.\n");
1074 return WINED3DERR_INVALIDCALL;
1075 }
1076
1077 /* Handle declarations. */
1078 if (ins.handler_idx == WINED3DSIH_DCL
1079 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1080 {
1081 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1082 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1083
1084 switch (semantic->reg.reg.type)
1085 {
1086 /* Mark input registers used. */
1087 case WINED3DSPR_INPUT:
1088 if (reg_idx >= MAX_REG_INPUT)
1089 {
1090 ERR("Invalid input register index %u.\n", reg_idx);
1091 break;
1092 }
1093 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1094 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1095 return WINED3DERR_INVALIDCALL;
1096 reg_maps->input_registers |= 1u << reg_idx;
1097 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1098 break;
1099
1100 /* Vertex shader: mark 3.0 output registers used, save token. */
1101 case WINED3DSPR_OUTPUT:
1102 if (reg_idx >= MAX_REG_OUTPUT)
1103 {
1104 ERR("Invalid output register index %u.\n", reg_idx);
1105 break;
1106 }
1107 reg_maps->output_registers |= 1u << reg_idx;
1108 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1109 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1110 reg_maps->fog = 1;
1111 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1112 reg_maps->point_size = 1;
1113 break;
1114
1115 case WINED3DSPR_SAMPLER:
1116 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1117 case WINED3DSPR_RESOURCE:
1118 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1119 {
1120 ERR("Invalid resource index %u.\n", reg_idx);
1121 break;
1122 }
1123 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1124 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1125 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1126 break;
1127
1128 case WINED3DSPR_UAV:
1129 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1130 {
1131 ERR("Invalid UAV resource index %u.\n", reg_idx);
1132 break;
1133 }
1134 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1135 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1136 if (ins.flags)
1137 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1138 break;
1139
1140 default:
1141 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1142 break;
1143 }
1144 }
1145 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1146 {
1147 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1148 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1149 {
1150 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1151 }
1152 else
1153 {
1154 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1155 wined3d_bitmap_set(&reg_maps->cb_map, reg->idx[0].offset);
1156 }
1157 }
1158 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1159 {
1160 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1161 {
1162 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1163 shader->u.ps.force_early_depth_stencil = TRUE;
1164 else
1165 FIXME("Invalid instruction %#x for shader type %#x.\n",
1166 ins.handler_idx, shader_version.type);
1167 }
1168 else
1169 {
1170 WARN("Ignoring global flags %#x.\n", ins.flags);
1171 }
1172 }
1173 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1174 {
1175 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1176 shader->u.gs.instance_count = ins.declaration.count;
1177 else
1178 FIXME("Invalid instruction %#x for shader type %#x.\n",
1179 ins.handler_idx, shader_version.type);
1180 }
1181 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1182 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1183 {
1184 if (phase)
1185 phase->instance_count = ins.declaration.count;
1186 else
1187 FIXME("Instruction %s outside of shader phase.\n",
1188 debug_d3dshaderinstructionhandler(ins.handler_idx));
1189 }
1190 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1191 {
1192 if (reg_maps->icb)
1193 FIXME("Multiple immediate constant buffers.\n");
1194 reg_maps->icb = ins.declaration.icb;
1195 }
1196 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1197 {
1198 if (phase)
1199 {
1200 FIXME("Indexable temporary registers not supported.\n");
1201 }
1202 else
1203 {
1204 struct wined3d_shader_indexable_temp *reg;
1205
1206 if (!(reg = heap_alloc(sizeof(*reg))))
1207 return E_OUTOFMEMORY;
1208
1209 *reg = ins.declaration.indexable_temp;
1210 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1211 }
1212 }
1213 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1214 {
1215 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1216 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1217 else
1218 FIXME("Invalid instruction %#x for shader type %#x.\n",
1219 ins.handler_idx, shader_version.type);
1220 }
1221 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1222 {
1223 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1224 if (reg_idx >= MAX_REG_INPUT)
1225 {
1226 ERR("Invalid register index %u.\n", reg_idx);
1227 break;
1228 }
1229 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1230 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1231 else
1232 FIXME("Invalid instruction %#x for shader type %#x.\n",
1233 ins.handler_idx, shader_version.type);
1234 }
1235 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1236 {
1237 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1238 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1239 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1240 {
1241 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1242 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1243 else
1244 FIXME("Invalid instruction %#x for shader type %#x.\n",
1245 ins.handler_idx, shader_version.type);
1246 }
1247 }
1248 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1249 {
1250 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1251 shader->u.hs.output_vertex_count = ins.declaration.count;
1252 else
1253 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1254 }
1255 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1256 {
1257 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1258 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1259 else
1260 FIXME("Invalid instruction %#x for shader type %#x.\n",
1261 ins.handler_idx, shader_version.type);
1262 }
1263 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1264 {
1265 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1266 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1267 {
1268 ERR("Invalid resource index %u.\n", reg_idx);
1269 break;
1270 }
1271 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1272 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1273 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1274 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1275 }
1276 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1277 {
1278 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1279 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1280 {
1281 ERR("Invalid resource index %u.\n", reg_idx);
1282 break;
1283 }
1284 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1285 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1286 reg_maps->resource_info[reg_idx].flags = 0;
1287 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1288 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1289 }
1290 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1291 {
1292 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1293 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1294 }
1295 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1296 {
1297 if (phase)
1298 phase->temporary_count = ins.declaration.count;
1299 else
1300 reg_maps->temporary_count = ins.declaration.count;
1301 }
1302 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1303 {
1304 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1305 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1306 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1307 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1308 }
1309 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1310 {
1311 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1312 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1313 else
1314 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1315 }
1316 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1317 {
1318 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1319 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1320 else
1321 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1322 }
1323 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1324 {
1325 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1326 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1327 return hr;
1328 }
1329 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1330 {
1331 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1332 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1333 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1334 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1335 return hr;
1336 }
1337 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1338 {
1339 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1340 {
1341 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1342 }
1343 else
1344 {
1345 FIXME("Invalid instruction %#x for shader type %#x.\n",
1346 ins.handler_idx, shader_version.type);
1347 }
1348 }
1349 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1350 {
1351 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1352 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1353 {
1354 ERR("Invalid UAV resource index %u.\n", reg_idx);
1355 break;
1356 }
1357 if (ins.flags)
1358 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1359 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1360 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1361 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1362 }
1363 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1364 {
1365 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1366 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1367 {
1368 ERR("Invalid UAV resource index %u.\n", reg_idx);
1369 break;
1370 }
1371 if (ins.flags)
1372 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1373 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1374 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1375 reg_maps->uav_resource_info[reg_idx].flags = 0;
1376 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1377 }
1378 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1379 {
1380 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1381 shader->u.gs.vertices_out = ins.declaration.count;
1382 else
1383 FIXME("Invalid instruction %#x for shader type %#x.\n",
1384 ins.handler_idx, shader_version.type);
1385 }
1386 else if (ins.handler_idx == WINED3DSIH_DEF)
1387 {
1388 struct wined3d_shader_lconst *lconst;
1389 float *value;
1390
1391 if (!(lconst = heap_alloc(sizeof(*lconst))))
1392 return E_OUTOFMEMORY;
1393
1394 lconst->idx = ins.dst[0].reg.idx[0].offset;
1395 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1396 value = (float *)lconst->value;
1397
1398 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1399 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1400 {
1401 if (value[0] < -1.0f) value[0] = -1.0f;
1402 else if (value[0] > 1.0f) value[0] = 1.0f;
1403 if (value[1] < -1.0f) value[1] = -1.0f;
1404 else if (value[1] > 1.0f) value[1] = 1.0f;
1405 if (value[2] < -1.0f) value[2] = -1.0f;
1406 else if (value[2] > 1.0f) value[2] = 1.0f;
1407 if (value[3] < -1.0f) value[3] = -1.0f;
1408 else if (value[3] > 1.0f) value[3] = 1.0f;
1409 }
1410
1411 list_add_head(&shader->constantsF, &lconst->entry);
1412
1413 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1414 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1415 {
1416 shader->lconst_inf_or_nan = TRUE;
1417 }
1418 }
1419 else if (ins.handler_idx == WINED3DSIH_DEFI)
1420 {
1421 struct wined3d_shader_lconst *lconst;
1422
1423 if (!(lconst = heap_alloc(sizeof(*lconst))))
1424 return E_OUTOFMEMORY;
1425
1426 lconst->idx = ins.dst[0].reg.idx[0].offset;
1427 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1428
1429 list_add_head(&shader->constantsI, &lconst->entry);
1430 reg_maps->local_int_consts |= (1u << lconst->idx);
1431 }
1432 else if (ins.handler_idx == WINED3DSIH_DEFB)
1433 {
1434 struct wined3d_shader_lconst *lconst;
1435
1436 if (!(lconst = heap_alloc(sizeof(*lconst))))
1437 return E_OUTOFMEMORY;
1438
1439 lconst->idx = ins.dst[0].reg.idx[0].offset;
1440 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1441
1442 list_add_head(&shader->constantsB, &lconst->entry);
1443 reg_maps->local_bool_consts |= (1u << lconst->idx);
1444 }
1445 /* Handle shader phases. */
1446 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1447 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1448 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1449 {
1450 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1451 return hr;
1452 }
1453 /* For subroutine prototypes. */
1454 else if (ins.handler_idx == WINED3DSIH_LABEL)
1455 {
1456 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1457 }
1458 /* Set texture, address, temporary registers. */
1459 else
1460 {
1461 BOOL color0_mov = FALSE;
1462 unsigned int i;
1463
1464 /* This will loop over all the registers and try to
1465 * make a bitmask of the ones we're interested in.
1466 *
1467 * Relative addressing tokens are ignored, but that's
1468 * okay, since we'll catch any address registers when
1469 * they are initialized (required by spec). */
1470 for (i = 0; i < ins.dst_count; ++i)
1471 {
1472 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1473 shader_version.type, constf_size))
1474 return WINED3DERR_INVALIDCALL;
1475
1476 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1477 {
1478 UINT idx = ins.dst[i].reg.idx[0].offset;
1479
1480 switch (ins.dst[i].reg.type)
1481 {
1482 case WINED3DSPR_RASTOUT:
1483 if (shader_version.major >= 3)
1484 break;
1485 switch (idx)
1486 {
1487 case 0: /* oPos */
1488 reg_maps->output_registers |= 1u << 10;
1489 shader_signature_from_usage(&output_signature_elements[10],
1490 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1491 break;
1492
1493 case 1: /* oFog */
1494 reg_maps->output_registers |= 1u << 11;
1495 shader_signature_from_usage(&output_signature_elements[11],
1496 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1497 break;
1498
1499 case 2: /* oPts */
1500 reg_maps->output_registers |= 1u << 11;
1501 shader_signature_from_usage(&output_signature_elements[11],
1502 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1503 break;
1504 }
1505 break;
1506
1507 case WINED3DSPR_ATTROUT:
1508 if (shader_version.major >= 3)
1509 break;
1510 if (idx < 2)
1511 {
1512 idx += 8;
1513 if (reg_maps->output_registers & (1u << idx))
1514 {
1515 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1516 }
1517 else
1518 {
1519 reg_maps->output_registers |= 1u << idx;
1520 shader_signature_from_usage(&output_signature_elements[idx],
1521 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1522 }
1523 }
1524 break;
1525
1526 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1527 if (shader_version.major >= 3)
1528 {
1529 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1530 {
1531 WARN("Invalid output register index %u.\n", idx);
1532 break;
1533 }
1534 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1535 break;
1536 }
1537 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1538 {
1539 WARN("Invalid texcoord index %u.\n", idx);
1540 break;
1541 }
1542 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1543 if (reg_maps->output_registers & (1u << idx))
1544 {
1545 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1546 }
1547 else
1548 {
1549 reg_maps->output_registers |= 1u << idx;
1550 shader_signature_from_usage(&output_signature_elements[idx],
1551 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1552 }
1553 break;
1554
1555 default:
1556 break;
1557 }
1558 }
1559
1560 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1561 {
1562 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1563 {
1564 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1565 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1566 * the mov and perform the sRGB write correction from the source register.
1567 *
1568 * However, if the mov is only partial, we can't do this, and if the write
1569 * comes from an instruction other than MOV it is hard to do as well. If
1570 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1571 shader->u.ps.color0_mov = FALSE;
1572 if (ins.handler_idx == WINED3DSIH_MOV
1573 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1574 {
1575 /* Used later when the source register is read. */
1576 color0_mov = TRUE;
1577 }
1578 }
1579 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1580 * end
1581 */
1582 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1583 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1584 {
1585 shader->u.ps.color0_mov = FALSE;
1586 }
1587 }
1588
1589 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1590 if (shader_version.major == 1
1591 && (ins.handler_idx == WINED3DSIH_TEX
1592 || ins.handler_idx == WINED3DSIH_TEXBEM
1593 || ins.handler_idx == WINED3DSIH_TEXBEML
1594 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1595 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1596 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1597 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1598 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1599 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1600 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1601 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1602 {
1603 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1604
1605 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1606 {
1607 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1608 continue;
1609 }
1610
1611 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1612 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1613 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1614 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1615 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1616
1617 /* texbem is only valid with < 1.4 pixel shaders */
1618 if (ins.handler_idx == WINED3DSIH_TEXBEM
1619 || ins.handler_idx == WINED3DSIH_TEXBEML)
1620 {
1621 reg_maps->bumpmat |= 1u << reg_idx;
1622 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1623 {
1624 reg_maps->luminanceparams |= 1u << reg_idx;
1625 }
1626 }
1627 }
1628 else if (ins.handler_idx == WINED3DSIH_BEM)
1629 {
1630 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1631 }
1632 }
1633
1634 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1635 {
1636 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1637 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1638 {
1639 ERR("Invalid UAV index %u.\n", reg_idx);
1640 break;
1641 }
1642 reg_maps->uav_counter_mask |= (1u << reg_idx);
1643 }
1644 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1645 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1646 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1647 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1648 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1649 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1650 {
1651 const struct wined3d_shader_register *reg;
1652
1653 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1654 reg = &ins.src[1].reg;
1655 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1656 reg = &ins.src[2].reg;
1657 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1658 reg = &ins.dst[0].reg;
1659 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1660 reg = &ins.src[0].reg;
1661 else
1662 reg = &ins.dst[1].reg;
1663
1664 if (reg->type == WINED3DSPR_UAV)
1665 {
1666 if (reg->idx[0].offset >= MAX_UNORDERED_ACCESS_VIEWS)
1667 {
1668 ERR("Invalid UAV index %u.\n", reg->idx[0].offset);
1669 break;
1670 }
1671 reg_maps->uav_read_mask |= (1u << reg->idx[0].offset);
1672 }
1673 }
1674 else if (ins.handler_idx == WINED3DSIH_NRM)
1675 {
1676 reg_maps->usesnrm = 1;
1677 }
1678 else if (ins.handler_idx == WINED3DSIH_DSY
1679 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1680 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1681 {
1682 reg_maps->usesdsy = 1;
1683 }
1684 else if (ins.handler_idx == WINED3DSIH_DSX
1685 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1686 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1687 {
1688 reg_maps->usesdsx = 1;
1689 }
1690 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1691 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1692 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1693 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1694 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1695 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1696 else if (ins.handler_idx == WINED3DSIH_LOOP
1697 || ins.handler_idx == WINED3DSIH_REP)
1698 {
1699 ++cur_loop_depth;
1700 if (cur_loop_depth > max_loop_depth)
1701 max_loop_depth = cur_loop_depth;
1702 }
1703 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1704 || ins.handler_idx == WINED3DSIH_ENDREP)
1705 {
1706 --cur_loop_depth;
1707 }
1708 else if (ins.handler_idx == WINED3DSIH_GATHER4
1709 || ins.handler_idx == WINED3DSIH_GATHER4_C
1710 || ins.handler_idx == WINED3DSIH_SAMPLE
1711 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1712 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1713 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1714 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1715 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1716 {
1717 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1718 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1719 }
1720 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1721 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1722 {
1723 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1724 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1725 }
1726 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1727 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1728 {
1729 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1730 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1731 }
1732 else if (ins.handler_idx == WINED3DSIH_LD
1733 || ins.handler_idx == WINED3DSIH_LD2DMS
1734 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1735 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1736 {
1737 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1738 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1739 }
1740 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1741 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1742 {
1743 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1744 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1745 }
1746
1747 if (ins.predicate)
1748 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1749 shader_version.type, constf_size))
1750 return WINED3DERR_INVALIDCALL;
1751
1752 for (i = 0; i < ins.src_count; ++i)
1753 {
1754 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1755 struct wined3d_shader_register reg = ins.src[i].reg;
1756
1757 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1758 shader_version.type, constf_size))
1759 return WINED3DERR_INVALIDCALL;
1760 while (count)
1761 {
1762 ++reg.idx[0].offset;
1763 if (!shader_record_register_usage(shader, reg_maps, &reg,
1764 shader_version.type, constf_size))
1765 return WINED3DERR_INVALIDCALL;
1766 --count;
1767 }
1768
1769 if (color0_mov)
1770 {
1771 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1772 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1773 {
1774 shader->u.ps.color0_mov = TRUE;
1775 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1776 }
1777 }
1778 }
1779 }
1780
1781 prev_ins = current_ins;
1782 }
1783 reg_maps->loop_depth = max_loop_depth;
1784
1785 if (phase)
1786 {
1787 phase->end = prev_ins;
1788 phase = NULL;
1789 }
1790
1791 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1792 * R0 is written to the render target. */
1793 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1794 reg_maps->rt_mask |= (1u << 0);
1795
1796 if (input_signature->elements)
1797 {
1798 for (i = 0; i < input_signature->element_count; ++i)
1799 {
1800 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1801 {
1802 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1803 {
1804 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1805 return WINED3DERR_INVALIDCALL;
1806 }
1807 }
1808 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1809 {
1810 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1811 reg_maps->vpos = 1;
1812 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1813 reg_maps->usesfacing = 1;
1814 }
1815 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1816 }
1817 }
1818 else if (!input_signature->elements && reg_maps->input_registers)
1819 {
1820 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1821 struct wined3d_shader_signature_element *e;
1822 unsigned int i;
1823
1824 if (!(input_signature->elements = heap_calloc(count, sizeof(*input_signature->elements))))
1825 return E_OUTOFMEMORY;
1826 input_signature->element_count = count;
1827
1828 e = input_signature->elements;
1829 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1830 {
1831 if (!(reg_maps->input_registers & (1u << i)))
1832 continue;
1833 input_signature_elements[i].register_idx = i;
1834 *e++ = input_signature_elements[i];
1835 }
1836 }
1837
1838 if (output_signature->elements)
1839 {
1840 if (FAILED(hr = shader_scan_output_signature(shader)))
1841 return hr;
1842 }
1843 else if (reg_maps->output_registers)
1844 {
1845 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1846 struct wined3d_shader_signature_element *e;
1847
1848 if (!(output_signature->elements = heap_calloc(count, sizeof(*output_signature->elements))))
1849 return E_OUTOFMEMORY;
1850 output_signature->element_count = count;
1851
1852 e = output_signature->elements;
1853 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1854 {
1855 if (!(reg_maps->output_registers & (1u << i)))
1856 continue;
1857 *e++ = output_signature_elements[i];
1858 }
1859 }
1860
1861 return WINED3D_OK;
1862 }
1863
1864 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1865 {
1866 struct wined3d_shader_indexable_temp *reg, *reg_next;
1867
1868 heap_free(reg_maps->constf);
1869 heap_free(reg_maps->sampler_map.entries);
1870
1871 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1872 heap_free(reg);
1873 list_init(&reg_maps->indexable_temps);
1874
1875 heap_free(reg_maps->tgsm);
1876 }
1877
1878 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1879 {
1880 DWORD map = 1u << max;
1881 map |= map - 1;
1882 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1883
1884 return wined3d_log2i(map);
1885 }
1886
1887 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1888 {
1889 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1890 {
1891 shader_addline(buffer, "refactoringAllowed");
1892 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1893 if (global_flags)
1894 shader_addline(buffer, " | ");
1895 }
1896
1897 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1898 {
1899 shader_addline(buffer, "forceEarlyDepthStencil");
1900 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1901 if (global_flags)
1902 shader_addline(buffer, " | ");
1903 }
1904
1905 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1906 {
1907 shader_addline(buffer, "enableRawAndStructuredBuffers");
1908 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1909 }
1910
1911 if (global_flags)
1912 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1913 }
1914
1915 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1916 {
1917 if (sync_flags & WINED3DSSF_GLOBAL_UAV)
1918 {
1919 shader_addline(buffer, "_uglobal");
1920 sync_flags &= ~WINED3DSSF_GLOBAL_UAV;
1921 }
1922 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1923 {
1924 shader_addline(buffer, "_g");
1925 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1926 }
1927 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1928 {
1929 shader_addline(buffer, "_t");
1930 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1931 }
1932
1933 if (sync_flags)
1934 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1935 }
1936
1937 static void shader_dump_precise_flags(struct wined3d_string_buffer *buffer, DWORD flags)
1938 {
1939 if (!(flags & WINED3DSI_PRECISE_XYZW))
1940 return;
1941
1942 shader_addline(buffer, " [precise");
1943 if (flags != WINED3DSI_PRECISE_XYZW)
1944 {
1945 shader_addline(buffer, "(%s%s%s%s)",
1946 flags & WINED3DSI_PRECISE_X ? "x" : "",
1947 flags & WINED3DSI_PRECISE_Y ? "y" : "",
1948 flags & WINED3DSI_PRECISE_Z ? "z" : "",
1949 flags & WINED3DSI_PRECISE_W ? "w" : "");
1950 }
1951 shader_addline(buffer, "]");
1952 }
1953
1954 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1955 {
1956 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1957 {
1958 shader_addline(buffer, "_glc");
1959 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1960 }
1961 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1962 {
1963 shader_addline(buffer, "_opc");
1964 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1965 }
1966
1967 if (uav_flags)
1968 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1969 }
1970
1971 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1972 enum wined3d_tessellator_domain domain)
1973 {
1974 switch (domain)
1975 {
1976 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1977 shader_addline(buffer, "line");
1978 break;
1979 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1980 shader_addline(buffer, "triangle");
1981 break;
1982 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1983 shader_addline(buffer, "quad");
1984 break;
1985 default:
1986 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1987 break;
1988 }
1989 }
1990
1991 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1992 enum wined3d_tessellator_output_primitive output_primitive)
1993 {
1994 switch (output_primitive)
1995 {
1996 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1997 shader_addline(buffer, "point");
1998 break;
1999 case WINED3D_TESSELLATOR_OUTPUT_LINE:
2000 shader_addline(buffer, "line");
2001 break;
2002 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
2003 shader_addline(buffer, "triangle_cw");
2004 break;
2005 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
2006 shader_addline(buffer, "triangle_ccw");
2007 break;
2008 default:
2009 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
2010 break;
2011 }
2012 }
2013
2014 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
2015 enum wined3d_tessellator_partitioning partitioning)
2016 {
2017 switch (partitioning)
2018 {
2019 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
2020 shader_addline(buffer, "integer");
2021 break;
2022 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
2023 shader_addline(buffer, "pow2");
2024 break;
2025 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
2026 shader_addline(buffer, "fractional_odd");
2027 break;
2028 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
2029 shader_addline(buffer, "fractional_even");
2030 break;
2031 default:
2032 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
2033 break;
2034 }
2035 }
2036
2037 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
2038 enum wined3d_shader_input_sysval_semantic semantic)
2039 {
2040 unsigned int i;
2041
2042 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
2043 {
2044 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
2045 {
2046 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
2047 return;
2048 }
2049 }
2050
2051 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
2052 }
2053
2054 static void shader_dump_resource_type(struct wined3d_string_buffer *buffer, enum wined3d_shader_resource_type type)
2055 {
2056 static const char *const resource_type_names[] =
2057 {
2058 /* WINED3D_SHADER_RESOURCE_NONE */ "none",
2059 /* WINED3D_SHADER_RESOURCE_BUFFER */ "buffer",
2060 /* WINED3D_SHADER_RESOURCE_TEXTURE_1D */ "texture1d",
2061 /* WINED3D_SHADER_RESOURCE_TEXTURE_2D */ "texture2d",
2062 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DMS */ "texture2dms",
2063 /* WINED3D_SHADER_RESOURCE_TEXTURE_3D */ "texture3d",
2064 /* WINED3D_SHADER_RESOURCE_TEXTURE_CUBE */ "texturecube",
2065 /* WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY */ "texture1darray",
2066 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY */ "texture2darray",
2067 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY */ "texture2dmsarray",
2068 /* WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY */ "texturecubearray",
2069 };
2070
2071 if (type < ARRAY_SIZE(resource_type_names))
2072 shader_addline(buffer, "%s", resource_type_names[type]);
2073 else
2074 shader_addline(buffer, "unknown");
2075 }
2076
2077 static void shader_dump_data_type(struct wined3d_string_buffer *buffer, enum wined3d_data_type type)
2078 {
2079 static const char *const data_type_names[] =
2080 {
2081 /* WINED3D_DATA_FLOAT */ "(float)",
2082 /* WINED3D_DATA_INT */ "(int)",
2083 /* WINED3D_DATA_RESOURCE */ "(resource)",
2084 /* WINED3D_DATA_SAMPLER */ "(sampler)",
2085 /* WINED3D_DATA_UAV */ "(uav)",
2086 /* WINED3D_DATA_UINT */ "(uint)",
2087 /* WINED3D_DATA_UNORM */ "(unorm)",
2088 /* WINED3D_DATA_SNORM */ "(snorm)",
2089 /* WINED3D_DATA_OPAQUE */ "(opaque)",
2090 };
2091
2092 if (type < ARRAY_SIZE(data_type_names))
2093 shader_addline(buffer, "%s", data_type_names[type]);
2094 else
2095 shader_addline(buffer, "(unknown)");
2096 }
2097
2098 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
2099 const struct wined3d_shader_semantic *semantic, unsigned int flags,
2100 const struct wined3d_shader_version *shader_version)
2101 {
2102 shader_addline(buffer, "dcl");
2103
2104 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
2105 {
2106 switch (semantic->resource_type)
2107 {
2108 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2109 shader_addline(buffer, "_2d");
2110 break;
2111
2112 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2113 shader_addline(buffer, "_3d");
2114 break;
2115
2116 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2117 shader_addline(buffer, "_cube");
2118 break;
2119
2120 default:
2121 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2122 break;
2123 }
2124 }
2125 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2126 {
2127 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2128 shader_addline(buffer, "_resource_");
2129 else
2130 shader_addline(buffer, "_uav_");
2131 shader_dump_resource_type(buffer, semantic->resource_type);
2132 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2133 shader_dump_uav_flags(buffer, flags);
2134 shader_dump_data_type(buffer, semantic->resource_data_type);
2135 }
2136 else
2137 {
2138 /* Pixel shaders 3.0 don't have usage semantics. */
2139 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2140 return;
2141 else
2142 shader_addline(buffer, "_");
2143
2144 switch (semantic->usage)
2145 {
2146 case WINED3D_DECL_USAGE_POSITION:
2147 shader_addline(buffer, "position%u", semantic->usage_idx);
2148 break;
2149
2150 case WINED3D_DECL_USAGE_BLEND_INDICES:
2151 shader_addline(buffer, "blend");
2152 break;
2153
2154 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2155 shader_addline(buffer, "weight");
2156 break;
2157
2158 case WINED3D_DECL_USAGE_NORMAL:
2159 shader_addline(buffer, "normal%u", semantic->usage_idx);
2160 break;
2161
2162 case WINED3D_DECL_USAGE_PSIZE:
2163 shader_addline(buffer, "psize");
2164 break;
2165
2166 case WINED3D_DECL_USAGE_COLOR:
2167 if (!semantic->usage_idx)
2168 shader_addline(buffer, "color");
2169 else
2170 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2171 break;
2172
2173 case WINED3D_DECL_USAGE_TEXCOORD:
2174 shader_addline(buffer, "texture%u", semantic->usage_idx);
2175 break;
2176
2177 case WINED3D_DECL_USAGE_TANGENT:
2178 shader_addline(buffer, "tangent");
2179 break;
2180
2181 case WINED3D_DECL_USAGE_BINORMAL:
2182 shader_addline(buffer, "binormal");
2183 break;
2184
2185 case WINED3D_DECL_USAGE_TESS_FACTOR:
2186 shader_addline(buffer, "tessfactor");
2187 break;
2188
2189 case WINED3D_DECL_USAGE_POSITIONT:
2190 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2191 break;
2192
2193 case WINED3D_DECL_USAGE_FOG:
2194 shader_addline(buffer, "fog");
2195 break;
2196
2197 case WINED3D_DECL_USAGE_DEPTH:
2198 shader_addline(buffer, "depth");
2199 break;
2200
2201 case WINED3D_DECL_USAGE_SAMPLE:
2202 shader_addline(buffer, "sample");
2203 break;
2204
2205 default:
2206 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2207 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2208 }
2209 }
2210 }
2211
2212 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2213 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2214 {
2215 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2216 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2217 UINT offset = reg->idx[0].offset;
2218
2219 switch (reg->type)
2220 {
2221 case WINED3DSPR_TEMP:
2222 shader_addline(buffer, "r");
2223 break;
2224
2225 case WINED3DSPR_INPUT:
2226 shader_addline(buffer, "v");
2227 break;
2228
2229 case WINED3DSPR_CONST:
2230 case WINED3DSPR_CONST2:
2231 case WINED3DSPR_CONST3:
2232 case WINED3DSPR_CONST4:
2233 shader_addline(buffer, "c");
2234 offset = shader_get_float_offset(reg->type, offset);
2235 break;
2236
2237 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2238 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2239 break;
2240
2241 case WINED3DSPR_RASTOUT:
2242 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2243 break;
2244
2245 case WINED3DSPR_COLOROUT:
2246 shader_addline(buffer, "oC");
2247 break;
2248
2249 case WINED3DSPR_DEPTHOUT:
2250 shader_addline(buffer, "oDepth");
2251 break;
2252
2253 case WINED3DSPR_DEPTHOUTGE:
2254 shader_addline(buffer, "oDepthGE");
2255 break;
2256
2257 case WINED3DSPR_DEPTHOUTLE:
2258 shader_addline(buffer, "oDepthLE");
2259 break;
2260
2261 case WINED3DSPR_ATTROUT:
2262 shader_addline(buffer, "oD");
2263 break;
2264
2265 case WINED3DSPR_TEXCRDOUT:
2266 /* Vertex shaders >= 3.0 use general purpose output registers
2267 * (WINED3DSPR_OUTPUT), which can include an address token. */
2268 if (shader_version->major >= 3)
2269 shader_addline(buffer, "o");
2270 else
2271 shader_addline(buffer, "oT");
2272 break;
2273
2274 case WINED3DSPR_CONSTINT:
2275 shader_addline(buffer, "i");
2276 break;
2277
2278 case WINED3DSPR_CONSTBOOL:
2279 shader_addline(buffer, "b");
2280 break;
2281
2282 case WINED3DSPR_LABEL:
2283 shader_addline(buffer, "l");
2284 break;
2285
2286 case WINED3DSPR_LOOP:
2287 shader_addline(buffer, "aL");
2288 break;
2289
2290 case WINED3DSPR_SAMPLER:
2291 shader_addline(buffer, "s");
2292 break;
2293
2294 case WINED3DSPR_MISCTYPE:
2295 if (offset > 1)
2296 {
2297 FIXME("Unhandled misctype register %u.\n", offset);
2298 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2299 }
2300 else
2301 {
2302 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2303 }
2304 break;
2305
2306 case WINED3DSPR_PREDICATE:
2307 shader_addline(buffer, "p");
2308 break;
2309
2310 case WINED3DSPR_IMMCONST:
2311 shader_addline(buffer, "l");
2312 break;
2313
2314 case WINED3DSPR_CONSTBUFFER:
2315 shader_addline(buffer, "cb");
2316 break;
2317
2318 case WINED3DSPR_IMMCONSTBUFFER:
2319 shader_addline(buffer, "icb");
2320 break;
2321
2322 case WINED3DSPR_PRIMID:
2323 shader_addline(buffer, "primID");
2324 break;
2325
2326 case WINED3DSPR_NULL:
2327 shader_addline(buffer, "null");
2328 break;
2329
2330 case WINED3DSPR_RASTERIZER:
2331 shader_addline(buffer, "rasterizer");
2332 break;
2333
2334 case WINED3DSPR_RESOURCE:
2335 shader_addline(buffer, "t");
2336 break;
2337
2338 case WINED3DSPR_UAV:
2339 shader_addline(buffer, "u");
2340 break;
2341
2342 case WINED3DSPR_OUTPOINTID:
2343 shader_addline(buffer, "vOutputControlPointID");
2344 break;
2345
2346 case WINED3DSPR_FORKINSTID:
2347 shader_addline(buffer, "vForkInstanceId");
2348 break;
2349
2350 case WINED3DSPR_JOININSTID:
2351 shader_addline(buffer, "vJoinInstanceId");
2352 break;
2353
2354 case WINED3DSPR_INCONTROLPOINT:
2355 shader_addline(buffer, "vicp");
2356 break;
2357
2358 case WINED3DSPR_OUTCONTROLPOINT:
2359 shader_addline(buffer, "vocp");
2360 break;
2361
2362 case WINED3DSPR_PATCHCONST:
2363 shader_addline(buffer, "vpc");
2364 break;
2365
2366 case WINED3DSPR_TESSCOORD:
2367 shader_addline(buffer, "vDomainLocation");
2368 break;
2369
2370 case WINED3DSPR_GROUPSHAREDMEM:
2371 shader_addline(buffer, "g");
2372 break;
2373
2374 case WINED3DSPR_THREADID:
2375 shader_addline(buffer, "vThreadID");
2376 break;
2377
2378 case WINED3DSPR_THREADGROUPID:
2379 shader_addline(buffer, "vThreadGroupID");
2380 break;
2381
2382 case WINED3DSPR_LOCALTHREADID:
2383 shader_addline(buffer, "vThreadIDInGroup");
2384 break;
2385
2386 case WINED3DSPR_LOCALTHREADINDEX:
2387 shader_addline(buffer, "vThreadIDInGroupFlattened");
2388 break;
2389
2390 case WINED3DSPR_IDXTEMP:
2391 shader_addline(buffer, "x");
2392 break;
2393
2394 case WINED3DSPR_STREAM:
2395 shader_addline(buffer, "m");
2396 break;
2397
2398 case WINED3DSPR_FUNCTIONBODY:
2399 shader_addline(buffer, "fb");
2400 break;
2401
2402 case WINED3DSPR_FUNCTIONPOINTER:
2403 shader_addline(buffer, "fp");
2404 break;
2405
2406 case WINED3DSPR_COVERAGE:
2407 shader_addline(buffer, "vCoverage");
2408 break;
2409
2410 case WINED3DSPR_SAMPLEMASK:
2411 shader_addline(buffer, "oMask");
2412 break;
2413
2414 case WINED3DSPR_GSINSTID:
2415 shader_addline(buffer, "vGSInstanceID");
2416 break;
2417
2418 default:
2419 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2420 break;
2421 }
2422
2423 if (reg->type == WINED3DSPR_IMMCONST)
2424 {
2425 shader_addline(buffer, "(");
2426 switch (reg->immconst_type)
2427 {
2428 case WINED3D_IMMCONST_SCALAR:
2429 switch (reg->data_type)
2430 {
2431 case WINED3D_DATA_FLOAT:
2432 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2433 break;
2434 case WINED3D_DATA_INT:
2435 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2436 break;
2437 case WINED3D_DATA_RESOURCE:
2438 case WINED3D_DATA_SAMPLER:
2439 case WINED3D_DATA_UINT:
2440 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2441 break;
2442 default:
2443 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2444 break;
2445 }
2446 break;
2447
2448 case WINED3D_IMMCONST_VEC4:
2449 switch (reg->data_type)
2450 {
2451 case WINED3D_DATA_FLOAT:
2452 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2453 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2454 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2455 break;
2456 case WINED3D_DATA_INT:
2457 shader_addline(buffer, "%d, %d, %d, %d",
2458 reg->u.immconst_data[0], reg->u.immconst_data[1],
2459 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2460 break;
2461 case WINED3D_DATA_RESOURCE:
2462 case WINED3D_DATA_SAMPLER:
2463 case WINED3D_DATA_UINT:
2464 shader_addline(buffer, "%u, %u, %u, %u",
2465 reg->u.immconst_data[0], reg->u.immconst_data[1],
2466 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2467 break;
2468 default:
2469 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2470 break;
2471 }
2472 break;
2473
2474 default:
2475 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2476 break;
2477 }
2478 shader_addline(buffer, ")");
2479 }
2480 else if (reg->type != WINED3DSPR_RASTOUT
2481 && reg->type != WINED3DSPR_MISCTYPE
2482 && reg->type != WINED3DSPR_NULL)
2483 {
2484 if (offset != ~0u)
2485 {
2486 shader_addline(buffer, "[");
2487 if (reg->idx[0].rel_addr)
2488 {
2489 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2490 shader_addline(buffer, " + ");
2491 }
2492 shader_addline(buffer, "%u]", offset);
2493
2494 if (reg->idx[1].offset != ~0u)
2495 {
2496 shader_addline(buffer, "[");
2497 if (reg->idx[1].rel_addr)
2498 {
2499 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2500 shader_addline(buffer, " + ");
2501 }
2502 shader_addline(buffer, "%u]", reg->idx[1].offset);
2503 }
2504 }
2505
2506 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2507 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2508 }
2509 }
2510
2511 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2512 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2513 {
2514 DWORD write_mask = param->write_mask;
2515
2516 shader_dump_register(buffer, &param->reg, shader_version);
2517
2518 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2519 {
2520 static const char write_mask_chars[] = "xyzw";
2521
2522 shader_addline(buffer, ".");
2523 if (write_mask & WINED3DSP_WRITEMASK_0)
2524 shader_addline(buffer, "%c", write_mask_chars[0]);
2525 if (write_mask & WINED3DSP_WRITEMASK_1)
2526 shader_addline(buffer, "%c", write_mask_chars[1]);
2527 if (write_mask & WINED3DSP_WRITEMASK_2)
2528 shader_addline(buffer, "%c", write_mask_chars[2]);
2529 if (write_mask & WINED3DSP_WRITEMASK_3)
2530 shader_addline(buffer, "%c", write_mask_chars[3]);
2531 }
2532 }
2533
2534 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2535 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2536 {
2537 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2538 DWORD swizzle = param->swizzle;
2539
2540 if (src_modifier == WINED3DSPSM_NEG
2541 || src_modifier == WINED3DSPSM_BIASNEG
2542 || src_modifier == WINED3DSPSM_SIGNNEG
2543 || src_modifier == WINED3DSPSM_X2NEG
2544 || src_modifier == WINED3DSPSM_ABSNEG)
2545 shader_addline(buffer, "-");
2546 else if (src_modifier == WINED3DSPSM_COMP)
2547 shader_addline(buffer, "1-");
2548 else if (src_modifier == WINED3DSPSM_NOT)
2549 shader_addline(buffer, "!");
2550
2551 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2552 shader_addline(buffer, "abs(");
2553
2554 shader_dump_register(buffer, &param->reg, shader_version);
2555
2556 switch (src_modifier)
2557 {
2558 case WINED3DSPSM_NONE: break;
2559 case WINED3DSPSM_NEG: break;
2560 case WINED3DSPSM_NOT: break;
2561 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2562 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2563 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2564 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2565 case WINED3DSPSM_COMP: break;
2566 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2567 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2568 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2569 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2570 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2571 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2572 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2573 }
2574
2575 if (swizzle != WINED3DSP_NOSWIZZLE)
2576 {
2577 static const char swizzle_chars[] = "xyzw";
2578 DWORD swizzle_x = swizzle & 0x03;
2579 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2580 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2581 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2582
2583 if (swizzle_x == swizzle_y
2584 && swizzle_x == swizzle_z
2585 && swizzle_x == swizzle_w)
2586 {
2587 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2588 }
2589 else
2590 {
2591 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2592 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2593 }
2594 }
2595 }
2596
2597 /* Shared code in order to generate the bulk of the shader string. */
2598 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2599 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2600 const DWORD *start, const DWORD *end)
2601 {
2602 struct wined3d_device *device = shader->device;
2603 const struct wined3d_shader_frontend *fe = shader->frontend;
2604 void *fe_data = shader->frontend_data;
2605 struct wined3d_shader_version shader_version;
2606 struct wined3d_shader_parser_state state;
2607 struct wined3d_shader_instruction ins;
2608 struct wined3d_shader_tex_mx tex_mx;
2609 struct wined3d_shader_context ctx;
2610 const DWORD *ptr;
2611
2612 /* Initialize current parsing state. */
2613 tex_mx.current_row = 0;
2614 state.current_loop_depth = 0;
2615 state.current_loop_reg = 0;
2616 state.in_subroutine = FALSE;
2617
2618 ctx.shader = shader;
2619 ctx.reg_maps = reg_maps;
2620 ctx.buffer = buffer;
2621 ctx.tex_mx = &tex_mx;
2622 ctx.state = &state;
2623 ctx.backend_data = backend_ctx;
2624 ins.ctx = &ctx;
2625
2626 fe->shader_read_header(fe_data, &ptr, &shader_version);
2627 if (start)
2628 ptr = start;
2629
2630 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2631 {
2632 /* Read opcode. */
2633 fe->shader_read_instruction(fe_data, &ptr, &ins);
2634
2635 /* Unknown opcode and its parameters. */
2636 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2637 {
2638 WARN("Encountered unrecognised or invalid instruction.\n");
2639 return WINED3DERR_INVALIDCALL;
2640 }
2641
2642 if (ins.predicate)
2643 FIXME("Predicates not implemented.\n");
2644
2645 /* Call appropriate function for output target */
2646 device->shader_backend->shader_handle_instruction(&ins);
2647 }
2648
2649 return WINED3D_OK;
2650 }
2651
2652 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2653 const struct wined3d_shader_dst_param *dst)
2654 {
2655 DWORD mmask = dst->modifiers;
2656
2657 switch (dst->shift)
2658 {
2659 case 0: break;
2660 case 13: shader_addline(buffer, "_d8"); break;
2661 case 14: shader_addline(buffer, "_d4"); break;
2662 case 15: shader_addline(buffer, "_d2"); break;
2663 case 1: shader_addline(buffer, "_x2"); break;
2664 case 2: shader_addline(buffer, "_x4"); break;
2665 case 3: shader_addline(buffer, "_x8"); break;
2666 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2667 }
2668
2669 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2670 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2671 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2672
2673 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2674 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2675 }
2676
2677 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2678 const struct wined3d_shader_primitive_type *primitive_type)
2679 {
2680 switch (primitive_type->type)
2681 {
2682 case WINED3D_PT_UNDEFINED:
2683 shader_addline(buffer, "undefined");
2684 break;
2685 case WINED3D_PT_POINTLIST:
2686 shader_addline(buffer, "pointlist");
2687 break;
2688 case WINED3D_PT_LINELIST:
2689 shader_addline(buffer, "linelist");
2690 break;
2691 case WINED3D_PT_LINESTRIP:
2692 shader_addline(buffer, "linestrip");
2693 break;
2694 case WINED3D_PT_TRIANGLELIST:
2695 shader_addline(buffer, "trianglelist");
2696 break;
2697 case WINED3D_PT_TRIANGLESTRIP:
2698 shader_addline(buffer, "trianglestrip");
2699 break;
2700 case WINED3D_PT_TRIANGLEFAN:
2701 shader_addline(buffer, "trianglefan");
2702 break;
2703 case WINED3D_PT_LINELIST_ADJ:
2704 shader_addline(buffer, "linelist_adj");
2705 break;
2706 case WINED3D_PT_LINESTRIP_ADJ:
2707 shader_addline(buffer, "linestrip_adj");
2708 break;
2709 case WINED3D_PT_TRIANGLELIST_ADJ:
2710 shader_addline(buffer, "trianglelist_adj");
2711 break;
2712 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2713 shader_addline(buffer, "trianglestrip_adj");
2714 break;
2715 case WINED3D_PT_PATCH:
2716 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2717 break;
2718 default:
2719 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2720 break;
2721 }
2722 }
2723
2724 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2725 enum wined3d_shader_interpolation_mode interpolation_mode)
2726 {
2727 switch (interpolation_mode)
2728 {
2729 case WINED3DSIM_CONSTANT:
2730 shader_addline(buffer, "constant");
2731 break;
2732 case WINED3DSIM_LINEAR:
2733 shader_addline(buffer, "linear");
2734 break;
2735 case WINED3DSIM_LINEAR_CENTROID:
2736 shader_addline(buffer, "linear centroid");
2737 break;
2738 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2739 shader_addline(buffer, "linear noperspective");
2740 break;
2741 case WINED3DSIM_LINEAR_SAMPLE:
2742 shader_addline(buffer, "linear sample");
2743 break;
2744 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2745 shader_addline(buffer, "linear noperspective centroid");
2746 break;
2747 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2748 shader_addline(buffer, "linear noperspective sample");
2749 break;
2750 default:
2751 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2752 break;
2753 }
2754 }
2755
2756 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2757 {
2758 struct wined3d_shader_version shader_version;
2759 struct wined3d_string_buffer buffer;
2760 const char *type_prefix;
2761 const char *p, *q;
2762 const DWORD *ptr;
2763 DWORD i;
2764
2765 if (!string_buffer_init(&buffer))
2766 {
2767 ERR("Failed to initialize string buffer.\n");
2768 return;
2769 }
2770
2771 fe->shader_read_header(fe_data, &ptr, &shader_version);
2772
2773 TRACE("Parsing %p.\n", ptr);
2774
2775 switch (shader_version.type)
2776 {
2777 case WINED3D_SHADER_TYPE_VERTEX:
2778 type_prefix = "vs";
2779 break;
2780
2781 case WINED3D_SHADER_TYPE_HULL:
2782 type_prefix = "hs";
2783 break;
2784
2785 case WINED3D_SHADER_TYPE_DOMAIN:
2786 type_prefix = "ds";
2787 break;
2788
2789 case WINED3D_SHADER_TYPE_GEOMETRY:
2790 type_prefix = "gs";
2791 break;
2792
2793 case WINED3D_SHADER_TYPE_PIXEL:
2794 type_prefix = "ps";
2795 break;
2796
2797 case WINED3D_SHADER_TYPE_COMPUTE:
2798 type_prefix = "cs";
2799 break;
2800
2801 default:
2802 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2803 type_prefix = "unknown";
2804 break;
2805 }
2806
2807 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2808
2809 while (!fe->shader_is_end(fe_data, &ptr))
2810 {
2811 struct wined3d_shader_instruction ins;
2812
2813 fe->shader_read_instruction(fe_data, &ptr, &ins);
2814 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2815 {
2816 WARN("Skipping unrecognized instruction.\n");
2817 shader_addline(&buffer, "<unrecognized instruction>\n");
2818 continue;
2819 }
2820
2821 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2822 {
2823 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2824 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2825 shader_addline(&buffer, " ");
2826 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2827 }
2828 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2829 {
2830 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2831 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2832 shader_addline(&buffer, ", %s",
2833 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2834 }
2835 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2836 {
2837 shader_addline(&buffer, "%s fb%u",
2838 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2839 }
2840 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2841 {
2842 shader_addline(&buffer, "%s ft%u = {...}",
2843 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2844 }
2845 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2846 {
2847 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2848 shader_dump_global_flags(&buffer, ins.flags);
2849 }
2850 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2851 {
2852 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2853 ins.declaration.max_tessellation_factor);
2854 }
2855 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2856 {
2857 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2858 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2859 {
2860 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2861 ins.declaration.icb->data[4 * i + 0],
2862 ins.declaration.icb->data[4 * i + 1],
2863 ins.declaration.icb->data[4 * i + 2],
2864 ins.declaration.icb->data[4 * i + 3]);
2865 }
2866 shader_addline(&buffer, "}");
2867 }
2868 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2869 {
2870 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2871 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2872 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2873 }
2874 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2875 {
2876 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2877 ins.declaration.indexable_temp.register_idx,
2878 ins.declaration.indexable_temp.register_size,
2879 ins.declaration.indexable_temp.component_count);
2880 }
2881 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2882 {
2883 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2884 shader_dump_interpolation_mode(&buffer, ins.flags);
2885 shader_addline(&buffer, " ");
2886 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2887 }
2888 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2889 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2890 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2891 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2892 {
2893 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2894 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2895 shader_addline(&buffer, ", ");
2896 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2897 }
2898 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2899 {
2900 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2901 shader_dump_interpolation_mode(&buffer, ins.flags);
2902 shader_addline(&buffer, " ");
2903 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2904 shader_addline(&buffer, ", ");
2905 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2906 }
2907 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2908 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2909 {
2910 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2911 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2912 }
2913 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2914 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2915 {
2916 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2917 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2918 }
2919 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2920 {
2921 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2922 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2923 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2924 }
2925 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2926 {
2927 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2928 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2929 }
2930 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2931 {
2932 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2933 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2934 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2935 }
2936 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2937 {
2938 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2939 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2940 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2941 shader_addline(&buffer, ", comparisonMode");
2942 }
2943 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2944 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2945 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2946 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2947 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2948 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2949 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2950 {
2951 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2952 }
2953 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2954 {
2955 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2956 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2957 }
2958 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2959 {
2960 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2961 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2962 }
2963 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2964 {
2965 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2966 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2967 }
2968 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2969 {
2970 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2971 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2972 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2973 }
2974 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2975 {
2976 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2977 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2978 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2979 ins.declaration.tgsm_structured.structure_count);
2980 }
2981 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2982 {
2983 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2984 ins.declaration.thread_group_size.x,
2985 ins.declaration.thread_group_size.y,
2986 ins.declaration.thread_group_size.z);
2987 }
2988 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2989 {
2990 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2991 shader_dump_uav_flags(&buffer, ins.flags);
2992 shader_addline(&buffer, " ");
2993 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2994 }
2995 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2996 {
2997 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2998 shader_dump_uav_flags(&buffer, ins.flags);
2999 shader_addline(&buffer, " ");
3000 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
3001 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
3002 }
3003 else if (ins.handler_idx == WINED3DSIH_DEF)
3004 {
3005 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
3006 ins.dst[0].reg.idx[0].offset),
3007 *(const float *)&ins.src[0].reg.u.immconst_data[0],
3008 *(const float *)&ins.src[0].reg.u.immconst_data[1],
3009 *(const float *)&ins.src[0].reg.u.immconst_data[2],
3010 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
3011 }
3012 else if (ins.handler_idx == WINED3DSIH_DEFI)
3013 {
3014 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
3015 ins.src[0].reg.u.immconst_data[0],
3016 ins.src[0].reg.u.immconst_data[1],
3017 ins.src[0].reg.u.immconst_data[2],
3018 ins.src[0].reg.u.immconst_data[3]);
3019 }
3020 else if (ins.handler_idx == WINED3DSIH_DEFB)
3021 {
3022 shader_addline(&buffer, "defb b%u = %s",
3023 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
3024 }
3025 else
3026 {
3027 if (ins.predicate)
3028 {
3029 shader_addline(&buffer, "(");
3030 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
3031 shader_addline(&buffer, ") ");
3032 }
3033
3034 /* PixWin marks instructions with the coissue flag with a '+' */
3035 if (ins.coissue)
3036 shader_addline(&buffer, "+");
3037
3038 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3039
3040 if (ins.handler_idx == WINED3DSIH_BREAKP
3041 || ins.handler_idx == WINED3DSIH_CONTINUEP
3042 || ins.handler_idx == WINED3DSIH_IF
3043 || ins.handler_idx == WINED3DSIH_RETP
3044 || ins.handler_idx == WINED3DSIH_TEXKILL)
3045 {
3046 switch (ins.flags)
3047 {
3048 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3049 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3050 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3051 }
3052 }
3053 else if (ins.handler_idx == WINED3DSIH_IFC
3054 || ins.handler_idx == WINED3DSIH_BREAKC)
3055 {
3056 switch (ins.flags)
3057 {
3058 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3059 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3060 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3061 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3062 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3063 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3064 default: shader_addline(&buffer, "_(%u)", ins.flags);
3065 }
3066 }
3067 else if (ins.handler_idx == WINED3DSIH_TEX
3068 && shader_version.major >= 2
3069 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3070 {
3071 shader_addline(&buffer, "p");
3072 }
3073 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3074 {
3075 switch (ins.flags)
3076 {
3077 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3078 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3079 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3080 }
3081 }
3082 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3083 {
3084 switch (ins.flags)
3085 {
3086 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3087 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3088 }
3089 }
3090 else if (ins.handler_idx == WINED3DSIH_SYNC)
3091 {
3092 shader_dump_sync_flags(&buffer, ins.flags);
3093 }
3094 else
3095 {
3096 shader_dump_precise_flags(&buffer, ins.flags);
3097 }
3098
3099 if (wined3d_shader_instruction_has_texel_offset(&ins))
3100 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3101
3102 if (ins.resource_type != WINED3D_SHADER_RESOURCE_NONE)
3103 {
3104 shader_addline(&buffer, "(");
3105 shader_dump_resource_type(&buffer, ins.resource_type);
3106 shader_addline(&buffer, ")");
3107 }
3108
3109 if (ins.resource_data_type != WINED3D_DATA_FLOAT)
3110 shader_dump_data_type(&buffer, ins.resource_data_type);
3111
3112 for (i = 0; i < ins.dst_count; ++i)
3113 {
3114 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
3115 shader_addline(&buffer, !i ? " " : ", ");
3116 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
3117 }
3118
3119 /* Other source tokens */
3120 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
3121 {
3122 shader_addline(&buffer, !i ? " " : ", ");
3123 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
3124 }
3125 }
3126 shader_addline(&buffer, "\n");
3127 }
3128
3129 for (p = buffer.buffer; *p; p = q)
3130 {
3131 if (!(q = strstr(p, "\n")))
3132 q = p + strlen(p);
3133 else
3134 ++q;
3135 TRACE(" %.*s", (int)(q - p), p);
3136 }
3137
3138 string_buffer_free(&buffer);
3139 }
3140
3141 static void shader_cleanup(struct wined3d_shader *shader)
3142 {
3143 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3144 {
3145 heap_free(shader->u.hs.phases.control_point);
3146 heap_free(shader->u.hs.phases.fork);
3147 heap_free(shader->u.hs.phases.join);
3148 }
3149
3150 heap_free(shader->patch_constant_signature.elements);
3151 heap_free(shader->output_signature.elements);
3152 heap_free(shader->input_signature.elements);
3153 shader->device->shader_backend->shader_destroy(shader);
3154 shader_cleanup_reg_maps(&shader->reg_maps);
3155 heap_free(shader->byte_code);
3156 shader_delete_constant_list(&shader->constantsF);
3157 shader_delete_constant_list(&shader->constantsB);
3158 shader_delete_constant_list(&shader->constantsI);
3159 list_remove(&shader->shader_list_entry);
3160
3161 if (shader->frontend && shader->frontend_data)
3162 shader->frontend->shader_free(shader->frontend_data);
3163 }
3164
3165 struct shader_none_priv
3166 {
3167 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3168 const struct wined3d_fragment_pipe_ops *fragment_pipe;
3169 BOOL ffp_proj_control;
3170 };
3171
3172 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3173 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3174 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3175 const struct wined3d_state *state) {}
3176 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3177 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3178 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3179 const struct wined3d_state *state) {}
3180 static void shader_none_destroy(struct wined3d_shader *shader) {}
3181 static void shader_none_free_context_data(struct wined3d_context *context) {}
3182 static void shader_none_init_context_state(struct wined3d_context *context) {}
3183
3184 /* Context activation is done by the caller. */
3185 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3186 const struct wined3d_state *state)
3187 {
3188 struct shader_none_priv *priv = shader_priv;
3189
3190 priv->vertex_pipe->vp_enable(context, !use_vs(state));
3191 priv->fragment_pipe->fp_enable(context, !use_ps(state));
3192 }
3193
3194 /* Context activation is done by the caller. */
3195 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3196 {
3197 struct shader_none_priv *priv = shader_priv;
3198
3199 priv->vertex_pipe->vp_enable(context, FALSE);
3200 priv->fragment_pipe->fp_enable(context, FALSE);
3201
3202 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3203 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3204 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3205 | (1u << WINED3D_SHADER_TYPE_HULL)
3206 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3207 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3208 }
3209
3210 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3211 const struct wined3d_fragment_pipe_ops *fragment_pipe)
3212 {
3213 struct fragment_caps fragment_caps;
3214 void *vertex_priv, *fragment_priv;
3215 struct shader_none_priv *priv;
3216
3217 if (!(priv = heap_alloc(sizeof(*priv))))
3218 return E_OUTOFMEMORY;
3219
3220 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3221 {
3222 ERR("Failed to initialize vertex pipe.\n");
3223 heap_free(priv);
3224 return E_FAIL;
3225 }
3226
3227 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3228 {
3229 ERR("Failed to initialize fragment pipe.\n");
3230 vertex_pipe->vp_free(device, NULL);
3231 heap_free(priv);
3232 return E_FAIL;
3233 }
3234
3235 priv->vertex_pipe = vertex_pipe;
3236 priv->fragment_pipe = fragment_pipe;
3237 fragment_pipe->get_caps(device->adapter, &fragment_caps);
3238 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3239
3240 device->vertex_priv = vertex_priv;
3241 device->fragment_priv = fragment_priv;
3242 device->shader_priv = priv;
3243
3244 return WINED3D_OK;
3245 }
3246
3247 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
3248 {
3249 struct shader_none_priv *priv = device->shader_priv;
3250
3251 priv->fragment_pipe->free_private(device, context);
3252 priv->vertex_pipe->vp_free(device, context);
3253 heap_free(priv);
3254 }
3255
3256 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3257 {
3258 return TRUE;
3259 }
3260
3261 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
3262 {
3263 /* Set the shader caps to 0 for the none shader backend */
3264 memset(caps, 0, sizeof(*caps));
3265 }
3266
3267 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3268 {
3269 /* We "support" every possible fixup, since we don't support any shader
3270 * model, and will never have to actually sample a texture. */
3271 return TRUE;
3272 }
3273
3274 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3275 {
3276 struct shader_none_priv *priv = shader_priv;
3277
3278 return priv->ffp_proj_control;
3279 }
3280
3281 static uint64_t shader_none_shader_compile(struct wined3d_context *context, const struct wined3d_shader_desc *shader_desc,
3282 enum wined3d_shader_type shader_type)
3283 {
3284 return 0;
3285 }
3286
3287 const struct wined3d_shader_backend_ops none_shader_backend =
3288 {
3289 shader_none_handle_instruction,
3290 shader_none_precompile,
3291 shader_none_select,
3292 shader_none_select_compute,
3293 shader_none_disable,
3294 shader_none_update_float_vertex_constants,
3295 shader_none_update_float_pixel_constants,
3296 shader_none_load_constants,
3297 shader_none_destroy,
3298 shader_none_alloc,
3299 shader_none_free,
3300 shader_none_allocate_context_data,
3301 shader_none_free_context_data,
3302 shader_none_init_context_state,
3303 shader_none_get_caps,
3304 shader_none_color_fixup_supported,
3305 shader_none_has_ffp_proj_control,
3306 shader_none_shader_compile,
3307 };
3308
3309 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
3310 {
3311 switch (level)
3312 {
3313 case WINED3D_FEATURE_LEVEL_11_1:
3314 case WINED3D_FEATURE_LEVEL_11:
3315 return 5;
3316 case WINED3D_FEATURE_LEVEL_10_1:
3317 case WINED3D_FEATURE_LEVEL_10:
3318 return 4;
3319 case WINED3D_FEATURE_LEVEL_9_3:
3320 return 3;
3321 case WINED3D_FEATURE_LEVEL_9_2:
3322 case WINED3D_FEATURE_LEVEL_9_1:
3323 return 2;
3324 default:
3325 return 1;
3326 }
3327 }
3328
3329 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
3330 enum wined3d_shader_type type, unsigned int float_const_count)
3331 {
3332 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3333 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3334 const struct wined3d_shader_version *version = &reg_maps->shader_version;
3335 const struct wined3d_shader_frontend *fe;
3336 unsigned int backend_version;
3337 HRESULT hr;
3338
3339 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
3340 shader, device, debug_shader_type(type), float_const_count);
3341
3342 fe = shader->frontend;
3343 if (!(shader->frontend_data = fe->shader_init(shader->function,
3344 shader->functionLength, &shader->output_signature)))
3345 {
3346 FIXME("Failed to initialize frontend.\n");
3347 return WINED3DERR_INVALIDCALL;
3348 }
3349
3350 /* First pass: trace shader. */
3351 if (TRACE_ON(d3d_shader))
3352 shader_trace_init(fe, shader->frontend_data);
3353
3354 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3355 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
3356 return hr;
3357
3358 if (version->type != type)
3359 {
3360 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
3361 return WINED3DERR_INVALIDCALL;
3362 }
3363 if (version->major > shader_max_version_from_feature_level(device->cs->c.state->feature_level))
3364 {
3365 WARN("Shader version %u not supported by this device.\n", version->major);
3366 return WINED3DERR_INVALIDCALL;
3367 }
3368 switch (type)
3369 {
3370 case WINED3D_SHADER_TYPE_VERTEX:
3371 backend_version = d3d_info->limits.vs_version;
3372 break;
3373 case WINED3D_SHADER_TYPE_HULL:
3374 backend_version = d3d_info->limits.hs_version;
3375 break;
3376 case WINED3D_SHADER_TYPE_DOMAIN:
3377 backend_version = d3d_info->limits.ds_version;
3378 break;
3379 case WINED3D_SHADER_TYPE_GEOMETRY:
3380 backend_version = d3d_info->limits.gs_version;
3381 break;
3382 case WINED3D_SHADER_TYPE_PIXEL:
3383 backend_version = d3d_info->limits.ps_version;
3384 break;
3385 case WINED3D_SHADER_TYPE_COMPUTE:
3386 backend_version = d3d_info->limits.cs_version;
3387 break;
3388 default:
3389 FIXME("No backend version-checking for this shader type.\n");
3390 backend_version = 0;
3391 }
3392 if (version->major > backend_version)
3393 {
3394 WARN("Shader version %u.%u not supported by the current shader backend.\n",
3395 version->major, version->minor);
3396 return WINED3DERR_INVALIDCALL;
3397 }
3398
3399 shader->load_local_constsF = shader->lconst_inf_or_nan;
3400
3401 return WINED3D_OK;
3402 }
3403
3404 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3405 {
3406 ULONG refcount = InterlockedIncrement(&shader->ref);
3407
3408 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3409
3410 return refcount;
3411 }
3412
3413 static void wined3d_shader_init_object(void *object)
3414 {
3415 struct wined3d_shader *shader = object;
3416 struct wined3d_device *device = shader->device;
3417
3418 TRACE("shader %p.\n", shader);
3419
3420 list_add_head(&device->shaders, &shader->shader_list_entry);
3421
3422 device->shader_backend->shader_precompile(device->shader_priv, shader);
3423 }
3424
3425 static void wined3d_shader_destroy_object(void *object)
3426 {
3427 TRACE("object %p.\n", object);
3428
3429 shader_cleanup(object);
3430 heap_free(object);
3431 }
3432
3433 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3434 {
3435 ULONG refcount = InterlockedDecrement(&shader->ref);
3436
3437 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3438
3439 if (!refcount)
3440 {
3441 wined3d_mutex_lock();
3442 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3443 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3444 wined3d_mutex_unlock();
3445 }
3446
3447 return refcount;
3448 }
3449
3450 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3451 {
3452 TRACE("shader %p.\n", shader);
3453
3454 return shader->parent;
3455 }
3456
3457 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3458 void *byte_code, UINT *byte_code_size)
3459 {
3460 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3461
3462 if (!byte_code)
3463 {
3464 *byte_code_size = shader->byte_code_size;
3465 return WINED3D_OK;
3466 }
3467
3468 if (*byte_code_size < shader->byte_code_size)
3469 {
3470 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3471 * than the required size we should write the required size and
3472 * return D3DERR_MOREDATA. That's not actually true. */
3473 return WINED3DERR_INVALIDCALL;
3474 }
3475
3476 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
3477
3478 return WINED3D_OK;
3479 }
3480
3481 /* Set local constants for d3d8 shaders. */
3482 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3483 UINT start_idx, const float *src_data, UINT count)
3484 {
3485 UINT end_idx = start_idx + count;
3486 UINT i;
3487
3488 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3489
3490 if (end_idx > shader->limits->constant_float)
3491 {
3492 WARN("end_idx %u > float constants limit %u.\n",
3493 end_idx, shader->limits->constant_float);
3494 end_idx = shader->limits->constant_float;
3495 }
3496
3497 for (i = start_idx; i < end_idx; ++i)
3498 {
3499 struct wined3d_shader_lconst *lconst;
3500 float *value;
3501
3502 if (!(lconst = heap_alloc(sizeof(*lconst))))
3503 return E_OUTOFMEMORY;
3504
3505 lconst->idx = i;
3506 value = (float *)lconst->value;
3507 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3508 list_add_head(&shader->constantsF, &lconst->entry);
3509
3510 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3511 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3512 {
3513 shader->lconst_inf_or_nan = TRUE;
3514 }
3515 }
3516
3517 return WINED3D_OK;
3518 }
3519
3520 static void init_interpolation_compile_args(DWORD *interpolation_args,
3521 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
3522 {
3523 if (!d3d_info->shader_output_interpolation || !pixel_shader
3524 || pixel_shader->reg_maps.shader_version.major < 4)
3525 {
3526 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3527 return;
3528 }
3529
3530 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3531 sizeof(pixel_shader->u.ps.interpolation_mode));
3532 }
3533
3534 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3535 struct vs_compile_args *args, const struct wined3d_context *context)
3536 {
3537 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3538 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3539 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3540 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3541 WORD swizzle_map = context->stream_info.swizzle_map;
3542
3543 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3544 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3545 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3546 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3547 args->point_size = state->primitive_type == WINED3D_PT_POINTLIST;
3548 args->per_vertex_point_size = shader->reg_maps.point_size;
3549 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
3550 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3551 if (shader->reg_maps.shader_version.major >= 4)
3552 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3553 : geometry_shader ? geometry_shader->limits->packed_input
3554 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3555 else
3556 args->next_shader_input_count = 0;
3557 args->swizzle_map = swizzle_map;
3558 if (d3d_info->emulated_flatshading)
3559 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3560 else
3561 args->flatshading = 0;
3562
3563 init_interpolation_compile_args(args->interpolation_mode,
3564 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
3565 }
3566
3567 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3568 {
3569 if (usage_idx1 != usage_idx2)
3570 return FALSE;
3571 if (usage1 == usage2)
3572 return TRUE;
3573 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3574 return TRUE;
3575 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3576 return TRUE;
3577
3578 return FALSE;
3579 }
3580
3581 BOOL vshader_get_input(const struct wined3d_shader *shader,
3582 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3583 {
3584 WORD map = shader->reg_maps.input_registers;
3585 unsigned int i;
3586
3587 for (i = 0; map; map >>= 1, ++i)
3588 {
3589 if (!(map & 1)) continue;
3590
3591 if (match_usage(shader->u.vs.attributes[i].usage,
3592 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3593 {
3594 *regnum = i;
3595 return TRUE;
3596 }
3597 }
3598 return FALSE;
3599 }
3600
3601 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3602 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3603 {
3604 HRESULT hr;
3605
3606 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
3607
3608 if (!desc->byte_code)
3609 return WINED3DERR_INVALIDCALL;
3610
3611 shader->ref = 1;
3612 shader->device = device;
3613 shader->parent = parent;
3614 shader->parent_ops = parent_ops;
3615
3616 list_init(&shader->linked_programs);
3617 list_init(&shader->constantsF);
3618 list_init(&shader->constantsB);
3619 list_init(&shader->constantsI);
3620 shader->lconst_inf_or_nan = FALSE;
3621 list_init(&shader->reg_maps.indexable_temps);
3622 list_init(&shader->shader_list_entry);
3623
3624 if (desc->byte_code_size == ~(size_t)0)
3625 {
3626 struct wined3d_shader_version shader_version;
3627 const struct wined3d_shader_frontend *fe;
3628 struct wined3d_shader_instruction ins;
3629 const DWORD *ptr;
3630 void *fe_data;
3631
3632 if (!(shader->frontend = shader_select_frontend(WINED3D_SHADER_BYTE_CODE_FORMAT_SM1)))
3633 {
3634 FIXME("Unable to find frontend for shader.\n");
3635 hr = WINED3DERR_INVALIDCALL;
3636 goto fail;
3637 }
3638
3639 fe = shader->frontend;
3640 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
3641 {
3642 WARN("Failed to initialise frontend data.\n");
3643 hr = WINED3DERR_INVALIDCALL;
3644 goto fail;
3645 }
3646
3647 fe->shader_read_header(fe_data, &ptr, &shader_version);
3648 while (!fe->shader_is_end(fe_data, &ptr))
3649 fe->shader_read_instruction(fe_data, &ptr, &ins);
3650
3651 fe->shader_free(fe_data);
3652
3653 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3654
3655 if (!(shader->byte_code = heap_alloc(shader->byte_code_size)))
3656 {
3657 hr = E_OUTOFMEMORY;
3658 goto fail;
3659 }
3660 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
3661
3662 shader->function = shader->byte_code;
3663 shader->functionLength = shader->byte_code_size;
3664 }
3665 else
3666 {
3667 enum wined3d_shader_byte_code_format format;
3668 unsigned int max_version;
3669
3670 if (!(shader->byte_code = heap_alloc(desc->byte_code_size)))
3671 {
3672 hr = E_OUTOFMEMORY;
3673 goto fail;
3674 }
3675 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
3676 shader->byte_code_size = desc->byte_code_size;
3677
3678 max_version = shader_max_version_from_feature_level(device->cs->c.state->feature_level);
3679 if (FAILED(hr = shader_extract_from_dxbc(shader, max_version, &format)))
3680 goto fail;
3681
3682 if (!(shader->frontend = shader_select_frontend(format)))
3683 {
3684 FIXME("Unable to find frontend for shader.\n");
3685 hr = WINED3DERR_INVALIDCALL;
3686 goto fail;
3687 }
3688 }
3689
3690 return WINED3D_OK;
3691
3692 fail:
3693 shader_cleanup(shader);
3694 return hr;
3695 }
3696
3697 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3698 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3699 {
3700 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3701 unsigned int i;
3702 HRESULT hr;
3703
3704 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3705 return hr;
3706
3707 if (FAILED(hr = shader_set_function(shader, device,
3708 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
3709 {
3710 shader_cleanup(shader);
3711 return hr;
3712 }
3713
3714 for (i = 0; i < shader->input_signature.element_count; ++i)
3715 {
3716 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3717
3718 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3719 continue;
3720
3721 shader->u.vs.attributes[input->register_idx].usage =
3722 shader_usage_from_semantic_name(input->semantic_name);
3723 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3724 }
3725
3726 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3727 shader->load_local_constsF = TRUE;
3728
3729 return WINED3D_OK;
3730 }
3731
3732 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
3733 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
3734 {
3735 struct wined3d_shader_signature_element *e = s->elements;
3736 unsigned int i;
3737
3738 for (i = 0; i < s->element_count; ++i)
3739 {
3740 if (e[i].stream_idx == stream_idx
3741 && !_strnicmp(e[i].semantic_name, semantic_name, -1)
3742 && e[i].semantic_idx == semantic_idx)
3743 return &e[i];
3744 }
3745
3746 return NULL;
3747 }
3748
3749 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
3750 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
3751 {
3752 const struct wined3d_shader_signature_element *output;
3753 unsigned int idx;
3754
3755 if (!(output = shader_find_signature_element(&shader->output_signature,
3756 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
3757 return FALSE;
3758
3759 for (idx = 0; idx < 4; ++idx)
3760 {
3761 if (output->mask & (1u << idx))
3762 break;
3763 }
3764 idx += so_element->component_idx;
3765
3766 *register_idx = output->register_idx;
3767 *component_idx = idx;
3768 return TRUE;
3769 }
3770
3771 static HRESULT geometry_shader_init_so_desc(struct wined3d_geometry_shader *gs, struct wined3d_device *device,
3772 const struct wined3d_stream_output_desc *so_desc)
3773 {
3774 struct wined3d_so_desc_entry *s;
3775 struct wine_rb_entry *entry;
3776 unsigned int i;
3777 size_t size;
3778 char *name;
3779
3780 if ((entry = wine_rb_get(&device->so_descs, so_desc)))
3781 {
3782 gs->so_desc = &WINE_RB_ENTRY_VALUE(entry, struct wined3d_so_desc_entry, entry)->desc;
3783 return WINED3D_OK;
3784 }
3785
3786 size = FIELD_OFFSET(struct wined3d_so_desc_entry, elements[so_desc->element_count]);
3787 for (i = 0; i < so_desc->element_count; ++i)
3788 {
3789 const char *n = so_desc->elements[i].semantic_name;
3790
3791 if (n)
3792 size += strlen(n) + 1;
3793 }
3794 if (!(s = heap_alloc(size)))
3795 return E_OUTOFMEMORY;
3796
3797 s->desc = *so_desc;
3798
3799 memcpy(s->elements, so_desc->elements, so_desc->element_count * sizeof(*s->elements));
3800 s->desc.elements = s->elements;
3801
3802 name = (char *)&s->elements[s->desc.element_count];
3803 for (i = 0; i < so_desc->element_count; ++i)
3804 {
3805 struct wined3d_stream_output_element *e = &s->elements[i];
3806
3807 if (!e->semantic_name)
3808 continue;
3809
3810 size = strlen(e->semantic_name) + 1;
3811 memcpy(name, e->semantic_name, size);
3812 e->semantic_name = name;
3813 name += size;
3814 }
3815
3816 if (wine_rb_put(&device->so_descs, &s->desc, &s->entry) == -1)
3817 {
3818 heap_free(s);
3819 return E_FAIL;
3820 }
3821 gs->so_desc = &s->desc;
3822
3823 return WINED3D_OK;
3824 }
3825
3826 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
3827 const struct wined3d_stream_output_desc *so_desc)
3828 {
3829 const struct wined3d_shader_frontend *fe = shader->frontend;
3830 const struct wined3d_shader_signature_element *output;
3831 unsigned int i, component_idx, register_idx, mask;
3832 struct wined3d_shader_version shader_version;
3833 const DWORD *ptr;
3834 void *fe_data;
3835 HRESULT hr;
3836
3837 if (!so_desc)
3838 return WINED3D_OK;
3839
3840 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
3841 {
3842 WARN("Failed to initialise frontend data.\n");
3843 return WINED3DERR_INVALIDCALL;
3844 }
3845 fe->shader_read_header(fe_data, &ptr, &shader_version);
3846 fe->shader_free(fe_data);
3847
3848 switch (shader_version.type)
3849 {
3850 case WINED3D_SHADER_TYPE_VERTEX:
3851 case WINED3D_SHADER_TYPE_DOMAIN:
3852 shader->function = NULL;
3853 shader->functionLength = 0;
3854 break;
3855 case WINED3D_SHADER_TYPE_GEOMETRY:
3856 break;
3857 default:
3858 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
3859 return E_INVALIDARG;
3860 }
3861
3862 if (!shader->function)
3863 {
3864 shader->reg_maps.shader_version = shader_version;
3865 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
3866 shader_set_limits(shader);
3867 if (FAILED(hr = shader_scan_output_signature(shader)))
3868 return hr;
3869 }
3870
3871 for (i = 0; i < so_desc->element_count; ++i)
3872 {
3873 const struct wined3d_stream_output_element *e = &so_desc->elements[i];
3874
3875 if (!e->semantic_name)
3876 continue;
3877 if (!(output = shader_find_signature_element(&shader->output_signature,
3878 e->stream_idx, e->semantic_name, e->semantic_idx))
3879 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
3880 {
3881 WARN("Failed to find output signature element for stream output entry.\n");
3882 return E_INVALIDARG;
3883 }
3884
3885 mask = wined3d_mask_from_size(e->component_count) << component_idx;
3886 if ((output->mask & 0xff & mask) != mask)
3887 {
3888 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
3889 component_idx, e->component_count, mask, output->mask & 0xff);
3890 return E_INVALIDARG;
3891 }
3892 }
3893
3894 if (FAILED(hr = geometry_shader_init_so_desc(&shader->u.gs, shader->device, so_desc)))
3895 {
3896 WARN("Failed to initialise stream output description, hr %#x.\n", hr);
3897 return hr;
3898 }
3899
3900 return WINED3D_OK;
3901 }
3902
3903 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3904 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3905 void *parent, const struct wined3d_parent_ops *parent_ops)
3906 {
3907 HRESULT hr;
3908
3909 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3910 return hr;
3911
3912 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
3913 goto fail;
3914
3915 if (shader->function
3916 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
3917 goto fail;
3918
3919 return WINED3D_OK;
3920
3921 fail:
3922 shader_cleanup(shader);
3923 return hr;
3924 }
3925
3926 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3927 struct ds_compile_args *args, const struct wined3d_context *context)
3928 {
3929 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3930 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3931 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3932
3933 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3934 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3935
3936 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3937 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3938 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3939
3940 args->render_offscreen = context->render_offscreen;
3941
3942 init_interpolation_compile_args(args->interpolation_mode,
3943 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
3944
3945 args->padding = 0;
3946 }
3947
3948 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3949 struct gs_compile_args *args, const struct wined3d_context *context)
3950 {
3951 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3952
3953 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3954
3955 if (!(args->primitive_type = shader->u.gs.input_type))
3956 args->primitive_type = state->primitive_type;
3957
3958 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
3959 }
3960
3961 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3962 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3963 {
3964 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3965 struct wined3d_texture *texture;
3966 unsigned int i;
3967
3968 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3969 if (!d3d_info->srgb_write_control && needs_srgb_write(d3d_info, state, &state->fb))
3970 {
3971 static unsigned int warned = 0;
3972
3973 args->srgb_correction = 1;
3974 if (state->blend_state && state->blend_state->desc.rt[0].enable && !warned++)
3975 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3976 "support, expect rendering artifacts.\n");
3977 }
3978
3979 if (shader->reg_maps.shader_version.major == 1
3980 && shader->reg_maps.shader_version.minor <= 3)
3981 {
3982 for (i = 0; i < shader->limits->sampler; ++i)
3983 {
3984 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3985
3986 if (flags & WINED3D_TTFF_PROJECTED)
3987 {
3988 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3989
3990 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3991 {
3992 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3993 unsigned int j;
3994 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3995 DWORD max_valid = WINED3D_TTFF_COUNT4;
3996
3997 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3998 {
3999 struct wined3d_vertex_declaration_element *element =
4000 &state->vertex_declaration->elements[j];
4001
4002 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
4003 && element->usage_idx == index)
4004 {
4005 max_valid = element->format->component_count;
4006 break;
4007 }
4008 }
4009 if (!tex_transform || tex_transform > max_valid)
4010 {
4011 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
4012 tex_transform, max_valid);
4013 tex_transform = max_valid;
4014 }
4015 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
4016 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
4017 && tex_transform > WINED3D_TTFF_COUNT2)
4018 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4019 && tex_transform > WINED3D_TTFF_COUNT3))
4020 tex_transform |= WINED3D_PSARGS_PROJECTED;
4021 else
4022 {
4023 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
4024 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
4025 i, tex_transform, resource_type);
4026 }
4027 }
4028 else
4029 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
4030
4031 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
4032 }
4033 }
4034 }
4035 if (shader->reg_maps.shader_version.major == 1
4036 && shader->reg_maps.shader_version.minor <= 4)
4037 {
4038 for (i = 0; i < shader->limits->sampler; ++i)
4039 {
4040 if (!shader->reg_maps.resource_info[i].type)
4041 continue;
4042
4043 /* Treat unbound textures as 2D. The dummy texture will provide
4044 * the proper sample value. The tex_types bitmap defaults to
4045 * 2D because of the memset. */
4046 if (!(texture = state->textures[i]))
4047 continue;
4048
4049 switch (wined3d_texture_gl(texture)->target)
4050 {
4051 /* RECT textures are distinguished from 2D textures via np2_fixup */
4052 default:
4053 break;
4054
4055 case GL_TEXTURE_3D:
4056 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4057 break;
4058
4059 case GL_TEXTURE_CUBE_MAP_ARB:
4060 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4061 break;
4062 }
4063 }
4064 }
4065 else if (shader->reg_maps.shader_version.major <= 3)
4066 {
4067 for (i = 0; i < shader->limits->sampler; ++i)
4068 {
4069 enum wined3d_shader_resource_type resource_type;
4070 enum wined3d_shader_tex_types tex_type;
4071
4072 if (!(resource_type = shader->reg_maps.resource_info[i].type))
4073 continue;
4074
4075 switch (resource_type)
4076 {
4077 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
4078 tex_type = WINED3D_SHADER_TEX_3D;
4079 break;
4080 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
4081 tex_type = WINED3D_SHADER_TEX_CUBE;
4082 break;
4083 default:
4084 tex_type = WINED3D_SHADER_TEX_2D;
4085 break;
4086 }
4087
4088 if ((texture = state->textures[i]))
4089 {
4090 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
4091 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4092 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
4093 tex_type = WINED3D_SHADER_TEX_2D;
4094 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
4095 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
4096 tex_type = WINED3D_SHADER_TEX_3D;
4097 }
4098 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4099 }
4100 }
4101
4102 if (shader->reg_maps.shader_version.major >= 4)
4103 {
4104 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
4105 args->shadow = 0;
4106 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4107 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4108 args->np2_fixup = 0;
4109 }
4110 else
4111 {
4112 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4113 {
4114 if (!shader->reg_maps.resource_info[i].type)
4115 continue;
4116
4117 texture = state->textures[i];
4118 if (!texture)
4119 {
4120 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4121 continue;
4122 }
4123 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
4124 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4125 else
4126 args->color_fixup[i] = texture->resource.format->color_fixup;
4127
4128 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
4129 args->shadow |= 1u << i;
4130
4131 /* Flag samplers that need NP2 texcoord fixup. */
4132 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
4133 args->np2_fixup |= (1u << i);
4134 }
4135 }
4136
4137 if (shader->reg_maps.shader_version.major >= 3)
4138 {
4139 if (position_transformed)
4140 args->vp_mode = WINED3D_VP_MODE_NONE;
4141 else if (use_vs(state))
4142 args->vp_mode = WINED3D_VP_MODE_SHADER;
4143 else
4144 args->vp_mode = WINED3D_VP_MODE_FF;
4145 args->fog = WINED3D_FFP_PS_FOG_OFF;
4146 }
4147 else
4148 {
4149 args->vp_mode = WINED3D_VP_MODE_SHADER;
4150 if (state->render_states[WINED3D_RS_FOGENABLE])
4151 {
4152 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
4153 {
4154 case WINED3D_FOG_NONE:
4155 if (position_transformed || use_vs(state))
4156 {
4157 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
4158 break;
4159 }
4160
4161 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
4162 {
4163 case WINED3D_FOG_NONE: /* Fall through. */
4164 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4165 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4166 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4167 }
4168 break;
4169
4170 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4171 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4172 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4173 }
4174 }
4175 else
4176 {
4177 args->fog = WINED3D_FFP_PS_FOG_OFF;
4178 }
4179 }
4180
4181 if (!d3d_info->full_ffp_varyings)
4182 {
4183 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
4184
4185 args->texcoords_initialized = 0;
4186 for (i = 0; i < WINED3D_MAX_TEXTURES; ++i)
4187 {
4188 if (vs)
4189 {
4190 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
4191 args->texcoords_initialized |= 1u << i;
4192 }
4193 else
4194 {
4195 const struct wined3d_stream_info *si = &context->stream_info;
4196 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
4197
4198 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
4199 & WINED3D_FFP_TCI_MASK
4200 || (coord_idx < WINED3D_MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
4201 args->texcoords_initialized |= 1u << i;
4202 }
4203 }
4204 }
4205 else
4206 {
4207 args->texcoords_initialized = wined3d_mask_from_size(WINED3D_MAX_TEXTURES);
4208 }
4209
4210 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
4211 && state->primitive_type == WINED3D_PT_POINTLIST;
4212
4213 if (d3d_info->ffp_alpha_test)
4214 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
4215 else
4216 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4217 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4218 : WINED3D_CMP_ALWAYS) - 1;
4219
4220 if (d3d_info->emulated_flatshading)
4221 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4222
4223 args->y_correction = (shader->reg_maps.vpos && d3d_info->frag_coord_correction)
4224 || (shader->reg_maps.usesdsy && wined3d_settings.offscreen_rendering_mode != ORM_FBO)
4225 ? !context->render_offscreen : 0;
4226
4227 for (i = 0; i < ARRAY_SIZE(state->fb.render_targets); ++i)
4228 {
4229 struct wined3d_rendertarget_view *rtv = state->fb.render_targets[i];
4230 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
4231 args->rt_alpha_swizzle |= 1u << i;
4232 }
4233
4234 args->dual_source_blend = state->blend_state && state->blend_state->dual_source;
4235 }
4236
4237 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4238 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4239 {
4240 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4241 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4242 HRESULT hr;
4243
4244 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
4245 return hr;
4246
4247 if (FAILED(hr = shader_set_function(shader, device,
4248 WINED3D_SHADER_TYPE_PIXEL, device->adapter->d3d_info.limits.ps_uniform_count)))
4249 {
4250 shader_cleanup(shader);
4251 return hr;
4252 }
4253
4254 for (i = 0; i < MAX_REG_INPUT; ++i)
4255 {
4256 if (shader->u.ps.input_reg_used & (1u << i))
4257 {
4258 ++num_regs_used;
4259 highest_reg_used = i;
4260 }
4261 }
4262
4263 /* Don't do any register mapping magic if it is not needed, or if we can't
4264 * achieve anything anyway */
4265 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4266 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4267 || shader->reg_maps.shader_version.major >= 4)
4268 {
4269 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4270 {
4271 /* This happens with relative addressing. The input mapper function
4272 * warns about this if the higher registers are declared too, so
4273 * don't write a FIXME here */
4274 WARN("More varying registers used than supported\n");
4275 }
4276
4277 for (i = 0; i < MAX_REG_INPUT; ++i)
4278 {
4279 shader->u.ps.input_reg_map[i] = i;
4280 }
4281
4282 shader->u.ps.declared_in_count = highest_reg_used + 1;
4283 }
4284 else
4285 {
4286 shader->u.ps.declared_in_count = 0;
4287 for (i = 0; i < MAX_REG_INPUT; ++i)
4288 {
4289 if (shader->u.ps.input_reg_used & (1u << i))
4290 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4291 else shader->u.ps.input_reg_map[i] = ~0U;
4292 }
4293 }
4294
4295 return WINED3D_OK;
4296 }
4297
4298 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
4299 unsigned int resource_idx, DWORD tex_types)
4300 {
4301 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
4302 {
4303 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
4304 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
4305 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
4306 };
4307
4308 unsigned int idx;
4309
4310 if (reg_maps->shader_version.major > 3)
4311 return reg_maps->resource_info[resource_idx].type;
4312
4313 if (!reg_maps->resource_info[resource_idx].type)
4314 return 0;
4315
4316 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
4317 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
4318 return shader_resource_type_from_shader_tex_types[idx];
4319 }
4320
4321 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4322 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4323 {
4324 struct wined3d_shader *object;
4325 HRESULT hr;
4326
4327 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4328 device, desc, parent, parent_ops, shader);
4329
4330 if (!(object = heap_alloc_zero(sizeof(*object))))
4331 return E_OUTOFMEMORY;
4332
4333 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4334 {
4335 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4336 heap_free(object);
4337 return hr;
4338 }
4339
4340 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
4341 {
4342 shader_cleanup(object);
4343 heap_free(object);
4344 return hr;
4345 }
4346
4347 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4348
4349 TRACE("Created compute shader %p.\n", object);
4350 *shader = object;
4351
4352 return WINED3D_OK;
4353 }
4354
4355 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4356 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4357 {
4358 struct wined3d_shader *object;
4359 HRESULT hr;
4360
4361 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4362 device, desc, parent, parent_ops, shader);
4363
4364 if (!(object = heap_alloc_zero(sizeof(*object))))
4365 return E_OUTOFMEMORY;
4366
4367 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4368 {
4369 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4370 heap_free(object);
4371 return hr;
4372 }
4373
4374 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
4375 {
4376 shader_cleanup(object);
4377 heap_free(object);
4378 return hr;
4379 }
4380
4381 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4382
4383 TRACE("Created domain shader %p.\n", object);
4384 *shader = object;
4385
4386 return WINED3D_OK;
4387 }
4388
4389 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4390 const struct wined3d_stream_output_desc *so_desc, void *parent,
4391 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4392 {
4393 struct wined3d_shader *object;
4394 HRESULT hr;
4395
4396 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4397 device, desc, so_desc, parent, parent_ops, shader);
4398
4399 if (!(object = heap_alloc_zero(sizeof(*object))))
4400 return E_OUTOFMEMORY;
4401
4402 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4403 {
4404 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4405 heap_free(object);
4406 return hr;
4407 }
4408
4409 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4410
4411 TRACE("Created geometry shader %p.\n", object);
4412 *shader = object;
4413
4414 return WINED3D_OK;
4415 }
4416
4417 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4418 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4419 {
4420 struct wined3d_shader *object;
4421 HRESULT hr;
4422
4423 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4424 device, desc, parent, parent_ops, shader);
4425
4426 if (!(object = heap_alloc_zero(sizeof(*object))))
4427 return E_OUTOFMEMORY;
4428
4429 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4430 {
4431 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4432 heap_free(object);
4433 return hr;
4434 }
4435
4436 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
4437 {
4438 shader_cleanup(object);
4439 heap_free(object);
4440 return hr;
4441 }
4442
4443 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4444
4445 TRACE("Created hull shader %p.\n", object);
4446 *shader = object;
4447
4448 return WINED3D_OK;
4449 }
4450
4451 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4452 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4453 {
4454 struct wined3d_shader *object;
4455 HRESULT hr;
4456
4457 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4458 device, desc, parent, parent_ops, shader);
4459
4460 if (!(object = heap_alloc_zero(sizeof(*object))))
4461 return E_OUTOFMEMORY;
4462
4463 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4464 {
4465 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4466 heap_free(object);
4467 return hr;
4468 }
4469
4470 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4471
4472 TRACE("Created pixel shader %p.\n", object);
4473 *shader = object;
4474
4475 return WINED3D_OK;
4476 }
4477
4478 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4479 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4480 {
4481 struct wined3d_shader *object;
4482 HRESULT hr;
4483
4484 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4485 device, desc, parent, parent_ops, shader);
4486
4487 if (!(object = heap_alloc_zero(sizeof(*object))))
4488 return E_OUTOFMEMORY;
4489
4490 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4491 {
4492 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4493 heap_free(object);
4494 return hr;
4495 }
4496
4497 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4498
4499 TRACE("Created vertex shader %p.\n", object);
4500 *shader = object;
4501
4502 return WINED3D_OK;
4503 }
Windows API implementation