?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
win32u: Move NtUserDrawCaptionTemp implementation from user32.
[wine.git] / dlls / wined3d / shader.c
blobc2e77021bb34050344b5eb10df15173b70a48a46
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include <stdio.h>
26 #include <string.h>
27
28 #include "wined3d_private.h"
29
30 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
31
32 const struct wined3d_vec4 wined3d_srgb_const[] =
33 {
34 /* pow, mul_high, sub_high, mul_low */
35 {0.41666f, 1.055f, 0.055f, 12.92f},
36 /* cmp */
37 {0.0031308f, 0.0f, 0.0f, 0.0f},
38 };
39
40 static const char * const shader_opcode_names[] =
41 {
42 /* WINED3DSIH_ABS */ "abs",
43 /* WINED3DSIH_ADD */ "add",
44 /* WINED3DSIH_AND */ "and",
45 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
46 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
47 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
48 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
49 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
50 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
51 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
52 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
53 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
54 /* WINED3DSIH_BEM */ "bem",
55 /* WINED3DSIH_BFI */ "bfi",
56 /* WINED3DSIH_BFREV */ "bfrev",
57 /* WINED3DSIH_BREAK */ "break",
58 /* WINED3DSIH_BREAKC */ "breakc",
59 /* WINED3DSIH_BREAKP */ "breakp",
60 /* WINED3DSIH_BUFINFO */ "bufinfo",
61 /* WINED3DSIH_CALL */ "call",
62 /* WINED3DSIH_CALLNZ */ "callnz",
63 /* WINED3DSIH_CASE */ "case",
64 /* WINED3DSIH_CMP */ "cmp",
65 /* WINED3DSIH_CND */ "cnd",
66 /* WINED3DSIH_CONTINUE */ "continue",
67 /* WINED3DSIH_CONTINUEP */ "continuec",
68 /* WINED3DSIH_COUNTBITS */ "countbits",
69 /* WINED3DSIH_CRS */ "crs",
70 /* WINED3DSIH_CUT */ "cut",
71 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
72 /* WINED3DSIH_DCL */ "dcl",
73 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
74 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
75 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
76 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
77 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
78 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
79 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
81 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
82 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
83 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
84 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
85 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
86 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
87 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
88 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
89 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
90 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
91 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
92 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
93 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
94 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
95 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
96 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
97 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
98 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
99 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
100 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
101 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
102 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
103 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
104 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
105 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
106 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
107 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
108 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
109 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
110 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
111 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
112 /* WINED3DSIH_DEF */ "def",
113 /* WINED3DSIH_DEFAULT */ "default",
114 /* WINED3DSIH_DEFB */ "defb",
115 /* WINED3DSIH_DEFI */ "defi",
116 /* WINED3DSIH_DIV */ "div",
117 /* WINED3DSIH_DP2 */ "dp2",
118 /* WINED3DSIH_DP2ADD */ "dp2add",
119 /* WINED3DSIH_DP3 */ "dp3",
120 /* WINED3DSIH_DP4 */ "dp4",
121 /* WINED3DSIH_DST */ "dst",
122 /* WINED3DSIH_DSX */ "dsx",
123 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
124 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
125 /* WINED3DSIH_DSY */ "dsy",
126 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
127 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
128 /* WINED3DSIH_ELSE */ "else",
129 /* WINED3DSIH_EMIT */ "emit",
130 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
131 /* WINED3DSIH_ENDIF */ "endif",
132 /* WINED3DSIH_ENDLOOP */ "endloop",
133 /* WINED3DSIH_ENDREP */ "endrep",
134 /* WINED3DSIH_ENDSWITCH */ "endswitch",
135 /* WINED3DSIH_EQ */ "eq",
136 /* WINED3DSIH_EVAL_CENTROID */ "eval_centroid",
137 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
138 /* WINED3DSIH_EXP */ "exp",
139 /* WINED3DSIH_EXPP */ "expp",
140 /* WINED3DSIH_F16TOF32 */ "f16tof32",
141 /* WINED3DSIH_F32TOF16 */ "f32tof16",
142 /* WINED3DSIH_FCALL */ "fcall",
143 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
144 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
145 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
146 /* WINED3DSIH_FRC */ "frc",
147 /* WINED3DSIH_FTOI */ "ftoi",
148 /* WINED3DSIH_FTOU */ "ftou",
149 /* WINED3DSIH_GATHER4 */ "gather4",
150 /* WINED3DSIH_GATHER4_C */ "gather4_c",
151 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
152 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
153 /* WINED3DSIH_GE */ "ge",
154 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
155 /* WINED3DSIH_HS_DECLS */ "hs_decls",
156 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
157 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
158 /* WINED3DSIH_IADD */ "iadd",
159 /* WINED3DSIH_IBFE */ "ibfe",
160 /* WINED3DSIH_IEQ */ "ieq",
161 /* WINED3DSIH_IF */ "if",
162 /* WINED3DSIH_IFC */ "ifc",
163 /* WINED3DSIH_IGE */ "ige",
164 /* WINED3DSIH_ILT */ "ilt",
165 /* WINED3DSIH_IMAD */ "imad",
166 /* WINED3DSIH_IMAX */ "imax",
167 /* WINED3DSIH_IMIN */ "imin",
168 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
169 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
170 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
171 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
172 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
173 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
174 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
175 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
176 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
177 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
178 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
179 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
180 /* WINED3DSIH_IMUL */ "imul",
181 /* WINED3DSIH_INE */ "ine",
182 /* WINED3DSIH_INEG */ "ineg",
183 /* WINED3DSIH_ISHL */ "ishl",
184 /* WINED3DSIH_ISHR */ "ishr",
185 /* WINED3DSIH_ITOF */ "itof",
186 /* WINED3DSIH_LABEL */ "label",
187 /* WINED3DSIH_LD */ "ld",
188 /* WINED3DSIH_LD2DMS */ "ld2dms",
189 /* WINED3DSIH_LD_RAW */ "ld_raw",
190 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
191 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
192 /* WINED3DSIH_LIT */ "lit",
193 /* WINED3DSIH_LOD */ "lod",
194 /* WINED3DSIH_LOG */ "log",
195 /* WINED3DSIH_LOGP */ "logp",
196 /* WINED3DSIH_LOOP */ "loop",
197 /* WINED3DSIH_LRP */ "lrp",
198 /* WINED3DSIH_LT */ "lt",
199 /* WINED3DSIH_M3x2 */ "m3x2",
200 /* WINED3DSIH_M3x3 */ "m3x3",
201 /* WINED3DSIH_M3x4 */ "m3x4",
202 /* WINED3DSIH_M4x3 */ "m4x3",
203 /* WINED3DSIH_M4x4 */ "m4x4",
204 /* WINED3DSIH_MAD */ "mad",
205 /* WINED3DSIH_MAX */ "max",
206 /* WINED3DSIH_MIN */ "min",
207 /* WINED3DSIH_MOV */ "mov",
208 /* WINED3DSIH_MOVA */ "mova",
209 /* WINED3DSIH_MOVC */ "movc",
210 /* WINED3DSIH_MUL */ "mul",
211 /* WINED3DSIH_NE */ "ne",
212 /* WINED3DSIH_NOP */ "nop",
213 /* WINED3DSIH_NOT */ "not",
214 /* WINED3DSIH_NRM */ "nrm",
215 /* WINED3DSIH_OR */ "or",
216 /* WINED3DSIH_PHASE */ "phase",
217 /* WINED3DSIH_POW */ "pow",
218 /* WINED3DSIH_RCP */ "rcp",
219 /* WINED3DSIH_REP */ "rep",
220 /* WINED3DSIH_RESINFO */ "resinfo",
221 /* WINED3DSIH_RET */ "ret",
222 /* WINED3DSIH_RETP */ "retp",
223 /* WINED3DSIH_ROUND_NE */ "round_ne",
224 /* WINED3DSIH_ROUND_NI */ "round_ni",
225 /* WINED3DSIH_ROUND_PI */ "round_pi",
226 /* WINED3DSIH_ROUND_Z */ "round_z",
227 /* WINED3DSIH_RSQ */ "rsq",
228 /* WINED3DSIH_SAMPLE */ "sample",
229 /* WINED3DSIH_SAMPLE_B */ "sample_b",
230 /* WINED3DSIH_SAMPLE_C */ "sample_c",
231 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
232 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
233 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
234 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
235 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
236 /* WINED3DSIH_SETP */ "setp",
237 /* WINED3DSIH_SGE */ "sge",
238 /* WINED3DSIH_SGN */ "sgn",
239 /* WINED3DSIH_SINCOS */ "sincos",
240 /* WINED3DSIH_SLT */ "slt",
241 /* WINED3DSIH_SQRT */ "sqrt",
242 /* WINED3DSIH_STORE_RAW */ "store_raw",
243 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
244 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
245 /* WINED3DSIH_SUB */ "sub",
246 /* WINED3DSIH_SWAPC */ "swapc",
247 /* WINED3DSIH_SWITCH */ "switch",
248 /* WINED3DSIH_SYNC */ "sync",
249 /* WINED3DSIH_TEX */ "texld",
250 /* WINED3DSIH_TEXBEM */ "texbem",
251 /* WINED3DSIH_TEXBEML */ "texbeml",
252 /* WINED3DSIH_TEXCOORD */ "texcrd",
253 /* WINED3DSIH_TEXDEPTH */ "texdepth",
254 /* WINED3DSIH_TEXDP3 */ "texdp3",
255 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
256 /* WINED3DSIH_TEXKILL */ "texkill",
257 /* WINED3DSIH_TEXLDD */ "texldd",
258 /* WINED3DSIH_TEXLDL */ "texldl",
259 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
260 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
261 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
262 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
263 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
264 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
265 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
266 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
267 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
268 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
269 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
270 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
271 /* WINED3DSIH_UBFE */ "ubfe",
272 /* WINED3DSIH_UDIV */ "udiv",
273 /* WINED3DSIH_UGE */ "uge",
274 /* WINED3DSIH_ULT */ "ult",
275 /* WINED3DSIH_UMAX */ "umax",
276 /* WINED3DSIH_UMIN */ "umin",
277 /* WINED3DSIH_UMUL */ "umul",
278 /* WINED3DSIH_USHR */ "ushr",
279 /* WINED3DSIH_UTOF */ "utof",
280 /* WINED3DSIH_XOR */ "xor",
281 };
282
283 static const char * const semantic_names[] =
284 {
285 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
286 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
287 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
288 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
289 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
290 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
291 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
292 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
293 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
294 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
295 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
296 /* WINED3D_DECL_USAGE_FOG */ "FOG",
297 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
298 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
299 };
300
301 static const struct
302 {
303 enum wined3d_shader_input_sysval_semantic sysval_semantic;
304 const char *sysval_name;
305 }
306 shader_input_sysval_semantic_names[] =
307 {
308 {WINED3D_SIV_POSITION, "position"},
309 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
310 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
311 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
312 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
313 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
314 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
315 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
316 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
317 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
318 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
319 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
320 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
322 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
323 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
324 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
325 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
326 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
327 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
328 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
329 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
330 };
331
332 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
333 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
334
335 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
336 {
337 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
338 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
339
340 return shader_opcode_names[handler_idx];
341 }
342
343 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
344 {
345 if (usage >= ARRAY_SIZE(semantic_names))
346 {
347 FIXME("Unrecognized usage %#x.\n", usage);
348 return "UNRECOGNIZED";
349 }
350
351 return semantic_names[usage];
352 }
353
354 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
355 {
356 unsigned int i;
357
358 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
359 {
360 if (!strcmp(name, semantic_names[i]))
361 return i;
362 }
363
364 return ~0U;
365 }
366
367 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
368 {
369 switch (usage)
370 {
371 case WINED3D_DECL_USAGE_POSITION:
372 return WINED3D_SV_POSITION;
373 default:
374 return 0;
375 }
376 }
377
378 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
379 {
380 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
381 }
382
383 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
384 const struct wined3d_shader_semantic *s)
385 {
386 e->semantic_name = shader_semantic_name_from_usage(s->usage);
387 e->semantic_idx = s->usage_idx;
388 e->stream_idx = 0;
389 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
390 e->component_type = WINED3D_TYPE_FLOAT;
391 e->register_idx = s->reg.reg.idx[0].offset;
392 e->mask = s->reg.write_mask;
393 }
394
395 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
396 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
397 {
398 e->semantic_name = shader_semantic_name_from_usage(usage);
399 e->semantic_idx = usage_idx;
400 e->stream_idx = 0;
401 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
402 e->component_type = WINED3D_TYPE_FLOAT;
403 e->register_idx = reg_idx;
404 e->mask = write_mask;
405 }
406
407 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
408 {
409 switch (format)
410 {
411 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
412 return &sm1_shader_frontend;
413
414 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
415 return &sm4_shader_frontend;
416
417 default:
418 WARN("Invalid byte code format %#x specified.\n", format);
419 return NULL;
420 }
421 }
422
423 void string_buffer_clear(struct wined3d_string_buffer *buffer)
424 {
425 buffer->buffer[0] = '\0';
426 buffer->content_size = 0;
427 }
428
429 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
430 {
431 buffer->buffer_size = 32;
432 if (!(buffer->buffer = heap_alloc(buffer->buffer_size)))
433 {
434 ERR("Failed to allocate shader buffer memory.\n");
435 return FALSE;
436 }
437
438 string_buffer_clear(buffer);
439 return TRUE;
440 }
441
442 void string_buffer_free(struct wined3d_string_buffer *buffer)
443 {
444 heap_free(buffer->buffer);
445 }
446
447 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
448 {
449 char *new_buffer;
450 unsigned int new_buffer_size = buffer->buffer_size * 2;
451
452 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
453 new_buffer_size *= 2;
454 if (!(new_buffer = heap_realloc(buffer->buffer, new_buffer_size)))
455 {
456 ERR("Failed to grow buffer.\n");
457 buffer->buffer[buffer->content_size] = '\0';
458 return FALSE;
459 }
460 buffer->buffer = new_buffer;
461 buffer->buffer_size = new_buffer_size;
462 return TRUE;
463 }
464
465 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
466 {
467 unsigned int rem;
468 int rc;
469
470 rem = buffer->buffer_size - buffer->content_size;
471 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
472 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
473 return rc;
474
475 buffer->content_size += rc;
476 return 0;
477 }
478
479 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
480 {
481 va_list args;
482 int ret;
483
484 for (;;)
485 {
486 va_start(args, format);
487 ret = shader_vaddline(buffer, format, args);
488 va_end(args);
489 if (!ret)
490 return ret;
491 if (!string_buffer_resize(buffer, ret))
492 return -1;
493 }
494 }
495
496 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
497 {
498 struct wined3d_string_buffer *buffer;
499
500 if (list_empty(&list->list))
501 {
502 buffer = heap_alloc(sizeof(*buffer));
503 if (!buffer || !string_buffer_init(buffer))
504 {
505 ERR("Couldn't allocate buffer for temporary string.\n");
506 heap_free(buffer);
507 return NULL;
508 }
509 }
510 else
511 {
512 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
513 list_remove(&buffer->entry);
514 }
515 string_buffer_clear(buffer);
516 return buffer;
517 }
518
519 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
520 {
521 if (!buffer)
522 return 0;
523 string_buffer_clear(buffer);
524 return shader_vaddline(buffer, format, args);
525 }
526
527 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
528 {
529 va_list args;
530 int ret;
531
532 for (;;)
533 {
534 va_start(args, format);
535 ret = string_buffer_vsprintf(buffer, format, args);
536 va_end(args);
537 if (!ret)
538 return;
539 if (!string_buffer_resize(buffer, ret))
540 return;
541 }
542 }
543
544 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
545 {
546 if (!buffer)
547 return;
548 list_add_head(&list->list, &buffer->entry);
549 }
550
551 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
552 {
553 list_init(&list->list);
554 }
555
556 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
557 {
558 struct wined3d_string_buffer *buffer, *buffer_next;
559
560 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
561 {
562 string_buffer_free(buffer);
563 heap_free(buffer);
564 }
565 list_init(&list->list);
566 }
567
568 /* Convert floating point offset relative to a register file to an absolute
569 * offset for float constants. */
570 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
571 {
572 switch (register_type)
573 {
574 case WINED3DSPR_CONST: return register_idx;
575 case WINED3DSPR_CONST2: return 2048 + register_idx;
576 case WINED3DSPR_CONST3: return 4096 + register_idx;
577 case WINED3DSPR_CONST4: return 6144 + register_idx;
578 default:
579 FIXME("Unsupported register type: %u.\n", register_type);
580 return register_idx;
581 }
582 }
583
584 static void shader_delete_constant_list(struct list *clist)
585 {
586 struct wined3d_shader_lconst *constant, *constant_next;
587
588 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
589 heap_free(constant);
590 list_init(clist);
591 }
592
593 static void shader_set_limits(struct wined3d_shader *shader)
594 {
595 static const struct limits_entry
596 {
597 unsigned int min_version;
598 unsigned int max_version;
599 struct wined3d_shader_limits limits;
600 }
601 vs_limits[] =
602 {
603 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
604 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
605 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
606 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
607 * even though they are capable of supporting much more (GL
608 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
609 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
610 * shaders to 256. */
611 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
612 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
613 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
614 {0}
615 },
616 hs_limits[] =
617 {
618 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
619 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
620 },
621 ds_limits[] =
622 {
623 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
624 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
625 },
626 gs_limits[] =
627 {
628 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
629 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
630 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
631 {0}
632 },
633 ps_limits[] =
634 {
635 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
636 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
637 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
638 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
639 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
640 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
641 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
642 {0}
643 },
644 cs_limits[] =
645 {
646 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
647 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
648 };
649 const struct limits_entry *limits_array;
650 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
651 shader->reg_maps.shader_version.minor);
652 int i = 0;
653
654 switch (shader->reg_maps.shader_version.type)
655 {
656 default:
657 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
658 /* Fall-through. */
659 case WINED3D_SHADER_TYPE_VERTEX:
660 limits_array = vs_limits;
661 break;
662 case WINED3D_SHADER_TYPE_HULL:
663 limits_array = hs_limits;
664 break;
665 case WINED3D_SHADER_TYPE_DOMAIN:
666 limits_array = ds_limits;
667 break;
668 case WINED3D_SHADER_TYPE_GEOMETRY:
669 limits_array = gs_limits;
670 break;
671 case WINED3D_SHADER_TYPE_PIXEL:
672 limits_array = ps_limits;
673 break;
674 case WINED3D_SHADER_TYPE_COMPUTE:
675 limits_array = cs_limits;
676 break;
677 }
678
679 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
680 {
681 if (shader_version <= limits_array[i].max_version)
682 {
683 shader->limits = &limits_array[i].limits;
684 break;
685 }
686 ++i;
687 }
688 if (!shader->limits)
689 {
690 FIXME("Unexpected shader version \"%u.%u\".\n",
691 shader->reg_maps.shader_version.major,
692 shader->reg_maps.shader_version.minor);
693 shader->limits = &limits_array[max(0, i - 1)].limits;
694 }
695 }
696
697 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
698 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
699 {
700 switch (reg->type)
701 {
702 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
703 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
704 reg_maps->texcoord |= 1u << reg->idx[0].offset;
705 else
706 reg_maps->address |= 1u << reg->idx[0].offset;
707 break;
708
709 case WINED3DSPR_TEMP:
710 reg_maps->temporary |= 1u << reg->idx[0].offset;
711 break;
712
713 case WINED3DSPR_INPUT:
714 if (reg->idx[0].rel_addr)
715 reg_maps->input_rel_addressing = 1;
716 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
717 {
718 /* If relative addressing is used, we must assume that all
719 * registers are used. Even if it is a construct like v3[aL],
720 * we can't assume that v0, v1 and v2 aren't read because aL
721 * can be negative. */
722 if (reg->idx[0].rel_addr)
723 shader->u.ps.input_reg_used = ~0u;
724 else
725 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
726 }
727 else
728 {
729 reg_maps->input_registers |= 1u << reg->idx[0].offset;
730 }
731 break;
732
733 case WINED3DSPR_RASTOUT:
734 if (reg->idx[0].offset == 1)
735 reg_maps->fog = 1;
736 if (reg->idx[0].offset == 2)
737 reg_maps->point_size = 1;
738 break;
739
740 case WINED3DSPR_MISCTYPE:
741 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
742 {
743 if (!reg->idx[0].offset)
744 reg_maps->vpos = 1;
745 else if (reg->idx[0].offset == 1)
746 reg_maps->usesfacing = 1;
747 }
748 break;
749
750 case WINED3DSPR_CONST:
751 if (reg->idx[0].rel_addr)
752 {
753 if (reg->idx[0].offset < reg_maps->min_rel_offset)
754 reg_maps->min_rel_offset = reg->idx[0].offset;
755 if (reg->idx[0].offset > reg_maps->max_rel_offset)
756 reg_maps->max_rel_offset = reg->idx[0].offset;
757 reg_maps->usesrelconstF = TRUE;
758 }
759 else
760 {
761 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
762 {
763 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
764 return FALSE;
765 }
766 else
767 {
768 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
769 }
770 }
771 break;
772
773 case WINED3DSPR_CONSTINT:
774 if (reg->idx[0].offset >= shader->limits->constant_int)
775 {
776 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
777 return FALSE;
778 }
779 else
780 {
781 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
782 }
783 break;
784
785 case WINED3DSPR_CONSTBOOL:
786 if (reg->idx[0].offset >= shader->limits->constant_bool)
787 {
788 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
789 return FALSE;
790 }
791 else
792 {
793 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
794 }
795 break;
796
797 case WINED3DSPR_COLOROUT:
798 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
799 break;
800
801 case WINED3DSPR_OUTCONTROLPOINT:
802 reg_maps->vocp = 1;
803 break;
804
805 case WINED3DSPR_SAMPLEMASK:
806 reg_maps->sample_mask = 1;
807 break;
808
809 default:
810 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
811 reg->type, reg->idx[0].offset, reg->idx[1].offset);
812 break;
813 }
814 return TRUE;
815 }
816
817 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
818 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
819 {
820 struct wined3d_shader_sampler_map_entry *entries, *entry;
821 struct wined3d_shader_sampler_map *map;
822 unsigned int i;
823
824 map = &reg_maps->sampler_map;
825 entries = map->entries;
826 for (i = 0; i < map->count; ++i)
827 {
828 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
829 return;
830 }
831
832 if (!map->size)
833 {
834 if (!(entries = heap_calloc(4, sizeof(*entries))))
835 {
836 ERR("Failed to allocate sampler map entries.\n");
837 return;
838 }
839 map->size = 4;
840 map->entries = entries;
841 }
842 else if (map->count == map->size)
843 {
844 size_t new_size = map->size * 2;
845
846 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
847 || !(entries = heap_realloc(entries, sizeof(*entries) * new_size)))
848 {
849 ERR("Failed to resize sampler map entries.\n");
850 return;
851 }
852 map->size = new_size;
853 map->entries = entries;
854 }
855
856 entry = &entries[map->count++];
857 entry->resource_idx = resource_idx;
858 entry->sampler_idx = sampler_idx;
859 entry->bind_idx = bind_idx;
860 }
861
862 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
863 {
864 switch (instr)
865 {
866 case WINED3DSIH_M4x4:
867 case WINED3DSIH_M3x4:
868 return param == 1 ? 3 : 0;
869
870 case WINED3DSIH_M4x3:
871 case WINED3DSIH_M3x3:
872 return param == 1 ? 2 : 0;
873
874 case WINED3DSIH_M3x2:
875 return param == 1 ? 1 : 0;
876
877 default:
878 return 0;
879 }
880 }
881
882 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
883 unsigned int register_idx, unsigned int size, unsigned int stride)
884 {
885 struct wined3d_shader_tgsm *tgsm;
886
887 if (register_idx >= MAX_TGSM_REGISTERS)
888 {
889 ERR("Invalid TGSM register index %u.\n", register_idx);
890 return S_OK;
891 }
892 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
893 {
894 FIXME("TGSM declarations are allowed only in compute shaders.\n");
895 return S_OK;
896 }
897
898 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
899 register_idx + 1, sizeof(*reg_maps->tgsm)))
900 return E_OUTOFMEMORY;
901
902 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
903 tgsm = &reg_maps->tgsm[register_idx];
904 tgsm->size = size;
905 tgsm->stride = stride;
906 return S_OK;
907 }
908
909 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
910 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
911 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
912 {
913 struct wined3d_shader_phase *phase;
914
915 if ((phase = *current_phase))
916 {
917 phase->end = previous_instruction_ptr;
918 *current_phase = NULL;
919 }
920
921 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
922 {
923 ERR("Unexpected shader type %s.\n", debug_shader_type(shader->reg_maps.shader_version.type));
924 return E_FAIL;
925 }
926
927 switch (ins->handler_idx)
928 {
929 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
930 if (shader->u.hs.phases.control_point)
931 {
932 FIXME("Multiple control point phases.\n");
933 heap_free(shader->u.hs.phases.control_point);
934 }
935 if (!(shader->u.hs.phases.control_point = heap_alloc_zero(sizeof(*shader->u.hs.phases.control_point))))
936 return E_OUTOFMEMORY;
937 phase = shader->u.hs.phases.control_point;
938 break;
939 case WINED3DSIH_HS_FORK_PHASE:
940 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
941 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
942 sizeof(*shader->u.hs.phases.fork)))
943 return E_OUTOFMEMORY;
944 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
945 break;
946 case WINED3DSIH_HS_JOIN_PHASE:
947 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
948 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
949 sizeof(*shader->u.hs.phases.join)))
950 return E_OUTOFMEMORY;
951 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
952 break;
953 default:
954 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
955 return E_FAIL;
956 }
957
958 phase->start = current_instruction_ptr;
959 *current_phase = phase;
960
961 return WINED3D_OK;
962 }
963
964 static HRESULT shader_calculate_clip_or_cull_distance_mask(
965 const struct wined3d_shader_signature_element *e, unsigned int *mask)
966 {
967 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
968 * the only allowed semantic indices.
969 */
970 if (e->semantic_idx >= WINED3D_MAX_CLIP_DISTANCES / 4)
971 {
972 *mask = 0;
973 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
974 return WINED3DERR_INVALIDCALL;
975 }
976
977 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
978 return WINED3D_OK;
979 }
980
981 static void wined3d_insert_interpolation_mode(DWORD *packed_interpolation_mode,
982 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
983 {
984 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
985 FIXME("Unexpected interpolation mode %#x.\n", mode);
986
987 wined3d_insert_bits(packed_interpolation_mode,
988 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
989 }
990
991 static HRESULT shader_scan_output_signature(struct wined3d_shader *shader)
992 {
993 const struct wined3d_shader_signature *output_signature = &shader->output_signature;
994 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
995 unsigned int i;
996 HRESULT hr;
997
998 for (i = 0; i < output_signature->element_count; ++i)
999 {
1000 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1001 unsigned int mask;
1002
1003 reg_maps->output_registers |= 1u << e->register_idx;
1004 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
1005 {
1006 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1007 return hr;
1008 reg_maps->clip_distance_mask |= mask;
1009 }
1010 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
1011 {
1012 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1013 return hr;
1014 reg_maps->cull_distance_mask |= mask;
1015 }
1016 else if (e->sysval_semantic == WINED3D_SV_VIEWPORT_ARRAY_INDEX)
1017 {
1018 reg_maps->viewport_array = 1;
1019 }
1020 }
1021
1022 return WINED3D_OK;
1023 }
1024
1025 /* Note that this does not count the loop register as an address register. */
1026 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, DWORD constf_size)
1027 {
1028 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
1029 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
1030 struct wined3d_shader_signature *output_signature = &shader->output_signature;
1031 struct wined3d_shader_signature *input_signature = &shader->input_signature;
1032 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1033 const struct wined3d_shader_frontend *fe = shader->frontend;
1034 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
1035 struct wined3d_shader_version shader_version;
1036 struct wined3d_shader_phase *phase = NULL;
1037 const DWORD *ptr, *prev_ins, *current_ins;
1038 void *fe_data = shader->frontend_data;
1039 unsigned int i;
1040 HRESULT hr;
1041
1042 memset(reg_maps, 0, sizeof(*reg_maps));
1043 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1044 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1045 reg_maps->min_rel_offset = ~0U;
1046 list_init(&reg_maps->indexable_temps);
1047
1048 fe->shader_read_header(fe_data, &ptr, &shader_version);
1049 prev_ins = current_ins = ptr;
1050 reg_maps->shader_version = shader_version;
1051
1052 shader_set_limits(shader);
1053
1054 if (!(reg_maps->constf = heap_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1055 sizeof(*reg_maps->constf))))
1056 {
1057 ERR("Failed to allocate constant map memory.\n");
1058 return E_OUTOFMEMORY;
1059 }
1060
1061 while (!fe->shader_is_end(fe_data, &ptr))
1062 {
1063 struct wined3d_shader_instruction ins;
1064
1065 current_ins = ptr;
1066 /* Fetch opcode. */
1067 fe->shader_read_instruction(fe_data, &ptr, &ins);
1068
1069 /* Unhandled opcode, and its parameters. */
1070 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1071 {
1072 WARN("Encountered unrecognised or invalid instruction.\n");
1073 return WINED3DERR_INVALIDCALL;
1074 }
1075
1076 /* Handle declarations. */
1077 if (ins.handler_idx == WINED3DSIH_DCL
1078 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1079 {
1080 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1081 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1082
1083 switch (semantic->reg.reg.type)
1084 {
1085 /* Mark input registers used. */
1086 case WINED3DSPR_INPUT:
1087 if (reg_idx >= MAX_REG_INPUT)
1088 {
1089 ERR("Invalid input register index %u.\n", reg_idx);
1090 break;
1091 }
1092 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1093 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1094 return WINED3DERR_INVALIDCALL;
1095 reg_maps->input_registers |= 1u << reg_idx;
1096 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1097 break;
1098
1099 /* Vertex shader: mark 3.0 output registers used, save token. */
1100 case WINED3DSPR_OUTPUT:
1101 if (reg_idx >= MAX_REG_OUTPUT)
1102 {
1103 ERR("Invalid output register index %u.\n", reg_idx);
1104 break;
1105 }
1106 reg_maps->output_registers |= 1u << reg_idx;
1107 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1108 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1109 reg_maps->fog = 1;
1110 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1111 reg_maps->point_size = 1;
1112 break;
1113
1114 case WINED3DSPR_SAMPLER:
1115 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1116 case WINED3DSPR_RESOURCE:
1117 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1118 {
1119 ERR("Invalid resource index %u.\n", reg_idx);
1120 break;
1121 }
1122 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1123 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1124 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1125 break;
1126
1127 case WINED3DSPR_UAV:
1128 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1129 {
1130 ERR("Invalid UAV resource index %u.\n", reg_idx);
1131 break;
1132 }
1133 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1134 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1135 if (ins.flags)
1136 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1137 break;
1138
1139 default:
1140 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1141 break;
1142 }
1143 }
1144 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1145 {
1146 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1147 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1148 {
1149 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1150 }
1151 else
1152 {
1153 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1154 wined3d_bitmap_set(&reg_maps->cb_map, reg->idx[0].offset);
1155 }
1156 }
1157 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1158 {
1159 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1160 {
1161 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1162 shader->u.ps.force_early_depth_stencil = TRUE;
1163 else
1164 FIXME("Invalid instruction %#x for shader type %#x.\n",
1165 ins.handler_idx, shader_version.type);
1166 }
1167 else
1168 {
1169 WARN("Ignoring global flags %#x.\n", ins.flags);
1170 }
1171 }
1172 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1173 {
1174 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1175 shader->u.gs.instance_count = ins.declaration.count;
1176 else
1177 FIXME("Invalid instruction %#x for shader type %#x.\n",
1178 ins.handler_idx, shader_version.type);
1179 }
1180 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1181 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1182 {
1183 if (phase)
1184 phase->instance_count = ins.declaration.count;
1185 else
1186 FIXME("Instruction %s outside of shader phase.\n",
1187 debug_d3dshaderinstructionhandler(ins.handler_idx));
1188 }
1189 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1190 {
1191 if (reg_maps->icb)
1192 FIXME("Multiple immediate constant buffers.\n");
1193 reg_maps->icb = ins.declaration.icb;
1194 }
1195 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1196 {
1197 if (phase)
1198 {
1199 FIXME("Indexable temporary registers not supported.\n");
1200 }
1201 else
1202 {
1203 struct wined3d_shader_indexable_temp *reg;
1204
1205 if (!(reg = heap_alloc(sizeof(*reg))))
1206 return E_OUTOFMEMORY;
1207
1208 *reg = ins.declaration.indexable_temp;
1209 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1210 }
1211 }
1212 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1213 {
1214 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1215 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1216 else
1217 FIXME("Invalid instruction %#x for shader type %#x.\n",
1218 ins.handler_idx, shader_version.type);
1219 }
1220 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1221 {
1222 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1223 if (reg_idx >= MAX_REG_INPUT)
1224 {
1225 ERR("Invalid register index %u.\n", reg_idx);
1226 break;
1227 }
1228 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1229 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1230 else
1231 FIXME("Invalid instruction %#x for shader type %#x.\n",
1232 ins.handler_idx, shader_version.type);
1233 }
1234 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1235 {
1236 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1237 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1238 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1239 {
1240 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1241 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1242 else
1243 FIXME("Invalid instruction %#x for shader type %#x.\n",
1244 ins.handler_idx, shader_version.type);
1245 }
1246 }
1247 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1248 {
1249 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1250 shader->u.hs.output_vertex_count = ins.declaration.count;
1251 else
1252 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1253 }
1254 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1255 {
1256 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1257 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1258 else
1259 FIXME("Invalid instruction %#x for shader type %#x.\n",
1260 ins.handler_idx, shader_version.type);
1261 }
1262 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1263 {
1264 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1265 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1266 {
1267 ERR("Invalid resource index %u.\n", reg_idx);
1268 break;
1269 }
1270 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1271 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1272 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1273 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1274 }
1275 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1276 {
1277 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1278 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1279 {
1280 ERR("Invalid resource index %u.\n", reg_idx);
1281 break;
1282 }
1283 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1284 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1285 reg_maps->resource_info[reg_idx].flags = 0;
1286 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1287 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1288 }
1289 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1290 {
1291 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1292 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1293 }
1294 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1295 {
1296 if (phase)
1297 phase->temporary_count = ins.declaration.count;
1298 else
1299 reg_maps->temporary_count = ins.declaration.count;
1300 }
1301 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1302 {
1303 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1304 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1305 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1306 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1307 }
1308 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1309 {
1310 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1311 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1312 else
1313 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1314 }
1315 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1316 {
1317 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1318 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1319 else
1320 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1321 }
1322 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1323 {
1324 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1325 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1326 return hr;
1327 }
1328 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1329 {
1330 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1331 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1332 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1333 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1334 return hr;
1335 }
1336 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1337 {
1338 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1339 {
1340 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1341 }
1342 else
1343 {
1344 FIXME("Invalid instruction %#x for shader type %#x.\n",
1345 ins.handler_idx, shader_version.type);
1346 }
1347 }
1348 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1349 {
1350 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1351 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1352 {
1353 ERR("Invalid UAV resource index %u.\n", reg_idx);
1354 break;
1355 }
1356 if (ins.flags)
1357 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1358 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1359 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1360 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1361 }
1362 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1363 {
1364 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1365 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1366 {
1367 ERR("Invalid UAV resource index %u.\n", reg_idx);
1368 break;
1369 }
1370 if (ins.flags)
1371 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1372 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1373 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1374 reg_maps->uav_resource_info[reg_idx].flags = 0;
1375 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1376 }
1377 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1378 {
1379 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1380 shader->u.gs.vertices_out = ins.declaration.count;
1381 else
1382 FIXME("Invalid instruction %#x for shader type %#x.\n",
1383 ins.handler_idx, shader_version.type);
1384 }
1385 else if (ins.handler_idx == WINED3DSIH_DEF)
1386 {
1387 struct wined3d_shader_lconst *lconst;
1388 float *value;
1389
1390 if (!(lconst = heap_alloc(sizeof(*lconst))))
1391 return E_OUTOFMEMORY;
1392
1393 lconst->idx = ins.dst[0].reg.idx[0].offset;
1394 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1395 value = (float *)lconst->value;
1396
1397 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1398 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1399 {
1400 if (value[0] < -1.0f) value[0] = -1.0f;
1401 else if (value[0] > 1.0f) value[0] = 1.0f;
1402 if (value[1] < -1.0f) value[1] = -1.0f;
1403 else if (value[1] > 1.0f) value[1] = 1.0f;
1404 if (value[2] < -1.0f) value[2] = -1.0f;
1405 else if (value[2] > 1.0f) value[2] = 1.0f;
1406 if (value[3] < -1.0f) value[3] = -1.0f;
1407 else if (value[3] > 1.0f) value[3] = 1.0f;
1408 }
1409
1410 list_add_head(&shader->constantsF, &lconst->entry);
1411
1412 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1413 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1414 {
1415 shader->lconst_inf_or_nan = TRUE;
1416 }
1417 }
1418 else if (ins.handler_idx == WINED3DSIH_DEFI)
1419 {
1420 struct wined3d_shader_lconst *lconst;
1421
1422 if (!(lconst = heap_alloc(sizeof(*lconst))))
1423 return E_OUTOFMEMORY;
1424
1425 lconst->idx = ins.dst[0].reg.idx[0].offset;
1426 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1427
1428 list_add_head(&shader->constantsI, &lconst->entry);
1429 reg_maps->local_int_consts |= (1u << lconst->idx);
1430 }
1431 else if (ins.handler_idx == WINED3DSIH_DEFB)
1432 {
1433 struct wined3d_shader_lconst *lconst;
1434
1435 if (!(lconst = heap_alloc(sizeof(*lconst))))
1436 return E_OUTOFMEMORY;
1437
1438 lconst->idx = ins.dst[0].reg.idx[0].offset;
1439 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1440
1441 list_add_head(&shader->constantsB, &lconst->entry);
1442 reg_maps->local_bool_consts |= (1u << lconst->idx);
1443 }
1444 /* Handle shader phases. */
1445 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1446 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1447 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1448 {
1449 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1450 return hr;
1451 }
1452 /* For subroutine prototypes. */
1453 else if (ins.handler_idx == WINED3DSIH_LABEL)
1454 {
1455 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1456 }
1457 /* Set texture, address, temporary registers. */
1458 else
1459 {
1460 BOOL color0_mov = FALSE;
1461 unsigned int i;
1462
1463 /* This will loop over all the registers and try to
1464 * make a bitmask of the ones we're interested in.
1465 *
1466 * Relative addressing tokens are ignored, but that's
1467 * okay, since we'll catch any address registers when
1468 * they are initialized (required by spec). */
1469 for (i = 0; i < ins.dst_count; ++i)
1470 {
1471 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1472 shader_version.type, constf_size))
1473 return WINED3DERR_INVALIDCALL;
1474
1475 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1476 {
1477 UINT idx = ins.dst[i].reg.idx[0].offset;
1478
1479 switch (ins.dst[i].reg.type)
1480 {
1481 case WINED3DSPR_RASTOUT:
1482 if (shader_version.major >= 3)
1483 break;
1484 switch (idx)
1485 {
1486 case 0: /* oPos */
1487 reg_maps->output_registers |= 1u << 10;
1488 shader_signature_from_usage(&output_signature_elements[10],
1489 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1490 break;
1491
1492 case 1: /* oFog */
1493 reg_maps->output_registers |= 1u << 11;
1494 shader_signature_from_usage(&output_signature_elements[11],
1495 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1496 break;
1497
1498 case 2: /* oPts */
1499 reg_maps->output_registers |= 1u << 11;
1500 shader_signature_from_usage(&output_signature_elements[11],
1501 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1502 break;
1503 }
1504 break;
1505
1506 case WINED3DSPR_ATTROUT:
1507 if (shader_version.major >= 3)
1508 break;
1509 if (idx < 2)
1510 {
1511 idx += 8;
1512 if (reg_maps->output_registers & (1u << idx))
1513 {
1514 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1515 }
1516 else
1517 {
1518 reg_maps->output_registers |= 1u << idx;
1519 shader_signature_from_usage(&output_signature_elements[idx],
1520 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1521 }
1522 }
1523 break;
1524
1525 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1526 if (shader_version.major >= 3)
1527 {
1528 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1529 {
1530 WARN("Invalid output register index %u.\n", idx);
1531 break;
1532 }
1533 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1534 break;
1535 }
1536 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1537 {
1538 WARN("Invalid texcoord index %u.\n", idx);
1539 break;
1540 }
1541 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1542 if (reg_maps->output_registers & (1u << idx))
1543 {
1544 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1545 }
1546 else
1547 {
1548 reg_maps->output_registers |= 1u << idx;
1549 shader_signature_from_usage(&output_signature_elements[idx],
1550 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1551 }
1552 break;
1553
1554 default:
1555 break;
1556 }
1557 }
1558
1559 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1560 {
1561 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1562 {
1563 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1564 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1565 * the mov and perform the sRGB write correction from the source register.
1566 *
1567 * However, if the mov is only partial, we can't do this, and if the write
1568 * comes from an instruction other than MOV it is hard to do as well. If
1569 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1570 shader->u.ps.color0_mov = FALSE;
1571 if (ins.handler_idx == WINED3DSIH_MOV
1572 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1573 {
1574 /* Used later when the source register is read. */
1575 color0_mov = TRUE;
1576 }
1577 }
1578 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1579 * end
1580 */
1581 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1582 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1583 {
1584 shader->u.ps.color0_mov = FALSE;
1585 }
1586 }
1587
1588 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1589 if (shader_version.major == 1
1590 && (ins.handler_idx == WINED3DSIH_TEX
1591 || ins.handler_idx == WINED3DSIH_TEXBEM
1592 || ins.handler_idx == WINED3DSIH_TEXBEML
1593 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1594 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1595 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1596 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1597 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1598 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1599 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1600 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1601 {
1602 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1603
1604 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1605 {
1606 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1607 continue;
1608 }
1609
1610 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1611 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1612 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1613 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1614 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1615
1616 /* texbem is only valid with < 1.4 pixel shaders */
1617 if (ins.handler_idx == WINED3DSIH_TEXBEM
1618 || ins.handler_idx == WINED3DSIH_TEXBEML)
1619 {
1620 reg_maps->bumpmat |= 1u << reg_idx;
1621 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1622 {
1623 reg_maps->luminanceparams |= 1u << reg_idx;
1624 }
1625 }
1626 }
1627 else if (ins.handler_idx == WINED3DSIH_BEM)
1628 {
1629 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1630 }
1631 }
1632
1633 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1634 {
1635 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1636 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1637 {
1638 ERR("Invalid UAV index %u.\n", reg_idx);
1639 break;
1640 }
1641 reg_maps->uav_counter_mask |= (1u << reg_idx);
1642 }
1643 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1644 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1645 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1646 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1647 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1648 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1649 {
1650 const struct wined3d_shader_register *reg;
1651
1652 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1653 reg = &ins.src[1].reg;
1654 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1655 reg = &ins.src[2].reg;
1656 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1657 reg = &ins.dst[0].reg;
1658 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1659 reg = &ins.src[0].reg;
1660 else
1661 reg = &ins.dst[1].reg;
1662
1663 if (reg->type == WINED3DSPR_UAV)
1664 {
1665 if (reg->idx[0].offset >= MAX_UNORDERED_ACCESS_VIEWS)
1666 {
1667 ERR("Invalid UAV index %u.\n", reg->idx[0].offset);
1668 break;
1669 }
1670 reg_maps->uav_read_mask |= (1u << reg->idx[0].offset);
1671 }
1672 }
1673 else if (ins.handler_idx == WINED3DSIH_NRM)
1674 {
1675 reg_maps->usesnrm = 1;
1676 }
1677 else if (ins.handler_idx == WINED3DSIH_DSY
1678 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1679 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1680 {
1681 reg_maps->usesdsy = 1;
1682 }
1683 else if (ins.handler_idx == WINED3DSIH_DSX
1684 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1685 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1686 {
1687 reg_maps->usesdsx = 1;
1688 }
1689 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1690 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1691 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1692 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1693 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1694 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1695 else if (ins.handler_idx == WINED3DSIH_LOOP
1696 || ins.handler_idx == WINED3DSIH_REP)
1697 {
1698 ++cur_loop_depth;
1699 if (cur_loop_depth > max_loop_depth)
1700 max_loop_depth = cur_loop_depth;
1701 }
1702 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1703 || ins.handler_idx == WINED3DSIH_ENDREP)
1704 {
1705 --cur_loop_depth;
1706 }
1707 else if (ins.handler_idx == WINED3DSIH_GATHER4
1708 || ins.handler_idx == WINED3DSIH_GATHER4_C
1709 || ins.handler_idx == WINED3DSIH_SAMPLE
1710 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1711 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1712 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1713 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1714 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1715 {
1716 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1717 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1718 }
1719 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1720 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1721 {
1722 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1723 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1724 }
1725 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1726 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1727 {
1728 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1729 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1730 }
1731 else if (ins.handler_idx == WINED3DSIH_LD
1732 || ins.handler_idx == WINED3DSIH_LD2DMS
1733 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1734 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1735 {
1736 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1737 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1738 }
1739 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1740 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1741 {
1742 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1743 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1744 }
1745
1746 if (ins.predicate)
1747 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1748 shader_version.type, constf_size))
1749 return WINED3DERR_INVALIDCALL;
1750
1751 for (i = 0; i < ins.src_count; ++i)
1752 {
1753 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1754 struct wined3d_shader_register reg = ins.src[i].reg;
1755
1756 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1757 shader_version.type, constf_size))
1758 return WINED3DERR_INVALIDCALL;
1759 while (count)
1760 {
1761 ++reg.idx[0].offset;
1762 if (!shader_record_register_usage(shader, reg_maps, &reg,
1763 shader_version.type, constf_size))
1764 return WINED3DERR_INVALIDCALL;
1765 --count;
1766 }
1767
1768 if (color0_mov)
1769 {
1770 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1771 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1772 {
1773 shader->u.ps.color0_mov = TRUE;
1774 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1775 }
1776 }
1777 }
1778 }
1779
1780 prev_ins = current_ins;
1781 }
1782 reg_maps->loop_depth = max_loop_depth;
1783
1784 if (phase)
1785 {
1786 phase->end = prev_ins;
1787 phase = NULL;
1788 }
1789
1790 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1791 * R0 is written to the render target. */
1792 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1793 reg_maps->rt_mask |= (1u << 0);
1794
1795 if (input_signature->elements)
1796 {
1797 for (i = 0; i < input_signature->element_count; ++i)
1798 {
1799 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1800 {
1801 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1802 {
1803 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1804 return WINED3DERR_INVALIDCALL;
1805 }
1806 }
1807 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1808 {
1809 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1810 reg_maps->vpos = 1;
1811 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1812 reg_maps->usesfacing = 1;
1813 }
1814 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1815 }
1816 }
1817 else if (!input_signature->elements && reg_maps->input_registers)
1818 {
1819 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1820 struct wined3d_shader_signature_element *e;
1821 unsigned int i;
1822
1823 if (!(input_signature->elements = heap_calloc(count, sizeof(*input_signature->elements))))
1824 return E_OUTOFMEMORY;
1825 input_signature->element_count = count;
1826
1827 e = input_signature->elements;
1828 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1829 {
1830 if (!(reg_maps->input_registers & (1u << i)))
1831 continue;
1832 input_signature_elements[i].register_idx = i;
1833 *e++ = input_signature_elements[i];
1834 }
1835 }
1836
1837 if (output_signature->elements)
1838 {
1839 if (FAILED(hr = shader_scan_output_signature(shader)))
1840 return hr;
1841 }
1842 else if (reg_maps->output_registers)
1843 {
1844 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1845 struct wined3d_shader_signature_element *e;
1846
1847 if (!(output_signature->elements = heap_calloc(count, sizeof(*output_signature->elements))))
1848 return E_OUTOFMEMORY;
1849 output_signature->element_count = count;
1850
1851 e = output_signature->elements;
1852 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1853 {
1854 if (!(reg_maps->output_registers & (1u << i)))
1855 continue;
1856 *e++ = output_signature_elements[i];
1857 }
1858 }
1859
1860 return WINED3D_OK;
1861 }
1862
1863 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1864 {
1865 struct wined3d_shader_indexable_temp *reg, *reg_next;
1866
1867 heap_free(reg_maps->constf);
1868 heap_free(reg_maps->sampler_map.entries);
1869
1870 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1871 heap_free(reg);
1872 list_init(&reg_maps->indexable_temps);
1873
1874 heap_free(reg_maps->tgsm);
1875 }
1876
1877 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1878 {
1879 DWORD map = 1u << max;
1880 map |= map - 1;
1881 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1882
1883 return wined3d_log2i(map);
1884 }
1885
1886 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1887 {
1888 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1889 {
1890 shader_addline(buffer, "refactoringAllowed");
1891 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1892 if (global_flags)
1893 shader_addline(buffer, " | ");
1894 }
1895
1896 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1897 {
1898 shader_addline(buffer, "forceEarlyDepthStencil");
1899 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1900 if (global_flags)
1901 shader_addline(buffer, " | ");
1902 }
1903
1904 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1905 {
1906 shader_addline(buffer, "enableRawAndStructuredBuffers");
1907 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1908 }
1909
1910 if (global_flags)
1911 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1912 }
1913
1914 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1915 {
1916 if (sync_flags & WINED3DSSF_GLOBAL_UAV)
1917 {
1918 shader_addline(buffer, "_uglobal");
1919 sync_flags &= ~WINED3DSSF_GLOBAL_UAV;
1920 }
1921 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1922 {
1923 shader_addline(buffer, "_g");
1924 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1925 }
1926 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1927 {
1928 shader_addline(buffer, "_t");
1929 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1930 }
1931
1932 if (sync_flags)
1933 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1934 }
1935
1936 static void shader_dump_precise_flags(struct wined3d_string_buffer *buffer, DWORD flags)
1937 {
1938 if (!(flags & WINED3DSI_PRECISE_XYZW))
1939 return;
1940
1941 shader_addline(buffer, " [precise");
1942 if (flags != WINED3DSI_PRECISE_XYZW)
1943 {
1944 shader_addline(buffer, "(%s%s%s%s)",
1945 flags & WINED3DSI_PRECISE_X ? "x" : "",
1946 flags & WINED3DSI_PRECISE_Y ? "y" : "",
1947 flags & WINED3DSI_PRECISE_Z ? "z" : "",
1948 flags & WINED3DSI_PRECISE_W ? "w" : "");
1949 }
1950 shader_addline(buffer, "]");
1951 }
1952
1953 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1954 {
1955 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1956 {
1957 shader_addline(buffer, "_glc");
1958 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1959 }
1960 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1961 {
1962 shader_addline(buffer, "_opc");
1963 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1964 }
1965
1966 if (uav_flags)
1967 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1968 }
1969
1970 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1971 enum wined3d_tessellator_domain domain)
1972 {
1973 switch (domain)
1974 {
1975 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1976 shader_addline(buffer, "line");
1977 break;
1978 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1979 shader_addline(buffer, "triangle");
1980 break;
1981 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1982 shader_addline(buffer, "quad");
1983 break;
1984 default:
1985 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1986 break;
1987 }
1988 }
1989
1990 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1991 enum wined3d_tessellator_output_primitive output_primitive)
1992 {
1993 switch (output_primitive)
1994 {
1995 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1996 shader_addline(buffer, "point");
1997 break;
1998 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1999 shader_addline(buffer, "line");
2000 break;
2001 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
2002 shader_addline(buffer, "triangle_cw");
2003 break;
2004 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
2005 shader_addline(buffer, "triangle_ccw");
2006 break;
2007 default:
2008 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
2009 break;
2010 }
2011 }
2012
2013 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
2014 enum wined3d_tessellator_partitioning partitioning)
2015 {
2016 switch (partitioning)
2017 {
2018 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
2019 shader_addline(buffer, "integer");
2020 break;
2021 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
2022 shader_addline(buffer, "pow2");
2023 break;
2024 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
2025 shader_addline(buffer, "fractional_odd");
2026 break;
2027 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
2028 shader_addline(buffer, "fractional_even");
2029 break;
2030 default:
2031 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
2032 break;
2033 }
2034 }
2035
2036 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
2037 enum wined3d_shader_input_sysval_semantic semantic)
2038 {
2039 unsigned int i;
2040
2041 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
2042 {
2043 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
2044 {
2045 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
2046 return;
2047 }
2048 }
2049
2050 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
2051 }
2052
2053 static void shader_dump_resource_type(struct wined3d_string_buffer *buffer, enum wined3d_shader_resource_type type)
2054 {
2055 static const char *const resource_type_names[] =
2056 {
2057 /* WINED3D_SHADER_RESOURCE_NONE */ "none",
2058 /* WINED3D_SHADER_RESOURCE_BUFFER */ "buffer",
2059 /* WINED3D_SHADER_RESOURCE_TEXTURE_1D */ "texture1d",
2060 /* WINED3D_SHADER_RESOURCE_TEXTURE_2D */ "texture2d",
2061 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DMS */ "texture2dms",
2062 /* WINED3D_SHADER_RESOURCE_TEXTURE_3D */ "texture3d",
2063 /* WINED3D_SHADER_RESOURCE_TEXTURE_CUBE */ "texturecube",
2064 /* WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY */ "texture1darray",
2065 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY */ "texture2darray",
2066 /* WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY */ "texture2dmsarray",
2067 /* WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY */ "texturecubearray",
2068 };
2069
2070 if (type < ARRAY_SIZE(resource_type_names))
2071 shader_addline(buffer, "%s", resource_type_names[type]);
2072 else
2073 shader_addline(buffer, "unknown");
2074 }
2075
2076 static void shader_dump_data_type(struct wined3d_string_buffer *buffer, enum wined3d_data_type type)
2077 {
2078 static const char *const data_type_names[] =
2079 {
2080 /* WINED3D_DATA_FLOAT */ "(float)",
2081 /* WINED3D_DATA_INT */ "(int)",
2082 /* WINED3D_DATA_RESOURCE */ "(resource)",
2083 /* WINED3D_DATA_SAMPLER */ "(sampler)",
2084 /* WINED3D_DATA_UAV */ "(uav)",
2085 /* WINED3D_DATA_UINT */ "(uint)",
2086 /* WINED3D_DATA_UNORM */ "(unorm)",
2087 /* WINED3D_DATA_SNORM */ "(snorm)",
2088 /* WINED3D_DATA_OPAQUE */ "(opaque)",
2089 };
2090
2091 if (type < ARRAY_SIZE(data_type_names))
2092 shader_addline(buffer, "%s", data_type_names[type]);
2093 else
2094 shader_addline(buffer, "(unknown)");
2095 }
2096
2097 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
2098 const struct wined3d_shader_semantic *semantic, unsigned int flags,
2099 const struct wined3d_shader_version *shader_version)
2100 {
2101 shader_addline(buffer, "dcl");
2102
2103 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
2104 {
2105 switch (semantic->resource_type)
2106 {
2107 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2108 shader_addline(buffer, "_2d");
2109 break;
2110
2111 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2112 shader_addline(buffer, "_3d");
2113 break;
2114
2115 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2116 shader_addline(buffer, "_cube");
2117 break;
2118
2119 default:
2120 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2121 break;
2122 }
2123 }
2124 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2125 {
2126 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2127 shader_addline(buffer, "_resource_");
2128 else
2129 shader_addline(buffer, "_uav_");
2130 shader_dump_resource_type(buffer, semantic->resource_type);
2131 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2132 shader_dump_uav_flags(buffer, flags);
2133 shader_dump_data_type(buffer, semantic->resource_data_type);
2134 }
2135 else
2136 {
2137 /* Pixel shaders 3.0 don't have usage semantics. */
2138 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2139 return;
2140 else
2141 shader_addline(buffer, "_");
2142
2143 switch (semantic->usage)
2144 {
2145 case WINED3D_DECL_USAGE_POSITION:
2146 shader_addline(buffer, "position%u", semantic->usage_idx);
2147 break;
2148
2149 case WINED3D_DECL_USAGE_BLEND_INDICES:
2150 shader_addline(buffer, "blend");
2151 break;
2152
2153 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2154 shader_addline(buffer, "weight");
2155 break;
2156
2157 case WINED3D_DECL_USAGE_NORMAL:
2158 shader_addline(buffer, "normal%u", semantic->usage_idx);
2159 break;
2160
2161 case WINED3D_DECL_USAGE_PSIZE:
2162 shader_addline(buffer, "psize");
2163 break;
2164
2165 case WINED3D_DECL_USAGE_COLOR:
2166 if (!semantic->usage_idx)
2167 shader_addline(buffer, "color");
2168 else
2169 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2170 break;
2171
2172 case WINED3D_DECL_USAGE_TEXCOORD:
2173 shader_addline(buffer, "texture%u", semantic->usage_idx);
2174 break;
2175
2176 case WINED3D_DECL_USAGE_TANGENT:
2177 shader_addline(buffer, "tangent");
2178 break;
2179
2180 case WINED3D_DECL_USAGE_BINORMAL:
2181 shader_addline(buffer, "binormal");
2182 break;
2183
2184 case WINED3D_DECL_USAGE_TESS_FACTOR:
2185 shader_addline(buffer, "tessfactor");
2186 break;
2187
2188 case WINED3D_DECL_USAGE_POSITIONT:
2189 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2190 break;
2191
2192 case WINED3D_DECL_USAGE_FOG:
2193 shader_addline(buffer, "fog");
2194 break;
2195
2196 case WINED3D_DECL_USAGE_DEPTH:
2197 shader_addline(buffer, "depth");
2198 break;
2199
2200 case WINED3D_DECL_USAGE_SAMPLE:
2201 shader_addline(buffer, "sample");
2202 break;
2203
2204 default:
2205 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2206 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2207 }
2208 }
2209 }
2210
2211 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2212 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2213 {
2214 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2215 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2216 UINT offset = reg->idx[0].offset;
2217
2218 switch (reg->type)
2219 {
2220 case WINED3DSPR_TEMP:
2221 shader_addline(buffer, "r");
2222 break;
2223
2224 case WINED3DSPR_INPUT:
2225 shader_addline(buffer, "v");
2226 break;
2227
2228 case WINED3DSPR_CONST:
2229 case WINED3DSPR_CONST2:
2230 case WINED3DSPR_CONST3:
2231 case WINED3DSPR_CONST4:
2232 shader_addline(buffer, "c");
2233 offset = shader_get_float_offset(reg->type, offset);
2234 break;
2235
2236 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2237 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2238 break;
2239
2240 case WINED3DSPR_RASTOUT:
2241 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2242 break;
2243
2244 case WINED3DSPR_COLOROUT:
2245 shader_addline(buffer, "oC");
2246 break;
2247
2248 case WINED3DSPR_DEPTHOUT:
2249 shader_addline(buffer, "oDepth");
2250 break;
2251
2252 case WINED3DSPR_DEPTHOUTGE:
2253 shader_addline(buffer, "oDepthGE");
2254 break;
2255
2256 case WINED3DSPR_DEPTHOUTLE:
2257 shader_addline(buffer, "oDepthLE");
2258 break;
2259
2260 case WINED3DSPR_ATTROUT:
2261 shader_addline(buffer, "oD");
2262 break;
2263
2264 case WINED3DSPR_TEXCRDOUT:
2265 /* Vertex shaders >= 3.0 use general purpose output registers
2266 * (WINED3DSPR_OUTPUT), which can include an address token. */
2267 if (shader_version->major >= 3)
2268 shader_addline(buffer, "o");
2269 else
2270 shader_addline(buffer, "oT");
2271 break;
2272
2273 case WINED3DSPR_CONSTINT:
2274 shader_addline(buffer, "i");
2275 break;
2276
2277 case WINED3DSPR_CONSTBOOL:
2278 shader_addline(buffer, "b");
2279 break;
2280
2281 case WINED3DSPR_LABEL:
2282 shader_addline(buffer, "l");
2283 break;
2284
2285 case WINED3DSPR_LOOP:
2286 shader_addline(buffer, "aL");
2287 break;
2288
2289 case WINED3DSPR_SAMPLER:
2290 shader_addline(buffer, "s");
2291 break;
2292
2293 case WINED3DSPR_MISCTYPE:
2294 if (offset > 1)
2295 {
2296 FIXME("Unhandled misctype register %u.\n", offset);
2297 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2298 }
2299 else
2300 {
2301 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2302 }
2303 break;
2304
2305 case WINED3DSPR_PREDICATE:
2306 shader_addline(buffer, "p");
2307 break;
2308
2309 case WINED3DSPR_IMMCONST:
2310 shader_addline(buffer, "l");
2311 break;
2312
2313 case WINED3DSPR_CONSTBUFFER:
2314 shader_addline(buffer, "cb");
2315 break;
2316
2317 case WINED3DSPR_IMMCONSTBUFFER:
2318 shader_addline(buffer, "icb");
2319 break;
2320
2321 case WINED3DSPR_PRIMID:
2322 shader_addline(buffer, "primID");
2323 break;
2324
2325 case WINED3DSPR_NULL:
2326 shader_addline(buffer, "null");
2327 break;
2328
2329 case WINED3DSPR_RASTERIZER:
2330 shader_addline(buffer, "rasterizer");
2331 break;
2332
2333 case WINED3DSPR_RESOURCE:
2334 shader_addline(buffer, "t");
2335 break;
2336
2337 case WINED3DSPR_UAV:
2338 shader_addline(buffer, "u");
2339 break;
2340
2341 case WINED3DSPR_OUTPOINTID:
2342 shader_addline(buffer, "vOutputControlPointID");
2343 break;
2344
2345 case WINED3DSPR_FORKINSTID:
2346 shader_addline(buffer, "vForkInstanceId");
2347 break;
2348
2349 case WINED3DSPR_JOININSTID:
2350 shader_addline(buffer, "vJoinInstanceId");
2351 break;
2352
2353 case WINED3DSPR_INCONTROLPOINT:
2354 shader_addline(buffer, "vicp");
2355 break;
2356
2357 case WINED3DSPR_OUTCONTROLPOINT:
2358 shader_addline(buffer, "vocp");
2359 break;
2360
2361 case WINED3DSPR_PATCHCONST:
2362 shader_addline(buffer, "vpc");
2363 break;
2364
2365 case WINED3DSPR_TESSCOORD:
2366 shader_addline(buffer, "vDomainLocation");
2367 break;
2368
2369 case WINED3DSPR_GROUPSHAREDMEM:
2370 shader_addline(buffer, "g");
2371 break;
2372
2373 case WINED3DSPR_THREADID:
2374 shader_addline(buffer, "vThreadID");
2375 break;
2376
2377 case WINED3DSPR_THREADGROUPID:
2378 shader_addline(buffer, "vThreadGroupID");
2379 break;
2380
2381 case WINED3DSPR_LOCALTHREADID:
2382 shader_addline(buffer, "vThreadIDInGroup");
2383 break;
2384
2385 case WINED3DSPR_LOCALTHREADINDEX:
2386 shader_addline(buffer, "vThreadIDInGroupFlattened");
2387 break;
2388
2389 case WINED3DSPR_IDXTEMP:
2390 shader_addline(buffer, "x");
2391 break;
2392
2393 case WINED3DSPR_STREAM:
2394 shader_addline(buffer, "m");
2395 break;
2396
2397 case WINED3DSPR_FUNCTIONBODY:
2398 shader_addline(buffer, "fb");
2399 break;
2400
2401 case WINED3DSPR_FUNCTIONPOINTER:
2402 shader_addline(buffer, "fp");
2403 break;
2404
2405 case WINED3DSPR_COVERAGE:
2406 shader_addline(buffer, "vCoverage");
2407 break;
2408
2409 case WINED3DSPR_SAMPLEMASK:
2410 shader_addline(buffer, "oMask");
2411 break;
2412
2413 case WINED3DSPR_GSINSTID:
2414 shader_addline(buffer, "vGSInstanceID");
2415 break;
2416
2417 default:
2418 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2419 break;
2420 }
2421
2422 if (reg->type == WINED3DSPR_IMMCONST)
2423 {
2424 shader_addline(buffer, "(");
2425 switch (reg->immconst_type)
2426 {
2427 case WINED3D_IMMCONST_SCALAR:
2428 switch (reg->data_type)
2429 {
2430 case WINED3D_DATA_FLOAT:
2431 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2432 break;
2433 case WINED3D_DATA_INT:
2434 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2435 break;
2436 case WINED3D_DATA_RESOURCE:
2437 case WINED3D_DATA_SAMPLER:
2438 case WINED3D_DATA_UINT:
2439 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2440 break;
2441 default:
2442 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2443 break;
2444 }
2445 break;
2446
2447 case WINED3D_IMMCONST_VEC4:
2448 switch (reg->data_type)
2449 {
2450 case WINED3D_DATA_FLOAT:
2451 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2452 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2453 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2454 break;
2455 case WINED3D_DATA_INT:
2456 shader_addline(buffer, "%d, %d, %d, %d",
2457 reg->u.immconst_data[0], reg->u.immconst_data[1],
2458 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2459 break;
2460 case WINED3D_DATA_RESOURCE:
2461 case WINED3D_DATA_SAMPLER:
2462 case WINED3D_DATA_UINT:
2463 shader_addline(buffer, "%u, %u, %u, %u",
2464 reg->u.immconst_data[0], reg->u.immconst_data[1],
2465 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2466 break;
2467 default:
2468 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2469 break;
2470 }
2471 break;
2472
2473 default:
2474 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2475 break;
2476 }
2477 shader_addline(buffer, ")");
2478 }
2479 else if (reg->type != WINED3DSPR_RASTOUT
2480 && reg->type != WINED3DSPR_MISCTYPE
2481 && reg->type != WINED3DSPR_NULL)
2482 {
2483 if (offset != ~0u)
2484 {
2485 shader_addline(buffer, "[");
2486 if (reg->idx[0].rel_addr)
2487 {
2488 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2489 shader_addline(buffer, " + ");
2490 }
2491 shader_addline(buffer, "%u]", offset);
2492
2493 if (reg->idx[1].offset != ~0u)
2494 {
2495 shader_addline(buffer, "[");
2496 if (reg->idx[1].rel_addr)
2497 {
2498 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2499 shader_addline(buffer, " + ");
2500 }
2501 shader_addline(buffer, "%u]", reg->idx[1].offset);
2502 }
2503 }
2504
2505 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2506 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2507 }
2508 }
2509
2510 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2511 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2512 {
2513 DWORD write_mask = param->write_mask;
2514
2515 shader_dump_register(buffer, &param->reg, shader_version);
2516
2517 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2518 {
2519 static const char write_mask_chars[] = "xyzw";
2520
2521 shader_addline(buffer, ".");
2522 if (write_mask & WINED3DSP_WRITEMASK_0)
2523 shader_addline(buffer, "%c", write_mask_chars[0]);
2524 if (write_mask & WINED3DSP_WRITEMASK_1)
2525 shader_addline(buffer, "%c", write_mask_chars[1]);
2526 if (write_mask & WINED3DSP_WRITEMASK_2)
2527 shader_addline(buffer, "%c", write_mask_chars[2]);
2528 if (write_mask & WINED3DSP_WRITEMASK_3)
2529 shader_addline(buffer, "%c", write_mask_chars[3]);
2530 }
2531 }
2532
2533 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2534 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2535 {
2536 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2537 DWORD swizzle = param->swizzle;
2538
2539 if (src_modifier == WINED3DSPSM_NEG
2540 || src_modifier == WINED3DSPSM_BIASNEG
2541 || src_modifier == WINED3DSPSM_SIGNNEG
2542 || src_modifier == WINED3DSPSM_X2NEG
2543 || src_modifier == WINED3DSPSM_ABSNEG)
2544 shader_addline(buffer, "-");
2545 else if (src_modifier == WINED3DSPSM_COMP)
2546 shader_addline(buffer, "1-");
2547 else if (src_modifier == WINED3DSPSM_NOT)
2548 shader_addline(buffer, "!");
2549
2550 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2551 shader_addline(buffer, "abs(");
2552
2553 shader_dump_register(buffer, &param->reg, shader_version);
2554
2555 switch (src_modifier)
2556 {
2557 case WINED3DSPSM_NONE: break;
2558 case WINED3DSPSM_NEG: break;
2559 case WINED3DSPSM_NOT: break;
2560 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2561 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2562 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2563 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2564 case WINED3DSPSM_COMP: break;
2565 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2566 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2567 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2568 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2569 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2570 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2571 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2572 }
2573
2574 if (swizzle != WINED3DSP_NOSWIZZLE)
2575 {
2576 static const char swizzle_chars[] = "xyzw";
2577 DWORD swizzle_x = swizzle & 0x03;
2578 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2579 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2580 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2581
2582 if (swizzle_x == swizzle_y
2583 && swizzle_x == swizzle_z
2584 && swizzle_x == swizzle_w)
2585 {
2586 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2587 }
2588 else
2589 {
2590 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2591 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2592 }
2593 }
2594 }
2595
2596 /* Shared code in order to generate the bulk of the shader string. */
2597 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2598 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2599 const DWORD *start, const DWORD *end)
2600 {
2601 struct wined3d_device *device = shader->device;
2602 const struct wined3d_shader_frontend *fe = shader->frontend;
2603 void *fe_data = shader->frontend_data;
2604 struct wined3d_shader_version shader_version;
2605 struct wined3d_shader_parser_state state;
2606 struct wined3d_shader_instruction ins;
2607 struct wined3d_shader_tex_mx tex_mx;
2608 struct wined3d_shader_context ctx;
2609 const DWORD *ptr;
2610
2611 /* Initialize current parsing state. */
2612 tex_mx.current_row = 0;
2613 state.current_loop_depth = 0;
2614 state.current_loop_reg = 0;
2615 state.in_subroutine = FALSE;
2616
2617 ctx.shader = shader;
2618 ctx.reg_maps = reg_maps;
2619 ctx.buffer = buffer;
2620 ctx.tex_mx = &tex_mx;
2621 ctx.state = &state;
2622 ctx.backend_data = backend_ctx;
2623 ins.ctx = &ctx;
2624
2625 fe->shader_read_header(fe_data, &ptr, &shader_version);
2626 if (start)
2627 ptr = start;
2628
2629 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2630 {
2631 /* Read opcode. */
2632 fe->shader_read_instruction(fe_data, &ptr, &ins);
2633
2634 /* Unknown opcode and its parameters. */
2635 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2636 {
2637 WARN("Encountered unrecognised or invalid instruction.\n");
2638 return WINED3DERR_INVALIDCALL;
2639 }
2640
2641 if (ins.predicate)
2642 FIXME("Predicates not implemented.\n");
2643
2644 /* Call appropriate function for output target */
2645 device->shader_backend->shader_handle_instruction(&ins);
2646 }
2647
2648 return WINED3D_OK;
2649 }
2650
2651 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2652 const struct wined3d_shader_dst_param *dst)
2653 {
2654 DWORD mmask = dst->modifiers;
2655
2656 switch (dst->shift)
2657 {
2658 case 0: break;
2659 case 13: shader_addline(buffer, "_d8"); break;
2660 case 14: shader_addline(buffer, "_d4"); break;
2661 case 15: shader_addline(buffer, "_d2"); break;
2662 case 1: shader_addline(buffer, "_x2"); break;
2663 case 2: shader_addline(buffer, "_x4"); break;
2664 case 3: shader_addline(buffer, "_x8"); break;
2665 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2666 }
2667
2668 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2669 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2670 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2671
2672 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2673 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2674 }
2675
2676 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2677 const struct wined3d_shader_primitive_type *primitive_type)
2678 {
2679 switch (primitive_type->type)
2680 {
2681 case WINED3D_PT_UNDEFINED:
2682 shader_addline(buffer, "undefined");
2683 break;
2684 case WINED3D_PT_POINTLIST:
2685 shader_addline(buffer, "pointlist");
2686 break;
2687 case WINED3D_PT_LINELIST:
2688 shader_addline(buffer, "linelist");
2689 break;
2690 case WINED3D_PT_LINESTRIP:
2691 shader_addline(buffer, "linestrip");
2692 break;
2693 case WINED3D_PT_TRIANGLELIST:
2694 shader_addline(buffer, "trianglelist");
2695 break;
2696 case WINED3D_PT_TRIANGLESTRIP:
2697 shader_addline(buffer, "trianglestrip");
2698 break;
2699 case WINED3D_PT_TRIANGLEFAN:
2700 shader_addline(buffer, "trianglefan");
2701 break;
2702 case WINED3D_PT_LINELIST_ADJ:
2703 shader_addline(buffer, "linelist_adj");
2704 break;
2705 case WINED3D_PT_LINESTRIP_ADJ:
2706 shader_addline(buffer, "linestrip_adj");
2707 break;
2708 case WINED3D_PT_TRIANGLELIST_ADJ:
2709 shader_addline(buffer, "trianglelist_adj");
2710 break;
2711 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2712 shader_addline(buffer, "trianglestrip_adj");
2713 break;
2714 case WINED3D_PT_PATCH:
2715 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2716 break;
2717 default:
2718 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2719 break;
2720 }
2721 }
2722
2723 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2724 enum wined3d_shader_interpolation_mode interpolation_mode)
2725 {
2726 switch (interpolation_mode)
2727 {
2728 case WINED3DSIM_CONSTANT:
2729 shader_addline(buffer, "constant");
2730 break;
2731 case WINED3DSIM_LINEAR:
2732 shader_addline(buffer, "linear");
2733 break;
2734 case WINED3DSIM_LINEAR_CENTROID:
2735 shader_addline(buffer, "linear centroid");
2736 break;
2737 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2738 shader_addline(buffer, "linear noperspective");
2739 break;
2740 case WINED3DSIM_LINEAR_SAMPLE:
2741 shader_addline(buffer, "linear sample");
2742 break;
2743 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2744 shader_addline(buffer, "linear noperspective centroid");
2745 break;
2746 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2747 shader_addline(buffer, "linear noperspective sample");
2748 break;
2749 default:
2750 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2751 break;
2752 }
2753 }
2754
2755 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2756 {
2757 struct wined3d_shader_version shader_version;
2758 struct wined3d_string_buffer buffer;
2759 const char *type_prefix;
2760 const char *p, *q;
2761 const DWORD *ptr;
2762 DWORD i;
2763
2764 if (!string_buffer_init(&buffer))
2765 {
2766 ERR("Failed to initialize string buffer.\n");
2767 return;
2768 }
2769
2770 fe->shader_read_header(fe_data, &ptr, &shader_version);
2771
2772 TRACE("Parsing %p.\n", ptr);
2773
2774 switch (shader_version.type)
2775 {
2776 case WINED3D_SHADER_TYPE_VERTEX:
2777 type_prefix = "vs";
2778 break;
2779
2780 case WINED3D_SHADER_TYPE_HULL:
2781 type_prefix = "hs";
2782 break;
2783
2784 case WINED3D_SHADER_TYPE_DOMAIN:
2785 type_prefix = "ds";
2786 break;
2787
2788 case WINED3D_SHADER_TYPE_GEOMETRY:
2789 type_prefix = "gs";
2790 break;
2791
2792 case WINED3D_SHADER_TYPE_PIXEL:
2793 type_prefix = "ps";
2794 break;
2795
2796 case WINED3D_SHADER_TYPE_COMPUTE:
2797 type_prefix = "cs";
2798 break;
2799
2800 default:
2801 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2802 type_prefix = "unknown";
2803 break;
2804 }
2805
2806 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2807
2808 while (!fe->shader_is_end(fe_data, &ptr))
2809 {
2810 struct wined3d_shader_instruction ins;
2811
2812 fe->shader_read_instruction(fe_data, &ptr, &ins);
2813 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2814 {
2815 WARN("Skipping unrecognized instruction.\n");
2816 shader_addline(&buffer, "<unrecognized instruction>\n");
2817 continue;
2818 }
2819
2820 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2821 {
2822 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2823 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2824 shader_addline(&buffer, " ");
2825 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2826 }
2827 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2828 {
2829 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2830 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2831 shader_addline(&buffer, ", %s",
2832 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2833 }
2834 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2835 {
2836 shader_addline(&buffer, "%s fb%u",
2837 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2838 }
2839 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2840 {
2841 shader_addline(&buffer, "%s ft%u = {...}",
2842 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2843 }
2844 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2845 {
2846 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2847 shader_dump_global_flags(&buffer, ins.flags);
2848 }
2849 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2850 {
2851 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2852 ins.declaration.max_tessellation_factor);
2853 }
2854 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2855 {
2856 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2857 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2858 {
2859 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2860 ins.declaration.icb->data[4 * i + 0],
2861 ins.declaration.icb->data[4 * i + 1],
2862 ins.declaration.icb->data[4 * i + 2],
2863 ins.declaration.icb->data[4 * i + 3]);
2864 }
2865 shader_addline(&buffer, "}");
2866 }
2867 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2868 {
2869 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2870 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2871 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2872 }
2873 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2874 {
2875 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2876 ins.declaration.indexable_temp.register_idx,
2877 ins.declaration.indexable_temp.register_size,
2878 ins.declaration.indexable_temp.component_count);
2879 }
2880 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2881 {
2882 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2883 shader_dump_interpolation_mode(&buffer, ins.flags);
2884 shader_addline(&buffer, " ");
2885 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2886 }
2887 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2888 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2889 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2890 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2891 {
2892 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2893 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2894 shader_addline(&buffer, ", ");
2895 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2896 }
2897 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2898 {
2899 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2900 shader_dump_interpolation_mode(&buffer, ins.flags);
2901 shader_addline(&buffer, " ");
2902 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2903 shader_addline(&buffer, ", ");
2904 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2905 }
2906 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2907 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2908 {
2909 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2910 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2911 }
2912 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2913 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2914 {
2915 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2916 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2917 }
2918 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2919 {
2920 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2921 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2922 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2923 }
2924 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2925 {
2926 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2927 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2928 }
2929 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2930 {
2931 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2932 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2933 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2934 }
2935 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2936 {
2937 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2938 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2939 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2940 shader_addline(&buffer, ", comparisonMode");
2941 }
2942 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2943 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2944 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2945 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2946 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2947 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2948 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2949 {
2950 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2951 }
2952 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2953 {
2954 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2955 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2956 }
2957 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2958 {
2959 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2960 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2961 }
2962 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2963 {
2964 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2965 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2966 }
2967 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2968 {
2969 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2970 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2971 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2972 }
2973 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2974 {
2975 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2976 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2977 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2978 ins.declaration.tgsm_structured.structure_count);
2979 }
2980 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2981 {
2982 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2983 ins.declaration.thread_group_size.x,
2984 ins.declaration.thread_group_size.y,
2985 ins.declaration.thread_group_size.z);
2986 }
2987 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2988 {
2989 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2990 shader_dump_uav_flags(&buffer, ins.flags);
2991 shader_addline(&buffer, " ");
2992 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2993 }
2994 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2995 {
2996 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2997 shader_dump_uav_flags(&buffer, ins.flags);
2998 shader_addline(&buffer, " ");
2999 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
3000 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
3001 }
3002 else if (ins.handler_idx == WINED3DSIH_DEF)
3003 {
3004 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
3005 ins.dst[0].reg.idx[0].offset),
3006 *(const float *)&ins.src[0].reg.u.immconst_data[0],
3007 *(const float *)&ins.src[0].reg.u.immconst_data[1],
3008 *(const float *)&ins.src[0].reg.u.immconst_data[2],
3009 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
3010 }
3011 else if (ins.handler_idx == WINED3DSIH_DEFI)
3012 {
3013 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
3014 ins.src[0].reg.u.immconst_data[0],
3015 ins.src[0].reg.u.immconst_data[1],
3016 ins.src[0].reg.u.immconst_data[2],
3017 ins.src[0].reg.u.immconst_data[3]);
3018 }
3019 else if (ins.handler_idx == WINED3DSIH_DEFB)
3020 {
3021 shader_addline(&buffer, "defb b%u = %s",
3022 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
3023 }
3024 else
3025 {
3026 if (ins.predicate)
3027 {
3028 shader_addline(&buffer, "(");
3029 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
3030 shader_addline(&buffer, ") ");
3031 }
3032
3033 /* PixWin marks instructions with the coissue flag with a '+' */
3034 if (ins.coissue)
3035 shader_addline(&buffer, "+");
3036
3037 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3038
3039 if (ins.handler_idx == WINED3DSIH_BREAKP
3040 || ins.handler_idx == WINED3DSIH_CONTINUEP
3041 || ins.handler_idx == WINED3DSIH_IF
3042 || ins.handler_idx == WINED3DSIH_RETP
3043 || ins.handler_idx == WINED3DSIH_TEXKILL)
3044 {
3045 switch (ins.flags)
3046 {
3047 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3048 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3049 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3050 }
3051 }
3052 else if (ins.handler_idx == WINED3DSIH_IFC
3053 || ins.handler_idx == WINED3DSIH_BREAKC)
3054 {
3055 switch (ins.flags)
3056 {
3057 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3058 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3059 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3060 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3061 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3062 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3063 default: shader_addline(&buffer, "_(%u)", ins.flags);
3064 }
3065 }
3066 else if (ins.handler_idx == WINED3DSIH_TEX
3067 && shader_version.major >= 2
3068 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3069 {
3070 shader_addline(&buffer, "p");
3071 }
3072 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3073 {
3074 switch (ins.flags)
3075 {
3076 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3077 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3078 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3079 }
3080 }
3081 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3082 {
3083 switch (ins.flags)
3084 {
3085 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3086 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3087 }
3088 }
3089 else if (ins.handler_idx == WINED3DSIH_SYNC)
3090 {
3091 shader_dump_sync_flags(&buffer, ins.flags);
3092 }
3093 else
3094 {
3095 shader_dump_precise_flags(&buffer, ins.flags);
3096 }
3097
3098 if (wined3d_shader_instruction_has_texel_offset(&ins))
3099 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3100
3101 if (ins.resource_type != WINED3D_SHADER_RESOURCE_NONE)
3102 {
3103 shader_addline(&buffer, "(");
3104 shader_dump_resource_type(&buffer, ins.resource_type);
3105 shader_addline(&buffer, ")");
3106 }
3107
3108 if (ins.resource_data_type != WINED3D_DATA_FLOAT)
3109 shader_dump_data_type(&buffer, ins.resource_data_type);
3110
3111 for (i = 0; i < ins.dst_count; ++i)
3112 {
3113 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
3114 shader_addline(&buffer, !i ? " " : ", ");
3115 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
3116 }
3117
3118 /* Other source tokens */
3119 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
3120 {
3121 shader_addline(&buffer, !i ? " " : ", ");
3122 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
3123 }
3124 }
3125 shader_addline(&buffer, "\n");
3126 }
3127
3128 for (p = buffer.buffer; *p; p = q)
3129 {
3130 if (!(q = strstr(p, "\n")))
3131 q = p + strlen(p);
3132 else
3133 ++q;
3134 TRACE(" %.*s", (int)(q - p), p);
3135 }
3136
3137 string_buffer_free(&buffer);
3138 }
3139
3140 static void shader_cleanup(struct wined3d_shader *shader)
3141 {
3142 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3143 {
3144 heap_free(shader->u.hs.phases.control_point);
3145 heap_free(shader->u.hs.phases.fork);
3146 heap_free(shader->u.hs.phases.join);
3147 }
3148
3149 heap_free(shader->patch_constant_signature.elements);
3150 heap_free(shader->output_signature.elements);
3151 heap_free(shader->input_signature.elements);
3152 shader->device->shader_backend->shader_destroy(shader);
3153 shader_cleanup_reg_maps(&shader->reg_maps);
3154 heap_free(shader->byte_code);
3155 shader_delete_constant_list(&shader->constantsF);
3156 shader_delete_constant_list(&shader->constantsB);
3157 shader_delete_constant_list(&shader->constantsI);
3158 list_remove(&shader->shader_list_entry);
3159
3160 if (shader->frontend && shader->frontend_data)
3161 shader->frontend->shader_free(shader->frontend_data);
3162 }
3163
3164 struct shader_none_priv
3165 {
3166 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3167 const struct wined3d_fragment_pipe_ops *fragment_pipe;
3168 BOOL ffp_proj_control;
3169 };
3170
3171 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3172 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3173 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3174 const struct wined3d_state *state) {}
3175 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3176 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3177 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3178 const struct wined3d_state *state) {}
3179 static void shader_none_destroy(struct wined3d_shader *shader) {}
3180 static void shader_none_free_context_data(struct wined3d_context *context) {}
3181 static void shader_none_init_context_state(struct wined3d_context *context) {}
3182
3183 /* Context activation is done by the caller. */
3184 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3185 const struct wined3d_state *state)
3186 {
3187 struct shader_none_priv *priv = shader_priv;
3188
3189 priv->vertex_pipe->vp_enable(context, !use_vs(state));
3190 priv->fragment_pipe->fp_enable(context, !use_ps(state));
3191 }
3192
3193 /* Context activation is done by the caller. */
3194 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3195 {
3196 struct shader_none_priv *priv = shader_priv;
3197
3198 priv->vertex_pipe->vp_enable(context, FALSE);
3199 priv->fragment_pipe->fp_enable(context, FALSE);
3200
3201 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3202 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3203 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3204 | (1u << WINED3D_SHADER_TYPE_HULL)
3205 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3206 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3207 }
3208
3209 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3210 const struct wined3d_fragment_pipe_ops *fragment_pipe)
3211 {
3212 struct fragment_caps fragment_caps;
3213 void *vertex_priv, *fragment_priv;
3214 struct shader_none_priv *priv;
3215
3216 if (!(priv = heap_alloc(sizeof(*priv))))
3217 return E_OUTOFMEMORY;
3218
3219 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3220 {
3221 ERR("Failed to initialize vertex pipe.\n");
3222 heap_free(priv);
3223 return E_FAIL;
3224 }
3225
3226 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3227 {
3228 ERR("Failed to initialize fragment pipe.\n");
3229 vertex_pipe->vp_free(device, NULL);
3230 heap_free(priv);
3231 return E_FAIL;
3232 }
3233
3234 priv->vertex_pipe = vertex_pipe;
3235 priv->fragment_pipe = fragment_pipe;
3236 fragment_pipe->get_caps(device->adapter, &fragment_caps);
3237 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3238
3239 device->vertex_priv = vertex_priv;
3240 device->fragment_priv = fragment_priv;
3241 device->shader_priv = priv;
3242
3243 return WINED3D_OK;
3244 }
3245
3246 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
3247 {
3248 struct shader_none_priv *priv = device->shader_priv;
3249
3250 priv->fragment_pipe->free_private(device, context);
3251 priv->vertex_pipe->vp_free(device, context);
3252 heap_free(priv);
3253 }
3254
3255 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3256 {
3257 return TRUE;
3258 }
3259
3260 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
3261 {
3262 /* Set the shader caps to 0 for the none shader backend */
3263 memset(caps, 0, sizeof(*caps));
3264 }
3265
3266 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3267 {
3268 /* We "support" every possible fixup, since we don't support any shader
3269 * model, and will never have to actually sample a texture. */
3270 return TRUE;
3271 }
3272
3273 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3274 {
3275 struct shader_none_priv *priv = shader_priv;
3276
3277 return priv->ffp_proj_control;
3278 }
3279
3280 static uint64_t shader_none_shader_compile(struct wined3d_context *context, const struct wined3d_shader_desc *shader_desc,
3281 enum wined3d_shader_type shader_type)
3282 {
3283 return 0;
3284 }
3285
3286 const struct wined3d_shader_backend_ops none_shader_backend =
3287 {
3288 shader_none_handle_instruction,
3289 shader_none_precompile,
3290 shader_none_select,
3291 shader_none_select_compute,
3292 shader_none_disable,
3293 shader_none_update_float_vertex_constants,
3294 shader_none_update_float_pixel_constants,
3295 shader_none_load_constants,
3296 shader_none_destroy,
3297 shader_none_alloc,
3298 shader_none_free,
3299 shader_none_allocate_context_data,
3300 shader_none_free_context_data,
3301 shader_none_init_context_state,
3302 shader_none_get_caps,
3303 shader_none_color_fixup_supported,
3304 shader_none_has_ffp_proj_control,
3305 shader_none_shader_compile,
3306 };
3307
3308 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
3309 {
3310 switch (level)
3311 {
3312 case WINED3D_FEATURE_LEVEL_11_1:
3313 case WINED3D_FEATURE_LEVEL_11:
3314 return 5;
3315 case WINED3D_FEATURE_LEVEL_10_1:
3316 case WINED3D_FEATURE_LEVEL_10:
3317 return 4;
3318 case WINED3D_FEATURE_LEVEL_9_3:
3319 return 3;
3320 case WINED3D_FEATURE_LEVEL_9_2:
3321 case WINED3D_FEATURE_LEVEL_9_1:
3322 return 2;
3323 default:
3324 return 1;
3325 }
3326 }
3327
3328 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
3329 enum wined3d_shader_type type, unsigned int float_const_count)
3330 {
3331 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3332 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3333 const struct wined3d_shader_version *version = &reg_maps->shader_version;
3334 const struct wined3d_shader_frontend *fe;
3335 unsigned int backend_version;
3336 HRESULT hr;
3337
3338 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
3339 shader, device, debug_shader_type(type), float_const_count);
3340
3341 fe = shader->frontend;
3342 if (!(shader->frontend_data = fe->shader_init(shader->function,
3343 shader->functionLength, &shader->output_signature)))
3344 {
3345 FIXME("Failed to initialize frontend.\n");
3346 return WINED3DERR_INVALIDCALL;
3347 }
3348
3349 /* First pass: trace shader. */
3350 if (TRACE_ON(d3d_shader))
3351 shader_trace_init(fe, shader->frontend_data);
3352
3353 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3354 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
3355 return hr;
3356
3357 if (version->type != type)
3358 {
3359 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
3360 return WINED3DERR_INVALIDCALL;
3361 }
3362 if (version->major > shader_max_version_from_feature_level(device->cs->c.state->feature_level))
3363 {
3364 WARN("Shader version %u not supported by this device.\n", version->major);
3365 return WINED3DERR_INVALIDCALL;
3366 }
3367 switch (type)
3368 {
3369 case WINED3D_SHADER_TYPE_VERTEX:
3370 backend_version = d3d_info->limits.vs_version;
3371 break;
3372 case WINED3D_SHADER_TYPE_HULL:
3373 backend_version = d3d_info->limits.hs_version;
3374 break;
3375 case WINED3D_SHADER_TYPE_DOMAIN:
3376 backend_version = d3d_info->limits.ds_version;
3377 break;
3378 case WINED3D_SHADER_TYPE_GEOMETRY:
3379 backend_version = d3d_info->limits.gs_version;
3380 break;
3381 case WINED3D_SHADER_TYPE_PIXEL:
3382 backend_version = d3d_info->limits.ps_version;
3383 break;
3384 case WINED3D_SHADER_TYPE_COMPUTE:
3385 backend_version = d3d_info->limits.cs_version;
3386 break;
3387 default:
3388 FIXME("No backend version-checking for this shader type.\n");
3389 backend_version = 0;
3390 }
3391 if (version->major > backend_version)
3392 {
3393 WARN("Shader version %u.%u not supported by the current shader backend.\n",
3394 version->major, version->minor);
3395 return WINED3DERR_INVALIDCALL;
3396 }
3397
3398 shader->load_local_constsF = shader->lconst_inf_or_nan;
3399
3400 return WINED3D_OK;
3401 }
3402
3403 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3404 {
3405 ULONG refcount = InterlockedIncrement(&shader->ref);
3406
3407 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3408
3409 return refcount;
3410 }
3411
3412 static void wined3d_shader_init_object(void *object)
3413 {
3414 struct wined3d_shader *shader = object;
3415 struct wined3d_device *device = shader->device;
3416
3417 TRACE("shader %p.\n", shader);
3418
3419 list_add_head(&device->shaders, &shader->shader_list_entry);
3420
3421 device->shader_backend->shader_precompile(device->shader_priv, shader);
3422 }
3423
3424 static void wined3d_shader_destroy_object(void *object)
3425 {
3426 TRACE("object %p.\n", object);
3427
3428 shader_cleanup(object);
3429 heap_free(object);
3430 }
3431
3432 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3433 {
3434 ULONG refcount = InterlockedDecrement(&shader->ref);
3435
3436 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3437
3438 if (!refcount)
3439 {
3440 wined3d_mutex_lock();
3441 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3442 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3443 wined3d_mutex_unlock();
3444 }
3445
3446 return refcount;
3447 }
3448
3449 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3450 {
3451 TRACE("shader %p.\n", shader);
3452
3453 return shader->parent;
3454 }
3455
3456 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3457 void *byte_code, UINT *byte_code_size)
3458 {
3459 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3460
3461 if (!byte_code)
3462 {
3463 *byte_code_size = shader->byte_code_size;
3464 return WINED3D_OK;
3465 }
3466
3467 if (*byte_code_size < shader->byte_code_size)
3468 {
3469 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3470 * than the required size we should write the required size and
3471 * return D3DERR_MOREDATA. That's not actually true. */
3472 return WINED3DERR_INVALIDCALL;
3473 }
3474
3475 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
3476
3477 return WINED3D_OK;
3478 }
3479
3480 /* Set local constants for d3d8 shaders. */
3481 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3482 UINT start_idx, const float *src_data, UINT count)
3483 {
3484 UINT end_idx = start_idx + count;
3485 UINT i;
3486
3487 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3488
3489 if (end_idx > shader->limits->constant_float)
3490 {
3491 WARN("end_idx %u > float constants limit %u.\n",
3492 end_idx, shader->limits->constant_float);
3493 end_idx = shader->limits->constant_float;
3494 }
3495
3496 for (i = start_idx; i < end_idx; ++i)
3497 {
3498 struct wined3d_shader_lconst *lconst;
3499 float *value;
3500
3501 if (!(lconst = heap_alloc(sizeof(*lconst))))
3502 return E_OUTOFMEMORY;
3503
3504 lconst->idx = i;
3505 value = (float *)lconst->value;
3506 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3507 list_add_head(&shader->constantsF, &lconst->entry);
3508
3509 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3510 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3511 {
3512 shader->lconst_inf_or_nan = TRUE;
3513 }
3514 }
3515
3516 return WINED3D_OK;
3517 }
3518
3519 static void init_interpolation_compile_args(DWORD *interpolation_args,
3520 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
3521 {
3522 if (!d3d_info->shader_output_interpolation || !pixel_shader
3523 || pixel_shader->reg_maps.shader_version.major < 4)
3524 {
3525 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3526 return;
3527 }
3528
3529 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3530 sizeof(pixel_shader->u.ps.interpolation_mode));
3531 }
3532
3533 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3534 struct vs_compile_args *args, const struct wined3d_context *context)
3535 {
3536 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3537 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3538 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3539 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3540 WORD swizzle_map = context->stream_info.swizzle_map;
3541
3542 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3543 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3544 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3545 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3546 args->point_size = state->primitive_type == WINED3D_PT_POINTLIST;
3547 args->per_vertex_point_size = shader->reg_maps.point_size;
3548 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
3549 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3550 if (shader->reg_maps.shader_version.major >= 4)
3551 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3552 : geometry_shader ? geometry_shader->limits->packed_input
3553 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3554 else
3555 args->next_shader_input_count = 0;
3556 args->swizzle_map = swizzle_map;
3557 if (d3d_info->emulated_flatshading)
3558 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3559 else
3560 args->flatshading = 0;
3561
3562 init_interpolation_compile_args(args->interpolation_mode,
3563 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
3564 }
3565
3566 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3567 {
3568 if (usage_idx1 != usage_idx2)
3569 return FALSE;
3570 if (usage1 == usage2)
3571 return TRUE;
3572 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3573 return TRUE;
3574 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3575 return TRUE;
3576
3577 return FALSE;
3578 }
3579
3580 bool vshader_get_input(const struct wined3d_shader *shader,
3581 uint8_t usage_req, uint8_t usage_idx_req, unsigned int *regnum)
3582 {
3583 uint32_t map = shader->reg_maps.input_registers & 0xffff;
3584 unsigned int i;
3585
3586 while (map)
3587 {
3588 i = wined3d_bit_scan(&map);
3589 if (match_usage(shader->u.vs.attributes[i].usage,
3590 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3591 {
3592 *regnum = i;
3593 return true;
3594 }
3595 }
3596
3597 return false;
3598 }
3599
3600 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3601 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3602 {
3603 HRESULT hr;
3604
3605 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
3606
3607 if (!desc->byte_code)
3608 return WINED3DERR_INVALIDCALL;
3609
3610 shader->ref = 1;
3611 shader->device = device;
3612 shader->parent = parent;
3613 shader->parent_ops = parent_ops;
3614
3615 list_init(&shader->linked_programs);
3616 list_init(&shader->constantsF);
3617 list_init(&shader->constantsB);
3618 list_init(&shader->constantsI);
3619 shader->lconst_inf_or_nan = FALSE;
3620 list_init(&shader->reg_maps.indexable_temps);
3621 list_init(&shader->shader_list_entry);
3622
3623 if (desc->byte_code_size == ~(size_t)0)
3624 {
3625 struct wined3d_shader_version shader_version;
3626 const struct wined3d_shader_frontend *fe;
3627 struct wined3d_shader_instruction ins;
3628 const DWORD *ptr;
3629 void *fe_data;
3630
3631 if (!(shader->frontend = shader_select_frontend(WINED3D_SHADER_BYTE_CODE_FORMAT_SM1)))
3632 {
3633 FIXME("Unable to find frontend for shader.\n");
3634 hr = WINED3DERR_INVALIDCALL;
3635 goto fail;
3636 }
3637
3638 fe = shader->frontend;
3639 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
3640 {
3641 WARN("Failed to initialise frontend data.\n");
3642 hr = WINED3DERR_INVALIDCALL;
3643 goto fail;
3644 }
3645
3646 fe->shader_read_header(fe_data, &ptr, &shader_version);
3647 while (!fe->shader_is_end(fe_data, &ptr))
3648 fe->shader_read_instruction(fe_data, &ptr, &ins);
3649
3650 fe->shader_free(fe_data);
3651
3652 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3653
3654 if (!(shader->byte_code = heap_alloc(shader->byte_code_size)))
3655 {
3656 hr = E_OUTOFMEMORY;
3657 goto fail;
3658 }
3659 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
3660
3661 shader->function = shader->byte_code;
3662 shader->functionLength = shader->byte_code_size;
3663 }
3664 else
3665 {
3666 enum wined3d_shader_byte_code_format format;
3667 unsigned int max_version;
3668
3669 if (!(shader->byte_code = heap_alloc(desc->byte_code_size)))
3670 {
3671 hr = E_OUTOFMEMORY;
3672 goto fail;
3673 }
3674 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
3675 shader->byte_code_size = desc->byte_code_size;
3676
3677 max_version = shader_max_version_from_feature_level(device->cs->c.state->feature_level);
3678 if (FAILED(hr = shader_extract_from_dxbc(shader, max_version, &format)))
3679 goto fail;
3680
3681 if (!(shader->frontend = shader_select_frontend(format)))
3682 {
3683 FIXME("Unable to find frontend for shader.\n");
3684 hr = WINED3DERR_INVALIDCALL;
3685 goto fail;
3686 }
3687 }
3688
3689 return WINED3D_OK;
3690
3691 fail:
3692 shader_cleanup(shader);
3693 return hr;
3694 }
3695
3696 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3697 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3698 {
3699 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3700 unsigned int i;
3701 HRESULT hr;
3702
3703 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3704 return hr;
3705
3706 if (FAILED(hr = shader_set_function(shader, device,
3707 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
3708 {
3709 shader_cleanup(shader);
3710 return hr;
3711 }
3712
3713 for (i = 0; i < shader->input_signature.element_count; ++i)
3714 {
3715 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3716
3717 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3718 continue;
3719
3720 shader->u.vs.attributes[input->register_idx].usage =
3721 shader_usage_from_semantic_name(input->semantic_name);
3722 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3723 }
3724
3725 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3726 shader->load_local_constsF = TRUE;
3727
3728 return WINED3D_OK;
3729 }
3730
3731 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
3732 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
3733 {
3734 struct wined3d_shader_signature_element *e = s->elements;
3735 unsigned int i;
3736
3737 for (i = 0; i < s->element_count; ++i)
3738 {
3739 if (e[i].stream_idx == stream_idx
3740 && !stricmp(e[i].semantic_name, semantic_name)
3741 && e[i].semantic_idx == semantic_idx)
3742 return &e[i];
3743 }
3744
3745 return NULL;
3746 }
3747
3748 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
3749 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
3750 {
3751 const struct wined3d_shader_signature_element *output;
3752 unsigned int idx;
3753
3754 if (!(output = shader_find_signature_element(&shader->output_signature,
3755 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
3756 return FALSE;
3757
3758 for (idx = 0; idx < 4; ++idx)
3759 {
3760 if (output->mask & (1u << idx))
3761 break;
3762 }
3763 idx += so_element->component_idx;
3764
3765 *register_idx = output->register_idx;
3766 *component_idx = idx;
3767 return TRUE;
3768 }
3769
3770 static HRESULT geometry_shader_init_so_desc(struct wined3d_geometry_shader *gs, struct wined3d_device *device,
3771 const struct wined3d_stream_output_desc *so_desc)
3772 {
3773 struct wined3d_so_desc_entry *s;
3774 struct wine_rb_entry *entry;
3775 unsigned int i;
3776 size_t size;
3777 char *name;
3778
3779 if ((entry = wine_rb_get(&device->so_descs, so_desc)))
3780 {
3781 gs->so_desc = &WINE_RB_ENTRY_VALUE(entry, struct wined3d_so_desc_entry, entry)->desc;
3782 return WINED3D_OK;
3783 }
3784
3785 size = FIELD_OFFSET(struct wined3d_so_desc_entry, elements[so_desc->element_count]);
3786 for (i = 0; i < so_desc->element_count; ++i)
3787 {
3788 const char *n = so_desc->elements[i].semantic_name;
3789
3790 if (n)
3791 size += strlen(n) + 1;
3792 }
3793 if (!(s = heap_alloc(size)))
3794 return E_OUTOFMEMORY;
3795
3796 s->desc = *so_desc;
3797
3798 memcpy(s->elements, so_desc->elements, so_desc->element_count * sizeof(*s->elements));
3799 s->desc.elements = s->elements;
3800
3801 name = (char *)&s->elements[s->desc.element_count];
3802 for (i = 0; i < so_desc->element_count; ++i)
3803 {
3804 struct wined3d_stream_output_element *e = &s->elements[i];
3805
3806 if (!e->semantic_name)
3807 continue;
3808
3809 size = strlen(e->semantic_name) + 1;
3810 memcpy(name, e->semantic_name, size);
3811 e->semantic_name = name;
3812 name += size;
3813 }
3814
3815 if (wine_rb_put(&device->so_descs, &s->desc, &s->entry) == -1)
3816 {
3817 heap_free(s);
3818 return E_FAIL;
3819 }
3820 gs->so_desc = &s->desc;
3821
3822 return WINED3D_OK;
3823 }
3824
3825 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
3826 const struct wined3d_stream_output_desc *so_desc)
3827 {
3828 const struct wined3d_shader_frontend *fe = shader->frontend;
3829 const struct wined3d_shader_signature_element *output;
3830 unsigned int i, component_idx, register_idx, mask;
3831 struct wined3d_shader_version shader_version;
3832 const DWORD *ptr;
3833 void *fe_data;
3834 HRESULT hr;
3835
3836 if (!so_desc)
3837 return WINED3D_OK;
3838
3839 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
3840 {
3841 WARN("Failed to initialise frontend data.\n");
3842 return WINED3DERR_INVALIDCALL;
3843 }
3844 fe->shader_read_header(fe_data, &ptr, &shader_version);
3845 fe->shader_free(fe_data);
3846
3847 switch (shader_version.type)
3848 {
3849 case WINED3D_SHADER_TYPE_VERTEX:
3850 case WINED3D_SHADER_TYPE_DOMAIN:
3851 shader->function = NULL;
3852 shader->functionLength = 0;
3853 break;
3854 case WINED3D_SHADER_TYPE_GEOMETRY:
3855 break;
3856 default:
3857 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
3858 return E_INVALIDARG;
3859 }
3860
3861 if (!shader->function)
3862 {
3863 shader->reg_maps.shader_version = shader_version;
3864 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
3865 shader_set_limits(shader);
3866 if (FAILED(hr = shader_scan_output_signature(shader)))
3867 return hr;
3868 }
3869
3870 for (i = 0; i < so_desc->element_count; ++i)
3871 {
3872 const struct wined3d_stream_output_element *e = &so_desc->elements[i];
3873
3874 if (!e->semantic_name)
3875 continue;
3876 if (!(output = shader_find_signature_element(&shader->output_signature,
3877 e->stream_idx, e->semantic_name, e->semantic_idx))
3878 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
3879 {
3880 WARN("Failed to find output signature element for stream output entry.\n");
3881 return E_INVALIDARG;
3882 }
3883
3884 mask = wined3d_mask_from_size(e->component_count) << component_idx;
3885 if ((output->mask & 0xff & mask) != mask)
3886 {
3887 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
3888 component_idx, e->component_count, mask, output->mask & 0xff);
3889 return E_INVALIDARG;
3890 }
3891 }
3892
3893 if (FAILED(hr = geometry_shader_init_so_desc(&shader->u.gs, shader->device, so_desc)))
3894 {
3895 WARN("Failed to initialise stream output description, hr %#x.\n", hr);
3896 return hr;
3897 }
3898
3899 return WINED3D_OK;
3900 }
3901
3902 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3903 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3904 void *parent, const struct wined3d_parent_ops *parent_ops)
3905 {
3906 HRESULT hr;
3907
3908 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3909 return hr;
3910
3911 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
3912 goto fail;
3913
3914 if (shader->function
3915 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
3916 goto fail;
3917
3918 return WINED3D_OK;
3919
3920 fail:
3921 shader_cleanup(shader);
3922 return hr;
3923 }
3924
3925 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3926 struct ds_compile_args *args, const struct wined3d_context *context)
3927 {
3928 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3929 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3930 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3931
3932 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3933 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3934
3935 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3936 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3937 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3938
3939 args->render_offscreen = context->render_offscreen;
3940
3941 init_interpolation_compile_args(args->interpolation_mode,
3942 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
3943
3944 args->padding = 0;
3945 }
3946
3947 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3948 struct gs_compile_args *args, const struct wined3d_context *context)
3949 {
3950 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3951
3952 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3953
3954 if (!(args->primitive_type = shader->u.gs.input_type))
3955 args->primitive_type = state->primitive_type;
3956
3957 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
3958 }
3959
3960 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3961 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3962 {
3963 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3964 struct wined3d_texture *texture;
3965 unsigned int i;
3966
3967 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3968 if (!d3d_info->srgb_write_control && needs_srgb_write(d3d_info, state, &state->fb))
3969 {
3970 static unsigned int warned = 0;
3971
3972 args->srgb_correction = 1;
3973 if (state->blend_state && state->blend_state->desc.rt[0].enable && !warned++)
3974 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3975 "support, expect rendering artifacts.\n");
3976 }
3977
3978 if (shader->reg_maps.shader_version.major == 1
3979 && shader->reg_maps.shader_version.minor <= 3)
3980 {
3981 for (i = 0; i < shader->limits->sampler; ++i)
3982 {
3983 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3984
3985 if (flags & WINED3D_TTFF_PROJECTED)
3986 {
3987 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3988
3989 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3990 {
3991 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3992 unsigned int j;
3993 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3994 DWORD max_valid = WINED3D_TTFF_COUNT4;
3995
3996 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3997 {
3998 struct wined3d_vertex_declaration_element *element =
3999 &state->vertex_declaration->elements[j];
4000
4001 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
4002 && element->usage_idx == index)
4003 {
4004 max_valid = element->format->component_count;
4005 break;
4006 }
4007 }
4008 if (!tex_transform || tex_transform > max_valid)
4009 {
4010 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
4011 tex_transform, max_valid);
4012 tex_transform = max_valid;
4013 }
4014 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
4015 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
4016 && tex_transform > WINED3D_TTFF_COUNT2)
4017 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4018 && tex_transform > WINED3D_TTFF_COUNT3))
4019 tex_transform |= WINED3D_PSARGS_PROJECTED;
4020 else
4021 {
4022 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
4023 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
4024 i, tex_transform, resource_type);
4025 }
4026 }
4027 else
4028 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
4029
4030 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
4031 }
4032 }
4033 }
4034 if (shader->reg_maps.shader_version.major == 1
4035 && shader->reg_maps.shader_version.minor <= 4)
4036 {
4037 for (i = 0; i < shader->limits->sampler; ++i)
4038 {
4039 if (!shader->reg_maps.resource_info[i].type)
4040 continue;
4041
4042 /* Treat unbound textures as 2D. The dummy texture will provide
4043 * the proper sample value. The tex_types bitmap defaults to
4044 * 2D because of the memset. */
4045 if (!(texture = state->textures[i]))
4046 continue;
4047
4048 switch (wined3d_texture_gl(texture)->target)
4049 {
4050 /* RECT textures are distinguished from 2D textures via np2_fixup */
4051 default:
4052 break;
4053
4054 case GL_TEXTURE_3D:
4055 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4056 break;
4057
4058 case GL_TEXTURE_CUBE_MAP_ARB:
4059 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4060 break;
4061 }
4062 }
4063 }
4064 else if (shader->reg_maps.shader_version.major <= 3)
4065 {
4066 for (i = 0; i < shader->limits->sampler; ++i)
4067 {
4068 enum wined3d_shader_resource_type resource_type;
4069 enum wined3d_shader_tex_types tex_type;
4070
4071 if (!(resource_type = shader->reg_maps.resource_info[i].type))
4072 continue;
4073
4074 switch (resource_type)
4075 {
4076 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
4077 tex_type = WINED3D_SHADER_TEX_3D;
4078 break;
4079 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
4080 tex_type = WINED3D_SHADER_TEX_CUBE;
4081 break;
4082 default:
4083 tex_type = WINED3D_SHADER_TEX_2D;
4084 break;
4085 }
4086
4087 if ((texture = state->textures[i]))
4088 {
4089 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
4090 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
4091 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
4092 tex_type = WINED3D_SHADER_TEX_2D;
4093 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
4094 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
4095 tex_type = WINED3D_SHADER_TEX_3D;
4096 }
4097 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
4098 }
4099 }
4100
4101 if (shader->reg_maps.shader_version.major >= 4)
4102 {
4103 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
4104 args->shadow = 0;
4105 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4106 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4107 args->np2_fixup = 0;
4108 }
4109 else
4110 {
4111 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
4112 {
4113 if (!shader->reg_maps.resource_info[i].type)
4114 continue;
4115
4116 texture = state->textures[i];
4117 if (!texture)
4118 {
4119 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4120 continue;
4121 }
4122 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
4123 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
4124 else
4125 args->color_fixup[i] = texture->resource.format->color_fixup;
4126
4127 if (texture->resource.format_caps & WINED3D_FORMAT_CAP_SHADOW)
4128 args->shadow |= 1u << i;
4129
4130 /* Flag samplers that need NP2 texcoord fixup. */
4131 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
4132 args->np2_fixup |= (1u << i);
4133 }
4134 }
4135
4136 if (shader->reg_maps.shader_version.major >= 3)
4137 {
4138 if (position_transformed)
4139 args->vp_mode = WINED3D_VP_MODE_NONE;
4140 else if (use_vs(state))
4141 args->vp_mode = WINED3D_VP_MODE_SHADER;
4142 else
4143 args->vp_mode = WINED3D_VP_MODE_FF;
4144 args->fog = WINED3D_FFP_PS_FOG_OFF;
4145 }
4146 else
4147 {
4148 args->vp_mode = WINED3D_VP_MODE_SHADER;
4149 if (state->render_states[WINED3D_RS_FOGENABLE])
4150 {
4151 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
4152 {
4153 case WINED3D_FOG_NONE:
4154 if (position_transformed || use_vs(state))
4155 {
4156 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
4157 break;
4158 }
4159
4160 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
4161 {
4162 case WINED3D_FOG_NONE: /* Fall through. */
4163 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4164 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4165 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4166 }
4167 break;
4168
4169 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
4170 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
4171 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
4172 }
4173 }
4174 else
4175 {
4176 args->fog = WINED3D_FFP_PS_FOG_OFF;
4177 }
4178 }
4179
4180 if (!d3d_info->full_ffp_varyings)
4181 {
4182 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
4183
4184 args->texcoords_initialized = 0;
4185 for (i = 0; i < WINED3D_MAX_TEXTURES; ++i)
4186 {
4187 if (vs)
4188 {
4189 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
4190 args->texcoords_initialized |= 1u << i;
4191 }
4192 else
4193 {
4194 const struct wined3d_stream_info *si = &context->stream_info;
4195 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
4196
4197 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
4198 & WINED3D_FFP_TCI_MASK
4199 || (coord_idx < WINED3D_MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
4200 args->texcoords_initialized |= 1u << i;
4201 }
4202 }
4203 }
4204 else
4205 {
4206 args->texcoords_initialized = wined3d_mask_from_size(WINED3D_MAX_TEXTURES);
4207 }
4208
4209 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
4210 && state->primitive_type == WINED3D_PT_POINTLIST;
4211
4212 if (d3d_info->ffp_alpha_test)
4213 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
4214 else
4215 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4216 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4217 : WINED3D_CMP_ALWAYS) - 1;
4218
4219 if (d3d_info->emulated_flatshading)
4220 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4221
4222 args->y_correction = (shader->reg_maps.vpos && d3d_info->frag_coord_correction)
4223 || (shader->reg_maps.usesdsy && wined3d_settings.offscreen_rendering_mode != ORM_FBO)
4224 ? !context->render_offscreen : 0;
4225
4226 for (i = 0; i < ARRAY_SIZE(state->fb.render_targets); ++i)
4227 {
4228 struct wined3d_rendertarget_view *rtv = state->fb.render_targets[i];
4229 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
4230 args->rt_alpha_swizzle |= 1u << i;
4231 }
4232
4233 args->dual_source_blend = state->blend_state && state->blend_state->dual_source;
4234 }
4235
4236 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4237 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4238 {
4239 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4240 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4241 HRESULT hr;
4242
4243 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
4244 return hr;
4245
4246 if (FAILED(hr = shader_set_function(shader, device,
4247 WINED3D_SHADER_TYPE_PIXEL, device->adapter->d3d_info.limits.ps_uniform_count)))
4248 {
4249 shader_cleanup(shader);
4250 return hr;
4251 }
4252
4253 for (i = 0; i < MAX_REG_INPUT; ++i)
4254 {
4255 if (shader->u.ps.input_reg_used & (1u << i))
4256 {
4257 ++num_regs_used;
4258 highest_reg_used = i;
4259 }
4260 }
4261
4262 /* Don't do any register mapping magic if it is not needed, or if we can't
4263 * achieve anything anyway */
4264 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4265 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4266 || shader->reg_maps.shader_version.major >= 4)
4267 {
4268 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4269 {
4270 /* This happens with relative addressing. The input mapper function
4271 * warns about this if the higher registers are declared too, so
4272 * don't write a FIXME here */
4273 WARN("More varying registers used than supported\n");
4274 }
4275
4276 for (i = 0; i < MAX_REG_INPUT; ++i)
4277 {
4278 shader->u.ps.input_reg_map[i] = i;
4279 }
4280
4281 shader->u.ps.declared_in_count = highest_reg_used + 1;
4282 }
4283 else
4284 {
4285 shader->u.ps.declared_in_count = 0;
4286 for (i = 0; i < MAX_REG_INPUT; ++i)
4287 {
4288 if (shader->u.ps.input_reg_used & (1u << i))
4289 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4290 else shader->u.ps.input_reg_map[i] = ~0U;
4291 }
4292 }
4293
4294 return WINED3D_OK;
4295 }
4296
4297 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
4298 unsigned int resource_idx, DWORD tex_types)
4299 {
4300 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
4301 {
4302 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
4303 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
4304 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
4305 };
4306
4307 unsigned int idx;
4308
4309 if (reg_maps->shader_version.major > 3)
4310 return reg_maps->resource_info[resource_idx].type;
4311
4312 if (!reg_maps->resource_info[resource_idx].type)
4313 return 0;
4314
4315 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
4316 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
4317 return shader_resource_type_from_shader_tex_types[idx];
4318 }
4319
4320 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4321 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4322 {
4323 struct wined3d_shader *object;
4324 HRESULT hr;
4325
4326 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4327 device, desc, parent, parent_ops, shader);
4328
4329 if (!(object = heap_alloc_zero(sizeof(*object))))
4330 return E_OUTOFMEMORY;
4331
4332 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4333 {
4334 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4335 heap_free(object);
4336 return hr;
4337 }
4338
4339 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
4340 {
4341 shader_cleanup(object);
4342 heap_free(object);
4343 return hr;
4344 }
4345
4346 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4347
4348 TRACE("Created compute shader %p.\n", object);
4349 *shader = object;
4350
4351 return WINED3D_OK;
4352 }
4353
4354 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4355 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4356 {
4357 struct wined3d_shader *object;
4358 HRESULT hr;
4359
4360 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4361 device, desc, parent, parent_ops, shader);
4362
4363 if (!(object = heap_alloc_zero(sizeof(*object))))
4364 return E_OUTOFMEMORY;
4365
4366 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4367 {
4368 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4369 heap_free(object);
4370 return hr;
4371 }
4372
4373 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
4374 {
4375 shader_cleanup(object);
4376 heap_free(object);
4377 return hr;
4378 }
4379
4380 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4381
4382 TRACE("Created domain shader %p.\n", object);
4383 *shader = object;
4384
4385 return WINED3D_OK;
4386 }
4387
4388 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4389 const struct wined3d_stream_output_desc *so_desc, void *parent,
4390 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4391 {
4392 struct wined3d_shader *object;
4393 HRESULT hr;
4394
4395 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4396 device, desc, so_desc, parent, parent_ops, shader);
4397
4398 if (!(object = heap_alloc_zero(sizeof(*object))))
4399 return E_OUTOFMEMORY;
4400
4401 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4402 {
4403 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4404 heap_free(object);
4405 return hr;
4406 }
4407
4408 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4409
4410 TRACE("Created geometry shader %p.\n", object);
4411 *shader = object;
4412
4413 return WINED3D_OK;
4414 }
4415
4416 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4417 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4418 {
4419 struct wined3d_shader *object;
4420 HRESULT hr;
4421
4422 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4423 device, desc, parent, parent_ops, shader);
4424
4425 if (!(object = heap_alloc_zero(sizeof(*object))))
4426 return E_OUTOFMEMORY;
4427
4428 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
4429 {
4430 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4431 heap_free(object);
4432 return hr;
4433 }
4434
4435 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
4436 {
4437 shader_cleanup(object);
4438 heap_free(object);
4439 return hr;
4440 }
4441
4442 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4443
4444 TRACE("Created hull shader %p.\n", object);
4445 *shader = object;
4446
4447 return WINED3D_OK;
4448 }
4449
4450 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4451 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4452 {
4453 struct wined3d_shader *object;
4454 HRESULT hr;
4455
4456 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4457 device, desc, parent, parent_ops, shader);
4458
4459 if (!(object = heap_alloc_zero(sizeof(*object))))
4460 return E_OUTOFMEMORY;
4461
4462 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4463 {
4464 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4465 heap_free(object);
4466 return hr;
4467 }
4468
4469 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4470
4471 TRACE("Created pixel shader %p.\n", object);
4472 *shader = object;
4473
4474 return WINED3D_OK;
4475 }
4476
4477 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4478 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4479 {
4480 struct wined3d_shader *object;
4481 HRESULT hr;
4482
4483 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4484 device, desc, parent, parent_ops, shader);
4485
4486 if (!(object = heap_alloc_zero(sizeof(*object))))
4487 return E_OUTOFMEMORY;
4488
4489 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4490 {
4491 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4492 heap_free(object);
4493 return hr;
4494 }
4495
4496 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
4497
4498 TRACE("Created vertex shader %p.\n", object);
4499 *shader = object;
4500
4501 return WINED3D_OK;
4502 }
Windows API implementation