?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
winhttp: Avoid returning while holding a lock (Coverity).
[wine.git] / dlls / wined3d / shader.c
blob8ed2cb9ed97c4f48de864b7f64019d6a6ab6d300
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <math.h>
29 #include <stdio.h>
30 #include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
47 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
48 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
49 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
50 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
51 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
52 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
53 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
54 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
55 /* WINED3DSIH_BEM */ "bem",
56 /* WINED3DSIH_BFI */ "bfi",
57 /* WINED3DSIH_BFREV */ "bfrev",
58 /* WINED3DSIH_BREAK */ "break",
59 /* WINED3DSIH_BREAKC */ "breakc",
60 /* WINED3DSIH_BREAKP */ "breakp",
61 /* WINED3DSIH_BUFINFO */ "bufinfo",
62 /* WINED3DSIH_CALL */ "call",
63 /* WINED3DSIH_CALLNZ */ "callnz",
64 /* WINED3DSIH_CASE */ "case",
65 /* WINED3DSIH_CMP */ "cmp",
66 /* WINED3DSIH_CND */ "cnd",
67 /* WINED3DSIH_CONTINUE */ "continue",
68 /* WINED3DSIH_CONTINUEP */ "continuec",
69 /* WINED3DSIH_COUNTBITS */ "countbits",
70 /* WINED3DSIH_CRS */ "crs",
71 /* WINED3DSIH_CUT */ "cut",
72 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
73 /* WINED3DSIH_DCL */ "dcl",
74 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
75 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
76 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
77 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
78 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
79 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
81 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
82 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
83 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
84 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
85 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
86 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
87 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
88 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
89 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
90 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
91 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
92 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
93 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
94 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
95 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
96 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
97 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
98 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
99 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
100 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
101 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
102 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
103 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
104 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
105 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
106 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
107 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
108 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
109 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
110 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
111 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
112 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
113 /* WINED3DSIH_DEF */ "def",
114 /* WINED3DSIH_DEFAULT */ "default",
115 /* WINED3DSIH_DEFB */ "defb",
116 /* WINED3DSIH_DEFI */ "defi",
117 /* WINED3DSIH_DIV */ "div",
118 /* WINED3DSIH_DP2 */ "dp2",
119 /* WINED3DSIH_DP2ADD */ "dp2add",
120 /* WINED3DSIH_DP3 */ "dp3",
121 /* WINED3DSIH_DP4 */ "dp4",
122 /* WINED3DSIH_DST */ "dst",
123 /* WINED3DSIH_DSX */ "dsx",
124 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
125 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
126 /* WINED3DSIH_DSY */ "dsy",
127 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
128 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
129 /* WINED3DSIH_ELSE */ "else",
130 /* WINED3DSIH_EMIT */ "emit",
131 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
132 /* WINED3DSIH_ENDIF */ "endif",
133 /* WINED3DSIH_ENDLOOP */ "endloop",
134 /* WINED3DSIH_ENDREP */ "endrep",
135 /* WINED3DSIH_ENDSWITCH */ "endswitch",
136 /* WINED3DSIH_EQ */ "eq",
137 /* WINED3DSIH_EXP */ "exp",
138 /* WINED3DSIH_EXPP */ "expp",
139 /* WINED3DSIH_F16TOF32 */ "f16tof32",
140 /* WINED3DSIH_F32TOF16 */ "f32tof16",
141 /* WINED3DSIH_FCALL */ "fcall",
142 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
143 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
144 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
145 /* WINED3DSIH_FRC */ "frc",
146 /* WINED3DSIH_FTOI */ "ftoi",
147 /* WINED3DSIH_FTOU */ "ftou",
148 /* WINED3DSIH_GATHER4 */ "gather4",
149 /* WINED3DSIH_GATHER4_C */ "gather4_c",
150 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
151 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
152 /* WINED3DSIH_GE */ "ge",
153 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
154 /* WINED3DSIH_HS_DECLS */ "hs_decls",
155 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
156 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
157 /* WINED3DSIH_IADD */ "iadd",
158 /* WINED3DSIH_IBFE */ "ibfe",
159 /* WINED3DSIH_IEQ */ "ieq",
160 /* WINED3DSIH_IF */ "if",
161 /* WINED3DSIH_IFC */ "ifc",
162 /* WINED3DSIH_IGE */ "ige",
163 /* WINED3DSIH_ILT */ "ilt",
164 /* WINED3DSIH_IMAD */ "imad",
165 /* WINED3DSIH_IMAX */ "imax",
166 /* WINED3DSIH_IMIN */ "imin",
167 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
168 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
169 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
170 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
171 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
172 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
173 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
174 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
175 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
176 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
177 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
178 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
179 /* WINED3DSIH_IMUL */ "imul",
180 /* WINED3DSIH_INE */ "ine",
181 /* WINED3DSIH_INEG */ "ineg",
182 /* WINED3DSIH_ISHL */ "ishl",
183 /* WINED3DSIH_ISHR */ "ishr",
184 /* WINED3DSIH_ITOF */ "itof",
185 /* WINED3DSIH_LABEL */ "label",
186 /* WINED3DSIH_LD */ "ld",
187 /* WINED3DSIH_LD2DMS */ "ld2dms",
188 /* WINED3DSIH_LD_RAW */ "ld_raw",
189 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
190 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
191 /* WINED3DSIH_LIT */ "lit",
192 /* WINED3DSIH_LOD */ "lod",
193 /* WINED3DSIH_LOG */ "log",
194 /* WINED3DSIH_LOGP */ "logp",
195 /* WINED3DSIH_LOOP */ "loop",
196 /* WINED3DSIH_LRP */ "lrp",
197 /* WINED3DSIH_LT */ "lt",
198 /* WINED3DSIH_M3x2 */ "m3x2",
199 /* WINED3DSIH_M3x3 */ "m3x3",
200 /* WINED3DSIH_M3x4 */ "m3x4",
201 /* WINED3DSIH_M4x3 */ "m4x3",
202 /* WINED3DSIH_M4x4 */ "m4x4",
203 /* WINED3DSIH_MAD */ "mad",
204 /* WINED3DSIH_MAX */ "max",
205 /* WINED3DSIH_MIN */ "min",
206 /* WINED3DSIH_MOV */ "mov",
207 /* WINED3DSIH_MOVA */ "mova",
208 /* WINED3DSIH_MOVC */ "movc",
209 /* WINED3DSIH_MUL */ "mul",
210 /* WINED3DSIH_NE */ "ne",
211 /* WINED3DSIH_NOP */ "nop",
212 /* WINED3DSIH_NOT */ "not",
213 /* WINED3DSIH_NRM */ "nrm",
214 /* WINED3DSIH_OR */ "or",
215 /* WINED3DSIH_PHASE */ "phase",
216 /* WINED3DSIH_POW */ "pow",
217 /* WINED3DSIH_RCP */ "rcp",
218 /* WINED3DSIH_REP */ "rep",
219 /* WINED3DSIH_RESINFO */ "resinfo",
220 /* WINED3DSIH_RET */ "ret",
221 /* WINED3DSIH_RETP */ "retp",
222 /* WINED3DSIH_ROUND_NE */ "round_ne",
223 /* WINED3DSIH_ROUND_NI */ "round_ni",
224 /* WINED3DSIH_ROUND_PI */ "round_pi",
225 /* WINED3DSIH_ROUND_Z */ "round_z",
226 /* WINED3DSIH_RSQ */ "rsq",
227 /* WINED3DSIH_SAMPLE */ "sample",
228 /* WINED3DSIH_SAMPLE_B */ "sample_b",
229 /* WINED3DSIH_SAMPLE_C */ "sample_c",
230 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
231 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
232 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
233 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
234 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
235 /* WINED3DSIH_SETP */ "setp",
236 /* WINED3DSIH_SGE */ "sge",
237 /* WINED3DSIH_SGN */ "sgn",
238 /* WINED3DSIH_SINCOS */ "sincos",
239 /* WINED3DSIH_SLT */ "slt",
240 /* WINED3DSIH_SQRT */ "sqrt",
241 /* WINED3DSIH_STORE_RAW */ "store_raw",
242 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
243 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
244 /* WINED3DSIH_SUB */ "sub",
245 /* WINED3DSIH_SWAPC */ "swapc",
246 /* WINED3DSIH_SWITCH */ "switch",
247 /* WINED3DSIH_SYNC */ "sync",
248 /* WINED3DSIH_TEX */ "texld",
249 /* WINED3DSIH_TEXBEM */ "texbem",
250 /* WINED3DSIH_TEXBEML */ "texbeml",
251 /* WINED3DSIH_TEXCOORD */ "texcrd",
252 /* WINED3DSIH_TEXDEPTH */ "texdepth",
253 /* WINED3DSIH_TEXDP3 */ "texdp3",
254 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
255 /* WINED3DSIH_TEXKILL */ "texkill",
256 /* WINED3DSIH_TEXLDD */ "texldd",
257 /* WINED3DSIH_TEXLDL */ "texldl",
258 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
259 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
260 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
261 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
262 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
263 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
264 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
265 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
266 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
267 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
268 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
269 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
270 /* WINED3DSIH_UBFE */ "ubfe",
271 /* WINED3DSIH_UDIV */ "udiv",
272 /* WINED3DSIH_UGE */ "uge",
273 /* WINED3DSIH_ULT */ "ult",
274 /* WINED3DSIH_UMAX */ "umax",
275 /* WINED3DSIH_UMIN */ "umin",
276 /* WINED3DSIH_UMUL */ "umul",
277 /* WINED3DSIH_USHR */ "ushr",
278 /* WINED3DSIH_UTOF */ "utof",
279 /* WINED3DSIH_XOR */ "xor",
280 };
281
282 static const char * const semantic_names[] =
283 {
284 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
285 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
286 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
287 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
288 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
289 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
290 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
291 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
292 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
293 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
294 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
295 /* WINED3D_DECL_USAGE_FOG */ "FOG",
296 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
297 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
298 };
299
300 static const struct
301 {
302 enum wined3d_shader_input_sysval_semantic sysval_semantic;
303 const char *sysval_name;
304 }
305 shader_input_sysval_semantic_names[] =
306 {
307 {WINED3D_SIV_POSITION, "position"},
308 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
309 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
310 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
311 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
312 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
313 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
314 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
315 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
316 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
317 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
318 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
319 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
320 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
322 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
323 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
324 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
325 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
326 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
327 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
328 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
329 };
330
331 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
332 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
333
334 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
335 {
336 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
337 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
338
339 return shader_opcode_names[handler_idx];
340 }
341
342 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
343 {
344 if (usage >= ARRAY_SIZE(semantic_names))
345 {
346 FIXME("Unrecognized usage %#x.\n", usage);
347 return "UNRECOGNIZED";
348 }
349
350 return semantic_names[usage];
351 }
352
353 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
354 {
355 unsigned int i;
356
357 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
358 {
359 if (!strcmp(name, semantic_names[i]))
360 return i;
361 }
362
363 return ~0U;
364 }
365
366 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
367 {
368 switch (usage)
369 {
370 case WINED3D_DECL_USAGE_POSITION:
371 return WINED3D_SV_POSITION;
372 default:
373 return 0;
374 }
375 }
376
377 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
378 {
379 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
380 }
381
382 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
383 const struct wined3d_shader_semantic *s)
384 {
385 e->semantic_name = shader_semantic_name_from_usage(s->usage);
386 e->semantic_idx = s->usage_idx;
387 e->stream_idx = 0;
388 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
389 e->component_type = WINED3D_TYPE_FLOAT;
390 e->register_idx = s->reg.reg.idx[0].offset;
391 e->mask = s->reg.write_mask;
392 }
393
394 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
395 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
396 {
397 e->semantic_name = shader_semantic_name_from_usage(usage);
398 e->semantic_idx = usage_idx;
399 e->stream_idx = 0;
400 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
401 e->component_type = WINED3D_TYPE_FLOAT;
402 e->register_idx = reg_idx;
403 e->mask = write_mask;
404 }
405
406 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
407 {
408 switch (format)
409 {
410 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
411 return &sm1_shader_frontend;
412
413 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
414 return &sm4_shader_frontend;
415
416 default:
417 WARN("Invalid byte code format %#x specified.\n", format);
418 return NULL;
419 }
420 }
421
422 void string_buffer_clear(struct wined3d_string_buffer *buffer)
423 {
424 buffer->buffer[0] = '\0';
425 buffer->content_size = 0;
426 }
427
428 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
429 {
430 buffer->buffer_size = 32;
431 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
432 {
433 ERR("Failed to allocate shader buffer memory.\n");
434 return FALSE;
435 }
436
437 string_buffer_clear(buffer);
438 return TRUE;
439 }
440
441 void string_buffer_free(struct wined3d_string_buffer *buffer)
442 {
443 HeapFree(GetProcessHeap(), 0, buffer->buffer);
444 }
445
446 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
447 {
448 char *new_buffer;
449 unsigned int new_buffer_size = buffer->buffer_size * 2;
450
451 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
452 new_buffer_size *= 2;
453 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
454 {
455 ERR("Failed to grow buffer.\n");
456 buffer->buffer[buffer->content_size] = '\0';
457 return FALSE;
458 }
459 buffer->buffer = new_buffer;
460 buffer->buffer_size = new_buffer_size;
461 return TRUE;
462 }
463
464 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
465 {
466 unsigned int rem;
467 int rc;
468
469 rem = buffer->buffer_size - buffer->content_size;
470 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
471 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
472 return rc;
473
474 buffer->content_size += rc;
475 return 0;
476 }
477
478 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
479 {
480 va_list args;
481 int ret;
482
483 for (;;)
484 {
485 va_start(args, format);
486 ret = shader_vaddline(buffer, format, args);
487 va_end(args);
488 if (!ret)
489 return ret;
490 if (!string_buffer_resize(buffer, ret))
491 return -1;
492 }
493 }
494
495 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
496 {
497 struct wined3d_string_buffer *buffer;
498
499 if (list_empty(&list->list))
500 {
501 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
502 if (!buffer || !string_buffer_init(buffer))
503 {
504 ERR("Couldn't allocate buffer for temporary string.\n");
505 HeapFree(GetProcessHeap(), 0, buffer);
506 return NULL;
507 }
508 }
509 else
510 {
511 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
512 list_remove(&buffer->entry);
513 }
514 string_buffer_clear(buffer);
515 return buffer;
516 }
517
518 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
519 {
520 if (!buffer)
521 return 0;
522 string_buffer_clear(buffer);
523 return shader_vaddline(buffer, format, args);
524 }
525
526 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
527 {
528 va_list args;
529 int ret;
530
531 for (;;)
532 {
533 va_start(args, format);
534 ret = string_buffer_vsprintf(buffer, format, args);
535 va_end(args);
536 if (!ret)
537 return;
538 if (!string_buffer_resize(buffer, ret))
539 return;
540 }
541 }
542
543 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
544 {
545 if (!buffer)
546 return;
547 list_add_head(&list->list, &buffer->entry);
548 }
549
550 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
551 {
552 list_init(&list->list);
553 }
554
555 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
556 {
557 struct wined3d_string_buffer *buffer, *buffer_next;
558
559 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
560 {
561 string_buffer_free(buffer);
562 HeapFree(GetProcessHeap(), 0, buffer);
563 }
564 list_init(&list->list);
565 }
566
567 /* Convert floating point offset relative to a register file to an absolute
568 * offset for float constants. */
569 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
570 {
571 switch (register_type)
572 {
573 case WINED3DSPR_CONST: return register_idx;
574 case WINED3DSPR_CONST2: return 2048 + register_idx;
575 case WINED3DSPR_CONST3: return 4096 + register_idx;
576 case WINED3DSPR_CONST4: return 6144 + register_idx;
577 default:
578 FIXME("Unsupported register type: %u.\n", register_type);
579 return register_idx;
580 }
581 }
582
583 static void shader_delete_constant_list(struct list *clist)
584 {
585 struct wined3d_shader_lconst *constant, *constant_next;
586
587 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
588 HeapFree(GetProcessHeap(), 0, constant);
589 list_init(clist);
590 }
591
592 static void shader_set_limits(struct wined3d_shader *shader)
593 {
594 static const struct limits_entry
595 {
596 unsigned int min_version;
597 unsigned int max_version;
598 struct wined3d_shader_limits limits;
599 }
600 vs_limits[] =
601 {
602 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
603 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
604 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
605 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
606 * even though they are capable of supporting much more (GL
607 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
608 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
609 * shaders to 256. */
610 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
611 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
612 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
613 {0}
614 },
615 hs_limits[] =
616 {
617 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
618 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
619 },
620 ds_limits[] =
621 {
622 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
623 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
624 },
625 gs_limits[] =
626 {
627 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
628 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
629 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
630 {0}
631 },
632 ps_limits[] =
633 {
634 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
635 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
636 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
637 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
638 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
639 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
640 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
641 {0}
642 },
643 cs_limits[] =
644 {
645 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
646 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
647 };
648 const struct limits_entry *limits_array;
649 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
650 shader->reg_maps.shader_version.minor);
651 int i = 0;
652
653 switch (shader->reg_maps.shader_version.type)
654 {
655 default:
656 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
657 /* Fall-through. */
658 case WINED3D_SHADER_TYPE_VERTEX:
659 limits_array = vs_limits;
660 break;
661 case WINED3D_SHADER_TYPE_HULL:
662 limits_array = hs_limits;
663 break;
664 case WINED3D_SHADER_TYPE_DOMAIN:
665 limits_array = ds_limits;
666 break;
667 case WINED3D_SHADER_TYPE_GEOMETRY:
668 limits_array = gs_limits;
669 break;
670 case WINED3D_SHADER_TYPE_PIXEL:
671 limits_array = ps_limits;
672 break;
673 case WINED3D_SHADER_TYPE_COMPUTE:
674 limits_array = cs_limits;
675 break;
676 }
677
678 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
679 {
680 if (shader_version <= limits_array[i].max_version)
681 {
682 shader->limits = &limits_array[i].limits;
683 break;
684 }
685 ++i;
686 }
687 if (!shader->limits)
688 {
689 FIXME("Unexpected shader version \"%u.%u\".\n",
690 shader->reg_maps.shader_version.major,
691 shader->reg_maps.shader_version.minor);
692 shader->limits = &limits_array[max(0, i - 1)].limits;
693 }
694 }
695
696 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
697 {
698 DWORD idx, shift;
699 idx = bit >> 5;
700 shift = bit & 0x1f;
701 bitmap[idx] |= (1u << shift);
702 }
703
704 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
705 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
706 {
707 switch (reg->type)
708 {
709 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
710 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
711 reg_maps->texcoord |= 1u << reg->idx[0].offset;
712 else
713 reg_maps->address |= 1u << reg->idx[0].offset;
714 break;
715
716 case WINED3DSPR_TEMP:
717 reg_maps->temporary |= 1u << reg->idx[0].offset;
718 break;
719
720 case WINED3DSPR_INPUT:
721 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
722 {
723 /* If relative addressing is used, we must assume that all
724 * registers are used. Even if it is a construct like v3[aL],
725 * we can't assume that v0, v1 and v2 aren't read because aL
726 * can be negative. */
727 if (reg->idx[0].rel_addr)
728 shader->u.ps.input_reg_used = ~0u;
729 else
730 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
731 }
732 else
733 reg_maps->input_registers |= 1u << reg->idx[0].offset;
734 break;
735
736 case WINED3DSPR_RASTOUT:
737 if (reg->idx[0].offset == 1)
738 reg_maps->fog = 1;
739 if (reg->idx[0].offset == 2)
740 reg_maps->point_size = 1;
741 break;
742
743 case WINED3DSPR_MISCTYPE:
744 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
745 {
746 if (!reg->idx[0].offset)
747 reg_maps->vpos = 1;
748 else if (reg->idx[0].offset == 1)
749 reg_maps->usesfacing = 1;
750 }
751 break;
752
753 case WINED3DSPR_CONST:
754 if (reg->idx[0].rel_addr)
755 {
756 if (reg->idx[0].offset < reg_maps->min_rel_offset)
757 reg_maps->min_rel_offset = reg->idx[0].offset;
758 if (reg->idx[0].offset > reg_maps->max_rel_offset)
759 reg_maps->max_rel_offset = reg->idx[0].offset;
760 reg_maps->usesrelconstF = TRUE;
761 }
762 else
763 {
764 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
765 {
766 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
767 return FALSE;
768 }
769 else
770 {
771 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
772 }
773 }
774 break;
775
776 case WINED3DSPR_CONSTINT:
777 if (reg->idx[0].offset >= shader->limits->constant_int)
778 {
779 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
780 return FALSE;
781 }
782 else
783 {
784 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
785 }
786 break;
787
788 case WINED3DSPR_CONSTBOOL:
789 if (reg->idx[0].offset >= shader->limits->constant_bool)
790 {
791 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
792 return FALSE;
793 }
794 else
795 {
796 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
797 }
798 break;
799
800 case WINED3DSPR_COLOROUT:
801 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
802 break;
803
804 case WINED3DSPR_OUTCONTROLPOINT:
805 reg_maps->vocp = 1;
806 break;
807
808 default:
809 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
810 reg->type, reg->idx[0].offset, reg->idx[1].offset);
811 break;
812 }
813 return TRUE;
814 }
815
816 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
817 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
818 {
819 struct wined3d_shader_sampler_map_entry *entries, *entry;
820 struct wined3d_shader_sampler_map *map;
821 unsigned int i;
822
823 map = &reg_maps->sampler_map;
824 entries = map->entries;
825 for (i = 0; i < map->count; ++i)
826 {
827 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
828 return;
829 }
830
831 if (!map->size)
832 {
833 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
834 {
835 ERR("Failed to allocate sampler map entries.\n");
836 return;
837 }
838 map->size = 4;
839 map->entries = entries;
840 }
841 else if (map->count == map->size)
842 {
843 size_t new_size = map->size * 2;
844
845 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
846 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
847 {
848 ERR("Failed to resize sampler map entries.\n");
849 return;
850 }
851 map->size = new_size;
852 map->entries = entries;
853 }
854
855 entry = &entries[map->count++];
856 entry->resource_idx = resource_idx;
857 entry->sampler_idx = sampler_idx;
858 entry->bind_idx = bind_idx;
859 }
860
861 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
862 {
863 switch (instr)
864 {
865 case WINED3DSIH_M4x4:
866 case WINED3DSIH_M3x4:
867 return param == 1 ? 3 : 0;
868
869 case WINED3DSIH_M4x3:
870 case WINED3DSIH_M3x3:
871 return param == 1 ? 2 : 0;
872
873 case WINED3DSIH_M3x2:
874 return param == 1 ? 1 : 0;
875
876 default:
877 return 0;
878 }
879 }
880
881 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
882 unsigned int register_idx, unsigned int size, unsigned int stride)
883 {
884 struct wined3d_shader_tgsm *tgsm;
885
886 if (register_idx >= MAX_TGSM_REGISTERS)
887 {
888 ERR("Invalid TGSM register index %u.\n", register_idx);
889 return S_OK;
890 }
891 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
892 {
893 FIXME("TGSM declarations are allowed only in compute shaders.\n");
894 return S_OK;
895 }
896
897 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
898 register_idx + 1, sizeof(*reg_maps->tgsm)))
899 return E_OUTOFMEMORY;
900
901 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
902 tgsm = &reg_maps->tgsm[register_idx];
903 tgsm->size = size;
904 tgsm->stride = stride;
905 return S_OK;
906 }
907
908 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
909 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
910 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
911 {
912 struct wined3d_shader_phase *phase;
913
914 if ((phase = *current_phase))
915 {
916 phase->end = previous_instruction_ptr;
917 *current_phase = NULL;
918 }
919
920 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
921 {
922 ERR("Unexpected shader type %#x.\n", shader->reg_maps.shader_version.type);
923 return E_FAIL;
924 }
925
926 switch (ins->handler_idx)
927 {
928 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
929 if (shader->u.hs.phases.control_point)
930 {
931 FIXME("Multiple control point phases.\n");
932 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
933 }
934 if (!(shader->u.hs.phases.control_point = HeapAlloc(GetProcessHeap(),
935 HEAP_ZERO_MEMORY, 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 /* Note that this does not count the loop register as an address register. */
965 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
966 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
967 struct wined3d_shader_signature *output_signature, DWORD constf_size)
968 {
969 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
970 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
971 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
972 struct wined3d_shader_version shader_version;
973 struct wined3d_shader_phase *phase = NULL;
974 const DWORD *ptr, *prev_ins, *current_ins;
975 void *fe_data = shader->frontend_data;
976 unsigned int i;
977 HRESULT hr;
978
979 memset(reg_maps, 0, sizeof(*reg_maps));
980 memset(input_signature_elements, 0, sizeof(input_signature_elements));
981 memset(output_signature_elements, 0, sizeof(output_signature_elements));
982 reg_maps->min_rel_offset = ~0U;
983 list_init(&reg_maps->indexable_temps);
984
985 fe->shader_read_header(fe_data, &ptr, &shader_version);
986 prev_ins = current_ins = ptr;
987 reg_maps->shader_version = shader_version;
988
989 shader_set_limits(shader);
990
991 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
992 sizeof(*reg_maps->constf))))
993 {
994 ERR("Failed to allocate constant map memory.\n");
995 return E_OUTOFMEMORY;
996 }
997
998 while (!fe->shader_is_end(fe_data, &ptr))
999 {
1000 struct wined3d_shader_instruction ins;
1001
1002 current_ins = ptr;
1003 /* Fetch opcode. */
1004 fe->shader_read_instruction(fe_data, &ptr, &ins);
1005
1006 /* Unhandled opcode, and its parameters. */
1007 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1008 {
1009 WARN("Encountered unrecognised or invalid instruction.\n");
1010 return WINED3DERR_INVALIDCALL;
1011 }
1012
1013 /* Handle declarations. */
1014 if (ins.handler_idx == WINED3DSIH_DCL
1015 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1016 {
1017 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1018 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1019
1020 switch (semantic->reg.reg.type)
1021 {
1022 /* Mark input registers used. */
1023 case WINED3DSPR_INPUT:
1024 if (reg_idx >= MAX_REG_INPUT)
1025 {
1026 ERR("Invalid input register index %u.\n", reg_idx);
1027 break;
1028 }
1029 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1030 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1031 return WINED3DERR_INVALIDCALL;
1032 reg_maps->input_registers |= 1u << reg_idx;
1033 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1034 break;
1035
1036 /* Vertex shader: mark 3.0 output registers used, save token. */
1037 case WINED3DSPR_OUTPUT:
1038 if (reg_idx >= MAX_REG_OUTPUT)
1039 {
1040 ERR("Invalid output register index %u.\n", reg_idx);
1041 break;
1042 }
1043 reg_maps->output_registers |= 1u << reg_idx;
1044 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1045 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1046 reg_maps->fog = 1;
1047 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1048 reg_maps->point_size = 1;
1049 break;
1050
1051 case WINED3DSPR_SAMPLER:
1052 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1053 case WINED3DSPR_RESOURCE:
1054 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1055 {
1056 ERR("Invalid resource index %u.\n", reg_idx);
1057 break;
1058 }
1059 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1060 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1061 break;
1062
1063 case WINED3DSPR_UAV:
1064 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1065 {
1066 ERR("Invalid UAV resource index %u.\n", reg_idx);
1067 break;
1068 }
1069 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1070 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1071 if (ins.flags)
1072 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1073 break;
1074
1075 default:
1076 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1077 break;
1078 }
1079 }
1080 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1081 {
1082 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1083 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1084 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1085 else
1086 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1087 }
1088 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1089 {
1090 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1091 {
1092 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1093 shader->u.ps.force_early_depth_stencil = TRUE;
1094 else
1095 FIXME("Invalid instruction %#x for shader type %#x.\n",
1096 ins.handler_idx, shader_version.type);
1097 }
1098 else
1099 {
1100 WARN("Ignoring global flags %#x.\n", ins.flags);
1101 }
1102 }
1103 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1104 {
1105 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1106 shader->u.gs.instance_count = ins.declaration.count;
1107 else
1108 FIXME("Invalid instruction %#x for shader type %#x.\n",
1109 ins.handler_idx, shader_version.type);
1110 }
1111 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1112 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1113 {
1114 if (phase)
1115 phase->instance_count = ins.declaration.count;
1116 else
1117 FIXME("Instruction %s outside of shader phase.\n",
1118 debug_d3dshaderinstructionhandler(ins.handler_idx));
1119 }
1120 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1121 {
1122 if (reg_maps->icb)
1123 FIXME("Multiple immediate constant buffers.\n");
1124 reg_maps->icb = ins.declaration.icb;
1125 }
1126 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1127 {
1128 if (phase)
1129 {
1130 FIXME("Indexable temporary registers not supported.\n");
1131 }
1132 else
1133 {
1134 struct wined3d_shader_indexable_temp *reg;
1135
1136 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
1137 return E_OUTOFMEMORY;
1138
1139 *reg = ins.declaration.indexable_temp;
1140 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1141 }
1142 }
1143 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1144 {
1145 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1146 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1147 else
1148 FIXME("Invalid instruction %#x for shader type %#x.\n",
1149 ins.handler_idx, shader_version.type);
1150 }
1151 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1152 {
1153 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1154 shader->u.hs.output_vertex_count = ins.declaration.count;
1155 else
1156 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1157 }
1158 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1159 {
1160 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1161 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1162 else
1163 FIXME("Invalid instruction %#x for shader type %#x.\n",
1164 ins.handler_idx, shader_version.type);
1165 }
1166 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1167 {
1168 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1169 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1170 {
1171 ERR("Invalid resource index %u.\n", reg_idx);
1172 break;
1173 }
1174 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1175 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1176 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1177 }
1178 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1179 {
1180 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1181 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1182 {
1183 ERR("Invalid resource index %u.\n", reg_idx);
1184 break;
1185 }
1186 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1187 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1188 reg_maps->resource_info[reg_idx].flags = 0;
1189 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1190 }
1191 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1192 {
1193 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1194 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1195 }
1196 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1197 {
1198 if (phase)
1199 phase->temporary_count = ins.declaration.count;
1200 else
1201 reg_maps->temporary_count = ins.declaration.count;
1202 }
1203 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1204 {
1205 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1206 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1207 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1208 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1209 }
1210 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1211 {
1212 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1213 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1214 else
1215 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1216 }
1217 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1218 {
1219 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1220 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1221 else
1222 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1223 }
1224 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1225 {
1226 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1227 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1228 return hr;
1229 }
1230 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1231 {
1232 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1233 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1234 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1235 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1236 return hr;
1237 }
1238 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1239 {
1240 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1241 {
1242 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1243 }
1244 else
1245 {
1246 FIXME("Invalid instruction %#x for shader type %#x.\n",
1247 ins.handler_idx, shader_version.type);
1248 }
1249 }
1250 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1251 {
1252 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1253 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1254 {
1255 ERR("Invalid UAV resource index %u.\n", reg_idx);
1256 break;
1257 }
1258 if (ins.flags)
1259 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1260 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1261 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1262 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1263 }
1264 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1265 {
1266 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1267 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1268 {
1269 ERR("Invalid UAV resource index %u.\n", reg_idx);
1270 break;
1271 }
1272 if (ins.flags)
1273 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1274 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1275 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1276 reg_maps->uav_resource_info[reg_idx].flags = 0;
1277 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1278 }
1279 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1280 {
1281 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1282 shader->u.gs.vertices_out = ins.declaration.count;
1283 else
1284 FIXME("Invalid instruction %#x for shader type %#x.\n",
1285 ins.handler_idx, shader_version.type);
1286 }
1287 else if (ins.handler_idx == WINED3DSIH_DEF)
1288 {
1289 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1290 float *value;
1291 if (!lconst) return E_OUTOFMEMORY;
1292
1293 lconst->idx = ins.dst[0].reg.idx[0].offset;
1294 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1295 value = (float *)lconst->value;
1296
1297 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1298 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1299 {
1300 if (value[0] < -1.0f) value[0] = -1.0f;
1301 else if (value[0] > 1.0f) value[0] = 1.0f;
1302 if (value[1] < -1.0f) value[1] = -1.0f;
1303 else if (value[1] > 1.0f) value[1] = 1.0f;
1304 if (value[2] < -1.0f) value[2] = -1.0f;
1305 else if (value[2] > 1.0f) value[2] = 1.0f;
1306 if (value[3] < -1.0f) value[3] = -1.0f;
1307 else if (value[3] > 1.0f) value[3] = 1.0f;
1308 }
1309
1310 list_add_head(&shader->constantsF, &lconst->entry);
1311
1312 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1313 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1314 {
1315 shader->lconst_inf_or_nan = TRUE;
1316 }
1317 }
1318 else if (ins.handler_idx == WINED3DSIH_DEFI)
1319 {
1320 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1321 if (!lconst) return E_OUTOFMEMORY;
1322
1323 lconst->idx = ins.dst[0].reg.idx[0].offset;
1324 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1325
1326 list_add_head(&shader->constantsI, &lconst->entry);
1327 reg_maps->local_int_consts |= (1u << lconst->idx);
1328 }
1329 else if (ins.handler_idx == WINED3DSIH_DEFB)
1330 {
1331 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1332 if (!lconst) return E_OUTOFMEMORY;
1333
1334 lconst->idx = ins.dst[0].reg.idx[0].offset;
1335 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1336
1337 list_add_head(&shader->constantsB, &lconst->entry);
1338 reg_maps->local_bool_consts |= (1u << lconst->idx);
1339 }
1340 /* Handle shader phases. */
1341 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1342 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1343 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1344 {
1345 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1346 return hr;
1347 }
1348 /* For subroutine prototypes. */
1349 else if (ins.handler_idx == WINED3DSIH_LABEL)
1350 {
1351 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1352 }
1353 /* Set texture, address, temporary registers. */
1354 else
1355 {
1356 BOOL color0_mov = FALSE;
1357 unsigned int i;
1358
1359 /* This will loop over all the registers and try to
1360 * make a bitmask of the ones we're interested in.
1361 *
1362 * Relative addressing tokens are ignored, but that's
1363 * okay, since we'll catch any address registers when
1364 * they are initialized (required by spec). */
1365 for (i = 0; i < ins.dst_count; ++i)
1366 {
1367 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1368 shader_version.type, constf_size))
1369 return WINED3DERR_INVALIDCALL;
1370
1371 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1372 {
1373 UINT idx = ins.dst[i].reg.idx[0].offset;
1374
1375 switch (ins.dst[i].reg.type)
1376 {
1377 case WINED3DSPR_RASTOUT:
1378 if (shader_version.major >= 3)
1379 break;
1380 switch (idx)
1381 {
1382 case 0: /* oPos */
1383 reg_maps->output_registers |= 1u << 10;
1384 shader_signature_from_usage(&output_signature_elements[10],
1385 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1386 break;
1387
1388 case 1: /* oFog */
1389 reg_maps->output_registers |= 1u << 11;
1390 shader_signature_from_usage(&output_signature_elements[11],
1391 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1392 break;
1393
1394 case 2: /* oPts */
1395 reg_maps->output_registers |= 1u << 11;
1396 shader_signature_from_usage(&output_signature_elements[11],
1397 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1398 break;
1399 }
1400 break;
1401
1402 case WINED3DSPR_ATTROUT:
1403 if (shader_version.major >= 3)
1404 break;
1405 if (idx < 2)
1406 {
1407 idx += 8;
1408 if (reg_maps->output_registers & (1u << idx))
1409 {
1410 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1411 }
1412 else
1413 {
1414 reg_maps->output_registers |= 1u << idx;
1415 shader_signature_from_usage(&output_signature_elements[idx],
1416 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1417 }
1418 }
1419 break;
1420
1421 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1422 if (shader_version.major >= 3)
1423 {
1424 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1425 {
1426 WARN("Invalid output register index %u.\n", idx);
1427 break;
1428 }
1429 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1430 break;
1431 }
1432 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1433 {
1434 WARN("Invalid texcoord index %u.\n", idx);
1435 break;
1436 }
1437 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1438 if (reg_maps->output_registers & (1u << idx))
1439 {
1440 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1441 }
1442 else
1443 {
1444 reg_maps->output_registers |= 1u << idx;
1445 shader_signature_from_usage(&output_signature_elements[idx],
1446 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1447 }
1448 break;
1449
1450 default:
1451 break;
1452 }
1453 }
1454
1455 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1456 {
1457 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1458 {
1459 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1460 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1461 * the mov and perform the sRGB write correction from the source register.
1462 *
1463 * However, if the mov is only partial, we can't do this, and if the write
1464 * comes from an instruction other than MOV it is hard to do as well. If
1465 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1466 shader->u.ps.color0_mov = FALSE;
1467 if (ins.handler_idx == WINED3DSIH_MOV
1468 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1469 {
1470 /* Used later when the source register is read. */
1471 color0_mov = TRUE;
1472 }
1473 }
1474 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1475 * end
1476 */
1477 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1478 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1479 {
1480 shader->u.ps.color0_mov = FALSE;
1481 }
1482 }
1483
1484 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1485 if (shader_version.major == 1
1486 && (ins.handler_idx == WINED3DSIH_TEX
1487 || ins.handler_idx == WINED3DSIH_TEXBEM
1488 || ins.handler_idx == WINED3DSIH_TEXBEML
1489 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1490 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1491 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1492 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1493 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1494 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1495 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1496 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1497 {
1498 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1499
1500 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1501 {
1502 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1503 continue;
1504 }
1505
1506 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1507 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1508 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1509 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1510
1511 /* texbem is only valid with < 1.4 pixel shaders */
1512 if (ins.handler_idx == WINED3DSIH_TEXBEM
1513 || ins.handler_idx == WINED3DSIH_TEXBEML)
1514 {
1515 reg_maps->bumpmat |= 1u << reg_idx;
1516 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1517 {
1518 reg_maps->luminanceparams |= 1u << reg_idx;
1519 }
1520 }
1521 }
1522 else if (ins.handler_idx == WINED3DSIH_BEM)
1523 {
1524 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1525 }
1526 }
1527
1528 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1529 {
1530 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1531 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1532 {
1533 ERR("Invalid UAV index %u.\n", reg_idx);
1534 break;
1535 }
1536 reg_maps->uav_counter_mask |= (1u << reg_idx);
1537 }
1538 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1539 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1540 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1541 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1542 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1543 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1544 {
1545 unsigned int reg_idx;
1546 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1547 reg_idx = ins.src[1].reg.idx[0].offset;
1548 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1549 reg_idx = ins.src[2].reg.idx[0].offset;
1550 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1551 reg_idx = ins.dst[0].reg.idx[0].offset;
1552 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1553 reg_idx = ins.src[0].reg.idx[0].offset;
1554 else
1555 reg_idx = ins.dst[1].reg.idx[0].offset;
1556 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1557 {
1558 ERR("Invalid UAV index %u.\n", reg_idx);
1559 break;
1560 }
1561 reg_maps->uav_read_mask |= (1u << reg_idx);
1562 }
1563 else if (ins.handler_idx == WINED3DSIH_NRM)
1564 {
1565 reg_maps->usesnrm = 1;
1566 }
1567 else if (ins.handler_idx == WINED3DSIH_DSY
1568 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1569 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1570 {
1571 reg_maps->usesdsy = 1;
1572 }
1573 else if (ins.handler_idx == WINED3DSIH_DSX
1574 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1575 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1576 {
1577 reg_maps->usesdsx = 1;
1578 }
1579 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1580 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1581 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1582 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1583 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1584 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1585 else if (ins.handler_idx == WINED3DSIH_LOOP
1586 || ins.handler_idx == WINED3DSIH_REP)
1587 {
1588 ++cur_loop_depth;
1589 if (cur_loop_depth > max_loop_depth)
1590 max_loop_depth = cur_loop_depth;
1591 }
1592 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1593 || ins.handler_idx == WINED3DSIH_ENDREP)
1594 {
1595 --cur_loop_depth;
1596 }
1597 else if (ins.handler_idx == WINED3DSIH_GATHER4
1598 || ins.handler_idx == WINED3DSIH_GATHER4_C
1599 || ins.handler_idx == WINED3DSIH_SAMPLE
1600 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1601 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1602 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1603 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1604 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1605 {
1606 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1607 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1608 }
1609 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1610 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1611 {
1612 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1613 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1614 }
1615 else if (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1616 {
1617 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1618 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1619 }
1620 else if (ins.handler_idx == WINED3DSIH_LD
1621 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1622 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1623 {
1624 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1625 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1626 }
1627 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1628 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1629 {
1630 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1631 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1632 }
1633
1634 if (ins.predicate)
1635 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1636 shader_version.type, constf_size))
1637 return WINED3DERR_INVALIDCALL;
1638
1639 for (i = 0; i < ins.src_count; ++i)
1640 {
1641 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1642 struct wined3d_shader_register reg = ins.src[i].reg;
1643
1644 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1645 shader_version.type, constf_size))
1646 return WINED3DERR_INVALIDCALL;
1647 while (count)
1648 {
1649 ++reg.idx[0].offset;
1650 if (!shader_record_register_usage(shader, reg_maps, &reg,
1651 shader_version.type, constf_size))
1652 return WINED3DERR_INVALIDCALL;
1653 --count;
1654 }
1655
1656 if (color0_mov)
1657 {
1658 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1659 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1660 {
1661 shader->u.ps.color0_mov = TRUE;
1662 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1663 }
1664 }
1665 }
1666 }
1667
1668 prev_ins = current_ins;
1669 }
1670 reg_maps->loop_depth = max_loop_depth;
1671
1672 if (phase)
1673 {
1674 phase->end = prev_ins;
1675 phase = NULL;
1676 }
1677
1678 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1679 * R0 is written to the render target. */
1680 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1681 reg_maps->rt_mask |= (1u << 0);
1682
1683 if (input_signature->elements)
1684 {
1685 for (i = 0; i < input_signature->element_count; ++i)
1686 {
1687 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1688 {
1689 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1690 {
1691 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1692 return WINED3DERR_INVALIDCALL;
1693 }
1694 }
1695 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1696 {
1697 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1698 reg_maps->vpos = 1;
1699 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1700 reg_maps->usesfacing = 1;
1701 }
1702 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1703 }
1704 }
1705 else if (!input_signature->elements && reg_maps->input_registers)
1706 {
1707 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1708 struct wined3d_shader_signature_element *e;
1709 unsigned int i;
1710
1711 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1712 return E_OUTOFMEMORY;
1713 input_signature->element_count = count;
1714
1715 e = input_signature->elements;
1716 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1717 {
1718 if (!(reg_maps->input_registers & (1u << i)))
1719 continue;
1720 input_signature_elements[i].register_idx = i;
1721 *e++ = input_signature_elements[i];
1722 }
1723 }
1724
1725 if (output_signature->elements)
1726 {
1727 for (i = 0; i < output_signature->element_count; ++i)
1728 {
1729 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1730 }
1731 }
1732 else if (reg_maps->output_registers)
1733 {
1734 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1735 struct wined3d_shader_signature_element *e;
1736
1737 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1738 return E_OUTOFMEMORY;
1739 output_signature->element_count = count;
1740
1741 e = output_signature->elements;
1742 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1743 {
1744 if (!(reg_maps->output_registers & (1u << i)))
1745 continue;
1746 *e++ = output_signature_elements[i];
1747 }
1748 }
1749
1750 return WINED3D_OK;
1751 }
1752
1753 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1754 {
1755 struct wined3d_shader_indexable_temp *reg, *reg_next;
1756
1757 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1758 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1759
1760 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1761 HeapFree(GetProcessHeap(), 0, reg);
1762 list_init(&reg_maps->indexable_temps);
1763
1764 HeapFree(GetProcessHeap(), 0, reg_maps->tgsm);
1765 }
1766
1767 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1768 {
1769 DWORD map = 1u << max;
1770 map |= map - 1;
1771 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1772
1773 return wined3d_log2i(map);
1774 }
1775
1776 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1777 {
1778 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1779 {
1780 shader_addline(buffer, "refactoringAllowed");
1781 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1782 if (global_flags)
1783 shader_addline(buffer, " | ");
1784 }
1785
1786 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1787 {
1788 shader_addline(buffer, "forceEarlyDepthStencil");
1789 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1790 if (global_flags)
1791 shader_addline(buffer, " | ");
1792 }
1793
1794 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1795 {
1796 shader_addline(buffer, "enableRawAndStructuredBuffers");
1797 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1798 }
1799
1800 if (global_flags)
1801 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1802 }
1803
1804 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1805 {
1806 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1807 {
1808 shader_addline(buffer, "_g");
1809 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1810 }
1811 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1812 {
1813 shader_addline(buffer, "_t");
1814 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1815 }
1816
1817 if (sync_flags)
1818 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1819 }
1820
1821 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1822 {
1823 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1824 {
1825 shader_addline(buffer, "_glc");
1826 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1827 }
1828 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1829 {
1830 shader_addline(buffer, "_opc");
1831 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1832 }
1833
1834 if (uav_flags)
1835 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1836 }
1837
1838 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1839 enum wined3d_tessellator_domain domain)
1840 {
1841 switch (domain)
1842 {
1843 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1844 shader_addline(buffer, "line");
1845 break;
1846 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1847 shader_addline(buffer, "triangle");
1848 break;
1849 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1850 shader_addline(buffer, "quad");
1851 break;
1852 default:
1853 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1854 break;
1855 }
1856 }
1857
1858 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1859 enum wined3d_tessellator_output_primitive output_primitive)
1860 {
1861 switch (output_primitive)
1862 {
1863 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1864 shader_addline(buffer, "point");
1865 break;
1866 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1867 shader_addline(buffer, "line");
1868 break;
1869 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1870 shader_addline(buffer, "triangle_cw");
1871 break;
1872 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1873 shader_addline(buffer, "triangle_ccw");
1874 break;
1875 default:
1876 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1877 break;
1878 }
1879 }
1880
1881 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1882 enum wined3d_tessellator_partitioning partitioning)
1883 {
1884 switch (partitioning)
1885 {
1886 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1887 shader_addline(buffer, "integer");
1888 break;
1889 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1890 shader_addline(buffer, "pow2");
1891 break;
1892 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1893 shader_addline(buffer, "fractional_odd");
1894 break;
1895 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1896 shader_addline(buffer, "fractional_even");
1897 break;
1898 default:
1899 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1900 break;
1901 }
1902 }
1903
1904 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
1905 enum wined3d_shader_input_sysval_semantic semantic)
1906 {
1907 unsigned int i;
1908
1909 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
1910 {
1911 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
1912 {
1913 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
1914 return;
1915 }
1916 }
1917
1918 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
1919 }
1920
1921 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1922 const struct wined3d_shader_semantic *semantic, unsigned int flags,
1923 const struct wined3d_shader_version *shader_version)
1924 {
1925 shader_addline(buffer, "dcl");
1926
1927 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1928 {
1929 switch (semantic->resource_type)
1930 {
1931 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1932 shader_addline(buffer, "_2d");
1933 break;
1934
1935 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1936 shader_addline(buffer, "_3d");
1937 break;
1938
1939 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1940 shader_addline(buffer, "_cube");
1941 break;
1942
1943 default:
1944 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1945 break;
1946 }
1947 }
1948 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1949 {
1950 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1951 shader_addline(buffer, "_resource_");
1952 else
1953 shader_addline(buffer, "_uav_");
1954 switch (semantic->resource_type)
1955 {
1956 case WINED3D_SHADER_RESOURCE_BUFFER:
1957 shader_addline(buffer, "buffer");
1958 break;
1959
1960 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1961 shader_addline(buffer, "texture1d");
1962 break;
1963
1964 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1965 shader_addline(buffer, "texture2d");
1966 break;
1967
1968 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1969 shader_addline(buffer, "texture2dms");
1970 break;
1971
1972 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1973 shader_addline(buffer, "texture3d");
1974 break;
1975
1976 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1977 shader_addline(buffer, "texturecube");
1978 break;
1979
1980 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1981 shader_addline(buffer, "texture1darray");
1982 break;
1983
1984 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1985 shader_addline(buffer, "texture2darray");
1986 break;
1987
1988 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1989 shader_addline(buffer, "texture2dmsarray");
1990 break;
1991
1992 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
1993 shader_addline(buffer, "texturecubearray");
1994 break;
1995
1996 default:
1997 shader_addline(buffer, "unknown");
1998 break;
1999 }
2000 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2001 shader_dump_uav_flags(buffer, flags);
2002 switch (semantic->resource_data_type)
2003 {
2004 case WINED3D_DATA_FLOAT:
2005 shader_addline(buffer, " (float)");
2006 break;
2007
2008 case WINED3D_DATA_INT:
2009 shader_addline(buffer, " (int)");
2010 break;
2011
2012 case WINED3D_DATA_UINT:
2013 shader_addline(buffer, " (uint)");
2014 break;
2015
2016 case WINED3D_DATA_UNORM:
2017 shader_addline(buffer, " (unorm)");
2018 break;
2019
2020 case WINED3D_DATA_SNORM:
2021 shader_addline(buffer, " (snorm)");
2022 break;
2023
2024 default:
2025 shader_addline(buffer, " (unknown)");
2026 break;
2027 }
2028 }
2029 else
2030 {
2031 /* Pixel shaders 3.0 don't have usage semantics. */
2032 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2033 return;
2034 else
2035 shader_addline(buffer, "_");
2036
2037 switch (semantic->usage)
2038 {
2039 case WINED3D_DECL_USAGE_POSITION:
2040 shader_addline(buffer, "position%u", semantic->usage_idx);
2041 break;
2042
2043 case WINED3D_DECL_USAGE_BLEND_INDICES:
2044 shader_addline(buffer, "blend");
2045 break;
2046
2047 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2048 shader_addline(buffer, "weight");
2049 break;
2050
2051 case WINED3D_DECL_USAGE_NORMAL:
2052 shader_addline(buffer, "normal%u", semantic->usage_idx);
2053 break;
2054
2055 case WINED3D_DECL_USAGE_PSIZE:
2056 shader_addline(buffer, "psize");
2057 break;
2058
2059 case WINED3D_DECL_USAGE_COLOR:
2060 if (!semantic->usage_idx)
2061 shader_addline(buffer, "color");
2062 else
2063 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2064 break;
2065
2066 case WINED3D_DECL_USAGE_TEXCOORD:
2067 shader_addline(buffer, "texture%u", semantic->usage_idx);
2068 break;
2069
2070 case WINED3D_DECL_USAGE_TANGENT:
2071 shader_addline(buffer, "tangent");
2072 break;
2073
2074 case WINED3D_DECL_USAGE_BINORMAL:
2075 shader_addline(buffer, "binormal");
2076 break;
2077
2078 case WINED3D_DECL_USAGE_TESS_FACTOR:
2079 shader_addline(buffer, "tessfactor");
2080 break;
2081
2082 case WINED3D_DECL_USAGE_POSITIONT:
2083 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2084 break;
2085
2086 case WINED3D_DECL_USAGE_FOG:
2087 shader_addline(buffer, "fog");
2088 break;
2089
2090 case WINED3D_DECL_USAGE_DEPTH:
2091 shader_addline(buffer, "depth");
2092 break;
2093
2094 case WINED3D_DECL_USAGE_SAMPLE:
2095 shader_addline(buffer, "sample");
2096 break;
2097
2098 default:
2099 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2100 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2101 }
2102 }
2103 }
2104
2105 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2106 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2107 {
2108 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2109 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2110 UINT offset = reg->idx[0].offset;
2111
2112 switch (reg->type)
2113 {
2114 case WINED3DSPR_TEMP:
2115 shader_addline(buffer, "r");
2116 break;
2117
2118 case WINED3DSPR_INPUT:
2119 shader_addline(buffer, "v");
2120 break;
2121
2122 case WINED3DSPR_CONST:
2123 case WINED3DSPR_CONST2:
2124 case WINED3DSPR_CONST3:
2125 case WINED3DSPR_CONST4:
2126 shader_addline(buffer, "c");
2127 offset = shader_get_float_offset(reg->type, offset);
2128 break;
2129
2130 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2131 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2132 break;
2133
2134 case WINED3DSPR_RASTOUT:
2135 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2136 break;
2137
2138 case WINED3DSPR_COLOROUT:
2139 shader_addline(buffer, "oC");
2140 break;
2141
2142 case WINED3DSPR_DEPTHOUT:
2143 shader_addline(buffer, "oDepth");
2144 break;
2145
2146 case WINED3DSPR_ATTROUT:
2147 shader_addline(buffer, "oD");
2148 break;
2149
2150 case WINED3DSPR_TEXCRDOUT:
2151 /* Vertex shaders >= 3.0 use general purpose output registers
2152 * (WINED3DSPR_OUTPUT), which can include an address token. */
2153 if (shader_version->major >= 3)
2154 shader_addline(buffer, "o");
2155 else
2156 shader_addline(buffer, "oT");
2157 break;
2158
2159 case WINED3DSPR_CONSTINT:
2160 shader_addline(buffer, "i");
2161 break;
2162
2163 case WINED3DSPR_CONSTBOOL:
2164 shader_addline(buffer, "b");
2165 break;
2166
2167 case WINED3DSPR_LABEL:
2168 shader_addline(buffer, "l");
2169 break;
2170
2171 case WINED3DSPR_LOOP:
2172 shader_addline(buffer, "aL");
2173 break;
2174
2175 case WINED3DSPR_SAMPLER:
2176 shader_addline(buffer, "s");
2177 break;
2178
2179 case WINED3DSPR_MISCTYPE:
2180 if (offset > 1)
2181 {
2182 FIXME("Unhandled misctype register %u.\n", offset);
2183 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2184 }
2185 else
2186 {
2187 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2188 }
2189 break;
2190
2191 case WINED3DSPR_PREDICATE:
2192 shader_addline(buffer, "p");
2193 break;
2194
2195 case WINED3DSPR_IMMCONST:
2196 shader_addline(buffer, "l");
2197 break;
2198
2199 case WINED3DSPR_CONSTBUFFER:
2200 shader_addline(buffer, "cb");
2201 break;
2202
2203 case WINED3DSPR_IMMCONSTBUFFER:
2204 shader_addline(buffer, "icb");
2205 break;
2206
2207 case WINED3DSPR_PRIMID:
2208 shader_addline(buffer, "primID");
2209 break;
2210
2211 case WINED3DSPR_NULL:
2212 shader_addline(buffer, "null");
2213 break;
2214
2215 case WINED3DSPR_RESOURCE:
2216 shader_addline(buffer, "t");
2217 break;
2218
2219 case WINED3DSPR_UAV:
2220 shader_addline(buffer, "u");
2221 break;
2222
2223 case WINED3DSPR_OUTPOINTID:
2224 shader_addline(buffer, "vOutputControlPointID");
2225 break;
2226
2227 case WINED3DSPR_FORKINSTID:
2228 shader_addline(buffer, "vForkInstanceId");
2229 break;
2230
2231 case WINED3DSPR_JOININSTID:
2232 shader_addline(buffer, "vJoinInstanceId");
2233 break;
2234
2235 case WINED3DSPR_INCONTROLPOINT:
2236 shader_addline(buffer, "vicp");
2237 break;
2238
2239 case WINED3DSPR_OUTCONTROLPOINT:
2240 shader_addline(buffer, "vocp");
2241 break;
2242
2243 case WINED3DSPR_PATCHCONST:
2244 shader_addline(buffer, "vpc");
2245 break;
2246
2247 case WINED3DSPR_TESSCOORD:
2248 shader_addline(buffer, "vDomainLocation");
2249 break;
2250
2251 case WINED3DSPR_GROUPSHAREDMEM:
2252 shader_addline(buffer, "g");
2253 break;
2254
2255 case WINED3DSPR_THREADID:
2256 shader_addline(buffer, "vThreadID");
2257 break;
2258
2259 case WINED3DSPR_THREADGROUPID:
2260 shader_addline(buffer, "vThreadGroupID");
2261 break;
2262
2263 case WINED3DSPR_LOCALTHREADID:
2264 shader_addline(buffer, "vThreadIDInGroup");
2265 break;
2266
2267 case WINED3DSPR_LOCALTHREADINDEX:
2268 shader_addline(buffer, "vThreadIDInGroupFlattened");
2269 break;
2270
2271 case WINED3DSPR_IDXTEMP:
2272 shader_addline(buffer, "x");
2273 break;
2274
2275 case WINED3DSPR_STREAM:
2276 shader_addline(buffer, "m");
2277 break;
2278
2279 case WINED3DSPR_FUNCTIONBODY:
2280 shader_addline(buffer, "fb");
2281 break;
2282
2283 case WINED3DSPR_FUNCTIONPOINTER:
2284 shader_addline(buffer, "fp");
2285 break;
2286
2287 case WINED3DSPR_COVERAGE:
2288 shader_addline(buffer, "vCoverage");
2289 break;
2290
2291 case WINED3DSPR_SAMPLEMASK:
2292 shader_addline(buffer, "oMask");
2293 break;
2294
2295 case WINED3DSPR_GSINSTID:
2296 shader_addline(buffer, "vGSInstanceID");
2297 break;
2298
2299 default:
2300 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2301 break;
2302 }
2303
2304 if (reg->type == WINED3DSPR_IMMCONST)
2305 {
2306 shader_addline(buffer, "(");
2307 switch (reg->immconst_type)
2308 {
2309 case WINED3D_IMMCONST_SCALAR:
2310 switch (reg->data_type)
2311 {
2312 case WINED3D_DATA_FLOAT:
2313 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2314 break;
2315 case WINED3D_DATA_INT:
2316 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2317 break;
2318 case WINED3D_DATA_RESOURCE:
2319 case WINED3D_DATA_SAMPLER:
2320 case WINED3D_DATA_UINT:
2321 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2322 break;
2323 default:
2324 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2325 break;
2326 }
2327 break;
2328
2329 case WINED3D_IMMCONST_VEC4:
2330 switch (reg->data_type)
2331 {
2332 case WINED3D_DATA_FLOAT:
2333 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2334 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2335 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2336 break;
2337 case WINED3D_DATA_INT:
2338 shader_addline(buffer, "%d, %d, %d, %d",
2339 reg->u.immconst_data[0], reg->u.immconst_data[1],
2340 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2341 break;
2342 case WINED3D_DATA_RESOURCE:
2343 case WINED3D_DATA_SAMPLER:
2344 case WINED3D_DATA_UINT:
2345 shader_addline(buffer, "%u, %u, %u, %u",
2346 reg->u.immconst_data[0], reg->u.immconst_data[1],
2347 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2348 break;
2349 default:
2350 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2351 break;
2352 }
2353 break;
2354
2355 default:
2356 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2357 break;
2358 }
2359 shader_addline(buffer, ")");
2360 }
2361 else if (reg->type != WINED3DSPR_RASTOUT
2362 && reg->type != WINED3DSPR_MISCTYPE
2363 && reg->type != WINED3DSPR_NULL)
2364 {
2365 if (offset != ~0u)
2366 {
2367 shader_addline(buffer, "[");
2368 if (reg->idx[0].rel_addr)
2369 {
2370 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2371 shader_addline(buffer, " + ");
2372 }
2373 shader_addline(buffer, "%u]", offset);
2374
2375 if (reg->idx[1].offset != ~0u)
2376 {
2377 shader_addline(buffer, "[");
2378 if (reg->idx[1].rel_addr)
2379 {
2380 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2381 shader_addline(buffer, " + ");
2382 }
2383 shader_addline(buffer, "%u]", reg->idx[1].offset);
2384 }
2385 }
2386
2387 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2388 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2389 }
2390 }
2391
2392 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2393 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2394 {
2395 DWORD write_mask = param->write_mask;
2396
2397 shader_dump_register(buffer, &param->reg, shader_version);
2398
2399 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2400 {
2401 static const char write_mask_chars[] = "xyzw";
2402
2403 shader_addline(buffer, ".");
2404 if (write_mask & WINED3DSP_WRITEMASK_0)
2405 shader_addline(buffer, "%c", write_mask_chars[0]);
2406 if (write_mask & WINED3DSP_WRITEMASK_1)
2407 shader_addline(buffer, "%c", write_mask_chars[1]);
2408 if (write_mask & WINED3DSP_WRITEMASK_2)
2409 shader_addline(buffer, "%c", write_mask_chars[2]);
2410 if (write_mask & WINED3DSP_WRITEMASK_3)
2411 shader_addline(buffer, "%c", write_mask_chars[3]);
2412 }
2413 }
2414
2415 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2416 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2417 {
2418 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2419 DWORD swizzle = param->swizzle;
2420
2421 if (src_modifier == WINED3DSPSM_NEG
2422 || src_modifier == WINED3DSPSM_BIASNEG
2423 || src_modifier == WINED3DSPSM_SIGNNEG
2424 || src_modifier == WINED3DSPSM_X2NEG
2425 || src_modifier == WINED3DSPSM_ABSNEG)
2426 shader_addline(buffer, "-");
2427 else if (src_modifier == WINED3DSPSM_COMP)
2428 shader_addline(buffer, "1-");
2429 else if (src_modifier == WINED3DSPSM_NOT)
2430 shader_addline(buffer, "!");
2431
2432 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2433 shader_addline(buffer, "abs(");
2434
2435 shader_dump_register(buffer, &param->reg, shader_version);
2436
2437 switch (src_modifier)
2438 {
2439 case WINED3DSPSM_NONE: break;
2440 case WINED3DSPSM_NEG: break;
2441 case WINED3DSPSM_NOT: break;
2442 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2443 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2444 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2445 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2446 case WINED3DSPSM_COMP: break;
2447 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2448 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2449 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2450 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2451 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2452 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2453 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2454 }
2455
2456 if (swizzle != WINED3DSP_NOSWIZZLE)
2457 {
2458 static const char swizzle_chars[] = "xyzw";
2459 DWORD swizzle_x = swizzle & 0x03;
2460 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2461 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2462 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2463
2464 if (swizzle_x == swizzle_y
2465 && swizzle_x == swizzle_z
2466 && swizzle_x == swizzle_w)
2467 {
2468 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2469 }
2470 else
2471 {
2472 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2473 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2474 }
2475 }
2476 }
2477
2478 /* Shared code in order to generate the bulk of the shader string. */
2479 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2480 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2481 const DWORD *start, const DWORD *end)
2482 {
2483 struct wined3d_device *device = shader->device;
2484 const struct wined3d_shader_frontend *fe = shader->frontend;
2485 void *fe_data = shader->frontend_data;
2486 struct wined3d_shader_version shader_version;
2487 struct wined3d_shader_parser_state state;
2488 struct wined3d_shader_instruction ins;
2489 struct wined3d_shader_tex_mx tex_mx;
2490 struct wined3d_shader_context ctx;
2491 const DWORD *ptr;
2492
2493 /* Initialize current parsing state. */
2494 tex_mx.current_row = 0;
2495 state.current_loop_depth = 0;
2496 state.current_loop_reg = 0;
2497 state.in_subroutine = FALSE;
2498
2499 ctx.shader = shader;
2500 ctx.gl_info = &device->adapter->gl_info;
2501 ctx.reg_maps = reg_maps;
2502 ctx.buffer = buffer;
2503 ctx.tex_mx = &tex_mx;
2504 ctx.state = &state;
2505 ctx.backend_data = backend_ctx;
2506 ins.ctx = &ctx;
2507
2508 fe->shader_read_header(fe_data, &ptr, &shader_version);
2509 if (start)
2510 ptr = start;
2511
2512 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2513 {
2514 /* Read opcode. */
2515 fe->shader_read_instruction(fe_data, &ptr, &ins);
2516
2517 /* Unknown opcode and its parameters. */
2518 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2519 {
2520 WARN("Encountered unrecognised or invalid instruction.\n");
2521 return WINED3DERR_INVALIDCALL;
2522 }
2523
2524 if (ins.predicate)
2525 FIXME("Predicates not implemented.\n");
2526
2527 /* Call appropriate function for output target */
2528 device->shader_backend->shader_handle_instruction(&ins);
2529 }
2530
2531 return WINED3D_OK;
2532 }
2533
2534 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2535 const struct wined3d_shader_dst_param *dst)
2536 {
2537 DWORD mmask = dst->modifiers;
2538
2539 switch (dst->shift)
2540 {
2541 case 0: break;
2542 case 13: shader_addline(buffer, "_d8"); break;
2543 case 14: shader_addline(buffer, "_d4"); break;
2544 case 15: shader_addline(buffer, "_d2"); break;
2545 case 1: shader_addline(buffer, "_x2"); break;
2546 case 2: shader_addline(buffer, "_x4"); break;
2547 case 3: shader_addline(buffer, "_x8"); break;
2548 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2549 }
2550
2551 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2552 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2553 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2554
2555 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2556 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2557 }
2558
2559 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2560 const struct wined3d_shader_primitive_type *primitive_type)
2561 {
2562 switch (primitive_type->type)
2563 {
2564 case WINED3D_PT_UNDEFINED:
2565 shader_addline(buffer, "undefined");
2566 break;
2567 case WINED3D_PT_POINTLIST:
2568 shader_addline(buffer, "pointlist");
2569 break;
2570 case WINED3D_PT_LINELIST:
2571 shader_addline(buffer, "linelist");
2572 break;
2573 case WINED3D_PT_LINESTRIP:
2574 shader_addline(buffer, "linestrip");
2575 break;
2576 case WINED3D_PT_TRIANGLELIST:
2577 shader_addline(buffer, "trianglelist");
2578 break;
2579 case WINED3D_PT_TRIANGLESTRIP:
2580 shader_addline(buffer, "trianglestrip");
2581 break;
2582 case WINED3D_PT_TRIANGLEFAN:
2583 shader_addline(buffer, "trianglefan");
2584 break;
2585 case WINED3D_PT_LINELIST_ADJ:
2586 shader_addline(buffer, "linelist_adj");
2587 break;
2588 case WINED3D_PT_LINESTRIP_ADJ:
2589 shader_addline(buffer, "linestrip_adj");
2590 break;
2591 case WINED3D_PT_TRIANGLELIST_ADJ:
2592 shader_addline(buffer, "trianglelist_adj");
2593 break;
2594 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2595 shader_addline(buffer, "trianglestrip_adj");
2596 break;
2597 case WINED3D_PT_PATCH:
2598 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2599 break;
2600 default:
2601 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2602 break;
2603 }
2604 }
2605
2606 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2607 enum wined3d_shader_interpolation_mode interpolation_mode)
2608 {
2609 switch (interpolation_mode)
2610 {
2611 case WINED3DSIM_CONSTANT:
2612 shader_addline(buffer, "constant");
2613 break;
2614 case WINED3DSIM_LINEAR:
2615 shader_addline(buffer, "linear");
2616 break;
2617 case WINED3DSIM_LINEAR_CENTROID:
2618 shader_addline(buffer, "linear centroid");
2619 break;
2620 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2621 shader_addline(buffer, "linear noperspective");
2622 break;
2623 case WINED3DSIM_LINEAR_SAMPLE:
2624 shader_addline(buffer, "linear sample");
2625 break;
2626 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2627 shader_addline(buffer, "linear noperspective centroid");
2628 break;
2629 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2630 shader_addline(buffer, "linear noperspective sample");
2631 break;
2632 default:
2633 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2634 break;
2635 }
2636 }
2637
2638 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2639 {
2640 struct wined3d_shader_version shader_version;
2641 struct wined3d_string_buffer buffer;
2642 const char *type_prefix;
2643 const char *p, *q;
2644 const DWORD *ptr;
2645 DWORD i;
2646
2647 if (!string_buffer_init(&buffer))
2648 {
2649 ERR("Failed to initialize string buffer.\n");
2650 return;
2651 }
2652
2653 fe->shader_read_header(fe_data, &ptr, &shader_version);
2654
2655 TRACE("Parsing %p.\n", ptr);
2656
2657 switch (shader_version.type)
2658 {
2659 case WINED3D_SHADER_TYPE_VERTEX:
2660 type_prefix = "vs";
2661 break;
2662
2663 case WINED3D_SHADER_TYPE_HULL:
2664 type_prefix = "hs";
2665 break;
2666
2667 case WINED3D_SHADER_TYPE_DOMAIN:
2668 type_prefix = "ds";
2669 break;
2670
2671 case WINED3D_SHADER_TYPE_GEOMETRY:
2672 type_prefix = "gs";
2673 break;
2674
2675 case WINED3D_SHADER_TYPE_PIXEL:
2676 type_prefix = "ps";
2677 break;
2678
2679 case WINED3D_SHADER_TYPE_COMPUTE:
2680 type_prefix = "cs";
2681 break;
2682
2683 default:
2684 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2685 type_prefix = "unknown";
2686 break;
2687 }
2688
2689 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2690
2691 while (!fe->shader_is_end(fe_data, &ptr))
2692 {
2693 struct wined3d_shader_instruction ins;
2694
2695 fe->shader_read_instruction(fe_data, &ptr, &ins);
2696 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2697 {
2698 WARN("Skipping unrecognized instruction.\n");
2699 shader_addline(&buffer, "<unrecognized instruction>\n");
2700 continue;
2701 }
2702
2703 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2704 {
2705 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2706 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2707 shader_addline(&buffer, " ");
2708 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2709 }
2710 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2711 {
2712 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2713 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2714 shader_addline(&buffer, ", %s",
2715 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2716 }
2717 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2718 {
2719 shader_addline(&buffer, "%s fb%u",
2720 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2721 }
2722 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2723 {
2724 shader_addline(&buffer, "%s ft%u = {...}",
2725 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2726 }
2727 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2728 {
2729 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2730 shader_dump_global_flags(&buffer, ins.flags);
2731 }
2732 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2733 {
2734 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2735 ins.declaration.max_tessellation_factor);
2736 }
2737 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2738 {
2739 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2740 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2741 {
2742 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2743 ins.declaration.icb->data[4 * i + 0],
2744 ins.declaration.icb->data[4 * i + 1],
2745 ins.declaration.icb->data[4 * i + 2],
2746 ins.declaration.icb->data[4 * i + 3]);
2747 }
2748 shader_addline(&buffer, "}");
2749 }
2750 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2751 {
2752 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2753 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2754 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2755 }
2756 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2757 {
2758 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2759 ins.declaration.indexable_temp.register_idx,
2760 ins.declaration.indexable_temp.register_size,
2761 ins.declaration.indexable_temp.component_count);
2762 }
2763 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2764 {
2765 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2766 shader_dump_interpolation_mode(&buffer, ins.flags);
2767 shader_addline(&buffer, " ");
2768 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2769 }
2770 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2771 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2772 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2773 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2774 {
2775 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2776 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2777 shader_addline(&buffer, ", ");
2778 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2779 }
2780 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2781 {
2782 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2783 shader_dump_interpolation_mode(&buffer, ins.flags);
2784 shader_addline(&buffer, " ");
2785 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2786 shader_addline(&buffer, ", ");
2787 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2788 }
2789 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2790 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2791 {
2792 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2793 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2794 }
2795 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2796 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2797 {
2798 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2799 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2800 }
2801 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2802 {
2803 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2804 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2805 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2806 }
2807 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2808 {
2809 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2810 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2811 }
2812 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2813 {
2814 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2815 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2816 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2817 }
2818 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2819 {
2820 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2821 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2822 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2823 shader_addline(&buffer, ", comparisonMode");
2824 }
2825 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2826 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2827 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2828 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2829 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2830 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2831 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2832 {
2833 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2834 }
2835 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2836 {
2837 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2838 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2839 }
2840 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2841 {
2842 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2843 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2844 }
2845 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2846 {
2847 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2848 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2849 }
2850 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2851 {
2852 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2853 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2854 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2855 }
2856 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2857 {
2858 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2859 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2860 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2861 ins.declaration.tgsm_structured.structure_count);
2862 }
2863 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2864 {
2865 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2866 ins.declaration.thread_group_size.x,
2867 ins.declaration.thread_group_size.y,
2868 ins.declaration.thread_group_size.z);
2869 }
2870 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2871 {
2872 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2873 shader_dump_uav_flags(&buffer, ins.flags);
2874 shader_addline(&buffer, " ");
2875 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2876 }
2877 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2878 {
2879 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2880 shader_dump_uav_flags(&buffer, ins.flags);
2881 shader_addline(&buffer, " ");
2882 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2883 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2884 }
2885 else if (ins.handler_idx == WINED3DSIH_DEF)
2886 {
2887 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2888 ins.dst[0].reg.idx[0].offset),
2889 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2890 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2891 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2892 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2893 }
2894 else if (ins.handler_idx == WINED3DSIH_DEFI)
2895 {
2896 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2897 ins.src[0].reg.u.immconst_data[0],
2898 ins.src[0].reg.u.immconst_data[1],
2899 ins.src[0].reg.u.immconst_data[2],
2900 ins.src[0].reg.u.immconst_data[3]);
2901 }
2902 else if (ins.handler_idx == WINED3DSIH_DEFB)
2903 {
2904 shader_addline(&buffer, "defb b%u = %s",
2905 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2906 }
2907 else
2908 {
2909 if (ins.predicate)
2910 {
2911 shader_addline(&buffer, "(");
2912 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2913 shader_addline(&buffer, ") ");
2914 }
2915
2916 /* PixWin marks instructions with the coissue flag with a '+' */
2917 if (ins.coissue)
2918 shader_addline(&buffer, "+");
2919
2920 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2921
2922 if (ins.handler_idx == WINED3DSIH_BREAKP
2923 || ins.handler_idx == WINED3DSIH_CONTINUEP
2924 || ins.handler_idx == WINED3DSIH_IF
2925 || ins.handler_idx == WINED3DSIH_RETP)
2926 {
2927 switch (ins.flags)
2928 {
2929 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2930 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2931 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2932 }
2933 }
2934 else if (ins.handler_idx == WINED3DSIH_IFC
2935 || ins.handler_idx == WINED3DSIH_BREAKC)
2936 {
2937 switch (ins.flags)
2938 {
2939 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2940 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2941 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2942 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2943 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2944 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2945 default: shader_addline(&buffer, "_(%u)", ins.flags);
2946 }
2947 }
2948 else if (ins.handler_idx == WINED3DSIH_TEX
2949 && shader_version.major >= 2
2950 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2951 {
2952 shader_addline(&buffer, "p");
2953 }
2954 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2955 {
2956 switch (ins.flags)
2957 {
2958 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2959 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2960 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2961 }
2962 }
2963 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2964 {
2965 switch (ins.flags)
2966 {
2967 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2968 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2969 }
2970 }
2971 else if (ins.handler_idx == WINED3DSIH_SYNC)
2972 {
2973 shader_dump_sync_flags(&buffer, ins.flags);
2974 }
2975
2976 if (wined3d_shader_instruction_has_texel_offset(&ins))
2977 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2978
2979 for (i = 0; i < ins.dst_count; ++i)
2980 {
2981 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2982 shader_addline(&buffer, !i ? " " : ", ");
2983 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2984 }
2985
2986 /* Other source tokens */
2987 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2988 {
2989 shader_addline(&buffer, !i ? " " : ", ");
2990 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2991 }
2992 }
2993 shader_addline(&buffer, "\n");
2994 }
2995
2996 for (p = buffer.buffer; *p; p = q)
2997 {
2998 if (!(q = strstr(p, "\n")))
2999 q = p + strlen(p);
3000 else
3001 ++q;
3002 TRACE(" %.*s", (int)(q - p), p);
3003 }
3004
3005 string_buffer_free(&buffer);
3006 }
3007
3008 static void shader_cleanup(struct wined3d_shader *shader)
3009 {
3010 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3011 {
3012 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
3013 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.fork);
3014 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.join);
3015 }
3016 else if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
3017 {
3018 HeapFree(GetProcessHeap(), 0, shader->u.gs.so_desc.elements);
3019 }
3020
3021 HeapFree(GetProcessHeap(), 0, shader->patch_constant_signature.elements);
3022 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3023 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3024 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3025 shader->device->shader_backend->shader_destroy(shader);
3026 shader_cleanup_reg_maps(&shader->reg_maps);
3027 HeapFree(GetProcessHeap(), 0, shader->function);
3028 shader_delete_constant_list(&shader->constantsF);
3029 shader_delete_constant_list(&shader->constantsB);
3030 shader_delete_constant_list(&shader->constantsI);
3031 list_remove(&shader->shader_list_entry);
3032
3033 if (shader->frontend && shader->frontend_data)
3034 shader->frontend->shader_free(shader->frontend_data);
3035 }
3036
3037 struct shader_none_priv
3038 {
3039 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3040 const struct fragment_pipeline *fragment_pipe;
3041 BOOL ffp_proj_control;
3042 };
3043
3044 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3045 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3046 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3047 const struct wined3d_state *state) {}
3048 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3049 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3050 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3051 const struct wined3d_state *state) {}
3052 static void shader_none_destroy(struct wined3d_shader *shader) {}
3053 static void shader_none_free_context_data(struct wined3d_context *context) {}
3054 static void shader_none_init_context_state(struct wined3d_context *context) {}
3055
3056 /* Context activation is done by the caller. */
3057 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3058 const struct wined3d_state *state)
3059 {
3060 const struct wined3d_gl_info *gl_info = context->gl_info;
3061 struct shader_none_priv *priv = shader_priv;
3062
3063 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
3064 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
3065 }
3066
3067 /* Context activation is done by the caller. */
3068 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3069 {
3070 struct shader_none_priv *priv = shader_priv;
3071 const struct wined3d_gl_info *gl_info = context->gl_info;
3072
3073 priv->vertex_pipe->vp_enable(gl_info, FALSE);
3074 priv->fragment_pipe->enable_extension(gl_info, FALSE);
3075
3076 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3077 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3078 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3079 | (1u << WINED3D_SHADER_TYPE_HULL)
3080 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3081 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3082 }
3083
3084 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3085 const struct fragment_pipeline *fragment_pipe)
3086 {
3087 struct fragment_caps fragment_caps;
3088 void *vertex_priv, *fragment_priv;
3089 struct shader_none_priv *priv;
3090
3091 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
3092 return E_OUTOFMEMORY;
3093
3094 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3095 {
3096 ERR("Failed to initialize vertex pipe.\n");
3097 HeapFree(GetProcessHeap(), 0, priv);
3098 return E_FAIL;
3099 }
3100
3101 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3102 {
3103 ERR("Failed to initialize fragment pipe.\n");
3104 vertex_pipe->vp_free(device);
3105 HeapFree(GetProcessHeap(), 0, priv);
3106 return E_FAIL;
3107 }
3108
3109 priv->vertex_pipe = vertex_pipe;
3110 priv->fragment_pipe = fragment_pipe;
3111 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
3112 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3113
3114 device->vertex_priv = vertex_priv;
3115 device->fragment_priv = fragment_priv;
3116 device->shader_priv = priv;
3117
3118 return WINED3D_OK;
3119 }
3120
3121 static void shader_none_free(struct wined3d_device *device)
3122 {
3123 struct shader_none_priv *priv = device->shader_priv;
3124
3125 priv->fragment_pipe->free_private(device);
3126 priv->vertex_pipe->vp_free(device);
3127 HeapFree(GetProcessHeap(), 0, priv);
3128 }
3129
3130 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3131 {
3132 return TRUE;
3133 }
3134
3135 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
3136 {
3137 /* Set the shader caps to 0 for the none shader backend */
3138 caps->vs_version = 0;
3139 caps->hs_version = 0;
3140 caps->ds_version = 0;
3141 caps->gs_version = 0;
3142 caps->ps_version = 0;
3143 caps->cs_version = 0;
3144 caps->vs_uniform_count = 0;
3145 caps->ps_uniform_count = 0;
3146 caps->ps_1x_max_value = 0.0f;
3147 caps->varying_count = 0;
3148 caps->wined3d_caps = 0;
3149 }
3150
3151 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3152 {
3153 /* We "support" every possible fixup, since we don't support any shader
3154 * model, and will never have to actually sample a texture. */
3155 return TRUE;
3156 }
3157
3158 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3159 {
3160 struct shader_none_priv *priv = shader_priv;
3161
3162 return priv->ffp_proj_control;
3163 }
3164
3165 const struct wined3d_shader_backend_ops none_shader_backend =
3166 {
3167 shader_none_handle_instruction,
3168 shader_none_precompile,
3169 shader_none_select,
3170 shader_none_select_compute,
3171 shader_none_disable,
3172 shader_none_update_float_vertex_constants,
3173 shader_none_update_float_pixel_constants,
3174 shader_none_load_constants,
3175 shader_none_destroy,
3176 shader_none_alloc,
3177 shader_none_free,
3178 shader_none_allocate_context_data,
3179 shader_none_free_context_data,
3180 shader_none_init_context_state,
3181 shader_none_get_caps,
3182 shader_none_color_fixup_supported,
3183 shader_none_has_ffp_proj_control,
3184 };
3185
3186 static HRESULT shader_set_function(struct wined3d_shader *shader, DWORD float_const_count,
3187 enum wined3d_shader_type type, unsigned int max_version)
3188 {
3189 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3190 const struct wined3d_shader_frontend *fe;
3191 HRESULT hr;
3192 unsigned int backend_version;
3193 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3194
3195 TRACE("shader %p, float_const_count %u, type %#x, max_version %u.\n",
3196 shader, float_const_count, type, max_version);
3197
3198 fe = shader->frontend;
3199 if (!(shader->frontend_data = fe->shader_init(shader->function,
3200 shader->functionLength, &shader->output_signature)))
3201 {
3202 FIXME("Failed to initialize frontend.\n");
3203 return WINED3DERR_INVALIDCALL;
3204 }
3205
3206 /* First pass: trace shader. */
3207 if (TRACE_ON(d3d_shader))
3208 shader_trace_init(fe, shader->frontend_data);
3209
3210 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3211 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
3212 &shader->output_signature, float_const_count)))
3213 return hr;
3214
3215 if (reg_maps->shader_version.type != type)
3216 {
3217 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
3218 return WINED3DERR_INVALIDCALL;
3219 }
3220 if (reg_maps->shader_version.major > max_version)
3221 {
3222 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
3223 return WINED3DERR_INVALIDCALL;
3224 }
3225 switch (type)
3226 {
3227 case WINED3D_SHADER_TYPE_VERTEX:
3228 backend_version = d3d_info->limits.vs_version;
3229 break;
3230 case WINED3D_SHADER_TYPE_HULL:
3231 backend_version = d3d_info->limits.hs_version;
3232 break;
3233 case WINED3D_SHADER_TYPE_DOMAIN:
3234 backend_version = d3d_info->limits.ds_version;
3235 break;
3236 case WINED3D_SHADER_TYPE_GEOMETRY:
3237 backend_version = d3d_info->limits.gs_version;
3238 break;
3239 case WINED3D_SHADER_TYPE_PIXEL:
3240 backend_version = d3d_info->limits.ps_version;
3241 break;
3242 case WINED3D_SHADER_TYPE_COMPUTE:
3243 backend_version = d3d_info->limits.cs_version;
3244 break;
3245 default:
3246 FIXME("No backend version-checking for this shader type.\n");
3247 backend_version = 0;
3248 }
3249 if (reg_maps->shader_version.major > backend_version)
3250 {
3251 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
3252 reg_maps->shader_version.major, reg_maps->shader_version.minor);
3253 return WINED3DERR_INVALIDCALL;
3254 }
3255
3256 return WINED3D_OK;
3257 }
3258
3259 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3260 {
3261 ULONG refcount = InterlockedIncrement(&shader->ref);
3262
3263 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3264
3265 return refcount;
3266 }
3267
3268 static void wined3d_shader_init_object(void *object)
3269 {
3270 struct wined3d_shader *shader = object;
3271 struct wined3d_device *device = shader->device;
3272
3273 list_add_head(&device->shaders, &shader->shader_list_entry);
3274
3275 device->shader_backend->shader_precompile(device->shader_priv, shader);
3276 }
3277
3278 static void wined3d_shader_destroy_object(void *object)
3279 {
3280 shader_cleanup(object);
3281 HeapFree(GetProcessHeap(), 0, object);
3282 }
3283
3284 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3285 {
3286 ULONG refcount = InterlockedDecrement(&shader->ref);
3287
3288 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3289
3290 if (!refcount)
3291 {
3292 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3293 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3294 }
3295
3296 return refcount;
3297 }
3298
3299 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3300 {
3301 TRACE("shader %p.\n", shader);
3302
3303 return shader->parent;
3304 }
3305
3306 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3307 void *byte_code, UINT *byte_code_size)
3308 {
3309 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3310
3311 if (!byte_code)
3312 {
3313 *byte_code_size = shader->functionLength;
3314 return WINED3D_OK;
3315 }
3316
3317 if (*byte_code_size < shader->functionLength)
3318 {
3319 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3320 * than the required size we should write the required size and
3321 * return D3DERR_MOREDATA. That's not actually true. */
3322 return WINED3DERR_INVALIDCALL;
3323 }
3324
3325 memcpy(byte_code, shader->function, shader->functionLength);
3326
3327 return WINED3D_OK;
3328 }
3329
3330 /* Set local constants for d3d8 shaders. */
3331 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3332 UINT start_idx, const float *src_data, UINT count)
3333 {
3334 UINT end_idx = start_idx + count;
3335 UINT i;
3336
3337 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3338
3339 if (end_idx > shader->limits->constant_float)
3340 {
3341 WARN("end_idx %u > float constants limit %u.\n",
3342 end_idx, shader->limits->constant_float);
3343 end_idx = shader->limits->constant_float;
3344 }
3345
3346 for (i = start_idx; i < end_idx; ++i)
3347 {
3348 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
3349 float *value;
3350 if (!lconst)
3351 return E_OUTOFMEMORY;
3352
3353 lconst->idx = i;
3354 value = (float *)lconst->value;
3355 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3356 list_add_head(&shader->constantsF, &lconst->entry);
3357
3358 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3359 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3360 {
3361 shader->lconst_inf_or_nan = TRUE;
3362 }
3363 }
3364
3365 return WINED3D_OK;
3366 }
3367
3368 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3369 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
3370 {
3371 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3372 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3373 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3374 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3375 args->point_size = state->gl_primitive_type == GL_POINTS;
3376 args->per_vertex_point_size = shader->reg_maps.point_size;
3377 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
3378 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3379 : WINED3D_SHADER_TYPE_PIXEL;
3380 if (shader->reg_maps.shader_version.major >= 4)
3381 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
3382 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
3383 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
3384 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
3385 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
3386 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3387 else
3388 args->next_shader_input_count = 0;
3389 args->swizzle_map = swizzle_map;
3390 if (d3d_info->emulated_flatshading)
3391 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3392 else
3393 args->flatshading = 0;
3394 }
3395
3396 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3397 {
3398 if (usage_idx1 != usage_idx2)
3399 return FALSE;
3400 if (usage1 == usage2)
3401 return TRUE;
3402 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3403 return TRUE;
3404 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3405 return TRUE;
3406
3407 return FALSE;
3408 }
3409
3410 BOOL vshader_get_input(const struct wined3d_shader *shader,
3411 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3412 {
3413 WORD map = shader->reg_maps.input_registers;
3414 unsigned int i;
3415
3416 for (i = 0; map; map >>= 1, ++i)
3417 {
3418 if (!(map & 1)) continue;
3419
3420 if (match_usage(shader->u.vs.attributes[i].usage,
3421 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3422 {
3423 *regnum = i;
3424 return TRUE;
3425 }
3426 }
3427 return FALSE;
3428 }
3429
3430 static HRESULT shader_signature_calculate_strings_length(const struct wined3d_shader_signature *signature,
3431 SIZE_T *total)
3432 {
3433 struct wined3d_shader_signature_element *e;
3434 unsigned int i;
3435 SIZE_T len;
3436
3437 for (i = 0; i < signature->element_count; ++i)
3438 {
3439 e = &signature->elements[i];
3440 len = strlen(e->semantic_name);
3441 if (len >= ~(SIZE_T)0 - *total)
3442 return E_OUTOFMEMORY;
3443
3444 *total += len + 1;
3445 }
3446 return WINED3D_OK;
3447 }
3448
3449 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3450 const struct wined3d_shader_signature *src, char **signature_strings)
3451 {
3452 struct wined3d_shader_signature_element *e;
3453 unsigned int i;
3454 SIZE_T len;
3455 char *ptr;
3456
3457 if (!src->element_count)
3458 return WINED3D_OK;
3459
3460 ptr = *signature_strings;
3461
3462 dst->element_count = src->element_count;
3463 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3464 return E_OUTOFMEMORY;
3465
3466 for (i = 0; i < src->element_count; ++i)
3467 {
3468 e = &src->elements[i];
3469 dst->elements[i] = *e;
3470
3471 len = strlen(e->semantic_name);
3472 memcpy(ptr, e->semantic_name, len + 1);
3473 dst->elements[i].semantic_name = ptr;
3474 ptr += len + 1;
3475 }
3476
3477 *signature_strings = ptr;
3478
3479 return WINED3D_OK;
3480 }
3481
3482 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3483 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3484 void *parent, const struct wined3d_parent_ops *parent_ops)
3485 {
3486 size_t byte_code_size;
3487 SIZE_T total;
3488 HRESULT hr;
3489 char *ptr;
3490
3491 TRACE("byte_code %p, byte_code_size %#lx, format %#x, max_version %#x.\n",
3492 desc->byte_code, (long)desc->byte_code_size, desc->format, desc->max_version);
3493
3494 if (!desc->byte_code)
3495 return WINED3DERR_INVALIDCALL;
3496
3497 if (!(shader->frontend = shader_select_frontend(desc->format)))
3498 {
3499 FIXME("Unable to find frontend for shader.\n");
3500 return WINED3DERR_INVALIDCALL;
3501 }
3502
3503 shader->ref = 1;
3504 shader->device = device;
3505 shader->parent = parent;
3506 shader->parent_ops = parent_ops;
3507
3508 total = 0;
3509 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->input_signature, &total)))
3510 return hr;
3511 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->output_signature, &total)))
3512 return hr;
3513 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->patch_constant_signature, &total)))
3514 return hr;
3515 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3516 return E_OUTOFMEMORY;
3517 ptr = shader->signature_strings;
3518
3519 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3520 {
3521 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3522 return hr;
3523 }
3524 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3525 {
3526 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3527 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3528 return hr;
3529 }
3530 if (FAILED(hr = shader_signature_copy(&shader->patch_constant_signature, &desc->patch_constant_signature, &ptr)))
3531 {
3532 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3533 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3534 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3535 return hr;
3536 }
3537
3538 list_init(&shader->linked_programs);
3539 list_init(&shader->constantsF);
3540 list_init(&shader->constantsB);
3541 list_init(&shader->constantsI);
3542 shader->lconst_inf_or_nan = FALSE;
3543 list_init(&shader->reg_maps.indexable_temps);
3544 list_init(&shader->shader_list_entry);
3545
3546 byte_code_size = desc->byte_code_size;
3547 if (byte_code_size == ~(size_t)0)
3548 {
3549 const struct wined3d_shader_frontend *fe = shader->frontend;
3550 struct wined3d_shader_version shader_version;
3551 struct wined3d_shader_instruction ins;
3552 const DWORD *ptr;
3553 void *fe_data;
3554
3555 if (!(fe_data = fe->shader_init(desc->byte_code, byte_code_size, &shader->output_signature)))
3556 {
3557 WARN("Failed to initialise frontend data.\n");
3558 shader_cleanup(shader);
3559 return WINED3DERR_INVALIDCALL;
3560 }
3561
3562 fe->shader_read_header(fe_data, &ptr, &shader_version);
3563 while (!fe->shader_is_end(fe_data, &ptr))
3564 fe->shader_read_instruction(fe_data, &ptr, &ins);
3565
3566 fe->shader_free(fe_data);
3567
3568 byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3569 }
3570
3571 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, byte_code_size)))
3572 {
3573 shader_cleanup(shader);
3574 return E_OUTOFMEMORY;
3575 }
3576 memcpy(shader->function, desc->byte_code, byte_code_size);
3577 shader->functionLength = byte_code_size;
3578
3579 if (FAILED(hr = shader_set_function(shader, float_const_count, type, desc->max_version)))
3580 {
3581 WARN("Failed to set function, hr %#x.\n", hr);
3582 shader_cleanup(shader);
3583 return hr;
3584 }
3585
3586 shader->load_local_constsF = shader->lconst_inf_or_nan;
3587
3588 wined3d_cs_init_object(shader->device->cs, wined3d_shader_init_object, shader);
3589
3590 return hr;
3591 }
3592
3593 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3594 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3595 {
3596 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3597 unsigned int i;
3598 HRESULT hr;
3599
3600 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3601 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3602 return hr;
3603
3604 for (i = 0; i < shader->input_signature.element_count; ++i)
3605 {
3606 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3607
3608 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3609 continue;
3610
3611 shader->u.vs.attributes[input->register_idx].usage =
3612 shader_usage_from_semantic_name(input->semantic_name);
3613 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3614 }
3615
3616 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3617 shader->load_local_constsF = TRUE;
3618
3619 return WINED3D_OK;
3620 }
3621
3622 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3623 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3624 {
3625 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3626 }
3627
3628 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3629 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3630 {
3631 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3632 }
3633
3634 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3635 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3636 void *parent, const struct wined3d_parent_ops *parent_ops)
3637 {
3638 struct wined3d_stream_output_element *elements = NULL;
3639 HRESULT hr;
3640
3641 if (so_desc && !(elements = wined3d_calloc(so_desc->element_count, sizeof(*elements))))
3642 return E_OUTOFMEMORY;
3643
3644 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
3645 {
3646 HeapFree(GetProcessHeap(), 0, elements);
3647 return hr;
3648 }
3649
3650 if (so_desc)
3651 {
3652 shader->u.gs.so_desc = *so_desc;
3653 shader->u.gs.so_desc.elements = elements;
3654 memcpy(elements, so_desc->elements, so_desc->element_count * sizeof(*elements));
3655 }
3656
3657 return WINED3D_OK;
3658 }
3659
3660 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3661 struct ds_compile_args *args, const struct wined3d_context *context)
3662 {
3663 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3664
3665 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3666 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3667
3668 args->output_count = state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ?
3669 state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input :
3670 state->shader[WINED3D_SHADER_TYPE_PIXEL] ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input
3671 : shader->limits->packed_output;
3672 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3673 : WINED3D_SHADER_TYPE_PIXEL;
3674
3675 args->render_offscreen = context->render_offscreen;
3676
3677 args->padding = 0;
3678 }
3679
3680 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3681 struct gs_compile_args *args)
3682 {
3683 args->output_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3684 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : shader->limits->packed_output;
3685 }
3686
3687 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3688 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3689 {
3690 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3691 const struct wined3d_gl_info *gl_info = context->gl_info;
3692 const struct wined3d_texture *texture;
3693 unsigned int i;
3694
3695 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3696 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3697 {
3698 static unsigned int warned = 0;
3699
3700 args->srgb_correction = 1;
3701 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3702 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3703 "support, expect rendering artifacts.\n");
3704 }
3705
3706 if (shader->reg_maps.shader_version.major == 1
3707 && shader->reg_maps.shader_version.minor <= 3)
3708 {
3709 for (i = 0; i < shader->limits->sampler; ++i)
3710 {
3711 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3712
3713 if (flags & WINED3D_TTFF_PROJECTED)
3714 {
3715 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3716
3717 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3718 {
3719 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3720 unsigned int j;
3721 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3722 DWORD max_valid = WINED3D_TTFF_COUNT4;
3723
3724 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3725 {
3726 struct wined3d_vertex_declaration_element *element =
3727 &state->vertex_declaration->elements[j];
3728
3729 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3730 && element->usage_idx == index)
3731 {
3732 max_valid = element->format->component_count;
3733 break;
3734 }
3735 }
3736 if (!tex_transform || tex_transform > max_valid)
3737 {
3738 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3739 tex_transform, max_valid);
3740 tex_transform = max_valid;
3741 }
3742 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3743 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3744 && tex_transform > WINED3D_TTFF_COUNT2)
3745 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3746 && tex_transform > WINED3D_TTFF_COUNT3))
3747 tex_transform |= WINED3D_PSARGS_PROJECTED;
3748 else
3749 {
3750 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3751 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3752 i, tex_transform, resource_type);
3753 }
3754 }
3755 else
3756 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3757
3758 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3759 }
3760 }
3761 }
3762 if (shader->reg_maps.shader_version.major == 1
3763 && shader->reg_maps.shader_version.minor <= 4)
3764 {
3765 for (i = 0; i < shader->limits->sampler; ++i)
3766 {
3767 const struct wined3d_texture *texture = state->textures[i];
3768
3769 if (!shader->reg_maps.resource_info[i].type)
3770 continue;
3771
3772 /* Treat unbound textures as 2D. The dummy texture will provide
3773 * the proper sample value. The tex_types bitmap defaults to
3774 * 2D because of the memset. */
3775 if (!texture)
3776 continue;
3777
3778 switch (texture->target)
3779 {
3780 /* RECT textures are distinguished from 2D textures via np2_fixup */
3781 case GL_TEXTURE_RECTANGLE_ARB:
3782 case GL_TEXTURE_2D:
3783 break;
3784
3785 case GL_TEXTURE_3D:
3786 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3787 break;
3788
3789 case GL_TEXTURE_CUBE_MAP_ARB:
3790 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3791 break;
3792 }
3793 }
3794 }
3795
3796 if (shader->reg_maps.shader_version.major >= 4)
3797 {
3798 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3799 args->shadow = 0;
3800 for (i = 0 ; i < MAX_FRAGMENT_SAMPLERS; ++i)
3801 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3802 args->np2_fixup = 0;
3803 }
3804 else
3805 {
3806 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3807 {
3808 if (!shader->reg_maps.resource_info[i].type)
3809 continue;
3810
3811 texture = state->textures[i];
3812 if (!texture)
3813 {
3814 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3815 continue;
3816 }
3817 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3818 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3819 else
3820 args->color_fixup[i] = texture->resource.format->color_fixup;
3821
3822 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3823 args->shadow |= 1u << i;
3824
3825 /* Flag samplers that need NP2 texcoord fixup. */
3826 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3827 args->np2_fixup |= (1u << i);
3828 }
3829 }
3830
3831 if (shader->reg_maps.shader_version.major >= 3)
3832 {
3833 if (position_transformed)
3834 args->vp_mode = pretransformed;
3835 else if (use_vs(state))
3836 args->vp_mode = vertexshader;
3837 else
3838 args->vp_mode = fixedfunction;
3839 args->fog = WINED3D_FFP_PS_FOG_OFF;
3840 }
3841 else
3842 {
3843 args->vp_mode = vertexshader;
3844 if (state->render_states[WINED3D_RS_FOGENABLE])
3845 {
3846 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3847 {
3848 case WINED3D_FOG_NONE:
3849 if (position_transformed || use_vs(state))
3850 {
3851 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3852 break;
3853 }
3854
3855 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3856 {
3857 case WINED3D_FOG_NONE: /* Fall through. */
3858 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3859 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3860 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3861 }
3862 break;
3863
3864 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3865 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3866 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3867 }
3868 }
3869 else
3870 {
3871 args->fog = WINED3D_FFP_PS_FOG_OFF;
3872 }
3873 }
3874
3875 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3876 {
3877 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3878
3879 args->texcoords_initialized = 0;
3880 for (i = 0; i < MAX_TEXTURES; ++i)
3881 {
3882 if (vs)
3883 {
3884 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3885 args->texcoords_initialized |= 1u << i;
3886 }
3887 else
3888 {
3889 const struct wined3d_stream_info *si = &context->stream_info;
3890 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3891
3892 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3893 & WINED3D_FFP_TCI_MASK
3894 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3895 args->texcoords_initialized |= 1u << i;
3896 }
3897 }
3898 }
3899 else
3900 {
3901 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3902 }
3903
3904 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3905 && state->gl_primitive_type == GL_POINTS;
3906
3907 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3908 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3909 else
3910 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3911 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3912 : WINED3D_CMP_ALWAYS) - 1;
3913
3914 if (d3d_info->emulated_flatshading)
3915 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3916
3917 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3918 ? context->render_offscreen : 0;
3919 }
3920
3921 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3922 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3923 {
3924 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3925 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3926 HRESULT hr;
3927
3928 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3929 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3930 return hr;
3931
3932 for (i = 0; i < MAX_REG_INPUT; ++i)
3933 {
3934 if (shader->u.ps.input_reg_used & (1u << i))
3935 {
3936 ++num_regs_used;
3937 highest_reg_used = i;
3938 }
3939 }
3940
3941 /* Don't do any register mapping magic if it is not needed, or if we can't
3942 * achieve anything anyway */
3943 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3944 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3945 || shader->reg_maps.shader_version.major >= 4)
3946 {
3947 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3948 {
3949 /* This happens with relative addressing. The input mapper function
3950 * warns about this if the higher registers are declared too, so
3951 * don't write a FIXME here */
3952 WARN("More varying registers used than supported\n");
3953 }
3954
3955 for (i = 0; i < MAX_REG_INPUT; ++i)
3956 {
3957 shader->u.ps.input_reg_map[i] = i;
3958 }
3959
3960 shader->u.ps.declared_in_count = highest_reg_used + 1;
3961 }
3962 else
3963 {
3964 shader->u.ps.declared_in_count = 0;
3965 for (i = 0; i < MAX_REG_INPUT; ++i)
3966 {
3967 if (shader->u.ps.input_reg_used & (1u << i))
3968 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3969 else shader->u.ps.input_reg_map[i] = ~0U;
3970 }
3971 }
3972
3973 return WINED3D_OK;
3974 }
3975
3976 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3977 {
3978 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3979 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3980 unsigned int i;
3981
3982 if (reg_maps->shader_version.major != 1) return;
3983
3984 for (i = 0; i < shader->limits->sampler; ++i)
3985 {
3986 /* We don't sample from this sampler. */
3987 if (!resource_info[i].type)
3988 continue;
3989
3990 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3991 {
3992 case WINED3D_SHADER_TEX_2D:
3993 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3994 break;
3995
3996 case WINED3D_SHADER_TEX_3D:
3997 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3998 break;
3999
4000 case WINED3D_SHADER_TEX_CUBE:
4001 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
4002 break;
4003 }
4004 }
4005 }
4006
4007 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4008 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4009 {
4010 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
4011 }
4012
4013 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4014 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4015 {
4016 struct wined3d_shader *object;
4017 HRESULT hr;
4018
4019 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4020 device, desc, parent, parent_ops, shader);
4021
4022 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4023 return E_OUTOFMEMORY;
4024
4025 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
4026 {
4027 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4028 HeapFree(GetProcessHeap(), 0, object);
4029 return hr;
4030 }
4031
4032 TRACE("Created compute shader %p.\n", object);
4033 *shader = object;
4034
4035 return WINED3D_OK;
4036 }
4037
4038 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4039 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4040 {
4041 struct wined3d_shader *object;
4042 HRESULT hr;
4043
4044 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4045 device, desc, parent, parent_ops, shader);
4046
4047 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4048 return E_OUTOFMEMORY;
4049
4050 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
4051 {
4052 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4053 HeapFree(GetProcessHeap(), 0, object);
4054 return hr;
4055 }
4056
4057 TRACE("Created domain shader %p.\n", object);
4058 *shader = object;
4059
4060 return WINED3D_OK;
4061 }
4062
4063 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4064 const struct wined3d_stream_output_desc *so_desc, void *parent,
4065 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4066 {
4067 struct wined3d_shader *object;
4068 HRESULT hr;
4069
4070 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4071 device, desc, so_desc, parent, parent_ops, shader);
4072
4073 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4074 return E_OUTOFMEMORY;
4075
4076 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4077 {
4078 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4079 HeapFree(GetProcessHeap(), 0, object);
4080 return hr;
4081 }
4082
4083 TRACE("Created geometry shader %p.\n", object);
4084 *shader = object;
4085
4086 return WINED3D_OK;
4087 }
4088
4089 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4090 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4091 {
4092 struct wined3d_shader *object;
4093 HRESULT hr;
4094
4095 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4096 device, desc, parent, parent_ops, shader);
4097
4098 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4099 return E_OUTOFMEMORY;
4100
4101 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
4102 {
4103 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4104 HeapFree(GetProcessHeap(), 0, object);
4105 return hr;
4106 }
4107
4108 TRACE("Created hull shader %p.\n", object);
4109 *shader = object;
4110
4111 return WINED3D_OK;
4112 }
4113
4114 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4115 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4116 {
4117 struct wined3d_shader *object;
4118 HRESULT hr;
4119
4120 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4121 device, desc, parent, parent_ops, shader);
4122
4123 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4124 if (!object)
4125 return E_OUTOFMEMORY;
4126
4127 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4128 {
4129 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4130 HeapFree(GetProcessHeap(), 0, object);
4131 return hr;
4132 }
4133
4134 TRACE("Created pixel shader %p.\n", object);
4135 *shader = object;
4136
4137 return WINED3D_OK;
4138 }
4139
4140 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4141 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4142 {
4143 struct wined3d_shader *object;
4144 HRESULT hr;
4145
4146 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4147 device, desc, parent, parent_ops, shader);
4148
4149 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4150 if (!object)
4151 return E_OUTOFMEMORY;
4152
4153 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4154 {
4155 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4156 HeapFree(GetProcessHeap(), 0, object);
4157 return hr;
4158 }
4159
4160 TRACE("Created vertex shader %p.\n", object);
4161 *shader = object;
4162
4163 return WINED3D_OK;
4164 }
Windows API implementation