?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
winex11: Store the visual info in the OpenGL pixel format.
[wine.git] / dlls / wined3d / shader.c
blob424025cb8d369a692ec4bf86b50a56a2854989e4
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <stdio.h>
29 #include <string.h>
30
31 #include "wined3d_private.h"
32
33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
34
35 /* pow, mul_high, sub_high, mul_low */
36 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
37 /* cmp */
38 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
39
40 static const char * const shader_opcode_names[] =
41 {
42 /* WINED3DSIH_ABS */ "abs",
43 /* WINED3DSIH_ADD */ "add",
44 /* WINED3DSIH_AND */ "and",
45 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
46 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
47 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
48 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
49 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
50 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
51 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
52 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
53 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
54 /* WINED3DSIH_BEM */ "bem",
55 /* WINED3DSIH_BFI */ "bfi",
56 /* WINED3DSIH_BFREV */ "bfrev",
57 /* WINED3DSIH_BREAK */ "break",
58 /* WINED3DSIH_BREAKC */ "breakc",
59 /* WINED3DSIH_BREAKP */ "breakp",
60 /* WINED3DSIH_BUFINFO */ "bufinfo",
61 /* WINED3DSIH_CALL */ "call",
62 /* WINED3DSIH_CALLNZ */ "callnz",
63 /* WINED3DSIH_CASE */ "case",
64 /* WINED3DSIH_CMP */ "cmp",
65 /* WINED3DSIH_CND */ "cnd",
66 /* WINED3DSIH_CONTINUE */ "continue",
67 /* WINED3DSIH_CONTINUEP */ "continuec",
68 /* WINED3DSIH_COUNTBITS */ "countbits",
69 /* WINED3DSIH_CRS */ "crs",
70 /* WINED3DSIH_CUT */ "cut",
71 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
72 /* WINED3DSIH_DCL */ "dcl",
73 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
74 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
75 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
76 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
77 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
78 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
79 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
81 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
82 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
83 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
84 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
85 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
86 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
87 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
88 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
89 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
90 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
91 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
92 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
93 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
94 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
95 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
96 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
97 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
98 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
99 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
100 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
101 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
102 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
103 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
104 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
105 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
106 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
107 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
108 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
109 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
110 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
111 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
112 /* WINED3DSIH_DEF */ "def",
113 /* WINED3DSIH_DEFAULT */ "default",
114 /* WINED3DSIH_DEFB */ "defb",
115 /* WINED3DSIH_DEFI */ "defi",
116 /* WINED3DSIH_DIV */ "div",
117 /* WINED3DSIH_DP2 */ "dp2",
118 /* WINED3DSIH_DP2ADD */ "dp2add",
119 /* WINED3DSIH_DP3 */ "dp3",
120 /* WINED3DSIH_DP4 */ "dp4",
121 /* WINED3DSIH_DST */ "dst",
122 /* WINED3DSIH_DSX */ "dsx",
123 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
124 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
125 /* WINED3DSIH_DSY */ "dsy",
126 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
127 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
128 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
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 BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
697 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
698 {
699 switch (reg->type)
700 {
701 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
702 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
703 reg_maps->texcoord |= 1u << reg->idx[0].offset;
704 else
705 reg_maps->address |= 1u << reg->idx[0].offset;
706 break;
707
708 case WINED3DSPR_TEMP:
709 reg_maps->temporary |= 1u << reg->idx[0].offset;
710 break;
711
712 case WINED3DSPR_INPUT:
713 if (reg->idx[0].rel_addr)
714 reg_maps->input_rel_addressing = 1;
715 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
716 {
717 /* If relative addressing is used, we must assume that all
718 * registers are used. Even if it is a construct like v3[aL],
719 * we can't assume that v0, v1 and v2 aren't read because aL
720 * can be negative. */
721 if (reg->idx[0].rel_addr)
722 shader->u.ps.input_reg_used = ~0u;
723 else
724 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
725 }
726 else
727 {
728 reg_maps->input_registers |= 1u << reg->idx[0].offset;
729 }
730 break;
731
732 case WINED3DSPR_RASTOUT:
733 if (reg->idx[0].offset == 1)
734 reg_maps->fog = 1;
735 if (reg->idx[0].offset == 2)
736 reg_maps->point_size = 1;
737 break;
738
739 case WINED3DSPR_MISCTYPE:
740 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
741 {
742 if (!reg->idx[0].offset)
743 reg_maps->vpos = 1;
744 else if (reg->idx[0].offset == 1)
745 reg_maps->usesfacing = 1;
746 }
747 break;
748
749 case WINED3DSPR_CONST:
750 if (reg->idx[0].rel_addr)
751 {
752 if (reg->idx[0].offset < reg_maps->min_rel_offset)
753 reg_maps->min_rel_offset = reg->idx[0].offset;
754 if (reg->idx[0].offset > reg_maps->max_rel_offset)
755 reg_maps->max_rel_offset = reg->idx[0].offset;
756 reg_maps->usesrelconstF = TRUE;
757 }
758 else
759 {
760 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
761 {
762 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
763 return FALSE;
764 }
765 else
766 {
767 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
768 }
769 }
770 break;
771
772 case WINED3DSPR_CONSTINT:
773 if (reg->idx[0].offset >= shader->limits->constant_int)
774 {
775 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
776 return FALSE;
777 }
778 else
779 {
780 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
781 }
782 break;
783
784 case WINED3DSPR_CONSTBOOL:
785 if (reg->idx[0].offset >= shader->limits->constant_bool)
786 {
787 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
788 return FALSE;
789 }
790 else
791 {
792 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
793 }
794 break;
795
796 case WINED3DSPR_COLOROUT:
797 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
798 break;
799
800 case WINED3DSPR_OUTCONTROLPOINT:
801 reg_maps->vocp = 1;
802 break;
803
804 default:
805 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
806 reg->type, reg->idx[0].offset, reg->idx[1].offset);
807 break;
808 }
809 return TRUE;
810 }
811
812 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
813 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
814 {
815 struct wined3d_shader_sampler_map_entry *entries, *entry;
816 struct wined3d_shader_sampler_map *map;
817 unsigned int i;
818
819 map = &reg_maps->sampler_map;
820 entries = map->entries;
821 for (i = 0; i < map->count; ++i)
822 {
823 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
824 return;
825 }
826
827 if (!map->size)
828 {
829 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
830 {
831 ERR("Failed to allocate sampler map entries.\n");
832 return;
833 }
834 map->size = 4;
835 map->entries = entries;
836 }
837 else if (map->count == map->size)
838 {
839 size_t new_size = map->size * 2;
840
841 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
842 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
843 {
844 ERR("Failed to resize sampler map entries.\n");
845 return;
846 }
847 map->size = new_size;
848 map->entries = entries;
849 }
850
851 entry = &entries[map->count++];
852 entry->resource_idx = resource_idx;
853 entry->sampler_idx = sampler_idx;
854 entry->bind_idx = bind_idx;
855 }
856
857 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
858 {
859 switch (instr)
860 {
861 case WINED3DSIH_M4x4:
862 case WINED3DSIH_M3x4:
863 return param == 1 ? 3 : 0;
864
865 case WINED3DSIH_M4x3:
866 case WINED3DSIH_M3x3:
867 return param == 1 ? 2 : 0;
868
869 case WINED3DSIH_M3x2:
870 return param == 1 ? 1 : 0;
871
872 default:
873 return 0;
874 }
875 }
876
877 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
878 unsigned int register_idx, unsigned int size, unsigned int stride)
879 {
880 struct wined3d_shader_tgsm *tgsm;
881
882 if (register_idx >= MAX_TGSM_REGISTERS)
883 {
884 ERR("Invalid TGSM register index %u.\n", register_idx);
885 return S_OK;
886 }
887 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
888 {
889 FIXME("TGSM declarations are allowed only in compute shaders.\n");
890 return S_OK;
891 }
892
893 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
894 register_idx + 1, sizeof(*reg_maps->tgsm)))
895 return E_OUTOFMEMORY;
896
897 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
898 tgsm = &reg_maps->tgsm[register_idx];
899 tgsm->size = size;
900 tgsm->stride = stride;
901 return S_OK;
902 }
903
904 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
905 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
906 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
907 {
908 struct wined3d_shader_phase *phase;
909
910 if ((phase = *current_phase))
911 {
912 phase->end = previous_instruction_ptr;
913 *current_phase = NULL;
914 }
915
916 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
917 {
918 ERR("Unexpected shader type %#x.\n", shader->reg_maps.shader_version.type);
919 return E_FAIL;
920 }
921
922 switch (ins->handler_idx)
923 {
924 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
925 if (shader->u.hs.phases.control_point)
926 {
927 FIXME("Multiple control point phases.\n");
928 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
929 }
930 if (!(shader->u.hs.phases.control_point = HeapAlloc(GetProcessHeap(),
931 HEAP_ZERO_MEMORY, sizeof(*shader->u.hs.phases.control_point))))
932 return E_OUTOFMEMORY;
933 phase = shader->u.hs.phases.control_point;
934 break;
935 case WINED3DSIH_HS_FORK_PHASE:
936 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
937 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
938 sizeof(*shader->u.hs.phases.fork)))
939 return E_OUTOFMEMORY;
940 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
941 break;
942 case WINED3DSIH_HS_JOIN_PHASE:
943 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
944 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
945 sizeof(*shader->u.hs.phases.join)))
946 return E_OUTOFMEMORY;
947 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
948 break;
949 default:
950 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
951 return E_FAIL;
952 }
953
954 phase->start = current_instruction_ptr;
955 *current_phase = phase;
956
957 return WINED3D_OK;
958 }
959
960 static HRESULT shader_calculate_clip_or_cull_distance_mask(
961 const struct wined3d_shader_signature_element *e, unsigned int *mask)
962 {
963 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
964 * the only allowed semantic indices.
965 */
966 if (e->semantic_idx >= MAX_CLIP_DISTANCES / 4)
967 {
968 *mask = 0;
969 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
970 return WINED3DERR_INVALIDCALL;
971 }
972
973 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
974 return WINED3D_OK;
975 }
976
977 static void wined3d_insert_interpolation_mode(DWORD *packed_interpolation_mode,
978 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
979 {
980 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
981 FIXME("Unexpected interpolation mode %#x.\n", mode);
982
983 wined3d_insert_bits(packed_interpolation_mode,
984 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
985 }
986
987 /* Note that this does not count the loop register as an address register. */
988 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
989 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
990 struct wined3d_shader_signature *output_signature, DWORD constf_size)
991 {
992 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
993 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
994 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
995 struct wined3d_shader_version shader_version;
996 struct wined3d_shader_phase *phase = NULL;
997 const DWORD *ptr, *prev_ins, *current_ins;
998 void *fe_data = shader->frontend_data;
999 unsigned int i;
1000 HRESULT hr;
1001
1002 memset(reg_maps, 0, sizeof(*reg_maps));
1003 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1004 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1005 reg_maps->min_rel_offset = ~0U;
1006 list_init(&reg_maps->indexable_temps);
1007
1008 fe->shader_read_header(fe_data, &ptr, &shader_version);
1009 prev_ins = current_ins = ptr;
1010 reg_maps->shader_version = shader_version;
1011
1012 shader_set_limits(shader);
1013
1014 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1015 sizeof(*reg_maps->constf))))
1016 {
1017 ERR("Failed to allocate constant map memory.\n");
1018 return E_OUTOFMEMORY;
1019 }
1020
1021 while (!fe->shader_is_end(fe_data, &ptr))
1022 {
1023 struct wined3d_shader_instruction ins;
1024
1025 current_ins = ptr;
1026 /* Fetch opcode. */
1027 fe->shader_read_instruction(fe_data, &ptr, &ins);
1028
1029 /* Unhandled opcode, and its parameters. */
1030 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1031 {
1032 WARN("Encountered unrecognised or invalid instruction.\n");
1033 return WINED3DERR_INVALIDCALL;
1034 }
1035
1036 /* Handle declarations. */
1037 if (ins.handler_idx == WINED3DSIH_DCL
1038 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1039 {
1040 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1041 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1042
1043 switch (semantic->reg.reg.type)
1044 {
1045 /* Mark input registers used. */
1046 case WINED3DSPR_INPUT:
1047 if (reg_idx >= MAX_REG_INPUT)
1048 {
1049 ERR("Invalid input register index %u.\n", reg_idx);
1050 break;
1051 }
1052 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1053 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1054 return WINED3DERR_INVALIDCALL;
1055 reg_maps->input_registers |= 1u << reg_idx;
1056 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1057 break;
1058
1059 /* Vertex shader: mark 3.0 output registers used, save token. */
1060 case WINED3DSPR_OUTPUT:
1061 if (reg_idx >= MAX_REG_OUTPUT)
1062 {
1063 ERR("Invalid output register index %u.\n", reg_idx);
1064 break;
1065 }
1066 reg_maps->output_registers |= 1u << reg_idx;
1067 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1068 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1069 reg_maps->fog = 1;
1070 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1071 reg_maps->point_size = 1;
1072 break;
1073
1074 case WINED3DSPR_SAMPLER:
1075 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1076 case WINED3DSPR_RESOURCE:
1077 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1078 {
1079 ERR("Invalid resource index %u.\n", reg_idx);
1080 break;
1081 }
1082 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1083 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1084 break;
1085
1086 case WINED3DSPR_UAV:
1087 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1088 {
1089 ERR("Invalid UAV resource index %u.\n", reg_idx);
1090 break;
1091 }
1092 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1093 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1094 if (ins.flags)
1095 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1096 break;
1097
1098 default:
1099 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1100 break;
1101 }
1102 }
1103 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1104 {
1105 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1106 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1107 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1108 else
1109 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1110 }
1111 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1112 {
1113 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1114 {
1115 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1116 shader->u.ps.force_early_depth_stencil = TRUE;
1117 else
1118 FIXME("Invalid instruction %#x for shader type %#x.\n",
1119 ins.handler_idx, shader_version.type);
1120 }
1121 else
1122 {
1123 WARN("Ignoring global flags %#x.\n", ins.flags);
1124 }
1125 }
1126 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1127 {
1128 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1129 shader->u.gs.instance_count = ins.declaration.count;
1130 else
1131 FIXME("Invalid instruction %#x for shader type %#x.\n",
1132 ins.handler_idx, shader_version.type);
1133 }
1134 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1135 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1136 {
1137 if (phase)
1138 phase->instance_count = ins.declaration.count;
1139 else
1140 FIXME("Instruction %s outside of shader phase.\n",
1141 debug_d3dshaderinstructionhandler(ins.handler_idx));
1142 }
1143 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1144 {
1145 if (reg_maps->icb)
1146 FIXME("Multiple immediate constant buffers.\n");
1147 reg_maps->icb = ins.declaration.icb;
1148 }
1149 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1150 {
1151 if (phase)
1152 {
1153 FIXME("Indexable temporary registers not supported.\n");
1154 }
1155 else
1156 {
1157 struct wined3d_shader_indexable_temp *reg;
1158
1159 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
1160 return E_OUTOFMEMORY;
1161
1162 *reg = ins.declaration.indexable_temp;
1163 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1164 }
1165 }
1166 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1167 {
1168 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1169 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1170 else
1171 FIXME("Invalid instruction %#x for shader type %#x.\n",
1172 ins.handler_idx, shader_version.type);
1173 }
1174 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1175 {
1176 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1177 if (reg_idx >= MAX_REG_INPUT)
1178 {
1179 ERR("Invalid register index %u.\n", reg_idx);
1180 break;
1181 }
1182 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1183 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1184 else
1185 FIXME("Invalid instruction %#x for shader type %#x.\n",
1186 ins.handler_idx, shader_version.type);
1187 }
1188 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1189 {
1190 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1191 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1192 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1193 {
1194 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1195 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1196 else
1197 FIXME("Invalid instruction %#x for shader type %#x.\n",
1198 ins.handler_idx, shader_version.type);
1199 }
1200 }
1201 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1202 {
1203 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1204 shader->u.hs.output_vertex_count = ins.declaration.count;
1205 else
1206 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1207 }
1208 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1209 {
1210 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1211 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1212 else
1213 FIXME("Invalid instruction %#x for shader type %#x.\n",
1214 ins.handler_idx, shader_version.type);
1215 }
1216 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1217 {
1218 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1219 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1220 {
1221 ERR("Invalid resource index %u.\n", reg_idx);
1222 break;
1223 }
1224 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1225 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1226 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1227 }
1228 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1229 {
1230 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1231 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1232 {
1233 ERR("Invalid resource index %u.\n", reg_idx);
1234 break;
1235 }
1236 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1237 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1238 reg_maps->resource_info[reg_idx].flags = 0;
1239 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1240 }
1241 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1242 {
1243 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1244 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1245 }
1246 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1247 {
1248 if (phase)
1249 phase->temporary_count = ins.declaration.count;
1250 else
1251 reg_maps->temporary_count = ins.declaration.count;
1252 }
1253 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1254 {
1255 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1256 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1257 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1258 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1259 }
1260 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1261 {
1262 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1263 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1264 else
1265 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1266 }
1267 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1268 {
1269 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1270 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1271 else
1272 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1273 }
1274 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1275 {
1276 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1277 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1278 return hr;
1279 }
1280 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1281 {
1282 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1283 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1284 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1285 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1286 return hr;
1287 }
1288 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1289 {
1290 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1291 {
1292 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1293 }
1294 else
1295 {
1296 FIXME("Invalid instruction %#x for shader type %#x.\n",
1297 ins.handler_idx, shader_version.type);
1298 }
1299 }
1300 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1301 {
1302 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1303 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1304 {
1305 ERR("Invalid UAV resource index %u.\n", reg_idx);
1306 break;
1307 }
1308 if (ins.flags)
1309 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1310 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1311 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1312 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1313 }
1314 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1315 {
1316 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1317 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1318 {
1319 ERR("Invalid UAV resource index %u.\n", reg_idx);
1320 break;
1321 }
1322 if (ins.flags)
1323 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1324 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1325 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1326 reg_maps->uav_resource_info[reg_idx].flags = 0;
1327 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1328 }
1329 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1330 {
1331 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1332 shader->u.gs.vertices_out = ins.declaration.count;
1333 else
1334 FIXME("Invalid instruction %#x for shader type %#x.\n",
1335 ins.handler_idx, shader_version.type);
1336 }
1337 else if (ins.handler_idx == WINED3DSIH_DEF)
1338 {
1339 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1340 float *value;
1341 if (!lconst) return E_OUTOFMEMORY;
1342
1343 lconst->idx = ins.dst[0].reg.idx[0].offset;
1344 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1345 value = (float *)lconst->value;
1346
1347 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1348 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1349 {
1350 if (value[0] < -1.0f) value[0] = -1.0f;
1351 else if (value[0] > 1.0f) value[0] = 1.0f;
1352 if (value[1] < -1.0f) value[1] = -1.0f;
1353 else if (value[1] > 1.0f) value[1] = 1.0f;
1354 if (value[2] < -1.0f) value[2] = -1.0f;
1355 else if (value[2] > 1.0f) value[2] = 1.0f;
1356 if (value[3] < -1.0f) value[3] = -1.0f;
1357 else if (value[3] > 1.0f) value[3] = 1.0f;
1358 }
1359
1360 list_add_head(&shader->constantsF, &lconst->entry);
1361
1362 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1363 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1364 {
1365 shader->lconst_inf_or_nan = TRUE;
1366 }
1367 }
1368 else if (ins.handler_idx == WINED3DSIH_DEFI)
1369 {
1370 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1371 if (!lconst) return E_OUTOFMEMORY;
1372
1373 lconst->idx = ins.dst[0].reg.idx[0].offset;
1374 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1375
1376 list_add_head(&shader->constantsI, &lconst->entry);
1377 reg_maps->local_int_consts |= (1u << lconst->idx);
1378 }
1379 else if (ins.handler_idx == WINED3DSIH_DEFB)
1380 {
1381 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1382 if (!lconst) return E_OUTOFMEMORY;
1383
1384 lconst->idx = ins.dst[0].reg.idx[0].offset;
1385 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1386
1387 list_add_head(&shader->constantsB, &lconst->entry);
1388 reg_maps->local_bool_consts |= (1u << lconst->idx);
1389 }
1390 /* Handle shader phases. */
1391 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1392 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1393 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1394 {
1395 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1396 return hr;
1397 }
1398 /* For subroutine prototypes. */
1399 else if (ins.handler_idx == WINED3DSIH_LABEL)
1400 {
1401 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1402 }
1403 /* Set texture, address, temporary registers. */
1404 else
1405 {
1406 BOOL color0_mov = FALSE;
1407 unsigned int i;
1408
1409 /* This will loop over all the registers and try to
1410 * make a bitmask of the ones we're interested in.
1411 *
1412 * Relative addressing tokens are ignored, but that's
1413 * okay, since we'll catch any address registers when
1414 * they are initialized (required by spec). */
1415 for (i = 0; i < ins.dst_count; ++i)
1416 {
1417 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1418 shader_version.type, constf_size))
1419 return WINED3DERR_INVALIDCALL;
1420
1421 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1422 {
1423 UINT idx = ins.dst[i].reg.idx[0].offset;
1424
1425 switch (ins.dst[i].reg.type)
1426 {
1427 case WINED3DSPR_RASTOUT:
1428 if (shader_version.major >= 3)
1429 break;
1430 switch (idx)
1431 {
1432 case 0: /* oPos */
1433 reg_maps->output_registers |= 1u << 10;
1434 shader_signature_from_usage(&output_signature_elements[10],
1435 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1436 break;
1437
1438 case 1: /* oFog */
1439 reg_maps->output_registers |= 1u << 11;
1440 shader_signature_from_usage(&output_signature_elements[11],
1441 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1442 break;
1443
1444 case 2: /* oPts */
1445 reg_maps->output_registers |= 1u << 11;
1446 shader_signature_from_usage(&output_signature_elements[11],
1447 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1448 break;
1449 }
1450 break;
1451
1452 case WINED3DSPR_ATTROUT:
1453 if (shader_version.major >= 3)
1454 break;
1455 if (idx < 2)
1456 {
1457 idx += 8;
1458 if (reg_maps->output_registers & (1u << idx))
1459 {
1460 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1461 }
1462 else
1463 {
1464 reg_maps->output_registers |= 1u << idx;
1465 shader_signature_from_usage(&output_signature_elements[idx],
1466 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1467 }
1468 }
1469 break;
1470
1471 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1472 if (shader_version.major >= 3)
1473 {
1474 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1475 {
1476 WARN("Invalid output register index %u.\n", idx);
1477 break;
1478 }
1479 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1480 break;
1481 }
1482 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1483 {
1484 WARN("Invalid texcoord index %u.\n", idx);
1485 break;
1486 }
1487 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1488 if (reg_maps->output_registers & (1u << idx))
1489 {
1490 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1491 }
1492 else
1493 {
1494 reg_maps->output_registers |= 1u << idx;
1495 shader_signature_from_usage(&output_signature_elements[idx],
1496 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1497 }
1498 break;
1499
1500 default:
1501 break;
1502 }
1503 }
1504
1505 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1506 {
1507 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1508 {
1509 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1510 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1511 * the mov and perform the sRGB write correction from the source register.
1512 *
1513 * However, if the mov is only partial, we can't do this, and if the write
1514 * comes from an instruction other than MOV it is hard to do as well. If
1515 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1516 shader->u.ps.color0_mov = FALSE;
1517 if (ins.handler_idx == WINED3DSIH_MOV
1518 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1519 {
1520 /* Used later when the source register is read. */
1521 color0_mov = TRUE;
1522 }
1523 }
1524 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1525 * end
1526 */
1527 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1528 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1529 {
1530 shader->u.ps.color0_mov = FALSE;
1531 }
1532 }
1533
1534 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1535 if (shader_version.major == 1
1536 && (ins.handler_idx == WINED3DSIH_TEX
1537 || ins.handler_idx == WINED3DSIH_TEXBEM
1538 || ins.handler_idx == WINED3DSIH_TEXBEML
1539 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1540 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1541 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1542 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1543 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1544 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1545 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1546 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1547 {
1548 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1549
1550 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1551 {
1552 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1553 continue;
1554 }
1555
1556 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1557 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1558 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1559 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1560
1561 /* texbem is only valid with < 1.4 pixel shaders */
1562 if (ins.handler_idx == WINED3DSIH_TEXBEM
1563 || ins.handler_idx == WINED3DSIH_TEXBEML)
1564 {
1565 reg_maps->bumpmat |= 1u << reg_idx;
1566 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1567 {
1568 reg_maps->luminanceparams |= 1u << reg_idx;
1569 }
1570 }
1571 }
1572 else if (ins.handler_idx == WINED3DSIH_BEM)
1573 {
1574 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1575 }
1576 }
1577
1578 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1579 {
1580 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1581 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1582 {
1583 ERR("Invalid UAV index %u.\n", reg_idx);
1584 break;
1585 }
1586 reg_maps->uav_counter_mask |= (1u << reg_idx);
1587 }
1588 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1589 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1590 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1591 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1592 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1593 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1594 {
1595 unsigned int reg_idx;
1596 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1597 reg_idx = ins.src[1].reg.idx[0].offset;
1598 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1599 reg_idx = ins.src[2].reg.idx[0].offset;
1600 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1601 reg_idx = ins.dst[0].reg.idx[0].offset;
1602 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1603 reg_idx = ins.src[0].reg.idx[0].offset;
1604 else
1605 reg_idx = ins.dst[1].reg.idx[0].offset;
1606 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1607 {
1608 ERR("Invalid UAV index %u.\n", reg_idx);
1609 break;
1610 }
1611 reg_maps->uav_read_mask |= (1u << reg_idx);
1612 }
1613 else if (ins.handler_idx == WINED3DSIH_NRM)
1614 {
1615 reg_maps->usesnrm = 1;
1616 }
1617 else if (ins.handler_idx == WINED3DSIH_DSY
1618 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1619 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1620 {
1621 reg_maps->usesdsy = 1;
1622 }
1623 else if (ins.handler_idx == WINED3DSIH_DSX
1624 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1625 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1626 {
1627 reg_maps->usesdsx = 1;
1628 }
1629 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1630 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1631 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1632 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1633 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1634 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1635 else if (ins.handler_idx == WINED3DSIH_LOOP
1636 || ins.handler_idx == WINED3DSIH_REP)
1637 {
1638 ++cur_loop_depth;
1639 if (cur_loop_depth > max_loop_depth)
1640 max_loop_depth = cur_loop_depth;
1641 }
1642 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1643 || ins.handler_idx == WINED3DSIH_ENDREP)
1644 {
1645 --cur_loop_depth;
1646 }
1647 else if (ins.handler_idx == WINED3DSIH_GATHER4
1648 || ins.handler_idx == WINED3DSIH_GATHER4_C
1649 || ins.handler_idx == WINED3DSIH_SAMPLE
1650 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1651 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1652 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1653 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1654 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1655 {
1656 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1657 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1658 }
1659 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1660 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1661 {
1662 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1663 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1664 }
1665 else if (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1666 {
1667 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1668 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1669 }
1670 else if (ins.handler_idx == WINED3DSIH_LD
1671 || ins.handler_idx == WINED3DSIH_LD2DMS
1672 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1673 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1674 {
1675 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1676 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1677 }
1678 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1679 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1680 {
1681 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1682 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1683 }
1684
1685 if (ins.predicate)
1686 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1687 shader_version.type, constf_size))
1688 return WINED3DERR_INVALIDCALL;
1689
1690 for (i = 0; i < ins.src_count; ++i)
1691 {
1692 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1693 struct wined3d_shader_register reg = ins.src[i].reg;
1694
1695 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1696 shader_version.type, constf_size))
1697 return WINED3DERR_INVALIDCALL;
1698 while (count)
1699 {
1700 ++reg.idx[0].offset;
1701 if (!shader_record_register_usage(shader, reg_maps, &reg,
1702 shader_version.type, constf_size))
1703 return WINED3DERR_INVALIDCALL;
1704 --count;
1705 }
1706
1707 if (color0_mov)
1708 {
1709 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1710 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1711 {
1712 shader->u.ps.color0_mov = TRUE;
1713 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1714 }
1715 }
1716 }
1717 }
1718
1719 prev_ins = current_ins;
1720 }
1721 reg_maps->loop_depth = max_loop_depth;
1722
1723 if (phase)
1724 {
1725 phase->end = prev_ins;
1726 phase = NULL;
1727 }
1728
1729 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1730 * R0 is written to the render target. */
1731 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1732 reg_maps->rt_mask |= (1u << 0);
1733
1734 if (input_signature->elements)
1735 {
1736 for (i = 0; i < input_signature->element_count; ++i)
1737 {
1738 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1739 {
1740 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1741 {
1742 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1743 return WINED3DERR_INVALIDCALL;
1744 }
1745 }
1746 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1747 {
1748 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1749 reg_maps->vpos = 1;
1750 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1751 reg_maps->usesfacing = 1;
1752 }
1753 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1754 }
1755 }
1756 else if (!input_signature->elements && reg_maps->input_registers)
1757 {
1758 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1759 struct wined3d_shader_signature_element *e;
1760 unsigned int i;
1761
1762 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1763 return E_OUTOFMEMORY;
1764 input_signature->element_count = count;
1765
1766 e = input_signature->elements;
1767 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1768 {
1769 if (!(reg_maps->input_registers & (1u << i)))
1770 continue;
1771 input_signature_elements[i].register_idx = i;
1772 *e++ = input_signature_elements[i];
1773 }
1774 }
1775
1776 if (output_signature->elements)
1777 {
1778 for (i = 0; i < output_signature->element_count; ++i)
1779 {
1780 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1781 unsigned int mask;
1782
1783 reg_maps->output_registers |= 1u << e->register_idx;
1784 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
1785 {
1786 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1787 return hr;
1788 reg_maps->clip_distance_mask |= mask;
1789 }
1790 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
1791 {
1792 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1793 return hr;
1794 reg_maps->cull_distance_mask |= mask;
1795 }
1796 }
1797 }
1798 else if (reg_maps->output_registers)
1799 {
1800 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1801 struct wined3d_shader_signature_element *e;
1802
1803 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1804 return E_OUTOFMEMORY;
1805 output_signature->element_count = count;
1806
1807 e = output_signature->elements;
1808 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1809 {
1810 if (!(reg_maps->output_registers & (1u << i)))
1811 continue;
1812 *e++ = output_signature_elements[i];
1813 }
1814 }
1815
1816 return WINED3D_OK;
1817 }
1818
1819 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1820 {
1821 struct wined3d_shader_indexable_temp *reg, *reg_next;
1822
1823 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1824 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1825
1826 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1827 HeapFree(GetProcessHeap(), 0, reg);
1828 list_init(&reg_maps->indexable_temps);
1829
1830 HeapFree(GetProcessHeap(), 0, reg_maps->tgsm);
1831 }
1832
1833 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1834 {
1835 DWORD map = 1u << max;
1836 map |= map - 1;
1837 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1838
1839 return wined3d_log2i(map);
1840 }
1841
1842 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1843 {
1844 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1845 {
1846 shader_addline(buffer, "refactoringAllowed");
1847 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1848 if (global_flags)
1849 shader_addline(buffer, " | ");
1850 }
1851
1852 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1853 {
1854 shader_addline(buffer, "forceEarlyDepthStencil");
1855 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1856 if (global_flags)
1857 shader_addline(buffer, " | ");
1858 }
1859
1860 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1861 {
1862 shader_addline(buffer, "enableRawAndStructuredBuffers");
1863 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1864 }
1865
1866 if (global_flags)
1867 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1868 }
1869
1870 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1871 {
1872 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1873 {
1874 shader_addline(buffer, "_g");
1875 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1876 }
1877 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1878 {
1879 shader_addline(buffer, "_t");
1880 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1881 }
1882
1883 if (sync_flags)
1884 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1885 }
1886
1887 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1888 {
1889 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1890 {
1891 shader_addline(buffer, "_glc");
1892 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1893 }
1894 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1895 {
1896 shader_addline(buffer, "_opc");
1897 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1898 }
1899
1900 if (uav_flags)
1901 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1902 }
1903
1904 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1905 enum wined3d_tessellator_domain domain)
1906 {
1907 switch (domain)
1908 {
1909 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1910 shader_addline(buffer, "line");
1911 break;
1912 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1913 shader_addline(buffer, "triangle");
1914 break;
1915 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1916 shader_addline(buffer, "quad");
1917 break;
1918 default:
1919 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1920 break;
1921 }
1922 }
1923
1924 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1925 enum wined3d_tessellator_output_primitive output_primitive)
1926 {
1927 switch (output_primitive)
1928 {
1929 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1930 shader_addline(buffer, "point");
1931 break;
1932 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1933 shader_addline(buffer, "line");
1934 break;
1935 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1936 shader_addline(buffer, "triangle_cw");
1937 break;
1938 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1939 shader_addline(buffer, "triangle_ccw");
1940 break;
1941 default:
1942 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1943 break;
1944 }
1945 }
1946
1947 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1948 enum wined3d_tessellator_partitioning partitioning)
1949 {
1950 switch (partitioning)
1951 {
1952 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1953 shader_addline(buffer, "integer");
1954 break;
1955 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1956 shader_addline(buffer, "pow2");
1957 break;
1958 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1959 shader_addline(buffer, "fractional_odd");
1960 break;
1961 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1962 shader_addline(buffer, "fractional_even");
1963 break;
1964 default:
1965 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1966 break;
1967 }
1968 }
1969
1970 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
1971 enum wined3d_shader_input_sysval_semantic semantic)
1972 {
1973 unsigned int i;
1974
1975 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
1976 {
1977 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
1978 {
1979 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
1980 return;
1981 }
1982 }
1983
1984 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
1985 }
1986
1987 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1988 const struct wined3d_shader_semantic *semantic, unsigned int flags,
1989 const struct wined3d_shader_version *shader_version)
1990 {
1991 shader_addline(buffer, "dcl");
1992
1993 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1994 {
1995 switch (semantic->resource_type)
1996 {
1997 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1998 shader_addline(buffer, "_2d");
1999 break;
2000
2001 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2002 shader_addline(buffer, "_3d");
2003 break;
2004
2005 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2006 shader_addline(buffer, "_cube");
2007 break;
2008
2009 default:
2010 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2011 break;
2012 }
2013 }
2014 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2015 {
2016 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2017 shader_addline(buffer, "_resource_");
2018 else
2019 shader_addline(buffer, "_uav_");
2020 switch (semantic->resource_type)
2021 {
2022 case WINED3D_SHADER_RESOURCE_BUFFER:
2023 shader_addline(buffer, "buffer");
2024 break;
2025
2026 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
2027 shader_addline(buffer, "texture1d");
2028 break;
2029
2030 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2031 shader_addline(buffer, "texture2d");
2032 break;
2033
2034 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
2035 shader_addline(buffer, "texture2dms");
2036 break;
2037
2038 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2039 shader_addline(buffer, "texture3d");
2040 break;
2041
2042 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2043 shader_addline(buffer, "texturecube");
2044 break;
2045
2046 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
2047 shader_addline(buffer, "texture1darray");
2048 break;
2049
2050 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
2051 shader_addline(buffer, "texture2darray");
2052 break;
2053
2054 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
2055 shader_addline(buffer, "texture2dmsarray");
2056 break;
2057
2058 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
2059 shader_addline(buffer, "texturecubearray");
2060 break;
2061
2062 default:
2063 shader_addline(buffer, "unknown");
2064 break;
2065 }
2066 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2067 shader_dump_uav_flags(buffer, flags);
2068 switch (semantic->resource_data_type)
2069 {
2070 case WINED3D_DATA_FLOAT:
2071 shader_addline(buffer, " (float)");
2072 break;
2073
2074 case WINED3D_DATA_INT:
2075 shader_addline(buffer, " (int)");
2076 break;
2077
2078 case WINED3D_DATA_UINT:
2079 shader_addline(buffer, " (uint)");
2080 break;
2081
2082 case WINED3D_DATA_UNORM:
2083 shader_addline(buffer, " (unorm)");
2084 break;
2085
2086 case WINED3D_DATA_SNORM:
2087 shader_addline(buffer, " (snorm)");
2088 break;
2089
2090 default:
2091 shader_addline(buffer, " (unknown)");
2092 break;
2093 }
2094 }
2095 else
2096 {
2097 /* Pixel shaders 3.0 don't have usage semantics. */
2098 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2099 return;
2100 else
2101 shader_addline(buffer, "_");
2102
2103 switch (semantic->usage)
2104 {
2105 case WINED3D_DECL_USAGE_POSITION:
2106 shader_addline(buffer, "position%u", semantic->usage_idx);
2107 break;
2108
2109 case WINED3D_DECL_USAGE_BLEND_INDICES:
2110 shader_addline(buffer, "blend");
2111 break;
2112
2113 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2114 shader_addline(buffer, "weight");
2115 break;
2116
2117 case WINED3D_DECL_USAGE_NORMAL:
2118 shader_addline(buffer, "normal%u", semantic->usage_idx);
2119 break;
2120
2121 case WINED3D_DECL_USAGE_PSIZE:
2122 shader_addline(buffer, "psize");
2123 break;
2124
2125 case WINED3D_DECL_USAGE_COLOR:
2126 if (!semantic->usage_idx)
2127 shader_addline(buffer, "color");
2128 else
2129 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2130 break;
2131
2132 case WINED3D_DECL_USAGE_TEXCOORD:
2133 shader_addline(buffer, "texture%u", semantic->usage_idx);
2134 break;
2135
2136 case WINED3D_DECL_USAGE_TANGENT:
2137 shader_addline(buffer, "tangent");
2138 break;
2139
2140 case WINED3D_DECL_USAGE_BINORMAL:
2141 shader_addline(buffer, "binormal");
2142 break;
2143
2144 case WINED3D_DECL_USAGE_TESS_FACTOR:
2145 shader_addline(buffer, "tessfactor");
2146 break;
2147
2148 case WINED3D_DECL_USAGE_POSITIONT:
2149 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2150 break;
2151
2152 case WINED3D_DECL_USAGE_FOG:
2153 shader_addline(buffer, "fog");
2154 break;
2155
2156 case WINED3D_DECL_USAGE_DEPTH:
2157 shader_addline(buffer, "depth");
2158 break;
2159
2160 case WINED3D_DECL_USAGE_SAMPLE:
2161 shader_addline(buffer, "sample");
2162 break;
2163
2164 default:
2165 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2166 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2167 }
2168 }
2169 }
2170
2171 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2172 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2173 {
2174 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2175 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2176 UINT offset = reg->idx[0].offset;
2177
2178 switch (reg->type)
2179 {
2180 case WINED3DSPR_TEMP:
2181 shader_addline(buffer, "r");
2182 break;
2183
2184 case WINED3DSPR_INPUT:
2185 shader_addline(buffer, "v");
2186 break;
2187
2188 case WINED3DSPR_CONST:
2189 case WINED3DSPR_CONST2:
2190 case WINED3DSPR_CONST3:
2191 case WINED3DSPR_CONST4:
2192 shader_addline(buffer, "c");
2193 offset = shader_get_float_offset(reg->type, offset);
2194 break;
2195
2196 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2197 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2198 break;
2199
2200 case WINED3DSPR_RASTOUT:
2201 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2202 break;
2203
2204 case WINED3DSPR_COLOROUT:
2205 shader_addline(buffer, "oC");
2206 break;
2207
2208 case WINED3DSPR_DEPTHOUT:
2209 shader_addline(buffer, "oDepth");
2210 break;
2211
2212 case WINED3DSPR_DEPTHOUTGE:
2213 shader_addline(buffer, "oDepthGE");
2214 break;
2215
2216 case WINED3DSPR_DEPTHOUTLE:
2217 shader_addline(buffer, "oDepthLE");
2218 break;
2219
2220 case WINED3DSPR_ATTROUT:
2221 shader_addline(buffer, "oD");
2222 break;
2223
2224 case WINED3DSPR_TEXCRDOUT:
2225 /* Vertex shaders >= 3.0 use general purpose output registers
2226 * (WINED3DSPR_OUTPUT), which can include an address token. */
2227 if (shader_version->major >= 3)
2228 shader_addline(buffer, "o");
2229 else
2230 shader_addline(buffer, "oT");
2231 break;
2232
2233 case WINED3DSPR_CONSTINT:
2234 shader_addline(buffer, "i");
2235 break;
2236
2237 case WINED3DSPR_CONSTBOOL:
2238 shader_addline(buffer, "b");
2239 break;
2240
2241 case WINED3DSPR_LABEL:
2242 shader_addline(buffer, "l");
2243 break;
2244
2245 case WINED3DSPR_LOOP:
2246 shader_addline(buffer, "aL");
2247 break;
2248
2249 case WINED3DSPR_SAMPLER:
2250 shader_addline(buffer, "s");
2251 break;
2252
2253 case WINED3DSPR_MISCTYPE:
2254 if (offset > 1)
2255 {
2256 FIXME("Unhandled misctype register %u.\n", offset);
2257 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2258 }
2259 else
2260 {
2261 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2262 }
2263 break;
2264
2265 case WINED3DSPR_PREDICATE:
2266 shader_addline(buffer, "p");
2267 break;
2268
2269 case WINED3DSPR_IMMCONST:
2270 shader_addline(buffer, "l");
2271 break;
2272
2273 case WINED3DSPR_CONSTBUFFER:
2274 shader_addline(buffer, "cb");
2275 break;
2276
2277 case WINED3DSPR_IMMCONSTBUFFER:
2278 shader_addline(buffer, "icb");
2279 break;
2280
2281 case WINED3DSPR_PRIMID:
2282 shader_addline(buffer, "primID");
2283 break;
2284
2285 case WINED3DSPR_NULL:
2286 shader_addline(buffer, "null");
2287 break;
2288
2289 case WINED3DSPR_RESOURCE:
2290 shader_addline(buffer, "t");
2291 break;
2292
2293 case WINED3DSPR_UAV:
2294 shader_addline(buffer, "u");
2295 break;
2296
2297 case WINED3DSPR_OUTPOINTID:
2298 shader_addline(buffer, "vOutputControlPointID");
2299 break;
2300
2301 case WINED3DSPR_FORKINSTID:
2302 shader_addline(buffer, "vForkInstanceId");
2303 break;
2304
2305 case WINED3DSPR_JOININSTID:
2306 shader_addline(buffer, "vJoinInstanceId");
2307 break;
2308
2309 case WINED3DSPR_INCONTROLPOINT:
2310 shader_addline(buffer, "vicp");
2311 break;
2312
2313 case WINED3DSPR_OUTCONTROLPOINT:
2314 shader_addline(buffer, "vocp");
2315 break;
2316
2317 case WINED3DSPR_PATCHCONST:
2318 shader_addline(buffer, "vpc");
2319 break;
2320
2321 case WINED3DSPR_TESSCOORD:
2322 shader_addline(buffer, "vDomainLocation");
2323 break;
2324
2325 case WINED3DSPR_GROUPSHAREDMEM:
2326 shader_addline(buffer, "g");
2327 break;
2328
2329 case WINED3DSPR_THREADID:
2330 shader_addline(buffer, "vThreadID");
2331 break;
2332
2333 case WINED3DSPR_THREADGROUPID:
2334 shader_addline(buffer, "vThreadGroupID");
2335 break;
2336
2337 case WINED3DSPR_LOCALTHREADID:
2338 shader_addline(buffer, "vThreadIDInGroup");
2339 break;
2340
2341 case WINED3DSPR_LOCALTHREADINDEX:
2342 shader_addline(buffer, "vThreadIDInGroupFlattened");
2343 break;
2344
2345 case WINED3DSPR_IDXTEMP:
2346 shader_addline(buffer, "x");
2347 break;
2348
2349 case WINED3DSPR_STREAM:
2350 shader_addline(buffer, "m");
2351 break;
2352
2353 case WINED3DSPR_FUNCTIONBODY:
2354 shader_addline(buffer, "fb");
2355 break;
2356
2357 case WINED3DSPR_FUNCTIONPOINTER:
2358 shader_addline(buffer, "fp");
2359 break;
2360
2361 case WINED3DSPR_COVERAGE:
2362 shader_addline(buffer, "vCoverage");
2363 break;
2364
2365 case WINED3DSPR_SAMPLEMASK:
2366 shader_addline(buffer, "oMask");
2367 break;
2368
2369 case WINED3DSPR_GSINSTID:
2370 shader_addline(buffer, "vGSInstanceID");
2371 break;
2372
2373 default:
2374 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2375 break;
2376 }
2377
2378 if (reg->type == WINED3DSPR_IMMCONST)
2379 {
2380 shader_addline(buffer, "(");
2381 switch (reg->immconst_type)
2382 {
2383 case WINED3D_IMMCONST_SCALAR:
2384 switch (reg->data_type)
2385 {
2386 case WINED3D_DATA_FLOAT:
2387 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2388 break;
2389 case WINED3D_DATA_INT:
2390 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2391 break;
2392 case WINED3D_DATA_RESOURCE:
2393 case WINED3D_DATA_SAMPLER:
2394 case WINED3D_DATA_UINT:
2395 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2396 break;
2397 default:
2398 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2399 break;
2400 }
2401 break;
2402
2403 case WINED3D_IMMCONST_VEC4:
2404 switch (reg->data_type)
2405 {
2406 case WINED3D_DATA_FLOAT:
2407 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2408 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2409 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2410 break;
2411 case WINED3D_DATA_INT:
2412 shader_addline(buffer, "%d, %d, %d, %d",
2413 reg->u.immconst_data[0], reg->u.immconst_data[1],
2414 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2415 break;
2416 case WINED3D_DATA_RESOURCE:
2417 case WINED3D_DATA_SAMPLER:
2418 case WINED3D_DATA_UINT:
2419 shader_addline(buffer, "%u, %u, %u, %u",
2420 reg->u.immconst_data[0], reg->u.immconst_data[1],
2421 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2422 break;
2423 default:
2424 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2425 break;
2426 }
2427 break;
2428
2429 default:
2430 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2431 break;
2432 }
2433 shader_addline(buffer, ")");
2434 }
2435 else if (reg->type != WINED3DSPR_RASTOUT
2436 && reg->type != WINED3DSPR_MISCTYPE
2437 && reg->type != WINED3DSPR_NULL)
2438 {
2439 if (offset != ~0u)
2440 {
2441 shader_addline(buffer, "[");
2442 if (reg->idx[0].rel_addr)
2443 {
2444 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2445 shader_addline(buffer, " + ");
2446 }
2447 shader_addline(buffer, "%u]", offset);
2448
2449 if (reg->idx[1].offset != ~0u)
2450 {
2451 shader_addline(buffer, "[");
2452 if (reg->idx[1].rel_addr)
2453 {
2454 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2455 shader_addline(buffer, " + ");
2456 }
2457 shader_addline(buffer, "%u]", reg->idx[1].offset);
2458 }
2459 }
2460
2461 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2462 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2463 }
2464 }
2465
2466 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2467 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2468 {
2469 DWORD write_mask = param->write_mask;
2470
2471 shader_dump_register(buffer, &param->reg, shader_version);
2472
2473 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2474 {
2475 static const char write_mask_chars[] = "xyzw";
2476
2477 shader_addline(buffer, ".");
2478 if (write_mask & WINED3DSP_WRITEMASK_0)
2479 shader_addline(buffer, "%c", write_mask_chars[0]);
2480 if (write_mask & WINED3DSP_WRITEMASK_1)
2481 shader_addline(buffer, "%c", write_mask_chars[1]);
2482 if (write_mask & WINED3DSP_WRITEMASK_2)
2483 shader_addline(buffer, "%c", write_mask_chars[2]);
2484 if (write_mask & WINED3DSP_WRITEMASK_3)
2485 shader_addline(buffer, "%c", write_mask_chars[3]);
2486 }
2487 }
2488
2489 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2490 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2491 {
2492 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2493 DWORD swizzle = param->swizzle;
2494
2495 if (src_modifier == WINED3DSPSM_NEG
2496 || src_modifier == WINED3DSPSM_BIASNEG
2497 || src_modifier == WINED3DSPSM_SIGNNEG
2498 || src_modifier == WINED3DSPSM_X2NEG
2499 || src_modifier == WINED3DSPSM_ABSNEG)
2500 shader_addline(buffer, "-");
2501 else if (src_modifier == WINED3DSPSM_COMP)
2502 shader_addline(buffer, "1-");
2503 else if (src_modifier == WINED3DSPSM_NOT)
2504 shader_addline(buffer, "!");
2505
2506 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2507 shader_addline(buffer, "abs(");
2508
2509 shader_dump_register(buffer, &param->reg, shader_version);
2510
2511 switch (src_modifier)
2512 {
2513 case WINED3DSPSM_NONE: break;
2514 case WINED3DSPSM_NEG: break;
2515 case WINED3DSPSM_NOT: break;
2516 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2517 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2518 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2519 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2520 case WINED3DSPSM_COMP: break;
2521 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2522 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2523 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2524 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2525 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2526 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2527 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2528 }
2529
2530 if (swizzle != WINED3DSP_NOSWIZZLE)
2531 {
2532 static const char swizzle_chars[] = "xyzw";
2533 DWORD swizzle_x = swizzle & 0x03;
2534 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2535 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2536 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2537
2538 if (swizzle_x == swizzle_y
2539 && swizzle_x == swizzle_z
2540 && swizzle_x == swizzle_w)
2541 {
2542 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2543 }
2544 else
2545 {
2546 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2547 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2548 }
2549 }
2550 }
2551
2552 /* Shared code in order to generate the bulk of the shader string. */
2553 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2554 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2555 const DWORD *start, const DWORD *end)
2556 {
2557 struct wined3d_device *device = shader->device;
2558 const struct wined3d_shader_frontend *fe = shader->frontend;
2559 void *fe_data = shader->frontend_data;
2560 struct wined3d_shader_version shader_version;
2561 struct wined3d_shader_parser_state state;
2562 struct wined3d_shader_instruction ins;
2563 struct wined3d_shader_tex_mx tex_mx;
2564 struct wined3d_shader_context ctx;
2565 const DWORD *ptr;
2566
2567 /* Initialize current parsing state. */
2568 tex_mx.current_row = 0;
2569 state.current_loop_depth = 0;
2570 state.current_loop_reg = 0;
2571 state.in_subroutine = FALSE;
2572
2573 ctx.shader = shader;
2574 ctx.gl_info = &device->adapter->gl_info;
2575 ctx.reg_maps = reg_maps;
2576 ctx.buffer = buffer;
2577 ctx.tex_mx = &tex_mx;
2578 ctx.state = &state;
2579 ctx.backend_data = backend_ctx;
2580 ins.ctx = &ctx;
2581
2582 fe->shader_read_header(fe_data, &ptr, &shader_version);
2583 if (start)
2584 ptr = start;
2585
2586 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2587 {
2588 /* Read opcode. */
2589 fe->shader_read_instruction(fe_data, &ptr, &ins);
2590
2591 /* Unknown opcode and its parameters. */
2592 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2593 {
2594 WARN("Encountered unrecognised or invalid instruction.\n");
2595 return WINED3DERR_INVALIDCALL;
2596 }
2597
2598 if (ins.predicate)
2599 FIXME("Predicates not implemented.\n");
2600
2601 /* Call appropriate function for output target */
2602 device->shader_backend->shader_handle_instruction(&ins);
2603 }
2604
2605 return WINED3D_OK;
2606 }
2607
2608 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2609 const struct wined3d_shader_dst_param *dst)
2610 {
2611 DWORD mmask = dst->modifiers;
2612
2613 switch (dst->shift)
2614 {
2615 case 0: break;
2616 case 13: shader_addline(buffer, "_d8"); break;
2617 case 14: shader_addline(buffer, "_d4"); break;
2618 case 15: shader_addline(buffer, "_d2"); break;
2619 case 1: shader_addline(buffer, "_x2"); break;
2620 case 2: shader_addline(buffer, "_x4"); break;
2621 case 3: shader_addline(buffer, "_x8"); break;
2622 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2623 }
2624
2625 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2626 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2627 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2628
2629 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2630 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2631 }
2632
2633 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2634 const struct wined3d_shader_primitive_type *primitive_type)
2635 {
2636 switch (primitive_type->type)
2637 {
2638 case WINED3D_PT_UNDEFINED:
2639 shader_addline(buffer, "undefined");
2640 break;
2641 case WINED3D_PT_POINTLIST:
2642 shader_addline(buffer, "pointlist");
2643 break;
2644 case WINED3D_PT_LINELIST:
2645 shader_addline(buffer, "linelist");
2646 break;
2647 case WINED3D_PT_LINESTRIP:
2648 shader_addline(buffer, "linestrip");
2649 break;
2650 case WINED3D_PT_TRIANGLELIST:
2651 shader_addline(buffer, "trianglelist");
2652 break;
2653 case WINED3D_PT_TRIANGLESTRIP:
2654 shader_addline(buffer, "trianglestrip");
2655 break;
2656 case WINED3D_PT_TRIANGLEFAN:
2657 shader_addline(buffer, "trianglefan");
2658 break;
2659 case WINED3D_PT_LINELIST_ADJ:
2660 shader_addline(buffer, "linelist_adj");
2661 break;
2662 case WINED3D_PT_LINESTRIP_ADJ:
2663 shader_addline(buffer, "linestrip_adj");
2664 break;
2665 case WINED3D_PT_TRIANGLELIST_ADJ:
2666 shader_addline(buffer, "trianglelist_adj");
2667 break;
2668 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2669 shader_addline(buffer, "trianglestrip_adj");
2670 break;
2671 case WINED3D_PT_PATCH:
2672 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2673 break;
2674 default:
2675 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2676 break;
2677 }
2678 }
2679
2680 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2681 enum wined3d_shader_interpolation_mode interpolation_mode)
2682 {
2683 switch (interpolation_mode)
2684 {
2685 case WINED3DSIM_CONSTANT:
2686 shader_addline(buffer, "constant");
2687 break;
2688 case WINED3DSIM_LINEAR:
2689 shader_addline(buffer, "linear");
2690 break;
2691 case WINED3DSIM_LINEAR_CENTROID:
2692 shader_addline(buffer, "linear centroid");
2693 break;
2694 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2695 shader_addline(buffer, "linear noperspective");
2696 break;
2697 case WINED3DSIM_LINEAR_SAMPLE:
2698 shader_addline(buffer, "linear sample");
2699 break;
2700 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2701 shader_addline(buffer, "linear noperspective centroid");
2702 break;
2703 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2704 shader_addline(buffer, "linear noperspective sample");
2705 break;
2706 default:
2707 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2708 break;
2709 }
2710 }
2711
2712 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2713 {
2714 struct wined3d_shader_version shader_version;
2715 struct wined3d_string_buffer buffer;
2716 const char *type_prefix;
2717 const char *p, *q;
2718 const DWORD *ptr;
2719 DWORD i;
2720
2721 if (!string_buffer_init(&buffer))
2722 {
2723 ERR("Failed to initialize string buffer.\n");
2724 return;
2725 }
2726
2727 fe->shader_read_header(fe_data, &ptr, &shader_version);
2728
2729 TRACE("Parsing %p.\n", ptr);
2730
2731 switch (shader_version.type)
2732 {
2733 case WINED3D_SHADER_TYPE_VERTEX:
2734 type_prefix = "vs";
2735 break;
2736
2737 case WINED3D_SHADER_TYPE_HULL:
2738 type_prefix = "hs";
2739 break;
2740
2741 case WINED3D_SHADER_TYPE_DOMAIN:
2742 type_prefix = "ds";
2743 break;
2744
2745 case WINED3D_SHADER_TYPE_GEOMETRY:
2746 type_prefix = "gs";
2747 break;
2748
2749 case WINED3D_SHADER_TYPE_PIXEL:
2750 type_prefix = "ps";
2751 break;
2752
2753 case WINED3D_SHADER_TYPE_COMPUTE:
2754 type_prefix = "cs";
2755 break;
2756
2757 default:
2758 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2759 type_prefix = "unknown";
2760 break;
2761 }
2762
2763 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2764
2765 while (!fe->shader_is_end(fe_data, &ptr))
2766 {
2767 struct wined3d_shader_instruction ins;
2768
2769 fe->shader_read_instruction(fe_data, &ptr, &ins);
2770 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2771 {
2772 WARN("Skipping unrecognized instruction.\n");
2773 shader_addline(&buffer, "<unrecognized instruction>\n");
2774 continue;
2775 }
2776
2777 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2778 {
2779 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2780 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2781 shader_addline(&buffer, " ");
2782 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2783 }
2784 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2785 {
2786 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2787 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2788 shader_addline(&buffer, ", %s",
2789 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2790 }
2791 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2792 {
2793 shader_addline(&buffer, "%s fb%u",
2794 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2795 }
2796 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2797 {
2798 shader_addline(&buffer, "%s ft%u = {...}",
2799 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2800 }
2801 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2802 {
2803 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2804 shader_dump_global_flags(&buffer, ins.flags);
2805 }
2806 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2807 {
2808 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2809 ins.declaration.max_tessellation_factor);
2810 }
2811 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2812 {
2813 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2814 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2815 {
2816 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2817 ins.declaration.icb->data[4 * i + 0],
2818 ins.declaration.icb->data[4 * i + 1],
2819 ins.declaration.icb->data[4 * i + 2],
2820 ins.declaration.icb->data[4 * i + 3]);
2821 }
2822 shader_addline(&buffer, "}");
2823 }
2824 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2825 {
2826 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2827 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2828 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2829 }
2830 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2831 {
2832 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2833 ins.declaration.indexable_temp.register_idx,
2834 ins.declaration.indexable_temp.register_size,
2835 ins.declaration.indexable_temp.component_count);
2836 }
2837 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2838 {
2839 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2840 shader_dump_interpolation_mode(&buffer, ins.flags);
2841 shader_addline(&buffer, " ");
2842 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2843 }
2844 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2845 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2846 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2847 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2848 {
2849 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2850 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2851 shader_addline(&buffer, ", ");
2852 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2853 }
2854 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2855 {
2856 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2857 shader_dump_interpolation_mode(&buffer, ins.flags);
2858 shader_addline(&buffer, " ");
2859 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2860 shader_addline(&buffer, ", ");
2861 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2862 }
2863 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2864 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2865 {
2866 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2867 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2868 }
2869 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2870 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2871 {
2872 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2873 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2874 }
2875 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2876 {
2877 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2878 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2879 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2880 }
2881 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2882 {
2883 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2884 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2885 }
2886 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2887 {
2888 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2889 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2890 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2891 }
2892 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2893 {
2894 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2895 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2896 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2897 shader_addline(&buffer, ", comparisonMode");
2898 }
2899 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2900 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2901 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2902 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2903 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2904 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2905 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2906 {
2907 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2908 }
2909 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2910 {
2911 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2912 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2913 }
2914 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2915 {
2916 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2917 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2918 }
2919 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2920 {
2921 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2922 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2923 }
2924 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2925 {
2926 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2927 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2928 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2929 }
2930 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2931 {
2932 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2933 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2934 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2935 ins.declaration.tgsm_structured.structure_count);
2936 }
2937 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2938 {
2939 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2940 ins.declaration.thread_group_size.x,
2941 ins.declaration.thread_group_size.y,
2942 ins.declaration.thread_group_size.z);
2943 }
2944 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2945 {
2946 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2947 shader_dump_uav_flags(&buffer, ins.flags);
2948 shader_addline(&buffer, " ");
2949 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2950 }
2951 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2952 {
2953 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2954 shader_dump_uav_flags(&buffer, ins.flags);
2955 shader_addline(&buffer, " ");
2956 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2957 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2958 }
2959 else if (ins.handler_idx == WINED3DSIH_DEF)
2960 {
2961 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2962 ins.dst[0].reg.idx[0].offset),
2963 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2964 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2965 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2966 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2967 }
2968 else if (ins.handler_idx == WINED3DSIH_DEFI)
2969 {
2970 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2971 ins.src[0].reg.u.immconst_data[0],
2972 ins.src[0].reg.u.immconst_data[1],
2973 ins.src[0].reg.u.immconst_data[2],
2974 ins.src[0].reg.u.immconst_data[3]);
2975 }
2976 else if (ins.handler_idx == WINED3DSIH_DEFB)
2977 {
2978 shader_addline(&buffer, "defb b%u = %s",
2979 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2980 }
2981 else
2982 {
2983 if (ins.predicate)
2984 {
2985 shader_addline(&buffer, "(");
2986 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2987 shader_addline(&buffer, ") ");
2988 }
2989
2990 /* PixWin marks instructions with the coissue flag with a '+' */
2991 if (ins.coissue)
2992 shader_addline(&buffer, "+");
2993
2994 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2995
2996 if (ins.handler_idx == WINED3DSIH_BREAKP
2997 || ins.handler_idx == WINED3DSIH_CONTINUEP
2998 || ins.handler_idx == WINED3DSIH_IF
2999 || ins.handler_idx == WINED3DSIH_RETP
3000 || ins.handler_idx == WINED3DSIH_TEXKILL)
3001 {
3002 switch (ins.flags)
3003 {
3004 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3005 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3006 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3007 }
3008 }
3009 else if (ins.handler_idx == WINED3DSIH_IFC
3010 || ins.handler_idx == WINED3DSIH_BREAKC)
3011 {
3012 switch (ins.flags)
3013 {
3014 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3015 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3016 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3017 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3018 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3019 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3020 default: shader_addline(&buffer, "_(%u)", ins.flags);
3021 }
3022 }
3023 else if (ins.handler_idx == WINED3DSIH_TEX
3024 && shader_version.major >= 2
3025 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3026 {
3027 shader_addline(&buffer, "p");
3028 }
3029 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3030 {
3031 switch (ins.flags)
3032 {
3033 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3034 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3035 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3036 }
3037 }
3038 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3039 {
3040 switch (ins.flags)
3041 {
3042 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3043 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3044 }
3045 }
3046 else if (ins.handler_idx == WINED3DSIH_SYNC)
3047 {
3048 shader_dump_sync_flags(&buffer, ins.flags);
3049 }
3050
3051 if (wined3d_shader_instruction_has_texel_offset(&ins))
3052 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3053
3054 for (i = 0; i < ins.dst_count; ++i)
3055 {
3056 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
3057 shader_addline(&buffer, !i ? " " : ", ");
3058 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
3059 }
3060
3061 /* Other source tokens */
3062 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
3063 {
3064 shader_addline(&buffer, !i ? " " : ", ");
3065 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
3066 }
3067 }
3068 shader_addline(&buffer, "\n");
3069 }
3070
3071 for (p = buffer.buffer; *p; p = q)
3072 {
3073 if (!(q = strstr(p, "\n")))
3074 q = p + strlen(p);
3075 else
3076 ++q;
3077 TRACE(" %.*s", (int)(q - p), p);
3078 }
3079
3080 string_buffer_free(&buffer);
3081 }
3082
3083 static void shader_cleanup(struct wined3d_shader *shader)
3084 {
3085 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3086 {
3087 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
3088 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.fork);
3089 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.join);
3090 }
3091 else if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
3092 {
3093 HeapFree(GetProcessHeap(), 0, shader->u.gs.so_desc.elements);
3094 }
3095
3096 HeapFree(GetProcessHeap(), 0, shader->patch_constant_signature.elements);
3097 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3098 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3099 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3100 shader->device->shader_backend->shader_destroy(shader);
3101 shader_cleanup_reg_maps(&shader->reg_maps);
3102 HeapFree(GetProcessHeap(), 0, shader->function);
3103 shader_delete_constant_list(&shader->constantsF);
3104 shader_delete_constant_list(&shader->constantsB);
3105 shader_delete_constant_list(&shader->constantsI);
3106 list_remove(&shader->shader_list_entry);
3107
3108 if (shader->frontend && shader->frontend_data)
3109 shader->frontend->shader_free(shader->frontend_data);
3110 }
3111
3112 struct shader_none_priv
3113 {
3114 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3115 const struct fragment_pipeline *fragment_pipe;
3116 BOOL ffp_proj_control;
3117 };
3118
3119 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3120 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3121 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3122 const struct wined3d_state *state) {}
3123 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3124 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3125 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3126 const struct wined3d_state *state) {}
3127 static void shader_none_destroy(struct wined3d_shader *shader) {}
3128 static void shader_none_free_context_data(struct wined3d_context *context) {}
3129 static void shader_none_init_context_state(struct wined3d_context *context) {}
3130
3131 /* Context activation is done by the caller. */
3132 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3133 const struct wined3d_state *state)
3134 {
3135 const struct wined3d_gl_info *gl_info = context->gl_info;
3136 struct shader_none_priv *priv = shader_priv;
3137
3138 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
3139 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
3140 }
3141
3142 /* Context activation is done by the caller. */
3143 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3144 {
3145 struct shader_none_priv *priv = shader_priv;
3146 const struct wined3d_gl_info *gl_info = context->gl_info;
3147
3148 priv->vertex_pipe->vp_enable(gl_info, FALSE);
3149 priv->fragment_pipe->enable_extension(gl_info, FALSE);
3150
3151 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3152 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3153 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3154 | (1u << WINED3D_SHADER_TYPE_HULL)
3155 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3156 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3157 }
3158
3159 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3160 const struct fragment_pipeline *fragment_pipe)
3161 {
3162 struct fragment_caps fragment_caps;
3163 void *vertex_priv, *fragment_priv;
3164 struct shader_none_priv *priv;
3165
3166 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
3167 return E_OUTOFMEMORY;
3168
3169 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3170 {
3171 ERR("Failed to initialize vertex pipe.\n");
3172 HeapFree(GetProcessHeap(), 0, priv);
3173 return E_FAIL;
3174 }
3175
3176 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3177 {
3178 ERR("Failed to initialize fragment pipe.\n");
3179 vertex_pipe->vp_free(device);
3180 HeapFree(GetProcessHeap(), 0, priv);
3181 return E_FAIL;
3182 }
3183
3184 priv->vertex_pipe = vertex_pipe;
3185 priv->fragment_pipe = fragment_pipe;
3186 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
3187 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3188
3189 device->vertex_priv = vertex_priv;
3190 device->fragment_priv = fragment_priv;
3191 device->shader_priv = priv;
3192
3193 return WINED3D_OK;
3194 }
3195
3196 static void shader_none_free(struct wined3d_device *device)
3197 {
3198 struct shader_none_priv *priv = device->shader_priv;
3199
3200 priv->fragment_pipe->free_private(device);
3201 priv->vertex_pipe->vp_free(device);
3202 HeapFree(GetProcessHeap(), 0, priv);
3203 }
3204
3205 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3206 {
3207 return TRUE;
3208 }
3209
3210 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
3211 {
3212 /* Set the shader caps to 0 for the none shader backend */
3213 caps->vs_version = 0;
3214 caps->hs_version = 0;
3215 caps->ds_version = 0;
3216 caps->gs_version = 0;
3217 caps->ps_version = 0;
3218 caps->cs_version = 0;
3219 caps->vs_uniform_count = 0;
3220 caps->ps_uniform_count = 0;
3221 caps->ps_1x_max_value = 0.0f;
3222 caps->varying_count = 0;
3223 caps->wined3d_caps = 0;
3224 }
3225
3226 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3227 {
3228 /* We "support" every possible fixup, since we don't support any shader
3229 * model, and will never have to actually sample a texture. */
3230 return TRUE;
3231 }
3232
3233 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3234 {
3235 struct shader_none_priv *priv = shader_priv;
3236
3237 return priv->ffp_proj_control;
3238 }
3239
3240 const struct wined3d_shader_backend_ops none_shader_backend =
3241 {
3242 shader_none_handle_instruction,
3243 shader_none_precompile,
3244 shader_none_select,
3245 shader_none_select_compute,
3246 shader_none_disable,
3247 shader_none_update_float_vertex_constants,
3248 shader_none_update_float_pixel_constants,
3249 shader_none_load_constants,
3250 shader_none_destroy,
3251 shader_none_alloc,
3252 shader_none_free,
3253 shader_none_allocate_context_data,
3254 shader_none_free_context_data,
3255 shader_none_init_context_state,
3256 shader_none_get_caps,
3257 shader_none_color_fixup_supported,
3258 shader_none_has_ffp_proj_control,
3259 };
3260
3261 static HRESULT shader_set_function(struct wined3d_shader *shader, DWORD float_const_count,
3262 enum wined3d_shader_type type, unsigned int max_version)
3263 {
3264 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3265 const struct wined3d_shader_frontend *fe;
3266 HRESULT hr;
3267 unsigned int backend_version;
3268 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3269
3270 TRACE("shader %p, float_const_count %u, type %#x, max_version %u.\n",
3271 shader, float_const_count, type, max_version);
3272
3273 fe = shader->frontend;
3274 if (!(shader->frontend_data = fe->shader_init(shader->function,
3275 shader->functionLength, &shader->output_signature)))
3276 {
3277 FIXME("Failed to initialize frontend.\n");
3278 return WINED3DERR_INVALIDCALL;
3279 }
3280
3281 /* First pass: trace shader. */
3282 if (TRACE_ON(d3d_shader))
3283 shader_trace_init(fe, shader->frontend_data);
3284
3285 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3286 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
3287 &shader->output_signature, float_const_count)))
3288 return hr;
3289
3290 if (reg_maps->shader_version.type != type)
3291 {
3292 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
3293 return WINED3DERR_INVALIDCALL;
3294 }
3295 if (reg_maps->shader_version.major > max_version)
3296 {
3297 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
3298 return WINED3DERR_INVALIDCALL;
3299 }
3300 switch (type)
3301 {
3302 case WINED3D_SHADER_TYPE_VERTEX:
3303 backend_version = d3d_info->limits.vs_version;
3304 break;
3305 case WINED3D_SHADER_TYPE_HULL:
3306 backend_version = d3d_info->limits.hs_version;
3307 break;
3308 case WINED3D_SHADER_TYPE_DOMAIN:
3309 backend_version = d3d_info->limits.ds_version;
3310 break;
3311 case WINED3D_SHADER_TYPE_GEOMETRY:
3312 backend_version = d3d_info->limits.gs_version;
3313 break;
3314 case WINED3D_SHADER_TYPE_PIXEL:
3315 backend_version = d3d_info->limits.ps_version;
3316 break;
3317 case WINED3D_SHADER_TYPE_COMPUTE:
3318 backend_version = d3d_info->limits.cs_version;
3319 break;
3320 default:
3321 FIXME("No backend version-checking for this shader type.\n");
3322 backend_version = 0;
3323 }
3324 if (reg_maps->shader_version.major > backend_version)
3325 {
3326 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
3327 reg_maps->shader_version.major, reg_maps->shader_version.minor);
3328 return WINED3DERR_INVALIDCALL;
3329 }
3330
3331 return WINED3D_OK;
3332 }
3333
3334 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3335 {
3336 ULONG refcount = InterlockedIncrement(&shader->ref);
3337
3338 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3339
3340 return refcount;
3341 }
3342
3343 static void wined3d_shader_init_object(void *object)
3344 {
3345 struct wined3d_shader *shader = object;
3346 struct wined3d_device *device = shader->device;
3347
3348 list_add_head(&device->shaders, &shader->shader_list_entry);
3349
3350 device->shader_backend->shader_precompile(device->shader_priv, shader);
3351 }
3352
3353 static void wined3d_shader_destroy_object(void *object)
3354 {
3355 shader_cleanup(object);
3356 HeapFree(GetProcessHeap(), 0, object);
3357 }
3358
3359 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3360 {
3361 ULONG refcount = InterlockedDecrement(&shader->ref);
3362
3363 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3364
3365 if (!refcount)
3366 {
3367 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3368 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3369 }
3370
3371 return refcount;
3372 }
3373
3374 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3375 {
3376 TRACE("shader %p.\n", shader);
3377
3378 return shader->parent;
3379 }
3380
3381 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3382 void *byte_code, UINT *byte_code_size)
3383 {
3384 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3385
3386 if (!byte_code)
3387 {
3388 *byte_code_size = shader->functionLength;
3389 return WINED3D_OK;
3390 }
3391
3392 if (*byte_code_size < shader->functionLength)
3393 {
3394 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3395 * than the required size we should write the required size and
3396 * return D3DERR_MOREDATA. That's not actually true. */
3397 return WINED3DERR_INVALIDCALL;
3398 }
3399
3400 memcpy(byte_code, shader->function, shader->functionLength);
3401
3402 return WINED3D_OK;
3403 }
3404
3405 /* Set local constants for d3d8 shaders. */
3406 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3407 UINT start_idx, const float *src_data, UINT count)
3408 {
3409 UINT end_idx = start_idx + count;
3410 UINT i;
3411
3412 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3413
3414 if (end_idx > shader->limits->constant_float)
3415 {
3416 WARN("end_idx %u > float constants limit %u.\n",
3417 end_idx, shader->limits->constant_float);
3418 end_idx = shader->limits->constant_float;
3419 }
3420
3421 for (i = start_idx; i < end_idx; ++i)
3422 {
3423 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
3424 float *value;
3425 if (!lconst)
3426 return E_OUTOFMEMORY;
3427
3428 lconst->idx = i;
3429 value = (float *)lconst->value;
3430 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3431 list_add_head(&shader->constantsF, &lconst->entry);
3432
3433 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3434 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3435 {
3436 shader->lconst_inf_or_nan = TRUE;
3437 }
3438 }
3439
3440 return WINED3D_OK;
3441 }
3442
3443 static void init_interpolation_compile_args(DWORD *interpolation_args,
3444 const struct wined3d_shader *pixel_shader, const struct wined3d_gl_info *gl_info)
3445 {
3446 if (!needs_interpolation_qualifiers_for_shader_outputs(gl_info)
3447 || !pixel_shader || pixel_shader->reg_maps.shader_version.major < 4)
3448 {
3449 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3450 return;
3451 }
3452
3453 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3454 sizeof(pixel_shader->u.ps.interpolation_mode));
3455 }
3456
3457 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3458 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_context *context)
3459 {
3460 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3461 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3462 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3463 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3464 const struct wined3d_gl_info *gl_info = context->gl_info;
3465
3466 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3467 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3468 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3469 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3470 args->point_size = state->gl_primitive_type == GL_POINTS;
3471 args->per_vertex_point_size = shader->reg_maps.point_size;
3472 args->next_shader_type = hull_shader? WINED3D_SHADER_TYPE_HULL
3473 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3474 if (shader->reg_maps.shader_version.major >= 4)
3475 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3476 : geometry_shader ? geometry_shader->limits->packed_input
3477 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3478 else
3479 args->next_shader_input_count = 0;
3480 args->swizzle_map = swizzle_map;
3481 if (d3d_info->emulated_flatshading)
3482 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3483 else
3484 args->flatshading = 0;
3485
3486 init_interpolation_compile_args(args->interpolation_mode,
3487 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, gl_info);
3488 }
3489
3490 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3491 {
3492 if (usage_idx1 != usage_idx2)
3493 return FALSE;
3494 if (usage1 == usage2)
3495 return TRUE;
3496 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3497 return TRUE;
3498 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3499 return TRUE;
3500
3501 return FALSE;
3502 }
3503
3504 BOOL vshader_get_input(const struct wined3d_shader *shader,
3505 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3506 {
3507 WORD map = shader->reg_maps.input_registers;
3508 unsigned int i;
3509
3510 for (i = 0; map; map >>= 1, ++i)
3511 {
3512 if (!(map & 1)) continue;
3513
3514 if (match_usage(shader->u.vs.attributes[i].usage,
3515 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3516 {
3517 *regnum = i;
3518 return TRUE;
3519 }
3520 }
3521 return FALSE;
3522 }
3523
3524 static HRESULT shader_signature_calculate_strings_length(const struct wined3d_shader_signature *signature,
3525 SIZE_T *total)
3526 {
3527 struct wined3d_shader_signature_element *e;
3528 unsigned int i;
3529 SIZE_T len;
3530
3531 for (i = 0; i < signature->element_count; ++i)
3532 {
3533 e = &signature->elements[i];
3534 len = strlen(e->semantic_name);
3535 if (len >= ~(SIZE_T)0 - *total)
3536 return E_OUTOFMEMORY;
3537
3538 *total += len + 1;
3539 }
3540 return WINED3D_OK;
3541 }
3542
3543 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3544 const struct wined3d_shader_signature *src, char **signature_strings)
3545 {
3546 struct wined3d_shader_signature_element *e;
3547 unsigned int i;
3548 SIZE_T len;
3549 char *ptr;
3550
3551 if (!src->element_count)
3552 return WINED3D_OK;
3553
3554 ptr = *signature_strings;
3555
3556 dst->element_count = src->element_count;
3557 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3558 return E_OUTOFMEMORY;
3559
3560 for (i = 0; i < src->element_count; ++i)
3561 {
3562 e = &src->elements[i];
3563 dst->elements[i] = *e;
3564
3565 len = strlen(e->semantic_name);
3566 memcpy(ptr, e->semantic_name, len + 1);
3567 dst->elements[i].semantic_name = ptr;
3568 ptr += len + 1;
3569 }
3570
3571 *signature_strings = ptr;
3572
3573 return WINED3D_OK;
3574 }
3575
3576 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3577 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3578 void *parent, const struct wined3d_parent_ops *parent_ops)
3579 {
3580 size_t byte_code_size;
3581 SIZE_T total;
3582 HRESULT hr;
3583 char *ptr;
3584
3585 TRACE("byte_code %p, byte_code_size %#lx, format %#x, max_version %#x.\n",
3586 desc->byte_code, (long)desc->byte_code_size, desc->format, desc->max_version);
3587
3588 if (!desc->byte_code)
3589 return WINED3DERR_INVALIDCALL;
3590
3591 if (!(shader->frontend = shader_select_frontend(desc->format)))
3592 {
3593 FIXME("Unable to find frontend for shader.\n");
3594 return WINED3DERR_INVALIDCALL;
3595 }
3596
3597 shader->ref = 1;
3598 shader->device = device;
3599 shader->parent = parent;
3600 shader->parent_ops = parent_ops;
3601
3602 total = 0;
3603 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->input_signature, &total)))
3604 return hr;
3605 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->output_signature, &total)))
3606 return hr;
3607 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->patch_constant_signature, &total)))
3608 return hr;
3609 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3610 return E_OUTOFMEMORY;
3611 ptr = shader->signature_strings;
3612
3613 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3614 {
3615 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3616 return hr;
3617 }
3618 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3619 {
3620 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3621 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3622 return hr;
3623 }
3624 if (FAILED(hr = shader_signature_copy(&shader->patch_constant_signature, &desc->patch_constant_signature, &ptr)))
3625 {
3626 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3627 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3628 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3629 return hr;
3630 }
3631
3632 list_init(&shader->linked_programs);
3633 list_init(&shader->constantsF);
3634 list_init(&shader->constantsB);
3635 list_init(&shader->constantsI);
3636 shader->lconst_inf_or_nan = FALSE;
3637 list_init(&shader->reg_maps.indexable_temps);
3638 list_init(&shader->shader_list_entry);
3639
3640 byte_code_size = desc->byte_code_size;
3641 if (byte_code_size == ~(size_t)0)
3642 {
3643 const struct wined3d_shader_frontend *fe = shader->frontend;
3644 struct wined3d_shader_version shader_version;
3645 struct wined3d_shader_instruction ins;
3646 const DWORD *ptr;
3647 void *fe_data;
3648
3649 if (!(fe_data = fe->shader_init(desc->byte_code, byte_code_size, &shader->output_signature)))
3650 {
3651 WARN("Failed to initialise frontend data.\n");
3652 shader_cleanup(shader);
3653 return WINED3DERR_INVALIDCALL;
3654 }
3655
3656 fe->shader_read_header(fe_data, &ptr, &shader_version);
3657 while (!fe->shader_is_end(fe_data, &ptr))
3658 fe->shader_read_instruction(fe_data, &ptr, &ins);
3659
3660 fe->shader_free(fe_data);
3661
3662 byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3663 }
3664
3665 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, byte_code_size)))
3666 {
3667 shader_cleanup(shader);
3668 return E_OUTOFMEMORY;
3669 }
3670 memcpy(shader->function, desc->byte_code, byte_code_size);
3671 shader->functionLength = byte_code_size;
3672
3673 if (FAILED(hr = shader_set_function(shader, float_const_count, type, desc->max_version)))
3674 {
3675 WARN("Failed to set function, hr %#x.\n", hr);
3676 shader_cleanup(shader);
3677 return hr;
3678 }
3679
3680 shader->load_local_constsF = shader->lconst_inf_or_nan;
3681
3682 wined3d_cs_init_object(shader->device->cs, wined3d_shader_init_object, shader);
3683
3684 return hr;
3685 }
3686
3687 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3688 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3689 {
3690 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3691 unsigned int i;
3692 HRESULT hr;
3693
3694 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3695 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3696 return hr;
3697
3698 for (i = 0; i < shader->input_signature.element_count; ++i)
3699 {
3700 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3701
3702 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3703 continue;
3704
3705 shader->u.vs.attributes[input->register_idx].usage =
3706 shader_usage_from_semantic_name(input->semantic_name);
3707 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3708 }
3709
3710 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3711 shader->load_local_constsF = TRUE;
3712
3713 return WINED3D_OK;
3714 }
3715
3716 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3717 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3718 void *parent, const struct wined3d_parent_ops *parent_ops)
3719 {
3720 struct wined3d_stream_output_element *elements = NULL;
3721 HRESULT hr;
3722
3723 if (so_desc && !(elements = wined3d_calloc(so_desc->element_count, sizeof(*elements))))
3724 return E_OUTOFMEMORY;
3725
3726 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
3727 {
3728 HeapFree(GetProcessHeap(), 0, elements);
3729 return hr;
3730 }
3731
3732 if (so_desc)
3733 {
3734 shader->u.gs.so_desc = *so_desc;
3735 shader->u.gs.so_desc.elements = elements;
3736 memcpy(elements, so_desc->elements, so_desc->element_count * sizeof(*elements));
3737 }
3738
3739 return WINED3D_OK;
3740 }
3741
3742 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3743 struct ds_compile_args *args, const struct wined3d_context *context)
3744 {
3745 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3746 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3747 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3748 const struct wined3d_gl_info *gl_info = context->gl_info;
3749
3750 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3751 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3752
3753 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3754 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3755 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3756
3757 args->render_offscreen = context->render_offscreen;
3758
3759 init_interpolation_compile_args(args->interpolation_mode,
3760 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, gl_info);
3761
3762 args->padding = 0;
3763 }
3764
3765 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3766 struct gs_compile_args *args, const struct wined3d_context *context)
3767 {
3768 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3769 const struct wined3d_gl_info *gl_info = context->gl_info;
3770
3771 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3772
3773 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, gl_info);
3774 }
3775
3776 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3777 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3778 {
3779 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3780 const struct wined3d_gl_info *gl_info = context->gl_info;
3781 const struct wined3d_texture *texture;
3782 unsigned int i;
3783
3784 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3785 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3786 {
3787 static unsigned int warned = 0;
3788
3789 args->srgb_correction = 1;
3790 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3791 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3792 "support, expect rendering artifacts.\n");
3793 }
3794
3795 if (shader->reg_maps.shader_version.major == 1
3796 && shader->reg_maps.shader_version.minor <= 3)
3797 {
3798 for (i = 0; i < shader->limits->sampler; ++i)
3799 {
3800 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3801
3802 if (flags & WINED3D_TTFF_PROJECTED)
3803 {
3804 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3805
3806 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3807 {
3808 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3809 unsigned int j;
3810 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3811 DWORD max_valid = WINED3D_TTFF_COUNT4;
3812
3813 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3814 {
3815 struct wined3d_vertex_declaration_element *element =
3816 &state->vertex_declaration->elements[j];
3817
3818 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3819 && element->usage_idx == index)
3820 {
3821 max_valid = element->format->component_count;
3822 break;
3823 }
3824 }
3825 if (!tex_transform || tex_transform > max_valid)
3826 {
3827 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3828 tex_transform, max_valid);
3829 tex_transform = max_valid;
3830 }
3831 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3832 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3833 && tex_transform > WINED3D_TTFF_COUNT2)
3834 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3835 && tex_transform > WINED3D_TTFF_COUNT3))
3836 tex_transform |= WINED3D_PSARGS_PROJECTED;
3837 else
3838 {
3839 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3840 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3841 i, tex_transform, resource_type);
3842 }
3843 }
3844 else
3845 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3846
3847 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3848 }
3849 }
3850 }
3851 if (shader->reg_maps.shader_version.major == 1
3852 && shader->reg_maps.shader_version.minor <= 4)
3853 {
3854 for (i = 0; i < shader->limits->sampler; ++i)
3855 {
3856 const struct wined3d_texture *texture = state->textures[i];
3857
3858 if (!shader->reg_maps.resource_info[i].type)
3859 continue;
3860
3861 /* Treat unbound textures as 2D. The dummy texture will provide
3862 * the proper sample value. The tex_types bitmap defaults to
3863 * 2D because of the memset. */
3864 if (!texture)
3865 continue;
3866
3867 switch (texture->target)
3868 {
3869 /* RECT textures are distinguished from 2D textures via np2_fixup */
3870 case GL_TEXTURE_RECTANGLE_ARB:
3871 case GL_TEXTURE_2D:
3872 break;
3873
3874 case GL_TEXTURE_3D:
3875 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3876 break;
3877
3878 case GL_TEXTURE_CUBE_MAP_ARB:
3879 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3880 break;
3881 }
3882 }
3883 }
3884
3885 if (shader->reg_maps.shader_version.major >= 4)
3886 {
3887 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3888 args->shadow = 0;
3889 for (i = 0 ; i < MAX_FRAGMENT_SAMPLERS; ++i)
3890 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3891 args->np2_fixup = 0;
3892 }
3893 else
3894 {
3895 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3896 {
3897 if (!shader->reg_maps.resource_info[i].type)
3898 continue;
3899
3900 texture = state->textures[i];
3901 if (!texture)
3902 {
3903 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3904 continue;
3905 }
3906 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3907 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3908 else
3909 args->color_fixup[i] = texture->resource.format->color_fixup;
3910
3911 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3912 args->shadow |= 1u << i;
3913
3914 /* Flag samplers that need NP2 texcoord fixup. */
3915 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3916 args->np2_fixup |= (1u << i);
3917 }
3918 }
3919
3920 if (shader->reg_maps.shader_version.major >= 3)
3921 {
3922 if (position_transformed)
3923 args->vp_mode = pretransformed;
3924 else if (use_vs(state))
3925 args->vp_mode = vertexshader;
3926 else
3927 args->vp_mode = fixedfunction;
3928 args->fog = WINED3D_FFP_PS_FOG_OFF;
3929 }
3930 else
3931 {
3932 args->vp_mode = vertexshader;
3933 if (state->render_states[WINED3D_RS_FOGENABLE])
3934 {
3935 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3936 {
3937 case WINED3D_FOG_NONE:
3938 if (position_transformed || use_vs(state))
3939 {
3940 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3941 break;
3942 }
3943
3944 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3945 {
3946 case WINED3D_FOG_NONE: /* Fall through. */
3947 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3948 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3949 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3950 }
3951 break;
3952
3953 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3954 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3955 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3956 }
3957 }
3958 else
3959 {
3960 args->fog = WINED3D_FFP_PS_FOG_OFF;
3961 }
3962 }
3963
3964 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3965 {
3966 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3967
3968 args->texcoords_initialized = 0;
3969 for (i = 0; i < MAX_TEXTURES; ++i)
3970 {
3971 if (vs)
3972 {
3973 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3974 args->texcoords_initialized |= 1u << i;
3975 }
3976 else
3977 {
3978 const struct wined3d_stream_info *si = &context->stream_info;
3979 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3980
3981 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3982 & WINED3D_FFP_TCI_MASK
3983 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3984 args->texcoords_initialized |= 1u << i;
3985 }
3986 }
3987 }
3988 else
3989 {
3990 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3991 }
3992
3993 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3994 && state->gl_primitive_type == GL_POINTS;
3995
3996 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3997 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3998 else
3999 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4000 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4001 : WINED3D_CMP_ALWAYS) - 1;
4002
4003 if (d3d_info->emulated_flatshading)
4004 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4005
4006 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
4007 ? context->render_offscreen : 0;
4008 }
4009
4010 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4011 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4012 {
4013 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4014 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4015 HRESULT hr;
4016
4017 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
4018 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
4019 return hr;
4020
4021 for (i = 0; i < MAX_REG_INPUT; ++i)
4022 {
4023 if (shader->u.ps.input_reg_used & (1u << i))
4024 {
4025 ++num_regs_used;
4026 highest_reg_used = i;
4027 }
4028 }
4029
4030 /* Don't do any register mapping magic if it is not needed, or if we can't
4031 * achieve anything anyway */
4032 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4033 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4034 || shader->reg_maps.shader_version.major >= 4)
4035 {
4036 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4037 {
4038 /* This happens with relative addressing. The input mapper function
4039 * warns about this if the higher registers are declared too, so
4040 * don't write a FIXME here */
4041 WARN("More varying registers used than supported\n");
4042 }
4043
4044 for (i = 0; i < MAX_REG_INPUT; ++i)
4045 {
4046 shader->u.ps.input_reg_map[i] = i;
4047 }
4048
4049 shader->u.ps.declared_in_count = highest_reg_used + 1;
4050 }
4051 else
4052 {
4053 shader->u.ps.declared_in_count = 0;
4054 for (i = 0; i < MAX_REG_INPUT; ++i)
4055 {
4056 if (shader->u.ps.input_reg_used & (1u << i))
4057 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4058 else shader->u.ps.input_reg_map[i] = ~0U;
4059 }
4060 }
4061
4062 return WINED3D_OK;
4063 }
4064
4065 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
4066 {
4067 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
4068 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
4069 unsigned int i;
4070
4071 if (reg_maps->shader_version.major != 1) return;
4072
4073 for (i = 0; i < shader->limits->sampler; ++i)
4074 {
4075 /* We don't sample from this sampler. */
4076 if (!resource_info[i].type)
4077 continue;
4078
4079 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
4080 {
4081 case WINED3D_SHADER_TEX_2D:
4082 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
4083 break;
4084
4085 case WINED3D_SHADER_TEX_3D:
4086 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
4087 break;
4088
4089 case WINED3D_SHADER_TEX_CUBE:
4090 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
4091 break;
4092 }
4093 }
4094 }
4095
4096 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4097 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4098 {
4099 struct wined3d_shader *object;
4100 HRESULT hr;
4101
4102 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4103 device, desc, parent, parent_ops, shader);
4104
4105 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4106 return E_OUTOFMEMORY;
4107
4108 if (FAILED(hr = shader_init(object, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops)))
4109 {
4110 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4111 HeapFree(GetProcessHeap(), 0, object);
4112 return hr;
4113 }
4114
4115 TRACE("Created compute shader %p.\n", object);
4116 *shader = object;
4117
4118 return WINED3D_OK;
4119 }
4120
4121 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4122 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4123 {
4124 struct wined3d_shader *object;
4125 HRESULT hr;
4126
4127 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4128 device, desc, parent, parent_ops, shader);
4129
4130 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4131 return E_OUTOFMEMORY;
4132
4133 if (FAILED(hr = shader_init(object, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops)))
4134 {
4135 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4136 HeapFree(GetProcessHeap(), 0, object);
4137 return hr;
4138 }
4139
4140 TRACE("Created domain shader %p.\n", object);
4141 *shader = object;
4142
4143 return WINED3D_OK;
4144 }
4145
4146 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4147 const struct wined3d_stream_output_desc *so_desc, void *parent,
4148 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4149 {
4150 struct wined3d_shader *object;
4151 HRESULT hr;
4152
4153 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4154 device, desc, so_desc, parent, parent_ops, shader);
4155
4156 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4157 return E_OUTOFMEMORY;
4158
4159 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4160 {
4161 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4162 HeapFree(GetProcessHeap(), 0, object);
4163 return hr;
4164 }
4165
4166 TRACE("Created geometry shader %p.\n", object);
4167 *shader = object;
4168
4169 return WINED3D_OK;
4170 }
4171
4172 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4173 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4174 {
4175 struct wined3d_shader *object;
4176 HRESULT hr;
4177
4178 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4179 device, desc, parent, parent_ops, shader);
4180
4181 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4182 return E_OUTOFMEMORY;
4183
4184 if (FAILED(hr = shader_init(object, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops)))
4185 {
4186 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4187 HeapFree(GetProcessHeap(), 0, object);
4188 return hr;
4189 }
4190
4191 TRACE("Created hull shader %p.\n", object);
4192 *shader = object;
4193
4194 return WINED3D_OK;
4195 }
4196
4197 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4198 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4199 {
4200 struct wined3d_shader *object;
4201 HRESULT hr;
4202
4203 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4204 device, desc, parent, parent_ops, shader);
4205
4206 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4207 if (!object)
4208 return E_OUTOFMEMORY;
4209
4210 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4211 {
4212 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4213 HeapFree(GetProcessHeap(), 0, object);
4214 return hr;
4215 }
4216
4217 TRACE("Created pixel shader %p.\n", object);
4218 *shader = object;
4219
4220 return WINED3D_OK;
4221 }
4222
4223 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4224 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4225 {
4226 struct wined3d_shader *object;
4227 HRESULT hr;
4228
4229 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4230 device, desc, parent, parent_ops, shader);
4231
4232 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4233 if (!object)
4234 return E_OUTOFMEMORY;
4235
4236 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4237 {
4238 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4239 HeapFree(GetProcessHeap(), 0, object);
4240 return hr;
4241 }
4242
4243 TRACE("Created vertex shader %p.\n", object);
4244 *shader = object;
4245
4246 return WINED3D_OK;
4247 }
Windows API implementation