?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dxgi: Added support for DXGI_SWAP_CHAIN_FLAG_GDI_COMPATIBLE.
[wine.git] / dlls / wined3d / shader.c
blob490c985acf1e63ad0bd914b663a55577428ae921
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <math.h>
29 #include <stdio.h>
30 #include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
47 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
48 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
49 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
50 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
51 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
52 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
53 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
54 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
55 /* WINED3DSIH_BEM */ "bem",
56 /* WINED3DSIH_BFI */ "bfi",
57 /* WINED3DSIH_BFREV */ "bfrev",
58 /* WINED3DSIH_BREAK */ "break",
59 /* WINED3DSIH_BREAKC */ "breakc",
60 /* WINED3DSIH_BREAKP */ "breakp",
61 /* WINED3DSIH_BUFINFO */ "bufinfo",
62 /* WINED3DSIH_CALL */ "call",
63 /* WINED3DSIH_CALLNZ */ "callnz",
64 /* WINED3DSIH_CASE */ "case",
65 /* WINED3DSIH_CMP */ "cmp",
66 /* WINED3DSIH_CND */ "cnd",
67 /* WINED3DSIH_CONTINUE */ "continue",
68 /* WINED3DSIH_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_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
78 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
79 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
80 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
81 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
82 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
83 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
84 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
85 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
86 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
87 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
88 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
89 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
90 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
91 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
92 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
93 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
94 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
95 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
96 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
97 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
98 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
99 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
100 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
101 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
102 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
103 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
104 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
105 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
106 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
107 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
108 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
109 /* WINED3DSIH_DEF */ "def",
110 /* WINED3DSIH_DEFAULT */ "default",
111 /* WINED3DSIH_DEFB */ "defb",
112 /* WINED3DSIH_DEFI */ "defi",
113 /* WINED3DSIH_DIV */ "div",
114 /* WINED3DSIH_DP2 */ "dp2",
115 /* WINED3DSIH_DP2ADD */ "dp2add",
116 /* WINED3DSIH_DP3 */ "dp3",
117 /* WINED3DSIH_DP4 */ "dp4",
118 /* WINED3DSIH_DST */ "dst",
119 /* WINED3DSIH_DSX */ "dsx",
120 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
121 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
122 /* WINED3DSIH_DSY */ "dsy",
123 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
124 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
125 /* WINED3DSIH_ELSE */ "else",
126 /* WINED3DSIH_EMIT */ "emit",
127 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
128 /* WINED3DSIH_ENDIF */ "endif",
129 /* WINED3DSIH_ENDLOOP */ "endloop",
130 /* WINED3DSIH_ENDREP */ "endrep",
131 /* WINED3DSIH_ENDSWITCH */ "endswitch",
132 /* WINED3DSIH_EQ */ "eq",
133 /* WINED3DSIH_EXP */ "exp",
134 /* WINED3DSIH_EXPP */ "expp",
135 /* WINED3DSIH_F16TOF32 */ "f16tof32",
136 /* WINED3DSIH_F32TOF16 */ "f32tof16",
137 /* WINED3DSIH_FCALL */ "fcall",
138 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
139 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
140 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
141 /* WINED3DSIH_FRC */ "frc",
142 /* WINED3DSIH_FTOI */ "ftoi",
143 /* WINED3DSIH_FTOU */ "ftou",
144 /* WINED3DSIH_GATHER4 */ "gather4",
145 /* WINED3DSIH_GATHER4_C */ "gather4_c",
146 /* WINED3DSIH_GE */ "ge",
147 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
148 /* WINED3DSIH_HS_DECLS */ "hs_decls",
149 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
150 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
151 /* WINED3DSIH_IADD */ "iadd",
152 /* WINED3DSIH_IEQ */ "ieq",
153 /* WINED3DSIH_IF */ "if",
154 /* WINED3DSIH_IFC */ "ifc",
155 /* WINED3DSIH_IGE */ "ige",
156 /* WINED3DSIH_ILT */ "ilt",
157 /* WINED3DSIH_IMAD */ "imad",
158 /* WINED3DSIH_IMAX */ "imax",
159 /* WINED3DSIH_IMIN */ "imin",
160 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
161 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
162 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
163 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
164 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
165 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
166 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
167 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
168 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
169 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
170 /* WINED3DSIH_IMUL */ "imul",
171 /* WINED3DSIH_INE */ "ine",
172 /* WINED3DSIH_INEG */ "ineg",
173 /* WINED3DSIH_ISHL */ "ishl",
174 /* WINED3DSIH_ISHR */ "ishr",
175 /* WINED3DSIH_ITOF */ "itof",
176 /* WINED3DSIH_LABEL */ "label",
177 /* WINED3DSIH_LD */ "ld",
178 /* WINED3DSIH_LD2DMS */ "ld2dms",
179 /* WINED3DSIH_LD_RAW */ "ld_raw",
180 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
181 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
182 /* WINED3DSIH_LIT */ "lit",
183 /* WINED3DSIH_LOD */ "lod",
184 /* WINED3DSIH_LOG */ "log",
185 /* WINED3DSIH_LOGP */ "logp",
186 /* WINED3DSIH_LOOP */ "loop",
187 /* WINED3DSIH_LRP */ "lrp",
188 /* WINED3DSIH_LT */ "lt",
189 /* WINED3DSIH_M3x2 */ "m3x2",
190 /* WINED3DSIH_M3x3 */ "m3x3",
191 /* WINED3DSIH_M3x4 */ "m3x4",
192 /* WINED3DSIH_M4x3 */ "m4x3",
193 /* WINED3DSIH_M4x4 */ "m4x4",
194 /* WINED3DSIH_MAD */ "mad",
195 /* WINED3DSIH_MAX */ "max",
196 /* WINED3DSIH_MIN */ "min",
197 /* WINED3DSIH_MOV */ "mov",
198 /* WINED3DSIH_MOVA */ "mova",
199 /* WINED3DSIH_MOVC */ "movc",
200 /* WINED3DSIH_MUL */ "mul",
201 /* WINED3DSIH_NE */ "ne",
202 /* WINED3DSIH_NOP */ "nop",
203 /* WINED3DSIH_NOT */ "not",
204 /* WINED3DSIH_NRM */ "nrm",
205 /* WINED3DSIH_OR */ "or",
206 /* WINED3DSIH_PHASE */ "phase",
207 /* WINED3DSIH_POW */ "pow",
208 /* WINED3DSIH_RCP */ "rcp",
209 /* WINED3DSIH_REP */ "rep",
210 /* WINED3DSIH_RESINFO */ "resinfo",
211 /* WINED3DSIH_RET */ "ret",
212 /* WINED3DSIH_ROUND_NE */ "round_ne",
213 /* WINED3DSIH_ROUND_NI */ "round_ni",
214 /* WINED3DSIH_ROUND_PI */ "round_pi",
215 /* WINED3DSIH_ROUND_Z */ "round_z",
216 /* WINED3DSIH_RSQ */ "rsq",
217 /* WINED3DSIH_SAMPLE */ "sample",
218 /* WINED3DSIH_SAMPLE_B */ "sample_b",
219 /* WINED3DSIH_SAMPLE_C */ "sample_c",
220 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
221 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
222 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
223 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
224 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
225 /* WINED3DSIH_SETP */ "setp",
226 /* WINED3DSIH_SGE */ "sge",
227 /* WINED3DSIH_SGN */ "sgn",
228 /* WINED3DSIH_SINCOS */ "sincos",
229 /* WINED3DSIH_SLT */ "slt",
230 /* WINED3DSIH_SQRT */ "sqrt",
231 /* WINED3DSIH_STORE_RAW */ "store_raw",
232 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
233 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
234 /* WINED3DSIH_SUB */ "sub",
235 /* WINED3DSIH_SWAPC */ "swapc",
236 /* WINED3DSIH_SWITCH */ "switch",
237 /* WINED3DSIH_SYNC */ "sync",
238 /* WINED3DSIH_TEX */ "texld",
239 /* WINED3DSIH_TEXBEM */ "texbem",
240 /* WINED3DSIH_TEXBEML */ "texbeml",
241 /* WINED3DSIH_TEXCOORD */ "texcrd",
242 /* WINED3DSIH_TEXDEPTH */ "texdepth",
243 /* WINED3DSIH_TEXDP3 */ "texdp3",
244 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
245 /* WINED3DSIH_TEXKILL */ "texkill",
246 /* WINED3DSIH_TEXLDD */ "texldd",
247 /* WINED3DSIH_TEXLDL */ "texldl",
248 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
249 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
250 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
251 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
252 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
253 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
254 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
255 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
256 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
257 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
258 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
259 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
260 /* WINED3DSIH_UBFE */ "ubfe",
261 /* WINED3DSIH_UDIV */ "udiv",
262 /* WINED3DSIH_UGE */ "uge",
263 /* WINED3DSIH_ULT */ "ult",
264 /* WINED3DSIH_UMAX */ "umax",
265 /* WINED3DSIH_UMIN */ "umin",
266 /* WINED3DSIH_USHR */ "ushr",
267 /* WINED3DSIH_UTOF */ "utof",
268 /* WINED3DSIH_XOR */ "xor",
269 };
270
271 static const char * const semantic_names[] =
272 {
273 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
274 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
275 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
276 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
277 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
278 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
279 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
280 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
281 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
282 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
283 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
284 /* WINED3D_DECL_USAGE_FOG */ "FOG",
285 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
286 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
287 };
288
289 static const struct
290 {
291 enum wined3d_sysval_semantic sysval_semantic;
292 const char *sysval_name;
293 }
294 sysval_semantic_names[] =
295 {
296 {WINED3D_SV_POSITION, "SV_Position"},
297 {WINED3D_SV_CLIP_DISTANCE, "SV_ClipDistance"},
298 {WINED3D_SV_CULL_DISTANCE, "SV_CullDistance"},
299 {WINED3D_SV_RENDER_TARGET_ARRAY_INDEX, "SV_RenderTargetArrayIndex"},
300 {WINED3D_SV_VIEWPORT_ARRAY_INDEX, "SV_ViewportArrayIndex"},
301 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
302 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
303 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
304 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
305 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
306 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
307 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
308 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
309 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
310 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
311 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
312 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
313 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
314 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
315 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
316 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
317 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
318 };
319
320 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
321 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
322
323 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
324 {
325 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
326 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
327
328 return shader_opcode_names[handler_idx];
329 }
330
331 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
332 {
333 if (usage >= ARRAY_SIZE(semantic_names))
334 {
335 FIXME("Unrecognized usage %#x.\n", usage);
336 return "UNRECOGNIZED";
337 }
338
339 return semantic_names[usage];
340 }
341
342 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
343 {
344 unsigned int i;
345
346 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
347 {
348 if (!strcmp(name, semantic_names[i]))
349 return i;
350 }
351
352 return ~0U;
353 }
354
355 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
356 {
357 switch (usage)
358 {
359 case WINED3D_DECL_USAGE_POSITION:
360 return WINED3D_SV_POSITION;
361 default:
362 return 0;
363 }
364 }
365
366 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
367 {
368 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
369 }
370
371 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
372 const struct wined3d_shader_semantic *s)
373 {
374 e->semantic_name = shader_semantic_name_from_usage(s->usage);
375 e->semantic_idx = s->usage_idx;
376 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
377 e->component_type = WINED3D_TYPE_FLOAT;
378 e->register_idx = s->reg.reg.idx[0].offset;
379 e->mask = s->reg.write_mask;
380 }
381
382 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
383 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
384 {
385 e->semantic_name = shader_semantic_name_from_usage(usage);
386 e->semantic_idx = usage_idx;
387 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
388 e->component_type = WINED3D_TYPE_FLOAT;
389 e->register_idx = reg_idx;
390 e->mask = write_mask;
391 }
392
393 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
394 {
395 switch (version_token >> 16)
396 {
397 case WINED3D_SM1_VS:
398 case WINED3D_SM1_PS:
399 return &sm1_shader_frontend;
400
401 case WINED3D_SM4_PS:
402 case WINED3D_SM4_VS:
403 case WINED3D_SM4_GS:
404 case WINED3D_SM5_HS:
405 case WINED3D_SM5_DS:
406 case WINED3D_SM5_CS:
407 return &sm4_shader_frontend;
408
409 default:
410 FIXME("Unrecognised version token %#x.\n", version_token);
411 return NULL;
412 }
413 }
414
415 void string_buffer_clear(struct wined3d_string_buffer *buffer)
416 {
417 buffer->buffer[0] = '\0';
418 buffer->content_size = 0;
419 }
420
421 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
422 {
423 buffer->buffer_size = 32;
424 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
425 {
426 ERR("Failed to allocate shader buffer memory.\n");
427 return FALSE;
428 }
429
430 string_buffer_clear(buffer);
431 return TRUE;
432 }
433
434 void string_buffer_free(struct wined3d_string_buffer *buffer)
435 {
436 HeapFree(GetProcessHeap(), 0, buffer->buffer);
437 }
438
439 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
440 {
441 char *new_buffer;
442 unsigned int new_buffer_size = buffer->buffer_size * 2;
443
444 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
445 new_buffer_size *= 2;
446 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
447 {
448 ERR("Failed to grow buffer.\n");
449 buffer->buffer[buffer->content_size] = '\0';
450 return FALSE;
451 }
452 buffer->buffer = new_buffer;
453 buffer->buffer_size = new_buffer_size;
454 return TRUE;
455 }
456
457 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
458 {
459 unsigned int rem;
460 int rc;
461
462 rem = buffer->buffer_size - buffer->content_size;
463 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
464 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
465 return rc;
466
467 buffer->content_size += rc;
468 return 0;
469 }
470
471 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
472 {
473 va_list args;
474 int ret;
475
476 for (;;)
477 {
478 va_start(args, format);
479 ret = shader_vaddline(buffer, format, args);
480 va_end(args);
481 if (!ret)
482 return ret;
483 if (!string_buffer_resize(buffer, ret))
484 return -1;
485 }
486 }
487
488 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
489 {
490 struct wined3d_string_buffer *buffer;
491
492 if (list_empty(&list->list))
493 {
494 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
495 if (!buffer || !string_buffer_init(buffer))
496 {
497 ERR("Couldn't allocate buffer for temporary string.\n");
498 HeapFree(GetProcessHeap(), 0, buffer);
499 return NULL;
500 }
501 }
502 else
503 {
504 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
505 list_remove(&buffer->entry);
506 }
507 string_buffer_clear(buffer);
508 return buffer;
509 }
510
511 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
512 {
513 if (!buffer)
514 return 0;
515 string_buffer_clear(buffer);
516 return shader_vaddline(buffer, format, args);
517 }
518
519 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
520 {
521 va_list args;
522 int ret;
523
524 for (;;)
525 {
526 va_start(args, format);
527 ret = string_buffer_vsprintf(buffer, format, args);
528 va_end(args);
529 if (!ret)
530 return;
531 if (!string_buffer_resize(buffer, ret))
532 return;
533 }
534 }
535
536 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
537 {
538 if (!buffer)
539 return;
540 list_add_head(&list->list, &buffer->entry);
541 }
542
543 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
544 {
545 list_init(&list->list);
546 }
547
548 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
549 {
550 struct wined3d_string_buffer *buffer, *buffer_next;
551
552 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
553 {
554 string_buffer_free(buffer);
555 HeapFree(GetProcessHeap(), 0, buffer);
556 }
557 list_init(&list->list);
558 }
559
560 /* Convert floating point offset relative to a register file to an absolute
561 * offset for float constants. */
562 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
563 {
564 switch (register_type)
565 {
566 case WINED3DSPR_CONST: return register_idx;
567 case WINED3DSPR_CONST2: return 2048 + register_idx;
568 case WINED3DSPR_CONST3: return 4096 + register_idx;
569 case WINED3DSPR_CONST4: return 6144 + register_idx;
570 default:
571 FIXME("Unsupported register type: %u.\n", register_type);
572 return register_idx;
573 }
574 }
575
576 static void shader_delete_constant_list(struct list *clist)
577 {
578 struct wined3d_shader_lconst *constant, *constant_next;
579
580 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
581 HeapFree(GetProcessHeap(), 0, constant);
582 list_init(clist);
583 }
584
585 static void shader_set_limits(struct wined3d_shader *shader)
586 {
587 static const struct limits_entry
588 {
589 unsigned int min_version;
590 unsigned int max_version;
591 struct wined3d_shader_limits limits;
592 }
593 vs_limits[] =
594 {
595 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
596 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
597 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
598 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
599 * even though they are capable of supporting much more (GL
600 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
601 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
602 * shaders to 256. */
603 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
604 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
605 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
606 {0}
607 },
608 hs_limits[] =
609 {
610 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
611 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
612 },
613 ds_limits[] =
614 {
615 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
616 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
617 },
618 gs_limits[] =
619 {
620 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
621 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
622 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
623 {0}
624 },
625 ps_limits[] =
626 {
627 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
628 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
629 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
630 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
631 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
632 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
633 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
634 {0}
635 },
636 cs_limits[] =
637 {
638 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
639 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
640 };
641 const struct limits_entry *limits_array;
642 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
643 shader->reg_maps.shader_version.minor);
644 int i = 0;
645
646 switch (shader->reg_maps.shader_version.type)
647 {
648 default:
649 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
650 /* Fall-through. */
651 case WINED3D_SHADER_TYPE_VERTEX:
652 limits_array = vs_limits;
653 break;
654 case WINED3D_SHADER_TYPE_HULL:
655 limits_array = hs_limits;
656 break;
657 case WINED3D_SHADER_TYPE_DOMAIN:
658 limits_array = ds_limits;
659 break;
660 case WINED3D_SHADER_TYPE_GEOMETRY:
661 limits_array = gs_limits;
662 break;
663 case WINED3D_SHADER_TYPE_PIXEL:
664 limits_array = ps_limits;
665 break;
666 case WINED3D_SHADER_TYPE_COMPUTE:
667 limits_array = cs_limits;
668 break;
669 }
670
671 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
672 {
673 if (shader_version <= limits_array[i].max_version)
674 {
675 shader->limits = &limits_array[i].limits;
676 break;
677 }
678 ++i;
679 }
680 if (!shader->limits)
681 {
682 FIXME("Unexpected shader version \"%u.%u\".\n",
683 shader->reg_maps.shader_version.major,
684 shader->reg_maps.shader_version.minor);
685 shader->limits = &limits_array[max(0, i - 1)].limits;
686 }
687 }
688
689 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
690 {
691 DWORD idx, shift;
692 idx = bit >> 5;
693 shift = bit & 0x1f;
694 bitmap[idx] |= (1u << shift);
695 }
696
697 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
698 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
699 {
700 switch (reg->type)
701 {
702 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
703 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
704 reg_maps->texcoord |= 1u << reg->idx[0].offset;
705 else
706 reg_maps->address |= 1u << reg->idx[0].offset;
707 break;
708
709 case WINED3DSPR_TEMP:
710 reg_maps->temporary |= 1u << reg->idx[0].offset;
711 break;
712
713 case WINED3DSPR_INPUT:
714 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
715 {
716 if (reg->idx[0].rel_addr)
717 {
718 /* If relative addressing is used, we must assume that all registers
719 * are used. Even if it is a construct like v3[aL], we can't assume
720 * that v0, v1 and v2 aren't read because aL can be negative */
721 unsigned int i;
722 for (i = 0; i < MAX_REG_INPUT; ++i)
723 {
724 shader->u.ps.input_reg_used[i] = TRUE;
725 }
726 }
727 else
728 {
729 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
730 }
731 }
732 else
733 reg_maps->input_registers |= 1u << reg->idx[0].offset;
734 break;
735
736 case WINED3DSPR_RASTOUT:
737 if (reg->idx[0].offset == 1)
738 reg_maps->fog = 1;
739 if (reg->idx[0].offset == 2)
740 reg_maps->point_size = 1;
741 break;
742
743 case WINED3DSPR_MISCTYPE:
744 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
745 {
746 if (!reg->idx[0].offset)
747 reg_maps->vpos = 1;
748 else if (reg->idx[0].offset == 1)
749 reg_maps->usesfacing = 1;
750 }
751 break;
752
753 case WINED3DSPR_CONST:
754 if (reg->idx[0].rel_addr)
755 {
756 if (reg->idx[0].offset < reg_maps->min_rel_offset)
757 reg_maps->min_rel_offset = reg->idx[0].offset;
758 if (reg->idx[0].offset > reg_maps->max_rel_offset)
759 reg_maps->max_rel_offset = reg->idx[0].offset;
760 reg_maps->usesrelconstF = TRUE;
761 }
762 else
763 {
764 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
765 {
766 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
767 return FALSE;
768 }
769 else
770 {
771 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
772 }
773 }
774 break;
775
776 case WINED3DSPR_CONSTINT:
777 if (reg->idx[0].offset >= shader->limits->constant_int)
778 {
779 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
780 return FALSE;
781 }
782 else
783 {
784 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
785 }
786 break;
787
788 case WINED3DSPR_CONSTBOOL:
789 if (reg->idx[0].offset >= shader->limits->constant_bool)
790 {
791 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
792 return FALSE;
793 }
794 else
795 {
796 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
797 }
798 break;
799
800 case WINED3DSPR_COLOROUT:
801 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
802 break;
803
804 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 /* Note that this does not count the loop register as an address register. */
878 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
879 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
880 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
881 {
882 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
883 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
884 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
885 void *fe_data = shader->frontend_data;
886 struct wined3d_shader_version shader_version;
887 const DWORD *ptr = byte_code;
888 unsigned int i;
889
890 memset(reg_maps, 0, sizeof(*reg_maps));
891 memset(input_signature_elements, 0, sizeof(input_signature_elements));
892 memset(output_signature_elements, 0, sizeof(output_signature_elements));
893 reg_maps->min_rel_offset = ~0U;
894 list_init(&reg_maps->indexable_temps);
895
896 fe->shader_read_header(fe_data, &ptr, &shader_version);
897 reg_maps->shader_version = shader_version;
898
899 shader_set_limits(shader);
900
901 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
902 sizeof(*reg_maps->constf))))
903 {
904 ERR("Failed to allocate constant map memory.\n");
905 return E_OUTOFMEMORY;
906 }
907
908 while (!fe->shader_is_end(fe_data, &ptr))
909 {
910 struct wined3d_shader_instruction ins;
911
912 /* Fetch opcode. */
913 fe->shader_read_instruction(fe_data, &ptr, &ins);
914
915 /* Unhandled opcode, and its parameters. */
916 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
917 {
918 TRACE("Skipping unrecognized instruction.\n");
919 continue;
920 }
921
922 /* Handle declarations. */
923 if (ins.handler_idx == WINED3DSIH_DCL
924 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
925 {
926 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
927 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
928
929 switch (semantic->reg.reg.type)
930 {
931 /* Mark input registers used. */
932 case WINED3DSPR_INPUT:
933 if (reg_idx >= MAX_REG_INPUT)
934 {
935 ERR("Invalid input register index %u.\n", reg_idx);
936 break;
937 }
938 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
939 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
940 return WINED3DERR_INVALIDCALL;
941 reg_maps->input_registers |= 1u << reg_idx;
942 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
943 break;
944
945 /* Vertex shader: mark 3.0 output registers used, save token. */
946 case WINED3DSPR_OUTPUT:
947 if (reg_idx >= MAX_REG_OUTPUT)
948 {
949 ERR("Invalid output register index %u.\n", reg_idx);
950 break;
951 }
952 reg_maps->output_registers |= 1u << reg_idx;
953 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
954 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
955 reg_maps->fog = 1;
956 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
957 reg_maps->point_size = 1;
958 break;
959
960 case WINED3DSPR_SAMPLER:
961 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
962 case WINED3DSPR_RESOURCE:
963 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
964 {
965 ERR("Invalid resource index %u.\n", reg_idx);
966 break;
967 }
968 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
969 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
970 break;
971
972 case WINED3DSPR_UAV:
973 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
974 {
975 ERR("Invalid UAV resource index %u.\n", reg_idx);
976 break;
977 }
978 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
979 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
980 break;
981
982 default:
983 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
984 break;
985 }
986 }
987 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
988 {
989 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
990 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
991 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
992 else
993 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
994 }
995 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
996 {
997 if (reg_maps->icb)
998 FIXME("Multiple immediate constant buffers.\n");
999 reg_maps->icb = ins.declaration.icb;
1000 }
1001 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1002 {
1003 struct wined3d_shader_indexable_temp *reg;
1004
1005 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
1006 return E_OUTOFMEMORY;
1007
1008 *reg = ins.declaration.indexable_temp;
1009 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1010 }
1011 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1012 {
1013 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1014 shader->u.gs.input_type = ins.declaration.primitive_type;
1015 else
1016 FIXME("Invalid instruction %#x for shader type %#x.\n",
1017 ins.handler_idx, shader_version.type);
1018 }
1019 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1020 {
1021 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1022 shader->u.gs.output_type = ins.declaration.primitive_type;
1023 else
1024 FIXME("Invalid instruction %#x for shader type %#x.\n",
1025 ins.handler_idx, shader_version.type);
1026 }
1027 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1028 {
1029 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1030 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1031 }
1032 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1033 {
1034 reg_maps->temporary_count = ins.declaration.count;
1035 }
1036 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1037 {
1038 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1039 shader->u.gs.vertices_out = ins.declaration.count;
1040 else
1041 FIXME("Invalid instruction %#x for shader type %#x.\n",
1042 ins.handler_idx, shader_version.type);
1043 }
1044 else if (ins.handler_idx == WINED3DSIH_DEF)
1045 {
1046 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1047 float *value;
1048 if (!lconst) return E_OUTOFMEMORY;
1049
1050 lconst->idx = ins.dst[0].reg.idx[0].offset;
1051 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1052 value = (float *)lconst->value;
1053
1054 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1055 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1056 {
1057 if (value[0] < -1.0f) value[0] = -1.0f;
1058 else if (value[0] > 1.0f) value[0] = 1.0f;
1059 if (value[1] < -1.0f) value[1] = -1.0f;
1060 else if (value[1] > 1.0f) value[1] = 1.0f;
1061 if (value[2] < -1.0f) value[2] = -1.0f;
1062 else if (value[2] > 1.0f) value[2] = 1.0f;
1063 if (value[3] < -1.0f) value[3] = -1.0f;
1064 else if (value[3] > 1.0f) value[3] = 1.0f;
1065 }
1066
1067 list_add_head(&shader->constantsF, &lconst->entry);
1068
1069 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1070 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1071 {
1072 shader->lconst_inf_or_nan = TRUE;
1073 }
1074 }
1075 else if (ins.handler_idx == WINED3DSIH_DEFI)
1076 {
1077 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1078 if (!lconst) return E_OUTOFMEMORY;
1079
1080 lconst->idx = ins.dst[0].reg.idx[0].offset;
1081 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1082
1083 list_add_head(&shader->constantsI, &lconst->entry);
1084 reg_maps->local_int_consts |= (1u << lconst->idx);
1085 }
1086 else if (ins.handler_idx == WINED3DSIH_DEFB)
1087 {
1088 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1089 if (!lconst) return E_OUTOFMEMORY;
1090
1091 lconst->idx = ins.dst[0].reg.idx[0].offset;
1092 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1093
1094 list_add_head(&shader->constantsB, &lconst->entry);
1095 reg_maps->local_bool_consts |= (1u << lconst->idx);
1096 }
1097 /* For subroutine prototypes. */
1098 else if (ins.handler_idx == WINED3DSIH_LABEL)
1099 {
1100 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1101 }
1102 /* Set texture, address, temporary registers. */
1103 else
1104 {
1105 BOOL color0_mov = FALSE;
1106 unsigned int i;
1107
1108 /* This will loop over all the registers and try to
1109 * make a bitmask of the ones we're interested in.
1110 *
1111 * Relative addressing tokens are ignored, but that's
1112 * okay, since we'll catch any address registers when
1113 * they are initialized (required by spec). */
1114 for (i = 0; i < ins.dst_count; ++i)
1115 {
1116 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1117 shader_version.type, constf_size))
1118 return WINED3DERR_INVALIDCALL;
1119
1120 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1121 {
1122 UINT idx = ins.dst[i].reg.idx[0].offset;
1123
1124 switch (ins.dst[i].reg.type)
1125 {
1126 case WINED3DSPR_RASTOUT:
1127 if (shader_version.major >= 3)
1128 break;
1129 switch (idx)
1130 {
1131 case 0: /* oPos */
1132 reg_maps->output_registers |= 1u << 10;
1133 shader_signature_from_usage(&output_signature_elements[10],
1134 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1135 break;
1136
1137 case 1: /* oFog */
1138 reg_maps->output_registers |= 1u << 11;
1139 shader_signature_from_usage(&output_signature_elements[11],
1140 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1141 break;
1142
1143 case 2: /* oPts */
1144 reg_maps->output_registers |= 1u << 11;
1145 shader_signature_from_usage(&output_signature_elements[11],
1146 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1147 break;
1148 }
1149 break;
1150
1151 case WINED3DSPR_ATTROUT:
1152 if (shader_version.major >= 3)
1153 break;
1154 if (idx < 2)
1155 {
1156 idx += 8;
1157 if (reg_maps->output_registers & (1u << idx))
1158 {
1159 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1160 }
1161 else
1162 {
1163 reg_maps->output_registers |= 1u << idx;
1164 shader_signature_from_usage(&output_signature_elements[idx],
1165 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1166 }
1167 }
1168 break;
1169
1170 case WINED3DSPR_TEXCRDOUT:
1171 if (shader_version.major >= 3)
1172 {
1173 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1174 break;
1175 }
1176 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1177 if (reg_maps->output_registers & (1u << idx))
1178 {
1179 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1180 }
1181 else
1182 {
1183 reg_maps->output_registers |= 1u << idx;
1184 shader_signature_from_usage(&output_signature_elements[idx],
1185 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1186 }
1187 break;
1188
1189 default:
1190 break;
1191 }
1192 }
1193
1194 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1195 {
1196 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1197 {
1198 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1199 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1200 * the mov and perform the sRGB write correction from the source register.
1201 *
1202 * However, if the mov is only partial, we can't do this, and if the write
1203 * comes from an instruction other than MOV it is hard to do as well. If
1204 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1205 shader->u.ps.color0_mov = FALSE;
1206 if (ins.handler_idx == WINED3DSIH_MOV
1207 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1208 {
1209 /* Used later when the source register is read. */
1210 color0_mov = TRUE;
1211 }
1212 }
1213 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1214 * end
1215 */
1216 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1217 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1218 {
1219 shader->u.ps.color0_mov = FALSE;
1220 }
1221 }
1222
1223 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1224 if (shader_version.major == 1
1225 && (ins.handler_idx == WINED3DSIH_TEX
1226 || ins.handler_idx == WINED3DSIH_TEXBEM
1227 || ins.handler_idx == WINED3DSIH_TEXBEML
1228 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1229 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1230 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1231 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1232 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1233 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1234 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1235 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1236 {
1237 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1238
1239 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1240 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1241 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1242 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1243
1244 /* texbem is only valid with < 1.4 pixel shaders */
1245 if (ins.handler_idx == WINED3DSIH_TEXBEM
1246 || ins.handler_idx == WINED3DSIH_TEXBEML)
1247 {
1248 reg_maps->bumpmat |= 1u << reg_idx;
1249 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1250 {
1251 reg_maps->luminanceparams |= 1u << reg_idx;
1252 }
1253 }
1254 }
1255 else if (ins.handler_idx == WINED3DSIH_BEM)
1256 {
1257 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1258 }
1259 }
1260
1261 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD
1262 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED)
1263 {
1264 unsigned int reg_idx;
1265 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD)
1266 reg_idx = ins.dst[0].reg.idx[0].offset;
1267 else
1268 reg_idx = ins.src[1].reg.idx[0].offset;
1269 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1270 {
1271 ERR("Invalid UAV index %u.\n", reg_idx);
1272 break;
1273 }
1274 reg_maps->uav_read_mask |= (1u << reg_idx);
1275 }
1276 else if (ins.handler_idx == WINED3DSIH_NRM)
1277 {
1278 reg_maps->usesnrm = 1;
1279 }
1280 else if (ins.handler_idx == WINED3DSIH_DSY
1281 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1282 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1283 {
1284 reg_maps->usesdsy = 1;
1285 }
1286 else if (ins.handler_idx == WINED3DSIH_DSX
1287 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1288 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1289 {
1290 reg_maps->usesdsx = 1;
1291 }
1292 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1293 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1294 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1295 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1296 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1297 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1298 else if (ins.handler_idx == WINED3DSIH_LOOP
1299 || ins.handler_idx == WINED3DSIH_REP)
1300 {
1301 ++cur_loop_depth;
1302 if (cur_loop_depth > max_loop_depth)
1303 max_loop_depth = cur_loop_depth;
1304 }
1305 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1306 || ins.handler_idx == WINED3DSIH_ENDREP)
1307 {
1308 --cur_loop_depth;
1309 }
1310 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1311 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1312 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1313 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1314 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1315 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1316 {
1317 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1318 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1319 }
1320 else if (ins.handler_idx == WINED3DSIH_LD
1321 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1322 {
1323 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1324 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1325 }
1326
1327 if (ins.predicate)
1328 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1329 shader_version.type, constf_size))
1330 return WINED3DERR_INVALIDCALL;
1331
1332 for (i = 0; i < ins.src_count; ++i)
1333 {
1334 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1335 struct wined3d_shader_register reg = ins.src[i].reg;
1336
1337 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1338 shader_version.type, constf_size))
1339 return WINED3DERR_INVALIDCALL;
1340 while (count)
1341 {
1342 ++reg.idx[0].offset;
1343 if (!shader_record_register_usage(shader, reg_maps, &reg,
1344 shader_version.type, constf_size))
1345 return WINED3DERR_INVALIDCALL;
1346 --count;
1347 }
1348
1349 if (color0_mov)
1350 {
1351 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1352 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1353 {
1354 shader->u.ps.color0_mov = TRUE;
1355 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1356 }
1357 }
1358 }
1359 }
1360 }
1361 reg_maps->loop_depth = max_loop_depth;
1362
1363 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1364 * R0 is written to the render target. */
1365 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1366 reg_maps->rt_mask |= (1u << 0);
1367
1368 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1369
1370 if (input_signature->elements)
1371 {
1372 for (i = 0; i < input_signature->element_count; ++i)
1373 {
1374 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1375 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1376 {
1377 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1378 reg_maps->vpos = 1;
1379 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1380 reg_maps->usesfacing = 1;
1381 }
1382 }
1383 }
1384 else if (!input_signature->elements && reg_maps->input_registers)
1385 {
1386 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1387 struct wined3d_shader_signature_element *e;
1388 unsigned int i;
1389
1390 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1391 return E_OUTOFMEMORY;
1392 input_signature->element_count = count;
1393
1394 e = input_signature->elements;
1395 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1396 {
1397 if (!(reg_maps->input_registers & (1u << i)))
1398 continue;
1399 input_signature_elements[i].register_idx = i;
1400 *e++ = input_signature_elements[i];
1401 }
1402 }
1403
1404 if (output_signature->elements)
1405 {
1406 for (i = 0; i < output_signature->element_count; ++i)
1407 {
1408 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1409 }
1410 }
1411 else if (reg_maps->output_registers)
1412 {
1413 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1414 struct wined3d_shader_signature_element *e;
1415
1416 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1417 return E_OUTOFMEMORY;
1418 output_signature->element_count = count;
1419
1420 e = output_signature->elements;
1421 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1422 {
1423 if (!(reg_maps->output_registers & (1u << i)))
1424 continue;
1425 *e++ = output_signature_elements[i];
1426 }
1427 }
1428
1429 return WINED3D_OK;
1430 }
1431
1432 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1433 {
1434 struct wined3d_shader_indexable_temp *reg, *reg_next;
1435
1436 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1437 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1438
1439 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1440 HeapFree(GetProcessHeap(), 0, reg);
1441 list_init(&reg_maps->indexable_temps);
1442 }
1443
1444 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1445 {
1446 DWORD map = 1u << max;
1447 map |= map - 1;
1448 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1449
1450 return wined3d_log2i(map);
1451 }
1452
1453 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1454 {
1455 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1456 {
1457 shader_addline(buffer, "refactoringAllowed");
1458 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1459 if (global_flags)
1460 shader_addline(buffer, " | ");
1461 }
1462
1463 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1464 {
1465 shader_addline(buffer, "enableRawAndStructuredBuffers");
1466 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1467 }
1468
1469 if (global_flags)
1470 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1471 }
1472
1473 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1474 {
1475 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1476 {
1477 shader_addline(buffer, "_g");
1478 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1479 }
1480 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1481 {
1482 shader_addline(buffer, "_t");
1483 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1484 }
1485
1486 if (sync_flags)
1487 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1488 }
1489
1490 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1491 {
1492 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1493 {
1494 shader_addline(buffer, "_glc");
1495 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1496 }
1497 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1498 {
1499 shader_addline(buffer, "_opc");
1500 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1501 }
1502
1503 if (uav_flags)
1504 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1505 }
1506
1507 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1508 enum wined3d_tessellator_domain domain)
1509 {
1510 switch (domain)
1511 {
1512 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1513 shader_addline(buffer, "line");
1514 break;
1515 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1516 shader_addline(buffer, "triangle");
1517 break;
1518 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1519 shader_addline(buffer, "quad");
1520 break;
1521 default:
1522 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1523 break;
1524 }
1525 }
1526
1527 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1528 enum wined3d_tessellator_output_primitive output_primitive)
1529 {
1530 switch (output_primitive)
1531 {
1532 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1533 shader_addline(buffer, "point");
1534 break;
1535 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1536 shader_addline(buffer, "line");
1537 break;
1538 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1539 shader_addline(buffer, "triangle_cw");
1540 break;
1541 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1542 shader_addline(buffer, "triangle_ccw");
1543 break;
1544 default:
1545 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1546 break;
1547 }
1548 }
1549
1550 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1551 enum wined3d_tessellator_partitioning partitioning)
1552 {
1553 switch (partitioning)
1554 {
1555 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1556 shader_addline(buffer, "integer");
1557 break;
1558 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1559 shader_addline(buffer, "pow2");
1560 break;
1561 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1562 shader_addline(buffer, "fractional_odd");
1563 break;
1564 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1565 shader_addline(buffer, "fractional_even");
1566 break;
1567 default:
1568 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1569 break;
1570 }
1571 }
1572
1573 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1574 {
1575 unsigned int i;
1576
1577 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1578 {
1579 if (sysval_semantic_names[i].sysval_semantic == semantic)
1580 {
1581 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1582 return;
1583 }
1584 }
1585
1586 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1587 }
1588
1589 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1590 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1591 {
1592 shader_addline(buffer, "dcl");
1593
1594 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1595 {
1596 switch (semantic->resource_type)
1597 {
1598 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1599 shader_addline(buffer, "_2d");
1600 break;
1601
1602 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1603 shader_addline(buffer, "_3d");
1604 break;
1605
1606 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1607 shader_addline(buffer, "_cube");
1608 break;
1609
1610 default:
1611 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1612 break;
1613 }
1614 }
1615 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1616 {
1617 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1618 shader_addline(buffer, "_resource_");
1619 else
1620 shader_addline(buffer, "_uav_");
1621 switch (semantic->resource_type)
1622 {
1623 case WINED3D_SHADER_RESOURCE_BUFFER:
1624 shader_addline(buffer, "buffer");
1625 break;
1626
1627 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1628 shader_addline(buffer, "texture1d");
1629 break;
1630
1631 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1632 shader_addline(buffer, "texture2d");
1633 break;
1634
1635 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1636 shader_addline(buffer, "texture2dms");
1637 break;
1638
1639 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1640 shader_addline(buffer, "texture3d");
1641 break;
1642
1643 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1644 shader_addline(buffer, "texturecube");
1645 break;
1646
1647 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1648 shader_addline(buffer, "texture1darray");
1649 break;
1650
1651 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1652 shader_addline(buffer, "texture2darray");
1653 break;
1654
1655 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1656 shader_addline(buffer, "texture2dmsarray");
1657 break;
1658
1659 default:
1660 shader_addline(buffer, "unknown");
1661 break;
1662 }
1663 switch (semantic->resource_data_type)
1664 {
1665 case WINED3D_DATA_FLOAT:
1666 shader_addline(buffer, " (float)");
1667 break;
1668
1669 case WINED3D_DATA_INT:
1670 shader_addline(buffer, " (int)");
1671 break;
1672
1673 case WINED3D_DATA_UINT:
1674 shader_addline(buffer, " (uint)");
1675 break;
1676
1677 case WINED3D_DATA_UNORM:
1678 shader_addline(buffer, " (unorm)");
1679 break;
1680
1681 case WINED3D_DATA_SNORM:
1682 shader_addline(buffer, " (snorm)");
1683 break;
1684
1685 default:
1686 shader_addline(buffer, " (unknown)");
1687 break;
1688 }
1689 }
1690 else
1691 {
1692 /* Pixel shaders 3.0 don't have usage semantics. */
1693 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1694 return;
1695 else
1696 shader_addline(buffer, "_");
1697
1698 switch (semantic->usage)
1699 {
1700 case WINED3D_DECL_USAGE_POSITION:
1701 shader_addline(buffer, "position%u", semantic->usage_idx);
1702 break;
1703
1704 case WINED3D_DECL_USAGE_BLEND_INDICES:
1705 shader_addline(buffer, "blend");
1706 break;
1707
1708 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1709 shader_addline(buffer, "weight");
1710 break;
1711
1712 case WINED3D_DECL_USAGE_NORMAL:
1713 shader_addline(buffer, "normal%u", semantic->usage_idx);
1714 break;
1715
1716 case WINED3D_DECL_USAGE_PSIZE:
1717 shader_addline(buffer, "psize");
1718 break;
1719
1720 case WINED3D_DECL_USAGE_COLOR:
1721 if (!semantic->usage_idx)
1722 shader_addline(buffer, "color");
1723 else
1724 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1725 break;
1726
1727 case WINED3D_DECL_USAGE_TEXCOORD:
1728 shader_addline(buffer, "texture%u", semantic->usage_idx);
1729 break;
1730
1731 case WINED3D_DECL_USAGE_TANGENT:
1732 shader_addline(buffer, "tangent");
1733 break;
1734
1735 case WINED3D_DECL_USAGE_BINORMAL:
1736 shader_addline(buffer, "binormal");
1737 break;
1738
1739 case WINED3D_DECL_USAGE_TESS_FACTOR:
1740 shader_addline(buffer, "tessfactor");
1741 break;
1742
1743 case WINED3D_DECL_USAGE_POSITIONT:
1744 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1745 break;
1746
1747 case WINED3D_DECL_USAGE_FOG:
1748 shader_addline(buffer, "fog");
1749 break;
1750
1751 case WINED3D_DECL_USAGE_DEPTH:
1752 shader_addline(buffer, "depth");
1753 break;
1754
1755 case WINED3D_DECL_USAGE_SAMPLE:
1756 shader_addline(buffer, "sample");
1757 break;
1758
1759 default:
1760 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1761 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
1762 }
1763 }
1764 }
1765
1766 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1767 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1768 {
1769 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1770 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1771 UINT offset = reg->idx[0].offset;
1772
1773 switch (reg->type)
1774 {
1775 case WINED3DSPR_TEMP:
1776 shader_addline(buffer, "r");
1777 break;
1778
1779 case WINED3DSPR_INPUT:
1780 shader_addline(buffer, "v");
1781 break;
1782
1783 case WINED3DSPR_CONST:
1784 case WINED3DSPR_CONST2:
1785 case WINED3DSPR_CONST3:
1786 case WINED3DSPR_CONST4:
1787 shader_addline(buffer, "c");
1788 offset = shader_get_float_offset(reg->type, offset);
1789 break;
1790
1791 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1792 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1793 break;
1794
1795 case WINED3DSPR_RASTOUT:
1796 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1797 break;
1798
1799 case WINED3DSPR_COLOROUT:
1800 shader_addline(buffer, "oC");
1801 break;
1802
1803 case WINED3DSPR_DEPTHOUT:
1804 shader_addline(buffer, "oDepth");
1805 break;
1806
1807 case WINED3DSPR_ATTROUT:
1808 shader_addline(buffer, "oD");
1809 break;
1810
1811 case WINED3DSPR_TEXCRDOUT:
1812 /* Vertex shaders >= 3.0 use general purpose output registers
1813 * (WINED3DSPR_OUTPUT), which can include an address token. */
1814 if (shader_version->major >= 3)
1815 shader_addline(buffer, "o");
1816 else
1817 shader_addline(buffer, "oT");
1818 break;
1819
1820 case WINED3DSPR_CONSTINT:
1821 shader_addline(buffer, "i");
1822 break;
1823
1824 case WINED3DSPR_CONSTBOOL:
1825 shader_addline(buffer, "b");
1826 break;
1827
1828 case WINED3DSPR_LABEL:
1829 shader_addline(buffer, "l");
1830 break;
1831
1832 case WINED3DSPR_LOOP:
1833 shader_addline(buffer, "aL");
1834 break;
1835
1836 case WINED3DSPR_SAMPLER:
1837 shader_addline(buffer, "s");
1838 break;
1839
1840 case WINED3DSPR_MISCTYPE:
1841 if (offset > 1)
1842 {
1843 FIXME("Unhandled misctype register %u.\n", offset);
1844 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1845 }
1846 else
1847 {
1848 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1849 }
1850 break;
1851
1852 case WINED3DSPR_PREDICATE:
1853 shader_addline(buffer, "p");
1854 break;
1855
1856 case WINED3DSPR_IMMCONST:
1857 shader_addline(buffer, "l");
1858 break;
1859
1860 case WINED3DSPR_CONSTBUFFER:
1861 shader_addline(buffer, "cb");
1862 break;
1863
1864 case WINED3DSPR_IMMCONSTBUFFER:
1865 shader_addline(buffer, "icb");
1866 break;
1867
1868 case WINED3DSPR_PRIMID:
1869 shader_addline(buffer, "primID");
1870 break;
1871
1872 case WINED3DSPR_NULL:
1873 shader_addline(buffer, "null");
1874 break;
1875
1876 case WINED3DSPR_RESOURCE:
1877 shader_addline(buffer, "t");
1878 break;
1879
1880 case WINED3DSPR_UAV:
1881 shader_addline(buffer, "u");
1882 break;
1883
1884 case WINED3DSPR_OUTPOINTID:
1885 shader_addline(buffer, "vOutputControlPointID");
1886 break;
1887
1888 case WINED3DSPR_FORKINSTID:
1889 shader_addline(buffer, "vForkInstanceId");
1890 break;
1891
1892 case WINED3DSPR_INCONTROLPOINT:
1893 shader_addline(buffer, "vicp");
1894 break;
1895
1896 case WINED3DSPR_OUTCONTROLPOINT:
1897 shader_addline(buffer, "vocp");
1898 break;
1899
1900 case WINED3DSPR_PATCHCONST:
1901 shader_addline(buffer, "vpc");
1902 break;
1903
1904 case WINED3DSPR_TESSCOORD:
1905 shader_addline(buffer, "vDomainLocation");
1906 break;
1907
1908 case WINED3DSPR_GROUPSHAREDMEM:
1909 shader_addline(buffer, "g");
1910 break;
1911
1912 case WINED3DSPR_THREADID:
1913 shader_addline(buffer, "vThreadID");
1914 break;
1915
1916 case WINED3DSPR_THREADGROUPID:
1917 shader_addline(buffer, "vThreadGroupID");
1918 break;
1919
1920 case WINED3DSPR_LOCALTHREADID:
1921 shader_addline(buffer, "vThreadIDInGroup");
1922 break;
1923
1924 case WINED3DSPR_LOCALTHREADINDEX:
1925 shader_addline(buffer, "vThreadIDInGroupFlattened");
1926 break;
1927
1928 case WINED3DSPR_IDXTEMP:
1929 shader_addline(buffer, "x");
1930 break;
1931
1932 case WINED3DSPR_STREAM:
1933 shader_addline(buffer, "m");
1934 break;
1935
1936 case WINED3DSPR_FUNCTIONBODY:
1937 shader_addline(buffer, "fb");
1938 break;
1939
1940 case WINED3DSPR_FUNCTIONPOINTER:
1941 shader_addline(buffer, "fp");
1942 break;
1943
1944 case WINED3DSPR_COVERAGE:
1945 shader_addline(buffer, "vCoverage");
1946 break;
1947
1948 case WINED3DSPR_SAMPLEMASK:
1949 shader_addline(buffer, "oMask");
1950 break;
1951
1952 default:
1953 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1954 break;
1955 }
1956
1957 if (reg->type == WINED3DSPR_IMMCONST)
1958 {
1959 shader_addline(buffer, "(");
1960 switch (reg->immconst_type)
1961 {
1962 case WINED3D_IMMCONST_SCALAR:
1963 switch (reg->data_type)
1964 {
1965 case WINED3D_DATA_FLOAT:
1966 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
1967 break;
1968 case WINED3D_DATA_INT:
1969 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
1970 break;
1971 case WINED3D_DATA_RESOURCE:
1972 case WINED3D_DATA_SAMPLER:
1973 case WINED3D_DATA_UINT:
1974 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
1975 break;
1976 default:
1977 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1978 break;
1979 }
1980 break;
1981
1982 case WINED3D_IMMCONST_VEC4:
1983 switch (reg->data_type)
1984 {
1985 case WINED3D_DATA_FLOAT:
1986 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1987 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
1988 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
1989 break;
1990 case WINED3D_DATA_INT:
1991 shader_addline(buffer, "%d, %d, %d, %d",
1992 reg->u.immconst_data[0], reg->u.immconst_data[1],
1993 reg->u.immconst_data[2], reg->u.immconst_data[3]);
1994 break;
1995 case WINED3D_DATA_RESOURCE:
1996 case WINED3D_DATA_SAMPLER:
1997 case WINED3D_DATA_UINT:
1998 shader_addline(buffer, "%u, %u, %u, %u",
1999 reg->u.immconst_data[0], reg->u.immconst_data[1],
2000 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2001 break;
2002 default:
2003 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2004 break;
2005 }
2006 break;
2007
2008 default:
2009 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2010 break;
2011 }
2012 shader_addline(buffer, ")");
2013 }
2014 else if (reg->type != WINED3DSPR_RASTOUT
2015 && reg->type != WINED3DSPR_MISCTYPE
2016 && reg->type != WINED3DSPR_NULL)
2017 {
2018 if (offset != ~0u)
2019 {
2020 shader_addline(buffer, "[");
2021 if (reg->idx[0].rel_addr)
2022 {
2023 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2024 shader_addline(buffer, " + ");
2025 }
2026 shader_addline(buffer, "%u]", offset);
2027
2028 if (reg->idx[1].offset != ~0u)
2029 {
2030 shader_addline(buffer, "[");
2031 if (reg->idx[1].rel_addr)
2032 {
2033 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2034 shader_addline(buffer, " + ");
2035 }
2036 shader_addline(buffer, "%u]", reg->idx[1].offset);
2037 }
2038 }
2039
2040 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2041 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2042 }
2043 }
2044
2045 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2046 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2047 {
2048 DWORD write_mask = param->write_mask;
2049
2050 shader_dump_register(buffer, &param->reg, shader_version);
2051
2052 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2053 {
2054 static const char write_mask_chars[] = "xyzw";
2055
2056 shader_addline(buffer, ".");
2057 if (write_mask & WINED3DSP_WRITEMASK_0)
2058 shader_addline(buffer, "%c", write_mask_chars[0]);
2059 if (write_mask & WINED3DSP_WRITEMASK_1)
2060 shader_addline(buffer, "%c", write_mask_chars[1]);
2061 if (write_mask & WINED3DSP_WRITEMASK_2)
2062 shader_addline(buffer, "%c", write_mask_chars[2]);
2063 if (write_mask & WINED3DSP_WRITEMASK_3)
2064 shader_addline(buffer, "%c", write_mask_chars[3]);
2065 }
2066 }
2067
2068 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2069 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2070 {
2071 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2072 DWORD swizzle = param->swizzle;
2073
2074 if (src_modifier == WINED3DSPSM_NEG
2075 || src_modifier == WINED3DSPSM_BIASNEG
2076 || src_modifier == WINED3DSPSM_SIGNNEG
2077 || src_modifier == WINED3DSPSM_X2NEG
2078 || src_modifier == WINED3DSPSM_ABSNEG)
2079 shader_addline(buffer, "-");
2080 else if (src_modifier == WINED3DSPSM_COMP)
2081 shader_addline(buffer, "1-");
2082 else if (src_modifier == WINED3DSPSM_NOT)
2083 shader_addline(buffer, "!");
2084
2085 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2086 shader_addline(buffer, "abs(");
2087
2088 shader_dump_register(buffer, &param->reg, shader_version);
2089
2090 switch (src_modifier)
2091 {
2092 case WINED3DSPSM_NONE: break;
2093 case WINED3DSPSM_NEG: break;
2094 case WINED3DSPSM_NOT: break;
2095 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2096 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2097 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2098 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2099 case WINED3DSPSM_COMP: break;
2100 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2101 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2102 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2103 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2104 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2105 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2106 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2107 }
2108
2109 if (swizzle != WINED3DSP_NOSWIZZLE)
2110 {
2111 static const char swizzle_chars[] = "xyzw";
2112 DWORD swizzle_x = swizzle & 0x03;
2113 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2114 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2115 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2116
2117 if (swizzle_x == swizzle_y
2118 && swizzle_x == swizzle_z
2119 && swizzle_x == swizzle_w)
2120 {
2121 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2122 }
2123 else
2124 {
2125 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2126 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2127 }
2128 }
2129 }
2130
2131 /* Shared code in order to generate the bulk of the shader string.
2132 * NOTE: A description of how to parse tokens can be found on MSDN. */
2133 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2134 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
2135 {
2136 struct wined3d_device *device = shader->device;
2137 const struct wined3d_shader_frontend *fe = shader->frontend;
2138 void *fe_data = shader->frontend_data;
2139 struct wined3d_shader_version shader_version;
2140 struct wined3d_shader_parser_state state;
2141 struct wined3d_shader_instruction ins;
2142 struct wined3d_shader_tex_mx tex_mx;
2143 struct wined3d_shader_context ctx;
2144 const DWORD *ptr = byte_code;
2145
2146 /* Initialize current parsing state. */
2147 tex_mx.current_row = 0;
2148 state.current_loop_depth = 0;
2149 state.current_loop_reg = 0;
2150 state.in_subroutine = FALSE;
2151
2152 ctx.shader = shader;
2153 ctx.gl_info = &device->adapter->gl_info;
2154 ctx.reg_maps = reg_maps;
2155 ctx.buffer = buffer;
2156 ctx.tex_mx = &tex_mx;
2157 ctx.state = &state;
2158 ctx.backend_data = backend_ctx;
2159 ins.ctx = &ctx;
2160
2161 fe->shader_read_header(fe_data, &ptr, &shader_version);
2162
2163 while (!fe->shader_is_end(fe_data, &ptr))
2164 {
2165 /* Read opcode. */
2166 fe->shader_read_instruction(fe_data, &ptr, &ins);
2167
2168 /* Unknown opcode and its parameters. */
2169 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2170 {
2171 TRACE("Skipping unrecognized instruction.\n");
2172 continue;
2173 }
2174
2175 if (ins.predicate)
2176 FIXME("Predicates not implemented.\n");
2177
2178 /* Call appropriate function for output target */
2179 device->shader_backend->shader_handle_instruction(&ins);
2180 }
2181 }
2182
2183 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2184 const struct wined3d_shader_dst_param *dst)
2185 {
2186 DWORD mmask = dst->modifiers;
2187
2188 switch (dst->shift)
2189 {
2190 case 0: break;
2191 case 13: shader_addline(buffer, "_d8"); break;
2192 case 14: shader_addline(buffer, "_d4"); break;
2193 case 15: shader_addline(buffer, "_d2"); break;
2194 case 1: shader_addline(buffer, "_x2"); break;
2195 case 2: shader_addline(buffer, "_x4"); break;
2196 case 3: shader_addline(buffer, "_x8"); break;
2197 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2198 }
2199
2200 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2201 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2202 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2203
2204 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2205 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2206 }
2207
2208 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2209 enum wined3d_primitive_type primitive_type)
2210 {
2211 switch (primitive_type)
2212 {
2213 case WINED3D_PT_UNDEFINED:
2214 shader_addline(buffer, "undefined");
2215 break;
2216 case WINED3D_PT_POINTLIST:
2217 shader_addline(buffer, "pointlist");
2218 break;
2219 case WINED3D_PT_LINELIST:
2220 shader_addline(buffer, "linelist");
2221 break;
2222 case WINED3D_PT_LINESTRIP:
2223 shader_addline(buffer, "linestrip");
2224 break;
2225 case WINED3D_PT_TRIANGLELIST:
2226 shader_addline(buffer, "trianglelist");
2227 break;
2228 case WINED3D_PT_TRIANGLESTRIP:
2229 shader_addline(buffer, "trianglestrip");
2230 break;
2231 case WINED3D_PT_TRIANGLEFAN:
2232 shader_addline(buffer, "trianglefan");
2233 break;
2234 case WINED3D_PT_LINELIST_ADJ:
2235 shader_addline(buffer, "linelist_adj");
2236 break;
2237 case WINED3D_PT_LINESTRIP_ADJ:
2238 shader_addline(buffer, "linestrip_adj");
2239 break;
2240 case WINED3D_PT_TRIANGLELIST_ADJ:
2241 shader_addline(buffer, "trianglelist_adj");
2242 break;
2243 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2244 shader_addline(buffer, "trianglestrip_adj");
2245 break;
2246 default:
2247 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2248 break;
2249 }
2250 }
2251
2252 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2253 enum wined3d_shader_interpolation_mode interpolation_mode)
2254 {
2255 switch (interpolation_mode)
2256 {
2257 case WINED3DSIM_CONSTANT:
2258 shader_addline(buffer, "constant");
2259 break;
2260 case WINED3DSIM_LINEAR:
2261 shader_addline(buffer, "linear");
2262 break;
2263 case WINED3DSIM_LINEAR_CENTROID:
2264 shader_addline(buffer, "linear centroid");
2265 break;
2266 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2267 shader_addline(buffer, "linear noperspective");
2268 break;
2269 case WINED3DSIM_LINEAR_SAMPLE:
2270 shader_addline(buffer, "linear sample");
2271 break;
2272 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2273 shader_addline(buffer, "linear noperspective centroid");
2274 break;
2275 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2276 shader_addline(buffer, "linear noperspective sample");
2277 break;
2278 default:
2279 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2280 break;
2281 }
2282 }
2283
2284 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2285 {
2286 struct wined3d_shader_version shader_version;
2287 struct wined3d_string_buffer buffer;
2288 const DWORD *ptr = byte_code;
2289 const char *type_prefix;
2290 const char *p, *q;
2291 DWORD i;
2292
2293 if (!string_buffer_init(&buffer))
2294 {
2295 ERR("Failed to initialize string buffer.\n");
2296 return;
2297 }
2298
2299 TRACE("Parsing %p.\n", byte_code);
2300
2301 fe->shader_read_header(fe_data, &ptr, &shader_version);
2302
2303 switch (shader_version.type)
2304 {
2305 case WINED3D_SHADER_TYPE_VERTEX:
2306 type_prefix = "vs";
2307 break;
2308
2309 case WINED3D_SHADER_TYPE_HULL:
2310 type_prefix = "hs";
2311 break;
2312
2313 case WINED3D_SHADER_TYPE_DOMAIN:
2314 type_prefix = "ds";
2315 break;
2316
2317 case WINED3D_SHADER_TYPE_GEOMETRY:
2318 type_prefix = "gs";
2319 break;
2320
2321 case WINED3D_SHADER_TYPE_PIXEL:
2322 type_prefix = "ps";
2323 break;
2324
2325 case WINED3D_SHADER_TYPE_COMPUTE:
2326 type_prefix = "cs";
2327 break;
2328
2329 default:
2330 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2331 type_prefix = "unknown";
2332 break;
2333 }
2334
2335 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2336
2337 while (!fe->shader_is_end(fe_data, &ptr))
2338 {
2339 struct wined3d_shader_instruction ins;
2340
2341 fe->shader_read_instruction(fe_data, &ptr, &ins);
2342 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2343 {
2344 WARN("Skipping unrecognized instruction.\n");
2345 shader_addline(&buffer, "<unrecognized instruction>\n");
2346 continue;
2347 }
2348
2349 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2350 {
2351 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2352 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2353 shader_addline(&buffer, " ");
2354 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2355 }
2356 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2357 {
2358 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2359 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2360 shader_addline(&buffer, ", %s",
2361 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2362 }
2363 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2364 {
2365 shader_addline(&buffer, "%s fb%u",
2366 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2367 }
2368 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2369 {
2370 shader_addline(&buffer, "%s ft%u = {...}",
2371 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2372 }
2373 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2374 {
2375 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2376 shader_dump_global_flags(&buffer, ins.flags);
2377 }
2378 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2379 {
2380 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2381 ins.declaration.max_tessellation_factor);
2382 }
2383 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2384 {
2385 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2386 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2387 {
2388 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2389 ins.declaration.icb->data[4 * i + 0],
2390 ins.declaration.icb->data[4 * i + 1],
2391 ins.declaration.icb->data[4 * i + 2],
2392 ins.declaration.icb->data[4 * i + 3]);
2393 }
2394 shader_addline(&buffer, "}");
2395 }
2396 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2397 {
2398 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2399 ins.declaration.indexable_temp.register_idx,
2400 ins.declaration.indexable_temp.register_size,
2401 ins.declaration.indexable_temp.component_count);
2402 }
2403 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2404 {
2405 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2406 shader_dump_interpolation_mode(&buffer, ins.flags);
2407 shader_addline(&buffer, " ");
2408 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2409 }
2410 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2411 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2412 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2413 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2414 {
2415 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2416 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2417 shader_addline(&buffer, ", ");
2418 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2419 }
2420 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2421 {
2422 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2423 shader_dump_interpolation_mode(&buffer, ins.flags);
2424 shader_addline(&buffer, " ");
2425 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2426 shader_addline(&buffer, ", ");
2427 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2428 }
2429 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2430 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2431 {
2432 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2433 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2434 }
2435 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2436 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2437 {
2438 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2439 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2440 }
2441 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2442 {
2443 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2444 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2445 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2446 }
2447 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2448 {
2449 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2450 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2451 }
2452 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2453 {
2454 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2455 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2456 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2457 }
2458 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2459 {
2460 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2461 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2462 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2463 shader_addline(&buffer, ", comparisonMode");
2464 }
2465 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2466 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2467 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2468 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2469 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2470 {
2471 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2472 }
2473 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2474 {
2475 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2476 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2477 }
2478 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2479 {
2480 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2481 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2482 }
2483 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2484 {
2485 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2486 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2487 }
2488 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2489 {
2490 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2491 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2492 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2493 }
2494 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2495 {
2496 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2497 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2498 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2499 ins.declaration.tgsm_structured.structure_count);
2500 }
2501 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2502 {
2503 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2504 ins.declaration.thread_group_size.x,
2505 ins.declaration.thread_group_size.y,
2506 ins.declaration.thread_group_size.z);
2507 }
2508 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2509 {
2510 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2511 shader_dump_uav_flags(&buffer, ins.flags);
2512 shader_addline(&buffer, " ");
2513 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2514 }
2515 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2516 {
2517 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2518 shader_dump_uav_flags(&buffer, ins.flags);
2519 shader_addline(&buffer, " ");
2520 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2521 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2522 }
2523 else if (ins.handler_idx == WINED3DSIH_DEF)
2524 {
2525 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2526 ins.dst[0].reg.idx[0].offset),
2527 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2528 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2529 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2530 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2531 }
2532 else if (ins.handler_idx == WINED3DSIH_DEFI)
2533 {
2534 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2535 ins.src[0].reg.u.immconst_data[0],
2536 ins.src[0].reg.u.immconst_data[1],
2537 ins.src[0].reg.u.immconst_data[2],
2538 ins.src[0].reg.u.immconst_data[3]);
2539 }
2540 else if (ins.handler_idx == WINED3DSIH_DEFB)
2541 {
2542 shader_addline(&buffer, "defb b%u = %s",
2543 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2544 }
2545 else
2546 {
2547 if (ins.predicate)
2548 {
2549 shader_addline(&buffer, "(");
2550 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2551 shader_addline(&buffer, ") ");
2552 }
2553
2554 /* PixWin marks instructions with the coissue flag with a '+' */
2555 if (ins.coissue)
2556 shader_addline(&buffer, "+");
2557
2558 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2559
2560 if (ins.handler_idx == WINED3DSIH_BREAKP
2561 || ins.handler_idx == WINED3DSIH_IF)
2562 {
2563 switch (ins.flags)
2564 {
2565 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2566 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2567 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2568 }
2569 }
2570 else if (ins.handler_idx == WINED3DSIH_IFC
2571 || ins.handler_idx == WINED3DSIH_BREAKC)
2572 {
2573 switch (ins.flags)
2574 {
2575 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2576 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2577 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2578 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2579 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2580 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2581 default: shader_addline(&buffer, "_(%u)", ins.flags);
2582 }
2583 }
2584 else if (ins.handler_idx == WINED3DSIH_TEX
2585 && shader_version.major >= 2
2586 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2587 {
2588 shader_addline(&buffer, "p");
2589 }
2590 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2591 {
2592 switch (ins.flags)
2593 {
2594 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2595 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2596 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2597 }
2598 }
2599 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2600 {
2601 switch (ins.flags)
2602 {
2603 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2604 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2605 }
2606 }
2607 else if (ins.handler_idx == WINED3DSIH_SYNC)
2608 {
2609 shader_dump_sync_flags(&buffer, ins.flags);
2610 }
2611
2612 if (wined3d_shader_instruction_has_texel_offset(&ins))
2613 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2614
2615 for (i = 0; i < ins.dst_count; ++i)
2616 {
2617 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2618 shader_addline(&buffer, !i ? " " : ", ");
2619 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2620 }
2621
2622 /* Other source tokens */
2623 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2624 {
2625 shader_addline(&buffer, !i ? " " : ", ");
2626 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2627 }
2628 }
2629 shader_addline(&buffer, "\n");
2630 }
2631
2632 for (p = buffer.buffer; *p; p = q)
2633 {
2634 if (!(q = strstr(p, "\n")))
2635 q = p + strlen(p);
2636 else
2637 ++q;
2638 TRACE(" %.*s", (int)(q - p), p);
2639 }
2640
2641 string_buffer_free(&buffer);
2642 }
2643
2644 static void shader_cleanup(struct wined3d_shader *shader)
2645 {
2646 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2647 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2648 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2649 shader->device->shader_backend->shader_destroy(shader);
2650 shader_cleanup_reg_maps(&shader->reg_maps);
2651 HeapFree(GetProcessHeap(), 0, shader->function);
2652 shader_delete_constant_list(&shader->constantsF);
2653 shader_delete_constant_list(&shader->constantsB);
2654 shader_delete_constant_list(&shader->constantsI);
2655 list_remove(&shader->shader_list_entry);
2656
2657 if (shader->frontend && shader->frontend_data)
2658 shader->frontend->shader_free(shader->frontend_data);
2659 }
2660
2661 struct shader_none_priv
2662 {
2663 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2664 const struct fragment_pipeline *fragment_pipe;
2665 BOOL ffp_proj_control;
2666 };
2667
2668 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2669 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
2670 const struct wined3d_state *state) {}
2671 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2672 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2673 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2674 const struct wined3d_state *state) {}
2675 static void shader_none_destroy(struct wined3d_shader *shader) {}
2676 static void shader_none_free_context_data(struct wined3d_context *context) {}
2677 static void shader_none_init_context_state(struct wined3d_context *context) {}
2678
2679 /* Context activation is done by the caller. */
2680 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2681 const struct wined3d_state *state)
2682 {
2683 const struct wined3d_gl_info *gl_info = context->gl_info;
2684 struct shader_none_priv *priv = shader_priv;
2685
2686 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2687 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2688 }
2689
2690 /* Context activation is done by the caller. */
2691 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2692 {
2693 struct shader_none_priv *priv = shader_priv;
2694 const struct wined3d_gl_info *gl_info = context->gl_info;
2695
2696 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2697 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2698
2699 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2700 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2701 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2702 | (1u << WINED3D_SHADER_TYPE_HULL)
2703 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2704 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2705 }
2706
2707 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2708 const struct fragment_pipeline *fragment_pipe)
2709 {
2710 struct fragment_caps fragment_caps;
2711 void *vertex_priv, *fragment_priv;
2712 struct shader_none_priv *priv;
2713
2714 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2715 return E_OUTOFMEMORY;
2716
2717 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2718 {
2719 ERR("Failed to initialize vertex pipe.\n");
2720 HeapFree(GetProcessHeap(), 0, priv);
2721 return E_FAIL;
2722 }
2723
2724 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2725 {
2726 ERR("Failed to initialize fragment pipe.\n");
2727 vertex_pipe->vp_free(device);
2728 HeapFree(GetProcessHeap(), 0, priv);
2729 return E_FAIL;
2730 }
2731
2732 priv->vertex_pipe = vertex_pipe;
2733 priv->fragment_pipe = fragment_pipe;
2734 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2735 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2736
2737 device->vertex_priv = vertex_priv;
2738 device->fragment_priv = fragment_priv;
2739 device->shader_priv = priv;
2740
2741 return WINED3D_OK;
2742 }
2743
2744 static void shader_none_free(struct wined3d_device *device)
2745 {
2746 struct shader_none_priv *priv = device->shader_priv;
2747
2748 priv->fragment_pipe->free_private(device);
2749 priv->vertex_pipe->vp_free(device);
2750 HeapFree(GetProcessHeap(), 0, priv);
2751 }
2752
2753 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2754 {
2755 return TRUE;
2756 }
2757
2758 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2759 {
2760 /* Set the shader caps to 0 for the none shader backend */
2761 caps->vs_version = 0;
2762 caps->hs_version = 0;
2763 caps->ds_version = 0;
2764 caps->gs_version = 0;
2765 caps->ps_version = 0;
2766 caps->cs_version = 0;
2767 caps->vs_uniform_count = 0;
2768 caps->ps_uniform_count = 0;
2769 caps->ps_1x_max_value = 0.0f;
2770 caps->varying_count = 0;
2771 caps->wined3d_caps = 0;
2772 }
2773
2774 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2775 {
2776 /* We "support" every possible fixup, since we don't support any shader
2777 * model, and will never have to actually sample a texture. */
2778 return TRUE;
2779 }
2780
2781 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2782 {
2783 struct shader_none_priv *priv = shader_priv;
2784
2785 return priv->ffp_proj_control;
2786 }
2787
2788 const struct wined3d_shader_backend_ops none_shader_backend =
2789 {
2790 shader_none_handle_instruction,
2791 shader_none_select,
2792 shader_none_select_compute,
2793 shader_none_disable,
2794 shader_none_update_float_vertex_constants,
2795 shader_none_update_float_pixel_constants,
2796 shader_none_load_constants,
2797 shader_none_destroy,
2798 shader_none_alloc,
2799 shader_none_free,
2800 shader_none_allocate_context_data,
2801 shader_none_free_context_data,
2802 shader_none_init_context_state,
2803 shader_none_get_caps,
2804 shader_none_color_fixup_supported,
2805 shader_none_has_ffp_proj_control,
2806 };
2807
2808 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2809 DWORD float_const_count, enum wined3d_shader_type type, unsigned int max_version)
2810 {
2811 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2812 const struct wined3d_shader_frontend *fe;
2813 HRESULT hr;
2814 unsigned int backend_version;
2815 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2816
2817 TRACE("shader %p, byte_code %p, float_const_count %u, type %#x, max_version %u.\n",
2818 shader, byte_code, float_const_count, type, max_version);
2819
2820 list_init(&shader->constantsF);
2821 list_init(&shader->constantsB);
2822 list_init(&shader->constantsI);
2823 shader->lconst_inf_or_nan = FALSE;
2824 list_init(&reg_maps->indexable_temps);
2825
2826 fe = shader_select_frontend(*byte_code);
2827 if (!fe)
2828 {
2829 FIXME("Unable to find frontend for shader.\n");
2830 return WINED3DERR_INVALIDCALL;
2831 }
2832 shader->frontend = fe;
2833 shader->frontend_data = fe->shader_init(byte_code, &shader->output_signature);
2834 if (!shader->frontend_data)
2835 {
2836 FIXME("Failed to initialize frontend.\n");
2837 return WINED3DERR_INVALIDCALL;
2838 }
2839
2840 /* First pass: trace shader. */
2841 if (TRACE_ON(d3d_shader))
2842 shader_trace_init(fe, shader->frontend_data, byte_code);
2843
2844 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2845 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2846 &shader->output_signature, byte_code, float_const_count)))
2847 return hr;
2848
2849 if (reg_maps->shader_version.type != type)
2850 {
2851 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2852 return WINED3DERR_INVALIDCALL;
2853 }
2854 if (reg_maps->shader_version.major > max_version)
2855 {
2856 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2857 return WINED3DERR_INVALIDCALL;
2858 }
2859 switch (type)
2860 {
2861 case WINED3D_SHADER_TYPE_VERTEX:
2862 backend_version = d3d_info->limits.vs_version;
2863 break;
2864 case WINED3D_SHADER_TYPE_HULL:
2865 backend_version = d3d_info->limits.hs_version;
2866 break;
2867 case WINED3D_SHADER_TYPE_DOMAIN:
2868 backend_version = d3d_info->limits.ds_version;
2869 break;
2870 case WINED3D_SHADER_TYPE_GEOMETRY:
2871 backend_version = d3d_info->limits.gs_version;
2872 break;
2873 case WINED3D_SHADER_TYPE_PIXEL:
2874 backend_version = d3d_info->limits.ps_version;
2875 break;
2876 case WINED3D_SHADER_TYPE_COMPUTE:
2877 backend_version = d3d_info->limits.cs_version;
2878 break;
2879 default:
2880 FIXME("No backend version-checking for this shader type.\n");
2881 backend_version = 0;
2882 }
2883 if (reg_maps->shader_version.major > backend_version)
2884 {
2885 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2886 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2887 return WINED3DERR_INVALIDCALL;
2888 }
2889
2890 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength)))
2891 return E_OUTOFMEMORY;
2892 memcpy(shader->function, byte_code, shader->functionLength);
2893
2894 return WINED3D_OK;
2895 }
2896
2897 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2898 {
2899 ULONG refcount = InterlockedIncrement(&shader->ref);
2900
2901 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2902
2903 return refcount;
2904 }
2905
2906 static void wined3d_shader_destroy_object(void *object)
2907 {
2908 shader_cleanup(object);
2909 HeapFree(GetProcessHeap(), 0, object);
2910 }
2911
2912 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2913 {
2914 ULONG refcount = InterlockedDecrement(&shader->ref);
2915
2916 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2917
2918 if (!refcount)
2919 {
2920 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2921 wined3d_cs_emit_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2922 }
2923
2924 return refcount;
2925 }
2926
2927 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2928 {
2929 TRACE("shader %p.\n", shader);
2930
2931 return shader->parent;
2932 }
2933
2934 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2935 void *byte_code, UINT *byte_code_size)
2936 {
2937 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2938
2939 if (!byte_code)
2940 {
2941 *byte_code_size = shader->functionLength;
2942 return WINED3D_OK;
2943 }
2944
2945 if (*byte_code_size < shader->functionLength)
2946 {
2947 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2948 * than the required size we should write the required size and
2949 * return D3DERR_MOREDATA. That's not actually true. */
2950 return WINED3DERR_INVALIDCALL;
2951 }
2952
2953 memcpy(byte_code, shader->function, shader->functionLength);
2954
2955 return WINED3D_OK;
2956 }
2957
2958 /* Set local constants for d3d8 shaders. */
2959 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2960 UINT start_idx, const float *src_data, UINT count)
2961 {
2962 UINT end_idx = start_idx + count;
2963 UINT i;
2964
2965 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2966
2967 if (end_idx > shader->limits->constant_float)
2968 {
2969 WARN("end_idx %u > float constants limit %u.\n",
2970 end_idx, shader->limits->constant_float);
2971 end_idx = shader->limits->constant_float;
2972 }
2973
2974 for (i = start_idx; i < end_idx; ++i)
2975 {
2976 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2977 float *value;
2978 if (!lconst)
2979 return E_OUTOFMEMORY;
2980
2981 lconst->idx = i;
2982 value = (float *)lconst->value;
2983 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2984 list_add_head(&shader->constantsF, &lconst->entry);
2985
2986 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2987 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2988 {
2989 shader->lconst_inf_or_nan = TRUE;
2990 }
2991 }
2992
2993 return WINED3D_OK;
2994 }
2995
2996 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2997 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2998 {
2999 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3000 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3001 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3002 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3003 args->point_size = state->gl_primitive_type == GL_POINTS;
3004 args->per_vertex_point_size = shader->reg_maps.point_size;
3005 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
3006 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3007 : WINED3D_SHADER_TYPE_PIXEL;
3008 if (shader->reg_maps.shader_version.major >= 4)
3009 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
3010 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
3011 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
3012 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
3013 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
3014 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3015 else
3016 args->next_shader_input_count = 0;
3017 args->swizzle_map = swizzle_map;
3018 if (d3d_info->emulated_flatshading)
3019 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3020 else
3021 args->flatshading = 0;
3022 }
3023
3024 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3025 {
3026 if (usage_idx1 != usage_idx2)
3027 return FALSE;
3028 if (usage1 == usage2)
3029 return TRUE;
3030 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3031 return TRUE;
3032 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3033 return TRUE;
3034
3035 return FALSE;
3036 }
3037
3038 BOOL vshader_get_input(const struct wined3d_shader *shader,
3039 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3040 {
3041 WORD map = shader->reg_maps.input_registers;
3042 unsigned int i;
3043
3044 for (i = 0; map; map >>= 1, ++i)
3045 {
3046 if (!(map & 1)) continue;
3047
3048 if (match_usage(shader->u.vs.attributes[i].usage,
3049 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3050 {
3051 *regnum = i;
3052 return TRUE;
3053 }
3054 }
3055 return FALSE;
3056 }
3057
3058 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3059 const struct wined3d_shader_signature *src, char **signature_strings)
3060 {
3061 struct wined3d_shader_signature_element *e;
3062 unsigned int i;
3063 SIZE_T len;
3064 char *ptr;
3065
3066 if (!src->element_count)
3067 return WINED3D_OK;
3068
3069 ptr = *signature_strings;
3070
3071 dst->element_count = src->element_count;
3072 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3073 return E_OUTOFMEMORY;
3074
3075 for (i = 0; i < src->element_count; ++i)
3076 {
3077 e = &src->elements[i];
3078 dst->elements[i] = *e;
3079
3080 len = strlen(e->semantic_name);
3081 memcpy(ptr, e->semantic_name, len + 1);
3082 dst->elements[i].semantic_name = ptr;
3083 ptr += len + 1;
3084 }
3085
3086 *signature_strings = ptr;
3087
3088 return WINED3D_OK;
3089 }
3090
3091 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3092 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3093 void *parent, const struct wined3d_parent_ops *parent_ops)
3094 {
3095 struct wined3d_shader_signature_element *e;
3096 SIZE_T total, len;
3097 unsigned int i;
3098 HRESULT hr;
3099 char *ptr;
3100
3101 if (!desc->byte_code)
3102 return WINED3DERR_INVALIDCALL;
3103
3104 shader->ref = 1;
3105 shader->device = device;
3106 shader->parent = parent;
3107 shader->parent_ops = parent_ops;
3108
3109 total = 0;
3110 for (i = 0; i < desc->input_signature.element_count; ++i)
3111 {
3112 e = &desc->input_signature.elements[i];
3113 len = strlen(e->semantic_name);
3114 if (len >= ~(SIZE_T)0 - total)
3115 return E_OUTOFMEMORY;
3116
3117 total += len + 1;
3118 }
3119 for (i = 0; i < desc->output_signature.element_count; ++i)
3120 {
3121 e = &desc->output_signature.elements[i];
3122 len = strlen(e->semantic_name);
3123 if (len >= ~(SIZE_T)0 - total)
3124 return E_OUTOFMEMORY;
3125
3126 total += len + 1;
3127 }
3128 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3129 return E_OUTOFMEMORY;
3130 ptr = shader->signature_strings;
3131
3132 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3133 {
3134 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3135 return hr;
3136 }
3137 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3138 {
3139 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3140 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3141 return hr;
3142 }
3143
3144 list_init(&shader->linked_programs);
3145 list_add_head(&device->shaders, &shader->shader_list_entry);
3146
3147 if (FAILED(hr = shader_set_function(shader, desc->byte_code,
3148 float_const_count, type, desc->max_version)))
3149 {
3150 WARN("Failed to set function, hr %#x.\n", hr);
3151 shader_cleanup(shader);
3152 }
3153
3154 shader->load_local_constsF = shader->lconst_inf_or_nan;
3155
3156 return hr;
3157 }
3158
3159 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3160 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3161 {
3162 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3163 unsigned int i;
3164 HRESULT hr;
3165
3166 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3167 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3168 return hr;
3169
3170 for (i = 0; i < shader->input_signature.element_count; ++i)
3171 {
3172 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3173
3174 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3175 continue;
3176
3177 shader->u.vs.attributes[input->register_idx].usage =
3178 shader_usage_from_semantic_name(input->semantic_name);
3179 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3180 }
3181
3182 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3183 shader->load_local_constsF = TRUE;
3184
3185 return WINED3D_OK;
3186 }
3187
3188 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3189 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3190 {
3191 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3192 }
3193
3194 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3195 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3196 {
3197 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3198 }
3199
3200 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3201 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3202 {
3203 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3204 }
3205
3206 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3207 struct gs_compile_args *args)
3208 {
3209 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3210 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3211 }
3212
3213 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3214 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3215 {
3216 const struct wined3d_gl_info *gl_info = context->gl_info;
3217 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3218 const struct wined3d_texture *texture;
3219 UINT i;
3220
3221 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3222 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3223 {
3224 static unsigned int warned = 0;
3225
3226 args->srgb_correction = 1;
3227 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3228 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3229 "support, expect rendering artifacts.\n");
3230 }
3231
3232 if (shader->reg_maps.shader_version.major == 1
3233 && shader->reg_maps.shader_version.minor <= 3)
3234 {
3235 for (i = 0; i < shader->limits->sampler; ++i)
3236 {
3237 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3238
3239 if (flags & WINED3D_TTFF_PROJECTED)
3240 {
3241 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3242
3243 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3244 {
3245 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3246 unsigned int j;
3247 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3248 DWORD max_valid = WINED3D_TTFF_COUNT4;
3249
3250 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3251 {
3252 struct wined3d_vertex_declaration_element *element =
3253 &state->vertex_declaration->elements[j];
3254
3255 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3256 && element->usage_idx == index)
3257 {
3258 max_valid = element->format->component_count;
3259 break;
3260 }
3261 }
3262 if (!tex_transform || tex_transform > max_valid)
3263 {
3264 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3265 tex_transform, max_valid);
3266 tex_transform = max_valid;
3267 }
3268 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3269 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3270 && tex_transform > WINED3D_TTFF_COUNT2)
3271 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3272 && tex_transform > WINED3D_TTFF_COUNT3))
3273 tex_transform |= WINED3D_PSARGS_PROJECTED;
3274 else
3275 {
3276 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3277 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3278 i, tex_transform, resource_type);
3279 }
3280 }
3281 else
3282 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3283
3284 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3285 }
3286 }
3287 }
3288 if (shader->reg_maps.shader_version.major == 1
3289 && shader->reg_maps.shader_version.minor <= 4)
3290 {
3291 for (i = 0; i < shader->limits->sampler; ++i)
3292 {
3293 const struct wined3d_texture *texture = state->textures[i];
3294
3295 if (!shader->reg_maps.resource_info[i].type)
3296 continue;
3297
3298 /* Treat unbound textures as 2D. The dummy texture will provide
3299 * the proper sample value. The tex_types bitmap defaults to
3300 * 2D because of the memset. */
3301 if (!texture)
3302 continue;
3303
3304 switch (texture->target)
3305 {
3306 /* RECT textures are distinguished from 2D textures via np2_fixup */
3307 case GL_TEXTURE_RECTANGLE_ARB:
3308 case GL_TEXTURE_2D:
3309 break;
3310
3311 case GL_TEXTURE_3D:
3312 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3313 break;
3314
3315 case GL_TEXTURE_CUBE_MAP_ARB:
3316 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3317 break;
3318 }
3319 }
3320 }
3321
3322 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3323 {
3324 if (!shader->reg_maps.resource_info[i].type)
3325 continue;
3326
3327 texture = state->textures[i];
3328 if (!texture)
3329 {
3330 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3331 continue;
3332 }
3333 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3334 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3335 else
3336 args->color_fixup[i] = texture->resource.format->color_fixup;
3337
3338 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3339 args->shadow |= 1u << i;
3340
3341 /* Flag samplers that need NP2 texcoord fixup. */
3342 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3343 args->np2_fixup |= (1u << i);
3344 }
3345
3346 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3347 if (shader->reg_maps.shader_version.major >= 4)
3348 args->shadow = 0;
3349
3350 if (shader->reg_maps.shader_version.major >= 3)
3351 {
3352 if (position_transformed)
3353 args->vp_mode = pretransformed;
3354 else if (use_vs(state))
3355 args->vp_mode = vertexshader;
3356 else
3357 args->vp_mode = fixedfunction;
3358 args->fog = WINED3D_FFP_PS_FOG_OFF;
3359 }
3360 else
3361 {
3362 args->vp_mode = vertexshader;
3363 if (state->render_states[WINED3D_RS_FOGENABLE])
3364 {
3365 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3366 {
3367 case WINED3D_FOG_NONE:
3368 if (position_transformed || use_vs(state))
3369 {
3370 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3371 break;
3372 }
3373
3374 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3375 {
3376 case WINED3D_FOG_NONE: /* Fall through. */
3377 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3378 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3379 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3380 }
3381 break;
3382
3383 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3384 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3385 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3386 }
3387 }
3388 else
3389 {
3390 args->fog = WINED3D_FFP_PS_FOG_OFF;
3391 }
3392 }
3393
3394 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3395 {
3396 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3397
3398 args->texcoords_initialized = 0;
3399 for (i = 0; i < MAX_TEXTURES; ++i)
3400 {
3401 if (vs)
3402 {
3403 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3404 args->texcoords_initialized |= 1u << i;
3405 }
3406 else
3407 {
3408 const struct wined3d_stream_info *si = &context->stream_info;
3409 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3410
3411 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3412 & WINED3D_FFP_TCI_MASK
3413 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3414 args->texcoords_initialized |= 1u << i;
3415 }
3416 }
3417 }
3418 else
3419 {
3420 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3421 }
3422
3423 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3424 && state->gl_primitive_type == GL_POINTS;
3425
3426 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3427 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3428 else
3429 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3430 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3431 : WINED3D_CMP_ALWAYS) - 1;
3432
3433 if (d3d_info->emulated_flatshading)
3434 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3435
3436 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3437 ? context->render_offscreen : 0;
3438 }
3439
3440 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3441 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3442 {
3443 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3444 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3445 HRESULT hr;
3446
3447 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3448 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3449 return hr;
3450
3451 for (i = 0; i < MAX_REG_INPUT; ++i)
3452 {
3453 if (shader->u.ps.input_reg_used[i])
3454 {
3455 ++num_regs_used;
3456 highest_reg_used = i;
3457 }
3458 }
3459
3460 /* Don't do any register mapping magic if it is not needed, or if we can't
3461 * achieve anything anyway */
3462 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3463 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3464 || shader->reg_maps.shader_version.major >= 4)
3465 {
3466 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3467 {
3468 /* This happens with relative addressing. The input mapper function
3469 * warns about this if the higher registers are declared too, so
3470 * don't write a FIXME here */
3471 WARN("More varying registers used than supported\n");
3472 }
3473
3474 for (i = 0; i < MAX_REG_INPUT; ++i)
3475 {
3476 shader->u.ps.input_reg_map[i] = i;
3477 }
3478
3479 shader->u.ps.declared_in_count = highest_reg_used + 1;
3480 }
3481 else
3482 {
3483 shader->u.ps.declared_in_count = 0;
3484 for (i = 0; i < MAX_REG_INPUT; ++i)
3485 {
3486 if (shader->u.ps.input_reg_used[i])
3487 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3488 else shader->u.ps.input_reg_map[i] = ~0U;
3489 }
3490 }
3491
3492 return WINED3D_OK;
3493 }
3494
3495 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3496 {
3497 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3498 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3499 unsigned int i;
3500
3501 if (reg_maps->shader_version.major != 1) return;
3502
3503 for (i = 0; i < shader->limits->sampler; ++i)
3504 {
3505 /* We don't sample from this sampler. */
3506 if (!resource_info[i].type)
3507 continue;
3508
3509 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3510 {
3511 case WINED3D_SHADER_TEX_2D:
3512 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3513 break;
3514
3515 case WINED3D_SHADER_TEX_3D:
3516 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3517 break;
3518
3519 case WINED3D_SHADER_TEX_CUBE:
3520 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3521 break;
3522 }
3523 }
3524 }
3525
3526 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3527 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3528 {
3529 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3530 }
3531
3532 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3533 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3534 {
3535 struct wined3d_shader *object;
3536 HRESULT hr;
3537
3538 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3539 device, desc, parent, parent_ops, shader);
3540
3541 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3542 return E_OUTOFMEMORY;
3543
3544 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3545 {
3546 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3547 HeapFree(GetProcessHeap(), 0, object);
3548 return hr;
3549 }
3550
3551 TRACE("Created compute shader %p.\n", object);
3552 *shader = object;
3553
3554 return WINED3D_OK;
3555 }
3556
3557 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3558 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3559 {
3560 struct wined3d_shader *object;
3561 HRESULT hr;
3562
3563 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3564 device, desc, parent, parent_ops, shader);
3565
3566 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3567 return E_OUTOFMEMORY;
3568
3569 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3570 {
3571 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3572 HeapFree(GetProcessHeap(), 0, object);
3573 return hr;
3574 }
3575
3576 TRACE("Created domain shader %p.\n", object);
3577 *shader = object;
3578
3579 return WINED3D_OK;
3580 }
3581
3582 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3583 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3584 {
3585 struct wined3d_shader *object;
3586 HRESULT hr;
3587
3588 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3589 device, desc, parent, parent_ops, shader);
3590
3591 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3592 if (!object)
3593 return E_OUTOFMEMORY;
3594
3595 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3596 {
3597 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3598 HeapFree(GetProcessHeap(), 0, object);
3599 return hr;
3600 }
3601
3602 TRACE("Created geometry shader %p.\n", object);
3603 *shader = object;
3604
3605 return WINED3D_OK;
3606 }
3607
3608 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3609 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3610 {
3611 struct wined3d_shader *object;
3612 HRESULT hr;
3613
3614 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3615 device, desc, parent, parent_ops, shader);
3616
3617 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3618 return E_OUTOFMEMORY;
3619
3620 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3621 {
3622 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3623 HeapFree(GetProcessHeap(), 0, object);
3624 return hr;
3625 }
3626
3627 TRACE("Created hull shader %p.\n", object);
3628 *shader = object;
3629
3630 return WINED3D_OK;
3631 }
3632
3633 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3634 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3635 {
3636 struct wined3d_shader *object;
3637 HRESULT hr;
3638
3639 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3640 device, desc, parent, parent_ops, shader);
3641
3642 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3643 if (!object)
3644 return E_OUTOFMEMORY;
3645
3646 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3647 {
3648 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3649 HeapFree(GetProcessHeap(), 0, object);
3650 return hr;
3651 }
3652
3653 TRACE("Created pixel shader %p.\n", object);
3654 *shader = object;
3655
3656 return WINED3D_OK;
3657 }
3658
3659 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3660 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3661 {
3662 struct wined3d_shader *object;
3663 HRESULT hr;
3664
3665 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3666 device, desc, parent, parent_ops, shader);
3667
3668 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3669 if (!object)
3670 return E_OUTOFMEMORY;
3671
3672 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3673 {
3674 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3675 HeapFree(GetProcessHeap(), 0, object);
3676 return hr;
3677 }
3678
3679 TRACE("Created vertex shader %p.\n", object);
3680 *shader = object;
3681
3682 return WINED3D_OK;
3683 }
Windows API implementation