?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
reg/tests: Test import with non-standard registry file headers.
[wine.git] / dlls / wined3d / shader.c
blob3085248f91fe26ef63fdc063963275a62223278e
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_IMAX */ "imm_atomic_imax",
167 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
168 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
169 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
170 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
171 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
172 /* WINED3DSIH_IMUL */ "imul",
173 /* WINED3DSIH_INE */ "ine",
174 /* WINED3DSIH_INEG */ "ineg",
175 /* WINED3DSIH_ISHL */ "ishl",
176 /* WINED3DSIH_ISHR */ "ishr",
177 /* WINED3DSIH_ITOF */ "itof",
178 /* WINED3DSIH_LABEL */ "label",
179 /* WINED3DSIH_LD */ "ld",
180 /* WINED3DSIH_LD2DMS */ "ld2dms",
181 /* WINED3DSIH_LD_RAW */ "ld_raw",
182 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
183 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
184 /* WINED3DSIH_LIT */ "lit",
185 /* WINED3DSIH_LOD */ "lod",
186 /* WINED3DSIH_LOG */ "log",
187 /* WINED3DSIH_LOGP */ "logp",
188 /* WINED3DSIH_LOOP */ "loop",
189 /* WINED3DSIH_LRP */ "lrp",
190 /* WINED3DSIH_LT */ "lt",
191 /* WINED3DSIH_M3x2 */ "m3x2",
192 /* WINED3DSIH_M3x3 */ "m3x3",
193 /* WINED3DSIH_M3x4 */ "m3x4",
194 /* WINED3DSIH_M4x3 */ "m4x3",
195 /* WINED3DSIH_M4x4 */ "m4x4",
196 /* WINED3DSIH_MAD */ "mad",
197 /* WINED3DSIH_MAX */ "max",
198 /* WINED3DSIH_MIN */ "min",
199 /* WINED3DSIH_MOV */ "mov",
200 /* WINED3DSIH_MOVA */ "mova",
201 /* WINED3DSIH_MOVC */ "movc",
202 /* WINED3DSIH_MUL */ "mul",
203 /* WINED3DSIH_NE */ "ne",
204 /* WINED3DSIH_NOP */ "nop",
205 /* WINED3DSIH_NOT */ "not",
206 /* WINED3DSIH_NRM */ "nrm",
207 /* WINED3DSIH_OR */ "or",
208 /* WINED3DSIH_PHASE */ "phase",
209 /* WINED3DSIH_POW */ "pow",
210 /* WINED3DSIH_RCP */ "rcp",
211 /* WINED3DSIH_REP */ "rep",
212 /* WINED3DSIH_RESINFO */ "resinfo",
213 /* WINED3DSIH_RET */ "ret",
214 /* WINED3DSIH_ROUND_NE */ "round_ne",
215 /* WINED3DSIH_ROUND_NI */ "round_ni",
216 /* WINED3DSIH_ROUND_PI */ "round_pi",
217 /* WINED3DSIH_ROUND_Z */ "round_z",
218 /* WINED3DSIH_RSQ */ "rsq",
219 /* WINED3DSIH_SAMPLE */ "sample",
220 /* WINED3DSIH_SAMPLE_B */ "sample_b",
221 /* WINED3DSIH_SAMPLE_C */ "sample_c",
222 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
223 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
224 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
225 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
226 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
227 /* WINED3DSIH_SETP */ "setp",
228 /* WINED3DSIH_SGE */ "sge",
229 /* WINED3DSIH_SGN */ "sgn",
230 /* WINED3DSIH_SINCOS */ "sincos",
231 /* WINED3DSIH_SLT */ "slt",
232 /* WINED3DSIH_SQRT */ "sqrt",
233 /* WINED3DSIH_STORE_RAW */ "store_raw",
234 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
235 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
236 /* WINED3DSIH_SUB */ "sub",
237 /* WINED3DSIH_SWAPC */ "swapc",
238 /* WINED3DSIH_SWITCH */ "switch",
239 /* WINED3DSIH_SYNC */ "sync",
240 /* WINED3DSIH_TEX */ "texld",
241 /* WINED3DSIH_TEXBEM */ "texbem",
242 /* WINED3DSIH_TEXBEML */ "texbeml",
243 /* WINED3DSIH_TEXCOORD */ "texcrd",
244 /* WINED3DSIH_TEXDEPTH */ "texdepth",
245 /* WINED3DSIH_TEXDP3 */ "texdp3",
246 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
247 /* WINED3DSIH_TEXKILL */ "texkill",
248 /* WINED3DSIH_TEXLDD */ "texldd",
249 /* WINED3DSIH_TEXLDL */ "texldl",
250 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
251 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
252 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
253 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
254 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
255 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
256 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
257 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
258 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
259 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
260 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
261 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
262 /* WINED3DSIH_UBFE */ "ubfe",
263 /* WINED3DSIH_UDIV */ "udiv",
264 /* WINED3DSIH_UGE */ "uge",
265 /* WINED3DSIH_ULT */ "ult",
266 /* WINED3DSIH_UMAX */ "umax",
267 /* WINED3DSIH_UMIN */ "umin",
268 /* WINED3DSIH_USHR */ "ushr",
269 /* WINED3DSIH_UTOF */ "utof",
270 /* WINED3DSIH_XOR */ "xor",
271 };
272
273 static const char * const semantic_names[] =
274 {
275 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
276 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
277 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
278 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
279 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
280 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
281 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
282 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
283 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
284 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
285 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
286 /* WINED3D_DECL_USAGE_FOG */ "FOG",
287 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
288 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
289 };
290
291 static const struct
292 {
293 enum wined3d_sysval_semantic sysval_semantic;
294 const char *sysval_name;
295 }
296 sysval_semantic_names[] =
297 {
298 {WINED3D_SV_POSITION, "SV_Position"},
299 {WINED3D_SV_CLIP_DISTANCE, "SV_ClipDistance"},
300 {WINED3D_SV_CULL_DISTANCE, "SV_CullDistance"},
301 {WINED3D_SV_RENDER_TARGET_ARRAY_INDEX, "SV_RenderTargetArrayIndex"},
302 {WINED3D_SV_VIEWPORT_ARRAY_INDEX, "SV_ViewportArrayIndex"},
303 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
304 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
305 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
306 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
307 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
308 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
309 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
310 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
311 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
312 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
313 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
314 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
315 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
316 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
317 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
318 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
319 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
320 };
321
322 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
323 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
324
325 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
326 {
327 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
328 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
329
330 return shader_opcode_names[handler_idx];
331 }
332
333 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
334 {
335 if (usage >= ARRAY_SIZE(semantic_names))
336 {
337 FIXME("Unrecognized usage %#x.\n", usage);
338 return "UNRECOGNIZED";
339 }
340
341 return semantic_names[usage];
342 }
343
344 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
345 {
346 unsigned int i;
347
348 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
349 {
350 if (!strcmp(name, semantic_names[i]))
351 return i;
352 }
353
354 return ~0U;
355 }
356
357 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
358 {
359 switch (usage)
360 {
361 case WINED3D_DECL_USAGE_POSITION:
362 return WINED3D_SV_POSITION;
363 default:
364 return 0;
365 }
366 }
367
368 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
369 {
370 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
371 }
372
373 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
374 const struct wined3d_shader_semantic *s)
375 {
376 e->semantic_name = shader_semantic_name_from_usage(s->usage);
377 e->semantic_idx = s->usage_idx;
378 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
379 e->component_type = WINED3D_TYPE_FLOAT;
380 e->register_idx = s->reg.reg.idx[0].offset;
381 e->mask = s->reg.write_mask;
382 }
383
384 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
385 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
386 {
387 e->semantic_name = shader_semantic_name_from_usage(usage);
388 e->semantic_idx = usage_idx;
389 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
390 e->component_type = WINED3D_TYPE_FLOAT;
391 e->register_idx = reg_idx;
392 e->mask = write_mask;
393 }
394
395 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
396 {
397 switch (format)
398 {
399 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
400 return &sm1_shader_frontend;
401
402 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
403 return &sm4_shader_frontend;
404
405 default:
406 WARN("Invalid byte code format %#x specified.\n", format);
407 return NULL;
408 }
409 }
410
411 void string_buffer_clear(struct wined3d_string_buffer *buffer)
412 {
413 buffer->buffer[0] = '\0';
414 buffer->content_size = 0;
415 }
416
417 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
418 {
419 buffer->buffer_size = 32;
420 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
421 {
422 ERR("Failed to allocate shader buffer memory.\n");
423 return FALSE;
424 }
425
426 string_buffer_clear(buffer);
427 return TRUE;
428 }
429
430 void string_buffer_free(struct wined3d_string_buffer *buffer)
431 {
432 HeapFree(GetProcessHeap(), 0, buffer->buffer);
433 }
434
435 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
436 {
437 char *new_buffer;
438 unsigned int new_buffer_size = buffer->buffer_size * 2;
439
440 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
441 new_buffer_size *= 2;
442 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
443 {
444 ERR("Failed to grow buffer.\n");
445 buffer->buffer[buffer->content_size] = '\0';
446 return FALSE;
447 }
448 buffer->buffer = new_buffer;
449 buffer->buffer_size = new_buffer_size;
450 return TRUE;
451 }
452
453 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
454 {
455 unsigned int rem;
456 int rc;
457
458 rem = buffer->buffer_size - buffer->content_size;
459 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
460 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
461 return rc;
462
463 buffer->content_size += rc;
464 return 0;
465 }
466
467 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
468 {
469 va_list args;
470 int ret;
471
472 for (;;)
473 {
474 va_start(args, format);
475 ret = shader_vaddline(buffer, format, args);
476 va_end(args);
477 if (!ret)
478 return ret;
479 if (!string_buffer_resize(buffer, ret))
480 return -1;
481 }
482 }
483
484 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
485 {
486 struct wined3d_string_buffer *buffer;
487
488 if (list_empty(&list->list))
489 {
490 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
491 if (!buffer || !string_buffer_init(buffer))
492 {
493 ERR("Couldn't allocate buffer for temporary string.\n");
494 HeapFree(GetProcessHeap(), 0, buffer);
495 return NULL;
496 }
497 }
498 else
499 {
500 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
501 list_remove(&buffer->entry);
502 }
503 string_buffer_clear(buffer);
504 return buffer;
505 }
506
507 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
508 {
509 if (!buffer)
510 return 0;
511 string_buffer_clear(buffer);
512 return shader_vaddline(buffer, format, args);
513 }
514
515 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
516 {
517 va_list args;
518 int ret;
519
520 for (;;)
521 {
522 va_start(args, format);
523 ret = string_buffer_vsprintf(buffer, format, args);
524 va_end(args);
525 if (!ret)
526 return;
527 if (!string_buffer_resize(buffer, ret))
528 return;
529 }
530 }
531
532 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
533 {
534 if (!buffer)
535 return;
536 list_add_head(&list->list, &buffer->entry);
537 }
538
539 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
540 {
541 list_init(&list->list);
542 }
543
544 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
545 {
546 struct wined3d_string_buffer *buffer, *buffer_next;
547
548 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
549 {
550 string_buffer_free(buffer);
551 HeapFree(GetProcessHeap(), 0, buffer);
552 }
553 list_init(&list->list);
554 }
555
556 /* Convert floating point offset relative to a register file to an absolute
557 * offset for float constants. */
558 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
559 {
560 switch (register_type)
561 {
562 case WINED3DSPR_CONST: return register_idx;
563 case WINED3DSPR_CONST2: return 2048 + register_idx;
564 case WINED3DSPR_CONST3: return 4096 + register_idx;
565 case WINED3DSPR_CONST4: return 6144 + register_idx;
566 default:
567 FIXME("Unsupported register type: %u.\n", register_type);
568 return register_idx;
569 }
570 }
571
572 static void shader_delete_constant_list(struct list *clist)
573 {
574 struct wined3d_shader_lconst *constant, *constant_next;
575
576 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
577 HeapFree(GetProcessHeap(), 0, constant);
578 list_init(clist);
579 }
580
581 static void shader_set_limits(struct wined3d_shader *shader)
582 {
583 static const struct limits_entry
584 {
585 unsigned int min_version;
586 unsigned int max_version;
587 struct wined3d_shader_limits limits;
588 }
589 vs_limits[] =
590 {
591 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
592 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
593 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
594 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
595 * even though they are capable of supporting much more (GL
596 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
597 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
598 * shaders to 256. */
599 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
600 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
601 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
602 {0}
603 },
604 hs_limits[] =
605 {
606 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
607 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
608 },
609 ds_limits[] =
610 {
611 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
612 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
613 },
614 gs_limits[] =
615 {
616 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
617 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
618 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
619 {0}
620 },
621 ps_limits[] =
622 {
623 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
624 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
625 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
626 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
627 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
628 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
629 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
630 {0}
631 },
632 cs_limits[] =
633 {
634 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
635 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
636 };
637 const struct limits_entry *limits_array;
638 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
639 shader->reg_maps.shader_version.minor);
640 int i = 0;
641
642 switch (shader->reg_maps.shader_version.type)
643 {
644 default:
645 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
646 /* Fall-through. */
647 case WINED3D_SHADER_TYPE_VERTEX:
648 limits_array = vs_limits;
649 break;
650 case WINED3D_SHADER_TYPE_HULL:
651 limits_array = hs_limits;
652 break;
653 case WINED3D_SHADER_TYPE_DOMAIN:
654 limits_array = ds_limits;
655 break;
656 case WINED3D_SHADER_TYPE_GEOMETRY:
657 limits_array = gs_limits;
658 break;
659 case WINED3D_SHADER_TYPE_PIXEL:
660 limits_array = ps_limits;
661 break;
662 case WINED3D_SHADER_TYPE_COMPUTE:
663 limits_array = cs_limits;
664 break;
665 }
666
667 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
668 {
669 if (shader_version <= limits_array[i].max_version)
670 {
671 shader->limits = &limits_array[i].limits;
672 break;
673 }
674 ++i;
675 }
676 if (!shader->limits)
677 {
678 FIXME("Unexpected shader version \"%u.%u\".\n",
679 shader->reg_maps.shader_version.major,
680 shader->reg_maps.shader_version.minor);
681 shader->limits = &limits_array[max(0, i - 1)].limits;
682 }
683 }
684
685 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
686 {
687 DWORD idx, shift;
688 idx = bit >> 5;
689 shift = bit & 0x1f;
690 bitmap[idx] |= (1u << shift);
691 }
692
693 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
694 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
695 {
696 switch (reg->type)
697 {
698 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
699 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
700 reg_maps->texcoord |= 1u << reg->idx[0].offset;
701 else
702 reg_maps->address |= 1u << reg->idx[0].offset;
703 break;
704
705 case WINED3DSPR_TEMP:
706 reg_maps->temporary |= 1u << reg->idx[0].offset;
707 break;
708
709 case WINED3DSPR_INPUT:
710 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
711 {
712 /* If relative addressing is used, we must assume that all
713 * registers are used. Even if it is a construct like v3[aL],
714 * we can't assume that v0, v1 and v2 aren't read because aL
715 * can be negative. */
716 if (reg->idx[0].rel_addr)
717 shader->u.ps.input_reg_used = ~0u;
718 else
719 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
720 }
721 else
722 reg_maps->input_registers |= 1u << reg->idx[0].offset;
723 break;
724
725 case WINED3DSPR_RASTOUT:
726 if (reg->idx[0].offset == 1)
727 reg_maps->fog = 1;
728 if (reg->idx[0].offset == 2)
729 reg_maps->point_size = 1;
730 break;
731
732 case WINED3DSPR_MISCTYPE:
733 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
734 {
735 if (!reg->idx[0].offset)
736 reg_maps->vpos = 1;
737 else if (reg->idx[0].offset == 1)
738 reg_maps->usesfacing = 1;
739 }
740 break;
741
742 case WINED3DSPR_CONST:
743 if (reg->idx[0].rel_addr)
744 {
745 if (reg->idx[0].offset < reg_maps->min_rel_offset)
746 reg_maps->min_rel_offset = reg->idx[0].offset;
747 if (reg->idx[0].offset > reg_maps->max_rel_offset)
748 reg_maps->max_rel_offset = reg->idx[0].offset;
749 reg_maps->usesrelconstF = TRUE;
750 }
751 else
752 {
753 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
754 {
755 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
756 return FALSE;
757 }
758 else
759 {
760 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
761 }
762 }
763 break;
764
765 case WINED3DSPR_CONSTINT:
766 if (reg->idx[0].offset >= shader->limits->constant_int)
767 {
768 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
769 return FALSE;
770 }
771 else
772 {
773 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
774 }
775 break;
776
777 case WINED3DSPR_CONSTBOOL:
778 if (reg->idx[0].offset >= shader->limits->constant_bool)
779 {
780 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
781 return FALSE;
782 }
783 else
784 {
785 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
786 }
787 break;
788
789 case WINED3DSPR_COLOROUT:
790 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
791 break;
792
793 default:
794 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
795 reg->type, reg->idx[0].offset, reg->idx[1].offset);
796 break;
797 }
798 return TRUE;
799 }
800
801 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
802 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
803 {
804 struct wined3d_shader_sampler_map_entry *entries, *entry;
805 struct wined3d_shader_sampler_map *map;
806 unsigned int i;
807
808 map = &reg_maps->sampler_map;
809 entries = map->entries;
810 for (i = 0; i < map->count; ++i)
811 {
812 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
813 return;
814 }
815
816 if (!map->size)
817 {
818 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
819 {
820 ERR("Failed to allocate sampler map entries.\n");
821 return;
822 }
823 map->size = 4;
824 map->entries = entries;
825 }
826 else if (map->count == map->size)
827 {
828 size_t new_size = map->size * 2;
829
830 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
831 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
832 {
833 ERR("Failed to resize sampler map entries.\n");
834 return;
835 }
836 map->size = new_size;
837 map->entries = entries;
838 }
839
840 entry = &entries[map->count++];
841 entry->resource_idx = resource_idx;
842 entry->sampler_idx = sampler_idx;
843 entry->bind_idx = bind_idx;
844 }
845
846 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
847 {
848 switch (instr)
849 {
850 case WINED3DSIH_M4x4:
851 case WINED3DSIH_M3x4:
852 return param == 1 ? 3 : 0;
853
854 case WINED3DSIH_M4x3:
855 case WINED3DSIH_M3x3:
856 return param == 1 ? 2 : 0;
857
858 case WINED3DSIH_M3x2:
859 return param == 1 ? 1 : 0;
860
861 default:
862 return 0;
863 }
864 }
865
866 /* Note that this does not count the loop register as an address register. */
867 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
868 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
869 struct wined3d_shader_signature *output_signature, DWORD constf_size)
870 {
871 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
872 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
873 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
874 void *fe_data = shader->frontend_data;
875 struct wined3d_shader_version shader_version;
876 const DWORD *ptr;
877 unsigned int i;
878
879 memset(reg_maps, 0, sizeof(*reg_maps));
880 memset(input_signature_elements, 0, sizeof(input_signature_elements));
881 memset(output_signature_elements, 0, sizeof(output_signature_elements));
882 reg_maps->min_rel_offset = ~0U;
883 list_init(&reg_maps->indexable_temps);
884
885 fe->shader_read_header(fe_data, &ptr, &shader_version);
886 reg_maps->shader_version = shader_version;
887
888 shader_set_limits(shader);
889
890 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
891 sizeof(*reg_maps->constf))))
892 {
893 ERR("Failed to allocate constant map memory.\n");
894 return E_OUTOFMEMORY;
895 }
896
897 while (!fe->shader_is_end(fe_data, &ptr))
898 {
899 struct wined3d_shader_instruction ins;
900
901 /* Fetch opcode. */
902 fe->shader_read_instruction(fe_data, &ptr, &ins);
903
904 /* Unhandled opcode, and its parameters. */
905 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
906 {
907 TRACE("Skipping unrecognized instruction.\n");
908 continue;
909 }
910
911 /* Handle declarations. */
912 if (ins.handler_idx == WINED3DSIH_DCL
913 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
914 {
915 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
916 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
917
918 switch (semantic->reg.reg.type)
919 {
920 /* Mark input registers used. */
921 case WINED3DSPR_INPUT:
922 if (reg_idx >= MAX_REG_INPUT)
923 {
924 ERR("Invalid input register index %u.\n", reg_idx);
925 break;
926 }
927 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
928 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
929 return WINED3DERR_INVALIDCALL;
930 reg_maps->input_registers |= 1u << reg_idx;
931 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
932 break;
933
934 /* Vertex shader: mark 3.0 output registers used, save token. */
935 case WINED3DSPR_OUTPUT:
936 if (reg_idx >= MAX_REG_OUTPUT)
937 {
938 ERR("Invalid output register index %u.\n", reg_idx);
939 break;
940 }
941 reg_maps->output_registers |= 1u << reg_idx;
942 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
943 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
944 reg_maps->fog = 1;
945 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
946 reg_maps->point_size = 1;
947 break;
948
949 case WINED3DSPR_SAMPLER:
950 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
951 case WINED3DSPR_RESOURCE:
952 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
953 {
954 ERR("Invalid resource index %u.\n", reg_idx);
955 break;
956 }
957 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
958 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
959 break;
960
961 case WINED3DSPR_UAV:
962 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
963 {
964 ERR("Invalid UAV resource index %u.\n", reg_idx);
965 break;
966 }
967 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
968 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
969 break;
970
971 default:
972 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
973 break;
974 }
975 }
976 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
977 {
978 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
979 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
980 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
981 else
982 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
983 }
984 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
985 {
986 if (reg_maps->icb)
987 FIXME("Multiple immediate constant buffers.\n");
988 reg_maps->icb = ins.declaration.icb;
989 }
990 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
991 {
992 struct wined3d_shader_indexable_temp *reg;
993
994 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
995 return E_OUTOFMEMORY;
996
997 *reg = ins.declaration.indexable_temp;
998 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
999 }
1000 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1001 {
1002 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1003 shader->u.gs.input_type = ins.declaration.primitive_type;
1004 else
1005 FIXME("Invalid instruction %#x for shader type %#x.\n",
1006 ins.handler_idx, shader_version.type);
1007 }
1008 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1009 {
1010 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1011 shader->u.gs.output_type = ins.declaration.primitive_type;
1012 else
1013 FIXME("Invalid instruction %#x for shader type %#x.\n",
1014 ins.handler_idx, shader_version.type);
1015 }
1016 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1017 {
1018 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1019 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1020 {
1021 ERR("Invalid resource index %u.\n", reg_idx);
1022 break;
1023 }
1024 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1025 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1026 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1027 }
1028 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1029 {
1030 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1031 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1032 }
1033 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1034 {
1035 reg_maps->temporary_count = ins.declaration.count;
1036 }
1037 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1038 {
1039 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1040 {
1041 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1042 }
1043 else
1044 {
1045 FIXME("Invalid instruction %#x for shader type %#x.\n",
1046 ins.handler_idx, shader_version.type);
1047 }
1048 }
1049 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1050 {
1051 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1052 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1053 {
1054 ERR("Invalid UAV resource index %u.\n", reg_idx);
1055 break;
1056 }
1057 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1058 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1059 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1060 }
1061 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1062 {
1063 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1064 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1065 {
1066 ERR("Invalid UAV resource index %u.\n", reg_idx);
1067 break;
1068 }
1069 if (ins.flags)
1070 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1071 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1072 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1073 reg_maps->uav_resource_info[reg_idx].flags = 0;
1074 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1075 }
1076 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1077 {
1078 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1079 shader->u.gs.vertices_out = ins.declaration.count;
1080 else
1081 FIXME("Invalid instruction %#x for shader type %#x.\n",
1082 ins.handler_idx, shader_version.type);
1083 }
1084 else if (ins.handler_idx == WINED3DSIH_DEF)
1085 {
1086 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1087 float *value;
1088 if (!lconst) return E_OUTOFMEMORY;
1089
1090 lconst->idx = ins.dst[0].reg.idx[0].offset;
1091 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1092 value = (float *)lconst->value;
1093
1094 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1095 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1096 {
1097 if (value[0] < -1.0f) value[0] = -1.0f;
1098 else if (value[0] > 1.0f) value[0] = 1.0f;
1099 if (value[1] < -1.0f) value[1] = -1.0f;
1100 else if (value[1] > 1.0f) value[1] = 1.0f;
1101 if (value[2] < -1.0f) value[2] = -1.0f;
1102 else if (value[2] > 1.0f) value[2] = 1.0f;
1103 if (value[3] < -1.0f) value[3] = -1.0f;
1104 else if (value[3] > 1.0f) value[3] = 1.0f;
1105 }
1106
1107 list_add_head(&shader->constantsF, &lconst->entry);
1108
1109 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1110 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1111 {
1112 shader->lconst_inf_or_nan = TRUE;
1113 }
1114 }
1115 else if (ins.handler_idx == WINED3DSIH_DEFI)
1116 {
1117 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1118 if (!lconst) return E_OUTOFMEMORY;
1119
1120 lconst->idx = ins.dst[0].reg.idx[0].offset;
1121 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1122
1123 list_add_head(&shader->constantsI, &lconst->entry);
1124 reg_maps->local_int_consts |= (1u << lconst->idx);
1125 }
1126 else if (ins.handler_idx == WINED3DSIH_DEFB)
1127 {
1128 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1129 if (!lconst) return E_OUTOFMEMORY;
1130
1131 lconst->idx = ins.dst[0].reg.idx[0].offset;
1132 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1133
1134 list_add_head(&shader->constantsB, &lconst->entry);
1135 reg_maps->local_bool_consts |= (1u << lconst->idx);
1136 }
1137 /* For subroutine prototypes. */
1138 else if (ins.handler_idx == WINED3DSIH_LABEL)
1139 {
1140 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1141 }
1142 /* Set texture, address, temporary registers. */
1143 else
1144 {
1145 BOOL color0_mov = FALSE;
1146 unsigned int i;
1147
1148 /* This will loop over all the registers and try to
1149 * make a bitmask of the ones we're interested in.
1150 *
1151 * Relative addressing tokens are ignored, but that's
1152 * okay, since we'll catch any address registers when
1153 * they are initialized (required by spec). */
1154 for (i = 0; i < ins.dst_count; ++i)
1155 {
1156 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1157 shader_version.type, constf_size))
1158 return WINED3DERR_INVALIDCALL;
1159
1160 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1161 {
1162 UINT idx = ins.dst[i].reg.idx[0].offset;
1163
1164 switch (ins.dst[i].reg.type)
1165 {
1166 case WINED3DSPR_RASTOUT:
1167 if (shader_version.major >= 3)
1168 break;
1169 switch (idx)
1170 {
1171 case 0: /* oPos */
1172 reg_maps->output_registers |= 1u << 10;
1173 shader_signature_from_usage(&output_signature_elements[10],
1174 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1175 break;
1176
1177 case 1: /* oFog */
1178 reg_maps->output_registers |= 1u << 11;
1179 shader_signature_from_usage(&output_signature_elements[11],
1180 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1181 break;
1182
1183 case 2: /* oPts */
1184 reg_maps->output_registers |= 1u << 11;
1185 shader_signature_from_usage(&output_signature_elements[11],
1186 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1187 break;
1188 }
1189 break;
1190
1191 case WINED3DSPR_ATTROUT:
1192 if (shader_version.major >= 3)
1193 break;
1194 if (idx < 2)
1195 {
1196 idx += 8;
1197 if (reg_maps->output_registers & (1u << idx))
1198 {
1199 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1200 }
1201 else
1202 {
1203 reg_maps->output_registers |= 1u << idx;
1204 shader_signature_from_usage(&output_signature_elements[idx],
1205 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1206 }
1207 }
1208 break;
1209
1210 case WINED3DSPR_TEXCRDOUT:
1211 if (shader_version.major >= 3)
1212 {
1213 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1214 {
1215 WARN("Invalid output register index %u.\n", idx);
1216 break;
1217 }
1218 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1219 break;
1220 }
1221 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1222 {
1223 WARN("Invalid texcoord index %u.\n", idx);
1224 break;
1225 }
1226 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1227 if (reg_maps->output_registers & (1u << idx))
1228 {
1229 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1230 }
1231 else
1232 {
1233 reg_maps->output_registers |= 1u << idx;
1234 shader_signature_from_usage(&output_signature_elements[idx],
1235 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1236 }
1237 break;
1238
1239 default:
1240 break;
1241 }
1242 }
1243
1244 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1245 {
1246 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1247 {
1248 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1249 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1250 * the mov and perform the sRGB write correction from the source register.
1251 *
1252 * However, if the mov is only partial, we can't do this, and if the write
1253 * comes from an instruction other than MOV it is hard to do as well. If
1254 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1255 shader->u.ps.color0_mov = FALSE;
1256 if (ins.handler_idx == WINED3DSIH_MOV
1257 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1258 {
1259 /* Used later when the source register is read. */
1260 color0_mov = TRUE;
1261 }
1262 }
1263 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1264 * end
1265 */
1266 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1267 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1268 {
1269 shader->u.ps.color0_mov = FALSE;
1270 }
1271 }
1272
1273 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1274 if (shader_version.major == 1
1275 && (ins.handler_idx == WINED3DSIH_TEX
1276 || ins.handler_idx == WINED3DSIH_TEXBEM
1277 || ins.handler_idx == WINED3DSIH_TEXBEML
1278 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1279 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1280 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1281 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1282 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1283 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1284 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1285 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1286 {
1287 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1288
1289 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1290 {
1291 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1292 continue;
1293 }
1294
1295 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1296 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1297 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1298 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1299
1300 /* texbem is only valid with < 1.4 pixel shaders */
1301 if (ins.handler_idx == WINED3DSIH_TEXBEM
1302 || ins.handler_idx == WINED3DSIH_TEXBEML)
1303 {
1304 reg_maps->bumpmat |= 1u << reg_idx;
1305 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1306 {
1307 reg_maps->luminanceparams |= 1u << reg_idx;
1308 }
1309 }
1310 }
1311 else if (ins.handler_idx == WINED3DSIH_BEM)
1312 {
1313 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1314 }
1315 }
1316
1317 if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1318 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx
1319 && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR
1320 && ins.handler_idx != WINED3DSIH_IMM_ATOMIC_CONSUME)
1321 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED)
1322 {
1323 unsigned int reg_idx;
1324 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED)
1325 reg_idx = ins.src[1].reg.idx[0].offset;
1326 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1327 reg_idx = ins.dst[0].reg.idx[0].offset;
1328 else
1329 reg_idx = ins.dst[1].reg.idx[0].offset;
1330 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1331 {
1332 ERR("Invalid UAV index %u.\n", reg_idx);
1333 break;
1334 }
1335 reg_maps->uav_read_mask |= (1u << reg_idx);
1336 }
1337 else if (ins.handler_idx == WINED3DSIH_NRM)
1338 {
1339 reg_maps->usesnrm = 1;
1340 }
1341 else if (ins.handler_idx == WINED3DSIH_DSY
1342 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1343 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1344 {
1345 reg_maps->usesdsy = 1;
1346 }
1347 else if (ins.handler_idx == WINED3DSIH_DSX
1348 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1349 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1350 {
1351 reg_maps->usesdsx = 1;
1352 }
1353 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1354 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1355 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1356 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1357 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1358 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1359 else if (ins.handler_idx == WINED3DSIH_LOOP
1360 || ins.handler_idx == WINED3DSIH_REP)
1361 {
1362 ++cur_loop_depth;
1363 if (cur_loop_depth > max_loop_depth)
1364 max_loop_depth = cur_loop_depth;
1365 }
1366 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1367 || ins.handler_idx == WINED3DSIH_ENDREP)
1368 {
1369 --cur_loop_depth;
1370 }
1371 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1372 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1373 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1374 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1375 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1376 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1377 {
1378 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1379 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1380 }
1381 else if (ins.handler_idx == WINED3DSIH_LD
1382 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1383 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1384 {
1385 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1386 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1387 }
1388
1389 if (ins.predicate)
1390 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1391 shader_version.type, constf_size))
1392 return WINED3DERR_INVALIDCALL;
1393
1394 for (i = 0; i < ins.src_count; ++i)
1395 {
1396 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1397 struct wined3d_shader_register reg = ins.src[i].reg;
1398
1399 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1400 shader_version.type, constf_size))
1401 return WINED3DERR_INVALIDCALL;
1402 while (count)
1403 {
1404 ++reg.idx[0].offset;
1405 if (!shader_record_register_usage(shader, reg_maps, &reg,
1406 shader_version.type, constf_size))
1407 return WINED3DERR_INVALIDCALL;
1408 --count;
1409 }
1410
1411 if (color0_mov)
1412 {
1413 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1414 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1415 {
1416 shader->u.ps.color0_mov = TRUE;
1417 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1418 }
1419 }
1420 }
1421 }
1422 }
1423 reg_maps->loop_depth = max_loop_depth;
1424
1425 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1426 * R0 is written to the render target. */
1427 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1428 reg_maps->rt_mask |= (1u << 0);
1429
1430 if (input_signature->elements)
1431 {
1432 for (i = 0; i < input_signature->element_count; ++i)
1433 {
1434 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1435 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1436 {
1437 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1438 reg_maps->vpos = 1;
1439 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1440 reg_maps->usesfacing = 1;
1441 }
1442 }
1443 }
1444 else if (!input_signature->elements && reg_maps->input_registers)
1445 {
1446 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1447 struct wined3d_shader_signature_element *e;
1448 unsigned int i;
1449
1450 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1451 return E_OUTOFMEMORY;
1452 input_signature->element_count = count;
1453
1454 e = input_signature->elements;
1455 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1456 {
1457 if (!(reg_maps->input_registers & (1u << i)))
1458 continue;
1459 input_signature_elements[i].register_idx = i;
1460 *e++ = input_signature_elements[i];
1461 }
1462 }
1463
1464 if (output_signature->elements)
1465 {
1466 for (i = 0; i < output_signature->element_count; ++i)
1467 {
1468 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1469 }
1470 }
1471 else if (reg_maps->output_registers)
1472 {
1473 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1474 struct wined3d_shader_signature_element *e;
1475
1476 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1477 return E_OUTOFMEMORY;
1478 output_signature->element_count = count;
1479
1480 e = output_signature->elements;
1481 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1482 {
1483 if (!(reg_maps->output_registers & (1u << i)))
1484 continue;
1485 *e++ = output_signature_elements[i];
1486 }
1487 }
1488
1489 return WINED3D_OK;
1490 }
1491
1492 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1493 {
1494 struct wined3d_shader_indexable_temp *reg, *reg_next;
1495
1496 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1497 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1498
1499 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1500 HeapFree(GetProcessHeap(), 0, reg);
1501 list_init(&reg_maps->indexable_temps);
1502 }
1503
1504 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1505 {
1506 DWORD map = 1u << max;
1507 map |= map - 1;
1508 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1509
1510 return wined3d_log2i(map);
1511 }
1512
1513 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1514 {
1515 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1516 {
1517 shader_addline(buffer, "refactoringAllowed");
1518 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1519 if (global_flags)
1520 shader_addline(buffer, " | ");
1521 }
1522
1523 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1524 {
1525 shader_addline(buffer, "enableRawAndStructuredBuffers");
1526 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1527 }
1528
1529 if (global_flags)
1530 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1531 }
1532
1533 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1534 {
1535 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1536 {
1537 shader_addline(buffer, "_g");
1538 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1539 }
1540 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1541 {
1542 shader_addline(buffer, "_t");
1543 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1544 }
1545
1546 if (sync_flags)
1547 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1548 }
1549
1550 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1551 {
1552 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1553 {
1554 shader_addline(buffer, "_glc");
1555 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1556 }
1557 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1558 {
1559 shader_addline(buffer, "_opc");
1560 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1561 }
1562
1563 if (uav_flags)
1564 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1565 }
1566
1567 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1568 enum wined3d_tessellator_domain domain)
1569 {
1570 switch (domain)
1571 {
1572 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1573 shader_addline(buffer, "line");
1574 break;
1575 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1576 shader_addline(buffer, "triangle");
1577 break;
1578 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1579 shader_addline(buffer, "quad");
1580 break;
1581 default:
1582 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1583 break;
1584 }
1585 }
1586
1587 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1588 enum wined3d_tessellator_output_primitive output_primitive)
1589 {
1590 switch (output_primitive)
1591 {
1592 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1593 shader_addline(buffer, "point");
1594 break;
1595 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1596 shader_addline(buffer, "line");
1597 break;
1598 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1599 shader_addline(buffer, "triangle_cw");
1600 break;
1601 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1602 shader_addline(buffer, "triangle_ccw");
1603 break;
1604 default:
1605 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1606 break;
1607 }
1608 }
1609
1610 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1611 enum wined3d_tessellator_partitioning partitioning)
1612 {
1613 switch (partitioning)
1614 {
1615 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1616 shader_addline(buffer, "integer");
1617 break;
1618 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1619 shader_addline(buffer, "pow2");
1620 break;
1621 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1622 shader_addline(buffer, "fractional_odd");
1623 break;
1624 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1625 shader_addline(buffer, "fractional_even");
1626 break;
1627 default:
1628 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1629 break;
1630 }
1631 }
1632
1633 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1634 {
1635 unsigned int i;
1636
1637 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1638 {
1639 if (sysval_semantic_names[i].sysval_semantic == semantic)
1640 {
1641 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1642 return;
1643 }
1644 }
1645
1646 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1647 }
1648
1649 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1650 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1651 {
1652 shader_addline(buffer, "dcl");
1653
1654 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1655 {
1656 switch (semantic->resource_type)
1657 {
1658 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1659 shader_addline(buffer, "_2d");
1660 break;
1661
1662 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1663 shader_addline(buffer, "_3d");
1664 break;
1665
1666 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1667 shader_addline(buffer, "_cube");
1668 break;
1669
1670 default:
1671 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1672 break;
1673 }
1674 }
1675 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1676 {
1677 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1678 shader_addline(buffer, "_resource_");
1679 else
1680 shader_addline(buffer, "_uav_");
1681 switch (semantic->resource_type)
1682 {
1683 case WINED3D_SHADER_RESOURCE_BUFFER:
1684 shader_addline(buffer, "buffer");
1685 break;
1686
1687 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1688 shader_addline(buffer, "texture1d");
1689 break;
1690
1691 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1692 shader_addline(buffer, "texture2d");
1693 break;
1694
1695 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1696 shader_addline(buffer, "texture2dms");
1697 break;
1698
1699 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1700 shader_addline(buffer, "texture3d");
1701 break;
1702
1703 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1704 shader_addline(buffer, "texturecube");
1705 break;
1706
1707 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1708 shader_addline(buffer, "texture1darray");
1709 break;
1710
1711 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1712 shader_addline(buffer, "texture2darray");
1713 break;
1714
1715 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1716 shader_addline(buffer, "texture2dmsarray");
1717 break;
1718
1719 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
1720 shader_addline(buffer, "texturecubearray");
1721 break;
1722
1723 default:
1724 shader_addline(buffer, "unknown");
1725 break;
1726 }
1727 switch (semantic->resource_data_type)
1728 {
1729 case WINED3D_DATA_FLOAT:
1730 shader_addline(buffer, " (float)");
1731 break;
1732
1733 case WINED3D_DATA_INT:
1734 shader_addline(buffer, " (int)");
1735 break;
1736
1737 case WINED3D_DATA_UINT:
1738 shader_addline(buffer, " (uint)");
1739 break;
1740
1741 case WINED3D_DATA_UNORM:
1742 shader_addline(buffer, " (unorm)");
1743 break;
1744
1745 case WINED3D_DATA_SNORM:
1746 shader_addline(buffer, " (snorm)");
1747 break;
1748
1749 default:
1750 shader_addline(buffer, " (unknown)");
1751 break;
1752 }
1753 }
1754 else
1755 {
1756 /* Pixel shaders 3.0 don't have usage semantics. */
1757 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1758 return;
1759 else
1760 shader_addline(buffer, "_");
1761
1762 switch (semantic->usage)
1763 {
1764 case WINED3D_DECL_USAGE_POSITION:
1765 shader_addline(buffer, "position%u", semantic->usage_idx);
1766 break;
1767
1768 case WINED3D_DECL_USAGE_BLEND_INDICES:
1769 shader_addline(buffer, "blend");
1770 break;
1771
1772 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1773 shader_addline(buffer, "weight");
1774 break;
1775
1776 case WINED3D_DECL_USAGE_NORMAL:
1777 shader_addline(buffer, "normal%u", semantic->usage_idx);
1778 break;
1779
1780 case WINED3D_DECL_USAGE_PSIZE:
1781 shader_addline(buffer, "psize");
1782 break;
1783
1784 case WINED3D_DECL_USAGE_COLOR:
1785 if (!semantic->usage_idx)
1786 shader_addline(buffer, "color");
1787 else
1788 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1789 break;
1790
1791 case WINED3D_DECL_USAGE_TEXCOORD:
1792 shader_addline(buffer, "texture%u", semantic->usage_idx);
1793 break;
1794
1795 case WINED3D_DECL_USAGE_TANGENT:
1796 shader_addline(buffer, "tangent");
1797 break;
1798
1799 case WINED3D_DECL_USAGE_BINORMAL:
1800 shader_addline(buffer, "binormal");
1801 break;
1802
1803 case WINED3D_DECL_USAGE_TESS_FACTOR:
1804 shader_addline(buffer, "tessfactor");
1805 break;
1806
1807 case WINED3D_DECL_USAGE_POSITIONT:
1808 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1809 break;
1810
1811 case WINED3D_DECL_USAGE_FOG:
1812 shader_addline(buffer, "fog");
1813 break;
1814
1815 case WINED3D_DECL_USAGE_DEPTH:
1816 shader_addline(buffer, "depth");
1817 break;
1818
1819 case WINED3D_DECL_USAGE_SAMPLE:
1820 shader_addline(buffer, "sample");
1821 break;
1822
1823 default:
1824 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1825 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
1826 }
1827 }
1828 }
1829
1830 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1831 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1832 {
1833 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1834 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1835 UINT offset = reg->idx[0].offset;
1836
1837 switch (reg->type)
1838 {
1839 case WINED3DSPR_TEMP:
1840 shader_addline(buffer, "r");
1841 break;
1842
1843 case WINED3DSPR_INPUT:
1844 shader_addline(buffer, "v");
1845 break;
1846
1847 case WINED3DSPR_CONST:
1848 case WINED3DSPR_CONST2:
1849 case WINED3DSPR_CONST3:
1850 case WINED3DSPR_CONST4:
1851 shader_addline(buffer, "c");
1852 offset = shader_get_float_offset(reg->type, offset);
1853 break;
1854
1855 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1856 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1857 break;
1858
1859 case WINED3DSPR_RASTOUT:
1860 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1861 break;
1862
1863 case WINED3DSPR_COLOROUT:
1864 shader_addline(buffer, "oC");
1865 break;
1866
1867 case WINED3DSPR_DEPTHOUT:
1868 shader_addline(buffer, "oDepth");
1869 break;
1870
1871 case WINED3DSPR_ATTROUT:
1872 shader_addline(buffer, "oD");
1873 break;
1874
1875 case WINED3DSPR_TEXCRDOUT:
1876 /* Vertex shaders >= 3.0 use general purpose output registers
1877 * (WINED3DSPR_OUTPUT), which can include an address token. */
1878 if (shader_version->major >= 3)
1879 shader_addline(buffer, "o");
1880 else
1881 shader_addline(buffer, "oT");
1882 break;
1883
1884 case WINED3DSPR_CONSTINT:
1885 shader_addline(buffer, "i");
1886 break;
1887
1888 case WINED3DSPR_CONSTBOOL:
1889 shader_addline(buffer, "b");
1890 break;
1891
1892 case WINED3DSPR_LABEL:
1893 shader_addline(buffer, "l");
1894 break;
1895
1896 case WINED3DSPR_LOOP:
1897 shader_addline(buffer, "aL");
1898 break;
1899
1900 case WINED3DSPR_SAMPLER:
1901 shader_addline(buffer, "s");
1902 break;
1903
1904 case WINED3DSPR_MISCTYPE:
1905 if (offset > 1)
1906 {
1907 FIXME("Unhandled misctype register %u.\n", offset);
1908 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1909 }
1910 else
1911 {
1912 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1913 }
1914 break;
1915
1916 case WINED3DSPR_PREDICATE:
1917 shader_addline(buffer, "p");
1918 break;
1919
1920 case WINED3DSPR_IMMCONST:
1921 shader_addline(buffer, "l");
1922 break;
1923
1924 case WINED3DSPR_CONSTBUFFER:
1925 shader_addline(buffer, "cb");
1926 break;
1927
1928 case WINED3DSPR_IMMCONSTBUFFER:
1929 shader_addline(buffer, "icb");
1930 break;
1931
1932 case WINED3DSPR_PRIMID:
1933 shader_addline(buffer, "primID");
1934 break;
1935
1936 case WINED3DSPR_NULL:
1937 shader_addline(buffer, "null");
1938 break;
1939
1940 case WINED3DSPR_RESOURCE:
1941 shader_addline(buffer, "t");
1942 break;
1943
1944 case WINED3DSPR_UAV:
1945 shader_addline(buffer, "u");
1946 break;
1947
1948 case WINED3DSPR_OUTPOINTID:
1949 shader_addline(buffer, "vOutputControlPointID");
1950 break;
1951
1952 case WINED3DSPR_FORKINSTID:
1953 shader_addline(buffer, "vForkInstanceId");
1954 break;
1955
1956 case WINED3DSPR_INCONTROLPOINT:
1957 shader_addline(buffer, "vicp");
1958 break;
1959
1960 case WINED3DSPR_OUTCONTROLPOINT:
1961 shader_addline(buffer, "vocp");
1962 break;
1963
1964 case WINED3DSPR_PATCHCONST:
1965 shader_addline(buffer, "vpc");
1966 break;
1967
1968 case WINED3DSPR_TESSCOORD:
1969 shader_addline(buffer, "vDomainLocation");
1970 break;
1971
1972 case WINED3DSPR_GROUPSHAREDMEM:
1973 shader_addline(buffer, "g");
1974 break;
1975
1976 case WINED3DSPR_THREADID:
1977 shader_addline(buffer, "vThreadID");
1978 break;
1979
1980 case WINED3DSPR_THREADGROUPID:
1981 shader_addline(buffer, "vThreadGroupID");
1982 break;
1983
1984 case WINED3DSPR_LOCALTHREADID:
1985 shader_addline(buffer, "vThreadIDInGroup");
1986 break;
1987
1988 case WINED3DSPR_LOCALTHREADINDEX:
1989 shader_addline(buffer, "vThreadIDInGroupFlattened");
1990 break;
1991
1992 case WINED3DSPR_IDXTEMP:
1993 shader_addline(buffer, "x");
1994 break;
1995
1996 case WINED3DSPR_STREAM:
1997 shader_addline(buffer, "m");
1998 break;
1999
2000 case WINED3DSPR_FUNCTIONBODY:
2001 shader_addline(buffer, "fb");
2002 break;
2003
2004 case WINED3DSPR_FUNCTIONPOINTER:
2005 shader_addline(buffer, "fp");
2006 break;
2007
2008 case WINED3DSPR_COVERAGE:
2009 shader_addline(buffer, "vCoverage");
2010 break;
2011
2012 case WINED3DSPR_SAMPLEMASK:
2013 shader_addline(buffer, "oMask");
2014 break;
2015
2016 default:
2017 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2018 break;
2019 }
2020
2021 if (reg->type == WINED3DSPR_IMMCONST)
2022 {
2023 shader_addline(buffer, "(");
2024 switch (reg->immconst_type)
2025 {
2026 case WINED3D_IMMCONST_SCALAR:
2027 switch (reg->data_type)
2028 {
2029 case WINED3D_DATA_FLOAT:
2030 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2031 break;
2032 case WINED3D_DATA_INT:
2033 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2034 break;
2035 case WINED3D_DATA_RESOURCE:
2036 case WINED3D_DATA_SAMPLER:
2037 case WINED3D_DATA_UINT:
2038 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2039 break;
2040 default:
2041 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2042 break;
2043 }
2044 break;
2045
2046 case WINED3D_IMMCONST_VEC4:
2047 switch (reg->data_type)
2048 {
2049 case WINED3D_DATA_FLOAT:
2050 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2051 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2052 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2053 break;
2054 case WINED3D_DATA_INT:
2055 shader_addline(buffer, "%d, %d, %d, %d",
2056 reg->u.immconst_data[0], reg->u.immconst_data[1],
2057 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2058 break;
2059 case WINED3D_DATA_RESOURCE:
2060 case WINED3D_DATA_SAMPLER:
2061 case WINED3D_DATA_UINT:
2062 shader_addline(buffer, "%u, %u, %u, %u",
2063 reg->u.immconst_data[0], reg->u.immconst_data[1],
2064 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2065 break;
2066 default:
2067 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2068 break;
2069 }
2070 break;
2071
2072 default:
2073 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2074 break;
2075 }
2076 shader_addline(buffer, ")");
2077 }
2078 else if (reg->type != WINED3DSPR_RASTOUT
2079 && reg->type != WINED3DSPR_MISCTYPE
2080 && reg->type != WINED3DSPR_NULL)
2081 {
2082 if (offset != ~0u)
2083 {
2084 shader_addline(buffer, "[");
2085 if (reg->idx[0].rel_addr)
2086 {
2087 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2088 shader_addline(buffer, " + ");
2089 }
2090 shader_addline(buffer, "%u]", offset);
2091
2092 if (reg->idx[1].offset != ~0u)
2093 {
2094 shader_addline(buffer, "[");
2095 if (reg->idx[1].rel_addr)
2096 {
2097 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2098 shader_addline(buffer, " + ");
2099 }
2100 shader_addline(buffer, "%u]", reg->idx[1].offset);
2101 }
2102 }
2103
2104 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2105 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2106 }
2107 }
2108
2109 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2110 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2111 {
2112 DWORD write_mask = param->write_mask;
2113
2114 shader_dump_register(buffer, &param->reg, shader_version);
2115
2116 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2117 {
2118 static const char write_mask_chars[] = "xyzw";
2119
2120 shader_addline(buffer, ".");
2121 if (write_mask & WINED3DSP_WRITEMASK_0)
2122 shader_addline(buffer, "%c", write_mask_chars[0]);
2123 if (write_mask & WINED3DSP_WRITEMASK_1)
2124 shader_addline(buffer, "%c", write_mask_chars[1]);
2125 if (write_mask & WINED3DSP_WRITEMASK_2)
2126 shader_addline(buffer, "%c", write_mask_chars[2]);
2127 if (write_mask & WINED3DSP_WRITEMASK_3)
2128 shader_addline(buffer, "%c", write_mask_chars[3]);
2129 }
2130 }
2131
2132 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2133 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2134 {
2135 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2136 DWORD swizzle = param->swizzle;
2137
2138 if (src_modifier == WINED3DSPSM_NEG
2139 || src_modifier == WINED3DSPSM_BIASNEG
2140 || src_modifier == WINED3DSPSM_SIGNNEG
2141 || src_modifier == WINED3DSPSM_X2NEG
2142 || src_modifier == WINED3DSPSM_ABSNEG)
2143 shader_addline(buffer, "-");
2144 else if (src_modifier == WINED3DSPSM_COMP)
2145 shader_addline(buffer, "1-");
2146 else if (src_modifier == WINED3DSPSM_NOT)
2147 shader_addline(buffer, "!");
2148
2149 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2150 shader_addline(buffer, "abs(");
2151
2152 shader_dump_register(buffer, &param->reg, shader_version);
2153
2154 switch (src_modifier)
2155 {
2156 case WINED3DSPSM_NONE: break;
2157 case WINED3DSPSM_NEG: break;
2158 case WINED3DSPSM_NOT: break;
2159 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2160 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2161 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2162 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2163 case WINED3DSPSM_COMP: break;
2164 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2165 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2166 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2167 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2168 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2169 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2170 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2171 }
2172
2173 if (swizzle != WINED3DSP_NOSWIZZLE)
2174 {
2175 static const char swizzle_chars[] = "xyzw";
2176 DWORD swizzle_x = swizzle & 0x03;
2177 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2178 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2179 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2180
2181 if (swizzle_x == swizzle_y
2182 && swizzle_x == swizzle_z
2183 && swizzle_x == swizzle_w)
2184 {
2185 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2186 }
2187 else
2188 {
2189 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2190 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2191 }
2192 }
2193 }
2194
2195 /* Shared code in order to generate the bulk of the shader string.
2196 * NOTE: A description of how to parse tokens can be found on MSDN. */
2197 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2198 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx)
2199 {
2200 struct wined3d_device *device = shader->device;
2201 const struct wined3d_shader_frontend *fe = shader->frontend;
2202 void *fe_data = shader->frontend_data;
2203 struct wined3d_shader_version shader_version;
2204 struct wined3d_shader_parser_state state;
2205 struct wined3d_shader_instruction ins;
2206 struct wined3d_shader_tex_mx tex_mx;
2207 struct wined3d_shader_context ctx;
2208 const DWORD *ptr;
2209
2210 /* Initialize current parsing state. */
2211 tex_mx.current_row = 0;
2212 state.current_loop_depth = 0;
2213 state.current_loop_reg = 0;
2214 state.in_subroutine = FALSE;
2215
2216 ctx.shader = shader;
2217 ctx.gl_info = &device->adapter->gl_info;
2218 ctx.reg_maps = reg_maps;
2219 ctx.buffer = buffer;
2220 ctx.tex_mx = &tex_mx;
2221 ctx.state = &state;
2222 ctx.backend_data = backend_ctx;
2223 ins.ctx = &ctx;
2224
2225 fe->shader_read_header(fe_data, &ptr, &shader_version);
2226
2227 while (!fe->shader_is_end(fe_data, &ptr))
2228 {
2229 /* Read opcode. */
2230 fe->shader_read_instruction(fe_data, &ptr, &ins);
2231
2232 /* Unknown opcode and its parameters. */
2233 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2234 {
2235 TRACE("Skipping unrecognized instruction.\n");
2236 continue;
2237 }
2238
2239 if (ins.predicate)
2240 FIXME("Predicates not implemented.\n");
2241
2242 /* Call appropriate function for output target */
2243 device->shader_backend->shader_handle_instruction(&ins);
2244 }
2245 }
2246
2247 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2248 const struct wined3d_shader_dst_param *dst)
2249 {
2250 DWORD mmask = dst->modifiers;
2251
2252 switch (dst->shift)
2253 {
2254 case 0: break;
2255 case 13: shader_addline(buffer, "_d8"); break;
2256 case 14: shader_addline(buffer, "_d4"); break;
2257 case 15: shader_addline(buffer, "_d2"); break;
2258 case 1: shader_addline(buffer, "_x2"); break;
2259 case 2: shader_addline(buffer, "_x4"); break;
2260 case 3: shader_addline(buffer, "_x8"); break;
2261 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2262 }
2263
2264 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2265 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2266 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2267
2268 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2269 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2270 }
2271
2272 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2273 enum wined3d_primitive_type primitive_type)
2274 {
2275 switch (primitive_type)
2276 {
2277 case WINED3D_PT_UNDEFINED:
2278 shader_addline(buffer, "undefined");
2279 break;
2280 case WINED3D_PT_POINTLIST:
2281 shader_addline(buffer, "pointlist");
2282 break;
2283 case WINED3D_PT_LINELIST:
2284 shader_addline(buffer, "linelist");
2285 break;
2286 case WINED3D_PT_LINESTRIP:
2287 shader_addline(buffer, "linestrip");
2288 break;
2289 case WINED3D_PT_TRIANGLELIST:
2290 shader_addline(buffer, "trianglelist");
2291 break;
2292 case WINED3D_PT_TRIANGLESTRIP:
2293 shader_addline(buffer, "trianglestrip");
2294 break;
2295 case WINED3D_PT_TRIANGLEFAN:
2296 shader_addline(buffer, "trianglefan");
2297 break;
2298 case WINED3D_PT_LINELIST_ADJ:
2299 shader_addline(buffer, "linelist_adj");
2300 break;
2301 case WINED3D_PT_LINESTRIP_ADJ:
2302 shader_addline(buffer, "linestrip_adj");
2303 break;
2304 case WINED3D_PT_TRIANGLELIST_ADJ:
2305 shader_addline(buffer, "trianglelist_adj");
2306 break;
2307 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2308 shader_addline(buffer, "trianglestrip_adj");
2309 break;
2310 default:
2311 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2312 break;
2313 }
2314 }
2315
2316 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2317 enum wined3d_shader_interpolation_mode interpolation_mode)
2318 {
2319 switch (interpolation_mode)
2320 {
2321 case WINED3DSIM_CONSTANT:
2322 shader_addline(buffer, "constant");
2323 break;
2324 case WINED3DSIM_LINEAR:
2325 shader_addline(buffer, "linear");
2326 break;
2327 case WINED3DSIM_LINEAR_CENTROID:
2328 shader_addline(buffer, "linear centroid");
2329 break;
2330 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2331 shader_addline(buffer, "linear noperspective");
2332 break;
2333 case WINED3DSIM_LINEAR_SAMPLE:
2334 shader_addline(buffer, "linear sample");
2335 break;
2336 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2337 shader_addline(buffer, "linear noperspective centroid");
2338 break;
2339 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2340 shader_addline(buffer, "linear noperspective sample");
2341 break;
2342 default:
2343 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2344 break;
2345 }
2346 }
2347
2348 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2349 {
2350 struct wined3d_shader_version shader_version;
2351 struct wined3d_string_buffer buffer;
2352 const char *type_prefix;
2353 const char *p, *q;
2354 const DWORD *ptr;
2355 DWORD i;
2356
2357 if (!string_buffer_init(&buffer))
2358 {
2359 ERR("Failed to initialize string buffer.\n");
2360 return;
2361 }
2362
2363 fe->shader_read_header(fe_data, &ptr, &shader_version);
2364
2365 TRACE("Parsing %p.\n", ptr);
2366
2367 switch (shader_version.type)
2368 {
2369 case WINED3D_SHADER_TYPE_VERTEX:
2370 type_prefix = "vs";
2371 break;
2372
2373 case WINED3D_SHADER_TYPE_HULL:
2374 type_prefix = "hs";
2375 break;
2376
2377 case WINED3D_SHADER_TYPE_DOMAIN:
2378 type_prefix = "ds";
2379 break;
2380
2381 case WINED3D_SHADER_TYPE_GEOMETRY:
2382 type_prefix = "gs";
2383 break;
2384
2385 case WINED3D_SHADER_TYPE_PIXEL:
2386 type_prefix = "ps";
2387 break;
2388
2389 case WINED3D_SHADER_TYPE_COMPUTE:
2390 type_prefix = "cs";
2391 break;
2392
2393 default:
2394 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2395 type_prefix = "unknown";
2396 break;
2397 }
2398
2399 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2400
2401 while (!fe->shader_is_end(fe_data, &ptr))
2402 {
2403 struct wined3d_shader_instruction ins;
2404
2405 fe->shader_read_instruction(fe_data, &ptr, &ins);
2406 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2407 {
2408 WARN("Skipping unrecognized instruction.\n");
2409 shader_addline(&buffer, "<unrecognized instruction>\n");
2410 continue;
2411 }
2412
2413 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2414 {
2415 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2416 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2417 shader_addline(&buffer, " ");
2418 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2419 }
2420 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2421 {
2422 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2423 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2424 shader_addline(&buffer, ", %s",
2425 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2426 }
2427 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2428 {
2429 shader_addline(&buffer, "%s fb%u",
2430 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2431 }
2432 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2433 {
2434 shader_addline(&buffer, "%s ft%u = {...}",
2435 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2436 }
2437 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2438 {
2439 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2440 shader_dump_global_flags(&buffer, ins.flags);
2441 }
2442 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2443 {
2444 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2445 ins.declaration.max_tessellation_factor);
2446 }
2447 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2448 {
2449 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2450 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2451 {
2452 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2453 ins.declaration.icb->data[4 * i + 0],
2454 ins.declaration.icb->data[4 * i + 1],
2455 ins.declaration.icb->data[4 * i + 2],
2456 ins.declaration.icb->data[4 * i + 3]);
2457 }
2458 shader_addline(&buffer, "}");
2459 }
2460 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2461 {
2462 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2463 ins.declaration.indexable_temp.register_idx,
2464 ins.declaration.indexable_temp.register_size,
2465 ins.declaration.indexable_temp.component_count);
2466 }
2467 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2468 {
2469 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2470 shader_dump_interpolation_mode(&buffer, ins.flags);
2471 shader_addline(&buffer, " ");
2472 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2473 }
2474 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2475 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2476 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2477 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2478 {
2479 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2480 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2481 shader_addline(&buffer, ", ");
2482 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2483 }
2484 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2485 {
2486 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2487 shader_dump_interpolation_mode(&buffer, ins.flags);
2488 shader_addline(&buffer, " ");
2489 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2490 shader_addline(&buffer, ", ");
2491 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2492 }
2493 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2494 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2495 {
2496 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2497 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2498 }
2499 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2500 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2501 {
2502 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2503 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2504 }
2505 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2506 {
2507 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2508 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2509 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2510 }
2511 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2512 {
2513 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2514 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2515 }
2516 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2517 {
2518 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2519 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2520 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2521 }
2522 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2523 {
2524 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2525 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2526 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2527 shader_addline(&buffer, ", comparisonMode");
2528 }
2529 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2530 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2531 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2532 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2533 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2534 {
2535 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2536 }
2537 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2538 {
2539 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2540 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2541 }
2542 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2543 {
2544 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2545 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2546 }
2547 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2548 {
2549 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2550 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2551 }
2552 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2553 {
2554 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2555 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2556 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2557 }
2558 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2559 {
2560 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2561 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2562 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2563 ins.declaration.tgsm_structured.structure_count);
2564 }
2565 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2566 {
2567 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2568 ins.declaration.thread_group_size.x,
2569 ins.declaration.thread_group_size.y,
2570 ins.declaration.thread_group_size.z);
2571 }
2572 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2573 {
2574 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2575 shader_dump_uav_flags(&buffer, ins.flags);
2576 shader_addline(&buffer, " ");
2577 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2578 }
2579 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2580 {
2581 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2582 shader_dump_uav_flags(&buffer, ins.flags);
2583 shader_addline(&buffer, " ");
2584 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2585 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2586 }
2587 else if (ins.handler_idx == WINED3DSIH_DEF)
2588 {
2589 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2590 ins.dst[0].reg.idx[0].offset),
2591 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2592 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2593 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2594 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2595 }
2596 else if (ins.handler_idx == WINED3DSIH_DEFI)
2597 {
2598 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2599 ins.src[0].reg.u.immconst_data[0],
2600 ins.src[0].reg.u.immconst_data[1],
2601 ins.src[0].reg.u.immconst_data[2],
2602 ins.src[0].reg.u.immconst_data[3]);
2603 }
2604 else if (ins.handler_idx == WINED3DSIH_DEFB)
2605 {
2606 shader_addline(&buffer, "defb b%u = %s",
2607 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2608 }
2609 else
2610 {
2611 if (ins.predicate)
2612 {
2613 shader_addline(&buffer, "(");
2614 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2615 shader_addline(&buffer, ") ");
2616 }
2617
2618 /* PixWin marks instructions with the coissue flag with a '+' */
2619 if (ins.coissue)
2620 shader_addline(&buffer, "+");
2621
2622 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2623
2624 if (ins.handler_idx == WINED3DSIH_BREAKP
2625 || ins.handler_idx == WINED3DSIH_IF)
2626 {
2627 switch (ins.flags)
2628 {
2629 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2630 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2631 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2632 }
2633 }
2634 else if (ins.handler_idx == WINED3DSIH_IFC
2635 || ins.handler_idx == WINED3DSIH_BREAKC)
2636 {
2637 switch (ins.flags)
2638 {
2639 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2640 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2641 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2642 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2643 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2644 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2645 default: shader_addline(&buffer, "_(%u)", ins.flags);
2646 }
2647 }
2648 else if (ins.handler_idx == WINED3DSIH_TEX
2649 && shader_version.major >= 2
2650 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2651 {
2652 shader_addline(&buffer, "p");
2653 }
2654 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2655 {
2656 switch (ins.flags)
2657 {
2658 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2659 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2660 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2661 }
2662 }
2663 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2664 {
2665 switch (ins.flags)
2666 {
2667 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2668 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2669 }
2670 }
2671 else if (ins.handler_idx == WINED3DSIH_SYNC)
2672 {
2673 shader_dump_sync_flags(&buffer, ins.flags);
2674 }
2675
2676 if (wined3d_shader_instruction_has_texel_offset(&ins))
2677 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2678
2679 for (i = 0; i < ins.dst_count; ++i)
2680 {
2681 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2682 shader_addline(&buffer, !i ? " " : ", ");
2683 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2684 }
2685
2686 /* Other source tokens */
2687 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2688 {
2689 shader_addline(&buffer, !i ? " " : ", ");
2690 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2691 }
2692 }
2693 shader_addline(&buffer, "\n");
2694 }
2695
2696 for (p = buffer.buffer; *p; p = q)
2697 {
2698 if (!(q = strstr(p, "\n")))
2699 q = p + strlen(p);
2700 else
2701 ++q;
2702 TRACE(" %.*s", (int)(q - p), p);
2703 }
2704
2705 string_buffer_free(&buffer);
2706 }
2707
2708 static void shader_cleanup(struct wined3d_shader *shader)
2709 {
2710 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2711 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2712 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2713 shader->device->shader_backend->shader_destroy(shader);
2714 shader_cleanup_reg_maps(&shader->reg_maps);
2715 HeapFree(GetProcessHeap(), 0, shader->function);
2716 shader_delete_constant_list(&shader->constantsF);
2717 shader_delete_constant_list(&shader->constantsB);
2718 shader_delete_constant_list(&shader->constantsI);
2719 list_remove(&shader->shader_list_entry);
2720
2721 if (shader->frontend && shader->frontend_data)
2722 shader->frontend->shader_free(shader->frontend_data);
2723 }
2724
2725 struct shader_none_priv
2726 {
2727 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2728 const struct fragment_pipeline *fragment_pipe;
2729 BOOL ffp_proj_control;
2730 };
2731
2732 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2733 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
2734 const struct wined3d_state *state) {}
2735 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2736 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2737 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2738 const struct wined3d_state *state) {}
2739 static void shader_none_destroy(struct wined3d_shader *shader) {}
2740 static void shader_none_free_context_data(struct wined3d_context *context) {}
2741 static void shader_none_init_context_state(struct wined3d_context *context) {}
2742
2743 /* Context activation is done by the caller. */
2744 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2745 const struct wined3d_state *state)
2746 {
2747 const struct wined3d_gl_info *gl_info = context->gl_info;
2748 struct shader_none_priv *priv = shader_priv;
2749
2750 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2751 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2752 }
2753
2754 /* Context activation is done by the caller. */
2755 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2756 {
2757 struct shader_none_priv *priv = shader_priv;
2758 const struct wined3d_gl_info *gl_info = context->gl_info;
2759
2760 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2761 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2762
2763 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2764 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2765 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2766 | (1u << WINED3D_SHADER_TYPE_HULL)
2767 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2768 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2769 }
2770
2771 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2772 const struct fragment_pipeline *fragment_pipe)
2773 {
2774 struct fragment_caps fragment_caps;
2775 void *vertex_priv, *fragment_priv;
2776 struct shader_none_priv *priv;
2777
2778 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2779 return E_OUTOFMEMORY;
2780
2781 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2782 {
2783 ERR("Failed to initialize vertex pipe.\n");
2784 HeapFree(GetProcessHeap(), 0, priv);
2785 return E_FAIL;
2786 }
2787
2788 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2789 {
2790 ERR("Failed to initialize fragment pipe.\n");
2791 vertex_pipe->vp_free(device);
2792 HeapFree(GetProcessHeap(), 0, priv);
2793 return E_FAIL;
2794 }
2795
2796 priv->vertex_pipe = vertex_pipe;
2797 priv->fragment_pipe = fragment_pipe;
2798 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2799 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2800
2801 device->vertex_priv = vertex_priv;
2802 device->fragment_priv = fragment_priv;
2803 device->shader_priv = priv;
2804
2805 return WINED3D_OK;
2806 }
2807
2808 static void shader_none_free(struct wined3d_device *device)
2809 {
2810 struct shader_none_priv *priv = device->shader_priv;
2811
2812 priv->fragment_pipe->free_private(device);
2813 priv->vertex_pipe->vp_free(device);
2814 HeapFree(GetProcessHeap(), 0, priv);
2815 }
2816
2817 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2818 {
2819 return TRUE;
2820 }
2821
2822 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2823 {
2824 /* Set the shader caps to 0 for the none shader backend */
2825 caps->vs_version = 0;
2826 caps->hs_version = 0;
2827 caps->ds_version = 0;
2828 caps->gs_version = 0;
2829 caps->ps_version = 0;
2830 caps->cs_version = 0;
2831 caps->vs_uniform_count = 0;
2832 caps->ps_uniform_count = 0;
2833 caps->ps_1x_max_value = 0.0f;
2834 caps->varying_count = 0;
2835 caps->wined3d_caps = 0;
2836 }
2837
2838 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2839 {
2840 /* We "support" every possible fixup, since we don't support any shader
2841 * model, and will never have to actually sample a texture. */
2842 return TRUE;
2843 }
2844
2845 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2846 {
2847 struct shader_none_priv *priv = shader_priv;
2848
2849 return priv->ffp_proj_control;
2850 }
2851
2852 const struct wined3d_shader_backend_ops none_shader_backend =
2853 {
2854 shader_none_handle_instruction,
2855 shader_none_select,
2856 shader_none_select_compute,
2857 shader_none_disable,
2858 shader_none_update_float_vertex_constants,
2859 shader_none_update_float_pixel_constants,
2860 shader_none_load_constants,
2861 shader_none_destroy,
2862 shader_none_alloc,
2863 shader_none_free,
2864 shader_none_allocate_context_data,
2865 shader_none_free_context_data,
2866 shader_none_init_context_state,
2867 shader_none_get_caps,
2868 shader_none_color_fixup_supported,
2869 shader_none_has_ffp_proj_control,
2870 };
2871
2872 static HRESULT shader_set_function(struct wined3d_shader *shader, DWORD float_const_count,
2873 enum wined3d_shader_type type, unsigned int max_version)
2874 {
2875 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2876 const struct wined3d_shader_frontend *fe;
2877 HRESULT hr;
2878 unsigned int backend_version;
2879 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2880
2881 TRACE("shader %p, float_const_count %u, type %#x, max_version %u.\n",
2882 shader, float_const_count, type, max_version);
2883
2884 fe = shader->frontend;
2885 if (!(shader->frontend_data = fe->shader_init(shader->function,
2886 shader->functionLength, &shader->output_signature)))
2887 {
2888 FIXME("Failed to initialize frontend.\n");
2889 return WINED3DERR_INVALIDCALL;
2890 }
2891
2892 /* First pass: trace shader. */
2893 if (TRACE_ON(d3d_shader))
2894 shader_trace_init(fe, shader->frontend_data);
2895
2896 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2897 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2898 &shader->output_signature, float_const_count)))
2899 return hr;
2900
2901 if (reg_maps->shader_version.type != type)
2902 {
2903 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2904 return WINED3DERR_INVALIDCALL;
2905 }
2906 if (reg_maps->shader_version.major > max_version)
2907 {
2908 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2909 return WINED3DERR_INVALIDCALL;
2910 }
2911 switch (type)
2912 {
2913 case WINED3D_SHADER_TYPE_VERTEX:
2914 backend_version = d3d_info->limits.vs_version;
2915 break;
2916 case WINED3D_SHADER_TYPE_HULL:
2917 backend_version = d3d_info->limits.hs_version;
2918 break;
2919 case WINED3D_SHADER_TYPE_DOMAIN:
2920 backend_version = d3d_info->limits.ds_version;
2921 break;
2922 case WINED3D_SHADER_TYPE_GEOMETRY:
2923 backend_version = d3d_info->limits.gs_version;
2924 break;
2925 case WINED3D_SHADER_TYPE_PIXEL:
2926 backend_version = d3d_info->limits.ps_version;
2927 break;
2928 case WINED3D_SHADER_TYPE_COMPUTE:
2929 backend_version = d3d_info->limits.cs_version;
2930 break;
2931 default:
2932 FIXME("No backend version-checking for this shader type.\n");
2933 backend_version = 0;
2934 }
2935 if (reg_maps->shader_version.major > backend_version)
2936 {
2937 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2938 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2939 return WINED3DERR_INVALIDCALL;
2940 }
2941
2942 return WINED3D_OK;
2943 }
2944
2945 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2946 {
2947 ULONG refcount = InterlockedIncrement(&shader->ref);
2948
2949 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2950
2951 return refcount;
2952 }
2953
2954 static void wined3d_shader_destroy_object(void *object)
2955 {
2956 shader_cleanup(object);
2957 HeapFree(GetProcessHeap(), 0, object);
2958 }
2959
2960 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2961 {
2962 ULONG refcount = InterlockedDecrement(&shader->ref);
2963
2964 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2965
2966 if (!refcount)
2967 {
2968 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2969 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2970 }
2971
2972 return refcount;
2973 }
2974
2975 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2976 {
2977 TRACE("shader %p.\n", shader);
2978
2979 return shader->parent;
2980 }
2981
2982 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2983 void *byte_code, UINT *byte_code_size)
2984 {
2985 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2986
2987 if (!byte_code)
2988 {
2989 *byte_code_size = shader->functionLength;
2990 return WINED3D_OK;
2991 }
2992
2993 if (*byte_code_size < shader->functionLength)
2994 {
2995 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2996 * than the required size we should write the required size and
2997 * return D3DERR_MOREDATA. That's not actually true. */
2998 return WINED3DERR_INVALIDCALL;
2999 }
3000
3001 memcpy(byte_code, shader->function, shader->functionLength);
3002
3003 return WINED3D_OK;
3004 }
3005
3006 /* Set local constants for d3d8 shaders. */
3007 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3008 UINT start_idx, const float *src_data, UINT count)
3009 {
3010 UINT end_idx = start_idx + count;
3011 UINT i;
3012
3013 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3014
3015 if (end_idx > shader->limits->constant_float)
3016 {
3017 WARN("end_idx %u > float constants limit %u.\n",
3018 end_idx, shader->limits->constant_float);
3019 end_idx = shader->limits->constant_float;
3020 }
3021
3022 for (i = start_idx; i < end_idx; ++i)
3023 {
3024 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
3025 float *value;
3026 if (!lconst)
3027 return E_OUTOFMEMORY;
3028
3029 lconst->idx = i;
3030 value = (float *)lconst->value;
3031 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3032 list_add_head(&shader->constantsF, &lconst->entry);
3033
3034 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3035 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3036 {
3037 shader->lconst_inf_or_nan = TRUE;
3038 }
3039 }
3040
3041 return WINED3D_OK;
3042 }
3043
3044 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3045 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
3046 {
3047 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3048 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3049 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3050 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3051 args->point_size = state->gl_primitive_type == GL_POINTS;
3052 args->per_vertex_point_size = shader->reg_maps.point_size;
3053 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
3054 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3055 : WINED3D_SHADER_TYPE_PIXEL;
3056 if (shader->reg_maps.shader_version.major >= 4)
3057 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
3058 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
3059 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
3060 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
3061 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
3062 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3063 else
3064 args->next_shader_input_count = 0;
3065 args->swizzle_map = swizzle_map;
3066 if (d3d_info->emulated_flatshading)
3067 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3068 else
3069 args->flatshading = 0;
3070 }
3071
3072 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3073 {
3074 if (usage_idx1 != usage_idx2)
3075 return FALSE;
3076 if (usage1 == usage2)
3077 return TRUE;
3078 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3079 return TRUE;
3080 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3081 return TRUE;
3082
3083 return FALSE;
3084 }
3085
3086 BOOL vshader_get_input(const struct wined3d_shader *shader,
3087 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3088 {
3089 WORD map = shader->reg_maps.input_registers;
3090 unsigned int i;
3091
3092 for (i = 0; map; map >>= 1, ++i)
3093 {
3094 if (!(map & 1)) continue;
3095
3096 if (match_usage(shader->u.vs.attributes[i].usage,
3097 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3098 {
3099 *regnum = i;
3100 return TRUE;
3101 }
3102 }
3103 return FALSE;
3104 }
3105
3106 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3107 const struct wined3d_shader_signature *src, char **signature_strings)
3108 {
3109 struct wined3d_shader_signature_element *e;
3110 unsigned int i;
3111 SIZE_T len;
3112 char *ptr;
3113
3114 if (!src->element_count)
3115 return WINED3D_OK;
3116
3117 ptr = *signature_strings;
3118
3119 dst->element_count = src->element_count;
3120 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3121 return E_OUTOFMEMORY;
3122
3123 for (i = 0; i < src->element_count; ++i)
3124 {
3125 e = &src->elements[i];
3126 dst->elements[i] = *e;
3127
3128 len = strlen(e->semantic_name);
3129 memcpy(ptr, e->semantic_name, len + 1);
3130 dst->elements[i].semantic_name = ptr;
3131 ptr += len + 1;
3132 }
3133
3134 *signature_strings = ptr;
3135
3136 return WINED3D_OK;
3137 }
3138
3139 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3140 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3141 void *parent, const struct wined3d_parent_ops *parent_ops)
3142 {
3143 struct wined3d_shader_signature_element *e;
3144 size_t byte_code_size;
3145 SIZE_T total, len;
3146 unsigned int i;
3147 HRESULT hr;
3148 char *ptr;
3149
3150 TRACE("byte_code %p, byte_code_size %#lx, format %#x, max_version %#x.\n",
3151 desc->byte_code, (long)desc->byte_code_size, desc->format, desc->max_version);
3152
3153 if (!desc->byte_code)
3154 return WINED3DERR_INVALIDCALL;
3155
3156 if (!(shader->frontend = shader_select_frontend(desc->format)))
3157 {
3158 FIXME("Unable to find frontend for shader.\n");
3159 return WINED3DERR_INVALIDCALL;
3160 }
3161
3162 shader->ref = 1;
3163 shader->device = device;
3164 shader->parent = parent;
3165 shader->parent_ops = parent_ops;
3166
3167 total = 0;
3168 for (i = 0; i < desc->input_signature.element_count; ++i)
3169 {
3170 e = &desc->input_signature.elements[i];
3171 len = strlen(e->semantic_name);
3172 if (len >= ~(SIZE_T)0 - total)
3173 return E_OUTOFMEMORY;
3174
3175 total += len + 1;
3176 }
3177 for (i = 0; i < desc->output_signature.element_count; ++i)
3178 {
3179 e = &desc->output_signature.elements[i];
3180 len = strlen(e->semantic_name);
3181 if (len >= ~(SIZE_T)0 - total)
3182 return E_OUTOFMEMORY;
3183
3184 total += len + 1;
3185 }
3186 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3187 return E_OUTOFMEMORY;
3188 ptr = shader->signature_strings;
3189
3190 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3191 {
3192 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3193 return hr;
3194 }
3195 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3196 {
3197 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3198 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3199 return hr;
3200 }
3201
3202 list_init(&shader->linked_programs);
3203 list_init(&shader->constantsF);
3204 list_init(&shader->constantsB);
3205 list_init(&shader->constantsI);
3206 shader->lconst_inf_or_nan = FALSE;
3207 list_init(&shader->reg_maps.indexable_temps);
3208 list_add_head(&device->shaders, &shader->shader_list_entry);
3209
3210 byte_code_size = desc->byte_code_size;
3211 if (byte_code_size == ~(size_t)0)
3212 {
3213 const struct wined3d_shader_frontend *fe = shader->frontend;
3214 struct wined3d_shader_version shader_version;
3215 struct wined3d_shader_instruction ins;
3216 const DWORD *ptr;
3217 void *fe_data;
3218
3219 if (!(fe_data = fe->shader_init(desc->byte_code, byte_code_size, &shader->output_signature)))
3220 {
3221 WARN("Failed to initialise frontend data.\n");
3222 shader_cleanup(shader);
3223 return WINED3DERR_INVALIDCALL;
3224 }
3225
3226 fe->shader_read_header(fe_data, &ptr, &shader_version);
3227 while (!fe->shader_is_end(fe_data, &ptr))
3228 fe->shader_read_instruction(fe_data, &ptr, &ins);
3229
3230 fe->shader_free(fe_data);
3231
3232 byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3233 }
3234
3235 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, byte_code_size)))
3236 {
3237 shader_cleanup(shader);
3238 return E_OUTOFMEMORY;
3239 }
3240 memcpy(shader->function, desc->byte_code, byte_code_size);
3241 shader->functionLength = byte_code_size;
3242
3243 if (FAILED(hr = shader_set_function(shader, float_const_count, type, desc->max_version)))
3244 {
3245 WARN("Failed to set function, hr %#x.\n", hr);
3246 shader_cleanup(shader);
3247 }
3248
3249 shader->load_local_constsF = shader->lconst_inf_or_nan;
3250
3251 return hr;
3252 }
3253
3254 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3255 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3256 {
3257 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3258 unsigned int i;
3259 HRESULT hr;
3260
3261 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3262 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3263 return hr;
3264
3265 for (i = 0; i < shader->input_signature.element_count; ++i)
3266 {
3267 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3268
3269 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3270 continue;
3271
3272 shader->u.vs.attributes[input->register_idx].usage =
3273 shader_usage_from_semantic_name(input->semantic_name);
3274 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3275 }
3276
3277 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3278 shader->load_local_constsF = TRUE;
3279
3280 return WINED3D_OK;
3281 }
3282
3283 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3284 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3285 {
3286 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3287 }
3288
3289 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3290 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3291 {
3292 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3293 }
3294
3295 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3296 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3297 {
3298 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3299 }
3300
3301 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3302 struct gs_compile_args *args)
3303 {
3304 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3305 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3306 }
3307
3308 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3309 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3310 {
3311 const struct wined3d_gl_info *gl_info = context->gl_info;
3312 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3313 const struct wined3d_texture *texture;
3314 UINT i;
3315
3316 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3317 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3318 {
3319 static unsigned int warned = 0;
3320
3321 args->srgb_correction = 1;
3322 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3323 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3324 "support, expect rendering artifacts.\n");
3325 }
3326
3327 if (shader->reg_maps.shader_version.major == 1
3328 && shader->reg_maps.shader_version.minor <= 3)
3329 {
3330 for (i = 0; i < shader->limits->sampler; ++i)
3331 {
3332 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3333
3334 if (flags & WINED3D_TTFF_PROJECTED)
3335 {
3336 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3337
3338 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3339 {
3340 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3341 unsigned int j;
3342 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3343 DWORD max_valid = WINED3D_TTFF_COUNT4;
3344
3345 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3346 {
3347 struct wined3d_vertex_declaration_element *element =
3348 &state->vertex_declaration->elements[j];
3349
3350 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3351 && element->usage_idx == index)
3352 {
3353 max_valid = element->format->component_count;
3354 break;
3355 }
3356 }
3357 if (!tex_transform || tex_transform > max_valid)
3358 {
3359 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3360 tex_transform, max_valid);
3361 tex_transform = max_valid;
3362 }
3363 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3364 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3365 && tex_transform > WINED3D_TTFF_COUNT2)
3366 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3367 && tex_transform > WINED3D_TTFF_COUNT3))
3368 tex_transform |= WINED3D_PSARGS_PROJECTED;
3369 else
3370 {
3371 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3372 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3373 i, tex_transform, resource_type);
3374 }
3375 }
3376 else
3377 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3378
3379 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3380 }
3381 }
3382 }
3383 if (shader->reg_maps.shader_version.major == 1
3384 && shader->reg_maps.shader_version.minor <= 4)
3385 {
3386 for (i = 0; i < shader->limits->sampler; ++i)
3387 {
3388 const struct wined3d_texture *texture = state->textures[i];
3389
3390 if (!shader->reg_maps.resource_info[i].type)
3391 continue;
3392
3393 /* Treat unbound textures as 2D. The dummy texture will provide
3394 * the proper sample value. The tex_types bitmap defaults to
3395 * 2D because of the memset. */
3396 if (!texture)
3397 continue;
3398
3399 switch (texture->target)
3400 {
3401 /* RECT textures are distinguished from 2D textures via np2_fixup */
3402 case GL_TEXTURE_RECTANGLE_ARB:
3403 case GL_TEXTURE_2D:
3404 break;
3405
3406 case GL_TEXTURE_3D:
3407 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3408 break;
3409
3410 case GL_TEXTURE_CUBE_MAP_ARB:
3411 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3412 break;
3413 }
3414 }
3415 }
3416
3417 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3418 {
3419 if (!shader->reg_maps.resource_info[i].type)
3420 continue;
3421
3422 texture = state->textures[i];
3423 if (!texture)
3424 {
3425 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3426 continue;
3427 }
3428 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3429 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3430 else
3431 args->color_fixup[i] = texture->resource.format->color_fixup;
3432
3433 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3434 args->shadow |= 1u << i;
3435
3436 /* Flag samplers that need NP2 texcoord fixup. */
3437 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3438 args->np2_fixup |= (1u << i);
3439 }
3440
3441 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3442 if (shader->reg_maps.shader_version.major >= 4)
3443 args->shadow = 0;
3444
3445 if (shader->reg_maps.shader_version.major >= 3)
3446 {
3447 if (position_transformed)
3448 args->vp_mode = pretransformed;
3449 else if (use_vs(state))
3450 args->vp_mode = vertexshader;
3451 else
3452 args->vp_mode = fixedfunction;
3453 args->fog = WINED3D_FFP_PS_FOG_OFF;
3454 }
3455 else
3456 {
3457 args->vp_mode = vertexshader;
3458 if (state->render_states[WINED3D_RS_FOGENABLE])
3459 {
3460 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3461 {
3462 case WINED3D_FOG_NONE:
3463 if (position_transformed || use_vs(state))
3464 {
3465 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3466 break;
3467 }
3468
3469 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3470 {
3471 case WINED3D_FOG_NONE: /* Fall through. */
3472 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3473 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3474 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3475 }
3476 break;
3477
3478 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3479 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3480 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3481 }
3482 }
3483 else
3484 {
3485 args->fog = WINED3D_FFP_PS_FOG_OFF;
3486 }
3487 }
3488
3489 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3490 {
3491 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3492
3493 args->texcoords_initialized = 0;
3494 for (i = 0; i < MAX_TEXTURES; ++i)
3495 {
3496 if (vs)
3497 {
3498 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3499 args->texcoords_initialized |= 1u << i;
3500 }
3501 else
3502 {
3503 const struct wined3d_stream_info *si = &context->stream_info;
3504 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3505
3506 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3507 & WINED3D_FFP_TCI_MASK
3508 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3509 args->texcoords_initialized |= 1u << i;
3510 }
3511 }
3512 }
3513 else
3514 {
3515 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3516 }
3517
3518 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3519 && state->gl_primitive_type == GL_POINTS;
3520
3521 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3522 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3523 else
3524 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3525 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3526 : WINED3D_CMP_ALWAYS) - 1;
3527
3528 if (d3d_info->emulated_flatshading)
3529 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3530
3531 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3532 ? context->render_offscreen : 0;
3533 }
3534
3535 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3536 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3537 {
3538 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3539 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3540 HRESULT hr;
3541
3542 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3543 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3544 return hr;
3545
3546 for (i = 0; i < MAX_REG_INPUT; ++i)
3547 {
3548 if (shader->u.ps.input_reg_used & (1u << i))
3549 {
3550 ++num_regs_used;
3551 highest_reg_used = i;
3552 }
3553 }
3554
3555 /* Don't do any register mapping magic if it is not needed, or if we can't
3556 * achieve anything anyway */
3557 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3558 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3559 || shader->reg_maps.shader_version.major >= 4)
3560 {
3561 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3562 {
3563 /* This happens with relative addressing. The input mapper function
3564 * warns about this if the higher registers are declared too, so
3565 * don't write a FIXME here */
3566 WARN("More varying registers used than supported\n");
3567 }
3568
3569 for (i = 0; i < MAX_REG_INPUT; ++i)
3570 {
3571 shader->u.ps.input_reg_map[i] = i;
3572 }
3573
3574 shader->u.ps.declared_in_count = highest_reg_used + 1;
3575 }
3576 else
3577 {
3578 shader->u.ps.declared_in_count = 0;
3579 for (i = 0; i < MAX_REG_INPUT; ++i)
3580 {
3581 if (shader->u.ps.input_reg_used & (1u << i))
3582 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3583 else shader->u.ps.input_reg_map[i] = ~0U;
3584 }
3585 }
3586
3587 return WINED3D_OK;
3588 }
3589
3590 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3591 {
3592 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3593 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3594 unsigned int i;
3595
3596 if (reg_maps->shader_version.major != 1) return;
3597
3598 for (i = 0; i < shader->limits->sampler; ++i)
3599 {
3600 /* We don't sample from this sampler. */
3601 if (!resource_info[i].type)
3602 continue;
3603
3604 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3605 {
3606 case WINED3D_SHADER_TEX_2D:
3607 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3608 break;
3609
3610 case WINED3D_SHADER_TEX_3D:
3611 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3612 break;
3613
3614 case WINED3D_SHADER_TEX_CUBE:
3615 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3616 break;
3617 }
3618 }
3619 }
3620
3621 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3622 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3623 {
3624 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3625 }
3626
3627 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3628 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3629 {
3630 struct wined3d_shader *object;
3631 HRESULT hr;
3632
3633 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3634 device, desc, parent, parent_ops, shader);
3635
3636 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3637 return E_OUTOFMEMORY;
3638
3639 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3640 {
3641 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3642 HeapFree(GetProcessHeap(), 0, object);
3643 return hr;
3644 }
3645
3646 TRACE("Created compute shader %p.\n", object);
3647 *shader = object;
3648
3649 return WINED3D_OK;
3650 }
3651
3652 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3653 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3654 {
3655 struct wined3d_shader *object;
3656 HRESULT hr;
3657
3658 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3659 device, desc, parent, parent_ops, shader);
3660
3661 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3662 return E_OUTOFMEMORY;
3663
3664 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3665 {
3666 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3667 HeapFree(GetProcessHeap(), 0, object);
3668 return hr;
3669 }
3670
3671 TRACE("Created domain shader %p.\n", object);
3672 *shader = object;
3673
3674 return WINED3D_OK;
3675 }
3676
3677 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3678 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3679 {
3680 struct wined3d_shader *object;
3681 HRESULT hr;
3682
3683 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3684 device, desc, parent, parent_ops, shader);
3685
3686 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3687 if (!object)
3688 return E_OUTOFMEMORY;
3689
3690 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3691 {
3692 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3693 HeapFree(GetProcessHeap(), 0, object);
3694 return hr;
3695 }
3696
3697 TRACE("Created geometry shader %p.\n", object);
3698 *shader = object;
3699
3700 return WINED3D_OK;
3701 }
3702
3703 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3704 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3705 {
3706 struct wined3d_shader *object;
3707 HRESULT hr;
3708
3709 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3710 device, desc, parent, parent_ops, shader);
3711
3712 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3713 return E_OUTOFMEMORY;
3714
3715 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3716 {
3717 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3718 HeapFree(GetProcessHeap(), 0, object);
3719 return hr;
3720 }
3721
3722 TRACE("Created hull shader %p.\n", object);
3723 *shader = object;
3724
3725 return WINED3D_OK;
3726 }
3727
3728 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3729 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3730 {
3731 struct wined3d_shader *object;
3732 HRESULT hr;
3733
3734 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3735 device, desc, parent, parent_ops, shader);
3736
3737 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3738 if (!object)
3739 return E_OUTOFMEMORY;
3740
3741 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3742 {
3743 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3744 HeapFree(GetProcessHeap(), 0, object);
3745 return hr;
3746 }
3747
3748 TRACE("Created pixel shader %p.\n", object);
3749 *shader = object;
3750
3751 return WINED3D_OK;
3752 }
3753
3754 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3755 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3756 {
3757 struct wined3d_shader *object;
3758 HRESULT hr;
3759
3760 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3761 device, desc, parent, parent_ops, shader);
3762
3763 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3764 if (!object)
3765 return E_OUTOFMEMORY;
3766
3767 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3768 {
3769 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3770 HeapFree(GetProcessHeap(), 0, object);
3771 return hr;
3772 }
3773
3774 TRACE("Created vertex shader %p.\n", object);
3775 *shader = object;
3776
3777 return WINED3D_OK;
3778 }
Windows API implementation