?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Add cube array sampler support.
[wine.git] / dlls / wined3d / shader.c
blobdcf77e93c799088157351c33ba95dd67e48eeb2c
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <math.h>
29 #include <stdio.h>
30 #include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
47 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
48 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
49 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
50 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
51 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
52 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
53 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
54 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
55 /* WINED3DSIH_BEM */ "bem",
56 /* WINED3DSIH_BFI */ "bfi",
57 /* WINED3DSIH_BFREV */ "bfrev",
58 /* WINED3DSIH_BREAK */ "break",
59 /* WINED3DSIH_BREAKC */ "breakc",
60 /* WINED3DSIH_BREAKP */ "breakp",
61 /* WINED3DSIH_BUFINFO */ "bufinfo",
62 /* WINED3DSIH_CALL */ "call",
63 /* WINED3DSIH_CALLNZ */ "callnz",
64 /* WINED3DSIH_CASE */ "case",
65 /* WINED3DSIH_CMP */ "cmp",
66 /* WINED3DSIH_CND */ "cnd",
67 /* WINED3DSIH_CONTINUE */ "continue",
68 /* WINED3DSIH_COUNTBITS */ "countbits",
69 /* WINED3DSIH_CRS */ "crs",
70 /* WINED3DSIH_CUT */ "cut",
71 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
72 /* WINED3DSIH_DCL */ "dcl",
73 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
74 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
75 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
76 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
77 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
78 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
79 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
80 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
81 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
82 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
83 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
84 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
85 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
86 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
87 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
88 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
89 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
90 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
91 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
92 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
93 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
94 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
95 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
96 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
97 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
98 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
99 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
100 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
101 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
102 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
103 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
104 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
105 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
106 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
107 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
108 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
109 /* WINED3DSIH_DEF */ "def",
110 /* WINED3DSIH_DEFAULT */ "default",
111 /* WINED3DSIH_DEFB */ "defb",
112 /* WINED3DSIH_DEFI */ "defi",
113 /* WINED3DSIH_DIV */ "div",
114 /* WINED3DSIH_DP2 */ "dp2",
115 /* WINED3DSIH_DP2ADD */ "dp2add",
116 /* WINED3DSIH_DP3 */ "dp3",
117 /* WINED3DSIH_DP4 */ "dp4",
118 /* WINED3DSIH_DST */ "dst",
119 /* WINED3DSIH_DSX */ "dsx",
120 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
121 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
122 /* WINED3DSIH_DSY */ "dsy",
123 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
124 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
125 /* WINED3DSIH_ELSE */ "else",
126 /* WINED3DSIH_EMIT */ "emit",
127 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
128 /* WINED3DSIH_ENDIF */ "endif",
129 /* WINED3DSIH_ENDLOOP */ "endloop",
130 /* WINED3DSIH_ENDREP */ "endrep",
131 /* WINED3DSIH_ENDSWITCH */ "endswitch",
132 /* WINED3DSIH_EQ */ "eq",
133 /* WINED3DSIH_EXP */ "exp",
134 /* WINED3DSIH_EXPP */ "expp",
135 /* WINED3DSIH_F16TOF32 */ "f16tof32",
136 /* WINED3DSIH_F32TOF16 */ "f32tof16",
137 /* WINED3DSIH_FCALL */ "fcall",
138 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
139 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
140 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
141 /* WINED3DSIH_FRC */ "frc",
142 /* WINED3DSIH_FTOI */ "ftoi",
143 /* WINED3DSIH_FTOU */ "ftou",
144 /* WINED3DSIH_GATHER4 */ "gather4",
145 /* WINED3DSIH_GATHER4_C */ "gather4_c",
146 /* WINED3DSIH_GE */ "ge",
147 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
148 /* WINED3DSIH_HS_DECLS */ "hs_decls",
149 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
150 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
151 /* WINED3DSIH_IADD */ "iadd",
152 /* WINED3DSIH_IEQ */ "ieq",
153 /* WINED3DSIH_IF */ "if",
154 /* WINED3DSIH_IFC */ "ifc",
155 /* WINED3DSIH_IGE */ "ige",
156 /* WINED3DSIH_ILT */ "ilt",
157 /* WINED3DSIH_IMAD */ "imad",
158 /* WINED3DSIH_IMAX */ "imax",
159 /* WINED3DSIH_IMIN */ "imin",
160 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
161 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
162 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
163 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
164 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
165 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
166 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
167 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
168 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
169 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
170 /* WINED3DSIH_IMUL */ "imul",
171 /* WINED3DSIH_INE */ "ine",
172 /* WINED3DSIH_INEG */ "ineg",
173 /* WINED3DSIH_ISHL */ "ishl",
174 /* WINED3DSIH_ISHR */ "ishr",
175 /* WINED3DSIH_ITOF */ "itof",
176 /* WINED3DSIH_LABEL */ "label",
177 /* WINED3DSIH_LD */ "ld",
178 /* WINED3DSIH_LD2DMS */ "ld2dms",
179 /* WINED3DSIH_LD_RAW */ "ld_raw",
180 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
181 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
182 /* WINED3DSIH_LIT */ "lit",
183 /* WINED3DSIH_LOD */ "lod",
184 /* WINED3DSIH_LOG */ "log",
185 /* WINED3DSIH_LOGP */ "logp",
186 /* WINED3DSIH_LOOP */ "loop",
187 /* WINED3DSIH_LRP */ "lrp",
188 /* WINED3DSIH_LT */ "lt",
189 /* WINED3DSIH_M3x2 */ "m3x2",
190 /* WINED3DSIH_M3x3 */ "m3x3",
191 /* WINED3DSIH_M3x4 */ "m3x4",
192 /* WINED3DSIH_M4x3 */ "m4x3",
193 /* WINED3DSIH_M4x4 */ "m4x4",
194 /* WINED3DSIH_MAD */ "mad",
195 /* WINED3DSIH_MAX */ "max",
196 /* WINED3DSIH_MIN */ "min",
197 /* WINED3DSIH_MOV */ "mov",
198 /* WINED3DSIH_MOVA */ "mova",
199 /* WINED3DSIH_MOVC */ "movc",
200 /* WINED3DSIH_MUL */ "mul",
201 /* WINED3DSIH_NE */ "ne",
202 /* WINED3DSIH_NOP */ "nop",
203 /* WINED3DSIH_NOT */ "not",
204 /* WINED3DSIH_NRM */ "nrm",
205 /* WINED3DSIH_OR */ "or",
206 /* WINED3DSIH_PHASE */ "phase",
207 /* WINED3DSIH_POW */ "pow",
208 /* WINED3DSIH_RCP */ "rcp",
209 /* WINED3DSIH_REP */ "rep",
210 /* WINED3DSIH_RESINFO */ "resinfo",
211 /* WINED3DSIH_RET */ "ret",
212 /* WINED3DSIH_ROUND_NE */ "round_ne",
213 /* WINED3DSIH_ROUND_NI */ "round_ni",
214 /* WINED3DSIH_ROUND_PI */ "round_pi",
215 /* WINED3DSIH_ROUND_Z */ "round_z",
216 /* WINED3DSIH_RSQ */ "rsq",
217 /* WINED3DSIH_SAMPLE */ "sample",
218 /* WINED3DSIH_SAMPLE_B */ "sample_b",
219 /* WINED3DSIH_SAMPLE_C */ "sample_c",
220 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
221 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
222 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
223 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
224 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
225 /* WINED3DSIH_SETP */ "setp",
226 /* WINED3DSIH_SGE */ "sge",
227 /* WINED3DSIH_SGN */ "sgn",
228 /* WINED3DSIH_SINCOS */ "sincos",
229 /* WINED3DSIH_SLT */ "slt",
230 /* WINED3DSIH_SQRT */ "sqrt",
231 /* WINED3DSIH_STORE_RAW */ "store_raw",
232 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
233 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
234 /* WINED3DSIH_SUB */ "sub",
235 /* WINED3DSIH_SWAPC */ "swapc",
236 /* WINED3DSIH_SWITCH */ "switch",
237 /* WINED3DSIH_SYNC */ "sync",
238 /* WINED3DSIH_TEX */ "texld",
239 /* WINED3DSIH_TEXBEM */ "texbem",
240 /* WINED3DSIH_TEXBEML */ "texbeml",
241 /* WINED3DSIH_TEXCOORD */ "texcrd",
242 /* WINED3DSIH_TEXDEPTH */ "texdepth",
243 /* WINED3DSIH_TEXDP3 */ "texdp3",
244 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
245 /* WINED3DSIH_TEXKILL */ "texkill",
246 /* WINED3DSIH_TEXLDD */ "texldd",
247 /* WINED3DSIH_TEXLDL */ "texldl",
248 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
249 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
250 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
251 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
252 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
253 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
254 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
255 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
256 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
257 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
258 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
259 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
260 /* WINED3DSIH_UBFE */ "ubfe",
261 /* WINED3DSIH_UDIV */ "udiv",
262 /* WINED3DSIH_UGE */ "uge",
263 /* WINED3DSIH_ULT */ "ult",
264 /* WINED3DSIH_UMAX */ "umax",
265 /* WINED3DSIH_UMIN */ "umin",
266 /* WINED3DSIH_USHR */ "ushr",
267 /* WINED3DSIH_UTOF */ "utof",
268 /* WINED3DSIH_XOR */ "xor",
269 };
270
271 static const char * const semantic_names[] =
272 {
273 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
274 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
275 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
276 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
277 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
278 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
279 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
280 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
281 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
282 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
283 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
284 /* WINED3D_DECL_USAGE_FOG */ "FOG",
285 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
286 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
287 };
288
289 static const struct
290 {
291 enum wined3d_sysval_semantic sysval_semantic;
292 const char *sysval_name;
293 }
294 sysval_semantic_names[] =
295 {
296 {WINED3D_SV_POSITION, "SV_Position"},
297 {WINED3D_SV_CLIP_DISTANCE, "SV_ClipDistance"},
298 {WINED3D_SV_CULL_DISTANCE, "SV_CullDistance"},
299 {WINED3D_SV_RENDER_TARGET_ARRAY_INDEX, "SV_RenderTargetArrayIndex"},
300 {WINED3D_SV_VIEWPORT_ARRAY_INDEX, "SV_ViewportArrayIndex"},
301 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
302 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
303 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
304 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
305 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
306 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
307 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
308 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
309 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
310 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
311 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
312 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
313 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
314 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
315 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
316 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
317 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
318 };
319
320 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
321 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
322
323 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
324 {
325 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
326 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
327
328 return shader_opcode_names[handler_idx];
329 }
330
331 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
332 {
333 if (usage >= ARRAY_SIZE(semantic_names))
334 {
335 FIXME("Unrecognized usage %#x.\n", usage);
336 return "UNRECOGNIZED";
337 }
338
339 return semantic_names[usage];
340 }
341
342 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
343 {
344 unsigned int i;
345
346 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
347 {
348 if (!strcmp(name, semantic_names[i]))
349 return i;
350 }
351
352 return ~0U;
353 }
354
355 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
356 {
357 switch (usage)
358 {
359 case WINED3D_DECL_USAGE_POSITION:
360 return WINED3D_SV_POSITION;
361 default:
362 return 0;
363 }
364 }
365
366 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
367 {
368 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
369 }
370
371 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
372 const struct wined3d_shader_semantic *s)
373 {
374 e->semantic_name = shader_semantic_name_from_usage(s->usage);
375 e->semantic_idx = s->usage_idx;
376 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
377 e->component_type = WINED3D_TYPE_FLOAT;
378 e->register_idx = s->reg.reg.idx[0].offset;
379 e->mask = s->reg.write_mask;
380 }
381
382 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
383 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
384 {
385 e->semantic_name = shader_semantic_name_from_usage(usage);
386 e->semantic_idx = usage_idx;
387 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
388 e->component_type = WINED3D_TYPE_FLOAT;
389 e->register_idx = reg_idx;
390 e->mask = write_mask;
391 }
392
393 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
394 {
395 switch (version_token >> 16)
396 {
397 case WINED3D_SM1_VS:
398 case WINED3D_SM1_PS:
399 return &sm1_shader_frontend;
400
401 case WINED3D_SM4_PS:
402 case WINED3D_SM4_VS:
403 case WINED3D_SM4_GS:
404 case WINED3D_SM5_HS:
405 case WINED3D_SM5_DS:
406 case WINED3D_SM5_CS:
407 return &sm4_shader_frontend;
408
409 default:
410 FIXME("Unrecognised version token %#x.\n", version_token);
411 return NULL;
412 }
413 }
414
415 void string_buffer_clear(struct wined3d_string_buffer *buffer)
416 {
417 buffer->buffer[0] = '\0';
418 buffer->content_size = 0;
419 }
420
421 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
422 {
423 buffer->buffer_size = 32;
424 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
425 {
426 ERR("Failed to allocate shader buffer memory.\n");
427 return FALSE;
428 }
429
430 string_buffer_clear(buffer);
431 return TRUE;
432 }
433
434 void string_buffer_free(struct wined3d_string_buffer *buffer)
435 {
436 HeapFree(GetProcessHeap(), 0, buffer->buffer);
437 }
438
439 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
440 {
441 char *new_buffer;
442 unsigned int new_buffer_size = buffer->buffer_size * 2;
443
444 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
445 new_buffer_size *= 2;
446 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
447 {
448 ERR("Failed to grow buffer.\n");
449 buffer->buffer[buffer->content_size] = '\0';
450 return FALSE;
451 }
452 buffer->buffer = new_buffer;
453 buffer->buffer_size = new_buffer_size;
454 return TRUE;
455 }
456
457 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
458 {
459 unsigned int rem;
460 int rc;
461
462 rem = buffer->buffer_size - buffer->content_size;
463 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
464 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
465 return rc;
466
467 buffer->content_size += rc;
468 return 0;
469 }
470
471 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
472 {
473 va_list args;
474 int ret;
475
476 for (;;)
477 {
478 va_start(args, format);
479 ret = shader_vaddline(buffer, format, args);
480 va_end(args);
481 if (!ret)
482 return ret;
483 if (!string_buffer_resize(buffer, ret))
484 return -1;
485 }
486 }
487
488 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
489 {
490 struct wined3d_string_buffer *buffer;
491
492 if (list_empty(&list->list))
493 {
494 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
495 if (!buffer || !string_buffer_init(buffer))
496 {
497 ERR("Couldn't allocate buffer for temporary string.\n");
498 HeapFree(GetProcessHeap(), 0, buffer);
499 return NULL;
500 }
501 }
502 else
503 {
504 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
505 list_remove(&buffer->entry);
506 }
507 string_buffer_clear(buffer);
508 return buffer;
509 }
510
511 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
512 {
513 if (!buffer)
514 return 0;
515 string_buffer_clear(buffer);
516 return shader_vaddline(buffer, format, args);
517 }
518
519 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
520 {
521 va_list args;
522 int ret;
523
524 for (;;)
525 {
526 va_start(args, format);
527 ret = string_buffer_vsprintf(buffer, format, args);
528 va_end(args);
529 if (!ret)
530 return;
531 if (!string_buffer_resize(buffer, ret))
532 return;
533 }
534 }
535
536 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
537 {
538 if (!buffer)
539 return;
540 list_add_head(&list->list, &buffer->entry);
541 }
542
543 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
544 {
545 list_init(&list->list);
546 }
547
548 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
549 {
550 struct wined3d_string_buffer *buffer, *buffer_next;
551
552 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
553 {
554 string_buffer_free(buffer);
555 HeapFree(GetProcessHeap(), 0, buffer);
556 }
557 list_init(&list->list);
558 }
559
560 /* Convert floating point offset relative to a register file to an absolute
561 * offset for float constants. */
562 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
563 {
564 switch (register_type)
565 {
566 case WINED3DSPR_CONST: return register_idx;
567 case WINED3DSPR_CONST2: return 2048 + register_idx;
568 case WINED3DSPR_CONST3: return 4096 + register_idx;
569 case WINED3DSPR_CONST4: return 6144 + register_idx;
570 default:
571 FIXME("Unsupported register type: %u.\n", register_type);
572 return register_idx;
573 }
574 }
575
576 static void shader_delete_constant_list(struct list *clist)
577 {
578 struct wined3d_shader_lconst *constant, *constant_next;
579
580 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
581 HeapFree(GetProcessHeap(), 0, constant);
582 list_init(clist);
583 }
584
585 static void shader_set_limits(struct wined3d_shader *shader)
586 {
587 static const struct limits_entry
588 {
589 unsigned int min_version;
590 unsigned int max_version;
591 struct wined3d_shader_limits limits;
592 }
593 vs_limits[] =
594 {
595 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
596 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
597 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
598 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
599 * even though they are capable of supporting much more (GL
600 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
601 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
602 * shaders to 256. */
603 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
604 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
605 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
606 {0}
607 },
608 hs_limits[] =
609 {
610 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
611 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
612 },
613 ds_limits[] =
614 {
615 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
616 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
617 },
618 gs_limits[] =
619 {
620 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
621 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
622 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
623 {0}
624 },
625 ps_limits[] =
626 {
627 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
628 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
629 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
630 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
631 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
632 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
633 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
634 {0}
635 },
636 cs_limits[] =
637 {
638 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
639 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
640 };
641 const struct limits_entry *limits_array;
642 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
643 shader->reg_maps.shader_version.minor);
644 int i = 0;
645
646 switch (shader->reg_maps.shader_version.type)
647 {
648 default:
649 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
650 /* Fall-through. */
651 case WINED3D_SHADER_TYPE_VERTEX:
652 limits_array = vs_limits;
653 break;
654 case WINED3D_SHADER_TYPE_HULL:
655 limits_array = hs_limits;
656 break;
657 case WINED3D_SHADER_TYPE_DOMAIN:
658 limits_array = ds_limits;
659 break;
660 case WINED3D_SHADER_TYPE_GEOMETRY:
661 limits_array = gs_limits;
662 break;
663 case WINED3D_SHADER_TYPE_PIXEL:
664 limits_array = ps_limits;
665 break;
666 case WINED3D_SHADER_TYPE_COMPUTE:
667 limits_array = cs_limits;
668 break;
669 }
670
671 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
672 {
673 if (shader_version <= limits_array[i].max_version)
674 {
675 shader->limits = &limits_array[i].limits;
676 break;
677 }
678 ++i;
679 }
680 if (!shader->limits)
681 {
682 FIXME("Unexpected shader version \"%u.%u\".\n",
683 shader->reg_maps.shader_version.major,
684 shader->reg_maps.shader_version.minor);
685 shader->limits = &limits_array[max(0, i - 1)].limits;
686 }
687 }
688
689 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
690 {
691 DWORD idx, shift;
692 idx = bit >> 5;
693 shift = bit & 0x1f;
694 bitmap[idx] |= (1u << shift);
695 }
696
697 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
698 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
699 {
700 switch (reg->type)
701 {
702 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
703 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
704 reg_maps->texcoord |= 1u << reg->idx[0].offset;
705 else
706 reg_maps->address |= 1u << reg->idx[0].offset;
707 break;
708
709 case WINED3DSPR_TEMP:
710 reg_maps->temporary |= 1u << reg->idx[0].offset;
711 break;
712
713 case WINED3DSPR_INPUT:
714 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
715 {
716 if (reg->idx[0].rel_addr)
717 {
718 /* If relative addressing is used, we must assume that all registers
719 * are used. Even if it is a construct like v3[aL], we can't assume
720 * that v0, v1 and v2 aren't read because aL can be negative */
721 unsigned int i;
722 for (i = 0; i < MAX_REG_INPUT; ++i)
723 {
724 shader->u.ps.input_reg_used[i] = TRUE;
725 }
726 }
727 else
728 {
729 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
730 }
731 }
732 else
733 reg_maps->input_registers |= 1u << reg->idx[0].offset;
734 break;
735
736 case WINED3DSPR_RASTOUT:
737 if (reg->idx[0].offset == 1)
738 reg_maps->fog = 1;
739 if (reg->idx[0].offset == 2)
740 reg_maps->point_size = 1;
741 break;
742
743 case WINED3DSPR_MISCTYPE:
744 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
745 {
746 if (!reg->idx[0].offset)
747 reg_maps->vpos = 1;
748 else if (reg->idx[0].offset == 1)
749 reg_maps->usesfacing = 1;
750 }
751 break;
752
753 case WINED3DSPR_CONST:
754 if (reg->idx[0].rel_addr)
755 {
756 if (reg->idx[0].offset < reg_maps->min_rel_offset)
757 reg_maps->min_rel_offset = reg->idx[0].offset;
758 if (reg->idx[0].offset > reg_maps->max_rel_offset)
759 reg_maps->max_rel_offset = reg->idx[0].offset;
760 reg_maps->usesrelconstF = TRUE;
761 }
762 else
763 {
764 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
765 {
766 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
767 return FALSE;
768 }
769 else
770 {
771 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
772 }
773 }
774 break;
775
776 case WINED3DSPR_CONSTINT:
777 if (reg->idx[0].offset >= shader->limits->constant_int)
778 {
779 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
780 return FALSE;
781 }
782 else
783 {
784 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
785 }
786 break;
787
788 case WINED3DSPR_CONSTBOOL:
789 if (reg->idx[0].offset >= shader->limits->constant_bool)
790 {
791 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
792 return FALSE;
793 }
794 else
795 {
796 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
797 }
798 break;
799
800 case WINED3DSPR_COLOROUT:
801 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
802 break;
803
804 default:
805 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
806 reg->type, reg->idx[0].offset, reg->idx[1].offset);
807 break;
808 }
809 return TRUE;
810 }
811
812 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
813 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
814 {
815 struct wined3d_shader_sampler_map_entry *entries, *entry;
816 struct wined3d_shader_sampler_map *map;
817 unsigned int i;
818
819 map = &reg_maps->sampler_map;
820 entries = map->entries;
821 for (i = 0; i < map->count; ++i)
822 {
823 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
824 return;
825 }
826
827 if (!map->size)
828 {
829 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
830 {
831 ERR("Failed to allocate sampler map entries.\n");
832 return;
833 }
834 map->size = 4;
835 map->entries = entries;
836 }
837 else if (map->count == map->size)
838 {
839 size_t new_size = map->size * 2;
840
841 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
842 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
843 {
844 ERR("Failed to resize sampler map entries.\n");
845 return;
846 }
847 map->size = new_size;
848 map->entries = entries;
849 }
850
851 entry = &entries[map->count++];
852 entry->resource_idx = resource_idx;
853 entry->sampler_idx = sampler_idx;
854 entry->bind_idx = bind_idx;
855 }
856
857 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
858 {
859 switch (instr)
860 {
861 case WINED3DSIH_M4x4:
862 case WINED3DSIH_M3x4:
863 return param == 1 ? 3 : 0;
864
865 case WINED3DSIH_M4x3:
866 case WINED3DSIH_M3x3:
867 return param == 1 ? 2 : 0;
868
869 case WINED3DSIH_M3x2:
870 return param == 1 ? 1 : 0;
871
872 default:
873 return 0;
874 }
875 }
876
877 /* Note that this does not count the loop register as an address register. */
878 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
879 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
880 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
881 {
882 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
883 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
884 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
885 void *fe_data = shader->frontend_data;
886 struct wined3d_shader_version shader_version;
887 const DWORD *ptr = byte_code;
888 unsigned int i;
889
890 memset(reg_maps, 0, sizeof(*reg_maps));
891 memset(input_signature_elements, 0, sizeof(input_signature_elements));
892 memset(output_signature_elements, 0, sizeof(output_signature_elements));
893 reg_maps->min_rel_offset = ~0U;
894 list_init(&reg_maps->indexable_temps);
895
896 fe->shader_read_header(fe_data, &ptr, &shader_version);
897 reg_maps->shader_version = shader_version;
898
899 shader_set_limits(shader);
900
901 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
902 sizeof(*reg_maps->constf))))
903 {
904 ERR("Failed to allocate constant map memory.\n");
905 return E_OUTOFMEMORY;
906 }
907
908 while (!fe->shader_is_end(fe_data, &ptr))
909 {
910 struct wined3d_shader_instruction ins;
911
912 /* Fetch opcode. */
913 fe->shader_read_instruction(fe_data, &ptr, &ins);
914
915 /* Unhandled opcode, and its parameters. */
916 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
917 {
918 TRACE("Skipping unrecognized instruction.\n");
919 continue;
920 }
921
922 /* Handle declarations. */
923 if (ins.handler_idx == WINED3DSIH_DCL
924 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
925 {
926 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
927 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
928
929 switch (semantic->reg.reg.type)
930 {
931 /* Mark input registers used. */
932 case WINED3DSPR_INPUT:
933 if (reg_idx >= MAX_REG_INPUT)
934 {
935 ERR("Invalid input register index %u.\n", reg_idx);
936 break;
937 }
938 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
939 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
940 return WINED3DERR_INVALIDCALL;
941 reg_maps->input_registers |= 1u << reg_idx;
942 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
943 break;
944
945 /* Vertex shader: mark 3.0 output registers used, save token. */
946 case WINED3DSPR_OUTPUT:
947 if (reg_idx >= MAX_REG_OUTPUT)
948 {
949 ERR("Invalid output register index %u.\n", reg_idx);
950 break;
951 }
952 reg_maps->output_registers |= 1u << reg_idx;
953 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
954 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
955 reg_maps->fog = 1;
956 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
957 reg_maps->point_size = 1;
958 break;
959
960 case WINED3DSPR_SAMPLER:
961 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
962 case WINED3DSPR_RESOURCE:
963 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
964 {
965 ERR("Invalid resource index %u.\n", reg_idx);
966 break;
967 }
968 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
969 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
970 break;
971
972 case WINED3DSPR_UAV:
973 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
974 {
975 ERR("Invalid UAV resource index %u.\n", reg_idx);
976 break;
977 }
978 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
979 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
980 break;
981
982 default:
983 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
984 break;
985 }
986 }
987 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
988 {
989 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
990 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
991 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
992 else
993 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
994 }
995 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
996 {
997 if (reg_maps->icb)
998 FIXME("Multiple immediate constant buffers.\n");
999 reg_maps->icb = ins.declaration.icb;
1000 }
1001 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1002 {
1003 struct wined3d_shader_indexable_temp *reg;
1004
1005 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
1006 return E_OUTOFMEMORY;
1007
1008 *reg = ins.declaration.indexable_temp;
1009 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1010 }
1011 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1012 {
1013 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1014 shader->u.gs.input_type = ins.declaration.primitive_type;
1015 else
1016 FIXME("Invalid instruction %#x for shader type %#x.\n",
1017 ins.handler_idx, shader_version.type);
1018 }
1019 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1020 {
1021 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1022 shader->u.gs.output_type = ins.declaration.primitive_type;
1023 else
1024 FIXME("Invalid instruction %#x for shader type %#x.\n",
1025 ins.handler_idx, shader_version.type);
1026 }
1027 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1028 {
1029 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1030 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1031 }
1032 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1033 {
1034 reg_maps->temporary_count = ins.declaration.count;
1035 }
1036 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1037 {
1038 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1039 {
1040 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1041 }
1042 else
1043 {
1044 FIXME("Invalid instruction %#x for shader type %#x.\n",
1045 ins.handler_idx, shader_version.type);
1046 }
1047 }
1048 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1049 {
1050 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1051 shader->u.gs.vertices_out = ins.declaration.count;
1052 else
1053 FIXME("Invalid instruction %#x for shader type %#x.\n",
1054 ins.handler_idx, shader_version.type);
1055 }
1056 else if (ins.handler_idx == WINED3DSIH_DEF)
1057 {
1058 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1059 float *value;
1060 if (!lconst) return E_OUTOFMEMORY;
1061
1062 lconst->idx = ins.dst[0].reg.idx[0].offset;
1063 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1064 value = (float *)lconst->value;
1065
1066 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1067 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1068 {
1069 if (value[0] < -1.0f) value[0] = -1.0f;
1070 else if (value[0] > 1.0f) value[0] = 1.0f;
1071 if (value[1] < -1.0f) value[1] = -1.0f;
1072 else if (value[1] > 1.0f) value[1] = 1.0f;
1073 if (value[2] < -1.0f) value[2] = -1.0f;
1074 else if (value[2] > 1.0f) value[2] = 1.0f;
1075 if (value[3] < -1.0f) value[3] = -1.0f;
1076 else if (value[3] > 1.0f) value[3] = 1.0f;
1077 }
1078
1079 list_add_head(&shader->constantsF, &lconst->entry);
1080
1081 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1082 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1083 {
1084 shader->lconst_inf_or_nan = TRUE;
1085 }
1086 }
1087 else if (ins.handler_idx == WINED3DSIH_DEFI)
1088 {
1089 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1090 if (!lconst) return E_OUTOFMEMORY;
1091
1092 lconst->idx = ins.dst[0].reg.idx[0].offset;
1093 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1094
1095 list_add_head(&shader->constantsI, &lconst->entry);
1096 reg_maps->local_int_consts |= (1u << lconst->idx);
1097 }
1098 else if (ins.handler_idx == WINED3DSIH_DEFB)
1099 {
1100 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1101 if (!lconst) return E_OUTOFMEMORY;
1102
1103 lconst->idx = ins.dst[0].reg.idx[0].offset;
1104 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1105
1106 list_add_head(&shader->constantsB, &lconst->entry);
1107 reg_maps->local_bool_consts |= (1u << lconst->idx);
1108 }
1109 /* For subroutine prototypes. */
1110 else if (ins.handler_idx == WINED3DSIH_LABEL)
1111 {
1112 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1113 }
1114 /* Set texture, address, temporary registers. */
1115 else
1116 {
1117 BOOL color0_mov = FALSE;
1118 unsigned int i;
1119
1120 /* This will loop over all the registers and try to
1121 * make a bitmask of the ones we're interested in.
1122 *
1123 * Relative addressing tokens are ignored, but that's
1124 * okay, since we'll catch any address registers when
1125 * they are initialized (required by spec). */
1126 for (i = 0; i < ins.dst_count; ++i)
1127 {
1128 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1129 shader_version.type, constf_size))
1130 return WINED3DERR_INVALIDCALL;
1131
1132 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1133 {
1134 UINT idx = ins.dst[i].reg.idx[0].offset;
1135
1136 switch (ins.dst[i].reg.type)
1137 {
1138 case WINED3DSPR_RASTOUT:
1139 if (shader_version.major >= 3)
1140 break;
1141 switch (idx)
1142 {
1143 case 0: /* oPos */
1144 reg_maps->output_registers |= 1u << 10;
1145 shader_signature_from_usage(&output_signature_elements[10],
1146 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1147 break;
1148
1149 case 1: /* oFog */
1150 reg_maps->output_registers |= 1u << 11;
1151 shader_signature_from_usage(&output_signature_elements[11],
1152 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1153 break;
1154
1155 case 2: /* oPts */
1156 reg_maps->output_registers |= 1u << 11;
1157 shader_signature_from_usage(&output_signature_elements[11],
1158 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1159 break;
1160 }
1161 break;
1162
1163 case WINED3DSPR_ATTROUT:
1164 if (shader_version.major >= 3)
1165 break;
1166 if (idx < 2)
1167 {
1168 idx += 8;
1169 if (reg_maps->output_registers & (1u << idx))
1170 {
1171 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1172 }
1173 else
1174 {
1175 reg_maps->output_registers |= 1u << idx;
1176 shader_signature_from_usage(&output_signature_elements[idx],
1177 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1178 }
1179 }
1180 break;
1181
1182 case WINED3DSPR_TEXCRDOUT:
1183 if (shader_version.major >= 3)
1184 {
1185 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1186 break;
1187 }
1188 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1189 if (reg_maps->output_registers & (1u << idx))
1190 {
1191 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1192 }
1193 else
1194 {
1195 reg_maps->output_registers |= 1u << idx;
1196 shader_signature_from_usage(&output_signature_elements[idx],
1197 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1198 }
1199 break;
1200
1201 default:
1202 break;
1203 }
1204 }
1205
1206 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1207 {
1208 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1209 {
1210 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1211 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1212 * the mov and perform the sRGB write correction from the source register.
1213 *
1214 * However, if the mov is only partial, we can't do this, and if the write
1215 * comes from an instruction other than MOV it is hard to do as well. If
1216 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1217 shader->u.ps.color0_mov = FALSE;
1218 if (ins.handler_idx == WINED3DSIH_MOV
1219 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1220 {
1221 /* Used later when the source register is read. */
1222 color0_mov = TRUE;
1223 }
1224 }
1225 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1226 * end
1227 */
1228 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1229 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1230 {
1231 shader->u.ps.color0_mov = FALSE;
1232 }
1233 }
1234
1235 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1236 if (shader_version.major == 1
1237 && (ins.handler_idx == WINED3DSIH_TEX
1238 || ins.handler_idx == WINED3DSIH_TEXBEM
1239 || ins.handler_idx == WINED3DSIH_TEXBEML
1240 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1241 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1242 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1243 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1244 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1245 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1246 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1247 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1248 {
1249 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1250
1251 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1252 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1253 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1254 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1255
1256 /* texbem is only valid with < 1.4 pixel shaders */
1257 if (ins.handler_idx == WINED3DSIH_TEXBEM
1258 || ins.handler_idx == WINED3DSIH_TEXBEML)
1259 {
1260 reg_maps->bumpmat |= 1u << reg_idx;
1261 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1262 {
1263 reg_maps->luminanceparams |= 1u << reg_idx;
1264 }
1265 }
1266 }
1267 else if (ins.handler_idx == WINED3DSIH_BEM)
1268 {
1269 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1270 }
1271 }
1272
1273 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD
1274 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED)
1275 {
1276 unsigned int reg_idx;
1277 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD)
1278 reg_idx = ins.dst[0].reg.idx[0].offset;
1279 else
1280 reg_idx = ins.src[1].reg.idx[0].offset;
1281 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1282 {
1283 ERR("Invalid UAV index %u.\n", reg_idx);
1284 break;
1285 }
1286 reg_maps->uav_read_mask |= (1u << reg_idx);
1287 }
1288 else if (ins.handler_idx == WINED3DSIH_NRM)
1289 {
1290 reg_maps->usesnrm = 1;
1291 }
1292 else if (ins.handler_idx == WINED3DSIH_DSY
1293 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1294 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1295 {
1296 reg_maps->usesdsy = 1;
1297 }
1298 else if (ins.handler_idx == WINED3DSIH_DSX
1299 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1300 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1301 {
1302 reg_maps->usesdsx = 1;
1303 }
1304 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1305 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1306 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1307 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1308 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1309 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1310 else if (ins.handler_idx == WINED3DSIH_LOOP
1311 || ins.handler_idx == WINED3DSIH_REP)
1312 {
1313 ++cur_loop_depth;
1314 if (cur_loop_depth > max_loop_depth)
1315 max_loop_depth = cur_loop_depth;
1316 }
1317 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1318 || ins.handler_idx == WINED3DSIH_ENDREP)
1319 {
1320 --cur_loop_depth;
1321 }
1322 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1323 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1324 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1325 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1326 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1327 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1328 {
1329 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1330 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1331 }
1332 else if (ins.handler_idx == WINED3DSIH_LD
1333 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1334 {
1335 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1336 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1337 }
1338
1339 if (ins.predicate)
1340 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1341 shader_version.type, constf_size))
1342 return WINED3DERR_INVALIDCALL;
1343
1344 for (i = 0; i < ins.src_count; ++i)
1345 {
1346 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1347 struct wined3d_shader_register reg = ins.src[i].reg;
1348
1349 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1350 shader_version.type, constf_size))
1351 return WINED3DERR_INVALIDCALL;
1352 while (count)
1353 {
1354 ++reg.idx[0].offset;
1355 if (!shader_record_register_usage(shader, reg_maps, &reg,
1356 shader_version.type, constf_size))
1357 return WINED3DERR_INVALIDCALL;
1358 --count;
1359 }
1360
1361 if (color0_mov)
1362 {
1363 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1364 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1365 {
1366 shader->u.ps.color0_mov = TRUE;
1367 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1368 }
1369 }
1370 }
1371 }
1372 }
1373 reg_maps->loop_depth = max_loop_depth;
1374
1375 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1376 * R0 is written to the render target. */
1377 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1378 reg_maps->rt_mask |= (1u << 0);
1379
1380 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1381
1382 if (input_signature->elements)
1383 {
1384 for (i = 0; i < input_signature->element_count; ++i)
1385 {
1386 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1387 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1388 {
1389 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1390 reg_maps->vpos = 1;
1391 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1392 reg_maps->usesfacing = 1;
1393 }
1394 }
1395 }
1396 else if (!input_signature->elements && reg_maps->input_registers)
1397 {
1398 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1399 struct wined3d_shader_signature_element *e;
1400 unsigned int i;
1401
1402 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1403 return E_OUTOFMEMORY;
1404 input_signature->element_count = count;
1405
1406 e = input_signature->elements;
1407 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1408 {
1409 if (!(reg_maps->input_registers & (1u << i)))
1410 continue;
1411 input_signature_elements[i].register_idx = i;
1412 *e++ = input_signature_elements[i];
1413 }
1414 }
1415
1416 if (output_signature->elements)
1417 {
1418 for (i = 0; i < output_signature->element_count; ++i)
1419 {
1420 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1421 }
1422 }
1423 else if (reg_maps->output_registers)
1424 {
1425 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1426 struct wined3d_shader_signature_element *e;
1427
1428 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1429 return E_OUTOFMEMORY;
1430 output_signature->element_count = count;
1431
1432 e = output_signature->elements;
1433 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1434 {
1435 if (!(reg_maps->output_registers & (1u << i)))
1436 continue;
1437 *e++ = output_signature_elements[i];
1438 }
1439 }
1440
1441 return WINED3D_OK;
1442 }
1443
1444 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1445 {
1446 struct wined3d_shader_indexable_temp *reg, *reg_next;
1447
1448 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1449 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1450
1451 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1452 HeapFree(GetProcessHeap(), 0, reg);
1453 list_init(&reg_maps->indexable_temps);
1454 }
1455
1456 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1457 {
1458 DWORD map = 1u << max;
1459 map |= map - 1;
1460 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1461
1462 return wined3d_log2i(map);
1463 }
1464
1465 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1466 {
1467 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1468 {
1469 shader_addline(buffer, "refactoringAllowed");
1470 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1471 if (global_flags)
1472 shader_addline(buffer, " | ");
1473 }
1474
1475 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1476 {
1477 shader_addline(buffer, "enableRawAndStructuredBuffers");
1478 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1479 }
1480
1481 if (global_flags)
1482 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1483 }
1484
1485 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1486 {
1487 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1488 {
1489 shader_addline(buffer, "_g");
1490 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1491 }
1492 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1493 {
1494 shader_addline(buffer, "_t");
1495 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1496 }
1497
1498 if (sync_flags)
1499 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1500 }
1501
1502 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1503 {
1504 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1505 {
1506 shader_addline(buffer, "_glc");
1507 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1508 }
1509 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1510 {
1511 shader_addline(buffer, "_opc");
1512 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1513 }
1514
1515 if (uav_flags)
1516 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1517 }
1518
1519 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1520 enum wined3d_tessellator_domain domain)
1521 {
1522 switch (domain)
1523 {
1524 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1525 shader_addline(buffer, "line");
1526 break;
1527 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1528 shader_addline(buffer, "triangle");
1529 break;
1530 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1531 shader_addline(buffer, "quad");
1532 break;
1533 default:
1534 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1535 break;
1536 }
1537 }
1538
1539 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1540 enum wined3d_tessellator_output_primitive output_primitive)
1541 {
1542 switch (output_primitive)
1543 {
1544 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1545 shader_addline(buffer, "point");
1546 break;
1547 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1548 shader_addline(buffer, "line");
1549 break;
1550 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1551 shader_addline(buffer, "triangle_cw");
1552 break;
1553 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1554 shader_addline(buffer, "triangle_ccw");
1555 break;
1556 default:
1557 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1558 break;
1559 }
1560 }
1561
1562 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1563 enum wined3d_tessellator_partitioning partitioning)
1564 {
1565 switch (partitioning)
1566 {
1567 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1568 shader_addline(buffer, "integer");
1569 break;
1570 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1571 shader_addline(buffer, "pow2");
1572 break;
1573 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1574 shader_addline(buffer, "fractional_odd");
1575 break;
1576 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1577 shader_addline(buffer, "fractional_even");
1578 break;
1579 default:
1580 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1581 break;
1582 }
1583 }
1584
1585 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1586 {
1587 unsigned int i;
1588
1589 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1590 {
1591 if (sysval_semantic_names[i].sysval_semantic == semantic)
1592 {
1593 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1594 return;
1595 }
1596 }
1597
1598 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1599 }
1600
1601 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1602 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1603 {
1604 shader_addline(buffer, "dcl");
1605
1606 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1607 {
1608 switch (semantic->resource_type)
1609 {
1610 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1611 shader_addline(buffer, "_2d");
1612 break;
1613
1614 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1615 shader_addline(buffer, "_3d");
1616 break;
1617
1618 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1619 shader_addline(buffer, "_cube");
1620 break;
1621
1622 default:
1623 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1624 break;
1625 }
1626 }
1627 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1628 {
1629 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1630 shader_addline(buffer, "_resource_");
1631 else
1632 shader_addline(buffer, "_uav_");
1633 switch (semantic->resource_type)
1634 {
1635 case WINED3D_SHADER_RESOURCE_BUFFER:
1636 shader_addline(buffer, "buffer");
1637 break;
1638
1639 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1640 shader_addline(buffer, "texture1d");
1641 break;
1642
1643 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1644 shader_addline(buffer, "texture2d");
1645 break;
1646
1647 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1648 shader_addline(buffer, "texture2dms");
1649 break;
1650
1651 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1652 shader_addline(buffer, "texture3d");
1653 break;
1654
1655 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1656 shader_addline(buffer, "texturecube");
1657 break;
1658
1659 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1660 shader_addline(buffer, "texture1darray");
1661 break;
1662
1663 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1664 shader_addline(buffer, "texture2darray");
1665 break;
1666
1667 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1668 shader_addline(buffer, "texture2dmsarray");
1669 break;
1670
1671 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
1672 shader_addline(buffer, "texturecubearray");
1673 break;
1674
1675 default:
1676 shader_addline(buffer, "unknown");
1677 break;
1678 }
1679 switch (semantic->resource_data_type)
1680 {
1681 case WINED3D_DATA_FLOAT:
1682 shader_addline(buffer, " (float)");
1683 break;
1684
1685 case WINED3D_DATA_INT:
1686 shader_addline(buffer, " (int)");
1687 break;
1688
1689 case WINED3D_DATA_UINT:
1690 shader_addline(buffer, " (uint)");
1691 break;
1692
1693 case WINED3D_DATA_UNORM:
1694 shader_addline(buffer, " (unorm)");
1695 break;
1696
1697 case WINED3D_DATA_SNORM:
1698 shader_addline(buffer, " (snorm)");
1699 break;
1700
1701 default:
1702 shader_addline(buffer, " (unknown)");
1703 break;
1704 }
1705 }
1706 else
1707 {
1708 /* Pixel shaders 3.0 don't have usage semantics. */
1709 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1710 return;
1711 else
1712 shader_addline(buffer, "_");
1713
1714 switch (semantic->usage)
1715 {
1716 case WINED3D_DECL_USAGE_POSITION:
1717 shader_addline(buffer, "position%u", semantic->usage_idx);
1718 break;
1719
1720 case WINED3D_DECL_USAGE_BLEND_INDICES:
1721 shader_addline(buffer, "blend");
1722 break;
1723
1724 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1725 shader_addline(buffer, "weight");
1726 break;
1727
1728 case WINED3D_DECL_USAGE_NORMAL:
1729 shader_addline(buffer, "normal%u", semantic->usage_idx);
1730 break;
1731
1732 case WINED3D_DECL_USAGE_PSIZE:
1733 shader_addline(buffer, "psize");
1734 break;
1735
1736 case WINED3D_DECL_USAGE_COLOR:
1737 if (!semantic->usage_idx)
1738 shader_addline(buffer, "color");
1739 else
1740 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1741 break;
1742
1743 case WINED3D_DECL_USAGE_TEXCOORD:
1744 shader_addline(buffer, "texture%u", semantic->usage_idx);
1745 break;
1746
1747 case WINED3D_DECL_USAGE_TANGENT:
1748 shader_addline(buffer, "tangent");
1749 break;
1750
1751 case WINED3D_DECL_USAGE_BINORMAL:
1752 shader_addline(buffer, "binormal");
1753 break;
1754
1755 case WINED3D_DECL_USAGE_TESS_FACTOR:
1756 shader_addline(buffer, "tessfactor");
1757 break;
1758
1759 case WINED3D_DECL_USAGE_POSITIONT:
1760 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1761 break;
1762
1763 case WINED3D_DECL_USAGE_FOG:
1764 shader_addline(buffer, "fog");
1765 break;
1766
1767 case WINED3D_DECL_USAGE_DEPTH:
1768 shader_addline(buffer, "depth");
1769 break;
1770
1771 case WINED3D_DECL_USAGE_SAMPLE:
1772 shader_addline(buffer, "sample");
1773 break;
1774
1775 default:
1776 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1777 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
1778 }
1779 }
1780 }
1781
1782 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1783 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1784 {
1785 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1786 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1787 UINT offset = reg->idx[0].offset;
1788
1789 switch (reg->type)
1790 {
1791 case WINED3DSPR_TEMP:
1792 shader_addline(buffer, "r");
1793 break;
1794
1795 case WINED3DSPR_INPUT:
1796 shader_addline(buffer, "v");
1797 break;
1798
1799 case WINED3DSPR_CONST:
1800 case WINED3DSPR_CONST2:
1801 case WINED3DSPR_CONST3:
1802 case WINED3DSPR_CONST4:
1803 shader_addline(buffer, "c");
1804 offset = shader_get_float_offset(reg->type, offset);
1805 break;
1806
1807 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1808 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1809 break;
1810
1811 case WINED3DSPR_RASTOUT:
1812 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1813 break;
1814
1815 case WINED3DSPR_COLOROUT:
1816 shader_addline(buffer, "oC");
1817 break;
1818
1819 case WINED3DSPR_DEPTHOUT:
1820 shader_addline(buffer, "oDepth");
1821 break;
1822
1823 case WINED3DSPR_ATTROUT:
1824 shader_addline(buffer, "oD");
1825 break;
1826
1827 case WINED3DSPR_TEXCRDOUT:
1828 /* Vertex shaders >= 3.0 use general purpose output registers
1829 * (WINED3DSPR_OUTPUT), which can include an address token. */
1830 if (shader_version->major >= 3)
1831 shader_addline(buffer, "o");
1832 else
1833 shader_addline(buffer, "oT");
1834 break;
1835
1836 case WINED3DSPR_CONSTINT:
1837 shader_addline(buffer, "i");
1838 break;
1839
1840 case WINED3DSPR_CONSTBOOL:
1841 shader_addline(buffer, "b");
1842 break;
1843
1844 case WINED3DSPR_LABEL:
1845 shader_addline(buffer, "l");
1846 break;
1847
1848 case WINED3DSPR_LOOP:
1849 shader_addline(buffer, "aL");
1850 break;
1851
1852 case WINED3DSPR_SAMPLER:
1853 shader_addline(buffer, "s");
1854 break;
1855
1856 case WINED3DSPR_MISCTYPE:
1857 if (offset > 1)
1858 {
1859 FIXME("Unhandled misctype register %u.\n", offset);
1860 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1861 }
1862 else
1863 {
1864 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1865 }
1866 break;
1867
1868 case WINED3DSPR_PREDICATE:
1869 shader_addline(buffer, "p");
1870 break;
1871
1872 case WINED3DSPR_IMMCONST:
1873 shader_addline(buffer, "l");
1874 break;
1875
1876 case WINED3DSPR_CONSTBUFFER:
1877 shader_addline(buffer, "cb");
1878 break;
1879
1880 case WINED3DSPR_IMMCONSTBUFFER:
1881 shader_addline(buffer, "icb");
1882 break;
1883
1884 case WINED3DSPR_PRIMID:
1885 shader_addline(buffer, "primID");
1886 break;
1887
1888 case WINED3DSPR_NULL:
1889 shader_addline(buffer, "null");
1890 break;
1891
1892 case WINED3DSPR_RESOURCE:
1893 shader_addline(buffer, "t");
1894 break;
1895
1896 case WINED3DSPR_UAV:
1897 shader_addline(buffer, "u");
1898 break;
1899
1900 case WINED3DSPR_OUTPOINTID:
1901 shader_addline(buffer, "vOutputControlPointID");
1902 break;
1903
1904 case WINED3DSPR_FORKINSTID:
1905 shader_addline(buffer, "vForkInstanceId");
1906 break;
1907
1908 case WINED3DSPR_INCONTROLPOINT:
1909 shader_addline(buffer, "vicp");
1910 break;
1911
1912 case WINED3DSPR_OUTCONTROLPOINT:
1913 shader_addline(buffer, "vocp");
1914 break;
1915
1916 case WINED3DSPR_PATCHCONST:
1917 shader_addline(buffer, "vpc");
1918 break;
1919
1920 case WINED3DSPR_TESSCOORD:
1921 shader_addline(buffer, "vDomainLocation");
1922 break;
1923
1924 case WINED3DSPR_GROUPSHAREDMEM:
1925 shader_addline(buffer, "g");
1926 break;
1927
1928 case WINED3DSPR_THREADID:
1929 shader_addline(buffer, "vThreadID");
1930 break;
1931
1932 case WINED3DSPR_THREADGROUPID:
1933 shader_addline(buffer, "vThreadGroupID");
1934 break;
1935
1936 case WINED3DSPR_LOCALTHREADID:
1937 shader_addline(buffer, "vThreadIDInGroup");
1938 break;
1939
1940 case WINED3DSPR_LOCALTHREADINDEX:
1941 shader_addline(buffer, "vThreadIDInGroupFlattened");
1942 break;
1943
1944 case WINED3DSPR_IDXTEMP:
1945 shader_addline(buffer, "x");
1946 break;
1947
1948 case WINED3DSPR_STREAM:
1949 shader_addline(buffer, "m");
1950 break;
1951
1952 case WINED3DSPR_FUNCTIONBODY:
1953 shader_addline(buffer, "fb");
1954 break;
1955
1956 case WINED3DSPR_FUNCTIONPOINTER:
1957 shader_addline(buffer, "fp");
1958 break;
1959
1960 case WINED3DSPR_COVERAGE:
1961 shader_addline(buffer, "vCoverage");
1962 break;
1963
1964 case WINED3DSPR_SAMPLEMASK:
1965 shader_addline(buffer, "oMask");
1966 break;
1967
1968 default:
1969 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1970 break;
1971 }
1972
1973 if (reg->type == WINED3DSPR_IMMCONST)
1974 {
1975 shader_addline(buffer, "(");
1976 switch (reg->immconst_type)
1977 {
1978 case WINED3D_IMMCONST_SCALAR:
1979 switch (reg->data_type)
1980 {
1981 case WINED3D_DATA_FLOAT:
1982 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
1983 break;
1984 case WINED3D_DATA_INT:
1985 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
1986 break;
1987 case WINED3D_DATA_RESOURCE:
1988 case WINED3D_DATA_SAMPLER:
1989 case WINED3D_DATA_UINT:
1990 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
1991 break;
1992 default:
1993 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1994 break;
1995 }
1996 break;
1997
1998 case WINED3D_IMMCONST_VEC4:
1999 switch (reg->data_type)
2000 {
2001 case WINED3D_DATA_FLOAT:
2002 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2003 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2004 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2005 break;
2006 case WINED3D_DATA_INT:
2007 shader_addline(buffer, "%d, %d, %d, %d",
2008 reg->u.immconst_data[0], reg->u.immconst_data[1],
2009 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2010 break;
2011 case WINED3D_DATA_RESOURCE:
2012 case WINED3D_DATA_SAMPLER:
2013 case WINED3D_DATA_UINT:
2014 shader_addline(buffer, "%u, %u, %u, %u",
2015 reg->u.immconst_data[0], reg->u.immconst_data[1],
2016 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2017 break;
2018 default:
2019 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2020 break;
2021 }
2022 break;
2023
2024 default:
2025 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2026 break;
2027 }
2028 shader_addline(buffer, ")");
2029 }
2030 else if (reg->type != WINED3DSPR_RASTOUT
2031 && reg->type != WINED3DSPR_MISCTYPE
2032 && reg->type != WINED3DSPR_NULL)
2033 {
2034 if (offset != ~0u)
2035 {
2036 shader_addline(buffer, "[");
2037 if (reg->idx[0].rel_addr)
2038 {
2039 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2040 shader_addline(buffer, " + ");
2041 }
2042 shader_addline(buffer, "%u]", offset);
2043
2044 if (reg->idx[1].offset != ~0u)
2045 {
2046 shader_addline(buffer, "[");
2047 if (reg->idx[1].rel_addr)
2048 {
2049 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2050 shader_addline(buffer, " + ");
2051 }
2052 shader_addline(buffer, "%u]", reg->idx[1].offset);
2053 }
2054 }
2055
2056 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2057 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2058 }
2059 }
2060
2061 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2062 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2063 {
2064 DWORD write_mask = param->write_mask;
2065
2066 shader_dump_register(buffer, &param->reg, shader_version);
2067
2068 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2069 {
2070 static const char write_mask_chars[] = "xyzw";
2071
2072 shader_addline(buffer, ".");
2073 if (write_mask & WINED3DSP_WRITEMASK_0)
2074 shader_addline(buffer, "%c", write_mask_chars[0]);
2075 if (write_mask & WINED3DSP_WRITEMASK_1)
2076 shader_addline(buffer, "%c", write_mask_chars[1]);
2077 if (write_mask & WINED3DSP_WRITEMASK_2)
2078 shader_addline(buffer, "%c", write_mask_chars[2]);
2079 if (write_mask & WINED3DSP_WRITEMASK_3)
2080 shader_addline(buffer, "%c", write_mask_chars[3]);
2081 }
2082 }
2083
2084 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2085 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2086 {
2087 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2088 DWORD swizzle = param->swizzle;
2089
2090 if (src_modifier == WINED3DSPSM_NEG
2091 || src_modifier == WINED3DSPSM_BIASNEG
2092 || src_modifier == WINED3DSPSM_SIGNNEG
2093 || src_modifier == WINED3DSPSM_X2NEG
2094 || src_modifier == WINED3DSPSM_ABSNEG)
2095 shader_addline(buffer, "-");
2096 else if (src_modifier == WINED3DSPSM_COMP)
2097 shader_addline(buffer, "1-");
2098 else if (src_modifier == WINED3DSPSM_NOT)
2099 shader_addline(buffer, "!");
2100
2101 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2102 shader_addline(buffer, "abs(");
2103
2104 shader_dump_register(buffer, &param->reg, shader_version);
2105
2106 switch (src_modifier)
2107 {
2108 case WINED3DSPSM_NONE: break;
2109 case WINED3DSPSM_NEG: break;
2110 case WINED3DSPSM_NOT: break;
2111 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2112 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2113 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2114 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2115 case WINED3DSPSM_COMP: break;
2116 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2117 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2118 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2119 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2120 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2121 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2122 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2123 }
2124
2125 if (swizzle != WINED3DSP_NOSWIZZLE)
2126 {
2127 static const char swizzle_chars[] = "xyzw";
2128 DWORD swizzle_x = swizzle & 0x03;
2129 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2130 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2131 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2132
2133 if (swizzle_x == swizzle_y
2134 && swizzle_x == swizzle_z
2135 && swizzle_x == swizzle_w)
2136 {
2137 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2138 }
2139 else
2140 {
2141 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2142 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2143 }
2144 }
2145 }
2146
2147 /* Shared code in order to generate the bulk of the shader string.
2148 * NOTE: A description of how to parse tokens can be found on MSDN. */
2149 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2150 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
2151 {
2152 struct wined3d_device *device = shader->device;
2153 const struct wined3d_shader_frontend *fe = shader->frontend;
2154 void *fe_data = shader->frontend_data;
2155 struct wined3d_shader_version shader_version;
2156 struct wined3d_shader_parser_state state;
2157 struct wined3d_shader_instruction ins;
2158 struct wined3d_shader_tex_mx tex_mx;
2159 struct wined3d_shader_context ctx;
2160 const DWORD *ptr = byte_code;
2161
2162 /* Initialize current parsing state. */
2163 tex_mx.current_row = 0;
2164 state.current_loop_depth = 0;
2165 state.current_loop_reg = 0;
2166 state.in_subroutine = FALSE;
2167
2168 ctx.shader = shader;
2169 ctx.gl_info = &device->adapter->gl_info;
2170 ctx.reg_maps = reg_maps;
2171 ctx.buffer = buffer;
2172 ctx.tex_mx = &tex_mx;
2173 ctx.state = &state;
2174 ctx.backend_data = backend_ctx;
2175 ins.ctx = &ctx;
2176
2177 fe->shader_read_header(fe_data, &ptr, &shader_version);
2178
2179 while (!fe->shader_is_end(fe_data, &ptr))
2180 {
2181 /* Read opcode. */
2182 fe->shader_read_instruction(fe_data, &ptr, &ins);
2183
2184 /* Unknown opcode and its parameters. */
2185 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2186 {
2187 TRACE("Skipping unrecognized instruction.\n");
2188 continue;
2189 }
2190
2191 if (ins.predicate)
2192 FIXME("Predicates not implemented.\n");
2193
2194 /* Call appropriate function for output target */
2195 device->shader_backend->shader_handle_instruction(&ins);
2196 }
2197 }
2198
2199 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2200 const struct wined3d_shader_dst_param *dst)
2201 {
2202 DWORD mmask = dst->modifiers;
2203
2204 switch (dst->shift)
2205 {
2206 case 0: break;
2207 case 13: shader_addline(buffer, "_d8"); break;
2208 case 14: shader_addline(buffer, "_d4"); break;
2209 case 15: shader_addline(buffer, "_d2"); break;
2210 case 1: shader_addline(buffer, "_x2"); break;
2211 case 2: shader_addline(buffer, "_x4"); break;
2212 case 3: shader_addline(buffer, "_x8"); break;
2213 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2214 }
2215
2216 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2217 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2218 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2219
2220 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2221 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2222 }
2223
2224 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2225 enum wined3d_primitive_type primitive_type)
2226 {
2227 switch (primitive_type)
2228 {
2229 case WINED3D_PT_UNDEFINED:
2230 shader_addline(buffer, "undefined");
2231 break;
2232 case WINED3D_PT_POINTLIST:
2233 shader_addline(buffer, "pointlist");
2234 break;
2235 case WINED3D_PT_LINELIST:
2236 shader_addline(buffer, "linelist");
2237 break;
2238 case WINED3D_PT_LINESTRIP:
2239 shader_addline(buffer, "linestrip");
2240 break;
2241 case WINED3D_PT_TRIANGLELIST:
2242 shader_addline(buffer, "trianglelist");
2243 break;
2244 case WINED3D_PT_TRIANGLESTRIP:
2245 shader_addline(buffer, "trianglestrip");
2246 break;
2247 case WINED3D_PT_TRIANGLEFAN:
2248 shader_addline(buffer, "trianglefan");
2249 break;
2250 case WINED3D_PT_LINELIST_ADJ:
2251 shader_addline(buffer, "linelist_adj");
2252 break;
2253 case WINED3D_PT_LINESTRIP_ADJ:
2254 shader_addline(buffer, "linestrip_adj");
2255 break;
2256 case WINED3D_PT_TRIANGLELIST_ADJ:
2257 shader_addline(buffer, "trianglelist_adj");
2258 break;
2259 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2260 shader_addline(buffer, "trianglestrip_adj");
2261 break;
2262 default:
2263 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2264 break;
2265 }
2266 }
2267
2268 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2269 enum wined3d_shader_interpolation_mode interpolation_mode)
2270 {
2271 switch (interpolation_mode)
2272 {
2273 case WINED3DSIM_CONSTANT:
2274 shader_addline(buffer, "constant");
2275 break;
2276 case WINED3DSIM_LINEAR:
2277 shader_addline(buffer, "linear");
2278 break;
2279 case WINED3DSIM_LINEAR_CENTROID:
2280 shader_addline(buffer, "linear centroid");
2281 break;
2282 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2283 shader_addline(buffer, "linear noperspective");
2284 break;
2285 case WINED3DSIM_LINEAR_SAMPLE:
2286 shader_addline(buffer, "linear sample");
2287 break;
2288 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2289 shader_addline(buffer, "linear noperspective centroid");
2290 break;
2291 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2292 shader_addline(buffer, "linear noperspective sample");
2293 break;
2294 default:
2295 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2296 break;
2297 }
2298 }
2299
2300 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2301 {
2302 struct wined3d_shader_version shader_version;
2303 struct wined3d_string_buffer buffer;
2304 const DWORD *ptr = byte_code;
2305 const char *type_prefix;
2306 const char *p, *q;
2307 DWORD i;
2308
2309 if (!string_buffer_init(&buffer))
2310 {
2311 ERR("Failed to initialize string buffer.\n");
2312 return;
2313 }
2314
2315 TRACE("Parsing %p.\n", byte_code);
2316
2317 fe->shader_read_header(fe_data, &ptr, &shader_version);
2318
2319 switch (shader_version.type)
2320 {
2321 case WINED3D_SHADER_TYPE_VERTEX:
2322 type_prefix = "vs";
2323 break;
2324
2325 case WINED3D_SHADER_TYPE_HULL:
2326 type_prefix = "hs";
2327 break;
2328
2329 case WINED3D_SHADER_TYPE_DOMAIN:
2330 type_prefix = "ds";
2331 break;
2332
2333 case WINED3D_SHADER_TYPE_GEOMETRY:
2334 type_prefix = "gs";
2335 break;
2336
2337 case WINED3D_SHADER_TYPE_PIXEL:
2338 type_prefix = "ps";
2339 break;
2340
2341 case WINED3D_SHADER_TYPE_COMPUTE:
2342 type_prefix = "cs";
2343 break;
2344
2345 default:
2346 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2347 type_prefix = "unknown";
2348 break;
2349 }
2350
2351 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2352
2353 while (!fe->shader_is_end(fe_data, &ptr))
2354 {
2355 struct wined3d_shader_instruction ins;
2356
2357 fe->shader_read_instruction(fe_data, &ptr, &ins);
2358 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2359 {
2360 WARN("Skipping unrecognized instruction.\n");
2361 shader_addline(&buffer, "<unrecognized instruction>\n");
2362 continue;
2363 }
2364
2365 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2366 {
2367 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2368 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2369 shader_addline(&buffer, " ");
2370 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2371 }
2372 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2373 {
2374 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2375 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2376 shader_addline(&buffer, ", %s",
2377 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2378 }
2379 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2380 {
2381 shader_addline(&buffer, "%s fb%u",
2382 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2383 }
2384 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2385 {
2386 shader_addline(&buffer, "%s ft%u = {...}",
2387 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2388 }
2389 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2390 {
2391 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2392 shader_dump_global_flags(&buffer, ins.flags);
2393 }
2394 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2395 {
2396 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2397 ins.declaration.max_tessellation_factor);
2398 }
2399 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2400 {
2401 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2402 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2403 {
2404 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2405 ins.declaration.icb->data[4 * i + 0],
2406 ins.declaration.icb->data[4 * i + 1],
2407 ins.declaration.icb->data[4 * i + 2],
2408 ins.declaration.icb->data[4 * i + 3]);
2409 }
2410 shader_addline(&buffer, "}");
2411 }
2412 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2413 {
2414 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2415 ins.declaration.indexable_temp.register_idx,
2416 ins.declaration.indexable_temp.register_size,
2417 ins.declaration.indexable_temp.component_count);
2418 }
2419 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2420 {
2421 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2422 shader_dump_interpolation_mode(&buffer, ins.flags);
2423 shader_addline(&buffer, " ");
2424 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2425 }
2426 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2427 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2428 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2429 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2430 {
2431 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2432 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2433 shader_addline(&buffer, ", ");
2434 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2435 }
2436 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2437 {
2438 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2439 shader_dump_interpolation_mode(&buffer, ins.flags);
2440 shader_addline(&buffer, " ");
2441 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2442 shader_addline(&buffer, ", ");
2443 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2444 }
2445 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2446 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2447 {
2448 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2449 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2450 }
2451 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2452 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2453 {
2454 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2455 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2456 }
2457 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2458 {
2459 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2460 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2461 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2462 }
2463 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2464 {
2465 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2466 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2467 }
2468 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2469 {
2470 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2471 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2472 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2473 }
2474 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2475 {
2476 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2477 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2478 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2479 shader_addline(&buffer, ", comparisonMode");
2480 }
2481 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2482 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2483 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2484 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2485 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2486 {
2487 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2488 }
2489 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2490 {
2491 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2492 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2493 }
2494 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2495 {
2496 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2497 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2498 }
2499 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2500 {
2501 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2502 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2503 }
2504 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2505 {
2506 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2507 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2508 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2509 }
2510 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2511 {
2512 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2513 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2514 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2515 ins.declaration.tgsm_structured.structure_count);
2516 }
2517 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2518 {
2519 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2520 ins.declaration.thread_group_size.x,
2521 ins.declaration.thread_group_size.y,
2522 ins.declaration.thread_group_size.z);
2523 }
2524 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2525 {
2526 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2527 shader_dump_uav_flags(&buffer, ins.flags);
2528 shader_addline(&buffer, " ");
2529 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2530 }
2531 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2532 {
2533 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2534 shader_dump_uav_flags(&buffer, ins.flags);
2535 shader_addline(&buffer, " ");
2536 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2537 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2538 }
2539 else if (ins.handler_idx == WINED3DSIH_DEF)
2540 {
2541 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2542 ins.dst[0].reg.idx[0].offset),
2543 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2544 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2545 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2546 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2547 }
2548 else if (ins.handler_idx == WINED3DSIH_DEFI)
2549 {
2550 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2551 ins.src[0].reg.u.immconst_data[0],
2552 ins.src[0].reg.u.immconst_data[1],
2553 ins.src[0].reg.u.immconst_data[2],
2554 ins.src[0].reg.u.immconst_data[3]);
2555 }
2556 else if (ins.handler_idx == WINED3DSIH_DEFB)
2557 {
2558 shader_addline(&buffer, "defb b%u = %s",
2559 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2560 }
2561 else
2562 {
2563 if (ins.predicate)
2564 {
2565 shader_addline(&buffer, "(");
2566 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2567 shader_addline(&buffer, ") ");
2568 }
2569
2570 /* PixWin marks instructions with the coissue flag with a '+' */
2571 if (ins.coissue)
2572 shader_addline(&buffer, "+");
2573
2574 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2575
2576 if (ins.handler_idx == WINED3DSIH_BREAKP
2577 || ins.handler_idx == WINED3DSIH_IF)
2578 {
2579 switch (ins.flags)
2580 {
2581 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2582 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2583 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2584 }
2585 }
2586 else if (ins.handler_idx == WINED3DSIH_IFC
2587 || ins.handler_idx == WINED3DSIH_BREAKC)
2588 {
2589 switch (ins.flags)
2590 {
2591 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2592 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2593 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2594 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2595 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2596 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2597 default: shader_addline(&buffer, "_(%u)", ins.flags);
2598 }
2599 }
2600 else if (ins.handler_idx == WINED3DSIH_TEX
2601 && shader_version.major >= 2
2602 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2603 {
2604 shader_addline(&buffer, "p");
2605 }
2606 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2607 {
2608 switch (ins.flags)
2609 {
2610 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2611 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2612 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2613 }
2614 }
2615 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2616 {
2617 switch (ins.flags)
2618 {
2619 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2620 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2621 }
2622 }
2623 else if (ins.handler_idx == WINED3DSIH_SYNC)
2624 {
2625 shader_dump_sync_flags(&buffer, ins.flags);
2626 }
2627
2628 if (wined3d_shader_instruction_has_texel_offset(&ins))
2629 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2630
2631 for (i = 0; i < ins.dst_count; ++i)
2632 {
2633 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2634 shader_addline(&buffer, !i ? " " : ", ");
2635 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2636 }
2637
2638 /* Other source tokens */
2639 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2640 {
2641 shader_addline(&buffer, !i ? " " : ", ");
2642 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2643 }
2644 }
2645 shader_addline(&buffer, "\n");
2646 }
2647
2648 for (p = buffer.buffer; *p; p = q)
2649 {
2650 if (!(q = strstr(p, "\n")))
2651 q = p + strlen(p);
2652 else
2653 ++q;
2654 TRACE(" %.*s", (int)(q - p), p);
2655 }
2656
2657 string_buffer_free(&buffer);
2658 }
2659
2660 static void shader_cleanup(struct wined3d_shader *shader)
2661 {
2662 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2663 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2664 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2665 shader->device->shader_backend->shader_destroy(shader);
2666 shader_cleanup_reg_maps(&shader->reg_maps);
2667 HeapFree(GetProcessHeap(), 0, shader->function);
2668 shader_delete_constant_list(&shader->constantsF);
2669 shader_delete_constant_list(&shader->constantsB);
2670 shader_delete_constant_list(&shader->constantsI);
2671 list_remove(&shader->shader_list_entry);
2672
2673 if (shader->frontend && shader->frontend_data)
2674 shader->frontend->shader_free(shader->frontend_data);
2675 }
2676
2677 struct shader_none_priv
2678 {
2679 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2680 const struct fragment_pipeline *fragment_pipe;
2681 BOOL ffp_proj_control;
2682 };
2683
2684 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2685 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
2686 const struct wined3d_state *state) {}
2687 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2688 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2689 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2690 const struct wined3d_state *state) {}
2691 static void shader_none_destroy(struct wined3d_shader *shader) {}
2692 static void shader_none_free_context_data(struct wined3d_context *context) {}
2693 static void shader_none_init_context_state(struct wined3d_context *context) {}
2694
2695 /* Context activation is done by the caller. */
2696 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2697 const struct wined3d_state *state)
2698 {
2699 const struct wined3d_gl_info *gl_info = context->gl_info;
2700 struct shader_none_priv *priv = shader_priv;
2701
2702 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2703 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2704 }
2705
2706 /* Context activation is done by the caller. */
2707 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2708 {
2709 struct shader_none_priv *priv = shader_priv;
2710 const struct wined3d_gl_info *gl_info = context->gl_info;
2711
2712 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2713 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2714
2715 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2716 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2717 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2718 | (1u << WINED3D_SHADER_TYPE_HULL)
2719 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2720 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2721 }
2722
2723 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2724 const struct fragment_pipeline *fragment_pipe)
2725 {
2726 struct fragment_caps fragment_caps;
2727 void *vertex_priv, *fragment_priv;
2728 struct shader_none_priv *priv;
2729
2730 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2731 return E_OUTOFMEMORY;
2732
2733 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2734 {
2735 ERR("Failed to initialize vertex pipe.\n");
2736 HeapFree(GetProcessHeap(), 0, priv);
2737 return E_FAIL;
2738 }
2739
2740 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2741 {
2742 ERR("Failed to initialize fragment pipe.\n");
2743 vertex_pipe->vp_free(device);
2744 HeapFree(GetProcessHeap(), 0, priv);
2745 return E_FAIL;
2746 }
2747
2748 priv->vertex_pipe = vertex_pipe;
2749 priv->fragment_pipe = fragment_pipe;
2750 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2751 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2752
2753 device->vertex_priv = vertex_priv;
2754 device->fragment_priv = fragment_priv;
2755 device->shader_priv = priv;
2756
2757 return WINED3D_OK;
2758 }
2759
2760 static void shader_none_free(struct wined3d_device *device)
2761 {
2762 struct shader_none_priv *priv = device->shader_priv;
2763
2764 priv->fragment_pipe->free_private(device);
2765 priv->vertex_pipe->vp_free(device);
2766 HeapFree(GetProcessHeap(), 0, priv);
2767 }
2768
2769 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2770 {
2771 return TRUE;
2772 }
2773
2774 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2775 {
2776 /* Set the shader caps to 0 for the none shader backend */
2777 caps->vs_version = 0;
2778 caps->hs_version = 0;
2779 caps->ds_version = 0;
2780 caps->gs_version = 0;
2781 caps->ps_version = 0;
2782 caps->cs_version = 0;
2783 caps->vs_uniform_count = 0;
2784 caps->ps_uniform_count = 0;
2785 caps->ps_1x_max_value = 0.0f;
2786 caps->varying_count = 0;
2787 caps->wined3d_caps = 0;
2788 }
2789
2790 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2791 {
2792 /* We "support" every possible fixup, since we don't support any shader
2793 * model, and will never have to actually sample a texture. */
2794 return TRUE;
2795 }
2796
2797 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2798 {
2799 struct shader_none_priv *priv = shader_priv;
2800
2801 return priv->ffp_proj_control;
2802 }
2803
2804 const struct wined3d_shader_backend_ops none_shader_backend =
2805 {
2806 shader_none_handle_instruction,
2807 shader_none_select,
2808 shader_none_select_compute,
2809 shader_none_disable,
2810 shader_none_update_float_vertex_constants,
2811 shader_none_update_float_pixel_constants,
2812 shader_none_load_constants,
2813 shader_none_destroy,
2814 shader_none_alloc,
2815 shader_none_free,
2816 shader_none_allocate_context_data,
2817 shader_none_free_context_data,
2818 shader_none_init_context_state,
2819 shader_none_get_caps,
2820 shader_none_color_fixup_supported,
2821 shader_none_has_ffp_proj_control,
2822 };
2823
2824 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2825 DWORD float_const_count, enum wined3d_shader_type type, unsigned int max_version)
2826 {
2827 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2828 const struct wined3d_shader_frontend *fe;
2829 HRESULT hr;
2830 unsigned int backend_version;
2831 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2832
2833 TRACE("shader %p, byte_code %p, float_const_count %u, type %#x, max_version %u.\n",
2834 shader, byte_code, float_const_count, type, max_version);
2835
2836 list_init(&shader->constantsF);
2837 list_init(&shader->constantsB);
2838 list_init(&shader->constantsI);
2839 shader->lconst_inf_or_nan = FALSE;
2840 list_init(&reg_maps->indexable_temps);
2841
2842 fe = shader_select_frontend(*byte_code);
2843 if (!fe)
2844 {
2845 FIXME("Unable to find frontend for shader.\n");
2846 return WINED3DERR_INVALIDCALL;
2847 }
2848 shader->frontend = fe;
2849 shader->frontend_data = fe->shader_init(byte_code, &shader->output_signature);
2850 if (!shader->frontend_data)
2851 {
2852 FIXME("Failed to initialize frontend.\n");
2853 return WINED3DERR_INVALIDCALL;
2854 }
2855
2856 /* First pass: trace shader. */
2857 if (TRACE_ON(d3d_shader))
2858 shader_trace_init(fe, shader->frontend_data, byte_code);
2859
2860 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2861 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2862 &shader->output_signature, byte_code, float_const_count)))
2863 return hr;
2864
2865 if (reg_maps->shader_version.type != type)
2866 {
2867 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2868 return WINED3DERR_INVALIDCALL;
2869 }
2870 if (reg_maps->shader_version.major > max_version)
2871 {
2872 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2873 return WINED3DERR_INVALIDCALL;
2874 }
2875 switch (type)
2876 {
2877 case WINED3D_SHADER_TYPE_VERTEX:
2878 backend_version = d3d_info->limits.vs_version;
2879 break;
2880 case WINED3D_SHADER_TYPE_HULL:
2881 backend_version = d3d_info->limits.hs_version;
2882 break;
2883 case WINED3D_SHADER_TYPE_DOMAIN:
2884 backend_version = d3d_info->limits.ds_version;
2885 break;
2886 case WINED3D_SHADER_TYPE_GEOMETRY:
2887 backend_version = d3d_info->limits.gs_version;
2888 break;
2889 case WINED3D_SHADER_TYPE_PIXEL:
2890 backend_version = d3d_info->limits.ps_version;
2891 break;
2892 case WINED3D_SHADER_TYPE_COMPUTE:
2893 backend_version = d3d_info->limits.cs_version;
2894 break;
2895 default:
2896 FIXME("No backend version-checking for this shader type.\n");
2897 backend_version = 0;
2898 }
2899 if (reg_maps->shader_version.major > backend_version)
2900 {
2901 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2902 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2903 return WINED3DERR_INVALIDCALL;
2904 }
2905
2906 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength)))
2907 return E_OUTOFMEMORY;
2908 memcpy(shader->function, byte_code, shader->functionLength);
2909
2910 return WINED3D_OK;
2911 }
2912
2913 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2914 {
2915 ULONG refcount = InterlockedIncrement(&shader->ref);
2916
2917 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2918
2919 return refcount;
2920 }
2921
2922 static void wined3d_shader_destroy_object(void *object)
2923 {
2924 shader_cleanup(object);
2925 HeapFree(GetProcessHeap(), 0, object);
2926 }
2927
2928 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2929 {
2930 ULONG refcount = InterlockedDecrement(&shader->ref);
2931
2932 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2933
2934 if (!refcount)
2935 {
2936 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2937 wined3d_cs_emit_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2938 }
2939
2940 return refcount;
2941 }
2942
2943 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2944 {
2945 TRACE("shader %p.\n", shader);
2946
2947 return shader->parent;
2948 }
2949
2950 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2951 void *byte_code, UINT *byte_code_size)
2952 {
2953 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2954
2955 if (!byte_code)
2956 {
2957 *byte_code_size = shader->functionLength;
2958 return WINED3D_OK;
2959 }
2960
2961 if (*byte_code_size < shader->functionLength)
2962 {
2963 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2964 * than the required size we should write the required size and
2965 * return D3DERR_MOREDATA. That's not actually true. */
2966 return WINED3DERR_INVALIDCALL;
2967 }
2968
2969 memcpy(byte_code, shader->function, shader->functionLength);
2970
2971 return WINED3D_OK;
2972 }
2973
2974 /* Set local constants for d3d8 shaders. */
2975 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2976 UINT start_idx, const float *src_data, UINT count)
2977 {
2978 UINT end_idx = start_idx + count;
2979 UINT i;
2980
2981 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2982
2983 if (end_idx > shader->limits->constant_float)
2984 {
2985 WARN("end_idx %u > float constants limit %u.\n",
2986 end_idx, shader->limits->constant_float);
2987 end_idx = shader->limits->constant_float;
2988 }
2989
2990 for (i = start_idx; i < end_idx; ++i)
2991 {
2992 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2993 float *value;
2994 if (!lconst)
2995 return E_OUTOFMEMORY;
2996
2997 lconst->idx = i;
2998 value = (float *)lconst->value;
2999 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3000 list_add_head(&shader->constantsF, &lconst->entry);
3001
3002 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3003 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3004 {
3005 shader->lconst_inf_or_nan = TRUE;
3006 }
3007 }
3008
3009 return WINED3D_OK;
3010 }
3011
3012 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3013 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
3014 {
3015 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3016 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3017 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3018 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3019 args->point_size = state->gl_primitive_type == GL_POINTS;
3020 args->per_vertex_point_size = shader->reg_maps.point_size;
3021 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
3022 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3023 : WINED3D_SHADER_TYPE_PIXEL;
3024 if (shader->reg_maps.shader_version.major >= 4)
3025 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
3026 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
3027 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
3028 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
3029 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
3030 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3031 else
3032 args->next_shader_input_count = 0;
3033 args->swizzle_map = swizzle_map;
3034 if (d3d_info->emulated_flatshading)
3035 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3036 else
3037 args->flatshading = 0;
3038 }
3039
3040 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3041 {
3042 if (usage_idx1 != usage_idx2)
3043 return FALSE;
3044 if (usage1 == usage2)
3045 return TRUE;
3046 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3047 return TRUE;
3048 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3049 return TRUE;
3050
3051 return FALSE;
3052 }
3053
3054 BOOL vshader_get_input(const struct wined3d_shader *shader,
3055 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3056 {
3057 WORD map = shader->reg_maps.input_registers;
3058 unsigned int i;
3059
3060 for (i = 0; map; map >>= 1, ++i)
3061 {
3062 if (!(map & 1)) continue;
3063
3064 if (match_usage(shader->u.vs.attributes[i].usage,
3065 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3066 {
3067 *regnum = i;
3068 return TRUE;
3069 }
3070 }
3071 return FALSE;
3072 }
3073
3074 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3075 const struct wined3d_shader_signature *src, char **signature_strings)
3076 {
3077 struct wined3d_shader_signature_element *e;
3078 unsigned int i;
3079 SIZE_T len;
3080 char *ptr;
3081
3082 if (!src->element_count)
3083 return WINED3D_OK;
3084
3085 ptr = *signature_strings;
3086
3087 dst->element_count = src->element_count;
3088 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3089 return E_OUTOFMEMORY;
3090
3091 for (i = 0; i < src->element_count; ++i)
3092 {
3093 e = &src->elements[i];
3094 dst->elements[i] = *e;
3095
3096 len = strlen(e->semantic_name);
3097 memcpy(ptr, e->semantic_name, len + 1);
3098 dst->elements[i].semantic_name = ptr;
3099 ptr += len + 1;
3100 }
3101
3102 *signature_strings = ptr;
3103
3104 return WINED3D_OK;
3105 }
3106
3107 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3108 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3109 void *parent, const struct wined3d_parent_ops *parent_ops)
3110 {
3111 struct wined3d_shader_signature_element *e;
3112 SIZE_T total, len;
3113 unsigned int i;
3114 HRESULT hr;
3115 char *ptr;
3116
3117 if (!desc->byte_code)
3118 return WINED3DERR_INVALIDCALL;
3119
3120 shader->ref = 1;
3121 shader->device = device;
3122 shader->parent = parent;
3123 shader->parent_ops = parent_ops;
3124
3125 total = 0;
3126 for (i = 0; i < desc->input_signature.element_count; ++i)
3127 {
3128 e = &desc->input_signature.elements[i];
3129 len = strlen(e->semantic_name);
3130 if (len >= ~(SIZE_T)0 - total)
3131 return E_OUTOFMEMORY;
3132
3133 total += len + 1;
3134 }
3135 for (i = 0; i < desc->output_signature.element_count; ++i)
3136 {
3137 e = &desc->output_signature.elements[i];
3138 len = strlen(e->semantic_name);
3139 if (len >= ~(SIZE_T)0 - total)
3140 return E_OUTOFMEMORY;
3141
3142 total += len + 1;
3143 }
3144 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3145 return E_OUTOFMEMORY;
3146 ptr = shader->signature_strings;
3147
3148 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3149 {
3150 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3151 return hr;
3152 }
3153 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3154 {
3155 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3156 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3157 return hr;
3158 }
3159
3160 list_init(&shader->linked_programs);
3161 list_add_head(&device->shaders, &shader->shader_list_entry);
3162
3163 if (FAILED(hr = shader_set_function(shader, desc->byte_code,
3164 float_const_count, type, desc->max_version)))
3165 {
3166 WARN("Failed to set function, hr %#x.\n", hr);
3167 shader_cleanup(shader);
3168 }
3169
3170 shader->load_local_constsF = shader->lconst_inf_or_nan;
3171
3172 return hr;
3173 }
3174
3175 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3176 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3177 {
3178 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3179 unsigned int i;
3180 HRESULT hr;
3181
3182 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3183 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3184 return hr;
3185
3186 for (i = 0; i < shader->input_signature.element_count; ++i)
3187 {
3188 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3189
3190 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3191 continue;
3192
3193 shader->u.vs.attributes[input->register_idx].usage =
3194 shader_usage_from_semantic_name(input->semantic_name);
3195 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3196 }
3197
3198 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3199 shader->load_local_constsF = TRUE;
3200
3201 return WINED3D_OK;
3202 }
3203
3204 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3205 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3206 {
3207 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3208 }
3209
3210 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3211 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3212 {
3213 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3214 }
3215
3216 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3217 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3218 {
3219 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3220 }
3221
3222 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3223 struct gs_compile_args *args)
3224 {
3225 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3226 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3227 }
3228
3229 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3230 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3231 {
3232 const struct wined3d_gl_info *gl_info = context->gl_info;
3233 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3234 const struct wined3d_texture *texture;
3235 UINT i;
3236
3237 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3238 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3239 {
3240 static unsigned int warned = 0;
3241
3242 args->srgb_correction = 1;
3243 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3244 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3245 "support, expect rendering artifacts.\n");
3246 }
3247
3248 if (shader->reg_maps.shader_version.major == 1
3249 && shader->reg_maps.shader_version.minor <= 3)
3250 {
3251 for (i = 0; i < shader->limits->sampler; ++i)
3252 {
3253 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3254
3255 if (flags & WINED3D_TTFF_PROJECTED)
3256 {
3257 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3258
3259 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3260 {
3261 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3262 unsigned int j;
3263 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3264 DWORD max_valid = WINED3D_TTFF_COUNT4;
3265
3266 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3267 {
3268 struct wined3d_vertex_declaration_element *element =
3269 &state->vertex_declaration->elements[j];
3270
3271 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3272 && element->usage_idx == index)
3273 {
3274 max_valid = element->format->component_count;
3275 break;
3276 }
3277 }
3278 if (!tex_transform || tex_transform > max_valid)
3279 {
3280 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3281 tex_transform, max_valid);
3282 tex_transform = max_valid;
3283 }
3284 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3285 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3286 && tex_transform > WINED3D_TTFF_COUNT2)
3287 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3288 && tex_transform > WINED3D_TTFF_COUNT3))
3289 tex_transform |= WINED3D_PSARGS_PROJECTED;
3290 else
3291 {
3292 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3293 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3294 i, tex_transform, resource_type);
3295 }
3296 }
3297 else
3298 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3299
3300 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3301 }
3302 }
3303 }
3304 if (shader->reg_maps.shader_version.major == 1
3305 && shader->reg_maps.shader_version.minor <= 4)
3306 {
3307 for (i = 0; i < shader->limits->sampler; ++i)
3308 {
3309 const struct wined3d_texture *texture = state->textures[i];
3310
3311 if (!shader->reg_maps.resource_info[i].type)
3312 continue;
3313
3314 /* Treat unbound textures as 2D. The dummy texture will provide
3315 * the proper sample value. The tex_types bitmap defaults to
3316 * 2D because of the memset. */
3317 if (!texture)
3318 continue;
3319
3320 switch (texture->target)
3321 {
3322 /* RECT textures are distinguished from 2D textures via np2_fixup */
3323 case GL_TEXTURE_RECTANGLE_ARB:
3324 case GL_TEXTURE_2D:
3325 break;
3326
3327 case GL_TEXTURE_3D:
3328 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3329 break;
3330
3331 case GL_TEXTURE_CUBE_MAP_ARB:
3332 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3333 break;
3334 }
3335 }
3336 }
3337
3338 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3339 {
3340 if (!shader->reg_maps.resource_info[i].type)
3341 continue;
3342
3343 texture = state->textures[i];
3344 if (!texture)
3345 {
3346 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3347 continue;
3348 }
3349 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3350 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3351 else
3352 args->color_fixup[i] = texture->resource.format->color_fixup;
3353
3354 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3355 args->shadow |= 1u << i;
3356
3357 /* Flag samplers that need NP2 texcoord fixup. */
3358 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3359 args->np2_fixup |= (1u << i);
3360 }
3361
3362 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3363 if (shader->reg_maps.shader_version.major >= 4)
3364 args->shadow = 0;
3365
3366 if (shader->reg_maps.shader_version.major >= 3)
3367 {
3368 if (position_transformed)
3369 args->vp_mode = pretransformed;
3370 else if (use_vs(state))
3371 args->vp_mode = vertexshader;
3372 else
3373 args->vp_mode = fixedfunction;
3374 args->fog = WINED3D_FFP_PS_FOG_OFF;
3375 }
3376 else
3377 {
3378 args->vp_mode = vertexshader;
3379 if (state->render_states[WINED3D_RS_FOGENABLE])
3380 {
3381 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3382 {
3383 case WINED3D_FOG_NONE:
3384 if (position_transformed || use_vs(state))
3385 {
3386 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3387 break;
3388 }
3389
3390 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3391 {
3392 case WINED3D_FOG_NONE: /* Fall through. */
3393 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3394 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3395 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3396 }
3397 break;
3398
3399 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3400 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3401 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3402 }
3403 }
3404 else
3405 {
3406 args->fog = WINED3D_FFP_PS_FOG_OFF;
3407 }
3408 }
3409
3410 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3411 {
3412 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3413
3414 args->texcoords_initialized = 0;
3415 for (i = 0; i < MAX_TEXTURES; ++i)
3416 {
3417 if (vs)
3418 {
3419 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3420 args->texcoords_initialized |= 1u << i;
3421 }
3422 else
3423 {
3424 const struct wined3d_stream_info *si = &context->stream_info;
3425 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3426
3427 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3428 & WINED3D_FFP_TCI_MASK
3429 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3430 args->texcoords_initialized |= 1u << i;
3431 }
3432 }
3433 }
3434 else
3435 {
3436 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3437 }
3438
3439 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3440 && state->gl_primitive_type == GL_POINTS;
3441
3442 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3443 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3444 else
3445 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3446 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3447 : WINED3D_CMP_ALWAYS) - 1;
3448
3449 if (d3d_info->emulated_flatshading)
3450 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3451
3452 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3453 ? context->render_offscreen : 0;
3454 }
3455
3456 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3457 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3458 {
3459 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3460 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3461 HRESULT hr;
3462
3463 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3464 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3465 return hr;
3466
3467 for (i = 0; i < MAX_REG_INPUT; ++i)
3468 {
3469 if (shader->u.ps.input_reg_used[i])
3470 {
3471 ++num_regs_used;
3472 highest_reg_used = i;
3473 }
3474 }
3475
3476 /* Don't do any register mapping magic if it is not needed, or if we can't
3477 * achieve anything anyway */
3478 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3479 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3480 || shader->reg_maps.shader_version.major >= 4)
3481 {
3482 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3483 {
3484 /* This happens with relative addressing. The input mapper function
3485 * warns about this if the higher registers are declared too, so
3486 * don't write a FIXME here */
3487 WARN("More varying registers used than supported\n");
3488 }
3489
3490 for (i = 0; i < MAX_REG_INPUT; ++i)
3491 {
3492 shader->u.ps.input_reg_map[i] = i;
3493 }
3494
3495 shader->u.ps.declared_in_count = highest_reg_used + 1;
3496 }
3497 else
3498 {
3499 shader->u.ps.declared_in_count = 0;
3500 for (i = 0; i < MAX_REG_INPUT; ++i)
3501 {
3502 if (shader->u.ps.input_reg_used[i])
3503 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3504 else shader->u.ps.input_reg_map[i] = ~0U;
3505 }
3506 }
3507
3508 return WINED3D_OK;
3509 }
3510
3511 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3512 {
3513 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3514 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3515 unsigned int i;
3516
3517 if (reg_maps->shader_version.major != 1) return;
3518
3519 for (i = 0; i < shader->limits->sampler; ++i)
3520 {
3521 /* We don't sample from this sampler. */
3522 if (!resource_info[i].type)
3523 continue;
3524
3525 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3526 {
3527 case WINED3D_SHADER_TEX_2D:
3528 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3529 break;
3530
3531 case WINED3D_SHADER_TEX_3D:
3532 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3533 break;
3534
3535 case WINED3D_SHADER_TEX_CUBE:
3536 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3537 break;
3538 }
3539 }
3540 }
3541
3542 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3543 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3544 {
3545 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3546 }
3547
3548 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3549 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3550 {
3551 struct wined3d_shader *object;
3552 HRESULT hr;
3553
3554 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3555 device, desc, parent, parent_ops, shader);
3556
3557 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3558 return E_OUTOFMEMORY;
3559
3560 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3561 {
3562 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3563 HeapFree(GetProcessHeap(), 0, object);
3564 return hr;
3565 }
3566
3567 TRACE("Created compute shader %p.\n", object);
3568 *shader = object;
3569
3570 return WINED3D_OK;
3571 }
3572
3573 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3574 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3575 {
3576 struct wined3d_shader *object;
3577 HRESULT hr;
3578
3579 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3580 device, desc, parent, parent_ops, shader);
3581
3582 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3583 return E_OUTOFMEMORY;
3584
3585 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3586 {
3587 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3588 HeapFree(GetProcessHeap(), 0, object);
3589 return hr;
3590 }
3591
3592 TRACE("Created domain shader %p.\n", object);
3593 *shader = object;
3594
3595 return WINED3D_OK;
3596 }
3597
3598 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3599 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3600 {
3601 struct wined3d_shader *object;
3602 HRESULT hr;
3603
3604 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3605 device, desc, parent, parent_ops, shader);
3606
3607 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3608 if (!object)
3609 return E_OUTOFMEMORY;
3610
3611 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3612 {
3613 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3614 HeapFree(GetProcessHeap(), 0, object);
3615 return hr;
3616 }
3617
3618 TRACE("Created geometry shader %p.\n", object);
3619 *shader = object;
3620
3621 return WINED3D_OK;
3622 }
3623
3624 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3625 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3626 {
3627 struct wined3d_shader *object;
3628 HRESULT hr;
3629
3630 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3631 device, desc, parent, parent_ops, shader);
3632
3633 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3634 return E_OUTOFMEMORY;
3635
3636 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3637 {
3638 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3639 HeapFree(GetProcessHeap(), 0, object);
3640 return hr;
3641 }
3642
3643 TRACE("Created hull shader %p.\n", object);
3644 *shader = object;
3645
3646 return WINED3D_OK;
3647 }
3648
3649 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3650 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3651 {
3652 struct wined3d_shader *object;
3653 HRESULT hr;
3654
3655 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3656 device, desc, parent, parent_ops, shader);
3657
3658 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3659 if (!object)
3660 return E_OUTOFMEMORY;
3661
3662 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3663 {
3664 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3665 HeapFree(GetProcessHeap(), 0, object);
3666 return hr;
3667 }
3668
3669 TRACE("Created pixel shader %p.\n", object);
3670 *shader = object;
3671
3672 return WINED3D_OK;
3673 }
3674
3675 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3676 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3677 {
3678 struct wined3d_shader *object;
3679 HRESULT hr;
3680
3681 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3682 device, desc, parent, parent_ops, shader);
3683
3684 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3685 if (!object)
3686 return E_OUTOFMEMORY;
3687
3688 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3689 {
3690 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3691 HeapFree(GetProcessHeap(), 0, object);
3692 return hr;
3693 }
3694
3695 TRACE("Created vertex shader %p.\n", object);
3696 *shader = object;
3697
3698 return WINED3D_OK;
3699 }
Windows API implementation