?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Recognize SM4.1 oMask output register.
[wine.git] / dlls / wined3d / shader.c
blobfeaa6b987568178f013dd4b5161fa7565811bbdc
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_OR */ "atomic_or",
50 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
51 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
52 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
53 /* WINED3DSIH_BEM */ "bem",
54 /* WINED3DSIH_BFI */ "bfi",
55 /* WINED3DSIH_BFREV */ "bfrev",
56 /* WINED3DSIH_BREAK */ "break",
57 /* WINED3DSIH_BREAKC */ "breakc",
58 /* WINED3DSIH_BREAKP */ "breakp",
59 /* WINED3DSIH_BUFINFO */ "bufinfo",
60 /* WINED3DSIH_CALL */ "call",
61 /* WINED3DSIH_CALLNZ */ "callnz",
62 /* WINED3DSIH_CASE */ "case",
63 /* WINED3DSIH_CMP */ "cmp",
64 /* WINED3DSIH_CND */ "cnd",
65 /* WINED3DSIH_CONTINUE */ "continue",
66 /* WINED3DSIH_CRS */ "crs",
67 /* WINED3DSIH_CUT */ "cut",
68 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
69 /* WINED3DSIH_DCL */ "dcl",
70 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
71 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
72 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
73 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
74 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
75 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
76 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
77 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
78 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
79 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
80 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
81 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
82 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
83 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
84 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
85 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
86 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
87 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
88 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
89 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
90 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
91 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
92 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
93 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
94 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
95 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
96 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
97 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
98 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
99 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
100 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
101 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
102 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
103 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
104 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
105 /* WINED3DSIH_DEF */ "def",
106 /* WINED3DSIH_DEFAULT */ "default",
107 /* WINED3DSIH_DEFB */ "defb",
108 /* WINED3DSIH_DEFI */ "defi",
109 /* WINED3DSIH_DIV */ "div",
110 /* WINED3DSIH_DP2 */ "dp2",
111 /* WINED3DSIH_DP2ADD */ "dp2add",
112 /* WINED3DSIH_DP3 */ "dp3",
113 /* WINED3DSIH_DP4 */ "dp4",
114 /* WINED3DSIH_DST */ "dst",
115 /* WINED3DSIH_DSX */ "dsx",
116 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
117 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
118 /* WINED3DSIH_DSY */ "dsy",
119 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
120 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
121 /* WINED3DSIH_ELSE */ "else",
122 /* WINED3DSIH_EMIT */ "emit",
123 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
124 /* WINED3DSIH_ENDIF */ "endif",
125 /* WINED3DSIH_ENDLOOP */ "endloop",
126 /* WINED3DSIH_ENDREP */ "endrep",
127 /* WINED3DSIH_ENDSWITCH */ "endswitch",
128 /* WINED3DSIH_EQ */ "eq",
129 /* WINED3DSIH_EXP */ "exp",
130 /* WINED3DSIH_EXPP */ "expp",
131 /* WINED3DSIH_FCALL */ "fcall",
132 /* WINED3DSIH_FRC */ "frc",
133 /* WINED3DSIH_FTOI */ "ftoi",
134 /* WINED3DSIH_FTOU */ "ftou",
135 /* WINED3DSIH_GATHER4 */ "gather4",
136 /* WINED3DSIH_GATHER4_C */ "gather4_c",
137 /* WINED3DSIH_GE */ "ge",
138 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
139 /* WINED3DSIH_HS_DECLS */ "hs_decls",
140 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
141 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
142 /* WINED3DSIH_IADD */ "iadd",
143 /* WINED3DSIH_IEQ */ "ieq",
144 /* WINED3DSIH_IF */ "if",
145 /* WINED3DSIH_IFC */ "ifc",
146 /* WINED3DSIH_IGE */ "ige",
147 /* WINED3DSIH_ILT */ "ilt",
148 /* WINED3DSIH_IMAD */ "imad",
149 /* WINED3DSIH_IMAX */ "imax",
150 /* WINED3DSIH_IMIN */ "imin",
151 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
152 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
153 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
154 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
155 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
156 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
157 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
158 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
159 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
160 /* WINED3DSIH_IMUL */ "imul",
161 /* WINED3DSIH_INE */ "ine",
162 /* WINED3DSIH_INEG */ "ineg",
163 /* WINED3DSIH_ISHL */ "ishl",
164 /* WINED3DSIH_ISHR */ "ishr",
165 /* WINED3DSIH_ITOF */ "itof",
166 /* WINED3DSIH_LABEL */ "label",
167 /* WINED3DSIH_LD */ "ld",
168 /* WINED3DSIH_LD2DMS */ "ld2dms",
169 /* WINED3DSIH_LD_RAW */ "ld_raw",
170 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
171 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
172 /* WINED3DSIH_LIT */ "lit",
173 /* WINED3DSIH_LOD */ "lod",
174 /* WINED3DSIH_LOG */ "log",
175 /* WINED3DSIH_LOGP */ "logp",
176 /* WINED3DSIH_LOOP */ "loop",
177 /* WINED3DSIH_LRP */ "lrp",
178 /* WINED3DSIH_LT */ "lt",
179 /* WINED3DSIH_M3x2 */ "m3x2",
180 /* WINED3DSIH_M3x3 */ "m3x3",
181 /* WINED3DSIH_M3x4 */ "m3x4",
182 /* WINED3DSIH_M4x3 */ "m4x3",
183 /* WINED3DSIH_M4x4 */ "m4x4",
184 /* WINED3DSIH_MAD */ "mad",
185 /* WINED3DSIH_MAX */ "max",
186 /* WINED3DSIH_MIN */ "min",
187 /* WINED3DSIH_MOV */ "mov",
188 /* WINED3DSIH_MOVA */ "mova",
189 /* WINED3DSIH_MOVC */ "movc",
190 /* WINED3DSIH_MUL */ "mul",
191 /* WINED3DSIH_NE */ "ne",
192 /* WINED3DSIH_NOP */ "nop",
193 /* WINED3DSIH_NOT */ "not",
194 /* WINED3DSIH_NRM */ "nrm",
195 /* WINED3DSIH_OR */ "or",
196 /* WINED3DSIH_PHASE */ "phase",
197 /* WINED3DSIH_POW */ "pow",
198 /* WINED3DSIH_RCP */ "rcp",
199 /* WINED3DSIH_REP */ "rep",
200 /* WINED3DSIH_RESINFO */ "resinfo",
201 /* WINED3DSIH_RET */ "ret",
202 /* WINED3DSIH_ROUND_NE */ "round_ne",
203 /* WINED3DSIH_ROUND_NI */ "round_ni",
204 /* WINED3DSIH_ROUND_PI */ "round_pi",
205 /* WINED3DSIH_ROUND_Z */ "round_z",
206 /* WINED3DSIH_RSQ */ "rsq",
207 /* WINED3DSIH_SAMPLE */ "sample",
208 /* WINED3DSIH_SAMPLE_B */ "sample_b",
209 /* WINED3DSIH_SAMPLE_C */ "sample_c",
210 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
211 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
212 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
213 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
214 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
215 /* WINED3DSIH_SETP */ "setp",
216 /* WINED3DSIH_SGE */ "sge",
217 /* WINED3DSIH_SGN */ "sgn",
218 /* WINED3DSIH_SINCOS */ "sincos",
219 /* WINED3DSIH_SLT */ "slt",
220 /* WINED3DSIH_SQRT */ "sqrt",
221 /* WINED3DSIH_STORE_RAW */ "store_raw",
222 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
223 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
224 /* WINED3DSIH_SUB */ "sub",
225 /* WINED3DSIH_SWAPC */ "swapc",
226 /* WINED3DSIH_SWITCH */ "switch",
227 /* WINED3DSIH_SYNC */ "sync",
228 /* WINED3DSIH_TEX */ "texld",
229 /* WINED3DSIH_TEXBEM */ "texbem",
230 /* WINED3DSIH_TEXBEML */ "texbeml",
231 /* WINED3DSIH_TEXCOORD */ "texcrd",
232 /* WINED3DSIH_TEXDEPTH */ "texdepth",
233 /* WINED3DSIH_TEXDP3 */ "texdp3",
234 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
235 /* WINED3DSIH_TEXKILL */ "texkill",
236 /* WINED3DSIH_TEXLDD */ "texldd",
237 /* WINED3DSIH_TEXLDL */ "texldl",
238 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
239 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
240 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
241 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
242 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
243 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
244 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
245 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
246 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
247 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
248 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
249 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
250 /* WINED3DSIH_UBFE */ "ubfe",
251 /* WINED3DSIH_UDIV */ "udiv",
252 /* WINED3DSIH_UGE */ "uge",
253 /* WINED3DSIH_ULT */ "ult",
254 /* WINED3DSIH_UMAX */ "umax",
255 /* WINED3DSIH_UMIN */ "umin",
256 /* WINED3DSIH_USHR */ "ushr",
257 /* WINED3DSIH_UTOF */ "utof",
258 /* WINED3DSIH_XOR */ "xor",
259 };
260
261 static const char * const semantic_names[] =
262 {
263 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
264 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
265 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
266 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
267 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
268 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
269 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
270 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
271 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
272 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
273 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
274 /* WINED3D_DECL_USAGE_FOG */ "FOG",
275 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
276 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
277 };
278
279 static const struct
280 {
281 enum wined3d_sysval_semantic sysval_semantic;
282 const char *sysval_name;
283 }
284 sysval_semantic_names[] =
285 {
286 {WINED3D_SV_POSITION, "SV_Position"},
287 {WINED3D_SV_CLIP_DISTANCE, "SV_ClipDistance"},
288 {WINED3D_SV_CULL_DISTANCE, "SV_CullDistance"},
289 {WINED3D_SV_RENDER_TARGET_ARRAY_INDEX, "SV_RenderTargetArrayIndex"},
290 {WINED3D_SV_VIEWPORT_ARRAY_INDEX, "SV_ViewportArrayIndex"},
291 {WINED3D_SV_VERTEX_ID, "SV_VertexID"},
292 {WINED3D_SV_INSTANCE_ID, "SV_InstanceID"},
293 {WINED3D_SV_PRIMITIVE_ID, "SV_PrimitiveID"},
294 {WINED3D_SV_IS_FRONT_FACE, "SV_IsFrontFace"},
295 {WINED3D_SV_SAMPLE_INDEX, "SV_SampleIndex"},
296 {WINED3D_SV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
297 {WINED3D_SV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
298 {WINED3D_SV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
299 {WINED3D_SV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
300 {WINED3D_SV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
301 {WINED3D_SV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
302 {WINED3D_SV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
303 {WINED3D_SV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
304 {WINED3D_SV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
305 {WINED3D_SV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
306 {WINED3D_SV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
307 {WINED3D_SV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
308 };
309
310 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
311 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
312
313 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
314 {
315 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
316 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
317
318 return shader_opcode_names[handler_idx];
319 }
320
321 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
322 {
323 if (usage >= ARRAY_SIZE(semantic_names))
324 {
325 FIXME("Unrecognized usage %#x.\n", usage);
326 return "UNRECOGNIZED";
327 }
328
329 return semantic_names[usage];
330 }
331
332 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
333 {
334 unsigned int i;
335
336 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
337 {
338 if (!strcmp(name, semantic_names[i]))
339 return i;
340 }
341
342 return ~0U;
343 }
344
345 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
346 {
347 switch (usage)
348 {
349 case WINED3D_DECL_USAGE_POSITION:
350 return WINED3D_SV_POSITION;
351 default:
352 return 0;
353 }
354 }
355
356 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
357 {
358 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
359 }
360
361 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
362 const struct wined3d_shader_semantic *s)
363 {
364 e->semantic_name = shader_semantic_name_from_usage(s->usage);
365 e->semantic_idx = s->usage_idx;
366 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
367 e->component_type = WINED3D_TYPE_FLOAT;
368 e->register_idx = s->reg.reg.idx[0].offset;
369 e->mask = s->reg.write_mask;
370 }
371
372 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
373 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
374 {
375 e->semantic_name = shader_semantic_name_from_usage(usage);
376 e->semantic_idx = usage_idx;
377 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
378 e->component_type = WINED3D_TYPE_FLOAT;
379 e->register_idx = reg_idx;
380 e->mask = write_mask;
381 }
382
383 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
384 {
385 switch (version_token >> 16)
386 {
387 case WINED3D_SM1_VS:
388 case WINED3D_SM1_PS:
389 return &sm1_shader_frontend;
390
391 case WINED3D_SM4_PS:
392 case WINED3D_SM4_VS:
393 case WINED3D_SM4_GS:
394 case WINED3D_SM5_HS:
395 case WINED3D_SM5_DS:
396 case WINED3D_SM5_CS:
397 return &sm4_shader_frontend;
398
399 default:
400 FIXME("Unrecognised version token %#x.\n", version_token);
401 return NULL;
402 }
403 }
404
405 void string_buffer_clear(struct wined3d_string_buffer *buffer)
406 {
407 buffer->buffer[0] = '\0';
408 buffer->content_size = 0;
409 }
410
411 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
412 {
413 buffer->buffer_size = 32;
414 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
415 {
416 ERR("Failed to allocate shader buffer memory.\n");
417 return FALSE;
418 }
419
420 string_buffer_clear(buffer);
421 return TRUE;
422 }
423
424 void string_buffer_free(struct wined3d_string_buffer *buffer)
425 {
426 HeapFree(GetProcessHeap(), 0, buffer->buffer);
427 }
428
429 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
430 {
431 char *new_buffer;
432 unsigned int new_buffer_size = buffer->buffer_size * 2;
433
434 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
435 new_buffer_size *= 2;
436 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
437 {
438 ERR("Failed to grow buffer.\n");
439 buffer->buffer[buffer->content_size] = '\0';
440 return FALSE;
441 }
442 buffer->buffer = new_buffer;
443 buffer->buffer_size = new_buffer_size;
444 return TRUE;
445 }
446
447 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
448 {
449 unsigned int rem;
450 int rc;
451
452 rem = buffer->buffer_size - buffer->content_size;
453 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
454 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
455 return rc;
456
457 buffer->content_size += rc;
458 return 0;
459 }
460
461 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
462 {
463 va_list args;
464 int ret;
465
466 for (;;)
467 {
468 va_start(args, format);
469 ret = shader_vaddline(buffer, format, args);
470 va_end(args);
471 if (!ret)
472 return ret;
473 if (!string_buffer_resize(buffer, ret))
474 return -1;
475 }
476 }
477
478 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
479 {
480 struct wined3d_string_buffer *buffer;
481
482 if (list_empty(&list->list))
483 {
484 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
485 if (!buffer || !string_buffer_init(buffer))
486 {
487 ERR("Couldn't allocate buffer for temporary string.\n");
488 HeapFree(GetProcessHeap(), 0, buffer);
489 return NULL;
490 }
491 }
492 else
493 {
494 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
495 list_remove(&buffer->entry);
496 }
497 string_buffer_clear(buffer);
498 return buffer;
499 }
500
501 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
502 {
503 if (!buffer)
504 return 0;
505 string_buffer_clear(buffer);
506 return shader_vaddline(buffer, format, args);
507 }
508
509 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
510 {
511 va_list args;
512 int ret;
513
514 for (;;)
515 {
516 va_start(args, format);
517 ret = string_buffer_vsprintf(buffer, format, args);
518 va_end(args);
519 if (!ret)
520 return;
521 if (!string_buffer_resize(buffer, ret))
522 return;
523 }
524 }
525
526 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
527 {
528 if (!buffer)
529 return;
530 list_add_head(&list->list, &buffer->entry);
531 }
532
533 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
534 {
535 list_init(&list->list);
536 }
537
538 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
539 {
540 struct wined3d_string_buffer *buffer, *buffer_next;
541
542 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
543 {
544 string_buffer_free(buffer);
545 HeapFree(GetProcessHeap(), 0, buffer);
546 }
547 list_init(&list->list);
548 }
549
550 /* Convert floating point offset relative to a register file to an absolute
551 * offset for float constants. */
552 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
553 {
554 switch (register_type)
555 {
556 case WINED3DSPR_CONST: return register_idx;
557 case WINED3DSPR_CONST2: return 2048 + register_idx;
558 case WINED3DSPR_CONST3: return 4096 + register_idx;
559 case WINED3DSPR_CONST4: return 6144 + register_idx;
560 default:
561 FIXME("Unsupported register type: %u.\n", register_type);
562 return register_idx;
563 }
564 }
565
566 static void shader_delete_constant_list(struct list *clist)
567 {
568 struct wined3d_shader_lconst *constant, *constant_next;
569
570 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
571 HeapFree(GetProcessHeap(), 0, constant);
572 list_init(clist);
573 }
574
575 static void shader_set_limits(struct wined3d_shader *shader)
576 {
577 static const struct limits_entry
578 {
579 unsigned int min_version;
580 unsigned int max_version;
581 struct wined3d_shader_limits limits;
582 }
583 vs_limits[] =
584 {
585 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
586 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
587 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
588 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
589 * even though they are capable of supporting much more (GL
590 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
591 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
592 * shaders to 256. */
593 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
594 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
595 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
596 {0}
597 },
598 hs_limits[] =
599 {
600 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
601 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
602 },
603 ds_limits[] =
604 {
605 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
606 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
607 },
608 gs_limits[] =
609 {
610 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
611 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
612 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
613 {0}
614 },
615 ps_limits[] =
616 {
617 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
618 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
619 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
620 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
621 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
622 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
623 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
624 {0}
625 },
626 cs_limits[] =
627 {
628 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
629 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
630 };
631 const struct limits_entry *limits_array;
632 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
633 shader->reg_maps.shader_version.minor);
634 int i = 0;
635
636 switch (shader->reg_maps.shader_version.type)
637 {
638 default:
639 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
640 /* Fall-through. */
641 case WINED3D_SHADER_TYPE_VERTEX:
642 limits_array = vs_limits;
643 break;
644 case WINED3D_SHADER_TYPE_HULL:
645 limits_array = hs_limits;
646 break;
647 case WINED3D_SHADER_TYPE_DOMAIN:
648 limits_array = ds_limits;
649 break;
650 case WINED3D_SHADER_TYPE_GEOMETRY:
651 limits_array = gs_limits;
652 break;
653 case WINED3D_SHADER_TYPE_PIXEL:
654 limits_array = ps_limits;
655 break;
656 case WINED3D_SHADER_TYPE_COMPUTE:
657 limits_array = cs_limits;
658 break;
659 }
660
661 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
662 {
663 if (shader_version <= limits_array[i].max_version)
664 {
665 shader->limits = &limits_array[i].limits;
666 break;
667 }
668 ++i;
669 }
670 if (!shader->limits)
671 {
672 FIXME("Unexpected shader version \"%u.%u\".\n",
673 shader->reg_maps.shader_version.major,
674 shader->reg_maps.shader_version.minor);
675 shader->limits = &limits_array[max(0, i - 1)].limits;
676 }
677 }
678
679 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
680 {
681 DWORD idx, shift;
682 idx = bit >> 5;
683 shift = bit & 0x1f;
684 bitmap[idx] |= (1u << shift);
685 }
686
687 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
688 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
689 {
690 switch (reg->type)
691 {
692 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
693 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
694 reg_maps->texcoord |= 1u << reg->idx[0].offset;
695 else
696 reg_maps->address |= 1u << reg->idx[0].offset;
697 break;
698
699 case WINED3DSPR_TEMP:
700 reg_maps->temporary |= 1u << reg->idx[0].offset;
701 break;
702
703 case WINED3DSPR_INPUT:
704 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
705 {
706 if (reg->idx[0].rel_addr)
707 {
708 /* If relative addressing is used, we must assume that all registers
709 * are used. Even if it is a construct like v3[aL], we can't assume
710 * that v0, v1 and v2 aren't read because aL can be negative */
711 unsigned int i;
712 for (i = 0; i < MAX_REG_INPUT; ++i)
713 {
714 shader->u.ps.input_reg_used[i] = TRUE;
715 }
716 }
717 else
718 {
719 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
720 }
721 }
722 else
723 reg_maps->input_registers |= 1u << reg->idx[0].offset;
724 break;
725
726 case WINED3DSPR_RASTOUT:
727 if (reg->idx[0].offset == 1)
728 reg_maps->fog = 1;
729 if (reg->idx[0].offset == 2)
730 reg_maps->point_size = 1;
731 break;
732
733 case WINED3DSPR_MISCTYPE:
734 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
735 {
736 if (!reg->idx[0].offset)
737 reg_maps->vpos = 1;
738 else if (reg->idx[0].offset == 1)
739 reg_maps->usesfacing = 1;
740 }
741 break;
742
743 case WINED3DSPR_CONST:
744 if (reg->idx[0].rel_addr)
745 {
746 if (reg->idx[0].offset < reg_maps->min_rel_offset)
747 reg_maps->min_rel_offset = reg->idx[0].offset;
748 if (reg->idx[0].offset > reg_maps->max_rel_offset)
749 reg_maps->max_rel_offset = reg->idx[0].offset;
750 reg_maps->usesrelconstF = TRUE;
751 }
752 else
753 {
754 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
755 {
756 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
757 return FALSE;
758 }
759 else
760 {
761 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
762 }
763 }
764 break;
765
766 case WINED3DSPR_CONSTINT:
767 if (reg->idx[0].offset >= shader->limits->constant_int)
768 {
769 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
770 return FALSE;
771 }
772 else
773 {
774 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
775 }
776 break;
777
778 case WINED3DSPR_CONSTBOOL:
779 if (reg->idx[0].offset >= shader->limits->constant_bool)
780 {
781 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
782 return FALSE;
783 }
784 else
785 {
786 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
787 }
788 break;
789
790 case WINED3DSPR_COLOROUT:
791 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
792 break;
793
794 default:
795 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
796 reg->type, reg->idx[0].offset, reg->idx[1].offset);
797 break;
798 }
799 return TRUE;
800 }
801
802 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
803 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
804 {
805 struct wined3d_shader_sampler_map_entry *entries, *entry;
806 struct wined3d_shader_sampler_map *map;
807 unsigned int i;
808
809 map = &reg_maps->sampler_map;
810 entries = map->entries;
811 for (i = 0; i < map->count; ++i)
812 {
813 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
814 return;
815 }
816
817 if (!map->size)
818 {
819 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
820 {
821 ERR("Failed to allocate sampler map entries.\n");
822 return;
823 }
824 map->size = 4;
825 map->entries = entries;
826 }
827 else if (map->count == map->size)
828 {
829 size_t new_size = map->size * 2;
830
831 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
832 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
833 {
834 ERR("Failed to resize sampler map entries.\n");
835 return;
836 }
837 map->size = new_size;
838 map->entries = entries;
839 }
840
841 entry = &entries[map->count++];
842 entry->resource_idx = resource_idx;
843 entry->sampler_idx = sampler_idx;
844 entry->bind_idx = bind_idx;
845 }
846
847 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
848 {
849 switch (instr)
850 {
851 case WINED3DSIH_M4x4:
852 case WINED3DSIH_M3x4:
853 return param == 1 ? 3 : 0;
854
855 case WINED3DSIH_M4x3:
856 case WINED3DSIH_M3x3:
857 return param == 1 ? 2 : 0;
858
859 case WINED3DSIH_M3x2:
860 return param == 1 ? 1 : 0;
861
862 default:
863 return 0;
864 }
865 }
866
867 /* Note that this does not count the loop register as an address register. */
868 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
869 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
870 struct wined3d_shader_signature *output_signature, const DWORD *byte_code, DWORD constf_size)
871 {
872 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
873 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
874 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
875 void *fe_data = shader->frontend_data;
876 struct wined3d_shader_version shader_version;
877 const DWORD *ptr = byte_code;
878 unsigned int i;
879
880 memset(reg_maps, 0, sizeof(*reg_maps));
881 memset(input_signature_elements, 0, sizeof(input_signature_elements));
882 memset(output_signature_elements, 0, sizeof(output_signature_elements));
883 reg_maps->min_rel_offset = ~0U;
884 list_init(&reg_maps->indexable_temps);
885
886 fe->shader_read_header(fe_data, &ptr, &shader_version);
887 reg_maps->shader_version = shader_version;
888
889 shader_set_limits(shader);
890
891 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
892 sizeof(*reg_maps->constf))))
893 {
894 ERR("Failed to allocate constant map memory.\n");
895 return E_OUTOFMEMORY;
896 }
897
898 while (!fe->shader_is_end(fe_data, &ptr))
899 {
900 struct wined3d_shader_instruction ins;
901
902 /* Fetch opcode. */
903 fe->shader_read_instruction(fe_data, &ptr, &ins);
904
905 /* Unhandled opcode, and its parameters. */
906 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
907 {
908 TRACE("Skipping unrecognized instruction.\n");
909 continue;
910 }
911
912 /* Handle declarations. */
913 if (ins.handler_idx == WINED3DSIH_DCL
914 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
915 {
916 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
917 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
918
919 switch (semantic->reg.reg.type)
920 {
921 /* Mark input registers used. */
922 case WINED3DSPR_INPUT:
923 if (reg_idx >= MAX_REG_INPUT)
924 {
925 ERR("Invalid input register index %u.\n", reg_idx);
926 break;
927 }
928 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
929 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
930 return WINED3DERR_INVALIDCALL;
931 reg_maps->input_registers |= 1u << reg_idx;
932 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
933 break;
934
935 /* Vertex shader: mark 3.0 output registers used, save token. */
936 case WINED3DSPR_OUTPUT:
937 if (reg_idx >= MAX_REG_OUTPUT)
938 {
939 ERR("Invalid output register index %u.\n", reg_idx);
940 break;
941 }
942 reg_maps->output_registers |= 1u << reg_idx;
943 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
944 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
945 reg_maps->fog = 1;
946 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
947 reg_maps->point_size = 1;
948 break;
949
950 case WINED3DSPR_SAMPLER:
951 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
952 case WINED3DSPR_RESOURCE:
953 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
954 {
955 ERR("Invalid resource index %u.\n", reg_idx);
956 break;
957 }
958 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
959 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
960 break;
961
962 case WINED3DSPR_UAV:
963 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
964 {
965 ERR("Invalid UAV resource index %u.\n", reg_idx);
966 break;
967 }
968 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
969 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
970 break;
971
972 default:
973 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
974 break;
975 }
976 }
977 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
978 {
979 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
980 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
981 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
982 else
983 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
984 }
985 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
986 {
987 if (reg_maps->icb)
988 FIXME("Multiple immediate constant buffers.\n");
989 reg_maps->icb = ins.declaration.icb;
990 }
991 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
992 {
993 struct wined3d_shader_indexable_temp *reg;
994
995 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
996 return E_OUTOFMEMORY;
997
998 *reg = ins.declaration.indexable_temp;
999 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1000 }
1001 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1002 {
1003 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1004 shader->u.gs.input_type = ins.declaration.primitive_type;
1005 else
1006 FIXME("Invalid instruction %#x for shader type %#x.\n",
1007 ins.handler_idx, shader_version.type);
1008 }
1009 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1010 {
1011 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1012 shader->u.gs.output_type = ins.declaration.primitive_type;
1013 else
1014 FIXME("Invalid instruction %#x for shader type %#x.\n",
1015 ins.handler_idx, shader_version.type);
1016 }
1017 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1018 {
1019 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1020 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1021 }
1022 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1023 {
1024 reg_maps->temporary_count = ins.declaration.count;
1025 }
1026 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1027 {
1028 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1029 shader->u.gs.vertices_out = ins.declaration.count;
1030 else
1031 FIXME("Invalid instruction %#x for shader type %#x.\n",
1032 ins.handler_idx, shader_version.type);
1033 }
1034 else if (ins.handler_idx == WINED3DSIH_DEF)
1035 {
1036 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1037 float *value;
1038 if (!lconst) return E_OUTOFMEMORY;
1039
1040 lconst->idx = ins.dst[0].reg.idx[0].offset;
1041 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1042 value = (float *)lconst->value;
1043
1044 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1045 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1046 {
1047 if (value[0] < -1.0f) value[0] = -1.0f;
1048 else if (value[0] > 1.0f) value[0] = 1.0f;
1049 if (value[1] < -1.0f) value[1] = -1.0f;
1050 else if (value[1] > 1.0f) value[1] = 1.0f;
1051 if (value[2] < -1.0f) value[2] = -1.0f;
1052 else if (value[2] > 1.0f) value[2] = 1.0f;
1053 if (value[3] < -1.0f) value[3] = -1.0f;
1054 else if (value[3] > 1.0f) value[3] = 1.0f;
1055 }
1056
1057 list_add_head(&shader->constantsF, &lconst->entry);
1058
1059 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1060 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1061 {
1062 shader->lconst_inf_or_nan = TRUE;
1063 }
1064 }
1065 else if (ins.handler_idx == WINED3DSIH_DEFI)
1066 {
1067 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1068 if (!lconst) return E_OUTOFMEMORY;
1069
1070 lconst->idx = ins.dst[0].reg.idx[0].offset;
1071 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1072
1073 list_add_head(&shader->constantsI, &lconst->entry);
1074 reg_maps->local_int_consts |= (1u << lconst->idx);
1075 }
1076 else if (ins.handler_idx == WINED3DSIH_DEFB)
1077 {
1078 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1079 if (!lconst) return E_OUTOFMEMORY;
1080
1081 lconst->idx = ins.dst[0].reg.idx[0].offset;
1082 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1083
1084 list_add_head(&shader->constantsB, &lconst->entry);
1085 reg_maps->local_bool_consts |= (1u << lconst->idx);
1086 }
1087 /* For subroutine prototypes. */
1088 else if (ins.handler_idx == WINED3DSIH_LABEL)
1089 {
1090 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1091 }
1092 /* Set texture, address, temporary registers. */
1093 else
1094 {
1095 BOOL color0_mov = FALSE;
1096 unsigned int i;
1097
1098 /* This will loop over all the registers and try to
1099 * make a bitmask of the ones we're interested in.
1100 *
1101 * Relative addressing tokens are ignored, but that's
1102 * okay, since we'll catch any address registers when
1103 * they are initialized (required by spec). */
1104 for (i = 0; i < ins.dst_count; ++i)
1105 {
1106 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1107 shader_version.type, constf_size))
1108 return WINED3DERR_INVALIDCALL;
1109
1110 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1111 {
1112 UINT idx = ins.dst[i].reg.idx[0].offset;
1113
1114 switch (ins.dst[i].reg.type)
1115 {
1116 case WINED3DSPR_RASTOUT:
1117 if (shader_version.major >= 3)
1118 break;
1119 switch (idx)
1120 {
1121 case 0: /* oPos */
1122 reg_maps->output_registers |= 1u << 10;
1123 shader_signature_from_usage(&output_signature_elements[10],
1124 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1125 break;
1126
1127 case 1: /* oFog */
1128 reg_maps->output_registers |= 1u << 11;
1129 shader_signature_from_usage(&output_signature_elements[11],
1130 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1131 break;
1132
1133 case 2: /* oPts */
1134 reg_maps->output_registers |= 1u << 11;
1135 shader_signature_from_usage(&output_signature_elements[11],
1136 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1137 break;
1138 }
1139 break;
1140
1141 case WINED3DSPR_ATTROUT:
1142 if (shader_version.major >= 3)
1143 break;
1144 if (idx < 2)
1145 {
1146 idx += 8;
1147 if (reg_maps->output_registers & (1u << idx))
1148 {
1149 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1150 }
1151 else
1152 {
1153 reg_maps->output_registers |= 1u << idx;
1154 shader_signature_from_usage(&output_signature_elements[idx],
1155 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1156 }
1157 }
1158 break;
1159
1160 case WINED3DSPR_TEXCRDOUT:
1161 if (shader_version.major >= 3)
1162 {
1163 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1164 break;
1165 }
1166 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1167 if (reg_maps->output_registers & (1u << idx))
1168 {
1169 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1170 }
1171 else
1172 {
1173 reg_maps->output_registers |= 1u << idx;
1174 shader_signature_from_usage(&output_signature_elements[idx],
1175 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1176 }
1177 break;
1178
1179 default:
1180 break;
1181 }
1182 }
1183
1184 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1185 {
1186 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1187 {
1188 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1189 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1190 * the mov and perform the sRGB write correction from the source register.
1191 *
1192 * However, if the mov is only partial, we can't do this, and if the write
1193 * comes from an instruction other than MOV it is hard to do as well. If
1194 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1195 shader->u.ps.color0_mov = FALSE;
1196 if (ins.handler_idx == WINED3DSIH_MOV
1197 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1198 {
1199 /* Used later when the source register is read. */
1200 color0_mov = TRUE;
1201 }
1202 }
1203 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1204 * end
1205 */
1206 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1207 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1208 {
1209 shader->u.ps.color0_mov = FALSE;
1210 }
1211 }
1212
1213 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1214 if (shader_version.major == 1
1215 && (ins.handler_idx == WINED3DSIH_TEX
1216 || ins.handler_idx == WINED3DSIH_TEXBEM
1217 || ins.handler_idx == WINED3DSIH_TEXBEML
1218 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1219 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1220 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1221 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1222 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1223 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1224 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1225 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1226 {
1227 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1228
1229 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1230 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1231 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1232 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1233
1234 /* texbem is only valid with < 1.4 pixel shaders */
1235 if (ins.handler_idx == WINED3DSIH_TEXBEM
1236 || ins.handler_idx == WINED3DSIH_TEXBEML)
1237 {
1238 reg_maps->bumpmat |= 1u << reg_idx;
1239 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1240 {
1241 reg_maps->luminanceparams |= 1u << reg_idx;
1242 }
1243 }
1244 }
1245 else if (ins.handler_idx == WINED3DSIH_BEM)
1246 {
1247 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1248 }
1249 }
1250
1251 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD
1252 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED)
1253 {
1254 unsigned int reg_idx;
1255 if (ins.handler_idx == WINED3DSIH_ATOMIC_IADD)
1256 reg_idx = ins.dst[0].reg.idx[0].offset;
1257 else
1258 reg_idx = ins.src[1].reg.idx[0].offset;
1259 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1260 {
1261 ERR("Invalid UAV index %u.\n", reg_idx);
1262 break;
1263 }
1264 reg_maps->uav_read_mask |= (1u << reg_idx);
1265 }
1266 else if (ins.handler_idx == WINED3DSIH_NRM)
1267 {
1268 reg_maps->usesnrm = 1;
1269 }
1270 else if (ins.handler_idx == WINED3DSIH_DSY
1271 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1272 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1273 {
1274 reg_maps->usesdsy = 1;
1275 }
1276 else if (ins.handler_idx == WINED3DSIH_DSX
1277 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1278 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1279 {
1280 reg_maps->usesdsx = 1;
1281 }
1282 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1283 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1284 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1285 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1286 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1287 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1288 else if (ins.handler_idx == WINED3DSIH_LOOP
1289 || ins.handler_idx == WINED3DSIH_REP)
1290 {
1291 ++cur_loop_depth;
1292 if (cur_loop_depth > max_loop_depth)
1293 max_loop_depth = cur_loop_depth;
1294 }
1295 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1296 || ins.handler_idx == WINED3DSIH_ENDREP)
1297 {
1298 --cur_loop_depth;
1299 }
1300 else if (ins.handler_idx == WINED3DSIH_SAMPLE
1301 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1302 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1303 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1304 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1305 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1306 {
1307 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1308 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1309 }
1310 else if (ins.handler_idx == WINED3DSIH_LD
1311 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1312 {
1313 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1314 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1315 }
1316
1317 if (ins.predicate)
1318 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1319 shader_version.type, constf_size))
1320 return WINED3DERR_INVALIDCALL;
1321
1322 for (i = 0; i < ins.src_count; ++i)
1323 {
1324 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1325 struct wined3d_shader_register reg = ins.src[i].reg;
1326
1327 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1328 shader_version.type, constf_size))
1329 return WINED3DERR_INVALIDCALL;
1330 while (count)
1331 {
1332 ++reg.idx[0].offset;
1333 if (!shader_record_register_usage(shader, reg_maps, &reg,
1334 shader_version.type, constf_size))
1335 return WINED3DERR_INVALIDCALL;
1336 --count;
1337 }
1338
1339 if (color0_mov)
1340 {
1341 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1342 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1343 {
1344 shader->u.ps.color0_mov = TRUE;
1345 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1346 }
1347 }
1348 }
1349 }
1350 }
1351 reg_maps->loop_depth = max_loop_depth;
1352
1353 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1354 * R0 is written to the render target. */
1355 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1356 reg_maps->rt_mask |= (1u << 0);
1357
1358 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
1359
1360 if (input_signature->elements)
1361 {
1362 for (i = 0; i < input_signature->element_count; ++i)
1363 {
1364 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1365 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1366 {
1367 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1368 reg_maps->vpos = 1;
1369 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1370 reg_maps->usesfacing = 1;
1371 }
1372 }
1373 }
1374 else if (!input_signature->elements && reg_maps->input_registers)
1375 {
1376 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1377 struct wined3d_shader_signature_element *e;
1378 unsigned int i;
1379
1380 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1381 return E_OUTOFMEMORY;
1382 input_signature->element_count = count;
1383
1384 e = input_signature->elements;
1385 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1386 {
1387 if (!(reg_maps->input_registers & (1u << i)))
1388 continue;
1389 input_signature_elements[i].register_idx = i;
1390 *e++ = input_signature_elements[i];
1391 }
1392 }
1393
1394 if (output_signature->elements)
1395 {
1396 for (i = 0; i < output_signature->element_count; ++i)
1397 {
1398 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1399 }
1400 }
1401 else if (reg_maps->output_registers)
1402 {
1403 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1404 struct wined3d_shader_signature_element *e;
1405
1406 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1407 return E_OUTOFMEMORY;
1408 output_signature->element_count = count;
1409
1410 e = output_signature->elements;
1411 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1412 {
1413 if (!(reg_maps->output_registers & (1u << i)))
1414 continue;
1415 *e++ = output_signature_elements[i];
1416 }
1417 }
1418
1419 return WINED3D_OK;
1420 }
1421
1422 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1423 {
1424 struct wined3d_shader_indexable_temp *reg, *reg_next;
1425
1426 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1427 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1428
1429 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1430 HeapFree(GetProcessHeap(), 0, reg);
1431 list_init(&reg_maps->indexable_temps);
1432 }
1433
1434 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1435 {
1436 DWORD map = 1u << max;
1437 map |= map - 1;
1438 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1439
1440 return wined3d_log2i(map);
1441 }
1442
1443 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1444 {
1445 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1446 {
1447 shader_addline(buffer, "refactoringAllowed");
1448 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1449 if (global_flags)
1450 shader_addline(buffer, " | ");
1451 }
1452
1453 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1454 {
1455 shader_addline(buffer, "enableRawAndStructuredBuffers");
1456 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1457 }
1458
1459 if (global_flags)
1460 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1461 }
1462
1463 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1464 {
1465 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1466 {
1467 shader_addline(buffer, "_g");
1468 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1469 }
1470 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1471 {
1472 shader_addline(buffer, "_t");
1473 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1474 }
1475
1476 if (sync_flags)
1477 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1478 }
1479
1480 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1481 {
1482 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1483 {
1484 shader_addline(buffer, "_glc");
1485 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1486 }
1487 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1488 {
1489 shader_addline(buffer, "_opc");
1490 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1491 }
1492
1493 if (uav_flags)
1494 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1495 }
1496
1497 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1498 enum wined3d_tessellator_domain domain)
1499 {
1500 switch (domain)
1501 {
1502 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1503 shader_addline(buffer, "line");
1504 break;
1505 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1506 shader_addline(buffer, "triangle");
1507 break;
1508 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1509 shader_addline(buffer, "quad");
1510 break;
1511 default:
1512 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1513 break;
1514 }
1515 }
1516
1517 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1518 enum wined3d_tessellator_output_primitive output_primitive)
1519 {
1520 switch (output_primitive)
1521 {
1522 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1523 shader_addline(buffer, "point");
1524 break;
1525 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1526 shader_addline(buffer, "line");
1527 break;
1528 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1529 shader_addline(buffer, "triangle_cw");
1530 break;
1531 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1532 shader_addline(buffer, "triangle_ccw");
1533 break;
1534 default:
1535 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1536 break;
1537 }
1538 }
1539
1540 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1541 enum wined3d_tessellator_partitioning partitioning)
1542 {
1543 switch (partitioning)
1544 {
1545 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1546 shader_addline(buffer, "integer");
1547 break;
1548 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1549 shader_addline(buffer, "pow2");
1550 break;
1551 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1552 shader_addline(buffer, "fractional_odd");
1553 break;
1554 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1555 shader_addline(buffer, "fractional_even");
1556 break;
1557 default:
1558 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1559 break;
1560 }
1561 }
1562
1563 static void shader_dump_sysval_semantic(struct wined3d_string_buffer *buffer, enum wined3d_sysval_semantic semantic)
1564 {
1565 unsigned int i;
1566
1567 for (i = 0; i < ARRAY_SIZE(sysval_semantic_names); ++i)
1568 {
1569 if (sysval_semantic_names[i].sysval_semantic == semantic)
1570 {
1571 shader_addline(buffer, "%s", sysval_semantic_names[i].sysval_name);
1572 return;
1573 }
1574 }
1575
1576 shader_addline(buffer, "unknown_sysval_semantic(%#x)", semantic);
1577 }
1578
1579 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1580 const struct wined3d_shader_semantic *semantic, const struct wined3d_shader_version *shader_version)
1581 {
1582 shader_addline(buffer, "dcl");
1583
1584 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1585 {
1586 switch (semantic->resource_type)
1587 {
1588 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1589 shader_addline(buffer, "_2d");
1590 break;
1591
1592 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1593 shader_addline(buffer, "_3d");
1594 break;
1595
1596 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1597 shader_addline(buffer, "_cube");
1598 break;
1599
1600 default:
1601 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1602 break;
1603 }
1604 }
1605 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1606 {
1607 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1608 shader_addline(buffer, "_resource_");
1609 else
1610 shader_addline(buffer, "_uav_");
1611 switch (semantic->resource_type)
1612 {
1613 case WINED3D_SHADER_RESOURCE_BUFFER:
1614 shader_addline(buffer, "buffer");
1615 break;
1616
1617 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1618 shader_addline(buffer, "texture1d");
1619 break;
1620
1621 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1622 shader_addline(buffer, "texture2d");
1623 break;
1624
1625 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1626 shader_addline(buffer, "texture2dms");
1627 break;
1628
1629 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1630 shader_addline(buffer, "texture3d");
1631 break;
1632
1633 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1634 shader_addline(buffer, "texturecube");
1635 break;
1636
1637 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1638 shader_addline(buffer, "texture1darray");
1639 break;
1640
1641 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1642 shader_addline(buffer, "texture2darray");
1643 break;
1644
1645 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1646 shader_addline(buffer, "texture2dmsarray");
1647 break;
1648
1649 default:
1650 shader_addline(buffer, "unknown");
1651 break;
1652 }
1653 switch (semantic->resource_data_type)
1654 {
1655 case WINED3D_DATA_FLOAT:
1656 shader_addline(buffer, " (float)");
1657 break;
1658
1659 case WINED3D_DATA_INT:
1660 shader_addline(buffer, " (int)");
1661 break;
1662
1663 case WINED3D_DATA_UINT:
1664 shader_addline(buffer, " (uint)");
1665 break;
1666
1667 case WINED3D_DATA_UNORM:
1668 shader_addline(buffer, " (unorm)");
1669 break;
1670
1671 case WINED3D_DATA_SNORM:
1672 shader_addline(buffer, " (snorm)");
1673 break;
1674
1675 default:
1676 shader_addline(buffer, " (unknown)");
1677 break;
1678 }
1679 }
1680 else
1681 {
1682 /* Pixel shaders 3.0 don't have usage semantics. */
1683 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
1684 return;
1685 else
1686 shader_addline(buffer, "_");
1687
1688 switch (semantic->usage)
1689 {
1690 case WINED3D_DECL_USAGE_POSITION:
1691 shader_addline(buffer, "position%u", semantic->usage_idx);
1692 break;
1693
1694 case WINED3D_DECL_USAGE_BLEND_INDICES:
1695 shader_addline(buffer, "blend");
1696 break;
1697
1698 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1699 shader_addline(buffer, "weight");
1700 break;
1701
1702 case WINED3D_DECL_USAGE_NORMAL:
1703 shader_addline(buffer, "normal%u", semantic->usage_idx);
1704 break;
1705
1706 case WINED3D_DECL_USAGE_PSIZE:
1707 shader_addline(buffer, "psize");
1708 break;
1709
1710 case WINED3D_DECL_USAGE_COLOR:
1711 if (!semantic->usage_idx)
1712 shader_addline(buffer, "color");
1713 else
1714 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
1715 break;
1716
1717 case WINED3D_DECL_USAGE_TEXCOORD:
1718 shader_addline(buffer, "texture%u", semantic->usage_idx);
1719 break;
1720
1721 case WINED3D_DECL_USAGE_TANGENT:
1722 shader_addline(buffer, "tangent");
1723 break;
1724
1725 case WINED3D_DECL_USAGE_BINORMAL:
1726 shader_addline(buffer, "binormal");
1727 break;
1728
1729 case WINED3D_DECL_USAGE_TESS_FACTOR:
1730 shader_addline(buffer, "tessfactor");
1731 break;
1732
1733 case WINED3D_DECL_USAGE_POSITIONT:
1734 shader_addline(buffer, "positionT%u", semantic->usage_idx);
1735 break;
1736
1737 case WINED3D_DECL_USAGE_FOG:
1738 shader_addline(buffer, "fog");
1739 break;
1740
1741 case WINED3D_DECL_USAGE_DEPTH:
1742 shader_addline(buffer, "depth");
1743 break;
1744
1745 case WINED3D_DECL_USAGE_SAMPLE:
1746 shader_addline(buffer, "sample");
1747 break;
1748
1749 default:
1750 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
1751 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
1752 }
1753 }
1754 }
1755
1756 static void shader_dump_register(struct wined3d_string_buffer *buffer,
1757 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
1758 {
1759 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1760 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1761 UINT offset = reg->idx[0].offset;
1762
1763 switch (reg->type)
1764 {
1765 case WINED3DSPR_TEMP:
1766 shader_addline(buffer, "r");
1767 break;
1768
1769 case WINED3DSPR_INPUT:
1770 shader_addline(buffer, "v");
1771 break;
1772
1773 case WINED3DSPR_CONST:
1774 case WINED3DSPR_CONST2:
1775 case WINED3DSPR_CONST3:
1776 case WINED3DSPR_CONST4:
1777 shader_addline(buffer, "c");
1778 offset = shader_get_float_offset(reg->type, offset);
1779 break;
1780
1781 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1782 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1783 break;
1784
1785 case WINED3DSPR_RASTOUT:
1786 shader_addline(buffer, "%s", rastout_reg_names[offset]);
1787 break;
1788
1789 case WINED3DSPR_COLOROUT:
1790 shader_addline(buffer, "oC");
1791 break;
1792
1793 case WINED3DSPR_DEPTHOUT:
1794 shader_addline(buffer, "oDepth");
1795 break;
1796
1797 case WINED3DSPR_ATTROUT:
1798 shader_addline(buffer, "oD");
1799 break;
1800
1801 case WINED3DSPR_TEXCRDOUT:
1802 /* Vertex shaders >= 3.0 use general purpose output registers
1803 * (WINED3DSPR_OUTPUT), which can include an address token. */
1804 if (shader_version->major >= 3)
1805 shader_addline(buffer, "o");
1806 else
1807 shader_addline(buffer, "oT");
1808 break;
1809
1810 case WINED3DSPR_CONSTINT:
1811 shader_addline(buffer, "i");
1812 break;
1813
1814 case WINED3DSPR_CONSTBOOL:
1815 shader_addline(buffer, "b");
1816 break;
1817
1818 case WINED3DSPR_LABEL:
1819 shader_addline(buffer, "l");
1820 break;
1821
1822 case WINED3DSPR_LOOP:
1823 shader_addline(buffer, "aL");
1824 break;
1825
1826 case WINED3DSPR_SAMPLER:
1827 shader_addline(buffer, "s");
1828 break;
1829
1830 case WINED3DSPR_MISCTYPE:
1831 if (offset > 1)
1832 {
1833 FIXME("Unhandled misctype register %u.\n", offset);
1834 shader_addline(buffer, "<unhandled misctype %#x>", offset);
1835 }
1836 else
1837 {
1838 shader_addline(buffer, "%s", misctype_reg_names[offset]);
1839 }
1840 break;
1841
1842 case WINED3DSPR_PREDICATE:
1843 shader_addline(buffer, "p");
1844 break;
1845
1846 case WINED3DSPR_IMMCONST:
1847 shader_addline(buffer, "l");
1848 break;
1849
1850 case WINED3DSPR_CONSTBUFFER:
1851 shader_addline(buffer, "cb");
1852 break;
1853
1854 case WINED3DSPR_IMMCONSTBUFFER:
1855 shader_addline(buffer, "icb");
1856 break;
1857
1858 case WINED3DSPR_PRIMID:
1859 shader_addline(buffer, "primID");
1860 break;
1861
1862 case WINED3DSPR_NULL:
1863 shader_addline(buffer, "null");
1864 break;
1865
1866 case WINED3DSPR_RESOURCE:
1867 shader_addline(buffer, "t");
1868 break;
1869
1870 case WINED3DSPR_UAV:
1871 shader_addline(buffer, "u");
1872 break;
1873
1874 case WINED3DSPR_OUTPOINTID:
1875 shader_addline(buffer, "vOutputControlPointID");
1876 break;
1877
1878 case WINED3DSPR_FORKINSTID:
1879 shader_addline(buffer, "vForkInstanceId");
1880 break;
1881
1882 case WINED3DSPR_INCONTROLPOINT:
1883 shader_addline(buffer, "vicp");
1884 break;
1885
1886 case WINED3DSPR_OUTCONTROLPOINT:
1887 shader_addline(buffer, "vocp");
1888 break;
1889
1890 case WINED3DSPR_PATCHCONST:
1891 shader_addline(buffer, "vpc");
1892 break;
1893
1894 case WINED3DSPR_TESSCOORD:
1895 shader_addline(buffer, "vDomainLocation");
1896 break;
1897
1898 case WINED3DSPR_GROUPSHAREDMEM:
1899 shader_addline(buffer, "g");
1900 break;
1901
1902 case WINED3DSPR_THREADID:
1903 shader_addline(buffer, "vThreadID");
1904 break;
1905
1906 case WINED3DSPR_THREADGROUPID:
1907 shader_addline(buffer, "vThreadGroupID");
1908 break;
1909
1910 case WINED3DSPR_LOCALTHREADID:
1911 shader_addline(buffer, "vThreadIDInGroup");
1912 break;
1913
1914 case WINED3DSPR_LOCALTHREADINDEX:
1915 shader_addline(buffer, "vThreadIDInGroupFlattened");
1916 break;
1917
1918 case WINED3DSPR_IDXTEMP:
1919 shader_addline(buffer, "x");
1920 break;
1921
1922 case WINED3DSPR_STREAM:
1923 shader_addline(buffer, "m");
1924 break;
1925
1926 case WINED3DSPR_FUNCTIONBODY:
1927 shader_addline(buffer, "fb");
1928 break;
1929
1930 case WINED3DSPR_FUNCTIONPOINTER:
1931 shader_addline(buffer, "fp");
1932 break;
1933
1934 case WINED3DSPR_COVERAGE:
1935 shader_addline(buffer, "vCoverage");
1936 break;
1937
1938 case WINED3DSPR_SAMPLEMASK:
1939 shader_addline(buffer, "oMask");
1940 break;
1941
1942 default:
1943 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
1944 break;
1945 }
1946
1947 if (reg->type == WINED3DSPR_IMMCONST)
1948 {
1949 shader_addline(buffer, "(");
1950 switch (reg->immconst_type)
1951 {
1952 case WINED3D_IMMCONST_SCALAR:
1953 switch (reg->data_type)
1954 {
1955 case WINED3D_DATA_FLOAT:
1956 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
1957 break;
1958 case WINED3D_DATA_INT:
1959 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
1960 break;
1961 case WINED3D_DATA_RESOURCE:
1962 case WINED3D_DATA_SAMPLER:
1963 case WINED3D_DATA_UINT:
1964 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
1965 break;
1966 default:
1967 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1968 break;
1969 }
1970 break;
1971
1972 case WINED3D_IMMCONST_VEC4:
1973 switch (reg->data_type)
1974 {
1975 case WINED3D_DATA_FLOAT:
1976 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
1977 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
1978 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
1979 break;
1980 case WINED3D_DATA_INT:
1981 shader_addline(buffer, "%d, %d, %d, %d",
1982 reg->u.immconst_data[0], reg->u.immconst_data[1],
1983 reg->u.immconst_data[2], reg->u.immconst_data[3]);
1984 break;
1985 case WINED3D_DATA_RESOURCE:
1986 case WINED3D_DATA_SAMPLER:
1987 case WINED3D_DATA_UINT:
1988 shader_addline(buffer, "%u, %u, %u, %u",
1989 reg->u.immconst_data[0], reg->u.immconst_data[1],
1990 reg->u.immconst_data[2], reg->u.immconst_data[3]);
1991 break;
1992 default:
1993 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
1994 break;
1995 }
1996 break;
1997
1998 default:
1999 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2000 break;
2001 }
2002 shader_addline(buffer, ")");
2003 }
2004 else if (reg->type != WINED3DSPR_RASTOUT
2005 && reg->type != WINED3DSPR_MISCTYPE
2006 && reg->type != WINED3DSPR_NULL)
2007 {
2008 if (offset != ~0u)
2009 {
2010 shader_addline(buffer, "[");
2011 if (reg->idx[0].rel_addr)
2012 {
2013 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2014 shader_addline(buffer, " + ");
2015 }
2016 shader_addline(buffer, "%u]", offset);
2017
2018 if (reg->idx[1].offset != ~0u)
2019 {
2020 shader_addline(buffer, "[");
2021 if (reg->idx[1].rel_addr)
2022 {
2023 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2024 shader_addline(buffer, " + ");
2025 }
2026 shader_addline(buffer, "%u]", reg->idx[1].offset);
2027 }
2028 }
2029
2030 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2031 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2032 }
2033 }
2034
2035 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2036 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2037 {
2038 DWORD write_mask = param->write_mask;
2039
2040 shader_dump_register(buffer, &param->reg, shader_version);
2041
2042 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2043 {
2044 static const char write_mask_chars[] = "xyzw";
2045
2046 shader_addline(buffer, ".");
2047 if (write_mask & WINED3DSP_WRITEMASK_0)
2048 shader_addline(buffer, "%c", write_mask_chars[0]);
2049 if (write_mask & WINED3DSP_WRITEMASK_1)
2050 shader_addline(buffer, "%c", write_mask_chars[1]);
2051 if (write_mask & WINED3DSP_WRITEMASK_2)
2052 shader_addline(buffer, "%c", write_mask_chars[2]);
2053 if (write_mask & WINED3DSP_WRITEMASK_3)
2054 shader_addline(buffer, "%c", write_mask_chars[3]);
2055 }
2056 }
2057
2058 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2059 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2060 {
2061 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2062 DWORD swizzle = param->swizzle;
2063
2064 if (src_modifier == WINED3DSPSM_NEG
2065 || src_modifier == WINED3DSPSM_BIASNEG
2066 || src_modifier == WINED3DSPSM_SIGNNEG
2067 || src_modifier == WINED3DSPSM_X2NEG
2068 || src_modifier == WINED3DSPSM_ABSNEG)
2069 shader_addline(buffer, "-");
2070 else if (src_modifier == WINED3DSPSM_COMP)
2071 shader_addline(buffer, "1-");
2072 else if (src_modifier == WINED3DSPSM_NOT)
2073 shader_addline(buffer, "!");
2074
2075 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2076 shader_addline(buffer, "abs(");
2077
2078 shader_dump_register(buffer, &param->reg, shader_version);
2079
2080 switch (src_modifier)
2081 {
2082 case WINED3DSPSM_NONE: break;
2083 case WINED3DSPSM_NEG: break;
2084 case WINED3DSPSM_NOT: break;
2085 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2086 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2087 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2088 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2089 case WINED3DSPSM_COMP: break;
2090 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2091 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2092 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2093 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2094 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2095 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2096 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2097 }
2098
2099 if (swizzle != WINED3DSP_NOSWIZZLE)
2100 {
2101 static const char swizzle_chars[] = "xyzw";
2102 DWORD swizzle_x = swizzle & 0x03;
2103 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2104 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2105 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2106
2107 if (swizzle_x == swizzle_y
2108 && swizzle_x == swizzle_z
2109 && swizzle_x == swizzle_w)
2110 {
2111 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2112 }
2113 else
2114 {
2115 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2116 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2117 }
2118 }
2119 }
2120
2121 /* Shared code in order to generate the bulk of the shader string.
2122 * NOTE: A description of how to parse tokens can be found on MSDN. */
2123 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2124 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
2125 {
2126 struct wined3d_device *device = shader->device;
2127 const struct wined3d_shader_frontend *fe = shader->frontend;
2128 void *fe_data = shader->frontend_data;
2129 struct wined3d_shader_version shader_version;
2130 struct wined3d_shader_parser_state state;
2131 struct wined3d_shader_instruction ins;
2132 struct wined3d_shader_tex_mx tex_mx;
2133 struct wined3d_shader_context ctx;
2134 const DWORD *ptr = byte_code;
2135
2136 /* Initialize current parsing state. */
2137 tex_mx.current_row = 0;
2138 state.current_loop_depth = 0;
2139 state.current_loop_reg = 0;
2140 state.in_subroutine = FALSE;
2141
2142 ctx.shader = shader;
2143 ctx.gl_info = &device->adapter->gl_info;
2144 ctx.reg_maps = reg_maps;
2145 ctx.buffer = buffer;
2146 ctx.tex_mx = &tex_mx;
2147 ctx.state = &state;
2148 ctx.backend_data = backend_ctx;
2149 ins.ctx = &ctx;
2150
2151 fe->shader_read_header(fe_data, &ptr, &shader_version);
2152
2153 while (!fe->shader_is_end(fe_data, &ptr))
2154 {
2155 /* Read opcode. */
2156 fe->shader_read_instruction(fe_data, &ptr, &ins);
2157
2158 /* Unknown opcode and its parameters. */
2159 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2160 {
2161 TRACE("Skipping unrecognized instruction.\n");
2162 continue;
2163 }
2164
2165 if (ins.predicate)
2166 FIXME("Predicates not implemented.\n");
2167
2168 /* Call appropriate function for output target */
2169 device->shader_backend->shader_handle_instruction(&ins);
2170 }
2171 }
2172
2173 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2174 const struct wined3d_shader_dst_param *dst)
2175 {
2176 DWORD mmask = dst->modifiers;
2177
2178 switch (dst->shift)
2179 {
2180 case 0: break;
2181 case 13: shader_addline(buffer, "_d8"); break;
2182 case 14: shader_addline(buffer, "_d4"); break;
2183 case 15: shader_addline(buffer, "_d2"); break;
2184 case 1: shader_addline(buffer, "_x2"); break;
2185 case 2: shader_addline(buffer, "_x4"); break;
2186 case 3: shader_addline(buffer, "_x8"); break;
2187 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2188 }
2189
2190 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2191 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2192 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2193
2194 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2195 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2196 }
2197
2198 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2199 enum wined3d_primitive_type primitive_type)
2200 {
2201 switch (primitive_type)
2202 {
2203 case WINED3D_PT_UNDEFINED:
2204 shader_addline(buffer, "undefined");
2205 break;
2206 case WINED3D_PT_POINTLIST:
2207 shader_addline(buffer, "pointlist");
2208 break;
2209 case WINED3D_PT_LINELIST:
2210 shader_addline(buffer, "linelist");
2211 break;
2212 case WINED3D_PT_LINESTRIP:
2213 shader_addline(buffer, "linestrip");
2214 break;
2215 case WINED3D_PT_TRIANGLELIST:
2216 shader_addline(buffer, "trianglelist");
2217 break;
2218 case WINED3D_PT_TRIANGLESTRIP:
2219 shader_addline(buffer, "trianglestrip");
2220 break;
2221 case WINED3D_PT_TRIANGLEFAN:
2222 shader_addline(buffer, "trianglefan");
2223 break;
2224 case WINED3D_PT_LINELIST_ADJ:
2225 shader_addline(buffer, "linelist_adj");
2226 break;
2227 case WINED3D_PT_LINESTRIP_ADJ:
2228 shader_addline(buffer, "linestrip_adj");
2229 break;
2230 case WINED3D_PT_TRIANGLELIST_ADJ:
2231 shader_addline(buffer, "trianglelist_adj");
2232 break;
2233 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2234 shader_addline(buffer, "trianglestrip_adj");
2235 break;
2236 default:
2237 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type);
2238 break;
2239 }
2240 }
2241
2242 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2243 enum wined3d_shader_interpolation_mode interpolation_mode)
2244 {
2245 switch (interpolation_mode)
2246 {
2247 case WINED3DSIM_CONSTANT:
2248 shader_addline(buffer, "constant");
2249 break;
2250 case WINED3DSIM_LINEAR:
2251 shader_addline(buffer, "linear");
2252 break;
2253 case WINED3DSIM_LINEAR_CENTROID:
2254 shader_addline(buffer, "linear centroid");
2255 break;
2256 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2257 shader_addline(buffer, "linear noperspective");
2258 break;
2259 case WINED3DSIM_LINEAR_SAMPLE:
2260 shader_addline(buffer, "linear sample");
2261 break;
2262 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2263 shader_addline(buffer, "linear noperspective centroid");
2264 break;
2265 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2266 shader_addline(buffer, "linear noperspective sample");
2267 break;
2268 default:
2269 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2270 break;
2271 }
2272 }
2273
2274 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
2275 {
2276 struct wined3d_shader_version shader_version;
2277 struct wined3d_string_buffer buffer;
2278 const DWORD *ptr = byte_code;
2279 const char *type_prefix;
2280 const char *p, *q;
2281 DWORD i;
2282
2283 if (!string_buffer_init(&buffer))
2284 {
2285 ERR("Failed to initialize string buffer.\n");
2286 return;
2287 }
2288
2289 TRACE("Parsing %p.\n", byte_code);
2290
2291 fe->shader_read_header(fe_data, &ptr, &shader_version);
2292
2293 switch (shader_version.type)
2294 {
2295 case WINED3D_SHADER_TYPE_VERTEX:
2296 type_prefix = "vs";
2297 break;
2298
2299 case WINED3D_SHADER_TYPE_HULL:
2300 type_prefix = "hs";
2301 break;
2302
2303 case WINED3D_SHADER_TYPE_DOMAIN:
2304 type_prefix = "ds";
2305 break;
2306
2307 case WINED3D_SHADER_TYPE_GEOMETRY:
2308 type_prefix = "gs";
2309 break;
2310
2311 case WINED3D_SHADER_TYPE_PIXEL:
2312 type_prefix = "ps";
2313 break;
2314
2315 case WINED3D_SHADER_TYPE_COMPUTE:
2316 type_prefix = "cs";
2317 break;
2318
2319 default:
2320 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2321 type_prefix = "unknown";
2322 break;
2323 }
2324
2325 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2326
2327 while (!fe->shader_is_end(fe_data, &ptr))
2328 {
2329 struct wined3d_shader_instruction ins;
2330
2331 fe->shader_read_instruction(fe_data, &ptr, &ins);
2332 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2333 {
2334 WARN("Skipping unrecognized instruction.\n");
2335 shader_addline(&buffer, "<unrecognized instruction>\n");
2336 continue;
2337 }
2338
2339 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2340 {
2341 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, &shader_version);
2342 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2343 shader_addline(&buffer, " ");
2344 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2345 }
2346 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2347 {
2348 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2349 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2350 shader_addline(&buffer, ", %s",
2351 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2352 }
2353 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2354 {
2355 shader_addline(&buffer, "%s fb%u",
2356 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2357 }
2358 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2359 {
2360 shader_addline(&buffer, "%s ft%u = {...}",
2361 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2362 }
2363 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2364 {
2365 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2366 shader_dump_global_flags(&buffer, ins.flags);
2367 }
2368 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2369 {
2370 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2371 ins.declaration.max_tessellation_factor);
2372 }
2373 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2374 {
2375 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2376 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2377 {
2378 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2379 ins.declaration.icb->data[4 * i + 0],
2380 ins.declaration.icb->data[4 * i + 1],
2381 ins.declaration.icb->data[4 * i + 2],
2382 ins.declaration.icb->data[4 * i + 3]);
2383 }
2384 shader_addline(&buffer, "}");
2385 }
2386 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2387 {
2388 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2389 ins.declaration.indexable_temp.register_idx,
2390 ins.declaration.indexable_temp.register_size,
2391 ins.declaration.indexable_temp.component_count);
2392 }
2393 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2394 {
2395 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2396 shader_dump_interpolation_mode(&buffer, ins.flags);
2397 shader_addline(&buffer, " ");
2398 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2399 }
2400 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2401 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2402 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2403 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2404 {
2405 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2406 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2407 shader_addline(&buffer, ", ");
2408 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2409 }
2410 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2411 {
2412 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2413 shader_dump_interpolation_mode(&buffer, ins.flags);
2414 shader_addline(&buffer, " ");
2415 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2416 shader_addline(&buffer, ", ");
2417 shader_dump_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2418 }
2419 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2420 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2421 {
2422 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2423 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2424 }
2425 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2426 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2427 {
2428 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2429 shader_dump_primitive_type(&buffer, ins.declaration.primitive_type);
2430 }
2431 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2432 {
2433 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2434 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2435 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2436 }
2437 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2438 {
2439 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2440 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2441 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2442 }
2443 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2444 {
2445 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2446 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2447 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2448 shader_addline(&buffer, ", comparisonMode");
2449 }
2450 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2451 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT
2452 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2453 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2454 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
2455 {
2456 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2457 }
2458 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2459 {
2460 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2461 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2462 }
2463 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2464 {
2465 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2466 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2467 }
2468 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2469 {
2470 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2471 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2472 }
2473 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2474 {
2475 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2476 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2477 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2478 }
2479 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2480 {
2481 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2482 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2483 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2484 ins.declaration.tgsm_structured.structure_count);
2485 }
2486 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2487 {
2488 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2489 ins.declaration.thread_group_size.x,
2490 ins.declaration.thread_group_size.y,
2491 ins.declaration.thread_group_size.z);
2492 }
2493 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2494 {
2495 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2496 shader_dump_uav_flags(&buffer, ins.flags);
2497 shader_addline(&buffer, " ");
2498 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2499 }
2500 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2501 {
2502 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2503 shader_dump_uav_flags(&buffer, ins.flags);
2504 shader_addline(&buffer, " ");
2505 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2506 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2507 }
2508 else if (ins.handler_idx == WINED3DSIH_DEF)
2509 {
2510 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2511 ins.dst[0].reg.idx[0].offset),
2512 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2513 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2514 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2515 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2516 }
2517 else if (ins.handler_idx == WINED3DSIH_DEFI)
2518 {
2519 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2520 ins.src[0].reg.u.immconst_data[0],
2521 ins.src[0].reg.u.immconst_data[1],
2522 ins.src[0].reg.u.immconst_data[2],
2523 ins.src[0].reg.u.immconst_data[3]);
2524 }
2525 else if (ins.handler_idx == WINED3DSIH_DEFB)
2526 {
2527 shader_addline(&buffer, "defb b%u = %s",
2528 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2529 }
2530 else
2531 {
2532 if (ins.predicate)
2533 {
2534 shader_addline(&buffer, "(");
2535 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2536 shader_addline(&buffer, ") ");
2537 }
2538
2539 /* PixWin marks instructions with the coissue flag with a '+' */
2540 if (ins.coissue)
2541 shader_addline(&buffer, "+");
2542
2543 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2544
2545 if (ins.handler_idx == WINED3DSIH_BREAKP
2546 || ins.handler_idx == WINED3DSIH_IF)
2547 {
2548 switch (ins.flags)
2549 {
2550 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2551 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2552 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2553 }
2554 }
2555 else if (ins.handler_idx == WINED3DSIH_IFC
2556 || ins.handler_idx == WINED3DSIH_BREAKC)
2557 {
2558 switch (ins.flags)
2559 {
2560 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2561 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2562 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2563 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2564 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2565 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2566 default: shader_addline(&buffer, "_(%u)", ins.flags);
2567 }
2568 }
2569 else if (ins.handler_idx == WINED3DSIH_TEX
2570 && shader_version.major >= 2
2571 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2572 {
2573 shader_addline(&buffer, "p");
2574 }
2575 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2576 {
2577 switch (ins.flags)
2578 {
2579 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2580 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2581 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2582 }
2583 }
2584 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2585 {
2586 switch (ins.flags)
2587 {
2588 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2589 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2590 }
2591 }
2592 else if (ins.handler_idx == WINED3DSIH_SYNC)
2593 {
2594 shader_dump_sync_flags(&buffer, ins.flags);
2595 }
2596
2597 if (wined3d_shader_instruction_has_texel_offset(&ins))
2598 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2599
2600 for (i = 0; i < ins.dst_count; ++i)
2601 {
2602 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2603 shader_addline(&buffer, !i ? " " : ", ");
2604 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2605 }
2606
2607 /* Other source tokens */
2608 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2609 {
2610 shader_addline(&buffer, !i ? " " : ", ");
2611 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2612 }
2613 }
2614 shader_addline(&buffer, "\n");
2615 }
2616
2617 for (p = buffer.buffer; *p; p = q)
2618 {
2619 if (!(q = strstr(p, "\n")))
2620 q = p + strlen(p);
2621 else
2622 ++q;
2623 TRACE(" %.*s", (int)(q - p), p);
2624 }
2625
2626 string_buffer_free(&buffer);
2627 }
2628
2629 static void shader_cleanup(struct wined3d_shader *shader)
2630 {
2631 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2632 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2633 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2634 shader->device->shader_backend->shader_destroy(shader);
2635 shader_cleanup_reg_maps(&shader->reg_maps);
2636 HeapFree(GetProcessHeap(), 0, shader->function);
2637 shader_delete_constant_list(&shader->constantsF);
2638 shader_delete_constant_list(&shader->constantsB);
2639 shader_delete_constant_list(&shader->constantsI);
2640 list_remove(&shader->shader_list_entry);
2641
2642 if (shader->frontend && shader->frontend_data)
2643 shader->frontend->shader_free(shader->frontend_data);
2644 }
2645
2646 struct shader_none_priv
2647 {
2648 const struct wined3d_vertex_pipe_ops *vertex_pipe;
2649 const struct fragment_pipeline *fragment_pipe;
2650 BOOL ffp_proj_control;
2651 };
2652
2653 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
2654 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
2655 enum wined3d_gl_resource_type tex_type, const SIZE *ds_mask_size) {}
2656 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
2657 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
2658 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
2659 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
2660 const struct wined3d_state *state) {}
2661 static void shader_none_destroy(struct wined3d_shader *shader) {}
2662 static void shader_none_free_context_data(struct wined3d_context *context) {}
2663 static void shader_none_init_context_state(struct wined3d_context *context) {}
2664
2665 /* Context activation is done by the caller. */
2666 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
2667 const struct wined3d_state *state)
2668 {
2669 const struct wined3d_gl_info *gl_info = context->gl_info;
2670 struct shader_none_priv *priv = shader_priv;
2671
2672 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
2673 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
2674 }
2675
2676 /* Context activation is done by the caller. */
2677 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
2678 {
2679 struct shader_none_priv *priv = shader_priv;
2680 const struct wined3d_gl_info *gl_info = context->gl_info;
2681
2682 priv->vertex_pipe->vp_enable(gl_info, FALSE);
2683 priv->fragment_pipe->enable_extension(gl_info, FALSE);
2684
2685 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
2686 | (1u << WINED3D_SHADER_TYPE_VERTEX)
2687 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
2688 | (1u << WINED3D_SHADER_TYPE_HULL)
2689 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
2690 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
2691 }
2692
2693 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
2694 const struct fragment_pipeline *fragment_pipe)
2695 {
2696 struct fragment_caps fragment_caps;
2697 void *vertex_priv, *fragment_priv;
2698 struct shader_none_priv *priv;
2699
2700 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
2701 return E_OUTOFMEMORY;
2702
2703 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
2704 {
2705 ERR("Failed to initialize vertex pipe.\n");
2706 HeapFree(GetProcessHeap(), 0, priv);
2707 return E_FAIL;
2708 }
2709
2710 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2711 {
2712 ERR("Failed to initialize fragment pipe.\n");
2713 vertex_pipe->vp_free(device);
2714 HeapFree(GetProcessHeap(), 0, priv);
2715 return E_FAIL;
2716 }
2717
2718 priv->vertex_pipe = vertex_pipe;
2719 priv->fragment_pipe = fragment_pipe;
2720 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
2721 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
2722
2723 device->vertex_priv = vertex_priv;
2724 device->fragment_priv = fragment_priv;
2725 device->shader_priv = priv;
2726
2727 return WINED3D_OK;
2728 }
2729
2730 static void shader_none_free(struct wined3d_device *device)
2731 {
2732 struct shader_none_priv *priv = device->shader_priv;
2733
2734 priv->fragment_pipe->free_private(device);
2735 priv->vertex_pipe->vp_free(device);
2736 HeapFree(GetProcessHeap(), 0, priv);
2737 }
2738
2739 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2740 {
2741 return TRUE;
2742 }
2743
2744 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
2745 {
2746 /* Set the shader caps to 0 for the none shader backend */
2747 caps->vs_version = 0;
2748 caps->hs_version = 0;
2749 caps->ds_version = 0;
2750 caps->gs_version = 0;
2751 caps->ps_version = 0;
2752 caps->cs_version = 0;
2753 caps->vs_uniform_count = 0;
2754 caps->ps_uniform_count = 0;
2755 caps->ps_1x_max_value = 0.0f;
2756 caps->varying_count = 0;
2757 caps->wined3d_caps = 0;
2758 }
2759
2760 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2761 {
2762 /* We "support" every possible fixup, since we don't support any shader
2763 * model, and will never have to actually sample a texture. */
2764 return TRUE;
2765 }
2766
2767 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2768 {
2769 struct shader_none_priv *priv = shader_priv;
2770
2771 return priv->ffp_proj_control;
2772 }
2773
2774 const struct wined3d_shader_backend_ops none_shader_backend =
2775 {
2776 shader_none_handle_instruction,
2777 shader_none_select,
2778 shader_none_disable,
2779 shader_none_select_depth_blt,
2780 shader_none_deselect_depth_blt,
2781 shader_none_update_float_vertex_constants,
2782 shader_none_update_float_pixel_constants,
2783 shader_none_load_constants,
2784 shader_none_destroy,
2785 shader_none_alloc,
2786 shader_none_free,
2787 shader_none_allocate_context_data,
2788 shader_none_free_context_data,
2789 shader_none_init_context_state,
2790 shader_none_get_caps,
2791 shader_none_color_fixup_supported,
2792 shader_none_has_ffp_proj_control,
2793 };
2794
2795 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
2796 DWORD float_const_count, enum wined3d_shader_type type, unsigned int max_version)
2797 {
2798 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2799 const struct wined3d_shader_frontend *fe;
2800 HRESULT hr;
2801 unsigned int backend_version;
2802 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2803
2804 TRACE("shader %p, byte_code %p, float_const_count %u, type %#x, max_version %u.\n",
2805 shader, byte_code, float_const_count, type, max_version);
2806
2807 list_init(&shader->constantsF);
2808 list_init(&shader->constantsB);
2809 list_init(&shader->constantsI);
2810 shader->lconst_inf_or_nan = FALSE;
2811 list_init(&reg_maps->indexable_temps);
2812
2813 fe = shader_select_frontend(*byte_code);
2814 if (!fe)
2815 {
2816 FIXME("Unable to find frontend for shader.\n");
2817 return WINED3DERR_INVALIDCALL;
2818 }
2819 shader->frontend = fe;
2820 shader->frontend_data = fe->shader_init(byte_code, &shader->output_signature);
2821 if (!shader->frontend_data)
2822 {
2823 FIXME("Failed to initialize frontend.\n");
2824 return WINED3DERR_INVALIDCALL;
2825 }
2826
2827 /* First pass: trace shader. */
2828 if (TRACE_ON(d3d_shader))
2829 shader_trace_init(fe, shader->frontend_data, byte_code);
2830
2831 /* Second pass: figure out which registers are used, what the semantics are, etc. */
2832 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
2833 &shader->output_signature, byte_code, float_const_count)))
2834 return hr;
2835
2836 if (reg_maps->shader_version.type != type)
2837 {
2838 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
2839 return WINED3DERR_INVALIDCALL;
2840 }
2841 if (reg_maps->shader_version.major > max_version)
2842 {
2843 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
2844 return WINED3DERR_INVALIDCALL;
2845 }
2846 switch (type)
2847 {
2848 case WINED3D_SHADER_TYPE_VERTEX:
2849 backend_version = d3d_info->limits.vs_version;
2850 break;
2851 case WINED3D_SHADER_TYPE_HULL:
2852 backend_version = d3d_info->limits.hs_version;
2853 break;
2854 case WINED3D_SHADER_TYPE_DOMAIN:
2855 backend_version = d3d_info->limits.ds_version;
2856 break;
2857 case WINED3D_SHADER_TYPE_GEOMETRY:
2858 backend_version = d3d_info->limits.gs_version;
2859 break;
2860 case WINED3D_SHADER_TYPE_PIXEL:
2861 backend_version = d3d_info->limits.ps_version;
2862 break;
2863 case WINED3D_SHADER_TYPE_COMPUTE:
2864 backend_version = d3d_info->limits.cs_version;
2865 break;
2866 default:
2867 FIXME("No backend version-checking for this shader type.\n");
2868 backend_version = 0;
2869 }
2870 if (reg_maps->shader_version.major > backend_version)
2871 {
2872 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
2873 reg_maps->shader_version.major, reg_maps->shader_version.minor);
2874 return WINED3DERR_INVALIDCALL;
2875 }
2876
2877 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength)))
2878 return E_OUTOFMEMORY;
2879 memcpy(shader->function, byte_code, shader->functionLength);
2880
2881 return WINED3D_OK;
2882 }
2883
2884 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2885 {
2886 ULONG refcount = InterlockedIncrement(&shader->ref);
2887
2888 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2889
2890 return refcount;
2891 }
2892
2893 static void wined3d_shader_destroy_object(void *object)
2894 {
2895 shader_cleanup(object);
2896 HeapFree(GetProcessHeap(), 0, object);
2897 }
2898
2899 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2900 {
2901 ULONG refcount = InterlockedDecrement(&shader->ref);
2902
2903 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2904
2905 if (!refcount)
2906 {
2907 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2908 wined3d_cs_emit_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2909 }
2910
2911 return refcount;
2912 }
2913
2914 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2915 {
2916 TRACE("shader %p.\n", shader);
2917
2918 return shader->parent;
2919 }
2920
2921 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2922 void *byte_code, UINT *byte_code_size)
2923 {
2924 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2925
2926 if (!byte_code)
2927 {
2928 *byte_code_size = shader->functionLength;
2929 return WINED3D_OK;
2930 }
2931
2932 if (*byte_code_size < shader->functionLength)
2933 {
2934 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2935 * than the required size we should write the required size and
2936 * return D3DERR_MOREDATA. That's not actually true. */
2937 return WINED3DERR_INVALIDCALL;
2938 }
2939
2940 memcpy(byte_code, shader->function, shader->functionLength);
2941
2942 return WINED3D_OK;
2943 }
2944
2945 /* Set local constants for d3d8 shaders. */
2946 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2947 UINT start_idx, const float *src_data, UINT count)
2948 {
2949 UINT end_idx = start_idx + count;
2950 UINT i;
2951
2952 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2953
2954 if (end_idx > shader->limits->constant_float)
2955 {
2956 WARN("end_idx %u > float constants limit %u.\n",
2957 end_idx, shader->limits->constant_float);
2958 end_idx = shader->limits->constant_float;
2959 }
2960
2961 for (i = start_idx; i < end_idx; ++i)
2962 {
2963 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
2964 float *value;
2965 if (!lconst)
2966 return E_OUTOFMEMORY;
2967
2968 lconst->idx = i;
2969 value = (float *)lconst->value;
2970 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2971 list_add_head(&shader->constantsF, &lconst->entry);
2972
2973 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2974 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2975 {
2976 shader->lconst_inf_or_nan = TRUE;
2977 }
2978 }
2979
2980 return WINED3D_OK;
2981 }
2982
2983 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2984 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
2985 {
2986 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2987 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2988 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2989 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2990 args->point_size = state->gl_primitive_type == GL_POINTS;
2991 args->per_vertex_point_size = shader->reg_maps.point_size;
2992 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
2993 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
2994 : WINED3D_SHADER_TYPE_PIXEL;
2995 if (shader->reg_maps.shader_version.major >= 4)
2996 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
2997 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
2998 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
2999 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
3000 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
3001 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3002 else
3003 args->next_shader_input_count = 0;
3004 args->swizzle_map = swizzle_map;
3005 if (d3d_info->emulated_flatshading)
3006 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3007 else
3008 args->flatshading = 0;
3009 }
3010
3011 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3012 {
3013 if (usage_idx1 != usage_idx2)
3014 return FALSE;
3015 if (usage1 == usage2)
3016 return TRUE;
3017 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3018 return TRUE;
3019 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3020 return TRUE;
3021
3022 return FALSE;
3023 }
3024
3025 BOOL vshader_get_input(const struct wined3d_shader *shader,
3026 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3027 {
3028 WORD map = shader->reg_maps.input_registers;
3029 unsigned int i;
3030
3031 for (i = 0; map; map >>= 1, ++i)
3032 {
3033 if (!(map & 1)) continue;
3034
3035 if (match_usage(shader->u.vs.attributes[i].usage,
3036 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3037 {
3038 *regnum = i;
3039 return TRUE;
3040 }
3041 }
3042 return FALSE;
3043 }
3044
3045 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3046 const struct wined3d_shader_signature *src, char **signature_strings)
3047 {
3048 struct wined3d_shader_signature_element *e;
3049 unsigned int i;
3050 SIZE_T len;
3051 char *ptr;
3052
3053 if (!src->element_count)
3054 return WINED3D_OK;
3055
3056 ptr = *signature_strings;
3057
3058 dst->element_count = src->element_count;
3059 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3060 return E_OUTOFMEMORY;
3061
3062 for (i = 0; i < src->element_count; ++i)
3063 {
3064 e = &src->elements[i];
3065 dst->elements[i] = *e;
3066
3067 len = strlen(e->semantic_name);
3068 memcpy(ptr, e->semantic_name, len + 1);
3069 dst->elements[i].semantic_name = ptr;
3070 ptr += len + 1;
3071 }
3072
3073 *signature_strings = ptr;
3074
3075 return WINED3D_OK;
3076 }
3077
3078 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3079 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3080 void *parent, const struct wined3d_parent_ops *parent_ops)
3081 {
3082 struct wined3d_shader_signature_element *e;
3083 SIZE_T total, len;
3084 unsigned int i;
3085 HRESULT hr;
3086 char *ptr;
3087
3088 if (!desc->byte_code)
3089 return WINED3DERR_INVALIDCALL;
3090
3091 shader->ref = 1;
3092 shader->device = device;
3093 shader->parent = parent;
3094 shader->parent_ops = parent_ops;
3095
3096 total = 0;
3097 for (i = 0; i < desc->input_signature.element_count; ++i)
3098 {
3099 e = &desc->input_signature.elements[i];
3100 len = strlen(e->semantic_name);
3101 if (len >= ~(SIZE_T)0 - total)
3102 return E_OUTOFMEMORY;
3103
3104 total += len + 1;
3105 }
3106 for (i = 0; i < desc->output_signature.element_count; ++i)
3107 {
3108 e = &desc->output_signature.elements[i];
3109 len = strlen(e->semantic_name);
3110 if (len >= ~(SIZE_T)0 - total)
3111 return E_OUTOFMEMORY;
3112
3113 total += len + 1;
3114 }
3115 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3116 return E_OUTOFMEMORY;
3117 ptr = shader->signature_strings;
3118
3119 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3120 {
3121 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3122 return hr;
3123 }
3124 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3125 {
3126 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3127 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3128 return hr;
3129 }
3130
3131 list_init(&shader->linked_programs);
3132 list_add_head(&device->shaders, &shader->shader_list_entry);
3133
3134 if (FAILED(hr = shader_set_function(shader, desc->byte_code,
3135 float_const_count, type, desc->max_version)))
3136 {
3137 WARN("Failed to set function, hr %#x.\n", hr);
3138 shader_cleanup(shader);
3139 }
3140
3141 shader->load_local_constsF = shader->lconst_inf_or_nan;
3142
3143 return hr;
3144 }
3145
3146 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3147 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3148 {
3149 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3150 unsigned int i;
3151 HRESULT hr;
3152
3153 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3154 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3155 return hr;
3156
3157 for (i = 0; i < shader->input_signature.element_count; ++i)
3158 {
3159 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3160
3161 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3162 continue;
3163
3164 shader->u.vs.attributes[input->register_idx].usage =
3165 shader_usage_from_semantic_name(input->semantic_name);
3166 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3167 }
3168
3169 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3170 shader->load_local_constsF = TRUE;
3171
3172 return WINED3D_OK;
3173 }
3174
3175 static HRESULT domain_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 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3179 }
3180
3181 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3182 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3183 {
3184 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3185 }
3186
3187 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3188 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3189 {
3190 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops);
3191 }
3192
3193 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3194 struct gs_compile_args *args)
3195 {
3196 args->ps_input_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3197 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3198 }
3199
3200 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3201 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3202 {
3203 const struct wined3d_gl_info *gl_info = context->gl_info;
3204 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3205 const struct wined3d_texture *texture;
3206 UINT i;
3207
3208 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3209 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3210 {
3211 static unsigned int warned = 0;
3212
3213 args->srgb_correction = 1;
3214 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3215 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3216 "support, expect rendering artifacts.\n");
3217 }
3218
3219 if (shader->reg_maps.shader_version.major == 1
3220 && shader->reg_maps.shader_version.minor <= 3)
3221 {
3222 for (i = 0; i < shader->limits->sampler; ++i)
3223 {
3224 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3225
3226 if (flags & WINED3D_TTFF_PROJECTED)
3227 {
3228 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3229
3230 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3231 {
3232 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3233 unsigned int j;
3234 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3235 DWORD max_valid = WINED3D_TTFF_COUNT4;
3236
3237 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3238 {
3239 struct wined3d_vertex_declaration_element *element =
3240 &state->vertex_declaration->elements[j];
3241
3242 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3243 && element->usage_idx == index)
3244 {
3245 max_valid = element->format->component_count;
3246 break;
3247 }
3248 }
3249 if (!tex_transform || tex_transform > max_valid)
3250 {
3251 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3252 tex_transform, max_valid);
3253 tex_transform = max_valid;
3254 }
3255 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3256 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3257 && tex_transform > WINED3D_TTFF_COUNT2)
3258 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3259 && tex_transform > WINED3D_TTFF_COUNT3))
3260 tex_transform |= WINED3D_PSARGS_PROJECTED;
3261 else
3262 {
3263 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3264 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3265 i, tex_transform, resource_type);
3266 }
3267 }
3268 else
3269 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3270
3271 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3272 }
3273 }
3274 }
3275 if (shader->reg_maps.shader_version.major == 1
3276 && shader->reg_maps.shader_version.minor <= 4)
3277 {
3278 for (i = 0; i < shader->limits->sampler; ++i)
3279 {
3280 const struct wined3d_texture *texture = state->textures[i];
3281
3282 if (!shader->reg_maps.resource_info[i].type)
3283 continue;
3284
3285 /* Treat unbound textures as 2D. The dummy texture will provide
3286 * the proper sample value. The tex_types bitmap defaults to
3287 * 2D because of the memset. */
3288 if (!texture)
3289 continue;
3290
3291 switch (texture->target)
3292 {
3293 /* RECT textures are distinguished from 2D textures via np2_fixup */
3294 case GL_TEXTURE_RECTANGLE_ARB:
3295 case GL_TEXTURE_2D:
3296 break;
3297
3298 case GL_TEXTURE_3D:
3299 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3300 break;
3301
3302 case GL_TEXTURE_CUBE_MAP_ARB:
3303 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3304 break;
3305 }
3306 }
3307 }
3308
3309 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3310 {
3311 if (!shader->reg_maps.resource_info[i].type)
3312 continue;
3313
3314 texture = state->textures[i];
3315 if (!texture)
3316 {
3317 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3318 continue;
3319 }
3320 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3321 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3322 else
3323 args->color_fixup[i] = texture->resource.format->color_fixup;
3324
3325 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3326 args->shadow |= 1u << i;
3327
3328 /* Flag samplers that need NP2 texcoord fixup. */
3329 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3330 args->np2_fixup |= (1u << i);
3331 }
3332
3333 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3334 if (shader->reg_maps.shader_version.major >= 4)
3335 args->shadow = 0;
3336
3337 if (shader->reg_maps.shader_version.major >= 3)
3338 {
3339 if (position_transformed)
3340 args->vp_mode = pretransformed;
3341 else if (use_vs(state))
3342 args->vp_mode = vertexshader;
3343 else
3344 args->vp_mode = fixedfunction;
3345 args->fog = WINED3D_FFP_PS_FOG_OFF;
3346 }
3347 else
3348 {
3349 args->vp_mode = vertexshader;
3350 if (state->render_states[WINED3D_RS_FOGENABLE])
3351 {
3352 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3353 {
3354 case WINED3D_FOG_NONE:
3355 if (position_transformed || use_vs(state))
3356 {
3357 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3358 break;
3359 }
3360
3361 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3362 {
3363 case WINED3D_FOG_NONE: /* Fall through. */
3364 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3365 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3366 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3367 }
3368 break;
3369
3370 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3371 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3372 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3373 }
3374 }
3375 else
3376 {
3377 args->fog = WINED3D_FFP_PS_FOG_OFF;
3378 }
3379 }
3380
3381 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3382 {
3383 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3384
3385 args->texcoords_initialized = 0;
3386 for (i = 0; i < MAX_TEXTURES; ++i)
3387 {
3388 if (vs)
3389 {
3390 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3391 args->texcoords_initialized |= 1u << i;
3392 }
3393 else
3394 {
3395 const struct wined3d_stream_info *si = &context->stream_info;
3396 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3397
3398 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3399 & WINED3D_FFP_TCI_MASK
3400 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3401 args->texcoords_initialized |= 1u << i;
3402 }
3403 }
3404 }
3405 else
3406 {
3407 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3408 }
3409
3410 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3411 && state->gl_primitive_type == GL_POINTS;
3412
3413 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3414 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3415 else
3416 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3417 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3418 : WINED3D_CMP_ALWAYS) - 1;
3419
3420 if (d3d_info->emulated_flatshading)
3421 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3422
3423 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3424 ? context->render_offscreen : 0;
3425 }
3426
3427 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3428 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3429 {
3430 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3431 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3432 HRESULT hr;
3433
3434 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3435 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3436 return hr;
3437
3438 for (i = 0; i < MAX_REG_INPUT; ++i)
3439 {
3440 if (shader->u.ps.input_reg_used[i])
3441 {
3442 ++num_regs_used;
3443 highest_reg_used = i;
3444 }
3445 }
3446
3447 /* Don't do any register mapping magic if it is not needed, or if we can't
3448 * achieve anything anyway */
3449 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3450 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3451 || shader->reg_maps.shader_version.major >= 4)
3452 {
3453 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3454 {
3455 /* This happens with relative addressing. The input mapper function
3456 * warns about this if the higher registers are declared too, so
3457 * don't write a FIXME here */
3458 WARN("More varying registers used than supported\n");
3459 }
3460
3461 for (i = 0; i < MAX_REG_INPUT; ++i)
3462 {
3463 shader->u.ps.input_reg_map[i] = i;
3464 }
3465
3466 shader->u.ps.declared_in_count = highest_reg_used + 1;
3467 }
3468 else
3469 {
3470 shader->u.ps.declared_in_count = 0;
3471 for (i = 0; i < MAX_REG_INPUT; ++i)
3472 {
3473 if (shader->u.ps.input_reg_used[i])
3474 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3475 else shader->u.ps.input_reg_map[i] = ~0U;
3476 }
3477 }
3478
3479 return WINED3D_OK;
3480 }
3481
3482 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3483 {
3484 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3485 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3486 unsigned int i;
3487
3488 if (reg_maps->shader_version.major != 1) return;
3489
3490 for (i = 0; i < shader->limits->sampler; ++i)
3491 {
3492 /* We don't sample from this sampler. */
3493 if (!resource_info[i].type)
3494 continue;
3495
3496 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3497 {
3498 case WINED3D_SHADER_TEX_2D:
3499 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3500 break;
3501
3502 case WINED3D_SHADER_TEX_3D:
3503 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3504 break;
3505
3506 case WINED3D_SHADER_TEX_CUBE:
3507 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3508 break;
3509 }
3510 }
3511 }
3512
3513 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3514 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3515 {
3516 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3517 }
3518
3519 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3520 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3521 {
3522 struct wined3d_shader *object;
3523 HRESULT hr;
3524
3525 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3526 device, desc, parent, parent_ops, shader);
3527
3528 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3529 return E_OUTOFMEMORY;
3530
3531 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3532 {
3533 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3534 HeapFree(GetProcessHeap(), 0, object);
3535 return hr;
3536 }
3537
3538 TRACE("Created compute shader %p.\n", object);
3539 *shader = object;
3540
3541 return WINED3D_OK;
3542 }
3543
3544 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3545 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3546 {
3547 struct wined3d_shader *object;
3548 HRESULT hr;
3549
3550 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3551 device, desc, parent, parent_ops, shader);
3552
3553 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3554 return E_OUTOFMEMORY;
3555
3556 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
3557 {
3558 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
3559 HeapFree(GetProcessHeap(), 0, object);
3560 return hr;
3561 }
3562
3563 TRACE("Created domain shader %p.\n", object);
3564 *shader = object;
3565
3566 return WINED3D_OK;
3567 }
3568
3569 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3570 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3571 {
3572 struct wined3d_shader *object;
3573 HRESULT hr;
3574
3575 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3576 device, desc, parent, parent_ops, shader);
3577
3578 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3579 if (!object)
3580 return E_OUTOFMEMORY;
3581
3582 if (FAILED(hr = geometry_shader_init(object, device, desc, parent, parent_ops)))
3583 {
3584 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
3585 HeapFree(GetProcessHeap(), 0, object);
3586 return hr;
3587 }
3588
3589 TRACE("Created geometry shader %p.\n", object);
3590 *shader = object;
3591
3592 return WINED3D_OK;
3593 }
3594
3595 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3596 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3597 {
3598 struct wined3d_shader *object;
3599 HRESULT hr;
3600
3601 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3602 device, desc, parent, parent_ops, shader);
3603
3604 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3605 return E_OUTOFMEMORY;
3606
3607 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
3608 {
3609 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
3610 HeapFree(GetProcessHeap(), 0, object);
3611 return hr;
3612 }
3613
3614 TRACE("Created hull shader %p.\n", object);
3615 *shader = object;
3616
3617 return WINED3D_OK;
3618 }
3619
3620 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3621 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3622 {
3623 struct wined3d_shader *object;
3624 HRESULT hr;
3625
3626 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3627 device, desc, parent, parent_ops, shader);
3628
3629 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3630 if (!object)
3631 return E_OUTOFMEMORY;
3632
3633 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3634 {
3635 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
3636 HeapFree(GetProcessHeap(), 0, object);
3637 return hr;
3638 }
3639
3640 TRACE("Created pixel shader %p.\n", object);
3641 *shader = object;
3642
3643 return WINED3D_OK;
3644 }
3645
3646 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3647 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3648 {
3649 struct wined3d_shader *object;
3650 HRESULT hr;
3651
3652 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3653 device, desc, parent, parent_ops, shader);
3654
3655 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
3656 if (!object)
3657 return E_OUTOFMEMORY;
3658
3659 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3660 {
3661 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
3662 HeapFree(GetProcessHeap(), 0, object);
3663 return hr;
3664 }
3665
3666 TRACE("Created vertex shader %p.\n", object);
3667 *shader = object;
3668
3669 return WINED3D_OK;
3670 }
Windows API implementation