?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Recognize SM5 eval_sample_index instruction.
[wine.git] / dlls / wined3d / shader.c
blobe59b9e5bfbe138fb57ad9062713615357ffb18d3
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26 #include "wine/port.h"
27
28 #include <stdio.h>
29 #include <string.h>
30
31 #include "wined3d_private.h"
32
33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
34
35 /* pow, mul_high, sub_high, mul_low */
36 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
37 /* cmp */
38 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
39
40 static const char * const shader_opcode_names[] =
41 {
42 /* WINED3DSIH_ABS */ "abs",
43 /* WINED3DSIH_ADD */ "add",
44 /* WINED3DSIH_AND */ "and",
45 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
46 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
47 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
48 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
49 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
50 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
51 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
52 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
53 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
54 /* WINED3DSIH_BEM */ "bem",
55 /* WINED3DSIH_BFI */ "bfi",
56 /* WINED3DSIH_BFREV */ "bfrev",
57 /* WINED3DSIH_BREAK */ "break",
58 /* WINED3DSIH_BREAKC */ "breakc",
59 /* WINED3DSIH_BREAKP */ "breakp",
60 /* WINED3DSIH_BUFINFO */ "bufinfo",
61 /* WINED3DSIH_CALL */ "call",
62 /* WINED3DSIH_CALLNZ */ "callnz",
63 /* WINED3DSIH_CASE */ "case",
64 /* WINED3DSIH_CMP */ "cmp",
65 /* WINED3DSIH_CND */ "cnd",
66 /* WINED3DSIH_CONTINUE */ "continue",
67 /* WINED3DSIH_CONTINUEP */ "continuec",
68 /* WINED3DSIH_COUNTBITS */ "countbits",
69 /* WINED3DSIH_CRS */ "crs",
70 /* WINED3DSIH_CUT */ "cut",
71 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
72 /* WINED3DSIH_DCL */ "dcl",
73 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
74 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
75 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
76 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
77 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
78 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
79 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
81 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
82 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
83 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
84 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
85 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
86 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
87 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
88 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
89 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
90 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
91 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
92 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
93 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
94 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
95 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
96 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
97 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
98 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
99 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
100 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
101 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
102 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
103 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
104 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
105 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
106 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
107 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
108 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
109 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
110 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
111 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
112 /* WINED3DSIH_DEF */ "def",
113 /* WINED3DSIH_DEFAULT */ "default",
114 /* WINED3DSIH_DEFB */ "defb",
115 /* WINED3DSIH_DEFI */ "defi",
116 /* WINED3DSIH_DIV */ "div",
117 /* WINED3DSIH_DP2 */ "dp2",
118 /* WINED3DSIH_DP2ADD */ "dp2add",
119 /* WINED3DSIH_DP3 */ "dp3",
120 /* WINED3DSIH_DP4 */ "dp4",
121 /* WINED3DSIH_DST */ "dst",
122 /* WINED3DSIH_DSX */ "dsx",
123 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
124 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
125 /* WINED3DSIH_DSY */ "dsy",
126 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
127 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
128 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
129 /* WINED3DSIH_ELSE */ "else",
130 /* WINED3DSIH_EMIT */ "emit",
131 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
132 /* WINED3DSIH_ENDIF */ "endif",
133 /* WINED3DSIH_ENDLOOP */ "endloop",
134 /* WINED3DSIH_ENDREP */ "endrep",
135 /* WINED3DSIH_ENDSWITCH */ "endswitch",
136 /* WINED3DSIH_EQ */ "eq",
137 /* WINED3DSIH_EXP */ "exp",
138 /* WINED3DSIH_EXPP */ "expp",
139 /* WINED3DSIH_F16TOF32 */ "f16tof32",
140 /* WINED3DSIH_F32TOF16 */ "f32tof16",
141 /* WINED3DSIH_FCALL */ "fcall",
142 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
143 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
144 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
145 /* WINED3DSIH_FRC */ "frc",
146 /* WINED3DSIH_FTOI */ "ftoi",
147 /* WINED3DSIH_FTOU */ "ftou",
148 /* WINED3DSIH_GATHER4 */ "gather4",
149 /* WINED3DSIH_GATHER4_C */ "gather4_c",
150 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
151 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
152 /* WINED3DSIH_GE */ "ge",
153 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
154 /* WINED3DSIH_HS_DECLS */ "hs_decls",
155 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
156 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
157 /* WINED3DSIH_IADD */ "iadd",
158 /* WINED3DSIH_IBFE */ "ibfe",
159 /* WINED3DSIH_IEQ */ "ieq",
160 /* WINED3DSIH_IF */ "if",
161 /* WINED3DSIH_IFC */ "ifc",
162 /* WINED3DSIH_IGE */ "ige",
163 /* WINED3DSIH_ILT */ "ilt",
164 /* WINED3DSIH_IMAD */ "imad",
165 /* WINED3DSIH_IMAX */ "imax",
166 /* WINED3DSIH_IMIN */ "imin",
167 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
168 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
169 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
170 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
171 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
172 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
173 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
174 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
175 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
176 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
177 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
178 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
179 /* WINED3DSIH_IMUL */ "imul",
180 /* WINED3DSIH_INE */ "ine",
181 /* WINED3DSIH_INEG */ "ineg",
182 /* WINED3DSIH_ISHL */ "ishl",
183 /* WINED3DSIH_ISHR */ "ishr",
184 /* WINED3DSIH_ITOF */ "itof",
185 /* WINED3DSIH_LABEL */ "label",
186 /* WINED3DSIH_LD */ "ld",
187 /* WINED3DSIH_LD2DMS */ "ld2dms",
188 /* WINED3DSIH_LD_RAW */ "ld_raw",
189 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
190 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
191 /* WINED3DSIH_LIT */ "lit",
192 /* WINED3DSIH_LOD */ "lod",
193 /* WINED3DSIH_LOG */ "log",
194 /* WINED3DSIH_LOGP */ "logp",
195 /* WINED3DSIH_LOOP */ "loop",
196 /* WINED3DSIH_LRP */ "lrp",
197 /* WINED3DSIH_LT */ "lt",
198 /* WINED3DSIH_M3x2 */ "m3x2",
199 /* WINED3DSIH_M3x3 */ "m3x3",
200 /* WINED3DSIH_M3x4 */ "m3x4",
201 /* WINED3DSIH_M4x3 */ "m4x3",
202 /* WINED3DSIH_M4x4 */ "m4x4",
203 /* WINED3DSIH_MAD */ "mad",
204 /* WINED3DSIH_MAX */ "max",
205 /* WINED3DSIH_MIN */ "min",
206 /* WINED3DSIH_MOV */ "mov",
207 /* WINED3DSIH_MOVA */ "mova",
208 /* WINED3DSIH_MOVC */ "movc",
209 /* WINED3DSIH_MUL */ "mul",
210 /* WINED3DSIH_NE */ "ne",
211 /* WINED3DSIH_NOP */ "nop",
212 /* WINED3DSIH_NOT */ "not",
213 /* WINED3DSIH_NRM */ "nrm",
214 /* WINED3DSIH_OR */ "or",
215 /* WINED3DSIH_PHASE */ "phase",
216 /* WINED3DSIH_POW */ "pow",
217 /* WINED3DSIH_RCP */ "rcp",
218 /* WINED3DSIH_REP */ "rep",
219 /* WINED3DSIH_RESINFO */ "resinfo",
220 /* WINED3DSIH_RET */ "ret",
221 /* WINED3DSIH_RETP */ "retp",
222 /* WINED3DSIH_ROUND_NE */ "round_ne",
223 /* WINED3DSIH_ROUND_NI */ "round_ni",
224 /* WINED3DSIH_ROUND_PI */ "round_pi",
225 /* WINED3DSIH_ROUND_Z */ "round_z",
226 /* WINED3DSIH_RSQ */ "rsq",
227 /* WINED3DSIH_SAMPLE */ "sample",
228 /* WINED3DSIH_SAMPLE_B */ "sample_b",
229 /* WINED3DSIH_SAMPLE_C */ "sample_c",
230 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
231 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
232 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
233 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
234 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
235 /* WINED3DSIH_SETP */ "setp",
236 /* WINED3DSIH_SGE */ "sge",
237 /* WINED3DSIH_SGN */ "sgn",
238 /* WINED3DSIH_SINCOS */ "sincos",
239 /* WINED3DSIH_SLT */ "slt",
240 /* WINED3DSIH_SQRT */ "sqrt",
241 /* WINED3DSIH_STORE_RAW */ "store_raw",
242 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
243 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
244 /* WINED3DSIH_SUB */ "sub",
245 /* WINED3DSIH_SWAPC */ "swapc",
246 /* WINED3DSIH_SWITCH */ "switch",
247 /* WINED3DSIH_SYNC */ "sync",
248 /* WINED3DSIH_TEX */ "texld",
249 /* WINED3DSIH_TEXBEM */ "texbem",
250 /* WINED3DSIH_TEXBEML */ "texbeml",
251 /* WINED3DSIH_TEXCOORD */ "texcrd",
252 /* WINED3DSIH_TEXDEPTH */ "texdepth",
253 /* WINED3DSIH_TEXDP3 */ "texdp3",
254 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
255 /* WINED3DSIH_TEXKILL */ "texkill",
256 /* WINED3DSIH_TEXLDD */ "texldd",
257 /* WINED3DSIH_TEXLDL */ "texldl",
258 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
259 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
260 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
261 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
262 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
263 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
264 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
265 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
266 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
267 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
268 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
269 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
270 /* WINED3DSIH_UBFE */ "ubfe",
271 /* WINED3DSIH_UDIV */ "udiv",
272 /* WINED3DSIH_UGE */ "uge",
273 /* WINED3DSIH_ULT */ "ult",
274 /* WINED3DSIH_UMAX */ "umax",
275 /* WINED3DSIH_UMIN */ "umin",
276 /* WINED3DSIH_UMUL */ "umul",
277 /* WINED3DSIH_USHR */ "ushr",
278 /* WINED3DSIH_UTOF */ "utof",
279 /* WINED3DSIH_XOR */ "xor",
280 };
281
282 static const char * const semantic_names[] =
283 {
284 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
285 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
286 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
287 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
288 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
289 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
290 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
291 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
292 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
293 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
294 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
295 /* WINED3D_DECL_USAGE_FOG */ "FOG",
296 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
297 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
298 };
299
300 static const struct
301 {
302 enum wined3d_shader_input_sysval_semantic sysval_semantic;
303 const char *sysval_name;
304 }
305 shader_input_sysval_semantic_names[] =
306 {
307 {WINED3D_SIV_POSITION, "position"},
308 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
309 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
310 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
311 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
312 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
313 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
314 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
315 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
316 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
317 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
318 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
319 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
320 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
322 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
323 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
324 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
325 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
326 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
327 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
328 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
329 };
330
331 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
332 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
333
334 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
335 {
336 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
337 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
338
339 return shader_opcode_names[handler_idx];
340 }
341
342 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
343 {
344 if (usage >= ARRAY_SIZE(semantic_names))
345 {
346 FIXME("Unrecognized usage %#x.\n", usage);
347 return "UNRECOGNIZED";
348 }
349
350 return semantic_names[usage];
351 }
352
353 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
354 {
355 unsigned int i;
356
357 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
358 {
359 if (!strcmp(name, semantic_names[i]))
360 return i;
361 }
362
363 return ~0U;
364 }
365
366 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
367 {
368 switch (usage)
369 {
370 case WINED3D_DECL_USAGE_POSITION:
371 return WINED3D_SV_POSITION;
372 default:
373 return 0;
374 }
375 }
376
377 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
378 {
379 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
380 }
381
382 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
383 const struct wined3d_shader_semantic *s)
384 {
385 e->semantic_name = shader_semantic_name_from_usage(s->usage);
386 e->semantic_idx = s->usage_idx;
387 e->stream_idx = 0;
388 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
389 e->component_type = WINED3D_TYPE_FLOAT;
390 e->register_idx = s->reg.reg.idx[0].offset;
391 e->mask = s->reg.write_mask;
392 }
393
394 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
395 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
396 {
397 e->semantic_name = shader_semantic_name_from_usage(usage);
398 e->semantic_idx = usage_idx;
399 e->stream_idx = 0;
400 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
401 e->component_type = WINED3D_TYPE_FLOAT;
402 e->register_idx = reg_idx;
403 e->mask = write_mask;
404 }
405
406 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
407 {
408 switch (format)
409 {
410 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
411 return &sm1_shader_frontend;
412
413 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
414 return &sm4_shader_frontend;
415
416 default:
417 WARN("Invalid byte code format %#x specified.\n", format);
418 return NULL;
419 }
420 }
421
422 void string_buffer_clear(struct wined3d_string_buffer *buffer)
423 {
424 buffer->buffer[0] = '\0';
425 buffer->content_size = 0;
426 }
427
428 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
429 {
430 buffer->buffer_size = 32;
431 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
432 {
433 ERR("Failed to allocate shader buffer memory.\n");
434 return FALSE;
435 }
436
437 string_buffer_clear(buffer);
438 return TRUE;
439 }
440
441 void string_buffer_free(struct wined3d_string_buffer *buffer)
442 {
443 HeapFree(GetProcessHeap(), 0, buffer->buffer);
444 }
445
446 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
447 {
448 char *new_buffer;
449 unsigned int new_buffer_size = buffer->buffer_size * 2;
450
451 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
452 new_buffer_size *= 2;
453 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
454 {
455 ERR("Failed to grow buffer.\n");
456 buffer->buffer[buffer->content_size] = '\0';
457 return FALSE;
458 }
459 buffer->buffer = new_buffer;
460 buffer->buffer_size = new_buffer_size;
461 return TRUE;
462 }
463
464 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
465 {
466 unsigned int rem;
467 int rc;
468
469 rem = buffer->buffer_size - buffer->content_size;
470 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
471 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
472 return rc;
473
474 buffer->content_size += rc;
475 return 0;
476 }
477
478 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
479 {
480 va_list args;
481 int ret;
482
483 for (;;)
484 {
485 va_start(args, format);
486 ret = shader_vaddline(buffer, format, args);
487 va_end(args);
488 if (!ret)
489 return ret;
490 if (!string_buffer_resize(buffer, ret))
491 return -1;
492 }
493 }
494
495 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
496 {
497 struct wined3d_string_buffer *buffer;
498
499 if (list_empty(&list->list))
500 {
501 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
502 if (!buffer || !string_buffer_init(buffer))
503 {
504 ERR("Couldn't allocate buffer for temporary string.\n");
505 HeapFree(GetProcessHeap(), 0, buffer);
506 return NULL;
507 }
508 }
509 else
510 {
511 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
512 list_remove(&buffer->entry);
513 }
514 string_buffer_clear(buffer);
515 return buffer;
516 }
517
518 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
519 {
520 if (!buffer)
521 return 0;
522 string_buffer_clear(buffer);
523 return shader_vaddline(buffer, format, args);
524 }
525
526 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
527 {
528 va_list args;
529 int ret;
530
531 for (;;)
532 {
533 va_start(args, format);
534 ret = string_buffer_vsprintf(buffer, format, args);
535 va_end(args);
536 if (!ret)
537 return;
538 if (!string_buffer_resize(buffer, ret))
539 return;
540 }
541 }
542
543 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
544 {
545 if (!buffer)
546 return;
547 list_add_head(&list->list, &buffer->entry);
548 }
549
550 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
551 {
552 list_init(&list->list);
553 }
554
555 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
556 {
557 struct wined3d_string_buffer *buffer, *buffer_next;
558
559 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
560 {
561 string_buffer_free(buffer);
562 HeapFree(GetProcessHeap(), 0, buffer);
563 }
564 list_init(&list->list);
565 }
566
567 /* Convert floating point offset relative to a register file to an absolute
568 * offset for float constants. */
569 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
570 {
571 switch (register_type)
572 {
573 case WINED3DSPR_CONST: return register_idx;
574 case WINED3DSPR_CONST2: return 2048 + register_idx;
575 case WINED3DSPR_CONST3: return 4096 + register_idx;
576 case WINED3DSPR_CONST4: return 6144 + register_idx;
577 default:
578 FIXME("Unsupported register type: %u.\n", register_type);
579 return register_idx;
580 }
581 }
582
583 static void shader_delete_constant_list(struct list *clist)
584 {
585 struct wined3d_shader_lconst *constant, *constant_next;
586
587 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
588 HeapFree(GetProcessHeap(), 0, constant);
589 list_init(clist);
590 }
591
592 static void shader_set_limits(struct wined3d_shader *shader)
593 {
594 static const struct limits_entry
595 {
596 unsigned int min_version;
597 unsigned int max_version;
598 struct wined3d_shader_limits limits;
599 }
600 vs_limits[] =
601 {
602 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
603 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
604 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
605 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
606 * even though they are capable of supporting much more (GL
607 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
608 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
609 * shaders to 256. */
610 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
611 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
612 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
613 {0}
614 },
615 hs_limits[] =
616 {
617 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
618 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
619 },
620 ds_limits[] =
621 {
622 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
623 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
624 },
625 gs_limits[] =
626 {
627 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
628 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
629 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
630 {0}
631 },
632 ps_limits[] =
633 {
634 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
635 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
636 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
637 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
638 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
639 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
640 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
641 {0}
642 },
643 cs_limits[] =
644 {
645 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
646 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
647 };
648 const struct limits_entry *limits_array;
649 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
650 shader->reg_maps.shader_version.minor);
651 int i = 0;
652
653 switch (shader->reg_maps.shader_version.type)
654 {
655 default:
656 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
657 /* Fall-through. */
658 case WINED3D_SHADER_TYPE_VERTEX:
659 limits_array = vs_limits;
660 break;
661 case WINED3D_SHADER_TYPE_HULL:
662 limits_array = hs_limits;
663 break;
664 case WINED3D_SHADER_TYPE_DOMAIN:
665 limits_array = ds_limits;
666 break;
667 case WINED3D_SHADER_TYPE_GEOMETRY:
668 limits_array = gs_limits;
669 break;
670 case WINED3D_SHADER_TYPE_PIXEL:
671 limits_array = ps_limits;
672 break;
673 case WINED3D_SHADER_TYPE_COMPUTE:
674 limits_array = cs_limits;
675 break;
676 }
677
678 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
679 {
680 if (shader_version <= limits_array[i].max_version)
681 {
682 shader->limits = &limits_array[i].limits;
683 break;
684 }
685 ++i;
686 }
687 if (!shader->limits)
688 {
689 FIXME("Unexpected shader version \"%u.%u\".\n",
690 shader->reg_maps.shader_version.major,
691 shader->reg_maps.shader_version.minor);
692 shader->limits = &limits_array[max(0, i - 1)].limits;
693 }
694 }
695
696 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
697 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
698 {
699 switch (reg->type)
700 {
701 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
702 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
703 reg_maps->texcoord |= 1u << reg->idx[0].offset;
704 else
705 reg_maps->address |= 1u << reg->idx[0].offset;
706 break;
707
708 case WINED3DSPR_TEMP:
709 reg_maps->temporary |= 1u << reg->idx[0].offset;
710 break;
711
712 case WINED3DSPR_INPUT:
713 if (reg->idx[0].rel_addr)
714 reg_maps->input_rel_addressing = 1;
715 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
716 {
717 /* If relative addressing is used, we must assume that all
718 * registers are used. Even if it is a construct like v3[aL],
719 * we can't assume that v0, v1 and v2 aren't read because aL
720 * can be negative. */
721 if (reg->idx[0].rel_addr)
722 shader->u.ps.input_reg_used = ~0u;
723 else
724 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
725 }
726 else
727 {
728 reg_maps->input_registers |= 1u << reg->idx[0].offset;
729 }
730 break;
731
732 case WINED3DSPR_RASTOUT:
733 if (reg->idx[0].offset == 1)
734 reg_maps->fog = 1;
735 if (reg->idx[0].offset == 2)
736 reg_maps->point_size = 1;
737 break;
738
739 case WINED3DSPR_MISCTYPE:
740 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
741 {
742 if (!reg->idx[0].offset)
743 reg_maps->vpos = 1;
744 else if (reg->idx[0].offset == 1)
745 reg_maps->usesfacing = 1;
746 }
747 break;
748
749 case WINED3DSPR_CONST:
750 if (reg->idx[0].rel_addr)
751 {
752 if (reg->idx[0].offset < reg_maps->min_rel_offset)
753 reg_maps->min_rel_offset = reg->idx[0].offset;
754 if (reg->idx[0].offset > reg_maps->max_rel_offset)
755 reg_maps->max_rel_offset = reg->idx[0].offset;
756 reg_maps->usesrelconstF = TRUE;
757 }
758 else
759 {
760 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
761 {
762 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
763 return FALSE;
764 }
765 else
766 {
767 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
768 }
769 }
770 break;
771
772 case WINED3DSPR_CONSTINT:
773 if (reg->idx[0].offset >= shader->limits->constant_int)
774 {
775 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
776 return FALSE;
777 }
778 else
779 {
780 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
781 }
782 break;
783
784 case WINED3DSPR_CONSTBOOL:
785 if (reg->idx[0].offset >= shader->limits->constant_bool)
786 {
787 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
788 return FALSE;
789 }
790 else
791 {
792 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
793 }
794 break;
795
796 case WINED3DSPR_COLOROUT:
797 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
798 break;
799
800 case WINED3DSPR_OUTCONTROLPOINT:
801 reg_maps->vocp = 1;
802 break;
803
804 default:
805 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
806 reg->type, reg->idx[0].offset, reg->idx[1].offset);
807 break;
808 }
809 return TRUE;
810 }
811
812 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
813 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
814 {
815 struct wined3d_shader_sampler_map_entry *entries, *entry;
816 struct wined3d_shader_sampler_map *map;
817 unsigned int i;
818
819 map = &reg_maps->sampler_map;
820 entries = map->entries;
821 for (i = 0; i < map->count; ++i)
822 {
823 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
824 return;
825 }
826
827 if (!map->size)
828 {
829 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
830 {
831 ERR("Failed to allocate sampler map entries.\n");
832 return;
833 }
834 map->size = 4;
835 map->entries = entries;
836 }
837 else if (map->count == map->size)
838 {
839 size_t new_size = map->size * 2;
840
841 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
842 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
843 {
844 ERR("Failed to resize sampler map entries.\n");
845 return;
846 }
847 map->size = new_size;
848 map->entries = entries;
849 }
850
851 entry = &entries[map->count++];
852 entry->resource_idx = resource_idx;
853 entry->sampler_idx = sampler_idx;
854 entry->bind_idx = bind_idx;
855 }
856
857 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
858 {
859 switch (instr)
860 {
861 case WINED3DSIH_M4x4:
862 case WINED3DSIH_M3x4:
863 return param == 1 ? 3 : 0;
864
865 case WINED3DSIH_M4x3:
866 case WINED3DSIH_M3x3:
867 return param == 1 ? 2 : 0;
868
869 case WINED3DSIH_M3x2:
870 return param == 1 ? 1 : 0;
871
872 default:
873 return 0;
874 }
875 }
876
877 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
878 unsigned int register_idx, unsigned int size, unsigned int stride)
879 {
880 struct wined3d_shader_tgsm *tgsm;
881
882 if (register_idx >= MAX_TGSM_REGISTERS)
883 {
884 ERR("Invalid TGSM register index %u.\n", register_idx);
885 return S_OK;
886 }
887 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
888 {
889 FIXME("TGSM declarations are allowed only in compute shaders.\n");
890 return S_OK;
891 }
892
893 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
894 register_idx + 1, sizeof(*reg_maps->tgsm)))
895 return E_OUTOFMEMORY;
896
897 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
898 tgsm = &reg_maps->tgsm[register_idx];
899 tgsm->size = size;
900 tgsm->stride = stride;
901 return S_OK;
902 }
903
904 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
905 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
906 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
907 {
908 struct wined3d_shader_phase *phase;
909
910 if ((phase = *current_phase))
911 {
912 phase->end = previous_instruction_ptr;
913 *current_phase = NULL;
914 }
915
916 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
917 {
918 ERR("Unexpected shader type %#x.\n", shader->reg_maps.shader_version.type);
919 return E_FAIL;
920 }
921
922 switch (ins->handler_idx)
923 {
924 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
925 if (shader->u.hs.phases.control_point)
926 {
927 FIXME("Multiple control point phases.\n");
928 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
929 }
930 if (!(shader->u.hs.phases.control_point = HeapAlloc(GetProcessHeap(),
931 HEAP_ZERO_MEMORY, sizeof(*shader->u.hs.phases.control_point))))
932 return E_OUTOFMEMORY;
933 phase = shader->u.hs.phases.control_point;
934 break;
935 case WINED3DSIH_HS_FORK_PHASE:
936 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
937 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
938 sizeof(*shader->u.hs.phases.fork)))
939 return E_OUTOFMEMORY;
940 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
941 break;
942 case WINED3DSIH_HS_JOIN_PHASE:
943 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
944 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
945 sizeof(*shader->u.hs.phases.join)))
946 return E_OUTOFMEMORY;
947 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
948 break;
949 default:
950 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
951 return E_FAIL;
952 }
953
954 phase->start = current_instruction_ptr;
955 *current_phase = phase;
956
957 return WINED3D_OK;
958 }
959
960 static HRESULT shader_calculate_clip_or_cull_distance_mask(
961 const struct wined3d_shader_signature_element *e, DWORD *mask)
962 {
963 unsigned int i;
964
965 *mask = 0;
966
967 /* Cull and clip distances are packed in 4 component registers. 0 and 1 are
968 * the only allowed semantic indices.
969 */
970 if (e->semantic_idx >= MAX_CLIP_DISTANCES / 4)
971 {
972 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
973 return WINED3DERR_INVALIDCALL;
974 }
975
976 for (i = 0; i < 4; ++i)
977 {
978 if (e->mask & (WINED3DSP_WRITEMASK_0 << i))
979 *mask |= 1u << (4 * e->semantic_idx + i);
980 }
981
982 return WINED3D_OK;
983 }
984
985 static void wined3d_insert_interpolation_mode(DWORD *packed_interpolation_mode,
986 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
987 {
988 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
989 FIXME("Unexpected interpolation mode %#x.\n", mode);
990
991 wined3d_insert_bits(packed_interpolation_mode,
992 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
993 }
994
995 /* Note that this does not count the loop register as an address register. */
996 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
997 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
998 struct wined3d_shader_signature *output_signature, DWORD constf_size)
999 {
1000 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
1001 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
1002 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
1003 struct wined3d_shader_version shader_version;
1004 struct wined3d_shader_phase *phase = NULL;
1005 const DWORD *ptr, *prev_ins, *current_ins;
1006 void *fe_data = shader->frontend_data;
1007 unsigned int i;
1008 HRESULT hr;
1009
1010 memset(reg_maps, 0, sizeof(*reg_maps));
1011 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1012 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1013 reg_maps->min_rel_offset = ~0U;
1014 list_init(&reg_maps->indexable_temps);
1015
1016 fe->shader_read_header(fe_data, &ptr, &shader_version);
1017 prev_ins = current_ins = ptr;
1018 reg_maps->shader_version = shader_version;
1019
1020 shader_set_limits(shader);
1021
1022 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1023 sizeof(*reg_maps->constf))))
1024 {
1025 ERR("Failed to allocate constant map memory.\n");
1026 return E_OUTOFMEMORY;
1027 }
1028
1029 while (!fe->shader_is_end(fe_data, &ptr))
1030 {
1031 struct wined3d_shader_instruction ins;
1032
1033 current_ins = ptr;
1034 /* Fetch opcode. */
1035 fe->shader_read_instruction(fe_data, &ptr, &ins);
1036
1037 /* Unhandled opcode, and its parameters. */
1038 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1039 {
1040 WARN("Encountered unrecognised or invalid instruction.\n");
1041 return WINED3DERR_INVALIDCALL;
1042 }
1043
1044 /* Handle declarations. */
1045 if (ins.handler_idx == WINED3DSIH_DCL
1046 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1047 {
1048 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1049 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1050
1051 switch (semantic->reg.reg.type)
1052 {
1053 /* Mark input registers used. */
1054 case WINED3DSPR_INPUT:
1055 if (reg_idx >= MAX_REG_INPUT)
1056 {
1057 ERR("Invalid input register index %u.\n", reg_idx);
1058 break;
1059 }
1060 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1061 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1062 return WINED3DERR_INVALIDCALL;
1063 reg_maps->input_registers |= 1u << reg_idx;
1064 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1065 break;
1066
1067 /* Vertex shader: mark 3.0 output registers used, save token. */
1068 case WINED3DSPR_OUTPUT:
1069 if (reg_idx >= MAX_REG_OUTPUT)
1070 {
1071 ERR("Invalid output register index %u.\n", reg_idx);
1072 break;
1073 }
1074 reg_maps->output_registers |= 1u << reg_idx;
1075 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1076 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1077 reg_maps->fog = 1;
1078 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1079 reg_maps->point_size = 1;
1080 break;
1081
1082 case WINED3DSPR_SAMPLER:
1083 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1084 case WINED3DSPR_RESOURCE:
1085 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1086 {
1087 ERR("Invalid resource index %u.\n", reg_idx);
1088 break;
1089 }
1090 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1091 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1092 break;
1093
1094 case WINED3DSPR_UAV:
1095 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1096 {
1097 ERR("Invalid UAV resource index %u.\n", reg_idx);
1098 break;
1099 }
1100 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1101 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1102 if (ins.flags)
1103 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1104 break;
1105
1106 default:
1107 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1108 break;
1109 }
1110 }
1111 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1112 {
1113 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1114 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1115 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1116 else
1117 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1118 }
1119 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1120 {
1121 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1122 {
1123 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1124 shader->u.ps.force_early_depth_stencil = TRUE;
1125 else
1126 FIXME("Invalid instruction %#x for shader type %#x.\n",
1127 ins.handler_idx, shader_version.type);
1128 }
1129 else
1130 {
1131 WARN("Ignoring global flags %#x.\n", ins.flags);
1132 }
1133 }
1134 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1135 {
1136 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1137 shader->u.gs.instance_count = ins.declaration.count;
1138 else
1139 FIXME("Invalid instruction %#x for shader type %#x.\n",
1140 ins.handler_idx, shader_version.type);
1141 }
1142 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1143 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1144 {
1145 if (phase)
1146 phase->instance_count = ins.declaration.count;
1147 else
1148 FIXME("Instruction %s outside of shader phase.\n",
1149 debug_d3dshaderinstructionhandler(ins.handler_idx));
1150 }
1151 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1152 {
1153 if (reg_maps->icb)
1154 FIXME("Multiple immediate constant buffers.\n");
1155 reg_maps->icb = ins.declaration.icb;
1156 }
1157 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1158 {
1159 if (phase)
1160 {
1161 FIXME("Indexable temporary registers not supported.\n");
1162 }
1163 else
1164 {
1165 struct wined3d_shader_indexable_temp *reg;
1166
1167 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
1168 return E_OUTOFMEMORY;
1169
1170 *reg = ins.declaration.indexable_temp;
1171 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1172 }
1173 }
1174 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1175 {
1176 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1177 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1178 else
1179 FIXME("Invalid instruction %#x for shader type %#x.\n",
1180 ins.handler_idx, shader_version.type);
1181 }
1182 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1183 {
1184 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1185 if (reg_idx >= MAX_REG_INPUT)
1186 {
1187 ERR("Invalid register index %u.\n", reg_idx);
1188 break;
1189 }
1190 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1191 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1192 else
1193 FIXME("Invalid instruction %#x for shader type %#x.\n",
1194 ins.handler_idx, shader_version.type);
1195 }
1196 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1197 {
1198 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1199 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1200 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1201 {
1202 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1203 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1204 else
1205 FIXME("Invalid instruction %#x for shader type %#x.\n",
1206 ins.handler_idx, shader_version.type);
1207 }
1208 }
1209 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1210 {
1211 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1212 shader->u.hs.output_vertex_count = ins.declaration.count;
1213 else
1214 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1215 }
1216 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1217 {
1218 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1219 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1220 else
1221 FIXME("Invalid instruction %#x for shader type %#x.\n",
1222 ins.handler_idx, shader_version.type);
1223 }
1224 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1225 {
1226 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1227 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1228 {
1229 ERR("Invalid resource index %u.\n", reg_idx);
1230 break;
1231 }
1232 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1233 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1234 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1235 }
1236 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1237 {
1238 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1239 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1240 {
1241 ERR("Invalid resource index %u.\n", reg_idx);
1242 break;
1243 }
1244 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1245 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1246 reg_maps->resource_info[reg_idx].flags = 0;
1247 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1248 }
1249 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1250 {
1251 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1252 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1253 }
1254 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1255 {
1256 if (phase)
1257 phase->temporary_count = ins.declaration.count;
1258 else
1259 reg_maps->temporary_count = ins.declaration.count;
1260 }
1261 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1262 {
1263 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1264 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1265 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1266 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1267 }
1268 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1269 {
1270 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1271 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1272 else
1273 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1274 }
1275 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1276 {
1277 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1278 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1279 else
1280 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1281 }
1282 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1283 {
1284 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1285 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1286 return hr;
1287 }
1288 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1289 {
1290 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1291 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1292 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1293 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1294 return hr;
1295 }
1296 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1297 {
1298 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1299 {
1300 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1301 }
1302 else
1303 {
1304 FIXME("Invalid instruction %#x for shader type %#x.\n",
1305 ins.handler_idx, shader_version.type);
1306 }
1307 }
1308 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1309 {
1310 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1311 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1312 {
1313 ERR("Invalid UAV resource index %u.\n", reg_idx);
1314 break;
1315 }
1316 if (ins.flags)
1317 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1318 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1319 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1320 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1321 }
1322 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1323 {
1324 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1325 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1326 {
1327 ERR("Invalid UAV resource index %u.\n", reg_idx);
1328 break;
1329 }
1330 if (ins.flags)
1331 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1332 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1333 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1334 reg_maps->uav_resource_info[reg_idx].flags = 0;
1335 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1336 }
1337 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1338 {
1339 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1340 shader->u.gs.vertices_out = ins.declaration.count;
1341 else
1342 FIXME("Invalid instruction %#x for shader type %#x.\n",
1343 ins.handler_idx, shader_version.type);
1344 }
1345 else if (ins.handler_idx == WINED3DSIH_DEF)
1346 {
1347 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1348 float *value;
1349 if (!lconst) return E_OUTOFMEMORY;
1350
1351 lconst->idx = ins.dst[0].reg.idx[0].offset;
1352 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1353 value = (float *)lconst->value;
1354
1355 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1356 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1357 {
1358 if (value[0] < -1.0f) value[0] = -1.0f;
1359 else if (value[0] > 1.0f) value[0] = 1.0f;
1360 if (value[1] < -1.0f) value[1] = -1.0f;
1361 else if (value[1] > 1.0f) value[1] = 1.0f;
1362 if (value[2] < -1.0f) value[2] = -1.0f;
1363 else if (value[2] > 1.0f) value[2] = 1.0f;
1364 if (value[3] < -1.0f) value[3] = -1.0f;
1365 else if (value[3] > 1.0f) value[3] = 1.0f;
1366 }
1367
1368 list_add_head(&shader->constantsF, &lconst->entry);
1369
1370 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1371 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1372 {
1373 shader->lconst_inf_or_nan = TRUE;
1374 }
1375 }
1376 else if (ins.handler_idx == WINED3DSIH_DEFI)
1377 {
1378 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1379 if (!lconst) return E_OUTOFMEMORY;
1380
1381 lconst->idx = ins.dst[0].reg.idx[0].offset;
1382 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1383
1384 list_add_head(&shader->constantsI, &lconst->entry);
1385 reg_maps->local_int_consts |= (1u << lconst->idx);
1386 }
1387 else if (ins.handler_idx == WINED3DSIH_DEFB)
1388 {
1389 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1390 if (!lconst) return E_OUTOFMEMORY;
1391
1392 lconst->idx = ins.dst[0].reg.idx[0].offset;
1393 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1394
1395 list_add_head(&shader->constantsB, &lconst->entry);
1396 reg_maps->local_bool_consts |= (1u << lconst->idx);
1397 }
1398 /* Handle shader phases. */
1399 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1400 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1401 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1402 {
1403 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1404 return hr;
1405 }
1406 /* For subroutine prototypes. */
1407 else if (ins.handler_idx == WINED3DSIH_LABEL)
1408 {
1409 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1410 }
1411 /* Set texture, address, temporary registers. */
1412 else
1413 {
1414 BOOL color0_mov = FALSE;
1415 unsigned int i;
1416
1417 /* This will loop over all the registers and try to
1418 * make a bitmask of the ones we're interested in.
1419 *
1420 * Relative addressing tokens are ignored, but that's
1421 * okay, since we'll catch any address registers when
1422 * they are initialized (required by spec). */
1423 for (i = 0; i < ins.dst_count; ++i)
1424 {
1425 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1426 shader_version.type, constf_size))
1427 return WINED3DERR_INVALIDCALL;
1428
1429 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1430 {
1431 UINT idx = ins.dst[i].reg.idx[0].offset;
1432
1433 switch (ins.dst[i].reg.type)
1434 {
1435 case WINED3DSPR_RASTOUT:
1436 if (shader_version.major >= 3)
1437 break;
1438 switch (idx)
1439 {
1440 case 0: /* oPos */
1441 reg_maps->output_registers |= 1u << 10;
1442 shader_signature_from_usage(&output_signature_elements[10],
1443 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1444 break;
1445
1446 case 1: /* oFog */
1447 reg_maps->output_registers |= 1u << 11;
1448 shader_signature_from_usage(&output_signature_elements[11],
1449 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1450 break;
1451
1452 case 2: /* oPts */
1453 reg_maps->output_registers |= 1u << 11;
1454 shader_signature_from_usage(&output_signature_elements[11],
1455 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1456 break;
1457 }
1458 break;
1459
1460 case WINED3DSPR_ATTROUT:
1461 if (shader_version.major >= 3)
1462 break;
1463 if (idx < 2)
1464 {
1465 idx += 8;
1466 if (reg_maps->output_registers & (1u << idx))
1467 {
1468 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1469 }
1470 else
1471 {
1472 reg_maps->output_registers |= 1u << idx;
1473 shader_signature_from_usage(&output_signature_elements[idx],
1474 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1475 }
1476 }
1477 break;
1478
1479 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1480 if (shader_version.major >= 3)
1481 {
1482 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1483 {
1484 WARN("Invalid output register index %u.\n", idx);
1485 break;
1486 }
1487 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1488 break;
1489 }
1490 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1491 {
1492 WARN("Invalid texcoord index %u.\n", idx);
1493 break;
1494 }
1495 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1496 if (reg_maps->output_registers & (1u << idx))
1497 {
1498 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1499 }
1500 else
1501 {
1502 reg_maps->output_registers |= 1u << idx;
1503 shader_signature_from_usage(&output_signature_elements[idx],
1504 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1505 }
1506 break;
1507
1508 default:
1509 break;
1510 }
1511 }
1512
1513 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1514 {
1515 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1516 {
1517 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1518 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1519 * the mov and perform the sRGB write correction from the source register.
1520 *
1521 * However, if the mov is only partial, we can't do this, and if the write
1522 * comes from an instruction other than MOV it is hard to do as well. If
1523 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1524 shader->u.ps.color0_mov = FALSE;
1525 if (ins.handler_idx == WINED3DSIH_MOV
1526 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1527 {
1528 /* Used later when the source register is read. */
1529 color0_mov = TRUE;
1530 }
1531 }
1532 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1533 * end
1534 */
1535 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1536 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1537 {
1538 shader->u.ps.color0_mov = FALSE;
1539 }
1540 }
1541
1542 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1543 if (shader_version.major == 1
1544 && (ins.handler_idx == WINED3DSIH_TEX
1545 || ins.handler_idx == WINED3DSIH_TEXBEM
1546 || ins.handler_idx == WINED3DSIH_TEXBEML
1547 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1548 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1549 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1550 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1551 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1552 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1553 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1554 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1555 {
1556 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1557
1558 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1559 {
1560 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1561 continue;
1562 }
1563
1564 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1565 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1566 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1567 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1568
1569 /* texbem is only valid with < 1.4 pixel shaders */
1570 if (ins.handler_idx == WINED3DSIH_TEXBEM
1571 || ins.handler_idx == WINED3DSIH_TEXBEML)
1572 {
1573 reg_maps->bumpmat |= 1u << reg_idx;
1574 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1575 {
1576 reg_maps->luminanceparams |= 1u << reg_idx;
1577 }
1578 }
1579 }
1580 else if (ins.handler_idx == WINED3DSIH_BEM)
1581 {
1582 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1583 }
1584 }
1585
1586 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1587 {
1588 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1589 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1590 {
1591 ERR("Invalid UAV index %u.\n", reg_idx);
1592 break;
1593 }
1594 reg_maps->uav_counter_mask |= (1u << reg_idx);
1595 }
1596 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1597 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1598 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1599 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1600 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1601 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1602 {
1603 unsigned int reg_idx;
1604 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1605 reg_idx = ins.src[1].reg.idx[0].offset;
1606 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1607 reg_idx = ins.src[2].reg.idx[0].offset;
1608 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1609 reg_idx = ins.dst[0].reg.idx[0].offset;
1610 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1611 reg_idx = ins.src[0].reg.idx[0].offset;
1612 else
1613 reg_idx = ins.dst[1].reg.idx[0].offset;
1614 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1615 {
1616 ERR("Invalid UAV index %u.\n", reg_idx);
1617 break;
1618 }
1619 reg_maps->uav_read_mask |= (1u << reg_idx);
1620 }
1621 else if (ins.handler_idx == WINED3DSIH_NRM)
1622 {
1623 reg_maps->usesnrm = 1;
1624 }
1625 else if (ins.handler_idx == WINED3DSIH_DSY
1626 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1627 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1628 {
1629 reg_maps->usesdsy = 1;
1630 }
1631 else if (ins.handler_idx == WINED3DSIH_DSX
1632 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1633 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1634 {
1635 reg_maps->usesdsx = 1;
1636 }
1637 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1638 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1639 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1640 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1641 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1642 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1643 else if (ins.handler_idx == WINED3DSIH_LOOP
1644 || ins.handler_idx == WINED3DSIH_REP)
1645 {
1646 ++cur_loop_depth;
1647 if (cur_loop_depth > max_loop_depth)
1648 max_loop_depth = cur_loop_depth;
1649 }
1650 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1651 || ins.handler_idx == WINED3DSIH_ENDREP)
1652 {
1653 --cur_loop_depth;
1654 }
1655 else if (ins.handler_idx == WINED3DSIH_GATHER4
1656 || ins.handler_idx == WINED3DSIH_GATHER4_C
1657 || ins.handler_idx == WINED3DSIH_SAMPLE
1658 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1659 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1660 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1661 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1662 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1663 {
1664 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1665 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1666 }
1667 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1668 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1669 {
1670 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1671 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1672 }
1673 else if (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1674 {
1675 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1676 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1677 }
1678 else if (ins.handler_idx == WINED3DSIH_LD
1679 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1680 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1681 {
1682 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1683 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1684 }
1685 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1686 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1687 {
1688 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1689 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1690 }
1691
1692 if (ins.predicate)
1693 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1694 shader_version.type, constf_size))
1695 return WINED3DERR_INVALIDCALL;
1696
1697 for (i = 0; i < ins.src_count; ++i)
1698 {
1699 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1700 struct wined3d_shader_register reg = ins.src[i].reg;
1701
1702 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1703 shader_version.type, constf_size))
1704 return WINED3DERR_INVALIDCALL;
1705 while (count)
1706 {
1707 ++reg.idx[0].offset;
1708 if (!shader_record_register_usage(shader, reg_maps, &reg,
1709 shader_version.type, constf_size))
1710 return WINED3DERR_INVALIDCALL;
1711 --count;
1712 }
1713
1714 if (color0_mov)
1715 {
1716 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1717 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1718 {
1719 shader->u.ps.color0_mov = TRUE;
1720 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1721 }
1722 }
1723 }
1724 }
1725
1726 prev_ins = current_ins;
1727 }
1728 reg_maps->loop_depth = max_loop_depth;
1729
1730 if (phase)
1731 {
1732 phase->end = prev_ins;
1733 phase = NULL;
1734 }
1735
1736 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1737 * R0 is written to the render target. */
1738 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1739 reg_maps->rt_mask |= (1u << 0);
1740
1741 if (input_signature->elements)
1742 {
1743 for (i = 0; i < input_signature->element_count; ++i)
1744 {
1745 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1746 {
1747 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1748 {
1749 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1750 return WINED3DERR_INVALIDCALL;
1751 }
1752 }
1753 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1754 {
1755 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1756 reg_maps->vpos = 1;
1757 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1758 reg_maps->usesfacing = 1;
1759 }
1760 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1761 }
1762 }
1763 else if (!input_signature->elements && reg_maps->input_registers)
1764 {
1765 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1766 struct wined3d_shader_signature_element *e;
1767 unsigned int i;
1768
1769 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1770 return E_OUTOFMEMORY;
1771 input_signature->element_count = count;
1772
1773 e = input_signature->elements;
1774 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1775 {
1776 if (!(reg_maps->input_registers & (1u << i)))
1777 continue;
1778 input_signature_elements[i].register_idx = i;
1779 *e++ = input_signature_elements[i];
1780 }
1781 }
1782
1783 if (output_signature->elements)
1784 {
1785 for (i = 0; i < output_signature->element_count; ++i)
1786 {
1787 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1788 DWORD mask;
1789
1790 reg_maps->output_registers |= 1u << e->register_idx;
1791 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
1792 {
1793 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1794 return hr;
1795 reg_maps->clip_distance_mask |= mask;
1796 }
1797 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
1798 {
1799 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
1800 return hr;
1801 reg_maps->cull_distance_mask |= mask;
1802 }
1803 }
1804 }
1805 else if (reg_maps->output_registers)
1806 {
1807 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1808 struct wined3d_shader_signature_element *e;
1809
1810 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1811 return E_OUTOFMEMORY;
1812 output_signature->element_count = count;
1813
1814 e = output_signature->elements;
1815 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1816 {
1817 if (!(reg_maps->output_registers & (1u << i)))
1818 continue;
1819 *e++ = output_signature_elements[i];
1820 }
1821 }
1822
1823 return WINED3D_OK;
1824 }
1825
1826 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1827 {
1828 struct wined3d_shader_indexable_temp *reg, *reg_next;
1829
1830 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1831 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1832
1833 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1834 HeapFree(GetProcessHeap(), 0, reg);
1835 list_init(&reg_maps->indexable_temps);
1836
1837 HeapFree(GetProcessHeap(), 0, reg_maps->tgsm);
1838 }
1839
1840 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1841 {
1842 DWORD map = 1u << max;
1843 map |= map - 1;
1844 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1845
1846 return wined3d_log2i(map);
1847 }
1848
1849 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1850 {
1851 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1852 {
1853 shader_addline(buffer, "refactoringAllowed");
1854 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1855 if (global_flags)
1856 shader_addline(buffer, " | ");
1857 }
1858
1859 if (global_flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1860 {
1861 shader_addline(buffer, "forceEarlyDepthStencil");
1862 global_flags &= ~WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL;
1863 if (global_flags)
1864 shader_addline(buffer, " | ");
1865 }
1866
1867 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1868 {
1869 shader_addline(buffer, "enableRawAndStructuredBuffers");
1870 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1871 }
1872
1873 if (global_flags)
1874 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1875 }
1876
1877 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1878 {
1879 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1880 {
1881 shader_addline(buffer, "_g");
1882 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1883 }
1884 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1885 {
1886 shader_addline(buffer, "_t");
1887 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1888 }
1889
1890 if (sync_flags)
1891 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1892 }
1893
1894 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1895 {
1896 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1897 {
1898 shader_addline(buffer, "_glc");
1899 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1900 }
1901 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1902 {
1903 shader_addline(buffer, "_opc");
1904 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1905 }
1906
1907 if (uav_flags)
1908 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1909 }
1910
1911 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1912 enum wined3d_tessellator_domain domain)
1913 {
1914 switch (domain)
1915 {
1916 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1917 shader_addline(buffer, "line");
1918 break;
1919 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1920 shader_addline(buffer, "triangle");
1921 break;
1922 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1923 shader_addline(buffer, "quad");
1924 break;
1925 default:
1926 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1927 break;
1928 }
1929 }
1930
1931 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1932 enum wined3d_tessellator_output_primitive output_primitive)
1933 {
1934 switch (output_primitive)
1935 {
1936 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1937 shader_addline(buffer, "point");
1938 break;
1939 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1940 shader_addline(buffer, "line");
1941 break;
1942 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1943 shader_addline(buffer, "triangle_cw");
1944 break;
1945 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1946 shader_addline(buffer, "triangle_ccw");
1947 break;
1948 default:
1949 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1950 break;
1951 }
1952 }
1953
1954 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1955 enum wined3d_tessellator_partitioning partitioning)
1956 {
1957 switch (partitioning)
1958 {
1959 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1960 shader_addline(buffer, "integer");
1961 break;
1962 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1963 shader_addline(buffer, "pow2");
1964 break;
1965 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1966 shader_addline(buffer, "fractional_odd");
1967 break;
1968 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1969 shader_addline(buffer, "fractional_even");
1970 break;
1971 default:
1972 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1973 break;
1974 }
1975 }
1976
1977 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
1978 enum wined3d_shader_input_sysval_semantic semantic)
1979 {
1980 unsigned int i;
1981
1982 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
1983 {
1984 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
1985 {
1986 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
1987 return;
1988 }
1989 }
1990
1991 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
1992 }
1993
1994 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1995 const struct wined3d_shader_semantic *semantic, unsigned int flags,
1996 const struct wined3d_shader_version *shader_version)
1997 {
1998 shader_addline(buffer, "dcl");
1999
2000 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
2001 {
2002 switch (semantic->resource_type)
2003 {
2004 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2005 shader_addline(buffer, "_2d");
2006 break;
2007
2008 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2009 shader_addline(buffer, "_3d");
2010 break;
2011
2012 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2013 shader_addline(buffer, "_cube");
2014 break;
2015
2016 default:
2017 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
2018 break;
2019 }
2020 }
2021 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
2022 {
2023 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
2024 shader_addline(buffer, "_resource_");
2025 else
2026 shader_addline(buffer, "_uav_");
2027 switch (semantic->resource_type)
2028 {
2029 case WINED3D_SHADER_RESOURCE_BUFFER:
2030 shader_addline(buffer, "buffer");
2031 break;
2032
2033 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
2034 shader_addline(buffer, "texture1d");
2035 break;
2036
2037 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
2038 shader_addline(buffer, "texture2d");
2039 break;
2040
2041 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
2042 shader_addline(buffer, "texture2dms");
2043 break;
2044
2045 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2046 shader_addline(buffer, "texture3d");
2047 break;
2048
2049 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2050 shader_addline(buffer, "texturecube");
2051 break;
2052
2053 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
2054 shader_addline(buffer, "texture1darray");
2055 break;
2056
2057 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
2058 shader_addline(buffer, "texture2darray");
2059 break;
2060
2061 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
2062 shader_addline(buffer, "texture2dmsarray");
2063 break;
2064
2065 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
2066 shader_addline(buffer, "texturecubearray");
2067 break;
2068
2069 default:
2070 shader_addline(buffer, "unknown");
2071 break;
2072 }
2073 if (semantic->reg.reg.type == WINED3DSPR_UAV)
2074 shader_dump_uav_flags(buffer, flags);
2075 switch (semantic->resource_data_type)
2076 {
2077 case WINED3D_DATA_FLOAT:
2078 shader_addline(buffer, " (float)");
2079 break;
2080
2081 case WINED3D_DATA_INT:
2082 shader_addline(buffer, " (int)");
2083 break;
2084
2085 case WINED3D_DATA_UINT:
2086 shader_addline(buffer, " (uint)");
2087 break;
2088
2089 case WINED3D_DATA_UNORM:
2090 shader_addline(buffer, " (unorm)");
2091 break;
2092
2093 case WINED3D_DATA_SNORM:
2094 shader_addline(buffer, " (snorm)");
2095 break;
2096
2097 default:
2098 shader_addline(buffer, " (unknown)");
2099 break;
2100 }
2101 }
2102 else
2103 {
2104 /* Pixel shaders 3.0 don't have usage semantics. */
2105 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2106 return;
2107 else
2108 shader_addline(buffer, "_");
2109
2110 switch (semantic->usage)
2111 {
2112 case WINED3D_DECL_USAGE_POSITION:
2113 shader_addline(buffer, "position%u", semantic->usage_idx);
2114 break;
2115
2116 case WINED3D_DECL_USAGE_BLEND_INDICES:
2117 shader_addline(buffer, "blend");
2118 break;
2119
2120 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2121 shader_addline(buffer, "weight");
2122 break;
2123
2124 case WINED3D_DECL_USAGE_NORMAL:
2125 shader_addline(buffer, "normal%u", semantic->usage_idx);
2126 break;
2127
2128 case WINED3D_DECL_USAGE_PSIZE:
2129 shader_addline(buffer, "psize");
2130 break;
2131
2132 case WINED3D_DECL_USAGE_COLOR:
2133 if (!semantic->usage_idx)
2134 shader_addline(buffer, "color");
2135 else
2136 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2137 break;
2138
2139 case WINED3D_DECL_USAGE_TEXCOORD:
2140 shader_addline(buffer, "texture%u", semantic->usage_idx);
2141 break;
2142
2143 case WINED3D_DECL_USAGE_TANGENT:
2144 shader_addline(buffer, "tangent");
2145 break;
2146
2147 case WINED3D_DECL_USAGE_BINORMAL:
2148 shader_addline(buffer, "binormal");
2149 break;
2150
2151 case WINED3D_DECL_USAGE_TESS_FACTOR:
2152 shader_addline(buffer, "tessfactor");
2153 break;
2154
2155 case WINED3D_DECL_USAGE_POSITIONT:
2156 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2157 break;
2158
2159 case WINED3D_DECL_USAGE_FOG:
2160 shader_addline(buffer, "fog");
2161 break;
2162
2163 case WINED3D_DECL_USAGE_DEPTH:
2164 shader_addline(buffer, "depth");
2165 break;
2166
2167 case WINED3D_DECL_USAGE_SAMPLE:
2168 shader_addline(buffer, "sample");
2169 break;
2170
2171 default:
2172 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2173 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2174 }
2175 }
2176 }
2177
2178 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2179 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2180 {
2181 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2182 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2183 UINT offset = reg->idx[0].offset;
2184
2185 switch (reg->type)
2186 {
2187 case WINED3DSPR_TEMP:
2188 shader_addline(buffer, "r");
2189 break;
2190
2191 case WINED3DSPR_INPUT:
2192 shader_addline(buffer, "v");
2193 break;
2194
2195 case WINED3DSPR_CONST:
2196 case WINED3DSPR_CONST2:
2197 case WINED3DSPR_CONST3:
2198 case WINED3DSPR_CONST4:
2199 shader_addline(buffer, "c");
2200 offset = shader_get_float_offset(reg->type, offset);
2201 break;
2202
2203 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2204 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2205 break;
2206
2207 case WINED3DSPR_RASTOUT:
2208 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2209 break;
2210
2211 case WINED3DSPR_COLOROUT:
2212 shader_addline(buffer, "oC");
2213 break;
2214
2215 case WINED3DSPR_DEPTHOUT:
2216 shader_addline(buffer, "oDepth");
2217 break;
2218
2219 case WINED3DSPR_DEPTHOUTGE:
2220 shader_addline(buffer, "oDepthGE");
2221 break;
2222
2223 case WINED3DSPR_DEPTHOUTLE:
2224 shader_addline(buffer, "oDepthLE");
2225 break;
2226
2227 case WINED3DSPR_ATTROUT:
2228 shader_addline(buffer, "oD");
2229 break;
2230
2231 case WINED3DSPR_TEXCRDOUT:
2232 /* Vertex shaders >= 3.0 use general purpose output registers
2233 * (WINED3DSPR_OUTPUT), which can include an address token. */
2234 if (shader_version->major >= 3)
2235 shader_addline(buffer, "o");
2236 else
2237 shader_addline(buffer, "oT");
2238 break;
2239
2240 case WINED3DSPR_CONSTINT:
2241 shader_addline(buffer, "i");
2242 break;
2243
2244 case WINED3DSPR_CONSTBOOL:
2245 shader_addline(buffer, "b");
2246 break;
2247
2248 case WINED3DSPR_LABEL:
2249 shader_addline(buffer, "l");
2250 break;
2251
2252 case WINED3DSPR_LOOP:
2253 shader_addline(buffer, "aL");
2254 break;
2255
2256 case WINED3DSPR_SAMPLER:
2257 shader_addline(buffer, "s");
2258 break;
2259
2260 case WINED3DSPR_MISCTYPE:
2261 if (offset > 1)
2262 {
2263 FIXME("Unhandled misctype register %u.\n", offset);
2264 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2265 }
2266 else
2267 {
2268 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2269 }
2270 break;
2271
2272 case WINED3DSPR_PREDICATE:
2273 shader_addline(buffer, "p");
2274 break;
2275
2276 case WINED3DSPR_IMMCONST:
2277 shader_addline(buffer, "l");
2278 break;
2279
2280 case WINED3DSPR_CONSTBUFFER:
2281 shader_addline(buffer, "cb");
2282 break;
2283
2284 case WINED3DSPR_IMMCONSTBUFFER:
2285 shader_addline(buffer, "icb");
2286 break;
2287
2288 case WINED3DSPR_PRIMID:
2289 shader_addline(buffer, "primID");
2290 break;
2291
2292 case WINED3DSPR_NULL:
2293 shader_addline(buffer, "null");
2294 break;
2295
2296 case WINED3DSPR_RESOURCE:
2297 shader_addline(buffer, "t");
2298 break;
2299
2300 case WINED3DSPR_UAV:
2301 shader_addline(buffer, "u");
2302 break;
2303
2304 case WINED3DSPR_OUTPOINTID:
2305 shader_addline(buffer, "vOutputControlPointID");
2306 break;
2307
2308 case WINED3DSPR_FORKINSTID:
2309 shader_addline(buffer, "vForkInstanceId");
2310 break;
2311
2312 case WINED3DSPR_JOININSTID:
2313 shader_addline(buffer, "vJoinInstanceId");
2314 break;
2315
2316 case WINED3DSPR_INCONTROLPOINT:
2317 shader_addline(buffer, "vicp");
2318 break;
2319
2320 case WINED3DSPR_OUTCONTROLPOINT:
2321 shader_addline(buffer, "vocp");
2322 break;
2323
2324 case WINED3DSPR_PATCHCONST:
2325 shader_addline(buffer, "vpc");
2326 break;
2327
2328 case WINED3DSPR_TESSCOORD:
2329 shader_addline(buffer, "vDomainLocation");
2330 break;
2331
2332 case WINED3DSPR_GROUPSHAREDMEM:
2333 shader_addline(buffer, "g");
2334 break;
2335
2336 case WINED3DSPR_THREADID:
2337 shader_addline(buffer, "vThreadID");
2338 break;
2339
2340 case WINED3DSPR_THREADGROUPID:
2341 shader_addline(buffer, "vThreadGroupID");
2342 break;
2343
2344 case WINED3DSPR_LOCALTHREADID:
2345 shader_addline(buffer, "vThreadIDInGroup");
2346 break;
2347
2348 case WINED3DSPR_LOCALTHREADINDEX:
2349 shader_addline(buffer, "vThreadIDInGroupFlattened");
2350 break;
2351
2352 case WINED3DSPR_IDXTEMP:
2353 shader_addline(buffer, "x");
2354 break;
2355
2356 case WINED3DSPR_STREAM:
2357 shader_addline(buffer, "m");
2358 break;
2359
2360 case WINED3DSPR_FUNCTIONBODY:
2361 shader_addline(buffer, "fb");
2362 break;
2363
2364 case WINED3DSPR_FUNCTIONPOINTER:
2365 shader_addline(buffer, "fp");
2366 break;
2367
2368 case WINED3DSPR_COVERAGE:
2369 shader_addline(buffer, "vCoverage");
2370 break;
2371
2372 case WINED3DSPR_SAMPLEMASK:
2373 shader_addline(buffer, "oMask");
2374 break;
2375
2376 case WINED3DSPR_GSINSTID:
2377 shader_addline(buffer, "vGSInstanceID");
2378 break;
2379
2380 default:
2381 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2382 break;
2383 }
2384
2385 if (reg->type == WINED3DSPR_IMMCONST)
2386 {
2387 shader_addline(buffer, "(");
2388 switch (reg->immconst_type)
2389 {
2390 case WINED3D_IMMCONST_SCALAR:
2391 switch (reg->data_type)
2392 {
2393 case WINED3D_DATA_FLOAT:
2394 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2395 break;
2396 case WINED3D_DATA_INT:
2397 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2398 break;
2399 case WINED3D_DATA_RESOURCE:
2400 case WINED3D_DATA_SAMPLER:
2401 case WINED3D_DATA_UINT:
2402 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2403 break;
2404 default:
2405 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2406 break;
2407 }
2408 break;
2409
2410 case WINED3D_IMMCONST_VEC4:
2411 switch (reg->data_type)
2412 {
2413 case WINED3D_DATA_FLOAT:
2414 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2415 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2416 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2417 break;
2418 case WINED3D_DATA_INT:
2419 shader_addline(buffer, "%d, %d, %d, %d",
2420 reg->u.immconst_data[0], reg->u.immconst_data[1],
2421 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2422 break;
2423 case WINED3D_DATA_RESOURCE:
2424 case WINED3D_DATA_SAMPLER:
2425 case WINED3D_DATA_UINT:
2426 shader_addline(buffer, "%u, %u, %u, %u",
2427 reg->u.immconst_data[0], reg->u.immconst_data[1],
2428 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2429 break;
2430 default:
2431 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2432 break;
2433 }
2434 break;
2435
2436 default:
2437 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2438 break;
2439 }
2440 shader_addline(buffer, ")");
2441 }
2442 else if (reg->type != WINED3DSPR_RASTOUT
2443 && reg->type != WINED3DSPR_MISCTYPE
2444 && reg->type != WINED3DSPR_NULL)
2445 {
2446 if (offset != ~0u)
2447 {
2448 shader_addline(buffer, "[");
2449 if (reg->idx[0].rel_addr)
2450 {
2451 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2452 shader_addline(buffer, " + ");
2453 }
2454 shader_addline(buffer, "%u]", offset);
2455
2456 if (reg->idx[1].offset != ~0u)
2457 {
2458 shader_addline(buffer, "[");
2459 if (reg->idx[1].rel_addr)
2460 {
2461 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2462 shader_addline(buffer, " + ");
2463 }
2464 shader_addline(buffer, "%u]", reg->idx[1].offset);
2465 }
2466 }
2467
2468 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2469 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2470 }
2471 }
2472
2473 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2474 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2475 {
2476 DWORD write_mask = param->write_mask;
2477
2478 shader_dump_register(buffer, &param->reg, shader_version);
2479
2480 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2481 {
2482 static const char write_mask_chars[] = "xyzw";
2483
2484 shader_addline(buffer, ".");
2485 if (write_mask & WINED3DSP_WRITEMASK_0)
2486 shader_addline(buffer, "%c", write_mask_chars[0]);
2487 if (write_mask & WINED3DSP_WRITEMASK_1)
2488 shader_addline(buffer, "%c", write_mask_chars[1]);
2489 if (write_mask & WINED3DSP_WRITEMASK_2)
2490 shader_addline(buffer, "%c", write_mask_chars[2]);
2491 if (write_mask & WINED3DSP_WRITEMASK_3)
2492 shader_addline(buffer, "%c", write_mask_chars[3]);
2493 }
2494 }
2495
2496 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2497 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2498 {
2499 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2500 DWORD swizzle = param->swizzle;
2501
2502 if (src_modifier == WINED3DSPSM_NEG
2503 || src_modifier == WINED3DSPSM_BIASNEG
2504 || src_modifier == WINED3DSPSM_SIGNNEG
2505 || src_modifier == WINED3DSPSM_X2NEG
2506 || src_modifier == WINED3DSPSM_ABSNEG)
2507 shader_addline(buffer, "-");
2508 else if (src_modifier == WINED3DSPSM_COMP)
2509 shader_addline(buffer, "1-");
2510 else if (src_modifier == WINED3DSPSM_NOT)
2511 shader_addline(buffer, "!");
2512
2513 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2514 shader_addline(buffer, "abs(");
2515
2516 shader_dump_register(buffer, &param->reg, shader_version);
2517
2518 switch (src_modifier)
2519 {
2520 case WINED3DSPSM_NONE: break;
2521 case WINED3DSPSM_NEG: break;
2522 case WINED3DSPSM_NOT: break;
2523 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2524 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2525 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2526 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2527 case WINED3DSPSM_COMP: break;
2528 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2529 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2530 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2531 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2532 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2533 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2534 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2535 }
2536
2537 if (swizzle != WINED3DSP_NOSWIZZLE)
2538 {
2539 static const char swizzle_chars[] = "xyzw";
2540 DWORD swizzle_x = swizzle & 0x03;
2541 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2542 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2543 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2544
2545 if (swizzle_x == swizzle_y
2546 && swizzle_x == swizzle_z
2547 && swizzle_x == swizzle_w)
2548 {
2549 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2550 }
2551 else
2552 {
2553 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2554 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2555 }
2556 }
2557 }
2558
2559 /* Shared code in order to generate the bulk of the shader string. */
2560 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2561 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2562 const DWORD *start, const DWORD *end)
2563 {
2564 struct wined3d_device *device = shader->device;
2565 const struct wined3d_shader_frontend *fe = shader->frontend;
2566 void *fe_data = shader->frontend_data;
2567 struct wined3d_shader_version shader_version;
2568 struct wined3d_shader_parser_state state;
2569 struct wined3d_shader_instruction ins;
2570 struct wined3d_shader_tex_mx tex_mx;
2571 struct wined3d_shader_context ctx;
2572 const DWORD *ptr;
2573
2574 /* Initialize current parsing state. */
2575 tex_mx.current_row = 0;
2576 state.current_loop_depth = 0;
2577 state.current_loop_reg = 0;
2578 state.in_subroutine = FALSE;
2579
2580 ctx.shader = shader;
2581 ctx.gl_info = &device->adapter->gl_info;
2582 ctx.reg_maps = reg_maps;
2583 ctx.buffer = buffer;
2584 ctx.tex_mx = &tex_mx;
2585 ctx.state = &state;
2586 ctx.backend_data = backend_ctx;
2587 ins.ctx = &ctx;
2588
2589 fe->shader_read_header(fe_data, &ptr, &shader_version);
2590 if (start)
2591 ptr = start;
2592
2593 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2594 {
2595 /* Read opcode. */
2596 fe->shader_read_instruction(fe_data, &ptr, &ins);
2597
2598 /* Unknown opcode and its parameters. */
2599 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2600 {
2601 WARN("Encountered unrecognised or invalid instruction.\n");
2602 return WINED3DERR_INVALIDCALL;
2603 }
2604
2605 if (ins.predicate)
2606 FIXME("Predicates not implemented.\n");
2607
2608 /* Call appropriate function for output target */
2609 device->shader_backend->shader_handle_instruction(&ins);
2610 }
2611
2612 return WINED3D_OK;
2613 }
2614
2615 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2616 const struct wined3d_shader_dst_param *dst)
2617 {
2618 DWORD mmask = dst->modifiers;
2619
2620 switch (dst->shift)
2621 {
2622 case 0: break;
2623 case 13: shader_addline(buffer, "_d8"); break;
2624 case 14: shader_addline(buffer, "_d4"); break;
2625 case 15: shader_addline(buffer, "_d2"); break;
2626 case 1: shader_addline(buffer, "_x2"); break;
2627 case 2: shader_addline(buffer, "_x4"); break;
2628 case 3: shader_addline(buffer, "_x8"); break;
2629 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2630 }
2631
2632 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2633 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2634 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2635
2636 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2637 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2638 }
2639
2640 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2641 const struct wined3d_shader_primitive_type *primitive_type)
2642 {
2643 switch (primitive_type->type)
2644 {
2645 case WINED3D_PT_UNDEFINED:
2646 shader_addline(buffer, "undefined");
2647 break;
2648 case WINED3D_PT_POINTLIST:
2649 shader_addline(buffer, "pointlist");
2650 break;
2651 case WINED3D_PT_LINELIST:
2652 shader_addline(buffer, "linelist");
2653 break;
2654 case WINED3D_PT_LINESTRIP:
2655 shader_addline(buffer, "linestrip");
2656 break;
2657 case WINED3D_PT_TRIANGLELIST:
2658 shader_addline(buffer, "trianglelist");
2659 break;
2660 case WINED3D_PT_TRIANGLESTRIP:
2661 shader_addline(buffer, "trianglestrip");
2662 break;
2663 case WINED3D_PT_TRIANGLEFAN:
2664 shader_addline(buffer, "trianglefan");
2665 break;
2666 case WINED3D_PT_LINELIST_ADJ:
2667 shader_addline(buffer, "linelist_adj");
2668 break;
2669 case WINED3D_PT_LINESTRIP_ADJ:
2670 shader_addline(buffer, "linestrip_adj");
2671 break;
2672 case WINED3D_PT_TRIANGLELIST_ADJ:
2673 shader_addline(buffer, "trianglelist_adj");
2674 break;
2675 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2676 shader_addline(buffer, "trianglestrip_adj");
2677 break;
2678 case WINED3D_PT_PATCH:
2679 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2680 break;
2681 default:
2682 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2683 break;
2684 }
2685 }
2686
2687 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2688 enum wined3d_shader_interpolation_mode interpolation_mode)
2689 {
2690 switch (interpolation_mode)
2691 {
2692 case WINED3DSIM_CONSTANT:
2693 shader_addline(buffer, "constant");
2694 break;
2695 case WINED3DSIM_LINEAR:
2696 shader_addline(buffer, "linear");
2697 break;
2698 case WINED3DSIM_LINEAR_CENTROID:
2699 shader_addline(buffer, "linear centroid");
2700 break;
2701 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2702 shader_addline(buffer, "linear noperspective");
2703 break;
2704 case WINED3DSIM_LINEAR_SAMPLE:
2705 shader_addline(buffer, "linear sample");
2706 break;
2707 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2708 shader_addline(buffer, "linear noperspective centroid");
2709 break;
2710 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2711 shader_addline(buffer, "linear noperspective sample");
2712 break;
2713 default:
2714 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2715 break;
2716 }
2717 }
2718
2719 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2720 {
2721 struct wined3d_shader_version shader_version;
2722 struct wined3d_string_buffer buffer;
2723 const char *type_prefix;
2724 const char *p, *q;
2725 const DWORD *ptr;
2726 DWORD i;
2727
2728 if (!string_buffer_init(&buffer))
2729 {
2730 ERR("Failed to initialize string buffer.\n");
2731 return;
2732 }
2733
2734 fe->shader_read_header(fe_data, &ptr, &shader_version);
2735
2736 TRACE("Parsing %p.\n", ptr);
2737
2738 switch (shader_version.type)
2739 {
2740 case WINED3D_SHADER_TYPE_VERTEX:
2741 type_prefix = "vs";
2742 break;
2743
2744 case WINED3D_SHADER_TYPE_HULL:
2745 type_prefix = "hs";
2746 break;
2747
2748 case WINED3D_SHADER_TYPE_DOMAIN:
2749 type_prefix = "ds";
2750 break;
2751
2752 case WINED3D_SHADER_TYPE_GEOMETRY:
2753 type_prefix = "gs";
2754 break;
2755
2756 case WINED3D_SHADER_TYPE_PIXEL:
2757 type_prefix = "ps";
2758 break;
2759
2760 case WINED3D_SHADER_TYPE_COMPUTE:
2761 type_prefix = "cs";
2762 break;
2763
2764 default:
2765 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2766 type_prefix = "unknown";
2767 break;
2768 }
2769
2770 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2771
2772 while (!fe->shader_is_end(fe_data, &ptr))
2773 {
2774 struct wined3d_shader_instruction ins;
2775
2776 fe->shader_read_instruction(fe_data, &ptr, &ins);
2777 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2778 {
2779 WARN("Skipping unrecognized instruction.\n");
2780 shader_addline(&buffer, "<unrecognized instruction>\n");
2781 continue;
2782 }
2783
2784 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2785 {
2786 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2787 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2788 shader_addline(&buffer, " ");
2789 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2790 }
2791 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2792 {
2793 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2794 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2795 shader_addline(&buffer, ", %s",
2796 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2797 }
2798 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2799 {
2800 shader_addline(&buffer, "%s fb%u",
2801 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2802 }
2803 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2804 {
2805 shader_addline(&buffer, "%s ft%u = {...}",
2806 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2807 }
2808 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2809 {
2810 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2811 shader_dump_global_flags(&buffer, ins.flags);
2812 }
2813 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2814 {
2815 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2816 ins.declaration.max_tessellation_factor);
2817 }
2818 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2819 {
2820 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2821 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2822 {
2823 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2824 ins.declaration.icb->data[4 * i + 0],
2825 ins.declaration.icb->data[4 * i + 1],
2826 ins.declaration.icb->data[4 * i + 2],
2827 ins.declaration.icb->data[4 * i + 3]);
2828 }
2829 shader_addline(&buffer, "}");
2830 }
2831 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2832 {
2833 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2834 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2835 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2836 }
2837 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2838 {
2839 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2840 ins.declaration.indexable_temp.register_idx,
2841 ins.declaration.indexable_temp.register_size,
2842 ins.declaration.indexable_temp.component_count);
2843 }
2844 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2845 {
2846 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2847 shader_dump_interpolation_mode(&buffer, ins.flags);
2848 shader_addline(&buffer, " ");
2849 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2850 }
2851 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2852 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2853 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2854 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2855 {
2856 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2857 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2858 shader_addline(&buffer, ", ");
2859 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2860 }
2861 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2862 {
2863 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2864 shader_dump_interpolation_mode(&buffer, ins.flags);
2865 shader_addline(&buffer, " ");
2866 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2867 shader_addline(&buffer, ", ");
2868 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2869 }
2870 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2871 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2872 {
2873 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2874 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2875 }
2876 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2877 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2878 {
2879 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2880 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2881 }
2882 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2883 {
2884 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2885 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2886 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2887 }
2888 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2889 {
2890 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2891 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2892 }
2893 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2894 {
2895 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2896 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2897 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2898 }
2899 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2900 {
2901 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2902 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2903 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2904 shader_addline(&buffer, ", comparisonMode");
2905 }
2906 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2907 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2908 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2909 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2910 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2911 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2912 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2913 {
2914 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2915 }
2916 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2917 {
2918 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2919 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2920 }
2921 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2922 {
2923 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2924 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2925 }
2926 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2927 {
2928 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2929 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2930 }
2931 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2932 {
2933 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2934 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2935 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2936 }
2937 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2938 {
2939 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2940 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2941 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2942 ins.declaration.tgsm_structured.structure_count);
2943 }
2944 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2945 {
2946 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2947 ins.declaration.thread_group_size.x,
2948 ins.declaration.thread_group_size.y,
2949 ins.declaration.thread_group_size.z);
2950 }
2951 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2952 {
2953 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2954 shader_dump_uav_flags(&buffer, ins.flags);
2955 shader_addline(&buffer, " ");
2956 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2957 }
2958 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2959 {
2960 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2961 shader_dump_uav_flags(&buffer, ins.flags);
2962 shader_addline(&buffer, " ");
2963 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2964 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2965 }
2966 else if (ins.handler_idx == WINED3DSIH_DEF)
2967 {
2968 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2969 ins.dst[0].reg.idx[0].offset),
2970 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2971 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2972 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2973 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2974 }
2975 else if (ins.handler_idx == WINED3DSIH_DEFI)
2976 {
2977 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2978 ins.src[0].reg.u.immconst_data[0],
2979 ins.src[0].reg.u.immconst_data[1],
2980 ins.src[0].reg.u.immconst_data[2],
2981 ins.src[0].reg.u.immconst_data[3]);
2982 }
2983 else if (ins.handler_idx == WINED3DSIH_DEFB)
2984 {
2985 shader_addline(&buffer, "defb b%u = %s",
2986 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2987 }
2988 else
2989 {
2990 if (ins.predicate)
2991 {
2992 shader_addline(&buffer, "(");
2993 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2994 shader_addline(&buffer, ") ");
2995 }
2996
2997 /* PixWin marks instructions with the coissue flag with a '+' */
2998 if (ins.coissue)
2999 shader_addline(&buffer, "+");
3000
3001 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
3002
3003 if (ins.handler_idx == WINED3DSIH_BREAKP
3004 || ins.handler_idx == WINED3DSIH_CONTINUEP
3005 || ins.handler_idx == WINED3DSIH_IF
3006 || ins.handler_idx == WINED3DSIH_RETP
3007 || ins.handler_idx == WINED3DSIH_TEXKILL)
3008 {
3009 switch (ins.flags)
3010 {
3011 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
3012 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
3013 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
3014 }
3015 }
3016 else if (ins.handler_idx == WINED3DSIH_IFC
3017 || ins.handler_idx == WINED3DSIH_BREAKC)
3018 {
3019 switch (ins.flags)
3020 {
3021 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
3022 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
3023 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
3024 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
3025 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
3026 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
3027 default: shader_addline(&buffer, "_(%u)", ins.flags);
3028 }
3029 }
3030 else if (ins.handler_idx == WINED3DSIH_TEX
3031 && shader_version.major >= 2
3032 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
3033 {
3034 shader_addline(&buffer, "p");
3035 }
3036 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
3037 {
3038 switch (ins.flags)
3039 {
3040 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
3041 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
3042 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3043 }
3044 }
3045 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
3046 {
3047 switch (ins.flags)
3048 {
3049 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
3050 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
3051 }
3052 }
3053 else if (ins.handler_idx == WINED3DSIH_SYNC)
3054 {
3055 shader_dump_sync_flags(&buffer, ins.flags);
3056 }
3057
3058 if (wined3d_shader_instruction_has_texel_offset(&ins))
3059 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
3060
3061 for (i = 0; i < ins.dst_count; ++i)
3062 {
3063 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
3064 shader_addline(&buffer, !i ? " " : ", ");
3065 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
3066 }
3067
3068 /* Other source tokens */
3069 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
3070 {
3071 shader_addline(&buffer, !i ? " " : ", ");
3072 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
3073 }
3074 }
3075 shader_addline(&buffer, "\n");
3076 }
3077
3078 for (p = buffer.buffer; *p; p = q)
3079 {
3080 if (!(q = strstr(p, "\n")))
3081 q = p + strlen(p);
3082 else
3083 ++q;
3084 TRACE(" %.*s", (int)(q - p), p);
3085 }
3086
3087 string_buffer_free(&buffer);
3088 }
3089
3090 static void shader_cleanup(struct wined3d_shader *shader)
3091 {
3092 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
3093 {
3094 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
3095 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.fork);
3096 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.join);
3097 }
3098 else if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
3099 {
3100 HeapFree(GetProcessHeap(), 0, shader->u.gs.so_desc.elements);
3101 }
3102
3103 HeapFree(GetProcessHeap(), 0, shader->patch_constant_signature.elements);
3104 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3105 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3106 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3107 shader->device->shader_backend->shader_destroy(shader);
3108 shader_cleanup_reg_maps(&shader->reg_maps);
3109 HeapFree(GetProcessHeap(), 0, shader->function);
3110 shader_delete_constant_list(&shader->constantsF);
3111 shader_delete_constant_list(&shader->constantsB);
3112 shader_delete_constant_list(&shader->constantsI);
3113 list_remove(&shader->shader_list_entry);
3114
3115 if (shader->frontend && shader->frontend_data)
3116 shader->frontend->shader_free(shader->frontend_data);
3117 }
3118
3119 struct shader_none_priv
3120 {
3121 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3122 const struct fragment_pipeline *fragment_pipe;
3123 BOOL ffp_proj_control;
3124 };
3125
3126 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3127 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3128 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3129 const struct wined3d_state *state) {}
3130 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3131 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3132 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3133 const struct wined3d_state *state) {}
3134 static void shader_none_destroy(struct wined3d_shader *shader) {}
3135 static void shader_none_free_context_data(struct wined3d_context *context) {}
3136 static void shader_none_init_context_state(struct wined3d_context *context) {}
3137
3138 /* Context activation is done by the caller. */
3139 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3140 const struct wined3d_state *state)
3141 {
3142 const struct wined3d_gl_info *gl_info = context->gl_info;
3143 struct shader_none_priv *priv = shader_priv;
3144
3145 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
3146 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
3147 }
3148
3149 /* Context activation is done by the caller. */
3150 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3151 {
3152 struct shader_none_priv *priv = shader_priv;
3153 const struct wined3d_gl_info *gl_info = context->gl_info;
3154
3155 priv->vertex_pipe->vp_enable(gl_info, FALSE);
3156 priv->fragment_pipe->enable_extension(gl_info, FALSE);
3157
3158 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3159 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3160 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3161 | (1u << WINED3D_SHADER_TYPE_HULL)
3162 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3163 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3164 }
3165
3166 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3167 const struct fragment_pipeline *fragment_pipe)
3168 {
3169 struct fragment_caps fragment_caps;
3170 void *vertex_priv, *fragment_priv;
3171 struct shader_none_priv *priv;
3172
3173 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
3174 return E_OUTOFMEMORY;
3175
3176 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3177 {
3178 ERR("Failed to initialize vertex pipe.\n");
3179 HeapFree(GetProcessHeap(), 0, priv);
3180 return E_FAIL;
3181 }
3182
3183 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3184 {
3185 ERR("Failed to initialize fragment pipe.\n");
3186 vertex_pipe->vp_free(device);
3187 HeapFree(GetProcessHeap(), 0, priv);
3188 return E_FAIL;
3189 }
3190
3191 priv->vertex_pipe = vertex_pipe;
3192 priv->fragment_pipe = fragment_pipe;
3193 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
3194 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3195
3196 device->vertex_priv = vertex_priv;
3197 device->fragment_priv = fragment_priv;
3198 device->shader_priv = priv;
3199
3200 return WINED3D_OK;
3201 }
3202
3203 static void shader_none_free(struct wined3d_device *device)
3204 {
3205 struct shader_none_priv *priv = device->shader_priv;
3206
3207 priv->fragment_pipe->free_private(device);
3208 priv->vertex_pipe->vp_free(device);
3209 HeapFree(GetProcessHeap(), 0, priv);
3210 }
3211
3212 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3213 {
3214 return TRUE;
3215 }
3216
3217 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
3218 {
3219 /* Set the shader caps to 0 for the none shader backend */
3220 caps->vs_version = 0;
3221 caps->hs_version = 0;
3222 caps->ds_version = 0;
3223 caps->gs_version = 0;
3224 caps->ps_version = 0;
3225 caps->cs_version = 0;
3226 caps->vs_uniform_count = 0;
3227 caps->ps_uniform_count = 0;
3228 caps->ps_1x_max_value = 0.0f;
3229 caps->varying_count = 0;
3230 caps->wined3d_caps = 0;
3231 }
3232
3233 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3234 {
3235 /* We "support" every possible fixup, since we don't support any shader
3236 * model, and will never have to actually sample a texture. */
3237 return TRUE;
3238 }
3239
3240 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3241 {
3242 struct shader_none_priv *priv = shader_priv;
3243
3244 return priv->ffp_proj_control;
3245 }
3246
3247 const struct wined3d_shader_backend_ops none_shader_backend =
3248 {
3249 shader_none_handle_instruction,
3250 shader_none_precompile,
3251 shader_none_select,
3252 shader_none_select_compute,
3253 shader_none_disable,
3254 shader_none_update_float_vertex_constants,
3255 shader_none_update_float_pixel_constants,
3256 shader_none_load_constants,
3257 shader_none_destroy,
3258 shader_none_alloc,
3259 shader_none_free,
3260 shader_none_allocate_context_data,
3261 shader_none_free_context_data,
3262 shader_none_init_context_state,
3263 shader_none_get_caps,
3264 shader_none_color_fixup_supported,
3265 shader_none_has_ffp_proj_control,
3266 };
3267
3268 static HRESULT shader_set_function(struct wined3d_shader *shader, DWORD float_const_count,
3269 enum wined3d_shader_type type, unsigned int max_version)
3270 {
3271 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3272 const struct wined3d_shader_frontend *fe;
3273 HRESULT hr;
3274 unsigned int backend_version;
3275 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3276
3277 TRACE("shader %p, float_const_count %u, type %#x, max_version %u.\n",
3278 shader, float_const_count, type, max_version);
3279
3280 fe = shader->frontend;
3281 if (!(shader->frontend_data = fe->shader_init(shader->function,
3282 shader->functionLength, &shader->output_signature)))
3283 {
3284 FIXME("Failed to initialize frontend.\n");
3285 return WINED3DERR_INVALIDCALL;
3286 }
3287
3288 /* First pass: trace shader. */
3289 if (TRACE_ON(d3d_shader))
3290 shader_trace_init(fe, shader->frontend_data);
3291
3292 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3293 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
3294 &shader->output_signature, float_const_count)))
3295 return hr;
3296
3297 if (reg_maps->shader_version.type != type)
3298 {
3299 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
3300 return WINED3DERR_INVALIDCALL;
3301 }
3302 if (reg_maps->shader_version.major > max_version)
3303 {
3304 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
3305 return WINED3DERR_INVALIDCALL;
3306 }
3307 switch (type)
3308 {
3309 case WINED3D_SHADER_TYPE_VERTEX:
3310 backend_version = d3d_info->limits.vs_version;
3311 break;
3312 case WINED3D_SHADER_TYPE_HULL:
3313 backend_version = d3d_info->limits.hs_version;
3314 break;
3315 case WINED3D_SHADER_TYPE_DOMAIN:
3316 backend_version = d3d_info->limits.ds_version;
3317 break;
3318 case WINED3D_SHADER_TYPE_GEOMETRY:
3319 backend_version = d3d_info->limits.gs_version;
3320 break;
3321 case WINED3D_SHADER_TYPE_PIXEL:
3322 backend_version = d3d_info->limits.ps_version;
3323 break;
3324 case WINED3D_SHADER_TYPE_COMPUTE:
3325 backend_version = d3d_info->limits.cs_version;
3326 break;
3327 default:
3328 FIXME("No backend version-checking for this shader type.\n");
3329 backend_version = 0;
3330 }
3331 if (reg_maps->shader_version.major > backend_version)
3332 {
3333 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
3334 reg_maps->shader_version.major, reg_maps->shader_version.minor);
3335 return WINED3DERR_INVALIDCALL;
3336 }
3337
3338 return WINED3D_OK;
3339 }
3340
3341 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3342 {
3343 ULONG refcount = InterlockedIncrement(&shader->ref);
3344
3345 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3346
3347 return refcount;
3348 }
3349
3350 static void wined3d_shader_init_object(void *object)
3351 {
3352 struct wined3d_shader *shader = object;
3353 struct wined3d_device *device = shader->device;
3354
3355 list_add_head(&device->shaders, &shader->shader_list_entry);
3356
3357 device->shader_backend->shader_precompile(device->shader_priv, shader);
3358 }
3359
3360 static void wined3d_shader_destroy_object(void *object)
3361 {
3362 shader_cleanup(object);
3363 HeapFree(GetProcessHeap(), 0, object);
3364 }
3365
3366 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3367 {
3368 ULONG refcount = InterlockedDecrement(&shader->ref);
3369
3370 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3371
3372 if (!refcount)
3373 {
3374 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3375 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3376 }
3377
3378 return refcount;
3379 }
3380
3381 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3382 {
3383 TRACE("shader %p.\n", shader);
3384
3385 return shader->parent;
3386 }
3387
3388 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3389 void *byte_code, UINT *byte_code_size)
3390 {
3391 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3392
3393 if (!byte_code)
3394 {
3395 *byte_code_size = shader->functionLength;
3396 return WINED3D_OK;
3397 }
3398
3399 if (*byte_code_size < shader->functionLength)
3400 {
3401 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3402 * than the required size we should write the required size and
3403 * return D3DERR_MOREDATA. That's not actually true. */
3404 return WINED3DERR_INVALIDCALL;
3405 }
3406
3407 memcpy(byte_code, shader->function, shader->functionLength);
3408
3409 return WINED3D_OK;
3410 }
3411
3412 /* Set local constants for d3d8 shaders. */
3413 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3414 UINT start_idx, const float *src_data, UINT count)
3415 {
3416 UINT end_idx = start_idx + count;
3417 UINT i;
3418
3419 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3420
3421 if (end_idx > shader->limits->constant_float)
3422 {
3423 WARN("end_idx %u > float constants limit %u.\n",
3424 end_idx, shader->limits->constant_float);
3425 end_idx = shader->limits->constant_float;
3426 }
3427
3428 for (i = start_idx; i < end_idx; ++i)
3429 {
3430 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
3431 float *value;
3432 if (!lconst)
3433 return E_OUTOFMEMORY;
3434
3435 lconst->idx = i;
3436 value = (float *)lconst->value;
3437 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3438 list_add_head(&shader->constantsF, &lconst->entry);
3439
3440 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3441 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3442 {
3443 shader->lconst_inf_or_nan = TRUE;
3444 }
3445 }
3446
3447 return WINED3D_OK;
3448 }
3449
3450 static void init_interpolation_compile_args(DWORD *interpolation_args,
3451 const struct wined3d_shader *pixel_shader, const struct wined3d_gl_info *gl_info)
3452 {
3453 if (!needs_interpolation_qualifiers_for_shader_outputs(gl_info)
3454 || !pixel_shader || pixel_shader->reg_maps.shader_version.major < 4)
3455 {
3456 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
3457 return;
3458 }
3459
3460 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
3461 sizeof(pixel_shader->u.ps.interpolation_mode));
3462 }
3463
3464 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3465 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_context *context)
3466 {
3467 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3468 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3469 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3470 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3471 const struct wined3d_gl_info *gl_info = context->gl_info;
3472
3473 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3474 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3475 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3476 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3477 args->point_size = state->gl_primitive_type == GL_POINTS;
3478 args->per_vertex_point_size = shader->reg_maps.point_size;
3479 args->next_shader_type = hull_shader? WINED3D_SHADER_TYPE_HULL
3480 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3481 if (shader->reg_maps.shader_version.major >= 4)
3482 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
3483 : geometry_shader ? geometry_shader->limits->packed_input
3484 : pixel_shader ? pixel_shader->limits->packed_input : 0;
3485 else
3486 args->next_shader_input_count = 0;
3487 args->swizzle_map = swizzle_map;
3488 if (d3d_info->emulated_flatshading)
3489 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3490 else
3491 args->flatshading = 0;
3492
3493 init_interpolation_compile_args(args->interpolation_mode,
3494 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, gl_info);
3495 }
3496
3497 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3498 {
3499 if (usage_idx1 != usage_idx2)
3500 return FALSE;
3501 if (usage1 == usage2)
3502 return TRUE;
3503 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3504 return TRUE;
3505 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3506 return TRUE;
3507
3508 return FALSE;
3509 }
3510
3511 BOOL vshader_get_input(const struct wined3d_shader *shader,
3512 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3513 {
3514 WORD map = shader->reg_maps.input_registers;
3515 unsigned int i;
3516
3517 for (i = 0; map; map >>= 1, ++i)
3518 {
3519 if (!(map & 1)) continue;
3520
3521 if (match_usage(shader->u.vs.attributes[i].usage,
3522 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3523 {
3524 *regnum = i;
3525 return TRUE;
3526 }
3527 }
3528 return FALSE;
3529 }
3530
3531 static HRESULT shader_signature_calculate_strings_length(const struct wined3d_shader_signature *signature,
3532 SIZE_T *total)
3533 {
3534 struct wined3d_shader_signature_element *e;
3535 unsigned int i;
3536 SIZE_T len;
3537
3538 for (i = 0; i < signature->element_count; ++i)
3539 {
3540 e = &signature->elements[i];
3541 len = strlen(e->semantic_name);
3542 if (len >= ~(SIZE_T)0 - *total)
3543 return E_OUTOFMEMORY;
3544
3545 *total += len + 1;
3546 }
3547 return WINED3D_OK;
3548 }
3549
3550 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3551 const struct wined3d_shader_signature *src, char **signature_strings)
3552 {
3553 struct wined3d_shader_signature_element *e;
3554 unsigned int i;
3555 SIZE_T len;
3556 char *ptr;
3557
3558 if (!src->element_count)
3559 return WINED3D_OK;
3560
3561 ptr = *signature_strings;
3562
3563 dst->element_count = src->element_count;
3564 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3565 return E_OUTOFMEMORY;
3566
3567 for (i = 0; i < src->element_count; ++i)
3568 {
3569 e = &src->elements[i];
3570 dst->elements[i] = *e;
3571
3572 len = strlen(e->semantic_name);
3573 memcpy(ptr, e->semantic_name, len + 1);
3574 dst->elements[i].semantic_name = ptr;
3575 ptr += len + 1;
3576 }
3577
3578 *signature_strings = ptr;
3579
3580 return WINED3D_OK;
3581 }
3582
3583 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3584 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3585 void *parent, const struct wined3d_parent_ops *parent_ops)
3586 {
3587 size_t byte_code_size;
3588 SIZE_T total;
3589 HRESULT hr;
3590 char *ptr;
3591
3592 TRACE("byte_code %p, byte_code_size %#lx, format %#x, max_version %#x.\n",
3593 desc->byte_code, (long)desc->byte_code_size, desc->format, desc->max_version);
3594
3595 if (!desc->byte_code)
3596 return WINED3DERR_INVALIDCALL;
3597
3598 if (!(shader->frontend = shader_select_frontend(desc->format)))
3599 {
3600 FIXME("Unable to find frontend for shader.\n");
3601 return WINED3DERR_INVALIDCALL;
3602 }
3603
3604 shader->ref = 1;
3605 shader->device = device;
3606 shader->parent = parent;
3607 shader->parent_ops = parent_ops;
3608
3609 total = 0;
3610 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->input_signature, &total)))
3611 return hr;
3612 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->output_signature, &total)))
3613 return hr;
3614 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->patch_constant_signature, &total)))
3615 return hr;
3616 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3617 return E_OUTOFMEMORY;
3618 ptr = shader->signature_strings;
3619
3620 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3621 {
3622 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3623 return hr;
3624 }
3625 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3626 {
3627 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3628 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3629 return hr;
3630 }
3631 if (FAILED(hr = shader_signature_copy(&shader->patch_constant_signature, &desc->patch_constant_signature, &ptr)))
3632 {
3633 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3634 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3635 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3636 return hr;
3637 }
3638
3639 list_init(&shader->linked_programs);
3640 list_init(&shader->constantsF);
3641 list_init(&shader->constantsB);
3642 list_init(&shader->constantsI);
3643 shader->lconst_inf_or_nan = FALSE;
3644 list_init(&shader->reg_maps.indexable_temps);
3645 list_init(&shader->shader_list_entry);
3646
3647 byte_code_size = desc->byte_code_size;
3648 if (byte_code_size == ~(size_t)0)
3649 {
3650 const struct wined3d_shader_frontend *fe = shader->frontend;
3651 struct wined3d_shader_version shader_version;
3652 struct wined3d_shader_instruction ins;
3653 const DWORD *ptr;
3654 void *fe_data;
3655
3656 if (!(fe_data = fe->shader_init(desc->byte_code, byte_code_size, &shader->output_signature)))
3657 {
3658 WARN("Failed to initialise frontend data.\n");
3659 shader_cleanup(shader);
3660 return WINED3DERR_INVALIDCALL;
3661 }
3662
3663 fe->shader_read_header(fe_data, &ptr, &shader_version);
3664 while (!fe->shader_is_end(fe_data, &ptr))
3665 fe->shader_read_instruction(fe_data, &ptr, &ins);
3666
3667 fe->shader_free(fe_data);
3668
3669 byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3670 }
3671
3672 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, byte_code_size)))
3673 {
3674 shader_cleanup(shader);
3675 return E_OUTOFMEMORY;
3676 }
3677 memcpy(shader->function, desc->byte_code, byte_code_size);
3678 shader->functionLength = byte_code_size;
3679
3680 if (FAILED(hr = shader_set_function(shader, float_const_count, type, desc->max_version)))
3681 {
3682 WARN("Failed to set function, hr %#x.\n", hr);
3683 shader_cleanup(shader);
3684 return hr;
3685 }
3686
3687 shader->load_local_constsF = shader->lconst_inf_or_nan;
3688
3689 wined3d_cs_init_object(shader->device->cs, wined3d_shader_init_object, shader);
3690
3691 return hr;
3692 }
3693
3694 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3695 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3696 {
3697 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3698 unsigned int i;
3699 HRESULT hr;
3700
3701 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3702 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3703 return hr;
3704
3705 for (i = 0; i < shader->input_signature.element_count; ++i)
3706 {
3707 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3708
3709 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3710 continue;
3711
3712 shader->u.vs.attributes[input->register_idx].usage =
3713 shader_usage_from_semantic_name(input->semantic_name);
3714 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3715 }
3716
3717 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3718 shader->load_local_constsF = TRUE;
3719
3720 return WINED3D_OK;
3721 }
3722
3723 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3724 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3725 void *parent, const struct wined3d_parent_ops *parent_ops)
3726 {
3727 struct wined3d_stream_output_element *elements = NULL;
3728 HRESULT hr;
3729
3730 if (so_desc && !(elements = wined3d_calloc(so_desc->element_count, sizeof(*elements))))
3731 return E_OUTOFMEMORY;
3732
3733 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
3734 {
3735 HeapFree(GetProcessHeap(), 0, elements);
3736 return hr;
3737 }
3738
3739 if (so_desc)
3740 {
3741 shader->u.gs.so_desc = *so_desc;
3742 shader->u.gs.so_desc.elements = elements;
3743 memcpy(elements, so_desc->elements, so_desc->element_count * sizeof(*elements));
3744 }
3745
3746 return WINED3D_OK;
3747 }
3748
3749 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3750 struct ds_compile_args *args, const struct wined3d_context *context)
3751 {
3752 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
3753 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3754 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3755 const struct wined3d_gl_info *gl_info = context->gl_info;
3756
3757 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3758 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3759
3760 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
3761 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3762 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
3763
3764 args->render_offscreen = context->render_offscreen;
3765
3766 init_interpolation_compile_args(args->interpolation_mode,
3767 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, gl_info);
3768
3769 args->padding = 0;
3770 }
3771
3772 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3773 struct gs_compile_args *args, const struct wined3d_context *context)
3774 {
3775 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
3776 const struct wined3d_gl_info *gl_info = context->gl_info;
3777
3778 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
3779
3780 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, gl_info);
3781 }
3782
3783 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3784 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3785 {
3786 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3787 const struct wined3d_gl_info *gl_info = context->gl_info;
3788 const struct wined3d_texture *texture;
3789 unsigned int i;
3790
3791 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3792 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3793 {
3794 static unsigned int warned = 0;
3795
3796 args->srgb_correction = 1;
3797 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3798 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3799 "support, expect rendering artifacts.\n");
3800 }
3801
3802 if (shader->reg_maps.shader_version.major == 1
3803 && shader->reg_maps.shader_version.minor <= 3)
3804 {
3805 for (i = 0; i < shader->limits->sampler; ++i)
3806 {
3807 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3808
3809 if (flags & WINED3D_TTFF_PROJECTED)
3810 {
3811 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3812
3813 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3814 {
3815 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3816 unsigned int j;
3817 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3818 DWORD max_valid = WINED3D_TTFF_COUNT4;
3819
3820 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3821 {
3822 struct wined3d_vertex_declaration_element *element =
3823 &state->vertex_declaration->elements[j];
3824
3825 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3826 && element->usage_idx == index)
3827 {
3828 max_valid = element->format->component_count;
3829 break;
3830 }
3831 }
3832 if (!tex_transform || tex_transform > max_valid)
3833 {
3834 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3835 tex_transform, max_valid);
3836 tex_transform = max_valid;
3837 }
3838 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3839 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3840 && tex_transform > WINED3D_TTFF_COUNT2)
3841 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3842 && tex_transform > WINED3D_TTFF_COUNT3))
3843 tex_transform |= WINED3D_PSARGS_PROJECTED;
3844 else
3845 {
3846 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3847 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3848 i, tex_transform, resource_type);
3849 }
3850 }
3851 else
3852 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3853
3854 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3855 }
3856 }
3857 }
3858 if (shader->reg_maps.shader_version.major == 1
3859 && shader->reg_maps.shader_version.minor <= 4)
3860 {
3861 for (i = 0; i < shader->limits->sampler; ++i)
3862 {
3863 const struct wined3d_texture *texture = state->textures[i];
3864
3865 if (!shader->reg_maps.resource_info[i].type)
3866 continue;
3867
3868 /* Treat unbound textures as 2D. The dummy texture will provide
3869 * the proper sample value. The tex_types bitmap defaults to
3870 * 2D because of the memset. */
3871 if (!texture)
3872 continue;
3873
3874 switch (texture->target)
3875 {
3876 /* RECT textures are distinguished from 2D textures via np2_fixup */
3877 case GL_TEXTURE_RECTANGLE_ARB:
3878 case GL_TEXTURE_2D:
3879 break;
3880
3881 case GL_TEXTURE_3D:
3882 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3883 break;
3884
3885 case GL_TEXTURE_CUBE_MAP_ARB:
3886 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3887 break;
3888 }
3889 }
3890 }
3891
3892 if (shader->reg_maps.shader_version.major >= 4)
3893 {
3894 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3895 args->shadow = 0;
3896 for (i = 0 ; i < MAX_FRAGMENT_SAMPLERS; ++i)
3897 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3898 args->np2_fixup = 0;
3899 }
3900 else
3901 {
3902 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3903 {
3904 if (!shader->reg_maps.resource_info[i].type)
3905 continue;
3906
3907 texture = state->textures[i];
3908 if (!texture)
3909 {
3910 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3911 continue;
3912 }
3913 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3914 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3915 else
3916 args->color_fixup[i] = texture->resource.format->color_fixup;
3917
3918 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3919 args->shadow |= 1u << i;
3920
3921 /* Flag samplers that need NP2 texcoord fixup. */
3922 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3923 args->np2_fixup |= (1u << i);
3924 }
3925 }
3926
3927 if (shader->reg_maps.shader_version.major >= 3)
3928 {
3929 if (position_transformed)
3930 args->vp_mode = pretransformed;
3931 else if (use_vs(state))
3932 args->vp_mode = vertexshader;
3933 else
3934 args->vp_mode = fixedfunction;
3935 args->fog = WINED3D_FFP_PS_FOG_OFF;
3936 }
3937 else
3938 {
3939 args->vp_mode = vertexshader;
3940 if (state->render_states[WINED3D_RS_FOGENABLE])
3941 {
3942 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3943 {
3944 case WINED3D_FOG_NONE:
3945 if (position_transformed || use_vs(state))
3946 {
3947 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3948 break;
3949 }
3950
3951 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3952 {
3953 case WINED3D_FOG_NONE: /* Fall through. */
3954 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3955 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3956 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3957 }
3958 break;
3959
3960 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3961 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3962 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3963 }
3964 }
3965 else
3966 {
3967 args->fog = WINED3D_FFP_PS_FOG_OFF;
3968 }
3969 }
3970
3971 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3972 {
3973 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3974
3975 args->texcoords_initialized = 0;
3976 for (i = 0; i < MAX_TEXTURES; ++i)
3977 {
3978 if (vs)
3979 {
3980 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3981 args->texcoords_initialized |= 1u << i;
3982 }
3983 else
3984 {
3985 const struct wined3d_stream_info *si = &context->stream_info;
3986 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3987
3988 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3989 & WINED3D_FFP_TCI_MASK
3990 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3991 args->texcoords_initialized |= 1u << i;
3992 }
3993 }
3994 }
3995 else
3996 {
3997 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3998 }
3999
4000 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
4001 && state->gl_primitive_type == GL_POINTS;
4002
4003 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
4004 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
4005 else
4006 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
4007 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
4008 : WINED3D_CMP_ALWAYS) - 1;
4009
4010 if (d3d_info->emulated_flatshading)
4011 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
4012
4013 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
4014 ? context->render_offscreen : 0;
4015 }
4016
4017 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
4018 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
4019 {
4020 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4021 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
4022 HRESULT hr;
4023
4024 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
4025 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
4026 return hr;
4027
4028 for (i = 0; i < MAX_REG_INPUT; ++i)
4029 {
4030 if (shader->u.ps.input_reg_used & (1u << i))
4031 {
4032 ++num_regs_used;
4033 highest_reg_used = i;
4034 }
4035 }
4036
4037 /* Don't do any register mapping magic if it is not needed, or if we can't
4038 * achieve anything anyway */
4039 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
4040 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
4041 || shader->reg_maps.shader_version.major >= 4)
4042 {
4043 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
4044 {
4045 /* This happens with relative addressing. The input mapper function
4046 * warns about this if the higher registers are declared too, so
4047 * don't write a FIXME here */
4048 WARN("More varying registers used than supported\n");
4049 }
4050
4051 for (i = 0; i < MAX_REG_INPUT; ++i)
4052 {
4053 shader->u.ps.input_reg_map[i] = i;
4054 }
4055
4056 shader->u.ps.declared_in_count = highest_reg_used + 1;
4057 }
4058 else
4059 {
4060 shader->u.ps.declared_in_count = 0;
4061 for (i = 0; i < MAX_REG_INPUT; ++i)
4062 {
4063 if (shader->u.ps.input_reg_used & (1u << i))
4064 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
4065 else shader->u.ps.input_reg_map[i] = ~0U;
4066 }
4067 }
4068
4069 return WINED3D_OK;
4070 }
4071
4072 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
4073 {
4074 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
4075 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
4076 unsigned int i;
4077
4078 if (reg_maps->shader_version.major != 1) return;
4079
4080 for (i = 0; i < shader->limits->sampler; ++i)
4081 {
4082 /* We don't sample from this sampler. */
4083 if (!resource_info[i].type)
4084 continue;
4085
4086 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
4087 {
4088 case WINED3D_SHADER_TEX_2D:
4089 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
4090 break;
4091
4092 case WINED3D_SHADER_TEX_3D:
4093 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
4094 break;
4095
4096 case WINED3D_SHADER_TEX_CUBE:
4097 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
4098 break;
4099 }
4100 }
4101 }
4102
4103 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4104 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4105 {
4106 struct wined3d_shader *object;
4107 HRESULT hr;
4108
4109 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4110 device, desc, parent, parent_ops, shader);
4111
4112 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4113 return E_OUTOFMEMORY;
4114
4115 if (FAILED(hr = shader_init(object, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops)))
4116 {
4117 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
4118 HeapFree(GetProcessHeap(), 0, object);
4119 return hr;
4120 }
4121
4122 TRACE("Created compute shader %p.\n", object);
4123 *shader = object;
4124
4125 return WINED3D_OK;
4126 }
4127
4128 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4129 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4130 {
4131 struct wined3d_shader *object;
4132 HRESULT hr;
4133
4134 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4135 device, desc, parent, parent_ops, shader);
4136
4137 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4138 return E_OUTOFMEMORY;
4139
4140 if (FAILED(hr = shader_init(object, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops)))
4141 {
4142 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4143 HeapFree(GetProcessHeap(), 0, object);
4144 return hr;
4145 }
4146
4147 TRACE("Created domain shader %p.\n", object);
4148 *shader = object;
4149
4150 return WINED3D_OK;
4151 }
4152
4153 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4154 const struct wined3d_stream_output_desc *so_desc, void *parent,
4155 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4156 {
4157 struct wined3d_shader *object;
4158 HRESULT hr;
4159
4160 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4161 device, desc, so_desc, parent, parent_ops, shader);
4162
4163 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4164 return E_OUTOFMEMORY;
4165
4166 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4167 {
4168 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4169 HeapFree(GetProcessHeap(), 0, object);
4170 return hr;
4171 }
4172
4173 TRACE("Created geometry shader %p.\n", object);
4174 *shader = object;
4175
4176 return WINED3D_OK;
4177 }
4178
4179 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4180 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4181 {
4182 struct wined3d_shader *object;
4183 HRESULT hr;
4184
4185 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4186 device, desc, parent, parent_ops, shader);
4187
4188 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4189 return E_OUTOFMEMORY;
4190
4191 if (FAILED(hr = shader_init(object, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops)))
4192 {
4193 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4194 HeapFree(GetProcessHeap(), 0, object);
4195 return hr;
4196 }
4197
4198 TRACE("Created hull shader %p.\n", object);
4199 *shader = object;
4200
4201 return WINED3D_OK;
4202 }
4203
4204 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4205 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4206 {
4207 struct wined3d_shader *object;
4208 HRESULT hr;
4209
4210 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4211 device, desc, parent, parent_ops, shader);
4212
4213 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4214 if (!object)
4215 return E_OUTOFMEMORY;
4216
4217 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4218 {
4219 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4220 HeapFree(GetProcessHeap(), 0, object);
4221 return hr;
4222 }
4223
4224 TRACE("Created pixel shader %p.\n", object);
4225 *shader = object;
4226
4227 return WINED3D_OK;
4228 }
4229
4230 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4231 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4232 {
4233 struct wined3d_shader *object;
4234 HRESULT hr;
4235
4236 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4237 device, desc, parent, parent_ops, shader);
4238
4239 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4240 if (!object)
4241 return E_OUTOFMEMORY;
4242
4243 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4244 {
4245 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4246 HeapFree(GetProcessHeap(), 0, object);
4247 return hr;
4248 }
4249
4250 TRACE("Created vertex shader %p.\n", object);
4251 *shader = object;
4252
4253 return WINED3D_OK;
4254 }
Windows API implementation