?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wmic: Use DOS line endings.
[wine.git] / dlls / wined3d / shader_sm4.c
blob5f9dbecfebaacc2d27786a43ac2b1a3497a0259e
1 /*
2 * Copyright 2009 Henri Verbeet for CodeWeavers
3 *
4 * This library is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2.1 of the License, or (at your option) any later version.
8 *
9 * This library is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
13 *
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with this library; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
17 */
18
19 #include "config.h"
20 #include "wine/port.h"
21
22 #include "wined3d_private.h"
23
24 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
25 WINE_DECLARE_DEBUG_CHANNEL(d3d_bytecode);
26
27 #define WINED3D_SM4_INSTRUCTION_MODIFIER (0x1u << 31)
28
29 #define WINED3D_SM4_MODIFIER_AOFFIMMI 0x1
30 #define WINED3D_SM4_AOFFIMMI_U_SHIFT 9
31 #define WINED3D_SM4_AOFFIMMI_U_MASK (0xfu << WINED3D_SM4_AOFFIMMI_U_SHIFT)
32 #define WINED3D_SM4_AOFFIMMI_V_SHIFT 13
33 #define WINED3D_SM4_AOFFIMMI_V_MASK (0xfu << WINED3D_SM4_AOFFIMMI_V_SHIFT)
34 #define WINED3D_SM4_AOFFIMMI_W_SHIFT 17
35 #define WINED3D_SM4_AOFFIMMI_W_MASK (0xfu << WINED3D_SM4_AOFFIMMI_W_SHIFT)
36
37 #define WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT 24
38 #define WINED3D_SM4_INSTRUCTION_LENGTH_MASK (0x1fu << WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT)
39
40 #define WINED3D_SM4_INSTRUCTION_FLAGS_SHIFT 11
41 #define WINED3D_SM4_INSTRUCTION_FLAGS_MASK (0x7u << WINED3D_SM4_INSTRUCTION_FLAGS_SHIFT)
42
43 #define WINED3D_SM4_RESOURCE_TYPE_SHIFT 11
44 #define WINED3D_SM4_RESOURCE_TYPE_MASK (0xfu << WINED3D_SM4_RESOURCE_TYPE_SHIFT)
45
46 #define WINED3D_SM4_PRIMITIVE_TYPE_SHIFT 11
47 #define WINED3D_SM4_PRIMITIVE_TYPE_MASK (0x3fu << WINED3D_SM4_PRIMITIVE_TYPE_SHIFT)
48
49 #define WINED3D_SM4_INDEX_TYPE_SHIFT 11
50 #define WINED3D_SM4_INDEX_TYPE_MASK (0x1u << WINED3D_SM4_INDEX_TYPE_SHIFT)
51
52 #define WINED3D_SM4_SAMPLER_MODE_SHIFT 11
53 #define WINED3D_SM4_SAMPLER_MODE_MASK (0xfu << WINED3D_SM4_SAMPLER_MODE_SHIFT)
54
55 #define WINED3D_SM4_SHADER_DATA_TYPE_SHIFT 11
56 #define WINED3D_SM4_SHADER_DATA_TYPE_MASK (0xfu << WINED3D_SM4_SHADER_DATA_TYPE_SHIFT)
57
58 #define WINED3D_SM4_INTERPOLATION_MODE_SHIFT 11
59 #define WINED3D_SM4_INTERPOLATION_MODE_MASK (0xfu << WINED3D_SM4_INTERPOLATION_MODE_SHIFT)
60
61 #define WINED3D_SM4_GLOBAL_FLAGS_SHIFT 11
62 #define WINED3D_SM4_GLOBAL_FLAGS_MASK (0xffu << WINED3D_SM4_GLOBAL_FLAGS_SHIFT)
63
64 #define WINED3D_SM5_PRECISE_SHIFT 19
65 #define WINED3D_SM5_PRECISE_MASK (0xfu << WINED3D_SM5_PRECISE_SHIFT)
66
67 #define WINED3D_SM5_CONTROL_POINT_COUNT_SHIFT 11
68 #define WINED3D_SM5_CONTROL_POINT_COUNT_MASK (0xffu << WINED3D_SM5_CONTROL_POINT_COUNT_SHIFT)
69
70 #define WINED3D_SM5_FP_ARRAY_SIZE_SHIFT 16
71 #define WINED3D_SM5_FP_TABLE_COUNT_MASK 0xffffu
72
73 #define WINED3D_SM5_UAV_FLAGS_SHIFT 15
74 #define WINED3D_SM5_UAV_FLAGS_MASK (0x1ffu << WINED3D_SM5_UAV_FLAGS_SHIFT)
75
76 #define WINED3D_SM5_SYNC_FLAGS_SHIFT 11
77 #define WINED3D_SM5_SYNC_FLAGS_MASK (0xffu << WINED3D_SM5_SYNC_FLAGS_SHIFT)
78
79 #define WINED3D_SM5_TESSELLATOR_SHIFT 11
80 #define WINED3D_SM5_TESSELLATOR_MASK (0xfu << WINED3D_SM5_TESSELLATOR_SHIFT)
81
82 #define WINED3D_SM4_OPCODE_MASK 0xff
83
84 #define WINED3D_SM4_REGISTER_MODIFIER (0x1u << 31)
85
86 #define WINED3D_SM4_ADDRESSING_SHIFT1 25
87 #define WINED3D_SM4_ADDRESSING_MASK1 (0x3u << WINED3D_SM4_ADDRESSING_SHIFT1)
88
89 #define WINED3D_SM4_ADDRESSING_SHIFT0 22
90 #define WINED3D_SM4_ADDRESSING_MASK0 (0x3u << WINED3D_SM4_ADDRESSING_SHIFT0)
91
92 #define WINED3D_SM4_REGISTER_ORDER_SHIFT 20
93 #define WINED3D_SM4_REGISTER_ORDER_MASK (0x3u << WINED3D_SM4_REGISTER_ORDER_SHIFT)
94
95 #define WINED3D_SM4_REGISTER_TYPE_SHIFT 12
96 #define WINED3D_SM4_REGISTER_TYPE_MASK (0xffu << WINED3D_SM4_REGISTER_TYPE_SHIFT)
97
98 #define WINED3D_SM4_SWIZZLE_TYPE_SHIFT 2
99 #define WINED3D_SM4_SWIZZLE_TYPE_MASK (0x3u << WINED3D_SM4_SWIZZLE_TYPE_SHIFT)
100
101 #define WINED3D_SM4_IMMCONST_TYPE_SHIFT 0
102 #define WINED3D_SM4_IMMCONST_TYPE_MASK (0x3u << WINED3D_SM4_IMMCONST_TYPE_SHIFT)
103
104 #define WINED3D_SM4_WRITEMASK_SHIFT 4
105 #define WINED3D_SM4_WRITEMASK_MASK (0xfu << WINED3D_SM4_WRITEMASK_SHIFT)
106
107 #define WINED3D_SM4_SWIZZLE_SHIFT 4
108 #define WINED3D_SM4_SWIZZLE_MASK (0xffu << WINED3D_SM4_SWIZZLE_SHIFT)
109
110 #define WINED3D_SM4_VERSION_MAJOR(version) (((version) >> 4) & 0xf)
111 #define WINED3D_SM4_VERSION_MINOR(version) (((version) >> 0) & 0xf)
112
113 #define WINED3D_SM4_ADDRESSING_RELATIVE 0x2
114 #define WINED3D_SM4_ADDRESSING_OFFSET 0x1
115
116 #define WINED3D_SM4_INSTRUCTION_FLAG_SATURATE 0x4
117
118 #define WINED3D_SM4_CONDITIONAL_NZ (0x1u << 18)
119
120 enum wined3d_sm4_opcode
121 {
122 WINED3D_SM4_OP_ADD = 0x00,
123 WINED3D_SM4_OP_AND = 0x01,
124 WINED3D_SM4_OP_BREAK = 0x02,
125 WINED3D_SM4_OP_BREAKC = 0x03,
126 WINED3D_SM4_OP_CASE = 0x06,
127 WINED3D_SM4_OP_CONTINUE = 0x07,
128 WINED3D_SM4_OP_CONTINUEC = 0x08,
129 WINED3D_SM4_OP_CUT = 0x09,
130 WINED3D_SM4_OP_DEFAULT = 0x0a,
131 WINED3D_SM4_OP_DERIV_RTX = 0x0b,
132 WINED3D_SM4_OP_DERIV_RTY = 0x0c,
133 WINED3D_SM4_OP_DISCARD = 0x0d,
134 WINED3D_SM4_OP_DIV = 0x0e,
135 WINED3D_SM4_OP_DP2 = 0x0f,
136 WINED3D_SM4_OP_DP3 = 0x10,
137 WINED3D_SM4_OP_DP4 = 0x11,
138 WINED3D_SM4_OP_ELSE = 0x12,
139 WINED3D_SM4_OP_EMIT = 0x13,
140 WINED3D_SM4_OP_ENDIF = 0x15,
141 WINED3D_SM4_OP_ENDLOOP = 0x16,
142 WINED3D_SM4_OP_ENDSWITCH = 0x17,
143 WINED3D_SM4_OP_EQ = 0x18,
144 WINED3D_SM4_OP_EXP = 0x19,
145 WINED3D_SM4_OP_FRC = 0x1a,
146 WINED3D_SM4_OP_FTOI = 0x1b,
147 WINED3D_SM4_OP_FTOU = 0x1c,
148 WINED3D_SM4_OP_GE = 0x1d,
149 WINED3D_SM4_OP_IADD = 0x1e,
150 WINED3D_SM4_OP_IF = 0x1f,
151 WINED3D_SM4_OP_IEQ = 0x20,
152 WINED3D_SM4_OP_IGE = 0x21,
153 WINED3D_SM4_OP_ILT = 0x22,
154 WINED3D_SM4_OP_IMAD = 0x23,
155 WINED3D_SM4_OP_IMAX = 0x24,
156 WINED3D_SM4_OP_IMIN = 0x25,
157 WINED3D_SM4_OP_IMUL = 0x26,
158 WINED3D_SM4_OP_INE = 0x27,
159 WINED3D_SM4_OP_INEG = 0x28,
160 WINED3D_SM4_OP_ISHL = 0x29,
161 WINED3D_SM4_OP_ISHR = 0x2a,
162 WINED3D_SM4_OP_ITOF = 0x2b,
163 WINED3D_SM4_OP_LABEL = 0x2c,
164 WINED3D_SM4_OP_LD = 0x2d,
165 WINED3D_SM4_OP_LD2DMS = 0x2e,
166 WINED3D_SM4_OP_LOG = 0x2f,
167 WINED3D_SM4_OP_LOOP = 0x30,
168 WINED3D_SM4_OP_LT = 0x31,
169 WINED3D_SM4_OP_MAD = 0x32,
170 WINED3D_SM4_OP_MIN = 0x33,
171 WINED3D_SM4_OP_MAX = 0x34,
172 WINED3D_SM4_OP_SHADER_DATA = 0x35,
173 WINED3D_SM4_OP_MOV = 0x36,
174 WINED3D_SM4_OP_MOVC = 0x37,
175 WINED3D_SM4_OP_MUL = 0x38,
176 WINED3D_SM4_OP_NE = 0x39,
177 WINED3D_SM4_OP_NOP = 0x3a,
178 WINED3D_SM4_OP_NOT = 0x3b,
179 WINED3D_SM4_OP_OR = 0x3c,
180 WINED3D_SM4_OP_RESINFO = 0x3d,
181 WINED3D_SM4_OP_RET = 0x3e,
182 WINED3D_SM4_OP_RETC = 0x3f,
183 WINED3D_SM4_OP_ROUND_NE = 0x40,
184 WINED3D_SM4_OP_ROUND_NI = 0x41,
185 WINED3D_SM4_OP_ROUND_PI = 0x42,
186 WINED3D_SM4_OP_ROUND_Z = 0x43,
187 WINED3D_SM4_OP_RSQ = 0x44,
188 WINED3D_SM4_OP_SAMPLE = 0x45,
189 WINED3D_SM4_OP_SAMPLE_C = 0x46,
190 WINED3D_SM4_OP_SAMPLE_C_LZ = 0x47,
191 WINED3D_SM4_OP_SAMPLE_LOD = 0x48,
192 WINED3D_SM4_OP_SAMPLE_GRAD = 0x49,
193 WINED3D_SM4_OP_SAMPLE_B = 0x4a,
194 WINED3D_SM4_OP_SQRT = 0x4b,
195 WINED3D_SM4_OP_SWITCH = 0x4c,
196 WINED3D_SM4_OP_SINCOS = 0x4d,
197 WINED3D_SM4_OP_UDIV = 0x4e,
198 WINED3D_SM4_OP_ULT = 0x4f,
199 WINED3D_SM4_OP_UGE = 0x50,
200 WINED3D_SM4_OP_UMUL = 0x51,
201 WINED3D_SM4_OP_UMAX = 0x53,
202 WINED3D_SM4_OP_UMIN = 0x54,
203 WINED3D_SM4_OP_USHR = 0x55,
204 WINED3D_SM4_OP_UTOF = 0x56,
205 WINED3D_SM4_OP_XOR = 0x57,
206 WINED3D_SM4_OP_DCL_RESOURCE = 0x58,
207 WINED3D_SM4_OP_DCL_CONSTANT_BUFFER = 0x59,
208 WINED3D_SM4_OP_DCL_SAMPLER = 0x5a,
209 WINED3D_SM4_OP_DCL_INDEX_RANGE = 0x5b,
210 WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY = 0x5c,
211 WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE = 0x5d,
212 WINED3D_SM4_OP_DCL_VERTICES_OUT = 0x5e,
213 WINED3D_SM4_OP_DCL_INPUT = 0x5f,
214 WINED3D_SM4_OP_DCL_INPUT_SGV = 0x60,
215 WINED3D_SM4_OP_DCL_INPUT_SIV = 0x61,
216 WINED3D_SM4_OP_DCL_INPUT_PS = 0x62,
217 WINED3D_SM4_OP_DCL_INPUT_PS_SGV = 0x63,
218 WINED3D_SM4_OP_DCL_INPUT_PS_SIV = 0x64,
219 WINED3D_SM4_OP_DCL_OUTPUT = 0x65,
220 WINED3D_SM4_OP_DCL_OUTPUT_SIV = 0x67,
221 WINED3D_SM4_OP_DCL_TEMPS = 0x68,
222 WINED3D_SM4_OP_DCL_INDEXABLE_TEMP = 0x69,
223 WINED3D_SM4_OP_DCL_GLOBAL_FLAGS = 0x6a,
224 WINED3D_SM4_OP_LOD = 0x6c,
225 WINED3D_SM4_OP_GATHER4 = 0x6d,
226 WINED3D_SM4_OP_SAMPLE_POS = 0x6e,
227 WINED3D_SM4_OP_SAMPLE_INFO = 0x6f,
228 WINED3D_SM5_OP_HS_DECLS = 0x71,
229 WINED3D_SM5_OP_HS_CONTROL_POINT_PHASE = 0x72,
230 WINED3D_SM5_OP_HS_FORK_PHASE = 0x73,
231 WINED3D_SM5_OP_HS_JOIN_PHASE = 0x74,
232 WINED3D_SM5_OP_EMIT_STREAM = 0x75,
233 WINED3D_SM5_OP_CUT_STREAM = 0x76,
234 WINED3D_SM5_OP_FCALL = 0x78,
235 WINED3D_SM5_OP_BUFINFO = 0x79,
236 WINED3D_SM5_OP_DERIV_RTX_COARSE = 0x7a,
237 WINED3D_SM5_OP_DERIV_RTX_FINE = 0x7b,
238 WINED3D_SM5_OP_DERIV_RTY_COARSE = 0x7c,
239 WINED3D_SM5_OP_DERIV_RTY_FINE = 0x7d,
240 WINED3D_SM5_OP_GATHER4_C = 0x7e,
241 WINED3D_SM5_OP_GATHER4_PO = 0x7f,
242 WINED3D_SM5_OP_GATHER4_PO_C = 0x80,
243 WINED3D_SM5_OP_RCP = 0x81,
244 WINED3D_SM5_OP_F32TOF16 = 0x82,
245 WINED3D_SM5_OP_F16TOF32 = 0x83,
246 WINED3D_SM5_OP_COUNTBITS = 0x86,
247 WINED3D_SM5_OP_FIRSTBIT_HI = 0x87,
248 WINED3D_SM5_OP_FIRSTBIT_LO = 0x88,
249 WINED3D_SM5_OP_FIRSTBIT_SHI = 0x89,
250 WINED3D_SM5_OP_UBFE = 0x8a,
251 WINED3D_SM5_OP_IBFE = 0x8b,
252 WINED3D_SM5_OP_BFI = 0x8c,
253 WINED3D_SM5_OP_BFREV = 0x8d,
254 WINED3D_SM5_OP_SWAPC = 0x8e,
255 WINED3D_SM5_OP_DCL_STREAM = 0x8f,
256 WINED3D_SM5_OP_DCL_FUNCTION_BODY = 0x90,
257 WINED3D_SM5_OP_DCL_FUNCTION_TABLE = 0x91,
258 WINED3D_SM5_OP_DCL_INTERFACE = 0x92,
259 WINED3D_SM5_OP_DCL_INPUT_CONTROL_POINT_COUNT = 0x93,
260 WINED3D_SM5_OP_DCL_OUTPUT_CONTROL_POINT_COUNT = 0x94,
261 WINED3D_SM5_OP_DCL_TESSELLATOR_DOMAIN = 0x95,
262 WINED3D_SM5_OP_DCL_TESSELLATOR_PARTITIONING = 0x96,
263 WINED3D_SM5_OP_DCL_TESSELLATOR_OUTPUT_PRIMITIVE = 0x97,
264 WINED3D_SM5_OP_DCL_HS_MAX_TESSFACTOR = 0x98,
265 WINED3D_SM5_OP_DCL_HS_FORK_PHASE_INSTANCE_COUNT = 0x99,
266 WINED3D_SM5_OP_DCL_HS_JOIN_PHASE_INSTANCE_COUNT = 0x9a,
267 WINED3D_SM5_OP_DCL_THREAD_GROUP = 0x9b,
268 WINED3D_SM5_OP_DCL_UAV_TYPED = 0x9c,
269 WINED3D_SM5_OP_DCL_UAV_RAW = 0x9d,
270 WINED3D_SM5_OP_DCL_UAV_STRUCTURED = 0x9e,
271 WINED3D_SM5_OP_DCL_TGSM_RAW = 0x9f,
272 WINED3D_SM5_OP_DCL_TGSM_STRUCTURED = 0xa0,
273 WINED3D_SM5_OP_DCL_RESOURCE_RAW = 0xa1,
274 WINED3D_SM5_OP_DCL_RESOURCE_STRUCTURED = 0xa2,
275 WINED3D_SM5_OP_LD_UAV_TYPED = 0xa3,
276 WINED3D_SM5_OP_STORE_UAV_TYPED = 0xa4,
277 WINED3D_SM5_OP_LD_RAW = 0xa5,
278 WINED3D_SM5_OP_STORE_RAW = 0xa6,
279 WINED3D_SM5_OP_LD_STRUCTURED = 0xa7,
280 WINED3D_SM5_OP_STORE_STRUCTURED = 0xa8,
281 WINED3D_SM5_OP_ATOMIC_AND = 0xa9,
282 WINED3D_SM5_OP_ATOMIC_OR = 0xaa,
283 WINED3D_SM5_OP_ATOMIC_XOR = 0xab,
284 WINED3D_SM5_OP_ATOMIC_CMP_STORE = 0xac,
285 WINED3D_SM5_OP_ATOMIC_IADD = 0xad,
286 WINED3D_SM5_OP_ATOMIC_IMAX = 0xae,
287 WINED3D_SM5_OP_ATOMIC_IMIN = 0xaf,
288 WINED3D_SM5_OP_ATOMIC_UMAX = 0xb0,
289 WINED3D_SM5_OP_ATOMIC_UMIN = 0xb1,
290 WINED3D_SM5_OP_IMM_ATOMIC_ALLOC = 0xb2,
291 WINED3D_SM5_OP_IMM_ATOMIC_CONSUME = 0xb3,
292 WINED3D_SM5_OP_IMM_ATOMIC_IADD = 0xb4,
293 WINED3D_SM5_OP_IMM_ATOMIC_AND = 0xb5,
294 WINED3D_SM5_OP_IMM_ATOMIC_OR = 0xb6,
295 WINED3D_SM5_OP_IMM_ATOMIC_XOR = 0xb7,
296 WINED3D_SM5_OP_IMM_ATOMIC_EXCH = 0xb8,
297 WINED3D_SM5_OP_IMM_ATOMIC_CMP_EXCH = 0xb9,
298 WINED3D_SM5_OP_IMM_ATOMIC_IMAX = 0xba,
299 WINED3D_SM5_OP_IMM_ATOMIC_IMIN = 0xbb,
300 WINED3D_SM5_OP_IMM_ATOMIC_UMAX = 0xbc,
301 WINED3D_SM5_OP_IMM_ATOMIC_UMIN = 0xbd,
302 WINED3D_SM5_OP_SYNC = 0xbe,
303 WINED3D_SM5_OP_EVAL_SAMPLE_INDEX = 0xcc,
304 WINED3D_SM5_OP_DCL_GS_INSTANCES = 0xce,
305 };
306
307 enum wined3d_sm4_register_type
308 {
309 WINED3D_SM4_RT_TEMP = 0x00,
310 WINED3D_SM4_RT_INPUT = 0x01,
311 WINED3D_SM4_RT_OUTPUT = 0x02,
312 WINED3D_SM4_RT_INDEXABLE_TEMP = 0x03,
313 WINED3D_SM4_RT_IMMCONST = 0x04,
314 WINED3D_SM4_RT_SAMPLER = 0x06,
315 WINED3D_SM4_RT_RESOURCE = 0x07,
316 WINED3D_SM4_RT_CONSTBUFFER = 0x08,
317 WINED3D_SM4_RT_IMMCONSTBUFFER = 0x09,
318 WINED3D_SM4_RT_PRIMID = 0x0b,
319 WINED3D_SM4_RT_DEPTHOUT = 0x0c,
320 WINED3D_SM4_RT_NULL = 0x0d,
321 WINED3D_SM4_RT_RASTERIZER = 0x0e,
322 WINED3D_SM4_RT_OMASK = 0x0f,
323 WINED3D_SM5_RT_STREAM = 0x10,
324 WINED3D_SM5_RT_FUNCTION_BODY = 0x11,
325 WINED3D_SM5_RT_FUNCTION_POINTER = 0x13,
326 WINED3D_SM5_RT_OUTPUT_CONTROL_POINT_ID = 0x16,
327 WINED3D_SM5_RT_FORK_INSTANCE_ID = 0x17,
328 WINED3D_SM5_RT_JOIN_INSTANCE_ID = 0x18,
329 WINED3D_SM5_RT_INPUT_CONTROL_POINT = 0x19,
330 WINED3D_SM5_RT_OUTPUT_CONTROL_POINT = 0x1a,
331 WINED3D_SM5_RT_PATCH_CONSTANT_DATA = 0x1b,
332 WINED3D_SM5_RT_DOMAIN_LOCATION = 0x1c,
333 WINED3D_SM5_RT_UAV = 0x1e,
334 WINED3D_SM5_RT_SHARED_MEMORY = 0x1f,
335 WINED3D_SM5_RT_THREAD_ID = 0x20,
336 WINED3D_SM5_RT_THREAD_GROUP_ID = 0x21,
337 WINED3D_SM5_RT_LOCAL_THREAD_ID = 0x22,
338 WINED3D_SM5_RT_COVERAGE = 0x23,
339 WINED3D_SM5_RT_LOCAL_THREAD_INDEX = 0x24,
340 WINED3D_SM5_RT_GS_INSTANCE_ID = 0x25,
341 WINED3D_SM5_RT_DEPTHOUT_GREATER_EQUAL = 0x26,
342 WINED3D_SM5_RT_DEPTHOUT_LESS_EQUAL = 0x27,
343 };
344
345 enum wined3d_sm4_output_primitive_type
346 {
347 WINED3D_SM4_OUTPUT_PT_POINTLIST = 0x1,
348 WINED3D_SM4_OUTPUT_PT_LINESTRIP = 0x3,
349 WINED3D_SM4_OUTPUT_PT_TRIANGLESTRIP = 0x5,
350 };
351
352 enum wined3d_sm4_input_primitive_type
353 {
354 WINED3D_SM4_INPUT_PT_POINT = 0x01,
355 WINED3D_SM4_INPUT_PT_LINE = 0x02,
356 WINED3D_SM4_INPUT_PT_TRIANGLE = 0x03,
357 WINED3D_SM4_INPUT_PT_LINEADJ = 0x06,
358 WINED3D_SM4_INPUT_PT_TRIANGLEADJ = 0x07,
359 WINED3D_SM5_INPUT_PT_PATCH1 = 0x08,
360 WINED3D_SM5_INPUT_PT_PATCH2 = 0x09,
361 WINED3D_SM5_INPUT_PT_PATCH3 = 0x0a,
362 WINED3D_SM5_INPUT_PT_PATCH4 = 0x0b,
363 WINED3D_SM5_INPUT_PT_PATCH5 = 0x0c,
364 WINED3D_SM5_INPUT_PT_PATCH6 = 0x0d,
365 WINED3D_SM5_INPUT_PT_PATCH7 = 0x0e,
366 WINED3D_SM5_INPUT_PT_PATCH8 = 0x0f,
367 WINED3D_SM5_INPUT_PT_PATCH9 = 0x10,
368 WINED3D_SM5_INPUT_PT_PATCH10 = 0x11,
369 WINED3D_SM5_INPUT_PT_PATCH11 = 0x12,
370 WINED3D_SM5_INPUT_PT_PATCH12 = 0x13,
371 WINED3D_SM5_INPUT_PT_PATCH13 = 0x14,
372 WINED3D_SM5_INPUT_PT_PATCH14 = 0x15,
373 WINED3D_SM5_INPUT_PT_PATCH15 = 0x16,
374 WINED3D_SM5_INPUT_PT_PATCH16 = 0x17,
375 WINED3D_SM5_INPUT_PT_PATCH17 = 0x18,
376 WINED3D_SM5_INPUT_PT_PATCH18 = 0x19,
377 WINED3D_SM5_INPUT_PT_PATCH19 = 0x1a,
378 WINED3D_SM5_INPUT_PT_PATCH20 = 0x1b,
379 WINED3D_SM5_INPUT_PT_PATCH21 = 0x1c,
380 WINED3D_SM5_INPUT_PT_PATCH22 = 0x1d,
381 WINED3D_SM5_INPUT_PT_PATCH23 = 0x1e,
382 WINED3D_SM5_INPUT_PT_PATCH24 = 0x1f,
383 WINED3D_SM5_INPUT_PT_PATCH25 = 0x20,
384 WINED3D_SM5_INPUT_PT_PATCH26 = 0x21,
385 WINED3D_SM5_INPUT_PT_PATCH27 = 0x22,
386 WINED3D_SM5_INPUT_PT_PATCH28 = 0x23,
387 WINED3D_SM5_INPUT_PT_PATCH29 = 0x24,
388 WINED3D_SM5_INPUT_PT_PATCH30 = 0x25,
389 WINED3D_SM5_INPUT_PT_PATCH31 = 0x26,
390 WINED3D_SM5_INPUT_PT_PATCH32 = 0x27,
391 };
392
393 enum wined3d_sm4_swizzle_type
394 {
395 WINED3D_SM4_SWIZZLE_NONE = 0x0,
396 WINED3D_SM4_SWIZZLE_VEC4 = 0x1,
397 WINED3D_SM4_SWIZZLE_SCALAR = 0x2,
398 };
399
400 enum wined3d_sm4_immconst_type
401 {
402 WINED3D_SM4_IMMCONST_SCALAR = 0x1,
403 WINED3D_SM4_IMMCONST_VEC4 = 0x2,
404 };
405
406 enum wined3d_sm4_resource_type
407 {
408 WINED3D_SM4_RESOURCE_BUFFER = 0x1,
409 WINED3D_SM4_RESOURCE_TEXTURE_1D = 0x2,
410 WINED3D_SM4_RESOURCE_TEXTURE_2D = 0x3,
411 WINED3D_SM4_RESOURCE_TEXTURE_2DMS = 0x4,
412 WINED3D_SM4_RESOURCE_TEXTURE_3D = 0x5,
413 WINED3D_SM4_RESOURCE_TEXTURE_CUBE = 0x6,
414 WINED3D_SM4_RESOURCE_TEXTURE_1DARRAY = 0x7,
415 WINED3D_SM4_RESOURCE_TEXTURE_2DARRAY = 0x8,
416 WINED3D_SM4_RESOURCE_TEXTURE_2DMSARRAY = 0x9,
417 WINED3D_SM4_RESOURCE_TEXTURE_CUBEARRAY = 0xa,
418 };
419
420 enum wined3d_sm4_data_type
421 {
422 WINED3D_SM4_DATA_UNORM = 0x1,
423 WINED3D_SM4_DATA_SNORM = 0x2,
424 WINED3D_SM4_DATA_INT = 0x3,
425 WINED3D_SM4_DATA_UINT = 0x4,
426 WINED3D_SM4_DATA_FLOAT = 0x5,
427 };
428
429 enum wined3d_sm4_sampler_mode
430 {
431 WINED3D_SM4_SAMPLER_DEFAULT = 0x0,
432 WINED3D_SM4_SAMPLER_COMPARISON = 0x1,
433 };
434
435 enum wined3d_sm4_shader_data_type
436 {
437 WINED3D_SM4_SHADER_DATA_IMMEDIATE_CONSTANT_BUFFER = 0x3,
438 WINED3D_SM4_SHADER_DATA_MESSAGE = 0x4,
439 };
440
441 struct wined3d_shader_src_param_entry
442 {
443 struct list entry;
444 struct wined3d_shader_src_param param;
445 };
446
447 struct wined3d_sm4_data
448 {
449 struct wined3d_shader_version shader_version;
450 const DWORD *start, *end;
451
452 unsigned int output_map[MAX_REG_OUTPUT];
453
454 struct wined3d_shader_src_param src_param[5];
455 struct wined3d_shader_dst_param dst_param[2];
456 struct list src_free;
457 struct list src;
458 struct wined3d_shader_immediate_constant_buffer icb;
459 };
460
461 struct wined3d_sm4_opcode_info
462 {
463 enum wined3d_sm4_opcode opcode;
464 enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx;
465 const char *dst_info;
466 const char *src_info;
467 void (*read_opcode_func)(struct wined3d_shader_instruction *ins,
468 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
469 struct wined3d_sm4_data *priv);
470 };
471
472 static const enum wined3d_primitive_type output_primitive_type_table[] =
473 {
474 /* UNKNOWN */ WINED3D_PT_UNDEFINED,
475 /* WINED3D_SM4_OUTPUT_PT_POINTLIST */ WINED3D_PT_POINTLIST,
476 /* UNKNOWN */ WINED3D_PT_UNDEFINED,
477 /* WINED3D_SM4_OUTPUT_PT_LINESTRIP */ WINED3D_PT_LINESTRIP,
478 /* UNKNOWN */ WINED3D_PT_UNDEFINED,
479 /* WINED3D_SM4_OUTPUT_PT_TRIANGLESTRIP */ WINED3D_PT_TRIANGLESTRIP,
480 };
481
482 static const enum wined3d_primitive_type input_primitive_type_table[] =
483 {
484 /* UNKNOWN */ WINED3D_PT_UNDEFINED,
485 /* WINED3D_SM4_INPUT_PT_POINT */ WINED3D_PT_POINTLIST,
486 /* WINED3D_SM4_INPUT_PT_LINE */ WINED3D_PT_LINELIST,
487 /* WINED3D_SM4_INPUT_PT_TRIANGLE */ WINED3D_PT_TRIANGLELIST,
488 /* UNKNOWN */ WINED3D_PT_UNDEFINED,
489 /* UNKNOWN */ WINED3D_PT_UNDEFINED,
490 /* WINED3D_SM4_INPUT_PT_LINEADJ */ WINED3D_PT_LINELIST_ADJ,
491 /* WINED3D_SM4_INPUT_PT_TRIANGLEADJ */ WINED3D_PT_TRIANGLELIST_ADJ,
492 };
493
494 static const enum wined3d_shader_resource_type resource_type_table[] =
495 {
496 /* 0 */ WINED3D_SHADER_RESOURCE_NONE,
497 /* WINED3D_SM4_RESOURCE_BUFFER */ WINED3D_SHADER_RESOURCE_BUFFER,
498 /* WINED3D_SM4_RESOURCE_TEXTURE_1D */ WINED3D_SHADER_RESOURCE_TEXTURE_1D,
499 /* WINED3D_SM4_RESOURCE_TEXTURE_2D */ WINED3D_SHADER_RESOURCE_TEXTURE_2D,
500 /* WINED3D_SM4_RESOURCE_TEXTURE_2DMS */ WINED3D_SHADER_RESOURCE_TEXTURE_2DMS,
501 /* WINED3D_SM4_RESOURCE_TEXTURE_3D */ WINED3D_SHADER_RESOURCE_TEXTURE_3D,
502 /* WINED3D_SM4_RESOURCE_TEXTURE_CUBE */ WINED3D_SHADER_RESOURCE_TEXTURE_CUBE,
503 /* WINED3D_SM4_RESOURCE_TEXTURE_1DARRAY */ WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY,
504 /* WINED3D_SM4_RESOURCE_TEXTURE_2DARRAY */ WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY,
505 /* WINED3D_SM4_RESOURCE_TEXTURE_2DMSARRAY */ WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY,
506 /* WINED3D_SM4_RESOURCE_TEXTURE_CUBEARRAY */ WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY,
507 };
508
509 static const enum wined3d_data_type data_type_table[] =
510 {
511 /* 0 */ WINED3D_DATA_FLOAT,
512 /* WINED3D_SM4_DATA_UNORM */ WINED3D_DATA_UNORM,
513 /* WINED3D_SM4_DATA_SNORM */ WINED3D_DATA_SNORM,
514 /* WINED3D_SM4_DATA_INT */ WINED3D_DATA_INT,
515 /* WINED3D_SM4_DATA_UINT */ WINED3D_DATA_UINT,
516 /* WINED3D_SM4_DATA_FLOAT */ WINED3D_DATA_FLOAT,
517 };
518
519 static BOOL shader_sm4_read_src_param(struct wined3d_sm4_data *priv, const DWORD **ptr, const DWORD *end,
520 enum wined3d_data_type data_type, struct wined3d_shader_src_param *src_param);
521 static BOOL shader_sm4_read_dst_param(struct wined3d_sm4_data *priv, const DWORD **ptr, const DWORD *end,
522 enum wined3d_data_type data_type, struct wined3d_shader_dst_param *dst_param);
523
524 static void shader_sm4_read_conditional_op(struct wined3d_shader_instruction *ins,
525 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
526 struct wined3d_sm4_data *priv)
527 {
528 shader_sm4_read_src_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_UINT, &priv->src_param[0]);
529 ins->flags = (opcode_token & WINED3D_SM4_CONDITIONAL_NZ) ?
530 WINED3D_SHADER_CONDITIONAL_OP_NZ : WINED3D_SHADER_CONDITIONAL_OP_Z;
531 }
532
533 static void shader_sm4_read_shader_data(struct wined3d_shader_instruction *ins,
534 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
535 struct wined3d_sm4_data *priv)
536 {
537 enum wined3d_sm4_shader_data_type type;
538 unsigned int icb_size;
539
540 type = (opcode_token & WINED3D_SM4_SHADER_DATA_TYPE_MASK) >> WINED3D_SM4_SHADER_DATA_TYPE_SHIFT;
541 if (type != WINED3D_SM4_SHADER_DATA_IMMEDIATE_CONSTANT_BUFFER)
542 {
543 FIXME("Ignoring shader data type %#x.\n", type);
544 ins->handler_idx = WINED3DSIH_NOP;
545 return;
546 }
547
548 ++tokens;
549 icb_size = token_count - 1;
550 if (icb_size % 4 || icb_size > MAX_IMMEDIATE_CONSTANT_BUFFER_SIZE)
551 {
552 FIXME("Unexpected immediate constant buffer size %u.\n", icb_size);
553 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
554 return;
555 }
556
557 priv->icb.vec4_count = icb_size / 4;
558 memcpy(priv->icb.data, tokens, sizeof(*tokens) * icb_size);
559 ins->declaration.icb = &priv->icb;
560 }
561
562 static void shader_sm4_read_dcl_resource(struct wined3d_shader_instruction *ins,
563 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
564 struct wined3d_sm4_data *priv)
565 {
566 enum wined3d_sm4_resource_type resource_type;
567 enum wined3d_sm4_data_type data_type;
568 enum wined3d_data_type reg_data_type;
569 DWORD components;
570
571 resource_type = (opcode_token & WINED3D_SM4_RESOURCE_TYPE_MASK) >> WINED3D_SM4_RESOURCE_TYPE_SHIFT;
572 if (!resource_type || (resource_type >= ARRAY_SIZE(resource_type_table)))
573 {
574 FIXME("Unhandled resource type %#x.\n", resource_type);
575 ins->declaration.semantic.resource_type = WINED3D_SHADER_RESOURCE_NONE;
576 }
577 else
578 {
579 ins->declaration.semantic.resource_type = resource_type_table[resource_type];
580 }
581 reg_data_type = opcode == WINED3D_SM4_OP_DCL_RESOURCE ? WINED3D_DATA_RESOURCE : WINED3D_DATA_UAV;
582 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], reg_data_type, &ins->declaration.semantic.reg);
583
584 components = *tokens++;
585 if ((components & 0xfff0) != (components & 0xf) * 0x1110)
586 FIXME("Components (%#x) have different data types.\n", components);
587 data_type = components & 0xf;
588
589 if (!data_type || (data_type >= ARRAY_SIZE(data_type_table)))
590 {
591 FIXME("Unhandled data type %#x.\n", data_type);
592 ins->declaration.semantic.resource_data_type = WINED3D_DATA_FLOAT;
593 }
594 else
595 {
596 ins->declaration.semantic.resource_data_type = data_type_table[data_type];
597 }
598
599 if (reg_data_type == WINED3D_DATA_UAV)
600 ins->flags = (opcode_token & WINED3D_SM5_UAV_FLAGS_MASK) >> WINED3D_SM5_UAV_FLAGS_SHIFT;
601 }
602
603 static void shader_sm4_read_dcl_constant_buffer(struct wined3d_shader_instruction *ins,
604 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
605 struct wined3d_sm4_data *priv)
606 {
607 shader_sm4_read_src_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT, &ins->declaration.src);
608 if (opcode_token & WINED3D_SM4_INDEX_TYPE_MASK)
609 ins->flags |= WINED3DSI_INDEXED_DYNAMIC;
610 }
611
612 static void shader_sm4_read_dcl_sampler(struct wined3d_shader_instruction *ins,
613 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
614 struct wined3d_sm4_data *priv)
615 {
616 ins->flags = (opcode_token & WINED3D_SM4_SAMPLER_MODE_MASK) >> WINED3D_SM4_SAMPLER_MODE_SHIFT;
617 if (ins->flags & ~WINED3D_SM4_SAMPLER_COMPARISON)
618 FIXME("Unhandled sampler mode %#x.\n", ins->flags);
619 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_SAMPLER, &ins->declaration.dst);
620 }
621
622 static void shader_sm4_read_dcl_index_range(struct wined3d_shader_instruction *ins,
623 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
624 struct wined3d_sm4_data *priv)
625 {
626 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_OPAQUE,
627 &ins->declaration.index_range.first_register);
628 ins->declaration.index_range.last_register = *tokens;
629 }
630
631 static void shader_sm4_read_dcl_output_topology(struct wined3d_shader_instruction *ins,
632 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
633 struct wined3d_sm4_data *priv)
634 {
635 enum wined3d_sm4_output_primitive_type primitive_type;
636
637 primitive_type = (opcode_token & WINED3D_SM4_PRIMITIVE_TYPE_MASK) >> WINED3D_SM4_PRIMITIVE_TYPE_SHIFT;
638 if (primitive_type >= ARRAY_SIZE(output_primitive_type_table))
639 ins->declaration.primitive_type.type = WINED3D_PT_UNDEFINED;
640 else
641 ins->declaration.primitive_type.type = output_primitive_type_table[primitive_type];
642
643 if (ins->declaration.primitive_type.type == WINED3D_PT_UNDEFINED)
644 FIXME("Unhandled output primitive type %#x.\n", primitive_type);
645 }
646
647 static void shader_sm4_read_dcl_input_primitive(struct wined3d_shader_instruction *ins,
648 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
649 struct wined3d_sm4_data *priv)
650 {
651 enum wined3d_sm4_input_primitive_type primitive_type;
652
653 primitive_type = (opcode_token & WINED3D_SM4_PRIMITIVE_TYPE_MASK) >> WINED3D_SM4_PRIMITIVE_TYPE_SHIFT;
654 if (WINED3D_SM5_INPUT_PT_PATCH1 <= primitive_type && primitive_type <= WINED3D_SM5_INPUT_PT_PATCH32)
655 {
656 ins->declaration.primitive_type.type = WINED3D_PT_PATCH;
657 ins->declaration.primitive_type.patch_vertex_count = primitive_type - WINED3D_SM5_INPUT_PT_PATCH1 + 1;
658 }
659 else if (primitive_type >= ARRAY_SIZE(input_primitive_type_table))
660 {
661 ins->declaration.primitive_type.type = WINED3D_PT_UNDEFINED;
662 }
663 else
664 {
665 ins->declaration.primitive_type.type = input_primitive_type_table[primitive_type];
666 }
667
668 if (ins->declaration.primitive_type.type == WINED3D_PT_UNDEFINED)
669 FIXME("Unhandled input primitive type %#x.\n", primitive_type);
670 }
671
672 static void shader_sm4_read_declaration_count(struct wined3d_shader_instruction *ins,
673 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
674 struct wined3d_sm4_data *priv)
675 {
676 ins->declaration.count = *tokens;
677 }
678
679 static void shader_sm4_read_declaration_dst(struct wined3d_shader_instruction *ins,
680 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
681 struct wined3d_sm4_data *priv)
682 {
683 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT, &ins->declaration.dst);
684 }
685
686 static void shader_sm4_read_declaration_register_semantic(struct wined3d_shader_instruction *ins,
687 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
688 struct wined3d_sm4_data *priv)
689 {
690 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT,
691 &ins->declaration.register_semantic.reg);
692 ins->declaration.register_semantic.sysval_semantic = *tokens;
693 }
694
695 static void shader_sm4_read_dcl_input_ps(struct wined3d_shader_instruction *ins,
696 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
697 struct wined3d_sm4_data *priv)
698 {
699 ins->flags = (opcode_token & WINED3D_SM4_INTERPOLATION_MODE_MASK) >> WINED3D_SM4_INTERPOLATION_MODE_SHIFT;
700 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT, &ins->declaration.dst);
701 }
702
703 static void shader_sm4_read_dcl_input_ps_siv(struct wined3d_shader_instruction *ins,
704 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
705 struct wined3d_sm4_data *priv)
706 {
707 ins->flags = (opcode_token & WINED3D_SM4_INTERPOLATION_MODE_MASK) >> WINED3D_SM4_INTERPOLATION_MODE_SHIFT;
708 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT,
709 &ins->declaration.register_semantic.reg);
710 ins->declaration.register_semantic.sysval_semantic = *tokens;
711 }
712
713 static void shader_sm4_read_dcl_indexable_temp(struct wined3d_shader_instruction *ins,
714 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
715 struct wined3d_sm4_data *priv)
716 {
717 ins->declaration.indexable_temp.register_idx = *tokens++;
718 ins->declaration.indexable_temp.register_size = *tokens++;
719 ins->declaration.indexable_temp.component_count = *tokens;
720 }
721
722 static void shader_sm4_read_dcl_global_flags(struct wined3d_shader_instruction *ins,
723 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
724 struct wined3d_sm4_data *priv)
725 {
726 ins->flags = (opcode_token & WINED3D_SM4_GLOBAL_FLAGS_MASK) >> WINED3D_SM4_GLOBAL_FLAGS_SHIFT;
727 }
728
729 static void shader_sm5_read_fcall(struct wined3d_shader_instruction *ins,
730 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
731 struct wined3d_sm4_data *priv)
732 {
733 priv->src_param[0].reg.u.fp_body_idx = *tokens++;
734 shader_sm4_read_src_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_OPAQUE, &priv->src_param[0]);
735 }
736
737 static void shader_sm5_read_dcl_function_body(struct wined3d_shader_instruction *ins,
738 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
739 struct wined3d_sm4_data *priv)
740 {
741 ins->declaration.index = *tokens;
742 }
743
744 static void shader_sm5_read_dcl_function_table(struct wined3d_shader_instruction *ins,
745 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
746 struct wined3d_sm4_data *priv)
747 {
748 ins->declaration.index = *tokens++;
749 FIXME("Ignoring set of function bodies (count %u).\n", *tokens);
750 }
751
752 static void shader_sm5_read_dcl_interface(struct wined3d_shader_instruction *ins,
753 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
754 struct wined3d_sm4_data *priv)
755 {
756 ins->declaration.fp.index = *tokens++;
757 ins->declaration.fp.body_count = *tokens++;
758 ins->declaration.fp.array_size = *tokens >> WINED3D_SM5_FP_ARRAY_SIZE_SHIFT;
759 ins->declaration.fp.table_count = *tokens++ & WINED3D_SM5_FP_TABLE_COUNT_MASK;
760 FIXME("Ignoring set of function tables (count %u).\n", ins->declaration.fp.table_count);
761 }
762
763 static void shader_sm5_read_control_point_count(struct wined3d_shader_instruction *ins,
764 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
765 struct wined3d_sm4_data *priv)
766 {
767 ins->declaration.count = (opcode_token & WINED3D_SM5_CONTROL_POINT_COUNT_MASK)
768 >> WINED3D_SM5_CONTROL_POINT_COUNT_SHIFT;
769 }
770
771 static void shader_sm5_read_dcl_tessellator_domain(struct wined3d_shader_instruction *ins,
772 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
773 struct wined3d_sm4_data *priv)
774 {
775 ins->declaration.tessellator_domain = (opcode_token & WINED3D_SM5_TESSELLATOR_MASK)
776 >> WINED3D_SM5_TESSELLATOR_SHIFT;
777 }
778
779 static void shader_sm5_read_dcl_tessellator_partitioning(struct wined3d_shader_instruction *ins,
780 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
781 struct wined3d_sm4_data *priv)
782 {
783 ins->declaration.tessellator_partitioning = (opcode_token & WINED3D_SM5_TESSELLATOR_MASK)
784 >> WINED3D_SM5_TESSELLATOR_SHIFT;
785 }
786
787 static void shader_sm5_read_dcl_tessellator_output_primitive(struct wined3d_shader_instruction *ins,
788 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
789 struct wined3d_sm4_data *priv)
790 {
791 ins->declaration.tessellator_output_primitive = (opcode_token & WINED3D_SM5_TESSELLATOR_MASK)
792 >> WINED3D_SM5_TESSELLATOR_SHIFT;
793 }
794
795 static void shader_sm5_read_dcl_hs_max_tessfactor(struct wined3d_shader_instruction *ins,
796 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
797 struct wined3d_sm4_data *priv)
798 {
799 ins->declaration.max_tessellation_factor = *(float *)tokens;
800 }
801
802 static void shader_sm5_read_dcl_thread_group(struct wined3d_shader_instruction *ins,
803 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
804 struct wined3d_sm4_data *priv)
805 {
806 ins->declaration.thread_group_size.x = *tokens++;
807 ins->declaration.thread_group_size.y = *tokens++;
808 ins->declaration.thread_group_size.z = *tokens++;
809 }
810
811 static void shader_sm5_read_dcl_uav_raw(struct wined3d_shader_instruction *ins,
812 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
813 struct wined3d_sm4_data *priv)
814 {
815 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_UAV, &ins->declaration.dst);
816 ins->flags = (opcode_token & WINED3D_SM5_UAV_FLAGS_MASK) >> WINED3D_SM5_UAV_FLAGS_SHIFT;
817 }
818
819 static void shader_sm5_read_dcl_uav_structured(struct wined3d_shader_instruction *ins,
820 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
821 struct wined3d_sm4_data *priv)
822 {
823 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_UAV,
824 &ins->declaration.structured_resource.reg);
825 ins->flags = (opcode_token & WINED3D_SM5_UAV_FLAGS_MASK) >> WINED3D_SM5_UAV_FLAGS_SHIFT;
826 ins->declaration.structured_resource.byte_stride = *tokens;
827 if (ins->declaration.structured_resource.byte_stride % 4)
828 FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.structured_resource.byte_stride);
829 }
830
831 static void shader_sm5_read_dcl_tgsm_raw(struct wined3d_shader_instruction *ins,
832 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
833 struct wined3d_sm4_data *priv)
834 {
835 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT, &ins->declaration.tgsm_raw.reg);
836 ins->declaration.tgsm_raw.byte_count = *tokens;
837 if (ins->declaration.tgsm_raw.byte_count % 4)
838 FIXME("Byte count %u is not multiple of 4.\n", ins->declaration.tgsm_raw.byte_count);
839 }
840
841 static void shader_sm5_read_dcl_tgsm_structured(struct wined3d_shader_instruction *ins,
842 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
843 struct wined3d_sm4_data *priv)
844 {
845 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_FLOAT,
846 &ins->declaration.tgsm_structured.reg);
847 ins->declaration.tgsm_structured.byte_stride = *tokens++;
848 ins->declaration.tgsm_structured.structure_count = *tokens;
849 if (ins->declaration.tgsm_structured.byte_stride % 4)
850 FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.tgsm_structured.byte_stride);
851 }
852
853 static void shader_sm5_read_dcl_resource_structured(struct wined3d_shader_instruction *ins,
854 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
855 struct wined3d_sm4_data *priv)
856 {
857 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_RESOURCE,
858 &ins->declaration.structured_resource.reg);
859 ins->declaration.structured_resource.byte_stride = *tokens;
860 if (ins->declaration.structured_resource.byte_stride % 4)
861 FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.structured_resource.byte_stride);
862 }
863
864 static void shader_sm5_read_dcl_resource_raw(struct wined3d_shader_instruction *ins,
865 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
866 struct wined3d_sm4_data *priv)
867 {
868 shader_sm4_read_dst_param(priv, &tokens, &tokens[token_count], WINED3D_DATA_RESOURCE, &ins->declaration.dst);
869 }
870
871 static void shader_sm5_read_sync(struct wined3d_shader_instruction *ins,
872 DWORD opcode, DWORD opcode_token, const DWORD *tokens, unsigned int token_count,
873 struct wined3d_sm4_data *priv)
874 {
875 ins->flags = (opcode_token & WINED3D_SM5_SYNC_FLAGS_MASK) >> WINED3D_SM5_SYNC_FLAGS_SHIFT;
876 }
877
878 /*
879 * f -> WINED3D_DATA_FLOAT
880 * i -> WINED3D_DATA_INT
881 * u -> WINED3D_DATA_UINT
882 * O -> WINED3D_DATA_OPAQUE
883 * R -> WINED3D_DATA_RESOURCE
884 * S -> WINED3D_DATA_SAMPLER
885 * U -> WINED3D_DATA_UAV
886 */
887 static const struct wined3d_sm4_opcode_info opcode_table[] =
888 {
889 {WINED3D_SM4_OP_ADD, WINED3DSIH_ADD, "f", "ff"},
890 {WINED3D_SM4_OP_AND, WINED3DSIH_AND, "u", "uu"},
891 {WINED3D_SM4_OP_BREAK, WINED3DSIH_BREAK, "", ""},
892 {WINED3D_SM4_OP_BREAKC, WINED3DSIH_BREAKP, "", "u",
893 shader_sm4_read_conditional_op},
894 {WINED3D_SM4_OP_CASE, WINED3DSIH_CASE, "", "u"},
895 {WINED3D_SM4_OP_CONTINUE, WINED3DSIH_CONTINUE, "", ""},
896 {WINED3D_SM4_OP_CONTINUEC, WINED3DSIH_CONTINUEP, "", "u",
897 shader_sm4_read_conditional_op},
898 {WINED3D_SM4_OP_CUT, WINED3DSIH_CUT, "", ""},
899 {WINED3D_SM4_OP_DEFAULT, WINED3DSIH_DEFAULT, "", ""},
900 {WINED3D_SM4_OP_DERIV_RTX, WINED3DSIH_DSX, "f", "f"},
901 {WINED3D_SM4_OP_DERIV_RTY, WINED3DSIH_DSY, "f", "f"},
902 {WINED3D_SM4_OP_DISCARD, WINED3DSIH_TEXKILL, "", "u",
903 shader_sm4_read_conditional_op},
904 {WINED3D_SM4_OP_DIV, WINED3DSIH_DIV, "f", "ff"},
905 {WINED3D_SM4_OP_DP2, WINED3DSIH_DP2, "f", "ff"},
906 {WINED3D_SM4_OP_DP3, WINED3DSIH_DP3, "f", "ff"},
907 {WINED3D_SM4_OP_DP4, WINED3DSIH_DP4, "f", "ff"},
908 {WINED3D_SM4_OP_ELSE, WINED3DSIH_ELSE, "", ""},
909 {WINED3D_SM4_OP_EMIT, WINED3DSIH_EMIT, "", ""},
910 {WINED3D_SM4_OP_ENDIF, WINED3DSIH_ENDIF, "", ""},
911 {WINED3D_SM4_OP_ENDLOOP, WINED3DSIH_ENDLOOP, "", ""},
912 {WINED3D_SM4_OP_ENDSWITCH, WINED3DSIH_ENDSWITCH, "", ""},
913 {WINED3D_SM4_OP_EQ, WINED3DSIH_EQ, "u", "ff"},
914 {WINED3D_SM4_OP_EXP, WINED3DSIH_EXP, "f", "f"},
915 {WINED3D_SM4_OP_FRC, WINED3DSIH_FRC, "f", "f"},
916 {WINED3D_SM4_OP_FTOI, WINED3DSIH_FTOI, "i", "f"},
917 {WINED3D_SM4_OP_FTOU, WINED3DSIH_FTOU, "u", "f"},
918 {WINED3D_SM4_OP_GE, WINED3DSIH_GE, "u", "ff"},
919 {WINED3D_SM4_OP_IADD, WINED3DSIH_IADD, "i", "ii"},
920 {WINED3D_SM4_OP_IF, WINED3DSIH_IF, "", "u",
921 shader_sm4_read_conditional_op},
922 {WINED3D_SM4_OP_IEQ, WINED3DSIH_IEQ, "u", "ii"},
923 {WINED3D_SM4_OP_IGE, WINED3DSIH_IGE, "u", "ii"},
924 {WINED3D_SM4_OP_ILT, WINED3DSIH_ILT, "u", "ii"},
925 {WINED3D_SM4_OP_IMAD, WINED3DSIH_IMAD, "i", "iii"},
926 {WINED3D_SM4_OP_IMAX, WINED3DSIH_IMAX, "i", "ii"},
927 {WINED3D_SM4_OP_IMIN, WINED3DSIH_IMIN, "i", "ii"},
928 {WINED3D_SM4_OP_IMUL, WINED3DSIH_IMUL, "ii", "ii"},
929 {WINED3D_SM4_OP_INE, WINED3DSIH_INE, "u", "ii"},
930 {WINED3D_SM4_OP_INEG, WINED3DSIH_INEG, "i", "i"},
931 {WINED3D_SM4_OP_ISHL, WINED3DSIH_ISHL, "i", "ii"},
932 {WINED3D_SM4_OP_ISHR, WINED3DSIH_ISHR, "i", "ii"},
933 {WINED3D_SM4_OP_ITOF, WINED3DSIH_ITOF, "f", "i"},
934 {WINED3D_SM4_OP_LABEL, WINED3DSIH_LABEL, "", "O"},
935 {WINED3D_SM4_OP_LD, WINED3DSIH_LD, "u", "iR"},
936 {WINED3D_SM4_OP_LD2DMS, WINED3DSIH_LD2DMS, "u", "iRi"},
937 {WINED3D_SM4_OP_LOG, WINED3DSIH_LOG, "f", "f"},
938 {WINED3D_SM4_OP_LOOP, WINED3DSIH_LOOP, "", ""},
939 {WINED3D_SM4_OP_LT, WINED3DSIH_LT, "u", "ff"},
940 {WINED3D_SM4_OP_MAD, WINED3DSIH_MAD, "f", "fff"},
941 {WINED3D_SM4_OP_MIN, WINED3DSIH_MIN, "f", "ff"},
942 {WINED3D_SM4_OP_MAX, WINED3DSIH_MAX, "f", "ff"},
943 {WINED3D_SM4_OP_SHADER_DATA, WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER, "", "",
944 shader_sm4_read_shader_data},
945 {WINED3D_SM4_OP_MOV, WINED3DSIH_MOV, "f", "f"},
946 {WINED3D_SM4_OP_MOVC, WINED3DSIH_MOVC, "f", "uff"},
947 {WINED3D_SM4_OP_MUL, WINED3DSIH_MUL, "f", "ff"},
948 {WINED3D_SM4_OP_NE, WINED3DSIH_NE, "u", "ff"},
949 {WINED3D_SM4_OP_NOP, WINED3DSIH_NOP, "", ""},
950 {WINED3D_SM4_OP_NOT, WINED3DSIH_NOT, "u", "u"},
951 {WINED3D_SM4_OP_OR, WINED3DSIH_OR, "u", "uu"},
952 {WINED3D_SM4_OP_RESINFO, WINED3DSIH_RESINFO, "f", "iR"},
953 {WINED3D_SM4_OP_RET, WINED3DSIH_RET, "", ""},
954 {WINED3D_SM4_OP_RETC, WINED3DSIH_RETP, "", "u",
955 shader_sm4_read_conditional_op},
956 {WINED3D_SM4_OP_ROUND_NE, WINED3DSIH_ROUND_NE, "f", "f"},
957 {WINED3D_SM4_OP_ROUND_NI, WINED3DSIH_ROUND_NI, "f", "f"},
958 {WINED3D_SM4_OP_ROUND_PI, WINED3DSIH_ROUND_PI, "f", "f"},
959 {WINED3D_SM4_OP_ROUND_Z, WINED3DSIH_ROUND_Z, "f", "f"},
960 {WINED3D_SM4_OP_RSQ, WINED3DSIH_RSQ, "f", "f"},
961 {WINED3D_SM4_OP_SAMPLE, WINED3DSIH_SAMPLE, "u", "fRS"},
962 {WINED3D_SM4_OP_SAMPLE_C, WINED3DSIH_SAMPLE_C, "f", "fRSf"},
963 {WINED3D_SM4_OP_SAMPLE_C_LZ, WINED3DSIH_SAMPLE_C_LZ, "f", "fRSf"},
964 {WINED3D_SM4_OP_SAMPLE_LOD, WINED3DSIH_SAMPLE_LOD, "u", "fRSf"},
965 {WINED3D_SM4_OP_SAMPLE_GRAD, WINED3DSIH_SAMPLE_GRAD, "u", "fRSff"},
966 {WINED3D_SM4_OP_SAMPLE_B, WINED3DSIH_SAMPLE_B, "u", "fRSf"},
967 {WINED3D_SM4_OP_SQRT, WINED3DSIH_SQRT, "f", "f"},
968 {WINED3D_SM4_OP_SWITCH, WINED3DSIH_SWITCH, "", "u"},
969 {WINED3D_SM4_OP_SINCOS, WINED3DSIH_SINCOS, "ff", "f"},
970 {WINED3D_SM4_OP_UDIV, WINED3DSIH_UDIV, "uu", "uu"},
971 {WINED3D_SM4_OP_ULT, WINED3DSIH_ULT, "u", "uu"},
972 {WINED3D_SM4_OP_UGE, WINED3DSIH_UGE, "u", "uu"},
973 {WINED3D_SM4_OP_UMUL, WINED3DSIH_UMUL, "uu", "uu"},
974 {WINED3D_SM4_OP_UMAX, WINED3DSIH_UMAX, "u", "uu"},
975 {WINED3D_SM4_OP_UMIN, WINED3DSIH_UMIN, "u", "uu"},
976 {WINED3D_SM4_OP_USHR, WINED3DSIH_USHR, "u", "uu"},
977 {WINED3D_SM4_OP_UTOF, WINED3DSIH_UTOF, "f", "u"},
978 {WINED3D_SM4_OP_XOR, WINED3DSIH_XOR, "u", "uu"},
979 {WINED3D_SM4_OP_DCL_RESOURCE, WINED3DSIH_DCL, "R", "",
980 shader_sm4_read_dcl_resource},
981 {WINED3D_SM4_OP_DCL_CONSTANT_BUFFER, WINED3DSIH_DCL_CONSTANT_BUFFER, "", "",
982 shader_sm4_read_dcl_constant_buffer},
983 {WINED3D_SM4_OP_DCL_SAMPLER, WINED3DSIH_DCL_SAMPLER, "", "",
984 shader_sm4_read_dcl_sampler},
985 {WINED3D_SM4_OP_DCL_INDEX_RANGE, WINED3DSIH_DCL_INDEX_RANGE, "", "",
986 shader_sm4_read_dcl_index_range},
987 {WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY, WINED3DSIH_DCL_OUTPUT_TOPOLOGY, "", "",
988 shader_sm4_read_dcl_output_topology},
989 {WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE, WINED3DSIH_DCL_INPUT_PRIMITIVE, "", "",
990 shader_sm4_read_dcl_input_primitive},
991 {WINED3D_SM4_OP_DCL_VERTICES_OUT, WINED3DSIH_DCL_VERTICES_OUT, "", "",
992 shader_sm4_read_declaration_count},
993 {WINED3D_SM4_OP_DCL_INPUT, WINED3DSIH_DCL_INPUT, "", "",
994 shader_sm4_read_declaration_dst},
995 {WINED3D_SM4_OP_DCL_INPUT_SGV, WINED3DSIH_DCL_INPUT_SGV, "", "",
996 shader_sm4_read_declaration_register_semantic},
997 {WINED3D_SM4_OP_DCL_INPUT_SIV, WINED3DSIH_DCL_INPUT_SIV, "", "",
998 shader_sm4_read_declaration_register_semantic},
999 {WINED3D_SM4_OP_DCL_INPUT_PS, WINED3DSIH_DCL_INPUT_PS, "", "",
1000 shader_sm4_read_dcl_input_ps},
1001 {WINED3D_SM4_OP_DCL_INPUT_PS_SGV, WINED3DSIH_DCL_INPUT_PS_SGV, "", "",
1002 shader_sm4_read_declaration_register_semantic},
1003 {WINED3D_SM4_OP_DCL_INPUT_PS_SIV, WINED3DSIH_DCL_INPUT_PS_SIV, "", "",
1004 shader_sm4_read_dcl_input_ps_siv},
1005 {WINED3D_SM4_OP_DCL_OUTPUT, WINED3DSIH_DCL_OUTPUT, "", "",
1006 shader_sm4_read_declaration_dst},
1007 {WINED3D_SM4_OP_DCL_OUTPUT_SIV, WINED3DSIH_DCL_OUTPUT_SIV, "", "",
1008 shader_sm4_read_declaration_register_semantic},
1009 {WINED3D_SM4_OP_DCL_TEMPS, WINED3DSIH_DCL_TEMPS, "", "",
1010 shader_sm4_read_declaration_count},
1011 {WINED3D_SM4_OP_DCL_INDEXABLE_TEMP, WINED3DSIH_DCL_INDEXABLE_TEMP, "", "",
1012 shader_sm4_read_dcl_indexable_temp},
1013 {WINED3D_SM4_OP_DCL_GLOBAL_FLAGS, WINED3DSIH_DCL_GLOBAL_FLAGS, "", "",
1014 shader_sm4_read_dcl_global_flags},
1015 {WINED3D_SM4_OP_LOD, WINED3DSIH_LOD, "f", "fRS"},
1016 {WINED3D_SM4_OP_GATHER4, WINED3DSIH_GATHER4, "u", "fRS"},
1017 {WINED3D_SM4_OP_SAMPLE_POS, WINED3DSIH_SAMPLE_POS, "f", "Ru"},
1018 {WINED3D_SM4_OP_SAMPLE_INFO, WINED3DSIH_SAMPLE_INFO, "f", "R"},
1019 {WINED3D_SM5_OP_HS_DECLS, WINED3DSIH_HS_DECLS, "", ""},
1020 {WINED3D_SM5_OP_HS_CONTROL_POINT_PHASE, WINED3DSIH_HS_CONTROL_POINT_PHASE, "", ""},
1021 {WINED3D_SM5_OP_HS_FORK_PHASE, WINED3DSIH_HS_FORK_PHASE, "", ""},
1022 {WINED3D_SM5_OP_HS_JOIN_PHASE, WINED3DSIH_HS_JOIN_PHASE, "", ""},
1023 {WINED3D_SM5_OP_EMIT_STREAM, WINED3DSIH_EMIT_STREAM, "", "f"},
1024 {WINED3D_SM5_OP_CUT_STREAM, WINED3DSIH_CUT_STREAM, "", "f"},
1025 {WINED3D_SM5_OP_FCALL, WINED3DSIH_FCALL, "", "O",
1026 shader_sm5_read_fcall},
1027 {WINED3D_SM5_OP_BUFINFO, WINED3DSIH_BUFINFO, "i", "U"},
1028 {WINED3D_SM5_OP_DERIV_RTX_COARSE, WINED3DSIH_DSX_COARSE, "f", "f"},
1029 {WINED3D_SM5_OP_DERIV_RTX_FINE, WINED3DSIH_DSX_FINE, "f", "f"},
1030 {WINED3D_SM5_OP_DERIV_RTY_COARSE, WINED3DSIH_DSY_COARSE, "f", "f"},
1031 {WINED3D_SM5_OP_DERIV_RTY_FINE, WINED3DSIH_DSY_FINE, "f", "f"},
1032 {WINED3D_SM5_OP_GATHER4_C, WINED3DSIH_GATHER4_C, "f", "fRSf"},
1033 {WINED3D_SM5_OP_GATHER4_PO, WINED3DSIH_GATHER4_PO, "f", "fiRS"},
1034 {WINED3D_SM5_OP_GATHER4_PO_C, WINED3DSIH_GATHER4_PO_C, "f", "fiRSf"},
1035 {WINED3D_SM5_OP_RCP, WINED3DSIH_RCP, "f", "f"},
1036 {WINED3D_SM5_OP_F32TOF16, WINED3DSIH_F32TOF16, "u", "f"},
1037 {WINED3D_SM5_OP_F16TOF32, WINED3DSIH_F16TOF32, "f", "u"},
1038 {WINED3D_SM5_OP_COUNTBITS, WINED3DSIH_COUNTBITS, "u", "u"},
1039 {WINED3D_SM5_OP_FIRSTBIT_HI, WINED3DSIH_FIRSTBIT_HI, "u", "u"},
1040 {WINED3D_SM5_OP_FIRSTBIT_LO, WINED3DSIH_FIRSTBIT_LO, "u", "u"},
1041 {WINED3D_SM5_OP_FIRSTBIT_SHI, WINED3DSIH_FIRSTBIT_SHI, "u", "i"},
1042 {WINED3D_SM5_OP_UBFE, WINED3DSIH_UBFE, "u", "iiu"},
1043 {WINED3D_SM5_OP_IBFE, WINED3DSIH_IBFE, "i", "iii"},
1044 {WINED3D_SM5_OP_BFI, WINED3DSIH_BFI, "u", "iiuu"},
1045 {WINED3D_SM5_OP_BFREV, WINED3DSIH_BFREV, "u", "u"},
1046 {WINED3D_SM5_OP_SWAPC, WINED3DSIH_SWAPC, "ff", "uff"},
1047 {WINED3D_SM5_OP_DCL_STREAM, WINED3DSIH_DCL_STREAM, "", "O"},
1048 {WINED3D_SM5_OP_DCL_FUNCTION_BODY, WINED3DSIH_DCL_FUNCTION_BODY, "", "",
1049 shader_sm5_read_dcl_function_body},
1050 {WINED3D_SM5_OP_DCL_FUNCTION_TABLE, WINED3DSIH_DCL_FUNCTION_TABLE, "", "",
1051 shader_sm5_read_dcl_function_table},
1052 {WINED3D_SM5_OP_DCL_INTERFACE, WINED3DSIH_DCL_INTERFACE, "", "",
1053 shader_sm5_read_dcl_interface},
1054 {WINED3D_SM5_OP_DCL_INPUT_CONTROL_POINT_COUNT, WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT, "", "",
1055 shader_sm5_read_control_point_count},
1056 {WINED3D_SM5_OP_DCL_OUTPUT_CONTROL_POINT_COUNT, WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT, "", "",
1057 shader_sm5_read_control_point_count},
1058 {WINED3D_SM5_OP_DCL_TESSELLATOR_DOMAIN, WINED3DSIH_DCL_TESSELLATOR_DOMAIN, "", "",
1059 shader_sm5_read_dcl_tessellator_domain},
1060 {WINED3D_SM5_OP_DCL_TESSELLATOR_PARTITIONING, WINED3DSIH_DCL_TESSELLATOR_PARTITIONING, "", "",
1061 shader_sm5_read_dcl_tessellator_partitioning},
1062 {WINED3D_SM5_OP_DCL_TESSELLATOR_OUTPUT_PRIMITIVE, WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE, "", "",
1063 shader_sm5_read_dcl_tessellator_output_primitive},
1064 {WINED3D_SM5_OP_DCL_HS_MAX_TESSFACTOR, WINED3DSIH_DCL_HS_MAX_TESSFACTOR, "", "",
1065 shader_sm5_read_dcl_hs_max_tessfactor},
1066 {WINED3D_SM5_OP_DCL_HS_FORK_PHASE_INSTANCE_COUNT, WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT, "", "",
1067 shader_sm4_read_declaration_count},
1068 {WINED3D_SM5_OP_DCL_HS_JOIN_PHASE_INSTANCE_COUNT, WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT, "", "",
1069 shader_sm4_read_declaration_count},
1070 {WINED3D_SM5_OP_DCL_THREAD_GROUP, WINED3DSIH_DCL_THREAD_GROUP, "", "",
1071 shader_sm5_read_dcl_thread_group},
1072 {WINED3D_SM5_OP_DCL_UAV_TYPED, WINED3DSIH_DCL_UAV_TYPED, "", "",
1073 shader_sm4_read_dcl_resource},
1074 {WINED3D_SM5_OP_DCL_UAV_RAW, WINED3DSIH_DCL_UAV_RAW, "", "",
1075 shader_sm5_read_dcl_uav_raw},
1076 {WINED3D_SM5_OP_DCL_UAV_STRUCTURED, WINED3DSIH_DCL_UAV_STRUCTURED, "", "",
1077 shader_sm5_read_dcl_uav_structured},
1078 {WINED3D_SM5_OP_DCL_TGSM_RAW, WINED3DSIH_DCL_TGSM_RAW, "", "",
1079 shader_sm5_read_dcl_tgsm_raw},
1080 {WINED3D_SM5_OP_DCL_TGSM_STRUCTURED, WINED3DSIH_DCL_TGSM_STRUCTURED, "", "",
1081 shader_sm5_read_dcl_tgsm_structured},
1082 {WINED3D_SM5_OP_DCL_RESOURCE_RAW, WINED3DSIH_DCL_RESOURCE_RAW, "", "",
1083 shader_sm5_read_dcl_resource_raw},
1084 {WINED3D_SM5_OP_DCL_RESOURCE_STRUCTURED, WINED3DSIH_DCL_RESOURCE_STRUCTURED, "", "",
1085 shader_sm5_read_dcl_resource_structured},
1086 {WINED3D_SM5_OP_LD_UAV_TYPED, WINED3DSIH_LD_UAV_TYPED, "u", "iU"},
1087 {WINED3D_SM5_OP_STORE_UAV_TYPED, WINED3DSIH_STORE_UAV_TYPED, "U", "iu"},
1088 {WINED3D_SM5_OP_LD_RAW, WINED3DSIH_LD_RAW, "u", "iU"},
1089 {WINED3D_SM5_OP_STORE_RAW, WINED3DSIH_STORE_RAW, "U", "iu"},
1090 {WINED3D_SM5_OP_LD_STRUCTURED, WINED3DSIH_LD_STRUCTURED, "u", "iiR"},
1091 {WINED3D_SM5_OP_STORE_STRUCTURED, WINED3DSIH_STORE_STRUCTURED, "U", "iiu"},
1092 {WINED3D_SM5_OP_ATOMIC_AND, WINED3DSIH_ATOMIC_AND, "U", "iu"},
1093 {WINED3D_SM5_OP_ATOMIC_OR, WINED3DSIH_ATOMIC_OR, "U", "iu"},
1094 {WINED3D_SM5_OP_ATOMIC_XOR, WINED3DSIH_ATOMIC_XOR, "U", "iu"},
1095 {WINED3D_SM5_OP_ATOMIC_CMP_STORE, WINED3DSIH_ATOMIC_CMP_STORE, "U", "iuu"},
1096 {WINED3D_SM5_OP_ATOMIC_IADD, WINED3DSIH_ATOMIC_IADD, "U", "ii"},
1097 {WINED3D_SM5_OP_ATOMIC_IMAX, WINED3DSIH_ATOMIC_IMAX, "U", "ii"},
1098 {WINED3D_SM5_OP_ATOMIC_IMIN, WINED3DSIH_ATOMIC_IMIN, "U", "ii"},
1099 {WINED3D_SM5_OP_ATOMIC_UMAX, WINED3DSIH_ATOMIC_UMAX, "U", "iu"},
1100 {WINED3D_SM5_OP_ATOMIC_UMIN, WINED3DSIH_ATOMIC_UMIN, "U", "iu"},
1101 {WINED3D_SM5_OP_IMM_ATOMIC_ALLOC, WINED3DSIH_IMM_ATOMIC_ALLOC, "u", "U"},
1102 {WINED3D_SM5_OP_IMM_ATOMIC_CONSUME, WINED3DSIH_IMM_ATOMIC_CONSUME, "u", "U"},
1103 {WINED3D_SM5_OP_IMM_ATOMIC_IADD, WINED3DSIH_IMM_ATOMIC_IADD, "uU", "ii"},
1104 {WINED3D_SM5_OP_IMM_ATOMIC_AND, WINED3DSIH_IMM_ATOMIC_AND, "uU", "iu"},
1105 {WINED3D_SM5_OP_IMM_ATOMIC_OR, WINED3DSIH_IMM_ATOMIC_OR, "uU", "iu"},
1106 {WINED3D_SM5_OP_IMM_ATOMIC_XOR, WINED3DSIH_IMM_ATOMIC_XOR, "uU", "iu"},
1107 {WINED3D_SM5_OP_IMM_ATOMIC_EXCH, WINED3DSIH_IMM_ATOMIC_EXCH, "uU", "iu"},
1108 {WINED3D_SM5_OP_IMM_ATOMIC_CMP_EXCH, WINED3DSIH_IMM_ATOMIC_CMP_EXCH, "uU", "iuu"},
1109 {WINED3D_SM5_OP_IMM_ATOMIC_IMAX, WINED3DSIH_IMM_ATOMIC_IMAX, "iU", "ii"},
1110 {WINED3D_SM5_OP_IMM_ATOMIC_IMIN, WINED3DSIH_IMM_ATOMIC_IMIN, "iU", "ii"},
1111 {WINED3D_SM5_OP_IMM_ATOMIC_UMAX, WINED3DSIH_IMM_ATOMIC_UMAX, "uU", "iu"},
1112 {WINED3D_SM5_OP_IMM_ATOMIC_UMIN, WINED3DSIH_IMM_ATOMIC_UMIN, "uU", "iu"},
1113 {WINED3D_SM5_OP_SYNC, WINED3DSIH_SYNC, "", "",
1114 shader_sm5_read_sync},
1115 {WINED3D_SM5_OP_EVAL_SAMPLE_INDEX, WINED3DSIH_EVAL_SAMPLE_INDEX, "f", "fi"},
1116 {WINED3D_SM5_OP_DCL_GS_INSTANCES, WINED3DSIH_DCL_GS_INSTANCES, "", "",
1117 shader_sm4_read_declaration_count},
1118 };
1119
1120 static const enum wined3d_shader_register_type register_type_table[] =
1121 {
1122 /* WINED3D_SM4_RT_TEMP */ WINED3DSPR_TEMP,
1123 /* WINED3D_SM4_RT_INPUT */ WINED3DSPR_INPUT,
1124 /* WINED3D_SM4_RT_OUTPUT */ WINED3DSPR_OUTPUT,
1125 /* WINED3D_SM4_RT_INDEXABLE_TEMP */ WINED3DSPR_IDXTEMP,
1126 /* WINED3D_SM4_RT_IMMCONST */ WINED3DSPR_IMMCONST,
1127 /* UNKNOWN */ ~0u,
1128 /* WINED3D_SM4_RT_SAMPLER */ WINED3DSPR_SAMPLER,
1129 /* WINED3D_SM4_RT_RESOURCE */ WINED3DSPR_RESOURCE,
1130 /* WINED3D_SM4_RT_CONSTBUFFER */ WINED3DSPR_CONSTBUFFER,
1131 /* WINED3D_SM4_RT_IMMCONSTBUFFER */ WINED3DSPR_IMMCONSTBUFFER,
1132 /* UNKNOWN */ ~0u,
1133 /* WINED3D_SM4_RT_PRIMID */ WINED3DSPR_PRIMID,
1134 /* WINED3D_SM4_RT_DEPTHOUT */ WINED3DSPR_DEPTHOUT,
1135 /* WINED3D_SM4_RT_NULL */ WINED3DSPR_NULL,
1136 /* WINED3D_SM4_RT_RASTERIZER */ WINED3DSPR_RASTERIZER,
1137 /* WINED3D_SM4_RT_OMASK */ WINED3DSPR_SAMPLEMASK,
1138 /* WINED3D_SM5_RT_STREAM */ WINED3DSPR_STREAM,
1139 /* WINED3D_SM5_RT_FUNCTION_BODY */ WINED3DSPR_FUNCTIONBODY,
1140 /* UNKNOWN */ ~0u,
1141 /* WINED3D_SM5_RT_FUNCTION_POINTER */ WINED3DSPR_FUNCTIONPOINTER,
1142 /* UNKNOWN */ ~0u,
1143 /* UNKNOWN */ ~0u,
1144 /* WINED3D_SM5_RT_OUTPUT_CONTROL_POINT_ID */ WINED3DSPR_OUTPOINTID,
1145 /* WINED3D_SM5_RT_FORK_INSTANCE_ID */ WINED3DSPR_FORKINSTID,
1146 /* WINED3D_SM5_RT_JOIN_INSTANCE_ID */ WINED3DSPR_JOININSTID,
1147 /* WINED3D_SM5_RT_INPUT_CONTROL_POINT */ WINED3DSPR_INCONTROLPOINT,
1148 /* WINED3D_SM5_RT_OUTPUT_CONTROL_POINT */ WINED3DSPR_OUTCONTROLPOINT,
1149 /* WINED3D_SM5_RT_PATCH_CONSTANT_DATA */ WINED3DSPR_PATCHCONST,
1150 /* WINED3D_SM5_RT_DOMAIN_LOCATION */ WINED3DSPR_TESSCOORD,
1151 /* UNKNOWN */ ~0u,
1152 /* WINED3D_SM5_RT_UAV */ WINED3DSPR_UAV,
1153 /* WINED3D_SM5_RT_SHARED_MEMORY */ WINED3DSPR_GROUPSHAREDMEM,
1154 /* WINED3D_SM5_RT_THREAD_ID */ WINED3DSPR_THREADID,
1155 /* WINED3D_SM5_RT_THREAD_GROUP_ID */ WINED3DSPR_THREADGROUPID,
1156 /* WINED3D_SM5_RT_LOCAL_THREAD_ID */ WINED3DSPR_LOCALTHREADID,
1157 /* WINED3D_SM5_RT_COVERAGE */ WINED3DSPR_COVERAGE,
1158 /* WINED3D_SM5_RT_LOCAL_THREAD_INDEX */ WINED3DSPR_LOCALTHREADINDEX,
1159 /* WINED3D_SM5_RT_GS_INSTANCE_ID */ WINED3DSPR_GSINSTID,
1160 /* WINED3D_SM5_RT_DEPTHOUT_GREATER_EQUAL */ WINED3DSPR_DEPTHOUTGE,
1161 /* WINED3D_SM5_RT_DEPTHOUT_LESS_EQUAL */ WINED3DSPR_DEPTHOUTLE,
1162 };
1163
1164 static const struct wined3d_sm4_opcode_info *get_opcode_info(enum wined3d_sm4_opcode opcode)
1165 {
1166 unsigned int i;
1167
1168 for (i = 0; i < ARRAY_SIZE(opcode_table); ++i)
1169 {
1170 if (opcode == opcode_table[i].opcode) return &opcode_table[i];
1171 }
1172
1173 return NULL;
1174 }
1175
1176 static void map_register(const struct wined3d_sm4_data *priv, struct wined3d_shader_register *reg)
1177 {
1178 switch (priv->shader_version.type)
1179 {
1180 case WINED3D_SHADER_TYPE_PIXEL:
1181 if (reg->type == WINED3DSPR_OUTPUT)
1182 {
1183 unsigned int reg_idx = reg->idx[0].offset;
1184
1185 if (reg_idx >= ARRAY_SIZE(priv->output_map))
1186 {
1187 ERR("Invalid output index %u.\n", reg_idx);
1188 break;
1189 }
1190
1191 reg->type = WINED3DSPR_COLOROUT;
1192 reg->idx[0].offset = priv->output_map[reg_idx];
1193 }
1194 break;
1195
1196 default:
1197 break;
1198 }
1199 }
1200
1201 static enum wined3d_data_type map_data_type(char t)
1202 {
1203 switch (t)
1204 {
1205 case 'f':
1206 return WINED3D_DATA_FLOAT;
1207 case 'i':
1208 return WINED3D_DATA_INT;
1209 case 'u':
1210 return WINED3D_DATA_UINT;
1211 case 'O':
1212 return WINED3D_DATA_OPAQUE;
1213 case 'R':
1214 return WINED3D_DATA_RESOURCE;
1215 case 'S':
1216 return WINED3D_DATA_SAMPLER;
1217 case 'U':
1218 return WINED3D_DATA_UAV;
1219 default:
1220 ERR("Invalid data type '%c'.\n", t);
1221 return WINED3D_DATA_FLOAT;
1222 }
1223 }
1224
1225 static enum wined3d_shader_type wined3d_get_sm4_shader_type(const DWORD *byte_code, size_t byte_code_size)
1226 {
1227 DWORD shader_type;
1228
1229 if (byte_code_size / sizeof(*byte_code) < 1)
1230 {
1231 WARN("Invalid byte code size %lu.\n", (long)byte_code_size);
1232 return WINED3D_SHADER_TYPE_INVALID;
1233 }
1234
1235 shader_type = byte_code[0] >> 16;
1236 switch (shader_type)
1237 {
1238 case WINED3D_SM4_PS:
1239 return WINED3D_SHADER_TYPE_PIXEL;
1240 break;
1241 case WINED3D_SM4_VS:
1242 return WINED3D_SHADER_TYPE_VERTEX;
1243 break;
1244 case WINED3D_SM4_GS:
1245 return WINED3D_SHADER_TYPE_GEOMETRY;
1246 break;
1247 case WINED3D_SM5_HS:
1248 return WINED3D_SHADER_TYPE_HULL;
1249 break;
1250 case WINED3D_SM5_DS:
1251 return WINED3D_SHADER_TYPE_DOMAIN;
1252 break;
1253 case WINED3D_SM5_CS:
1254 return WINED3D_SHADER_TYPE_COMPUTE;
1255 break;
1256 default:
1257 FIXME("Unrecognised shader type %#x.\n", shader_type);
1258 return WINED3D_SHADER_TYPE_INVALID;
1259 }
1260 }
1261
1262 static void *shader_sm4_init(const DWORD *byte_code, size_t byte_code_size,
1263 const struct wined3d_shader_signature *output_signature)
1264 {
1265 DWORD version_token, token_count;
1266 struct wined3d_sm4_data *priv;
1267 unsigned int i;
1268
1269 if (byte_code_size / sizeof(*byte_code) < 2)
1270 {
1271 WARN("Invalid byte code size %lu.\n", (long)byte_code_size);
1272 return NULL;
1273 }
1274
1275 version_token = byte_code[0];
1276 TRACE("Version: 0x%08x.\n", version_token);
1277 token_count = byte_code[1];
1278 TRACE("Token count: %u.\n", token_count);
1279
1280 if (token_count < 2 || byte_code_size / sizeof(*byte_code) < token_count)
1281 {
1282 WARN("Invalid token count %u.\n", token_count);
1283 return NULL;
1284 }
1285
1286 if (!(priv = heap_alloc(sizeof(*priv))))
1287 {
1288 ERR("Failed to allocate private data\n");
1289 return NULL;
1290 }
1291
1292 priv->start = &byte_code[2];
1293 priv->end = &byte_code[token_count];
1294
1295 priv->shader_version.type = wined3d_get_sm4_shader_type(byte_code, byte_code_size);
1296 if (priv->shader_version.type == WINED3D_SHADER_TYPE_INVALID)
1297 {
1298 heap_free(priv);
1299 return NULL;
1300 }
1301
1302 priv->shader_version.major = WINED3D_SM4_VERSION_MAJOR(version_token);
1303 priv->shader_version.minor = WINED3D_SM4_VERSION_MINOR(version_token);
1304
1305 memset(priv->output_map, 0xff, sizeof(priv->output_map));
1306 for (i = 0; i < output_signature->element_count; ++i)
1307 {
1308 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1309
1310 if (e->register_idx >= ARRAY_SIZE(priv->output_map))
1311 {
1312 WARN("Invalid output index %u.\n", e->register_idx);
1313 continue;
1314 }
1315
1316 priv->output_map[e->register_idx] = e->semantic_idx;
1317 }
1318
1319 list_init(&priv->src_free);
1320 list_init(&priv->src);
1321
1322 return priv;
1323 }
1324
1325 static void shader_sm4_free(void *data)
1326 {
1327 struct wined3d_shader_src_param_entry *e1, *e2;
1328 struct wined3d_sm4_data *priv = data;
1329
1330 list_move_head(&priv->src_free, &priv->src);
1331 LIST_FOR_EACH_ENTRY_SAFE(e1, e2, &priv->src_free, struct wined3d_shader_src_param_entry, entry)
1332 {
1333 heap_free(e1);
1334 }
1335 heap_free(priv);
1336 }
1337
1338 static struct wined3d_shader_src_param *get_src_param(struct wined3d_sm4_data *priv)
1339 {
1340 struct wined3d_shader_src_param_entry *e;
1341 struct list *elem;
1342
1343 if (!list_empty(&priv->src_free))
1344 {
1345 elem = list_head(&priv->src_free);
1346 list_remove(elem);
1347 }
1348 else
1349 {
1350 if (!(e = heap_alloc(sizeof(*e))))
1351 return NULL;
1352 elem = &e->entry;
1353 }
1354
1355 list_add_tail(&priv->src, elem);
1356 e = LIST_ENTRY(elem, struct wined3d_shader_src_param_entry, entry);
1357 return &e->param;
1358 }
1359
1360 static void shader_sm4_read_header(void *data, const DWORD **ptr, struct wined3d_shader_version *shader_version)
1361 {
1362 struct wined3d_sm4_data *priv = data;
1363
1364 *ptr = priv->start;
1365 *shader_version = priv->shader_version;
1366 }
1367
1368 static BOOL shader_sm4_read_reg_idx(struct wined3d_sm4_data *priv, const DWORD **ptr, const DWORD *end,
1369 DWORD addressing, struct wined3d_shader_register_index *reg_idx)
1370 {
1371 if (addressing & WINED3D_SM4_ADDRESSING_RELATIVE)
1372 {
1373 struct wined3d_shader_src_param *rel_addr = get_src_param(priv);
1374
1375 if (!(reg_idx->rel_addr = rel_addr))
1376 {
1377 ERR("Failed to get src param for relative addressing.\n");
1378 return FALSE;
1379 }
1380
1381 if (addressing & WINED3D_SM4_ADDRESSING_OFFSET)
1382 reg_idx->offset = *(*ptr)++;
1383 else
1384 reg_idx->offset = 0;
1385 shader_sm4_read_src_param(priv, ptr, end, WINED3D_DATA_INT, rel_addr);
1386 }
1387 else
1388 {
1389 reg_idx->rel_addr = NULL;
1390 reg_idx->offset = *(*ptr)++;
1391 }
1392
1393 return TRUE;
1394 }
1395
1396 static BOOL shader_sm4_read_param(struct wined3d_sm4_data *priv, const DWORD **ptr, const DWORD *end,
1397 enum wined3d_data_type data_type, struct wined3d_shader_register *param,
1398 enum wined3d_shader_src_modifier *modifier)
1399 {
1400 enum wined3d_sm4_register_type register_type;
1401 DWORD token, order;
1402
1403 if (*ptr >= end)
1404 {
1405 WARN("Invalid ptr %p >= end %p.\n", *ptr, end);
1406 return FALSE;
1407 }
1408 token = *(*ptr)++;
1409
1410 register_type = (token & WINED3D_SM4_REGISTER_TYPE_MASK) >> WINED3D_SM4_REGISTER_TYPE_SHIFT;
1411 if (register_type >= ARRAY_SIZE(register_type_table)
1412 || register_type_table[register_type] == ~0u)
1413 {
1414 FIXME("Unhandled register type %#x.\n", register_type);
1415 param->type = WINED3DSPR_TEMP;
1416 }
1417 else
1418 {
1419 param->type = register_type_table[register_type];
1420 }
1421 param->data_type = data_type;
1422
1423 if (token & WINED3D_SM4_REGISTER_MODIFIER)
1424 {
1425 DWORD m;
1426
1427 if (*ptr >= end)
1428 {
1429 WARN("Invalid ptr %p >= end %p.\n", *ptr, end);
1430 return FALSE;
1431 }
1432 m = *(*ptr)++;
1433
1434 switch (m)
1435 {
1436 case 0x41:
1437 *modifier = WINED3DSPSM_NEG;
1438 break;
1439
1440 case 0x81:
1441 *modifier = WINED3DSPSM_ABS;
1442 break;
1443
1444 case 0xc1:
1445 *modifier = WINED3DSPSM_ABSNEG;
1446 break;
1447
1448 default:
1449 FIXME("Skipping modifier 0x%08x.\n", m);
1450 *modifier = WINED3DSPSM_NONE;
1451 break;
1452 }
1453 }
1454 else
1455 {
1456 *modifier = WINED3DSPSM_NONE;
1457 }
1458
1459 order = (token & WINED3D_SM4_REGISTER_ORDER_MASK) >> WINED3D_SM4_REGISTER_ORDER_SHIFT;
1460
1461 if (order < 1)
1462 param->idx[0].offset = ~0U;
1463 else
1464 {
1465 DWORD addressing = (token & WINED3D_SM4_ADDRESSING_MASK0) >> WINED3D_SM4_ADDRESSING_SHIFT0;
1466 if (!(shader_sm4_read_reg_idx(priv, ptr, end, addressing, &param->idx[0])))
1467 {
1468 ERR("Failed to read register index.\n");
1469 return FALSE;
1470 }
1471 }
1472
1473 if (order < 2)
1474 param->idx[1].offset = ~0U;
1475 else
1476 {
1477 DWORD addressing = (token & WINED3D_SM4_ADDRESSING_MASK1) >> WINED3D_SM4_ADDRESSING_SHIFT1;
1478 if (!(shader_sm4_read_reg_idx(priv, ptr, end, addressing, &param->idx[1])))
1479 {
1480 ERR("Failed to read register index.\n");
1481 return FALSE;
1482 }
1483 }
1484
1485 if (order > 2)
1486 FIXME("Unhandled order %u.\n", order);
1487
1488 if (register_type == WINED3D_SM4_RT_IMMCONST)
1489 {
1490 enum wined3d_sm4_immconst_type immconst_type =
1491 (token & WINED3D_SM4_IMMCONST_TYPE_MASK) >> WINED3D_SM4_IMMCONST_TYPE_SHIFT;
1492
1493 switch (immconst_type)
1494 {
1495 case WINED3D_SM4_IMMCONST_SCALAR:
1496 param->immconst_type = WINED3D_IMMCONST_SCALAR;
1497 if (end - *ptr < 1)
1498 {
1499 WARN("Invalid ptr %p, end %p.\n", *ptr, end);
1500 return FALSE;
1501 }
1502 memcpy(param->u.immconst_data, *ptr, 1 * sizeof(DWORD));
1503 *ptr += 1;
1504 break;
1505
1506 case WINED3D_SM4_IMMCONST_VEC4:
1507 param->immconst_type = WINED3D_IMMCONST_VEC4;
1508 if (end - *ptr < 4)
1509 {
1510 WARN("Invalid ptr %p, end %p.\n", *ptr, end);
1511 return FALSE;
1512 }
1513 memcpy(param->u.immconst_data, *ptr, 4 * sizeof(DWORD));
1514 *ptr += 4;
1515 break;
1516
1517 default:
1518 FIXME("Unhandled immediate constant type %#x.\n", immconst_type);
1519 break;
1520 }
1521 }
1522
1523 map_register(priv, param);
1524
1525 return TRUE;
1526 }
1527
1528 static BOOL shader_sm4_read_src_param(struct wined3d_sm4_data *priv, const DWORD **ptr, const DWORD *end,
1529 enum wined3d_data_type data_type, struct wined3d_shader_src_param *src_param)
1530 {
1531 DWORD token;
1532
1533 if (*ptr >= end)
1534 {
1535 WARN("Invalid ptr %p >= end %p.\n", *ptr, end);
1536 return FALSE;
1537 }
1538 token = **ptr;
1539
1540 if (!shader_sm4_read_param(priv, ptr, end, data_type, &src_param->reg, &src_param->modifiers))
1541 {
1542 ERR("Failed to read parameter.\n");
1543 return FALSE;
1544 }
1545
1546 if (src_param->reg.type == WINED3DSPR_IMMCONST)
1547 {
1548 src_param->swizzle = WINED3DSP_NOSWIZZLE;
1549 }
1550 else
1551 {
1552 enum wined3d_sm4_swizzle_type swizzle_type =
1553 (token & WINED3D_SM4_SWIZZLE_TYPE_MASK) >> WINED3D_SM4_SWIZZLE_TYPE_SHIFT;
1554
1555 switch (swizzle_type)
1556 {
1557 case WINED3D_SM4_SWIZZLE_NONE:
1558 src_param->swizzle = WINED3DSP_NOSWIZZLE;
1559 break;
1560
1561 case WINED3D_SM4_SWIZZLE_SCALAR:
1562 src_param->swizzle = (token & WINED3D_SM4_SWIZZLE_MASK) >> WINED3D_SM4_SWIZZLE_SHIFT;
1563 src_param->swizzle = (src_param->swizzle & 0x3) * 0x55;
1564 break;
1565
1566 case WINED3D_SM4_SWIZZLE_VEC4:
1567 src_param->swizzle = (token & WINED3D_SM4_SWIZZLE_MASK) >> WINED3D_SM4_SWIZZLE_SHIFT;
1568 break;
1569
1570 default:
1571 FIXME("Unhandled swizzle type %#x.\n", swizzle_type);
1572 break;
1573 }
1574 }
1575
1576 return TRUE;
1577 }
1578
1579 static BOOL shader_sm4_read_dst_param(struct wined3d_sm4_data *priv, const DWORD **ptr, const DWORD *end,
1580 enum wined3d_data_type data_type, struct wined3d_shader_dst_param *dst_param)
1581 {
1582 enum wined3d_shader_src_modifier modifier;
1583 DWORD token;
1584
1585 if (*ptr >= end)
1586 {
1587 WARN("Invalid ptr %p >= end %p.\n", *ptr, end);
1588 return FALSE;
1589 }
1590 token = **ptr;
1591
1592 if (!shader_sm4_read_param(priv, ptr, end, data_type, &dst_param->reg, &modifier))
1593 {
1594 ERR("Failed to read parameter.\n");
1595 return FALSE;
1596 }
1597
1598 if (modifier != WINED3DSPSM_NONE)
1599 {
1600 ERR("Invalid source modifier %#x on destination register.\n", modifier);
1601 return FALSE;
1602 }
1603
1604 dst_param->write_mask = (token & WINED3D_SM4_WRITEMASK_MASK) >> WINED3D_SM4_WRITEMASK_SHIFT;
1605 dst_param->modifiers = 0;
1606 dst_param->shift = 0;
1607
1608 return TRUE;
1609 }
1610
1611 static void shader_sm4_read_instruction_modifier(DWORD modifier, struct wined3d_shader_instruction *ins)
1612 {
1613 static const DWORD recognized_bits = WINED3D_SM4_INSTRUCTION_MODIFIER
1614 | WINED3D_SM4_MODIFIER_AOFFIMMI
1615 | WINED3D_SM4_AOFFIMMI_U_MASK
1616 | WINED3D_SM4_AOFFIMMI_V_MASK
1617 | WINED3D_SM4_AOFFIMMI_W_MASK;
1618
1619 if (modifier & ~recognized_bits)
1620 {
1621 FIXME("Unhandled modifier 0x%08x.\n", modifier);
1622 }
1623 else
1624 {
1625 /* Bit fields are used for sign extension */
1626 struct
1627 {
1628 int u : 4;
1629 int v : 4;
1630 int w : 4;
1631 }
1632 aoffimmi;
1633 aoffimmi.u = (modifier & WINED3D_SM4_AOFFIMMI_U_MASK) >> WINED3D_SM4_AOFFIMMI_U_SHIFT;
1634 aoffimmi.v = (modifier & WINED3D_SM4_AOFFIMMI_V_MASK) >> WINED3D_SM4_AOFFIMMI_V_SHIFT;
1635 aoffimmi.w = (modifier & WINED3D_SM4_AOFFIMMI_W_MASK) >> WINED3D_SM4_AOFFIMMI_W_SHIFT;
1636 ins->texel_offset.u = aoffimmi.u;
1637 ins->texel_offset.v = aoffimmi.v;
1638 ins->texel_offset.w = aoffimmi.w;
1639 }
1640 }
1641
1642 static void shader_sm4_read_instruction(void *data, const DWORD **ptr, struct wined3d_shader_instruction *ins)
1643 {
1644 const struct wined3d_sm4_opcode_info *opcode_info;
1645 DWORD opcode_token, opcode, previous_token;
1646 struct wined3d_sm4_data *priv = data;
1647 unsigned int i, len;
1648 SIZE_T remaining;
1649 const DWORD *p;
1650 DWORD precise;
1651
1652 list_move_head(&priv->src_free, &priv->src);
1653
1654 if (*ptr >= priv->end)
1655 {
1656 WARN("End of byte-code, failed to read opcode.\n");
1657 goto fail;
1658 }
1659 remaining = priv->end - *ptr;
1660
1661 opcode_token = *(*ptr)++;
1662 opcode = opcode_token & WINED3D_SM4_OPCODE_MASK;
1663
1664 len = ((opcode_token & WINED3D_SM4_INSTRUCTION_LENGTH_MASK) >> WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT);
1665 if (!len)
1666 {
1667 if (remaining < 2)
1668 {
1669 WARN("End of byte-code, failed to read length token.\n");
1670 goto fail;
1671 }
1672 len = **ptr;
1673 }
1674 if (!len || remaining < len)
1675 {
1676 WARN("Read invalid length %u (remaining %lu).\n", len, remaining);
1677 goto fail;
1678 }
1679 --len;
1680
1681 if (TRACE_ON(d3d_bytecode))
1682 {
1683 TRACE_(d3d_bytecode)("[ %08x ", opcode_token);
1684 for (i = 0; i < len; ++i)
1685 {
1686 TRACE_(d3d_bytecode)("%08x ", (*ptr)[i]);
1687 }
1688 TRACE_(d3d_bytecode)("]\n");
1689 }
1690
1691 if (!(opcode_info = get_opcode_info(opcode)))
1692 {
1693 FIXME("Unrecognized opcode %#x, opcode_token 0x%08x.\n", opcode, opcode_token);
1694 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
1695 *ptr += len;
1696 return;
1697 }
1698
1699 ins->handler_idx = opcode_info->handler_idx;
1700 ins->flags = 0;
1701 ins->coissue = 0;
1702 ins->predicate = NULL;
1703 ins->dst_count = strlen(opcode_info->dst_info);
1704 ins->dst = priv->dst_param;
1705 ins->src_count = strlen(opcode_info->src_info);
1706 ins->src = priv->src_param;
1707 memset(&ins->texel_offset, 0, sizeof(ins->texel_offset));
1708
1709 p = *ptr;
1710 *ptr += len;
1711
1712 if (opcode_info->read_opcode_func)
1713 {
1714 opcode_info->read_opcode_func(ins, opcode, opcode_token, p, len, priv);
1715 }
1716 else
1717 {
1718 enum wined3d_shader_dst_modifier instruction_dst_modifier = WINED3DSPDM_NONE;
1719
1720 previous_token = opcode_token;
1721 while (previous_token & WINED3D_SM4_INSTRUCTION_MODIFIER && p != *ptr)
1722 shader_sm4_read_instruction_modifier(previous_token = *p++, ins);
1723
1724 ins->flags = (opcode_token & WINED3D_SM4_INSTRUCTION_FLAGS_MASK) >> WINED3D_SM4_INSTRUCTION_FLAGS_SHIFT;
1725 if (ins->flags & WINED3D_SM4_INSTRUCTION_FLAG_SATURATE)
1726 {
1727 ins->flags &= ~WINED3D_SM4_INSTRUCTION_FLAG_SATURATE;
1728 instruction_dst_modifier = WINED3DSPDM_SATURATE;
1729 }
1730 precise = (opcode_token & WINED3D_SM5_PRECISE_MASK) >> WINED3D_SM5_PRECISE_SHIFT;
1731 ins->flags |= precise << WINED3DSI_PRECISE_SHIFT;
1732
1733 for (i = 0; i < ins->dst_count; ++i)
1734 {
1735 if (!(shader_sm4_read_dst_param(priv, &p, *ptr, map_data_type(opcode_info->dst_info[i]),
1736 &priv->dst_param[i])))
1737 {
1738 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
1739 return;
1740 }
1741 priv->dst_param[i].modifiers |= instruction_dst_modifier;
1742 }
1743
1744 for (i = 0; i < ins->src_count; ++i)
1745 {
1746 if (!(shader_sm4_read_src_param(priv, &p, *ptr, map_data_type(opcode_info->src_info[i]),
1747 &priv->src_param[i])))
1748 {
1749 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
1750 return;
1751 }
1752 }
1753 }
1754
1755 return;
1756
1757 fail:
1758 *ptr = priv->end;
1759 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
1760 return;
1761 }
1762
1763 static BOOL shader_sm4_is_end(void *data, const DWORD **ptr)
1764 {
1765 struct wined3d_sm4_data *priv = data;
1766 return *ptr == priv->end;
1767 }
1768
1769 const struct wined3d_shader_frontend sm4_shader_frontend =
1770 {
1771 shader_sm4_init,
1772 shader_sm4_free,
1773 shader_sm4_read_header,
1774 shader_sm4_read_instruction,
1775 shader_sm4_is_end,
1776 };
1777
1778 #define TAG_AON9 WINEMAKEFOURCC('A', 'o', 'n', '9')
1779 #define TAG_DXBC WINEMAKEFOURCC('D', 'X', 'B', 'C')
1780 #define TAG_ISGN WINEMAKEFOURCC('I', 'S', 'G', 'N')
1781 #define TAG_OSG5 WINEMAKEFOURCC('O', 'S', 'G', '5')
1782 #define TAG_OSGN WINEMAKEFOURCC('O', 'S', 'G', 'N')
1783 #define TAG_PCSG WINEMAKEFOURCC('P', 'C', 'S', 'G')
1784 #define TAG_SHDR WINEMAKEFOURCC('S', 'H', 'D', 'R')
1785 #define TAG_SHEX WINEMAKEFOURCC('S', 'H', 'E', 'X')
1786
1787 struct aon9_header
1788 {
1789 DWORD chunk_size;
1790 DWORD shader_version;
1791 DWORD unknown;
1792 DWORD byte_code_offset;
1793 };
1794
1795 struct shader_handler_context
1796 {
1797 struct wined3d_shader *shader;
1798 enum wined3d_shader_byte_code_format *format;
1799 unsigned int max_version;
1800 };
1801
1802 static void read_dword(const char **ptr, DWORD *d)
1803 {
1804 memcpy(d, *ptr, sizeof(*d));
1805 *ptr += sizeof(*d);
1806 }
1807
1808 static BOOL require_space(size_t offset, size_t count, size_t size, size_t data_size)
1809 {
1810 return !count || (data_size - offset) / count >= size;
1811 }
1812
1813 static void skip_dword_unknown(const char **ptr, unsigned int count)
1814 {
1815 unsigned int i;
1816 DWORD d;
1817
1818 WARN("Skipping %u unknown DWORDs:\n", count);
1819 for (i = 0; i < count; ++i)
1820 {
1821 read_dword(ptr, &d);
1822 WARN("\t0x%08x\n", d);
1823 }
1824 }
1825
1826 static HRESULT parse_dxbc(const char *data, SIZE_T data_size,
1827 HRESULT (*chunk_handler)(const char *data, DWORD data_size, DWORD tag, void *ctx), void *ctx)
1828 {
1829 const char *ptr = data;
1830 HRESULT hr = S_OK;
1831 DWORD chunk_count;
1832 DWORD total_size;
1833 unsigned int i;
1834 DWORD version;
1835 DWORD tag;
1836
1837 read_dword(&ptr, &tag);
1838 TRACE("tag: %s.\n", debugstr_an((const char *)&tag, 4));
1839
1840 if (tag != TAG_DXBC)
1841 {
1842 WARN("Wrong tag.\n");
1843 return E_INVALIDARG;
1844 }
1845
1846 WARN("Ignoring DXBC checksum.\n");
1847 skip_dword_unknown(&ptr, 4);
1848
1849 read_dword(&ptr, &version);
1850 TRACE("version: %#x.\n", version);
1851 if (version != 0x00000001)
1852 {
1853 WARN("Got unexpected DXBC version %#x.\n", version);
1854 return E_INVALIDARG;
1855 }
1856
1857 read_dword(&ptr, &total_size);
1858 TRACE("total size: %#x\n", total_size);
1859
1860 read_dword(&ptr, &chunk_count);
1861 TRACE("chunk count: %#x\n", chunk_count);
1862
1863 for (i = 0; i < chunk_count; ++i)
1864 {
1865 DWORD chunk_tag, chunk_size;
1866 const char *chunk_ptr;
1867 DWORD chunk_offset;
1868
1869 read_dword(&ptr, &chunk_offset);
1870 TRACE("chunk %u at offset %#x\n", i, chunk_offset);
1871
1872 if (chunk_offset >= data_size || !require_space(chunk_offset, 2, sizeof(DWORD), data_size))
1873 {
1874 WARN("Invalid chunk offset %#x (data size %#lx).\n", chunk_offset, data_size);
1875 return E_FAIL;
1876 }
1877
1878 chunk_ptr = data + chunk_offset;
1879
1880 read_dword(&chunk_ptr, &chunk_tag);
1881 read_dword(&chunk_ptr, &chunk_size);
1882
1883 if (!require_space(chunk_ptr - data, 1, chunk_size, data_size))
1884 {
1885 WARN("Invalid chunk size %#x (data size %#lx, chunk offset %#x).\n",
1886 chunk_size, data_size, chunk_offset);
1887 return E_FAIL;
1888 }
1889
1890 if (FAILED(hr = chunk_handler(chunk_ptr, chunk_size, chunk_tag, ctx)))
1891 break;
1892 }
1893
1894 return hr;
1895 }
1896
1897 static const char *shader_get_string(const char *data, size_t data_size, DWORD offset)
1898 {
1899 size_t len, max_len;
1900
1901 if (offset >= data_size)
1902 {
1903 WARN("Invalid offset %#x (data size %#lx).\n", offset, (long)data_size);
1904 return NULL;
1905 }
1906
1907 max_len = data_size - offset;
1908 len = strnlen(data + offset, max_len);
1909
1910 if (len == max_len)
1911 return NULL;
1912
1913 return data + offset;
1914 }
1915
1916 static HRESULT shader_parse_signature(DWORD tag, const char *data, DWORD data_size,
1917 struct wined3d_shader_signature *s)
1918 {
1919 struct wined3d_shader_signature_element *e;
1920 const char *ptr = data;
1921 unsigned int i;
1922 DWORD count;
1923
1924 if (!require_space(0, 2, sizeof(DWORD), data_size))
1925 {
1926 WARN("Invalid data size %#x.\n", data_size);
1927 return E_INVALIDARG;
1928 }
1929
1930 read_dword(&ptr, &count);
1931 TRACE("%u elements.\n", count);
1932
1933 skip_dword_unknown(&ptr, 1); /* It seems to always be 0x00000008. */
1934
1935 if (!require_space(ptr - data, count, 6 * sizeof(DWORD), data_size))
1936 {
1937 WARN("Invalid count %#x (data size %#x).\n", count, data_size);
1938 return E_INVALIDARG;
1939 }
1940
1941 if (!(e = heap_calloc(count, sizeof(*e))))
1942 {
1943 ERR("Failed to allocate input signature memory.\n");
1944 return E_OUTOFMEMORY;
1945 }
1946
1947 for (i = 0; i < count; ++i)
1948 {
1949 DWORD name_offset;
1950
1951 if (tag == TAG_OSG5)
1952 read_dword(&ptr, &e[i].stream_idx);
1953 else
1954 e[i].stream_idx = 0;
1955 read_dword(&ptr, &name_offset);
1956 if (!(e[i].semantic_name = shader_get_string(data, data_size, name_offset)))
1957 {
1958 WARN("Invalid name offset %#x (data size %#x).\n", name_offset, data_size);
1959 heap_free(e);
1960 return E_INVALIDARG;
1961 }
1962 read_dword(&ptr, &e[i].semantic_idx);
1963 read_dword(&ptr, &e[i].sysval_semantic);
1964 read_dword(&ptr, &e[i].component_type);
1965 read_dword(&ptr, &e[i].register_idx);
1966 read_dword(&ptr, &e[i].mask);
1967
1968 TRACE("Stream: %u, semantic: %s, semantic idx: %u, sysval_semantic %#x, "
1969 "type %u, register idx: %u, use_mask %#x, input_mask %#x.\n",
1970 e[i].stream_idx, debugstr_a(e[i].semantic_name), e[i].semantic_idx, e[i].sysval_semantic,
1971 e[i].component_type, e[i].register_idx, (e[i].mask >> 8) & 0xff, e[i].mask & 0xff);
1972 }
1973
1974 s->elements = e;
1975 s->element_count = count;
1976
1977 return S_OK;
1978 }
1979
1980 static HRESULT shader_dxbc_chunk_handler(const char *data, DWORD data_size, DWORD tag, void *context)
1981 {
1982 struct shader_handler_context *ctx = context;
1983 struct wined3d_shader *shader = ctx->shader;
1984 HRESULT hr;
1985
1986 switch (tag)
1987 {
1988 case TAG_ISGN:
1989 if (ctx->max_version < 4)
1990 {
1991 TRACE("Skipping shader input signature.\n");
1992 break;
1993 }
1994 if (shader->input_signature.elements)
1995 {
1996 FIXME("Multiple input signatures.\n");
1997 break;
1998 }
1999 if (FAILED(hr = shader_parse_signature(tag, data, data_size, &shader->input_signature)))
2000 return hr;
2001 break;
2002
2003 case TAG_OSGN:
2004 case TAG_OSG5:
2005 if (ctx->max_version < 4)
2006 {
2007 TRACE("Skipping shader output signature.\n");
2008 break;
2009 }
2010 if (shader->output_signature.elements)
2011 {
2012 FIXME("Multiple output signatures.\n");
2013 break;
2014 }
2015 if (FAILED(hr = shader_parse_signature(tag, data, data_size, &shader->output_signature)))
2016 return hr;
2017 break;
2018
2019 case TAG_PCSG:
2020 if (shader->patch_constant_signature.elements)
2021 {
2022 FIXME("Multiple patch constant signatures.\n");
2023 break;
2024 }
2025 if (FAILED(hr = shader_parse_signature(tag, data, data_size, &shader->patch_constant_signature)))
2026 return hr;
2027 break;
2028
2029 case TAG_SHDR:
2030 case TAG_SHEX:
2031 if (ctx->max_version < 4)
2032 {
2033 TRACE("Skipping SM4+ shader.\n");
2034 break;
2035 }
2036 if (shader->function)
2037 FIXME("Multiple shader code chunks.\n");
2038 shader->function = (const DWORD *)data;
2039 shader->functionLength = data_size;
2040 *ctx->format = WINED3D_SHADER_BYTE_CODE_FORMAT_SM4;
2041 break;
2042
2043 case TAG_AON9:
2044 if (ctx->max_version < 4)
2045 {
2046 const struct aon9_header *header = (const struct aon9_header *)data;
2047 unsigned int unknown_dword_count;
2048 const char *byte_code;
2049
2050 if (data_size < sizeof(*header))
2051 {
2052 WARN("Invalid Aon9 data size %#x.\n", data_size);
2053 return E_FAIL;
2054 }
2055 byte_code = data + header->byte_code_offset;
2056 unknown_dword_count = (header->byte_code_offset - sizeof(*header)) / sizeof(DWORD);
2057
2058 if (data_size - 2 * sizeof(DWORD) < header->byte_code_offset)
2059 {
2060 WARN("Invalid byte code offset %#x (size %#x).\n", header->byte_code_offset, data_size);
2061 return E_FAIL;
2062 }
2063 FIXME("Skipping %u unknown DWORDs.\n", unknown_dword_count);
2064
2065 if (shader->function)
2066 FIXME("Multiple shader code chunks.\n");
2067 shader->function = (const DWORD *)byte_code;
2068 shader->functionLength = data_size - header->byte_code_offset;
2069 *ctx->format = WINED3D_SHADER_BYTE_CODE_FORMAT_SM1;
2070 TRACE("Feature level 9 shader version 0%08x, 0%08x.\n",
2071 header->shader_version, *shader->function);
2072 }
2073 else
2074 {
2075 TRACE("Skipping feature level 9 shader code.\n");
2076 }
2077 break;
2078
2079 default:
2080 TRACE("Skipping chunk %s.\n", debugstr_an((const char *)&tag, 4));
2081 break;
2082 }
2083
2084 return S_OK;
2085 }
2086
2087 HRESULT shader_extract_from_dxbc(struct wined3d_shader *shader,
2088 unsigned int max_shader_version, enum wined3d_shader_byte_code_format *format)
2089 {
2090 struct shader_handler_context ctx;
2091 HRESULT hr;
2092
2093 ctx.shader = shader;
2094 ctx.format = format;
2095 ctx.max_version = max_shader_version;
2096
2097 hr = parse_dxbc(shader->byte_code, shader->byte_code_size, shader_dxbc_chunk_handler, &ctx);
2098 if (!shader->function)
2099 hr = E_INVALIDARG;
2100
2101 if (FAILED(hr))
2102 WARN("Failed to parse DXBC, hr %#x.\n", hr);
2103
2104 return hr;
2105 }
Windows API implementation