24
24
#include "libavcodec/cabac.h"
25
25
#include "libavutil/attributes.h"
26
#include "libavutil/x86_cpu.h"
26
#include "libavutil/x86/asm.h"
27
#include "libavutil/internal.h"
27
28
#include "config.h"
30
#define BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, lowword, range, tmp)\
32
#ifdef BROKEN_RELOCATIONS
33
#define TABLES_ARG , "r"(tables)
36
#define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
37
"cmp "low" , "tmp" \n\t"\
38
"cmova %%ecx , "range" \n\t"\
39
"sbb %%rcx , %%rcx \n\t"\
40
"and %%ecx , "tmp" \n\t"\
41
"xor %%rcx , "retq" \n\t"\
42
"sub "tmp" , "low" \n\t"
43
#else /* HAVE_FAST_CMOV */
44
#define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
45
/* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \
46
"sub "low" , "tmp" \n\t"\
47
"sar $31 , "tmp" \n\t"\
48
"sub %%ecx , "range" \n\t"\
49
"and "tmp" , "range" \n\t"\
50
"add %%ecx , "range" \n\t"\
51
"shl $17 , %%ecx \n\t"\
52
"and "tmp" , %%ecx \n\t"\
53
"sub %%ecx , "low" \n\t"\
54
"xor "tmp" , "ret" \n\t"\
55
"movslq "ret" , "retq" \n\t"
56
#endif /* HAVE_FAST_CMOV */
58
#define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
59
"movzbl "statep" , "ret" \n\t"\
60
"mov "range" , "tmp" \n\t"\
61
"and $0xC0 , "range" \n\t"\
62
"lea ("ret", "range", 2), %%ecx \n\t"\
63
"movzbl "lps_off"("tables", %%rcx), "range" \n\t"\
64
"sub "range" , "tmp" \n\t"\
65
"mov "tmp" , %%ecx \n\t"\
66
"shl $17 , "tmp" \n\t"\
67
BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
68
"movzbl "norm_off"("tables", "rangeq"), %%ecx \n\t"\
69
"shl %%cl , "range" \n\t"\
70
"movzbl "mlps_off"+128("tables", "retq"), "tmp" \n\t"\
71
"shl %%cl , "low" \n\t"\
72
"mov "tmpbyte" , "statep" \n\t"\
73
"test "lowword" , "lowword" \n\t"\
75
"mov "byte" , %%"REG_c" \n\t"\
76
"cmp "end" , %%"REG_c" \n\t"\
78
"add"OPSIZE" $2 , "byte" \n\t"\
80
"movzwl (%%"REG_c") , "tmp" \n\t"\
81
"lea -1("low") , %%ecx \n\t"\
82
"xor "low" , %%ecx \n\t"\
83
"shr $15 , %%ecx \n\t"\
85
"shr $15 , "tmp" \n\t"\
86
"movzbl "norm_off"("tables", %%rcx), %%ecx \n\t"\
87
"sub $0xFFFF , "tmp" \n\t"\
89
"add $7 , %%ecx \n\t"\
90
"shl %%cl , "tmp" \n\t"\
91
"add "tmp" , "low" \n\t"\
94
#else /* BROKEN_RELOCATIONS */
98
#define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
31
99
"mov "tmp" , %%ecx \n\t"\
32
100
"shl $17 , "tmp" \n\t"\
33
101
"cmp "low" , "tmp" \n\t"\
51
119
"xor "tmp" , "ret" \n\t"
52
120
#endif /* HAVE_FAST_CMOV */
54
#define BRANCHLESS_GET_CABAC(ret, statep, low, lowword, range, tmp, tmpbyte, byte) \
122
#define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
55
123
"movzbl "statep" , "ret" \n\t"\
56
124
"mov "range" , "tmp" \n\t"\
57
125
"and $0xC0 , "range" \n\t"\
58
"movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
126
"movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
59
127
"sub "range" , "tmp" \n\t"\
60
BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, lowword, range, tmp) \
61
"movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx \n\t"\
128
BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp) \
129
"movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx \n\t"\
62
130
"shl %%cl , "range" \n\t"\
63
"movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp" \n\t"\
131
"movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp" \n\t"\
64
132
"shl %%cl , "low" \n\t"\
65
133
"mov "tmpbyte" , "statep" \n\t"\
66
134
"test "lowword" , "lowword" \n\t"\
68
136
"mov "byte" , %%"REG_c" \n\t"\
137
"cmp "end" , %%"REG_c" \n\t"\
69
139
"add"OPSIZE" $2 , "byte" \n\t"\
70
141
"movzwl (%%"REG_c") , "tmp" \n\t"\
71
142
"lea -1("low") , %%ecx \n\t"\
72
143
"xor "low" , %%ecx \n\t"\
73
144
"shr $15 , %%ecx \n\t"\
74
145
"bswap "tmp" \n\t"\
75
146
"shr $15 , "tmp" \n\t"\
76
"movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t"\
147
"movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
77
148
"sub $0xFFFF , "tmp" \n\t"\
79
150
"add $7 , %%ecx \n\t"\
80
151
"shl %%cl , "tmp" \n\t"\
81
152
"add "tmp" , "low" \n\t"\
84
#if HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
155
#endif /* BROKEN_RELOCATIONS */
85
159
#define get_cabac_inline get_cabac_inline_x86
86
160
static av_always_inline int get_cabac_inline_x86(CABACContext *c,
87
161
uint8_t *const state)
92
BRANCHLESS_GET_CABAC("%0", "(%5)", "%1", "%w1", "%2",
94
:"=&r"(bit), "+&r"(c->low), "+&r"(c->range), "=&q"(tmp),
164
#ifdef BROKEN_RELOCATIONS
168
"lea "MANGLE(ff_h264_cabac_tables)", %0 \n\t"
174
BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1",
175
"%2", "%q2", "%3", "%b3",
176
"%c6(%5)", "%c7(%5)",
177
AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
178
AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
179
AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
181
: "=&r"(bit), "+&r"(c->low), "+&r"(c->range), "=&q"(tmp)
182
: "r"(state), "r"(c),
183
"i"(offsetof(CABACContext, bytestream)),
184
"i"(offsetof(CABACContext, bytestream_end))
97
186
: "%"REG_c, "memory"
101
#endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
190
#endif /* HAVE_7REGS */
103
192
#define get_cabac_bypass_sign get_cabac_bypass_sign_x86
104
193
static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
107
196
__asm__ volatile(
109
"movl %2, %%eax \n\t"
111
"add %%eax, %%eax \n\t"
112
"sub %k1, %%eax \n\t"
114
"and %%edx, %k1 \n\t"
115
"add %k1, %%eax \n\t"
116
"xor %%edx, %%ecx \n\t"
117
"sub %%edx, %%ecx \n\t"
118
"test %%ax, %%ax \n\t"
121
"subl $0xFFFF, %%eax \n\t"
122
"movzwl (%1), %%edx \n\t"
124
"shrl $15, %%edx \n\t"
126
"addl %%edx, %%eax \n\t"
129
"movl %%eax, %2 \n\t"
197
"movl %c6(%2), %k1 \n\t"
198
"movl %c3(%2), %%eax \n\t"
200
"add %%eax, %%eax \n\t"
201
"sub %k1, %%eax \n\t"
203
"and %%edx, %k1 \n\t"
204
"add %k1, %%eax \n\t"
205
"xor %%edx, %%ecx \n\t"
206
"sub %%edx, %%ecx \n\t"
207
"test %%ax, %%ax \n\t"
209
"mov %c4(%2), %1 \n\t"
210
"subl $0xFFFF, %%eax \n\t"
211
"movzwl (%1), %%edx \n\t"
213
"shrl $15, %%edx \n\t"
214
"addl %%edx, %%eax \n\t"
215
"cmp %c5(%2), %1 \n\t"
217
"add"OPSIZE" $2, %c4(%2) \n\t"
219
"movl %%eax, %c3(%2) \n\t"
131
:"+c"(val), "=&r"(tmp), "+m"(c->low), "+m"(c->bytestream)
221
: "+c"(val), "=&r"(tmp)
223
"i"(offsetof(CABACContext, low)),
224
"i"(offsetof(CABACContext, bytestream)),
225
"i"(offsetof(CABACContext, bytestream_end)),
226
"i"(offsetof(CABACContext, range))
227
: "%eax", "%edx", "memory"
232
#endif /* HAVE_INLINE_ASM */
138
233
#endif /* AVCODEC_X86_CABAC_H */