Skip to content

Commit 38e5363

Browse files
ZERICO2005ZERICO2005
committed
�optimized llshrs, added routines to <stdbit.h> and <bit> that directly correspond to builtins
1 parent 1e85adb commit 38e5363

File tree

5 files changed

+458
-4
lines changed

5 files changed

+458
-4
lines changed

‎src/crt/llshrs.src‎

Lines changed: 12 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,9 @@
11
assume adl=1
22

33
section .text
4+
45
public __llshrs
6+
57
__llshrs:
68
; Suboptimal for large shift amounts
79
push af
@@ -19,14 +21,20 @@ __llshrs:
1921
.hijack_llshru:
2022
rr c
2123
rr (iy - 1)
22-
rr (iy - 2)
23-
rr (iy - 3)
24+
rr d
25+
rr e
2426
rr (iy - 4)
25-
rr (iy - 5)
26-
rr (iy - 6)
27+
rr h
28+
rr l
2729
dec a
2830
jr nz, .loop
31+
32+
ld (iy - 3), e
33+
ld (iy - 2), d
34+
ex de, hl
2935
pop hl
36+
ld l, e
37+
ld h, d
3038
pop de
3139
public __llshrs.finish
3240
.finish:

‎src/libc/header_test.c‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -14,6 +14,7 @@
1414
#include <setjmp.h>
1515
#include <stdalign.h>
1616
#include <stdarg.h>
17+
#include <stdbit.h>
1718
#include <stdbool.h>
1819
#include <stddef.h>
1920
#include <stdint.h>

‎src/libc/include/stdbit.h‎

Lines changed: 242 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,242 @@
1+
#ifndef _STDBIT_H
2+
#define _STDBIT_H
3+
4+
#include <cdefs.h>
5+
#include <stdbool.h>
6+
7+
#ifdef __cplusplus
8+
extern "C" {
9+
#endif
10+
11+
// #define __STDC_VERSION_STDBIT_H__ 202311L
12+
13+
#define __STDC_ENDIAN_LITTLE__ __ORDER_LITTLE_ENDIAN__
14+
#define __STDC_ENDIAN_BIG__ __ORDER_BIG_ENDIAN__
15+
#define __STDC_ENDIAN_NATIVE__ __BYTE_ORDER__
16+
17+
//------------------------------------------------------------------------------
18+
// stdc_leading_zeros
19+
//------------------------------------------------------------------------------
20+
21+
// #define stdc_leading_zeros_uc(x)
22+
23+
#define stdc_leading_zeros_us(x) \
24+
__builtin_clzs(x)
25+
26+
#define stdc_leading_zeros_ui(x) \
27+
__builtin_clz(x)
28+
29+
#define stdc_leading_zeros_ul(x) \
30+
__builtin_clzl(x)
31+
32+
// #define stdc_leading_zeros_ui48(x)
33+
34+
#define stdc_leading_zeros_ull(x) \
35+
__builtin_clzll(x)
36+
37+
#define stdc_leading_zeros(x) _Generic((x), \
38+
/* unsigned char: stdc_leading_zeros_uc, */ \
39+
unsigned short: stdc_leading_zeros_us, \
40+
unsigned int: stdc_leading_zeros_ui, \
41+
unsigned long: stdc_leading_zeros_ul, \
42+
/* unsigned __int48: stdc_leading_zeros_ui48, */ \
43+
unsigned long long: stdc_leading_zeros_ull \
44+
)(x)
45+
46+
//------------------------------------------------------------------------------
47+
// stdc_leading_ones
48+
//------------------------------------------------------------------------------
49+
50+
51+
52+
//------------------------------------------------------------------------------
53+
// stdc_trailing_zeros
54+
//------------------------------------------------------------------------------
55+
56+
// #define stdc_trailing_zeros_uc(x)
57+
58+
#define stdc_trailing_zeros_us(x) \
59+
__builtin_ctzs(x)
60+
61+
#define stdc_trailing_zeros_ui(x) \
62+
__builtin_ctz(x)
63+
64+
#define stdc_trailing_zeros_ul(x) \
65+
__builtin_ctzl(x)
66+
67+
// #define stdc_trailing_zeros_ui48(x)
68+
69+
#define stdc_trailing_zeros_ull(x) \
70+
__builtin_ctzll(x)
71+
72+
#define stdc_trailing_zeros(x) _Generic((x), \
73+
/* unsigned char: stdc_trailing_zeros_uc, */ \
74+
unsigned short: stdc_trailing_zeros_us, \
75+
unsigned int: stdc_trailing_zeros_ui, \
76+
unsigned long: stdc_trailing_zeros_ul, \
77+
/* unsigned __int48: stdc_trailing_zeros_ui48, */ \
78+
unsigned long long: stdc_trailing_zeros_ull \
79+
)(x)
80+
81+
//------------------------------------------------------------------------------
82+
// stdc_trailing_ones
83+
//------------------------------------------------------------------------------
84+
85+
86+
87+
//------------------------------------------------------------------------------
88+
// stdc_first_leading_zero
89+
//------------------------------------------------------------------------------
90+
91+
92+
93+
//------------------------------------------------------------------------------
94+
// stdc_first_trailing_one
95+
//------------------------------------------------------------------------------
96+
97+
#define stdc_first_trailing_one_uc(x) \
98+
__builtin_ffs((unsigned int)x) /** @todo fix this inefficiency */
99+
100+
#define stdc_first_trailing_one_us(x) \
101+
__builtin_ffs((unsigned int)x) /** @todo fix this inefficiency */
102+
103+
#define stdc_first_trailing_one_ui(x) \
104+
__builtin_ffs(x)
105+
106+
#define stdc_first_trailing_one_ul(x) \
107+
__builtin_ffsl(x)
108+
109+
#define stdc_first_trailing_one_ui48(x) \
110+
__builtin_ffsll((unsigned long long)x) /** @todo fix this inefficiency */
111+
112+
#define stdc_first_trailing_one_ull(x) \
113+
__builtin_ffsll(x)
114+
115+
#define stdc_first_trailing_one(x) _Generic((x), \
116+
unsigned char: stdc_first_trailing_one_uc, \
117+
unsigned short: stdc_first_trailing_one_us, \
118+
unsigned int: stdc_first_trailing_one_ui, \
119+
unsigned long: stdc_first_trailing_one_ul, \
120+
unsigned __int48: stdc_first_trailing_one_ui48, \
121+
unsigned long long: stdc_first_trailing_one_ull \
122+
)(x)
123+
124+
//------------------------------------------------------------------------------
125+
// stdc_first_trailing_zero
126+
//------------------------------------------------------------------------------
127+
128+
129+
130+
//------------------------------------------------------------------------------
131+
// stdc_count_ones
132+
//------------------------------------------------------------------------------
133+
134+
#define stdc_count_ones_uc(x) \
135+
__builtin_popcount((unsigned int)x) /** @todo fix this inefficiency */
136+
#define stdc_count_ones_us(x) \
137+
__builtin_popcount((unsigned int)x) /** @todo fix this inefficiency */
138+
#define stdc_count_ones_ui(x) \
139+
__builtin_popcount(x)
140+
#define stdc_count_ones_ul(x) \
141+
__builtin_popcountl(x)
142+
#define stdc_count_ones_ui48(x) \
143+
__builtin_popcountll((unsigned long long)x) /** @todo fix this inefficiency */
144+
#define stdc_count_ones_ull(x) \
145+
__builtin_popcountll(x)
146+
147+
#define stdc_count_ones(x) _Generic((x), \
148+
unsigned char: stdc_count_ones_uc, \
149+
unsigned short: stdc_count_ones_us, \
150+
unsigned int: stdc_count_ones_ui, \
151+
unsigned long: stdc_count_ones_ul, \
152+
unsigned __int48: stdc_count_ones_ui48, \
153+
unsigned long long: stdc_count_ones_ull \
154+
)(x)
155+
156+
//------------------------------------------------------------------------------
157+
// stdc_count_zeros
158+
//------------------------------------------------------------------------------
159+
160+
161+
162+
//------------------------------------------------------------------------------
163+
// stdc_has_single_bit
164+
//------------------------------------------------------------------------------
165+
166+
167+
168+
//------------------------------------------------------------------------------
169+
// stdc_bit_width
170+
//------------------------------------------------------------------------------
171+
172+
173+
174+
//------------------------------------------------------------------------------
175+
// stdc_bit_floor
176+
//------------------------------------------------------------------------------
177+
178+
179+
180+
//------------------------------------------------------------------------------
181+
// stdc_rotate_left
182+
//------------------------------------------------------------------------------
183+
184+
#define stdc_rotate_left_uc(x, s) \
185+
__builtin_rotateleft8(x, s)
186+
187+
#define stdc_rotate_left_us(x, s) \
188+
__builtin_rotateleft16(x, s)
189+
190+
// #define stdc_rotate_left_ui(x, s)
191+
192+
#define stdc_rotate_left_ul(x, s) \
193+
__builtin_rotateleft32(x, s)
194+
195+
// #define stdc_rotate_left_ui48(x, s)
196+
197+
#define stdc_rotate_left_ull(x, s) \
198+
__builtin_rotateleft64(x, s)
199+
200+
#define stdc_rotate_left(x, s) _Generic((x), \
201+
unsigned char: stdc_rotate_left_uc, \
202+
unsigned short: stdc_rotate_left_us, \
203+
/* unsigned int: stdc_rotate_left_ui, */ \
204+
unsigned long: stdc_rotate_left_ul, \
205+
/* unsigned __int48: stdc_rotate_left_ui48, */ \
206+
unsigned long long: stdc_rotate_left_ull \
207+
)(x, s)
208+
209+
//------------------------------------------------------------------------------
210+
// stdc_rotate_right
211+
//------------------------------------------------------------------------------
212+
213+
#define stdc_rotate_right_uc(x, s) \
214+
__builtin_rotateright8(x, s)
215+
216+
#define stdc_rotate_right_us(x, s) \
217+
__builtin_rotateright16(x, s)
218+
219+
// #define stdc_rotate_right_ui(x, s)
220+
221+
#define stdc_rotate_right_ul(x, s) \
222+
__builtin_rotateright32(x, s)
223+
224+
// #define stdc_rotate_right_ui48(x, s)
225+
226+
#define stdc_rotate_right_ull(x, s) \
227+
__builtin_rotateright64(x, s)
228+
229+
#define stdc_rotate_right(x, s) _Generic((x), \
230+
unsigned char: stdc_rotate_right_uc, \
231+
unsigned short: stdc_rotate_right_us, \
232+
/* unsigned int: stdc_rotate_right_ui, */ \
233+
unsigned long: stdc_rotate_right_ul, \
234+
/* unsigned __int48: stdc_rotate_right_ui48, */ \
235+
unsigned long long: stdc_rotate_right_ull \
236+
)(x, s)
237+
238+
#ifdef __cplusplus
239+
}
240+
#endif
241+
242+
#endif /* _STDBIT_H */

‎src/libcxx/header_test.cpp‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -52,6 +52,7 @@
5252
#include <setjmp.h>
5353
#include <stdalign.h>
5454
#include <stdarg.h>
55+
#include <stdbit.h>
5556
#include <stdbool.h>
5657
#include <stddef.h>
5758
#include <stdint.h>

0 commit comments

Comments
 (0)