Hi, David
From: David Laight
Sent: 27 August 2023 22:52
...
Of course, we can also use the __stringify() trick to do so, but it is expensive (bigger size, worse performance) to unstringify and get the number again, the expensive atoi() 'works' for the numeric __NR_*, but not work for (__NR_*_base + offset) like __NR_* definitions (used by ARM and MIPS), a simple interpreter is required for such cases and it is more expensive than atoi().
/* not for ARM and MIPS */ static int atoi(const char *s); #define __get_nr(name) __nr_atoi(__stringify(__NR_##name)) #define __nr_atoi(str) (str[0] == '_' ? -1L : ___nr_atoi(str)) #define ___nr_atoi(str) (str[0] == '(' ? -1L : atoi(str))Welcome more discussion or let's simply throw away this direction ;-)
While it will look horrid the it ought to be possible to get the compiler to evaluate the string.
...
So something that starts: #define dig(c) (c < '0' || c > '9' ? 999999 : c - '0') str[0] == '_' ? -1 : str[0] != '(' ? str[1] == ' ' ? dig(str[0]) : str[2] == '1' ? (dig(str[0]) * 10 + dig(str[1]) : Any unexpected character will expand the 99999 and generate an over-large result.
See https://godbolt.org/z/rear4c1hj
That will convert "1234" or "(1234 + 5678)" (or shorter numbers) as a compile-time constant.
Thanks very much, it works perfectly.
I tuned it for more complicated cases, including ((0x900000+0x0f0000)+5) used by ARM+OABI (not used by nolibc), now, it should work for all of the architectures: https://godbolt.org/z/a7hxWj83E ;-)
To get fast building, we can provide different versions for different architectures. A simple test shows, only two versions (as you mentioned above, one is "1234" converting, another is "(1234 + 5678)" calculating) are enough for current nolibc supported architectures and the building of nolibc-test.c is not slow.
With the __stringify() based __is_nr_defined() macro and this new __nrtoi() macro based __get_nr() macro, there is no need to redefine the old NOLIBC__NR_* macros, as a result, all of the duplicated -ENOSYS return lines and even all of the #ifdef's from sys.h could be dropped and even no need to add them for new future syscalls, and also, the old syscall() macro can return -ENOSYS at the runtime instead of any compiling failures.
For the sys_* definitions, to avoid forgetting passing the arguments, instead of using __VA_ARGS__, perhaps we should simply passing all of the arguments.
Best Regards, Zhangjin
David
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)