On 9/24/20 9:33 PM, Peter Maydell wrote: > On Thu, 24 Sep 2020 at 20:30, Richard Henderson > <richard.hender...@linaro.org> wrote: >> >> On 9/23/20 3:30 AM, Philippe Mathieu-Daudé wrote: >>> On 9/23/20 12:16 PM, Peter Maydell wrote: >>>> I suppose if we wanted to just use the __builtin_bswap* >>>> implementation on all hosts (ie drop the special casing >>>> of CONFIG_MACHINE_BSWAP_H/FreeBSD/Haiku) we'd need to >>>> rename our macros to avoid potential conflicts with the >>>> versions in the system headers there in case they were >>>> pulled in via some other path ? >>> >>> Yes, this is why I couldn't get ride of everything >>> >>> After reading Paolo's suggestion to use qatomic*, I >>> am tempted to suggest qbswap* but I am still looking >>> for better alternatives... >> >> Hum, maybe. It's pretty ugly. We could just leave those alone and hope the >> system version is decent enough. > > Mostly what I would like is to be able to be rid of the > configure machinery so we could just use a single > portable implementation.
Good news. Using the Haiku image provided by Alexander, GCC is pretty recent and supports builtin atomic: > uname -a Haiku shredder 1 hrev54154+115 Jun 30 2020 15:20 x86_64 x86_64 Haiku > gcc --version gcc (2019_05_24) 8.3.0 Copyright (C) 2018 Free Software Foundation, Inc. The <endian.h> include (/system/develop/headers/posix/endian.h) only defines BYTE_ORDER. Looking at FreeBSD, bswap*() are defined as macro: https://github.com/freebsd/freebsd/blob/master/sys/sys/endian.h#L58 So we could eventually undefine them... but the manpage describe them as functions [*]: DESCRIPTION The bswap16(), bswap32(), and bswap64() functions return a byte order swapped integer. So theoretically they could change the prototype still respecting the documentation contract, and undefines wouldn't work anymore. Would this be acceptable? Any clever idea? [*] https://www.freebsd.org/cgi/man.cgi?query=bswap32&apropos=0&sektion=9 > > thanks > -- PMM >
