Re: [PATCH 3/4] hurd: Microoptimize mmap ()
Sergey Bugaev, le mar. 25 avril 2023 00:09:58 +0300, a ecrit: > What I should rather look into is marking __hurd_fail and friends with > __attribute__((cold)); that would take care of all the error branches > everywhere automatically without having to mark things up. Yes, that'd probably be great :) > But I did a quick grep and found nothing using __attribute__((cold)) > yet, so I don't know what the right way of using it would be > (and maybe it's not being used intentionally?). It's probably that just nobody thought about adding it. > I'm thinking it should probably go into misc/sys/cdefs.h as __COLD (or > __attribute_cold?). Something like this: > > #if __glibc_has_attribute (cold) > #define __COLD __attribute__ ((cold)) > #else > #define __COLD > #endif > > What do you think? Yes! Though you can even make it #if __GNUC_PREREQ (4,3) || __glibc_has_attribute (__cold__) Samuel
Re: [PATCH 3/4] hurd: Microoptimize mmap ()
On Mon, Apr 24, 2023 at 11:47 PM Samuel Thibault wrote: > Is it really worth making the code a bit obscure? No, not really. What happened here was I looked at what my mask computation compiled to, to verify it does the right thing on both x86_64 and i686, and then I saw how the error branches are compiled, and next thing you know there are new __glibc_unlikely's in my tree :) What I should rather look into is marking __hurd_fail and friends with __attribute__((cold)); that would take care of all the error branches everywhere automatically without having to mark things up. But I did a quick grep and found nothing using __attribute__((cold)) yet, so I don't know what the right way of using it would be (and maybe it's not being used intentionally?). I'm thinking it should probably go into misc/sys/cdefs.h as __COLD (or __attribute_cold?). Something like this: #if __glibc_has_attribute (cold) #define __COLD __attribute__ ((cold)) #else #define __COLD #endif What do you think? Sergey
Re: [PATCH 3/4] hurd: Microoptimize mmap ()
Hello,
Is it really worth making the code a bit obscure? The mapping RPC will
be way more expensive than branch misprediction...
Samuel
Sergey Bugaev, le lun. 24 avril 2023 00:55:25 +0300, a ecrit:
> Signed-off-by: Sergey Bugaev
> ---
> sysdeps/mach/hurd/mmap.c | 8 +---
> 1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/sysdeps/mach/hurd/mmap.c b/sysdeps/mach/hurd/mmap.c
> index 790eb238..d570be24 100644
> --- a/sysdeps/mach/hurd/mmap.c
> +++ b/sysdeps/mach/hurd/mmap.c
> @@ -42,7 +42,8 @@ __mmap (void *addr, size_t len, int prot, int flags, int
> fd, off_t offset)
>mapaddr = (vm_address_t) addr;
>
>/* ADDR and OFFSET must be page-aligned. */
> - if ((mapaddr & (__vm_page_size - 1)) || (offset & (__vm_page_size - 1)))
> + if (__glibc_unlikely ((mapaddr & (__vm_page_size - 1))
> + || (offset & (__vm_page_size - 1
> return (void *) (long int) __hurd_fail (EINVAL);
>
>vmprot = VM_PROT_NONE;
> @@ -73,7 +74,8 @@ __mmap (void *addr, size_t len, int prot, int flags, int
> fd, off_t offset)
> mach_port_t robj, wobj;
> if (err = HURD_DPORT_USE (fd, __io_map (port, &robj, &wobj)))
> {
> - if (err == MIG_BAD_ID || err == EOPNOTSUPP || err == ENOSYS)
> + if (__glibc_unlikely (err == MIG_BAD_ID || err == EOPNOTSUPP
> + || err == ENOSYS))
> err = ENODEV; /* File descriptor doesn't support mmap. */
> return (void *) (long int) __hurd_dfail (fd, err);
> }
> @@ -173,7 +175,7 @@ __mmap (void *addr, size_t len, int prot, int flags, int
> fd, off_t offset)
>if (err == KERN_PROTECTION_FAILURE)
> err = EACCES;
>
> - if (err)
> + if (__glibc_unlikely (err))
> return (void *) (long int) __hurd_fail (err);
>
>return (void *) mapaddr;
> --
> 2.40.0
