linux: Allow avoiding va_list for generic syscall and use it for LoongArch
Checks
Context |
Check |
Description |
dj/TryBot-apply_patch |
success
|
Patch applied to master at the time it was sent
|
dj/TryBot-32bit |
success
|
Build for i686
|
Commit Message
Currently GCC generates highly sub-optimal code on architectures where
the calling convention prefers registers for arugment passing. This is
GCC PR100955. While it's technically a missed-optimization in GCC, it
seems not trivial to fix (I've not seen any compiler which can optimize
this properly yet).
As the generic Linux syscall actually uses a fixed number of arguments,
we can avoid va_list if possible and make the compiler do right thing.
Add a macro __ASSUME_SYSCALL_NAMED_WORKS which should be defined if:
* The generic Linux syscall implementation will be used.
* The calling convention has at least 7 GARs.
* The kernel ABI uses registers for syscall numbers and arguments.
LoongArch is benefited from this (saving about 430 CPU cycles per
syscall, though I won't call it a significant improvement because
syscall is "slow" in nature). And in the future we may switch more
ports to use the generic syscall without a performance regression,
reducing the number of target-specific syscall.{c,S} files we need to
maintain.
---
.../sysv/linux/loongarch/kernel-features.h | 25 +++++++++++++
sysdeps/unix/sysv/linux/syscall.c | 35 ++++++++++++++-----
2 files changed, 52 insertions(+), 8 deletions(-)
create mode 100644 sysdeps/unix/sysv/linux/loongarch/kernel-features.h
Comments
* Xi Ruoyao via Libc-alpha:
> Currently GCC generates highly sub-optimal code on architectures where
> the calling convention prefers registers for arugment passing. This is
Typo: ar[gu]ment
> LoongArch is benefited from this (saving about 430 CPU cycles per
> syscall, though I won't call it a significant improvement because
> syscall is "slow" in nature). And in the future we may switch more
> ports to use the generic syscall without a performance regression,
> reducing the number of target-specific syscall.{c,S} files we need to
> maintain.
Does this impact the open* and fcntl* wrappers as well?
On Fri, 2023-03-24 at 10:47 +0100, Florian Weimer wrote:
> * Xi Ruoyao via Libc-alpha:
>
> > Currently GCC generates highly sub-optimal code on architectures
> > where
> > the calling convention prefers registers for arugment passing. This
> > is
>
> Typo: ar[gu]ment
>
> > LoongArch is benefited from this (saving about 430 CPU cycles per
> > syscall, though I won't call it a significant improvement because
> > syscall is "slow" in nature). And in the future we may switch more
> > ports to use the generic syscall without a performance regression,
> > reducing the number of target-specific syscall.{c,S} files we need
> > to
> > maintain.
>
> Does this impact the open* and fcntl* wrappers as well?
Yes, on LoongArch all GARs are saved :(. I'll see if it's possible to
avoid using va_list for them too.
* Xi Ruoyao:
> On Fri, 2023-03-24 at 10:47 +0100, Florian Weimer wrote:
>> * Xi Ruoyao via Libc-alpha:
>>
>> > Currently GCC generates highly sub-optimal code on architectures
>> > where
>> > the calling convention prefers registers for arugment passing. This
>> > is
>>
>> Typo: ar[gu]ment
>>
>> > LoongArch is benefited from this (saving about 430 CPU cycles per
>> > syscall, though I won't call it a significant improvement because
>> > syscall is "slow" in nature). And in the future we may switch more
>> > ports to use the generic syscall without a performance regression,
>> > reducing the number of target-specific syscall.{c,S} files we need
>> > to
>> > maintain.
>>
>> Does this impact the open* and fcntl* wrappers as well?
>
> Yes, on LoongArch all GARs are saved :(. I'll see if it's possible to
> avoid using va_list for them too.
If we had overread the argument list in open*, that certainly would
have avoided all the grief we had with O_TMPFILE (where the original
glibc implementation did not pass the mode argument to the kernel). I
think fcntl* already overreads the argument list, so it should be
safe.
new file mode 100644
@@ -0,0 +1,25 @@
+/* Set flags signalling availability of kernel features based on given
+ kernel version number.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include_next <kernel-features.h>
+
+/* Define this if the calling convention for passing 7 named arguments is
+ same as passing one named argument and 6 variable arguments, and the
+ kernel ABI uses registers for syscall number and arguments. */
+#define __ASSUME_SYSCALL_NAMED_WORKS 1
@@ -16,9 +16,33 @@
License along with the GNU C Library. If not, see
<https://www.gnu.org/licenses/>. */
-#include <stdarg.h>
#include <sysdep.h>
+#ifndef __ASSUME_SYSCALL_NAMED_WORKS
+#include <stdarg.h>
+#endif
+
+static inline long int
+__syscall (long int number, long int a0, long int a1, long int a2, long int a3,
+ long int a4, long int a5)
+{
+ long int r = INTERNAL_SYSCALL_NCS_CALL (number, a0, a1, a2, a3, a4, a5);
+ if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (r)))
+ {
+ __set_errno (-r);
+ return -1;
+ }
+ return r;
+}
+
+#ifdef __ASSUME_SYSCALL_NAMED_WORKS
+long int
+syscall (long int number, long int a0, long int a1, long int a2, long int a3,
+ long int a4, long int a5)
+{
+ return __syscall (number, a0, a1, a2, a3, a4, a5);
+}
+#else
long int
syscall (long int number, ...)
{
@@ -33,11 +57,6 @@ syscall (long int number, ...)
long int a5 = va_arg (args, long int);
va_end (args);
- long int r = INTERNAL_SYSCALL_NCS_CALL (number, a0, a1, a2, a3, a4, a5);
- if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (r)))
- {
- __set_errno (-r);
- return -1;
- }
- return r;
+ return __syscall (number, a0, a1, a2, a3, a4, a5);
}
+#endif