x86/ACPI/boot: Improve __acpi_acquire_global_lock
PATCH v2 1/1 x86/ACPI/boot: Improve __acpi_acquire_global_lock
Improve __acpi_acquire_global_lock by using a temporary variable. This enables compiler to perform if-conversion and improves generated code from:
...
72a: d1 ea shr %edx 72c: 83 e1 fc and $0xfffffffc,%ecx 72f: 83 e2 01 and $0x1,%edx 732: 09 ca or %ecx,%edx 734: 83 c2 02 add $0x2,%edx 737: f0 0f b1 17 lock cmpxchg %edx,(%rdi) 73b: 75 e9 jne 726 <__acpi_acquire_global_lock+0x6> 73d: 83 e2 03 and $0x3,%edx 740: 31 c0 xor %eax,%eax 742: 83 fa 03 cmp $0x3,%edx 745: 0f 95 c0 setne %al 748: f7 d8 neg %eax
to:
...
72a: d1 e9 shr %ecx 72c: 83 e2 fc and $0xfffffffc,%edx 72f: 83 e1 01 and $0x1,%ecx 732: 09 ca or %ecx,%edx 734: 83 c2 02 add $0x2,%edx 737: f0 0f b1 17 lock cmpxchg %edx,(%rdi) 73b: 75 e9 jne 726 <__acpi_acquire_global_lock+0x6> 73d: 8d 41 ff lea -0x1(%rcx),%eax
BTW: the compiler could generate:
lea 0x2(%rcx,%rdx,1),%edx
instead of:
or %ecx,%edx add $0x2,%edx
but unwated conversion from add to or when bits are known to be zero prevents this improvement. This is GCC PR108477.
No functional change intended.
---
v2: Expand return statement.
--- arch/x86/kernel/acpi/boot.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)
arch/x86/kernel/acpi/boot.c +8 -3
@@ -1853,13 +1853,18 @@ early_param("acpi_sci", setup_acpi_sci);int __acpi_acquire_global_lock (unsigned int *lock) { -unsigned int old, new; +unsigned int old, new, val; old =READ_ONCE (*lock);do { - new = (((old & ~0x3 ) +2 ) + ((old >>1 ) &0x1 )); + val = (old >>1 ) &0x1 ; + new = (old & ~0x3 ) +2 + val; }while (!try_cmpxchg (lock, &old, new)); -return ((new &0x3 ) <3 ) ?-1 :0 ; + +if (val) +return 0 ; + +return -1 ; }int __acpi_release_global_lock (unsigned int *lock)
Signature · 1 line
2.39.2
Quoted 53 lines · Uros Bizjak
Improve
__acpi_acquire_global_lockby using a temporary variable. This enables compiler to perform if-conversion and improves generated code from:...
72a: d1 ea shr %edx 72c: 83 e1 fc and $0xfffffffc,%ecx 72f: 83 e2 01 and $0x1,%edx 732: 09 ca or %ecx,%edx 734: 83 c2 02 add $0x2,%edx 737: f0 0f b1 17 lock cmpxchg %edx,(%rdi) 73b: 75 e9 jne 726 <__acpi_acquire_global_lock+0x6> 73d: 83 e2 03 and $0x3,%edx 740: 31 c0 xor %eax,%eax 742: 83 fa 03 cmp $0x3,%edx 745: 0f 95 c0 setne %al 748: f7 d8 neg %eaxto:
...
72a: d1 e9 shr %ecx 72c: 83 e2 fc and $0xfffffffc,%edx 72f: 83 e1 01 and $0x1,%ecx 732: 09 ca or %ecx,%edx 734: 83 c2 02 add $0x2,%edx 737: f0 0f b1 17 lock cmpxchg %edx,(%rdi) 73b: 75 e9 jne 726 <__acpi_acquire_global_lock+0x6> 73d: 8d 41 ff lea -0x1(%rcx),%eaxBTW: the compiler could generate:
lea 0x2(%rcx,%rdx,1),%edxinstead of:
or %ecx,%edx add $0x2,%edxbut unwated conversion from add to or when bits are known to be zero prevents this improvement. This is GCC PR108477.
No functional change intended.
Signed-off-by Uros Bizjak <[email protected]>Cc "Rafael J. Wysocki" <[email protected]>Cc Len Brown <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc Borislav Petkov <[email protected]>Cc Dave Hansen <[email protected]>Cc "H. Peter Anvin" <[email protected]>
Quoted 2 lines
---
v2: Expand return statement.
or please let me know if you want me to pick this up (in which case it will require an ACK from one of the x86 maintainers).
Quoted 33 lines
--- arch/x86/kernel/acpi/boot.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)arch/x86/kernel/acpi/boot.c +8
-4@@ -1853,13 +1853,18 @@ early_param("acpi_sci", setup_acpi_sci);int __acpi_acquire_global_lock (unsigned int *lock) { -unsigned int old, new; +unsigned int old, new, val; old =READ_ONCE (*lock);do { - new = (((old & ~0x3 ) +2 ) + ((old >>1 ) &0x1 )); + val = (old >>1 ) &0x1 ; + new = (old & ~0x3 ) +2 + val; }while (!try_cmpxchg (lock, &old, new)); -return ((new &0x3 ) <3 ) ?-1 :0 ; + +if (val) +return 0 ; + +return -1 ; }int __acpi_release_global_lock (unsigned int *lock) --2.39.2
Quoted 4 lines · Rafael J. Wysocki
or please let me know if you want me to pick this up (in which case it will require an ACK from one of the x86 maintainers).
I'll pull it into x86/acpi. I'm kinda shocked the compiler is so clueless, but this makes the C code more readable anyway. Win/win, I guess.
Quoted 9 lines · Dave Hansen
Quoted 4 lines · Rafael J. Wysocki
or please let me know if you want me to pick this up (in which case it will require an ACK from one of the x86 maintainers).
I'll pull it into x86/acpi. I'm kinda shocked the compiler is so clueless, but this makes the C code more readable anyway. Win/win, I guess.
Please note that the return form __acpi_{acquire,release}_global_lock is actually used as bool:
acenv.h:
int __acpi_acquire_global_lock(unsigned int *lock); int __acpi_release_global_lock(unsigned int *lock);
#define ACPI_ACQUIRE_GLOBAL_LOCK(facs, Acq) \
((Acq) = __acpi_acquire_global_lock(&facs->global_lock))
#define ACPI_RELEASE_GLOBAL_LOCK(facs, Acq) \
((Acq) = __acpi_release_global_lock(&facs->global_lock))
evglock.c:
acpi_status acpi_ev_acquire_global_lock(u16 timeout)
{
...
u8 acquired = FALSE;
...
ACPI_ACQUIRE_GLOBAL_LOCK(acpi_gbl_FACS, acquired);
if (acquired) ...
}acpi_status acpi_ev_release_global_lock(void)
{
u8 pending = FALSE;
... ACPI_RELEASE_GLOBAL_LOCK(acpi_gbl_FACS, pending); if (pending) ...
}
These functions are also defined for ia64, so I didn't want to change the return value. But ia64 is going to be retired, and this opens the optimization opportunity.
Uros.