Messages in this thread | | | From | Vineet Gupta <> | Subject | [PATCH 0/3] atomic_or() related changes | Date | Thu, 9 Jul 2015 13:43:16 +0530 |
| |
Hi,
This started off as an effort to convert a cmpxchg based loop in arc/kernel/smp.c to an API which is more LLOCK/SCOND friendly.
e.g. do { new = old = ACCESS_ONCE(*ipi_data_ptr); new |= 1U << msg; } while (cmpxchg(ipi_data_ptr, old, new) != old);
The generated code is horrible. There are 2 useless branches here and a LD/LLOCK to same address all inside a loop.
8015cefc: ld_s r2,[r3,0] 8015cefe: or r5,r2,r1 8015cf02: llock r4,[r3] 8015cf06: brne r4,r2,8015cf12 8015cf0a: scond r5,[r3] 8015cf0e: bnz 8015cf02 8015cf12: brne r2,r4,8015cefc
An atomic_or() kind of API is better suited to generate something like below
8015cf02: llock r4,[r3] 8015cf06: or r5,r2,r1 8015cf0a: scond r5,[r3] 8015cf0e: bnz 8015cf02
Although this doesn't work for the specific instance I wanted to fix as ipi_data_ptr is not atomic_t, I did run into a few things which could be improved, hence this series.
Compile tested on ARC, ARM, x86.
I do have some concern about mixing long and int on 64 bit arch, which I've captured inline in patch 2/3. It is most likely a lack of understand on my part, but worth asking..
Thx, -Vineet
Vineet Gupta (3): asm-generic/atomic.h: ARCH_HAS_ATOMIC_OR -> CONFIG_ARCH_HAS_ATOMIC_OR brcmfmac: dhd_sdio.c: use existing atomic_or primitive ARC: provide atomic_or() and define ARCH_HAS_ATOMIC_OR
arch/arc/include/asm/atomic.h | 9 +++++++++ drivers/net/wireless/brcm80211/brcmfmac/sdio.c | 13 ++----------- include/asm-generic/atomic.h | 2 +- include/linux/atomic.h | 4 ++-- 4 files changed, 14 insertions(+), 14 deletions(-)
-- 1.9.1
| |