https://github.com/torvalds/linux
Revision e658a6f14d7c0243205f035979d0ecf6c12a036f authored by Chris Metcalf on 16 November 2016, 16:18:05 UTC, committed by Chris Metcalf on 23 November 2016, 20:28:54 UTC
For large values of "mult" and long uptimes, the intermediate result of "cycles * mult" can overflow 64 bits. For example, the tile platform calls clocksource_cyc2ns with a 1.2 GHz clock; we have mult = 853, and after 208.5 days, we overflow 64 bits. Since clocksource_cyc2ns() is intended to be used for relative cycle counts, not absolute cycle counts, performance is more importance than accepting a wider range of cycle values. So, just use mult_frac() directly in tile's sched_clock(). Commit 4cecf6d401a0 ("sched, x86: Avoid unnecessary overflow in sched_clock") by Salman Qazi results in essentially the same generated code for x86 as this change does for tile. In fact, a follow-on change by Salman introduced mult_frac() and switched to using it, so the C code was largely identical at that point too. Peter Zijlstra then added mul_u64_u32_shr() and switched x86 to use it. This is, in principle, better; by optimizing the 64x64->64 multiplies to be 32x32->64 multiplies we can potentially save some time. However, the compiler piplines the 64x64->64 multiplies pretty well, and the conditional branch in the generic mul_u64_u32_shr() causes some bubbles in execution, with the result that it's pretty much a wash. If tilegx provided its own implementation of mul_u64_u32_shr() without the conditional branch, we could potentially save 3 cycles, but that seems like small gain for a fair amount of additional build scaffolding; no other platform currently provides a mul_u64_u32_shr() override, and tile doesn't currently have an <asm/div64.h> header to put the override in. Additionally, gcc currently has an optimization bug that prevents it from recognizing the opportunity to use a 32x32->64 multiply, and so the result would be no better than the existing mult_frac() until such time as the compiler is fixed. For now, just using mult_frac() seems like the right answer. Cc: stable@kernel.org [v3.4+] Signed-off-by: Chris Metcalf <cmetcalf@mellanox.com>
1 parent ded9b5d
Tip revision: e658a6f14d7c0243205f035979d0ecf6c12a036f authored by Chris Metcalf on 16 November 2016, 16:18:05 UTC
tile: avoid using clocksource_cyc2ns with absolute cycle count
tile: avoid using clocksource_cyc2ns with absolute cycle count
Tip revision: e658a6f
File | Mode | Size |
---|---|---|
Makefile | -rw-r--r-- | 377 bytes |
compat.c | -rw-r--r-- | 19.1 KB |
compat_mq.c | -rw-r--r-- | 3.9 KB |
ipc_sysctl.c | -rw-r--r-- | 5.4 KB |
mq_sysctl.c | -rw-r--r-- | 2.9 KB |
mqueue.c | -rw-r--r-- | 35.7 KB |
msg.c | -rw-r--r-- | 23.3 KB |
msgutil.c | -rw-r--r-- | 3.6 KB |
namespace.c | -rw-r--r-- | 4.5 KB |
sem.c | -rw-r--r-- | 56.3 KB |
shm.c | -rw-r--r-- | 33.3 KB |
syscall.c | -rw-r--r-- | 2.3 KB |
util.c | -rw-r--r-- | 21.1 KB |
util.h | -rw-r--r-- | 6.4 KB |
Computing file changes ...