target-arm: emulate LL/SC using cmpxchg helpers
Emulating LL/SC with cmpxchg is not correct, since it can
suffer from the ABA problem. Portable parallel code, however,
is written assuming only cmpxchg--and not LL/SC--is available.
This means that in practice emulating LL/SC with cmpxchg is
a viable alternative.
The appended emulates LL/SC pairs in ARM with cmpxchg helpers.
This works in both user and system mode. In usermode, it avoids
pausing all other CPUs to perform the LL/SC pair. The subsequent
performance and scalability improvement is significant, as the
plots below show. They plot the throughput of atomic_add-bench
compiled for ARM and executed on a 64-core x86 machine.
Hi-res plots: http://imgur.com/a/aNQpB
atomic_add-bench:
1000000 ops/thread, [0,1] range
9 ++---------+----------+----------+----------+----------+----------+---++
+cmpxchg +-E--+ + + + + + |
8 +Emaster +-H--+ ++
| | |
7 ++E ++
| | |
6 ++++ ++
| | |
5 ++ | ++
4 ++ | ++
| | |
3 ++ | ++
| | |
2 ++ | ++
|H++E+--- +++ ---+E+------+E+------+E|
1 +++ +E+-----+E+------+E+------+E+------+E+-- +++ +++ ++
++H+ + +++ + +++ ++++ + + + |
0 ++--H----H-+-----H----+----------+----------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
atomic_add-bench:
1000000 ops/thread, [0,2] range
16 ++---------+----------+---------+----------+----------+----------+---++
+cmpxchg +-E--+ + + + + + |
14 ++master +-H--+ ++
| | |
12 ++| ++
| E |
10 ++| ++
| | |
8 ++++ ++
|E+| |
| | |
6 ++ | ++
| | |
4 ++ | ++
| +E+--- +++ +++ +++ ---+E+------+E|
2 +H+ +E+------E-------+E+-----+E+------+E+------+E+-- +++
+ | + +++ + ++++ + + + |
0 ++H-H----H-+-----H----+---------+----------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
atomic_add-bench:
1000000 ops/thread, [0,128] range
70 ++---------+----------+---------+----------+----------+----------+---++
+cmpxchg +-E--+ + + + ++++ + |
60 ++master +-H--+ ----E------+E+-------++
| -+E+--- +++ +++ +E|
| +++ ---- +++ ++|
50 ++ +++ ---+E+- ++
| -E--- |
40 ++ ---+++ ++
| +++--- |
| -+E+ |
30 ++ +++---- ++
| +E+ |
20 ++ +++-- ++
| +E+ |
|+E+ |
10 +E+ ++
+ + + + + + + |
0 +HH-H----H-+-----H----+---------+----------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
atomic_add-bench:
1000000 ops/thread, [0,1024] range
120 ++---------+---------+----------+---------+----------+----------+---++
+cmpxchg +-E--+ + + + + + |
| master +-H--+ ++|
100 ++ ----E+
| +++ ---+E+--- ++|
| --E--- +++ |
80 ++ ---- +++ ++
| ---+E+- |
60 ++ -+E+-- ++
| +++ ---- +++ |
| -+E+- |
40 ++ +++---- ++
| +++ ---+E+ |
| -+E+--- |
20 ++ +E+ ++
|+E+++ |
+E+ + + + + + + |
0 +HH-H---H--+-----H---+----------+---------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
[rth: Enforce alignment for ldrexd.]
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Message-Id: <
1467054136-10430-23-git-send-email-cota@braap.org>
Signed-off-by: Richard Henderson <rth@twiddle.net>