Fix computation of cache-line size on Power #6515
Conversation
keithc-ca
force-pushed
the
ppc_cache_line_size
branch
from
d538b0a
to
df47457
Compare
port/unix/omrcpu.c
Outdated
| :"r"((void *) &buf[512])); | ||
| memset(buf, 255, sizeof(buf)); | ||
|
|
||
| __asm__ __volatile__( |
There was a problem hiding this comment.
I suspect the __volatile__ isn't needed. The port library copy doesn't use it.
There was a problem hiding this comment.
Ah nm, I see you just added it here. I guess it doesn't hurt.
There was a problem hiding this comment.
It was used in both places in j9memclr.cpp and won't hurt. In fact, it may save us from a future compiler optimization.
|
jenkins build all |
|
The changes LGTM, but can the cache line size detection be factored out into a single place? |
I'll look into whether there's a clear path to that. |
keithc-ca
force-pushed
the
ppc_cache_line_size
branch
from
df47457
to
18814e6
Compare
|
Updated as suggested, reusing |
|
The description of the functional fix (i.e. the memory clobber) seems to have gone missing from the commit message. It's probably more important now than it was before, since there are more changes |
The compiler needs to be told that our uses of dcbz clobber memory, otherwise the optimizer can reasonably assume in getCacheLineSize(), that each element of buf still has the value (255) assigned by memset(). * reuse getCacheLineSize() in omrcpu_startup() * use uint32_t consistently for cache line size Signed-off-by: Keith W. Campbell <[email protected]>
keithc-ca
force-pushed
the
ppc_cache_line_size
branch
from
18814e6
to
8b3bf5b
Compare
|
That description was never in the commit message itself (only in the description here); it's now also in the commit message. |
|
Jenkins build all |
This allows the compiler to do the widening operation once, rather than repeatedly in loops, fixing a performance regression introduced in eclipse#6515. Signed-off-by: Keith W. Campbell <[email protected]>
This allows the compiler to do the widening operation once, rather than repeatedly in loops, fixing a performance regression introduced in eclipse#6515. Signed-off-by: Keith W. Campbell <[email protected]>
We need to inform the compiler that our uses of
dcbzclobber memory, otherwise the optimizer can reasonably assume, e.g. inomrcpu_startup(),thatbuf[0] == 255at the beginning of the for loop.