https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
Andrew Pinski changed:
What|Removed |Added
Resolution|--- |WONTFIX
Status|UNCONFIRMED
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #9 from Andi Kleen 2013-01-14
22:34:16 UTC ---
Also i need to look more closely, but most likely the C++ atomic code should be
changed to avoid this situation. This would give much better code on x86 in any
case even without el
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #8 from Andi Kleen 2013-01-14
22:32:06 UTC ---
forbid = give warning and drop bit
It's a hint, but in a good implementation it should not be silently dropped or
code generated that has no chance to elide. It's a quality of im
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
Andrew Macleod changed:
What|Removed |Added
CC||amacleod at redhat dot com
---
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #6 from Andi Kleen 2013-01-14
22:05:38 UTC ---
Hmm that's true. x86 doesn't have xand, x_or, x_xor, only xadd
Maybe cmpxchg is the only way?
For some special cases it can be done (like and single bit-> btr, or single bit
->
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #5 from Uros Bizjak 2013-01-14 21:25:13
UTC ---
Following testcase will expand to a cmpxchg loop:
int hle_and (int *p, int v)
{
return __atomic_fetch_and_4 (p, v, __ATOMIC_ACQUIRE | __ATOMIC_HLE_ACQUIRE);
}
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #4 from Uros Bizjak 2013-01-14 21:22:35
UTC ---
The problem is, that in failed case maybe_emit_op() gets target register to
return the result to, so with after=false, it expands via
optab->mem_fetch_before.
Unfortunately, at
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
Andi Kleen changed:
What|Removed |Added
Attachment #29163|0 |1
is obsolete|
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #2 from Andi Kleen 2013-01-14
19:52:03 UTC ---
Created attachment 29163
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29163
preprocessed testcase
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55966
--- Comment #1 from Andi Kleen 2013-01-14
19:06:02 UTC ---
Here's a test case. This requires the libstdc++ HLE patch from
http://gcc.gnu.org/ml/gcc-patches/2013-01/msg00673.html
g++ -std=gnu++0x
#include
#define ACQ memory_order_a
10 matches
Mail list logo