On 03/03/16 09:42, Ramana Radhakrishnan wrote:
On 03/03/16 09:35, Kyrill Tkachov wrote:
Hi all,
In this PR shrink-wrapping ends up duplicating the load-exclusive part of a
load-exclusive/store-exclusive loop
used to implement an atomic compare exchange operation. Look in bugzilla for
the kind of sequences generated.
The load-exclusive paths are duplicated, but not the store-exclusive ones, with
some very counter-intuitive
control flow between them.
The suggested solution is to catch the load exclusive instructions in
TARGET_CANNOT_COPY_INSN_P so that
the RTL optimisers don't do any funny tricks with them. Since these
instructions only really appear when
implementing atomic operations it's unlikely that the user will want
particularly fancy transformations involving
duplication of exclusive access sequences anyway (though examples to the
contrary would certainly be appreciated)
and this fix seems like the safest at this point. This is a regression fix from
GCC 5 as shrink-wrapping got more
aggressive for GCC 6.
With this patch the sequence is not shrink-wrapped and is the same as the
sequence for GCC 5 as shown in the PR.
Bootstrapped and tested on arm-none-linux-gnueabihf.
Ok for trunk?
While reading the patch again I wondered if we should also catch store
exclusives here but the cases we are interested in preventing are copies of
blocks containing ldrex instructions which almost always are paired with an
strex instruction i.e. I don't think an strex instruction can exist without a
corresponding ldrex instruction. Thus your patch looks sufficient.
I had considered that too and came to the same conclusion.
It would also be good to audit the aarch64 backend for any such sequences but
that need not hold up this patch.
Thus, Ok for trunk.
Thanks, I'll commit shortly.
Kyrill
Thanks,
Ramana
Thanks,
Kyrill
2016-03-03 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
PR rtl-optimization/69904
* config/arm/arm.c (arm_cannot_copy_insn_p):
Return true for load-exclusive instructions.
2016-03-03 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
PR rtl-optimization/69904
* gcc.target/arm/pr69904.c: New test.
