On 30/07/2022 20:57, Jeff Law via Gcc-patches wrote:
On 7/29/2022 7:52 AM, Richard Earnshaw via Gcc-patches wrote:
A SET operation that writes memory may have the same value as an
earlier store but if the alias sets of the new and earlier store do
not conflict then the set is not truly redundant. This can happen,
for example, if objects of different types share a stack slot.
To fix this we define a new function in cselib that first checks for
equality and if that is successful then finds the earlier store in the
value history and checks the alias sets.
The routine is used in two places elsewhere in the compiler. Firstly
in cfgcleanup and secondly in postreload.
gcc/ChangeLog:
* alias.h (mems_same_for_tbaa_p): Declare.
* alias.cc (mems_same_for_tbaa_p): New function.
* dse.cc (record_store): Use it instead of open-coding
alias check.
* cselib.h (cselib_redundant_set_p): Declare.
* cselib.cc: Include alias.h
(cselib_redundant_set_p): New function.
* cfgcleanup.cc: (mark_effect): Use cselib_redundant_set_p instead
of rtx_equal_for_cselib_p.
* postreload.c (reload_cse_simplify): Use cselib_redundant_set_p.
(reload_cse_noop_set_p): Delete.
Seems quite reasonable. The only question I would have would be
whether or not you considered including the aliasing info into the
hashing used by cselib. You'd probably still need the bulk of this
patch as well since we could presumably still get a hash conflict with
two stores of the same value to the same location, but with different
alias sets (it's just much less likely), so perhaps it doesn't really
buy us anything.
I thought about this, but if the alias set were included in the hash,
then surely you'd get every alias set in a different value. Then you'd
miss the cases where the alias sets do conflict even though they are not
the same. Anyway, the values /are/ the same so in some circumstances
you might want to know that.
Ideally this would include a testcase. You might be able to turn that
non-executawble reduced case into something useful by scanning the
post-reload dumps.
I considered this as well, but the testcase I have is far too fragile, I
think. The existing test only fails on Arm, only fails on 11.2 (not
11.3 or gcc-12 onwards), relies on two objects with the same value being
in distinct alias sets but still assigned to the same stack slot and for
some copy dance to end up trying to write back the original value to the
same slot but with a non-conflicting set. And finally, the scheduler
has to then try to move a load past the non-aliasing store.
To get anywhere close to this I think we'd need something akin to the
gimple reader but for RTL so that we could set up all the conditions for
the failure without the risk of an earlier transform blowing the test
away.
I even considered whether we could start with a gimple dump and
bypassing all the tree/gimple transformations, but even that would be
still at the mercy of the stack-slot allocation algorithm.
Jeff
R.
