On 2/9/24 03:26, Richard Biener wrote:
The following allows a base term to be derived from an existing
MEM_EXPR, notably the points-to set of a MEM_REF base. For the
testcase in the PR this helps RTL DSE elide stores to a stack
temporary. This covers pointers to NONLOCAL which can be mapped
to arg_base_value, helping to disambiguate against other special
bases (ADDRESS) as well as PARM_DECL accesses.
I like it and as you note later, it's extendable.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
This is an attempt to recover some of the losses from dumbing down
find_base_{term,value}. I did give my ideas how to properly do
this during stage1 a start, I will post a short incomplete RFC series
later today.
I saw those, but set them aside for gcc-15.
OK for trunk?
I've included all languages in testing and also tested with -m32 but
details of RTL alias analysis might escape me ...
Thanks,
Richard.
PR rtl-optimization/113597
* alias.cc (find_base_term): Add argument for the whole mem
and derive a base term from its MEM_EXPR.
(true_dependence_1): Pass down the MEMs to find_base_term.
(write_dependence_p): Likewise.
(may_alias_p): Likewise.
I'd lean ever so slightly against including this. Not because I see
anything wrong, more so because we don't have a lot of time for this to
shake out if there are any problems. But I wouldn't go as far as to say
I object to including it.
So OK for the trunk if you want to go forward now. Or defer if you want
to take the somewhat safer route of waiting to gcc-15 to tackle this.
jeff
