This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: About the is_gimple_min_invariant predicate


On 7/4/07, Diego Novillo <dnovillo@google.com> wrote:
On 7/4/07 6:24 PM, Eric Botcazou wrote:
>> The problem is that in GIMPLE we only allow TREE_INVARIANT as a gimple
>> value for ADDR_EXPRs.  We still require a TREE_CONSTANT as an array
>> index.  So,  ARRAY[TREE_INVARIANT] is not valid GIMPLE.
>
> OK, my understanding is that is_gimple_min_invariant as currently written is
> correct as "non-recursive" predicate (for the updated grammar), i.e. during
> gimplification.
>
>> FRE should refuse to substitute a TREE_INVARIANT expression as an
>> ARRAY_REF index.
>
> The problem is that FRE (and other optimization passes) uses it as an absolute
> predicate.  This mostly works because it essentially encompasses terminals of
> the grammar, but fails for the non-terminal case.
>
> How would you fix that?  By making it recursive, e.g.:
>
> or by uncoupling the 2 functions, i.e creating a recursive predicate for use
> outside of gimplification?

Passes ought to distinguish GIMPLE values from GIMPLE invariants
according to context.  As this test case shows, a GIMPLE invariant is
not always the right value to replace as it cannot be used as a valid
ARRAY_REF index.

My proposal is that we add a predicate is_gimple_const() and
is_gimple_invariant() such that is_gimple_invariant() is implemented as:

bool
is_gimple_invariant (tree expr)
{
   if (is_gimple_const (expr))
      return true;
   else
      return TREE_INVARIANT (expr);
}

And is_gimple_const() is the current is_gimple_min_invariant() without
the ARRAY_REF case.

Okay, then i would think we should change things like FRE/PRE simply use is_gimple_const, since ...


So my proposal is that PRE/FRE and the value propagators need to make sure that they propagate is_gimple_invariant() except inside ARRAY_REF indexes.

... this is somewhat ugly to do for *PRE* (for FRE, it's not really) :)



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]