On 13/08/2014 16:07, Roman Gareev wrote:
If I’m not mistaken all tree nodes, which have pointer type, can be
divided into two groups: the type is a is a pointer to data member
(TYPE_PTRMEM_P is true for it), the type is a pointer type, and the
pointee is not a data member (TYPE_PTR_P is true for it). I think that
we’re interested in disabling of the second group handling. It can
also be divided into two groups: the type is a pointer to function
type (TYPE_PTRFN_P is true for it), the type is a pointer to object
type (TYPE_PTROB_P is true for it). In my opinion, the second one is
worth to be considered. It contains, for example, nop_expr (these
nodes are used to represent conversions that do not require any
code-generation) integer_cst (these nodes represent integer
constants), ssa_name. I haven’t found all types, which are contained
in it. However, I think that we could disable other types, if it is
necessary in the future.
What we should do later is to build a run-time check that ensures
no pointer overflow is happening and then just create code without it.
I think that it is better to do this when the pointer handling is completed.
I’ve attached a Change_Log, which corresponds to the previous patch.
Are they fine for trunk? Could we give a headsup on the GCC mailing
list and ask other people to try the new isl support in case this
patch is committed?
Two minor thinks:
- I assume you verified this passes all graphite tests.
- Please add a brief comment in the source code regarding why we
do not want to detect such SCoPs.
Otherwise LGTM.
Cheers,
Tobias
