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: SSA usage question


    Whatever.  Then perhaps you want to a BIT_FIELD_EXPR with
    variable parameters.

    What we will NOT do is have implicit variable references.
    Of any kind.  This is not open to discussion.

Of course it's open for discussion!  It, however, should be been discussed
before the merge.

It's simply not an acceptable approach to throw away all typing, aliasing and
array structure implementation just because there's a variable size, lower
bound, or position involved.

Doing that will cause the generation of far *worse* code than when we were
able to propagate this information to RTL.  Losing all the array stucture
information will destroy all the loop optimization possibilities and it's
exactly the variable cases that will benefit the most from loop optimizations.

I must say I'm quite disturbed to find out all this at this late stage.
Properly dealing with variable position and size objects is a major part of
the compiler and one of the trickiest areas.  It's always easy to do work
when you can ignore all the hard cases!

I don't share your pessimism here and still don't understand what the
problem with adding dependencies on the needed variables are.  But if you're
right that it can't be done, then I would suggest that the merge was
premature since it removed support for a large part of the tree structure.


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