This is the mail archive of the 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: Patch: PR testsuite/21865 (really a middle-end vector bug)

On Tue, Jun 28, 2005 at 10:14:22AM -0700, Steve Ellcey wrote:
> Maybe Richard will offer an opinion since both the code I am changing
> and the testcase were added by him.

While I conceptually agree that there ought to be some upper bound
to the alignment that we attempt to get for vectors, most hardware
vendors tend to come out with brand new vector units that require
more alignment than was ever required before.  Thus BIGGEST_ALIGNMENT
tends to be lower than is really needed.

To compound this, BIGGEST_ALIGNMENT has effects beyond merely being
an upper bound of what alignment a target requires.  Thus is cannot
be changed without breaking an ABI.

Thus I'm unwilling to lower this alignment unless you can promise
me that no vendor will ever introduce a new even wider vector type
on an existing architecture.

Or convince me I'm totally out to lunch with the issues here.

Personally I think you should simply xfail the one test and be done
with it.


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