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: ia64-hpux fix compat/struct-by-value-5

Jim Wilson <> writes:

> On Tue, 2004-02-17 at 00:16, Zack Weinberg wrote:
>> ia64-hpux bootstrap in progress; comments?
> I forgot to comment on the patch.  I think the BYTES_BIG_ENDIAN checks
> are unnecessary but harmless.  If you always create the DImode parallel,
> then we get the right result for both big and little endian.  So the
> checks can either be there or not be there, your choice.  I think we get
> slightly more efficient code for little-endian if the BYTES_BIG_ENDIAN
> checks are there, so I don't mind them being there.

I did it this way mainly because it's done that way in the block
above, but if we generate more efficient little-endian code this
way then I vote to keep it.

> Isn't this going to break handling of regular SFmode arguments?  I think
> you only want to do this when hfa_mode is SFmode.
> I think the second part of the patch is unnecessary.  The special hfa
> code at the top handles all cases except prototype && !named, so only
> the first part of your patch is doing something useful.  The second part
> is unreachable code.

According to the "Itanium(TM) Software Conventions and Runtime
Architecture Guide", bare SFmode [in fact, all floating point]
quantities are passed in general registers when they're varargs
parameters.  At which point the same alignment rule applies to
them.  So I think that the patch is correct as is, and that both
chunks are necessary (the second part handles the case where an
unprototyped function is called with float arguments).


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