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: Fortran/PR19303 PATCH: Runtime selection of record markers forunformatted sequential io

François-Xavier Coudert wrote:
> I'm somehow deep into things above my understanding here, but I may have a 
> suggestion. What happens if you write a 96 bits header, even if the size 
> would fit in 32 bits? That is, you always write -1 in the first 32 bits, and 
> then the full size in the next 64. I know this isn't how HP does it, but can 
> it understand it this way? That would remove the need for buffering and the 
> performance issue, wouldn't it?

That would probably work, but it would break backwards compatibility with 32
bit record markers which is the reason why this weird format exists in the
first place.  I don't think we should care about performance so much as to
introduce probably incompatible new record marker formats.  If someone wants
fast HP record markers for > 2 Gig records, they can enhance the library
themselves afterall.  It should certainly be no requirement for inclusion of
Janne's patch.

Also, if HP's compiler writes to a buffer first I don't see how it can be much
faster than an implementation which shifts the record in the worst case --
writing several gigabytes will still involve a disk-to-disk operation, because
the buffer won't fit into memory.

- Tobi

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