This is the mail archive of the mailing list for the Cygwin project.

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

Re: 8 bit alignnment

Yes, the performance of cygwin-compiled code using 64-bit data (C double, Fortran double precision,..) is abysmal with the standard 32-bit or 4 byte alignment.  I have been running for some time with binutils rebuilt for 64-bit alignment.  The results with gcc and g77 are excellent (doesn't even interfere with rebuilding gcc), but it breaks important parts of the existing g++ and objc run-time support binaries.  I suppose that I have introduced struct/class padding mis-match.  I'm hoping to try rebuilding the libraries to see if that will cure this.

The .p2align code and long double data alignments used by the gcc aren't fully effective without 128-bit alignment, which doesn't appear to be possible with cygwin.  Going from 32-bit to 64-bit alignment does, I believe, greatly reduce the number of incidents where .p2align hurts rather than helping.

With the change to 64-bit alignment, cygwin/gcc out-performs some big name Windows C++ and linux gcc compilers on certain tests of the effect of alignment on memory bandwidth. And the big name Windows compilers cheat a bit with their 8-byte long doubles.

Yes, binutils, combined with the cygwin run-time, keep the stack aligned to the selected boundaries.  The stack is supposed to be 128-bit aligned in linux with the current glibc.  In principle, you could make a function which would set the stack up to your selected alignment each time you call it.  Not a good practice in general.

This alignment thing has been around at least 30 years, and it's not going away, with the greater use of 64-bit architectures.

McNulty Junior Bobby <> wrote:
> Have you guys thought of compiling the code with 8 bit
alignments? This will bring Cygwin to 64 bits.
I know. I did this with Csound, and it is acting
I read how to optimize the programs that I use and the
alignment was one of them.
Actually, i think it was the stack boundary.

Want to unsubscribe from this list?
Send a message to

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