Re: Re: mips n64 porting review

[Rich Felker <dalias@libc.org>]

On Thu, Feb 04, 2016 at 06:52:47PM -0500, Rich Felker wrote:
> On Wed, Feb 03, 2016 at 06:36:57PM -0500, Rich Felker wrote:
> > On Wed, Feb 03, 2016 at 03:41:13PM +0000, Mahesh Bodapati wrote:
> > > Hi Rich,
> > > I have attached the patch which has all the MIPS n64 porting work. I
> > > have created mips64port remote branch on GitHub and the repository
> > > is https://github.com/MaheshBodapati/musl/tree/mips64port which has
> > > the broken down patches and the base revision on which I have
> > > prepared patch is v1.1.12-41-g3abb094.
> > 
> > Some preliminary review:
> 
> One more thing that came up in reviewing syscall_cp.s was actually a
> bug copied from existing code in musl, which is fixed by this commit:
> 
> http://git.musl-libc.org/cgit/musl/commit/?id=756c8af8589265e99e454fe3adcda1d0bc5e1963
> 
> In practice the code seemed to work but it was wrong with respect to
> ABI requirements.
> 
> I think the way you're saving $gp on the stack in sigsetjmp.s is also
> invalid since that part of the stack will have been clobbered by the
> time setjmp returns a second time. You could save it inside an unused
> part of the jump buffer, but it might be better to just avoid the $gp
> register and instead use temp registers and possibly some pc-relative
> address computations.

Likewise in pipe.s, it's overly complicated by the fact that it's
saving $gp on the stack and thereby can't make a tail call to
__syscall_ret like it's intended to. If you just use a different
register instead of $gp that's call-clobbered you can tail-call just
like the 32-bit mips version. Alternatively it's possible now to write
src/unistd/mips64/pipe.c instead using inline asm and let the compiler
handle the nastiness of the calling convention. I'd be happy with
either approach; the C version is probably mildly nicer, especially if
the compiler still does the tail-call right and doesn't generate come
bloated monstrosity. (Note that you can't use C for sigsetjmp though
because it can't use the stack.)

At this point I think I've looked through all the mips64 files at
least once and caught all the important issues.

Rich