Re: [PATCH 2/2] add powerpc64 port

[Rich Felker <dalias@libc.org>]

On Sat, Apr 02, 2016 at 12:02:28PM -0500, Bobby Bingham wrote:
> > One thing that's not clear to me is whether the jumps/calls from asm,
> > as you've written them, go to the local entry point or the global
> > entry point of the callee. Since you're not loading r12 I assume the
> > intent is to go to the local entry point, but that only works if the
> > callee is hidden. Or does the PLT take care of converting that for you
> > (as long as r2 is set)?
> 
> Calls that go through the PLT go to the global entry point.  Otherwise,
> they go to the local entry point.
> 
> The PLT stub saves r2 to a slot in the stack from required by the ABI to
> be reserved for it, loads the global entry point address to r12, and
> calls it.  Calls that might go through the PLT are required to be made
> with a "bl; nop" sequence, so that the linker can replace the nop with
> an instruction to reload r2 from the stack where the PLT stub saved it.

Wow. That's clever but hideous. Thanks for explaining it, though.

> > One place that looks wrong to me is sigsetjmp. I would expect r2 to be
> > lost/clobbered when setjmp is called, but maybe not since it looks
> > like setjmp saves it in the jmp_buf, despite it not being a call-saved
> > register. Is that your expectation?
> 
> r2 is call-saved when using the local entry point.  When using the global
> entry point it's technically call-clobbered, but the PLT stub and linker
> conspire together to save/restore it.

OK, I see now.

> > > > > diff --git a/src/signal/powerpc64/sigsetjmp.s b/src/signal/powerpc64/sigsetjmp.s
> > > > > new file mode 100644
> > > > > index 0000000..ce59b60
> > > > > --- /dev/null
> > > > > +++ b/src/signal/powerpc64/sigsetjmp.s
> > > > > @@ -0,0 +1,30 @@
> > > > > +	.global sigsetjmp
> > > > > +	.global __sigsetjmp
> > > > > +	.type sigsetjmp,%function
> > > > > +	.type __sigsetjmp,%function
> > > > > +	.hidden ___setjmp
> > > > > +sigsetjmp:
> > > > > +__sigsetjmp:
> > > > > +	addis 2, 12, .TOC.-__sigsetjmp@ha
> > > > > +	addi  2,  2, .TOC.-__sigsetjmp@l
> > > > > +	.localentry sigsetjmp,.-sigsetjmp
> > > > > +	.localentry __sigsetjmp,.-__sigsetjmp
> > > > 
> > > > Again I don't see what the purpose of these insns is; if the resulting
> > > > value is needed, are you aware of how that interacts with ___setjmp
> > > > returning twice?
> > > 
> > > This sets up r2 with the TOC pointer, as is required by the ABI in order
> > > to call setjmp's local entry point.  Since setjmp is also written in asm,
> > > we could do away with this here.
> > > 
> > > I don't think the fact that setjmp returns twice matters for this.
> > 
> > When setjmp returns the second time, all registers it did not save
> > have been clobbered (by arbitrary code that ran after the first return
> > from setjmp). However despite not being a call-saved register
> > (AFAICT), r2 is saved by setjmp, so it's probably okay.
> 
> r2 is call-saved when calling to the local entry point, so setjmp needs
> to save it.

OK, I see how this works for local calls to setjmp. But how does the
linker PLT magic work for setjmp?

After the first return, the caller's stack slot where r2 was saved
belongs to the caller, and the compiler can clobber it. Upon the
second return, it would load junk into r2. Does longjmp have to do
something special (its own store to this stack slot, matching what a
PLT thunk would do) so that the caller loads the correct value?

Rich