Re: [PATCH] Expand PIC calls without PLT with -fno-plt

[Rich Felker <dalias at libc dot org>]

On Wed, May 06, 2015 at 11:44:57AM -0700, H.J. Lu wrote:
> On Wed, May 6, 2015 at 11:37 AM, Rich Felker <dalias@libc.org> wrote:
> > On Wed, May 06, 2015 at 11:26:29AM -0700, H.J. Lu wrote:
> >> On Wed, May 6, 2015 at 10:35 AM, Rich Felker <dalias@libc.org> wrote:
> >> > On Wed, May 06, 2015 at 07:43:58PM +0300, Alexander Monakov wrote:
> >> >> On Wed, 6 May 2015, Jakub Jelinek wrote:
> >> >> > The linker would know very well what kind of relocations are used for
> >> >> > particular PLT slot, and for the new relocations which would resolve to the
> >> >> > address of the .got.plt slot it could just tweak corresponding 3rd insn
> >> >> > in the slot, to not jump to first plt slot - 16, but a few bytes before that
> >> >> > that would just load the address of _G_O_T_ into %ebx and then fallthru
> >> >> > into the 0x4c2b7310 snippet above.  The lazy binding would be a few ticks
> >> >> > slower in that case, but no requirement on %ebx to contain _G_O_T_.
> >> >>
> >> >> No, %ebx is callee-saved, so you can't outright overwrite it in the PLT stub.
> >> >
> >> > Indeed. And the situation is the same on almost all targets. The only
> >> > exceptions are those with direct PC-relative addressing (like x86_64)
> >> > and those with reserved inter-procedural linkage registers and
> >> > efficient PC-relative address loading via them (like ARM and AArch64).
> >> > MIPS (o32) is also an interesting exception in that the normal ABI is
> >> > already PLT-free, and while callees need a PIC register loaded, it's a
> >> > call-clobbered register, not a call-saved one, so it doesn't make the
> >> > same kind of trouble,
> >> >
> >> > I really don't see a need to make no-PLT code gen support lazy binding
> >> > when it's necessarily going to be costly to do so, and precludes most
> >> > of the benefits of the no-PLT approach. Anyone still wanting/needing
> >> > lazy binding semantics can use PLT, and can even choose on a per-TU
> >> > basis (or maybe even more fine-grained with pragmas/attributes?).
> >> > Those of us who are suffering the cost of PLT with no benefits
> >> > (because we use -Wl,-z,relro -Wl,-z,now) can just be rid of it (by
> >> > adding -fno-plt) and enjoy something like a 10% performance boost in
> >> > PIC/PIE.
> >> >
> >>
> >> There are things compiler can do for performance and correctness
> >> if it is told what options will be passed to linker.  -z now is one and
> >> -Bsymbolic is another one:
> >>
> >> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65886
> >>
> >> I think we should add -fnow and -fsymbolic.  Together with LTO,
> >> we can generate faster executables as well as shared libraries.
> >
> > I don't see how knowing about -Bsymbolic can help the compiler
> > optimize. Without visibility, it can't know whether the symbols will
> > be defined in the same DSO. With visibility, it can already do the
> > equivalent hints. Perhaps it helps in the case where the symbol is
> > already defined (and non-weak) in the same TU, but I think in this
> > case it should already be optimizing the reference. Symbol
> > interposition over top of a non-weak symbol from the same TU is always
> > invalid and the compiler should not be pessimizing code to make it
> > work.
> 
> -Bsymbolic will bind all references to local definitions in shared libraries,
> with and without visibility, weak or non-weak.  Compiler can use it
> in binds_tls_local_p and we can generate much better codes in shared
> libraries.

Yes, I'm aware of what it does. But at compile-time the compiler can't
know whether the referenced symbol will be defined in the same DSO
unless this is visibility annotation telling it. Even when linking a
shared library using -Bsymbolic, the library code can still make calls
(or data references) to symbols in other DSOs.

> > As for -fnow, I haven't thought about it much but I also don't see
> > many places where it could help. The only benefit that comes to mind
> > is on targets with weak memory order, where it would eliminate some of
> > the cost of synchronizing TLSDESC lazy bindings (see Szabolcs Nagy's
> > work on AArch64). It might also benefit PLT calls on such targets, but
> > you would get a lot more benefit from -fno-plt, and in that case -fnow
> > would not allow any further optimization.
> 
> -fno-plt doesn't work with lazy binding.  -fnow tells compiler that
> lazy binding is not used and it can optimize without PLT.  With
> -flto -fnow, compiler can make much better choices.

Ah, I see now you had LTO in mind. In that case the compiler does know
when the symbol is defined in the same DSO for -Bsymbolic. So that
clears up the usefulness of your proposed -fsymbolic. I still don't
see how -fnow would have a lot of practical usefulness, but I'm
certainly not opposed to it.

Rich