LTO Optimization with asm renames of intrinsic functions

LTO Optimization with asm renames of intrinsic functions

[Paul E Murphy]

I ran into a slight divergence of behavior when enabling LTO using GCC
10.2 or newer when building libdfp on x86 and ppc64.

Libdfp builds a copy of libdecnumber which exports its own copy of ABI
as specified for _Decimal types in the GCC documentation.  Some of this
ABI is implicitly used in libdfp to implement various facilities.

These intrinsic calls generate PLT calls which should resolve to
symbols provided by libdfp.  However, libdfp tries to avoid these
by inserting asm renames of the form:

   asm ("__bid_floatsidd  = __GI___bid_floatsidd");

Where the libdfp copy of such symbols are prefixed with __GI_ using
the trick inherited from glibc.

This creative workaround has worked with lto disabled, but
when enabled, it is effectively ignored.  Is there a more robust way
to bypass plt calls for intrinsic functions when using LTO?

Re: LTO Optimization with asm renames of intrinsic functions

[Jeff Law]

On Wed, 2020-08-12 at 17:26 -0500, Paul E Murphy via Gcc-help wrote:
> I ran into a slight divergence of behavior when enabling LTO using GCC
> 10.2 or newer when building libdfp on x86 and ppc64.
> 
> Libdfp builds a copy of libdecnumber which exports its own copy of ABI
> as specified for _Decimal types in the GCC documentation.  Some of this
> ABI is implicitly used in libdfp to implement various facilities.
> 
> These intrinsic calls generate PLT calls which should resolve to
> symbols provided by libdfp.  However, libdfp tries to avoid these
> by inserting asm renames of the form:
> 
>    asm ("__bid_floatsidd  = __GI___bid_floatsidd");
If you've got an ASM like this, then you can't use LTO.  In simplest terms LTO
doesn't look inside ASMs for symbol references and can't inherently do the right
thing.

You might be able to use the symver attribute to do renaming, but I don't know if
anyone's tried it yet for anything other than specifying symbol versioning.

jeff
> 
> 

Re: LTO Optimization with asm renames of intrinsic functions

[Segher Boessenkool]

On Wed, Aug 12, 2020 at 05:26:35PM -0500, Paul E Murphy via Gcc-help wrote:
>   asm ("__bid_floatsidd  = __GI___bid_floatsidd");

> This creative workaround has worked with lto disabled, but
> when enabled, it is effectively ignored.  Is there a more robust way
> to bypass plt calls for intrinsic functions when using LTO?

Do linker scripts work with LTO?  You can do this in a linker script
(which is neater than doing it in assembler, anyway :-) )


Segher

Re: LTO Optimization with asm renames of intrinsic functions

[Jeff Law]

On Thu, 2020-08-13 at 18:07 -0500, Segher Boessenkool wrote:
> On Wed, Aug 12, 2020 at 05:26:35PM -0500, Paul E Murphy via Gcc-help wrote:
> >   asm ("__bid_floatsidd  = __GI___bid_floatsidd");
> > This creative workaround has worked with lto disabled, but
> > when enabled, it is effectively ignored.  Is there a more robust way
> > to bypass plt calls for intrinsic functions when using LTO?
> 
> Do linker scripts work with LTO?  You can do this in a linker script
> (which is neater than doing it in assembler, anyway :-) )
I wouldn't think that anything with symbol manipulation would work since LTO
wouldn't know about anything the linker script did.


jeff