[musl] [PATCH] arm: Fix _init/_fini alignment in crti.o.

[musl] [PATCH] arm: Fix _init/_fini alignment in crti.o.

[Alex Rønne Petersen]

This is just cbf59dd6 applied to arm. It manifested when compiling for thumb2
with zig cc (clang) + lld at -O1 or above, but not at -O0.
---
 crt/arm/crti.s | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/crt/arm/crti.s b/crt/arm/crti.s
index 18dc1e41..cccda3ea 100644
--- a/crt/arm/crti.s
+++ b/crt/arm/crti.s
@@ -3,11 +3,13 @@
 .section .init
 .global _init
 .type _init,%function
+.align 2
 _init:
 	push {r0,lr}
 
 .section .fini
 .global _fini
 .type _fini,%function
+.align 2
 _fini:
 	push {r0,lr}
-- 
2.40.1

Re: [musl] [PATCH] arm: Fix _init/_fini alignment in crti.o.

[Rich Felker]

On Thu, Oct 10, 2024 at 10:50:46PM +0200, Alex Rønne Petersen wrote:
> This is just cbf59dd6 applied to arm. It manifested when compiling for thumb2
> with zig cc (clang) + lld at -O1 or above, but not at -O0.
> ---
>  crt/arm/crti.s | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/crt/arm/crti.s b/crt/arm/crti.s
> index 18dc1e41..cccda3ea 100644
> --- a/crt/arm/crti.s
> +++ b/crt/arm/crti.s
> @@ -3,11 +3,13 @@
>  .section .init
>  .global _init
>  .type _init,%function
> +.align 2
>  _init:
>  	push {r0,lr}
>  
>  .section .fini
>  .global _fini
>  .type _fini,%function
> +.align 2
>  _fini:
>  	push {r0,lr}
> -- 
> 2.40.1

No objection to this fix, but I wonder if after applying it we should
just remove these files entirely. We have NO_LEGACY_INITFINI defined
on arm due to a historical mistake, so these will never be called in
dynamic-linked programs. I guess theoretically they could be used with
static linking, but all functional toolchains will be using init/fini
arrays instead or they would be incompatible with dynamic linking.

Rich

Re: [musl] [PATCH] arm: Fix _init/_fini alignment in crti.o.

[Alex Rønne Petersen]

On Thu, Oct 10, 2024 at 10:57 PM Rich Felker <dalias@libc.org> wrote:
>
> On Thu, Oct 10, 2024 at 10:50:46PM +0200, Alex Rønne Petersen wrote:
> > This is just cbf59dd6 applied to arm. It manifested when compiling for thumb2
> > with zig cc (clang) + lld at -O1 or above, but not at -O0.
> > ---
> >  crt/arm/crti.s | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/crt/arm/crti.s b/crt/arm/crti.s
> > index 18dc1e41..cccda3ea 100644
> > --- a/crt/arm/crti.s
> > +++ b/crt/arm/crti.s
> > @@ -3,11 +3,13 @@
> >  .section .init
> >  .global _init
> >  .type _init,%function
> > +.align 2
> >  _init:
> >       push {r0,lr}
> >
> >  .section .fini
> >  .global _fini
> >  .type _fini,%function
> > +.align 2
> >  _fini:
> >       push {r0,lr}
> > --
> > 2.40.1
>
> No objection to this fix, but I wonder if after applying it we should
> just remove these files entirely. We have NO_LEGACY_INITFINI defined
> on arm due to a historical mistake, so these will never be called in
> dynamic-linked programs. I guess theoretically they could be used with
> static linking, but all functional toolchains will be using init/fini
> arrays instead or they would be incompatible with dynamic linking.

FWIW, most musl usage with Zig is statically linked, including the
conditions under which this issue occurred.

That said, we unconditionally use init/fini arrays for Zig code, and
it appears that clang, and by extension zig cc, uses them by default
as well (you *can* opt out... not sure why you would though).

>
> Rich