Re: armv5 userland broken

[Nick Hudson <nick.hudson%gmx.co.uk@localhost>]

Great work!

On 29/06/2021 08:13, Rin Okuyama wrote:
> I've examined assembler codes for userland libraries for arm:
> common/lib/libc, csu, libc, compiler_rt, and libm, and found that
> some routines in compiler_rt do not align sp correctly:
>
> http://www.netbsd.org/~rin/arm_stack_align_compiler_rt.patch
>
> Also, some routines in libc do not also in Thumb-mode userland:
>
> http://www.netbsd.org/~rin/arm_stack_align_libc_thumb.patch
I think it's better / less ugly to push even numbers of registers rather
than adjusting sp (especially for thumb)


> By fixing these, armv5 system works just fine, except for new
> regression in tests/lib/libc/tls/t_tls_dlopen; the alignment fault
> occurs in the above mentioned ldrd in ld.elf_so!
>
> This seems a bug in GCC itself, since testf() in the test attempts to
> call EABI runtime routines before sp is aligned:
>
> ----
> t_tls_dlopen:     file format elf32-littlearm
>
> ...
>
> 00010ca0 <testf>:
>    10ca0:       e92d40f0        push    {r4, r5, r6, r7, lr}
>    10ca4:       ebffff4e        bl      109e4 <__aeabi_read_tp@plt>
>    10ca8:       e59f41d0        ldr     r4, [pc, #464]  ; 10e80
> <testf+0x1e0>
>    10cac:       e24dd00c        sub     sp, sp, #12
> ...
> ----
>
> This seems a GCC bug for armv5 and prior, that needs function call to
> obtain TLS pointer. armv6 and later use coprocessor instruction to
> get TLS pointer; coprocessor instruction can be always safely used
> before sp is adjusted.
>
> It seems that there are problem in load_tp_soft{,_fdpic} in arm.md in
> GCC, but I've never found how to fix it yet...
hmm


Nick