> If this is not satisfactory I'll have to think about updating DECL_ALIGN
> after the function's attribute are processed and keep using it in
> assemble_start_function.
> 
> Come back later
> 

Indeed, I found that resetting DECL_ALIGN in the
arm_set_current_function hooks is much cleaner than my previous patch,
and, as you said, preserve the case where DECL_USER_ALIGN is 1.

First I thought it was weird to change a tree fndecl field in the
targetm.set_current_function (fndecl) hook as I haven't seen in any
other target and that only the restore_target_globals or
cl_target_option_restore functions would switch the states between
functions (like align_functions_log, target_flags) thanks to
target_reinit. But it seems to be the less intrusive one line fix I
found without touching the middle end parts. Other suggestions welcome
though.

I'm testing with the attached patch to replace the previous (2/2) one
and I'll discuss it next week with the arm maintainers after more testing.
sorry for the noise.

thanks