Re: Question about ARM -mapcs-frame documentation

[Richard Earnshaw <rearnsha at arm dot com>]

> The documentation for the ARM -mapcs-frame option says this:
> 
>     @item -mapcs-frame
>     @opindex mapcs-frame
>     Generate a stack frame that is compliant with the ARM Procedure Call
>     Standard for all functions, even if this is not strictly necessary for
>     correct execution of the code.  Specifying @option{-fomit-frame-pointer}
>     with this option will cause the stack frames not to be generated for
>     leaf functions.  The default is @option{-mno-apcs-frame}.
> 
> This says that the default is -mno-apcs-frame.  However, the default
> definition of TARGET_DEFAULT, and many of the overrides, include
> ARM_FLAG_APCS_FRAME.  The exceptions I see are:
>     arm*-*-linux*
>     arm*-*-uclinux*
>     arm*-wince-pe*
>     arm-*-pe*
>     strongarm-*-pe
>     arm-semi-aof
>     armel-semi-aof
> 
> So for other targets, I think -mapcs-frame is currently the default.
> These other targets are:
>     arm-*-coff*
>     armel-*-coff*
>     arm*-*-freebsd*
>     strongarm*-*-freebsd*
>     arm*-*-netbsdelf*
>     arm*-*-netbsd*
>     arm*-*-ecos-elf
>     arm*-*-rtems*
>     arm*-*-elf
>     ep9312-*-elf
>     arm*-*-kaos*
> 
> Is this an error in the documentation or an error in the code?

You are in a twisty little maze of options, all different.

The documentation is wrong, the default is target-specific.  If there are 
any targets in error, then it is probably those that default to 
-mno-apcs-frame, since on these it is hard to debug with gdb unless you 
have full dwarf unwind tables (support on ARM is only just coming); and 
the gcc guide says that the frame pointer should only be omitted if it 
does not affect debugging.

However, using a frame pointer is very costly, which is why some targets 
prefer to default to not using one.

The trend will increasingly be towards configurations that do not use a 
frame pointer; indeed that ATPCS and AAPCS (in preparation) make no use of 
a frame pointer and instead say that dwarf unwind information must always 
be emitted, even for non-debuggable code.

R.