Re: [RFC PATCH] MIPS: Kconfig: Select ARCH_WANT_FRAME_POINTERS

[Tiezhu Yang]

On 11/21/2020 06:37 AM, Maciej W. Rozycki wrote:
> On Mon, 16 Nov 2020, Tiezhu Yang wrote:
>
> > Select ARCH_WANT_FRAME_POINTERS to fix the following build error under
> > CONFIG_DEBUG_ATOMIC_SLEEP:
> >
> >    CC      arch/mips/kernel/signal.o
> > {standard input}: Assembler messages:
> > {standard input}:1775: Error: Unable to parse register name $fp
> > scripts/Makefile.build:283: recipe for target 'arch/mips/kernel/signal.o' failed
> > make[2]: *** [arch/mips/kernel/signal.o] Error 1
> > scripts/Makefile.build:500: recipe for target 'arch/mips/kernel' failed
> > make[1]: *** [arch/mips/kernel] Error 2
> > Makefile:1799: recipe for target 'arch/mips' failed
> > make: *** [arch/mips] Error 2
>   Your change description does not explain to me what is going on here I am
> afraid, and based on it I am unable to determine if it is fit for purpose.
>
>   It seems to me like your change papers over an issue by changing code
> generation somehow with the kernel configuration option selected so that
> invalid assembly is not produced anymore while invalid assembly should not
> happen in the first place regardless of the configuration.
>
>   In particular `$fp' is a standard assembly alias for `$30' aka `$s8' and
> it is expected to work where `$30' or indeed any general-purpose register
> would:
>
> #define SYMBOLIC_REGISTER_NAMES \
> [...]
>      {"$s8",	RTYPE_GP | 30}, \
>      {"$fp",	RTYPE_GP | 30}, \
>      {"$ra",	RTYPE_GP | 31}
>
> (from gas/config/tc-mips.c) so please show us what the assembly line GAS
> chokes on looks like in your case.
>
> > Documentation/dev-tools/kgdb.rst
> > This option inserts code to into the compiled executable which saves
> > the frame information in registers or on the stack at different points
> > which allows a debugger such as gdb to more accurately construct stack
> > back traces while debugging the kernel.
>   Hmm, this is what DWARF debug information is for in the context of GDB,
> and I certainly used to use GDB to debug standard MIPS/Linux kernels built
> without the use of a separate frame pointer register (which there wasn't a
> kernel configuration option for back then, though which you obviously
> still could try to enforce with the use of `-fno-omit-frame-pointer' via
> CFLAGS) using JTAG probes or simulation some 15 years ago.
>
>   And given the variable layout of the MIPS stack frame (unlike with some
> psABIs, e.g. Power) the use of `$fp' alone does not let you reconstruct a
> backtrace, because you cannot infer from the value of `$fp' where to
> retrieve the value of `$ra' from.  For that you need debug information.
>
>   So the information you quote seems misleading or missing the context.
>
>   NB hardly any MIPS software uses the frame pointer register and all is
> debuggable regardless; the only actual use for $fp is `alloca', VLAs or
> similar dynamic frame arrangements.
>
>   So what actual problem are you trying to solve, except for the assembly
> error, and what is your use case for `$fp' with MIPS kernel debugging?

Hi Maciej,

Thank you very much for your reply and detailed explanation.

The initial aim of this patch is to fix the build error.
I found this build error used with gcc 4.9.4.
I try it used with gcc 7.3.1 and it has no problem.
We can use new gcc version to avoid this build error.

The other commit message about config and kgdb seems no
related with the above build error, just give more info
to discuss.

As I see it now, this RFC patch is meaningless, so please
ignore it and thank you again.

Thanks,
Tiezhu

>    Maciej