Re: [patch, arm] Consistent display of registers in corefile

[Mark Kettenis <mark dot kettenis at xs4all dot nl>]

> Date: Fri, 10 Dec 2010 21:37:30 +0800
> From: Yao Qi <yao@codesourcery.com>
> 
> GDB trunk has a test failure on ARM,
> 
>    FAIL: gdb.base/gcore.exp: corefile restored general registers
> 
> In short, this failure is caused by output of 'info registers' before
> coredump doesn't match output of 'info registers' when corefole is
> loaded again, there are mainly two differences, [1] and [2].
> 
> Output before coredump,
> r0             0x12008  73736^M
> r1             0xbea1f0c0       -1096683328^M
> [...]
> sp             0xbea1f0a4       0xbea1f0a4^M
> lr             0x849b   33947^M
> pc             0x83fc   0x83fc <terminal_func+4>^M
> cpsr           0x20000030       536870960^M
> 
> Output when corefile is loaded,
> r0             0x12008  73736^M
> r1             0xbea1f0c0       3198283968^M  // <---- [1]
> [...]
> sp             0xbea1f0a4       0xbea1f0a4^M
> lr             0x849b   33947^M
> pc             0x83fc   0x83fc <terminal_func+4>^M
> fps            0x727a622f       1920623151^M  // <---- [2]
> cpsr           0x20000030       536870960^M
> 
> The difference [1] is caused by different register types, uint32 vs.
> int32.  In tdesc, the type of general register is "int", while in
> arm_register_type, it is regarded as builtin_uint32.  This can be fixed
> when register type is handled in a consistent way (in reg_type.patch).
> 
> The difference [2] is about displaying "fps" in output of "info
> registers".  In default_register_reggroup_p, the group of register is
> determined by the type of register, which is not very precise.  FPS
> should be in float group, but its type is INT.  This can be fixed by
> defining ARM's own register_reggroup_p to override
> default_register_reggroup_p (in arm_fps_group.patch).
> 
> Regression tested with combination of
> "\{-mthumb,-marm\}\{-fstack-protector,-fno-stack-protector}\{-march=armv7-a,-march=armv5t\}".
> 
> OK for mainline?

I would suspect that the proper thing to do would be to align the
tdesc with the code instead of the other way around.  The arm-core.xml
file seems to underspecify things by omitting the type=xxx clause on
many registers.  Whoever wrote arm_register_type() at least had to
make a conscious decision about the signedness of the type used for
the general purpose registers.