Re: [PATCH] Skip gdb.base/func-ptrs.exp in thumb mode

[Pedro Alves <palves at redhat dot com>]

On 06/29/2016 05:25 PM, Yao Qi wrote:
> On Wed, Jun 29, 2016 at 12:55 PM, Pedro Alves <palves@redhat.com> wrote:
>>
>> I think the question is what happens at run time, when a function
>> pointer is set to a thumb function.  Say, with something like this:
>>
>> typedef void (func_ptr) (void);
>>
>> void arm_function (void);
>> void thumb_function (void); // add whatever __attribute__ necessary
>>                             // for thumb.
>>
>> func_ptr *ptr;
>>
>> void foo (void)
>> {
>>    ptr = thumb_function;
>>    ptr ();  // stop here
>> }
>>
>> Does "ptr" hold an address with the LSB set at run time, somehow?
>> I assume it does.
> 
> Yes, it does.
> 
>>
>> Then, doesn't this in gdb:
>>
>>  (gdb) print ptr = thumb_function
>>
>> end up with the same value (LSB set)?  If not, seems like that's a bug.
>>
> 
> ptr doesn't have LSB set, because the value of symbol "thumb_function"
> doesn't have LSB set in GDB.
> 
>> Note that from the original commit log, one can't tell what
>> exactly goes wrong in the test that requires skipping it.  It talks
>> in terms of Maciej's patch, and how some prototype based on that
>> behaves, and not in terms of what is wrong with the current support, or
>> if there's something the test is doing that is fundamentally impossible
>> to do on arm.
> 
> The test has nothing wrong, but current approach, added by Maciej, can't
> be extended for ARM thumb mode, as a result of my investigation and
> prototype.  In current approach, we propagate ISA bit of address into symbols
> and line tables, and that works for MIPS.  In ARM, the LSB of address
> only indicates that instruction set will be changed to thumb, and
> function is still
> 2-byte aligned.  For example, thumb_function's address is 0x00010510, and
> it really starts from there.  We only set LSB of address, or set it to
> 0x00010511,
> when it is the target address of branch with changing mode from arm to thumb.
> Set LSB of address in other places make no sense to ARM.

Agreed.

> 
> I also find that gdb.base/func-ptrs.exp fails on ppc64, due to the function
> descriptor.  GDB should assign function descriptor to the pointer rather than
> the function address.  I am thinking maybe we can use a different approach
> for arm and ppc64, that is, when we do function pointer assignment, use a
> gdbarch method to adjust the result, on arm, if the result is the entry of a
> thumb function, set its LSB; on ppc64, if the result is the entry of a function,
> set it to the address of function descriptor.  However, I don't know where is
> the best place (in expression evaluation?) to adjust the result of function
> pointer assignment.

Off hand, I'd think somewhere around value_assign / value_cast / value_cast_pointers.
 
Thanks,
Pedro Alves