Re: [PATCH] Make ftrace tests work with remote targets

[Pedro Alves <palves at redhat dot com>]

On 03/31/2016 09:24 PM, Simon Marchi wrote:
> On 16-03-29 06:46 PM, Pedro Alves wrote:
>> On 03/03/2016 09:22 PM, Simon Marchi wrote:
>>> --- a/gdb/testsuite/lib/gdb.exp
>>> +++ b/gdb/testsuite/lib/gdb.exp
>>> @@ -3374,7 +3374,7 @@ proc gdb_compile {source dest type options} {
>>>      # dynamically load one by basename, we must specify rpath.  If we
>>>      # are using a remote host, DejaGNU will link to the shared library
>>>      # using a relative path, so again we must specify an rpath.
>>> -    if { $shlib_load || ($shlib_found && [is_remote target]) } {
>>> +    if { $shlib_load || $shlib_found } {
>>
>> I think the comment above needs updating.
> 
> What about:
> 
>      # Because we link with libraries using their basename, we may need
>      # (depending on the platform) to set a special rpath value, to allow
>      # the executable to find the libraries it depends on.

That's great.

> 
>>>  	    lappend link_options "additional_flags=-Wl,--out-implib,${name}.a"
>>> -	} elseif [is_remote target] {
>>> +	} else {
>>>  	    # By default, we do not set the soname.  This causes the linker
>>
>> Likewise.
> 
> What about:
> 
>              # Set the soname of the library.  This causes the linker on ELF
>              # systems to create the DT_NEEDED entry in the executable referring
>              # to the soname of the library, and not its absolute path.  This
>              # (using the absolute path) would be problem when testing on a
>              # remote target.
>              #
>              # In conjunction with setting the soname, we add the special
>              # rpath=$ORIGIN value when building the executable, so that it's
>              # able to find the library in its own directory.

Perfect.

> 
> 
>>>  proc gdb_load_shlibs { args } {
>>> -    if {![is_remote target]} {
>>> -	return
>>> -    }
>>> +    if {[is_remote target]} {
>>> +	foreach file $args {
>>> +	    # When the target is remote, we simply send the file to the target.
>>> +	    gdb_download [shlib_target_file $file]
>>> +	}
>>> +    } else {
>>> +	foreach from $args {
>>> +	    # When the target is native, we copy the files to the test directory
>>> +	    # (next to the executable), except if that's already where it is.
>>> +	    set to [standard_output_file [file tail $from]]
>>
>> I'd think it better to make the gdb_download path work for native
>> testing as well.  WDYT?  E.g., make shlib_target_file default to
>> return the test directory path?
> 
> If I understand correctly, gdb_download would take a local file path and take
> care of transferring that file to the "runtime test directory", whether it's
> local or remote.

Right.  I'm not sure whether it's the right level or not.

  If the target is local, it checks whether $src == $dest, and
> copies the file if not.  Basically, the same as my implementation of
> gdb_load_shlibs, just not limited to shlibs.

*nod*

Now that I grep for gdb_download, sounds like we'll need to
change mi_load_shlibs similarly.  Not sure offhand why the jit*.so
hits don't use gdb_load_shlibs instead or gdb_download 

> 
> I am not sure if "make shlib_target_file default to return the test directory path?"
> would be ok.  I think it's meant to return the local file path to download to the
> target.

You're right.  I added those for SymbianOS years ago, since that target 
had a post link step to convert elf to pe/dll, and it'd be the latter that'd
be copied to the target.  IIRC.


>  So gdb_load_shlibs would then become:
> 
> proc gdb_load_shlibs { args } {
>     foreach file $args {
>         gdb_download [shlib_target_file $file]
>     }
> }
> 
> I just tested it quickly and it seems to work fine.  But I also noticed that there
> might be some overlap in feature between gdb_download and gdb_remote_download, so
> perhaps I can try to merge them, WDYT?
> 

Indeed.

>>> # Even if the target supplies full paths for shared libraries,
>>> # they may not be paths for this system.
>>> gdb_test "set solib-search-path [file dirname [lindex $args 0]]" "" ""
>>
>> Can we skip this command on native testing?
>>
>> I'm worried that that command might paper over, or cause, issues with
>> sysroot / dso path lookup.  Normal native debugging usage will not specify
>> that command, so if we could avoid it, I'd prefer it, on grounds of
>> testing what users normally use.  
> 
> Makes sense.  I'll add that, with a comment explaining why.
> 
> Thanks!
> 

Thank you!

-- 
Pedro Alves