Re: [RFC 3/3] Test on solib load and unload

[Yao Qi <yao@codesourcery.com>]

On 09/23/2013 08:14 AM, Doug Evans wrote:
> I think it is the test framework's responsibility to provide the utilities to.
> Large tests (of the size we need to collect data for) are best not
> written by hand, and if we're going to machine generate source, I
> would rather such generators come from the framework than always be
> hardcoded into every such test.  [Obviously some tests may have unique
> needs though.]
>

Doug,
Generating source is easy in this test case.  However, I am not sure it 
is easy to generate source for other perf test cases, like symbols and 
types.  Supposing we want to generate source files have 1 million 
classes, with some hierarchies, the generation script can't be simple, 
IMO.  On the other hand, I don't know how representative the generated 
program is, compared with the real large applications, such as 
openoffice, clang, etc.

>> >We can add a new proc gdb_produce_source with two parameters, NAME and
>> >SOURCES.  NAME is the file name and SOURCES is a list of lines of source
>> >code we want to write to file NAME.  For instance,
>> >
>> >    gdb_produce_source $src { "int shr$i (void) {return 0;}" }
>> >
>> >It can be used here and replace some code in gdb.exp.
> Here's an incomplete list of some of the axes we need to test (in random order):
> - # threads
> - # shared libs

They are not hard to generate.

> - # ELF symbols
> - # object files
> - # types (e.g., # DWARF type units)

I am not familiar with type and symbols, but I assume we need some 
scripts to generate source files having a large number of different 
symbols and types, which looks hard.

> - stack depth

It is not hard to generate either.

> - # pretty-printers?
>

I am OK to add utilities to generate sources for shared libs and stack 
depth, but I am still unable to find an approach to generate source 
files for perf tests on symbols and types.

-- 
Yao (齐尧)