Re: GIMPLE FE

[Prasad Ghangal <prasad dot ghangal at gmail dot com>]

On 27 June 2016 at 14:16, Richard Biener <richard.guenther@gmail.com> wrote:
> On Sun, Jun 26, 2016 at 5:55 PM, Prasad Ghangal
> <prasad.ghangal@gmail.com> wrote:
>> Hi Richard,
>>
>> For the first stage of the project, I have completed the following things -
>> parsed -- assign-statement, labeled-statement, if-statement,
>> switch-statement, goto-statement, return-statement;
>> handled local declarations and
>> attempted hacking pass manager
>>
>> Link for github repository is : https://github.com/PrasadG193/gcc_gimple_fe
>
> I've CCed the gcc list - please do so in future as well.
>
> Looking at this now.  Do you have a set of test inputs that work?  I'm
> struggling
> to create even basic ones with local declarations.
>
> This means a good next step is to add testcases to the testsuite!
>
> For example I can add
>
>> cat gcc/testsuite/gcc.dg/gimplefe-1.c
> /* { dg-do compile } */
> /* { dg-options "-fgimple" } */
>
> int i;
> void __GIMPLE () foo ()
> {
>   i = 1;
> }
>
> and then run it from inside gcc/ in the build tree like so:
>
> obj-gimplefe-g/gcc> make check-gcc RUNTESTFLAGS="dg.exp=gimplefe*.c"
>
> which then reports
>
> ...
> Running target unix
> Using /usr/share/dejagnu/baseboards/unix.exp as board description file
> for target.
> Using /usr/share/dejagnu/config/unix.exp as generic interface file for target.
> Using /space/rguenther/src/gcc_gimple_fe/gcc/testsuite/config/default.exp
> as tool-and-target-specific interface file.
> Running /space/rguenther/src/gcc_gimple_fe/gcc/testsuite/gcc.dg/dg.exp ...
> FAIL: gcc.dg/gimplefe-1.c (test for excess errors)
>
>                 === gcc Summary ===
>
> # of unexpected failures        1
> /home/abuild/rguenther/obj-gimplefe-g/gcc/xgcc  version 7.0.0 20160531
> (experimental) (GCC)
>
> make[1]: Leaving directory `/home/abuild/rguenther/obj-gimplefe-g/gcc'
>
> the FAIL is due to the debug output you added (might be a good idea to
> maybe guard it with
> some extra flag).
>

I have added -fgimple-debug option which enable producing debug output.
I have also added some test cases to the testsuite and tried running
make check-gcc RUNTESTFLAGS="dg.exp=gimplefe*.c" which gives -

        === gcc tests ===

Schedule of variations:
    unix

Running target unix
Using /usr/share/dejagnu/baseboards/unix.exp as board description file
for target.
Using /usr/share/dejagnu/config/unix.exp as generic interface file for target.
Using /home/prasad/Workspace/gcc/trunk/gcc/testsuite/config/default.exp
as tool-and-target-specific interface file.
Running /home/prasad/Workspace/gcc/trunk/gcc/testsuite/gcc.dg/dg.exp ...

        === gcc Summary ===

# of expected passes        6
/home/prasad/Workspace/gcc/build1/gcc/xgcc  version 7.0.0 20160531
(experimental) (GCC)

make[1]: Leaving directory '/home/prasad/Workspace/gcc/build1/gcc'


> Playing with the input somewhat I can get the following to parse:
>
> int i;
> void __GIMPLE () foo ()
> {
>   int tem;
>   tem = i;
>   if (tem)
>     goto a;
>   else
>     goto b;
>
> a:
>   i = 2;
>   return;
>
> b:
>   i = 3;
> }
>
> int main()
> {
>   i = 1;
>   foo ();
>   if (i != 2)
>     __builtin_abort ();
>   return 0;
> }
>
> and it seems to work ok.  The above can even use { dg-do run } and run
> the produced
> executable successfully.
>
> Currently if you supply invalid GIMPLE as input it will be the gimple
> verification that
> will lead to an internal compiler error instead of the FE providing a
> nice error.  I think
> that's ok for now.
>
> So please add the various test inputs you must have accumulated as testcases
> in gcc/testsuite/gcc.dg/ as gimplefe-{1,2,3...}.c.
>
> Thanks,
> Richard.
>
>>
>>
>> Thanks,
>> Prasad Ghangal
>>
>> On 17 June 2016 at 03:34, Prasad Ghangal <prasad.ghangal@gmail.com> wrote:
>>> Hi Richard,
>>>
>>> Please follow the link for github repository -
>>> https://github.com/PrasadG193/gcc_gimple_fe
>>>
>>> Thanks,
>>> Prasad
>>>
>>>
>>> On 15/06/2016, Richard Biener <richard.guenther@gmail.com> wrote:
>>>> On Wed, Jun 15, 2016 at 8:47 AM, Prasad Ghangal
>>>> <prasad.ghangal@gmail.com> wrote:
>>>>> Hi Richard,
>>>>>
>>>>> Sorry I couldn't to mail you yesterday because I was travelling and I
>>>>> had no access to my pc and internet.
>>>>>
>>>>> Till now I am able to parse basic block (i.e bb_1: ) and extended
>>>>> gimple expression parsing. Currently I am trying to hack pass manager.
>>>>> My basic idea is to add custom pass list to function structure like
>>>>> gimple_opt_pass **pass_list ; and execute the list instead of all
>>>>> passes in execute_pass_list function. Please tell me if there is other
>>>>> better way to do this.
>>>>
>>>> For final unit-testing this might be the most flexible thing indeed, say
>>>>
>>>> __GIMPLE (execute ("ccp", "fre")) foo ()
>>>> {
>>>> ...
>>>> }
>>>>
>>>> exactly naming passes to execute.  That's best for unit testing that
>>>> in the end just inspects dump-files.
>>>>
>>>> For correctness testcases we'd need to generate code eventually
>>>> and thus there we only want to specify the entry point into the
>>>> optimization pipeline, say
>>>>
>>>> __GIMPLE (startwith ("fre1")) foo ()
>>>> {
>>>> ...
>>>> }
>>>>
>>>> You still have to go through the lowering phase if you do not build
>>>> a CFG right from the parser of course.
>>>>
>>>> I think doing execute ("ccp", "fre") is easier and using a custom
>>>> pass-list would work.  startwith would essentially need to perform
>>>> the equivalent of -fdisable-tree-X for each pass before the pass
>>>> to start with.  Eventually providing overrides for all the global
>>>> entries into the pass pipeline would work as well - see calls to
>>>> execute_ipa_pass_list and execute_pass_list.  Unfortunately the
>>>> pass manager doesn't have full control over the compilation but
>>>> much of it is driven via ad-hoc code.
>>>>
>>>> Btw, please make your work accessible via git on github.
>>>>
>>>> Thanks,
>>>> Richard.
>>>>
>>>>>
>>>>>
>>>>> On 14 June 2016 at 19:50, Richard Biener <richard.guenther@gmail.com>
>>>>> wrote:
>>>>>> Hi Prasad,
>>>>>>
>>>>>> can you please point me to the git you use for development and/or send an
>>>>>> update
>>>>>> on how the work goes along?  I will be on vacation starting from
>>>>>> Thursday, June 16
>>>>>> and will return only on Monday, June 27 which is the deadline for the
>>>>>> midterm
>>>>>> evaluation.  Currently I am not impressed with how you report progress
>>>>>> or (fail to)
>>>>>> ask questions.
>>>>>>
>>>>>> Thus please make sure I have something to look at when I return from
>>>>>> vacation.
>>>>>>
>>>>>> Thanks,
>>>>>> Richard.
>>>>
>>>
>>>
>>> --
>>> Thanks and Regards,
>>> Prasad Ghangal