This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [GSoC] generation of Gimple code from isl_ast_node_if


On 26/07/2014 15:48, Roman Gareev wrote:
What do you mean by wrong answer? Is there still a bug in the code
generation or the AST is just more complex as expected.

I mean that the value of res is wrong. I think it is because of the
wrong id of pbb->domain and pbb->transformed inherited from S_3 by S_9
(It was correct for S_3, but it is incorrect for S_9).

Sorry Roman. I am still confused. Are we looking for a test case or are we still trying to understand an issue. Specifically, do we still incorrectly transform the code even after your isl_id_for_pbb() patch?

To my understanding bb copies are introduced in case reductions are seen.
You could try to just initialize an array (maybe a little bit more complex,
but without += statements):

for i:
   A[i] = ...

You could do the summation/verfication outside of the scop e.g. in the main
function.

It seems that it doesn't help (at least for now).

Help for what? I was looking to create a simple test case. Is there still an open bug?

I am looking for a simple test case that does _not_ contain a reduction in addition to the test case you already have.

However, I've found out the following example:

static int __attribute__((noinline))
foo ()
{
   int i, res = 0;

   for (i = 0; i < 50; i++)
     {
       if (i < 5)
         res += 1;
     }

This one still has a reduction.

It gives the correct result, inspite of duplicating of pbbs and
generates the following ISL AST

{
   for (int c1 = 0; c1 <= 49; c1 += 1) {
     for (int c2 = 0; c2 <= 1; c2 += 1)
       S_3(c1);
     if (c1 <= 4)
       S_4(c1);
     S_5(c1);
   }
   S_7();

And this one still duplicates pbbs. So this is not the simple test case I am looking for. It introduces several new statements I do not understand, so it does not seem to be a trivial test case.

Maybe we could use it. What do you think about this?

What do you want to use it for?

Cheers,
Tobias


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]