pet_loopback and onriginal version of medley/deriche give different results

[Emil Vatai]

Hello!

While checking correctness of the implementation of my library which uses ISL/PET, I was running my code with polybench, and the medley/deriche benchmark was always failing (giving wrong/different results than the original). As a sanity check, I ran it through PET's `pet_loopback` which is supposed to give the code back (more or less) as is, but the results are different even with pet_loopback, i.e. without any transformations

(-DPOLYBENCH_DUMP_ARRAYS -DSMALL_DATASET)

I didn't have time to investigate it in detail, just wanted to check here first if someone knows about this issue.

Best,

Emil

[Sven Verdoolaege]

The default loop iterator names conflict with some variables
in the original code. PPCG takes care to use non-conflicting names,
pet_loopback does not.

skimo

[Sven Verdoolaege]

Now it does.

skimo

[Emil Vatai]

Thanks.

I think fixing that in Tadashi was one thing I asked a clanker (AI) to
do and I think it did a decent job:
https://github.com/vatai/tadashi/blob/ef51cd6ad1e534b9662182dc1f2aecabed3b465c/tadashi/src/codegen.c#L501

Ahhh... But I see you use pet_scop_get_ids()... I didn't know about
that! To quote Kintaro Oe: So educational!

Best,
E

--
Emil Vatai

[Sven Verdoolaege]

On Sat, Jun 20, 2026 at 08:34:10PM +0900, Emil Vatai wrote:
> Ahhh... But I see you use pet_scop_get_ids()... I didn't know about
> that! To quote Kintaro Oe: So educational!

I added it for this very purpose...
(even though it was mostly copied from ppcg)

skimo