- User Since
- May 4 2017, 4:51 AM (151 w, 6 d)
Oct 1 2017
More cleanups. Also fixed a pathologic edge case...
Jul 30 2017
Another round of fixes:
- All unit tests pass now. The last remaining test was changed to reflect the newly generated preconditions
- Types are now limited to 63 bits to leave space for the sign bit
- Fixed preconditions meaning the exact opposite of what was intended
- Generate preconditions and parts of polly-acc generated code (e.g. host <-> device memory transfer) using fixed 64 bits
- More code cleanup
Jul 25 2017
Another round of fixes:
- Fix last miscompiles in polly unittests. This brings us down to two testcases where the generated IR changed (GPU test cases excluded).
- Prepare PPCGCodeGen. It's still not working (memory accesses need to be handled), but move initialization of options on isl_ctx so that they are set early enough for this. Also fix the addditional RTC generated by PPCGCodeGen.
- Small cleanups
Jul 22 2017
Another round of cleanups:
Jul 18 2017
First round of cleanups:
- Down to four test cases that fail code generation
- Adressed most comments
- Added an option to isl to not print bounds on values that polly enables per default
Jul 17 2017
No, this is not supposed to only change isl inside of polly. Among the things that were noted when I originally wanted to upstream the code into isl is a lack of motivation for the interface, since at that time it was a big patch to isl only. After discussion with Sven, we agreed on changing the original interface (store information in AST (expression) nodes instead of directly inside expressions) but before cleaning up and preparing another round of patches and resubmitting it seems like a good idea to show the whole code in one place so that we can agree on what and how we'd like isl to provide.
Jul 16 2017
Note: Modifications in codegen.c are only for debugging.
Jul 9 2017
I looked into providing a testcase and it seems to be infeasible, because the only test cases I can reliably reproduce this on are
when running with "-polly-process-unprofitable". Those are all big multi-source tests and bugpoint also seems unable to reduce them by much. This seems to be kind of logical since this is essentially a heap corruption which is detected because glibc notices the corrupted free list, therefore we're most likely to detect it only if we're finding lots of ScOPs and then only as a side-effect, so I'm not sure how I would create a test that's suitable for our regression test suite in this case. I mean, I can include one of the above, but they are about 250KB.
POLLY_TEST_USE_VG results in make check-polly complaining, but as far as I can see due to a different issue that I have yet to track down...
Jul 3 2017
Update isl to isl-0.18-768-g033b61ae which includes an additional regression fix.
Jul 1 2017
Jun 30 2017
May 18 2017