Validate failures on the new codegen
Edward Z. Yang
ezyang at MIT.EDU
Fri Apr 8 14:09:47 CEST 2011
The T* failures are all memory allocation failures on the part of
GHC, probably because Hoopl generates a lot more garbage than the
old code generator.
The space leak one is really interesting, because it doesn't show up
when I do normal tests with a devel2 built stage2. Maybe I fixed it
with my optimization pass.
Excerpts from Simon Marlow's message of Fri Apr 08 04:41:24 -0400 2011:
> On 07/04/2011 17:07, Edward Z. Yang wrote:
> > Amazingly enough, a GHC fully built with the new code generator
> > and with it set to default only fails five tests on validate:
> > OVERALL SUMMARY for test run started at Thu Apr 7 11:53:31 EDT 2011
> > 2712 total tests, which gave rise to
> > 9094 test cases, of which
> > 0 caused framework failures
> > 6691 were skipped
> > 2317 expected passes
> > 81 expected failures
> > 0 unexpected passes
> > 5 unexpected failures
> > Unexpected failures:
> > T1969(normal)
> > T3064(normal)
> > T3294(normal)
> > T4801(normal)
> > space_leak_001(normal)
> > Which appear to be performance problems.
> Interesting - in what way do they fail? We hope that the new code
> generator doesn't affect allocation or stack use.
More information about the Cvs-ghc