Chapel Code

Moved to GitHub

Archive Chapel at SourceForge SVN trunk on 2017-04-06

Large "whitespace" checkin to passes/scopeResolve.cpp

I am preparing to do some functional changes to scopeResolve.cpp

a) a few small cleanup operations

b) modest changes to support the continued relocation of module init

This is a large whitespace change ahead of that. Functions have
been shuffled so that they are generally top to bottom but then
grouped by the virtual "sub-passes" that exist within this pass.

I ran a full paratest to increase confidence but there were no
non-whitespace changes to any code whatsoever.

Compiled: clang on STD gcc on STD/LLVM
Tested: passes paratest
Review: none

Add trivial test that will be used to introduce team to git.

The test itself has nothing to do with git. The team will use it as an exercise
when learning git...

Early-in-week pass through REGRESSIONS file

I started my standard full sweep of REGRESSIONS against nightly logs (whether
reporting failures or not) to try and bring things up-to-date with the
present and found several cases that had not been noted or retired properly.
This should catch us up to today, approximately, the deltas being:

regress :(
----------
- undefined reference to chpl_bitops_debruijn64 with PGI compiler for new caching test
(should now be fixed)
- Qthreads "worker unit does not exist" error on threadring (should now be fixed)
- verification failures for new --about tests (should now be fixed)
- freadComplex glibc failure 'invalid next size'
- portability issues with filerator test for most PrgEnvs
- asserteof failure for PGI (???)
- stable double free or corruption errors with freadIntFailed, fwriteIntFailed
- gcc warnings about assuming strict overflow

progress!
---------
+ retired regressions related to empty directories in git
+ removed long-since fixed ptrans timeouts from performance testing
+ removed long-since fixed pidigits-pretty2 memory leak issue
+ removed fast regressions related to gcc 4.3 (they probably still fail, but if we're
not seeing them in testing, I don't think they belong here)
+ retired sudoku-reduce2 bug, which looks more TRIAGE than REGRESSION
+ retired build failures which are no longer failing
+ retired gcc 4.8 -Waggressive-loop-optimmizations problems which seem resolved
+ retired a number of gasnet.qthreads.numa issues which seem resolved

other
-----
* added a new section for VM-based performance runs
* re-sorted failures to match order in cron mailings
* ascribed memory leaks increase to r23794
* generally removed most timeouts since we've decided they belong more in TRIAGE,
unless they were sufficiently consistent to raise concerns
* made gasnet.qthreads.flat inherit from GASNet rather than repeating known issues

Seeing what happens with .perfexecenv removed

[discussed with Lydia]

In discussions today, we've been having trouble precisely why this
.perfexecenv was used, and our best guess was that Lydia thought she
saw performance improvements with it. To that end, we wanted to
remove it for a few nights, to see if that's the case; the ultimate
goal being to reduce the special settings that tests require (in
general and for our shootout entry).

Greg points out that a benefit of this file is that it'll turn a
deadlock into an error message if the system resources are exceeded;
but given that the testing system has timeouts and the setting isn't
helpful outside of the performance testing system (and we haven't seen
systems on which we hit those limits), I'd prefer not to have the
file, if there are no other downsides.

If, after some performance testing, we see a change, we can discuss
whether that's surprising, and should reinstate it. I'm open to
more discussion on Greg's take as well, but not enough to not see
what the performance impacts are for a few days.

Fixing a PGI regression

[stenographing for Kyle on behalf of Michael]

PGI was failing on this C test because one of the operations in use
isn't supported by intrinsics, and the lookup table to support it is
in a .c rather than .h file. Added the .c file to the .compopts which
provides the table and should make this test pass on the next PGI
compile.

Do not set svn:ignore props when committing perf data.

The perf data is migrating to github, but we are going to use the svn interface
to simplify the transition. The repos already have .gitignore files and they
will not understand svn:ignore properties anyway.

Removing bulkget in favor of standard IO's readstring

Do some cleaning up.

[Not reviewed.]

There are three changes here:
- Change chpl_localeID to chpl_nodeID. We changed the name of this
global variable many months ago. We missed these uses, and were not
reminded of them because they occur in debug code and we don't build
qthreads with that enabled.
- Change some <TAB> characters to spaces.
- Fix some indentation. The r23836 changes indented by 2 columns, but
qthreads uses 4-space indentation.