Age | Commit message (Collapse) | Author | Files | Lines |
|
|
|
Adding perf tests for the single precision math functions in System.MathF.
|
|
1) Add LinkBench to the xunit-perf benchmark runs, as a scenario test
2) Some improvements to run-xunit-perf to
- Handle tests that perform size measurement
- Parametrization of test group
- Tests running multiple scenarios
CoreFX testing is currently disabled because it fails in the lab
due to a path length limitation.
|
|
tests to have the correct expected results.
|
|
precision when printing failure messages.
|
|
netcoreapp1.1 instead of netstandard1.4
|
|
|
|
* Disable tail call stress in GH_10780 if ZapDisable is enabled.
Tail call stress does not mix well with ZapDisable due to the issues
described in #11408.
Fixes #11648.
* Add a missing semicolon.
|
|
Turn on tests which are now passing.
|
|
(#11622)
|
|
This also adds metadata for all tests which are to be excluded
because they do not work with core.
|
|
This bug was yet another case of a transformation being performed during
remorphing that preserved a tree's value number without preserving its
associated semantics. This fix disables the problematic transformations
in morph.
|
|
(#11596)
|
|
(#11575)
|
|
(#11544)
|
|
[Arm64] Disable new test regressions
|
|
Disable tests based on #11533, #11534.
|
|
Morph simplifies (ind (addr (lclVar))) if the type of the load matches
the type of the lclVar node. This simplification is not valid, however,
if the lclVar must be normalized upon load: in that case, the lclVar
node must be normalized appropriately as part of the transformation.
This change fixes the simplification to perform this normalization if
necessary by calling `fgMorphLclVar` on the result if the result is a
lclVar node.
Fixes issue #11508.
|
|
|
|
Report the linker size reduction metric as a ratio
(linked filesize / unlinked filesize) instead of as a percentage.
|
|
Windows Arm64 - Build/Run tests in CI
|
|
Build tests with arm64 runs and upload/run them to the arm64 test
machines.
|
|
1. Move the testing to the latest version of ILLink package
(for all benchmarks except for Roslyn, which is still on netcoreapp1.1)
2. Support for LinkBench to run individual benchmarks in the suite --
helps local testing.
3. Support for running without clean, cloning, build etc --
for local runs during continuous development.
4. Support to obtain the correct version(s) of DotNet CLI required to run
the benchmarks
5. Get-Crossgen.ps1 is required to crossgen MusicStore binaries (before/after linking).
Since it was recently dropped from JitBench repo, it is added to linkBench as an asset.
6. Add the --perf:runid parameter to generate output file names suitable
to upload the results to BenchView.
|
|
Disable a few long-running tests.
|
|
|
|
These tests consistently time out under GC stress or heap verification.
Each disabled test is known to complete successfully given adequate
time.
|
|
Look at #11476 for the description.
|
|
Stop hoisting statics above cctors
|
|
|
|
Extend the logic in optLoopHoist around `CLS_VAR` nodes to ensure we don't
hoist any tree that contains a `CLS_VAR` for a field with the flag
`CORINFO_FLG_FIELD_INITCLASS` set, unless we also (identify and) hoist the
corresponding static init helper call. The previous logic was
insufficient in that it blocked hoisting of singleton `CLS_VAR` nodes, but
not hoisting of trees that contain `CLS_VAR` nodes as sub-trees.
Add flag `GTF_FLD_INITCLASS`/`GTF_CLS_VAR_INITCLASS` so that optLoopHoist
can recall which fields are paired with static init helper calls, and only
block the hoisting of their trees.
Fixes #10780.
|
|
|
|
Because it causes out of memory exception in crossgen mode.
|
|
(#11427)
|
|
|
|
|
|
When I merged the PR that enabled the lean runs I added another call to
our run script. This run script had an assumption that it would only
ever be called once and as such would try and copy two files that had
the same name. This would cause it to prompt to overwrite or not and
fail the run. This change makes each copy unique based on the
collection flags that were used for collection.
|
|
* Add GC Reliability Framework CI job
* Three changes:
1) To address offline feedback, update the CI trigger phase document
and the GC testing instructions
2) Set the timeout for GC RF CI jobs to 24 hours, so they don't time
out (they will run for 15 hours by default)
3) Refactor "scenario == 'gc_reliability_framework' to its own method
so that new GC RF scenarios can easily and cleanly be added if
desired in the future
|
|
|
|
In the xunit performance harness we added support for a "lean" run which
only runs the stopwatch to collect duration and none of the ETW
providers are turned on. This adds support in the automation for doing
runs that use this lean feature.
|
|
(#11384)
|
|
|
|
Add new test for COMPlus_TailcallStress=1 bug, currently disabled
|
|
|
|
(#11354)
|
|
This adds a run that will fire on the set of Azure VMs and will only run
for one interation. This along with the changes to fail a performance
run if a test fails will allow us to start gating checkins on the
performance tests still running.
|
|
|
|
|
|
Rather than disabling tailcall tests entirely, surgically remove the
failing cases so the remaining cases can be enabled.
|
|
(#11326)
|
|
|