I agree with what Daniel said. Even without explicit @build tags (as in
the reproducer in CODETOOLS-790198), if you use something like
@run main RedefineClassHelper
that would cause an implicit invocation of "@build test/lib", because
RedefineClassHelper.java is part of test/lib
So it's not possible to avoid @build altogether (even if you're not
using reflection).
=====
More explanation of the jtreg bug:
In the CODETOOLS-790198 reproducer's case, another test
(ModifyAnonymous.java) uses jdk.test.lib.compiler.InMemoryJavaCompiler
without an explicit @build.
Later, when RedefineRunningMethodsWithResolutionErrors.java is executed
and runs "@run main RedefineClassHelper", the jtreg bug causes
classes/test/lib to be partially compiled -- RedefineClassHelper.class
is there, but InMemoryJavaCompiler.class is missing.
Sure, according to the jtreg docs, "@build
jdk.test.lib.compiler.InMemoryJavaCompiler" should be added to
ModifyAnonymous.java. However, when you test fails because ANOTHER TEST
forgets to add an @build, and you are looking at a sea of over 1000 test
cases, you're completely lost.
So what we have is a jtreg rule that says "you should ...", but there's
no enforcement (every test runs perfectly fine by itself), and when
failure happens there's no diagnostic that tells you who's to blame.
Seems like a perfect recipe for anarchy.
- Ioi
On 6/2/17 2:19 AM, Daniel Fuchs wrote:
Hi guys,
The jtreg bug really needs to be fixed.
What I hear is that adding an explicit @build in one test
can make an unrelated test that depends on the same library
but doesn't have the explicit @build fail (and possibly
randomly and intermittently depending of the order in
which tests are run).
This is very unintuitive, and the 'obvious' (thouhj maybe
wrong) fix for anyone stumbling on the issue would be to fix
the failing test by adding the explicit @build - not grep
the whole test base in search for a test that might have an
explicit @build, which as pointed elsewhere might well be
legitimate if the test is using reflection.
So until the jtreg bug is fixed, I am not at all sure that
removing all the explicit @build is the correct thing to do,
as it's still bound to make existing unrelated tests fail
randomly if new tests with an explicit @build are added
later on...
my2c
-- daniel
On 01/06/2017 23:37, Ioi Lam wrote:
On 6/1/17 1:17 PM, Igor Ignatyev wrote:
On Jun 1, 2017, at 1:20 AM, Chris Hegarty
<chris.hega...@oracle.com> wrote:
Igor,
On 1 Jun 2017, at 04:32, Igor Ignatyev <igor.ignat...@oracle.com>
wrote:
Hi Felix,
I have suggested the exact opposite change[1-3] to fix the same
problem.
I’m sorry, but this is all just too confusing. After your change,
who, or what, is
responsible for building/compiling the test library dependencies?
jtreg is responsible, there is an implicit build for each @run, and
jtreg will analyze a test class to get transitive closure for static
dependencies, hence you have to have @build only for classes which
are not in constant pool, e.g. used only by reflection or whose
classnames are only used to spawn a new java instance.
I suspect the problem is caused by a long standing bug in jtreg that
results in library classes being partially compiled. Please see my
evaluation in
https://bugs.openjdk.java.net/browse/CODETOOLS-7901986
In the bug report, there is test case that can reliably reproduce the
NoClassDefFoundError problem.
I think adding all the @build commands in the tests are just
band-aids. Things will break unless every test explicitly uses @build
to build every class in every library that they use, including all
the private classes that are not directly accessible by the test cases.
For example: doing this may be enough for now:
* @build jdk.test.lib.process.*
But what if in the future, jdk.test.lib.process is restructured to
have a private package jdk.test.lib.process.hidden? To work around
CODETOOLS-7901986, all the test cases that must be modified to the
following, which unnecessarily exposes library implementation details
to the library users:
* @build jdk.test.lib.process.* jdk.test.lib.process.hidden.*
Just imagine this -- "in order to use malloc() you must explicitly
build not only malloc(), but also sbrk() ... and every other function
in libc". That seems unreasonable to me.
By the way, we made a fix in the HotSpot tests (see
https://bugs.openjdk.java.net/browse/JDK-8157957) that got rid of
many (but not all) of the NoClassDefFoundErrors by *removing* the
@build lines .....
My proposal is, instead of just adding @build for band-aid, we should
fix CODETOOLS-7901986 instead.
Thanks
- Ioi
Test library code has no @modules tags, so does not explicitly
declare its
module dependencies. Instead module dependencies, required by test
library code, are declared in the test using the library. If we
wildcard, or
otherwise leave broad build dependencies, from tests then there is no
way to know what new module dependencies may be added in the future.
That is, one of, the reason(s) I asked Felix to be explicit about
the build
dependencies.
having explicit builds does not really help w/ module dependency, if
someone change a testlibrary class so it starts to depend on another
testlibrary class, jtreg will implicitly build it and if this class
has some module dependencies, you will have to reflect them in the
test.
generally speaking, I don't like having explicit build actions
because build actions themselves are implicit, so they don't really
help, it's still will be hard to spot missed explicit builds. not
having (unneeded) explicit builds is an easy rule to follow and we
can easily find all places which don't follow this rule by grep.
-- Igor
-Chris.
[1] https://bugs.openjdk.java.net/browse/JDK-8181391
[2]
http://mail.openjdk.java.net/pipermail/core-libs-dev/2017-June/048012.html
[3]
http://cr.openjdk.java.net/~iignatyev//8181391/webrev.00/index.html
