RFR: JDK-8200358 Remove mapfiles for JDK executables
Magnus Ihse Bursie
magnus.ihse.bursie at oracle.com
Wed Apr 4 07:42:16 UTC 2018
On 2018-04-03 23:16, Erik Joelsson wrote:
> Looks good to me at least. Exporting symbols from executables seems
> wrong so applying hidden as default seems good to me.
Unfortunately, it was not this easy. :-(
Out of paranoia, I also started a test run on Windows, even though it
should not have been affected. Well, "should not". The added JNIEXPORT
turns into a __declspec(dllexport) on Windows, which causes the
Microsoft linker to behave like when linking a dll, so it creates a .lib
and .exp for each binary, in the bin directory.
I see two ways out of this.
1) Make the JNIEXPORT conditional on OS. Either by #ifdef before the
main function, or by declaring something like EXEC_EXPORT instead, and
let it be empty on Windows.
2) Let the part of SetupNativeCompilation that handles executables
behave more like the one for shared libraries, and setting -implib, etc.
I'm not sure what happens if you pass in -implib when linking an
executable which does *not* have any dllexport decorations. If it
breaks, then this does not really seem like a way forward. Otherwise,
it's probably the safest choice, since it will make sure we never leak
any *.lib or *.exp for a potential future executable.
What do you think?
> On 2018-04-03 13:42, Magnus Ihse Bursie wrote:
>> Here's an updated webrev that uses the same pattern as for native
>> shared libraries to hide non-exported symbols, for executables as well.
>> I hope you're happy now :-)
>> I know there's a bit of redundancy now for setting
>> -fvisibility=hidden and friends. I'll clean that up going forward.
>> On 2018-03-29 07:48, Martin Buchholz wrote:
>>> On Wed, Mar 28, 2018 at 2:48 PM, Magnus Ihse Bursie
>>> <magnus.ihse.bursie at oracle.com
>>> <mailto:magnus.ihse.bursie at oracle.com>> wrote:
>>> On 2018-03-28 22:39, Martin Buchholz wrote:
>>>> On Wed, Mar 28, 2018 at 12:07 PM, Magnus Ihse Bursie
>>>> <magnus.ihse.bursie at oracle.com
>>>> <mailto:magnus.ihse.bursie at oracle.com>> wrote:
>>>> Anyway, with this patch all symbols in executables will be
>>>> visible, so there should be no problem anyway.
>>>> The symbols visible in the main executable are a sort-of-public
>>>> interface. The difference is visible via e.g. nm -D, or any
>>>> native code that does dlsym(NULL, symbolName) (yes, we do
>>>> this!). The behavior of native code is likely to be affected if
>>>> there is a symbol conflict. The larger the exported symbol
>>>> table, the more overhead there will be at startup (probably). The
>>>> theory is that changing an interface requires a CSR.
>>> If I understand your objections correctly, you are claiming
>>> (correct me if I'm misunderstanding you):
>>> 1) Removing the mapfiles will affect performance negatively
>>> 2) The exported symbols from executables are a public API and the
>>> change therefore require a CSR.
>>> To this I reply:
>>> 1) While theoretically this might affect startup time, I can't for
>>> the life of me think this would even be measurable. I think any
>>> uncertainities in the measurement of the startup of "java" will
>>> dwarf any changes due to loading with a different set of exported
>>> symbols, in several orders of magnitude. If you claim otherwise, I
>>> challenge you to do the measurements.
>>> It's true the performance loss here is very small - every java
>>> program might be a microsecond slower to start up.
>>> 2) If this is a public API, then show me the documentation. If
>>> there is no documentation, then this is not a public interface.
>>> Just the fact that you might have used "nm" to locate symbols in a
>>> native file and use it, does not mean it's a public interface that
>>> requires a CSR to change. If that would be the case, then we could
>>> not ever do any change to any native file without filing a CSR,
>>> which is obviously absurd.
>>> Jigsaw team just spent 10 years working to prevent people from
>>> accessing Java internals. But here the proposal for ELF symbols is
>>> "just make everything public" Every ELF symbol that is needlessly
>>> exported is something that someone might build a dependency on or
>>> might cause a name conflict. ELF files don't have much
>>> encapsulation - all they have is public and private.
>>> If you have code that are dependent on a certain set of symbols or
>>> whatnot, and you want it to keep functioning, then I recommend
>>> that you file a bug and submit a patch to get it into mainline. If
>>> you're just collecting a bunch of downstream patches, and this
>>> change make your life harder, well, then, sorry, that's not my
>>> No, actually making everything public/exporting all symbols will
>>> probably make Google local changes easier to maintain - no map files!
>>> I would prefer if build team worked on generating map files with
>>> minimal symbols exported, instead of removing them entirely.
More information about the build-dev