@thakis thanks for taking care so quickly!

In D76780#1942190, @bryanpkc wrote:

In D76780#1942144, @bryanpkc wrote:

Why doesn't the KMP_MB after the store to th_next_waiting guarantee that the unblocked thread sees that store?

Answering my own question, the ARM architecture reference manual states that the DMB instruction ensures that all affected memory accesses by the PE executing the DMB that appear in program order before the DMB and those which originate from a different PE, ...which have been Observed-by the PE before the DMB is executed, are Observed-by each PE, ...before any affected memory accesses that appear in program order after the DMB are Observed-by that PE.

I think this means, in this case, that the store of FALSE to th_spin_here is observed in the correct order only by the releasing thread that issued the DMB. Other threads (e.g. the spinning thread) could still see the updates to th_spin_here and th_next_waiting out of order, unless they also issue DMB, which is the fix in this patch.

Yes, this looks related to memory consistency. As far as I understand the threads synchronize on th_spin_here, so this is guaranteed to be updated. Any other write before this in __kmp_release_queuing_lock is not guaranteed to be synchronized by a weak memory model. This includes th_next_waiting (which triggers the assertion), but also writes by the user application. That's particularly bad because this should be taken care of by the runtime!

In D75001#1922789, @jdoerfert wrote:

In D75001#1891284, @Hahnfeld wrote:

In D75001#1887876, @jdoerfert wrote:

I like this way better. I was hoping we could do it in our cmake only :)

Give others a day or so to comment before your commit but I'm fine with this.

Well, that doesn't really work if openmp-commits is only subscribed on commit. That said, the solution is a bit ugly but I don't have an alternative right now.

What is the problem with openmp-commits here? I got the emails, didn't you?

In D75001#1887876, @jdoerfert wrote:

I like this way better. I was hoping we could do it in our cmake only :)

Give others a day or so to comment before your commit but I'm fine with this.

In D74258#1865532, @JonChesterfield wrote:

Unless there are good reasons to switch, I'd object because it requires newer versions of compilers than many HPC systems have.

I disagree. The use case we would be supporting is:

• building llvm libomptarget from source
• using it with a binary llvm dist, or some other toolchain
• won't download or build a newer toolchain

This should have a RFC on openmp-dev. As far as I understand the linked patch, the aim is to use std::make_unique which is just syntactic sugar IMO. Unless there are good reasons to switch, I'd object because it requires newer versions of compilers than many HPC systems have.

LGTM from the code, I didn't test that the issue is indeed fixed on i386 though

@protze.joachim for 64 bit, we still need at least 5 threads IIRC

As a general comment, I think this patch should be split into at least two separate ones:

1. Load plugins from the same directory as libomptarget.so is found.
2. Quick fail mechanism, although I'm not sure how much this saves? In most cases, we won't have the plugins for a foreign architecture (PPC on x86) anyway, that might only be a concern for accelerators.

In D73249#1835551, @mgorny wrote:

In D73249#1835526, @Hahnfeld wrote:

I'm not sure LIBOMP_OMPT_SUPPORT is guaranteed to be there yet when defined in runtime/.

If by 'guaranteed' you mean in the current code, then I've tested this patch with it being both true and false.

I'm not sure LIBOMP_OMPT_SUPPORT is guaranteed to be there yet when defined in runtime/. Maybe it would be better to move the option to the top-level openmp/CMakeList.txt and rename it to OPENMP_OMPT_SUPPORT (with compatibility support for LIBOMP_OMPT_SUPPORT)

LG

In D70412#1821136, @protze.joachim wrote:

In D70412#1820406, @Hahnfeld wrote:

This revision was not accepted before being committed!

It was a misunderstanding on my side. Johannes told me that he is fine with the changes before I pushed. I didn't check that the accepted flag was not on the patch.
What procedure do you suggest?

This revision was not accepted before being committed!

Can you please also submit a patch to libc++ where we copied this code from?

In D71988#1799064, @jdoerfert wrote:

I would expect that some versioning is needed..

Versioning is fine with me. Versioning the runtime would be even preferable.

Yes, this must not be changed for compatibility reasons. If this really brings a benefit, we need new entry functions and update the compiler(s).

In D71340#1779446, @simoll wrote:

To clarify: this patch is not intended to be committed as-is but gives an overview of the required changes. It's the a patch of a series for the VE backend and includes all the earlier patches.

In D70804#1763443, @bryanpkc wrote:

In D70804#1763289, @ABataev wrote:

Tests are required

The tests for this were added in D30644, but they are currently failing on AArch64 due to D32035. This patch allows those tests to pass again. What further testing should I add?

Random drive-by comments to parts of the patch, I won't review in full.

Testing of openmp is currently very lacking, but I don't think this change will apply cleanly anymore...

Going through my old revisions, is this still something that we want?

I won't continue work on this.

Heads up that this change broke compatibility with old binaries. Not sure if that is something worth to fix, but it should probably at least be spelt out somewhere...

In D66107#1666702, @Dor1s wrote:

In D66107#1666574, @Hahnfeld wrote:

I think the bots are also green, so it might be just related to how I build Clang (with libc++, for example). I'm half way through building ToT with GCC, that should give insight whether it's related to my system or my configuration.

Actually, some ARM bots got broken, so I've uploaded a fix: https://reviews.llvm.org/D67458

Just curious: So the number of new features and coverage edges does not need to be the same? I also have "15 new freatures" for CHECK2.

Yes, features include coverage edges + additional signal (e.g. value profiling), that's why the number is different and why number of features can differ between platforms.

In D66107#1666569, @Dor1s wrote:

Hm, doesn't fail for me, but I guess the feature detection might be platform-dependent to some extent, so I'm fine with replacing the number of the features with a regex. Do you want to upload a change, or should I?

The new test is failing for me because CHECK1 is not satisfied. Instead the line says MERGE-OUTER: 3 new files with 12 new features added; 11 new coverage edges (instead of 11 new features). I'm currently investigating what's wrong here, let me know if you have an idea.

LGTM

LG iff the following is correct: The leak is when dyn_cast can't cast and returns null, right? If so, please add into the summary (and do so for future revisions, it will speed up review).

The code itself looks good to me (it should probably be clang-formated). If others have comments, please raise them now.

In D67298#1662222, @serge-sans-paille wrote:

Successfully tested on my box! I'd just use

diff
 - UNSUPPORTED: asan-dynamic-runtime
 +REQUIRES: x86_64-target-arch && !android && !asan-dynamic-runtime

if that works on your side too.

For the record, I don't think this would have worked: I do have libstdc++ available on my system (I just choose SANITIZER_CXX_ABI=libcxxabi for other reasons) and passing -stdlib=libstdc++ when building the test doesn't help as I've already said in D67298. Moreover, there's nothing in the test that requires libstdc++, it just happens to pass due to many chained implications, and restricting a test without understanding the root cause is probably no good idea.

In D67298#1662107, @serge-sans-paille wrote:

And that's what I don't understand

My fault for not being clear enough. Let me elaborate the original faulty use case:

You've got an application written in C, compiled with asan, that dlopens a library written in C++, linked to libstdc++
When the application starts, asan installs a few handler based on the loaded symbols. It used to install a handler that points to itself for __cxa_throw because the symbol is not present in the binary.

I think I just figured out that the original problem is about static runtime, so it doesn't make sense to run the test with dynamic asan.

In D67298#1661672, @serge-sans-paille wrote:

Mmmmh the whole point of the test is to test how asan behaves when meeting a symbol like __cxa_throw in a dlopened library when the calling compilation unit is not linked against libstdc++. Previous behavior was a segfault, we should now get a reasonable error.

In D67298#1661672, @serge-sans-paille wrote:

Mmmmh the whole point of the test is to test how asan behaves when meeting a symbol like __cxa_throw in a dlopened library when the calling compilation unit is not linked against libstdc++. Previous behavior was a segfault, we should now get a reasonable error. By defining __cxa_throw in the c file, you remove the origin of the bug, so sure it works, but it no longer tests the faulty situation. Maybe there's a symbol caught by asan and defined in both libcxx and libstdc++?

I've posted a change that makes the test work for me in D67298, please take a look.

Also, there should be a summary of the changes in here.

In D63877#1656860, @serge-sans-paille wrote:

Is that summary correct?

It is :-)

In D63877#1656135, @serge-sans-paille wrote:

@Hahnfeld : I understand your concern. However this test case is meant to test handling of stdlib exceptions, so it doesn't look strange to disable it if stdlib is not used, right? [...]

This changes error messages, so I'd say it's not NFC.

In D63877#1641054, @serge-sans-paille wrote:

@Hahnfeld : I failed to reproduce the issue locally, but I assume the following patch may solve your issue:

--- a/compiler-rt/test/asan/TestCases/Linux/dlopen-mixed-c-cxx.c
+++ b/compiler-rt/test/asan/TestCases/Linux/dlopen-mixed-c-cxx.c
@@ -5,7 +5,7 @@
 //
 // CHECK: {{.*}}AddressSanitizer: failed to intercept '__cxa_{{.*}}throw{{.*}}'
 //
-// REQUIRES: x86_64-target-arch && !android
+// REQUIRES: x86_64-target-arch && !android && !cxxabi

can you give it a try?

LG, thanks for the changes.

LG

Please submit the test changes unrelated to the code changes in a separate patch!

In D65868#1625642, @jlpeyton wrote:

Unfortunately, the Windows atomic code will not compile with this change. See in kmp_atomic.h the different atomic types for Windows. The complaints are regarding assignment operations for the complex types.

../../src/kmp_atomic.cpp(1325): error: class "__kmp_cmplx32_t" has no suitable assignment operator
  ATOMIC_CMPXCHG(cmplx4, add, kmp_cmplx32, 64, +, 8c, 7,
  ^

The code changes look good to me, but the test doesn't pass on x86. We've faced the same problem when clang-offload-bundler was initially committed and the current testing is the best we were able to do.

Can we close this after D65341 landed?

Okay, so I wasn't happy with the current explanations because I don't like "fixing" an issue without understanding the problem. Here's a small reproducer:

$  cat main.cpp 
#include "test.h"

Jon, please subscribe to openmp-commits so that commit emails get through immediately. Thanks!

In D65819#1627638, @ABataev wrote:

In D65819#1627631, @Hahnfeld wrote:

In D65819#1627620, @ABataev wrote:

In D65819#1627600, @Hahnfeld wrote:

In D65819#1627036, @ABataev wrote:

In D65819#1627032, @Hahnfeld wrote:

In D65819#1617736, @Hahnfeld wrote:

Will this patch change the ability to consume a bundled object file without calling the unbundler? Using known ELF tools on the produced object files was an important design decision and IIRC was somewhat important for using build systems that are unaware of the bundled format.

Ping.

Missed this. We still produce correct object files, so all the tools will work with this.

I agree on a technical level that it's still a "correct" object, but not what I was looking for: The host object file will only be in the bundled section, so you cannot examine it without unbundling.

For example, with a small test file and clang -fopenmp -fopenmp-targets=x86_64 -c test.c I saw the following:

$ nm test.o                                       
0000000000000000 t .omp_offloading.requires_reg
0000000000000000 T test
                 U __tgt_register_requires

After applying this patch, the output is empty which might be a problem in certain cases.

Unfortunately, this is the only possible solution I see. There are already 2 reports that bundled objects does not work correctly after unbundling.

Can you please again share what exactly is the problem, with a small example? I saw discussions on openmp-dev, but that project was huge, and above you were quoting a man page and hinted to global constructors.

I don't have a small reproducer, unfortunately, only the big one.
Here is the message from the user:

Hi,
I am revisiting this possible compiler bug in Clang 8.0.0.

In the source code I am developing, there's a global static variable, nest::sli_neuron::recordablesMap_ put in the BSS section and it is expected to be fully initialized by the time nest::sli_neuron::sli_neuron() gets called, however in a gdb session:

(gdb) p nest::sli_neuron::recordablesMap_
Python Exception <type 'exceptions.AttributeError'> 'gdb.Type' object has no attribute 'name':
$1 = {<std::map<Name, double (nest::sli_neuron::*)() const, std::less<Name>, std::allocator<std::pair<Name const, double (nest::sli_neuron::*)() const> > >> = std::map with 0 elements, _vptr$RecordablesMap = 0x0}

this doesn't happen when -fopenmp-targets is _not_ used, is it not trivial to come up a reproducer, thus I am sending a work-in-progress report hoping someone will shed some light on this.

Thanks,
Itaru.

Another one:

Hi,
I am seeing a link error shown below:

`.text.startup' referenced in section `.init_array.0' of /tmp/event_delivery_manager-02f392.o: defined in discarded section `.text.startup[_ZN4nest18DataSecondaryEventIdNS_16GapJunctionEventEE18supported_syn_ids_E]' of /tmp/event_delivery_manager-02f392.o
clang-10: error: linker command failed with exit code 1 (use -v to see invocation)

I am not sure how to tackle this as the part is referenced isn't what I am
working on. I am using the latest trunk Clang 10 at this moment.

Steps to reproduce:

Steps to produce:
$ git clone https://https://github.com/nest/nest-simulator/
$ mkdir /tmp/build/nest-clang-offload/
$ cd /tmp/build/nest-clang-offload/
$ cmake -DCMAKE_TOOLCHAIN_FILE=Platform/JURON_Clang -DCMAKE_INSTALL_PREFIX=/path/to/opt/nest-clang -Dwith-gsl=ON -Dwith-openmp=ON -Dwith-mpi=OFF -Dwith-python=OFF -Dwith-offload=ON /path/to/nest-simulator/

In D65819#1627620, @ABataev wrote:

In D65819#1627600, @Hahnfeld wrote:

In D65819#1627036, @ABataev wrote:

In D65819#1627032, @Hahnfeld wrote:

In D65819#1617736, @Hahnfeld wrote:

Will this patch change the ability to consume a bundled object file without calling the unbundler? Using known ELF tools on the produced object files was an important design decision and IIRC was somewhat important for using build systems that are unaware of the bundled format.

Ping.

Missed this. We still produce correct object files, so all the tools will work with this.

I agree on a technical level that it's still a "correct" object, but not what I was looking for: The host object file will only be in the bundled section, so you cannot examine it without unbundling.

For example, with a small test file and clang -fopenmp -fopenmp-targets=x86_64 -c test.c I saw the following:

$ nm test.o                                       
0000000000000000 t .omp_offloading.requires_reg
0000000000000000 T test
                 U __tgt_register_requires

After applying this patch, the output is empty which might be a problem in certain cases.

Unfortunately, this is the only possible solution I see. There are already 2 reports that bundled objects does not work correctly after unbundling.

In D65819#1627036, @ABataev wrote:

In D65819#1627032, @Hahnfeld wrote:

In D65819#1617736, @Hahnfeld wrote:

Will this patch change the ability to consume a bundled object file without calling the unbundler? Using known ELF tools on the produced object files was an important design decision and IIRC was somewhat important for using build systems that are unaware of the bundled format.

Ping.

Missed this. We still produce correct object files, so all the tools will work with this.

Rebase.

Change check_c_compiler_flag() to check_cxx_compiler_flag().