In order to fix PR44560 and to prepare for loop transformations we now
finalize a function late, which will also do the outlining late. The
logic is as before but the actual outlining step happens now after the
function was fully constructed. Once we have loop transformations we
can apply them in the finalize step before the outlining.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
Patch looks good with the above nits.
I'm not totally sure about the callback vs running a separate IR pass after the finalize() call, but when the callback is this simple it looks fine. I like that this preserves the current semantics.
llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h | ||
---|---|---|
272 | Calling getNewOutlineInfo will invalidate the references previously returned. That's not a bug in this patch but looks like it'll be easy to get wrong in future. Perhaps the backing store should be a linked list, such that push_back doesn't invalidate any existing references? | |
llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp | ||
668 | s/parallel/outlined? |
clang/lib/CodeGen/CodeGenFunction.cpp | ||
---|---|---|
113 | Does this mean that finalize() is being called twice everytime, here and ~OpenMPIRBuilder()? |
clang/lib/CodeGen/CodeGenFunction.cpp | ||
---|---|---|
113 | It's called twice now, not that it matters. The basic rule is, call it at least once before you finalize the module. I'll remove the destructor for now to make it explicit only. | |
llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h | ||
272 | The address of these objects is not stable but without that requirement we don't need a list. We make this function take a OutlineInfo object that we than move into the vector if that is preferred. | |
llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp | ||
114 | Agreed, will be moved. | |
659–660 | I simplify forgot the remove them. Thx for noticing. | |
666 | I'll call it AfterIP again. | |
668 | this is in the omp parallel callback so "parallel" is still fine. |
clang/lib/CodeGen/CodeGenFunction.cpp | ||
---|---|---|
113 | Right now it doesn't matter, I agree. However, if later finalize() is modified for any reason, I worry this may introduce problems unintentionally. In any case, if you decide to keep only one of them, I prefer to keep the one in ~OpenMPIRBuilder(). This way the OMPBuilder is self contained, so regardless of the frontend using it, we can be certain it's always called at the end. You can indicate explicitly in the description comment for finalize() that it is being called in the destructor, if you want. |
clang/lib/CodeGen/CodeGenFunction.cpp | ||
---|---|---|
113 | That doesn't work, otherwise I would not call finalize here. At least it doesn't work without extra changes. As it is right now, the OMPBuilder in CodeGenModule is deleted *after* the IR is generated and emitted, so any changes at destructor time are too late. |
Can this delete NeedWorkaroundForOpenMPIRBuilderBug from llvm/lib/Transforms/Utils/CodeExtractor.cpp?
I didn't know about the workaround but that was the plan. I'll verify and add it to the patch.
Thanks :). IIRC check-clang is enough to exercise the relevant code path, as we get an assert in CodeExtractor without the workaround. (side note: please don't take my comments here as blocking, I just wanted to see if we could delete some cruft)
clang/lib/CodeGen/CodeGenFunction.cpp | ||
---|---|---|
113 | Got it. Then a minor suggestion is to just put a note in finalize() description about whatever you decide to do (i.e. keep both, remove the one in destructor, or something else entirely). This way, whoever is using/modifying the code is aware of it. |
Can you verify I deleted the workaround? I was not sure if the workaround is all of it or only the OMP... part.
I run the test you mentioned w/ and w/o the debug flag you mentioned in the PR, both times it passes w/o assertion:
PASS: Clang :: OpenMP/parallel_codegen.cpp
llvm/lib/Transforms/Utils/CodeExtractor.cpp | ||
---|---|---|
1408 | This part lgtm, thanks. |
I think this fixes the bug and allows loop optimizations later, let me know if there is anything else
Hello @jdoerfert,
OpenMPIRBuilderTest.ParallelCancelBarrier is failed on Windows builders:
http://lab.llvm.org:8011/builders/llvm-clang-x86_64-win-fast/builds/11242/steps/test-check-llvm-unit/logs/stdio
http://lab.llvm.org:8011/builders/llvm-clang-win-x-armv7l/builds/4537
http://lab.llvm.org:8011/builders/llvm-clang-win-x-aarch64/builds/4534
C:\buildbot\as-builder-3\llvm-clang-x86_64-win-fast\llvm-project\llvm\unittests\Frontend\OpenMPIRBuilderTest.cpp(612): error: Expected: CI->getNextNode()->getSuccessor(0) Which is: 000000DC6479BEC0 To be equal to: ExitBB Which is: 000000DC6479B140 [ FAILED ] OpenMPIRBuilderTest.ParallelCancelBarrier (4 ms)
Would you you fix the test or revert this commit?
it also gets failed on the Linux builders
http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-ubuntu/builds/2843
http://lab.llvm.org:8011/builders/lld-x86_64-ubuntu-fast/builds/11581
Could this be reverted for the time being so that bots can go green? There appear to be an awful lot of commits piling up on top of this one.
Nice, thank you! https://github.com/llvm/llvm-project/commit/3f3ec9c40b2574929b41b93ac38484081b49837b did the trick.
Does this mean that finalize() is being called twice everytime, here and ~OpenMPIRBuilder()?
if yes, is there a reason why you want that?