Backports for 1.12-rc2 #59061
Open
Backports for 1.12-rc2 #59061
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This was in DAECompiler.jl code found by @serenity4. He also mentioned that writing up how one might go and fix a bug like this so i'll give a quick writeup (this was a very simple bug so it might not be too interesting) The original crash which looked something like > %19 = alloca [10 x i64], align 8 %155 = insertelement <4 x ptr> poison, ptr %19, i32 0 Unexpected instruction > [898844] signal 6 (-6): Aborted in expression starting at /home/gbaraldi/DAECompiler.jl/test/reflection.jl:28 pthread_kill at /lib/x86_64-linux-gnu/libc.so.6 (unknown line) gsignal at /lib/x86_64-linux-gnu/libc.so.6 (unknown line) abort at /lib/x86_64-linux-gnu/libc.so.6 (unknown line) RecursivelyVisit<llvm::IntrinsicInst, LateLowerGCFrame::PlaceRootsAndUpdateCalls(llvm::ArrayRef<int>, int, State&, std::map<llvm::Value*, std::pair<int, int> >)::<lambda(llvm::AllocaInst*&)>::<lambda(llvm::Use&)> > at /home/gbaraldi/julia4/src/llvm-late-gc-lowering.cpp:803 operator() at /home/gbaraldi/julia4/src/llvm-late-gc-lowering.cpp:2560 [inlined] PlaceRootsAndUpdateCalls at /home/gbaraldi/julia4/src/llvm-late-gc-lowering.cpp:2576 runOnFunction at /home/gbaraldi/julia4/src/llvm-late-gc-lowering.cpp:2638 run at /home/gbaraldi/julia4/src/llvm-late-gc-lowering.cpp:2675 run at /home/gbaraldi/julia4/usr/include/llvm/IR/PassManagerInternal.h:91 which means it was crashing inside of late-gc-lowering, so the first thing I did was ran julia and the same test with LLVM_ASSERTIONS=1 and FORCE_ASSERTIONS=1 to see if LLVM complained about a malformed module, and both were fine. Next step was trying to get the failing code out for inspection. Easiest way is to do `export JULIA_LLVM_ARGS="--print-before=LateLowerGCFrame --print-module-scope"` and pipe the output to a file. The file is huge, but since it's a crash in LLVM we know that the last thing is what we want, and that gave me the IR I wanted. To verify that this is failing I did `make -C src install-analysis-deps` to install the LLVM machinery (opt...). That gets put in the `tools` directory of a julia build. Then I checked if this crashed outside of julia by doing `./opt -load-pass-plugin=../lib/libjulia-codegen.dylib --passes=LateLowerGCFrame -S test.ll -o tmp3.ll `. This is run from inside the tools dir so your paths might vary (the -S is so LLVM doesn't generate bitcode) and my code did crash, however it was over 500 lines of IR which makes it harder to debug and to write a test. Next step then is to minimize the crash by doing [`llvm-reduce`](https://llvm.org/docs/CommandGuide/llvm-reduce.html) over it (it's basically creduce but optimized for LLVM IR) which gave me a 2 line reproducer (in this case apparently just having the insertelement was enough for the pass to fail). One thing to be wary is that llvm-reduce will usually make very weird code, so it might be useful to modify the code slightly so it doesn't look odd (it will have unreachable basic-blocks and such). After the cleanup fixing the bug here wasn't interesting but this doesn't apply generally. And also always transform your reduced IR into a test to put in llvmpasses. (cherry picked from commit 906d348)
Fixes #58229 (LLVM JITLink stack overflow issue) I tried submitting this promise/future implementation upstream (llvm/llvm-project@main...vtjnash:llvm-project:jn/cowait-jit) so that I would not need to duplicate nearly as much code here to fix this bug, but upstream is currently opposed to fixing this bug and instead insists it is preferable for each downstream project to implement this fix themselves adding extra maintenance burden for us for now. Sigh. (cherry picked from commit 00351da)
We observe an abort on Windows on Revise master CI, where a free'd handle is passed to jl_close_uv. The root cause is that uv_fseventscb_file called uvfinalize earlier, but did not set the handle to NULL, so when the actual finalizer ran later, it would see corrupted state. (cherry picked from commit b45b429)
Store full method interference relationship graph in interferences field
of Method to avoid expensive morespecific calls during dispatch. This
provides significant performance improvements:
- Replace method comparisons with precomputed interference lookup.
- Optimize ml_matches minmax computation using interference lookups.
- Optimize sort_mlmatches for large return sets by iterating over
interferences instead of all matching methods.
- Add method_morespecific_via_interferences in both C and Julia.
This representation may exclude some edges that are implied by
transitivity since sort_mlmatches will ensure the correct result by
following strong edges. Ambiguous edges are guaranteed to be checkable
without recursion.
Also fix a variety of bugs along the way:
- Builtins signature would cause them to try to discard all other
methods during `sort_mlmatches`.
- Some ambiguities were over-estimated, which now are improved upon.
- Setting lim==-1 now gives the same limited list of methods as lim>0,
since that is actually faster now than attempting to give the
unsorted list. This provides a better fix to #53814 than #57837 and
fixes #58766.
- Reverts recent METHOD_SIG_LATEST_HAS_NOTMORESPECIFIC attempt (though
not the whole commit), since I found a significant problem with any
usage of that bit during testing: it only tracks methods that
intersect with a target, but new methods do not necessarily intersect
with any existing target.
This provides a decent performance improvement to `methods` calls, which
implies a decent speed up to package loading also (e.g. ModelingToolkit
loads in about 4 seconds instead of 5 seconds).
(cherry picked from commit 59a7bb3)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Backported PRs:
Need manual backport:
Contains multiple commits, manual intervention needed:
Non-merged PRs with backport label:
--trace-compile#58535transcode: prevent Windows sysimage invalidation #58038@nospecializeforstring_index_err#57604maxthreadidfromThreads#57490