Commit b3bc56a
Make late gc lower handle insertelement of alloca use. (#58637)
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)1 parent 9aeee1b commit b3bc56a
File tree
2 files changed
+15
-1
lines changed- src
- test/llvmpasses
2 files changed
+15
-1
lines changed| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
1115 | 1115 | | |
1116 | 1116 | | |
1117 | 1117 | | |
1118 | | - | |
| 1118 | + | |
1119 | 1119 | | |
1120 | 1120 | | |
1121 | 1121 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
226 | 226 | | |
227 | 227 | | |
228 | 228 | | |
| 229 | + | |
| 230 | + | |
| 231 | + | |
| 232 | + | |
| 233 | + | |
| 234 | + | |
| 235 | + | |
| 236 | + | |
| 237 | + | |
| 238 | + | |
| 239 | + | |
| 240 | + | |
| 241 | + | |
| 242 | + | |
229 | 243 | | |
230 | 244 | | |
231 | 245 | | |
| |||
0 commit comments