@boomanaiden154 Fix for the existing reported issue has been merged. Can you see if this pass is causing any more breakage for you?

I started testing. I think I should have a decent chunk of results by tomorrow. I'll update the thread with results once I have them.

I am seeing some failures that look like miscompiles still. No crashes from what I've seen yet, and there were one or two before.

I'll try and spend some time debugging tomorrow, along with digging through the rest of the results. The one failure I've seen so far is in https://github.com/google/xls in partial_ops_test if anyone else wants to try and take a stab at it before I get to it.

https://gist.github.com/boomanaiden154/445779ab0e9bd122b9c8c15ce00db48f

Reproduces the issue:

aidengrossman@aidengrossman:/work/llvm-project/build$ ./bin/clang++ --target=x86_64 -Wno-override-module -O3 ./seed-renamed.ll -mllvm -enable-dfa-jump-thread=true -o /tmp/dfa_jump_thread
aidengrossman@aidengrossman:/work/llvm-project/build$ /tmp/dfa_jump_thread
aidengrossman@aidengrossman:/work/llvm-project/build$ echo $?
1
aidengrossman@aidengrossman:/work/llvm-project/build$ ./bin/clang++ --target=x86_64 -Wno-override-module -O3 ./seed-renamed.ll -mllvm -enable-dfa-jump-thread=false -o /tmp/normal
aidengrossman@aidengrossman:/work/llvm-project/build$ /tmp/normal
aidengrossman@aidengrossman:/work/llvm-project/build$ echo $?
0
aidengrossman@aidengrossman:/work/llvm-project/build$ git log --oneline --max-count=1
e71f243a8d0e (HEAD -> main, origin/main, origin/HEAD) [TableGen] Simplify MachineValueTypeSet::iterator::find_from_pos. NFC (#169227)

Sorry for the largeish reproducer. llvm-reduce kept reducing the test case in such a way that the interestingness test began to fail.

https://gist.github.com/boomanaiden154/445779ab0e9bd122b9c8c15ce00db48f

I tried to reduce it with https://github.com/dtcxzyw/llvm-ub-aware-interpreter, but it contains immediate UB:

UB triggered: load from poison pointer
Exited with immediate UB.
Stacktrace:
    %load = load ptr, ptr poison, align 8 at @hoge.107
    call void @hoge.107() #7 at @pluto
    call void @pluto(ptr noundef nonnull align 8 dereferenceable(24) %alloca1, ptr %load, i64 %load16) at @main

I think that's just how llvm-reduce does things. quux is the function that contains the bad transformation (not sure how things look after inlining).

Do you have the original unreduced IR file? I am going to try to reduce a UB-free testcase.

Do you have the original unreduced IR file? I am going to try to reduce a UB-free testcase.

Not one I can just post here given this comes from our internal build system..