Closed ellishg closed 4 months ago
Thanks for the reproducer @ellishg.
Is this a bug in hotcoldsplit? I'm skeptical because @llvm.dbg.assign() and !DIAssignID !12 are in two different blocks, so I'm wondering if this is valid IR
In this case it is a "bug" (/incomplete implementation). DIAssignID
uses and attachments should be unique between functions but it's ok and expected for them to get spread across different blocks within a function.
@llvm/issue-subscribers-debuginfo
Author: Ellis Hoag (ellishg)
While investigating a clang crash https://github.com/apple/llvm-project/issues/8741, I was able to find a reduced IR that fails to verify after
hotcoldsplit
.input.ll
The crash goes away if I add
-hotcoldsplit-threshold=200
, which tells me that splitting_foo()
is causing the verifier to fail.Is this a bug in
hotcoldsplit
? I'm skeptical because@llvm.dbg.assign()
and!DIAssignID !12
are in two different blocks, so I'm wondering if this is valid IR. If this is not valid, then some prior pass must have broken it while I was usingllvm-reduce
to find a minimal reproducer.CC @SLTozer becase you seem to have recent commits with
DIAssignID
.