While testing my own AA, I found it being queried for values that
didn't exist previously. Luckily there's a pass built into LLVM for
catching such things, "Debug-AA", and I used that to confirm that GVN
does this fairly frequently.
I've attached a test case that triggers the bug (reduced from
Stanford/Treesort, for whatever that's worth), this is on LLVM 2.7.
Should this be filed as a bug....?
Additionally it just seems.. /unfortunate/ that we have a pass
specifically designed to catch bugs like this, yet it's not being
used. Don't suppose this is something someone would feel motivated to
add this to some LLVM testing code?
Thanks for your time,
DebugAA-GVN.ll (2.95 KB)