https://llvm.org/bugs/show_bug.cgi?id=27273
Bug ID: 27273
Summary: Missing Verifier check for debug info intrinsics that
precede their SSA operands
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: Core LLVM classes
Assignee: unassignedb...@nondot.org
Reporter: dexonsm...@apple.com
CC: apra...@apple.com, llvm-bugs@lists.llvm.org
Classification: Unclassified
r265631, reverted in r265637, uncovered a bug in -reassociate that breaks the
order of @llvm.dbg.value instructions compared to the SSA values they
reference.
The current Verifier check doesn't look through the
MetadataAsValue/LocalAsMetadata operands that debug info intrinsics use, such
as `metadata i32 %x` in:
--
call void @llvm.dbg.value(metadata i32 %x, ...)
%x = add i32 0, %arg
--
As a result, we need a hack in llvm::MapMetadata to return `!{}` instead of
`null` when RF_IgnoreMissingLocals is turned off to prevent a crash in
llvm::RemapInstruction.
We should add a verifier check; fix the bugs; write an upgrade to drop invalid
instructions and commit the verifier check; and then fix llvm::MapMetadata to
ignore RF_IgnoreMissingLocals (as r265631 tried to do).
There's a more complete write-up of the issue on llvm-commits (in the reply to
r265631):
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160404/345617.html
For now I'm going to push forward with a modified version of r265631 that
documents what's going on. I'll add a comment with the revision.
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
