metadata linking bug or by design

Project Infrastructure LLVM Dev List Archives
1

Hi,

It seems that current linker "drops" updated llvm::Value * references
in metadata. Here goes an example. We have a two-line source file
a.c:

static int x;
void foo() { x = 1; }

and a similar b.c:

static int x;
void bar() { x = 1; }

They both have an internal variable x; one of them will be renamed (to
x1) during linking.

clang -emit-llvm -c -g a.c
clang -emit-llvm -c -g b.c
llvm-link -o t.bc a.o b.o

Before linking we can see the following lines in both bitcode files
from a.c and b.c, where the last element of !5 refers to the internal
variable x.

!llvm.dbg.sp = !{!1}
!5 = metadata !{i32 655412, i32 0, metadata !0, metadata !"x",
metadata !"x", metadata !"", metadata !2, i32 1, metadata !6, i32 1,
i32 1, i32* @x} ; [ DW_TAG_variable ]

However after linking the last element of !11 becomes null. I was
expecting that it would refer to the renamed x1.

!llvm.dbg.gv = !{!9, !11}
!9 = metadata !{i32 655412, i32 0, metadata !0, metadata !"x",
metadata !"x", metadata !"", metadata !3, i32 1, metadata !10, i32 1,
i32 1, i32* @x} ; [ DW_TAG_variable ]
!11 = metadata !{i32 655412, i32 0, metadata !1, metadata !"x",
metadata !"x", metadata !"", metadata !7, i32 1, metadata !12, i32 1,
i32 1, null} ; [ DW_TAG_variable ]

Is it intentional or a bug? Thanks.

- xi

2

It's a bug, probably a regression. Please file a bugzilla report.
Thanks,

3

I looked at an earlier version. NamedMDNodes were linked after global
values. Current trunk version links NamedMDNodes before that, though
the comment says otherwise: "We do this after linking GlobalValues so
that MDNodes that reference GlobalValues are properly remapped". see
ModuleLinker::run in lib/Linker/LinkModules.cpp.

If I move the call to linkNamedMDNodes in ModuleLinker::run to the
last step (just before return false), everything looks good. Not sure
if linkNamedMDNodes must be called earlier.

- xi

4

I made that change in response to Chris's change ValueMapper. Metadata mapping is being investigated as part of PR 10553 anyway. So Ill fix this particular regression you have found.
Thanks!

5

Great! Thanks!

I noticed this when I was creating my own metadata nodes, which
contained references to functions --- after linking some elements that
referred to functions magically became null.

- xi