diff options
| author | Bill Wendling <isanbard@gmail.com> | 2010-03-11 01:18:13 +0000 |
|---|---|---|
| committer | Bill Wendling <isanbard@gmail.com> | 2010-03-11 01:18:13 +0000 |
| commit | e8e79524d2af67e1dcb518a74ba5598146fdc4d2 (patch) | |
| tree | 3d2fa551aa749b5e7b9be45bf117d672813ba64d /llvm/lib/CodeGen/BranchFolding.h | |
| parent | f5e81aeba50eab64a947274018108326308e011b (diff) | |
| download | bcm5719-llvm-e8e79524d2af67e1dcb518a74ba5598146fdc4d2.tar.gz bcm5719-llvm-e8e79524d2af67e1dcb518a74ba5598146fdc4d2.zip | |
When outputing a non-lazy pointer for a stub, we may need to fill in the value
for the NLP because the object it's pointing to may be internal to the file.
This seems counter-intuitive, but bear with me. When we place the LSDA into the
TEXT section, the type info pointers need to be indirect and pc-rel. We
accomplish this by using NLPs. However, sometimes the types are local to the
file. GCC gets around this by not using a NLP in this case, but a "regular"
indirection like this:
GCC_except_tbl:
.long Lfoo-.
__ZTIA: @ This is local
...
Lfoo:
.long __ZTIA
LLVM prefers NLPs on Darwin. In fact, it's more optimal for load performance to
use them.
llvm-svn: 98218
Diffstat (limited to 'llvm/lib/CodeGen/BranchFolding.h')
0 files changed, 0 insertions, 0 deletions

