Fix SegFault on LLVM 3.8.0 #16503
Fix SegFault on LLVM 3.8.0 #16503
Conversation
|
We have a few days left to nominate patches for 3.8.1, just need to email the release manager and llvm-commits |
| @@ -728,6 +728,8 @@ $(eval $(call LLVM_PATCH,llvm-3.8.0_winshlib)) | |||
| $(eval $(call LLVM_PATCH,llvm-3.8.0_threads)) | |||
| # fix replutil test on unix | |||
| $(eval $(call LLVM_PATCH,llvm-D17165-D18583)) | |||
| # Segfault for aggregate load | |||
| $(eval $(call LLVM_PATCH,llvm-D17326_unpack_load)) | |||
There was a problem hiding this comment.
touches the same file as llvm-D14260.patch, should have a dependency to avoid a race condition
This is due to a LLVM bug which sets the wrong alignment on a scalar load unpacked from an aligned aggregate load.
I left a comment on http://reviews.llvm.org/D17326#436371 . How to contact the release manager? |
|
email llvm-commits, cc the code owner and tom stellard http://lists.llvm.org/pipermail/llvm-dev/2016-May/099946.html |
|
Confirmed fixes the segfault in the subarray test on win64. edit: also fixes the bitarray test if I apply this patch at d6df2a0 |
|
Was David Majnemer the code owner there? May be good to remind them that you don't have commit access (how many patches have you had upstream now? what's their commit access nomination process?) |
He's the one I find for InstCombine.
Following the thread you linked, I'm waiting for Tom's reply and will ask him to help committing. |
|
You may want to include @Keno in the conversation since he's got LLVM commit access. |
|
Backporting is generally done by the release manager, I don't even know how to do it, even though I probably have the permissions. |
This patch is backported from upstream, see JuliaLang/julia#16503 for more information.
This is due to a LLVM bug which sets the wrong alignment on a scalar load
unpacked from an aligned aggregate load.
This is very similar to the issue in #15417 (comment) although the transformation is somehow visible in
code_llvmand not onlycode_nativeso more tests are needed on windows (all tests passes on linux for me with this patch).The bug was introduced by http://reviews.llvm.org/D17158 which was applied on 3.8 branch in llvm-mirror/llvm@7a0ec46. This is fixed on the LLVM master in llvm-mirror/llvm@7aec751 5 days after but was not backported....