[Bug]: Miscompilation on armv7 with NDK r25c #1860
Assignees
Labels
Comments
|
I don't think we're going to be doing an r25d, so we'll get this "for free" when we update the compiler in r26. Leaving this open since we don't actually have a newer working compiler for r26 yet. |
|
Oh, and thanks for the detailed investigation :) That saves us a ton of time. |
|
Ok, we'll stay on r24 until r26 is released. I've also bisected which change between r24 and r25 caused the crash which lead to llvm/llvm-project@2caf85a (https://reviews.llvm.org/D112811). |
foxsen
pushed a commit
to android-la64/platform-ndk
that referenced
this issue
Oct 10, 2024
Includes a bunch of test `#include` fixes for the reason explained in the changelog. Also updates the ASan test. The ASan failure message has been rephrased so the old test no longer matched the death message correctly. As best as I can tell the test we have is a copy of the upstream test with all the supporting functions from the various sanitizer test support headers copied into it. I've replaced the guts of the test and copied the new required supporting functions. This also disables the TSan smoke test. It hangs indefinitely, and we don't have timeout support (`timeout` is not available on old devices). The test was already an xfail, so we can just disable it for now. The test will still be built, so we can be sure that it won't suffer build- time regressions. A few libc++ tests were disabled because the tests (either the failing asserts or whole tests) were removed upstream. Bug: android/ndk#1298 Bug: android/ndk#1677 Bug: android/ndk#1834 Bug: android/ndk#1860 Bug: android/ndk#1862 Test: checkbuild.py && run_tests.py Change-Id: Id467537487908cad99099bfeb2cfacf4787d875c
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Description
When we tried to switch from NDK r24 to r25c, our application has started to crash when compiled to armv7.
Unfortunately, I couldn't easily minimize our code to submit it here, but I managed to reproduce the same crash on armv7 Linux using upstream clang 15.0.7. Updating clang to 16.0.1 solved the crash, so I bisected the llvm-project repo to find that the necessary commit for the fix is llvm/llvm-project@8a09875. The fix itself is not armv7 specific, I assume it can cause problems on any platform.
I found a discussion here https://discourse.llvm.org/t/instcombine-struct-w-padding-aggregate-loads-stores-as-scalars/67035/5 that describes the issue and has small testcases that can reproduce the miscompilation, but on CUDA target. I don't have a small isolated testcase for Android, I hope the information I'm providing here is enough.
I tried to cherry-pick the commit onto llvm/llvm-project@282c83c which clang-r450784d1 is based on which is included in NDK r25c, there is a conflict in the test case that this fix modifies but the fix itself in llvm/lib/Analysis/InstructionSimplify.cpp is ok and works, it no longer crashes.
Please cherry-pick the llvm/llvm-project@8a09875 commit into the next NDK release.
Upstream bug
llvm/llvm-project#59156
Commit to cherry-pick
llvm/llvm-project@8a09875
Affected versions
r25
Canary version
No response
Host OS
Linux
Host OS version
Debian 11
Affected ABIs
armeabi-v7a
The text was updated successfully, but these errors were encountered: