[BUG] clang++ optimizer crash (segmentation fault) with neon intrinsics

[SimonSuckut]

Description

I encountered a bug in the optimizer of clang++ which leads to a segmentation fault when working with neon intrinsic. This only happens when I run a release build in android studio. I played a bit with the clang++ command and found out that this happens due to the optimization level (-Oz) of the release build.

I provided a small test case that shows the issue:
https://github.com/SimonSuckut/ndk_clang-_bug

with build_debug.sh (contains the clang++ command of the debug build) the compilation works perfectly fine, with build.sh (contains the clang++ command of the release build) the compile process ends with an segmentation fault:

1. <eof> parser at end of file 2. Code generation 3. Running pass 'Function Pass Manager' on module '/home/simon/AndroidStudioProjects/llvm_bug/llvm_bug.cpp'. 4. Running pass 'ARM Instruction Selection' on function '@_GLOBAL__sub_I_llvm_bug.cpp' #0 0x0000000002e6e2b5 llvm::sys::PrintStackTrace(llvm::raw_ostream&) (/home/simon/Dokumente/Android/Sdk/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang+++0x2e6e2b5) #1 0x0000000002e6e63e llvm::sys::RunSignalHandlers() (/home/simon/Dokumente/Android/Sdk/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang+++0x2e6e63e) #2 0x0000000002e6e4e8 (/home/simon/Dokumente/Android/Sdk/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang+++0x2e6e4e8) #3 0x00007f5b9b5453c0 __restore_rt (/lib/x86_64-linux-gnu/libpthread.so.0+0x153c0) #4 0x000000000266a480 llvm::TargetLowering::DAGCombinerInfo::CombineTo(llvm::SDNode*, llvm::ArrayRef<llvm::SDValue>, bool) (/home/simon/Dokumente/Android/Sdk/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang+++0x266a480) #5 0x0000000004bc6968 clang++: error: unable to execute command: Segmentation fault (core dumped) clang++: error: clang frontend command failed due to signal (use -v to see invocation) Android (5900059 based on r365631c) clang version 9.0.8 (https://android.googlesource.com/toolchain/llvm-project 207d7abc1a2abf3ef8d4301736d6a7ebc224a290) (based on LLVM 9.0.8svn) Target: armv7-none-linux-android19 Thread model: posix

Environment Details

• NDK Version: 21.3.6528147
• Build system: clang++
• Host OS: Ubuntu 20.04
• ABI: armeabi-v7a
• NDK API level: 19

[DanAlbert]

Do you know if this crashed on r20?

[DanAlbert]

(I'll check if you don't know, but lmk if you do)

[SimonSuckut]

It seem to compile successfully on r20 (tested with 20.0.5594570).

[DanAlbert]

Thanks, triaging for r21d then.

[stephenhines]

This is a crash on invalid code actually. If I run with a more recent clang, I get the following diagnostic:

srhines@ringworld:__:aosp-master-plus-llvm$ /disk/llvm/build/bin/clang++ --target=armv7-none-linux-androideabi19 -c -Oz llvm_bug.cpp 
llvm_bug.cpp:38:31: error: argument value 2 is outside the valid range [0, 1]
    result.smid_data.val[0] = vmlaq_lane_f32(result.smid_data.val[0], m1.smid_data.val[3], vget_high_f32(m2.smid_data.val[0]), 2);
                              ^                                                                                                ~
/disk/llvm/build/lib/clang/11.0.0/include/arm_neon.h:14507:34: note: expanded from macro 'vmlaq_lane_f32'
  __ret_46 = __s0_46 + __s1_46 * splatq_lane_f32(__s2_46, __p3_46); \
                                 ^                        ~~~~~~~
/disk/llvm/build/lib/clang/11.0.0/include/arm_neon.h:703:25: note: expanded from macro 'splatq_lane_f32'
  __ret = (float32x4_t) __builtin_neon_splatq_lane_v((int8x8_t)__s0, __p1, 9); \
                        ^                                            ~~~~
1 error generated.

We can backport the fix for this, but the actual code as written uses an invalid index. I'll continue to track that down, but I wanted to post this in case you feel blocked on writing your own code with the current NDK.

[SimonSuckut]

Oh thanks, thats really nice to know.
There is actually a typo in there. Only an index in the range [0, 1] makes sense here.
But when I chagne to the correct index, which is 1 in this case, the crash is still there.

[stephenhines]

Reduced and filed https://bugs.llvm.org/show_bug.cgi?id=46615 upstream with the ARM maintainers.

[stephenhines]

This needs more than a prebuilt update, @DanAlbert . I do need to cherry-pick this for r21e if we're going to take it there. This is only in our toolchain-testing branch (and r22) right now.

[stephenhines]

Actually, it's in r22. It's only missing from r21e.

[stephenhines]

https://android-review.googlesource.com/q/topic:%22fix_1303%22+(status:open%20OR%20status:merged) cherry-picks the fix. We then need to check-in prebuilts so that you can get them for the LTS NDK.