SIGSEGV in getrf on OS X 10.8

[mlubin]

Moving the issue over from jump-dev/JuMP.jl#50, I get a crash in dgetrf on OS X 10.8:

* thread #1: tid = 0x5bda4, 0x00000001063e3e20 libopenblas.dylib`.L11 + 48, queue = 'com.apple.main-thread, stop reason = signal SIGSEGV
    frame #0: 0x00000001063e3e20 libopenblas.dylib`.L11 + 48
libopenblas.dylib`.L11 + 48:
-> 0x1063e3e20:  movapd 48(%rsi), %xmm7
   0x1063e3e25:  andpd  %xmm15, %xmm7
   0x1063e3e2a:  maxpd  %xmm7, %xmm3
   0x1063e3e2e:  prefetcht0 1472(%rsi)
(lldb) bt
* thread #1: tid = 0x5bda4, 0x00000001063e3e20 libopenblas.dylib`.L11 + 48, queue = 'com.apple.main-thread, stop reason = signal SIGSEGV
    frame #0: 0x00000001063e3e20 libopenblas.dylib`.L11 + 48
    frame #1: 0x0000000104df9397 libopenblas.dylib`sgetf2_k(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fb87040, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 2647 at getf2_k.c:97
    frame #2: 0x0000000104dffb11 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fb87040, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 801 at getrf_parallel.c:423
    frame #3: 0x0000000104dffbd1 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fb9ad40, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 993 at getrf_parallel.c:435
    frame #4: 0x0000000104dffbd1 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fbaea40, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 993 at getrf_parallel.c:435
    frame #5: 0x0000000104dffbd1 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fbc2740, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 993 at getrf_parallel.c:435
    frame #6: 0x0000000104dffbd1 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fbd6440, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 993 at getrf_parallel.c:435
    frame #7: 0x0000000104dffbd1 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x00007fff5fbea140, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 993 at getrf_parallel.c:435
    frame #8: 0x0000000104dffbd1 libopenblas.dylib`dgetrf_parallel(args=0x00007fff5fbea2b0, range_m=0x0000000000000000, range_n=0x0000000000000000, sa=0x000000010fd00000, sb=0x000000010fe00000, myid=0) + 993 at getrf_parallel.c:435
    frame #9: 0x0000000104829b94 libopenblas.dylib`dgetrf_(M=0x000000010d9b9d2c, N=0x000000010d9b9d2c, a=0x000000010d9bb4c0, ldA=0x000000010d9b9d2c, ipiv=0x000000010d9bb720, Info=0x00007fff5fbea40c) + 660 at getrf.c:108
    frame #10: 0x000000010f2b2c99 libCoinUtils.3.dylib`CoinDenseFactorization::factor(this=0x0000000100000000) + 201 at CoinDenseFactorization.cpp:202

Unfortunately I don't have a standalone example. Openblas (v0.2.8) is built by julia using clang with DEBUG=1.

$ system_profiler SPHardwareDataType
Hardware:

    Hardware Overview:

      Model Name: MacBook Pro
      Model Identifier: MacBookPro10,2
      Processor Name: Intel Core i5
      Processor Speed: 2.6 GHz
      Number of Processors: 1
      Total Number of Cores: 2
      L2 Cache (per Core): 256 KB
      L3 Cache: 3 MB
      Memory: 8 GB
      Boot ROM Version: MBP102.0106.B03
      SMC Version (system): 2.6f59
      Serial Number (system): ***
      Hardware UUID: ***
      Sudden Motion Sensor:
          State: Enabled

I can try compiling with gcc also if that would be helpful.

[xianyi]

Could you try to build OpenBLAS with USE_OPENMP=1?

[mlubin]

On 10.8, the default compiler is clang, so USE_OPENMP=1 fails. If I force it to build with gcc, I get the following compile error:

gcc-4.8 -O2 -DEXPRECISION -m128bit-long-double -fopenmp -Wall -m64 -DF_INTERFACE_GFORT -fPIC  -DSMP_SERVER -DUSE_OPENMP -DMAX_CPU_NUMBER=4 -DASMNAME=_sgemm_otcopy -DASMFNAME=_sgemm_otcopy_ -DNAME=sgemm_otcopy_ -DCNAME=sgemm_otcopy -DCHAR_NAME=\"sgemm_otcopy_\" -DCHAR_CNAME=\"sgemm_otcopy\" -DNO_AFFINITY -I.. -UDOUBLE  -UCOMPLEX -c -UDOUBLE -UCOMPLEX ../kernel/x86_64/../generic/gemm_tcopy_8.c -o sgemm_otcopy.o
../kernel/x86_64/sgemm_kernel_8x8_sandy.S:241:no such instruction: `vzeroupper'
../kernel/x86_64/sgemm_kernel_8x8_sandy.S:243:no such instruction: `vmovlps %xmm0, 48(%rsp)'
../kernel/x86_64/sgemm_kernel_8x8_sandy.S:286:no such instruction: `vxorps %ymm15, %ymm15,%ymm15'
../kernel/x86_64/sgemm_kernel_8x8_sandy.S:288:no such instruction: `vxorps %ymm14, %ymm14,%ymm14'
../kernel/x86_64/sgemm_kernel_8x8_sandy.S:290:no such instruction: `vxorps %ymm13, %ymm13,%ymm13'
../kernel/x86_64/sgemm_kernel_8x8_sandy.S:292:no such instruction: `vxorps %ymm12, %ymm12,%ymm12'
...

$ gcc-4.8 --version
gcc-4.8 (GCC) 4.8.1

[mlubin]

I ran the build with: make CC=gcc-4.8 USE_OPENMP=1

[xianyi]

Please use this script to replace /usr/bin/as
https://gist.github.com/xianyi/2957847

[mlubin]

It crashes as well, even when I set OMP_NUM_THREADS=1

* thread #1: tid = 0x217da1, 0x00000001044c66f5 libopenblas_sandybridgep-r0.2.8.dylib`.L11 + 69, queue = 'com.apple.main-thread, stop reason = signal SIGSEGV
    frame #0: 0x00000001044c66f5 libopenblas_sandybridgep-r0.2.8.dylib`.L11 + 69
libopenblas_sandybridgep-r0.2.8.dylib`.L11 + 69:
-> 0x1044c66f5:  movapd 64(%rsi), %xmm4
   0x1044c66fa:  andpd  %xmm15, %xmm4
   0x1044c66ff:  maxpd  %xmm4, %xmm0
   0x1044c6703:  movapd 80(%rsi), %xmm5
(lldb) bt
error: unable to find CIE at 0x00000018 for cie_id = 0x00000004 for entry at 0x00000018.
* thread #1: tid = 0x217da1, 0x00000001044c66f5 libopenblas_sandybridgep-r0.2.8.dylib`.L11 + 69, queue = 'com.apple.main-thread, stop reason = signal SIGSEGV
    frame #0: 0x00000001044c66f5 libopenblas_sandybridgep-r0.2.8.dylib`.L11 + 69
    frame #1: 0x00000001046c158f libopenblas_sandybridgep-r0.2.8.dylib`sgetf2_k(args=<unavailable>, range_m=<unavailable>, range_n=<unavailable>, sa=<unavailable>, sb=<unavailable>, myid=<unavailable>) + 2027 at getf2_k.c:97

[mlubin]

This was probably a 32-/64-bit ABI mismatch. Closing