Recently I have been re-organizing and re-compiling all third-party dependencies of Nebula Graph, an open-source distributed graph database. And I have come across two interesting issues and would like to share with you.

Flex Segmentation Fault — — Segmentation fault (core dumped)

Segmentation fault happened upon compiling Flex：

make[2]: Entering directory '/home/dutor/flex-2.6.4/src'
./stage1flex   -o stage1scan.c ./scan.l
make[2]: *** [Makefile:1696: stage1scan.c] Segmentation fault (core dumped)

Check coredump with gdb:

Core was generated by `./stage1flex -o stage1scan.c ./scan.l'.
Program terminated with signal SIGSEGV, Segmentation fault.
#0  flexinit (argc=4, argv=0x7ffd25bea718) at main.c:976
976             action_array[0] = '\0';
(gdb) disas
Dump of assembler code for function flexinit:
   0x0000556c1b1ae040 <+0>:     push   %r15
   0x0000556c1b1ae042 <+2>:     lea    0x140fd(%rip),%rax        # 0x556c1b1c2146
   ...
   0x0000556c1b1ae20f <+463>:   callq  0x556c1b1af460 <allocate_array> # Allocate buffer
   ...
=> 0x0000556c1b1ae24f <+527>:   movb   $0x0,(%rax) # Write to buffer[0], failed due to illegal address
   ...
(gdb) disas allocate_array
Dump of assembler code for function allocate_array:
   0x0000556c1b1af460 <+0>:     sub    $0x8,%rsp
   0x0000556c1b1af464 <+4>:     mov    %rsi,%rdx
   0x0000556c1b1af467 <+7>:     xor    %eax,%eax
   0x0000556c1b1af469 <+9>:     movslq %edi,%rsi
   0x0000556c1b1af46c <+12>:    xor    %edi,%edi
   0x0000556c1b1af46e <+14>:    callq  0x556c1b19a100 <reallocarray@plt> # Allocate buffer
   0x0000556c1b1af473 <+19>:    test   %eax,%eax  # Check if the result pointer is NULL
   0x0000556c1b1af475 <+21>:    je     0x556c1b1af47e <allocate_array+30># Jump to error handler if NULL 
   0x0000556c1b1af477 <+23>:    cltq   # Extend eax to rax, truncated
   0x0000556c1b1af479 <+25>:    add    $0x8,%rsp
   0x0000556c1b1af47d <+29>:    retq
   ...
End of assembler dump.

We can see from the assembly code above that the issue was caused by the allocate_array function. reallocarray returned a pointer, which should be saved in the 64-bit register rax. However, allocate_arraycalled reallocarray and returned the 32-bit register eax. Meanwhile it used instruction cltq to extend eaxto rax.

The possible reason could be that the prototype of reallocarray that allocate_array saw was different than the real prototype.

When looking at the compiing log, I did find such a warning, like implicit declaration of function _reallocarray'_.

This issue can be resolved by adding CFLAGS=-D_GNU_SOURCE at the configure stage.

Please note that this issue is not supposed to appear every time. However, enabling compiling/link option -pie and core parameter kernel.randomize_va_space helps produce the issue.

Takeaways:

1. The return type of an implicit declarative function is int in C
2. Pay attention to compiler warnings with -Wall and -Wextra enabled. Better enable -Werror under development mode

GCC Illegal Instruction — — internal compiler error: Illegal instruction

A while ago I’ve received feedback from Nebula Graph users that they encountered a compiler error: illeggal instruction. See the details in this pull request: https://github.com/vesoft-inc/nebula/issues/978.

Below is the error message:

Scanning dependencies of target base_obj_gch
[ 0%] Generating Base.h.gch
In file included from /opt/nebula/gcc/include/c++/8.2.0/chrono:40,
from /opt/nebula/gcc/include/c++/8.2.0/thread:38,
from /home/zkzy/nebula/nebula/src/common/base/Base.h:15:
/opt/nebula/gcc/include/c++/8.2.0/limits:1599:7: internal compiler error: Illegal instruction
min() _GLIBCXX_USE_NOEXCEPT { return FLT_MIN; }
^~~
0xb48c5f crash_signal
../.././gcc/toplev.c:325
Please submit a full bug report,
with preprocessed source if appropriate.

Since it’s an internal compiler error, my assumption would be that an illegal instruction was encountered in g++ itself. To locate the specific illegal instruction set and the component it belongs to, we need to reproduce the error.

Luckily, the code snippet below can do the magic:

#include <thread>
int main() 
{
    return 0;
}

Illegal instruction is sure to trigger SIGIL. Since g++ acts only as the entrance of the compiler, the real compiler is cc1plus.

We can use gdb to perform the compiling process and catch the illegal instruction on spot:

$ gdb --args /opt/nebula/gcc/bin/g++ test.cpp
gdb> set follow-fork-mode child
gdb> run
Starting program: /opt/nebula/gcc/bin/g++ test.cpp
[New process 31172]
process 31172 is executing new program: /opt/nebula/gcc/libexec/gcc/x86_64-pc-linux-gnu/8.2.0/cc1plus
Thread 2.1 "cc1plus" received signal SIGILL, Illegal instruction.
[Switching to process 31172]
0x00000000013aa0fb in __gmpn_mul_1 ()
gdb> disas
...
0x00000000013aa086 <+38>: mulx (%rsi),%r10,%r8
...

Bingo！

mulx belongs to BMI2 instruction set and the CPU of the machine in error doesn't support this instruction set.

After a thorough investigation, I found that it was GMP, which is one of GCC’s dependencies, that introduced this instruction set. By default, GMP would detect the CPU type of the host machine at the configure stage to make use of the most recent instruction sets, which improves performance while sacrificing the portability of the binary.

To solve the issue, you can try to override two files in the GMP source tree, i.e. config.guess and config.sub with configfsf.guess and configfsf.sub respectively before**configure** .

Conclusion

• GCC won’t adopt new instruction set due to compatibility issue by default.
• To balance compatibility and performance, you need to do some extra work. For example, select and bind a specific instance for gllibc when it is running.

Finally, if you are interested in compiling the source code of Nebula Graph, please refer to the documentation here.