unable to reproduce divcc/gzip hello world with divine-4.2.1+2019.02.06

[Kamil Dudka]

I tried to reproduce ​https://divine.fi.muni.cz/2018/dios/#building-gzip with the latest nightly snapshot of Divine but it does not seem to work any more.

Steps to reproduce:

1. download ​https://divine.fi.muni.cz/download/snapshots/divine-4.2.1+2019.02.06.tar.gz
2. build it using the instructions at ​https://divine.fi.muni.cz/manual.html#installation
3. run ninja divcc in the build directory (as divcc was not built by default)
4. follow the instructions at ​https://divine.fi.muni.cz/2018/dios/#building-gzip

Actual result:

loading bitcode … LART … RR … constants … done
booting … done
states per second: 1000                                                   
state count: 28
mips: 1.4

error found: yes
error trace: |
  (0) Non-zero exit code: 1
  (0) FAULT: exit called with non-zero value
  [0] FATAL: unknown in userspace
  FAULT: invalid pointer passed to __vm_obj_free
  FATAL: fault handler called recursively

active stack:
  - symbol: void __dios::FaultBase::handler<__dios::Upcall<__dios::fs::VFS<__dios::ProcessManager<__dios::Fault<__dios::Scheduler<__dios::config::Base> > > > > >(_VM_Fault, _VM_Frame*, void (*)())
    location: /dios/include/dios/sys/fault.hpp:87
  - symbol: __dios_fault
    location: /dios/src/libc/sys/fault.c:14
  - symbol: _exit
    location: /dios/src/libc/sys/start.cpp:76
  - symbol: _PDCLIB_Exit
    location: /dios/src/libc/_PDCLIB/_PDCLIB_Exit.c:20
  - symbol: _Exit
    location: /dios/src/libc/stdlib/_Exit.c:18
  - symbol: exit
    location: /dios/src/libc/stdlib/exit.c:53
  - symbol: do_exit
    location: gzip.c:2102
  - symbol: abort_gzip
    location: gzip.c:2140
  - symbol: gzip_error
    location: util.c:429
  - symbol: unzip
    location: unzip.c:137
a report was written to gzip.report

Note that ./gzip -f -d - < hello.gz, as well as gzip -f -d - < hello.gz, return zero exit code when invoked off Divine.

[mornfall]

Turns out that gzip works OK when compiled with -O0 (i.e. ./configure CC=divcc CFLAGS=-O0). I have tracked down the problem to the VM mishandling a 'switch' instruction on a 2-bit integer that the optimizer now emits. I won't have time to fix this tomorrow, but next week should be quite doable. For now, I have added the recipe as an automated testcase so that we do not regress on this in the future.

PS: gzip-1.8-divine.diff is no longer required with recent versions of divine

[mornfall]

I have checked in the fix. Also turns out that LLVM codegen can't deal (on x86_64 at least) with vaarg of 'long double'. On debug builds, it triggers an assertion, on release builds, it silently generates bad code. I have added a workaround that makes the generated code fail at runtime (meaning the gzip test should pass now also with assertions enabled).

[Kamil Dudka]

I can confirm that the problem does not occur with ​https://divine.fi.muni.cz/download/snapshots/divine-4.2.1+2019.02.10.tar.gz Thanks for the quick fix!