Cilk bkend segfaults: Some remaining problems with shared library handling

[rrnewton]

In JITRuntime.hs we use withDL, but I'm seeing some suspecious segfaults even when the whole test suite claims to PASS. And valgrind says:

$ DEBUG=2 valgrind ./dist/build/test-accelerate-cpu-cilk/test-accelerate-cpu-cilk -t p18f
==8168== Memcheck, a memory error detector
==8168== Copyright (C) 2002-2012, and GNU GPL'd, by Julian Seward et al.
==8168== Using Valgrind-3.8.1 and LibVEX; rerun with -h for copyright info
==8168== Command: ./dist/build/test-accelerate-cpu-cilk/test-accelerate-cpu-cilk -t p18f
==8168== 
  [Note: passing through options to test-framework]: -t p18f
 [!] Testing backend: Cilk
[main] First checking that all requested tests can be found within 'allProgs'...
[main] Yep, all 70 tests are there.
run test 39 oneDim_p18f::
 ! Responding to env Var: DEBUG=2
 [dbg] engaging optional typecheck pass, AST size 7
COMPILETIME_phase0: 0.116065s
COMPILETIME_phase1: 0.253395s
 [dbg] engaging optional typecheck pass, AST size 9
 [dbg] engaging optional typecheck pass, AST size 9
 [dbg] engaging optional typecheck pass, AST size 11
 [dbg] engaging optional typecheck pass, AST size 11
 [dbg] engaging optional typecheck pass, AST size 11
 [dbg] engaging optional typecheck pass, AST size 12
 [dbg] engaging optional typecheck pass, AST size 13
 [dbg] engaging optional typecheck pass, AST size 13
!! Victory, inlineCheap: inlining reference to aLt0
!! Victory, inlineCheap: inlining reference to tmp_1
 [dbg] engaging optional typecheck pass, AST size 15
 [dbg] engaging optional typecheck pass, AST size 15
!! Victory: deadArrays, start an unraveling by eliminating: aLt0
!! Victory: deadArrays, start an unraveling by eliminating: tmp_1
 [dbg] engaging optional typecheck pass, AST size 9
[DBG OneDimensionalize]Computing 1D size for ndims 1 input exp EVr tmp_0_shape
 [dbg] engaging optional typecheck pass, AST size 15
 [dbg] engaging optional typecheck pass, AST size 15
COMPILETIME_phase2: 0.140196s
COMPILETIME_phase3: 0.002773s
COMPILETIME_emit: 0.421638s
[JIT] Invoking C compiler on: .genC_CilkParallel_oneDimp18f.c
[JIT] Found ICC (/l/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/bin/intel64/icc) Using it.
[JIT]   Compiling with: /l/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/bin/intel64/icc-fast -ww13397 -vec-report2 -g -lcilkrts -std=c99 -shared -fPIC .genC_CilkParallel_oneDimp18f.c -o .genC_CilkParallel_oneDimp18f.so
COMPILETIME_C: 0.302799s
.genC_CilkParallel_oneDimp18f.c(141): (col. 5) remark: loop was not vectorized: existence of vector dependence
.genC_CilkParallel_oneDimp18f.c(28): (col. 5) remark: LOOP WAS VECTORIZED
.genC_CilkParallel_oneDimp18f.c(14): (col. 1) remark: FUNCTION WAS VECTORIZED
.genC_CilkParallel_oneDimp18f.c(14): (col. 1) remark: FUNCTION WAS VECTORIZED
.genC_CilkParallel_oneDimp18f.c(14): (col. 1) remark: FUNCTION WAS VECTORIZED
.genC_CilkParallel_oneDimp18f.c(14): (col. 1) remark: FUNCTION WAS VECTORIZED
[JIT]: Working directory: /nfs/nfs3/home/rrnewton/working_copies/accelerate/accelerate-backend-kit/icc-opencl

 [dbg] Top lvl scalar binding: gensym_0 = 7
 [dbg] Top lvl scalar binding: tmp_0_shape = 7
 [dbg] Top lvl scalar binding: tmp_0_size = 7
==8168== Thread 3:
==8168== Invalid read of size 8
==8168==    at 0x7A0CC88: __intel_sse2_strlen (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A217A0: tbb::internal::dynamic_link(char const*, tbb::internal::dynamic_link_descriptor const*, unsigned long, unsigned long, void**) (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A21503: tbb::internal::rml::tbb_factory::open() (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A21CE4: __cilkrts_rml_startup (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A22158: __cilkrts_start_workers (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A25F9D: __cilkrts_init_internal (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A1550C: __cilkrts_bind_thread (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7A11213: void cilk_for_root<unsigned long, void (*)(void*, unsigned long, unsigned long)>(void (*)(void*, unsigned long, unsigned long), void*, unsigned long, int) (in /nfs/nfs5-insecure/home/insecure-ro/software/rhel6_x86_64/intel/composer_xe_2013_sp1.1.106/composer_xe_2013_sp1.1.106/compiler/lib/intel64/libcilkrts.so.5)
==8168==    by 0x7802350: MainProg (.genC_CilkParallel_oneDimp18f.c:27)
==8168==    by 0x4140E3: ??? (in /nfs/nfs3/home/rrnewton/working_copies/accelerate/accelerate-backend-kit/icc-opencl/dist/build/test-accelerate-cpu-cilk/test-accelerate-cpu-cilk)
==8168==  Address 0x760c5a8 is 104 bytes inside a block of size 107 alloc'd
==8168==    at 0x4A069EE: malloc (vg_replace_malloc.c:270)
==8168==    by 0x35A1A05BCD: open_path (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A08464: _dl_map_object (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A0C2D1: openaux (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A0E1B5: _dl_catch_error (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A0C9B4: _dl_map_object_deps (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A12AA0: dl_open_worker (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A0E1B5: _dl_catch_error (in /lib64/ld-2.12.so)
==8168==    by 0x35A1A124F9: _dl_open (in /lib64/ld-2.12.so)
==8168==    by 0x35A2200F65: dlopen_doit (in /lib64/libdl-2.12.so)
==8168==    by 0x35A1A0E1B5: _dl_catch_error (in /lib64/ld-2.12.so)
==8168==    by 0x35A220129B: _dlerror_run (in /lib64/libdl-2.12.so)
==8168== 
SELFTIMED: 0.110766s
[JIT] Finished executing dynamically loaded Acc computation!
[JIT] Fetched result ptr: tmp_0 = 0x0000000007648fc0 and size 7
 [dbg] !! Copying back 56 bytes (array len 7) into haskell heap!
 [dbg] !! Copying to 0x000000000511c320 from 0x0000000007648fc0
[JIT] Destroying args record: 0x0000000007611820
[JIT] Destroying results record: 0x0000000007611860
  : [OK]

         Test Cases  Total      
 Passed  1           1          
 Failed  0           0          
 Total   1           1          
==8168== Jump to the invalid address stated on the next line
==8168==    at 0x7A19DE0: ???
==8168==    by 0x35A2607A68: start_thread (in /lib64/libpthread-2.12.so)
==8168==    by 0x66016FF: ???
==8168==  Address 0x7a19de0 is not stack'd, malloc'd or (recently) free'd
==8168== 
==8168== 
==8168== Process terminating with default action of signal 11 (SIGSEGV)
==8168==  Access not within mapped region at address 0x7A19DE0
==8168==    at 0x7A19DE0: ???
==8168==    by 0x35A2607A68: start_thread (in /lib64/libpthread-2.12.so)
==8168==    by 0x66016FF: ???
==8168==  If you believe this happened as a result of a stack
==8168==  overflow in your program's main thread (unlikely but
==8168==  possible), you can try to increase the size of the
==8168==  main thread stack using the --main-stacksize= flag.
==8168==  The main thread stack size used in this run was 10485760.
==8168== 
==8168== HEAP SUMMARY:
==8168==     in use at exit: 227,288 bytes in 85 blocks
==8168==   total heap usage: 4,782 allocs, 4,697 frees, 6,578,989 bytes allocated
==8168== 
==8168== LEAK SUMMARY:
==8168==    definitely lost: 392 bytes in 4 blocks
==8168==    indirectly lost: 33 bytes in 3 blocks
==8168==      possibly lost: 576 bytes in 2 blocks
==8168==    still reachable: 226,287 bytes in 76 blocks
==8168==         suppressed: 0 bytes in 0 blocks
==8168== Rerun with --leak-check=full to see details of leaked memory
==8168== 
==8168== For counts of detected and suppressed errors, rerun with: -v
==8168== ERROR SUMMARY: 2 errors from 2 contexts (suppressed: 8 from 8)
Killed