jl_typename_type->type is uninitialized?

[nolta]

I've been seeing a lot of random crashes lately, so i ran julia through valgrind:

$ valgrind --track-origins=yes ./usr/bin/julia-debug-basic 
==25701== Memcheck, a memory error detector
==25701== Copyright (C) 2002-2011, and GNU GPL'd, by Julian Seward et al.
==25701== Using Valgrind-3.7.0 and LibVEX; rerun with -h for copyright info
==25701== Command: ./usr/bin/julia-debug-basic
==25701== 
--25701-- WARNING: Serious error when reading debug info
--25701-- When reading debug info from /scratch/r/rbond/nolta/julia/usr/lib/libjulia-debug.so:
--25701-- DWARF line info appears to be corrupt - the section is too small
==25701== Conditional jump or move depends on uninitialised value(s)
==25701==    at 0x54FB398: emit_getfield(_jl_value_t*, _jl_sym_t*, jl_codectx_t*) (codegen.cpp:699)
==25701==    by 0x54FDE87: emit_known_call(_jl_value_t*, _jl_value_t**, unsigned long, jl_codectx_t*, llvm::Value**, llvm::Value**, _jl_value_t*) (codegen.cpp:1084)
==25701==    by 0x54FE50D: emit_call(_jl_value_t**, unsigned long, jl_codectx_t*, _jl_value_t*) (codegen.cpp:1150)
==25701==    by 0x5500358: emit_expr(_jl_value_t*, jl_codectx_t*, bool, bool) (codegen.cpp:1463)
==25701==    by 0x54FBC81: emit_known_call(_jl_value_t*, _jl_value_t**, unsigned long, jl_codectx_t*, llvm::Value**, llvm::Value**, _jl_value_t*) (codegen.cpp:795)
==25701==    by 0x54FE50D: emit_call(_jl_value_t**, unsigned long, jl_codectx_t*, _jl_value_t*) (codegen.cpp:1150)
==25701==    by 0x5500358: emit_expr(_jl_value_t*, jl_codectx_t*, bool, bool) (codegen.cpp:1463)
==25701==    by 0x5503C2C: emit_function(_jl_lambda_info_t*, llvm::Function*) (codegen.cpp:2014)
==25701==    by 0x54E9EEC: to_function(_jl_lambda_info_t*) (codegen.cpp:170)
==25701==    by 0x54EA340: jl_compile (codegen.cpp:230)
==25701==    by 0x54E7759: jl_trampoline (builtins.c:669)
==25701==    by 0x54DDB48: jl_apply (julia.h:901)
==25701==  Uninitialised value was created by a heap allocation
==25701==    at 0x4C27223: malloc (vg_replace_malloc.c:263)
==25701==    by 0x55328BE: add_page (gc.c:319)
==25701==    by 0x55329C3: pool_alloc (gc.c:346)
==25701==    by 0x55337D7: allocobj (gc.c:767)
==25701==    by 0x5523884: newobj (newobj_internal.h:6)
==25701==    by 0x5524EC2: jl_new_uninitialized_struct_type (alloc.c:514)
==25701==    by 0x54DBBB1: jl_init_types (jltypes.c:2247)
==25701==    by 0x552810E: julia_init (init.c:159)
==25701==    by 0x402648: main (repl.c:292)
==25701== 
==25701== Conditional jump or move depends on uninitialised value(s)
==25701==    at 0x54DE044: mtcache_hash_lookup (gf.c:153)
==25701==    by 0x54DE57F: jl_method_table_assoc_exact (gf.c:266)
==25701==    by 0x54E1406: jl_apply_generic (gf.c:1309)
==25701==    by 0x404ACF7: ???
==25701==    by 0x54E4D9E: jl_apply (julia.h:901)
==25701==    by 0x54E77AE: jl_trampoline (builtins.c:672)
==25701==    by 0x54DDB48: jl_apply (julia.h:901)
==25701==    by 0x54E15A4: jl_apply_generic (gf.c:1339)
==25701==    by 0x401A22: jl_apply (julia.h:901)
==25701==    by 0x40242E: true_main (repl.c:249)
==25701==    by 0x5528816: julia_trampoline (init.c:290)
==25701==    by 0x402660: main (repl.c:293)
==25701==  Uninitialised value was created by a heap allocation
==25701==    at 0x4C27223: malloc (vg_replace_malloc.c:263)
==25701==    by 0x55328BE: add_page (gc.c:319)
==25701==    by 0x55329C3: pool_alloc (gc.c:346)
==25701==    by 0x55337D7: allocobj (gc.c:767)
==25701==    by 0x5523884: newobj (newobj_internal.h:6)
==25701==    by 0x5524EC2: jl_new_uninitialized_struct_type (alloc.c:514)
==25701==    by 0x54DBBB1: jl_init_types (jltypes.c:2247)
==25701==    by 0x552810E: julia_init (init.c:159)
==25701==    by 0x402648: main (repl.c:292)
==25701== 
               _
   _       _ _(_)_     |
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  A fresh approach to technical computing
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 0.0.0+97764241.r9a04
 _/ |\__'_|_|_|\__'_|  |  Commit 9a04865b89 (2012-09-27 12:23:07)
|__/                   |

[JeffBezanson]

As far as I can tell, jl_typename_type->type is initialized. This must be something else.

[JeffBezanson]

One thing that might help in cases like this is defining MEMDEBUG at the top of gc.c, which separately allocates all values so valgrind's origin info is better. Otherwise you get the allocation of the page containing the value.

[nolta]

Thanks, that fixed it. Not sure if this is still an error (it's compiled w/ MEMDEBUG):

$ valgrind --track-origins=yes ./usr/bin/julia-debug-basic 
==4388== Memcheck, a memory error detector
==4388== Copyright (C) 2002-2011, and GNU GPL'd, by Julian Seward et al.
==4388== Using Valgrind-3.7.0 and LibVEX; rerun with -h for copyright info
==4388== Command: ./usr/bin/julia-debug-basic
==4388== 
--4388-- WARNING: Serious error when reading debug info
--4388-- When reading debug info from /scratch/r/rbond/nolta/julia/usr/lib/libjulia-debug.so:
--4388-- DWARF line info appears to be corrupt - the section is too small
               _
   _       _ _(_)_     |
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  A fresh approach to technical computing
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 0.0.0+97770869.rf934.dirty
 _/ |\__'_|_|_|\__'_|  |  Commit f934b6b1b2 (2012-09-27 14:13:35)*
|__/                   |

julia> x=[1,2,3]
==4388== Conditional jump or move depends on uninitialised value(s)
==4388==    at 0x5533317: gc_mark_all (gc.c:597)
==4388==    by 0x55334D0: gc_mark (gc.c:651)
==4388==    by 0x5533618: jl_gc_collect (gc.c:705)
==4388==    by 0x55325BB: alloc_big (gc.c:225)
==4388==    by 0x5533721: alloc_2w (gc.c:773)
==4388==    by 0x550495A: jl_new_box (codegen.cpp:2096)
==4388==    by 0x404054F: ???
==4388==    by 0x403FF37: ???
==4388==    by 0x403FD9E: ???
==4388==    by 0x54DDB78: jl_apply (julia.h:901)
==4388==    by 0x54E15D4: jl_apply_generic (gf.c:1339)
==4388==    by 0x403B509: ???
==4388==  Uninitialised value was created by a heap allocation
==4388==    at 0x4C27223: malloc (vg_replace_malloc.c:263)
==4388==    by 0x5532609: alloc_big (gc.c:231)
==4388==    by 0x5533711: allocobj (gc.c:763)
==4388==    by 0x55238B4: newobj (newobj_internal.h:6)
==4388==    by 0x5524EF2: jl_new_uninitialized_struct_type (alloc.c:514)
==4388==    by 0x54DBC3E: jl_init_types (jltypes.c:2258)
==4388==    by 0x552813E: julia_init (init.c:159)
==4388==    by 0x402648: main (repl.c:292)
==4388== 
3-element Int64 Array:
 1
 2
 3

[vtjnash]

i think that is still an error:

jl_value_t *ctor_factory;
jl_value_t *instance;  // for singletons

are appear to also be uninitialized for jl_tag_kind