Building julia with clang: tests fail

[ViralBShah]

Building julia with clang segfaults when building sys0.ji. Using OS X 10.7.3, with the clang that ships with XCode 4.3.1 (Apple clang version 3.1 (tags/Apple/clang-318.0.54) (based on LLVM 3.1svn))

(gdb) set args -b stage0.jl
(gdb) r
Starting program: /Users/viral/julia/julia -b stage0.jl
Reading symbols for shared libraries ++++++................................................................................................ done
Reading symbols for shared libraries . done

Program received signal EXC_BAD_ACCESS, Could not access memory.
Reason: KERN_INVALID_ADDRESS at address: 0x00000000000000a0
0x000000010064c7f5 in fl_raise (e=4331668007) at flisp.c:175
175     while (readstate != (fl_readstate_t*)fl_ctx->rdst) {
(gdb) bt
#0  0x000000010064c7f5 in fl_raise (e=4331668007) at flisp.c:175
#1  0x000000010064cd6d in type_error (fname=0x10067d72e "apply", expected=0x1006ab117 "function", got=1) at flisp.c:217
#2  0x0000000100654e3b in _applyn (n=0) at flisp.c:637
#3  0x00000001006550c7 in fl_applyn (n=0, f=1) at flisp.c:674
#4  0x000000010001423f in jl_init_frontend () at ast.c:95
#5  0x0000000100060c6a in julia_init (imageFile=0x0) at init.c:112
#6  0x0000000100002122 in main ()
Current language:  auto; currently minimal
(gdb) 

[vtjnash]

Here's my backtrace, based on Apple clang version 2.1 (tags/Apple/clang-163.7.1) (based on LLVM 3.0svn). (note that sys0.ji built ok, it's stage1 that failed)

(lldb) run -J sys0.ji stage1.jl                                                                                                                 Process 70160 launched: '/Users/jameson/Documents/no-backup/julia/julia' (x86_64)
Process 70160 stopped
* thread #1: tid = 0x1f03, 0x00000001006658c3 julia`apply_cl + 99 at flisp.c:958, stop reason = EXC_BAD_ACCESS (code=1, address=0x8)
    frame #0: 0x00000001006658c3 julia`apply_cl + 99 at flisp.c:958
   955   apply_cl_top:
   956      captured = 0;
   957      func = Stack[SP-nargs-1];
-> 958      ip = cv_data((cvalue_t*)ptr(fn_bcode(func)));
   959      assert(!ismanaged((uptrint_t)ip));
   960      while (SP+GET_INT32(ip) > N_STACK) {
   961          grow_stack();
(lldb) 

And on Ubuntu clang version 2.9 (tags/RELEASE_29/final) i386-pc-linux-gnu

(gdb) run -b stage0.jl
Starting program: /home/jameson/julia-clang/julia-debug-readline -b stage0.jl
[Thread debugging using libthread_db enabled]

Program received signal SIGSEGV, Segmentation fault.
0x0879499d in apply_cl (nargs=1) at flisp.c:958
958         ip = cv_data((cvalue_t*)ptr(fn_bcode(func)));
(gdb) bt
#0  0x0879499d in apply_cl (nargs=1) at flisp.c:958
#1  0x0878af91 in _applyn (n=1) at flisp.c:629
#2  0x0878b158 in fl_applyn (n=1, f=3078179338) at flisp.c:674
#3  0x08066479 in jl_expand (expr=0x89fe160) at ast.c:446
#4  0x08069d7f in jl_toplevel_eval_flex (e=0x89fe160, fast=0, plineno=0xbfffd8a8) at builtins.c:362
#5  0x0806a5f2 in jl_load_file_expr (fname=0x89a7fa8 "/home/jameson/julia-clang/jl/base.jl", ast=0x89ec570) at builtins.c:446
#6  0x0806aac2 in jl_load (fname=0x89e3f40 "base.jl") at builtins.c:502
#7  0x0806ab1c in jl_load_ (str=0x89e6950) at builtins.c:509
#8  0xb7600383 in ?? ()
#9  0x08060085 in jl_apply (f=0x89eb190, args=0xbfffda54, nargs=1) at ./julia.h:780
#10 0x08061b6d in jl_apply_generic (F=0x89b3330, args=0xbfffda54, nargs=1) at gf.c:1185
#11 0x080c6fc5 in jl_apply (f=0x89b3330, args=0xbfffda54, nargs=1) at ./julia.h:780
#12 0x080c6f5f in do_call (f=0x89b3330, args=0x89f2d64, nargs=1, locals=0x0, nl=0) at interpreter.c:57
#13 0x080c5c7d in eval (e=0x89ec560, locals=0x0, nl=0) at interpreter.c:115
#14 0x080c58fe in jl_interpret_toplevel_expr (e=0x89ec560) at interpreter.c:20
#15 0x08069f26 in jl_toplevel_eval_flex (e=0x89eb1c0, fast=0, plineno=0xbfffe1d8) at builtins.c:387
#16 0x0806a5f2 in jl_load_file_expr (fname=0x899b410 "/home/jameson/julia-clang/jl/sysimg.jl", ast=0x89eb1b0) at builtins.c:446
#17 0x0806aac2 in jl_load (fname=0x8965aa0 "sysimg.jl") at builtins.c:502
#18 0x0806ab1c in jl_load_ (str=0x89cdcc8) at builtins.c:509
#19 0xb7600383 in ?? ()
#20 0x08069755 in jl_apply (f=0x89eb190, args=0xbfffe404, nargs=1) at ./julia.h:780
#21 0x0806ca7f in jl_trampoline (F=0x89eb190, args=0xbfffe404, nargs=1) at builtins.c:895
#22 0x08060085 in jl_apply (f=0x89eb190, args=0xbfffe404, nargs=1) at ./julia.h:780
#23 0x08061b6d in jl_apply_generic (F=0x89b3330, args=0xbfffe404, nargs=1) at gf.c:1185
#24 0x080c6fc5 in jl_apply (f=0x89b3330, args=0xbfffe404, nargs=1) at ./julia.h:780
#25 0x080c6f5f in do_call (f=0x89b3330, args=0x89e3ef4, nargs=1, locals=0x0, nl=0) at interpreter.c:57
#26 0x080c5c7d in eval (e=0x89eb180, locals=0x0, nl=0) at interpreter.c:115
#27 0x080c58fe in jl_interpret_toplevel_expr (e=0x89eb180) at interpreter.c:20
#28 0x08069f26 in jl_toplevel_eval_flex (e=0x89b3b90, fast=0, plineno=0xbfffed78) at builtins.c:387
#29 0x08069b10 in jl_eval_module_expr (ex=0x89b3480, plineno=0xbfffed78) at builtins.c:264
#30 0x08069ce1 in jl_toplevel_eval_flex (e=0x89b3480, fast=0, plineno=0xbfffed78) at builtins.c:349
#31 0x0806a5f2 in jl_load_file_expr (fname=0x89ae0b8 "/home/jameson/julia-clang/jl/stage0.jl", ast=0x89b3470) at builtins.c:446
#32 0x0806aac2 in jl_load (fname=0xbffff2f0 "stage0.jl") at builtins.c:502
#33 0x080513de in exec_program ()
#34 0x08050fc2 in true_main ()
#35 0x080cd11e in julia_trampoline (argc=1, argv=0xbffff15c, pmain=0x8050e00 <true_main>) at init.c:203
#36 0x0805166a in main ()

[vtjnash]

Out of curiosity, I built src/flisp and src/support with gcc, then everything else with clang (on my mac). Somewhat surprisingly (to me at least :), the backtrace was the same in my previous post.

[JeffBezanson]

Now that is fascinating... fl_applyn is the problem, and it is a varargs function. The backtrace here actually shows 1 less argument to it than is present in the source code. Must be calling convention related.

[ViralBShah]

I bet the ubuntu stacktrace will be the same as what I posted or the mac one, if you build flisp and support in debug mode also. Use src/Makefile.debug for that.

-viral

On 18-Mar-2012, at 10:34 AM, Jameson Nash wrote:

Here's my backtrace, based on Apple clang version 2.1 (tags/Apple/clang-163.7.1) (based on LLVM 3.0svn). (note that sys0.ji built ok, it's stage1 that failed)

(lldb) run -J sys0.ji stage1.jl                                                                                                                 Process 70160 launched: '/Users/jameson/Documents/no-backup/julia/julia' (x86_64)
Process 70160 stopped
* thread #1: tid = 0x1f03, 0x00000001006658c3 julia`apply_cl + 99 at flisp.c:958, stop reason = EXC_BAD_ACCESS (code=1, address=0x8)
   frame #0: 0x00000001006658c3 julia`apply_cl + 99 at flisp.c:958
  955      apply_cl_top:
  956         captured = 0;
  957         func = Stack[SP-nargs-1];
-> 958        ip = cv_data((cvalue_t*)ptr(fn_bcode(func)));
  959         assert(!ismanaged((uptrint_t)ip));
  960         while (SP+GET_INT32(ip) > N_STACK) {
  961             grow_stack();
(lldb) 

And on Ubuntu clang version 2.9 (tags/RELEASE_29/final) i386-pc-linux-gnu

(gdb) run -b stage0.jl
Starting program: /home/jameson/julia-clang/julia-debug-readline -b stage0.jl
[Thread debugging using libthread_db enabled]

Program received signal SIGSEGV, Segmentation fault.
0x0879499d in apply_cl (nargs=1) at flisp.c:958
958         ip = cv_data((cvalue_t*)ptr(fn_bcode(func)));
(gdb) bt
#0  0x0879499d in apply_cl (nargs=1) at flisp.c:958
#1  0x0878af91 in _applyn (n=1) at flisp.c:629
#2  0x0878b158 in fl_applyn (n=1, f=3078179338) at flisp.c:674
#3  0x08066479 in jl_expand (expr=0x89fe160) at ast.c:446
#4  0x08069d7f in jl_toplevel_eval_flex (e=0x89fe160, fast=0, plineno=0xbfffd8a8) at builtins.c:362
#5  0x0806a5f2 in jl_load_file_expr (fname=0x89a7fa8 "/home/jameson/julia-clang/jl/base.jl", ast=0x89ec570) at builtins.c:446
#6  0x0806aac2 in jl_load (fname=0x89e3f40 "base.jl") at builtins.c:502
#7  0x0806ab1c in jl_load_ (str=0x89e6950) at builtins.c:509
#8  0xb7600383 in ?? ()
#9  0x08060085 in jl_apply (f=0x89eb190, args=0xbfffda54, nargs=1) at ./julia.h:780
#10 0x08061b6d in jl_apply_generic (F=0x89b3330, args=0xbfffda54, nargs=1) at gf.c:1185
#11 0x080c6fc5 in jl_apply (f=0x89b3330, args=0xbfffda54, nargs=1) at ./julia.h:780
#12 0x080c6f5f in do_call (f=0x89b3330, args=0x89f2d64, nargs=1, locals=0x0, nl=0) at interpreter.c:57
#13 0x080c5c7d in eval (e=0x89ec560, locals=0x0, nl=0) at interpreter.c:115
#14 0x080c58fe in jl_interpret_toplevel_expr (e=0x89ec560) at interpreter.c:20
#15 0x08069f26 in jl_toplevel_eval_flex (e=0x89eb1c0, fast=0, plineno=0xbfffe1d8) at builtins.c:387
#16 0x0806a5f2 in jl_load_file_expr (fname=0x899b410 "/home/jameson/julia-clang/jl/sysimg.jl", ast=0x89eb1b0) at builtins.c:446
#17 0x0806aac2 in jl_load (fname=0x8965aa0 "sysimg.jl") at builtins.c:502
#18 0x0806ab1c in jl_load_ (str=0x89cdcc8) at builtins.c:509
#19 0xb7600383 in ?? ()
#20 0x08069755 in jl_apply (f=0x89eb190, args=0xbfffe404, nargs=1) at ./julia.h:780
#21 0x0806ca7f in jl_trampoline (F=0x89eb190, args=0xbfffe404, nargs=1) at builtins.c:895
#22 0x08060085 in jl_apply (f=0x89eb190, args=0xbfffe404, nargs=1) at ./julia.h:780
#23 0x08061b6d in jl_apply_generic (F=0x89b3330, args=0xbfffe404, nargs=1) at gf.c:1185
#24 0x080c6fc5 in jl_apply (f=0x89b3330, args=0xbfffe404, nargs=1) at ./julia.h:780
#25 0x080c6f5f in do_call (f=0x89b3330, args=0x89e3ef4, nargs=1, locals=0x0, nl=0) at interpreter.c:57
#26 0x080c5c7d in eval (e=0x89eb180, locals=0x0, nl=0) at interpreter.c:115
#27 0x080c58fe in jl_interpret_toplevel_expr (e=0x89eb180) at interpreter.c:20
#28 0x08069f26 in jl_toplevel_eval_flex (e=0x89b3b90, fast=0, plineno=0xbfffed78) at builtins.c:387
#29 0x08069b10 in jl_eval_module_expr (ex=0x89b3480, plineno=0xbfffed78) at builtins.c:264
#30 0x08069ce1 in jl_toplevel_eval_flex (e=0x89b3480, fast=0, plineno=0xbfffed78) at builtins.c:349
#31 0x0806a5f2 in jl_load_file_expr (fname=0x89ae0b8 "/home/jameson/julia-clang/jl/stage0.jl", ast=0x89b3470) at builtins.c:446
#32 0x0806aac2 in jl_load (fname=0xbffff2f0 "stage0.jl") at builtins.c:502
#33 0x080513de in exec_program ()
#34 0x08050fc2 in true_main ()
#35 0x080cd11e in julia_trampoline (argc=1, argv=0xbffff15c, pmain=0x8050e00 <true_main>) at init.c:203
#36 0x0805166a in main ()

---

Reply to this email directly or view it on GitHub:
#602 (comment)

[ViralBShah]

We use -std=gnu99 with gcc in Make.inc, which suggests some kind of dependency on a GNU standard. As per clang docs, it also defaults to the gnu99 standard.

[vtjnash]

I don't know if it helps, but here's a more complete traceback from my mac:
(I think gdb is just ignoring varags in the function count.)

frame #0: 0x00000001006658c3 julia`apply_cl + 99 at flisp.c:958
   955   apply_cl_top:
   956      captured = 0;
   957      func = Stack[SP-nargs-1];
-> 958      ip = cv_data((cvalue_t*)ptr(fn_bcode(func)));
   959      assert(!ismanaged((uptrint_t)ip));
   960      while (SP+GET_INT32(ip) > N_STACK) {
   961          grow_stack();
(lldb) frame variable
(uint32_t) nargs = 1
(uint32_t) top_frame = 0
(uint32_t) n = 0
(uint32_t) captured = 0
(uint32_t) bp = 32767
(const uint8_t *) ip = 0x0000000100c305f0
(fixnum_t) s = 4307748336
(fixnum_t) hi = 4307864080
(uint32_t) i = 0
(symbol_t *) sym = 0x0000000100654e6a
(lldb) up
frame select -r 1
frame #1: 0x000000010065c9ee julia`_applyn + 206 at flisp.c:629
   626          v = ((builtin_t*)ptr(f))[3](&Stack[SP-n], n);
   627      }
   628      else if (isfunction(f)) {
-> 629          v = apply_cl(n);
   630      }
   631      else if (isbuiltin(f)) {
   632          value_t tab = symbol_value(builtins_table_sym);
(lldb) frame variable
(uint32_t) n = 1
(value_t) f = 4334490402
(uint32_t) saveSP = 2
(value_t) v = 15855067719180737544
(lldb) up
frame select -r 1
frame #2: 0x000000010065cd39 julia`fl_applyn + 649 at flisp.c:674
   671          value_t a = va_arg(ap, value_t);
   672          PUSH(a);
   673      }
-> 674      value_t v = _applyn(n);
   675      POPN(n+1);
   676      va_end(ap);
   677      return v;
(lldb) frame variable
(uint32_t) n = 1
(value_t) f = 4334490402
(va_list) ap = {
  [0] = {
    gp_offset = 24
    fp_offset = 48
    overflow_arg_area = 0x00007fff5fbfd0f0
    reg_save_area = 0x00007fff5fbfcfe0
  }
}
(size_t) i = 1
(value_t) v = 140734799794368
(lldb) up
frame select -r 1
frame #3: 0x0000000100014fdb julia`jl_expand + 107 at ast.c:439
   436  {
   437      int np = jl_gc_n_preserved_values();
   438      value_t e = fl_applyn(1, symbol_value(symbol("jl-expand-to-thunk")),
-> 439                            julia_to_scm(expr));
   440      jl_value_t *result;
   441      if (e == FL_T || e == FL_F || e == FL_EOF) {
   442          result = NULL;
(lldb) frame variable
(jl_value_t *) expr = 0x00000001013ef028
(int) np = 0
(value_t) e = 4295364389
(jl_value_t *) result = 0x00007fff5fbfd140
(lldb) 

[JeffBezanson]

D'oh. julia_to_scm, of course, allocates scheme objects, causing the object returned by symbol_value to move if a gc happens. The compilers were evaluating the arguments in different orders.

[ViralBShah]

Compiles fine with clang for me now, but goes into infinite loop in the hashing tests on OS X. Does anyone else see this? Surprisingly, CPU usage is 30% only.

I am reopening this - with an updated title.

[JeffBezanson]

Excellent; our infinite loops are efficient :)

[JeffBezanson]

I get everything passing in debug build, and the hashing test segfaults in release build :(

[ViralBShah]

The debug build completes all the tests successfully. The release build does not.