problems with interpreter backtraces

[vtjnash]

Seems to be two, possibly disjoint problems observed here:

1. "repeated 387 more times" and the immediately subsequent "repeated 166 more times" are precisely the same stacktrace
2. the interpreter duplicates f28442 on the callstack

$ ./julia --compile=min -q
julia> begin
           f28442(c) = g28442(c + 1)
           g28442(c) = c > 2000 ? (return backtrace()) : f28442(c + 1)
           bt = f28442(1)
           Base.show_backtrace(stdout, bt)
       end

Stacktrace:
 [1] g28442(::Int64) at REPL[0]:3
 [2] f28442(::Int64) at REPL[0]:2
 ... (the last 2 lines are repeated 442 more times)
 [887] g28442(::Int64) at REPL[0]:3
 [888] f28442(::Int64) at REPL[0]:2 (repeats 2 times)
 [889] g28442(::Int64) at REPL[0]:3
 ... (the last 2 lines are repeated 387 more times)
 [1664] f28442(::Int64) at REPL[0]:2 (repeats 2 times)
 [1665] g28442(::Int64) at REPL[0]:3
 ... (the last 2 lines are repeated 166 more times)
 [1998] f28442(::Int64) at REPL[0]:2
 [1999] top-level scope at REPL[0]:4
 [2000] eval(::Module, ::Any) at ./boot.jl:319
 [2001] eval_user_input(::Any, ::REPL.REPLBackend) at REPL.jl:85
 [2002] macro expansion at REPL.jl:117 [inlined]

[c42f]

the interpreter duplicates f28442 on the callstack

Actually this appears to be the opposite: g28442 is "missing" from the callstack (actually present as the C frame Lenter_interpreter_frame_start_val). Here's the problematic part of the stacktrace with C frames turned on:

g28442(::Int64) at REPL[30]:3
jl_fptr_interpret_call at interpreter.c:861
jl_apply_generic at gf.c:2209
do_call at interpreter.c:323
eval_value at interpreter.c:411
eval_stmt_value at interpreter.c:362 [inlined]
eval_body at interpreter.c:757
jl_interpret_call_callback at interpreter.c:853

f28442(::Int64) at REPL[30]:2
jl_fptr_interpret_call at interpreter.c:861
jl_apply_generic at gf.c:2209
do_call at interpreter.c:323
eval_value at interpreter.c:411
eval_stmt_value at interpreter.c:362 [inlined]
eval_body at interpreter.c:757
jl_interpret_call_callback at interpreter.c:853

Lenter_interpreter_frame_start_val at libjulia.so.1:?
jl_fptr_interpret_call at interpreter.c:861
jl_apply_generic at gf.c:2209
do_call at interpreter.c:323
eval_value at interpreter.c:411
eval_stmt_value at interpreter.c:362 [inlined]
eval_body at interpreter.c:757
jl_interpret_call_callback at interpreter.c:853

f28442(::Int64) at REPL[30]:2
jl_fptr_interpret_call at interpreter.c:861
jl_apply_generic at gf.c:2209
do_call at interpreter.c:323
eval_value at interpreter.c:411
eval_stmt_value at interpreter.c:362 [inlined]
eval_body at interpreter.c:757
jl_interpret_call_callback at interpreter.c:853

Looking at the code, the reason for this is they way that jl_backtrace_from_here captures the stack trace in chunks; when there's an interpreter frame on the boundary of a chunk there's not enough space to capture it inside jl_unw_stepn, but instead of falling back to the next chunk the frame gets captured as a plain C frame instead.