Backtraces: some line numbers are still 0

[timholy]

@inline foo1(x) = foo2(x)
@inline foo2(x) = foo3(x)
@noinline foo3(x) = x ÷ 3

function bar(n)
    X = rand(1:10, n)
    Y = [foo1(x) for x in X]
end

bar(1)
@profile bar(10^7)
data = unique(Profile.fetch())
vsf = Base.StackTraces.lookup.(data)
for sfs in vsf
    for sf in sfs
        if sf.line == 0
            @show sf
        end
    end
end

Results:

sf =  in rand at random.jl:0 [inlined]
sf =  in rand at random.jl:0 [inlined]
sf =  in rand(::MersenneTwister, ::Base.Random.RangeGeneratorInt{Int64,UInt64}) at random.jl:0
sf =  in mt_setfull! at random.jl:0 [inlined]
sf =  in reserve_1 at random.jl:0 [inlined]
sf =  in rand_ui52_raw at random.jl:0 [inlined]
sf =  in rand at random.jl:0 [inlined]
sf =  in mt_pop! at random.jl:0 [inlined]
sf =  in rand_inbounds at random.jl:0 [inlined]
sf =  in rand_ui52_raw_inbounds at random.jl:0 [inlined]
sf =  in rand_ui52_raw at random.jl:0 [inlined]
sf =  in rand at random.jl:0 [inlined]
sf =  in rand(::MersenneTwister, ::Base.Random.RangeGeneratorInt{Int64,UInt64}) at random.jl:0
sf =  in rand(::MersenneTwister, ::Base.Random.RangeGeneratorInt{Int64,UInt64}) at random.jl:0
sf =  in gen_rand at random.jl:0 [inlined]
sf =  in reserve_1 at random.jl:0 [inlined]
sf =  in rand_ui52_raw at random.jl:0 [inlined]
sf =  in rand at random.jl:0 [inlined]
sf =  in foo3(::Int64) at testprof.jl:0

This is with a completely-fresh julia master,

julia> versioninfo()
Julia Version 0.5.0-dev+5188
Commit 36d36b0* (2016-07-07 03:28 UTC)
Platform Info:
  System: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Core(TM) i7-5500U CPU @ 2.40GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, broadwell)

[vtjnash]

Good find. Looks like this is broken:

                        else if (jl_is_long(arg))
                            inlined_func_lineno = jl_unbox_long(arg);

(lldb) p jl_(stmt_e)
Expr(:meta, :push_loc, :random.jl, :rand, 239)::Any
(lldb) p jl_(arg)
239
(lldb) p jl_(jl_typeof(arg))
Int32

[vtjnash]

There's also at least one place were we are emitting line number locations that we shouldn't be:

  %3 = call %jl_value_t*** @jl_get_ptls_states(), !dbg !28

And jl_gdblookup and disasm.cpp needs to be updated to show inlining.

[timholy]

Also worth noting that foo1 and foo2 don't seem to appear in the stackframes.

[vtjnash]

Note that even with my fixes, some line numbers are supposed to be 0 (to indicate that it is part of the function prologue).

The absence of foo1 / foo2 is a limitation of our current codegen and will be fixed when codegen is better at handling a linear IR.

[timholy]

Thanks a ton, @vtjnash!