Invalid `write(io::IO, s::String)` method

[yuyichao]

The method definition

write(io::IO, s::String) = unsafe_write(io, pointer(s), reinterpret(UInt, sizeof(s)))

is invalid since it is not guaranteed that the s will be kept alive when this function is inlined.

In particular, this cases @code_llvm show(IOBuffer(), 4) to be invalid

; Function show
; Location: show.jl
define void @jlsys_show_37268(%jl_value_t addrspace(10)* dereferenceable(48), i64) #0 {
top:
; Location: show.jl:353
  %2 = ashr i64 %1, 63
  %3 = add i64 %2, %1
  %4 = xor i64 %3, %2
  %.lobit = lshr i64 %1, 63
  %5 = trunc i64 %.lobit to i8
  %6 = call %jl_value_t addrspace(10)* @jlsys_dec_38356(i64 %4, i64 1, i8 %5)
  %7 = addrspacecast %jl_value_t addrspace(10)* %6 to %jl_value_t*
  %8 = ptrtoint %jl_value_t* %7 to i64
  %9 = add i64 %8, 8
  %10 = inttoptr i64 %9 to i8*
  %11 = bitcast %jl_value_t addrspace(10)* %6 to i64 addrspace(10)*
  %12 = load i64, i64 addrspace(10)* %11, align 8
  %13 = call i64 @jlsys_unsafe_write_37226(%jl_value_t addrspace(10)* %0, i8* %10, i64 %12)
  ret void
}

(Note that the return value of dec is passed to unsafe_write without being rooted).

This may show up as intermittent test failure with UndefVarError(t_...) since it can cause the "t_$i" to return the wrong value and messes up lowering. (Not sure if this shows up on CI but I've hit it multiple times in ~1 day of GC stress testing)

We can workaround this for now by adding @noinline although it'll be nice if we can fix this with some kind of GC preserve scope for a certain object.

[StefanKarpinski]

We should probably audit all uses of pointer in the code base.