Backtrace contains function names with MIR, but not on MSVC.

[eddyb]

From http://buildbot.rust-lang.org/builders/auto-win-msvc-64-opt-mir/builds/9/steps/test/logs/stdio:

thread '<main>' panicked at 'explicit panic', C:/bot/slave/auto-win-msvc-64-opt-mir/build/src/test/run-pass/backtrace.rs:23
stack backtrace:
   0:      0x7feeb726289 - sys::backtrace::Cleanup.Drop::drop::h9efecea147e63d6fECu
   1:      0x7feeb73da2e - panicking::Location<'a>::line::h0bafe181f48d2793YZz
   2:      0x7feeb73ce05 - panicking::set_hook::h43cadfa78fcc3418FWz
   3:      0x7feeb6d569c - sys_common::unwind::begin_unwind_inner::h944552b52ab4b9efoyt
   4:        0x13ffe1129 - __ImageBase
   5:        0x13ffe1089 - __ImageBase
   6:        0x13ffe347f - __ImageBase
   7:      0x7feeb73bba9 - rt::lang_start::h6c718470b597915boSz
   8:      0x7feeb7210d8 - sys_common::unwind::inner_try::hcf25cf847275bec0qvt
   9:      0x7feeb73b920 - rt::lang_start::h6c718470b597915boSz
  10:        0x13ffe7a02 - __tmainCRTStartup
                        at f:\dd\vctools\crt\crtw32\dllstuff\crtexe.c:626
  11:         0x771e652c - BaseThreadInitThunk

Instead of the actual function names, we get __ImageBase - which suggests internal functions don't get symbols on MSVC without debuginfo, but what part of debuginfo does it need? cc @michaelwoerister

The test also passes on linux w/ MIR - http://buildbot.rust-lang.org/builders/auto-linux-64-opt-mir/builds/10/steps/test/logs/stdio contains:

test [run-pass] run-pass/backtrace.rs ... ok

This is blocking gating on the auto-win-msvc-64-opt-mir. cc @alexcrichton

[retep998]

which suggests internal functions don't get symbols on MSVC without debuginfo

The function address -> symbol name information is in the .pdb produced by the linker. When Rust does not emit debug info it still tells the linker to emit a .pdb, but since the object files don't have any debug info, the only symbol names they have are for symbols that are publicly visible to other object files. Thus symbols internal to object files don't get resolved in backtraces..

[eddyb]

From @retep998 (aka WindowsBunny) on IRC:

<eddyb> WindowsBunny: so you're telling me if I don't have debuginfo for an instruction
<eddyb> it can't be resolved back to a function?
<WindowsBunny> yep
<WindowsBunny> eddyb: The names of functions internal to an object file aren't stored anywhere
<WindowsBunny> there is nothing for the linker to dig up to shove into the .pdb

So it seems we can't do anything about this other than to make it not use MIR on MSVC.

[eddyb]

The fix should be to add, at the top of the test:

#![feature(rustc_attrs)]

And before each of fn foo and fn double:

// FIXME #31005 MIR missing debuginfo currently.
#[cfg_attr(target_env = "msvc", rustc_no_mir)]

I can't test it myself as I have no MSVC setup currently and I don't want to waste buildbot time.

[alexcrichton]

Yeah this test just required debuginfo on MSVC for now, seems fine to disable MIR temporarily while we sort out that story for this test specifically.