Support -Zfunction-sections

[bjorn3]

This puts every function and data object in their own section. This allows the linker to omit unused functions and data objects with --gc-sections.

On linux this shrinks a hello world binary without optimizations (neither sysroot nor binary) from 17MB to 13MB. It shrinks a hello world binary with only sysroot optimizations from 14MB to 13MB. For comparison cg_llvm produces a 3.5MB debug mode hello world binary with an optimized sysroot. Cg_clif produces a 10MB debug mode hello world binary without an optimized sysroot.

Fixed #117

[antoyo]

Thanks!

[bjorn3]

While it helps, it helps not nearly as much as I would have hoped. Even with optimizations the executable is bigger than with cg_clif which doesn't do optimizations.

[bjorn3]

---- slice::test_array_chunks_count stdout ----
---- slice::test_array_chunks_count stderr ----
thread 'main' panicked at 'attempt to create unaligned or null slice', /home/runner/work/rustc_codegen_gcc/rustc_codegen_gcc/build_sysroot/sysroot_src/library/core/src/slice/raw.rs:141:9
stack backtrace:
   0: rust_begin_unwind
   1: core::panicking::panic_fmt
   2: core::panicking::panic
   3: core::slice::raw::debug_check_data_len::rt_check::<[i32; 2]>
   4: <core::slice::raw::debug_check_data_len::rt_check<[i32; 2]> as core::ops::function::FnOnce<(*const [i32; 2],)>>::call_once
   5: core::intrinsics::const_eval_select::<(*const [i32; 2],), core::slice::raw::debug_check_data_len::noop<[i32; 2]>, core::slice::raw::debug_check_data_len::rt_check<[i32; 2]>, ()>
   6: core::slice::raw::debug_check_data_len::<[i32; 2]>
   7: core::slice::raw::from_raw_parts::<[i32; 2]>
   8: <[i32]>::as_chunks_unchecked::<2>
   9: <[i32]>::as_chunks::<2>
  10: <core::slice::iter::ArrayChunks<i32, 2>>::new
  11: <[i32]>::array_chunks::<2>
  12: coretests::slice::test_array_chunks_count
  13: coretests::slice::test_array_chunks_count::{closure#0}
  14: <coretests::slice::test_array_chunks_count::{closure#0} as core::ops::function::FnOnce<()>>::call_once
  15: <fn() as core::ops::function::FnOnce<()>>::call_once
  16: <fn() as core::ops::function::FnOnce<()>>::call_once::{shim:vtable#0}
  17: <alloc::boxed::Box<dyn core::ops::function::FnOnce<(), Output = ()> + core::marker::Send> as core::ops::function::FnOnce<()>>::call_once
  18: test::run_test_in_spawned_subprocess
  19: test::test_main_static_abort
  20: coretests::main
  21: <fn() as core::ops::function::FnOnce<()>>::call_once
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.

I think this perturbed the test executable just enough to expose an existing bug.

[antoyo]

You already found a similar bug a few weeks ago. You had commented the test for unaligned slice.

I started working on this but wasn't able to fix the issue yet. Somehow, the global variable I needed to be aligned on 4 bytes is aligned on 1 byte because that's what rustc says. So, I still need to dig into this to figure out what's happening.

[antoyo]

Should be fixed in #121. The branches in my gcc fork are currently a mess. I'll fix the master branch if needed.

[antoyo]

I had forgotten to build the libgccjit artifact. Now it's done and the core tests pass.

[bjorn3]

Rebased to trigger CI.