ABI differs when generated by different architecture compilers

[alexcrichton]

For example, given this file:

// foo.rs
#![crate_type = "lib"]   
pub fn foo() {}          

We get compiles looking like this:

# Using an i686-unknown-linux-gnu host rustc
$ rustc foo.rs --emit obj
$ nm -g foo.o
00000000 T _ZN3foo20hcfee31b285b97994eaaE

# Using an x86_64-unknown-linux-gnu host rustc
$ rustc foo.rs --emit obj
$ nm -g foo.o
0000000000000000 T _ZN3foo20hf85c4a4e2bec9df7eaaE

The relevant parts being:

_ZN3foo20hcfee31b285b97994eaaE
_ZN3foo20hf85c4a4e2bec9df7eaaE

Surprisingly enough, the symbols are different!

This doesn't have too many implications today, the major one I can think of is #29228 which means that libraries generated by an x86_64 compiler cannot be used as plugins to an i686 compiler. I don't think there's any fundamental reason this is happening today, it just sounds like a bug where our hashing is architecture dependent somewhere.

[alexcrichton]

I'll also note that the only place I know of where hashing is pointer-size-dependent is when hashing usize, so there's probably just one of those or a slice/string somewhere along the road which we're hashing (the length could probably just be a u32 in this case).

[sorear]

I found one of the places this happens a couple weeks ago as part of what I did to #10389:

https://github.com/sorear/vow-rs/blob/master/src/main.rs#L121

rust/src/librustc/middle/ty/util.rs

Line 544 in 540fd3a

hash!(inner.len());

[chordowl]

Triage: seems fixed on stable
With the same file as in the original report. I get the following on stable-{i686,x86_64}-pc-windows-gnu:

# Using an i686-pc-windows-gnu host rustc
$ rustc foo.rs --emit obj
$ nm -g foo.o
00000000 T __ZN3foo3foo17h83583f8ff17ec21aE
# Using a x86_64-pc-windows-gnu host rustc
$ rustc foo.rs --emit obj
$ nm -g foo.o
0000000000000000 T _ZN3foo3foo17h83583f8ff17ec21aE

The relevant parts are identical now, save for the usual additional underscore on i686:

__ZN3foo3foo17h83583f8ff17ec21aE
 _ZN3foo3foo17h83583f8ff17ec21aE

Meta

rustc 1.16.0 (30cf806ef 2017-03-10)
binary: rustc
commit-hash: 30cf806ef8881c41821fbd43e5cf3699c5290c16
commit-date: 2017-03-10
host: i686-pc-windows-gnu
release: 1.16.0
LLVM version: 3.9

rustc 1.16.0 (30cf806ef 2017-03-10)
binary: rustc
commit-hash: 30cf806ef8881c41821fbd43e5cf3699c5290c16
commit-date: 2017-03-10
host: x86_64-pc-windows-gnu
release: 1.16.0
LLVM version: 3.9

[arielb1]

I don't think that symbols depending on the target ABI is a problem - different target ABIs are, well, ABI-incompatible.