Output of an --emit dependent on other outputs present

[cceckman]

Passing rustc the --emit=bc flag and passing --emit=bc,link may produce different .bc files (that is, outputs). This is true of rust-core (see below), though was not the case for a simpler ("Hello, world!") test file.

Test script:

#!/bin/bash

mkdir bc
mkdir bclink
git clone git@github.com:thestinger/rust-core

rustc --emit=bc --out-dir=bc rust-core/core/lib.rs
rustc --emit=bc,link --out-dir=bclink rust-core/core/lib.rs

diff bc/core.bc bclink/core.bc

Diff output indicates binary files differ.

Tested with rustc 0.10-pre (0a181a8 2014-03-17 09:57:06 -0700), rust-core on dd0584450. rustc compiled from master on Arch Linux.

If this is, in fact, expected behavior- sorry for the trouble, unexpected to me.

[alexcrichton]

Can you provide a diff between the outputs? This is likely a known issue, but I would like to be sure.

[cceckman]

.bc is a binary format, so diff just gives "Binary files bc/core.bc and bclink/core.bc differ" as its output. On my machine, bclink/core.bc is 19K and bc/core.bc is 62K.

I'm not familiar enough with the format to extract much more data.

[cceckman]

I note the difference persists when the second line is

rustc  --emit=bc --emit=link --out-dir=bclink rust-core/core/lib.rs

(two separate --emit flags), though both produce both artifacts.

[alexcrichton]

Can you pass emit=ir and diff those outputs?

[cceckman]

With emit flags as --emit=ir and --emit=ir,link respectively, there is no difference in bc/core.ll and bclink/core.ll (according to diff.)

[pczarn]

It seems that every bytecode is invalid only when a rlib is produced.

It stopped working between rust-git-0.9.1301 and 1362 (shortly before 329fcd4 2014-02-23 15:37:05 -0800).

Going back thorugh the log, I found #12448.

I believe that --emit=bc,link --crate-type=rlib creates deflated bytecode. Perhaps *.rlib(*.bc) should be stored with extension .bc.flate or flate.bc.