638149 - Need a way to disable -dead_strip on OS X

Status: RESOLVED FIXED

(Firefox Build System :: General, defect)

Description

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Attachment: add a --enable-dead-strip option to configure

The current OS X linker crashes if using both LTO and -dead_strip, so we need a way to disable -dead_strip.

Comment 1

[(not currently active) Ted Mielczarek]

I'd rather not add more configure options. Can you just detect that you're using LTO and not use dead strip in that case?

Comment 2

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Not sure if in a reliable way. There are many ways to enable it. With clang you can use -O4, -flto or -emit-llvm. GCC has -flto and I think -fwhopr.

For clang the most reliable is probably to test the type of file the compiler produces, but gcc uses native files with extra sections containing the IL. Should I just ignore newer GCCs on OS X for now?

Another option would be adding a test that crashes -dead_strip, but I never managed to get a small enough test to include in a configure.

I will also try xcode 4 to see if that linker is fixed. If it is, it might not be worth it to add the option.

Comment 3

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Unfortunately the linker in xcode4 still fails :-(

Comment 4

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Attachment: alternative patch

This patch adds a macro for checking for LTO and uses that to disable -dead_stip. It currently only handles LLVM based compilers, but could be extended if anyone is interested in trying gcc.

Comment 5

[(not currently active) Ted Mielczarek]

Comment on attachment 516506 [details] [diff] [review]
alternative patch

Funky, but sounds ok. Can we call it MOZ_LLVM_LTO_ENABLED or something like that? MOZ_CHECK_LLVM_LTO ? Also, is there an existing autoconf macro to get the assembly output from a compile? I guess it's hard to say since we're using such an ancient autoconf...

Comment 6

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

This is avoiding a bug in the linker, so I assume it could also hit gcc's lto. In which case they just have to extend this macro.

I can only do copy-and-paste programming with m4, so I don't know if there is an autoconf macro available, but I didn't notice one in the existing configure source code.

Comment 7

[(not currently active) Ted Mielczarek]

Okay. That's fine.

Comment 8

[(no longer active)]

I landed this on cedar <http://hg.mozilla.org/projects/cedar/rev/eb6edf77b7c8> but had to back it out immediately <http://hg.mozilla.org/projects/cedar/rev/224477bd4b3c> because it breaks the builds.

http://tinderbox.mozilla.org/showlog.cgi?log=Cedar/1301253927.1301254046.2775.gz

Comment 9

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Attachment: Patch including the aclocal that was missing

Sorry, the previous patch was missing a change to js/src/aclocal.m4

Comment 10

[Mitchell Field [:Mitch]]

http://hg.mozilla.org/projects/build-system/rev/f858cc85724a

Comment 11

[Mitchell Field [:Mitch]]

http://hg.mozilla.org/mozilla-central/rev/f858cc85724a