make sys.so redistributable or generate it at install time

[JeffBezanson]

Better startup time is one of the main features of 0.3. For it to work, we need to generate sys.so on the user's machine. @loladiro and I decided the best way to do this in time for 0.3 is to generate the .so ourselves (not calling llc) after install.

[ihnorton]

What would be the disadvantage of using llc? or is the linker the problem?

[Keno]

We don't want to ship llc. Though I did realize the other day that e.g. on Windows, we don't have a system linker, so this plan might not work anyway.

[blakejohnson]

Why not just run llc with a conservative march parameter choice on the build machine?

[kmsquire]

Let me add a wrinkle: on our cluster, we have at least 2 or 3 different computer architectures (multiple Intel and AMD processor types). Software installation varies:

1. some software is installed on each machine
2. some software is installed for use on all machines via an NFS mount
3. some software is installed in individual directories

For 2 or 3, it would be nice if I could

• build julia once and run everywhere, or
• build a different .so file for different machines, and have the correct one used

[Keno]

For the 0.3 release the idea is indeed to target a reasonable minimum (probably to be debated what that is), and if you have anything less than that we just won't load the precompiled .so. Eventually when we switch to LLVM 3.5 we may to do the generation at install time as outlined in my PR.

[StefanKarpinski]

Yes! This was one of the last things holding up a 0.3 release candidate :-)

[ihnorton]

@staticfloat are the Mac nightlies using JULIA_CPU_TARGET=core2? see https://groups.google.com/d/msg/julia-users/xK8iE2nwc_4/Ki2MJyTAE0wJ

[staticfloat]

I missed the wrap-up of that discussion, awesome! There's one further thing that needs to be done, however. To protect against people distributing a sys.{so,dylib,dll}, this line was added to Makefile. Does the closing of this issue allow us to remove it so that our binary distributions now ship with that file?

[ihnorton]

Yes, conditional on JULIA_CPU_TARGET.

[staticfloat]

Alright, done in 245782c

[tkelman]

You sure we should distribute sys.dll on Windows? Doesn't it makes backtraces worse? #6557 (comment)

Edit: yes, yes it does

$ usr/bin/julia -e 'include("nonexistent")'
ERROR: could not open file D:\code\msys64\home\Tony\julia\nonexistent

$ rm usr/lib/julia/sys.dll
$ usr/bin/julia -e 'include("nonexistent")'
ERROR: could not open file D:\code\msys64\home\Tony\julia\nonexistent
 in include at boot.jl:244
 in include_from_node1 at loading.jl:128

[ihnorton]

Right, we won't distribute it on Windows.

[tkelman]

staticfloat/julia-nightly-packaging@4e5f38d should probably be reverted then

[ihnorton]

I run on a minor fork. Mostly cosmetics, but also disabling the auto-update functionality :)

[staticfloat]

Ah, I misread a reply, I thought I was supposed to do it on all OS'es. :P I'll go ahead and revert now then.