Always use LLD to link when using clang to build #8204
Conversation
All recent distributions of clang also include lld, and it links much faster than most other linkers.
|
This does seem to speed up the clang builds on CI (e.g. clang-fuzztest) and makes a noticeable difference on a cloudtop (mine goes from 30s to 21s for a full build). |
| endif() | ||
| endif() | ||
|
|
||
| if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND (LINUX OR WINDOWS)) |
There was a problem hiding this comment.
Maybe add a comment why not MacOS?
There was a problem hiding this comment.
(I am guessing the reason is that it is always used there..?)
There was a problem hiding this comment.
Actually the opposite; lld historically hasn't had good support for Mach-O. I'm not sure what the current status is but the version of Clang in the SDK on the bots here doesn't even support -fuse-ld=lld, so I figured it wasn't really worth investigating.
|
Have you measured with mold btw? (might not be worth the hassle, but it is said to be much faster) |
|
No, I thought about mold, but lld is fast enough (and mature) so it didn't seem worth it as a default. But given that GNU ld -> lld made a noticeable difference, maybe using mold might be worth trying for really fast debug cycles. Combined with either a high-parallelism cloudtop or reclient, turnaround could be really fast. |
|
I tried mold locally on the default build of Binaryen. I think it was maybe a little faster than lld, but hard to say, might have been in the noise. I'll have to try it on llvm sometime. |
kripken
left a comment
kripken
left a comment
There was a problem hiding this comment.
Interesting about mold - yeah, maybe it needs a really big project to show benefits.
2 participants
All recent distributions of clang also include lld, and it links much
faster than most other linkers.