Thanks!
You are right, that comment is not needed. Removed and updated in place
(since it was just a comment).
/Erik
On 2019-01-28 14:24, Claes Redestad wrote:
Looks great, thanks for fixing!
Nit: I'd drop this comment "Ubuntu does this when building binutils so
should be safe."
/Claes
On 2019-01-28 23:17, Erik Joelsson wrote:
Hello,
While investigating performance with different linkers and linker
configuration, we discovered that the devkit linker on Linux (both
bfd and gold) was a lot slower than the native linker on Ubuntu (23s
vs 12s for bfd to link libjvm.so on my machine). They are both the
same version (2.30), so this was puzzling.
It turned out to be a difference in how binutils was built. On Ubuntu
(and I assume most major Linux vendors), they explicitly set
CFLAGS="-O2" when running the build for binutils. I had naively
assumed that binutils set reasonable default CFLAGS in their build so
have not modified them like this before. I tried adding -O2 for our
devkit, and that created a kit with comparable linking performance on
par with the Ubuntu linker. Since this seems to be widely done, I
assume the change to be quite safe.
With this patch, I'm modifying the devkit creation makefiles to add
-O2 when building binutils. I'm also adding --enable-threads so that
the user may run the gold linker with --threads (as well as updating
a download link which now requires https). For Oracle users, the
devkit versions are bumped in jib-profiles.js so the new devkits are
used.
Bug: https://bugs.openjdk.java.net/browse/JDK-8217910
Webrev: http://cr.openjdk.java.net/~erikj/8217910/webrev.01/index.html
/Erik
