Hi Martin,
Thanks for the info. I didn't know that this has been discussed so
recently. Should have checked befor I posted.
Anyway, I built the latest lld from GIT, but it's slower than gold on my
machine. Just linking libjvm.so is 5.8s vs 3.8s.
On 2/19/19 5:23 PM, Martin Buchholz wrote:
https://openjdk.markmail.org/thread/q3layufiyjivu42c
On Tue, Feb 19, 2019 at 5:03 PM Ioi Lam <ioi....@oracle.com
<mailto:ioi....@oracle.com>> wrote:
For the impatient engineers ....
I did some measurement between the default GNU linker "ld" vs
"ld.gold".
I am trying to get the fastest rebuild time after I modify one cpp
file.
With gold, it's down to about 1/3 of the original time.
slowdebug (~220MB libjvm.so)
recompile 1 cpp file | relink libjvm.so only
ld: 33 s 25 s
gold 1 thread: 16 s 9 s
gold 8 threads: 13 s 6 s
fastdebug (~360MB libjvm.so)
recompile 1 cpp file | relink libjvm.so only
ld: 35 s 25 s
gold 1 thread: 18 s 10 s
gold 8 threads: 15 s 6 s
Question: do we want to add built-in support for gold into the JDK
makefiles?
Notes:
To choose gold, run configure with something like:
--with-extra-ldflags='-fuse-ld=gold
-Wl,--threads,--thread-count,8'
I essentially do a "make hotspot" and then move the libjvm.so into
a JDK
image, instead of doing a full JDK image build.
"make hotspot" makes a copy of libjvm.so (from
support/modules_libs/java.base/server/ to jdk/lib/server/). I
hacked the
Makefile to make a hard link instead to avoid the unnecessary I/O.
libjvm.so is built with --with-native-debug-symbols=internal to avoid
the expensive invocations of objcopy and strip.
My environment:
I am using gcc7.3.0 on Ubuntu 16.04.5 on a 5 year old Dell Precision
T7600 with dual socket Xeon E5-2665 @ 2.40GHz, 64GB RAM, Samsung
840 PRO
SSD. I suspect gold can run even faster, but my slow SSD is
holding it back.
ld version = GNU ld (GNU Binutils) 2.30
gold version = GNU gold (GNU Binutils for Ubuntu 2.26.1) 1.11
(These are just the versions available to me on my machine, not
necessarily the best)