On Monday, 4 June 2018 at 18:47:02 UTC, Dennis wrote:
On Monday, 4 June 2018 at 18:11:47 UTC, Steven Schveighoffer
wrote:
BTW, do you have cross-module inlining on? I wonder if that
makes a difference if you didn't have it on before. (I'm
somewhat speaking from ignorance, as I've heard people talk
about this limitation, but am not sure exactly when it's
enabled)
Cross-module inlining is never implicitly enabled in LDC. Not
having it enabled is definitely something that hurts performance
of LDC generated code.
Enable it with: `-enable-cross-module-inlining`. (May lead to
missing symbols during linking when using templates with __FILE__
arguments.)
I don't know much about this either. Clang has link-time
optimization with -O4, but looking at the --help of LDC it
turns out -O4 is equivalent to -O3 for D. Maybe someone else
knows?
Clang and LDC treat `-O4` as `-O3`.
To enable LTO (cross-module inlining and other big perf gains),
you have to use `-flto=full`or `-flto=thin`, but it'll only give
cross module inlining for modules that have been compiled with
it. Notably: default Phobos/druntime is _not_ compiled with LTO
enabled. LDC 1.9.0 release packages on Github ship with a second
set of LTO Phobos/druntime libs. With LDC 1.9.0, you can do
`-flto=<thin|full> -defaultlib=phobos2-ldc-lto,druntime-ldc-lto`,
for maximum druntime/Phobos inlining.
-Johan
