Can't inline

[travisdowns]

Since PCG is oriented towards speed, I thought I'd mention that as far as performance goes, not being able to inline the core functions like pcg32_random_r has a significant impact on speed. I'm getting around this by simply doing #include "pcg-basic.c" - yes, including the implementation file!

A usual solution is to declare the key functions static inline in the header file, then you don't have to mess around with "out of line" versions.

The more fully featured C library is able to be inlined.

[lemire]

Can't LTO solve this? -flto? Gold linker, etc.

[travisdowns]

I don't think gold linker has anything to do with it (?), but yes LTO solves it. LTO is still a bit like PGO - some people know about it, few people use it. If you are writing a high performance library you don't want to rely on people using it. In a way LTO is harder to turn on than PGO even, because the old TU-at-a-time model hides certain types of bugs that are revealed when you enable LTO, because the compiler can now see that what you've been doing is UB after all, etc.

…

On Mon, Mar 4, 2019, 10:10 AM Daniel Lemire ***@***.***> wrote: Can't LTO solve this? -flto? Gold linker, etc. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#14 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACSswffZpNsyL45gj_hu3X3G5UMQkrH9ks5vTRs7gaJpZM4bbfKc> .