Slowdown with newer GCC versions

[AlexDaniel]

Found it when trying to bisect this issue: #2945

This is the code that I used for the benchmark:

(^50_000).grep(*.is-prime).tail.say; say (now - INIT now)

Using bisectable I found this:

¦1cb266b^^^: «49999␤7.3227272␤»
¦1cb266b   : «49999␤9.4255653␤»

So the slowdown was caused by one of these commits:

• 1cb266b8da^^
• 1cb266b8da^
• 1cb266b8da

Which obviously makes no sense at all. So I tested it again and again and again, trying to make sure that it is not some kind of noise. It's not.

On irc @timo jokingly suggested that the compiler itself can be an issue (these rakudo+moar builds were produced long time ago, and it's possible that it was done with different compilers). So I checked with objdump:

objdump --full-contents --section=.comment moar

And indeed:

GCC: (Debian 4.9.2-10) 4.9.2
GCC: (Debian 6.3.0-18) 6.3.0 20170516

That is, rakudo+moar built with newer gcc has worse performance. And not just a little bit worse, it's around 0.78x as fast. Weird, huh?

<timotimo> we should first check to see if the time difference lives inside libtommath or moarvm
<timotimo> probably not in the jit, since that shouldn't be dependent on the compiler
<timotimo> though a difference in compiler could change when jitting happens exactly

Though I'd say that we should simply bisect gcc and see what we get.

[taboege]

FWIW, I tested both revisions compiled with the same compiler gcc (GCC) 8.3.0 and indeed don't notice a difference:

# This is Rakudo version 2017.09-102-g9af5607d5 built on MoarVM version 2017.09.1-46-g27d3d01e5
49999
8.0941886

# This is Rakudo version 2017.09-105-g1cb266b8d built on MoarVM version 2017.09.1-46-g27d3d01e5
49999
8.101132

[AlexDaniel]

Oh yes, it is worth noting that both compilers in my tests are relatively old. Maybe we should test against gcc 9.