Trying Link-Time Optimizations On GCC 4.10

[phoronix]

Phoronix: Trying Link-Time Optimizations On GCC 4.10

This weekend I ran some quick and dirty link-time optimization (LTO) benchmarks from a GCC 4.10 compiler snapshot from earlier this month. Here's the results...

http://www.phoronix.com/vr.php?view=MTcyNzU

[log0]

I don't think you will see much performance increases in those benchmarks, as the hot paths are probably inlined already. It is something for more complex code, like firefox as an extreme example with ~6M LOC, where they probably won't waste time optimizing it manually.

[xeekei]

I don't think you will see much performance increases in those benchmarks, as the hot paths are probably inlined already. It is something for more complex code, like firefox as an extreme example with ~6M LOC, where they probably won't waste time optimizing it manually.

Man, imagine the build time for Firefox with these. Or LibreOffice.

[froyo]

Man, imagine the build time for Firefox with these. Or LibreOffice.

Well, you could turn LTO on only for release (or release candidate) builds.

[Akdor 1154]

Well, you could turn LTO on only for release (or release candidate) builds.

I currently compile the Dolphin emulator with -O3 and LTO using gcc4.9 and the increase in compilation time is noticeable but not horrible..

[Zan Lynx]

Man, imagine the build time for Firefox with these. Or LibreOffice.

Well, Firefox built for Windows needs a 64-bit build system because the linking (which includes LTO) takes more than 3GB of RAM. That's with Microsoft's compiler, but I bet GCC will have similar requirements.

[Caleb]

Why do you publish articles with zero content?

If you ran benchmarks on some pre-alpha build of GCC and found nothing interesting why would you still publish the article and waste people's time?

[jakubo]

Why do you publish articles with zero content?

If you ran benchmarks on some pre-alpha build of GCC and found nothing interesting why would you still publish the article and waste people's time?

why would you complain about things everybody knows and waste everybody's time?
Michael needs the clicks. What do you have for an excuse?

[Nexus6]

I don't think you will see much performance increases in those benchmarks, as the hot paths are probably inlined already. It is something for more complex code, like firefox as an extreme example with ~6M LOC, where they probably won't waste time optimizing it manually.

Knowing something of the various projects tested here, they make a reasonable amount of sense.

GraphicsMagick generally has its hot path in one c file for each utility (montage, convert, etc). So the same performance seen with/without LTO is expected.
Himeno I don't know, but 2% improvement isn't terrible given all the effort required was just a compile flag. Most benchmarks tend to be pretty small code segments.
LAME encoding is 4% better - that is worth having. As I work in Performance QA for my job, getting a 4% performance improvement is a huge win for a large project. That could equate to many months of effort by many developers on a mature project to acheive the same result with code changes.

[mufasa72]

Man, imagine the build time for Firefox with these. Or LibreOffice.

Why do you tempt me!?!