An Early Look At GCC 12 Compiler Performance On The Core i9 12900K

[phoronix]

Phoronix: An Early Look At GCC 12 Compiler Performance On The Core i9 12900K

With GCC 12 now onto stage four meaning that the major feature work is over, I've slowly begun running more tests on the GCC 12 compiler that is due for its stable introduction around April. First up is a look at the Core i9 12900K "Alder Lake" performance on GCC 12 in its near-final form compared to GCC 11.2 as the current stable release from last year.

https://www.phoronix.com/vr.php?view=30928

[max0x7ba]

The charts look stone age and are barely readable.

Use responsive design column charts with error bars, e.g. https://www.highcharts.com/docs/char...ror-bar-series

[birdie]

The charts look stone age and are barely readable.

Use responsive design column charts with error bars, e.g. https://www.highcharts.com/docs/char...ror-bar-series

PTS is open source, so you're welcome to contribute.

[DavidBrown]

The charts in the benchmarks only have one real flaw - the mix of "bigger is better" and "smaller is better" makes them far harder to understand. It would be a huge step forward to be consistent - bigger is better.

[dmalcolm]

Error bars would be a big step forward; it's hard to know to what extent results are affected by noise (especially with just N=3).

Benchmarking is hard; see e.g. https://speakerdeck.com/haypo/how-to...able-benchmark (disclosure: Victor is a colleague of mine).

From the "please may I have a pony" department: even better would be for the benchmark to detect where the hotspots in the profile are, and to compare the generated asm before/after for the pertinent hotspots, and file some kind of actionable bug report in our bug tracker (caveat: I'm a GCC developer, though I focus mostly on diagnostics rather than optimization).

[kylew77]

No complaints about the charts. They convey the information well enough. Michael works 100 hour weeks and probably doesn't make much money keeping this wonderful site going for over a decade now I think. He worked his butt off at the beginning of the year doing a *BSD flavor comparison I asked for just for me.

What is interesting is that -mach=native sometimes yielded no gains, sometimes substantial gains, and curiously enough sometimes huge performance losses!

[SigHunter]

are the produced binaries with -O3 -flto in some way measurably faster than the ones with -O2?

[Sin2x]

are the produced binaries with -O3 -flto in some way measurably faster than the ones with -O2?

Clear Linux consistently wins in benchmarks on Phoronix and they use O3 -- with a plethora of other optimizations enabled, though.

[CochainComplex]

PTS is open source, so you're welcome to contribute.

Best comment so far!

[ms178]

Very recently GCC12 started to be able to compile the Kernel successfully again. That was broken for months.