Slowdown observed in gcc 4.8.0-1 vs 4.7.2-4

Question

Not sure where the appropriate section is for this post.  Mods feel free to move elsewhere.
tl; dr Conclusions:
1) gcc 4.8.0 compiles ~ 5% slower than gcc 4.7.2 compiles.
2) gcc 4.8.0 produces kernels that run slower when compiling than kernels produced by gcc 4.7.2.
Experiment/Discussion:
I read the Intel i7 gcc 4.8 benchmark piece that Phoronix published a week ago wherein they noted some significant slow downs in the 4.8 series.  Now that 4.8.0-1 hit [testing] I wanted to see how the two version compared using a compilation based endpoint with `make -j8 bzImage` in the latest linux package being the test.  This run was repeated 10 times (threw out the first run as a warm up).  Results were analyzed in a oneway ANOVA.
Test machine: Intel i7-3370k @ 4.5 GHz/16 GB RAM running Arch 86_64.
Four test conditions:
1) linux-ck-3.8.5-1 (build with gcc-4.7.2-4) running the benchmark with gcc-4.7.2-4; blue in the graphic below.
2) linux-ck-3.8.5-1 (build with gcc-4.7.2-4) running the benchmark with gcc-4.8.0-1; green in the graphic below.
3) linux-ck-3.8.5-2 (build with gcc-4.8.0-1) running the benchmark with gcc-4.7.2-4; pink in the graphic below.
4) linux-ck-3.8.5-2 (build with gcc-4.8.0-1) running the benchmark with gcc-4.8.0-1; yellow in the graphic below.
Refer to the data plotted in this graphic below which is compile time on the y-axis and conditions on the x-axis.  This is a oneway so all 9 values are plotted as a distribution:
*Compare median blue to green or compare median pink to yellow; result is the 4.8 version of gcc gives ~5 % decrease in efficiency as measured by compilation time.
*Compare median green to yellow; result is that the kernel compiled with 4.8.0-1 runs slower as measured by this end-point.  Yes, it is a trivial slow down of ~272 ms but that slow down is statistically significant.
References:
Here is the script that did the work.
Here is the raw data I generated.
Last edited by graysky (2013-03-30 14:33:40)

Answer 1

It's the first stable version that uses c++, so I think that is not bad, actually. I saw this some days ago: http://rusty.ozlabs.org/?p=330