git: 23408297fbf3 - main - Merge llvm-project 12.0.1 rc2

[Dmitry Chagin]

On Sun, Jun 20, 2021 at 09:46:32PM +0200, Dimitry Andric wrote:
> On 20 Jun 2021, at 14:49, Dmitry Chagin <dchagin at freebsd.org> wrote:
> > 
> > On Sat, Jun 19, 2021 at 06:11:04PM +0000, Dimitry Andric wrote:
> >> The branch main has been updated by dim:
> >> 
> >> URL: https://cgit.FreeBSD.org/src/commit/?id=23408297fbf3089f0388a8873b02fa75ab3f5bb9
> >> 
> >> commit 23408297fbf3089f0388a8873b02fa75ab3f5bb9
> >> Merge: cac129e60300 e4bbddaec868
> >> Author:     Dimitry Andric <dim at FreeBSD.org>
> >> AuthorDate: 2021-06-19 10:06:00 +0000
> >> Commit:     Dimitry Andric <dim at FreeBSD.org>
> >> CommitDate: 2021-06-19 18:09:28 +0000
> >> 
> >>    Merge llvm-project 12.0.1 rc2
> >> 
> >>    This updates llvm, clang, compiler-rt, libc++, libunwind, lld, lldb and
> >>    openmp to llvmorg-12.0.1-rc2-0-ge7dac564cd0e, a.k.a. 12.0.1 rc2.
> >> 
> >>    PR:             255570
> >>    MFC after:      6 weeks
> >> 
> > hi! with 11 clang kernel build time is about 60 sec, with 12.0.0 and
> > 12.0.1 rc2 build time increased to 500 sec.
> 
> Did you previously have llvm/clang assertions turned off (via
> MK_LLVM_ASSERTIONS=no)?
> 
no,
I found the reason. I tried to build a kernel without optimization,
set CFLAGS=-O and forgot to  remove it before building the world.





> I did a few measurements of "make -j32 buildkernel", on
> main-n247475-9d1cafb3049, with clang 11.0.1 and clang 12.0.1 (both with
> assertions turned off, and statically linked:
> 
> clang 11.0.1 (n=5):
> 
> real user sys
> 138.65 3145.23 285.25
> 129.00 3240.73 290.61
> 126.38 3244.03 284.76
> 128.90 3250.95 268.71
> 136.42 3250.73 267.93
> 
> clang 12.0.1 (n=5):
> 
> real user sys
> 129.16 3306.13 291.84
> 128.18 3308.14 284.32
> 128.77 3301.35 281.90
> 141.37 3293.22 282.20
> 129.93 3304.08 284.47
> 
> $ ministat -A clang1101-real.txt clang1201-real.txt
> x clang1101-real.txt
> + clang1201-real.txt
>     N           Min           Max        Median           Avg        Stddev
> x   5        126.38        138.65           129        131.87     5.3354662
> +   5        128.18        141.37        129.16       131.482     5.5639707
> No difference proven at 95.0% confidence
> 
> $ ministat -A clang1101-user.txt clang1201-user.txt
> x clang1101-user.txt
> + clang1201-user.txt
>     N           Min           Max        Median           Avg        Stddev
> x   5       3145.23       3250.95       3244.03      3226.334     45.550423
> +   5       3293.22       3308.14       3304.08      3302.584     5.8070759
> Difference at 95.0% confidence
>         76.25 +/- 47.3552
>         2.36336% +/- 1.50191%
>         (Student's t, pooled s = 32.4697)
> 
> Before I merged 12.0.1, I did the same sort of measurement, but for
> "make -j32 buildworld buildkernel" (with CROSS_TOOLCHAIN set so the
> bootstrap compiler isn't built), and this showed roughly similar
> results:
> 
> $ ministat -A clang1101-real.txt clang1200-real.txt
> x clang1101-real.txt
> + clang1200-real.txt
>     N           Min           Max        Median           Avg        Stddev
> x   5       1410.24       1461.04        1434.2      1435.296     18.210247
> +   5       1461.54       1499.69       1462.07      1469.976     16.667199
> Difference at 95.0% confidence
>         34.68 +/- 25.4583
>         2.41623% +/- 1.79718%
>         (Student's t, pooled s = 17.4558)
> 
> $ ministat -A clang1101-user.txt clang1200-user.txt
> x clang1101-user.txt
> + clang1200-user.txt
>     N           Min           Max        Median           Avg        Stddev
> x   5      37495.09       37913.2      37750.64     37750.734     165.00886
> +   5      38133.12      38396.77      38390.77     38324.132     113.63363
> Difference at 95.0% confidence
>         573.398 +/- 206.617
>         1.51891% +/- 0.552972%
>         (Student's t, pooled s = 141.67)
> 
> So basically ~2% slower. It's would be a little tricky to find out if
> there is any particular upstream commit that might be responsible,
> as the difference is so small.
> 
> I think I'll also compare the performance with assertions enabled, and
> MK_SHARED_TOOLCHAIN=yes, which should be worse than the above, but still
> comparable between 11.0 and 12.0.
> 
> -Dimitry
>