RFR: 8210416: [linux] Poor StrictMath performance due to non-optimized compilation

[Severin Gehwolf]

Hi Joe,

On Wed, 2018-09-05 at 12:15 -0700, joe darcy wrote:
> On 9/5/2018 6:12 AM, Severin Gehwolf wrote:
> > Hi,
> > 
> > Cross-posting this review-thread on core-libs-dev and build-dev as
> > this
> > is a build change, but affects fdlibm which is core-libs.
> > 
> > With JDK-8170153 optimization for fdlibm code has been turned on
> > for
> > ppc64 s390 and aarch64. This patch intends to turn it on on all
> > arches
> > on Linux. I've not observed any precision issues. Is there a good
> > reason to not use -O3 -ffp-contract=off everywhere?
> > 
> > Bug:    https://bugs.openjdk.java.net/browse/JDK-8210416
> > webrev: 
> > http://cr.openjdk.java.net/~sgehwolf/webrevs/JDK-8210416/webrev.01/
> > 
> > Testing: - java/lang/Math, java/lang/StrictMath tests (all pass).
> >           - Currently running through submit repo.
> > 
> > A simple micro benchmark from JDK-8170153[1] shows these numbers
> > for
> > StrictMath:
> > 
> > Function      | before     | after
> > ----------------------------------------------
> > sin           | 0m33.382s  | 0m18.731s
> > cos           | 0m31.562s  | 0m18.796s
> > tan           | 0m33.657s  | 0m21.093s
> > atan          | 0m5.714s   | 0m4.902s
> > log           | 0m6.212s   | 0m4.439s
> > log10         | 0m7.946s   | 0m5.543s
> > sqrt          | 0m0.481s   | 0m0.449s
> > cbrt          | 0m5.295s   | 0m5.214s
> > tanh          | 0m1.404s   | 0m1.307s
> > log1p         | 0m6.457s   | 0m5.131s
> > IEEEremainder | 0m10.629s  | 0m6.048s
> > atan2         | 0m8.037s   | 0m5.668s
> > hypot         | 0m2.171s   | 0m2.147s
> > 
> > 
> 
> Note that pow (JDK-8134795), hypot (JDK-7130085), cbrt (JDK-8136799), 
> and exp (JDK-8139688), have been ported to Java as of JDK 9. The sqrt 
> method is commonly handled as an intrinsic.

OK thanks. Since ppc64/s390x/aarch64 uses this already on Linux do you
anticipate the same being applied to x86/x86_64 causing issues (modulo
compiler bugs of course)?

> Testing that was not geared toward finding precision/rounding issues 
> would be unlikely to find them.

Would running the TCK be geared towards precision/rounding issues? I
could ask someone to run the TCK on a test build on x86_64/x86 to find
out.

> I don't see the sources of the microbenchmark in JDK-8170153.

https://github.com/gromero/strictmath/

Thanks,
Severin