RFR: 8262950: Restructure compiler/intrinsics/TestRotate.java for easier compilation [v2]
Aleksey Shipilev
shade at openjdk.java.net
Wed Mar 3 21:23:39 UTC 2021
On Wed, 3 Mar 2021 20:23:09 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:
>> By "not enough" you probably mean "while we are at it, we can do even better"? Because current version definitely helps JIT compilation, as evidenced by test times and the absence of compilation aborts.
>>
>> Off the top of my head:
>> - Dropping `ITERS` one order of magnitude down to 50K does not improve test times, AFAICS.
>> - Verifying each on each step verifies OSR compilations as well, which must be good.
>> - Having a test call not in triplicate (how you suggest) but in singular (how it is in this PR) seems more readable.
>>
>> I can take a look at these later, but honestly, I think the current patch is an acceptable thing to make `tier1` run fast again.
>
> 500,000 vs 50,000 ITER with your changes:
> elapsed time (seconds): 22.664
> vs
> elapsed time (seconds): 6.683
Yup, that only seems to improve considerably when I run the test directly as Java class. jtreg adds a lot of constant overhead, which masks this improvement:
# 500K
$ time CONF=linux-x86_64-server-fastdebug make run-test TEST=compiler/intrinsics/TestRotate.java
...
real 0m24.330s
user 0m51.197s
sys 0m3.607s
# 50K
$ time CONF=linux-x86_64-server-fastdebug make run-test TEST=compiler/intrinsics/TestRotate.java
...
real 0m19.109s
user 0m45.391s
sys 0m3.787s
No matter, new patch dropped `ITERS` to `50K`.
-------------
PR: https://git.openjdk.java.net/jdk/pull/2811
More information about the hotspot-compiler-dev
mailing list