review (M) for 6862956: PhaseIdealLoop should have a CFG verification mode
Tom Rodriguez
Thomas.Rodriguez at Sun.COM
Wed Aug 12 13:55:54 PDT 2009
On Aug 12, 2009, at 1:14 PM, Tom Rodriguez wrote:
> No I haven't. It's certainly not free. I can add some TracePhases
> and do some runs. I also moved the first verify into the preceeding
> block of PhaseIdealLoop passes so we aren't doing doing multiple
> verifies for code without loops. I'll get some numbers.
I added another elapsedTimer and ran jbb fastdebug with -XX:
+TimeCompiler on an 8 way 2.7Ghz opteron. It reports about .33
seconds spent in verify out of 19 seconds total. Seems pretty
manageable. I've updated the webrev with the new timer code.
tom
Accumulated compiler times:
---------------------------
Total compilation: 19.067 sec.
method compilation : 19.057 sec/135529 bytes (7112 bytes per sec)
stub compilation : 0.009 sec.
Phases:
parse : 4.822 sec
optimizer : 2.652 sec
iterGVN : 0.138 sec
idealLoop : 1.456 sec
idealLoopVerify: 0.332 sec
ccp : 0.046 sec
iterGVN2 : 0.164 sec
graphReshape : 0.030 sec
subtotal : 1.670 sec, 62.95 %
matcher : 0.525 sec
scheduler : 0.629 sec
regalloc : 4.276 sec
ctorChaitin : 0.002 sec
buildIFG : 1.425 sec
computeLive : 0.759 sec
regAllocSplit: 1.260 sec
postAllocCopyRemoval: 0.140 sec
fixupSpills : 0.012 sec
subtotal : 3.598 sec, 84.14 %
macroExpand : 0.140 sec
blockOrdering: 0.049 sec
peephole : 0.007 sec
codeGen : 4.347 sec
install_code : 0.057 sec
------------ : ----------
total : 17.536 sec, 92.02 %
output : 4.348 sec
isched : 0.000 sec
bldOopMaps: 0.088 sec
>
> tom
>
> On Aug 12, 2009, at 12:53 PM, Vladimir Kozlov wrote:
>
>> Thank you, Tom, for doing this.
>>
>> Changes look good.
>> Did you measure how much time compiler spends in the verification
>> code
>> (in PhaseIdealLoop::verify(igvn))?
>>
>> Thanks,
>> Vladimir
>>
>> Tom Rodriguez wrote:
>>> http://cr.openjdk.java.net/~never/6862956/
>
More information about the hotspot-compiler-dev
mailing list