Hi Per,
Thanks for taking a look!
Inline.
Cheers
Markus
From: Per Liden
Sent: den 18 november 2013 13:22
To: Markus Gronlund; hotspot-runtime-...@openjdk.java.net;
hotspot-gc-...@openjdk.java.net; serviceability-dev@openjdk.java.net
serviceability-dev@openjdk.java.net
Subject: Re: RFR(M): 8028128: Add a type safe alternative for working with
counter based data
Hi Markus,
Looks like a nice cleanup. Naming it always hard, so I'll avoid that topic ;) A
few other minor comments, feel free to ignore.
[MG] i agree on the naming parts. I also agree that the current names are not
the best possible. I have assumed they will change.
101 class TicksToTimeHelper : public AllStatic {
102 public:
103 enum Unit {
104 SECONDS = 1,
105 MILLISECONDS = 1000
106 };
107 static double seconds(const Tickspan& span);
108 static jlong milliseconds(const Tickspan& span);
109 };
Do we really want a separate class for this instead of having
Tickspan::seconds() and Tickspan::milliseconds()? If moved inside Tickspan,
maybe value() should be ticks() to be consistent (and I guess Ticks::value() ->
Ticks::ticks() would follow)?
[MG] In this particular case I would prefer to keep the convenience functions
non-member, non-friends since a convenience member function really does not add
anything extra that cannot be retrieved from outside the type. I also think
that keeping the convenience functions outside/decoupled from the type itself
increases encapsulation and increases flexibility - esp. for this case very
much so where the convenience functions for at ticks/time representation is
very much open ended; I just whacked together a seconds() and milliseconds(),
but what about microseconds(), nanoseconds() and other timing convenience
functions that might be needed further? I think keeping it outside increases
the flexibility to add convenience functions without affecting the primary type.
43 inline bool operator!=(const Tickspan& lhs, const Tickspan& rhs) {
44 return !operator==(lhs,rhs);
45 }
In many of the operators you have code like the example above. Is there a
reason you want to spell it out like that instead of just "return !(lhs ==
rhs);"?
[MG] I use this way of writing to make it explicit that a particular non-member
operator is implemented in terms of another non-member operator. This point of
view might be considered a bit too "implementation" centric, but I have found
it useful sometimes as an easy way to directly point to the implementation -
granted, for a case such as !(lhs == rhs) that might already be obvious..
145 void register_gc_phase_start(const char* name, const Ticks& time =
Ticks::now());
There are quite a few places where you've added default values like the one
above. But as far as I can tell most (I think I saw at least one exception in
there) of these functions are either always called without the argument or
always called with the argument. Seems a bit unnecessary to provide this
flexibility when it's never used (as it opens up for incorrect use of the
interface)?
[MG] Most of the default parameters are there in order to avoid polluting all
the GC collectors with the knowledge about the Ticks/Tickspan types - I can
encapsulate the usages only to gcTrace and gcTimer modules. You are right about
this for the register_gc_phase_* functions though - they always pass a
parameter, providing a default parameter for them might be a bit gratuitous. I
will update the phase functions accordingly:
void register_gc_phase_start(const char* name, const Ticks& time);
void register_gc_phase_end(const Ticks& time);
Thanks Per!
I will send out an updated webrev.
cheers,
/Per
On 2013-11-12 12:17, Markus Gronlund wrote:
Greetings,
Kindly asking for reviews for the following changeset:
Webrev: HYPERLINK
"http://cr.openjdk.java.net/%7Emgronlun/8028128/webrev01/"http://cr.openjdk.java.net/~mgronlun/8028128/webrev01/
Bug: https://bugs.openjdk.java.net/browse/JDK-8028128
Description:
Currently, when working with counter based data, the primary data type
representation in the VM are raw jlong's. jlong's are employed to represent
both:
1. a single instance in time, for example a "now" timestamp
2. a time duration, a timespan, for example the duration between "now" - "then"
Having a raw jlong type represent both of these shifts the responsibility to
the programmer to always ensure the correct validity of the counter data,
without any assistance either at compile time or at runtime.
The event based tracing framework relies heavily on counter based data in order
to keep track of time and timing related information and we have seen counter
data errors where the representation sent to the tracing framework has
submitted the wrong time, i.e. a timespan when a time instant was expected and
vice versa.
We should therefore add an alternative to jlongs for representing counter based
data. We should enlist the help of the type system to enforce the correct data
types at compile time as well as introduce strong runtime asserts in order to
help with the correct usage of time and to reduce the number of potential bugs.
I will therefore add two new types in src/share/vm/utilities:
1. "Ticks" - represents a single counter based timestamp, a single instance in
time.
2. "Tickspan" - represents a counter based duration, a timespan, between two
"Ticks"
Please see the added files:
- src/share/vm/utilities/ticks.hpp
- src/share/vm/utilities/ticks.inline.hpp
- src/share/vm/utilities/ticks.cpp
I have also updated some of the existing event based tracing "event sites" to
employ usage of these new types as well as updated a few places in the event
based tracing framework implementation to accommodate these types.
Testing completed:
nsk.stress.testlist (UTE/Aurora)
GC nightlies (Aurora)
gc/gctests/* (UTE/Aurora)
nsk.gc.testlist (UTE/Aurora)
SpecJBB2013 (no regression)
SpecJBB2005 (no regression)
Kitchensink (locally Windows) runs for +> 9 days (still running.)
Thanks in advance
Markus
