guardIfConstant, the constant sniffer combinator

[Remi Forax]

----- Mail original -----
> De: "John Rose" <john.r.rose at oracle.com>
> À: "Da Vinci Machine Project" <mlvm-dev at openjdk.java.net>
> Cc: "valhalla-dev" <valhalla-dev at openjdk.java.net>
> Envoyé: Jeudi 11 Avril 2019 00:08:07
> Objet: Re: guardIfConstant, the constant sniffer combinator

> This is very similar to a "growable switch" combinator,
> which would call the provider for each distinct selector
> value.  A key difference is that non-constant values
> go through the fallback, where a "growable switch"
> doesn't need a fallback, since the provider MH is
> always free to just return a standard fallback as
> its result (as is the case with your combinator).
> 
> I wonder if the concepts can be combined somehow?

They are clearly related.
I'm trying to avoid the shared pairs of constant/target you need to maintain for a "growable switch" combinator by pushing these states neare to the generated assembly code where the inlining tree is resolved. 

> 
> Here's a test question:  When the JIT, after heroic
> effort, discovers that an argument is constant, can
> it run the provider (in the compiler thread???) to
> determine a custom handler for the newly discovered
> branch of constant code?  The answer is probably,
> "sorry, no", although the JIT might set up an uncommon
> trap (and/or an execution counter) that can revisit
> the question at some point.

yes,
the idea is to revisit later, my hope is that it can work well with tiered compilation or if there is only one JIT, to force a recompilation after some time.

> 
> Here's a close point of correspondence between the
> idea of a growable switch and your combinator:  Both
> need a memory.  Both want to remember the appearance
> of constants, so a later optimization phase can use the
> full historical knowledge.

yes !
and my hope is that we don't need a shared memory for that.   

> 
> — John

Rémi

> 
>> On Apr 10, 2019, at 2:47 PM, Remi Forax <forax at univ-mlv.fr> wrote:
>> 
>> The problem is the following,
>> with the java compiler intrinsic of amber, String.format() is optimized using
>> invokedynamic in the case the format (the first argument) is constant (and some
>> other conditions on the format), this is great, perf are like 25x in simple
>> benchmarks, and that all because in a lot of code, the format is not constant
>> for the Java compiler.
>> 
>> By example,
>>  class Logger {
>>    public static void log(String format, String message) {
>>      System.err.println(String.format(format, message));
>>    }
>>  }
>>  ...
>>  logger.log("%s", "hello");
>> 
>> The format is not a constant inside Logger::log for the Java compiler but when
>> the code is JITed, due to inlining, logger.log("hello") calls String.format()
>> with a constant.
>> 
>> 
>> I propose a way to fix that, by providing a method handle combiner
>> (guardIfConstant) that detects if an argument is a constant and do something
>> different if it's a constant or not.
>> It's a little more complex than that, we don't only want to have a separate path
>> if the argument is a constant, we also want to be able to build a method handle
>> tree depending on the value of that constant.
>> 
>>  MethodHandle guardIfConstant(int argument, MethodHandle targetProvider,
>>  MethodHandle fallback)
>> 
>> the semantics: if the nth argument is a constant, the target provider is called
>> with that argument and the return value, a method handle, is called with all
>> the arguments, otherwise the fallback is called.
>> 
>> in term of method type:
>>  - the method type of the return value of guardIfConstant is the same as fallback
>>  - the method type of targetProvider returns a MethodHandle and takes a single
>>  parameter which is the nth parameter type of the fallback method type,
>>    the returned method handle as to have the same method type as the fallback.
>> 
>> Rémi
>> 
>> _______________________________________________
>> mlvm-dev mailing list
>> mlvm-dev at openjdk.java.net
> > https://mail.openjdk.java.net/mailman/listinfo/mlvm-dev