Help with performance issues.

Mon Jul 30 09:59:00 GMT 2001

>>>>> "Andrew" == Andrew Haley <aph@redhat.com> writes:

>> I think we could consider doing this for any final method with fixed
>> semantics. For instance, anything in java.lang.Math (taking strictfp
>> concerns into account, of course). If other final methods are
>> important for performance, we could inline those too. For instance,
>> there might be situations where inlining System.arraycopy makes sense.

Andrew> All this falls down, though, when we need to inline a *native*
Andrew> method. That would be a good trick...
In the above case (final methods with fixed semantics), we can do
this. That's because we aren't "inlining" -- we are recognizing the
name of the function and replacing the call with a builtin (or with
the equivalent code). So for instance we could replace Math.cos with
__builtin_cos. Ditto Math.max(double,double), etc.
Inlining C++ into Java would be hard. We could either merge the
compilers, or we could do link-time inlining. These both seem
dauntingly hard.
Inlining Java into C++ can be done much more easily. We could replace
the trivial decompilation code in gcjh with a real decompiler.
Tom