vm_call_iseq_setup_normal_0start() is simple, however it has
some loops/conditions depends on ISeq::param.size and
ISeq::local_size (in vm_push_frame(), inlined into this function).
There are many simple methods which has a few parameters and local
variables. So that this patch introduces several special functions
generated in vm_call_iseq_optimized.inc by
tool/mk_call_iseq_optimized.rb.
This script makes
vm_call_iseq_setup_normal_0start_Xparams_Ylocals()
where X is 0 to 3 and Y is 1 to 6 (as current setting).
In this case, X * Y = 24 functions are created.
These functions creates fast method dispatch by inlining
vm_push_frame() with immediate params/locals sizes.
On my laptop, we can have the following results.
vm2_method* 1.083 (8.3% faster)
vm2_poly_method* 0.961 (3.4% slower)
It shows 8.3% faster for inner loop method dispatch (hit inline
cache), but 3.4% slower when inline cache miss because we need
to find a suitable call handler.
* common.mk: add a rule for vm_call_iseq_optimized.inc.
* tool/mk_call_iseq_optimized.rb: added.
* vm.c: include vm_call_iseq_optimized.inc.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@52254 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
nobu
svn
ko1