Conversation
|
After a couple more profiling runs of red/black, here's the findings I've come up with:
The fixes in the PR have eliminated |
|
Moving on to another benchmark, Sampled profile shows heavy use of I was surprised how much of an impact! A patch to simply BEFORE AFTER: We need better mechanisms to know whether methods like Here's the patch. diff --git a/core/src/main/java/org/jruby/RubyNumeric.java b/core/src/main/java/org/jruby/RubyNumeric.java
index f606ca9..9642153 100644
--- a/core/src/main/java/org/jruby/RubyNumeric.java
+++ b/core/src/main/java/org/jruby/RubyNumeric.java
@@ -39,6 +39,7 @@ import org.jruby.anno.JRubyMethod;
import org.jruby.ast.util.ArgsUtil;
import org.jruby.common.RubyWarnings;
import org.jruby.exceptions.RaiseException;
+import org.jruby.ext.bigdecimal.RubyBigDecimal;
import org.jruby.javasupport.JavaUtil;
import org.jruby.runtime.Block;
import org.jruby.runtime.ClassIndex;
@@ -275,6 +276,22 @@ public class RubyNumeric extends RubyObject {
throw runtime.newTypeError("can't convert " + num.getType() + " into Float");
}
+ if (num instanceof RubyFloat) {
+ return num;
+ }
+
+ if (num instanceof RubyFixnum) {
+ return ((RubyFixnum) num).to_f();
+ }
+
+ if (num instanceof RubyBignum) {
+ return ((RubyBignum) num).to_f();
+ }
+
+ if (num instanceof RubyBigDecimal) {
+ return ((RubyBigDecimal) num).to_f();
+ }
+
return TypeConverter.convertToType(num, runtime.getFloat(), "to_f");
}
|
|
The patch above for |
|
Here's some numbers on pythag compared with other impls, for a current baseline. This includes my patch for |
|
Er, that was "we need to find a better way to do dynamic calls from Java..." |
|
@headius thanks for these detailed write ups! |
|
I have merged latest master and committed the numericToFloat fix. I'll let it cook in CI and merge if it looks ok. |
|
Backed off the numeric optimization a bit. We'll get a better system in place for detecting core methods. |
2 participants
I'm doing a few passes over our favorite benchmarks looking for things to fix. As of creation, this PR contains the following:
[]=should require a frame. All other frame-aware method names compile as Call and not AttrAssign.BasicObject#!=with a Ruby version, so it caches==.These changes have a measurable effect on red/black, bringing average time per iter from high 0.4s to low 0.4s on 8u92.
More to come, probably.