Changeset 144973 in webkit for trunk/Source/JavaScriptCore/dfg/DFGDriver.cpp

Timestamp:

Mar 6, 2013, 1:23:39 PM (12 years ago)

Author:

[email protected]

Message:

DFG should not run full CSE after the optimization fixpoint, since it really just wants store elimination
​https://bugs.webkit.org/show_bug.cgi?id=111536

Reviewed by Oliver Hunt and Mark Hahnenberg.

The fixpoint will do aggressive load elimination and pure CSE. There's no need to do it after the fixpoint.
On the other hand, the fixpoint does not profit from doing store elimination (except for SetLocal/Flush).
Previously we had CSE do both, and had it avoid doing some store elimination during the fixpoint by querying
the fixpoint state. This changes CSE to be templated on mode - either NormalCSE or StoreElimination - so
that we explicitly put it into one of those modes depending on where we call it from. The goal is to reduce
time spent doing load elimination after the fixpoint, since that is just wasted cycles.

• dfg/DFGCSEPhase.cpp:

(JSC::DFG::CSEPhase::CSEPhase):
(JSC::DFG::CSEPhase::run):
(JSC::DFG::CSEPhase::performNodeCSE):
(JSC::DFG::CSEPhase::performBlockCSE):
(JSC::DFG::performCSE):
(DFG):
(JSC::DFG::performStoreElimination):

• dfg/DFGCSEPhase.h:

(DFG):

• dfg/DFGDriver.cpp:

(JSC::DFG::compile):

File:

• trunk/Source/JavaScriptCore/dfg/DFGDriver.cpp (modified) (1 diff)

trunk/Source/JavaScriptCore/dfg/DFGDriver.cpp

@@ -153 +153 @@
 
     dfg.m_fixpointState = FixpointConverged;
-    performCSE(dfg);
-    performCPSRethreading(dfg); // This should usually be a no-op since CSE rarely dethreads the graph.
+    performStoreElimination(dfg);
+    performCPSRethreading(dfg); // This should usually be a no-op since store elimination rarely dethreads the graph.
     performDCE(dfg);
     performVirtualRegisterAllocation(dfg);