Changeset 183718 in webkit

Timestamp:

May 2, 2015 10:33:36 AM (9 years ago)

Author:

fpizlo@apple.com

Message:

Unreviewed, add FIXMEs referencing ​https://bugs.webkit.org/show_bug.cgi?id=144524 and
​https://bugs.webkit.org/show_bug.cgi?id=144525.

• dfg/DFGLICMPhase.cpp:

(JSC::DFG::LICMPhase::attemptHoist):

• dfg/DFGPhantomInsertionPhase.cpp:

Location:

trunk/Source/JavaScriptCore

Files:

• ChangeLog (modified) (1 diff)
• dfg/DFGLICMPhase.cpp (modified) (1 diff)
• dfg/DFGPhantomInsertionPhase.cpp (modified) (1 diff)

trunk/Source/JavaScriptCore/ChangeLog

@@ +1 @@
+2015-05-02  Filip Pizlo  <fpizlo@apple.com>
+
+        Unreviewed, add FIXMEs referencing https://bugs.webkit.org/show_bug.cgi?id=144524 and
+        https://bugs.webkit.org/show_bug.cgi?id=144525.
+
+        * dfg/DFGLICMPhase.cpp:
+        (JSC::DFG::LICMPhase::attemptHoist):
+        * dfg/DFGPhantomInsertionPhase.cpp:
+
 2015-05-02  Yusuke Suzuki  <utatane.tea@gmail.com>
 

trunk/Source/JavaScriptCore/dfg/DFGLICMPhase.cpp

@@ -214 +214 @@
         }
         
+        // FIXME: At this point if the hoisting of the full node fails but the node has type checks,
+        // we could still hoist just the checks.
+        // https://bugs.webkit.org/show_bug.cgi?id=144525
+        
         if (readsOverlap(m_graph, node, data.writes)) {
             if (verbose) {

trunk/Source/JavaScriptCore/dfg/DFGPhantomInsertionPhase.cpp

@@ -82 +82 @@
     void handleBlock(BasicBlock* block)
     {
+        // FIXME: For blocks that have low register pressure, it would make the most sense to
+        // simply insert Phantoms at the last point possible since that would obviate the need to
+        // query bytecode liveness:
+        //
+        // - If we MovHint @x into loc42 then put a Phantom on the last MovHinted value in loc42.
+        // - At the end of the block put Phantoms for each MovHinted value.
+        //
+        // This will definitely not work if there are any phantom allocations. For those blocks
+        // where this would be legal, it remains to be seen how profitable it would be even if there
+        // was high register pressure. After all, a Phantom would cause a spill but it wouldn't
+        // cause a fill.
+        //
+        // https://bugs.webkit.org/show_bug.cgi?id=144524
+        
         m_values.fill(nullptr);