Changeset 145299 in webkit for trunk/Source/JavaScriptCore/dfg/DFGCPSRethreadingPhase.cpp

Timestamp:

Mar 8, 2013, 6:51:06 PM (12 years ago)

Author:

[email protected]

Message:

DFG overflow check elimination is too smart for its own good
​https://bugs.webkit.org/show_bug.cgi?id=111832

Source/JavaScriptCore:

Reviewed by Oliver Hunt and Gavin Barraclough.

This improves overflow check elimination in three ways:

1) It reduces the amount of time the compiler will spend doing it.

2) It fixes bugs where overflow check elimination was overzealous. Precisely, for a binary operation

over @a and @b where both @a and @b will type check that their inputs (@a->children, @b->children)
are int32's and then perform a possibly-overflowing operation, we must be careful not to assume
that @a's non-int32 parts don't matter if at the point that @a runs we have as yet not proved that
@b->children are int32's and that hence @b might produce a large enough result that doubles would
start chopping low bits. The specific implication of this is that for a binary operation to not
propagate that it cares about non-int32 parts (NodeUsedAsNumber), we must prove that at least one
of the inputs is guaranteed to produce a result within 2^{32 and that there won't be a tower of such
operations large enough to ultimately produce a double greater than 2}52 (roughly). We achieve the
latter by disabling this optimization for very large basic blocks. It's noteworthy that blocks that
large won't even make it into the DFG currently.

3) It makes the overflow check elimination more precise for cases where the inputs to an Add or Sub

are the outputs of a bit-op. For example in (@a + (@b | 0)) | 0, we don't need to propagate
NodeUsedAsNumber to either @a or @b.

This is neutral on V8v7 and a slight speed-up on compile time benchmarks.

• CMakeLists.txt:
• GNUmakefile.list.am:
• JavaScriptCore.xcodeproj/project.pbxproj:
• Target.pri:
• dfg/DFGArrayMode.cpp:

(JSC::DFG::ArrayMode::refine):

• dfg/DFGBackwardsPropagationPhase.cpp: Added.

(DFG):
(BackwardsPropagationPhase):
(JSC::DFG::BackwardsPropagationPhase::BackwardsPropagationPhase):
(JSC::DFG::BackwardsPropagationPhase::run):
(JSC::DFG::BackwardsPropagationPhase::isNotNegZero):
(JSC::DFG::BackwardsPropagationPhase::isNotZero):
(JSC::DFG::BackwardsPropagationPhase::isWithinPowerOfTwoForConstant):
(JSC::DFG::BackwardsPropagationPhase::isWithinPowerOfTwoNonRecursive):
(JSC::DFG::BackwardsPropagationPhase::isWithinPowerOfTwo):
(JSC::DFG::BackwardsPropagationPhase::mergeDefaultFlags):
(JSC::DFG::BackwardsPropagationPhase::propagate):
(JSC::DFG::performBackwardsPropagation):

• dfg/DFGBackwardsPropagationPhase.h: Added.

(DFG):

• dfg/DFGCPSRethreadingPhase.cpp:

(JSC::DFG::CPSRethreadingPhase::run):
(JSC::DFG::CPSRethreadingPhase::clearIsLoadedFrom):
(CPSRethreadingPhase):
(JSC::DFG::CPSRethreadingPhase::canonicalizeGetLocalFor):
(JSC::DFG::CPSRethreadingPhase::canonicalizeFlushOrPhantomLocalFor):

• dfg/DFGDriver.cpp:

(JSC::DFG::compile):

• dfg/DFGGraph.cpp:

(JSC::DFG::Graph::dump):

• dfg/DFGNodeFlags.cpp:

(JSC::DFG::dumpNodeFlags):
(DFG):

• dfg/DFGNodeFlags.h:

(DFG):

• dfg/DFGPredictionPropagationPhase.cpp:

(PredictionPropagationPhase):
(JSC::DFG::PredictionPropagationPhase::propagate):

• dfg/DFGUnificationPhase.cpp:

(JSC::DFG::UnificationPhase::run):

• dfg/DFGVariableAccessData.h:

(JSC::DFG::VariableAccessData::VariableAccessData):
(JSC::DFG::VariableAccessData::mergeIsLoadedFrom):
(VariableAccessData):
(JSC::DFG::VariableAccessData::setIsLoadedFrom):
(JSC::DFG::VariableAccessData::isLoadedFrom):

LayoutTests:

Reviewed by Oliver Hunt and Gavin Barraclough.

• fast/js/dfg-arith-add-overflow-check-elimination-predicted-but-not-proven-int-expected.txt: Added.
• fast/js/dfg-arith-add-overflow-check-elimination-predicted-but-not-proven-int.html: Added.
• fast/js/dfg-arith-add-overflow-check-elimination-tower-of-large-numbers-expected.txt: Added.
• fast/js/dfg-arith-add-overflow-check-elimination-tower-of-large-numbers.html: Added.
• fast/js/jsc-test-list:
• fast/js/script-tests/dfg-arith-add-overflow-check-elimination-predicted-but-not-proven-int.js: Added.

(foo):
(bar):

• fast/js/script-tests/dfg-arith-add-overflow-check-elimination-tower-of-large-numbers.js: Added.

(foo):
(bar):

File:

• trunk/Source/JavaScriptCore/dfg/DFGCPSRethreadingPhase.cpp (modified) (7 diffs)

trunk/Source/JavaScriptCore/dfg/DFGCPSRethreadingPhase.cpp

@@ -48 +48 @@
             return false;
         
+        clearIsLoadedFrom();
         freeUnnecessaryNodes();
         canonicalizeLocalsInBlocks();
@@ -58 +59 @@
 
 private:
+    
+    void clearIsLoadedFrom()
+    {
+        for (unsigned i = 0; i < m_graph.m_variableAccessData.size(); ++i)
+            m_graph.m_variableAccessData[i].setIsLoadedFrom(false);
+    }
     
     void freeUnnecessaryNodes()
@@ -171 +178 @@
             
             if (otherNode->op() == SetArgument) {
+                variable->setIsLoadedFrom(true);
                 node->children.setChild1(Edge(otherNode));
                 m_block->variablesAtTail.atFor<operandKind>(idx) = node;
@@ -177 +185 @@
             
             if (variable->isCaptured()) {
+                variable->setIsLoadedFrom(true);
                 if (otherNode->op() == GetLocal)
                     otherNode = otherNode->child1().node();
@@ -201 +210 @@
         }
         
+        variable->setIsLoadedFrom(true);
         Node* phi = addPhi<operandKind>(node->codeOrigin, variable, idx);
         node->children.setChild1(Edge(phi));
@@ -257 +267 @@
             }
             
+            variable->setIsLoadedFrom(true);
             // There is nothing wrong with having redundant Flush's. It just needs to
             // be linked appropriately. Note that if there had already been a previous
@@ -267 +278 @@
         }
         
+        variable->setIsLoadedFrom(true);
         node->children.setChild1(Edge(addPhi<operandKind>(node->codeOrigin, variable, idx)));
         m_block->variablesAtHead.atFor<operandKind>(idx) = node;