Changeset 180098 in webkit for trunk/Source/JavaScriptCore/dfg/DFGStrengthReductionPhase.cpp

Timestamp:

Feb 13, 2015, 8:20:21 PM (10 years ago)

Author:

[email protected]

Message:

Add a DFG node for the Pow Intrinsics
​https://bugs.webkit.org/show_bug.cgi?id=141540

Patch by Benjamin Poulain <​[email protected]> on 2015-02-13
Reviewed by Filip Pizlo.

Add a DFG Node for PowIntrinsic. This patch covers the basic cases
need to avoid massive regression. I will iterate over the node to cover
the missing types.

With this patch I get the following progressions on benchmarks:
-LongSpider's math-partial-sums: +5%.
-Kraken's imaging-darkroom: +17%
-AsmBench's cray.c: +6.6%
-CompressionBench: +2.2% globally.

• dfg/DFGAbstractInterpreterInlines.h:

(JSC::DFG::AbstractInterpreter<AbstractStateType>::executeEffects):
Cover a couple of trivial cases:
-If the exponent is zero, the result is always one, regardless of the base.
-If both arguments are constants, compute the result at compile time.

• dfg/DFGByteCodeParser.cpp:

(JSC::DFG::ByteCodeParser::handleIntrinsic):

• dfg/DFGClobberize.h:

(JSC::DFG::clobberize):

• dfg/DFGDoesGC.cpp:

(JSC::DFG::doesGC):

• dfg/DFGFixupPhase.cpp:

(JSC::DFG::FixupPhase::fixupNode):
We only support 2 basic cases at this time:
-Math.pow(double, int)
-Math.pow(double, double).

I'll cover Math.pow(int, int) in a follow up.

• dfg/DFGNode.h:

(JSC::DFG::Node::convertToArithSqrt):
(JSC::DFG::Node::arithNodeFlags):

• dfg/DFGNodeType.h:
• dfg/DFGPredictionPropagationPhase.cpp:

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

• dfg/DFGSafeToExecute.h:

(JSC::DFG::safeToExecute):

• dfg/DFGSpeculativeJIT.cpp:

(JSC::DFG::compileArithPowIntegerFastPath):
(JSC::DFG::SpeculativeJIT::compileArithPow):

• dfg/DFGSpeculativeJIT.h:
• dfg/DFGSpeculativeJIT32_64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

• dfg/DFGSpeculativeJIT64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

• dfg/DFGStrengthReductionPhase.cpp:

(JSC::DFG::StrengthReductionPhase::handleNode):

• dfg/DFGValidate.cpp:

(JSC::DFG::Validate::validate):

• ftl/FTLCapabilities.cpp:

(JSC::FTL::canCompile):

• ftl/FTLIntrinsicRepository.h:
• ftl/FTLLowerDFGToLLVM.cpp:

(JSC::FTL::LowerDFGToLLVM::compileNode):
(JSC::FTL::LowerDFGToLLVM::compileArithPow):

• ftl/FTLOutput.h:

(JSC::FTL::Output::doublePow):
(JSC::FTL::Output::doublePowi):

• jit/JITOperations.cpp:
• jit/JITOperations.h:
• runtime/MathObject.cpp:

(JSC::mathProtoFuncPow):
(JSC::isDenormal): Deleted.
(JSC::isEdgeCase): Deleted.
(JSC::mathPow): Deleted.

• tests/stress/math-pow-basics.js: Added.
• tests/stress/math-pow-integer-exponent-fastpath.js: Added.
• tests/stress/math-pow-nan-behaviors.js: Added.
• tests/stress/math-pow-with-constants.js: Added.

Start some basic testing of Math.pow().
Due to the various transform, the value change when the code tiers up,
I covered this by checking for approximate values.

File:

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

trunk/Source/JavaScriptCore/dfg/DFGStrengthReductionPhase.cpp

@@ -133 +133 @@
             }
             break;
-            
+
+        case ArithPow:
+            if (m_node->child2()->isNumberConstant()) {
+                double yOperandValue = m_node->child2()->asNumber();
+                if (yOperandValue == 1) {
+                    convertToIdentityOverChild1();
+                } else if (yOperandValue == 0.5) {
+                    m_insertionSet.insertNode(m_nodeIndex, SpecNone, Phantom, m_node->origin, m_node->children);
+                    m_node->convertToArithSqrt();
+                    m_changed = true;
+                }
+            }
+            break;
+
         case GetArrayLength:
             if (JSArrayBufferView* view = m_graph.tryGetFoldableViewForChild1(m_node))