Changeset 193781 in webkit for trunk/Source/JavaScriptCore/dfg/DFGPredictionPropagationPhase.cpp

Timestamp:

Dec 8, 2015, 1:44:12 PM (10 years ago)

Author:

[email protected]

Message:

Polymorphic operand types for DFG and FTL div.
​https://bugs.webkit.org/show_bug.cgi?id=151747

Reviewed by Geoffrey Garen.

Source/JavaScriptCore:

Perf on benchmarks is neutral. The new JSRegress ftl-object-div test shows
a speed up not from the div operator itself, but from the fact that the
polymorphic operand types support now allow the test function to run without OSR
exiting, thereby realizing the DFG and FTL's speed up on other work that the test
function does.

This patch has passed the layout tests on x86_64 with a debug build.
It passed the JSC tests with x86 and x86_64 debug builds.

• dfg/DFGAbstractInterpreterInlines.h:

(JSC::DFG::AbstractInterpreter<AbstractStateType>::executeEffects):

• dfg/DFGClobberize.h:

(JSC::DFG::clobberize):

• dfg/DFGFixupPhase.cpp:

(JSC::DFG::FixupPhase::fixupNode):

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

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

• dfg/DFGSpeculativeJIT.cpp:

(JSC::DFG::SpeculativeJIT::compileArithDiv):

• ftl/FTLCompileBinaryOp.cpp:

(JSC::FTL::generateBinaryArithOpFastPath):
(JSC::FTL::generateBinaryOpFastPath):

• ftl/FTLInlineCacheDescriptor.h:
• ftl/FTLInlineCacheDescriptorInlines.h:

(JSC::FTL::ArithDivDescriptor::ArithDivDescriptor):
(JSC::FTL::ArithDivDescriptor::icSize):

• ftl/FTLInlineCacheSize.cpp:

(JSC::FTL::sizeOfArithDiv):

• ftl/FTLInlineCacheSize.h:

• ftl/FTLLowerDFGToLLVM.cpp:

(JSC::FTL::DFG::LowerDFGToLLVM::lower):
(JSC::FTL::DFG::LowerDFGToLLVM::compileArithMul):

• Fixed a cut-paste bug where the op_mul IC was using the op_sub IC size. This bug is benign because the op_sub IC size turns out to be larger than op_mul needs.

(JSC::FTL::DFG::LowerDFGToLLVM::compileArithDiv):

• jit/JITArithmetic.cpp:

(JSC::JIT::emit_op_div):

• Fixed a bug where the scratchFPR was not allocated for the 64bit port. This bug is benign because the scratchFPR is only needed if we are using scratchGPR register (used for branchConvertDoubleToInt32()) is

  X86Registers::r8. Since we're always using regT2 for the scratchT2,

  the scratchFPR is never needed. However, we should fix this anyway to be correct.

• tests/stress/op_div.js:
• Fixed some test values.

LayoutTests:

• js/regress/ftl-object-div-expected.txt: Added.
• js/regress/ftl-object-div.html: Added.
• js/regress/script-tests/ftl-object-div.js: Added.

(o1.valueOf):
(foo):

File:

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

trunk/Source/JavaScriptCore/dfg/DFGPredictionPropagationPhase.cpp

@@ -355 +355 @@
             
             if (left && right) {
-                if (Node::shouldSpeculateInt32OrBooleanForArithmetic(node->child1().node(), node->child2().node())
-                    && node->canSpeculateInt32(m_pass))
-                    changed |= mergePrediction(SpecInt32);
-                else
-                    changed |= mergePrediction(SpecBytecodeDouble);
+                if (isFullNumberOrBooleanSpeculationExpectingDefined(left)
+                    && isFullNumberOrBooleanSpeculationExpectingDefined(right)) {
+                    if (Node::shouldSpeculateInt32OrBooleanForArithmetic(node->child1().node(), node->child2().node())
+                        && node->canSpeculateInt32(m_pass))
+                        changed |= mergePrediction(SpecInt32);
+                    else
+                        changed |= mergePrediction(SpecBytecodeDouble);
+                } else
+                    changed |= mergePrediction(SpecInt32 | SpecBytecodeDouble);
             }
             break;