DFG fixup phase should be responsible for inserting ValueToInt32's as needed and it should use Phantom to keep the original values alive in case of OSR exit
https://bugs.webkit.org/show_bug.cgi?id=126600
Reviewed by Michael Saboff.
This fixes an embarrassing OSR exit liveness bug. It also simplifies the code. We were
already using FixupPhase as the place where conversion nodes get inserted. ValueToInt32
was the only exception to that rule, and that was one of the reasons why we had this bug.
Henceforth ValueToInt32 is only inserted by FixupPhase, and only when it is necessary:
we have a BitOp that will want a ToInt32 conversion and the operand is not predicted to
already be an int32. If FixupPhase inserts any ValueToInt32's then the BitOp will no
longer appear to use the original operand, which will make OSR exit think that the
original operand is dead. We work around this they way we always do: insert a Phantom on
the original operands right after the BitOp. This ensures that any OSR exit in any of the
ValueToInt32's or in the BitOp itself will have values for the original inputs.
- dfg/DFGBackwardsPropagationPhase.cpp:
(JSC::DFG::BackwardsPropagationPhase::isWithinPowerOfTwo):
(JSC::DFG::BackwardsPropagationPhase::propagate):
- dfg/DFGByteCodeParser.cpp:
(JSC::DFG::ByteCodeParser::handleIntrinsic):
(JSC::DFG::ByteCodeParser::parseBlock):
(JSC::DFG::FixupPhase::fixupNode):
(JSC::DFG::FixupPhase::fixIntEdge):
(JSC::DFG::FixupPhase::fixBinaryIntEdges):
- dfg/DFGPredictionPropagationPhase.cpp:
(JSC::DFG::PredictionPropagationPhase::propagate):
- tests/stress/bit-op-value-to-int32-input-liveness.js: Added.
(foo):
