This doesn't make it much faster, but it reduces the debug output
volume. It also simplifies the code.
I've found this change necessary when I compute the full IR on a
Wireshark snapshot in QL4E. Without it, Eclipse runs out of memory
because the console log is too large.
The new predicate `isOrphan` gets inlined into
`ignoreExprAndDescendants`, whose performance improves from
TranslatedElement::ignoreExprAndDescendants#f .. 23.4s (executed 9 times)
to
TranslatedElement::ignoreExprAndDescendants#f ... 4.3s (executed 9 times)
This dramatic improvement is not only due to eliminating a type check in
the recursive case. Removing the type check from the other base cases
also enabled them to get better join orders.
The previous reccomentation changed the behaviour of the code.
A user following the advice might have broken her/his code:
With call-by-value, the original parameter is not changed.
With a call-by-reference, however, it may be changed. To be sure,
nothing breaks by blindly following the advice, suggest to pass a
const reference.
This relation was almost 40x the size it needed to be on Wireshark
because it lacked a restriction on the `tag` parameter. To implement
that restriction efficiently, I had to split the relation in two to
dictate the join order.
With the fix, `getInstruction` now computes the same as
`getInstructionTranslatedElementAndTag`, so the latter could be
simplified.
I made a corresponding change to `TranslatedElement.getTempVariable` for
the sake of consistency.
A part of `SSAConstruction.getInstructionOperandDefinition` was more
expensive than it had to be. On a ChakraCore snapshot, this changes the
tuple counts from
3020569 ~2% {3} r40 = JOIN OperandTag::TUnmodeledUseOperand#f WITH Instruction::Instruction::getFunction_dispred#ff CARTESIAN PRODUCT OUTPUT FIELDS {Instruction::Instruction::getFunction_dispred#ff.<0>,OperandTag::TUnmodeledUseOperand#f.<0>,Instruction::Instruction::getFunction_dispred#ff.<1>}
62405 ~0% {3} r41 = JOIN r40 WITH Instruction::UnmodeledUseInstruction#class#fffffff ON r40.<0>=Instruction::UnmodeledUseInstruction#class#fffffff.<0> OUTPUT FIELDS {r40.<2>,r40.<1>,r40.<0>}
2868421 ~1% {3} r42 = JOIN r41 WITH Instruction::Instruction::getFunction_dispred#ff_10#join_rhs ON r41.<0>=Instruction::Instruction::getFunction_dispred#ff_10#join_rhs.<0> OUTPUT FIELDS {Instruction::Instruction::getFunction_dispred#ff_10#join_rhs.<1>,r41.<1>,r41.<2>}
62405 ~0% {3} r43 = JOIN r42 WITH Instruction::UnmodeledDefinitionInstruction#class#fffffff ON r42.<0>=Instruction::UnmodeledDefinitionInstruction#class#fffffff.<0> OUTPUT FIELDS {r42.<2>,r42.<1>,r42.<0>}
to
(0s) Starting to evaluate predicate SSAConstruction::Cached::getUnmodeledUseInstruction#ff
(0s) Tuple counts:
62405 ~0% {2} r1 = JOIN Instruction::UnmodeledUseInstruction#class#fffffff WITH Instruction::Instruction::getFunction_dispred#ff ON Instruction::UnmodeledUseInstruction#class#fffffff.<0>=Instruction::Instruction::getFunction_dispred#ff.<0> OUTPUT FIELDS {Instruction::Instruction::getFunction_dispred#ff.<1>,Instruction::Instruction::getFunction_dispred#ff.<0>}
return r1
...
75716 ~0% {3} r40 = JOIN OperandTag::TUnmodeledUseOperand#f WITH FunctionIR::FunctionIR::getUnmodeledDefinitionInstruction#ff CARTESIAN PRODUCT OUTPUT FIELDS {FunctionIR::FunctionIR::getUnmodeledDefinitionInstruction#ff.<0>,OperandTag::TUnmodeledUseOperand#f.<0>,FunctionIR::FunctionIR::getUnmodeledDefinitionInstruction#ff.<1>}
62405 ~0% {3} r41 = JOIN r40 WITH FunctionIR::FunctionIR::getUnmodeledUseInstruction#ff ON r40.<0>=FunctionIR::FunctionIR::getUnmodeledUseInstruction#ff.<0> OUTPUT FIELDS {FunctionIR::FunctionIR::getUnmodeledUseInstruction#ff.<1>,r40.<1>,r40.<2>}
Now that we have `Expr.getParentWithConversions`, we can implement
`TranslatedElement.getRealParent` simpler. This implementation also
avoids recursion.
In theory this query will produce no results on C++ code; in practice, I
suspect the "cpp" suite is often run on code compiled as C, so it is
likely to be worth running anyways.