On some snapshots, notably ffmpeg, the IR `ValueNumbering` recursion
would generate billions of tuples and eventually run out of space.
It turns out it was fairly common for an `Instruction` to get more than
one `ValueNumber` in the base cases for `VariableAddressInstruction` and
`InitializeParameterInstruction`, and it could also happen in an
instruction with more than one operand of the same `OperandTag`. When a
binary operation was applied to an instruction with `m` value numbers
and another instruction with `n` value numbers, the result would get
`m * n` value numbers. This led to doubly-exponential growth in the
number of value numbers in rare cases.
The underlying reason why a `VariableAddressInstruction` could get
multiple value numbers is that it was keyed on the associated
`IRVariable`, and the `IRVariable` is defined in part by the type of its
underlying `Variable` (or other AST element). If the extractor defines a
variable to have multiple types because of linker ambiguity, this leads
to the creation of multiple `IRVariable`s. That should ideally be solved
in `TIRVariable.qll`, but for now I've put a workaround in
`ValueNumberingInternal.qll` instead.
To remove the problem with instructions having multiple operands, the
construction in `Operand.qll` will now filter out any such operand. It
wasn't enough to apply that filter to the `raw` stage, so I've applied
it to all three stages.
We were hitting a combinatorial explosion in `hasDefinitionAtRank` for functions that contain a large number of string literals. The problem was that every `Chi` instruction for `AliasedVirtualVariable` was treated as a definition of every string literal. We already mark string literals as `isReadOnly()`, but we were allowing `AliasedVirtualVariable` to define read-only locations so that the `AliasedDefinition` instruction would provide the initial definition for all string literals.
To fix this, I've introduced the new `InitializeNonLocal` instruction, which is inserted in the prologue of every function right after `AliasedDefinition`. It provides the initial definition for every non-stack memory location, including read-only locations, but is never written to anywhere else. It is the conterpart of the `AliasedUse` instruction in the function epilogue, which represents the use of all non-stack memory after the function returns. I considered renaming `AliasedUse` to `ReturnNonLocal`, to match the `InitializeXXX`/`ReturnXXX` pattern we already use for parameters and indirections, but held off to avoid unnecessary churn. Any thoughts on whether I should make this name change?
This change has a significant speedup in evaluation time for a few of our troublesome databases:
`attnam/ivan`: 13%
`awslabs/s2n`: 26%
`SinaMostafanejad/OpenRDM`: 7%
`zcoinofficial/zcoin`: 8%
Instead of tracking `IRVariable`s directly, alias analysis now tracks instances of the `Allocation` type provided by its `Configuration` parameter. For unaliased SSA, an `Allocation` is just an `IRAutomaticVariable`. For aliased SSA, an `Allocation` is either an `IRVariable` or the memory pointed to by an indirect parameter.
There was already a `WriteSideEffectInstruction` class that served as a
superclass for all the specific write side effects. This new class
serves the same purpose for read side effects.
When building SSA, we'll be assuming that stack variables do not escape, at least until we improve our alias analysis. I've added a new `IREscapeAnalysisConfiguration` class to allow the query to control this, and a new `UseSoundEscapeAnalysis.qll` module that can be imported to switch to the sound escape analysis. I've cloned the existing IR and SSA tests to have both sound and unsound versions. There were relatively few diffs in the IR dump tests, and the sanity tests still give the same results after one change described below.
Assuming that stack variables do not escape exposed an existing bug where we do not emit an `Uninitialized` instruction for the temporary variables used by `return` statements and `throw` expressions, even if the initializer is a constructor call or array initializer. I've refactored the code for handling elements that initialize a variable to share a common base class. I added a test case for returning an object initialized by constructor call, and ensured that the IR diffs for the existing `throw` test cases are correct.
