This PR adds better support for differentiating complex and imaginary floating-point types from real floating-point types, in both the AST and in the IR type system.
*AST Changes*
- Introduces the new class `TypeDomain`, which can be either `RealDomain`, `ImaginaryDomain` or `ComplexDomain`. "type domain" is the term used for this concept in the C standard, and I couldn't think of a better one.
- Introduces `FloatingPointType.getDomain()`, to get the type domain of the type.
- Introduces `FloatingPointType.getBase()`, to get the numeric base of the type (either 2 or 10).
- Introduces three new subtypes of `FloatingPointType`: `RealNumberType`, `ComplexNumberType`, and `ImaginaryNumberType`, which differentiate between the types based on their type domain. Note that the decimal types (e.g., `_Decimal32`) are included in `RealNumberType`.
- Introduces two new subtypes of `FloatingPointType`: `BinaryFloatingPointType` and `DecimalFloatingPointType`, which differentiate between the types based on their numeric base, independent of type domain.
*IR Changes*
- `IRFloatingPointType` now has two additional parameters: the base and the type domain.
- New test that ensures that C++ types get mapped to the correct IR types.
- New IR test that verifies the IR for some basic usage of complex FP types.
`Instruction.getDefinitionOverlap()` depends on `SSAConstruction::getMemoryOperandDefinition()`, which in turn depends on `SSAConstruction::hasMemoryOperandDefinition()`. When the definition in question came from a `Chi` instruction, `hasMemoryOperandDefinition()` incorrectly bound `overlap` to the overlap relationship between the original (non-`Chi`) instruction and the use. The fix is to make use of the `actualDefLocation` parameter to `getDefinitionOrChiInstruction()`, which specifies the location for the result of the `Chi` in that case.
The result of `getDefinitionOverlap()` should never be `MayPartiallyOverlap`, because if that were the case, we should have inserted as `Chi` instruction and hooked the definition up to that instead.
There are quite a few existing failures.
This predicate replaces `isChiForAllAliasedMemory`, which was always
intended to be temporary. A test is added to `IRSanity.qll` to verify
that the new predicate corresponds exactly with (a fixed version of) the
old one.
The implementation of the new predicate,
`Cached::hasConflatedMemoryResult` in `SSAConstruction.qll`, is faster
to compute than the old `isChiForAllAliasedMemory` because it uses
information that's readily available during SSA construction.
In the Unix ABI, `std::va_list` is defined as `typedef struct __va_list_tag { ... } va_list[1];`, which means that any `std::va_list` used as a function parameter decays to `struct __va_list_tag*`. Handling this actually made the QL code slightly cleaner. The only tricky bit is that we have to determine what type to use as the actual `va_list` type when loading, storing, or modifying a `std::va_list`. To do this, we look at the type of the argument to the `va_*` macro. A detailed QLDoc comment explains the details.
I added a test case for passing a `va_list` as an argument, and then manipulating that `va_list` in the callee.
This PR changes the IR we generate for functions that accept a variable argument list. Rather than simply using `BuiltInOperationInstruction` to model the various `va_*` macros as mysterious function-like operations, we now model them in more detail. The intent is to enable better alias analysis and taint flow through varargs.
The `va_start` macro now generates a unary `VarArgsStart` instruction that takes the address of the ellipsis pseudo-parameter as its operand, and returns a value of type `std::va_list`. This value is then stored into the actual `std::va_list` variable via a regular `Store`.
The `va_arg` macro now loads the `std::va_list` argument, then emits a `VarArg` instruction on the result. This returns the address of the vararg argument to be loaded. That address is later used as the address operand of a regular `Load` to return the value of the argument. To model the side effect of moving to the next argument, we emit a `NextVarArg` instruction that takes the previous `std::va_list` value and returns an updated one, which is then stored back into the `std::va_list` variable.
The `va_end` macro just emits a `VarArgsEnd` unary instruction that takes the address of the `std::va_list` argument and does nothing, since `va_end` doesn't really do anything on most compiler implementations anyway.
The `va_copy` macro is just modeled as a plain copy.
This change introduces a new synthesized `IRVariable` in every varargs function. This variable represents the entire set of arguments passed to the ellipsis by the caller. We give it an opaque type big enough hold all of the arguments passed by the largest vararg call in the database. It is treated just like any other parameter. It is initialized the same, it has indirect buffers, etc.
I had to introduce a couple new APIs to `Call` and `Function`. The QLDoc comments should explain these. I added tests for these new APIs as well.
The next step will be to change the IR generation for the `va_*` macros to manipulate the ellipsis parameter.
Added a new `StaticLocalVariable` class, which made several other pieces of the original change a bit cleaner.
Fixed test failures due to a mistake in the original `CFG.qll` change.
Added a test case for static local variables with constructors.
Removed the `Uninitialized` instruction from the initialization of a static local, because all objects with static storage duration are zero-initialized at startup.
Fixed expectations for `SignAnalysis.ql` to reflect that a bad result is now fixed.
Previously, the IR for the initialization of a static local variable ran the initialization unconditionally, every time the declaration was reached during execution. This means that we don't model the possibility that an access to the static variable fetches a value that was set on a previous execution of the function.
I've added some simple modelling of the correct behavior to the IR. For each static local variable that has a dynamic initializer, we synthesize a (static) `bool` variable to hold whether the initializer for the original variable has executed. When executing a declaration, we check the value of the synthesized variable, and skip the initialization code if it is `true`. If it is `false`, we execute the initialization code as before, and then set the flag to `true`. This doesn't capture the thread-safe nature of static initialization, but I think it's more than enough to handle anything we're likely to care about for the foreseeable future.
In `TranslatedDeclarationEntry.qll`, I split the translation of a static local variable declaration into two `TranslatedElement`s: one for the declaration itself, and one for the initialization. The declaration part handles the checking and setting of the flag; the initialization just does the initialization as before.
I've added an IR test case that has static variables with constant, zero, and dynamic initialization. I've also verified the new IR generated for @jbj's previous test cases for constant initialization.
I inverted the sense of the `hasConstantInitialization()` predicate to be `hasDynamicInitialization()`. Mostly this just made more sense to me, but I think it also fixed a potential bug where `hasConstantInitialization()` would not hold for a zero-initialized variable. Technically, constant initialization isn't the same as zero initialization, but I believe that most code really cares about the distinction between dynamic initialization and static initialization, where static initialization includes both constant and zero initialization.
I've fixed up the C# side of IR generation to continue working, but it doesn't use any of the dynamic initialization stuff. In theory, it could use something similar to model the initialization of static fields.
