High-level Overview of the Native-Code Generation Process

Representation of Pointers

When emitting code to an object file, pointers will be emitted as relocations. The deserialization code will ensure any object that pointed to one of these constants gets recreated and contains the right runtime pointer.

Otherwise, they will be emitted as literal constants.

To emit one of these objects, call literal_pointer_val. It'll handle tracking the Julia value and the LLVM global, ensuring they are valid both for the current runtime and after deserialization.

When emitted into the object file, these globals are stored as references in a large gvals table. This allows the deserializer to reference them by index, and implement a custom manual mechanism similar to a Global Offset Table (GOT) to restore them.

Function pointers are handled similarly. They are stored as values in a large fvals table. Like globals, this allows the deserializer to reference them by index.

Note that extern functions are handled separately, with names, via the usual symbol resolution mechanism in the linker.

Note too that ccall functions are also handled separately, via a manual GOT and Procedure Linkage Table (PLT).

Representation of Intermediate Values

Values are passed around in a jl_cgval_t struct. This represents an R-value, and includes enough information to determine how to assign or pass it somewhere.

They are created via one of the helper constructors, usually: mark_julia_type (for immediate values) and mark_julia_slot (for pointers to values).

The function convert_julia_type can transform between any two types. It returns an R-value with cgval.typ set to typ. It'll cast the object to the requested representation, making heap boxes, allocating stack copies, and computing tagged unions as needed to change the representation.

By contrast update_julia_type will change cgval.typ to typ, only if it can be done at zero-cost (i.e. without emitting any code).

Union representation

Inferred union types may be stack allocated via a tagged type representation.

The primitive routines that need to be able to handle tagged unions are:

  • mark-type
  • load-local
  • store-local
  • isa
  • is
  • emit_typeof
  • emit_sizeof
  • boxed
  • unbox
  • specialized cc-ret

Everything else should be possible to handle in inference by using these primitives to implement union-splitting.

The representation of the tagged-union is as a pair of < void* union, byte selector >. The selector is fixed-size as byte & 0x7f, and will union-tag the first 126 isbits. It records the one-based depth-first count into the type-union of the isbits objects inside. An index of zero indicates that the union* is actually a tagged heap-allocated jl_value_t*, and needs to be treated as normal for a boxed object rather than as a tagged union.

The high bit of the selector (byte & 0x80) can be tested to determine if the void* is actually a heap-allocated (jl_value_t*) box, thus avoiding the cost of re-allocating a box, while maintaining the ability to efficiently handle union-splitting based on the low bits.

It is guaranteed that byte & 0x7f is an exact test for the type, if the value can be represented by a tag – it will never be marked byte = 0x80. It is not necessary to also test the type-tag when testing isa.

The union* memory region may be allocated at any size. The only constraint is that it is big enough to contain the data currently specified by selector. It might not be big enough to contain the union of all types that could be stored there according to the associated Union type field. Use appropriate care when copying.

Specialized Calling Convention Signature Representation

A jl_returninfo_t object describes the calling convention details of any callable.

If any of the arguments or return type of a method can be represented unboxed, and the method is not varargs, it'll be given an optimized calling convention signature based on its specTypes and rettype fields.

The general principles are that:

  • Primitive types get passed in int/float registers.
  • Tuples of VecElement types get passed in vector registers.
  • Structs get passed on the stack.
  • Return values are handle similarly to arguments, with a size-cutoff at which they will instead be returned via a hidden sret argument.

The total logic for this is implemented by get_specsig_function and deserves_sret.

Additionally, if the return type is a union, it may be returned as a pair of values (a pointer and a tag). If the union values can be stack-allocated, then sufficient space to store them will also be passed as a hidden first argument. It is up to the callee whether the returned pointer will point to this space, a boxed object, or even other constant memory.