---

19 Representation selection

File: represent

Some types of object (such as single-float) have multiple possible representations. Multiple representations are useful mainly when there is a particularly efficient non-descriptor representation. In this case, there is the normal descriptor representation, and an alternate non-descriptor representation.

This possibility brings up two major issues:

• The compiler must decide which representation will be most efficient for any given value, and
• Representation conversion code must be inserted where the representation of a value is changed.

First, the representations for TNs are selected by examining all the TN references and attempting to minimize reference costs. Then representation conversion code is introduced.

This phase is in effect a pre-pass to register allocation. The main reason for its existence is that representation conversions may be farily complex (e.g. involving memory allocation), and thus must be discovered before register allocation.

VMR conversion leaves stubs for representation specific move operations. Representation selection recognizes move by name. Argument and return value passing for call VOPs is controlled by the :move-arguments option to define-vop.

Representation selection is also responsible for determining what functions use the number stack. If any representation is chosen which could involve packing into the non-descriptor-stack SB, then we allocate the NFP register throughout the component. As an optimization, permit the decision of whether a number stack frame needs to be allocated to be made on a per-function basis. If a function doesn’t use the number stack, and isn’t in the same tail-set as any function that uses the number stack, then it doesn’t need a number stack frame, even if other functions in the component do.