BURG
Program-Transformation.Org: The Program Transformation Wiki
BURG is a system for
CodeGeneration from
IntermediateRepresentation? expression trees developed by
ChristopherFraser,
ToddProebsting and others in the early 90's.
Papers
- C. W. Fraser, D. R. Hanson, and T. A. Proebsting. Engineering a simple, efficient code-generator generator. ACM Letters on Programming Languages and Systems, 1(3):213--226, September 1992.
- C. W. Fraser, R. R. Henry, and T. A. Proebsting. BURG---fast optimal instruction selection and tree parsing. ACM SIGPLAN Notices, 27(4):68--76, April 1992.
- T. A. Proebsting. BURS automata generation. ACM Transactions on Programming Languages and Systems, 17(3):461--486, May 1995.
Description
BURG is a system for
code generation from intermediate expression trees. A mapping from intermediate representation
trees to machine instruction is defined by means of a tree grammar. A
production of the form
n -> t (c)
defines a mapping of
tree pattern
t
to non-terminal
n
at cost
c
. Associated with each production is an action to take
when the production is selected. For example,
ToddProebsting
gives the following example grammar:
[1] goal -> reg (0) [5] reg -> Plus(reg, reg) (2)
[2] reg -> Reg (0) [6] addr -> reg (0)
[3] reg -> Int (1) [7] addr -> Int (0)
[4] reg -> Fetch(addr) (2) [8] addr -> Plus(reg, Int) (0)
According to this grammar, the term
Fetch(Fetch(Plus(Reg,Int)))
has two coverings
corresponding to the derivations
4(4(6(5(2,3))))
and
4(4(8(2)))
.
As illustrated by this example, more than one covering of a tree is
possible, corresponding to different ways to generate code. Each node
can have several different parses because of overlapping patterns and
chain rules. The costs associated with the productions express the
cost of executing the associated machine instruction. The goal of a
code generator is to find the lowest cost covering (i.e., lowest
execution time) of an intermediate representation expression tree.
According to bottom-up rewriting theory (BURS) an intermediate representation tree
can be translated to a sequence of instructions using the following
strategy. In a bottom-up traversal all lowest-cost patterns that
match each node are computed and associated with the node. This
involves matching the righ-hand sides of the productions to the tree,
taking into account earlier matches for sub-trees. Instructions are
then selected in a top-down traversal that is driven by the goal
non-terminal for the root of the tree.
This restricted form of rewriting can also be applied for
simple type inference problems, for checking tree formats, and for
tree simplifications.
--
EelcoVisser - 30 Apr 2001
CategorySystem