Iteration of transformation passes over attributed program trees

Abstract

Summary.Transformations of attributed program trees form an essential part of compiler optimizations. A strategy of repeatedly applying alternate attribute evaluation and tree transformation phases is discussed. An attribute evaluation phase consists of a sequence of passes over the tree. A tree transformation phase consists of a single pass, which is never interrupted to carry out a re-evaluation. Both phases can be performed in parallel. This strategy requires a distinction between consistent (i.e., correct) and approximate attribute values. Tree transformations can be considered safe if they guarantee that the attribute values everywhere in the program tree will remain consistent or will become at least approximations of the consistent values, so that subsequent transformations can be applied correctly.This attribute evaluation and tree transformation strategy shows similarities with the evaluation methods for circular attribute grammars.

Concurrent incremental attribute evaluation

“…This requires a different view of the correctness of attribute values of a syntax tree. The use of approximat¢ attribute values, as defined in [4], allows multiple transformations and multiple asynchronous reevaluations. asynchronous tree…”

“…To remove this restriction, we formulate criteria which permit a delay in calling the reevaluator. These criteria allow in particular the execution of tree transformations in different regions at the same time, and also a delay in calling the reevaluator until a sequence of tree transformations has been performed and several parts of a region have been affected [4]. A delay in calling the reevaluator requires a different view of the correctness of attribute values of a syntax tree.…”

“…In [4] it has been proved that, for monotonic attribute evaluation rules, every locally safe tree transformation rule is safe. These criteria may help the writer of the attribute evaluation and tree transformation rules to check the safety of his rules.…”

“…These criteria may help the writer of the attribute evaluation and tree transformation rules to check the safety of his rules. Another important conclusion from [4] is that, the use of both monotonic attribute evaluation rules and locally safe tree transformation rules yields an improvement of the attribute values at any local or global reevaluation.…”

Attributed tree transformations with delayed and smart re-evaluation

“…Instead, we formulate criteria which allow a delay in calling the re-evaluator until a sequence of tree transformations has been performed and several parts of the tree are assumed to be affected [4]. A delay in calling the re-evaluator requires a different view of the correctness of attribute values of a derivation tree.…”