Stateful transformation causes non-termination in Hoopl analysis.
It seems that Hoopl's analyses may apply rewrites to nodes multiple times without invoking the restart
function between applications. This causes havoc for code transformations that have side-effects, for example a transformation that introduces a new variable to hold some intermediate result. In such an analysis, a transformation may introduce a new variable, which leads to a changed fact. Later the transformation may be run again, leading to a new fact, and so on (when there is a loop in the control flow.
The attached example illustrates this problem. This example modifies the constant propagation pass provided in the testing directory of the hoopl library. The example added a new instruction called "Funny" that gets rewritten to an unconditional branch instruction and introduces a new variable. The main program P1 performs the analysis on a particular program that has a loop in the control flow graph and causes the hoopl analysis to run forever.
Is this a bug in Hoopl? Ie. should analyses that have side-effects involving checkpointed state be supported? If not, could it be changed to support these analyses?
Trac metadata
Trac field | Value |
---|---|
Version | 7.8.3 |
Type | Bug |
TypeOfFailure | OtherFailure |
Priority | normal |
Resolution | Unresolved |
Component | libraries/hoopl |
Test case | |
Differential revisions | |
BlockedBy | |
Related | |
Blocking | |
CC | |
Operating system | |
Architecture |