Enabling loop-interchange

[nikic]

@sebpop

DependenceAnalysis […] no longer uses the GEP source element type

There is only one use of GEP->getSourceElementType() in delinearization, and dozens in other passes. What’s the recommended way to replace it? Do you have a patch that does that in some other part of the compiler? Can you also please explain why delinearization should not access that info?

The problem is not that delinearization is using getSourceElementType(), it’s what it does with it. You are not allowed to drive optimization heuristics based on the structural type information in a GEP. You can only make use of the offset calculation that type implies. The type information will go away entirely in the future, once the ongoing migration to ptradd representation completes.

A typical way to replace the type access is to use collectOffset(), which converts it into an explicit offset representation. But this does lose the explicit array bounds information from the type – and losing it is intentional.

@sebpop

Loop interchange should not depend on AA when the code has no nested loops.

GCC’s pass manager has a “gate” function to determine whether the pass should run or not. Maybe we need something similar to avoid building all the required analyses.

LoopInterchange computes DependenceInfo itself, not the via pass manager, so it has full control over when (and whether) it gets computed.