This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH]: Fix places where we expect non-null loops in the looparray

On Wed, 8 Dec 2004, Zdenek Dvorak wrote:


get by accident (i.e. a lurking bug).

no, it is expected behavior. As for why I decided to do it this way -- so that the numbers of the loops do not change in the middle of an optimization (so you can for example store some information for loops in an array indexed by loop numbers). I do not think it is seriously used anywhere (except perhaps for scev, that store loop numbers in CHREC nodes).

SCEV also depends on the indexes being in a certain order (DFS order in this case). Is this behavior guaranteed?

it is in general preserved that a loop has higher number than all its superloops (except maybe in the "linear" pass -- I do not know whether you change the loop numbers when you swap them or not).

We can do more than just swap loops.
We can more or less arbitrarily change the loop nesting order.
To prevent problems caused by trying to figure out where a given loop went, and because the transformations involve more than just loop swapping, we never actually mess with the cfg.
We just rewrite the loop conditions, induction variables, and bodies so that they act as if they were swapped.
If we only did interchange, we of course would have just redirected some edges and changed the loop indexes.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]