create SCEVInfinite

[nlewycky]

Bugzilla Link	2546
Resolution	INVALID
Resolved on	Jul 15, 2008 00:34
Version	trunk
OS	All

Extended Description

Currently we use SCEVCouldNotCompute for cases where the analysis failed and when the loop was successfully analyzed as being infinite. We should split those apart and return a new SCEVInfinite type for loops that are provable infinite.

Internally, a SCEVInfinite early on should skip the brute force loop count computation. Once we're no longer wasting time brute forcing infinite loops, we can try cranking up the number of brute force iterations.

Also, some passes, like loop deletion, might want to see SCEVInfinite.

[lattner]

If the loop is provably infinite, it would be nice to get rid of the backedge comparison and make it an uncond branch. Does this really occur in practice? It seems that everyone uses:

while(1)...
for (;;) ...
etc, they don't write their infintite loops as:

while (unsigned i = ... ; i >= 0; --i)

-Chris

[lattner]

Even that case would be trivially foldable by instcombine. What provably infinite loops are you seeing?

[nlewycky]

Oh, wow, I didn't realize that. Indeed, I tried it out by modifying SCEV and there are none!