On Sun, Oct 01, 2017 at 09:22:56AM -0600, Jeff Law wrote: > > A short while ago Martin Liska posted a patch that lowered certain > switch statements into cascading conditionals. > > His work tripped two regressions in the testsuite, both cases where we > did not optimize as well as we should have. > > Upon investigation I realized a simple improvement to DOM would fix > things up. Further testing showed that the situation occurred > reasonably often in practice and that the situation did occur even when > VRP was enabled. > > What we want to do is detect cases where we have something like this in > the expression hash table > > TRUE = (i <= 1) > > And we're presented with the condition we want to optimize like (i >= 1) > > The only value of i that satisfies both is i == 1 So we can change the > conditional to (i == 1). > > The simplified conditional is useful because it exposes a constant in > one arm of the conditional which we can propagate. Furthermore the > equality test is easier for tree-ssa-uninit.c to consume. > > The implementation is pretty simple. For X GE/LE Y we lookup X LE/GE Y > and if we get a hit then we know we can optimize the conditional to X == > Y. If Y is a constant, we can handle GT/LT with trivial canonicalization. > > The testcase I've added is potentially overly simplistic -- it'll likely > be compromised by work from Aldy or Andrew at some point. We'll likely > have to reduce a new testcase at that time.
Wouldn't testing this with the gimple fe make this much less of an issue? Thanks Trev
