On Mon, Jul 13, 2015 at 10:48:19AM +0100, Kyrill Tkachov wrote: > For the testcase in the patch we were generating an extra neg instruction: > cmp w0, wzr > csneg w0, w0, w0, ge > neg w0, w0 > ret > > instead of the optimal: > cmp w0, wzr > csneg w0, w0, w0, lt > ret > > The reason is that combine tries to merge the operation into a negation of > an abs.
Before combine, you have two insns, a negation and an abs, so that is not so very strange :-) Some archs have actual nabs insns btw (for floating point, anyway). Archs without abs or conditional assignment, and with cheap branches, get a branch around a neg followed by another neg, at expand time. This then isn't optimised away either. So I'd say expand should be made a bit smarter for this. Failing that, your approach looks fine to me -- assuming you want to have a fake "abs" insn at all. On to the patch... > +;; Combine will try merging (c > 0 ? -x : x) into (-|x|). This isn't a good "x > 0" here. Segher
