In order to avoid a number of redundant sign/zero-extensions, we have
to custom combine AssertSext on our target and fold a generic pattern
in the target independent DAG combiner for the SIGN_EXTEND_INREG node.
Depends on D10970
vkalintiris on Jan 15 2016, 7:09 AM.Authored by
Well caught. It turns out we've been falling through to the O32 handler in some cases.
There's a few changes that don't look right to me and I've commented on these below. I'd like to explain one set of the comments in advance. Up until select-flt.ll I've pointed out a number of redundant instructions. These look like regressions but it's possible that we've always emitted some of them and have just not mentioned them in the test. Could you check whether these are regressions or not?