[PATCH] Constant fold undefined sqrt to undef instead of 0
resistor at mac.com
Tue Dec 17 20:52:19 PST 2013
It's more acceptable in the sense that it doesn't violate a universally quantified type constraint. I'd be a bit nervous that it might interact poorly with fast math mode, particularly the nnan bit.
> On Dec 17, 2013, at 6:17 PM, Matt Arsenault <Matthew.Arsenault at amd.com> wrote:
> Would NaN be a more acceptable alternative instead of undef? It matches the libm behavior and still is more obviously broken
More information about the llvm-commits