[PATCH] D149587: InstSimplify: Simplifications for ldexp

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Analysis/InstructionSimplify.cpp:6103
+  if (!IsStrict) {
+    // ldexp(undef, x) -> nan
+    if (Q.isUndefValue(Op0))
----------------
foad wrote:
> arsenm wrote:
> > arsenm wrote:
> > > foad wrote:
> > > > I still don't understand why this one isn't strictfp-safe, if you simplify -> qnan.
> > > if undef could be anything, it could have been a signaling nan that would demand quieting
> > Plus we evidently don't have agreement on how nan payload bits are supposed to work
> > it could have been a signaling nan
> That's like saying "it could have been 99.9". The point is we are *choosing* a particular value to refine it to, so why not choose a quiet nan?
> 
> > Plus we evidently don't have agreement on how nan payload bits are supposed to work
> OK, if we are unable to make a quiet nan because we don't know what bits to put in the payload, that seems like a good reason - but please add a comment to that effect since it is massively non-obvious. (Also if that is true then how does `C->makeQuiet()` below work??)
Make quiet just flips the bit of an existing nan which is obviously ok (ignoring old mips). This is synthesizing a new choice 


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D149587/new/

https://reviews.llvm.org/D149587