Pattern match a wrapper around llvm.canonicalize which
weakens the semantics to not require quieting signaling
nans. Depending on the denormal mode and FP type, we can
either drop the pattern entirely or reduce it only to
a canonicalize call. I'm inventing this pattern to deal
with LLVM's lax canonicalization model in math library
code.
The math library code currently has explicit checks for
the denormal mode, and conditionally canonicalizes the
result if there is flushing. Semantically, this could be
directly replaced with a simple call to llvm.canonicalize,
but doing so would incur an additional cost when using
standard IEEE behavior. If we do not care about quieting
a signaling nan, this should be a no-op unless the denormal
mode may flush. This will allow replacement of the
conditional code with a zero cost abstraction utility
function.
Note we need a standard LLVM floating-point operation
in the nan case to assert we do not care about preserving
the nan payload and sign bit. This could be any no-op fp
instruction; a normal choice would be fmul by 1.0. Using
that presents an ordering problem - since LLVM fp operations
are not required to canonicalize, instcombine would fold
out the fmul before reaching this select combine. The galaxy
brain solution here is to use fdiv 1.0, %x as the no-op.
This is not a no-op - it could potentially return infinity
if %x were 0 (or very close to 0) so it will not be dropped.
For the purposes here, that's fine since it's only being used
as a nan sink.
https://alive2.llvm.org/ce/z/QYS4en
7aaec28fde