fix: denormal flush of inputs

[regehr]

this fixes more false alarms seen while running riscv-tv on the LLVM test suite -- no fuzzer involved, so far!

the change here is from LangRef: input flushing is mandatory, not optional

[nunoplopes]

Uhm, I forgot about this one. Wasn't this correct?

[regehr]

Re-opening this, please take a look at the reasoning below Nuno.

Here's an example that I believe alive should reject-- our master branch accepts it.

define i1 @src() denormal_fpenv(ieee|positivezero) {
  %r = fcmp oeq float 0x380FFFFFC0000000, 0.0
  ret i1 %r
}

define i1 @tgt() denormal_fpenv(ieee|positivezero) {
  ret i1 false
}

Here's the chain of reasoning:

0x380FFFFFC0000000 comes out to be the largest float denormal.

LangRef says:

"If the input mode is preservesign, or positivezero, a floating-point operation must treat any input denormal value as zero."

The fcmp clearly has an input denormal value, so the interpretation of this seems completely clear-- this can be folded to true, but not to false.

[nunoplopes]

Ok, I traced down the issue. There was a change in LangRef when the input/output flushing modes were split.