[PATCH] D87391: [Intrinsics] define semantics for experimental fmax/fmin vector reductions

[Nikita Popov via Phabricator via llvm-commits]

nikic added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp:4805
     NeutralElem = DAG.getConstantFP(
-        std::numeric_limits<double>::infinity(), dl, ElemVT);
+        std::numeric_limits<double>::quiet_NaN(), dl, ElemVT);
     break;
----------------
spatel wrote:
> craig.topper wrote:
> > Do we need to drop nonan FMF then? Probably should have been dropping noinf before.
> > 
> > Do we have non-power of 2 tests for X86? X86 needs nonan to optimally lower fmaxnum/fminnum. But if you put a nan here then we shouldn't be using optimal lowering.
> Yes, we need to drop 'nnan' - otherwise this would create poison.
> No, we don't have non-pow-2 vector sizes in x86 tests from what I see. I'll add some.
Given how much X86 needs nnan for a decent lowering here, would it make sense to keep using +/- infinity if nnan is set, and only use qNan if it is not set?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D87391/new/

https://reviews.llvm.org/D87391