[llvm] X86: Improve cost model of fp16 conversion (PR #113195)

[Matthias Braun via llvm-commits]

================
@@ -3146,6 +3174,11 @@ InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
                             TTI::CastContextHint::None, CostKind);
   }
 
+  if (ISD == ISD::FP_ROUND && LTDest.second.getScalarType() == MVT::f16) {
+    // Conversion requires a libcall.
+    return InstructionCost::getInvalid();
----------------
MatzeB wrote:

I am not able to reproduce so far. Unfortunately the dump does not contain some of the referenced metadata so I have to make guesses for that. Then trying to run `opt -S -o - -passes=loop-vectorize /tmp/x.ll`  works just fine and I guess I need some target setup (I played with some `-mtriple=x86_64 -mattr=+avx512f,+f16c` but that doesn't repro either).

That said, could you try if replacing the `InstructionCost::getInvalid();` with `InstructionCost::getMax()` or if that doesn't work with a big number like 128 helps?

https://github.com/llvm/llvm-project/pull/113195