[llvm] [RISCV] Don't cost vector arithmetic fp ops as cheaper than scalar (PR #99594)

[Luke Lau via llvm-commits]

================
@@ -1739,8 +1738,14 @@ InstructionCost RISCVTTIImpl::getArithmeticInstrCost(
                                                            Op1Info, Op2Info,
                                                            Args, CxtI);
   }
-  return ConstantMatCost +
-         LT.first * getRISCVInstructionCost(Op, LT.second, CostKind);
+
+  InstructionCost InstrCost = getRISCVInstructionCost(Op, LT.second, CostKind);
+  // We use BasicTTIImpl to calculate scalar costs, which assumes floating point
+  // ops are twice as expensive as integer ops. Do the same for vectors so
+  // scalar floating point ops aren't cheaper than their vector equivalents.
+  if (Ty->isFPOrFPVectorTy())
----------------
lukel97 wrote:

That's what I did at first, but after grepping I found other non-arithmetic FP users of getRISCVInstructionCost like VFMV_V_F which had some observable cost model differences.

BasicTTIImpl only seems to double the cost for floats in getArithmeticInstrCost, so I wanted to keep it local to RISCVTTIImpl::getArithmeticInstrCost since that's the only place where we need the fp-cost-doubling to stay in sync with scalar costs

https://github.com/llvm/llvm-project/pull/99594