[llvm] [RISCV][CostModel] Estimate cost of llvm.vector.reduce.fmaximum/fminimum (PR #80697)

[Philip Reames via llvm-commits]

================
@@ -1001,6 +1001,53 @@ RISCVTTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty,
       return getArithmeticReductionCost(Instruction::And, Ty, FMF, CostKind);
   }
 
+  if (IID == Intrinsic::maximum || IID == Intrinsic::minimum) {
+    SmallVector<unsigned, 5> SplitOps;
+    SmallVector<unsigned, 3> Opcodes;
+    InstructionCost ExtraCost = 0;
+    switch (IID) {
+    case Intrinsic::maximum:
+      if (FMF.noNaNs()) {
+        SplitOps = {RISCV::VFMAX_VV};
+        Opcodes = {RISCV::VFREDMAX_VS, RISCV::VFMV_F_S};
+      } else {
+        SplitOps = {RISCV::VMFEQ_VV, RISCV::VMERGE_VVM, RISCV::VMFEQ_VV,
+                    RISCV::VMERGE_VVM, RISCV::VFMAX_VV};
+        Opcodes = {RISCV::VMFNE_VV, RISCV::VCPOP_M, RISCV::VFREDMAX_VS,
+                   RISCV::VFMV_F_S};
+        // Cost of Canonical Nan
+        // lui a0, 523264
+        // fmv.w.x fa0, a0
+        ExtraCost = 2;
----------------
preames wrote:

This depends on element type - a double is more expensive than the float.  Can we use generic immediate costing here?  (Not sure if we can - if not, 2 is not an unreasonable ballpark estimate.)

https://github.com/llvm/llvm-project/pull/80697