[llvm] [RISCV][CostModel] Add getRISCVInstructionCost() to TTI for Cost… (PR #73651)

[Craig Topper via llvm-commits]

================
@@ -34,6 +34,63 @@ static cl::opt<unsigned> SLPMaxVF(
         "exclusively by SLP vectorizer."),
     cl::Hidden);
 
+InstructionCost
+RISCVTTIImpl::getRISCVInstructionCost(ArrayRef<unsigned> OpCodes, MVT VT,
+                                      TTI::TargetCostKind CostKind) {
+  size_t NumInstr = OpCodes.size();
+  if (CostKind == TTI::TCK_CodeSize)
+    return NumInstr;
+  InstructionCost LMULCost = TLI->getLMULCost(VT);
+  if ((CostKind != TTI::TCK_RecipThroughput) && (CostKind != TTI::TCK_Latency))
+    return LMULCost * NumInstr;
+  InstructionCost Cost = 0;
+  for (auto Op : OpCodes) {
+    switch (Op) {
+    case RISCV::VRGATHER_VI:
+      Cost += TLI->getVRGatherVICost(VT);
+      break;
+    case RISCV::VRGATHER_VV:
+      Cost += TLI->getVRGatherVVCost(VT);
+      break;
+    case RISCV::VSLIDEUP_VI:
+    case RISCV::VSLIDEDOWN_VI:
+    case RISCV::VSLIDEUP_VX:
----------------
topperc wrote:

.vx and .vi slides are quite different. The .vi instructions know at decode time how far to slide so hardware can know early which source DLEN pieces are needed for each DLEN piece of the result.

.vx requires the sources to determined after looking at the scalar register.

https://github.com/llvm/llvm-project/pull/73651