[llvm] [LV] Vectorize histogram operations (PR #99851)

[Graham Hunter via llvm-commits]

================
@@ -6519,8 +6520,33 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I,
     // We've proven all lanes safe to speculate, fall through.
     [[fallthrough]];
   case Instruction::Add:
+  case Instruction::Sub: {
+    auto Info = Legal->getHistogramInfo(I);
+    if (Info && VF.isVector()) {
+      const HistogramInfo *HGram = Info.value();
+      // Assume that a non-constant update value (or a constant != 1) requires
+      // a multiply, and add that into the cost.
+      InstructionCost MulCost = TTI::TCC_Free;
+      ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1));
+      if (!RHS || RHS->getZExtValue() != 1)
+        MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
+
+      // Find the cost of the histogram operation itself.
+      Type *PtrTy = VectorType::get(HGram->Load->getPointerOperandType(), VF);
+      Type *ScalarTy = I->getType();
+      Type *MaskTy = VectorType::get(Type::getInt1Ty(I->getContext()), VF);
+      IntrinsicCostAttributes ICA(Intrinsic::experimental_vector_histogram_add,
+                                  Type::getVoidTy(I->getContext()),
+                                  {PtrTy, ScalarTy, MaskTy});
+
+      // Add the costs together with the add/sub operation.
+      return TTI.getIntrinsicInstrCost(
+                 ICA, TargetTransformInfo::TCK_RecipThroughput) +
+             MulCost + TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy);
+    }
+    [[fallthrough]];
+  }
----------------
huntergr-arm wrote:

Yes, that's still needed for now -- the two cost models disagree and assert if I remove that code.

https://github.com/llvm/llvm-project/pull/99851