GH-55277: [C++][Gandiva] Fix stack overflow in decimal evaluation with JIT optimization disabled

cpp/src/gandiva/decimal_ir.cc

@@ -243,12 +243,18 @@ DecimalIR::ValueWithOverflow DecimalIR::AddWithOverflowCheck(const ValueFull& x,
   return ValueWithOverflow(sum_descaled, overflow);
 }
 
+std::pair<llvm::AllocaInst*, llvm::AllocaInst*> DecimalIR::CreateEntryBlockReturnSlots() {
+  auto& entry = ir_builder()->GetInsertBlock()->getParent()->getEntryBlock();
+  llvm::IRBuilder<> entry_builder(&entry, entry.begin());
+  auto out_high_ptr = entry_builder.CreateAlloca(types()->i64_type(), nullptr, "out_hi");
+  auto out_low_ptr = entry_builder.CreateAlloca(types()->i64_type(), nullptr, "out_low");
+  return {out_high_ptr, out_low_ptr};
+}
+
 // This is pretty complex, so use CPP fns.
 llvm::Value* DecimalIR::AddLarge(const ValueFull& x, const ValueFull& y,
                                  const ValueFull& out) {
-  auto block = ir_builder()->GetInsertBlock();
-  auto out_high_ptr = new llvm::AllocaInst(types()->i64_type(), 0, "out_hi", block);
-  auto out_low_ptr = new llvm::AllocaInst(types()->i64_type(), 0, "out_low", block);
+  auto [out_high_ptr, out_low_ptr] = CreateEntryBlockReturnSlots();
   auto x_split = ValueSplit::MakeFromInt128(this, x.value());
   auto y_split = ValueSplit::MakeFromInt128(this, y.value());
 
@@ -454,9 +460,9 @@ llvm::Value* DecimalIR::CallDecimalFunction(const std::string& function_name,
   llvm::Value* result = nullptr;
   if (return_type == i128) {
     // for i128 ret, replace with two int64* args, and join them.
-    auto block = ir_builder()->GetInsertBlock();
-    auto out_high_ptr = new llvm::AllocaInst(i64, 0, "out_hi", block);
-    auto out_low_ptr = new llvm::AllocaInst(i64, 0, "out_low", block);
+    // Decimal calls can be emitted inside the record loop or a CASE branch.
+    // Allocate return slots once per invocation, rather than once per row.
+    auto [out_high_ptr, out_low_ptr] = CreateEntryBlockReturnSlots();
     dis_assembled_args.push_back(out_high_ptr);
     dis_assembled_args.push_back(out_low_ptr);
 

cpp/src/gandiva/decimal_ir.h

@@ -19,6 +19,7 @@
 
 #include <memory>
 #include <string>
+#include <utility>
 #include <vector>
 
 #include "gandiva/function_ir_builder.h"
@@ -137,6 +138,9 @@ class DecimalIR : public FunctionIRBuilder {
   ValueWithOverflow AddWithOverflowCheck(const ValueFull& x, const ValueFull& y,
                                          const ValueFull& out);
 
+  // Allocate decimal return slots in the current function's entry block.
+  std::pair<llvm::AllocaInst*, llvm::AllocaInst*> CreateEntryBlockReturnSlots();
+
   // Do addition of large integers (both positive and negative).
   llvm::Value* AddLarge(const ValueFull& x, const ValueFull& y, const ValueFull& out);
 

cpp/src/gandiva/tests/decimal_test.cc

@@ -67,6 +67,42 @@ std::vector<Decimal128> TestDecimal::MakeDecimalVector(std::vector<std::string>
   return ret;
 }
 
+TEST_F(TestDecimal, CastLargeBatch) {
+  // Allocating decimal return slots in the record loop can exhaust the stack
+  // when optimization is disabled. A single int32-to-decimal cast is sufficient.
+  constexpr int kNumRecords = 1 << 19;
+  auto input_field = arrow::field("large_batch_decimal", arrow::int32());
+  auto schema = arrow::schema({input_field});
+  auto decimal_type = arrow::decimal128(10, 0);
+  auto decimal = TreeExprBuilder::MakeFunction(
+      "castDECIMAL", {TreeExprBuilder::MakeField(input_field)}, decimal_type);
+  auto expression =
+      TreeExprBuilder::MakeExpression(decimal, arrow::field("result", decimal_type));
+
+  std::vector<int32_t> values(kNumRecords);
+  std::vector<bool> validity(kNumRecords);
+  std::vector<Decimal128> expected_values(kNumRecords);
+  for (int i = 0; i < kNumRecords; ++i) {
+    values[i] = i % 17 - 8;
+    validity[i] = i % 11 != 0;
+    expected_values[i] = Decimal128(values[i]);
+  }
+  auto batch = arrow::RecordBatch::Make(schema, kNumRecords,
+                                        {MakeArrowArrayInt32(values, validity)});
+  auto expected = MakeArrowArrayDecimal(decimal_type, expected_values, validity);
+
+  for (bool optimize : {false, true}) {
+    SCOPED_TRACE(optimize);
+    auto config = std::make_shared<Configuration>(optimize);
+    std::shared_ptr<Projector> projector;
+    ASSERT_OK(Projector::Make(schema, {expression}, config, &projector));
+    arrow::ArrayVector output;
+    ASSERT_OK(projector->Evaluate(*batch, pool_, &output));
+    ASSERT_EQ(output.size(), 1);
+    EXPECT_ARROW_ARRAY_EQUALS(expected, output[0]);
+  }
+}
+
 TEST_F(TestDecimal, TestSimple) {
   // schema for input fields
   constexpr int32_t precision = 36;