Testbench add zeonica_testbench submodule and some other tests

.gitmodules

@@ -2,3 +2,7 @@
 	path = test/benchmark/CGRA-Bench
 	url = https://github.com/tancheng/CGRA-Bench.git
 	ignore = all
+[submodule "test-benchmark-zeonica-testbench"]
+	path = test/benchmark/Zeonica_Testbench
+	url = https://github.com/sarchlab/Zeonica_Testbench.git
+	ignore = all

README.md

@@ -56,3 +56,15 @@ Build LLVM & Neura
  $ /workspace/llvm-project/build/bin/llvm-lit test.mlir -v
 ```
 
+Sync `test/e2e` outputs into Zeonica_Testbench (submodule)
+--------------------------------------------------------
+This repo vendors [`sarchlab/Zeonica_Testbench`](https://github.com/sarchlab/Zeonica_Testbench.git) as a git submodule at `test/benchmark/Zeonica_Testbench`.
+
+After running e2e compilation/tests, you can sync the generated artifacts into the testbench repo:
+```sh
+$ ./tools/sync_e2e_outputs_to_zeonica_testbench.sh
+```
+
+Mapping (per kernel `K`):
+- `test/e2e/K/tmp-generated-dfg.{dot,yaml}` -> `test/benchmark/Zeonica_Testbench/kernel/K/K-dfg.{dot,yaml}`
+- `test/e2e/K/tmp-generated-instructions.{asm,yaml}` -> `test/benchmark/Zeonica_Testbench/kernel/K/K-instructions.{asm,yaml}`

lib/NeuraDialect/Transforms/GenerateCodePass.cpp

@@ -595,8 +595,8 @@ struct GenerateCodePass
         setUniqueDestination(pi, producer_direction.str());
       else if (!regs.empty())
         setUniqueDestination(pi, "$" + std::to_string(regs.back().regId));
+      }
     }
-  }
 
   // Egress: on source tile at first_link_ts, move [$src_reg] -> [out_dir].
   void placeSrcEgress(const Topology &topology, int src_tile_id, int time_step,
@@ -706,7 +706,8 @@ struct GenerateCodePass
         const bool should_rewire_to_register =
             IsCtrl || (consumer_placement.time_step > deposit_time_step);
         if (should_rewire_to_register) {
-          setConsumerSourceExact(consumer_operation, value_at_consumer, "$" + std::to_string(register_id));
+          setConsumerSourceExact(consumer_operation, value_at_consumer,
+                                 "$" + std::to_string(register_id));
           return true;
         }
       }
@@ -804,7 +805,7 @@ struct GenerateCodePass
     } else {
       return collectDataMovConsumers(forwarder);
     }
-  }
+    }
 
   template<bool IsCtrl>
   void rewriteMovConsumers(Operation *forwarder,
@@ -1491,7 +1492,8 @@ struct GenerateCodePass
                          return a->time_step < b->time_step;
                        });
 
-      yaml_out << "            - index_per_ii: " << index_per_ii << "\n              operations:\n";
+      yaml_out << "            - index_per_ii: " << index_per_ii
+               << "\n              operations:\n";
       for (const Instruction *inst : operations) {
         yaml_out << "                - opcode: \"" << inst->opcode << "\"\n";
         if (inst->id >= 0)
@@ -1502,13 +1504,15 @@ struct GenerateCodePass
         if (!inst->src_operands.empty()) {
           yaml_out << "                  src_operands:\n";
           for (const Operand &opnd : inst->src_operands)
-            yaml_out << "                    - operand: \"" << opnd.operand << "\"\n                      color: \"" << opnd.color << "\"\n";
+            yaml_out << "                    - operand: \"" << opnd.operand
+                     << "\"\n                      color: \"" << opnd.color << "\"\n";
         }
         // destinations.
         if (!inst->dst_operands.empty()) {
           yaml_out << "                  dst_operands:\n";
           for (const Operand &opnd : inst->dst_operands)
-            yaml_out << "                    - operand: \"" << opnd.operand << "\"\n                      color: \"" << opnd.color << "\"\n";
+            yaml_out << "                    - operand: \"" << opnd.operand
+                     << "\"\n                      color: \"" << opnd.color << "\"\n";
         }
       }
     }
@@ -1577,7 +1581,8 @@ struct GenerateCodePass
       for (size_t i = 0; i < instructions.size(); ++i) {
         const Instruction *inst = instructions[i];
         asm_out << "  " << inst->opcode;
-        for (const Operand &operand : inst->src_operands) asm_out << ", " << formatOperand(operand);
+        for (const Operand &operand : inst->src_operands)
+          asm_out << ", " << formatOperand(operand);
         if (!inst->dst_operands.empty()) {
           asm_out << " -> ";
           for (size_t j = 0; j < inst->dst_operands.size(); ++j) {

test/benchmark/axpy/axpy_int.cpp

@@ -0,0 +1,17 @@
+// Keep the function name matching "kernel" so llvm-extract --rfunc=".*kernel.*" can find it.
+extern "C" void kernel_axpy_int(int n, int a, const int *x, int *y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = a * x[i] + y[i];
+  }
+}
+
+// Provide a tiny main so clang emits a complete TU; tests extract only the kernel anyway.
+int main() {
+  const int N = 16;
+  static int x[N];
+  static int y[N];
+  kernel_axpy_int(N, /*a=*/3, x, y);
+  return 0;
+}
+
+

test/benchmark/gemv/gemv_int.cpp

@@ -0,0 +1,22 @@
+// A simple int GEMV: y = A*x + y.
+// Use fixed small sizes to keep the generated IR stable-ish.
+extern "C" void kernel_gemv_int(const int *A, const int *x, int *y) {
+  const int N = 4;
+  for (int i = 0; i < N; ++i) {
+    int acc = 0;
+    for (int j = 0; j < N; ++j) {
+      acc += A[i * N + j] * x[j];
+    }
+    y[i] = acc;
+  }
+}
+
+int main() {
+  static int A[16];
+  static int x[4];
+  static int y[4];
+  kernel_gemv_int(A, x, y);
+  return 0;
+}
+
+