Avoid allocations when touching graph LRU entries

exec/CMakeLists.txt

@@ -17,6 +17,7 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Release)
 endif()
+include(CTest)
 
 find_package(CUDAToolkit REQUIRED)
 
@@ -33,6 +34,15 @@ target_include_directories(flashrt_exec
           ${CMAKE_CURRENT_SOURCE_DIR}/backend)
 target_link_libraries(flashrt_exec PUBLIC CUDA::cudart)
 
+if(BUILD_TESTING)
+  add_executable(test_graph_lru_alloc tests/test_graph_lru_alloc.cpp)
+  target_include_directories(test_graph_lru_alloc PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/src)
+  target_link_libraries(test_graph_lru_alloc PRIVATE flashrt_exec)
+  add_test(NAME graph_lru_alloc COMMAND test_graph_lru_alloc)
+endif()
+
 # --- pybind dev module (optional) ---
 find_package(Python3 COMPONENTS Interpreter Development.Module QUIET)
 find_package(pybind11 CONFIG QUIET)

exec/src/graph.cpp

@@ -4,7 +4,10 @@
 
 void frt_graph_s::touch(frt_shape_key key) {
     for (auto it = lru.begin(); it != lru.end(); ++it) {
-        if (*it == key) { lru.erase(it); break; }
+        if (*it == key) {
+            lru.splice(lru.end(), lru, it);
+            return;
+        }
     }
     lru.push_back(key);  // back = most recently used
 }

exec/tests/test_graph_lru_alloc.cpp

@@ -0,0 +1,61 @@
+#include "internal.h"
+
+#include <atomic>
+#include <cstdlib>
+#include <iostream>
+#include <new>
+
+namespace {
+
+std::atomic<bool> count_allocations{false};
+std::atomic<std::size_t> allocation_count{0};
+
+[[noreturn]] void fail(const char* expression, int line) {
+    std::cerr << "FAIL line " << line << ": " << expression << '\n';
+    std::abort();
+}
+
+#define CHECK(expression)                                \
+    do {                                                 \
+        if (!(expression)) fail(#expression, __LINE__); \
+    } while (false)
+
+}  // namespace
+
+void* operator new(std::size_t bytes) {
+    if (count_allocations.load(std::memory_order_relaxed)) {
+        allocation_count.fetch_add(1, std::memory_order_relaxed);
+    }
+    if (void* pointer = std::malloc(bytes)) return pointer;
+    throw std::bad_alloc();
+}
+
+void operator delete(void* pointer) noexcept { std::free(pointer); }
+void operator delete(void* pointer, std::size_t) noexcept {
+    std::free(pointer);
+}
+
+int main() {
+    frt_graph_s graph;
+    graph.lru = {1, 2, 3};
+
+    allocation_count.store(0, std::memory_order_relaxed);
+    count_allocations.store(true, std::memory_order_relaxed);
+    for (int i = 0; i < 1000; ++i) graph.touch((i % 3) + 1);
+    count_allocations.store(false, std::memory_order_relaxed);
+
+    CHECK(allocation_count.load(std::memory_order_relaxed) == 0);
+    CHECK(graph.lru.size() == 3);
+    CHECK(graph.lru.back() == 1);
+
+    allocation_count.store(0, std::memory_order_relaxed);
+    count_allocations.store(true, std::memory_order_relaxed);
+    graph.touch(4);
+    count_allocations.store(false, std::memory_order_relaxed);
+    CHECK(allocation_count.load(std::memory_order_relaxed) > 0);
+    CHECK(graph.lru.size() == 4);
+    CHECK(graph.lru.back() == 4);
+
+    std::cout << "PASS - graph LRU touch allocation contract\n";
+    return 0;
+}