Merge pull request #590 from Sonnexo/master

Introduce subsystems for memory and loop

DiscoPoP/DiscoPoP.cpp

@@ -2004,7 +2004,7 @@ void DiscoPoP::dp_reduction_insert_functions() {
   llvm::FunctionType *loop_incr_fn_type = llvm::FunctionType::get(
       llvm::Type::getVoidTy(*ctx_), loop_incr_fn_args, false);
   FunctionCallee incr_loop_counter_callee =
-      module_->getOrInsertFunction("incr_loop_counter", loop_incr_fn_type);
+      module_->getOrInsertFunction("__dp_loop_incr", loop_incr_fn_type);
 
   for (auto const &loop_info : loops_) {
     llvm::Value *val =
@@ -2019,11 +2019,11 @@ void DiscoPoP::dp_reduction_insert_functions() {
   loop_metadata_file.close();
 
   // add a function to output the final data
-  // loop_counter_output
+  // dp_loop_output
   llvm::FunctionType *output_fn_type =
       llvm::FunctionType::get(llvm::Type::getVoidTy(*ctx_), false);
   FunctionCallee loop_counter_output_callee =
-      module_->getOrInsertFunction("loop_counter_output", output_fn_type);
+      module_->getOrInsertFunction("__dp_loop_output", output_fn_type);
   FunctionCallee cu_taken_branch_counter_output_callee =
       module_->getOrInsertFunction("__dp_taken_branch_counter_output",
                                    output_fn_type);

benchmark/CMakeLists.txt

@@ -32,6 +32,7 @@ target_sources(
   DiscoPoP_BM
   PRIVATE 
   benchmarks.cpp
+  memory_region_tree/benchmark_memory_region_tree.cpp
   perfect_shadow/benchmark_perfect_shadow.cpp
   scope/benchmark_scope.cpp)
 

benchmark/memory_region_tree/benchmark_memory_region_tree.cpp

@@ -0,0 +1,193 @@
+/*
+ * This file is part of the DiscoPoP software
+ * (http://www.discopop.tu-darmstadt.de)
+ *
+ * Copyright (c) 2020, Technische Universitaet Darmstadt, Germany
+ *
+ * This software may be modified and distributed under the terms of
+ * the 3-Clause BSD License. See the LICENSE file in the package base
+ * directory for details.
+ *
+ */
+
+#include <benchmark/benchmark.h>
+
+#include <algorithm>
+#include <cstdint>
+#include <random>
+#include <vector>
+
+#include "../../rtlib/memory/MemoryRegionTree.hpp"
+
+// General functions
+
+static std::vector<ADDR> convert_to_address(const std::int64_t number_iterations) {
+    auto mt = std::mt19937{0};
+    auto uid = std::uniform_int_distribution<ADDR>{0, 0x7FFFFFFFFFFFFFFF};
+
+    auto addresses = std::vector<ADDR>{};
+    addresses.resize(number_iterations);
+
+    for (auto i = std::int64_t(0); i < number_iterations; i++) {
+        addresses[i] = uid(mt);
+    }
+
+    std::sort(addresses.begin(), addresses.end());
+
+    return addresses;
+}
+
+// Benchmarks for old version (i.e., establishing a base line)
+
+static void benchmark_mrt_allocate_region(benchmark::State& state) {
+    const auto number_iterations = state.range(0);
+
+    const auto addresses = convert_to_address(number_iterations * 2);
+
+    // This exists so that the destructor call does not interfere with the timing
+    auto dumping_ground = std::vector<__dp::MemoryRegionTree>{};
+
+    for (auto _ : state) {
+        state.PauseTiming();
+        auto tree = __dp::MemoryRegionTree{};
+        state.ResumeTiming();
+
+        for (auto i = 0; i < number_iterations * 2; i += 2) {
+            tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+        }
+
+        state.PauseTiming();
+        dumping_ground.emplace_back(std::move(tree));
+        state.ResumeTiming();
+    }
+}
+
+static void benchmark_mrt_get_memory_region_id(benchmark::State& state) {
+    const auto number_iterations = state.range(0);
+
+    const auto addresses = convert_to_address(number_iterations * 2);
+
+    auto tree = __dp::MemoryRegionTree{};
+
+    for (auto i = 0; i < number_iterations * 2; i += 2) {
+        tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+    }
+
+    for (auto _ : state) {
+        for (auto i = 0; i < number_iterations * 2; i++) {
+            benchmark::DoNotOptimize(tree.get_memory_region_id(addresses[i]));
+        }
+    }
+}
+
+static void benchmark_mrt_get_memory_region_id_string_found(benchmark::State& state) {
+    const auto number_iterations = state.range(0);
+
+    const auto addresses = convert_to_address(number_iterations * 2);
+
+    auto tree = __dp::MemoryRegionTree{};
+
+    for (auto i = 0; i < number_iterations * 2; i += 2) {
+        tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+    }
+
+    // This exists so that the destructor call does not interfere with the timing
+    auto dumping_ground = std::vector<std::string>{};
+    dumping_ground.reserve(number_iterations);
+
+    for (auto _ : state) {
+        for (auto i = 0; i < number_iterations * 2; i++) {
+            dumping_ground.emplace_back(tree.get_memory_region_id_string(addresses[i], "fallback"));
+        }
+    }
+}
+
+static void benchmark_mrt_get_memory_region_id_string_fallback(benchmark::State& state) {
+    const auto number_iterations = state.range(0);
+
+    const auto addresses = convert_to_address(number_iterations * 2);
+
+    auto tree = __dp::MemoryRegionTree{};
+
+    for (auto i = 0; i < number_iterations * 2; i += 2) {
+        tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+    }
+
+    // This exists so that the destructor call does not interfere with the timing
+    auto dumping_ground = std::vector<std::string>{};
+    dumping_ground.reserve(number_iterations);
+
+    for (auto _ : state) {
+        for (auto i = 0; i < number_iterations * 2; i++) {
+            const auto base_address = addresses[i];
+            const auto address = (i % 2 == 0) ? base_address - 1 : base_address + 1;
+
+            dumping_ground.emplace_back(tree.get_memory_region_id_string(address, "fallback"));
+        }
+    }
+}
+
+static void benchmark_mrt_destructor(benchmark::State& state) {
+    const auto number_iterations = state.range(0);
+
+    const auto addresses = convert_to_address(number_iterations * 2);
+
+    for (auto _ : state) {
+        state.PauseTiming();
+        auto tree = __dp::MemoryRegionTree{};
+        for (auto i = 0; i < number_iterations * 2; i += 2) {
+            tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+        }
+        state.ResumeTiming();
+    }
+}
+
+static void benchmark_mrt_free_region(benchmark::State& state) {
+    const auto number_iterations = state.range(0);
+
+    const auto addresses = convert_to_address(number_iterations * 2);
+
+    // This exists so that the destructor call does not interfere with the timing
+    auto dumping_ground = std::vector<__dp::MemoryRegionTree>{};
+
+    auto tree = __dp::MemoryRegionTree{};
+
+    for (auto i = 0; i < number_iterations * 2; i += 2) {
+        tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+    }
+
+    for (auto _ : state) {
+        state.PauseTiming();
+        auto tree = __dp::MemoryRegionTree{};
+        for (auto i = 0; i < number_iterations * 2; i += 2) {
+            tree.allocate_region(addresses[i], addresses[i + 1], i + 1);
+        }
+        state.ResumeTiming();
+
+        for (auto i = 0; i < number_iterations * 2; i += 2) {
+            tree.free_region(addresses[i]);
+        }
+
+        state.PauseTiming();
+        dumping_ground.emplace_back(std::move(tree));
+        state.ResumeTiming();
+    }
+}
+
+BENCHMARK(benchmark_mrt_allocate_region)->Unit(benchmark::kMillisecond)->Arg(32)->Iterations(100);
+BENCHMARK(benchmark_mrt_allocate_region)->Unit(benchmark::kMillisecond)->Arg(1024)->Iterations(100);
+
+BENCHMARK(benchmark_mrt_get_memory_region_id)->Unit(benchmark::kMillisecond)->Arg(32)->Iterations(100);
+BENCHMARK(benchmark_mrt_get_memory_region_id)->Unit(benchmark::kMillisecond)->Arg(1024)->Iterations(100);
+
+BENCHMARK(benchmark_mrt_get_memory_region_id_string_found)->Unit(benchmark::kMillisecond)->Arg(32)->Iterations(100);
+BENCHMARK(benchmark_mrt_get_memory_region_id_string_found)->Unit(benchmark::kMillisecond)->Arg(1024)->Iterations(100);
+
+BENCHMARK(benchmark_mrt_get_memory_region_id_string_fallback)->Unit(benchmark::kMillisecond)->Arg(32)->Iterations(100);
+BENCHMARK(benchmark_mrt_get_memory_region_id_string_fallback)->Unit(benchmark::kMillisecond)->Arg(1024)->Iterations(100);
+
+BENCHMARK(benchmark_mrt_destructor)->Unit(benchmark::kMillisecond)->Arg(32)->Iterations(100);
+BENCHMARK(benchmark_mrt_destructor)->Unit(benchmark::kMillisecond)->Arg(1024)->Iterations(100);
+
+BENCHMARK(benchmark_mrt_free_region)->Unit(benchmark::kMillisecond)->Arg(32)->Iterations(100);
+BENCHMARK(benchmark_mrt_free_region)->Unit(benchmark::kMillisecond)->Arg(1024)->Iterations(100);

benchmark/perfect_shadow/benchmark_perfect_shadow.cpp

@@ -15,7 +15,7 @@
 #include <cstdint>
 #include <vector>
 
-#include "../../rtlib/perfect_shadow.hpp"
+#include "../../rtlib/memory/PerfectShadow.hpp"
 
 // General functions
 

benchmark/scope/benchmark_scope.cpp

@@ -15,7 +15,7 @@
 #include <cstdint>
 #include <vector>
 
-#include "../../rtlib/scope.hpp"
+#include "../../rtlib/memory/Scope.hpp"
 
 // General functions
 

docs/setup/discopop.md

@@ -1,39 +1,39 @@
----
-layout: default
-title: DiscoPoP
-parent: Setup
-nav_order: 1
----
-
-# DiscoPoP Setup
-## Prerequisites
-- LLVM/clang version 11
-- Python version 3.6 or greater
-
-## Setup
-```
-git clone git@github.com:discopop-project/discopop.git
-cd discopop
-mkdir build
-```
-
-## Build libraries and install Python modules
-```
-cd build
-cmake .. <CMAKE_FLAGS>
-make
-cd ..
-```
-
-where `<CMAKE_FLAGS>` can consist of any combination of the following flags and commonly used CMAKE_FLAGS:
-- In case you want to use a specific Version of LLVM, it is possible to specify the `-DUSE_LLVM_VERSION=<version>` flag.
-- In case you want to use a specific LLVM installation, specify the location via the `-DLLVM_DIST_PATH=<llvm_base_dir>` flag.
-- In case your application uses PThreads, please specify `-DDP_PTHREAD_COMPATIBILITY_MODE=[0|1]`. Note, however, that this can influence the runtime of the profiling.
-- In case you require a more verbose output of the runtime library, specify the `-DDP_RTLIB_VERBOSE=[0|1]` flag.
-- In case you want to specify the number of Workers available for the profiling step, specify the `-DDP_NUM_WORKERS=<int>` flag. By default, `3` worker threads are used to analyze the observed memory accesses. `0` might be used to disable the creation of additional threads for the analysis.
-
-## Testing the installation
-To test the installation, it is possible to execute the provided set of unit tests.
-```
-python -m unittest -v
-```
+---
+layout: default
+title: DiscoPoP
+parent: Setup
+nav_order: 1
+---
+
+# DiscoPoP Setup
+## Prerequisites
+- LLVM/clang version 11
+- Python version 3.6 or greater
+
+## Setup
+```
+git clone git@github.com:discopop-project/discopop.git
+cd discopop
+mkdir build
+```
+
+## Build libraries and install Python modules
+```
+cd build
+cmake .. <CMAKE_FLAGS>
+make
+cd ..
+```
+
+where `<CMAKE_FLAGS>` can consist of any combination of the following flags and commonly used CMAKE_FLAGS:
+- In case you want to use a specific Version of LLVM, it is possible to specify the `-DUSE_LLVM_VERSION=<version>` flag.
+- In case you want to use a specific LLVM installation, specify the location via the `-DLLVM_DIST_PATH=<llvm_base_dir>` flag.
+- In case your application uses PThreads, please specify `-DDP_PTHREAD_COMPATIBILITY_MODE=[0|1]`. Note, however, that this can influence the runtime of the profiling.
+- In case you require a more verbose output of the runtime library, specify the `-DDP_RTLIB_VERBOSE=[0|1]` flag.
+- In case you want to specify the number of Workers available for the profiling step, specify the `-DDP_NUM_WORKERS=<int>` flag. By default, `3` worker threads are used to analyze the observed memory accesses. `0` might be used to disable the creation of additional threads for the analysis.
+
+## Testing the installation
+To test the installation, it is possible to execute the provided set of unit tests.
+```
+python -m unittest -v
+```

rtlib/CMakeLists.txt

@@ -14,28 +14,29 @@
 set(DiscoPoP_SOURCES
         iFunctions.cpp
         iFunctionsGlobals.cpp
-        iFunctionsTypes.cpp
-        signature.cpp
-        loop_counter.cpp
-        cu_taken_branch_counter.cpp
-        ../share/lib/timer.cpp
-        MemoryRegionTree.cpp
+
+        memory/MemoryManager.cpp
+        memory/Signature.cpp
 
-	functions/dp_add_bb_deps.cpp
-	functions/dp_alloca.cpp
-	functions/dp_call.cpp
-	functions/dp_decl.cpp
-	functions/dp_delete.cpp
-	functions/dp_finalize.cpp
-	functions/dp_func_entry.cpp
-	functions/dp_func_exit.cpp
-	functions/dp_loop_entry.cpp
-	functions/dp_loop_exit.cpp
-	functions/dp_new.cpp
-	functions/dp_read.cpp
-	functions/dp_report_bb.cpp
-	functions/dp_report_bb_pair.cpp
-	functions/dp_write.cpp
+        injected_functions/dp_add_bb_deps.cpp
+        injected_functions/dp_alloca.cpp
+        injected_functions/dp_call.cpp
+        injected_functions/dp_decl.cpp
+        injected_functions/dp_delete.cpp
+        injected_functions/dp_finalize.cpp
+        injected_functions/dp_func_entry.cpp
+        injected_functions/dp_func_exit.cpp
+        injected_functions/dp_incr_taken_branch_counter.cpp
+        injected_functions/dp_loop_entry.cpp
+        injected_functions/dp_loop_exit.cpp
+        injected_functions/dp_loop_incr.cpp
+        injected_functions/dp_loop_output.cpp
+        injected_functions/dp_new.cpp
+        injected_functions/dp_read.cpp
+        injected_functions/dp_report_bb.cpp
+        injected_functions/dp_report_bb_pair.cpp
+        injected_functions/dp_taken_branch_counter_output.cpp
+        injected_functions/dp_write.cpp
         )
 
 set(CMAKE_CXX_FLAGS
@@ -70,4 +71,4 @@ install(TARGETS DiscoPoP_RT ARCHIVE DESTINATION lib)
 # compile simple-alias-detection
 #exec_program(${CMAKE_CURRENT_SOURCE_DIR}/simple-alias-detection/compile.sh
 #    ARGS ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR}
-#)
+#)

rtlib/DPTypes.hpp

@@ -25,6 +25,12 @@
 
 // #define SKIP_DUP_INSTR 1
 
+// To manually enable/disable internal timing
+// #define DP_SKIP_INTERNAL_TIMER
+#ifndef DP_SKIP_INTERNAL_TIMER
+#define DP_INTERNAL_TIMER
+#endif
+
 typedef std::int64_t LID;
 typedef std::int64_t ADDR;
 typedef std::int64_t sigElement;