From 30ab31cc4194f57866ba48753aeceae40e823d81 Mon Sep 17 00:00:00 2001
From: nihui <nihuini@tencent.com>
Date: Wed, 20 Jul 2022 18:16:28 +0800
Subject: [PATCH] add address sanitizer ci, fix potential memory leak shouted
 by asan (#4058)

---
 .github/workflows/linux-x64-cpu-gcc-san.yml | 42 +++++++++++++++++++++
 CMakeLists.txt                              |  1 +
 src/CMakeLists.txt                          |  5 +++
 src/layer/x86/cast_fp16.h                   |  8 ++--
 4 files changed, 52 insertions(+), 4 deletions(-)
 create mode 100644 .github/workflows/linux-x64-cpu-gcc-san.yml

diff --git a/.github/workflows/linux-x64-cpu-gcc-san.yml b/.github/workflows/linux-x64-cpu-gcc-san.yml
new file mode 100644
index 00000000000..ae57b37bc54
--- /dev/null
+++ b/.github/workflows/linux-x64-cpu-gcc-san.yml
@@ -0,0 +1,42 @@
+name: linux-x64-cpu-gcc-san
+on:
+  push:
+    branches: [master]
+    paths:
+    - '.github/workflows/linux-x64-cpu-gcc-san.yml'
+    - 'CMakeLists.txt'
+    - 'cmake/**'
+    - 'src/*'
+    - 'src/layer/*'
+    - 'src/layer/x86/**'
+    - 'tests/**'
+  pull_request:
+    branches: [master]
+    paths:
+    - '.github/workflows/linux-x64-cpu-gcc-san.yml'
+    - 'CMakeLists.txt'
+    - 'cmake/**'
+    - 'src/*'
+    - 'src/layer/*'
+    - 'src/layer/x86/**'
+    - 'tests/**'
+concurrency:
+  group: linux-x64-cpu-gcc-san-${{ github.ref }}
+  cancel-in-progress: true
+permissions:
+  contents: read
+
+jobs:
+  linux-gcc-san:
+    runs-on: ubuntu-22.04
+    steps:
+    - uses: actions/checkout@v3
+    - name: build
+      run: |
+        mkdir build && cd build
+        cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_ASAN=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..
+        cmake --build . -j 2
+    - name: test
+      run: |
+        cd build
+        ctest --output-on-failure -j 2
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 187d3b7ba7e..85b11ed616a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -79,6 +79,7 @@ option(NCNN_RUNTIME_CPU "runtime dispatch cpu routines" ON)
 option(NCNN_DISABLE_PIC "disable position-independent code" OFF)
 option(NCNN_BUILD_TESTS "build tests" OFF)
 option(NCNN_COVERAGE "build for coverage" OFF)
+option(NCNN_ASAN "build for address sanitizer" OFF)
 option(NCNN_BUILD_BENCHMARK "build benchmark" ON)
 option(NCNN_PYTHON "build python api" OFF)
 option(NCNN_INT8 "int8 inference" ON)
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index a5582fc8b96..279f24be7b3 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -471,6 +471,11 @@ if(NCNN_COVERAGE)
     target_link_libraries(ncnn PUBLIC -coverage -lgcov)
 endif()
 
+if(NCNN_ASAN)
+    target_compile_options(ncnn PUBLIC -fsanitize=address)
+    target_link_libraries(ncnn PUBLIC -fsanitize=address)
+endif()
+
 add_dependencies(ncnn ncnn-generate-spirv)
 
 if(NCNN_INSTALL_SDK)
diff --git a/src/layer/x86/cast_fp16.h b/src/layer/x86/cast_fp16.h
index fb07a504d6f..8fba9748c3b 100644
--- a/src/layer/x86/cast_fp16.h
+++ b/src/layer/x86/cast_fp16.h
@@ -78,7 +78,7 @@ static void cast_fp32_to_fp16_sse(const Mat& bottom_blob, Mat& top_blob, const O
         {
             __m128 _v_fp32 = _mm_loadu_ps(ptr);
             __m128h _v_fp16 = _mm_cvtxps_ph(_v_fp32);
-            _mm_storeu_si128((__m128i*)outptr, (__m128i)_v_fp16);
+            _mm_storel_epi64((__m128i*)outptr, (__m128i)_v_fp16);
 
             ptr += 4;
             outptr += 4;
@@ -97,7 +97,7 @@ static void cast_fp32_to_fp16_sse(const Mat& bottom_blob, Mat& top_blob, const O
         {
             __m128 _v_fp32 = _mm_loadu_ps(ptr);
             __m128i _v_fp16 = _mm_cvtps_ph(_v_fp32, _MM_FROUND_TRUNC);
-            _mm_storeu_si128((__m128i*)outptr, _v_fp16);
+            _mm_storel_epi64((__m128i*)outptr, _v_fp16);
 
             ptr += 4;
             outptr += 4;
@@ -164,7 +164,7 @@ static void cast_fp16_to_fp32_sse(const Mat& bottom_blob, Mat& top_blob, const O
         }
         for (; i + 3 < size; i += 4)
         {
-            __m128h _v_fp16 = (__m128h)_mm_loadu_si128((const __m128i*)ptr);
+            __m128h _v_fp16 = (__m128h)_mm_loadl_epi64((const __m128i*)ptr);
             __m128 _v_fp32 = _mm_cvtxph_ps(_v_fp16);
             _mm_storeu_ps(outptr, _v_fp32);
 
@@ -183,7 +183,7 @@ static void cast_fp16_to_fp32_sse(const Mat& bottom_blob, Mat& top_blob, const O
         }
         for (; i + 3 < size; i += 4)
         {
-            __m128i _v_fp16 = _mm_loadu_si128((const __m128i*)ptr);
+            __m128i _v_fp16 = _mm_loadl_epi64((const __m128i*)ptr);
             __m128 _v_fp32 = _mm_cvtph_ps(_v_fp16);
             _mm_storeu_ps(outptr, _v_fp32);