Liuye Notebook
0%

Cpp-Trivial

发表于 更新于 分类于 阅读次数： Valine：

阅读更多

1 Compile Process

flowchart TD
    other_target(["其他目标代码"])
    linker[["链接器"]]
    source(["源代码"])
    compiler[["编译器"]]
    assembly(["汇编代码"])
    assembler[["汇编器"]]
    target(["目标代码"])
    lib(["库文件"])
    result_target(["目标代码"])
    exec(["可执行程序"])
    result_lib(["库文件"])

    other_target --> linker

    subgraph 编译过程
        source --> compiler
        compiler --> assembly
        assembly --> assembler
        assembler --> target
        end
        target --> linker
        lib --> linker

    linker --> result_target
    linker --> exec
    linker --> result_lib

2 Library

2.1 Static Library

后缀：*.a

编译时如何链接静态链接库：

  • -L：指定静态链接库的搜索路径
  • -l <static_lib_name>：
    • 假设静态链接库的名称是libgflags.a，那么<static_lib_name>既不需要lib前缀，也不需要.a后缀，即gflags

如何查看二进制的静态链接库：由于链接器在链接时，就会丢弃静态库的名字信息，因此，一般是看不到的

  • -Xlinker -Map=a.map：将链接时的信息记录到a.map中
  • nm/objdump/readelf/strings或许可以找到一些静态库相关的hint

2.2 Dynamic Library

后缀：*.so

如何查看二进制的动态链接库：ldd (list dynamic dependencies)

查看动态链接库绑定信息：ldconfig -v、ldconfig -p

2.2.1 Linux’s so version mechanism

本小节转载摘录自一文读懂Linux下动态链接库版本管理及查找加载方式

在/lib64、/usr/lib64、/usr/local/lib64目录下，会看到很多具有下列特征的软连接，其中x、y、z为数字, 那么这些软连接和他们后面的数字有什么用途呢？

1
2
3
libfoo.so    ->  libfoo.so.x
libfoo.so.x  ->  libfoo.so.x.y.z
libbar.so.x  ->  libbar.so.x.y

这里的x、y、z分别代表的是这个so的主版本号（MAJOR），次版本号（MINOR），以及发行版本号（RELEASE），对于这三个数字各自的含义，以及什么时候会进行增长，不同的文献上有不同的解释，不同的组织遵循的规定可能也有细微的差别，但有一个可以肯定的事情是：主版本号（MAJOR）不同的两个so库，所暴露出的API接口是不兼容的。而对于次版本号，和发行版本号，则有着不同定义，其中一种定义是：次要版本号表示API接口的定义发生了改变（比如参数的含义发生了变化），但是保持向前兼容；而发行版本号则是函数内部的一些功能调整、优化、BUG修复，不涉及API接口定义的修改

介绍一下在so查找过程中的几个名字

  • SONAME：一组具有兼容API的so库所共有的名字，其命名特征是lib<库名>.so.<数字>这种形式的
  • real name：是真实具有so库可执行代码的那个文件，之所以叫real是相对于SONAME和linker name而言的，因为另外两种名字一般都是一个软连接，这些软连接最终指向的文件都是具有real name命名形式的文件。real name的命名格式中，可能有2个数字尾缀，也可能有3个数字尾缀，但这不重要。你只要记住，真实的那个，不是以软连接形式存在的，就是一个real name
  • linker name：这个名字只是给编译工具链中的连接器使用的名字，和程序运行并没有什么关系，仅仅在链接得到可执行文件的过程中才会用到。它的命名特征是以lib开头，以.so结尾，不带任何数字后缀的格式

2.2.1.2 SONAME

假设在你的Linux系统中有3个不同版本的bar共享库，他们在磁盘上保存的文件名如下：

  • /usr/lib64/libbar.so.1.3
  • /usr/lib64/libbar.so.1.5
  • /usr/lib64/libbar.so.2.1

假设以上三个文件，都是真实的so库文件，而不是软连接，也就是说，上面列出的名字都是real name

根据我们之前对版本号的定义，我们可以知道：

  • libbar.so.1.3和libbar.so.1.5之间是互相兼容的
  • libbar.so.2.1和上述两个库之间互相不兼容

引入软连接的好处是什么呢？假设有一天，libbar.so.2.1库进行了升级，但API接口仍然保持兼容，升级后的库文件为libbar.so.2.2，这时候，我们只要将之前的软连接重新指向升级后的文件，然后重新启动B程序，B程序就可以使用全新版本的so库了，我们并不需要去重新编译链接来更新B程序

总结一下上面的逻辑：

  • 通常SONAME是一个指向real name的软连接
  • 应用程序中存储自己所依赖的so库的SONAME，也就是仅保证主版本号能匹配就行
  • 通过修改软连接的指向，可以让应用程序在互相兼容的so库中方便切换使用哪一个
  • 通常情况下，大家使用最新版本即可，除非是为了在特定版本下做一些调试、开发工作

2.2.1.3 linker name

上一节中我们提到，可执行文件里会存储精确到主版本号的SONAME，但是在编译生成可执行文件的过程中，编译器怎么知道应该用哪个主版本号呢？为了回答这个问题，我们从编译链接的过程来梳理一下

假设我们使用gcc编译生成一个依赖foo库的可执行文件A：gcc A.c -lfoo -o A

熟悉gcc编译的读者们肯定知道，上述的-l标记后跟随了foo参数，表示我们告诉gcc在编译的过程中需要用到一个外部的名为foo的库，但这里有一个问题，我们并没有说使用哪一个主版本，我们只给出了一个名字。为了解决这个问题，软链接再次发挥作用，具体流程如下：

根据linux下动态链接库的命名规范，gcc会根据-lfoo这个标识拼接出libfoo.so这个文件名，这个文件名就是linker name，然后去尝试读取这个文件，并将这个库链接到生成的可执行文件A中。在执行编译前，我们可以通过软链接的形式，将libfoo.so指向一个具体so库，也就是指向一个real name，在编译过程中，gcc会从这个真实的库中读取出SONAME并将它写入到生成的可执行文件A中。例如，若libfoo.so指向libfoo.so.1.5,则生成的可执行文件A使用主版本号为1的SONAME，即libfoo.so.1

在上述编译过程完成之后，SONAME已经被写入可执行文件A中了，因此可以看到linker name仅仅在编译的过程中，可以起到指定连接那个库版本的作用，除此之外，再无其他作用

总结一下上面的逻辑：

  • 通常linker name是一个指向real name的软连接
  • 通过修改软连接的指向，可以指定编译生成的可执行文件使用那个主版本号so库
  • 编译器从软链接指向的文件里找到其SONAME，并将SONAME写入到生成的可执行文件中
  • 通过改变linker name软连接的指向，可以将不同主版本号的SONAME写入到生成的可执行文件中

2.3 Search Order

2.3.1 Compilation Time

During compilation, the order in which the linker (such as GNU’s ld) searches for library files follows certain rules, ensuring that the linker can find and link the correct version of the library. The search order typically goes as follows:

  • Library files specified directly in the command line: If the compilation command directly specifies the full path to the library files, the linker will first use these specified paths. For example, gcc main.c /path/to/libmylibrary.a will link directly to the library at the specified path.
  • Directories specified with the -L option: The compilation command can include the -L option to add additional library search directories. The linker searches for library files in these directories in the order that the -L options appear. For example, gcc main.c -L/path/to/libs -lmylibrary instructs the linker to search for a library named libmylibrary in /path/to/libs.
  • The LIBRARY_PATH environment variable: If the LIBRARY_PATH environment variable is set, the linker also searches for library files in the directories specified by this variable. This environment variable can contain multiple directories and is commonly used to add search paths for non-standard libraries during compilation.
  • System default library search paths: If the linker does not find the library files in the paths specified above, it will turn to the system default library search paths. These default paths typically include standard library directories such as /lib and /usr/lib, as well as architecture-specific directories (e.g., /lib/x86_64-linux-gnu or /usr/lib/x86_64-linux-gnu).
    • Refer to ld --verbose | grep SEARCH_DIR for details

The final value of LIBRARY_PATH(gcc -v can see the actual value) may not necessarily be just the value you set due to several reasons:

  • Combination with Default and Built-in Paths: GCC and the linker (ld) combine the LIBRARY_PATH you set with a set of default and built-in paths. These built-in paths are determined by the GCC’s configuration during its installation and are intended to ensure that the compiler and linker can find standard libraries and headers necessary for compilation and linking, even if they are not in standard locations.
  • Augmentation by GCC: GCC might automatically augment LIBRARY_PATH with additional paths based on its internal logic or other environment variables. For instance, GCC might add paths related to its internal libraries or the target architecture’s standard library locations.
  • Security Restrictions: In some secure or restricted environments, modifications to environment variables like LIBRARY_PATH may be ignored or overridden by security policies. This is often seen in managed or shared computing environments where administrators want to control the software linking process strictly.

2.3.2 Runtime

The runtime search order for shared libraries (.so files) on Unix-like systems is determined by several factors and environment settings. Here’s a general overview of how the runtime search order works: (Refer to man ld.so for details)

  • RPATH and RUNPATH: When a binary is compiled, it can be linked with shared libraries using -rpath or -runpath linker options. These options embed paths directly into the binary where the dynamic linker (ld.so or similar) should look for shared libraries. RPATH is checked first, but if RUNPATH is also specified, it takes precedence over RPATH when the dynamic linker is configured to use DT_RUNPATH entries (a newer feature).
  • LD_LIBRARY_PATH Environment Variable: Before falling back to default system paths, the dynamic linker checks the directories listed in the LD_LIBRARY_PATH environment variable, if it is set. This allows users to override the system’s default library paths or the paths embedded in the binary. However, for security reasons, this variable is ignored for setuid/setgid executables.
  • Default System Library Paths: If the library is not found in any of the previously mentioned locations, the linker searches the default library paths. These typically include /lib, /usr/lib, and their architecture-specific counterparts (e.g., /lib/x86_64-linux-gnu on some Linux distributions). The exact default paths can vary between systems and are defined in the dynamic linker’s configuration file (usually /etc/ld.so.conf), which can include additional directories beyond the standard ones.
  • DT_RPATH and DT_RUNPATH of Used Libraries (Dependencies): If the shared library being loaded has dependencies on other shared libraries, the dynamic linker also searches the RPATH and RUNPATH of those dependencies. This step ensures that all nested dependencies are resolved correctly.
  • /etc/ld.so.cache: This is a compiled list of candidate libraries previously found in the configured system library paths. The dynamic linker uses this cache to speed up the lookup process. The cache can be updated with the ldconfig command, which scans the directories in /etc/ld.so.conf and the standard directories for libraries, then builds the cache file.
  • Built-in System Paths: Finally, if the library still hasn’t been found, the linker falls back to built-in system paths hardcoded into the dynamic linker. This usually includes the standard directories like /lib and /usr/lib.

2.3.2.1 How to update /etc/ld.so.cache

Adding content directly to /etc/ld.so.cache is not recommended because it’s a binary file generated by the ldconfig tool. Modifying it manually can lead to system instability or even make your system unbootable. The correct way to add new libraries to the cache involves placing the library files in a directory that is scanned by ldconfig and then running ldconfig to update the cache.

Here’s how you can add a new library to the system and update /etc/ld.so.cache:

Place the Library in a Standard Directory

First, you need to place your library files (.so files) in a standard directory that is recognized by the dynamic linker. Common directories include:

  • /lib
  • /usr/lib
  • For system-wide libraries not provided by the package manager.

Or you can use a custom directory, but you’ll need to tell ldconfig about it, as described below.

(Optional) Add a Custom Directory to the ld.so Configuration

If you’re using a custom directory (/custom/lib for example), you need to add it to the dynamic linker’s configuration. Do this by creating a new file in /etc/ld.so.conf.d/:

Create /etc/ld.so.conf.d/mylibs.conf, Add the path to your custom library directory in this file:

1
/custom/lib

Update the Cache

After placing your libraries in the appropriate directory and configuring any custom paths, you need to update /etc/ld.so.cache by running ldconfig. This step requires root privileges:

1
sudo ldconfig

This command scans the directories listed in /etc/ld.so.conf, its *.conf includes in /etc/ld.so.conf.d/, and the default directories (/lib, /usr/lib). It then updates /etc/ld.so.cache with the paths to the available libraries.

Why does the order in which libraries are linked sometimes cause errors in GCC?

示例：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
cat > a.cpp << 'EOF'
extern int a;
int main() {
    return a;
}
EOF

cat > b.cpp << 'EOF'
extern int b;
int a = b;
EOF

cat > d.cpp << 'EOF'
int b;
EOF

g++ -c b.cpp -o b.o
ar cr libb.a b.o
g++ -c d.cpp -o d.o
ar cr libd.a d.o

g++ -L. -ld -lb a.cpp # wrong order
g++ -L. -lb -ld a.cpp # wrong order
g++ a.cpp -L. -ld -lb # wrong order
g++ a.cpp -L. -lb -ld # right order

链接器解析过程如下：从左到右扫描目标文件、库文件

  • 记录未解析的符号
  • 若发现某个库文件可以解决「未解析符号表」中的某个符号，则将该符号从「未解析符号表」中删除

因此，假设libA依赖libB，那么需要将libA写前面，libB写后面

2.5 Environment Variables

Please refer to man ld.so for details:

2.5.1 LD_PRELOAD

Lists shared libraries that are loaded (preloaded) before any other shared libraries. This can be used to override functions in other shared objects.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
cat > sample.c << 'EOF'
#include <stdio.h>
int main(void) {
    printf("Calling the fopen() function...\n");
    FILE *fd = fopen("test.txt","r");
    if (!fd) {
        printf("fopen() returned NULL\n");
        return 1;
    }
    printf("fopen() succeeded\n");
    return 0;
}
EOF
gcc -o sample sample.c

./sample 
#-------------------------↓↓↓↓↓↓-------------------------
Calling the fopen() function...
fopen() returned NULL
#-------------------------↑↑↑↑↑↑-------------------------

touch test.txt
./sample
#-------------------------↓↓↓↓↓↓-------------------------
Calling the fopen() function...
fopen() succeeded
#-------------------------↑↑↑↑↑↑-------------------------

cat > myfopen.c << 'EOF'
#include <stdio.h>
FILE *fopen(const char *path, const char *mode) {
    printf("This is my fopen!\n");
    return NULL;
}
EOF

gcc -o myfopen.so myfopen.c -Wall -fPIC -shared

LD_PRELOAD=./myfopen.so ./sample
#-------------------------↓↓↓↓↓↓-------------------------
Calling the fopen() function...
This is my fopen!
fopen() returned NULL
#-------------------------↑↑↑↑↑↑-------------------------

2.5.2 LD_DEBUG

Usage: LD_DEBUG=<type> <binary>, here are all available types.

  1. all: Print all debugging information (except statistics and unused; see below).
  2. bindings: Display information about which definition each symbol is bound to.
  3. files: Display progress for input file.
  4. libs: Display library search paths.
  5. reloc: Display relocation processing.
  6. scopes: Display scope information.
  7. statistics: Display relocation statistics.
  8. symbols: Display search paths for each symbol look-up.
  9. unused: Determine unused DSOs.
  10. versions: Display version dependencies.

2.6 How to make library

2.6.1 How to make static library

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
cat > foo.cpp << 'EOF'
#include <iostream>

void foo() {
    std::cout << "hello, this is foo" << std::endl;
}
EOF

cat > main.cpp << 'EOF'
void foo();

int main() {
    foo();
    return 0;
}
EOF

gcc -o foo.o -c foo.cpp -O3 -Wall -fPIC
ar rcs libfoo.a foo.o

gcc -o main main.cpp -O3 -L . -lfoo -lstdc++
./main

2.6.2 How to make dynamic library

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
cat > foo.cpp << 'EOF'
#include <iostream>

void foo() {
    std::cout << "hello, this is foo" << std::endl;
}
EOF

cat > main.cpp << 'EOF'
void foo();

int main() {
    foo();
    return 0;
}
EOF

gcc -o foo.o -c foo.cpp -O3 -Wall -fPIC
gcc -shared -o libfoo.so foo.o

gcc -o main main.cpp -O3 -L . -lfoo -lstdc++
./main # ./main: error while loading shared libraries: libfoo.so: cannot open shared object file: No such file or directory
gcc -o main main.cpp -O3 -L . -Wl,-rpath=`pwd` -lfoo -lstdc++
./main

2.7 ABI

ABI全称是Application Binary Interface，它是两个二进制模块间的接口，二进制模块可以是lib，可以是操作系统的基础设施，或者一个正在运行的用户程序

An ABI defines how data structures or computational routines are accessed in machine code, which is a low-level, hardware-dependent format

abi

具体来说，ABI定义了如下内容：

  1. 指令集、寄存器、栈组织结构，内存访问类型等等
  2. 基本数据类型的size、layout、alignment
  3. 调用约定，比如参数如何传递、结果如何传递
    • 参数传递到栈中，还是传递到寄存器中
    • 如果传递到寄存器中的话，哪个入参对应哪个寄存器
    • 如果传递到栈中的话，第一个入参是先压栈还是后压栈
    • 返回值使用哪个寄存器
  4. 如何发起系统调用
  5. lib、object file等的文件格式

2.7.1 Language-Specific ABI

摘自What is C++ ABI?

Often, a platform will specify a “base ABI” that specifies how to use that platform’s basic services and that is often done in terms of C language capabilities. However, other programming languages like C++ may require support for additional mechanisms. That’s how you get to language-specific ABIs, including a variety of C++ ABIs. Among the concerns of a C++ ABI are the following:

  • How are classes with virtual functions represented? A C++ ABI will just about always extend the C layout rules for this, and specify implicit pointers in the objects that point to static tables (“vtables”) that themselves point to virtual functions.
  • How are base classes (including virtual base classes) laid out in class objects? Again, this usually starts with the C struct layout rules.
  • How are closure classes (for lambdas) organized?
  • How is RTTI information stored?
  • How are exceptions implemented?
  • How are overloaded functions and operators “named” in object files? This is where “name mangling” usually comes in: The type of the various functions is encoded in their object file names. That also handles the “overloading” that results from template instantiations.
  • How are spilled inline functions and template instantiations handled? After all, different object files might use/spill the same instances, which could lead to collisions.

2.7.2 Dual ABI

Dual ABI

In the GCC 5.1 release libstdc++ introduced a new library ABI that includes new implementations of std::string and std::list. These changes were necessary to conform to the 2011 C++ standard which forbids Copy-On-Write strings and requires lists to keep track of their size.

宏_GLIBCXX_USE_CXX11_ABI用于控制使用新版本还是老版本

  • _GLIBCXX_USE_CXX11_ABI = 0：老版本
  • _GLIBCXX_USE_CXX11_ABI = 1：新版本

其他参考：

2.7.3 ABI Mismatch Issue

1
2
3
4
5
6
7
8
9
# compile works fine
gcc -o main main.cpp -O0 -lstdc++

# Errors occur at runtime
./main

# ./main: /lib64/libstdc++.so.6: version `GLIBCXX_3.4.29' not found (required by ./main)
# ./main: /lib64/libstdc++.so.6: version `GLIBCXX_3.4.30' not found (required by ./main)
# ./main: /lib64/libstdc++.so.6: version `CXXABI_1.3.13' not found (required by ./main)

The errors you’re encountering at runtime indicate that your application (main) is linked against newer versions of the C++ standard library (libstdc++) and C++ ABI (CXXABI) than are available on your system. This typically happens when your application is compiled with a newer version of GCC than the one installed on the runtime system.

  • gcc -o main main.cpp -O0 -lstdc++ -fuse-ld=gold -Wl,--verbose: Check the dynamic lib path that used at link time.
  • ldd main: Check the dynamic lib path that used at runtime.

2.8 Commonly-used Librarys

libc/glibc/glib（man libc/glibc）

  • libc实现了C的标准库函数，例如strcpy()，以及POSIX函数，例如系统调用getpid()。此外，不是所有的C标准库函数都包含在libc中，比如大多数math相关的库函数都封装在libm中，大多数压缩相关的库函数都封装在libz中
    • 系统调用有别于普通函数，它无法被链接器解析。实现系统调用必须引入平台相关的汇编指令。我们可以通过手动实现这些汇编指令来完成系统调用，或者直接使用libc（它已经为我们封装好了）
  • glibc, GNU C Library可以看做是libc的另一种实现，它不仅包含libc的所有功能还包含libm以及其他核心库，比如libpthread
  • glib是Linux下C的一些工具库，和glibc没有关系

查看glibc的版本

  • ldd --version：ldd隶属于glibc，因此ldd的版本就是glibc的版本
  • getconf GNU_LIBC_VERSION
  • gcc -print-file-name=libc.so
  • strings /lib64/libc.so.6 | grep GLIBC：查看glibc的API的版本

其他常用动态库可以参考Library Interfaces and Headers中的介绍

  1. libdl：dynamic linking library
    • libdl主要作用是将那些早已存在于libc中的private dl functions对外露出，便于用户实现一些特殊的需求。dlopen in libc and libdl
    • 可以通过readelf -s /lib64/ld-linux-x86-64.so.2 | grep PRIVATE查看露出的这些方法
  2. libm：c math library
  3. libz：compression/decompression library
  4. libpthread：POSIX threads library

2.9 Reference

3 Memory Management

Virtual memory is allocated through system call brk and mmap, please refer to Linux-Memory-Management for details.

memory management libraries run in user mode and typically use low-level system calls for virtual memory allocation. Different libraries may employ different memory allocation algorithms to effectively manage allocated and deallocated memory blocks in the heap.

3.1 tcmalloc

tcmalloc

tcmalloc

特点：

  • Small object allocation
    • 每个线程都会有个ThreadCache，用于为当前线程分配小对象
    • 当其容量不足时，会从CentralCache获取额外的存储空间
  • CentralCache allocation management
    • 用于分配大对象，大对象通常指>32K
    • 当内存空间用完后，用sbrk/mmap从操作系统中分配内存
    • 在多线程高并发的场景中，CentralCache中的锁竞争很容易成为瓶颈
  • Recycle

如何安装：

1
yum -y install gperftools gperftools-devel

3.1.1 Heap-profile

main.cpp：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
#include <stdlib.h>

void* create(unsigned int size) {
    return malloc(size);
}

void create_destory(unsigned int size) {
    void* p = create(size);
    free(p);
}

int main(void) {
    const int loop = 4;
    char* a[loop];
    unsigned int mega = 1024 * 1024;

    for (int i = 0; i < loop; i++) {
        const unsigned int create_size = 1024 * mega;
        create(create_size);

        const unsigned int malloc_size = 1024 * mega;
        a[i] = (char*)malloc(malloc_size);

        const unsigned int create_destory_size = mega;
        create_destory(create_destory_size);
    }

    for (int i = 0; i < loop; i++) {
        free(a[i]);
    }

    return 0;
}

编译：

1
gcc -o main main.cpp -Wall -O3 -lstdc++ -ltcmalloc -std=gnu++17

运行：

1
2
3
4
# 开启 heap profile 功能
export HEAPPROFILE=/tmp/test-profile

./main

输出如下：

1
2
3
4
5
6
7
Starting tracking the heap
tcmalloc: large alloc 1073741824 bytes == 0x2c46000 @  0x7f6a8fd244ef 0x7f6a8fd43e76 0x400571 0x7f6a8f962555 0x4005bf
Dumping heap profile to /tmp/test-profile.0001.heap (1024 MB allocated cumulatively, 1024 MB currently in use)
Dumping heap profile to /tmp/test-profile.0002.heap (2048 MB allocated cumulatively, 2048 MB currently in use)
Dumping heap profile to /tmp/test-profile.0003.heap (3072 MB allocated cumulatively, 3072 MB currently in use)
Dumping heap profile to /tmp/test-profile.0004.heap (4096 MB allocated cumulatively, 4096 MB currently in use)
Dumping heap profile to /tmp/test-profile.0005.heap (Exiting, 0 bytes in use)

使用pprof分析内存：

1
2
3
4
5
# 文本格式
pprof --text ./main /tmp/test-profile.0001.heap | head -30

# 图片格式
pprof --svg ./main /tmp/test-profile.0001.heap > heap.svg

3.2 jemalloc

jemalloc

jemalloc

特点：

  • 在多核、多线程场景下，跨线程分配/释放的性能比较好
  • 大量分配小对象时，所占空间会比tcmalloc稍多一些
  • 对于大对象分配，所造成的的内存碎片会比tcmalloc少一些
  • 内存分类粒度更细，锁比tcmalloc更少

3.2.1 Install

jemalloc/INSTALL.md

1
2
3
4
5
6
7
8
git clone git@github.com:jemalloc/jemalloc.git
cd jemalloc
git checkout 5.3.0

./autogen.sh
./configure --prefix=/usr/local --enable-prof
make -j 16
sudo make install

3.2.2 Heap-profile

Getting Started

1
2
3
4
5
6
7
8
9
10
11
12
#include <cstdint>

void alloc_1M() {
    new uint8_t[1024 * 1024];
}

int main() {
    for (int i = 0; i < 100; i++) {
        alloc_1M();
    }
    return 0;
}

编译执行：

  • 方式1：隐式链接，用LD_PRELOAD

    1
    2
    gcc -o main main.cpp -O0 -lstdc++ -std=gnu++17
    MALLOC_CONF=prof_leak:true,lg_prof_sample:0,prof_final:true LD_PRELOAD=/usr/local/lib/libjemalloc.so.2 ./main

  • 方式2：显示链接

    1
    2
    gcc -o main main.cpp -O0 -lstdc++ -std=gnu++17 -L`jemalloc-config --libdir` -Wl,-rpath,`jemalloc-config --libdir` -ljemalloc `jemalloc-config --libs`
    MALLOC_CONF=prof_leak:true,lg_prof_sample:0,prof_final:true ./main

查看：jeprof --text main jeprof.145931.0.f.heap

  • 第一列：函数直接申请的内存大小，单位MB
  • 第二列：第一列占总内存的百分比
  • 第三列：第二列的累积值
  • 第四列：函数以及函数所有调用的函数申请的内存大小，单位MB
  • 第五列：第四列占总内存的百分比
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Using local file main.
Using local file jeprof.145931.0.f.heap.
Total: 100.1 MB
100.1 100.0% 100.0%    100.1 100.0% prof_backtrace_impl
    0.0   0.0% 100.0%      0.1   0.1% _GLOBAL__sub_I_eh_alloc.cc
    0.0   0.0% 100.0%    100.0  99.9% __libc_start_main
    0.0   0.0% 100.0%      0.1   0.1% __static_initialization_and_destruction_0 (inline)
    0.0   0.0% 100.0%      0.1   0.1% _dl_init_internal
    0.0   0.0% 100.0%      0.1   0.1% _dl_start_user
    0.0   0.0% 100.0%    100.0  99.9% _start
    0.0   0.0% 100.0%    100.0  99.9% alloc_1M
    0.0   0.0% 100.0%    100.1 100.0% imalloc (inline)
    0.0   0.0% 100.0%    100.1 100.0% imalloc_body (inline)
    0.0   0.0% 100.0%    100.1 100.0% je_malloc_default
    0.0   0.0% 100.0%    100.1 100.0% je_prof_backtrace
    0.0   0.0% 100.0%    100.1 100.0% je_prof_tctx_create
    0.0   0.0% 100.0%    100.0  99.9% main
    0.0   0.0% 100.0%      0.1   0.1% pool (inline)
    0.0   0.0% 100.0%    100.1 100.0% prof_alloc_prep (inline)
    0.0   0.0% 100.0%    100.0  99.9% void* fallback_impl

3.2.3 Work with http

源码如下：

  • yhirose/cpp-httplib
  • path是固定，即/pprof/heap以及/pprof/profile
    • /pprof/profile的实现估计有问题，拿不到预期的结果
1
2
3
4
5
sudo apt-get install libgoogle-perftools-dev

# or

sudo yum install gperftools-devel
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
#include <gperftools/profiler.h>
#include <jemalloc/jemalloc.h>
#include <stdlib.h>

#include <iterator>
#include <sstream>
#include <thread>

#include "httplib.h"

uint8_t* alloc_1M() {
    return new uint8_t[1024 * 1024];
}

int main(int argc, char** argv) {
    httplib::Server svr;

    svr.Get("/pprof/heap", [](const httplib::Request&, httplib::Response& res) {
        std::stringstream fname;
        fname << "./heap." << getpid() << "." << rand();
        auto fname_str = fname.str();
        const char* fname_cstr = fname_str.c_str();

        std::string content;
        if (mallctl("prof.dump", nullptr, nullptr, &fname_cstr, sizeof(const char*)) == 0) {
            std::ifstream f(fname_cstr);
            content = std::string(std::istreambuf_iterator<char>(f), std::istreambuf_iterator<char>());
        } else {
            content = "dump jemalloc prof file failed";
        }

        res.set_content(content, "text/plain");
    });

    svr.Get("/pprof/profile", [](const httplib::Request& req, httplib::Response& res) {
        int seconds = 30;
        const std::string& seconds_str = req.get_param_value("seconds");
        if (!seconds_str.empty()) {
            seconds = std::atoi(seconds_str.c_str());
        }

        std::ostringstream fname;
        fname << "./profile." << getpid() << "." << rand();
        auto fname_str = fname.str();
        const char* fname_cstr = fname_str.c_str();

        ProfilerStart(fname_cstr);
        sleep(seconds);
        ProfilerStop();

        std::ifstream f(fname_cstr, std::ios::in);
        std::string content;
        if (f.is_open()) {
            content = std::string(std::istreambuf_iterator<char>(f), std::istreambuf_iterator<char>());
        } else {
            content = "dump jemalloc prof file failed";
        }

        res.set_content(content, "text/plain");
    });

    std::thread t_alloc([]() {
        while (true) {
            alloc_1M();
            sleep(1);
        }
    });

    svr.listen("0.0.0.0", 16691);

    return 0;
}

Compile and run:

1
2
3
4
5
6
7
8
9
gcc -o main main.cpp -O0 -lstdc++ -std=gnu++17 -L`jemalloc-config --libdir` -Wl,-rpath,`jemalloc-config --libdir` -ljemalloc `jemalloc-config --libs` -lprofiler
MALLOC_CONF=prof_leak:true,lg_prof_sample:0,prof_final:true ./main

# In another terminal
jeprof main localhost:16691/pprof/heap # interactive mode
jeprof main localhost:16691/pprof/heap --text # text
jeprof main localhost:16691/pprof/heap --svg > main.svg # graph

jeprof main localhost:16691/pprof/profile --text --seconds=15

How to create a graph:

1
2
3
4
5
6
curl http://127.0.0.1:16691/pprof/heap > begin.heap
sleep 10
curl http://127.0.0.1:16691/pprof/heap > end.heap
jeprof --dot --base=begin.heap ./main end.heap > heap-profile.dot
dot -Tsvg heap-profile.dot -o heap-profile.svg
dot -Tpdf heap-profile.dot -o heap-profile.pdf

3.2.4 Reference

3.3 mimalloc

3.4 Comparison

tcmalloc and jemalloc are both memory allocators, and they are commonly used in software development to manage dynamic memory allocation in programs. However, they have different characteristics and were designed to address different issues. Here’s a comparison of the two:

  1. Design Philosophy:
    • jemalloc: Designed by Jason Evans, originally to enhance the performance of FreeBSD. jemalloc focuses on reducing memory fragmentation and improving memory allocation efficiency, especially in concurrent environments. It employs an advanced memory allocation strategy that reduces lock contention, thus improving performance in multithreaded applications.
    • tcmalloc: Developed by Google, standing for “Thread-Caching Malloc”. The key feature of tcmalloc is that it provides individual memory caches for each thread. This design reduces the reliance on a global memory allocation lock, thus offering better performance in multithreaded applications.
  2. Memory Allocation Strategy:
    • jemalloc: Uses size classes to manage memory allocations, which helps in reducing memory fragmentation. It also employs a delayed recycling strategy to further optimize memory usage.
    • tcmalloc: Manages memory allocations for each thread through Thread Local Storage (TLS), meaning each thread has its own small memory pool for quick allocation and deallocation.
  3. Memory Fragmentation Management:
    • jemalloc: Effectively reduces memory fragmentation through its sophisticated memory allocation strategy.
    • tcmalloc: While its thread caching mechanism can boost performance, it might lead to more memory fragmentation in some scenarios.
  4. Suitable Use Cases:
    • jemalloc: Widely used in applications requiring high-performance memory management, such as databases and large multithreaded applications.
    • tcmalloc: Particularly suitable for applications that frequently allocate and deallocate memory due to its high-performance thread caching feature.

3.5 Hook

3.5.1 Override the operator new

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
#include <cstddef>
#include <cstdio>
#include <cstdlib>
#include <iostream>
#include <new>

void* operator new(std::size_t size) {
    std::cout << "Custom new called with size: " << size << std::endl;
    return malloc(size);
}

void operator delete(void* ptr) noexcept {
    std::cout << "Custom delete called" << std::endl;
    free(ptr);
}

extern "C" {
void* __real_malloc(size_t c);

void* __wrap_malloc(size_t c) {
    printf("malloc called with %zu\n", c);
    return __real_malloc(c);
}
}

int main() {
    int* num1 = (int*)malloc(sizeof(int));
    int* num2 = new int;
    return *num1 * *num2;
}
1
2
gcc -o main main.cpp -Wl,-wrap=malloc -lstdc++ -std=gnu++17
./main

3.5.2 Wrap operator new by its mangled name

The mangled name for operator new with a size_t argument is _Znwm.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#include <cstddef>
#include <cstdio>
#include <cstdlib>

extern "C" {
void* __real_malloc(size_t c);

void* __wrap_malloc(size_t c) {
    printf("malloc called with %zu\n", c);
    return __real_malloc(c);
}

void* __real__Znwm(size_t c); // operator new

void* __wrap__Znwm(size_t c) {
    printf("operator new called with %zu\n", c);
    return __real__Znwm(c);
}
}

int main() {
    int* num1 = (int*)malloc(sizeof(int));
    int* num2 = new int;
    return *num1 * *num2;
}
1
2
gcc -o main main.cpp -Wl,-wrap=malloc -Wl,-wrap=_Znwm -lstdc++ -std=gnu++17
./main

3.5.3 Work With Library

foo.cpp is like this:

1
2
3
4
5
6
#include <iostream>

int* foo() {
    std::cout << "hello, this is foo" << std::endl;
    return (int*)malloc(100);
}

main.cpp is like this:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
#include <cstdio>
#include <cstdlib>

int* foo();

extern "C" {
void* __real_malloc(size_t c);

void* __wrap_malloc(size_t c) {
    printf("malloc called with %zu\n", c);
    return __real_malloc(c);
}
}

int main() {
    auto* res = foo();
    res[0] = 1;
    return 0;
}

3.5.3.1 Static Linking Library(Working)

And here are how to build the program.

1
2
3
4
5
gcc -o foo.o -c foo.cpp -O3 -Wall -fPIC
ar rcs libfoo.a foo.o
gcc -o main main.cpp -O3 -L . -Wl,-wrap=malloc -lfoo -lstdc++

./main

And it prints:

1
2
hello, this is foo
malloc called with 100

The reason why static linking works fine with the -Wl,-wrap=malloc hook, is because of how static linking works:

  • In static linking, the object code (or binary code) for the functions used from the static library is embedded directly into the final executable at link-time.
  • So, when you use the -Wl,-wrap=malloc linker option, it can intercept and replace all calls to malloc in both the main program and any static libraries, as they are all being linked together into a single executable at the same time.
  • Essentially, the linker sees the entire code (from the main program and the static library) as one unit and can replace all calls to malloc with calls to __wrap_malloc.

3.5.3.2 Dynamic Linking library(Not Working)

And here are how to build the program.

1
2
3
4
5
gcc -o foo.o -c foo.cpp -O3 -Wall -fPIC
gcc -shared -o libfoo.so foo.o
gcc -o main main.cpp -O3 -L . -Wl,-rpath=`pwd` -Wl,-wrap=malloc  -lfoo -lstdc++

./main

And it prints:

1
hello, this is foo

Here’s a step-by-step breakdown:

  1. When you compile foo.cpp into foo.o, there’s a call to malloc in the machine code, but it’s not yet resolved to an actual memory address. It’s just a placeholder that says “I want to call malloc”.
  2. When you link foo.o into libfoo.so, the call to malloc inside foo is linked. It’s resolved to the malloc function provided by the C library.
  3. Later, when you link main.cpp into an executable, you’re using the -Wl,-wrap=malloc option, but that only affects calls to malloc that are being linked at that time. The call to malloc inside foo was already linked in step 2, so it’s unaffected.

3.6 Reference

4 Name mangling

Name mangling, also known as name decoration or name encoding, is a technique used by compilers to encode the names of functions, variables, and other symbols in a program in a way that includes additional information about their types, namespaces, and sometimes other attributes. Name mangling is primarily used in languages like C++ to support features such as function overloading, function templates, and namespaces.

Here are some key points about name mangling:

  • Function Overloading: In C++, you can define multiple functions with the same name but different parameter lists. Name mangling ensures that these functions have distinct mangled names so the compiler can differentiate between them during the linking process.
  • Template Specialization: In C++, templates allow you to write generic code that works with different types. Name mangling encodes the template arguments and specializations so that the compiler can generate the correct code for each specialization.
  • Namespace Resolution: When you use namespaces to organize your code, name mangling includes the namespace information in the mangled name to avoid naming conflicts.
  • Type Information: Name mangling can encode information about the types of function parameters and return values, which can be used for type checking and to resolve overloaded function calls.
  • Demangling: While the mangled names generated by the compiler are not meant to be human-readable, tools and libraries exist to “demangle” these names, converting them back into their original, human-readable form for debugging and error messages.
  • Compiler-Specific: The specific rules for name mangling are compiler-dependent. Different C++ compilers, such as GCC, Clang, and MSVC, may use different name mangling schemes. These schemes are not standardized by the C++ language standard.

4.1 Encoding Elements

Name mangling (also known as name decoration) is a compiler-specific technique to encode C++ symbols into unique names. The exact format and rules for name mangling are dependent on the compiler and its version. Therefore, there isn’t a single standard for it. However, many C++ compilers follow (at least loosely) the Itanium C++ ABI for mangling names, especially on Unix-like systems.

Below are some common elements of name mangling based on the Itanium C++ ABI:

  1. Function Names:
    • _Z: Prefix for encoded names.
    • N: Begins a nested name.
    • E: Ends a nested name.
  2. Namespace and Class Names:
    • Length of the name followed by the actual name. E.g., 5Outer for Outer.
  3. Types:
    • i: int
    • l: long
    • s: short
    • d: double
    • f: float
    • c: char
    • b: bool
    • v: void
    • P: Pointer to a type (e.g., Pi is pointer to int)
    • R: Reference
    • O: rvalue reference
    • C1, C2: Constructor
    • D1, D2: Destructor
  4. Template Classes:
    • I: Marks the beginning of template arguments.
    • E: Marks the end of template arguments.
  5. Const/Volatile qualifiers:
    • K: const
    • V: volatile
    • r: restrict (from C99)
  6. Function Arguments:
    • Encoded in order, using their type encodings.
  7. Arrays:
    • The size followed by the type, e.g., A10_i for int[10].
  8. Modifiers:
    • U: Vendor-specific type qualifier.
  9. Special Names:
    • _vt: Virtual table
    • _qt: Virtual table for a qualified name
  10. Operators:
    • Most operators get their own special mangled name, e.g., _pl for operator+.
  11. CV Qualifiers:
    • K: const
    • V: volatile
  12. Calling Convention:
    • While the Itanium ABI doesn’t specify mangling for calling conventions (because it’s designed for architectures with a single calling convention), some other mangling schemes might have special symbols for this.
  13. Other Features:
    • S: String literal
    • And various encodings for other built-in types, custom types, etc.

4.2 Example

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
namespace Outer::Inner {

template <typename T>
class Foo {
public:
    Foo(T value) : data(value) {}
    void bar() {}
    void bar() const {}

private:
    T data;
};

void func_no_arg() {}
void func_void(void) {}

void func_one_arg(int x) {}
void func_one_arg(double x) {}
void func_one_arg(long& x) {}

void func2(int x, int y) {}

void func2(double x, double y) {}

} // namespace Outer::Inner

int main() {
    Outer::Inner::Foo<int> foo(5);
    foo.bar();
    const Outer::Inner::Foo<int> cfoo(6);
    cfoo.bar();
    return 0;
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# gcc-12.3.0
$ gcc -o main main.cpp -lstdc++ -std=gnu++17 -g -ggdb -O0 && nm main | grep '_Z'

0000000000001149 T _ZN5Outer5Inner11func_no_argEv
000000000000116d T _ZN5Outer5Inner12func_one_argEd
000000000000115f T _ZN5Outer5Inner12func_one_argEi
000000000000117d T _ZN5Outer5Inner12func_one_argERl
000000000000123e W _ZN5Outer5Inner3FooIiE3barEv
0000000000001222 W _ZN5Outer5Inner3FooIiEC1Ei
0000000000001222 W _ZN5Outer5Inner3FooIiEC2Ei
000000000000119d T _ZN5Outer5Inner5func2Edd
000000000000118c T _ZN5Outer5Inner5func2Eii
0000000000001154 T _ZN5Outer5Inner9func_voidEv
000000000000124e W _ZNK5Outer5Inner3FooIiE3barEv

5 Address Sanitizer

Address Sanitizer (ASan) has its origins in Google. The project was started as a way to improve the detection of memory-related bugs, which have historically been some of the most elusive and difficult-to-diagnose issues in software development.

Here’s a brief history of Address Sanitizer:

  1. Early 2010s Origin: Address Sanitizer was developed by Konstantin Serebryany, Dmitry Vyukov, and other contributors at Google. The initial focus was on detecting memory issues in C++ programs, but the tool’s capabilities expanded over time.
  2. Release and Integration with Major Compilers: After its development, Address Sanitizer was integrated into major compilers. Clang was one of the first compilers to support ASan, and later, it was incorporated into GCC (GNU Compiler Collection). The ease of use—simply adding a compiler flag—made it very attractive for developers to adopt.
  3. Chromium Adoption: One of the early and significant adoptions of ASan was in the Chromium project (the open-source foundation of the Chrome browser). Given the scale and complexity of this project, and the potential security implications of memory bugs in a web browser, ASan proved to be an invaluable tool for ensuring the robustness and safety of the code.
  4. Expanding Beyond Address Sanitizer: The success of ASan led to the development of other “sanitizers” such as Thread Sanitizer (TSan) for detecting data races, Memory Sanitizer (MSan) for uninitialized memory reads, and Undefined Behavior Sanitizer (UBSan) for undefined behavior checks.
  5. Integration in Development Ecosystems: As ASan’s popularity grew, it found its way into various development ecosystems. Debugging tools, continuous integration systems, and even some integrated development environments (IDEs) began offering first-class support for ASan and related tools.
  6. Widening Support: Beyond just C++ and C, ASan also found applicability in other languages and platforms where the underlying memory model had similarities, further broadening its reach and impact.

It is a tool that can help developers find memory-related issues in their programs, such as:

  1. Use-after-free: Detects when a program uses memory after it has been freed.
  2. Heap buffer overflow: Detects when a program writes past the end of an allocated object on the heap.
  3. Stack buffer overflow: Detects when a program writes past the end of an allocated object on the stack.
  4. Global buffer overflow: Similar to the above but for global variables.
  5. Memory leaks: Identifies instances where memory is allocated but not freed, resulting in memory being lost until the program ends.
  6. Use-after-scope: Detects the usage of local variables outside of their scope.
  7. Initialization errors: Finds reads from uninitialized memory locations.

Address Sanitizer achieves this by instrumenting the binary code. When compiling code with Address Sanitizer enabled, the compiler (e.g., Clang or GCC) will insert additional checks around memory operations. Additionally, Address Sanitizer uses a shadow memory mechanism to keep track of the status (e.g., allocated, deallocated) of each byte in the application’s memory.

Because of this instrumentation, programs compiled with Address Sanitizer tend to run slower and use more memory than their non-instrumented counterparts. However, the benefits in terms of finding and fixing memory-related bugs can be substantial.

To use Address Sanitizer, developers typically pass the -fsanitize=address flag (and sometimes additional flags) when compiling and linking their code.

Remember, Address Sanitizer is not a substitute for other types of testing or for manual code review. However, when combined with other testing methodologies, it can be a powerful tool in a developer’s arsenal for ensuring software reliability and security.

5.1 How it works

Here’s an in-depth breakdown of how it works:

  1. Compile-time Instrumentation:
    • When a program is compiled with ASan enabled (e.g., using the -fsanitize=address flag), the compiler inserts extra instructions around memory operations. These instructions help in checking for memory-related issues.
    • This instrumentation helps the runtime component of ASan understand memory accesses and allocations/deallocations.
  2. Shadow Memory:
    • ASan uses a concept called “shadow memory” to keep track of each byte of the application’s memory.
    • For every 8 bytes of application memory, there’s a corresponding byte in the shadow memory. This shadow byte holds metadata about the status of those 8 bytes (whether they are allocated, deallocated, part of a redzone, etc.).
    • Using shadow memory, ASan can quickly check the status of a memory location whenever the program accesses it. For example, if a program accesses memory that corresponds to a shadow byte indicating it’s deallocated, ASan can immediately flag a use-after-free error.
  3. Redzones:
    • ASan places “redzones” (extra bytes) around allocated memory. These redzones are used to detect out-of-bounds accesses.
    • If the program accesses memory inside a redzone, it indicates a buffer overflow or underflow.
  4. **Allocation and Deallocation:
    • ASan replaces the standard memory allocation and deallocation functions (like malloc and free in C).
    • When memory is allocated, ASan also sets up the corresponding shadow memory and redzones. When memory is deallocated, it is poisoned (marked in a way that any access will be flagged by ASan), helping detect use-after-free bugs.
  5. Memory Leak Detection:
    • ASan can also track allocations that were never deallocated. At the end of the program’s execution, it can report these as memory leaks.
  6. Reporting:
    • When ASan detects an error, it immediately halts the program and provides a detailed report. This report typically includes the type of error (e.g., buffer overflow), the memory address where the error occurred, and a stack trace, helping developers pinpoint and understand the cause.
  7. Performance Impact:
    • ASan’s checks and shadow memory mechanism introduce overhead. Programs compiled with ASan typically run slower (often around 2x) and consume more memory than their non-instrumented counterparts.
    • It’s important to understand that while ASan can detect a variety of memory errors, it’s not foolproof. For example, it might not detect all memory leaks, especially if they involve complex data structures. However, its ability to catch many common and critical memory errors makes it a valuable tool for developers.

When you compile your program with Address Sanitizer (ASan) enabled, the standard memory allocation and deallocation functions like malloc, free, new, and delete get intercepted or “wrapped” by ASan. This means that when your program calls malloc, it’s actually calling ASan’s version of malloc.

This interception is crucial for how ASan functions. Here’s why:

  1. Setting up Redzones: When memory is allocated, ASan’s version of malloc (or new for C++) can add “redzones” around the allocated memory. These redzones help in detecting out-of-bounds accesses.
  2. Shadow Memory Management: ASan maintains a shadow memory that maps to your program’s memory to track the state of each byte (e.g., allocated, deallocated). When memory is allocated or deallocated, ASan’s version of the memory functions update the shadow memory accordingly.
  3. Poisoning Memory: When memory is deallocated using ASan’s free or delete, the memory isn’t immediately returned to the system. Instead, it is “poisoned”. This means that any subsequent access to this memory will raise an error, helping in detecting use-after-free bugs.
  4. Memory Leak Detection: By intercepting these calls, ASan can also maintain a list of all current allocations. When the program exits, it checks this list to report any memory that hasn’t been deallocated, indicating a memory leak.

Here’s an example to illustrate the mechanism:

1
2
3
4
5
6
#include <iostream>

int main() {
    std::cout << "Hello AddressSanitizer!\n";
    return 0;
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# dynamic link
gcc -o main main.cpp -lstdc++ -std=gnu++17 -fsanitize=address

objdump -p main | grep NEEDED
  NEEDED               libasan.so.8
  NEEDED               libstdc++.so.6
  NEEDED               libc.so.6

# It shows that the call malloc was redirect to libasan.so.8
LD_DEBUG="bindings" ./main 2> bind.txt && cat bind.txt | grep malloc
Hello AddressSanitizer!
   2623584:	binding file /lib/x86_64-linux-gnu/libc.so.6 [0] to /lib/x86_64-linux-gnu/libasan.so.8 [0]: normal symbol `malloc' [GLIBC_2.2.5]
   2623584:	binding file ./main [0] to /lib/x86_64-linux-gnu/libasan.so.8 [0]: normal symbol `malloc' [GLIBC_2.2.5]
   2623584:	binding file /lib/x86_64-linux-gnu/libstdc++.so.6 [0] to /lib/x86_64-linux-gnu/libasan.so.8 [0]: normal symbol `malloc' [GLIBC_2.2.5]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# static link
gcc -o main main.cpp -lstdc++ -std=gnu++17 -fsanitize=address -static-libasan

objdump -p main | grep NEEDED
  NEEDED               libstdc++.so.6
  NEEDED               libm.so.6
  NEEDED               libgcc_s.so.1
  NEEDED               libc.so.6

# It shows that the call malloc was redirect to main itself
LD_DEBUG="bindings" ./main 2> bind.txt && cat bind.txt | grep malloc
Hello AddressSanitizer!
   2624551:	binding file /lib/x86_64-linux-gnu/libc.so.6 [0] to ./main [0]: normal symbol `malloc' [GLIBC_2.2.5]
   2624551:	binding file ./main [0] to ./main [0]: normal symbol `malloc' [GLIBC_2.2.5]
   2624551:	binding file /lib/x86_64-linux-gnu/libstdc++.so.6 [0] to ./main [0]: normal symbol `malloc' [GLIBC_2.2.5]

5.2 memory leak

1
2
3
4
5
6
7
8
9
10
11
12
cat > test_memory_leak.cpp << 'EOF'
#include <iostream>

int main() {
    int *p = new int(5);
    std::cout << *p << std::endl;
    return 0;
}
EOF

gcc test_memory_leak.cpp -o test_memory_leak -g -lstdc++ -fsanitize=address -static-libasan
./test_memory_leak

5.3 stack buffer underflow

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
cat > test_stack_buffer_underflow.cpp << 'EOF'
#include <iostream>

int main() {
    char buffer[5] = "";
    uint8_t num = 5;

    buffer[-1] = 7;

    std::cout << (void*)buffer << std::endl;
    std::cout << (void*)&buffer[-1] << std::endl;
    std::cout << (void*)&num << std::endl;
    std::cout << (uint32_t)num << std::endl;

    return 0;
}
EOF

# 非asan模式
gcc test_stack_buffer_underflow.cpp -o test_stack_buffer_underflow -g -lstdc++ 
./test_stack_buffer_underflow

# asan模式
gcc test_stack_buffer_underflow.cpp -o test_stack_buffer_underflow -g -lstdc++ -fsanitize=address -static-libasan
./test_stack_buffer_underflow

5.4 Reference

6 User-defined Thread

jump_fcontext and make_fcontext is copy from context.h and context.cpp

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
#include <cstddef>
#include <iostream>

typedef void* fcontext_t;

extern "C" {
intptr_t jump_fcontext(fcontext_t* ofc, fcontext_t nfc, intptr_t vp, bool preserve_fpu = false);
fcontext_t make_fcontext(void* sp, size_t size, void (*fn)(intptr_t));
}

__asm(".text\n"                         // Section for code
      ".globl jump_fcontext\n"          // Globally visible symbol
      ".type jump_fcontext,@function\n" // Specifies the symbol type
      ".align 16\n"                     // Aligns the code at 16-byte boundaries
      "jump_fcontext:\n"                // Start of function named 'jump_fcontext'
      "    pushq  %rbp  \n"             // Pushes base pointer onto the stack
      "    pushq  %rbx  \n"             // Pushes bx register onto the stack
      "    pushq  %r15  \n"             // Pushes r15 register onto the stack
      "    pushq  %r14  \n"             // Pushes r14 register onto the stack
      "    pushq  %r13  \n"             // Pushes r13 register onto the stack
      "    pushq  %r12  \n"             // Pushes r12 register onto the stack
      "    leaq  -0x8(%rsp), %rsp\n"    // Decrement the stack pointer by 8 bytes (allocate stack memory)
      "    cmp  $0, %rcx\n"             // Compare rcx register with 0
      "    je  1f\n"                    // If rcx == 0, jump to label 1
      "    stmxcsr  (%rsp)\n"           // Store the contents of the MXCSR register onto the stack
      "    fnstcw   0x4(%rsp)\n"        // Store the control word of FPU onto the stack
      "1:\n"
      "    movq  %rsp, (%rdi)\n" // Save current stack pointer to location pointed by rdi
      "    movq  %rsi, %rsp\n"   // Set stack pointer to value in rsi
      "    cmp  $0, %rcx\n"      // Compare rcx register with 0
      "    je  2f\n"             // If rcx == 0, jump to label 2
      "    ldmxcsr  (%rsp)\n"    // Load value from the stack into the MXCSR register
      "    fldcw  0x4(%rsp)\n"   // Load control word into FPU
      "2:\n"
      "    leaq  0x8(%rsp), %rsp\n"               // Increment the stack pointer by 8 bytes (deallocate stack memory)
      "    popq  %r12  \n"                        // Restore r12 register from the stack
      "    popq  %r13  \n"                        // Restore r13 register from the stack
      "    popq  %r14  \n"                        // Restore r14 register from the stack
      "    popq  %r15  \n"                        // Restore r15 register from the stack
      "    popq  %rbx  \n"                        // Restore bx register from the stack
      "    popq  %rbp  \n"                        // Restore base pointer from the stack
      "    popq  %r8\n"                           // Restore r8 register from the stack
      "    movq  %rdx, %rax\n"                    // Copy value from rdx to rax
      "    movq  %rdx, %rdi\n"                    // Copy value from rdx to rdi
      "    jmp  *%r8\n"                           // Jump to the address in r8
      ".size jump_fcontext,.-jump_fcontext\n"     // Specifies the size of jump_fcontext function
      ".section .note.GNU-stack,\"\",%progbits\n" // Stack properties note for linkers
      ".previous\n");                             // Return to previous section

__asm(".text\n"                         // Section for code
      ".globl make_fcontext\n"          // Globally visible symbol
      ".type make_fcontext,@function\n" // Specifies the symbol type
      ".align 16\n"                     // Aligns the code at 16-byte boundaries
      "make_fcontext:\n"                // Start of function named 'make_fcontext'
      "    movq  %rdi, %rax\n"          // Copy the value of rdi into rax
      "    andq  $-16, %rax\n"          // Align rax to the nearest lower 16-byte boundary
      "    leaq  -0x48(%rax), %rax\n"   // Decrement rax by 72 bytes (0x48) (allocate stack memory)
      "    movq  %rdx, 0x38(%rax)\n"    // Store the value of rdx at memory address rax + 56 bytes (0x38)
      "    stmxcsr  (%rax)\n"           // Store the contents of the MXCSR register at the address in rax
      "    fnstcw   0x4(%rax)\n"        // Store the control word of FPU at memory address rax + 4 bytes
      "    leaq  finish(%rip), %rcx\n"  // Load the address of the `finish` label into rcx
      "    movq  %rcx, 0x40(%rax)\n"    // Store the address from rcx at memory address rax + 64 bytes (0x40)
      "    ret \n"                      // Return from the function
      "finish:\n"                       // Label 'finish'
      "    xorq  %rdi, %rdi\n"          // Zero out the rdi register
      "    call  _exit@PLT\n"           // Call the exit function to terminate
      "    hlt\n"                       // Halt the processor (This should never be reached due to the above _exit call)
      ".size make_fcontext,.-make_fcontext\n"     // Specifies the size of make_fcontext function
      ".section .note.GNU-stack,\"\",%progbits\n" // Stack properties note for linkers
      ".previous\n");                             // Return to previous section

fcontext_t ctx_main;
fcontext_t ctx1;
fcontext_t ctx2;

void func1(intptr_t) {
    std::cout << "func1 step1, jump to func2" << std::endl;
    jump_fcontext(&ctx1, ctx2, 0, false);
    std::cout << "func1 step2, jump to func2" << std::endl;
    jump_fcontext(&ctx1, ctx2, 0, false);
    std::cout << "func1 step3, jump to func2" << std::endl;
    jump_fcontext(&ctx1, ctx2, 0, false);
    std::cout << "func1 step4, jump to func2" << std::endl;
    jump_fcontext(&ctx1, ctx2, 0, false);
    std::cout << "func1 step5, jump to func2" << std::endl;
    jump_fcontext(&ctx1, ctx2, 0, false);
}

void func2(intptr_t) {
    std::cout << "func2 step1, jump to func1" << std::endl;
    jump_fcontext(&ctx2, ctx1, 0, false);
    std::cout << "func2 step2, jump to func1" << std::endl;
    jump_fcontext(&ctx2, ctx1, 0, false);
    std::cout << "func2 step3, jump to func1" << std::endl;
    jump_fcontext(&ctx2, ctx1, 0, false);
    std::cout << "func2 step4, jump to func1" << std::endl;
    jump_fcontext(&ctx2, ctx1, 0, false);
    std::cout << "func2 step5, jump to main" << std::endl;
    jump_fcontext(&ctx2, ctx_main, 0, false);
}

int main() {
    char sp1[4096];
    char sp2[4096];
    std::cout << "create ctx1" << std::endl;
    ctx1 = make_fcontext(sp1 + sizeof(sp1), sizeof(sp1), &func1);
    std::cout << "create ctx2" << std::endl;
    ctx2 = make_fcontext(sp2 + sizeof(sp2), sizeof(sp2), &func2);

    std::cout << "start, jump to func1" << std::endl;
    jump_fcontext(&ctx_main, ctx1, 0, false);
    std::cout << "end, back to main" << std::endl;

    return 0;
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
create ctx1
create ctx2
start, jump to func1
func1 step1, jump to func2
func2 step1, jump to func1
func1 step2, jump to func2
func2 step2, jump to func1
func1 step3, jump to func2
func2 step3, jump to func1
func1 step4, jump to func2
func2 step4, jump to func1
func1 step5, jump to func2
func2 step5, jump to main
end, back to main

7 GNU Tools

  1. ld：the GNU linker
  2. as：the GNU assembler
  3. gold：a new, faster, ELF only linker
  4. addr2line：Converts addresses into filenames and line numbers
  5. ar：A utility for creating, modifying and extracting from archives
  6. c++filt - Filter to demangle encoded C++ symbols.
  7. dlltool - Creates files for building and using DLLs.
  8. elfedit - Allows alteration of ELF format files.
  9. gprof - Displays profiling information.
  10. nlmconv - Converts object code into an NLM.
  11. nm - Lists symbols from object files.
  12. objcopy - Copies and translates object files.
  13. objdump - Displays information from object files.
  14. ranlib - Generates an index to the contents of an archive.
  15. readelf - Displays information from any ELF format object file.
  16. size - Lists the section sizes of an object or archive file.
  17. strings - Lists printable strings from files.
  18. strip - Discards symbols.
  19. windmc - A Windows compatible message compiler.
  20. windres - A compiler for Windows resource files.

7.1 gcc

常用参数说明：

  1. 文件类型
    • -E：生成预处理文件（.i）
    • -S：生成汇编文件（.s）
      • -fverbose-asm：带上一些注释信息
    • -c：生成目标文件（.o）
    • 默认生成可执行文件
  2. 优化级别
    1. -O0（默认）：不做任何优化
    2. -O/-O1：在不影响编译速度的前提下，尽量采用一些优化算法降低代码大小和可执行代码的运行速度
    3. -O2：该优化选项会牺牲部分编译速度，除了执行-O1所执行的所有优化之外，还会采用几乎所有的目标配置支持的优化算法，用以提高目标代码的运行速度
    4. -O3：该选项除了执行-O2所有的优化选项之外，一般都是采取很多向量化算法，提高代码的并行执行程度，利用现代CPU中的流水线，Cache等
    • 不同优化等级对应开启的优化参数参考man page
  3. 调试
    • -gz[=type]：对于DWARF格式的文件中的调试部分按照指定的压缩方式进行压缩
    • -g：生成调试信息
    • -ggdb：生成gdb专用的调试信息
    • -gdwarf/-gdwarf-version：生成DWARF格式的调试信息，version的可选值有2/3/4/5，默认是4
  4. -print-search-dirs：打印搜索路径
  5. -I <path>：增加头文件搜索路径
    • 可以并列使用多个-I参数，例如-I path1 -I path2
  6. -L <path>：增加库文件搜索路径
  7. -l<lib_name>：增加库文件
    • 搜索指定名称的静态或者动态库，如果同时存在，默认选择动态库
    • 例如-lstdc++、-lpthread
  8. -std=<std_version>：指定标注库类型以及版本信息
    • 例如-std=gnu++17
  9. -W<xxx>：warning提示
    • -Wall：启用大部分warning提示（部分warning无法通过该参数默认启用）
    • -Wno<xxx>：关闭指定种类的warning提示
    • -Werror：所有warning变为error（会导致编译失败）
    • -Werror=<xxx>：指定某个warning变为error（会导致编译失败）
  10. -static：所有库都用静态链接，包括libc和libc++
  11. -D <macro_name>[=<macro_definition>]：定义宏
    • 例如-D MY_DEMO_MACRO、-D MY_DEMO_MACRO=2、-D 'MY_DEMO_MACRO="hello"'、-D 'ECHO(a)=(a)'
  12. -U <macro_name>：取消宏定义
  13. --coverage：覆盖率测试
  14. -fno-access-control：关闭访问控制，例如在类外可以直接访问某类的私有字段和方法，一般用于单元测试
  15. 向量化相关参数
    • -fopt-info-vec/-fopt-info-vec-optimized：当循环进行向量化优化时，输出详细信息
    • -fopt-info-vec-missed：当循环无法向量化时，输出详细信息
    • -fopt-info-vec-note：输出循环向量化优化的所有详细信息
    • -fopt-info-vec-all：开启所有输出向量化详细信息的参数
    • -fno-tree-vectorize：关闭向量化
    • 一般来说，需要指定参数后才能使用更大宽度的向量化寄存器
      • -mmmx
      • -msse、-msse2、-msse3、-mssse3、-msse4、-msse4a、-msse4.1、-msse4.2
      • -mavx、-mavx2、-mavx512f、-mavx512pf、-mavx512er、-mavx512cd、-mavx512vl、-mavx512bw、-mavx512dq、-mavx512ifma、-mavx512vbmi
      • …
  16. -fPIC：如果目标机器支持，则生成与位置无关的代码，适用于动态链接并避免对全局偏移表大小的任何限制
  17. -fomit-frame-pointer：必要的话，允许部分函数没有栈指针
    • -fno-omit-frame-pointer：所有函数必须包含栈指针
  18. -faligned-new
  19. -fsized-deallocation：启用接收size参数的delete运算符。C++ Sized Deallocation。现代内存分配器在给对象分配内存时，需要指定大小，出于空间利用率的考虑，不会在对象内存周围存储对象的大小信息。因此在释放对象时，需要查找对象占用的内存大小，查找的开销很大，因为通常不在缓存中。因此，编译器允许提供接受一个size参数的global delete operator，并用这个版本来对对象进行析构
  20. -mcmodel=small/medium/large: is an option used in compilers like GCC (GNU Compiler Collection) to specify the memory model for code generation. This option is particularly relevant in systems with large address spaces, such as 64-bit architectures, where how the program accesses memory can significantly impact performance and compatibility.
    • small
      • This is the default memory model.
      • Assumes that all symbols are within 2GB of each other.
      • Code and data are assumed to be close, which allows the compiler to use shorter and more efficient instructions for calls and references.
      • Suitable for most applications where the total memory usage (including code, data, and stack) does not exceed 2GB.
    • medium
      • Used for applications larger than 2GB but less than a certain threshold (often around tens of GBs).
      • Code is generated under the assumption that it will be within 2GB, but data may be farther away.
      • This model uses absolute addresses for data and thus can handle larger data sizes, but still has limitations on the size of the code.
    • large
      • Designed for applications where both the code and data are larger than the limits of the medium model.
      • Uses absolute addresses for both data and code, allowing for very large applications.
      • However, this comes at the cost of efficiency, as the generated code is less optimized compared to the small and medium models.

Use g++:

1
2
3
4
5
g++ -o main main.cpp -static-libstdc++

# Or

g++ -o main main.cpp -static

Use gcc:

1
2
3
4
5
gcc main.cpp -o main -std=gnu++17 -Wl,-Bstatic -lstdc++ -Wl,-Bdynamic

# Or

gcc main.cpp -o main -std=gnu++17 -static -lstdc++

Use cmake:

1
2
target_compile_options(<target> PRIVATE -static-libstdc++)
target_link_options(<target> PRIVATE -static-libstdc++)

7.2 ld

种类

  • GNU ld
  • GNU gold
  • LLVM lld
  • mold

gcc如何指定linker

  • -B/usr/bin: For GNU ld, setup searching directory
  • -fuse-ld=gold: For GNU gold
  • -fuse-ld=lld: For LLVM lld
  • -B/usr/local/bin/gcc-mold: For mold, setup searching directory

常用参数说明：

  • --verbose：打印链接时的详细信息，包括链接了哪些动态库
  • -l <name>：增加库文件，查找lib<name>.a或者lib<name>.so，如果都存在，默认使用so版本
  • -L <dir>：增加库文件搜索路径，其优先级会高于默认的搜索路径。允许指定多个，搜索顺序与其指定的顺序相同
  • -rpath=<dir>：增加运行时库文件搜索路径（务必用绝路径，否则二进制一旦换目录就无法运行了）。-L参数只在编译、链接期间生效，运行时仍然会找不到动态库文件，需要通过该参数指定。因此，对于位于非默认搜索路径下的动态库文件，-L与-Wl,-rpath=这两个参数通常是一起使用的
  • -Bstatic：修改默认行为，强制使用静态链接库，只对该参数之后出现的库有效。如果找不到对应的静态库会报错（即便有动态库）
  • -Bdynamic：修改默认行为，强制使用动态链接库，只对该参数之后出现的库有效。如果找不到对应的动态库会报错（即便有静态库）
  • --wrap=<symbol>
    • 任何指向<symbol>的未定义的符号都会被解析为__wrap_<symbol>
    • 任何指向__real_<symbol>的未定义的符号都会被解析为<symbol>
    • 于是，我们就可以在__wrap_<symbol>增加额外的逻辑，并最终调用__real_<symbol>来实现代理
    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    cat > proxy_malloc.c << 'EOF'
    #include <stddef.h>
    #include <stdio.h>
    #include <stdlib.h>

    void* __real_malloc(size_t c);

    void* __wrap_malloc(size_t c) {
        printf("malloc called with %zu\n", c);
        return __real_malloc(c);
    }

    int main() {
        int* num = (int*)malloc(sizeof(int));
        return *num;
    }
    EOF

    gcc -o proxy_malloc proxy_malloc.c -Wl,-wrap=malloc

    ./proxy_malloc

7.2.1 How to check default linker

1
2
3
ls -l $(which ld)

update-alternatives --display ld

7.2.2 How to print dynamic lib path when linking program

1
2
3
4
5
6
7
8
# default GNU ld
gcc -o your_program your_program.c -Wl,--verbose

# gold
gcc -o your_program your_program.c -fuse-ld=gold -Wl,--verbose

# lld
gcc -o your_program your_program.c -fuse-ld=lld -Wl,--verbose
1
2
3
4
5
# method 1
readelf -p .comment <binary_file>

# method 2
strings <binary_file> | grep <linker_name>

7.3 DWARF Problems

Try use different linker.

7.4 Reference

8 LLVM Tools

8.1 clang

Doc:

8.2 clang-format

如何安装clang-format

1
npm install -g clang-format

如何使用：在用户目录或者项目根目录中创建.clang-format文件用于指定格式化的方式，下面给一个示例

  • 优先使用项目根目录中的.clang-format；如果不存在，则使用用户目录中的~/.clang-format
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
---
Language: Cpp
BasedOnStyle: Google
AccessModifierOffset: -4
AllowShortFunctionsOnASingleLine: Inline
ColumnLimit: 120
ConstructorInitializerIndentWidth: 8 # double of IndentWidth
ContinuationIndentWidth: 8 # double of IndentWidth
DerivePointerAlignment: false # always use PointerAlignment
IndentCaseLabels: false
IndentWidth: 4
PointerAlignment: Left
ReflowComments: false
SortUsingDeclarations: false
SpacesBeforeTrailingComments: 1

注意：

  • 即便.clang-format相同，不同版本的clang-format格式化的结果也有差异

示例：

9 Assorted

9.1 Dynamic Analysis

analysis-tools

9.2 Header File Search Order

头文件#include "xxx.h"的搜索顺序

  1. 先搜索当前目录
  2. 然后搜索-I参数指定的目录
  3. 再搜索gcc的环境变量CPLUS_INCLUDE_PATH（C程序使用的是C_INCLUDE_PATH）
  4. 最后搜索gcc的内定目录，包括：
    • /usr/include
    • /usr/local/include
    • /usr/lib/gcc/x86_64-redhat-linux/<gcc version>/include（C头文件）或者/usr/include/c++/<gcc version>（C++头文件）

头文件#include <xxx.h>的搜索顺序

  1. 先搜索-I参数指定的目录
  2. 再搜索gcc的环境变量CPLUS_INCLUDE_PATH（C程序使用的是C_INCLUDE_PATH）
  3. 最后搜索gcc的内定目录，包括：
    • /usr/include
    • /usr/local/include
    • /usr/lib/gcc/x86_64-redhat-linux/<gcc version>/include（C头文件）或者/usr/include/c++/<gcc version>（C++头文件）

9.3 How to check the compile error message

Example:

1
2
3
4
5
6
7
8
9
10
11
In file included from /starrocks/be/src/exprs/agg/approx_top_k.h:20:
/starrocks/be/src/column/column_hash.h: In instantiation of ‘std::size_t starrocks::StdHashWithSeed<T, seed>::operator()(T) const [with T = std::vector<unsigned char, std::allocator<unsigned char> >; starrocks::PhmapSeed seed = starrocks::PhmapSeed1; std::size_t = long unsigned int]’:
/starrocks/be/src/util/phmap/phmap.h:1723:49:   required from ‘size_t phmap::priv::raw_hash_set<Policy, Hash, Eq, Alloc>::HashElement::operator()(const K&, Args&& ...) const [with K = std::vector<unsigned char, std::allocator<unsigned char> >; Args = {}; Policy = phmap::priv::FlatHashMapPolicy<std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::ApproxTopKState<starrocks::TYPE_VARCHAR>::Counter*>; Hash = starrocks::StdHashWithSeed<std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::PhmapSeed1>; Eq = phmap::EqualTo<std::vector<unsigned char, std::allocator<unsigned char> > >; Alloc = std::allocator<std::pair<const std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::ApproxTopKState<starrocks::TYPE_VARCHAR>::Counter*> >; size_t = long unsigned int]’
/starrocks/be/src/util/phmap/phmap.h:1689:39:   required from ‘size_t phmap::priv::raw_hash_set<Policy, Hash, Eq, Alloc>::hash(const K&) const [with K = std::vector<unsigned char, std::allocator<unsigned char> >; Policy = phmap::priv::FlatHashMapPolicy<std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::ApproxTopKState<starrocks::TYPE_VARCHAR>::Counter*>; Hash = starrocks::StdHashWithSeed<std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::PhmapSeed1>; Eq = phmap::EqualTo<std::vector<unsigned char, std::allocator<unsigned char> > >; Alloc = std::allocator<std::pair<const std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::ApproxTopKState<starrocks::TYPE_VARCHAR>::Counter*> >; size_t = long unsigned int]’
/starrocks/be/src/util/phmap/phmap.h:1625:36:   required from ‘phmap::priv::raw_hash_set<Policy, Hash, Eq, Alloc>::iterator phmap::priv::raw_hash_set<Policy, Hash, Eq, Alloc>::find(key_arg<K>&) [with K = std::vector<unsigned char, std::allocator<unsigned char> >; Policy = phmap::priv::FlatHashMapPolicy<std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::ApproxTopKState<starrocks::TYPE_VARCHAR>::Counter*>; Hash = starrocks::StdHashWithSeed<std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::PhmapSeed1>; Eq = phmap::EqualTo<std::vector<unsigned char, std::allocator<unsigned char> > >; Alloc = std::allocator<std::pair<const std::vector<unsigned char, std::allocator<unsigned char> >, starrocks::ApproxTopKState<starrocks::TYPE_VARCHAR>::Counter*> >; key_arg<K> = std::vector<unsigned char, std::allocator<unsigned char> >]’
/starrocks/be/src/exprs/agg/approx_top_k.h:80:23:   required from ‘void starrocks::ApproxTopKState<LT>::process(const CppType&, int64_t, bool) [with starrocks::LogicalType LT = starrocks::TYPE_VARCHAR; CppType = std::vector<unsigned char, std::allocator<unsigned char> >; int64_t = long int]’
/starrocks/be/src/exprs/agg/approx_top_k.h:75:13:   required from ‘void starrocks::ApproxTopKState<LT>::merge(std::vector<Counter>) [with starrocks::LogicalType LT = starrocks::TYPE_VARCHAR]’
/starrocks/be/src/exprs/agg/approx_top_k.h:239:32:   required from ‘void starrocks::ApproxTopKAggregateFunction<LT, T>::merge(starrocks::FunctionContext*, const starrocks::Column*, starrocks::AggDataPtr, size_t) const [with starrocks::LogicalType LT = starrocks::TYPE_VARCHAR; T = starrocks::Slice; starrocks::AggDataPtr = unsigned char*; size_t = long unsigned int]’
/starrocks/be/src/exprs/agg/approx_top_k.h:215:10:   required from here
/starrocks/be/src/column/column_hash.h:282:101: error: use of deleted function ‘std::hash<std::vector<unsigned char, std::allocator<unsigned char> > >::hash()’
  282 |     std::size_t operator()(T value) const { return phmap_mix_with_seed<sizeof(size_t), seed>()(std::hash<T>()(value)); }

When interpreting compiler error messages, especially those involving template instantiation (which can be particularly verbose and intricate), there’s a general approach that you can take. Here’s how to approach such error messages:

  1. Start At The End:
    • Often, the actual error (like a type mismatch, use of an undefined variable, etc.) is reported at the end of the message.
    • In your provided message, the core error was at the end, stating the use of a deleted function.
  2. Work Your Way Up:
    • Once you have the core error in mind, start moving upwards to see where in your code this problem originates. Often, the actual line of your code that triggers the error is one of the last things mentioned before the core error.
    • In your case, the direct line of code causing the error was provided in column_hash.h.
  3. Template Call Stack:
    • Template errors tend to provide a “call stack” of sorts which traces the sequence of template instantiations that led to the error. This can be useful to understand the flow of logic and how you arrived at the error.
    • In your message, this was the sequence starting from phmap.h:1723:49 and moving through various template functions/classes.
  4. Context:
    • The initial part of the message usually provides the broader context: where the error started, what file it originated from, and so on.
    • In your output, the initial message pointed to column_hash.h being included from approx_top_k.h, providing a starting point for the cascade of template instantiation.

In Summary: While the bottom-up approach is useful for quickly identifying the core error and the immediate lines of code causing it, you sometimes need to go top-down to fully understand the context and sequence of events leading to the error. With experience, you’ll develop an intuition for quickly scanning and pinpointing the most relevant parts of such error messages.

9.4 How to get coverage of code

Here’s how it works: gcov determines which files to analyze for coverage information based on the profile data files (*.gcda and *.gcno) that are generated when you compile and run your program with the appropriate GCC flags (-fprofile-arcs and -ftest-coverage). Here’s a breakdown of how gcov knows which files to load:

  1. Compilation and Execution:
    • When you compile your C/C++ program with GCC using -fprofile-arcs and -ftest-coverage, it produces two types of files for each source file:
    • *.gcno: These are generated during the compilation. They contain information about the source code’s structure (like basic blocks, branches, etc.).
    • *.gcda: These are generated when the compiled program is run. They contain the actual coverage data, such as how many times each line of code was executed.
  2. Locating Files:
    • When you invoke gcov, it looks for .gcno and .gcda files in the current directory by default. These files are named after the source files but with different extensions.
    • For example, if your source file is example.c, gcov will look for example.gcno and example.gcda in the current directory.
  3. Matching Source Files:
    • gcov uses the information in these files to match with the corresponding source file. It relies on the path and filename information stored in the .gcno and .gcda files to find the correct source file.
    • If your source files are in a different directory, you may need to specify the path when running gcov, or run gcov from the directory containing the .gcda and .gcno files.
  4. Generating the Report:
    • Once gcov has matched the .gcno and .gcda files with their corresponding source files, it processes these files to generate a coverage report. This report shows how many times each line in the source file was executed.
  5. Manual Specification:
    • You can also manually specify which source files to generate reports for by passing their names as arguments to gcov.

Here’s an example:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
#include <iostream>

void printMessage() {
    std::cout << "Hello, World!" << std::endl;
}

void unusedFunction() {
    std::cout << "This function is never called." << std::endl;
}

int main() {
    printMessage();
    return 0;
}
1
2
3
4
g++ --coverage -o example example.cpp
./example
gcov example.cpp
cat example.cpp.gcov

9.5 How to check standard library search path when compiling

Add -v option.

9.6 Document

  1. cpp reference
  2. cppman
    • 安装：pip install cppman
    • 示例：cppman vector::begin
    • 重建索引：cppman -r

9.7 Reference