寒假捉虫记——从一段损坏的调用栈开始折腾

　　放假在家，继续调试《家园》。目前的进度是MinGW上的编译链接都已通过，游戏程序也已经可以跑起来并进入主菜单界面，但加载关卡之后就会闪退。这让我想起了以前上中学时玩盗版游戏的日子。那个年代的单机游戏估计大多是用C/C++写的，一个不小心的内存操作就会让进程崩掉；而且那个年代的操作系统没现在稳定，可能破解技术也不够先进，从电脑城里买来的五六块钱的盗版游戏质量参差不齐。很多游戏跑着跑着就闪退，有的甚至连打都打不开，让人甚为恼火。如今源代码在手，并且我也是程序员了，可以对闪退的原因一探究竟，再也不用怕。

　　不过让人失望的是，用MinGW构建出的程序不会像Linux程序那样在崩溃时吐核。还好这回的闪退是可以必现的，所以就在gdb中运行程序，看看它崩在什么地方。

　　结果程序如预期崩溃后，调用栈成了下面这个样子：

Program received signal SIGSEGV, Segmentation fault.

0x0ddcf5c0 in ?? ()

(gdb) bt

#0  0x0ddcf5c0 in ?? ()

#1  0xabababab in ?? ()

#2  0x0000abab in ?? ()

#3  0x00000000 in ?? ()

(gdb)

　　看样子调用栈已经坏掉了。记得以前在老东家遇到过这种损坏的调用栈，但后来很喜感地发现原来是机器的内存坏了。我相信我家电脑还没有到如此风烛残年的地步。

　　在StackOverflow上搜到一篇帖子：http://stackoverflow.com/questions/9809810/gdb-corrupted-stack-frame-how-to-debug，正好是我想问的问题：如何在gdb中调试这种已经损坏的调用栈。帖子里的答案说这种情况99%是因为调用了非法的函数指针。在32位环境中可以用如下方法恢复调用栈：

(gdb) set $pc = *(void **)$esp

(gdb) set $esp = $esp + 4

　　可是我检查了一下esp寄存器指向的内存：

(gdb) p $esp

$1 = (void *) 0x22fa34

(gdb) x $esp

0x22fa34:       0x0000007f

(gdb) x/i 0x7f

0x7f:        Cannot access memory at address 0x7f

(gdb)

　　0x7f显然不可能是一个合法的指令地址。看来我落到剩下那1%的区间里了=A=。

　　把这事情分享到朋友圈里后，主程建议我让程序链接tcmalloc试试，看看能否让程序在应用代码进行非法内存操作时就崩溃，兴许那时调用栈还没损坏。可是试过后情况并无改变。不过这倒是提醒了我，以后不妨让自己的程序都链接tcmalloc，这样可以让很多问题都提前暴露。顺便写下，我的tcmalloc链接选项是-L/local/lib -ltcmalloc_minimal -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc
-fno-builtin-free。

　　无奈，最后还是通过打日志和单步调试的方法，通过应用代码本身的逻辑定位到了具体崩溃位置。原来程序崩在了一个OpenGL接口——glDrawElements的调用。多留日志和熟练掌握项目代码逻辑真是重要啊。

　　如果是在工作中，我的排查工作一般在这一步也就该结束了。因为我从未接触过OpenGL，所以此时我应该让有OpenGL经验的同事来帮忙处理。不过这回不是在工作而是在玩耍，所以我打算满足下自己用牛刀杀鸡、用导弹打蚊子的癖好，好好研究一番，一是看看能否在gdb中恢复出调用栈，二是研究下glDrawElements这个调用为啥会崩溃，趁机接触下OpenGL。

调用栈的恢复

　　先看看崩溃的直接原因是什么，看看崩溃时执行的汇编指令是什么：

(gdb) x/i $pc

=> 0xddcf5c0:   mov    (%esi),%edi

　　看来esi寄存器里存了一个非法内存地址。

(gdb) p/x $esi

$3 = 0xfeee4

(gdb) p *(void **)$esi

Cannot access memory at address 0xf00d4

(gdb)

　　果然如此。

　　再看看这之前还执行了什么指令。

(gdb) x/40i $pc-90

0xc54f7a6:   add    %al,(%eax)

0xc54f7a8:   add    %al,(%eax)

0xc54f7aa:   add    %al,(%eax)

0xc54f7ac:   add    %al,(%eax)

0xc54f7ae:   add    %al,(%eax)

0xc54f7b0:   add    %al,(%eax)

0xc54f7b2:   add    %al,(%eax)

0xc54f7b4:   add    %al,(%eax)

0xc54f7b6:   add    %al,(%eax)

0xc54f7b8:   cmp    $0xff,%bh

0xc54f7bb:   incl   0x550000f7(%eax)

0xc54f7c1:   mov    %esp,%ebp

0xc54f7c3:   push   %ebx

0xc54f7c4:   push   %esi

0xc54f7c5:   push   %edi

0xc54f7c6:   mov    0x8(%ebp),%ebx

0xc54f7c9:   mov    0xc(%ebp),%eax

0xc54f7cc:   mov    0x14(%ebp),%ebp

0xc54f7cf:   mov    %ebp,%edi

0xc54f7d1:   shl    $0x14,%edi

0xc54f7d4:   lea    0x40003640(%edi),%esi

0xc54f7da:   mov    %esi,(%eax)

0xc54f7dc:   add    $0x4,%eax

0xc54f7df:   mov    0x1c(%esp),%edx

0xc54f7e3:   lea    (%edx,%ebp,2),%ebp

0xc54f7e6:   mov    %ebp,0x20(%esp)

0xc54f7ea:   movzwl (%edx),%ecx

0xc54f7ed:   add    $0x2,%edx

0xc54f7f0:   mov    0xc509890,%esi

0xc54f7f6:   mov    0x4(%esi),%esi

0xc54f7f9:   mov    %ecx,%edi

0xc54f7fb:   shl    $0x4,%edi

0xc54f7fe:   add    %edi,%esi

=> 0xc54f800:   mov    (%esi),%edi

0xc54f802:   mov    0x4(%esi),%ebp

0xc54f805:   mov    %edi,(%eax)

0xc54f807:   mov    %ebp,0x4(%eax)

0xc54f80a:   mov    0x8(%esi),%edi

0xc54f80d:   mov    %edi,0x8(%eax)

0xc54f810:   mov    0xc509890,%esi

(gdb)

　　看样子在0xc54f7bb附近很可能有一个函数头。函数开头通常由两条汇编指令组成——第一条指令保存当前栈帧的帧底地址，第二条指令将当前的栈顶指为栈帧底，开启新栈帧：

push %ebp

move %esp %ebp

　　于是从0xc54f7bc开始，一路用x命令检查：

(gdb) x/40i 0xc54f7bc

...

(gdb) x/40i 0xc54f7bd

...

(gdb) x/40i 0xc54f7be

0xc54f7be:   add    %al,(%eax)

0xc54f7c0:   push   %ebp

0xc54f7c1:   mov    %esp,%ebp

   0xc54f7c3:   push   %ebx

0xc54f7c4:   push   %esi

0xc54f7c5:   push   %edi

0xc54f7c6:   mov    0x8(%ebp),%ebx

0xc54f7c9:   mov    0xc(%ebp),%eax

0xc54f7cc:   mov    0x14(%ebp),%ebp

0xc54f7cf:   mov    %ebp,%edi

0xc54f7d1:   shl    $0x14,%edi

0xc54f7d4:   lea    0x40003640(%edi),%esi

0xc54f7da:   mov    %esi,(%eax)

0xc54f7dc:   add    $0x4,%eax

0xc54f7df:   mov    0x1c(%esp),%edx

0xc54f7e3:   lea    (%edx,%ebp,2),%ebp

0xc54f7e6:   mov    %ebp,0x20(%esp)

0xc54f7ea:   movzwl (%edx),%ecx

0xc54f7ed:   add    $0x2,%edx

0xc54f7f0:   mov    0xc509890,%esi

0xc54f7f6:   mov    0x4(%esi),%esi

0xc54f7f9:   mov    %ecx,%edi

0xc54f7fb:   shl    $0x4,%edi

0xc54f7fe:   add    %edi,%esi

=> 0xc54f800:   mov    (%esi),%edi

0xc54f802:   mov    0x4(%esi),%ebp

0xc54f805:   mov    %edi,(%eax)

0xc54f807:   mov    %ebp,0x4(%eax)

0xc54f80a:   mov    0x8(%esi),%edi

0xc54f80d:   mov    %edi,0x8(%eax)

0xc54f810:   mov    0xc509890,%esi

0xc54f816:   mov    0x54(%esi),%esi

0xc54f819:   mov    %ecx,%edi

0xc54f81b:   shl    $0x4,%edi

0xc54f81e:   add    %edi,%esi

0xc54f820:   mov    (%esi),%edi

0xc54f822:   mov    %edi,0xc(%eax)

0xc54f825:   add    $0x10,%eax

0xc54f828:   cmp    0x20(%esp),%edx

0xc54f82c:   jne    0xc54f7ea

(gdb)

　　果真如此。地址0xc54f7c0和0xc54f7c1这两条指令就是典型的函数开头：

0xc54f7c0:   push   %ebp

0xc54f7c1:   mov    %esp,%ebp

　　从这之后到崩溃处0xc54f800，有两处修改ebp的指令：

0xc54f7cc:   mov    0x14(%ebp),%ebp

...

0xc54f7e3:   lea    (%edx,%ebp,2),%ebp

　　因此在0xc54f800: mov (%esi),%edi 崩溃的时候，寄存器中记录的就是错误的栈帧。也就是说，在gdb中查看的调用栈不正常是因为ebp被篡改了。从这两条指令还可以看出，崩溃时ebp的值取决于传入函数的参数。

　　在指令0xc54f7c1: mov %esp,%ebp刚执行之后，esp的值和ebp的值是相等的。在这之后直到崩溃前，只有三条压栈指令（0xc54f7c6至0xc54f7cc）会修改esp。它们会使esp自减3个word，即3*4=12字节。分析到这里，就有办法恢复ebp的值了：

(gdb) set $ebp = $esp + 12

　　通过检查esp指向的内存段可以进一步确认：

(gdb) x/8x $esp

0x22fa34:       0x0000007f      0x0c4f0000      0x0cbb6660      0x0022fa70

0x22fa44:       0x69a84ce2      0x0c4f0000      0x0cd3e900      0x0cbb6660

(gdb)

　　可见栈在内存段0x22fa??附近，0x002fa70想必就是上一个栈帧的帧底，0x69a84ce2就是函数调用前的指令地址，也就是函数的返回地址。

　　现在可以看到正确的调用栈了：

(gdb) bt

#0  0x0c54f800 in ?? ()

#1  0x69a84ce2 in nvoglv32!DrvPresentBuffers () from C:\Windows\system32\nvoglv32.dll

#2  0x69a85ed6 in nvoglv32!DrvPresentBuffers () from C:\Windows\system32\nvoglv32.dll

#3  0x69a86214 in nvoglv32!DrvPresentBuffers () from C:\Windows\system32\nvoglv32.dll

#4  0x695f8988 in ?? () from C:\Windows\system32\nvoglv32.dll

#5  0x0049dd33 in btgRender () at ../../../src/Game/BTG.c:1307

#6  0x00417d0c in rndBackgroundRender (radius=100000, camera=0x9259c0 <universe+32>, bDrawStars=1) at ../../../src/SDL/render.c:1287

#7  0x00419c37 in rndMainViewRenderFunction (camera=0x9259c0 <universe+32>) at ../../../src/SDL/render.c:2440

#8  0x0040d730 in mrRegionDraw (reg=0xa66a0a0) at ../../../src/SDL/mainrgn.c:5509

#9  0x0056508f in regFunctionsDraw () at ../../../src/Game/Region.c:1094

#10 0x0041c47a in rndRenderTask (taskContextPtr=0xa666710) at ../../../src/SDL/render.c:3869

#11 0x005b0bc7 in taskExecuteAllPending (ticks=4) at ../../../src/Game/Task.c:370

#12 0x0042bb26 in utyTasksDispatch () at ../../../src/SDL/utility.c:4721

#13 0x00402df6 in HWSDL_main (argc=4, argv=0x59315d8) at ../../../src/SDL/main.c:2252

#14 0x004013e0 in main (argc=4, argv=0x59315d8) at ../../src/homeworld.c:32

(gdb)

　　崩溃的应用程序代码（BTG.c:1307）与通过日志和单步调试分析出的结果完全一致。

　　看样子崩在了OpenGL内部。既然是崩在了glDrawElements里面，那就要研究下glDrawElements的使用，想必是API使用不当。

glDrawElements的使用

　　从glDrawElements的官方文档来看，这个接口的作用是批量绘制多个基本图元（如点、线、三角形和多边形）。不过这个接口并没有参数可以直接传入顶点数据，那个indices参数只是顶点数据的索引而已。这一点让我花了很长时间琢磨。用glDrawElements作关键词搜了很多文章，基本都能看懂，但还是不知道该如何从零开始用起来。

　　还是先把这函数放一放，从基本的OpenGL程序开始吧，先写个Helloworld。从网上的文章得知现代OpenGL和过去的OpenGL 1.x在用法上似乎有很大不同；而《家园》是很老的游戏了，早在1999年就已经发行，即便是HomeworldSDL的代码也非常老，更新很缓慢。所以我恐怕还得学习老式OpenGL的用法。幸运的是从这里搜到了Tutorial：http://en.wikibooks.org/wiki/OpenGL_Programming#Legacy_OpenGL_1.x。根据以下两节教程写出了一个Windows上运行的OpenGL小程序window.c：

http://en.wikibooks.org/wiki/OpenGL_Programming/GLStart/Tut1

http://en.wikibooks.org/wiki/OpenGL_Programming/GLStart/Tut2

#include <windows.h>
#include <GL/gl.h>
#include <GL/glu.h>

HDC hDC; //device context
HGLRC hglrc; //rendering context

void SetupPixels(HDC hDC)
{
    int pixelFormat;
    PIXELFORMATDESCRIPTOR pfd;
    pfd.nSize = sizeof(PIXELFORMATDESCRIPTOR);
    pfd.dwFlags = PFD_SUPPORT_OPENGL | PFD_DRAW_TO_WINDOW | PFD_DOUBLEBUFFER;
    pfd.nVersion = 1;
    pfd.iPixelType = PFD_TYPE_RGBA;
    pfd.cColorBits = 32;
    pfd.cDepthBits = 24;
    pixelFormat = ChoosePixelFormat(hDC, &pfd);
    if(!SetPixelFormat(hDC, pixelFormat, &pfd))
    {
         MessageBox(NULL,"Error setting up Pixel Format","ERROR",MB_OK);
         PostQuitMessage(0);
    }
}

void Resize(int width, int height)
{
    glViewport(0,0,(GLsizei)width,(GLsizei)height);
    glMatrixMode(GL_PROJECTION);
    glLoadIdentity();
    gluPerspective(45.0f,(GLfloat)width/(GLfloat)height,1.0f,1000.0f);
    glMatrixMode(GL_MODELVIEW);
    glLoadIdentity();
}

void Render()
{
    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
    glLoadIdentity();
    glTranslatef(0.0f,0.0f,-4.0f);
    glColor3f(0.0f,0.0f,1.0f);
    glBegin(GL_POLYGON);
    glVertex3f(1.0f,1.0f,0.0f);
    glVertex3f(-1.0f,1.0f,0.0f);
    glVertex3f(-1.0f,-1.0f,0.0f);
    glVertex3f(1.0f,-1.0f,0.0f);
    glEnd();
}

LRESULT CALLBACK WinProc(HWND hWnd,
                         UINT msg,
                         WPARAM wParam,
                         LPARAM lParam)
{
    int w,h;
    switch(msg)
    {
    case WM_CREATE:
        hDC = GetDC(hWnd);
        SetupPixels(hDC);
        hglrc = wglCreateContext(hDC);
        wglMakeCurrent(hDC, hglrc);
        break;
    case WM_DESTROY:
        wglMakeCurrent(hDC,NULL);
        wglDeleteContext(hglrc);
        PostQuitMessage(0);
        break;
    case WM_SIZE:
        w = LOWORD(lParam);
        h = HIWORD(lParam);
        Resize(w,h);
        break;
    default: break;
    }
    return DefWindowProc(hWnd,msg,wParam,lParam);
}

int WINAPI WinMain(HINSTANCE hInstance,
                   HINSTANCE hPrevInstance,
                   LPSTR lpCmdLine,
                   int nShowCmd)
{
    HWND hWnd;
    WNDCLASSEX wcex;
    MSG msg;

    wcex.cbSize = sizeof(WNDCLASSEX);
    wcex.style = CS_HREDRAW | CS_VREDRAW;
    wcex.lpfnWndProc = WinProc;
    wcex.cbClsExtra = 0;
    wcex.cbWndExtra = 0;
    wcex.hInstance = hInstance;
    wcex.hIcon = LoadIcon(NULL,IDI_APPLICATION);
    wcex.hCursor = LoadCursor(NULL,IDC_ARROW);
    wcex.hbrBackground = (HBRUSH) GetStockObject(GRAY_BRUSH);
    wcex.lpszMenuName = NULL;
    wcex.lpszClassName = "WinClass";
    wcex.hIconSm = NULL;

    RegisterClassEx(&wcex);

    hWnd = CreateWindow("WinClass","My Window",
        WS_OVERLAPPEDWINDOW,0,0,400,400,NULL,NULL,
        hInstance,NULL);

    if(hWnd == NULL)
    {
        MessageBox(NULL,"Error: Unable to create Window","ERROR",MB_OK);
        return -1;
    }

    ShowWindow(hWnd,nShowCmd);
    UpdateWindow(hWnd);
    hDC = GetDC(hWnd);
    glClearColor(0.0f,0.0f,0.0f,0.0f);

    while(1)
    {
        Render();
        SwapBuffers(hDC);
        if(PeekMessage(&msg,NULL,0,0,PM_REMOVE))
        {
            if(msg.message == WM_QUIT) break;
            TranslateMessage(&msg);
            DispatchMessage(&msg);
        }
    }

    return 0;
}

　　MinGW上的编译命令：gcc -o window.exe window.c -mwindows -lopengl32 -lglu32

　　运行结果如下：

　　程序中关键的绘制代码就是Render函数中从glBegin到glEnd的部分。

glBegin(GL_POLYGON);
glVertex3f(1.0f,1.0f,0.0f);
glVertex3f(-1.0f,1.0f,0.0f);
glVertex3f(-1.0f,-1.0f,0.0f);
glVertex3f(1.0f,-1.0f,0.0f);
glEnd();

　　这段代码绘制了四个顶点，从而绘制出一个正方形。这种用法是Tutorial 3中提到的Immediate
Mode。看来如果要用glDrawElements，关键就是将这段代码替换为glDrawElements。

　　现在关键是要知道如何将顶点数据传给glDrawElements。从HomeworldSDL的代码和glDrawElements相关的资料中，我注意到两个概念：Vertex Array Object（VAO）和Vertex Buffer Object（VBO）。Tutorial
3的末尾就简略地提及了这两个概念。简要地说，VAO就是我们要传给glDrawElements的顶点数据，这些数据是以数组形式存放的。glDrawElements被调用时可以从内存里拿这个数据，也可以从显存里拿，后一种方式的性能更好。如果是从显存里拿数据，那glDrawElements拿的数据就是VBO。HomeworldSDL崩溃时用的就是VBO。具体说明可以参考OpenGL官方文档中的Vertex
Specification。

　　于是归纳了一下，采用VBO方式使用glDrawElements的步骤大致如下：

1、初始化，准备好顶点数据

1）用glGenBuffers申请Buffer
Object的名字，也就是为即将分配的显存申请ID。

2）用glBindBuffer绑定Buffer
Object，这样在再次调用glBindBuffer之前，接下来的Buffer Object相关的操作都是针对当前绑定的Buffer Object。target参数需是GL_ARRAY_BUFFER。

3）用glBufferData分配并初始化一段显存，将顶点数据传进显存。

4）再次调用glBindBuffer解绑Buffer Object。

2、准备好索引数组，步骤和1类似，只是调用glBindBuffer时target需是GL_ELEMENT_ARRAY_BUFFER。

3、调用glVertexPointer，指定顶点数据。

4、调用glDrawElements进行绘制。

5、如果所有绘制工作完成，之前的显存不再需使用，就要调用glDeleteBuffers释放显存。

　　需要注明的是，以上只是一种简单的使用VBO的方式，并不是说glDrawElements一定要严格按照这个流程。在顶点和索引数据都不会变化的情况下，3~4两步可以反复执行。索引数组也不一定要放在显存里，可以在调用glDrawElements的时候直接通过参数将内存中的数组传进去，这时就不需要第2步。同理如果顶点数据也不用VBO的话，那第1步也省去了，glGenBuffers、glBindBuffer、glBufferData和glDeleteBuffers这几个函数都不用调用。

　　很快，我的glDrawElements版本的Helloworld出炉了：

#include <windows.h>
#include <GL/gl.h>
#include <GL/glu.h>

HDC hDC; //device context
HGLRC hglrc; //rendering context

#ifdef USE_VBO
GLfloat transVerts[] =
{
    1.0f, 1.0f, 0.0f,
    -1.0f, 1.0f, 0.0f,
    -1.0f,-1.0f, 0.0f,
    1.0f, -1.0f, 0.0f
};
GLuint vboTransVerts;
GLushort indices[] =
{
    0, 1, 2,
    3, 4, 5,
    6, 7, 8,
    9, 10, 11
};
GLuint vboIndices;
#endif

void SetupPixels(HDC hDC)
{
    int pixelFormat;
    PIXELFORMATDESCRIPTOR pfd;
    pfd.nSize = sizeof(PIXELFORMATDESCRIPTOR);
    pfd.dwFlags = PFD_SUPPORT_OPENGL | PFD_DRAW_TO_WINDOW | PFD_DOUBLEBUFFER;
    pfd.nVersion = 1;
    pfd.iPixelType = PFD_TYPE_RGBA;
    pfd.cColorBits = 32;
    pfd.cDepthBits = 24;
    pixelFormat = ChoosePixelFormat(hDC, &pfd);
    if(!SetPixelFormat(hDC, pixelFormat, &pfd))
    {
         MessageBox(NULL,"Error setting up Pixel Format","ERROR",MB_OK);
         PostQuitMessage(0);
    }
}

void Resize(int width, int height)
{
    glViewport(0,0,(GLsizei)width,(GLsizei)height);
    glMatrixMode(GL_PROJECTION);
    glLoadIdentity();
    gluPerspective(45.0f,(GLfloat)width/(GLfloat)height,1.0f,1000.0f);
    glMatrixMode(GL_MODELVIEW);
    glLoadIdentity();
}

#ifdef USE_VBO
void InitVBO()
{
    glGenBuffers(1, &vboTransVerts);
    glBindBuffer(GL_ARRAY_BUFFER, vboTransVerts);
    glBufferData(GL_ARRAY_BUFFER, sizeof(transVerts), transVerts, GL_STATIC_DRAW);
    glBindBuffer(GL_ARRAY_BUFFER, 0);
    glGenBuffers(1, &vboIndices);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vboIndices);
    glBufferData(GL_ELEMENT_ARRAY_BUFFER, sizeof(indices), indices, GL_STATIC_DRAW);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0);
}
#endif

void Render()
{
    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
    glLoadIdentity();
    glTranslatef(0.0f,0.0f,-4.0f);
    glColor3f(0.0f,0.0f,1.0f);
#ifdef USE_VBO
    glEnableClientState(GL_VERTEX_ARRAY);
    glBindBuffer(GL_ARRAY_BUFFER, vboTransVerts);
    glVertexPointer(3, GL_FLOAT, 0, 0);
    glBindBuffer(GL_ARRAY_BUFFER, 0);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vboIndices);
    glDrawElements(GL_POLYGON, sizeof(transVerts) / sizeof(transVerts[0]), GL_UNSIGNED_SHORT, 0);
#else
    glBegin(GL_POLYGON);
    glVertex3f(1.0f,1.0f,0.0f);
    glVertex3f(-1.0f,1.0f,0.0f);
    glVertex3f(-1.0f,-1.0f,0.0f);
    glVertex3f(1.0f,-1.0f,0.0f);
    glEnd();
#endif  // USE_VBO
}

LRESULT CALLBACK WinProc(HWND hWnd,
                         UINT msg,
                         WPARAM wParam,
                         LPARAM lParam)
{
    int w,h;
    switch(msg)
    {
    case WM_CREATE:
        hDC = GetDC(hWnd);
        SetupPixels(hDC);
        hglrc = wglCreateContext(hDC);
        wglMakeCurrent(hDC, hglrc);
#ifdef USE_VBO
        InitVBO();
#endif
        break;
    case WM_DESTROY:
#ifdef USE_VBO
        glDeleteBuffers(1, &vboTransVerts);
        glDeleteBuffers(1, &vboIndices);
#endif
        wglMakeCurrent(hDC,NULL);
        wglDeleteContext(hglrc);
        PostQuitMessage(0);
        break;
    case WM_SIZE:
        w = LOWORD(lParam);
        h = HIWORD(lParam);
        Resize(w,h);
        break;
    default: break;
    }
    return DefWindowProc(hWnd,msg,wParam,lParam);
}

int WINAPI WinMain(HINSTANCE hInstance,
                   HINSTANCE hPrevInstance,
                   LPSTR lpCmdLine,
                   int nShowCmd)
{
    HWND hWnd;
    WNDCLASSEX wcex;
    MSG msg;

    wcex.cbSize = sizeof(WNDCLASSEX);
    wcex.style = CS_HREDRAW | CS_VREDRAW;
    wcex.lpfnWndProc = WinProc;
    wcex.cbClsExtra = 0;
    wcex.cbWndExtra = 0;
    wcex.hInstance = hInstance;
    wcex.hIcon = LoadIcon(NULL,IDI_APPLICATION);
    wcex.hCursor = LoadCursor(NULL,IDC_ARROW);
    wcex.hbrBackground = (HBRUSH) GetStockObject(GRAY_BRUSH);
    wcex.lpszMenuName = NULL;
    wcex.lpszClassName = "WinClass";
    wcex.hIconSm = NULL;

    RegisterClassEx(&wcex);

    hWnd = CreateWindow("WinClass","My Window",
        WS_OVERLAPPEDWINDOW,0,0,400,400,NULL,NULL,
        hInstance,NULL);

    if(hWnd == NULL)
    {
        MessageBox(NULL,"Error: Unable to create Window","ERROR",MB_OK);
        return -1;
    }

    ShowWindow(hWnd,nShowCmd);
    UpdateWindow(hWnd);
    hDC = GetDC(hWnd);
    glClearColor(0.0f,0.0f,0.0f,0.0f);

    while(1)
    {
        Render();
        SwapBuffers(hDC);
        if(PeekMessage(&msg,NULL,0,0,PM_REMOVE))
        {
            if(msg.message == WM_QUIT) break;
            TranslateMessage(&msg);
            DispatchMessage(&msg);
        }
    }

    return 0;
}

　　MinGW上的编译命令：gcc -o window.exe window.c -DUSE_VBO -mwindows -lopengl32 -lglu32

　　不过先别高兴太早，报错了：

$ gcc -o window.exe window.c -DUSE_VBO -mwindows -lopengl32 -lglu32

window.c: In function ‘InitVBO‘:

window.c:60:18: error: ‘GL_ARRAY_BUFFER‘ undeclared (first use in this function)

glBindBuffer(GL_ARRAY_BUFFER, vboTransVerts);

^

window.c:60:18: note: each undeclared identifier is reported only once for each function it appears in

window.c:61:67: error: ‘GL_STATIC_DRAW‘ undeclared (first use in this function)

glBufferData(GL_ARRAY_BUFFER, sizeof(transVerts), transVerts, GL_STATIC_DRAW);

^

window.c:64:18: error: ‘GL_ELEMENT_ARRAY_BUFFER‘ undeclared (first use in this function)

glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vboIndices);

^

window.c: In function ‘Render‘:

window.c:78:18: error: ‘GL_ARRAY_BUFFER‘ undeclared (first use in this function)

glBindBuffer(GL_ARRAY_BUFFER, vboTransVerts);

^

window.c:81:18: error: ‘GL_ELEMENT_ARRAY_BUFFER‘ undeclared (first use in this function)

glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vboIndices);

　　原来GL/gl.h和GL/glu.h里都没有定义GL_ARRAY_BUFFER、GL_ELEMENT_ARRAY_BUFFER和GL_STATIC_DRAW。从两篇StackOverflow的帖子里得知，GL/gl.h（包括MinGW提供的）只提供了OpenGL 1.1（还是1.2？）的接口声明。不仅如此，通过nm工具，发现libopengl32.a和opengl32.dll并未导出glGenBuffers等VBO接口。

http://stackoverflow.com/questions/11951380/lots-of-undeclared-identifiers-related-to-opengl-using-glfw

http://stackoverflow.com/questions/679113/trouble-porting-opengl-app-to-windows

　　于是我只好参考SDL的源代码，写出了可以工作的glDrawElements版Helloworld：

#include <windows.h>
#include <GL/gl.h>
#include <GL/glu.h>

#ifdef USE_VBO  // copied from SDL_opengl.h
#define GL_ARRAY_BUFFER                   0x8892
#define GL_ELEMENT_ARRAY_BUFFER           0x8893
#define GL_STATIC_DRAW                    0x88E4
#endif

HDC hDC; //device context
HGLRC hglrc; //rendering context

#ifdef USE_VBO
typedef ptrdiff_t GLsizeiptr;

//void * (WINAPI *wglGetProcAddress)(const char *proc);  // May need this for VC++
GLAPI void (APIENTRY *glGenBuffers)(GLsizei, GLuint *);
GLAPI void (APIENTRY *glBindBuffer)(GLenum, GLuint);
GLAPI void (APIENTRY *glBufferData)(GLenum, GLsizeiptr, const GLvoid *, GLenum);
GLAPI void (APIENTRY *glDeleteBuffers)(GLsizei, const GLuint *);

GLfloat transVerts[] =
{
    1.0f, 1.0f, 0.0f,
    -1.0f, 1.0f, 0.0f,
    -1.0f,-1.0f, 0.0f,
    1.0f, -1.0f, 0.0f
};
GLuint vboTransVerts;
GLushort indices[] =
{
    0, 1, 2,
    3, 4, 5,
    6, 7, 8,
    9, 10, 11
};
GLuint vboIndices;
#endif

void SetupPixels(HDC hDC)
{
    int pixelFormat;
    PIXELFORMATDESCRIPTOR pfd;
    pfd.nSize = sizeof(PIXELFORMATDESCRIPTOR);
    pfd.dwFlags = PFD_SUPPORT_OPENGL | PFD_DRAW_TO_WINDOW | PFD_DOUBLEBUFFER;
    pfd.nVersion = 1;
    pfd.iPixelType = PFD_TYPE_RGBA;
    pfd.cColorBits = 32;
    pfd.cDepthBits = 24;
    pixelFormat = ChoosePixelFormat(hDC, &pfd);
    if(!SetPixelFormat(hDC, pixelFormat, &pfd))
    {
         MessageBox(NULL,"Error setting up Pixel Format","ERROR",MB_OK);
         PostQuitMessage(0);
    }
}

void Resize(int width, int height)
{
    glViewport(0,0,(GLsizei)width,(GLsizei)height);
    glMatrixMode(GL_PROJECTION);
    glLoadIdentity();
    gluPerspective(45.0f,(GLfloat)width/(GLfloat)height,1.0f,1000.0f);
    glMatrixMode(GL_MODELVIEW);
    glLoadIdentity();
}

#ifdef USE_VBO
void InitVBO()
{
#if 0  // May need the following code for VC++
    HMODULE hModule = LoadLibrary("opengl32.dll");
    if (hModule == NULL)
    {
        MessageBox(NULL, "Error: Unable to load opengl32.dll", "ERROR", MB_OK);
        exit(-1);
    }
    wglGetProcAddress = (void * (WINAPI *)(const char *)) GetProcAddress(hModule, "wglGetProcAddress");
#endif
    glGenBuffers = (GLAPI void (APIENTRY *)(GLsizei, GLuint *)) wglGetProcAddress("glGenBuffers");
    glBindBuffer = (GLAPI void (APIENTRY *)(GLenum, GLuint)) wglGetProcAddress("glBindBuffer");
    glBufferData = (GLAPI void (APIENTRY *)(GLenum, GLsizeiptr, const GLvoid *, GLenum)) wglGetProcAddress("glBufferData");
    glDeleteBuffers = (GLAPI void (APIENTRY *)(GLsizei, const GLuint *)) wglGetProcAddress("glDeleteBuffers");
    glGenBuffers(1, &vboTransVerts);
    glBindBuffer(GL_ARRAY_BUFFER, vboTransVerts);
    glBufferData(GL_ARRAY_BUFFER, sizeof(transVerts), transVerts, GL_STATIC_DRAW);
    glBindBuffer(GL_ARRAY_BUFFER, 0);
    glGenBuffers(1, &vboIndices);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vboIndices);
    glBufferData(GL_ELEMENT_ARRAY_BUFFER, sizeof(indices), indices, GL_STATIC_DRAW);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0);
}
#endif

void Render()
{
    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
    glLoadIdentity();
    glTranslatef(0.0f,0.0f,-4.0f);
    glColor3f(0.0f,0.0f,1.0f);
#ifdef USE_VBO
    glEnableClientState(GL_VERTEX_ARRAY);
    glBindBuffer(GL_ARRAY_BUFFER, vboTransVerts);
    glVertexPointer(3, GL_FLOAT, 0, 0);
    glBindBuffer(GL_ARRAY_BUFFER, 0);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vboIndices);
    glDrawElements(GL_POLYGON, sizeof(transVerts) / sizeof(transVerts[0]), GL_UNSIGNED_SHORT, 0);
    //glDrawElements(GL_POLYGON, sizeof(transVerts) / sizeof(transVerts[0]), GL_UNSIGNED_SHORT, transIndices);  // equivalent as above
    //glDrawArrays(GL_POLYGON, 0, sizeof(transVerts) / sizeof(transVerts[0]));  // equivalent as above
#else
    glBegin(GL_POLYGON);
    glVertex3f(1.0f,1.0f,0.0f);
    glVertex3f(-1.0f,1.0f,0.0f);
    glVertex3f(-1.0f,-1.0f,0.0f);
    glVertex3f(1.0f,-1.0f,0.0f);
    glEnd();
#endif  // USE_VBO
}

LRESULT CALLBACK WinProc(HWND hWnd,
                         UINT msg,
                         WPARAM wParam,
                         LPARAM lParam)
{
    int w,h;
    switch(msg)
    {
    case WM_CREATE:
        hDC = GetDC(hWnd);
        SetupPixels(hDC);
        hglrc = wglCreateContext(hDC);
        wglMakeCurrent(hDC, hglrc);
#ifdef USE_VBO
        InitVBO();
#endif
        break;
    case WM_DESTROY:
#ifdef USE_VBO
        glDeleteBuffers(1, &vboTransVerts);
        glDeleteBuffers(1, &vboIndices);
#endif
        wglMakeCurrent(hDC,NULL);
        wglDeleteContext(hglrc);
        PostQuitMessage(0);
        break;
    case WM_SIZE:
        w = LOWORD(lParam);
        h = HIWORD(lParam);
        Resize(w,h);
        break;
    default: break;
    }
    return DefWindowProc(hWnd,msg,wParam,lParam);
}
int WINAPI WinMain(HINSTANCE hInstance,
                   HINSTANCE hPrevInstance,
                   LPSTR lpCmdLine,
                   int nShowCmd)
{
    HWND hWnd;
    WNDCLASSEX wcex;
    MSG msg;
    wcex.cbSize = sizeof(WNDCLASSEX);
    wcex.style = CS_HREDRAW | CS_VREDRAW;
    wcex.lpfnWndProc = WinProc;
    wcex.cbClsExtra = 0;
    wcex.cbWndExtra = 0;
    wcex.hInstance = hInstance;
    wcex.hIcon = LoadIcon(NULL,IDI_APPLICATION);
    wcex.hCursor = LoadCursor(NULL,IDC_ARROW);
    wcex.hbrBackground = (HBRUSH) GetStockObject(GRAY_BRUSH);
    wcex.lpszMenuName = NULL;
    wcex.lpszClassName = "WinClass";
    wcex.hIconSm = NULL;
    RegisterClassEx(&wcex);
    hWnd = CreateWindow("WinClass","My Window",
        WS_OVERLAPPEDWINDOW,0,0,400,400,NULL,NULL,
        hInstance,NULL);
    if(hWnd == NULL)
    {
        MessageBox(NULL,"Error: Unable to create Window","ERROR",MB_OK);
        return -1;
    }
    ShowWindow(hWnd,nShowCmd);
    UpdateWindow(hWnd);
    hDC = GetDC(hWnd);
    glClearColor(0.0f,0.0f,0.0f,0.0f);
    while(1)
    {
        Render();
        SwapBuffers(hDC);
        if(PeekMessage(&msg,NULL,0,0,PM_REMOVE))
        {
            if(msg.message == WM_QUIT) break;
            TranslateMessage(&msg);
            DispatchMessage(&msg);
        }
    }
    return 0;
}

　　MinGW上的编译命令还是不变：gcc -o window.exe window.c -DUSE_VBO -mwindows -lopengl32 -lglu32

　　这下知道glDrawElements怎么用了。再回到HomeworldSDL的代码中，我发现程序没有调用glBufferData开辟显存空间就调用了glDrawElements。想必这就是崩溃的原因。于是我将我的window.c中的glBufferData调用删除，再运行程序，果然就崩溃了，并且崩溃处的汇编代码和之前查看的基本一模一样。

　　这虫子算是确确实实地捉到了。

　　说实话，作为完全没接触过OpenGL的新手，我到处搜资料，花了好些时间才掌握了glDrawElements的用法。虽然Unity和CUDA的经验让我在相关概念的理解上没什么大碍，但我花了一天左右时间才真正能用这函数写一个OpenGL小程序。这是因为还有好些其它OpenGL的函数要了解；而且到最后一步还发现，为了像崩溃的代码一样，在这函数中使用OpenGL的顶点缓存对象（vertex buffer object，简称vbo），OpenGL的初始化会麻烦很多。正如《关于游戏开发，学校没有教给我的十件事》一文写道：“了解和理解是不一样的……当你坐一个项目时，你可能会想‘我知道怎么做’。然而，除非你之前做过，否则你仅仅是有怎么做的想法。‘我知道’和‘我有一个如何做的想法’是不同的，它们的区别可能会让你头疼好几个小时。”想必这是每个做过实际项目的程序员的感受。