跟讀練習: Wrote My Own Memory Allocator (and ran DOOM with It) - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

151 句

如果句子過短或過長，請點選 Edit 進行調整。

One day I was outside, doing normal human activities like touching grass and listening to birds.

When the blazing sun gave me a strange idea.

What if I wrote my own malloc?

So as soon as I got home I started building.

And in this video I'll show you how I wrote my own memory allocator in PUC from scratch

and how I used to run a real application.

Before we start programming it's important to understand what malloc actually does.

The standard malloc function is used to allocate memory dynamically.

For example, you can use it to create a dynamic array, precisely known at runtime.

But because this is C, there is always some problem with memory hiding somewhere.

As you probably know, C doesn't have a built-in garbage collector, so managing memory is the programmer's responsibility.

For the allocation, C has a special free function.

It takes a pointer to a previously allocated memory block and marks that block as free.

So the system can reuse it later.

And that means I also need to create my own free function.

First I had to choose the names.

I remembered the name of my channel and decided to call them podalloc and pod3.

But before we continue, let's take a look at how memory allocation actually works.

Every process has its own memory space.

This space is divided into several zones.

At the top we have the stack, which is used for function calls and local variables.

Below it we have the heap, which is used exactly for dynamic user data.

And this is the part that malloc works with.

Below that there are other memory regions used for global variables, program code and other things that are not really important for us right now.

In this layout the stack grows downward, while the heap grows upward.

And it's better to keep these two zones from overlapping, unless you really want to meet a segmentation fold.

Another question is how we can actually use this heap for our needs.

We need some way to request memory from the operating system.

There are a few ways to do this, but for me the best is MMAP.

It works kinda similar to malloc.

You tell it how much memory you need and it's returned a pointer to the beginning of that memory region.

But the important difference is that malloc reserves memory inside an already existing area.

While MMAP asks the iOS for a new memory region of their requested size.

Now we can use this to write a simple version of the allocator.

To free allocated memory we need to know its size.

For that I added a mem block structure with a size field and a data pointer, which points to the beginning of the memory available to the user.

In POT-A-LOG we simply call a map to request some memory from the operating system.

Then at the beginning of the memory region we place our mem block structure.

It works like a header for the allocated block.

Later in POT-3 we use this header with monMap monMap does the opposite of mmap.

It releases the memory back to the OS.

For this to work it needs a pointer to the beginning of the memory region and its size, exactly what we keep in the header.

After this we can finally test how it actually works.

As you can see I allocated some memory, wrote some text into it and printed it.

And works fine!

But of course it can't be that easy.

Right now this is basically just a wrapper around M-map and Moonmap.

The main problem with this approach is time.

I wrote a simple test that allocates and frees memory, and then compares it with the malloc from C.

As you can see, the difference is really big.

My allocator is 435 times slower.

It's even slower than me understanding hints.

Using this in a real program would seriously slow it down.

The reason is M-Map and MoonMap.

Every allocation and every free requires a system call.

And they are pretty slow.

Sometimes they are so slow that you forget why you even call them.

To make our allocator faster, we need to reduce the number of M-Map calls as much as possible.

A common approach is to call a map once to allocate a large chunk of memory.

Usually at least one virtual memory page in your system.

If you're wondering what page I'm talking about, here's a quick reminder of what virtual memory is.

It's a memory abstraction provided by the operating system.

Every process gets its own virtual address space, which is divided into small, fixed-sized parts called pages.

The program works with virtual addresses, and the system translates these addresses to real physical memory in RAM.

Because of this, variables in different programs can even have the same addresses.

But that doesn't mean they are stored in the same place in RAM.

Each address just belongs to its own process virtual memory space.

So, with a new approach we allocate one big chunk of memory

and then filled step by step whenever someone calls for tolog.

When this page fills up, we allocate another one.

As you can see, in this example we have 8 memory allocations, but only 2 mmap calls.

This is one of those cases where the number doesn't matter, but matters how you use it.

To implement this I changed the mem block structure a little bit.

Now each block works like part of doubly linked list.

And I also added a free flag to show whether the block is available for allocation or not.

Also because we now use pages, I added a page structure.

It works like another list, but instead of linking memory blocks, it links different pages together.

This will come in handy later.

After this I change the allocator function.

Now it searches for a free block with enough space.

If it can't find one, we create another memory page.

You may notice that if the user asks for only 100 bytes, we allocate an entire page for it.

We use a much bigger block.

We are just wasting memory.

To deal with this, we split the memory block.

One part is given to the user and the other part remains free for future allocations.

With this new approach I also changed port 3.

If we split large memory blocks during allocation, we need to merge them back during their allocation.

Then the merged block can be reduced later.

And now we only return memory to the system when the whole page becomes free.

So this approach reduces main map calls too.

Well, let's test this with some allocations.

And as you can see it somehow become worse.

Now it is 1823 times slower, which doesn't really look like a better solution.

Without previous changes, it looks more like a waste of time.

But if we look at the number of mMAP calls, it becomes much smaller, so that means the problem is not the system calls.

After a bit of investigation, I found that most of the time is spent searching for a free block.

More precisely, it takes 99.3% of the total time.

On one hand, that's good, because system calls are no longer the main problem.

100

On other hand, now we need some fancy data structure to make it faster.

101

The current version searches through a list of all memory blocks.

102

It has to check each block until it finds a free one.

103

So this takes about 0 and time.

104

Looking at this layout, we can see that free blocks are scattered randomly around the page.

105

At the end of the page, there is also some unused space that has not been allocated yet.

106

What if we move free blocks into a separate list?

107

That's where we can quickly get a free block without searching through all allocated blocks.

108

But there is another problem.

109

We need to find a block with the right size.

110

For example, if we need a block of 100 bytes or more, we may have to skip smaller blocks before finding one that actually fits.

111

This can be fixed too.

112

Instead of using one free list, we create 8 of them.

113

Each list stores memory blocks of a certain size category.

114

When a block is freed, it's added to the appropriate list.

115

The benefit is that this is much faster than the previous approach.

116

Usually programs allocate small chunks of memory, so most allocation will fit into one of those eight categories.

117

Another question is how to do this in code.

118

First I added three list pointers to our structures.

119

Then I added functions for inserting blocks into three lists and removing blocks from them.

120

Now when a new page is created, its first free block is automatically added to the appropriate free list.

121

I also changed how splitting works.

122

Now when we split a block, we remove the original block from the free list, give one part to the user and add the remaining part back to the correct free list.

123

And in pod3 I now use the same insert and remove function to keep all free lists updated.

124

At this point it can already works, but there is always some hidden problem in a low level program.

125

And this time the problem is not even the program itself, it's the machine where the program runs.

126

Right now I can allocate 3 bytes and then 4 bytes.

127

And as you can see the second address is not aligned.

128

On some systems this can cause an error, because some processes don't like unaligned addresses.

129

Also, aligned memory is usually faster to read and write.

130

So, to fix this we need to round up the requested amount of memory.

131

We can do it with simple math, but that would be too slow and I need somehow to justify all the time I spent learning bit operators.

132

This version works much faster.

133

We start by adding this to move the number into the next alignment range.

134

For example, if the number is 12 and the alignment is 8, it becomes 19.

135

When we take that, which is 7, in binary it looks like this, exactly the lowest 3 bits.

136

After that we use the tilde operator to flip the bits and create a mask.

137

Finally we use the end operator to raise the lowest 3 bits.

138

As a result 19 becomes 16, which is perfectly aligned to 8 bytes.

139

Pretty easy, as you can see.

140

So after crawling through the depths of C, we can finally run it.

141

And this version actually works better than all previous ones.

142

Of course it's still not as fast as the original Mylock, but further improvements would be much harder and would probably give us a smaller performance gain.

143

Testing only with benchmarks is boring, so let's see how it works in a real program.

144

For the target I chose Doom.

145

I replaced its original allocation function with my own, added some logs to track memory usage and wrote a small python script to visualize it.

146

After playing for some time, everything worked fine for me, no crashes or anything like that.

147

For me, this was a really interesting project, And I hope you enjoy it too.

148

I recommend you trying to write your own allocator as well.

149

It really helps you understand how memory works and improves your understanding of low-level programming.

150

If you enjoyed this video and want to see more, please subscribe, leave a like and write a comment with your ideas or questions.

151

Thanks for watching!

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

FOR YOU

看YouTube學英語：從記憶體分配器到 Doom 的程式設計故事

這是一個程式設計師突發奇想的真實場景：在戶外感受自然時，被陽光激發靈感，決定自己編寫記憶體分配器，還成功用它運行了經典遊戲 Doom。影片裡不僅有技術細節，還有除錯過程中的笑與淚，是個兼具趣味性和知識性的英語學習素材。通過這樣的內容，你既能瞭解程式設計知識，又能練習英語聽力和口語，可謂一舉兩得。

實用表達與搭配

blazing sun：烈日。比如影片開頭提到“the blazing sun gave me a strange idea”，生動描繪了陽光強烈的場景。
dynamic array：動態陣列。這是程式設計中的常用術語，在描述 malloc 功能時頻繁出現。
segmentation fault：段錯誤。程式設計師常遇到的錯誤型別，影片中用“unless you really want to meet a segmentation fault”形象說明記憶體重疊的後果。
system call：系統呼叫。是作業系統提供的服務介面，影片裡解釋了 mmap 和 munmap 屬於系統呼叫。
doubly linked list：雙向連結串列。資料結構中的概念，用於最佳化記憶體分配器的效能。

你的 shadow speak 挑戰

現在就開啟影片，找到“After this, I changed the allocator function... we create another memory page”這段（約影片中間部分），開始 shadowing 練習。先聽一遍原音，注意說話人的語調、停頓和重音，尤其是“search for a free block”“fill up”等短語的發音。然後跟著錄音逐句模仿，儘量做到語速、節奏一致。重複練習 3-5 次後，用自己的話複述這段內容，檢查是否能流暢表達關鍵資訊。這個任務能有效提高英語發音和口語流暢度，快來試試吧！

通過看YouTube學英語，結合 shadow speak 練習，你會發現英語口語進步其實沒那麼難。下次遇到喜歡的影片，不妨都用這種方法試試，堅持下去必有收穫。

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	is divided · are stored · is given
定語從句 who / which + 從句 — 補充說明人或事物	stack, which is · heap, which is · space, which is

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: Wrote My Own Memory Allocator (and ran DOOM with It) - 通過影片學習英語口語

選擇練習模式

下載應用

推薦影片

Learning to code: How Pavel Durov learned to program | Lex Fridman Podcast Clips

Making RAM at Home

Rust in 100 Seconds

You are wasting memory

Can I remake those SH*TTY mobile games in 1 HOUR? ...again

A neuroscientist’s guide to protecting your brain, in 58 minutes | Lisa Genova: Full Interview

Watch this before you boot up another device

What is clean code? - Uncle Bob

JavaScript Visualized - Event Loop, Web APIs, (Micro)task Queue

World Models explained in 10min..

How to Make Hermes Agent 10x More Powerful (Memory & Skills)

The most interesting "hack" in history...

看YouTube學英語：從記憶體分配器到 Doom 的程式設計故事

實用表達與搭配

你的 shadow speak 挑戰

影片中的語法

什麼是跟讀法？