一个不太为人所知的区别是，在具有乐观内存分配的操作系统(如Linux)中，由malloc返回的指针直到程序实际接触它时才得到实际内存的支持。

calloc确实会接触内存(它会在内存上写0)，因此您可以确定操作系统正在用实际的RAM(或swap)支持分配。这也是为什么它比malloc慢的原因(它不仅必须将它归零，操作系统还必须通过交换其他进程来找到合适的内存区域)

例如，请参阅这个SO问题以进一步讨论malloc的行为

Calloc一般是malloc+memset为0

显式使用malloc+memset通常会稍微好一点，特别是当你在做以下事情时:

ptr=malloc(sizeof(Item));
memset(ptr, 0, sizeof(Item));

That is better because sizeof(Item) is know to the compiler at compile time and the compiler will in most cases replace it with the best possible instructions to zero memory. On the other hand if memset is happening in calloc, the parameter size of the allocation is not compiled in in the calloc code and real memset is often called, which would typically contain code to do byte-by-byte fill up until long boundary, than cycle to fill up memory in sizeof(long) chunks and finally byte-by-byte fill up of the remaining space. Even if the allocator is smart enough to call some aligned_memset it will still be a generic loop.

一个值得注意的例外是，当您对一个非常大的内存块(一些power__2kb)执行malloc/calloc时，在这种情况下，可以直接从内核进行分配。由于操作系统内核通常会出于安全原因将它们放弃的所有内存归零，足够聪明的calloc可能只返回内存，而不进行额外的归零。同样，如果你只是分配一些你知道很小的东西，那么在性能方面使用malloc+memset可能会更好。

摘自Georg Hager的博客上的一篇文章，用calloc()进行有趣的基准测试

When allocating memory using calloc(), the amount of memory requested is not allocated right away. Instead, all pages that belong to the memory block are connected to a single page containing all zeroes by some MMU magic (links below). If such pages are only read (which was true for arrays b, c and d in the original version of the benchmark), the data is provided from the single zero page, which – of course – fits into cache. So much for memory-bound loop kernels. If a page gets written to (no matter how), a fault occurs, the “real” page is mapped and the zero page is copied to memory. This is called copy-on-write, a well-known optimization approach (that I even have taught multiple times in my C++ lectures). After that, the zero-read trick does not work any more for that page and this is why performance was so much lower after inserting the – supposedly redundant – init loop.

块数: Malloc()分配请求的单个内存块， Calloc()为请求的内存分配多个块

初始化: Malloc() -不清除和初始化分配的内存。 Calloc() -将分配的内存初始化为0。

速度: Malloc()速度很快。 Calloc()比malloc()慢。

参数和语法: Malloc()接受1个参数:

字节 要分配的字节数

Calloc()有两个参数:

长度 要分配的内存块的数量 字节 在每个内存块上分配的字节数

void *malloc(size_t bytes);         
void *calloc(size_t length, size_t bytes);      

内存分配方式: malloc函数从可用堆中分配所需“大小”的内存。 calloc函数分配的内存大小等于' num *size '。

名称含义: malloc的意思是“内存分配”。 calloc的意思是“连续分配”。

还有一个没有提到的区别:大小限制

void *malloc(size_t size)只能分配到SIZE_MAX。

Void *calloc(size_t nmemb, size_t size);可以分配大约SIZE_MAX*SIZE_MAX。

在许多具有线性寻址的平台中，不经常使用此功能。这样的系统用nmemb * size <= SIZE_MAX限制calloc()。

考虑一种名为disk_sector的512字节类型，代码希望使用大量扇区。在这里，代码最多只能使用SIZE_MAX/sizeof disk_sector扇区。

size_t count = SIZE_MAX/sizeof disk_sector;
disk_sector *p = malloc(count * sizeof *p);

考虑下面允许更大分配的情况。

size_t count = something_in_the_range(SIZE_MAX/sizeof disk_sector + 1, SIZE_MAX)
disk_sector *p = calloc(count, sizeof *p);

现在，这样一个系统能否提供如此大的分配是另一回事。今天大多数人都不会。然而，当SIZE_MAX为65535时，这种情况已经发生了很多年。根据摩尔定律，这种情况将在2030年左右发生，某些内存模型SIZE_MAX == 4294967295，内存池为100 gb。

The documentation makes the calloc look like malloc, which just does zero-initialize the memory; this is not the primary difference! The idea of calloc is to abstract copy-on-write semantics for memory allocation. When you allocate memory with calloc it all maps to same physical page which is initialized to zero. When any of the pages of the allocated memory is written into a physical page is allocated. This is often used to make HUGE hash tables, for example since the parts of hash which are empty aren't backed by any extra memory (pages); they happily point to the single zero-initialized page, which can be even shared between processes.

任何对虚拟地址的写都被映射到一个页，如果该页是零页，则分配另一个物理页，将零页复制到那里，并将控制流返回给客户端进程。这与内存映射文件、虚拟内存等工作方式相同。它使用分页。

下面是一个关于这个主题的优化故事: http://blogs.fau.de/hager/2007/05/08/benchmarking-fun-with-calloc-and-zero-pages/