Agner`s CPU blog

With regards to static vs dynamic linking it is probably worth considering the following note by Ulrich Drepper

There are significant performance costs to DLLs and SOs which I want to avoid. Each DLL/SO has its own memory pages for code and data. You will be wasting a whole 4kB page even if the function uses only a few hundred bytes. The data section of a DLL/SO cannot be shared. The code section can be shared in most cases, but it is not shared if it contains relocated cross-references. The amount of memory that you are saving by sharing functions between multiple processes is very small compared to the amount of memory that you are wasting by loading a lot of other functions that your program is not using.

The dynamic libraries are scattered around in memory which means poor caching. Many functions are placed at the start of a memory page. This means that cache line sets with low numbers are used disproportionately more, or you need a more complicated cache design to distribute the load on cache line sets more evenly.

If a DLL contains many functions and some of these functions - not the ones you are using - link to other DLLs, then you will load these other DLLs unnecessarily, or suffer the unpredictable response time of lazy loading.

Unix-style shared objects have further performance losses due to the requirement of the so-called position-independent code, which actually involves something completely different, namely symbol interposition. 32-bit x86 shared objects have an additional performance problem because the instruction set has no position-independent addressing mode.