Drealmer's Tumblr

Whole Program Optimization

Not long ago, I was compiling some library with Visual C++ 2008 and I noticed that the release version was noticeably larger than the debug version. The culprit was “link time code generation” (LCTG), the mechanism behind the “whole program optimization” option of Visual C++. But what is it doing exactly?

When you build a project, Visual C++ usually goes through three stages:

1. The frontend, which parses the source code and builds a graph-like structure called the “intermediate representation” (IR).
2. The backend, which compiles the IR into assembly code.
3. The linker, which glues together all the compiled parts of the program.

Usually, a library is nothing more than a collection of compiled chunks, somewhere between steps 2 and 3. But LCTG allows libraries to store IR instead of compiled code. The advantage in keeping the structure is that cross-unit optimizations are still possible at link time (like function inlining).

This page explains the mechanism inside-out: Under The Hood - Link-time Code Generation (link)