Stage 2: a full compiler is produced by the bootstrap compiler.
Stage 3: a full compiler is produced by the stage 2 full compiler.
The full compiler is built twice in order to compare the outputs of the two stages. If they are different, either the bootstrap or the full compiler contains a bug.
Advantages
Bootstrapping a compiler has the following advantages:
it is a non-trivial test of the language being compiled, and as such is a form of dogfooding.
compiler developers and bug reporters only need to know the language being compiled.
compiler development can be performed in the higher-level language being compiled.
improvements to the compiler's back-end improve not only general-purpose programs but also the compiler itself.
it is a comprehensive consistency check as it should be able to reproduce its own object code.
Note that some of these points assume that the language runtime is also written in the same language.
Methods
If one needs to compile a compiler for language X written in language X, there is the issue of how the first compiler can be compiled. The different methods that are used in practice include:
Implementing an interpreter or compiler for language X in language Y. Niklaus Wirth reported that he wrote the first Pascal compiler in Fortran.
Another interpreter or compiler for X has already been written in another language Y; this is how Scheme is often bootstrapped.
Earlier versions of the compiler were written in a subset of X for which there existed some other compiler; this is how some supersets of Java, Haskell, and the initial Free Pascal compiler are bootstrapped.
A compiler supporting non-standard language extensions or optional language features can be written without using those extensions and features, to enable it being compiled with another compiler supporting the same base language but a different set of extensions and features. The main parts of the C++ compilerclang were written in a subset of C++ that can be compiled by both g++ and Microsoft Visual C++. Advanced features are written with some GCC extensions.
The compiler for X is cross compiled from another architecture where there exists a compiler for X; this is how compilers for C are usually ported to other platforms. Also this is the method used for Free Pascal after the initial bootstrap.
Writing the compiler in X; then hand-compiling it from source and running that on the code to get an optimized compiler. Donald Knuth used this for his WEB literate programming system.
Methods for distributing compilers in source code include providing a portable bytecode version of the compiler, so as to bootstrap the process of compiling the compiler with itself. The T-diagram is a notation used to explain these compiler bootstrap techniques. In some cases, the most convenient way to get a complicated compiler running on a system that has little or no software on it involves a series of ever more sophisticated assemblers and compilers.
History
Assemblers were the first language tools to bootstrap themselves. The first high-level language to provide such a bootstrap was NELIAC in 1958. The first widely used languages to do so were Burroughs B5000 Algol in 1961 and LISP in 1962. Hart and Levin wrote a LISP compiler in LISP at MIT in 1962, testing it inside an existing LISP interpreter. Once they had improved the compiler to the point where it could compile its own source code, it was self-hosting. This technique is only possible when an interpreter already exists for the very same language that is to be compiled. It borrows directly from the notion of running a program on itself as input, which is also used in various proofs in theoretical computer science, such as the proof that the halting problem is undecidable.
Current efforts
Due to security concerns regarding the Trusting Trust Attack and various attacks against binary trustworthiness, multiple projects are working to reduce the effort for not only bootstrapping from source but also allowing everyone to verify that source and executable correspond. These include the Bootstrappable builds project and the Reproducible builds project.