BCPL was designed so that small and simple compilers could be written for it; reputedly some compilers could be run in 16 kilobytes. Further, the original compiler, itself written in BCPL, was easily portable. BCPL was thus a popular choice for bootstrapping a system. A major reason for the compiler’s portability lay in its structure. It was split into two parts: the front end parsed the source and generated O-code, an intermediate language. The back end took the O-code and translated it into the machine code for the target machine. Only 1⁄5 of the compiler’s code needed to be rewritten to support a new machine, a task that usually took between 2 and 5 man-months. Soon afterwards this structure became fairly common practice, but the Richards BCPL compiler was the first to define a virtual machine for this purpose.
The language is unusual in having only one data type: a word, a fixed number of bits, usually chosen to align with the architecture’s machine word and of adequate capacity to represent any valid storage address. For many machines of the time, this data type was a 16-bit word. This choice later proved to be a significant problem when BCPL was used on machines in which the smallest addressable item was not a word but a byte or on machines with larger word sizes such as 32-bit or 64-bit.
The interpretation of any value was determined by the operators used to process the values. (For example, “+” added two values together, treating them as integers; “!” indirected through a value, effectively treating it as a pointer.) In order for this to work, the implementation provided no type checking. Hungarian notation was developed to help programmers avoid inadvertent type errors.
The mismatch between BCPL’s word orientation and byte-oriented hardware was addressed in several ways. One was by providing standard library routines for packing and unpacking words into byte strings. Later, two language features were added: the bit-field selection operator and the infix byte indirection operator (denoted by “%”).
All data shared between different compilation units comprises scalars and pointers to vectors stored in a pre-arranged place in the global vector. Thus the header files (files included during compilation using the “GET” directive) become the primary means of synchronizing global data between compilation units, containing “GLOBAL” directives that present lists of symbolic names, each paired with a number that associates the name with the corresponding numerically addressed word in the global vector. As well as variables, the global vector contains bindings for external procedures. This makes dynamic loading of compilation units very simple to achieve. Instead of relying on the link loader of the underlying implementation, effectively BCPL gives the programmer control of the linking process.
The global vector also made it very simple to replace or augment standard library routines. A program could save the pointer from the global vector to the original routine and replace it with a pointer to an alternative version. The alternative might call the original as part of its processing. This could be used as a quick ad hoc debugging aid.
BCPL was the first brace programming language and the braces survived the syntactical changes and have become a common means of denoting program source code statements. In practice, on limited keyboards of the day, source programs often used the sequences $( and $) instead of the symbols { and }. The single-line // comments of BCPL, which were not adopted by C, reappeared in C++ and later in C99.