Changes against the original release
The Aldor compiler was written a long time ago, when compilers were stupid and C was unportable. Nowadays, many operating system interfaces have been unified and C compilers have gained massive improvements. Computers are also much faster than in the old days.
This project aims to modernise and clean up the Aldor compiler. The following is a non-exhaustive list of changes. For a full list. see the git log.
- Removed garbage collector
The compiler itself no longer uses garbage collection. A precise GC is planned for Aldor code, but for now, the option to use Boehm's GC exists.
- Replaced macros with functions
The code is fast enough so we don't need macros everywhere. Many macros have been replaced with out-of-line functions. That way, profiling and error checking gets easier.
- Removed untestable OS support
The original compiler supports operating systems that we cannot test. Such systems include DOS, VMS, CMS and MacOS 7. These have been removed. In addition, the operating system dependent layer has been reduced and certain functions that can be implemented in pure C have been implemented thusly.
- Split up the compiler into many small libraries
The "generic" layer containing utilities and non-specific data structures (list, hash table, tree, buffer, ...) has been split into 30 smaller libraries. This is an ongoing effort, because proper splitting is not entirely possible in struct and phases. These layers have several issues, including the fact that struct depends on phases.
- Removed edit levels
All edit levels have been removed, one at a time. This history has been preserved in the git log. One edit level was kept and turned into a
configureoption (the new hash code algorithm). - Privatised data structures
Data hiding principles have been applied in several places (Foam, AbSyn, Buffer, ...). All accesses now go through get and set functions.
- Runtime type checking
Aldor uses two major polymorphic data types: AbSyn and Foam. Casting between instance types of this data type is undefined in the general case. There are some valid casts, though. Runtime checks have been implemented to ensure that the casts are valid. A few potential bugs have been fixed this way.
- Fixed bug in constant folding
There were some bugs in constant folding that caused 2 + 4 to equal 4, because argument 1 was added to argument 1.
- Removed SSA
The original compiler contains an unfinished implementation of a Foam to SSA pass, which doesn't even compile. There is enough existing code to work with, so this code was removed.
- Added unit tests
Unit tests and regression tests have been automated. We developed a continuous integration system for the Aldor compiler, so we can validate changes on several different platforms.
- Use the GNU build system
Instead of handwritten makefiles, we now use automake, autoconf and libtool.
- Generate more code
We want to move towards writing less code and achieving more. Domain specific languages were developed to aid in code generation, reducing redundancy and error-proneness.
- Stub out C library calls
No part of the compiler calls the C library directly, anymore (except the parts that do
setjmp/longjmp). All calls go through a C stub library, which can do additional argument checking and modification. It can supplement the C library with non-existent functions and reduce the dependency on a fulllibcimplementation, thereby increasing portability to embedded platforms. - Reduced knowledge
The files
cconfig.handcompiler.hcontained a lot of compiler specific knowledge. This has mostly been removed and we now use behaviour checks provided by autoconf, instead. Also, the dependency on endianness has mostly been lifted. We now use a possibly slower but more portable way. The performance hit is not noticable.