Language matters

Universal and existential quantification: see Section and Section respectively. Universal quantification was in in GHC 2.10 and later, but the syntax has changed: it's now forall a. ... instead of All a => .... Multi-paramter type classes. We have relaxed some of the rules since GHC 3.00. In particular -fallow-overlapping-instances allows overlapping instances -fallow-undecidable-instances allows you to write instance contexts that constrain non-type-variables Full details in Section . New Libraries

Documentation in . Dynamic Dynamic types. Exceptions The library Exception in -syslib exts provide an interface for exception handling. Calls to error, pattern matching failures and a bunch of other exception can be caught. Compiler internals

The intermediate language used by GHC has had a radical overhaul. The new Core language (coreSyn/CoreSyn.lhs) is much smaller and more uniform. The main transformation engine (the "simplifier") has been totally rewritten. The compiler is several thousand lines shorter as a result. It's also very much easier to understand and modify. (It's true. Honest!) Run time system

There is a completely new runtime system, aimed at integration with Hugs. Tons of cruft removed. Some changes to code generation; things typically run a bit faster as a result. An overview of the new RTS is available: . Storage Manager/Garbage Collector

The new storage manager features a dynamically resizing heap, so you won't need those pesky -H options anymore. In fact, the -H option is now ignored for compatibility with older versions. Stacks are now also expandable, and the -K option now specifies a 1M. The garbage collector now collects CAFs, so there's no more space leaks associated with these. If you think you have a CAF-related space leak, we'd like to hear about it. The storage manager current only has a two-space garbage collector, which will be slower than 3.02's generational collector when the amount of live data is large. A new generational collector is high on our priority list. For the other new tweakables, see Section . Profiling

There is a new profiler based on -prof -auto-all, then run it with +RTS -p and see what comes out (in the <prog>.prof file). The feature is still experimental, and the call graph may not be exactly what you were expecting. Also, we only have time/allocation profiling at the moment; heap profiling will follow shortly. Miscellaneous

Concurrent Haskell is now the default. No need to compile/download special libraries. The context switch interval isn't tweakable any more; threads just yield after allocating a fixed amount of heap (currently 4k). If you want a quicker context switch, that won't be possible: this decision was made for efficiency reasons (it reduces the penalty for runnning threaded code to almost zero). We might allow the context switch interval to be increased in the future; but also context switches are pretty fast (faster than before). forkIO now returns a ThreadId, and there's a new operation killThread which will terminate an existing thread. See Section for more details. You need -syslib concurrent to get the Concurrent library. The rewritten IO library from 3.03 has been ported to 4.00. New constructs: foriegn import and foreign export for H/Direct. Supported architectures: all the development has been done on x86(mainly FreeBSD/Linux) so this is the most stable environment. Sparc (Solaris) and x86/Win32 (cygwin32) have also been mildly tested, and an Alpha port is in the works. Everything else will need a little porting effort; we don't have machines here, so any help would be greatly appreciated. Code is faster and smaller, but programs might run slower due to the garbage collector (see "Storage Manager" above). Increasing the minimum allocation area with the -A RTS option can claw back some of this speed. We now use GMP 2.02, and attempt to use an already-installed copy if available. If you have GMP on your system but the configure script doesn't detect it, we'd like to know. seq works better than in 3.xx - you can now seq functions without crashing the program. The Eval class has been removed (i.e. every type is assumed to be in class Eval). This change has been made in anticipation of Haskell 98. The native code generator has had very little testing (it's not used on x86 yet), so Sparcs may have some trouble with it. Try -fvia-C if strange errors occur. The compiler is slightly faster, but sometimes needs more heap. If you have an example where this is the case, please let us know.