GitLab

inputReady can be an unsafe foreign import.

Move towards having the IO implementation be plat-independent
at the Haskell source code level.

- export hGetcBuffered, which is used by the compiler (StringBuffer)
- don't export commitBuffer.

More small performance improvements

Subvert GHC's full-laziness optimisation by explicitly lambda-lifting
a local lambda expression to the top level, and exporting it to ensure
it doesn't get inlined.  Without this hack, full-laziness will float
out several subexpressions, which turns out to be a pessimisation in
this case.

This is worth about 20% in hPutStr performance, so we now beat the old
I/O library on 'cat' ;-)

I'd like to MERGE TO STABLE this, but 5.02 is imminent so it might
have to wait until 5.02.1.

Fix a couple more cut-n-pastos in the line-buffered version of
hPutStr.  Also fill in the export list while I'm here.

MERGE TO STABLE

oops: hPutStr wasn't flushing a line-buffered handle properly at the
end of a line.

MERGE TO STABLE

Globally-uniquify the names of some _wrap functions so that the
previously-undetected global namespace clashes don't cause an
assertion failure in Hash.c when compiled -DDEBUG when starting GHCi
with various hslibs packages.

don't include <string.h> here, HsStd.h is the right place.

#include <string.h> for memcpy prototype.

Fixed assumption that int and HsInt are the same.

Don't use relative #include paths, as these files will be compiled from
inside cbits (they're now linked into the cbits library). All paths on which
include files may be found must now be given with -I directives.

Fix a bug in hGetContents, namely that it wasn't closing the handle
when the end of file was reached.  Also tried to tidy the code up a
bit while I was here.

Fix paths to HsStd.h.

Make hGetContents work on DuplexHandles, and some small cleanups.

- System.exitWith now raises a (new) exception,
	ExitException ExitCode.

- While I was there I cleaned up System.getArgs and
  System.getProgName, using foriegn label and removing progargs.c

minor performance improvement

I/O library rewrite
-------------------

This commit replaces the old C/Haskell I/O implementation with a new
Haskell-only one using the new FFI & hsc2hs.

main points:

   - lots of code deleted: we're about 3000 lines of C lighter,
     but the amount of Haskell code is about the same.

   - performance is ok: some operations are faster, others are
     slower.  There's still some tuning to do, though.

   - the new library is designed to handle read/write streams
     much better: a read/write stream gets a special kind of
     handle internally called a "DuplexHandle", which actually
     contains two separate handles, one for writing and one for
     reading.  The upshot is that you can do simultaneous reading
     and writing to/from a socket or FIFO without any locking
     problems.  The effect is similar to calling socketToHandle
     twice, except that finalization works properly (creating
     two separate Handles could lead to the socket being closed
     too early when one of the Handles is GC'd).

   - hConnectTo and withHandleFor are gone (no one responded to
     my mail on GHC users, but we can always bring 'em back if
     necessary).

   - I made a half-hearted attempt at keeping the system-specific
     code in one place: see PrelPosix.hsc.

   - I've rearranged the I/O tests and added lots more.
     ghc/tests/lib/IO now contains Haskell 98-only IO tests,
     ghc/test/lib/{IOExts, Directory, Time} now contain tests for
     the relevant libraries.  I haven't quite finished in here yet,
     the IO tests work but the others don't yet.

   - I haven't done anything about Unicode yet, but now we can
     start to discuss what needs doing here.  The new library
     is using MutableByteArrays for its buffers because that
     turned out to be a *lot* easier (and quicker) than malloc'd
     buffers - I hope this won't cause trouble for unicode
     translations though.

WARNING: Windows users refrain from updating until we've had a chance
to fix any issues that arise.

Testing: the basic H98 stuff has been pretty thoroughly tested, but
the new duplex handle stuff is still a little green.