Announce 4.06

urk, Solaris sh doesn't understand [^0-9] in patterns, it seems.

Add missing final paragraph of explaination about x86 FP trickery.

Minor improvements to x86 FP fake-to-real insn translation.

genCCall for x86, as supplied, used PUSH et al to move args onto the C
stack ready for the call.  Reasonable as this seems, it causes a
problem with spill code, since the spiller spills relative to %esp and
assumes that %esp doesn't move.  If the args of a ccall involved any
spilled values, the resulting code would be wrong.

The One True Way is to do it like a RISC: move args to the stack
without adjusting %esp for each argument, then adjust it all at once
immediately prior to the call insn and un-adjust it immediately
afterwards.  genCCall now does this.  In general, push/pop and other
C-stack effecting operations should not be generated for the same
reason.

Add -optCrts-M80m for older compilers.  Sigh.

Added intToWord to PrelAddr.
Use it instead of int2Word#-hacks.

Start a NOTES file, recording known but un-fixed nativeGen bugs.

Fix syntax errors in #ifdef'd Alpha/Sparc bits.

Insert large commit message re x86 FP rehash as a comment.

ARR_HDR_SIZE --> ARR_WORDS_HDR_SIZE, and derived quantities in
Constants.h, Constants.lhs et al are similarly renamed.

new constant ARR_PTRS_HDR_SIZE, with corresponding derivatives.

Changed default paper size for SGML output to A4 (%paper-type%).

Major reworking of the x86 floating point code generation.

Intel, in their infinite wisdom, selected a stack model for floating
point registers on x86.  That might have made sense back in 1979 --
nowadays we can see it for the nonsense it really is.  A stack model
fits poorly with the existing nativeGen infrastructure, which assumes
flat integer and FP register sets.  Prior to this commit, nativeGen
could not generate correct x86 FP code -- to do so would have meant
somehow working the register-stack paradigm into the register
allocator and spiller, which sounds very difficult.

We have decided to cheat, and go for a simple fix which requires no
infrastructure modifications, at the expense of generating ropey but
correct FP code.  All notions of the x86 FP stack and its insns have
been removed.  Instead, we pretend (to the instruction selector and
register allocator) that x86 has six floating point registers, %fake0
.. %fake5, which can be used in the usual flat manner.  We further
claim that x86 has floating point instructions very similar to SPARC
and Alpha, that is, a simple 3-operand register-register arrangement.
Code generation and register allocation proceed on this basis.

When we come to print out the final assembly, our convenient fiction
is converted to dismal reality.  Each fake instruction is
independently converted to a series of real x86 instructions.
%fake0 .. %fake5 are mapped to %st(0) .. %st(5).  To do reg-reg
arithmetic operations, the two operands are pushed onto the top of the
FP stack, the operation done, and the result copied back into the
relevant register.  There are only six %fake registers because 2 are
needed for the translation, and x86 has 8 in total.

The translation is inefficient but is simple and it works.  A cleverer
translation would handle a sequence of insns, simulating the FP stack
contents, would not impose a fixed mapping from %fake to %st regs, and
hopefully could avoid most of the redundant reg-reg moves of the
current translation.

Added table example.

Added autoconf magic for size/alignment of some more C types

Changed double quotes to “ and ”. Improvements to Windows
installation instructions.

Update the Hall of Fame.

Stylesheet used for processing SGML. Initially it's identical to the Cygnus
DocBook Tools stylesheet, with %section-autonumber% set to true for HTML, so
that all sections are numbered.

Set SRC_SGML2XXX_OPTS variables rather than SGML2XXX_OPTS

Wrote 4.06 release notes and updated version numbers in the rest of the user
guide.

Added SGMLSTYLESHEET variable and set SGML2XXX_OPTS to -d $(SGMLSTYLESHEET)

Extra install stuff

Increase the heap size for Parser.hs to 80M (for 4.04).

GHC now uses the "Hugs" split function, which is believed to have
better behaviour: it is not unsafe, is deterministic, and works
better with QuickCheck, a major client.

Rational for the Record, quoted from Mark Jones mail on the Hugs list:

[Mark Jones]
A couple of months ago, John Hughes sent me mail about a problem that
he had uncovered with the implementation of the Random library in Hugs.
He had been using the "split" function in an attempt to generate a
stream of random number generators, each of which he hoped would be
different from the others.  But instead he found that he actually
ended with many different copies of the *same* random number generator.
A disappointing, and frustratingly non-random result.

If you don't happen to recall, split is a member of the RandomGen class,
with type RandomGen g => g -> (g,g); it takes a single random number
generator as its argument, and returns a pair of two new generators as
its result.  The only thing that the specification requires is that the
two generators returned are (a) distinct and (b) `independently robust'
from a statistical point of view.  To the best of my knowledge, the
implementation in Hugs meets this modest specification.  Sadly, assuming
only this specification, you cannot write the function that John was
looking for and be sure that it will generate more than two different
generators.

For example, the specification allows even the following trivial
implementation for split:  split _ = (g1, g2), where g1 and g2 are some
arbitrary but constant pair of distinct, and independently robust
generators.  With this implementation, you can split as often as you
want and you'll never get more that two generators.

Hugs and GHC (as far as I can tell) both use definitions of the form:

   split g = (g, f g)

for some function f.  (My understanding of the code in GHC is that it
uses an unsafe function for f, breaking referential transparency; I hope
the optimizer knows about this.)  Note that this definition returns the
argument as a result; the specification doesn't prohibit that; all it
requires is that the two results returned be distinct.  But if you try
to generate a list of n different generators using:

   take n (iterate (fst . split) g)

then you will be sorely disappointed; you might as well have written
replicate n g.  (On the other hand, if you were lucky enough to have
used (snd . split), instead of (fst . split), then you wouldn't have
noticed the problem ...)

I know very little about the mathematics or pragmatics of random
number generators, so I'm not sure that I know how to fix this
problem.  However, starting from this position of ignorance, I have
hacked up a new version of "split" for the standard "StdGen" that
will appear in the next release of Hugs (real soon now!).  Judging
from the tests that I've tried so far, it seems to work much
better than the old version.  That said:

 - Take care if you use Random.split in your programs, because it
   may not do what you expect.

 - There should probably be an errata about this for the Haskell 98
   library report ... if somebody can figure out what it should say.

 - If you use Hugs, be aware that the implementation of Random.split
   was hacked up by someone who has no way of justifying that
   implementation, beyond some simple experiments.

 - If you know something about the mechanics of random number
   generators, here's an area where Haskell specifications and
   implementations could benefit from your knowledge!

All the best,
Mark
[end quote]

Fix bug in async exception handling: the target TSO may have been
relocated as a result of a stack overflow.

Introduce a new StgTSOWhatNext value "ThreadRelocated", which
indicates that this TSO has moved, and the new location is in the link
field.  The garbage collector shorts these out just like indirections.

We have to check for relocated TSOs in killThread# (and any other
primops which take a ThreadId# as an argument - there aren't any at
present).

- FreeBSD 3+ now known as i386-unknown-freebsd
- add alpha-unknown-{linux,freebsd}

freebsd3 ==> freebsd

freebsd3 ==> freebsd

- FreeBSD 3+ is now known as i386-unknown-freebsd (to accomodate
  forthcoming FreeBSD 4.X

-  add alpha-unknown-{linux,freebsd}

Remove potentially confusing references to concurrent bundles.

Fairly grotesque hacks to get the HTML docs, which are now generated
into a subdirectory by db2html, into a binary dist.

remove ghc/docs/libraries from the bindist, add hslibs/doc

trivialCode (x86), when fst arg is immediate, assumed you could reverse
the order of operands, but not true for eg subtract.  Fixed.

Make this test more comprehensive.  We now test

	+,-,div,mod,quot,rem,gcd,lcm

over a range of legal operands, for both Int and Integer.

- add some more test cases, now shows up the minInt bugs in Integer
division.

The minInt saga continues:

 - fix several Integer division-type functions for the
   case when the dividend == S# (-2147483648#).

Pointed out by: Marc Van Dongen <dongen@cs.ucc.ie>

amodeToStix: correctly compute offset for CHARLIKE_closure-s.

Add comment about code generation for debug tracing.

MachCode.stmt2Instrs, StFunBegin, x86 case only: for debugging,
generate trace code to print the name of each labelled code block.

genCCall for x86 assumed that all args were 4 bytes long :-(.
Now works with doubles too.