import wibbles

Explain the *other* thing AP_UPDs are used for (not just Hugs, but
also stack-freezing for raiseAsync()).

Wibble.

de-stodged

hc-file-bundle: new location of PrimWrappers.hs (nee PrelPrimOpWrappers.hs)

Along the lines of how 'unsafe' is handled, treat 'safe' and 'threadsafe' as special ids too

Urk, code generator assumed reversed info tables
in its implementation of dataToTag# - broke a
compiler built unregisterised rather weirdly
(and only, none of the code in testsuite/ nor
nofib/ showed up this bug.)

This commit assumes that unregisterised==
info_tables_not_next_to_code & calls upon
a trusty old PrimOps.h macro to locate the tag
in the info table (the reasons for doing it
this way is explained in AbsCUtils.lhs
comments).

To help the poor sod who has to debug a break
like this sometime in the future, I'm trying to
come up with a repro case smaller than
ghc/compiler/ -- not yet successful.

Anyway, this concludes the fixes to the
unregisterised bits; I hereby claim that it
is now working again.

re-introduce dataToTagzh macro (unreg builds only)

Add the CPP flags used by an unregisterised build
when compiling the .h generating utilities herein.
As was, info table size was wrongly computed in
the unregisterised case.

Add stg_ut_1_0_unreg_info proto (only)

- new utility return address, stg_ut_1_0_unreg_{info,ret}, which
  returns via Sp[1] (assuming R1 is in Sp[0], which it is in
  the unregisterised, no-REG_R1 case.

  The only 'consumer' of this entry point is the implementation
  of PrimOps.takeMVarzh_fast in an unregisterised build.

- gcdIntegerIntzh_fast: another primop with mismatched return
  convention.
- tidy up PerformTake() for non-REG_R1 case.
- forkzh_fast and myThreadIdzh_fast: I mistakenly had ThreadId#s
  down as something unboxed; they're not, but unpointed. Update
  primops to use the RET_P() macro upon return instead.

Addendum to previous commit message:

We can't tell whether the stg_sel and stg_ap thunks correspond to
static or dynamic thunks, and it's not worth duplicating them all just
so the ticky-ticky stats can be correct.  So I have played it safe
(for my purposes at least!) and counted them all as dynamic.  Yell if
this is a problem (njn?)

Fix bug in ticky-ticky stats (particularly obvious in
nofib/imaginary/queens) that was meaning some thunk entries were not
being counted in ENT_{STATIC,DYN}_THK_ctr.  Specifically, tick the
counter for the stg_sel and stg_ap thunks, as well as non-standard
thunks!

Also add TICK_UPDF_OMITTED into the stg_sel_noupd thunks (there are no
stg_ap_noupd thunks, or else I would have put it there too).

Also, fix a bug in the sense of the test for update squeezing when
printing (or not) ENT_PERM_IND: we don't print out ENT_PERM_IND unless
update squeezing is turned off, because it may be invalid (see note).
If someone needs this counter, please just remove the test entirely
and always print ENT_PERM_IND.

Minor bug in float-out

remove redundant check for size_t

Generalise a couple of regexps in the -monly-[23]-regs patching code.
Should fix problems with gcc 3.0.4.

stg_gc_unpt_r1_ret: if R1 isn't mapped to a real register,
keep it on the stack.

I confess to 'pushing symbols' here a bit, as the exact
nature (and function) of stg_gc_unpt_r1_info isn't clear
to me (documentation isn't abundant, for one.) However,
it will now behave in a manner consistent with what
the code generator emits in the unregisterised case.

Un-rot nofib-analyse, so it compiles with ghc-5.02.2.

* add a regexp for GNU time 1.7

* import Html hiding ((!)) now hides Html.! as well as !, so add an
  import qualified Html ((!)) as well.

* compile Printf with -fvia-c to work around bug in foreign import of
  snprintf (was always printing "0.00" no matter what input given).

make sure stdout is flushed; unused arg wibble

myThreadIdzh_fast: used bogus return convention

Huh? Surely stg_block_{take,put}mvar ought to be pushing info pointers and not direct return addresses

---------------------------------
	Fix a rather obscure bug in tcGen
	---------------------------------

This bug concerns deciding when a type variable "escapes",
and hence we can't generalise it.  Our new subsumption-checking
machinery for higher-ranked types requires a slightly
more general approach than I had before.  The main excitement
is in TcUnify.checkSigTyVars and its friends.

As usual, I moved functions around and cleaned things up a bit;
hence the multi-module commit.

module names are allowed to contain '.' these days.

- gcdIntzh_fast, cmpIntegerIntzh_fast, cmpIntegerzh_fast,
  integer2Intzh_fast, and integer2Wordzh_fast all used the
  unboxed tuple return convention, which isn't quite right
  since they're just returning unboxed values. Fixed.

- similarly for forkzh_fast, it was using an unboxed value
  return convention, ought to be doing an unboxed tuple
  return.


Both of these broke the unregisterised build -- it still not
sane, more debugging to be done.

typo

Escape space characters in filenames in the Makefile generated by ghc -M.

Escape space characters in filenames in the Makefile generated from ghc -M.

comments

When searching for files with multiple extensions in multiple paths,
search each path for all the possible files, rather than the other way
around.  Otherwise, a .hs file later in the path may override a .lhs
file earlier in the path, and similar strange things happen with
mkdependHS.

Reported-by: Duncan Coutts <duncan.coutts@worcester.oxford.ac.uk>

MERGE TO STABLE

Add notes about OS-thread-safety in f-x-{static,dynamic}, or more
precisely, the current lack thereof.

Complete changes to f-x-dynamic.  This brings it into line with what
is described in ghc/docs/comm/the-beast/fexport.html.

Add description of how f-x-{static,dynamic} are implemented.

add libraries to BinDistDirs

Change a couple of erroneous <code> to </code>.

Expand on the behaviour of fromIntegral a bit.

stg_gc_unbx_r1: fixed what was surely a typo in the 06/07/01
commit which instead of pushing an info pointer, started
pushing the entry code (stg_gc_unbx_r1_ret).

Unbreaks the mini-interpreter.

comment wibble

Fix sparc borkage.

Mention that you can say 'make way=foo' to build a particular way only
in the current directory.

Clean up some markup, and fix a markup error at the same time.