GitLab

Fix egregious bug in the unregisterised case for RET_N, which was
causing an unregisterised compiler to crash.

Comments and cosmetics about dataTcOccs

Comment case-alternative invariants

forgot one file

These test have all moved over to
fptools/testsuite/tests/ghc-regress/ghci.  The old test suite in
fptools/ghc/tests is now finally dead!

Commit an untested, but hopefully more workable and up-to-date version
of this script.

Another cost-centre-restoring fix.  Restoring the cost centre in an
unboxed-tuple case alternative was using the wrong stack offset.

When GhcUnregisterised=YES, we must tell genapply about it.

Fix a nasty bug in eq_ufVar, which meant that unfoldings were
comparing equal when non-local names in the unfolding had changed,
which could lead to erroneously avoiding recompilation.

This bug could cause bad behaviour along the lines of "my program was
segfaulting, but I did make clean; make and now it works!".

merge to STABLE

Oops, move include of stdlib.h after MachRegs.h, so that it comes
after register variable declarations.

Update the documentation for filenames and related options to reflect
the new story.  Now we talk about hierarchical modules from the
outset, and hopefully the whole description is somewhat clearer than
before.

Check for unknown flags before (rather than after) checking -o and -ohi
flags, avoiding confusing behaviour like

% ghc -bad-option -c x.hs
ghc-6.0: unrecognised flags: -bad-option

% ghc -bad-option -c x.hs -ohi x.hi
ghc-6.0: -ohi can only be used when compiling a single source file
Usage: For basic information, try the `--help' option.

% ghc -bad-option -c x.hs -o x.o
ghc-6.0: can't apply -o to multiple source files
Usage: For basic information, try the `--help' option.

(looks better with scores of flags)

please merge to STABLE

We use NULL in the headers (at least when -prof is on), so we better
#include <stdlib.h>.  We normally get it via HsBase.h anyway, but we
shouldn't rely on that.

Document that changing directories unloads all modules.

:cd now causes all modules to be unloaded.

This is the easy fix for a problem whereby changing directories will
cause GHC to try to demand-load objects from the wrong place, because
it is using relative pathnames.

It's not obvious what the "right" thing to do is if the user changes
directory with :cd and then does a :reload.  Perhaps the right thing
to do is to make the search path "sticky", so that it stays relative
to the original current directory.  This would probably require some
re-engineering of GHC to only use absolute paths internally.

Comment out the warning about filenames not matching module names.

Add support for the new AWAKEN_BQ_CLOSURE macro to the NCG.  Fixes
broken HEAD builds.

More on rules and specialisations

When restoring the cost centre in a let-no-escape, don't free the
stack slot containing it.  We might need the saved cost centre again
for a recursive call to this let-no-escape.

Should fix profiling a bit more.

Unbreak GHCi: Add stg_BLACKHOLE_BQ_info and awakenBlockedQueue to
RTS_SYMBOLS.

To accomodate the Hugs package conversion magic, use separate
variables for OpenGL libraries and GLUT libraries.

Restoring of cost centre in let-no-escape: we need to do this after
binding the args to stack locations, otherwise we end up grabbing the
wrong slot.

Should hopefully fix profiling crashes.  We still don't pay any
attention to the cost centre attached to the let-no-escape binding
itself, which looks wrong, but I don't intend to do anything about
that right now.

When doing update-in-place, there is a (small) chance that the thunk
may have been blackholed and another thread might be waiting on it.
We can't therefore just splat it with the value; we have to check
whether it is a blocking queue and awaken any threads that might be
waiting on it if so.

Should fix the scavenge_mutable_list crash reported recently.  If it
does, it will be merged to STABLE.

We don't need hacky filename comparisons in ppFilesFromSummaries now
that the real bug in SysTools has been fixed.

Revision to the filename policy changes of yesterday.  This fixes the
broken stage2 and library builds.

The new story is this:

  A Haskell source filename should take the form

	<path>/<basename>.<extension>

  where

	<path>      is any directory
	<basename>  is the name of the module, with dots replaced by slashes.
	<extension> is ".hs" or ".lhs".

  given a source filename in this form, GHC will produce object and
  interface files named respectively:

        <hidir>/<basename>.<hisuf>
	<odir>/<basename>.<osuf>

  where

	<hidir>   is the value of the -hidir flag, if one was given, or
		  the value of <path> from the source file otherwise.

	<odir>    the same, for the -odir flag.

        <osuf>	  the object suffix (settable with the -osuf flag)
	<hisuf>   the hi suffix (settable with the -hisuf flag)

For example, if I have a module A.B.C, in source file foo/A/B/C.hs,
then GHC will create foo/A/B/C.hi and foo/A/B/C.o.

If GHC is given a source filename which is *not* of the form
<path>/<basename>.<extension>, then it uses a different strategy.
This happens if the filename does not follow the module name.  In this
case, GHC will set <path> to be the directory in which the source file
resides, and <basename> to the module name with dots replaced by
slashes, and then use the rules above.

For example, if we put module A.B.C in foo/bar/baz.hs, then GHC will
produce foo/bar/A/B/C.o and foo/bar/A/B/C.hi.

oops, remove excess baggage in previous commit

Fix a blatant bug in cleanTempFilesExcept, which was causing
legitimate source files to be deleted.  The previous fixes for this
bug missed the real cause of the problem.

I take full blame for this bug, which has been here since the dawn of
GHCi (at least I traced it back to 5.00).

Filename-related cleanup & fixes
--------------------------------

This commit rationalises some of our filename policies.  The new story
is this:

  When compiling a Haskell module A.B.C:

    The object file is placed in <obj-path>/A/B/C.o
    The interface file is placed in <hi-path>/A/B/C.hi

    Where <objpath> is
	- the argument of the -odir flag, if one was given
	- the element of the search path in which the source file was found,
	  when in --make mode.
	- "." otherwise.

    Where <hipath> is
	- the argument of the -hidir flag, if one was given
	- the element of the search path in which the source file was found,
	  when in --make mode.
	- "." otherwise.

NOTE, in particular, that the name of the source file has no bearing
on the name of the object or interface file any more.  This is a
nchange from the previous semantics, where the name of the object file
would, under certain circumstances, follow the name of the source file.

eg. before, if you said

	ghc -c dir/foo.hs

you would get dir/foo.o.  Now, you get something like Main.o,
depending on what module is in foo.hs.  This means that the driver
pipeline machinery now needs to pass around a Maybe ModLocation, which
is filled in by the Hsc phase and used later on to figure out the name
of the object file (this was fairly painful, but seems to be the only
way to get the right behaviour).

Update comments to match reality.

- change prototype of async proc calls to

      typedef int (*DoProcProc)(void *param);

  i.e., have the proc return a result. Turned out that almost all
  uses of the primop ended up encoding a result via their 'param'.

- when adding new I/O requests, shorten the time the IOManager
  lock is held. Helps to keep down the size of the thread pool.

ppFilesFromSummaries.toPpFile: make it clear that file paths are being eq-tested here

Put a newline after showing an I/O error.

Respect -odir in one-shot compilations.

Arrow notation: add a new (more primitive) form of command:

	cmd ::= ... | cmd aexp

analogous to ordinary application, and also represented using HsApp.
To avoid an overlap, the syntax for combining forms is changed to

	(|aexp cmd1 ... cmdn|)

Add extra functions operating on outsized tuples (used by the translation
of arrow notation).

interruptible(): add missing BlockedOnDoProc case

awaitRequests(): BlockedOnDelay threads not expected on blocked_queue

unblockThread: barf() string wibble

[win32]delay#: set why_blocked to BlockedOnDoProc since we're appending the thread to blocked_queue

Fix the big-char-literal parsing issue in External Core, correctly this time