GitLab

More healing

Heal the HEAD

Add comments, and nuke strictness info in CoreTidy if totally boring

Dear Sigbjorn,

next time you turn an

if...endif

into an

if...else..endif

remember to check the sense of the test. Sometimes, it needs to be reversed
(there was in this case a hint in the surrounding code).

Love,

Reuben.

wibbles only

---------------------
	Clear up infelicities
	---------------------
	CorePrep, CoreUtils, SimplUtils
	LiberateCase (wibbles only)

* Previously CorePrep was floating LocalIds to top level, which
  breaks the invariant that after CorePrep all top level Ids are
  GlobalIds.  But it didn't really need to, and this pass makes it
  so.  It's much tidier now.

* Make CorePrep do eta expansion on partial applications
	x = foldr f y  ==>   x = \ys -> foldr f y ys
  (This used to be done in the simplifier, but now the
  simplifier only eta expands where there is at least one
  lambda already.)

* Omit CoreUtils.etaReduce.  (Never called.)

* Improve CoreUtils.etaExpand, so that it doesn't add gratuitous
  beta redexes.

----------------------------------------------
	Output curried functions for data constructors
	----------------------------------------------
			(incomplete)

The idea here is to output code for the *curried* version of
the worker of a data constructor, so that the worker can be
treated as a first-class citizen elsewhere in the compiler.
In particular, it doesn't need to be a "hasNoBinding" thing,
which are the subject of a number of nasty hacks.

These changes only do the business for the code gen route
via AbstractC.  Remaining to do: the byte code generator.

Idea: move the byte-code gen to STG code, and inject the
curried data-constructor workers at the STG stage.


	I hope the changes here won't make
	anything stop working.  For now, constructor
	workers remain "hasNoBinding" things.

CgConTbls, CodeGen, CoreTidy, CoreToStg

Tidy up ghc/includes/Constants and related things.

Now all the constants that the compiler needs to know, such as header
size, update frame size, info table size and so on are generated
automatically into a header file, DeriviedConstants.h, by a small C
program in the same way as NativeDefs.h.  The C code in the RTS is
expected to use sizeof() directly (it already does).

Also tidied up the constants in MachDeps.h - all the constants
representing the sizes of various types are named SIZEOF_<foo>, to
match the constants defined in config.h.  PrelStorable.lhs now doesn't
contain any special knowledge about GHC's conventions as regards the
size of certain types, this is all in MachDeps.h.

-------------------------------------------
	Don't bomb out on inlining big constructors
	-------------------------------------------

The unfolder bombs out if the expression get too big; no point in
computing its exact size when it's vast.  But for *constructors*
applied to a lot of args, it's worth keeping going, because we
get a big *disount* too
	$fFooInt = :CFoo a1 a2 a3 ... a50

We want to keep the inlining for $fFooInt in interface files, so
that importing guys can do the selection.


Solution: only bomb out when size-discount gets too big.

This file shouldn't be in the repository any more (there's a
PrelGHC.hi-boot.pp).

Why is it only me that notices these phantom leftovers? Why don't they cause
problems for other people?

In the interactive FFI, handle f-i'd fns which return char.

Remove import of unDosifyPath, and the comment saying that it is used.

minor wibble: use Maybe.isJust rather than maybeToBool

Don't print the "compilation IS NOT required" in quiet mode (-v0).

field: robustify - give a parse error (rather than p-match failure)
if LHS isn't known.

Wibble in eqString

Fix spurious sp increment in RET_{VEC_}_SMALL printing code.

MERGE TO STABLE

Dont try to float unboxed things to top level

Add a FAQ entry about using getContents in GHCi.

ILX ifdefery

Various updates to track the state of GHC and ILX.

Track reality a bit.

Implement -keep-il-file.

Implement -keep-il-file

Add -keep-il-file.

PrelGHC.dll -> PrelGHC.i_o, and a few other corrections to follow
changes to ILX.

Reflect reality ever-so-slightly better.

Remove unused import.

Removed unused imports.

Get rid of \rs at the end of lines in prescripts and postscripts to
work around Windows dodginess.

Export run_queue_hd to the dynamic linker; TkHaskell has a grotesque
hack which requires it.

Allow default RTS options to be specified by linking in an object file
which defines the symbol `ghc_rts_opts' to point to a string of RTS
options.

This is preferred to using defaultsHook(), if possible.  Perhaps we
could remove defaultsHook().

(won't work with DLL's; if we ever ressurrect them we'll
have to deal with this somehow).

Actually bomb out with an error message if the maximum heap size is
smaller than the minimum allocation area size.

Fix a bug in the heap size calculation, where a negative result wasn't
noticed because we're working with unsigned types.  We now explicitly
test that the heap has enough room for the minimum allocation area
size, otherwise a heap overflow is reported.

---------------------------
	Match rules before inlining
	---------------------------

This commit fulfils a long-standing wish by Manuel that RULES
matching occurs before inlining.  So if a RULE matches, it'll
get used, even if the function can also be inlined.

It's a bit dodgy to actually rely on this, because maybe the rule
doesn't match *yet* but will do after a bit more transformation.
But it does help with things like class operations.  Class ops are
simply selectors which pick a method out of a dictionary, so they
are inlined rather vigorously.  But we might want a RULE for a
class method (e.g. (==) [Char] = eqString), and such rules would
practically never fire if inlining took priority.

Stuff to make a RULE work for
	eqString "foo" "foo" = True
(etc.)  The rule is of course a BuiltinRule in PrelRules

Add a rule for equality on strings

Fix a long-standing lossage of rules attached to class operations
(A one-line fix to SimplCore.updateBinders.)

Rename 'doDiscard' to 'lexPragma', and make it EOF-resistant.
As was, it would run off the end of a StringBuffer (and SIGSEGV,
most likely) if a pragma wasn't properly closed.

Provide finer-grained control for turning off mk/target.mk's
'all', 'boot' and 'install' rules. i.e., instead of having
the variable NO_ALL_TARGETS control the defnition of rules
for all three, NO_ALL_TARGET, NO_BOOT_TARGET, and NO_INSTALL_TARGET
lets you individually control which ones you don't want.

Sub-projects (GC and HDirect, for example) have the need
to turn off the 'boot' rule, which is what motivated this
change.