GitLab

extend the scope of the doNothing() macro; can now be used in Stg headers

Put spaces round %, in preparation for splittable name supplies

unbreak tcModule.

Use (DefMeth Name) rather than (DefMeth Id) in ClassOpItem.  This not
only eliminates a space leak, because Names generally hold on to much
less stuff than Ids, but also turns out to be a minor cleanup.

Replace a lazy pattern match in tcGetInstLoc with a strict one (fixes
a space leak).

Adorn various constructor fields with strictness annotations - this
fixes at least one space leak, in the mi_usages field of a ModIface.

Collect up _scc_ expressions on the right hand side of a closure
definition and attach them directly to the closure using PushCC-type
cost centres, so that the allocation of the closure gets attributed to
the right place.

Fix a long-standing bug in the cost attribution of cost-center stacks.
The problem case is this:

	let z = _scc_ "z" f x
	in ... z ...

previously we were attributing the cost of allocating the closure 'z'
to the enclosing cost center stack (CCCS), when it should really be
attributed to "z":CCCS.  The effects are particularly visible with
retainer profiling, because the closure retaining 'f' and 'x' would
show up with the wrong CCS attached.

To fix this, we need a new form of CCS representation internally:
'PushCC CostCentre CostCentreStack' which subsumes (and therefore
replaces) SingletonCCS.  SingletonCCS is now represented by 'PushCC cc
NoCCS'.

The CCS argument to SET_HDR may now be an arbitrary expression, such
as PushCostCentre(CCCS,foo_cc), as may be the argument to CCS_ALLOC().
So we combine SET_HDR and CCS_ALLOC into a single macro, SET_HDR_, to
avoid repeated calls to PushCostCentre().

Call LDV_ENTER() on entry to a constructor in profiling mode.

Call LDV_ENTER() on entry to a thunk or function in profiling mode.

Add
	ldvEnter :: Code

for LDV profiling: it records that the closure pointed to by R1 has
just been entered.

We don't need to consider Hp as a volatile register across C calls; it
is already saved by the CALLER_SAVES_SYSTEM macro.

Fix compilation problems with Stats.c in the !DEBUG case.

Omit GC and SYSTEM from the cost centre summary at the top of the profile.

Add rts_getThreadId to the Linker's built in symbol table

Retainer Profiling / Lag-drag-void profiling.

This is mostly work by Sungwoo Park, who spent a summer internship at
MSR Cambridge this year implementing these two types of heap profiling
in GHC.

Relative to Sungwoo's original work, I've made some improvements to
the code:

   - it's now possible to apply constraints to retainer and LDV profiles
     in the same way as we do for other types of heap profile (eg.
     +RTS -hc{foo,bar} -hR -RTS gives you a retainer profiling considering
     only closures with cost centres 'foo' and 'bar').

   - the heap-profile timer implementation is cleaned up.

   - heap profiling no longer has to be run in a two-space heap.

   - general cleanup of the code and application of the SDM C coding
     style guidelines.

Profiling will be a little slower and require more space than before,
mainly because closures have an extra header word to support either
retainer profiling or LDV profiling (you can't do both at the same
time).

We've used the new profiling tools on GHC itself, with moderate
success.  Fixes for some space leaks in GHC to follow...

Inlining pragma buglet

Update this document.  Some of the implementation problems can be
solved in a cleaner way, and the document was confused about the
problems caused by slop in the heap (it's only a problem when
traversing the dead heap, not the live heap).

initialize_virtual_timer: Robustified win32 version; don't
assume that the minimal resolution is one millisec, consult
the underlying impl via timeGetDevCaps().

StablePtr indexes/counts are no longer weighted (i.e., sync wrt Stable.{c,h} changes)

#ifdef protect generated config.h file.

Turn the weighted reference count in a StablePtr into a simple counter
in the stable ptr table.  We never made use of the weighted count, and
it caused problems when making many StablePtrs to the same object.

In the future we could remove the reference counting altogether and
just make a new StablePtr each time, rather than attempting to use an
existing one.  This would save on the hash table lookup at creation
time.

Have prototypes and implementation of prel_PrelHandle_{write,read}() match
up with the stated Haskell FFI type in PrelHandle.hs, i.e., the size
parameter is a CInt, not an Int.

getTimes (win32): Win9x OSes don't provide per-process timing data, so
GetProcessTimes() doesn't do anything interesting. So, resort to using
system-time instead to measure user/process time iff running under Win9x.
(the precision is not as good, deltas were ~6 millisecs when testing
on a Win98 box, but it's better than nothing).

This change should make profiling on Win9x boxes finally work.

Nuke #include of MachDeps.h - nothing in there that cannot be
(better) gotten from the in-tree config.h

This is the only (last?) use of MachDeps.h, so if anyone won't argue
for its continued existence, I'll nuke it sometime soon.

Print to stdout instead of stderr.  I'm pretty ambivalent about this,
but Sungwoo Park made the changes as part of his retainer profiling
work so he must have had a good reason for it at the time...

Add constant-folding rules for Float# and Double#

If the configure script determined that the build-tree version of happy
is to be used, descend into happy/ and build it, if needs be.

- change the interpretation of FPTOOLS_TOP_ABS, it is now the UNIXy path.
  For Win32, it is of the form <drive>:<path> where '/' is the directory
  separator. Prior to this commit, the directory separator was '\'.
- for the (only) place we need to have a platform-native representation
  of FPTOOLS_TOP_ABS, use FPTOOLS_TOP_ABS_PLATFORM.

Tidy up imports

Improve error msg

--------------------------------------
	Yet another cut at the DmdAnal domains
	--------------------------------------

This version of the domain for demand analysis was developed
in discussion with Peter Sestoft, so I think it might at last
be more or less right!

Our idea is mentally to separate
	strictness analysis
from
	absence and boxity analysis

Then we combine them back into a single domain.  The latter
is all you see in the compiler (the Demand type, as before)
but we understand it better now.

Added a new section that covers lexing and parsing.

if IS_CBITS_LIB is YES: Extend SRC_HS2HS_OPTS in the same way that SRC_CC_OPTS is

---------------------------------------
	Add continuation splitting to Simplify
	---------------------------------------

When the simplifier finds a 'case', it calls mkDupableAlt
to make the "continuation" (that is, the context of the
case expression) duplicatable, so that it can push it into
the case branches.  This is crucial for the case-of-case
transformation.

But it turns out that it's a bad idea to do that when
the context is "I'm the argument of a strict function".  Consider

	f (case x of { True -> False; False -> True }) arg2

where f is a strict function.  Then we *could* (and were)
transforming to

	let $j a = f a arg2
	in
	case x of { True -> $j False; False -> $j True }

But this is in general a terribly bad thing to do.
See the example in comments with Simplify.mkDupableCont.

Correct docs about SPECIALIZE instance; MERGE TO STABLE

Don't restrict doc building to i386-linux.  please merge.

use SET_HDR rather than explicitly setting p->header.info.

Fix up the interpreter following the recent modifications to
suspendThread/resumeThread.  Someone should test that foreign imports
in the interpreter still work.

Change the way we do finalization for duplex handles.  Previously, we
arranged that the read side pointed to the right side via a special
handle type (ReadSideHandle _), and the finalizer points to the write
side.  This turned out to interact badly with hGetContents, which
likes to explicitly close the read side of the handle after it reads
EOF or gets an error, which resulted in double-closes for duplex
handles.

Now we store the pointer from the read side to the write side in the
Handle structure itself, so it doesn't get lost when hGetContents
changes the handle type to SemiClosedHandle.  Furthermore, in hClose
we no longer close the file descriptor associated with the read side
of a duplex handle - the actual close will have to wait until the
finalizer runs, because someone might still be using the write side.

Thanks to Volker Stolz for pointing out the problem.