mirrors/emacs
1
Fork 0
mirror of git://git.sv.gnu.org/emacs.git synced 2026-04-20 04:50:55 -07:00
Code Issues Projects Releases Wiki Activity

Analysis justifying the value of stackprobedepth.

Browse source 
Copied from Perforce
 Change: 187380
 ServerID: perforce.ravenbrook.com
This commit is contained in:
Gareth Rees 2014-10-24 10:16:28 +01:00
parent 5c8c708445
commit 0975646ca3
2 changed files with 65 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

9
mps/code/config.h
View file

@ -427,14 +427,11 @@
#define VM_ARENA_SIZE_DEFAULT ((Size)1 << 28)
/* Stack configuration */
/* Stack configuration -- see <code/sp*.c> */
/* Currently StackProbe has a useful implementation only on Windows. */
#if defined(PLATFORM_ANSI)
#define StackProbeDEPTH ((Size)0)
#elif defined(MPS_OS_W3) && defined(MPS_ARCH_I3)
#define StackProbeDEPTH ((Size)500)
#elif defined(MPS_OS_W3) && defined(MPS_ARCH_I6)
#if defined(MPS_OS_W3)
/* See <design/sp/#sol.depth.analysis> for a justification of this value. */
#define StackProbeDEPTH ((Size)500)
#else
#define StackProbeDEPTH ((Size)0)

68
mps/design/sp.txt
View file

@ -58,8 +58,61 @@ probes.
.. _Windows: http://support.microsoft.com/kb/100775
_`.sol.no-recursion`: In order to implement this design, the MPS must
have constant bounded stack depth, and therefore, no recursion.
_`.sol.depth.no-recursion`: In order to implement this design, the MPS
must have constant bounded stack depth, and therefore, no recursion.
_`.sol.depth.analysis`: Here's a table showing a deep call into the
MPS (in the master sources at changelevel 187378), starting in
``ArenaAccess()`` at the point where the arena ring lock is taken. The
access forces a scan of a segment in an AMC pool, which fixes a
reference to an object in an AMC pool's oldspace, which has to be
forwarded, and this overflows the forwarding buffer, which requires
the arena to allocate a new buffer in an appropriate zone, by
searching the splay tree representing free memory.
The "Args" column gives the number of arguments to the function (all
arguments to functions in the MPS are word-sized or smaller, since we
prohibit passing structures by value), and the "Locals" column gives
the number of words in local variables.
==== ====== ========================
Args Locals Function
==== ====== ========================
5 0 ``PoolAccess()``
5 0 ``PoolSegAccess()``
3 5 ``TraceSegAccess()``
4 1 ``traceScanSeg()``
4 8 ``traceScanSegRes()``
4 0 ``PoolScan()``
4 5 ``AMCScan()``
3 ??? ``format->scan()``
4 15 ``AMCFix()``
4 5 ``BufferFill()``
6 10 ``AMCBufferFill()``
6 9 ``PoolGenAlloc()``
7 5 ``SegAlloc()``
5 5 ``ArenaAlloc()``
5 5 ``arenaAllocPolicy()``
5 11 ``arenaAllocFromLand()``
7 1 ``LandFindInZones()``
7 16 ``cbsFindInZones()``
5 3 ``cbsFindFirst()``
6 7 ``SplayFindFirst()``
3 7 ``SplaySplay()``
4 8 ``SplaySplitDown()``
3 0 ``SplayZig()``
109 126 **Total**
==== ====== ========================
We expect that a compiler will often be able to share stack space
between function arguments and local variables, but in the worst case
where it cannot, this call requires 235 words of stack space, plus
whatever stack space is required by the client program's scanning
function.
This isn't necessarily the deepest call into the MPS, but it's
probably close. The value for ``StackProbeDEPTH`` is thus chosen to be
a round number that's comfortably larger than this.
Interface
@ -77,15 +130,18 @@ Issues
_`.issue.an`: The generic implementation is non-functional and so does
not meet `.req.complete`_.
_`.issue.depth`: The value for ``StackProbeDEPTH`` is just a guess:
there is no assurance that it is correct. See job003639_.
_`.issue.depth`: The analysis in `.sol.depth.analysis`_ does not
guarantee that the value for ``StackProbeDEPTH`` is correct.
.. _job003639: http://www.ravenbrook.com/project/mps/issue/job003639/
_`.issue.depth.client`: The design relies on the client program's
object format methods not using too much stack. This restriction needs
to be documented.
_`.issue.depth.constraint`: None of the implementations take account
of `.sol.depth.constraint`_. (This is harmless for now as
``StackProbeDEPTH`` is smaller than the number of words per page on
all supported platforms.)
all supported platforms, but there should at least be an assertion
somewhere.)
Implementations