Brian Norris [Wed, 6 Feb 2013 22:42:00 +0000 (14:42 -0800)]
cyclegraph: mergeNodes(): return early if we violate RMW
To be consistent, and to prevent unecessary merging work, we should
simply return from the mergeNodes() function as soon as we know it has
failed.
Brian Norris [Wed, 6 Feb 2013 22:41:12 +0000 (14:41 -0800)]
cyclegraph: refactor for indentation
Brian Norris [Wed, 6 Feb 2013 22:00:05 +0000 (14:00 -0800)]
Makefile / git: cleanup generated .dot and .pdf files
Brian Norris [Wed, 6 Feb 2013 02:27:35 +0000 (18:27 -0800)]
cyclegraph: add wrappers for some common functionality
Brian Norris [Wed, 6 Feb 2013 01:45:31 +0000 (17:45 -0800)]
model: fully utilize Promise nodes in CycleGraph
We can eliminate the post-promise-resolution computations now that
mo_graph is built up to include Promise nodes.
Brian Norris [Wed, 6 Feb 2013 01:39:36 +0000 (17:39 -0800)]
model: cosmetic improvements to resolve_promises()
Rename 'resolved' to 'haveResolved', since we will use the 'resolved'
name for something else.
Brian Norris [Wed, 6 Feb 2013 01:36:05 +0000 (17:36 -0800)]
cyclegraph: move function definitions out of header
I found out how to instantiate these functions from within the
implementation file (rather than the header). I can specifically list
those implementations that I want available.
Brian Norris [Wed, 6 Feb 2013 01:13:28 +0000 (17:13 -0800)]
cyclegraph: template-ize addRMWEdge()
Brian Norris [Tue, 5 Feb 2013 22:26:24 +0000 (14:26 -0800)]
cyclegraph: bugfix - pop edges properly
Now that CycleNode's have back-edges, we need to properly remove *both*
the forward and back edges when doing roll-back.
Brian Norris [Tue, 5 Feb 2013 22:10:06 +0000 (14:10 -0800)]
model: add write-to-promise edges
Assume that any time a promise exists, is compatible with a store, and
is exclusive to the same thread as the store, that it is mod-ordered
after that store. This should never produce cycles, until we decide to
begin satisfying promises. At that point, if there's a cycle, then we
must either merge nodes (i.e., the store must satisfy that promise) or
else we discard the execution for the moment and perform the
satisfaction at a later point in the search space.
Brian Norris [Tue, 5 Feb 2013 22:03:28 +0000 (14:03 -0800)]
promise: add is_compatible_exclusive()
Check if a Promise is both compatible *and* exclusive to a particular
thread. This allows ordering optimizations when we can guarantee that a
particular thread must fulfill the promise.
Brian Norris [Tue, 5 Feb 2013 22:01:34 +0000 (14:01 -0800)]
promise: add const
Brian Norris [Tue, 5 Feb 2013 21:47:20 +0000 (13:47 -0800)]
model: utilize CycleGraph 'addEdge()' status
The mo_graph is only modified if addEdge() returns true. Utilize this
information.
Brian Norris [Tue, 5 Feb 2013 21:44:32 +0000 (13:44 -0800)]
cyclegraph: return 'added' status for addEdge()
The ModelChecker would like to know if it's adding a new edge, or just
and existing edge. If the edge is not new, then there are fewer
resulting update checks needed.
Brian Norris [Tue, 5 Feb 2013 06:50:45 +0000 (22:50 -0800)]
cyclegraph: add documentation
Brian Norris [Tue, 5 Feb 2013 01:12:11 +0000 (17:12 -0800)]
cyclegraph: bugfix - only use concrete writes to eliminate threads
Brian Norris [Tue, 5 Feb 2013 00:24:36 +0000 (16:24 -0800)]
cyclegraph: edit template for checkReachable
Brian Norris [Tue, 5 Feb 2013 00:24:21 +0000 (16:24 -0800)]
cyclegraph: edit template for addEdge
Brian Norris [Tue, 5 Feb 2013 00:22:03 +0000 (16:22 -0800)]
cyclegraph: add overloaded getNode_noCreate()
Need a generic way to map a ModelAction/Promise to a CycleNode.
Brian Norris [Sat, 2 Feb 2013 01:51:56 +0000 (17:51 -0800)]
model: add promise-node edges for 'read' actions
Brian Norris [Sat, 2 Feb 2013 01:47:57 +0000 (17:47 -0800)]
cyclegraph: template-ize checkReachable()
ModelChecker will need to query the mo_graph regarding the modification
order of promised future writes.
Brian Norris [Sat, 2 Feb 2013 00:44:22 +0000 (16:44 -0800)]
model: template-ize 'r_modification_order'
We will need to add "read modification order" edges when 'rf' is either
a ModelAction or a Promise. This function works nicely with a template.
Brian Norris [Fri, 1 Feb 2013 23:33:38 +0000 (15:33 -0800)]
action: fixup printing for RMW/RMWR
Some read actions didn't really have a correct 'value' saved in the
ModelAction. Now that we have the 'promise' recorded, always retrieve
the value from reads_from or reads_from_promise.
Brian Norris [Fri, 1 Feb 2013 23:27:32 +0000 (15:27 -0800)]
action: store Promise in ModelAction
Brian Norris [Fri, 1 Feb 2013 23:11:13 +0000 (15:11 -0800)]
cyclegraph: add full promise resolution, node merging
We now support the interfaces needed for tracking Promise constraints in
CycleGraph.
Brian Norris [Fri, 1 Feb 2013 23:09:07 +0000 (15:09 -0800)]
cyclegraph: add CycleNode::resolvePromise(), CycleNode::getPromise()
Brian Norris [Fri, 1 Feb 2013 23:07:56 +0000 (15:07 -0800)]
promise: add is_compatible()
Brian Norris [Tue, 29 Jan 2013 02:48:46 +0000 (18:48 -0800)]
cyclegraph: map Promises to Promise nodes
Brian Norris [Tue, 29 Jan 2013 02:39:23 +0000 (18:39 -0800)]
cyclegraph: template for addEdge()
This function will be identical for all combinations of ModelAction /
Promise. So just make it a template.
Brian Norris [Tue, 29 Jan 2013 02:30:28 +0000 (18:30 -0800)]
cyclegraph: rename addEdge() to addNodeEdge()
This is a "low-level" internal function, where the other addEdge()
implementations will be high-level, external interfaces. Differentiate
the names, then.
Brian Norris [Sat, 26 Jan 2013 00:41:27 +0000 (16:41 -0800)]
cyclegraph: add removeEdge(), removeBackEdge()
For popping edges of a node that's about to be merged with another one.
Note that if using these functions, a user has to manage the
'CycleGraph::hasCycles' state. This only handles the inter-Node edge
lists.
Brian Norris [Sat, 26 Jan 2013 01:19:55 +0000 (17:19 -0800)]
cyclegraph: add Promise CycleNode
A very bare-bones constructor.
Brian Norris [Wed, 6 Feb 2013 21:42:24 +0000 (13:42 -0800)]
cyclegraph: trivial shortening
Brian Norris [Tue, 5 Feb 2013 23:16:03 +0000 (15:16 -0800)]
Makfile: add 'make pdfs' target to build PDF from .dot graph
Brian Norris [Wed, 6 Feb 2013 03:31:57 +0000 (19:31 -0800)]
model: don't print scheduler randomly
Brian Norris [Sat, 2 Feb 2013 02:08:39 +0000 (18:08 -0800)]
action: fix incorrect comment
Brian Norris [Sat, 2 Feb 2013 01:18:32 +0000 (17:18 -0800)]
Doxyfile: don't recurse into all directories
Brian Norris [Wed, 30 Jan 2013 18:19:21 +0000 (10:19 -0800)]
cyclegraph: use vector empty() and clear()
Brian Norris [Tue, 29 Jan 2013 18:15:44 +0000 (10:15 -0800)]
nodestack/model: spacing
Brian Norris [Sat, 26 Jan 2013 00:59:16 +0000 (16:59 -0800)]
promise: update names/comments
Make it more clear what 'act' is (it's a 'read' action).
Brian Norris [Fri, 25 Jan 2013 21:35:38 +0000 (13:35 -0800)]
cyclegraph: RMW atomicity violation must flag a cycle
Because we've removed some of the special-casing for RMW atomicity
violations, we don't need the separate 'hasRMWViolation' condition;
instead, we should explicitly flag atomicity violations as a cycle.
Previously, there was a subtle bug related to this issue, where cycles
were flagged only as RMW violations and did not show up to the
model-checker at the appropriate points.
Brian Norris [Fri, 25 Jan 2013 01:29:58 +0000 (17:29 -0800)]
cyclegraph: reformat, improve 'addRMWEdge' comments
Make it slightly clearer what this function's intent is. (It's only used
for the introduction of the first edge between 'from' and 'rmw' s.t.:
from->is_write() && rmw->is_rmw() && from --rf-> rmw
Brian Norris [Fri, 25 Jan 2013 01:26:28 +0000 (17:26 -0800)]
cyclegraph: separate an 'addEdge(CycleNode *, CycleNode *) function
We need to represent edge addition as a low-level operation on
CycleNodes, not just on ModelActions.
Brian Norris [Fri, 25 Jan 2013 01:18:06 +0000 (17:18 -0800)]
model: remove ModelChecker::is_infeasible_ignoreRMW()
We don't make any special exceptions for this condition anymore, so drop
the function.
Brian Norris [Thu, 10 Jan 2013 01:39:24 +0000 (17:39 -0800)]
model: add immediate future value for RMW reordering
When RMW atomicity would otherwise rule an execution infeasible, we may
need to pass a future value back 'immediately,' to allow the
model-checker to effectively reorder the RMW's. For such cases, we don't
wait for all current promises to be resolved.
Brian Norris [Wed, 9 Jan 2013 23:26:47 +0000 (15:26 -0800)]
cyclegraph: detect cycles immediately
Previously, we have used a special case in CycleGraph to delay detection
of a mo-cycle (actually, an RMW atomicity violation) long enough so that
a RMW can "reorder" across another RMW by passing its future value.
Unfortunately, this only works some of the time, as the execution
*may or may not* be a feasible prefix which allows the sending of the
future value.
Instead, we need to pull this special case out of CycleGraph and (in
subsequent commits) will add a more proper solution in model.cc.
Brian Norris [Thu, 24 Jan 2013 03:00:49 +0000 (19:00 -0800)]
promise: bugfix - a disabled thread should not be eliminated
Because we now only consider a subset of threads for satisfying
promises, we can have a disabled thread that still may satisfy a
promise; it could simply synchronize with another unrelated thread, then
continue to generate a write that can resolve the promise.
As a side effect, this change makes "has_failed()" much simpler.
Brian Norris [Thu, 24 Jan 2013 02:47:53 +0000 (18:47 -0800)]
model: modify promises on THREAD_{CREATE,FINISH}
When a thread is created it *must* pass its promise satisfiability to
its children.
When a thread finishes, we should remove it from all promises. This can
help prune a search early, if some read was depending only on the
finishing thread: we can then declare an unresolved promise.
Brian Norris [Thu, 24 Jan 2013 02:43:14 +0000 (18:43 -0800)]
promise: associate Promises with a set of threads
A Promise can be associated with a set of threads, so that we can
perform elimination more efficiently and precisely.
Brian Norris [Thu, 24 Jan 2013 02:05:28 +0000 (18:05 -0800)]
action: record future value
Future values don't currently show up in ModelAction (e.g., when
printing). We should set the value when we choose to read from a future
value.
Brian Norris [Thu, 24 Jan 2013 01:33:23 +0000 (17:33 -0800)]
future_value: add thread ID parameter
This just records the parameter for now, and makes it a distinction
between future values. Shouldn't affect model-checker behavior, except
for the introduction of extra future values: one for each thread which
writes the same value.
We calculate whether the writing thread exists at the time of the future
value and if not, find an ancestor thread which does.
Brian Norris [Thu, 24 Jan 2013 01:24:11 +0000 (17:24 -0800)]
nodestack: pass 'struct future_value' to add_future_value()
This makes more sense than passing individual parameters. This way, we
construct our future_value in ModelChecker, then simply pass it to the
NodeStack.
Brian Norris [Thu, 24 Jan 2013 01:21:19 +0000 (17:21 -0800)]
model: add const to get_thread(ModelAction *act)
Brian Norris [Thu, 24 Jan 2013 01:02:46 +0000 (17:02 -0800)]
model: fixup indentation, spelling in comment
Brian Norris [Thu, 24 Jan 2013 00:51:39 +0000 (16:51 -0800)]
model: style
Brian Norris [Thu, 24 Jan 2013 00:34:00 +0000 (16:34 -0800)]
model: remove local variable
Brian Norris [Wed, 23 Jan 2013 20:56:01 +0000 (12:56 -0800)]
promise: SnapshotAlloc for vector
We must ensure that any new memory allocated by the model-checker uses
our private snapshotting heap, not the user program's heap.
Brian Norris [Wed, 23 Jan 2013 20:50:57 +0000 (12:50 -0800)]
promise: refactor has_failed()
has_failed() can simply call the thread_is_eliminated() function which
does much of the same computation already. We can drop the
'((int)i != promise_tid)' clause, since the 'promise_tid' thread must
already be eliminated.
Brian Norris [Wed, 23 Jan 2013 20:49:01 +0000 (12:49 -0800)]
promise: move thread_is_eliminated()
Should be in the implementation file, not the header.
Brian Norris [Wed, 23 Jan 2013 20:36:28 +0000 (12:36 -0800)]
promise: rename has_sync_thread() -> thread_is_eliminated()
This function doesn't actually check for synchronization; there are
numerous ways to be eliminated.
Brian Norris [Wed, 23 Jan 2013 19:53:33 +0000 (11:53 -0800)]
promise: rename check_promise() -> has_failed()
The name "check_promise" doesn't really tell what this function is
doing. It is actually checking if the promise is unresolvable.
Brian Norris [Wed, 23 Jan 2013 19:49:01 +0000 (11:49 -0800)]
promise: refactor eliminate_thread()/check_promise()
eliminate_thread() and check_promise() have near-duplicate code, for
checking if the promise has failed. This change makes them consistent
(check_promise() now knows to eliminate its own thread immediately) and
allows one function to simply call the other.
Brian Norris [Wed, 23 Jan 2013 19:38:43 +0000 (11:38 -0800)]
promise: rename increment_threads() -> eliminate_thread()
This function name really doesn't describe what it does, in my opinion.
Update the name and add a description.
Brian Norris [Wed, 23 Jan 2013 19:10:20 +0000 (11:10 -0800)]
nodestack: add const
Brian Norris [Wed, 23 Jan 2013 19:04:59 +0000 (11:04 -0800)]
promise: construct with a 'future_value' struct
Don't pass a bunch of parameters individually; just pass the struct.
Brian Norris [Wed, 23 Jan 2013 19:02:38 +0000 (11:02 -0800)]
nodestack/promise: move future_value struct, update headers
struct future_value will be used more closely with class Promise, so put
it in the promise.h header. This requires some updates to other header
inclusions.
Brian Norris [Wed, 23 Jan 2013 18:43:48 +0000 (10:43 -0800)]
nodestack: pass writer ModelAction to add_future_value()
NodeStack will have to know a little more about who wrote the future
value, so just pass the ModelAction as a parameter.
Brian Norris [Wed, 23 Jan 2013 18:41:23 +0000 (10:41 -0800)]
nodestack: eliminate get_future_value_expiration()
We don't need separate accessors for each member of 'struct
future_value'. Just pass the entire struct.
Brian Norris [Wed, 23 Jan 2013 18:30:46 +0000 (10:30 -0800)]
nodestack: remove extra blank lines
Brian Norris [Wed, 23 Jan 2013 18:17:44 +0000 (10:17 -0800)]
remove #include <list>
STL lists aren't used here anymore.
Brian Norris [Wed, 23 Jan 2013 00:49:11 +0000 (16:49 -0800)]
nodestack: improve debug print() method
Print the contents of the future values and may-read-from set, if we're
printing the NodeStack for debugging purposes.
Brian Demsky [Wed, 23 Jan 2013 00:31:16 +0000 (16:31 -0800)]
fix bug in promise handling code...
if any promise for a thread resolved, we assumed that the corresponding
write happened after all promises for that threads...clearly doesn't
need to be the case...
[Brian Norris:] This fixes some bugs where we don't see all the expected
behaviors: e.g., with
./run.sh test/rmwprog.o -- 3
Now we see all 20 behaviors.
Brian Norris [Wed, 23 Jan 2013 00:24:44 +0000 (16:24 -0800)]
promise: fixup style
Brian Norris [Tue, 22 Jan 2013 19:08:06 +0000 (11:08 -0800)]
test: rmwprog: support command-line argument
This test can easily be extended to allow more than 2 atomic increments
per thread. For instance, it's interesting to see the model-checker
behavior for 4 increments per thread. Now, you can do this with a simple
numeric command-line argument that gets passed to the user program.
Example - run rmwprog with 4 increments per thread, viewing the
execution traces verbosely:
./run.sh test/rmwprog.o -v -- 4
Brian Norris [Sat, 19 Jan 2013 00:43:02 +0000 (16:43 -0800)]
nodestack: improve debug prints
Brian Norris [Sat, 19 Jan 2013 00:38:55 +0000 (16:38 -0800)]
schedule: improve scheduler printing
Print more detailed, informative info about the Scheduler, still on a
single line.
Brian Norris [Sat, 19 Jan 2013 00:30:26 +0000 (16:30 -0800)]
schedule: allow Schedule::print() even in non-DEBUG build
For instance, printing from w/in gdb
Brian Norris [Thu, 17 Jan 2013 01:42:57 +0000 (17:42 -0800)]
model: bugfix - sleep sets are NOT directly compatible with DPOR
We failed to incorporate changes from the POPL'05 DPOR paper's addendum,
preventing us from seeing all behaviors in some test programs. This
fixes that.
Brian Norris [Wed, 16 Jan 2013 08:24:48 +0000 (00:24 -0800)]
litmus: iriw: allow command-line switch 's' for seq_cst
The IRIW litmus test fails if we make it seq_cst; we don't see all 15
potential behaviors. Run with:
# release/acquire litmus test
./run.sh test/litmus/iriw.o -v
# seq_cst litmus test
./run.sh test/litmus/iriw.o -v -- s
Brian Norris [Wed, 16 Jan 2013 07:57:45 +0000 (23:57 -0800)]
litmus: iriw: use release/acquire, not release/relaxed
Brian Norris [Wed, 16 Jan 2013 07:43:44 +0000 (23:43 -0800)]
tests: litmus: link up Makefile properly
Now this builds the litmus tests properly.
Brian Norris [Tue, 15 Jan 2013 20:44:47 +0000 (12:44 -0800)]
test: litmus: add litmus tests from Nitpicking C++
This adds most of the litmus tests from Nitpicking C++ Concurrency. Note
that it leaves out the 'consume' tests (replacing it with the
also-mentioned release/acquire test). It doesn't test the seq_cst
results, although I could provide command-line switches to do so...
Brian Norris [Tue, 15 Jan 2013 20:30:33 +0000 (12:30 -0800)]
test: Makefile: add $(BASE) variable
Makes this Makefile a little more flexible
Brian Norris [Thu, 10 Jan 2013 18:54:02 +0000 (10:54 -0800)]
model: reformat execution trace prints
Bring the "INFEASIBLE" message together with the "Execution #" message,
to make the log more compact and readable (e.g., grep for "Execution"
will give you some regularly-patterned, useful information).
Brian Norris [Thu, 10 Jan 2013 01:50:57 +0000 (17:50 -0800)]
model: spelling mistake
Brian Norris [Wed, 9 Jan 2013 23:54:04 +0000 (15:54 -0800)]
model: fixup "infeasible" messages
These messages should be much more compact and targeted. Currently, we
either get no infeasibility messages (when verbose=0) or we get messages
every time is_infeasible() is executed. Neither is helpful for debugging,
since the infeasibility may be temporary, as we choose a new path for
execution on the fly. So instead, just print the infeasibility message
at the end of infeasible executions.
Also, rather than spread the messages across multiple lines, just put
them together into a single string.
Brian Norris [Wed, 9 Jan 2013 22:07:07 +0000 (14:07 -0800)]
cyclegraph: add documentation
Brian Norris [Wed, 9 Jan 2013 21:47:57 +0000 (13:47 -0800)]
cyclegraph: remove redundant code
The edgeCreatesCycle() function can simply be reduced to
checkReachable(), since checkReachable() will detect "to == from".
Brian Norris [Wed, 9 Jan 2013 00:40:29 +0000 (16:40 -0800)]
cyclegraph: add edgeCreatesCycle() function
This is repeated code, so make it a function.
Brian Norris [Wed, 9 Jan 2013 00:29:17 +0000 (16:29 -0800)]
cyclegraph: add const
Brian Norris [Tue, 8 Jan 2013 02:22:57 +0000 (18:22 -0800)]
model: add 'add_future_value()' wrapper
Encapsulate a little logic in a function here, so we can reuse this in
other places.
Brian Norris [Tue, 8 Jan 2013 01:38:25 +0000 (17:38 -0800)]
model: add PendingFutureValue constructor
Brian Demsky [Tue, 8 Jan 2013 00:55:00 +0000 (16:55 -0800)]
new test case
Brian Norris [Fri, 4 Jan 2013 19:05:10 +0000 (11:05 -0800)]
cyclegraph: add back edges to CycleNode
Brian Norris [Fri, 4 Jan 2013 18:09:29 +0000 (10:09 -0800)]
hashtable: style
Brian Norris [Fri, 4 Jan 2013 17:53:04 +0000 (09:53 -0800)]
cyclegraph: add putNode() helper
Brian Norris [Fri, 4 Jan 2013 01:01:09 +0000 (17:01 -0800)]
hashtable: style
Brian Norris [Fri, 4 Jan 2013 00:24:54 +0000 (16:24 -0800)]
promise: style fixup
Brian Norris [Fri, 4 Jan 2013 00:18:10 +0000 (16:18 -0800)]
cyclegraph: fix insignificant memory leak
Brian Norris [Thu, 3 Jan 2013 22:00:37 +0000 (14:00 -0800)]
cyclegraph: small cleanup