myself because I'm getting tired of seeing the red buildbots, which have
been red since 5:30PM PDT last night.
Proposed supplement to developer policy: committers should make sure to
be around to watch for buildbot failures after committing.
llvm-svn: 77785
- Off by default, you can test it with the --no-sh argument.
- For me it works for all but 3 tests, but there a number of FIXMEs and QOI
issues:
o Redirection isn't completely accurate -- in practice it can't portably
be, but I would like to error out if someone writes something which isn't
going to work. This is the source of the 3 test failures.
o Some pipe configurations have the potential to deadlock.
o It is significantly slower when multithreaded. I believe this is due to
locking happening under the hood, there is probably some kind of solution
but I haven't investigated yet.
o Log output is ugly.
llvm-svn: 77784
made to RegionStore (and related classes) in order to handle some
analyzer failures involving casts and manipulation of symbolic memory.
The root of the change is in StoreManager::CastRegion(). Instead of
using ad hoc heuristics to decide when to layer an ElementRegion on a
casted MemRegion, we now always layer an ElementRegion when the cast
type is different than the original type of the region. This carries
the current cast information associated with a region around without
resorting to the error prone recording of "casted types" in GRState.
Along with this new policy of layering ElementRegions, I added a new
algorithm to strip away existing ElementRegions when they simply
represented casts of a base memory object. This algorithm computes
the raw "byte offset" that an ElementRegion represents from the base
region, and allows the new ElementRegion to be based off that offset.
The added benefit is that this naturally handles a series of casts of
a MemRegion without building up a set of redundant ElementRegions
(thus canonicalizing the region view).
Other related changes that cascaded from this one (as tests were
failing in RegionStore):
- Revamped RegionStoreManager::InvalidateRegion() to completely remove
all bindings and default values from a region and all subregions.
Now invalidated fields are not bound directly to new symbolic
values; instead the base region has a "default" symbol value from
which "derived symbols" can be created. The main advantage of this
approach is that it allows us to invalidate a region hierarchy and
then lazily instantiate new values no matter how deep the hierarchy
went (i.e., regardless of the number of field accesses,
e.g. x->f->y->z->...). The previous approach did not do this.
- Slightly reworked RegionStoreManager::RemoveDeadBindings() to also
incorporate live symbols and live regions that do not have direct
bindings but also have "default values" used for lazy instantiation.
The changes to 'InvalidateRegion' revealed that these were necessary
in order to achieve lazy instantiation of values in the region store
with those bindings being removed too early.
- The changes to InvalidateRegion() and RemoveDeadBindings() revealed
a serious bug in 'getSubRegionMap()' where not all region -> subregion
relationships involved in actually bindings (explicit and implicit)
were being recorded. This has been fixed by using a worklist algorithm
to iteratively fill in the region map.
- Added special support to RegionStoreManager::Bind()/Retrieve() to handle
OSAtomicCompareAndSwap in light of the new 'CastRegion' changes and the
layering of ElementRegions.
- Fixed a bug in SymbolReaper::isLive() where derived symbols were not
being marked live if the symbol they were derived from was also live.
This fix was critical for getting lazy instantiation in RegionStore
to work.
- Tidied up the implementation of ValueManager::getXXXSymbolVal() methods
to use SymbolManager::canSymbolicate() to decide whether or not a
symbol should be symbolicated.
- 'test/Analysis/misc-ps-xfail.m' now passes; that test case has been
moved to 'test/Analysis/misc-ps.m'.
- Tweaked some pretty-printing of MemRegions, and implemented
'ElementRegion::getRawOffset()' for use with the CastRegion changes.
llvm-svn: 77782
instructions for calls since BL and BLX are always 32-bit long and BX is always
16-bit long.
Also, we should be using BLX to call external function stubs.
llvm-svn: 77756
in SelectionDAGLowering::visitTargetIntrinsic.
This removes a bit of special-case code for vector types. After staring
at it for a while, I managed to convince myself that it is not necessary.
The only case where TLI.getValueType() differs from MVT::getMVT is for iPTR,
so this code could potentially make a difference for a vector of pointers.
But, it looks like that is not supported. Calling TLI.getValueType() on
a vector of pointers leads to the following sequence of calls:
TargetLowering::getValueType
MVT::getMVT
MVT::getVectorVT(iPTR, num elements)
MVT::getExtendedVectorVT
MVT::getTypeForMVT for iPTR
assertion fails "Type is not extended!"
So, unless I'm really missing something, this bit of code is irrelevant to
the current version of LLVM, which is consistent with the fact that I don't
see this code in other similar places.
llvm-svn: 77747
- Operands which are just a label should be parsed as immediates, not memory
operands (from the assembler perspective).
- Match a few more flavors of immediates.
- Distinguish match functions for memory operands which don't take a segment
register.
- We match the .s for "hello world" now!
llvm-svn: 77745
Chris Lattner