Fixed an issue where our new accelerator tables could cause a crash
when we got a full 32 bit hash match, yet a C string mismatch.
We had a member variable in DWARFMappedHash::Prologue named
"min_hash_data_byte_size" the would compute the byte size of HashData
so we could skip hash data efficiently. It started out with a byte size
value of 4. When we read the table in from disk, we would clear the
atom array and read it from disk, and the byte size would still be set
to 4. We would then, as we read each atom from disk, increment this count.
So the byte size of the HashData was off, which means when we get a lookup
whose 32 bit hash does matches, but the C string does NOT match (which is
very very rare), then we try and skip the data for that hash and we would
add an incorrect offset and get off in our parsing of the hash data and
cause this crash.
To fix this I added a few safeguards:
1 - I now correctly clear the hash data size when we reset the atom array using the new DWARFMappedHash::Prologue::ClearAtoms() function.
2 - I now correctly always let the AppendAtom() calculate the byte size of the hash (before we were doing things manually some times, which was correct, but not good)
3 - I also track if the size of each HashData is a fixed byte size or not, and "do the right thing" when we need to skip the data.
4 - If we do get off in the weeds, then I make sure to return an error and stop any further parsing from happening.
llvm-svn: 147334
captured. This allows the tracker to look at the specific use, which may be
especially interesting for function calls.
Use this to fix 'nocapture' deduction in FunctionAttrs. The existing one does
not iterate until a fixpoint and does not guarantee that it produces the same
result regardless of iteration order. The new implementation builds up a graph
of how arguments are passed from function to function, and uses a bottom-up walk
on the argument-SCCs to assign nocapture. This gets us nocapture more often, and
does so rather efficiently and independent of iteration order.
llvm-svn: 147327
block-level expr. Currently CXXConstructExpr is always added as a block-level
expr. This caused two problems for the analyzer (and potentially for the
CFG-based codegen).
1. We have no way to know whether a ctor call is base or complete.
2. We have no way to know the destination object being contructed.
llvm-svn: 147306
mach_override library.
In this case chances are that functions from it will be called from
mach_override_ptr() during ASan initialization.
This may lead to crashes (if those functions are instrumented) or
incorrect behavior (if the implementations differ).
The attached patch renames mach_override_ptr() into
__asan_mach_override_ptr() and makes the rest of the mach_override
internals hidden.
The corresponding AddressSanitizer bug is
http://code.google.com/p/address-sanitizer/issues/detail?id=22
Patch by glider@google.com
llvm-svn: 147303
Kostya Serebryany