For some basic blocks, it is possible to generate many candidate pairs for
relatively few pairable instructions. When many (tens of thousands) of these pairs
are generated for a single instruction group, the time taken to generate and
rank the different vectorization plans can become quite large. As a result, we now
cap the number of candidate pairs within each instruction group. This is done by
closing out the group once the threshold is reached (set now at 3000 pairs).
Although this will limit the overall compile-time impact, this may not be the best
way to achieve this result. It might be better, for example, to prune excessive
candidate pairs after the fact the prevent the generation of short, but highly-connected
groups. We can experiment with this in the future.
This change reduces the overall compile-time slowdown of the csa.ll test case in
PR15222 to ~5x. If 5x is still considered too large, a lower limit can be
used as the default.
This represents a functionality change, but only for very large inputs
(thus, there is no regression test).
llvm-svn: 175251
to have it not named appropriately. Also in StopInfoMachException, we aren't testing for software or not software, just
whether the thing is a breakpoint we set. So don't use "software"...
llvm-svn: 175241
This just adds a very simple check that if a DerivedToBase CastExpr is
operating on a value with known C++ object type, and that type is not the
base type specified in the AST, then the cast is invalid and we should
return UnknownVal.
In the future, perhaps we can have a checker that specifies that this is
illegal, but we still shouldn't assert even if the user turns that checker
off.
PR14872
llvm-svn: 175239
not matter but makes it more gcc compatible which avoids possible subtle
problems. Also, turned back on a disabled check in helloworld.ll.
llvm-svn: 175237
...after a host of optimizations related to the use of LazyCompoundVals
(our implementation of aggregate binds).
Originally applied in r173951.
Reverted in r174069 because it was causing hangs.
Re-applied in r174212.
Reverted in r174265 because it was /still/ causing hangs.
If this needs to be reverted again it will be punted to far in the future.
llvm-svn: 175234
Previously, we were scanning the current store. Now, we properly scan the
store that the LazyCompoundVal came from, which may have very different
live symbols.
llvm-svn: 175232
Previously, whenever we had a LazyCompoundVal, we crawled through the
entire store snapshot looking for bindings within the LCV's region. Now, we
just ask for the subregion bindings of the lazy region and only visit those.
This is an optimization (so no test case), but it may allow us to clean up
more dead bindings than we were previously.
llvm-svn: 175230
This is going to be used in the next commit.
While I'm here, tighten up assumptions about symbolic offset
BindingKeys, and make offset calculation explicitly handle all
MemRegion kinds.
No functionality change.
llvm-svn: 175228
The SEL data formatter was working hard to ensure that pointers-to-selectors could be formatted by the same block of code. In that effort, we were taking the address-of a SEL.
This operation fails when the SEL lives in a register, and was causing problems.
The formatter has been fixed to work correctly without assuming &selector will be a valid object.
llvm-svn: 175227
- stop ignoring the error-codes in the 'error' variable
- removed out-of-bounds accesses with read-only array fields such as:
self.assertTrue(data2.uint8[6] == 0, 'binary 0 terminator')
Since SBData wraps a (6-character) python string literal, trying to read the
null-terminator raises an exception. Instead, I replaced the out-of-bounds
read with a length-check.
Other out-of-bounds reads (via accessor function like SBData.GetUnsignedInt8)
don't throw and are OK. I just added asserts that errors are set for these
negative cases.
llvm-svn: 175223
assembler should also accept a two arg form, as the docuemntation specifies that
the first (destination) register is optional.
This patch uses TwoOperandAliasConstraint to add the two argument form.
It also fixes an 80-column formatting problem in:
test/MC/ARM/neon-bitwise-encoding
<rdar://problem/12909419> Clang rejects ARM NEON assembly instructions
llvm-svn: 175221
1. Define and use function terminateSearch.
2. Use MachineBasicBlock::iterator instead of MachineBasicBlock::instr_iterator.
3. Delete the line which checks whether an instruction is a pseudo.
llvm-svn: 175219
All instances of std::multimap have now been replaced by
DenseMap<K, std::vector<V> >, and this yields a speedup of 5% on the
csa.ll test case from PR15222.
No functionality change intended.
llvm-svn: 175216
This is another commit on the road to removing std::multimap from
BBVectorize. This gives an ~1% speedup on the csa.ll test case
in PR15222.
No functionality change intended.
llvm-svn: 175215
The purpose of this change is to simplify creating non-atom sections.
Previously _contentType, _sectionKind and _order were used for multiple
purposes and collided in places. This moves all of the Atom specific logic down
into AtomSection and makes Section just have raw Elf_Shdr flags.
llvm-svn: 175207
This patch doesn't introduce any functionality changes.
It adds some new fields to the Hexagon instruction classes and
changes their layout to support instruction encoding.
llvm-svn: 175205
Alexey Samsonov