We used to call Verify before adding DICompileUnit to the list, and now we
remove the check and always add DICompileUnit to the list in DebugInfoFinder,
so we can verify them later on.
llvm-svn: 187237
robust. It now uses an InstVisitor and worklist to actually walk the
uses of the Alloca transitively and detect the pattern which we can
directly promote: loads & stores of the whole alloca and instructions we
can completely ignore.
Also, with this new implementation teach both the predicate for testing
whether we can promote and the promotion engine itself to use the same
code so we no longer have strange divergence between the two code paths.
I've added some silly test cases to demonstrate that we can handle
slightly more degenerate code patterns now. See the below for why this
is even interesting.
Performance impact: roughly 1% regression in the performance of SROA or
ScalarRepl on a large C++-ish test case where most of the allocas are
basically ready for promotion. The reason is because of silly redundant
work that I've left FIXMEs for and which I'll address in the next
commit. I wanted to separate this commit as it changes the behavior.
Once the redundant work in removing the dead uses of the alloca is
fixed, this code appears to be faster than the old version. =]
So why is this useful? Because the previous requirement for promotion
required a *specific* visit pattern of the uses of the alloca to verify:
we *had* to look for no more than 1 intervening use. The end goal is to
have SROA automatically detect when an alloca is already promotable and
directly hand it to the mem2reg machinery rather than trying to
partition and rewrite it. This is a 25% or more performance improvement
for SROA, and a significant chunk of the delta between it and
ScalarRepl. To get there, we need to make mem2reg actually capable of
promoting allocas which *look* promotable to SROA without have SROA do
tons of work to massage the code into just the right form.
This is actually the tip of the iceberg. There are tremendous potential
savings we can realize here by de-duplicating work between mem2reg and
SROA.
llvm-svn: 187191
Also avoid locals evicting locals just because they want a cheaper register.
Problem: MI Sched knows exactly how many registers we have and assumes
they can be colored. In cases where we have large blocks, usually from
unrolled loops, greedy coloring fails. This is a source of
"regressions" from the MI Scheduler on x86. I noticed this issue on
x86 where we have long chains of two-address defs in the same live
range. It's easy to see this in matrix multiplication benchmarks like
IRSmk and even the unit test misched-matmul.ll.
A fundamental difference between the LLVM register allocator and
conventional graph coloring is that in our model a live range can't
discover its neighbors, it can only verify its neighbors. That's why
we initially went for greedy coloring and added eviction to deal with
the hard cases. However, for singly defined and two-address live
ranges, we can optimally color without visiting neighbors simply by
processing the live ranges in instruction order.
Other beneficial side effects:
It is much easier to understand and debug regalloc for large blocks
when the live ranges are allocated in order. Yes, global allocation is
still very confusing, but it's nice to be able to comprehend what
happened locally.
Heuristics could be added to bias register assignment based on
instruction locality (think late register pairing, banks...).
Intuituvely this will make some test cases that are on the threshold
of register pressure more stable.
llvm-svn: 187139
Make sure the context and type fields are MDNodes. We will generate
verification errors if those fields are non-empty strings.
Fix testing cases to make them pass the verifier.
llvm-svn: 187106
The language reference says that:
"If a symbol appears in the @llvm.used list, then the compiler,
assembler, and linker are required to treat the symbol as if there is
a reference to the symbol that it cannot see"
Since even the linker cannot see the reference, we must assume that
the reference can be using the symbol table. For example, a user can add
__attribute__((used)) to a debug helper function like dump and use it from
a debugger.
llvm-svn: 187103
schedule an alloca for another iteration in SROA. This only showed up
with a mixture of promotable and unpromotable selects and phis. Added
a test case for this.
llvm-svn: 187031
pending speculation for a phi node. The problem here is that we were
using growth of the specluation set as an indicator of whether
speculation would occur, and if the phi node is already in the set we
don't see it grow. This is a symptom of the fact that this signal is
a total hack.
Unfortunately, I couldn't really come up with a non-hacky way of
signaling that promotion remains valid *after* speculation occurs, such
that we only speculate when all else looks good for promotion. In the
end, I went with at least a much more explicit approach of doing the
work of queuing inside the phi and select processing and setting
a preposterously named flag to convey that we're in the special state of
requiring speculating before promotion.
Thanks to Richard Trieu and Nick Lewycky for the excellent work reducing
a testcase for this from a pretty giant, nasty assert in a big
application. =] The testcase was excellent.
llvm-svn: 187029
MDNodes used by DbgDeclareInst and DbgValueInst.
Another 16 testing cases failed and they are disabled with
-disable-debug-info-verifier.
A total of 34 cases are disabled with -disable-debug-info-verifier and will be
corrected.
llvm-svn: 186902
GlobalOpt simplifies llvm.compiler.used by removing any members that are also
in the more strict llvm.used. Handle the special case where llvm.compiler.used
becomes empty.
llvm-svn: 186778
We were incorrectly using compiler_used instead of compiler.used. Unfortunately
the passes using the broken name had tests also using the broken name.
llvm-svn: 186705
SROA.
The crux of the issue is that now we track uses of a partition of the
alloca in two places: the iterators over the partitioning uses and the
previously collected split uses vector. We weren't accounting for the
fact that the split uses might invalidate integer widening in ways other
than due to their width (in this case due to being volatile).
Further reduced testcase added to the tests.
llvm-svn: 186655
end of a vector. This was found with ASan. I've had one other report of
a crasher, but thus far been unable to reproduce the crash. It may well
be fixed with this version, and if not I'd like to get more information
from the build bots about what is happening.
See r186316 for the full commit log for the new implementation of the
SROA algorithm.
llvm-svn: 186565
Duncan pointed out a mistake in my fix in r186425 when only one of the allocas
being compared had the target-default alignment. This is essentially his
suggested solution. Thanks!
llvm-svn: 186510
For safety, the inliner cannot decrease the allignment on an alloca when
merging it with another.
I've included two variants of the test case for this: one with DataLayout
available, and one without. When DataLayout is not available, if only one of
the allocas uses the default alignment (getAlignment() == 0), then they cannot
be safely merged.
llvm-svn: 186425
a bot.
This reverts the commit which introduced a new implementation of the
fancy SROA pass designed to reduce its overhead. I'll skip the huge
commit log here, refer to r186316 if you're looking for how this all
works and why it works that way.
llvm-svn: 186332
different core implementation strategy.
Previously, SROA would build a relatively elaborate partitioning of an
alloca, associate uses with each partition, and then rewrite the uses of
each partition in an attempt to break apart the alloca into chunks that
could be promoted. This was very wasteful in terms of memory and compile
time because regardless of how complex the alloca or how much we're able
to do in breaking it up, all of the datastructure work to analyze the
partitioning was done up front.
The new implementation attempts to form partitions of the alloca lazily
and on the fly, rewriting the uses that make up that partition as it
goes. This has a few significant effects:
1) Much simpler data structures are used throughout.
2) No more double walk of the recursive use graph of the alloca, only
walk it once.
3) No more complex algorithms for associating a particular use with
a particular partition.
4) PHI and Select speculation is simplified and happens lazily.
5) More precise information is available about a specific use of the
alloca, removing the need for some side datastructures.
Ultimately, I think this is a much better implementation. It removes
about 300 lines of code, but arguably removes more like 500 considering
that some code grew in the process of being factored apart and cleaned
up for this all to work.
I've re-used as much of the old implementation as possible, which
includes the lion's share of code in the form of the rewriting logic.
The interesting new logic centers around how the uses of a partition are
sorted, and split into actual partitions.
Each instruction using a pointer derived from the alloca gets
a 'Partition' entry. This name is totally wrong, but I'll do a rename in
a follow-up commit as there is already enough churn here. The entry
describes the offset range accessed and the nature of the access. Once
we have all of these entries we sort them in a very specific way:
increasing order of begin offset, followed by whether they are
splittable uses (memcpy, etc), followed by the end offset or whatever.
Sorting by splittability is important as it simplifies the collection of
uses into a partition.
Once we have these uses sorted, we walk from the beginning to the end
building up a range of uses that form a partition of the alloca.
Overlapping unsplittable uses are merged into a single partition while
splittable uses are broken apart and carried from one partition to the
next. A partition is also introduced to bridge splittable uses between
the unsplittable regions when necessary.
I've looked at the performance PRs fairly closely. PR15471 no longer
will even load (the module is invalid). Not sure what is up there.
PR15412 improves by between 5% and 10%, however it is nearly impossible
to know what is holding it up as SROA (the entire pass) takes less time
than reading the IR for that test case. The analysis takes the same time
as running mem2reg on the final allocas. I suspect (without much
evidence) that the new implementation will scale much better however,
and it is just the small nature of the test cases that makes the changes
small and noisy. Either way, it is still simpler and cleaner I think.
llvm-svn: 186316
This conversion was done with the following bash script:
find test/Transforms -name "*.ll" | \
while read NAME; do
echo "$NAME"
if ! grep -q "^; *RUN: *llc" $NAME; then
TEMP=`mktemp -t temp`
cp $NAME $TEMP
sed -n "s/^define [^@]*@\([A-Za-z0-9_]*\)(.*$/\1/p" < $NAME | \
while read FUNC; do
sed -i '' "s/;\(.*\)\([A-Za-z0-9_]*\):\( *\)define\([^@]*\)@$FUNC\([( ]*\)\$/;\1\2-LABEL:\3define\4@$FUNC(/g" $TEMP
done
mv $TEMP $NAME
fi
done
llvm-svn: 186269
This update was done with the following bash script:
find test/Transforms -name "*.ll" | \
while read NAME; do
echo "$NAME"
if ! grep -q "^; *RUN: *llc" $NAME; then
TEMP=`mktemp -t temp`
cp $NAME $TEMP
sed -n "s/^define [^@]*@\([A-Za-z0-9_]*\)(.*$/\1/p" < $NAME | \
while read FUNC; do
sed -i '' "s/;\(.*\)\([A-Za-z0-9_]*\):\( *\)@$FUNC\([( ]*\)\$/;\1\2-LABEL:\3@$FUNC(/g" $TEMP
done
mv $TEMP $NAME
fi
done
llvm-svn: 186268
If an outside loop user of the reduction value uses the header phi node we
cannot just reduce the vectorized phi value in the vector code epilog because
we would loose VF-1 reductions.
lp:
p = phi (0, lv)
lv = lv + 1
...
brcond , lp, outside
outside:
usr = add 0, p
(Say the loop iterates two times, the value of p coming out of the loop is one).
We cannot just transform this to:
vlp:
p = phi (<0,0>, lv)
lv = lv + <1,1>
..
brcond , lp, outside
outside:
p_reduced = p[0] + [1];
usr = add 0, p_reduced
(Because the original loop iterated two times the vectorized loop would iterate
one time, but p_reduced ends up being zero instead of one).
We would have to execute VF-1 iterations in the scalar remainder loop in such
cases. For now, just disable vectorization.
PR16522
llvm-svn: 186256
In general, one should always complete CFG modifications first, update
CFG-based analyses, like Dominatores and LoopInfo, then generate
instruction sequences.
LoopVectorizer was creating a new loop, calling SCEVExpander to
generate checks, then updating LoopInfo. I just changed the order.
llvm-svn: 186241
Fixes a 35% degradation compared to unvectorized code in
MiBench/automotive-susan and an equally serious regression on a private
image processing benchmark.
radar://14351991
llvm-svn: 186188