the default.
Based on the patch by Matt Arsenault, D1764!
I switched one place to use the more direct pointer type to compute the
desired address space, and I reworked the memcpy rewriting section to
reflect significant refactorings that this patch helped inspire.
Thanks to several of the folks who helped review and improve the patch
as well.
llvm-svn: 202247
to work independently for the slice side and the other side.
This allows us to only compute the minimum of the two when we actually
rewrite to a memcpy that needs to take the minimum, and preserve higher
alignment for one side or the other when rewriting to loads and stores.
This fix was inspired by seeing the result of some refactoring that
makes addrspace handling better.
llvm-svn: 202242
checking in SROA.
The primary change is to just rely on uge for checking that the offset
is within the allocation size. This removes the explicit checks against
isNegative which were terribly error prone (including the reversed logic
that led to PR18615) and prevented us from supporting stack allocations
larger than half the address space.... Ok, so maybe the latter isn't
*common* but it's a silly restriction to have.
Also, we used to try to support a PHI node which loaded from before the
start of the allocation if any of the loaded bytes were within the
allocation. This doesn't make any sense, we have never really supported
loading or storing *before* the allocation starts. The simplified logic
just doesn't care.
We continue to allow loading past the end of the allocation in part to
support cases where there is a PHI and some loads are larger than others
and the larger ones reach past the end of the allocation. We could solve
this a different and more conservative way, but I'm still somewhat
paranoid about this.
llvm-svn: 202224
ordering.
The fundamental problem that we're hitting here is that the use-def
chain ordering is *itself* not a stable thing to be relying on in the
rewriting for SROA. Further, we use a non-stable sort over the slices to
arrange them based on the section of the alloca they're operating on.
With a debugging STL implementation (or different implementations in
stage2 and stage3) this can cause stage2 != stage3.
The specific aspect of this problem fixed in this commit deals with the
rewriting and load-speculation around PHIs and Selects. This, like many
other aspects of the use-rewriting in SROA, is really part of the
"strong SSA-formation" that is doen by SROA where it works very hard to
canonicalize loads and stores in *just* the right way to satisfy the
needs of mem2reg[1]. When we have a select (or a PHI) with 2 uses of the
same alloca, we test that loads downstream of the select are
speculatable around it twice. If only one of the operands to the select
needs to be rewritten, then if we get lucky we rewrite that one first
and the select is immediately speculatable. This can cause the order of
operand visitation, and thus the order of slices to be rewritten, to
change an alloca from promotable to non-promotable and vice versa.
The fix is to defer all of the speculation until *after* the rewrite
phase is done. Once we've rewritten everything, we can accurately test
for whether speculation will work (once, instead of twice!) and the
order ceases to matter.
This also happens to simplify the other subtlety of speculation -- we
need to *not* speculate anything unless the result of speculating will
make the alloca fully promotable by mem2reg. I had a previous attempt at
simplifying this, but it was still pretty horrible.
There is actually already a *really* nice test case for this in
basictest.ll, but on multiple STL implementations and inputs, we just
got "lucky". Fortunately, the test case is very small and we can
essentially build it in exactly the opposite way to get reasonable
coverage in both directions even from normal STL implementations.
llvm-svn: 202092
intrinsics.
Reported on the list by Evan with a couple of attempts to fix, but it
took a while to dig down to the root cause. There are two overlapping
bugs here, both centering around the circumstance of discovering
a memcpy operand which is known to be completely outside the bounds of
the alloca.
First, we need to kill the *other* side of the memcpy if it was added to
this alloca. Otherwise we'll factor it into our slicing and try to
rewrite it even though we know for a fact that it is dead. This is made
more tricky because we can visit the sides in either order. So we have
to both kill the other side and skip instructions marked as dead. The
latter really should be goodness in every case, but here is a matter of
correctness.
Second, we need to actually remove the *uses* of the alloca by the
memcpy when queuing it for later deletion. Otherwise it may still be
using the alloca when we go to promote it (if the rewrite re-uses the
existing alloca instruction). Do this by factoring out the
use-clobbering used when for nixing a Phi argument and re-using it
across the operands of a to-be-deleted instruction.
llvm-svn: 199590
order of slices of the alloca which have exactly the same size and other
properties. This was found by a perniciously unstable sort
implementation used to flush out buggy uses of the algorithm.
The fundamental idea is that findCommonType should return the best
common type it can find across all of the slices in the range. There
were two bugs here previously:
1) We would accept an integer type smaller than a byte-width multiple,
and if there were different bit-width integer types, we would accept
the first one. This caused an actual failure in the testcase updated
here when the sort order changed.
2) If we found a bad combination of types or a non-load, non-store use
before an integer typed load or store we would bail, but if we found
the integere typed load or store, we would use it. The correct
behavior is to always use an integer typed operation which covers the
partition if one exists.
While a clever debugging sort algorithm found problem #1 in our existing
test cases, I have no useful test case ideas for #2. I spotted in by
inspection when looking at this code.
llvm-svn: 195118
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
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 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
This is an edge case that can happen if we modify a chain of multiple selects.
Update all operands in that case and remove the assert. PR15805.
llvm-svn: 179982
The fix for PR14972 in r177055 introduced a real think-o in the *store*
side, likely because I was much more focused on the load side. While we
can arbitrarily widen (or narrow) a loaded value, we can't arbitrarily
widen a value to be stored, as that changes the width of memory access!
Lock down the code path in the store rewriting which would do this to
only handle the intended circumstance.
All of the existing tests continue to pass, and I've added a test from
the PR.
llvm-svn: 178974
The fundamental problem is that SROA didn't allow for overly wide loads
where the bits past the end of the alloca were masked away and the load
was sufficiently aligned to ensure there is no risk of page fault, or
other trapping behavior. With such widened loads, SROA would delete the
load entirely rather than clamping it to the size of the alloca in order
to allow mem2reg to fire. This was exposed by a test case that neatly
arranged for GVN to run first, widening certain loads, followed by an
inline step, and then SROA which miscompiles the code. However, I see no
reason why this hasn't been plaguing us in other contexts. It seems
deeply broken.
Diagnosing all of the above took all of 10 minutes of debugging. The
really annoying aspect is that fixing this completely breaks the pass.
;] There was an implicit reliance on the fact that no loads or stores
extended past the alloca once we decided to rewrite them in the final
stage of SROA. This was used to encode information about whether the
loads and stores had been split across multiple partitions of the
original alloca. That required threading explicit tracking of whether
a *use* of a partition is split across multiple partitions.
Once that was done, another problem arose: we allowed splitting of
integer loads and stores iff they were loads and stores to the entire
alloca. This is a really arbitrary limitation, and splitting at least
some integer loads and stores is crucial to maximize promotion
opportunities. My first attempt was to start removing the restriction
entirely, but currently that does Very Bad Things by causing *many*
common alloca patterns to be fully decomposed into i8 operations and
lots of or-ing together to produce larger integers on demand. The code
bloat is terrifying. That is still the right end-goal, but substantial
work must be done to either merge partitions or ensure that small i8
values are eagerly merged in some other pass. Sadly, figuring all this
out took essentially all the time and effort here.
So the end result is that we allow splitting only when the load or store
at least covers the alloca. That ensures widened loads and stores don't
hurt SROA, and that we don't rampantly decompose operations more than we
have previously.
All of this was already fairly well tested, and so I've just updated the
tests to cover the wide load behavior. I can add a test that crafts the
pass ordering magic which caused the original PR, but that seems really
brittle and to provide little benefit. The fundamental problem is that
widened loads should Just Work.
llvm-svn: 177055
This was a silly oversight, we weren't pruning allocas which were used
by variable-length memory intrinsics from the set that could be widened
and promoted as integers. Fix that.
llvm-svn: 170353
When SROA was evaluating a mixture of i1 and i8 loads and stores, in
just a particular case, it would tickle a latent bug where we compared
bits to bytes rather than bits to bits. As a consequence of the latent
bug, we would allow integers through which were not byte-size multiples,
a situation the later rewriting code was never intended to handle.
In release builds this could trigger all manner of oddities, but the
reported issue in PR14548 was forming invalid bitcast instructions.
The only downside of this fix is that it makes it more clear that SROA
in its current form is not capable of handling mixed i1 and i8 loads and
stores. Sometimes with the previous code this would work by luck, but
usually it would crash, so I'm not terribly worried. I'll watch the LNT
numbers just to be sure.
llvm-svn: 169719
to properly handle the combinations of these with split integer loads
and stores. This essentially replaces Evan's r168227 by refactoring the
code in a different way, and trynig to mirror that refactoring in both
the load and store sides of the rewriting.
Generally speaking there was some really problematic duplicated code
here that led to poorly founded assumptions and then subtle bugs. Now
much of the code actually flows through and follows a more consistent
style and logical path. There is still a tiny bit of duplication on the
store side of things, but it is much less bad.
This also changes the logic to never re-use a load or store instruction
as that was simply too error prone in practice.
I've added a few tests (one a reduction of the one in Evan's original
patch, which happened to be the same as the report in PR14349). I'm
going to look at adding a few more tests for things I found and fixed in
passing (such as the volatile tests in the vectorizable predicate).
This patch has survived bootstrap, and modulo one bugfix survived
Duncan's test suite, but let me know if anything else explodes.
llvm-svn: 168346
smaller integer loads and stores.
The high-level motivation is that the frontend sometimes generates
a single whole-alloca integer load or store during ABI lowering of
splittable allocas. We need to be able to break this apart in order to
see the underlying elements and properly promote them to SSA values. The
hope is that this fixes some performance regressions on x86-32 with the
new SROA pass.
Unfortunately, this causes quite a bit of churn in the test cases, and
bloats some IR that comes out. When we see an alloca that consists soley
of bits and bytes being extracted and re-inserted, we now do some
splitting first, before building widened integer "bucket of bits"
representations. These are always well folded by instcombine however, so
this shouldn't actually result in missed opportunities.
If this splitting of all-integer allocas does cause problems (perhaps
due to smaller SSA values going into the RA), we could potentially go to
some extreme measures to only do this integer splitting trick when there
are non-integer component accesses of an alloca, but discovering this is
quite expensive: it adds yet another complete walk of the recursive use
tree of the alloca.
Either way, I will be watching build bots and LNT bots to see what
fallout there is here. If anyone gets x86-32 numbers before & after this
change, I would be very interested.
llvm-svn: 166662
a pointer. A very bad idea. Let's not do that. Fixes PR14105.
Note that this wasn't *that* glaring of an oversight. Originally, these
routines were only called on offsets within an alloca, which are
intrinsically positive. But over the evolution of the pass, they ended
up being called for arbitrary offsets, and things went downhill...
llvm-svn: 166095
includes extracting ints for copying elsewhere and inserting ints when
copying into the alloca. This should fix the CanSROA assertion coming
out of Clang's regression test suite.
llvm-svn: 165931
cases where we have partial integer loads and stores to an otherwise
promotable alloca to widen[1] those loads and stores to cover the entire
alloca and bitcast them into the appropriate type such that promotion
can proceed.
These partial loads and stores stem from an annoying confluence of ARM's
calling convention and ABI lowering and the FCA pre-splitting which
takes place in SROA. Clang lowers a { double, double } in-register
function argument as a [4 x i32] function argument to ensure it is
placed into integer 32-bit registers (a really unnerving implicit
contract between Clang and the ARM backend I would add). This results in
a FCA load of [4 x i32]* from the { double, double } alloca, and SROA
decomposes this into a sequence of i32 loads and stores. Inlining
proceeds, code gets folded, but at the end of the day, we still have i32
stores to the low and high halves of a double alloca. Widening these to
be i64 operations, and bitcasting them to double prior to loading or
storing allows promotion to proceed for these allocas.
I looked quite a bit changing the IR which Clang produces for this case
to be more friendly, but small changes seem unlikely to help. I think
the best representation we could use currently would be to pass 4 i32
arguments thereby avoiding any FCAs, but that would still require this
fix. It seems like it might eventually be nice to somehow encode the ABI
register selection choices outside of the parameter type system so that
the parameter can be a { double, double }, but the CC register
annotations indicate that this should be passed via 4 integer registers.
This patch does not address the second problem in PR14059, which is the
reverse: when a struct alloca is loaded as a *larger* single integer.
This patch also does not address some of the code quality issues with
the FCA-splitting. Those don't actually impede any optimizations really,
but they're on my list to clean up.
[1]: Pedantic footnote: for those concerned about memory model issues
here, this is safe. For the alloca to be promotable, it cannot escape or
have any use of its address that could allow these loads or stores to be
racing. Thus, widening is always safe.
llvm-svn: 165928
type coercion code, especially when targetting ARM. Things like [1
x i32] instead of i32 are very common there.
The goal of this logic is to ensure that when we are picking an alloca
type, we look through such wrapper aggregates and across any zero-length
aggregate elements to find the simplest type possible to form a type
partition.
This logic should (generally speaking) rarely fire. It only ends up
kicking in when an alloca is accessed using two different types (for
instance, i32 and float), and the underlying alloca type has wrapper
aggregates around it. I noticed a significant amount of this occurring
looking at stepanov_abstraction generated code for arm, and suspect it
happens elsewhere as well.
Note that this doesn't yet address truly heinous IR productions such as
PR14059 is concerning. Those result in mismatched *sizes* of types in
addition to mismatched access and alloca types.
llvm-svn: 165870
are in fact identity operations. We detect these and kill their
partitions so that even splitting is unaffected by them. This is
particularly important because Clang relies on emitting identity memcpy
operations for struct copies, and these fold away to constants very
often after inlining.
Fixes the last big performance FIXME I have on my plate.
llvm-svn: 165285
Currently, we re-visit allocas when something changes about the way they
might be *split* to allow better scalarization to take place. However,
we weren't handling the case when the *promotion* is what would change
the behavior of SROA. When an address derived from an alloca is stored
into another alloca, we consider the first to have escaped. If the
second is ever promoted to an SSA value, we will suddenly be able to run
the SROA pass on the first alloca.
This patch adds explicit support for this form if iteration. When we
detect a store of a pointer derived from an alloca, we flag the
underlying alloca for reprocessing after promotion. The logic works hard
to only do this when there is definitely going to be promotion and it
might remove impediments to the analysis of the alloca.
Thanks to Nick for the great test case and Benjamin for some sanity
check review.
llvm-svn: 165223
Sorry for this being broken so long. =/
As part of this, switch all of the existing tests to be Little Endian,
which is the behavior I was asserting in them anyways! Add in a new
big-endian test that checks the interesting behavior there.
Another part of this is to tighten the rules abotu when we perform the
full-integer promotion. This logic now rejects cases where there fully
promoted integer is a non-multiple-of-8 bitwidth or cases where the
loads or stores touch bits which are in the allocated space of the
alloca but are not loaded or stored when accessing the integer. Sadly,
these aren't really observable today as the rest of the pass will
already ensure the invariants hold. However, the latter situation is
likely to become a potential concern in the future.
Thanks to Benjamin and Duncan for early review of this patch. I'm still
looking into whether there are further endianness issues, please let me
know if anyone sees BE failures persisting past this.
llvm-svn: 165219
scheduled for processing on the worklist eventually gets deleted while
we are processing another alloca, fixing the original test case in
PR13990.
To facilitate this, add a remove_if helper to the SetVector abstraction.
It's not easy to use the standard abstractions for this because of the
specifics of SetVectors types and implementation.
Finally, a nice small test case is included. Thanks to Benjamin for the
fantastic reduced test case here! All I had to do was delete some empty
basic blocks!
llvm-svn: 165065
rewriter in SROA to carry a proper alignment. This involves
interrogating various sources of alignment, etc. This is a more complete
and principled fix to PR13920 as well as related bugs pointed out by Eli
in review and by inspection in the area.
Also by inspection fix the integer and vector promotion paths to create
aligned loads and stores. I still need to work up test cases for
these... Sorry for the delay, they were found purely by inspection.
llvm-svn: 164689
This should really, really fix PR13916. For real this time. The
underlying bug is... a bit more subtle than I had imagined.
The setup is a code pattern that leads to an @llvm.memcpy call with two
equal pointers to an alloca in the source and dest. Now, not any pattern
will do. The alloca needs to be formed just so, and both pointers should
be wrapped in different bitcasts etc. When this precise pattern hits,
a funny sequence of events transpires. First, we correctly detect the
potential for overlap, and correctly optimize the memcpy. The first
time. However, we do simplify the set of users of the alloca, and that
causes us to run the alloca back through the SROA pass in case there are
knock-on simplifications. At this point, a curious thing has happened.
If we happen to have an i8 alloca, we have direct i8 pointer values. So
we don't bother creating a cast, we rewrite the arguments to the memcpy
to dircetly refer to the alloca.
Now, in an unrelated area of the pass, we have clever logic which
ensures that when visiting each User of a particular pointer derived
from an alloca, we only visit that User once, and directly inspect all
of its operands which refer to that particular pointer value. However,
the mechanism used to detect memcpy's with the potential to overlap
relied upon getting visited once per *Use*, not once per *User*. This is
always true *unless* the same exact value is both source and dest. It
turns out that almost nothing actually produces that pattern though.
We can hand craft test cases that more directly test this behavior of
course, and those are included. Also, note that there is a significant
missed optimization here -- we prove in many cases that there is
a non-volatile memcpy call with identical source and dest addresses. We
shouldn't prevent splitting the alloca in that case, and in fact we
should just remove such memcpy calls eagerly. I'll address that in
a subsequent commit.
llvm-svn: 164669
only a missed optimization opportunity if the store is over-aligned, but a
miscompile if the store's new type has a higher natural alignment than the
memcpy did. Fixes PR13920!
llvm-svn: 164641
Chandler, it's not obvious that it's okay that this alloca gets into the list
twice to begin with. Please review and see whether this is the fix you really
want, but I wanted to get a fix checked in quickly.
llvm-svn: 164634
integer promotion analogous to vector promotion. When there is an
integer alloca being accessed both as its integer type and as a narrower
integer type, promote the narrower access to "insert" and "extract" the
smaller integer from the larger one, and make the integer alloca
a candidate for promotion.
In the new formulation, we don't care about target legal integer or use
thresholds to control things. Instead, we only perform this promotion to
an integer type which the frontend has already emitted a load or store
for. This bounds the scope and prevents optimization passes from
coalescing larger and larger entities into a single integer.
llvm-svn: 164479
across the uses of the alloca. It's entirely possible for negative
numbers to come up here, and in some rare cases simply doing the 2's
complement arithmetic isn't the correct decision. Notably, we can't zext
the index of the GEP. The definition of GEP is that these offsets are
sign extended or truncated to the size of the pointer, and then wrapping
2's complement arithmetic used.
This patch fixes an issue that comes up with *no* input from the
buildbots or bootstrap afaict. The only place where it manifested,
disturbingly, is Clang's own regression test suite. A reduced and
targeted collection of tests are added to cope with this. Note that I've
tried to pin down the potential cases of overflow, but may have missed
some cases. I've tried to add a few cases to test this, but its hard
because LLVM has quite limited support for >64bit constructs.
llvm-svn: 164475
from the dragonegg build bots when we turned on the full version of the
pass. Included a much reduced test case for this pesky bug, despite
bugpoint's uncooperative behavior.
Also, I audited all the similar code I could find and didn't spot any
other cases where this mistake cropped up.
llvm-svn: 164178
working on FCA splitting. Instead of refusing to form a common type when
there are uses of a subsection of the alloca as well as a use of the
entire alloca, just skip the subsection uses and continue looking for
a whole-alloca use with a type that we can use.
This produces slightly prettier IR I think, and also fixes the other
failure in the test.
llvm-svn: 164146
new one, and add support for running the new pass in that mode and in
that slot of the pass manager. With this the new pass can completely
replace the old one within the pipeline.
The strategy for enabling or disabling the SSAUpdater logic is to do it
by making the requirement of the domtree analysis optional. By default,
it is required and we get the standard mem2reg approach. This is usually
the desired strategy when run in stand-alone situations. Within the
CGSCC pass manager, we disable requiring of the domtree analysis and
consequentially trigger fallback to the SSAUpdater promotion.
In theory this would allow the pass to re-use a domtree if one happened
to be available even when run in a mode that doesn't require it. In
practice, it lets us have a single pass rather than two which was
simpler for me to wrap my head around.
There is a hidden flag to force the use of the SSAUpdater code path for
the purpose of testing. The primary testing strategy is just to run the
existing tests through that path. One notable difference is that it has
custom code to handle lifetime markers, and one of the tests has been
enhanced to exercise that code.
This has survived a bootstrap and the test suite without serious
correctness issues, however my run of the test suite produced *very*
alarming performance numbers. I don't entirely understand or trust them
though, so more investigation is on-going.
To aid my understanding of the performance impact of the new SROA now
that it runs throughout the optimization pipeline, I'm enabling it by
default in this commit, and will disable it again once the LNT bots have
picked up one iteration with it. I want to get those bots (which are
much more stable) to evaluate the impact of the change before I jump to
any conclusions.
NOTE: Several Clang tests will fail because they run -O3 and check the
result's order of output. They'll go back to passing once I disable it
again.
llvm-svn: 163965
This is essentially a ground up re-think of the SROA pass in LLVM. It
was initially inspired by a few problems with the existing pass:
- It is subject to the bane of my existence in optimizations: arbitrary
thresholds.
- It is overly conservative about which constructs can be split and
promoted.
- The vector value replacement aspect is separated from the splitting
logic, missing many opportunities where splitting and vector value
formation can work together.
- The splitting is entirely based around the underlying type of the
alloca, despite this type often having little to do with the reality
of how that memory is used. This is especially prevelant with unions
and base classes where we tail-pack derived members.
- When splitting fails (often due to the thresholds), the vector value
replacement (again because it is separate) can kick in for
preposterous cases where we simply should have split the value. This
results in forming i1024 and i2048 integer "bit vectors" that
tremendously slow down subsequnet IR optimizations (due to large
APInts) and impede the backend's lowering.
The new design takes an approach that fundamentally is not susceptible
to many of these problems. It is the result of a discusison between
myself and Duncan Sands over IRC about how to premptively avoid these
types of problems and how to do SROA in a more principled way. Since
then, it has evolved and grown, but this remains an important aspect: it
fixes real world problems with the SROA process today.
First, the transform of SROA actually has little to do with replacement.
It has more to do with splitting. The goal is to take an aggregate
alloca and form a composition of scalar allocas which can replace it and
will be most suitable to the eventual replacement by scalar SSA values.
The actual replacement is performed by mem2reg (and in the future
SSAUpdater).
The splitting is divided into four phases. The first phase is an
analysis of the uses of the alloca. This phase recursively walks uses,
building up a dense datastructure representing the ranges of the
alloca's memory actually used and checking for uses which inhibit any
aspects of the transform such as the escape of a pointer.
Once we have a mapping of the ranges of the alloca used by individual
operations, we compute a partitioning of the used ranges. Some uses are
inherently splittable (such as memcpy and memset), while scalar uses are
not splittable. The goal is to build a partitioning that has the minimum
number of splits while placing each unsplittable use in its own
partition. Overlapping unsplittable uses belong to the same partition.
This is the target split of the aggregate alloca, and it maximizes the
number of scalar accesses which become accesses to their own alloca and
candidates for promotion.
Third, we re-walk the uses of the alloca and assign each specific memory
access to all the partitions touched so that we have dense use-lists for
each partition.
Finally, we build a new, smaller alloca for each partition and rewrite
each use of that partition to use the new alloca. During this phase the
pass will also work very hard to transform uses of an alloca into a form
suitable for promotion, including forming vector operations, speculating
loads throguh PHI nodes and selects, etc.
After splitting is complete, each newly refined alloca that is
a candidate for promotion to a scalar SSA value is run through mem2reg.
There are lots of reasonably detailed comments in the source code about
the design and algorithms, and I'm going to be trying to improve them in
subsequent commits to ensure this is well documented, as the new pass is
in many ways more complex than the old one.
Some of this is still a WIP, but the current state is reasonbly stable.
It has passed bootstrap, the nightly test suite, and Duncan has run it
successfully through the ACATS and DragonEgg test suites. That said, it
remains behind a default-off flag until the last few pieces are in
place, and full testing can be done.
Specific areas I'm looking at next:
- Improved comments and some code cleanup from reviews.
- SSAUpdater and enabling this pass inside the CGSCC pass manager.
- Some datastructure tuning and compile-time measurements.
- More aggressive FCA splitting and vector formation.
Many thanks to Duncan Sands for the thorough final review, as well as
Benjamin Kramer for lots of review during the process of writing this
pass, and Daniel Berlin for reviewing the data structures and algorithms
and general theory of the pass. Also, several other people on IRC, over
lunch tables, etc for lots of feedback and advice.
llvm-svn: 163883
