Ideally only those transform passes that run at -O0 remain enabled,
in reality we get as close as we reasonably can.
Passes are responsible for disabling themselves, it's not the job of
the pass manager to do it for them.
llvm-svn: 200892
inconsistent results for different orderings of alloca slices. The
fundamental issue is that it is just always a mistake to return early
from this function. There is no effective early exit to leverage. This
patch stops trynig to do so and simplifies the code a bit as
a consequence.
Original diagnosis and patch by James Molloy with some name tweaks by me
in part reflecting feedback from Duncan Smith on the mailing list.
llvm-svn: 199771
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
can be used by both the new pass manager and the old.
This removes it from any of the virtual mess of the pass interfaces and
lets it derive cleanly from the DominatorTreeBase<> template. In turn,
tons of boilerplate interface can be nuked and it turns into a very
straightforward extension of the base DominatorTree interface.
The old analysis pass is now a simple wrapper. The names and style of
this split should match the split between CallGraph and
CallGraphWrapperPass. All of the users of DominatorTree have been
updated to match using many of the same tricks as with CallGraph. The
goal is that the common type remains the resulting DominatorTree rather
than the pass. This will make subsequent work toward the new pass
manager significantly easier.
Also in numerous places things became cleaner because I switched from
re-running the pass (!!! mid way through some other passes run!!!) to
directly recomputing the domtree.
llvm-svn: 199104
directory. These passes are already defined in the IR library, and it
doesn't make any sense to have the headers in Analysis.
Long term, I think there is going to be a much better way to divide
these matters. The dominators code should be fully separated into the
abstract graph algorithm and have that put in Support where it becomes
obvious that evn Clang's CFGBlock's can use it. Then the verifier can
manually construct dominance information from the Support-driven
interface while the Analysis library can provide a pass which both
caches, reconstructs, and supports a nice update API.
But those are very long term, and so I don't want to leave the really
confusing structure until that day arrives.
llvm-svn: 199082
All other uses of this macro in LLVM/clang have been moved to the function
definition so follow suite (and the usage advice) here too for consistency.
llvm-svn: 198516
The motivation is to mark dump methods as used in debug builds so that they can
be called from lldb, but to not do so in release builds so that they can be
dead-stripped.
There's lots of potential follow-up work suggested in the thread
"Should dump methods be LLVM_ATTRIBUTE_USED only in debug builds?" on cfe-dev,
but everyone seems to agreen on this subset.
Macro name chosen by fair coin toss.
llvm-svn: 198456
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 wants to convert any types of equivalent widths but it's not possible to
convert vectors of pointers to an integer scalar with a single cast. As a
workaround we add a bitcast to the corresponding int ptr type first. This type
of cast used to be an edge case but has become common with SLP vectorization.
Fixes PR17271.
llvm-svn: 191143
However, opt -O2 doesn't run mem2reg directly so nobody noticed until r188146
when SROA started sending more things directly down the PromoteMemToReg path.
In order to revert r187191, I also revert dependent revisions r187296, r187322
and r188146. Fixes PR16867. Does not add the testcases from that PR, but both
of them should get added for both mem2reg and sroa when this revert gets
unreverted.
llvm-svn: 188327
SROA-based analysis has enough information. This should work now that
both mem2reg *and* the SSAUpdater-based AllocaPromoter have been updated
to be able to promote the types of allocas that the SROA analysis
detects.
I've included tests for the AllocaPromoter that were only possible to
write once we fast-tracked promotable allocas without rewriting them.
This includes a test both for r187347 and r188145.
Original commit log for r187323:
"""
Now that mem2reg understands how to cope with a slightly wider set of uses of
an alloca, we can pre-compute promotability while analyzing an alloca for
splitting in SROA. That lets us short-circuit the common case of a bunch of
trivially promotable allocas. This cuts 20% to 30% off the run time of SROA for
typical frontend-generated IR sequneces I'm seeing. It gets the new SROA to
within 20% of ScalarRepl for such code. My current benchmark for these numbers
is PR15412, but it fits the general pattern of IR emitted by Clang so it should
be widely applicable.
"""
llvm-svn: 188146
the more general set of patterns that are now handled by mem2reg and that we
can detect quickly while doing SROA's initial analysis. Notably, this allows it
to promote through no-op bitcast and GEP sequences. A core part of the
SSAUpdater approach is the ability to test whether a particular instruction is
part of the set being promoted. Testing this becomes significantly more complex
in the world where the operand to every load and store isn't the alloca itself.
I ended up using the approach of walking up the def-chain until we find the
alloca. I benchmarked this against keeping a set of pointer operands and
keeping a set of the loads and stores we care about, and this one seemed faster
although the difference was very small.
No test case yet because currently the rewriting always "fixes" the inputs to
not require this. The next patch which re-enables early promotion of easy cases
in SROA will include a test case that specifically exercises this aspect of the
alloca promoter.
llvm-svn: 188145
our visiting datastructures in the AllocaPromoter/SSAUpdater path of
SROA. Also shift the order if clears around to be more consistent.
No functionality changed here, this is just a cleanup.
llvm-svn: 188144
infrastructure to do promotion without a domtree the same smarts about
looking through GEPs, bitcasts, etc., that I just taught mem2reg about.
This way, if SROA chooses to promote an alloca which still has some
noisy instructions this code can cope with them.
I've not used as principled of an approach here for two reasons:
1) This code doesn't really need it as we were already set up to zip
through the instructions used by the alloca.
2) I view the code here as more of a hack, and hopefully a temporary one.
The SSAUpdater path in SROA is a real sore point for me. It doesn't make
a lot of architectural sense for many reasons:
- We're likely to end up needing the domtree anyways in a subsequent
pass, so why not compute it earlier and use it.
- In the future we'll likely end up needing the domtree for parts of the
inliner itself.
- If we need to we could teach the inliner to preserve the domtree. Part
of the re-work of the pass manager will allow this to be very powerful
even in large SCCs with many functions.
- Ultimately, computing a domtree has gotten significantly faster since
the original SSAUpdater-using code went into ScalarRepl. We no longer
use domfrontiers, and much of domtree is lazily done based on queries
rather than eagerly.
- At this point keeping the SSAUpdater-based promotion saves a total of
0.7% on a build of the 'opt' tool for me. That's not a lot of
performance given the complexity!
So I'm leaving this a bit ugly in the hope that eventually we just
remove all of this nonsense.
I can't even readily test this because this code isn't reachable except
through SROA. When I re-instate the patch that fast-tracks allocas
already suitable for promotion, I'll add a testcase there that failed
before this change. Before that, SROA will fix any test case I give it.
llvm-svn: 187347
uses of an alloca, we can pre-compute promotability while analyzing an
alloca for splitting in SROA. That lets us short-circuit the common case
of a bunch of trivially promotable allocas. This cuts 20% to 30% off the
run time of SROA for typical frontend-generated IR sequneces I'm seeing.
It gets the new SROA to within 20% of ScalarRepl for such code. My
current benchmark for these numbers is PR15412, but it fits the general
pattern of IR emitted by Clang so it should be widely applicable.
llvm-svn: 187323
their being optimized out in debug mode. Realistically, this just isn't
going to be the slow part anyways. This also fixes unused variable
warnings that are breaking LLD build bots. =/ I didn't see these at
first, and kept losing track of the fact that they were broken.
llvm-svn: 187297
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
implementation of the SROA algorithm. We were using the term 'partition'
in many places that no longer ever represented an actual partition, but
rather just an arbitrary slice of an alloca.
No functionality change intended here. Mostly just renaming of types,
functions, variables, and rewording of comments. Several comments were
rewritten to make a lot more sense in the new structure of things.
The stats are still weird and not reflective of how this really works.
I'll fix those up in a separate patch as it is a touch more semantic of
a change...
llvm-svn: 186659
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
When a 1-element vector alloca is promoted, a store instruction can often be
rewritten without converting the value to a scalar and using an insertelement
instruction to stuff it into the new alloca. This patch just adds a check
to skip that conversion when it is unnecessary. This turns out to be really
important for some ARM Neon operations where <1 x i64> is used to get around
the fact that i64 is not a legal type.
llvm-svn: 184870
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 key part of this is ensuring that name prefixes remain in a Twine
form until we get to a point where we can nuke them under NDEBUG. This
is tricky using the old APIs as they played fast and loose with Twine,
which is prone to serious error. The inserter is much cleaner as it is
actually in the call stack leading to the setName call, and so has
a good opportunity to prepend the prefix.
This matters more than you might imagine because most runs over an
alloca find a single partition, and rewrite 3 or 4 instructions
referring to it. As a consequence doing this lazily and exclusively with
Twine allows the optimizer to delete more of it and shaves another 2% to
3% off of the release build's SROA run time for PR15412. I also think
the APIs are cleaner, and the use of Twine is more reliable, so
I consider it a win-win despite the churn required to reach this state.
llvm-svn: 177631
This is espcially important because the new SROA pass goes to great
lengths to provide helpful names for debugging, and as a consequence
they can become very slow to render.
Good for between 5% and 15% of the SROA runtime on some slow test cases
such as the one in PR15412.
llvm-svn: 177495
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
Rafael Espindola