When the trip count is -1, getSmallConstantTripMultiple could return zero,
and this would cause runtime loop unrolling to assert. Instead of returning
zero, one is now returned (consistent with the existing overflow cases).
Fixes PR14167.
llvm-svn: 166612
loads. It's not really profitable and may result in GVN going into an infinite
loop when it hits constructs like this:
%x = gep %some.type %x, ...
Found via an LTO build of LLVM.
llvm-svn: 166490
%V = mul i64 %N, 4
%t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
into
%t1 = getelementptr i32* %arr, i32 %N
%t = bitcast i32* %t1 to i8*
incorporating the multiplication into the getelementptr.
This happens all the time in dragonegg, for example for
int foo(int *A, int N) {
return A[N];
}
because gcc turns this into byte pointer arithmetic before it hits the plugin:
D.1590_2 = (long unsigned int) N_1(D);
D.1591_3 = D.1590_2 * 4;
D.1592_5 = A_4(D) + D.1591_3;
D.1589_6 = *D.1592_5;
return D.1589_6;
The D.1592_5 line is a POINTER_PLUS_EXPR, which is turned into a getelementptr
on a bitcast of A_4 to i8*, so this becomes exactly the kind of IR that the
transform fires on.
An analogous transform (with no testcases!) already existed for bitcasts of
arrays, so I rewrote it to share code with this one.
llvm-svn: 166474
Unreachable blocks can have invalid instructions. For example,
jump threading can produce self-referential instructions in
unreachable blocks. Also, we should not be spending time
optimizing unreachable code. Fixes PR14133.
llvm-svn: 166423
very small but very important bugfix:
bool shouldExplore(Use *U) {
Value *V = U->get();
if (isa<CallInst>(V) || isa<InvokeInst>(V))
[...]
should have read:
bool shouldExplore(Use *U) {
Value *V = U->getUser();
if (isa<CallInst>(V) || isa<InvokeInst>(V))
Fixes PR14143!
llvm-svn: 166407
This is important for vectors of pointers because only DataLayout,
not the underlying vector type, knows how to calculate the size
of the pointers in the vector. Fixes PR14138.
llvm-svn: 166401
It passes all tests, produces better results than the old code but uses the
wrong pass, LoopDependenceAnalysis, which is old and unmaintained. "Why is it
still in tree?", you might ask. The answer is obviously: "To confuse developers."
Just swapping in the new dependency pass sends the pass manager into an infinte
loop, I'll try to figure out why tomorrow.
llvm-svn: 166399
Requires a lot less code and complexity on loop-idiom's side and the more
precise analysis can catch more cases, like the one I included as a test case.
This also fixes the edge-case miscompilation from PR9481. I'm not entirely
sure that all cases are handled that the old checks handled but LDA will
certainly become smarter in the future.
llvm-svn: 166390
We used a SCEV to detect that A[X] is consecutive. We assumed that X was
the induction variable. But X can be any expression that uses the induction
for example: X = i + 2;
llvm-svn: 166388
This is important for nested-loop reductions such as :
In the innermost loop, the induction variable does not start with zero:
for (i = 0 .. n)
for (j = 0 .. m)
sum += ...
llvm-svn: 166387
If the pointer is consecutive then it is safe to read and write. If the pointer is non-loop-consecutive then
it is unsafe to vectorize it because we may hit an ordering issue.
llvm-svn: 166371
This patch migrates the strcpy optimizations from the simplify-libcalls pass
into the instcombine library call simplifier. Note also that StrCpyChkOpt
has been updated with a few simplifications that were being done in the
simplify-libcalls version of StrCpyOpt, but not in the migrated implementation
of StrCpyOpt. There is no reason to overload StrCpyOpt with fortified and
regular simplifications in the new model since there is already a dedicated
simplifier for __strcpy_chk.
llvm-svn: 166198
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
An obfuscated splat is where the frontend poorly generates code for a splat
using several different shuffles to create the splat, i.e.,
%A = load <4 x float>* %in_ptr, align 16
%B = shufflevector <4 x float> %A, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 undef, i32 undef>
%C = shufflevector <4 x float> %B, <4 x float> %A, <4 x i32> <i32 0, i32 1, i32 4, i32 undef>
%D = shufflevector <4 x float> %C, <4 x float> %A, <4 x i32> <i32 0, i32 1, i32 2, i32 4>
llvm-svn: 166061
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
This patch migrates the strcmp and strncmp optimizations from the
simplify-libcalls pass into the instcombine library call simplifier.
llvm-svn: 165915
This patch migrates the strchr and strrchr optimizations from the
simplify-libcalls pass into the instcombine library call simplifier.
llvm-svn: 165875
This patch migrates the strcat and strncat optimizations from the
simplify-libcalls pass into the instcombine library call simplifier.
llvm-svn: 165874
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
DeadArgumentElimination pass can replace one LLVM function with another,
invalidating a pointer stored in debug info metadata entry for this function.
To fix this, we collect debug info descriptors for functions before
running a DeadArgumentElimination pass and "patch" pointers in metadata nodes
if we replace a function.
llvm-svn: 165490
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
We conservatively only check the first use to avoid walking long use chains.
This catches the common case of having both a load and a store to a pointer
supplied by a PHI node.
llvm-svn: 165232
cpyDest can be mutated in some cases, which would then cause a crash later if
indeed the memory was underaligned. This brought down several buildbots, so
I guess the underaligned case is much more common than I thought!
llvm-svn: 165228
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
was less aligned than the old. In the testcase this results in an overaligned
memset: the memset alignment was correct for the original memory but is too much
for the new memory. Fix this by either increasing the alignment of the new
memory or bailing out if that isn't possible. Should fix the gcc-4.7 self-host
buildbot failure.
llvm-svn: 165220
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
a memcpy to reflect that '0' has a different meaning when applied to
a load or store. Now we correctly use underaligned loads and stores for
the test case added.
llvm-svn: 165101
necessary during rewriting. As part of this, fix a real think-o here
where we might have left off an alignment specification when the address
is in fact underaligned. I haven't come up with any way to trigger this,
as there is always some other factor that reduces the alignment, but it
certainly might have been an observable bug in some way I can't think
of. This also slightly changes the strategy for placing explicit
alignments on loads and stores to only do so when the alignment does not
match that required by the ABI. This causes a few redundant alignments
to go away from test cases.
I've also added a couple of tests that really push on the alignment that
we end up with on loads and stores. More to come here as I try to fix an
underlying bug I have conjectured and produced test cases for, although
it's not clear if this bug is the one currently hitting dragonegg's
gcc47 bootstrap.
llvm-svn: 165100
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
alignment requirements of the new alloca. As one consequence which was
reported as a bug by Duncan, we overaligned memcpy calls to ranges of
allocas after they were rewritten to types with lower alignment
requirements. Other consquences are possible, but I don't have any test
cases for them.
llvm-svn: 164937
a pair of instructions, one for the used pointer and the second for the
user. This simplifies the representation and also makes it more dense.
This was noticed because of the miscompile in PR13926. In that case, we
were running up against a fundamental "bad idea" in the speculation of
PHI and select instructions: the speculation and rewriting are
interleaved, which requires phi speculation to also perform load
rewriting! This is bad, and causes us to miss opportunities to do (for
example) vector rewriting only exposed after PHI speculation, etc etc.
It also, in the old system, required us to insert *new* load uses into
the current partition's use list, which would then be ignored during
rewriting because we had already extracted an end iterator for the use
list. The appending behavior (and much of the other oddities) stem from
the strange de-duplication strategy in the PartitionUse builder.
Amusingly, all this went without notice for so long because it could
only be triggered by having *different* GEPs into the same partition of
the same alloca, where both different GEPs were operands of a single
PHI, and where the GEP which was not encountered first also had multiple
uses within that same PHI node... Hence the insane steps required to
reproduce.
So, step one in fixing this fundamental bad idea is to make the
PartitionUse actually contain a Use*, and to make the builder do proper
deduplication instead of funky de-duplication. This is enough to remove
the appending behavior, and fix the miscompile in PR13926, but there is
more work to be done here. Subsequent commits will lift the speculation
into its own visitor. It'll be a useful step toward potentially
extracting all of the speculation logic into a generic utility
transform.
The existing PHI test case for repeated operands has been made more
extreme to catch even these issues. This test case, run through the old
pass, will exactly reproduce the miscompile from PR13926. ;] We were so
close here!
llvm-svn: 164925
alignment could lose it due to the alloca type moving down to a much
smaller alignment guarantee.
Now SROA will actively compute a proper alignment, factoring the target
data, any explicit alignment, and the offset within the struct. This
will in some cases lower the alignment requirements, but when we lower
them below those of the type, we drop the alignment entirely to give
freedom to the code generator to align it however is convenient.
Thanks to Duncan for the lovely test case that pinned this down. =]
llvm-svn: 164891
If the width is very large it gets truncated from uint64_t to uint32_t when
passed to TD->fitsInLegalInteger. The truncated value can fit in a register.
This manifested in massive memory usage or crashes (PR13946).
llvm-svn: 164784
- Put statistics in alphabetical order
- Don't use getZextValue when building TableInt, just use APInts
- Introduce Create{Z,S}ExtOrTrunc in IRBuilder.
llvm-svn: 164696
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
tables in bitmaps when they fit in a target-legal register.
This saves some space, and it also allows for building tables that would
otherwise be deemed too sparse.
One interesting case that this hits is example 7 from
http://blog.regehr.org/archives/320. We currently generate good code
for this when lowering the switch to the selection DAG: we build a
bitmask to decide whether to jump to one block or the other. My patch
will result in the same bitmask, but it removes the need for the jump,
as the return value can just be retrieved from the mask.
llvm-svn: 164684
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
to chains or cycles between PHIs and/or selects. Also add a couple of
really nice test cases reduced from Kostya's reports in PR13905 and
PR13906. Both are fixed by this patch.
llvm-svn: 164596
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
selects with a constant condition. This resulted in the operands
remaining live through the SROA rewriter. Most of the time, this just
caused some dead allocas to persist and get zapped by later passes, but
in one case found by Joerg, it caused a crash when we tried to *promote*
the alloca despite it having this dead use. We already have the
mechanisms in place to handle this, just wire select up to them.
llvm-svn: 164427
We already have HoistThenElseCodeToIf, this patch implements
SinkThenElseCodeToEnd. When END block has only two predecessors and each
predecessor terminates with unconditional branches, we compare instructions in
IF and ELSE blocks backwards and check whether we can sink the common
instructions down.
rdar://12191395
llvm-svn: 164325
two variables where the first variable is returned and the second
ignored.
I don't think this occurs in practice (other passes should have cleaned
up the unused phi node), but it should still be handled correctly.
Also make the logic for determining if we should return early less
sketchy.
llvm-svn: 164225
This is a follow-up from r163302, which added a transformation to
SimplifyCFG that turns some switches into loads from lookup tables.
It was pointed out that some targets, such as GPUs and deeply embedded
targets, might not find this appropriate, but SimplifyCFG doesn't have
enough information about the target to decide this.
This patch adds the reverse transformation to CodeGenPrep: it turns
loads from lookup tables back into switches for targets where we do not
build jump tables (assuming these are also the targets where lookup
tables are inappropriate).
Hopefully we will eventually get to have target information in
SimplifyCFG, and then this CodeGenPrep transformation can be removed.
llvm-svn: 164206
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
FCAs. This is essential in order to promote allocas that are used in
struct returns by frontends like Clang. The FCA load would block the
rest of the pass from firing, resulting is significant regressions with
the bullet benchmark in the nightly test suite.
Thanks to Duncan for repeated discussions about how best to do this, and
to both him and Benjamin for review.
This appears to have blocked many places where the pass tries to fire,
and so I'm expect somewhat different results with this fix added.
As with the last big patch, I'm including a change to enable the SROA by
default *temporarily*. Ben is going to remove this as soon as the LNT
bots pick up the patch. I'm just trying to get a round of LNT numbers
from the stable machines in the lab.
NOTE: Four clang tests are expected to fail in the brief window where
this is enabled. Sorry for the noise!
llvm-svn: 164119
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
destination.
Updated previous implementation to fix a case not covered:
// PBI: br i1 %x, TrueDest, BB
// BI: br i1 %y, TrueDest, FalseDest
The other case was handled correctly.
// PBI: br i1 %x, BB, FalseDest
// BI: br i1 %y, TrueDest, FalseDest
Also tried to use 64-bit arithmetic instead of APInt with scale to simplify the
computation. Let me know if you have other opinions about this.
llvm-svn: 163954
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
a pair of switch/branch where both depend on the value of the same variable and
the default case of the first switch/branch goes to the second switch/branch.
Code clean up and fixed a few issues:
1> handling the case where some cases of the 2nd switch are invalidated
2> correctly calculate the weight for the 2nd switch when it is a conditional eq
Testing case is modified from Alastair's original patch.
llvm-svn: 163635
The lookup tables did not get built in a deterministic order.
This makes them get built in the order that the corresponding phi nodes
were found.
llvm-svn: 163305
This adds a transformation to SimplifyCFG that attemps to turn switch
instructions into loads from lookup tables. It works on switches that
are only used to initialize one or more phi nodes in a common successor
basic block, for example:
int f(int x) {
switch (x) {
case 0: return 5;
case 1: return 4;
case 2: return -2;
case 5: return 7;
case 6: return 9;
default: return 42;
}
This speeds up the code by removing the hard-to-predict jump, and
reduces code size by removing the code for the jump targets.
llvm-svn: 163302
switch, make sure we include the value for the cases when calculating edge
value from switch to the default destination.
rdar://12241132
llvm-svn: 163270
pointers-to-strong-pointers may be in play. These can lead to retains and
releases happening in unstructured ways, foiling the optimizer. This fixes
rdar://12150909.
llvm-svn: 163180
Scan the body of the loop and find instructions that may trap.
Use this information when deciding if it is safe to hoist or sink instructions.
Notice that we can optimize the search of instructions that may throw in the case of nested loops.
rdar://11518836
llvm-svn: 163132
This code used to only handle malloc-like calls, which do not read memory.
r158919 changed it to check isNoAliasFn(), which includes strdup-like and
realloc-like calls, but it was not checking for dependencies on the memory
read by those calls.
llvm-svn: 163106
We update until we hit a fixpoint. This is probably slow but also
slightly simplifies the code. It should also fix the occasional
invalid domtrees observed when building with expensive checking.
I couldn't find a case where this had a measurable slowdown, but
if someone finds a pathological case where it does we may have
to find a cleverer way of updating dominators here.
Thanks to Duncan for the test case.
llvm-svn: 163091
The old PHI updating code in loop-rotate was replaced with SSAUpdater a while
ago, it has no problems with comples PHIs. What had to be fixed is detecting
whether a loop was already rotated and updating dominators when multiple exits
were present.
This change increases overall code size a bit, mostly due to additional loop
unrolling opportunities. Passes test-suite and selfhost with -verify-dom-info.
Fixes PR7447.
Thanks to Andy for the input on the domtree updating code.
llvm-svn: 162912
This disables malloc-specific optimization when -fno-builtin (or -ffreestanding)
is specified. This has been a problem for a long time but became more severe
with the recent memory builtin improvements.
Since the memory builtin functions are used everywhere, this required passing
TLI in many places. This means that functions that now have an optional TLI
argument, like RecursivelyDeleteTriviallyDeadFunctions, won't remove dead
mallocs anymore if the TLI argument is missing. I've updated most passes to do
the right thing.
Fixes PR13694 and probably others.
llvm-svn: 162841
This optimization is really just replacing allocas wholesale with
globals, there is no scalarization.
The underlying motivation for this patch is to simplify the SROA pass
and focus it on splitting and promoting allocas.
llvm-svn: 162271
The previous fix only checked for simple cycles, use a set to catch longer
cycles too.
Drop the broken check from the ObjectSizeOffsetEvaluator. The BoundsChecking
pass doesn't have to deal with invalid IR like InstCombine does.
llvm-svn: 162120
I really need to find a way to automate this, but I can't come up with a regex
that has no false positives while handling tricky cases like custom check
prefixes.
llvm-svn: 162097
where some fact lake a=b dominates a use in a phi, but doesn't dominate the
basic block itself.
This feature could also be implemented by splitting critical edges, but at least
with the current algorithm reasoning about the dominance directly is faster.
The time for running "opt -O2" in the testcase in pr10584 is 1.003 times slower
and on gcc as a single file it is 1.0007 times faster.
llvm-svn: 162023
- memcpy size is wrongly truncated into 32-bit and treat 8GB memcpy is
0-sized memcpy
- as 0-sized memcpy/memset is already removed before SimplifyMemTransfer
and SimplifyMemSet in visitCallInst, replace 0 checking with
assertions.
- replace getZExtValue() with getLimitedValue() according to
Eli Friedman
llvm-svn: 161923
and allow some optimizations to turn conditional branches into unconditional.
This commit adds a simple control-flow optimization which merges two consecutive
basic blocks which are connected by a single edge. This allows the codegen to
operate on larger basic blocks.
rdar://11973998
llvm-svn: 161852
multiple scalar promotions on a single loop. This also has the effect of
preserving the order of stores sunk out of loops, which is aesthetically
pleasing, and it happens to fix the testcase in PR13542, though it doesn't
fix the underlying problem.
llvm-svn: 161459
An unsigned value converted to floating-point will always be greater than
a negative constant. Unfortunately InstCombine reversed the check so that
unsigned values were being optimized to always be greater than all positive
floating-point constants. <rdar://problem/12029145>
llvm-svn: 161452
We give a bonus for every argument because the argument setup is not needed
anymore when the function is inlined. With this patch we interpret byval
arguments as a compact representation of many arguments. The byval argument
setup is implemented in the backend as an inline memcpy, so to model the
cost as accurately as possible we take the number of pointer-sized elements
in the byval argument and give a bonus of 2 instructions for every one of
those. The bonus is capped at 8 elements, which is the number of stores
at which the x86 backend switches from an expanded inline memcpy to a real
memcpy. It would be better to use the real memcpy threshold from the backend,
but it's not available via TargetData.
This change brings the performance of c-ray in line with gcc 4.7. The included
test case tries to reproduce the c-ray problem to catch regressions for this
benchmark early, its performance is dominated by the inline decision of a
specific call.
This only has a small impact on most code, more on x86 and arm than on x86_64
due to the way the ABI works. When building LLVM for x86 it gives a small
inline cost boost to virtually any function using StringRef or STL allocators,
but only a 0.01% increase in overall binary size. The size of gcc compiled by
clang actually shrunk by a couple bytes with this patch applied, but not
significantly.
llvm-svn: 161413
instsimplify+inline strategy.
The crux of the problem is that instsimplify was reasonably relying on
an invariant that is true within any single function, but is no longer
true mid-inline the way we use it. This invariant is that an argument
pointer != a local (alloca) pointer.
The fix is really light weight though, and allows instsimplify to be
resiliant to these situations: when checking the relation ships to
function arguments, ensure that the argumets come from the same
function. If they come from different functions, then none of these
assumptions hold. All credit to Benjamin Kramer for coming up with this
clever solution to the problem.
llvm-svn: 161410
Previously, MBP essentially aligned every branch target it could. This
bloats code quite a bit, especially non-looping code which has no real
reason to prefer aligned branch targets so heavily.
As Andy said in review, it's still a bit odd to do this without a real
cost model, but this at least has much more plausible heuristics.
Fixes PR13265.
llvm-svn: 161409
This can happen as long as the instruction is not reachable. Instcombine does generate these unreachable malformed selects when doing RAUW
llvm-svn: 160874
of an array element (rather than at the beginning of the element) and extended
into the next element, then the load from the second element was being handled
wrong due to incorrect updating of the notion of which byte to load next. This
fixes PR13442. Thanks to Chris Smowton for reporting the problem, analyzing it
and providing a fix.
llvm-svn: 160711
might be deliberate "one time" leaks, so that leak checkers can find them.
This is a reapply of r160602 with the fix that this time I'm committing the
code I thought I was committing last time; the I->eraseFromParent() goes
*after* the break out of the loop.
llvm-svn: 160664
r160529 that was subsequently reverted. The fix was to not call
GV->eraseFromParent() right before the caller does the same. The existing
testcases already caught this bug if run under valgrind.
llvm-svn: 160602
GetBestDestForJumpOnUndef() assumes there is at least 1 successor, which isn't
true if the block ends in an indirect branch with no successors. Fix this by
bailing out earlier in this case.
llvm-svn: 160546
Fixes PR13371: indvars pass incorrectly substitutes 'undef' values.
I do not like this fix. It's needed until/unless the meaning of undef
changes. It attempts to be complete according to the IR spec, but I
don't have much confidence in the implementation given the difficulty
testing undefined behavior. Worse, this invalidates some of my
hard-fought work on indvars and LSR to optimize pointer induction
variables. It results benchmark regressions, which I'll track
internally. On x86_64 no LTO I see:
-3% huffbench
-3% 400.perlbench
-8% fhourstones
My only suggestion for recovering is to change the meaning of
undef. If we could trust an arbitrary instruction to produce a some
real value that can be manipulated (e.g. incremented) according to
non-undef rules, then this case could be easily handled with SCEV.
llvm-svn: 160421
All SCEV expressions used by LSR formulae must be safe to
expand. i.e. they may not contain UDiv unless we can prove nonzero
denominator.
Fixes PR11356: LSR hoists UDiv.
llvm-svn: 160205
%shr = lshr i64 %key, 3
%0 = load i64* %val, align 8
%sub = add i64 %0, -1
%and = and i64 %sub, %shr
ret i64 %and
to:
%shr = lshr i64 %key, 3
%0 = load i64* %val, align 8
%sub = add i64 %0, 2305843009213693951
%and = and i64 %sub, %shr
ret i64 %and
The demanded bit optimization is actually a pessimization because add -1 would
be codegen'ed as a sub 1. Teach the demanded constant shrinking optimization
to check for negated constant to make sure it is actually reducing the width
of the constant.
rdar://11793464
llvm-svn: 160101
This patch removes ~70 lines in InstCombineLoadStoreAlloca.cpp and makes both functions a bit more aggressive than before :)
In theory, we can be more aggressive when removing an alloca than a malloc, because an alloca pointer should never escape, but we are not taking advantage of this anyway
llvm-svn: 159952
This means we can do cheap DSE for heap memory.
Nothing is done if the pointer excapes or has a load.
The churn in the tests is mostly due to objectsize, since we want to make sure we
don't delete the malloc call before evaluating the objectsize (otherwise it becomes -1/0)
llvm-svn: 159876
another mechanical change accomplished though the power of terrible Perl
scripts.
I have manually switched some "s to 's to make escaping simpler.
While I started this to fix tests that aren't run in all configurations,
the massive number of tests is due to a really frustrating fragility of
our testing infrastructure: things like 'grep -v', 'not grep', and
'expected failures' can mask broken tests all too easily.
Essentially, I'm deeply disturbed that I can change the testsuite so
radically without causing any change in results for most platforms. =/
llvm-svn: 159547
versions of Bash. In addition, I can back out the change to the lit
built-in shell test runner to support this.
This should fix the majority of fallout on Darwin, but I suspect there
will be a few straggling issues.
llvm-svn: 159544
This was done through the aid of a terrible Perl creation. I will not
paste any of the horrors here. Suffice to say, it require multiple
staged rounds of replacements, state carried between, and a few
nested-construct-parsing hacks that I'm not proud of. It happens, by
luck, to be able to deal with all the TCL-quoting patterns in evidence
in the LLVM test suite.
If anyone is maintaining large out-of-tree test trees, feel free to poke
me and I'll send you the steps I used to convert things, as well as
answer any painful questions etc. IRC works best for this type of thing
I find.
Once converted, switch the LLVM lit config to use ShTests the same as
Clang. In addition to being able to delete large amounts of Python code
from 'lit', this will also simplify the entire test suite and some of
lit's architecture.
Finally, the test suite runs 33% faster on Linux now. ;]
For my 16-hardware-thread (2x 4-core xeon e5520): 36s -> 24s
llvm-svn: 159525
