Currently there is a bug in SROA::presplitLoadsAndStores which causes assertion in
GEPOperator::accumulateConstantOffset.
Basically it does not consider the situation that the pointer operand of load or store
may be in a non-zero address space and its size may be different from the size of
a pointer in address space 0.
This patch fixes assertion when compiling Blender Cycles kernels for amdgpu backend.
Diffferential Revision: https://reviews.llvm.org/D33298
llvm-svn: 305107
Summary:
As shown in the test case, SROA was crashing when trying to split
stores (to the alloca) of loads (from anywhere), because it assumed
the pointer operand to the loads and stores had to have the same
address space. This isn't the case. Make sure to use the correct
pointer type for both the load and the store.
Reviewed By: yaxunl
Differential Revision: https://reviews.llvm.org/D32593
llvm-svn: 304585
LLVM makes several assumptions about address space 0. However,
alloca is presently constrained to always return this address space.
There's no real way to avoid using alloca, so without this
there is no way to opt out of these assumptions.
The problematic assumptions include:
- That the pointer size used for the stack is the same size as
the code size pointer, which is also the maximum sized pointer.
- That 0 is an invalid, non-dereferencable pointer value.
These are problems for AMDGPU because alloca is used to
implement the private address space, which uses a 32-bit
index as the pointer value. Other pointers are 64-bit
and behave more like LLVM's notion of generic address
space. By changing the address space used for allocas,
we can change our generic pointer type to be LLVM's generic
pointer type which does have similar properties.
llvm-svn: 299888
so we can stop using DW_OP_bit_piece with the wrong semantics.
The entire back story can be found here:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20161114/405934.html
The gist is that in LLVM we've been misinterpreting DW_OP_bit_piece's
offset field to mean the offset into the source variable rather than
the offset into the location at the top the DWARF expression stack. In
order to be able to fix this in a subsequent patch, this patch
introduces a dedicated DW_OP_LLVM_fragment operation with the
semantics that we used to apply to DW_OP_bit_piece, which is what we
actually need while inside of LLVM. This patch is complete with a
bitcode upgrade for expressions using the old format. It does not yet
fix the DWARF backend to use DW_OP_bit_piece correctly.
Implementation note: We discussed several options for implementing
this, including reserving a dedicated field in DIExpression for the
fragment size and offset, but using an custom operator at the end of
the expression works just fine and is more efficient because we then
only pay for it when we need it.
Differential Revision: https://reviews.llvm.org/D27361
rdar://problem/29335809
llvm-svn: 288683
Preserving lifetime markers isn't as important as allowing promotion,
so just drop the lifetime markers if necessary.
This also fixes an assertion failure where other parts of SROA assumed
that lifetime markers never block promotion.
Fixes https://llvm.org/bugs/show_bug.cgi?id=29139.
Differential Revision: https://reviews.llvm.org/D24854
llvm-svn: 288074
SROA doesn't preserve the llvm.mem.parallel_loop_access metadata when it
transforms loads/stores. This patch fixes a couple occurences of this
issue.
(Partially addresses PR28981).
Differential Revision: https://reviews.llvm.org/D23549
llvm-svn: 281960
pointing to the same addr space. This can prevent SROA from creating a bitcast
between pointers with different addr spaces.
Differential Revision: http://reviews.llvm.org/D19697
llvm-svn: 268424
Currently each Function points to a DISubprogram and DISubprogram has a
scope field. For member functions the scope is a DICompositeType. DIScopes
point to the DICompileUnit to facilitate type uniquing.
Distinct DISubprograms (with isDefinition: true) are not part of the type
hierarchy and cannot be uniqued. This change removes the subprograms
list from DICompileUnit and instead adds a pointer to the owning compile
unit to distinct DISubprograms. This would make it easy for ThinLTO to
strip unneeded DISubprograms and their transitively referenced debug info.
Motivation
----------
Materializing DISubprograms is currently the most expensive operation when
doing a ThinLTO build of clang.
We want the DISubprogram to be stored in a separate Bitcode block (or the
same block as the function body) so we can avoid having to expensively
deserialize all DISubprograms together with the global metadata. If a
function has been inlined into another subprogram we need to store a
reference the block containing the inlined subprogram.
Attached to https://llvm.org/bugs/show_bug.cgi?id=27284 is a python script
that updates LLVM IR testcases to the new format.
http://reviews.llvm.org/D19034
<rdar://problem/25256815>
llvm-svn: 266446
While not strictly necessary, since we don't support large integer
types, this avoids bugs due to silent truncation from uint64_t to a
32-bit unsigned (e.g. DL.isLegalInteger(DL.getTypeSizeInBits(Ty) )
This fixes PR26972.
Differential Revision: http://reviews.llvm.org/D18258
llvm-svn: 263850
of, and I misdiagnosed for months and months.
Andrea has had a patch for this forever, but I just couldn't see how
it was fixing the root cause of the problem. It didn't make sense to me,
even though the patch was perfectly good and the analysis of the actual
failure event was *fantastic*.
Well, I came back to it today because the patch has sat for *far* too
long and needs attention and decided I wouldn't let it go until I really
understood what was going on. After quite some time in the debugger,
I finally realized that in fact I had just missed an important case with
my previous attempt to fix PR22093 in r225149. Not only do we need to
handle loads that won't be split, but stores-of-loads that we won't
split. We *do* actually have enough logic in the presplitting to form
new slices for split stores.... *unless* we decided not to split them!
I'm so sorry that it took me this long to come to the realization that
this is the issue. It seems so obvious in hind sight (of course).
Anyways, the fix becomes *much* smaller and more focused. The fact that
we're left doing integer smashing is related to the FIXME in my original
commit: fundamentally, we're not aggressive about pre-splitting for
loads and stores to the same alloca. If we want to get aggressive about
this, it'll need both what Andrea had put into the proposed fix, but
also a *lot* more logic to essentially iteratively pre-split the alloca
until we can't do any more. As I said in that commit log, its really
unclear that this is the right call. Instead, the integer blending and
letting targets lower this to narrower stores seems slightly better. But
we definitely shouldn't really go down that path just to fix this bug.
Again, tons of thanks are owed to Andrea and others at Sony for working
on this bug. I really should have seen what was going on here and
re-directed them sooner. =////
llvm-svn: 263121
Summary: If SROA creates only one piece (e.g. because the other is not needed),
it still needs to create a bit_piece expression if that bit piece is smaller
than the original size of the alloca.
Reviewers: aprantl
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D16187
llvm-svn: 257795
Note, this was reviewed (and more details are in) http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20151109/312083.html
These intrinsics currently have an explicit alignment argument which is
required to be a constant integer. It represents the alignment of the
source and dest, and so must be the minimum of those.
This change allows source and dest to each have their own alignments
by using the alignment attribute on their arguments. The alignment
argument itself is removed.
There are a few places in the code for which the code needs to be
checked by an expert as to whether using only src/dest alignment is
safe. For those places, they currently take the minimum of src/dest
alignments which matches the current behaviour.
For example, code which used to read:
call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %src, i32 500, i32 8, i1 false)
will now read:
call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 8 %dest, i8* align 8 %src, i32 500, i1 false)
For out of tree owners, I was able to strip alignment from calls using sed by replacing:
(call.*llvm\.memset.*)i32\ [0-9]*\,\ i1 false\)
with:
$1i1 false)
and similarly for memmove and memcpy.
I then added back in alignment to test cases which needed it.
A similar commit will be made to clang which actually has many differences in alignment as now
IRBuilder can generate different source/dest alignments on calls.
In IRBuilder itself, a new argument was added. Instead of calling:
CreateMemCpy(Dst, Src, getInt64(Size), DstAlign, /* isVolatile */ false)
you now call
CreateMemCpy(Dst, Src, getInt64(Size), DstAlign, SrcAlign, /* isVolatile */ false)
There is a temporary class (IntegerAlignment) which takes the source alignment and rejects
implicit conversion from bool. This is to prevent isVolatile here from passing its default
parameter to the source alignment.
Note, changes in future can now be made to codegen. I didn't change anything here, but this
change should enable better memcpy code sequences.
Reviewed by Hal Finkel.
llvm-svn: 253511
In some ways this is a very boring port to the new pass manager as there
are no interesting analyses or dependencies or other oddities.
However, this does introduce the first good example of a transformation
pass with non-trivial state porting to the new pass manager. I've tried
to carve out patterns here to replicate elsewhere, and would appreciate
comments on whether folks like these patterns:
- A common need in the new pass manager is to effectively lift the pass
class and some of its state into a public header file. Prior to this,
LLVM used anonymous namespaces to provide "module private" types and
utilities, but that doesn't scale to cases where a public header file
is needed and the new pass manager will exacerbate that. The pattern
I've adopted here is to use the namespace-cased-name of the core pass
(what would be a module if we had them) as a module-private namespace.
Then utility and other code can be declared and defined in this
namespace. At some point in the future, we could even have
(conditionally compiled) code that used modules features when
available to do the same basic thing.
- I've split the actual pass run method in two in order to expose
a private method usable by the old pass manager to wrap the new class
with a minimum of duplicated code. I actually looked at a bunch of
ways to automate or generate these, but they are all quite terrible
IMO. The fundamental need is to extract the set of analyses which need
to cross this interface boundary, and that will end up being too
unpredictable to effectively encapsulate IMO. This is also
a relatively small amount of boiler plate that will live a relatively
short time, so I'm not too worried about the fact that it is boiler
plate.
The rest of the patch is totally boring but results in a massive diff
(sorry). It just moves code around and removes or adds qualifiers to
reflect the new name and nesting structure.
Differential Revision: http://reviews.llvm.org/D12773
llvm-svn: 247501
handle more allocas with loads past the end of the alloca.
I suspect there are some related crashers with slightly different
patterns, but I'll fix those and add test cases as I find them.
Thanks to David Majnemer for the excellent test case reduction here.
Made this super simple to debug and fix.
llvm-svn: 246289
This was only added to preserve the old ScalarRepl's use of SSAUpdater
which was originally to avoid use of dominance frontiers. Now, we only
need a domtree, and we'll need a domtree right after this pass as well
and so it makes perfect sense to always and only use the dom-tree
powered mem2reg. This was flag-flipper earlier and has stuck reasonably
so I wanted to gut the now-dead code out of SROA before we waste more
time with it. Among other things, this will make passmanager porting
easier.
llvm-svn: 246028
types and loads, loads or stores widened past the size of an alloca,
etc.
This started off with a bug report about big-endian behavior with
bitfields and loads and stores to a { i32, i24 } struct. An initial
attempt to fix this was sent for review in D10357, but that didn't
really get to the root of the problem.
The core issue was that canConvertValue and convertValue in SROA were
handling different bitwidth integers by doing a zext of the integer. It
wouldn't do a trunc though, only a zext! This would in turn lead SROA to
form an i24 load from an i24 alloca, zext it to i32, and then use it.
This would at least produce the wrong value for big-endian systems.
One of my many false starts here was to correct the computation for
big-endian systems by shifting. But this doesn't actually work because
the original code has a 64-bit store to the entire 8 bytes, and a 32-bit
load of the last 4 bytes, and because the alloc size is 8 bytes, we
can't lose that last (least significant if bigendian) byte! The real
problem here is that we're forming an i24 load in SROA which is actually
not sufficiently wide to load all of the necessary bits here. The source
has an i32 load, and SROA needs to form that as well.
The straightforward way to do this is to disable the zext logic in
canConvertValue and convertValue, forcing us to actually load all
32-bits. This seems like a really good change, but it in turn breaks
several other parts of SROA.
First in the chain of knock-on failures, we had places where we were
doing integer-widening promotion even though some of the integer loads
or stores extended *past the end* of the alloca's memory! There was even
a comment about preventing this, but it only prevented the case where
the type had a different bit size from its store size. So I added checks
to handle the cases where we actually have a widened load or store and
to avoid trying to special integer widening promotion in those cases.
Second, we actually rely on the ability to promote in the face of loads
past the end of an alloca! This is important so that we can (for
example) speculate loads around PHI nodes to do more promotion. The bits
loaded are garbage, but as long as they aren't used and the alignment is
suitable high (which it wasn't in the test case!) this is "fine". And we
can't stop promoting here, lots of things stop working well if we do. So
we need to add specific logic to handle the extension (and truncation)
case, but *only* where that extension or truncation are over bytes that
*are outside the alloca's allocated storage* and thus totally bogus to
load or store.
And of course, once we add back this correct handling of extension or
truncation, we need to correctly handle bigendian systems to avoid
re-introducing the exact bug that started us off on this chain of misery
in the first place, but this time even more subtle as it only happens
along speculated loads atop a PHI node.
I've ported an existing test for PHI speculation to the big-endian test
file and checked that we get that part correct, and I've added several
more interesting big-endian test cases that should help check that we're
getting this correct.
Fun times.
llvm-svn: 242869
Volatile loads and stores are made visible in global state regardless of
what memory is involved. It is not correct to disregard the ordering
and synchronization scope because it is possible to synchronize with
memory operations performed by hardware.
This partially addresses PR23737.
llvm-svn: 242126
PPC_FP128 is really the sum of two consecutive doubles, where the first double
is always stored first in memory, regardless of the target endianness. The
memory layout of i128, however, depends on the target endianness, and so we
can't fold this without target endianness information. As a result, we must not
do this folding in lib/IR/ConstantFold.cpp (it could be done instead in
Analysis/ConstantFolding.cpp, but that's not done now).
Fixes PR23026.
llvm-svn: 233481
Similar to gep (r230786) and load (r230794) changes.
Similar migration script can be used to update test cases, which
successfully migrated all of LLVM and Polly, but about 4 test cases
needed manually changes in Clang.
(this script will read the contents of stdin and massage it into stdout
- wrap it in the 'apply.sh' script shown in previous commits + xargs to
apply it over a large set of test cases)
import fileinput
import sys
import re
rep = re.compile(r"(getelementptr(?:\s+inbounds)?\s*\()((<\d*\s+x\s+)?([^@]*?)(|\s*addrspace\(\d+\))\s*\*(?(3)>)\s*)(?=$|%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|zeroinitializer|<|\[\[[a-zA-Z]|\{\{)", re.MULTILINE | re.DOTALL)
def conv(match):
line = match.group(1)
line += match.group(4)
line += ", "
line += match.group(2)
return line
line = sys.stdin.read()
off = 0
for match in re.finditer(rep, line):
sys.stdout.write(line[off:match.start()])
sys.stdout.write(conv(match))
off = match.end()
sys.stdout.write(line[off:])
llvm-svn: 232184
Essentially the same as the GEP change in r230786.
A similar migration script can be used to update test cases, though a few more
test case improvements/changes were required this time around: (r229269-r229278)
import fileinput
import sys
import re
pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)")
for line in sys.stdin:
sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line))
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7649
llvm-svn: 230794
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
llvm-svn: 230786
assert out of the new pre-splitting in SROA.
This fix makes the code do what was originally intended -- when we have
a store of a load both dealing in the same alloca, we force them to both
be pre-split with identical offsets. This is really quite hard to do
because we can keep discovering problems as we go along. We have to
track every load over the current alloca which for any resaon becomes
invalid for pre-splitting, and go back to remove all stores of those
loads. I've included a couple of test cases derived from PR22093 that
cover the different ways this can happen. While that PR only really
triggered the first of these two, its the same fundamental issue.
The other challenge here is documented in a FIXME now. We end up being
quite a bit more aggressive for pre-splitting when loads and stores
don't refer to the same alloca. This aggressiveness comes at the cost of
introducing potentially redundant loads. It isn't clear that this is the
right balance. It might be considerably better to require that we only
do pre-splitting when we can presplit every load and store involved in
the entire operation. That would give more consistent if conservative
results. Unfortunately, it requires a non-trivial change to the actual
pre-splitting operation in order to correctly handle cases where we end
up pre-splitting stores out-of-order. And it isn't 100% clear that this
is the right direction, although I'm starting to suspect that it is.
llvm-svn: 225149
a pre-splitting pass over loads and stores.
Historically, splitting could cause enough problems that I hamstrung the
entire process with a requirement that splittable integer loads and
stores must cover the entire alloca. All smaller loads and stores were
unsplittable to prevent chaos from ensuing. With the new pre-splitting
logic that does load/store pair splitting I introduced in r225061, we
can now very nicely handle arbitrarily splittable loads and stores. In
order to fully benefit from these smarts, we need to mark all of the
integer loads and stores as splittable.
However, we don't actually want to rewrite partitions with all integer
loads and stores marked as splittable. This will fail to extract scalar
integers from aggregates, which is kind of the point of SROA. =] In
order to resolve this, what we really want to do is only do
pre-splitting on the alloca slices with integer loads and stores fully
splittable. This allows us to uncover all non-integer uses of the alloca
that would benefit from a split in an integer load or store (and where
introducing the split is safe because it is just memory transfer from
a load to a store). Once done, we make all the non-whole-alloca integer
loads and stores unsplittable just as they have historically been,
repartition and rewrite.
The result is that when there are integer loads and stores anywhere
within an alloca (such as from a memcpy of a sub-object of a larger
object), we can split them up if there are non-integer components to the
aggregate hiding beneath. I've added the challenging test cases to
demonstrate how this is able to promote to scalars even a case where we
have even *partially* overlapping loads and stores.
This restores the single-store behavior for small arrays of i8s which is
really nice. I've restored both the little endian testing and big endian
testing for these exactly as they were prior to r225061. It also forced
me to be more aggressive in an alignment test to actually defeat SROA.
=] Without the added volatiles there, we actually split up the weird i16
loads and produce nice double allocas with better alignment.
This also uncovered a number of bugs where we failed to handle
splittable load and store slices which didn't have a begininng offset of
zero. Those fixes are included, and without them the existing test cases
explode in glorious fireworks. =]
I've kept support for leaving whole-alloca integer loads and stores as
splittable even for the purpose of rewriting, but I think that's likely
no longer needed. With the new pre-splitting, we might be able to remove
all the splitting support for loads and stores from the rewriter. Not
doing that in this patch to try to isolate any performance regressions
that causes in an easy to find and revert chunk.
llvm-svn: 225074
stores.
When there are accesses to an entire alloca with an integer
load or store as well as accesses to small pieces of the alloca, SROA
splits up the large integer accesses. In order to do that, it uses bit
math to merge the small accesses into large integers. While this is
effective, it produces insane IR that can cause significant problems in
the rest of the optimizer:
- It can cause load and store mismatches with GVN on the non-alloca side
where we end up loading an i64 (or some such) rather than loading
specific elements that are stored.
- We can't always get rid of the integer bit math, which is why we can't
always fix the loads and stores to work well with GVN.
- This is especially bad when we have operations that mix poorly with
integer bit math such as floating point operations.
- It will block things like the vectorizer which might be able to handle
the scalar stores that underly the aggregate.
At the same time, we can't just directly split up these loads and stores
in all cases. If there is actual integer arithmetic involved on the
values, then using integer bit math is actually the perfect lowering
because we can often combine it heavily with the surrounding math.
The solution this patch provides is to find places where SROA is
partitioning aggregates into small elements, and look for splittable
loads and stores that it can split all the way to some other adjacent
load and store. These are uniformly the cases where failing to split the
loads and stores hurts the optimizer that I have seen, and I've looked
extensively at the code produced both from more and less aggressive
approaches to this problem.
However, it is quite tricky to actually do this in SROA. We may have
loads and stores to the same alloca, or other complex patterns that are
hard to handle. This complexity leads to the somewhat subtle algorithm
implemented here. We have to do this entire process as a separate pass
over the partitioning of the alloca, and split up all of the loads prior
to splitting the stores so that we can handle safely the cases of
overlapping, including partially overlapping, loads and stores to the
same alloca. We also have to reconstitute the post-split slice
configuration so we can avoid iterating again over all the alloca uses
(the slow part of SROA). But we also have to ensure that when we split
up loads and stores to *other* allocas, we *do* re-iterate over them in
SROA to adapt to the more refined partitioning now required.
With this, I actually think we can fix a long-standing TODO in SROA
where I avoided splitting as many loads and stores as probably should be
splittable. This limitation historically mitigated the fallout of all
the bad things mentioned above. Now that we have more intelligent
handling, I plan to remove the FIXME and more aggressively mark integer
loads and stores as splittable. I'll do that in a follow-up patch to
help with bisecting any fallout.
The net result of this change should be more fine-grained and accurate
scalars being formed out of aggregates. At the very least, Clang now
generates perfect code for this high-level test case using
std::complex<float>:
#include <complex>
void g1(std::complex<float> &x, float a, float b) {
x += std::complex<float>(a, b);
}
void g2(std::complex<float> &x, float a, float b) {
x -= std::complex<float>(a, b);
}
void foo(const std::complex<float> &x, float a, float b,
std::complex<float> &x1, std::complex<float> &x2) {
std::complex<float> l1 = x;
g1(l1, a, b);
std::complex<float> l2 = x;
g2(l2, a, b);
x1 = l1;
x2 = l2;
}
This code isn't just hypothetical either. It was reduced out of the hot
inner loops of essentially every part of the Eigen math library when
using std::complex<float>. Those loops would consistently and
pervasively hop between the floating point unit and the integer unit due
to bit math extraction and insertion of floating point values that were
"stored" in a 64-bit integer register around the loop backedge.
So far, this change has passed a bootstrap and I have done some other
testing and so far, no issues. That doesn't mean there won't be though,
so I'll be prepared to help with any fallout. If you performance swings
in particular, please let me know. I'm very curious what all the impact
of this change will be. Stay tuned for the follow-up to also split more
integer loads and stores.
llvm-svn: 225061
The alloca's type is irrelevant, only those types which are used in a
load or store of the exact size of the slice should be considered.
This manifested as an assertion failure when we compared the various
types: we had a size mismatch.
This fixes PR21480.
llvm-svn: 222499
cases where the alloca type, the load types, and the store types used
all disagree.
Previously, the only way that vector-based promotion occured was if the
alloca type was a vector type. This was one of the *very* few remaining
uses of the alloca's type to guide SROA/mem2reg left in LLVM. It turns
out it was a bad idea.
The alloca type can change very easily based on the mixture of types
loaded and stored to that alloca. We shouldn't be relying on it as
a signal for very much. Instead, the source of truth should be loads and
stores. We should canonicalize the loads and stores as much as possible
and then rely on them exclusively in SROA.
When looking and loads and stores, we may find many different candidate
vector types. This change will let SROA try all of them to find a vector
type which is a viable way to promote the entire alloca to a vector
register.
With this change, it becomes possible to do better canonicalization and
optimization of loads and stores without breaking SROA in random ways,
and that should allow fixing a core source of performance loss in hot
numerical loops such as those in Eigen.
llvm-svn: 220116
SROA may decide that it needs to insert a bitcast and would set it's
insertion point before a PHI. This will create an invalid module
right quick.
Instead, choose the first insertion point in the basic block that holds
our PHI.
This fixes PR20822.
Differential Revision: http://reviews.llvm.org/D5141
llvm-svn: 216891
Summary:
Fixes PR20425.
During slice building, if all of the incoming values of a PHI node are the same, replace the PHI node with the common value. This simplification makes alloca's used by PHI nodes easier to promote.
Test Plan: Added three more tests in phi-and-select.ll
Reviewers: nlewycky, eliben, meheff, chandlerc
Reviewed By: chandlerc
Subscribers: zinovy.nis, hfinkel, baldrick, llvm-commits
Differential Revision: http://reviews.llvm.org/D4659
llvm-svn: 216299
In this case, we are creating an x86_fp80 slice for a union from C where
the padding bytes may contain real data. An x86_fp80 alloca is 16 bytes,
and that's just fine. We can't, however, use regular loads and stores to
access the slice, because the store size is only 10 bytes / 80 bits.
Instead, use memcpy and memset.
Fixes PR18726.
Reviewed By: chandlerc
Differential Revision: http://reviews.llvm.org/D5012
llvm-svn: 216248
this case, the code path dealing with vector promotion was missing the explicit
checks for lifetime intrinsics that were present on the corresponding integer
promotion path.
llvm-svn: 215148
r199771 accidently broke the logic that makes sure that SROA only splits
load on byte boundaries. If such a split happens, some bits get lost
when reassembling loads of wider types, causing data corruption.
Move the width check up to reject such splits early, avoiding the
corruption. Fixes PR19250.
Patch by: Björn Steinbrink <bsteinbr@gmail.com>
llvm-svn: 211082
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
