{code}
// loop body
... = a[i] (1)
... = a[i+1] (2)
.......
a[i+1] = .... (3)
a[i] = ... (4)
{code}
The algorithm tries to collect memory access candidates from AliasSetTracker, and then check memory dependences one another. The memory accesses are unique in AliasSetTracker, and a single memory access in AliasSetTracker may map to multiple entries in AccessAnalysis, which could cover both 'read' and 'write'. Originally the algorithm only checked 'write' entry in Accesses if only 'write' exists. This is incorrect and the consequence is it ignored all read access, and finally some RAW and WAR dependence are missed.
For the case given above, if we ignore two reads, the dependence between (1) and (3) would not be able to be captured, and finally this loop will be incorrectly vectorized.
The fix simply inserts a new loop to find all entries in Accesses. Since it will skip most of all other memory accesses by checking the Value pointer at the very beginning of the loop, it should not increase compile-time visibly.
llvm-svn: 225159
PPC has an instruction for ctlz with defined zero behavior, and our lowering of
cttz (provided by DAGCombine) is also efficient and branchless, so speculating
these makes sense.
llvm-svn: 225150
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
r225135 added the ability to materialize i64 constants using rotations in order
to reduce the instruction count. Sometimes we can use a rotation only with some
extra masking, so that we take advantage of the fact that generating a bunch of
extra higher-order 1 bits is easy using li/lis.
llvm-svn: 225147
The required functionality has been there for some time, but I never
managed to actually wire it into the command line registry of passes.
Let's do that.
llvm-svn: 225144
Materializing full 64-bit constants on PPC64 can be expensive, requiring up to
5 instructions depending on the locations of the non-zero bits. Sometimes
materializing a rotated constant, and then applying the inverse rotation, requires
fewer instructions than the direct method. If so, do that instead.
In r225132, I added support for forming constants using bit inversion. In
effect, this reverts that commit and replaces it with rotation support. The bit
inversion is useful for turning constants that are mostly ones into ones that
are mostly zeros (thus enabling a more-efficient shift-based materialization),
but the same effect can be obtained by using negative constants and a rotate,
and that is at least as efficient, if not more.
llvm-svn: 225135
Materializing full 64-bit constants on PPC64 can be expensive, requiring up to
5 instructions depending on the locations of the non-zero bits. Sometimes
materializing the bit-reversed constant, and then flipping the bits, requires
fewer instructions than the direct method. If so, do that instead.
llvm-svn: 225132
We assumed the output of a match was a Value, this would cause us to
assert because we would fail a cast<>. Instead, use a helper in the
Operator family to hide the distinction between Value and Constant.
This fixes PR22087.
llvm-svn: 225127
PHI nodes can have zero operands in the middle of a transform. It is
expected that utilities in Analysis don't freak out when this happens.
Note that it is considered invalid to allow these misshapen phi nodes to
make it to another pass.
This fixes PR22086.
llvm-svn: 225126
Weak externals are resolved statically, so we can actually generate the tail
call on PE/COFF targets without breaking the requirements. It is questionable
whether we want to propagate the current behaviour for MachO as the requirements
are part of the ARM ELF specifications, and it seems that prior to the SVN
r215890, we would have tail'ed the call. For now, be conservative and only
permit it on PE/COFF where the call will always be fully resolved.
llvm-svn: 225119
The existing code provided for specifying a global loop alignment preference.
However, the preferred loop alignment might depend on the loop itself. For
recent POWER cores, loops between 5 and 8 instructions should have 32-byte
alignment (while the others are better with 16-byte alignment) so that the
entire loop will fit in one i-cache line.
To support this, getPrefLoopAlignment has been made virtual, and can be
provided with an optional MachineLoop* so the target can inspect the loop
before answering the query. The default behavior, as before, is to return the
value set with setPrefLoopAlignment. MachineBlockPlacement now queries the
target for each loop instead of only once per function. There should be no
functional change for other targets.
llvm-svn: 225117
Most modern PowerPC cores prefer that functions and loops start on
16-byte-aligned boundaries (*), so instruct block placement, etc. to make this
happen. The branch selector has also been adjusted so account for the extra
nops that might now be inserted before loop headers.
(*) Some cores actually prefer other alignments for small loops, but that will
be addressed in a follow-up commit.
llvm-svn: 225115
Newer POWER cores, and the A2, support the cmpb instruction. This instruction
compares its operands, treating each of the 8 bytes in the GPRs separately,
returning a 'mask' result of 0 (for false) or -1 (for true) in each byte.
Code generation support is added, in the form of a PPCISelDAGToDAG
DAG-preprocessing routine, that recognizes patterns close to what the
instruction computes (either exactly, or related by a constant masking
operation), and generates the cmpb instruction (along with any necessary
constant masking operation). This can be expanded if use cases arise.
llvm-svn: 225106
This is necessary to allow the disassembler to be able to handle AdSize32 instructions in 64-bit mode when address size prefix is used.
Eventually we should probably also support 'addr32' and 'addr16' in the assembler to override the address size on some of these instructions. But for now we'll just use special operand types that will lookup the current mode size to select the right instruction.
llvm-svn: 225075
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
This is the second installment of improvements to instruction selection for "bit
permutation" instruction sequences. r224318 added logic for instruction
selection for 32-bit bit permutation sequences, and this adds lowering for
64-bit sequences. The 64-bit sequences are more complicated than the 32-bit
ones because:
a) the 64-bit versions of the 32-bit rotate-and-mask instructions
work by replicating the lower 32-bits of the value-to-be-rotated into the
upper 32 bits -- and integrating this into the cost modeling for the various
bit group operations is non-trivial
b) unlike the 32-bit instructions in 32-bit mode, the rotate-and-mask instructions
cannot, in one instruction, specify the
mask starting index, the mask ending index, and the rotation factor. Also,
forming arbitrary 64-bit constants is more complicated than in 32-bit mode
because the number of instructions necessary is value dependent.
Plus, support for 'late masking' was added: it is sometimes more efficient to
treat the overall value as if it had no mandatory zero bits when planning the
bit-group insertions, and then mask them in at the very end. Unfortunately, as
the structure of the bit groups is different in the two cases, the more
feasible implementation technique was to generate both instruction sequences,
and then pick the shorter one.
And finally, we now generate reasonable code for i64 bswap:
rldicl 5, 3, 16, 0
rldicl 4, 3, 8, 0
rldicl 6, 3, 24, 0
rldimi 4, 5, 8, 48
rldicl 5, 3, 32, 0
rldimi 4, 6, 16, 40
rldicl 6, 3, 48, 0
rldimi 4, 5, 24, 32
rldicl 5, 3, 56, 0
rldimi 4, 6, 40, 16
rldimi 4, 5, 48, 8
rldimi 4, 3, 56, 0
vs. what we used to produce:
li 4, 255
rldicl 5, 3, 24, 40
rldicl 6, 3, 40, 24
rldicl 7, 3, 56, 8
sldi 8, 3, 8
sldi 10, 3, 24
sldi 12, 3, 40
rldicl 0, 3, 8, 56
sldi 9, 4, 32
sldi 11, 4, 40
sldi 4, 4, 48
andi. 5, 5, 65280
andis. 6, 6, 255
andis. 7, 7, 65280
sldi 3, 3, 56
and 8, 8, 9
and 4, 12, 4
and 9, 10, 11
or 6, 7, 6
or 5, 5, 0
or 3, 3, 4
or 7, 9, 8
or 4, 6, 5
or 3, 3, 7
or 3, 3, 4
which is 12 instructions, instead of 25, and seems optimal (at least in terms
of code size).
llvm-svn: 225056
The issues was that AArch64 has additional restrictions on when local
relocations can be used. We have to take those into consideration when
deciding to put a L symbol in the symbol table or not.
Original message:
Remove doesSectionRequireSymbols.
In an assembly expression like
bar:
.long L0 + 1
the intended semantics is that bar will contain a pointer one byte past L0.
In sections that are merged by content (strings, 4 byte constants, etc), a
single position in the section doesn't give the linker enough information.
For example, it would not be able to tell a relocation must point to the
end of a string, since that would look just like the start of the next.
The solution used in ELF to use relocation with symbols if there is a non-zero
addend.
In MachO before this patch we would just keep all symbols in some sections.
This would miss some cases (only cstrings on x86_64 were implemented) and was
inefficient since most relocations have an addend of 0 and can be represented
without the symbol.
This patch implements the non-zero addend logic for MachO too.
llvm-svn: 225048
We are allowed to move the 'B' to the right hand side if we an prove
there is no signed overflow and if the comparison itself is signed.
llvm-svn: 225034
Too many different comment characters - instead of trying to account for
them all, instead disable the comments and just check for end-of-line
instead.
llvm-svn: 225020
GCC does this for non-zero discriminators and since GCC doesn't produce
column info, that was the only place it comes up there. For LLVM, since
we can emit discriminators and/or column info, it makes more sense to
invert the condition and just test for changes in line number.
This should resolve at least some of the GDB 7.5 test suite failures
created by recent Clang changes that increase the location fidelity
(which, since Clang defaults to including column info on Linux by
default created a bunch of cases that confused GDB).
In theory we could do this better/differently by grouping actual source
statements together in a similar manner to the way lexical scopes are
handled but given that GDB isn't really in a position to consume that (&
users are probably somewhat used to different lines being different
'statements') this seems the safest and cheapest change. (I'm concerned
that doing this 'right' would bloat the debugloc data even further -
something Duncan's working hard to address)
llvm-svn: 225011
Under the large code model, we cannot assume that __morestack lives within
2^31 bytes of the call site, so we cannot use pc-relative addressing. We
cannot perform the call via a temporary register, as the rax register may
be used to store the static chain, and all other suitable registers may be
either callee-save or used for parameter passing. We cannot use the stack
at this point either because __morestack manipulates the stack directly.
To avoid these issues, perform an indirect call via a read-only memory
location containing the address.
This solution is not perfect, as it assumes that the .rodata section
is laid out within 2^31 bytes of each function body, but this seems to
be sufficient for JIT.
Differential Revision: http://reviews.llvm.org/D6787
llvm-svn: 225003
If a linker directive is already quoted, don't try to quote it again, otherwise it creates a mess.
This pops up in places like:
#pragma comment(linker,"\"/foo bar'\"")
Differential Revision: http://reviews.llvm.org/D6792
llvm-svn: 224998
