This patch ensures that SHL/SRL/SRA shifts for i8 and i16 vectors avoid scalarization. It builds on the existing i8 SHL vectorized implementation of moving the shift bits up to the sign bit position and separating the 4, 2 & 1 bit shifts with several improvements:
1 - SSE41 targets can use (v)pblendvb directly with the sign bit instead of performing a comparison to feed into a VSELECT node.
2 - pre-SSE41 targets were masking + comparing with an 0x80 constant - we avoid this by using the fact that a set sign bit means a negative integer which can be compared against zero to then feed into VSELECT, avoiding the need for a constant mask (zero generation is much cheaper).
3 - SRA i8 needs to be unpacked to the upper byte of a i16 so that the i16 psraw instruction can be correctly used for sign extension - we have to do more work than for SHL/SRL but perf tests indicate that this is still beneficial.
The i16 implementation is similar but simpler than for i8 - we have to do 8, 4, 2 & 1 bit shifts but less shift masking is involved. SSE41 use of (v)pblendvb requires that the i16 shift amount is splatted to both bytes however.
Tested on SSE2, SSE41 and AVX machines.
Differential Revision: http://reviews.llvm.org/D9474
llvm-svn: 239509
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
First, don't combine bit masking into vector shuffles (even ones the
target can handle) once operation legalization has taken place. Custom
legalization of vector shuffles may exist for these patterns (making the
predicate return true) but that custom legalization may in some cases
produce the exact bit math this matches. We only really want to handle
this prior to operation legalization.
However, the x86 backend, in a fit of awesome, relied on this. What it
would do is mark VSELECTs as expand, which would turn them into
arithmetic, which this would then match back into vector shuffles, which
we would then lower properly. Amazing.
Instead, the second change is to teach the x86 backend to directly form
vector shuffles from VSELECT nodes with constant conditions, and to mark
all of the vector types we support lowering blends as shuffles as custom
VSELECT lowering. We still mark the forms which actually support
variable blends as *legal* so that the custom lowering is bypassed, and
the legal lowering can even be used by the vector shuffle legalization
(yes, i know, this is confusing. but that's how the patterns are
written).
This makes the VSELECT lowering much more sensible, and in fact should
fix a bunch of bugs with it. However, as you'll see in the test cases,
right now what it does is point out the *hilarious* deficiency of the
new vector shuffle lowering when it comes to blends. Fortunately, my
very next patch fixes that. I can't submit it yet, because that patch,
somewhat obviously, forms the exact and/or pattern that the DAG combine
is matching here! Without this patch, teaching the vector shuffle
lowering to produce the right code infloops in the DAG combiner. With
this patch alone, we produce terrible code but at least lower through
the right paths. With both patches, all the regressions here should be
fixed, and a bunch of the improvements (like using 2 shufps with no
memory loads instead of 2 andps with memory loads and an orps) will
stay. Win!
There is one other change worth noting here. We had hilariously wrong
vectorization cost estimates for vselect because we fell through to the
code path that assumed all "expand" vector operations are scalarized.
However, the "expand" lowering of VSELECT is vector bit math, most
definitely not scalarized. So now we go back to the correct if horribly
naive cost of "1" for "not scalarized". If anyone wants to add actual
modeling of shuffle costs, that would be cool, but this seems an
improvement on its own. Note the removal of 16 and 32 "costs" for doing
a blend. Even in SSE2 we can blend in fewer than 16 instructions. ;] Of
course, we don't right now because of OMG bad code, but I'm going to fix
that. Next patch. I promise.
llvm-svn: 229835
type erased interface and a single analysis pass rather than an
extremely complex analysis group.
The end result is that the TTI analysis can contain a type erased
implementation that supports the polymorphic TTI interface. We can build
one from a target-specific implementation or from a dummy one in the IR.
I've also factored all of the code into "mix-in"-able base classes,
including CRTP base classes to facilitate calling back up to the most
specialized form when delegating horizontally across the surface. These
aren't as clean as I would like and I'm planning to work on cleaning
some of this up, but I wanted to start by putting into the right form.
There are a number of reasons for this change, and this particular
design. The first and foremost reason is that an analysis group is
complete overkill, and the chaining delegation strategy was so opaque,
confusing, and high overhead that TTI was suffering greatly for it.
Several of the TTI functions had failed to be implemented in all places
because of the chaining-based delegation making there be no checking of
this. A few other functions were implemented with incorrect delegation.
The message to me was very clear working on this -- the delegation and
analysis group structure was too confusing to be useful here.
The other reason of course is that this is *much* more natural fit for
the new pass manager. This will lay the ground work for a type-erased
per-function info object that can look up the correct subtarget and even
cache it.
Yet another benefit is that this will significantly simplify the
interaction of the pass managers and the TargetMachine. See the future
work below.
The downside of this change is that it is very, very verbose. I'm going
to work to improve that, but it is somewhat an implementation necessity
in C++ to do type erasure. =/ I discussed this design really extensively
with Eric and Hal prior to going down this path, and afterward showed
them the result. No one was really thrilled with it, but there doesn't
seem to be a substantially better alternative. Using a base class and
virtual method dispatch would make the code much shorter, but as
discussed in the update to the programmer's manual and elsewhere,
a polymorphic interface feels like the more principled approach even if
this is perhaps the least compelling example of it. ;]
Ultimately, there is still a lot more to be done here, but this was the
huge chunk that I couldn't really split things out of because this was
the interface change to TTI. I've tried to minimize all the other parts
of this. The follow up work should include at least:
1) Improving the TargetMachine interface by having it directly return
a TTI object. Because we have a non-pass object with value semantics
and an internal type erasure mechanism, we can narrow the interface
of the TargetMachine to *just* do what we need: build and return
a TTI object that we can then insert into the pass pipeline.
2) Make the TTI object be fully specialized for a particular function.
This will include splitting off a minimal form of it which is
sufficient for the inliner and the old pass manager.
3) Add a new pass manager analysis which produces TTI objects from the
target machine for each function. This may actually be done as part
of #2 in order to use the new analysis to implement #2.
4) Work on narrowing the API between TTI and the targets so that it is
easier to understand and less verbose to type erase.
5) Work on narrowing the API between TTI and its clients so that it is
easier to understand and less verbose to forward.
6) Try to improve the CRTP-based delegation. I feel like this code is
just a bit messy and exacerbating the complexity of implementing
the TTI in each target.
Many thanks to Eric and Hal for their help here. I ended up blocked on
this somewhat more abruptly than I expected, and so I appreciate getting
it sorted out very quickly.
Differential Revision: http://reviews.llvm.org/D7293
llvm-svn: 227669
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
rdar://problem/18153096>
llvm-svn: 221657
The default implementation of getCmpSelInstrCost, which provides the cost of
icmp/fcmp/select instructions, did not deal sensibly with illegal vector types
that were scalarized. We'd ask for the legalization cost of the vector type,
which would return something like (4, f64) given an input of <4 x double>, and
we'd then check the TLI status of the ISD opcode on that scalar type. This would
result in querying (ISD::VSELECT, f64), for example. Amusingly enough,
ISD::VSELECT on scalar types is marked as Legal by default (as with most other
operations), and most backends never change this because VSELECT is never
generated on scalars. However, seeing the resulting operation as Legal, we'd
neglect to add the scalarization cost before returning. The result is that we'd
grossly under-estimate the cost of cmps/selects on illegal vector types.
Now, if type legalization clearly results in scalarization, we skip the early
return and add the scalarization cost.
llvm-svn: 217859
Cross-class copies being expensive is actually a trait of the microarchitecture, but as I haven't yet seen an example of a microarchitecture where they're cheap it seems best to just enable this by default, covering the non-mcpu build case.
llvm-svn: 217674
This patch:
1) Improves the cost model for x86 alternate shuffles (originally
added at revision 211339);
2) Teaches the Cost Model Analysis pass how to analyze alternate shuffles.
Alternate shuffles are a special kind of blend; on x86, we can often
easily lowered alternate shuffled into single blend
instruction (depending on the subtarget features).
The existing cost model didn't take into account subtarget features.
Also, it had a couple of "dead" entries for vector types that are never
legal (example: on x86 types v2i32 and v2f32 are not legal; those are
always either promoted or widened to 128-bit vector types).
The new x86 cost model takes into account what target features we have
before returning the shuffle cost (i.e. the number of instructions
after the blend is lowered/expanded).
This patch also teaches the Cost Model Analysis how to identify and analyze
alternate shuffles (i.e. 'SK_Alternate' shufflevector instructions):
- added function 'isAlternateVectorMask';
- added some logic to check if an instruction is a alternate shuffle and, in
case, call the target specific TTI to get the corresponding shuffle cost;
- added a test to verify the cost model analysis on alternate shuffles.
llvm-svn: 212296
This commit starts with a "git mv ARM64 AArch64" and continues out
from there, renaming the C++ classes, intrinsics, and other
target-local objects for consistency.
"ARM64" test directories are also moved, and tests that began their
life in ARM64 use an arm64 triple, those from AArch64 use an aarch64
triple. Both should be equivalent though.
This finishes the AArch64 merge, and everyone should feel free to
continue committing as normal now.
llvm-svn: 209577
BasicTTI::getMemoryOpCost must explicitly check for non-simple types; setting
AllowUnknown=true with TLI->getSimpleValueType is not sufficient because, for
example, non-power-of-two vector types return non-simple EVTs (not MVT::Other).
llvm-svn: 206150
This provides more realistic costs for the insert/extractelement instructions
(which are load/store pairs), accounts for the cheap unaligned Altivec load
sequence, and for unaligned VSX load/stores.
Bad news:
MultiSource/Applications/sgefa/sgefa - 35% slowdown (this will require more investigation)
SingleSource/Benchmarks/McGill/queens - 20% slowdown (we no longer vectorize this, but it was a constant store that was scalarized)
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2 - 2% slowdown
Good news:
SingleSource/Benchmarks/Shootout/ary3 - 54% speedup
SingleSource/Benchmarks/Shootout-C++/ary - 40% speedup
MultiSource/Benchmarks/Ptrdist/ks/ks - 35% speedup
MultiSource/Benchmarks/FreeBench/neural/neural - 30% speedup
MultiSource/Benchmarks/TSVC/Symbolics-flt/Symbolics-flt - 20% speedup
Unfortunately, estimating the costs of the stack-based scalarization sequences
is hard, and adjusting these costs is like a game of whac-a-mole :( I'll
revisit this again after we have better codegen for vector extloads and
truncstores and unaligned load/stores.
llvm-svn: 205658
When a vector type legalizes to a larger vector type, and the target does not
support the associated extending load (or truncating store), then legalization
will scalarize the load (or store) resulting in an associated scalarization
cost. BasicTTI::getMemoryOpCost needs to account for this.
Between this, and r205487, PowerPC on the P7 with VSX enabled shows:
MultiSource/Benchmarks/PAQ8p/paq8p: 43% speedup
SingleSource/Benchmarks/BenchmarkGame/puzzle: 51% speedup
SingleSource/UnitTests/Vectorizer/gcc-loops 28% speedup
(some of these are new; some of these, such as PAQ8p, just reverse regressions
that VSX support would trigger)
llvm-svn: 205495
For an cast (extension, etc.), the currently logic predicts a low cost if the
associated operation (keyed on the destination type) is legal (or promoted).
This is not true when the number of values required to legalize the type is
changing. For example, <8 x i16> being sign extended by <8 x i32> is not
generically cheap on PPC with VSX, even though sign extension to v4i32 is
legal, because two output v4i32 values are required compared to the single
v8i16 input value, and without custom logic in the target, this conversion will
scalarize.
llvm-svn: 205487
This adds a second implementation of the AArch64 architecture to LLVM,
accessible in parallel via the "arm64" triple. The plan over the
coming weeks & months is to merge the two into a single backend,
during which time thorough code review should naturally occur.
Everything will be easier with the target in-tree though, hence this
commit.
llvm-svn: 205090
the legalization cost must be included to get an accurate
estimation of the total cost of the scalarized vector.
The inaccurate cost triggered unprofitable SLP vectorization on
32-bit X86.
Summary:
Include legalization overhead when computing scalarization cost
Reviewers: hfinkel, nadav
CC: chandlerc, rnk, llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2992
llvm-svn: 203509
'OK_NonUniformConstValue' to identify operands which are constants but
not constant splats.
The cost model now allows returning 'OK_NonUniformConstValue'
for non splat operands that are instances of ConstantVector or
ConstantDataVector.
With this change, targets are now able to compute different costs
for instructions with non-uniform constant operands.
For example, On X86 the cost of a vector shift may vary depending on whether
the second operand is a uniform or non-uniform constant.
This patch applies the following changes:
- The cost model computation now takes into account non-uniform constants;
- The cost of vector shift instructions has been improved in
X86TargetTransformInfo analysis pass;
- BBVectorize, SLPVectorizer and LoopVectorize now know how to distinguish
between non-uniform and uniform constant operands.
Added a new test to verify that the output of opt
'-cost-model -analyze' is valid in the following configurations: SSE2,
SSE4.1, AVX, AVX2.
llvm-svn: 201272
The most important part of this is probably adding any cost at all for
operations like zext <8 x i8> to <8 x i32>. Before they were being
recorded as extremely costly (24, I believe) which made LLVM fall back
on a 4-wide vectorisation of a loop.
It also rebalances the values for sext, zext and trunc. Lacking any
other sane metric that might work across CPU microarchitectures I went
for instructions. This seems to be in reasonable accord with the rest
of the table (sitofp, ...) though no doubt at least one value is
sub-optimal for some bizarre reason.
Finally, separate AVX and AVX2 values are provided where appropriate.
The CodeGen is quite different in many cases.
rdar://problem/15981990
llvm-svn: 200928
Upcoming SLP vectorization improvements will want to be able to estimate costs
of horizontal reductions. Add infrastructure to support this.
We model reductions as a series of (shufflevector,add) tuples ultimately
followed by an extractelement. For example, for an add-reduction of <4 x float>
we could generate the following sequence:
(v0, v1, v2, v3)
\ \ / /
\ \ /
+ +
(v0+v2, v1+v3, undef, undef)
\ /
((v0+v2) + (v1+v3), undef, undef)
%rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef,
<4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
%bin.rdx = fadd <4 x float> %rdx, %rdx.shuf
%rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef,
<4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
%bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7
%r = extractelement <4 x float> %bin.rdx8, i32 0
This commit adds a cost model interface "getReductionCost(Opcode, Ty, Pairwise)"
that will allow clients to ask for the cost of such a reduction (as backends
might generate more efficient code than the cost of the individual instructions
summed up). This interface is excercised by the CostModel analysis pass which
looks for reduction patterns like the one above - starting at extractelements -
and if it sees a matching sequence will call the cost model interface.
We will also support a second form of pairwise reduction that is well supported
on common architectures (haddps, vpadd, faddp).
(v0, v1, v2, v3)
\ / \ /
(v0+v1, v2+v3, undef, undef)
\ /
((v0+v1)+(v2+v3), undef, undef, undef)
%rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef,
<4 x i32> <i32 0, i32 2 , i32 undef, i32 undef>
%rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef,
<4 x i32> <i32 1, i32 3, i32 undef, i32 undef>
%bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1
%rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef,
<4 x i32> <i32 0, i32 undef, i32 undef, i32 undef>
%rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef,
<4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
%bin.rdx.1 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1
%r = extractelement <4 x float> %bin.rdx.1, i32 0
llvm-svn: 190876
- Instead of setting the suffixes in a bunch of places, just set one master
list in the top-level config. We now only modify the suffix list in a few
suites that have one particular unique suffix (.ml, .mc, .yaml, .td, .py).
- Aside from removing the need for a bunch of lit.local.cfg files, this enables
4 tests that were inadvertently being skipped (one in
Transforms/BranchFolding, a .s file each in DebugInfo/AArch64 and
CodeGen/PowerPC, and one in CodeGen/SI which is now failing and has been
XFAILED).
- This commit also fixes a bunch of config files to use config.root instead of
older copy-pasted code.
llvm-svn: 188513
This patch fixes the multiple breakages on ARM test-suite after the SLP
vectorizer was introduced by default on O3. The problem was an illegal
vector type on ARMTTI::getCmpSelInstrCost() <3 x i1> which is not simple.
The guard protects this code from breaking (cause of the problems) but
doesn't fix the issue that is generating the odd vector in the first
place, which also needs to be investigated.
llvm-svn: 187658
This fixes an oversight that Intrinsic::nearbyint was not being mapped to
ISD::FNEARBYINT (thus fixing the over-optimistic cost we were assigning to
nearbyint calls for some targets).
llvm-svn: 185783
Rather than just splitting the input type and hoping for the best, apply
a bit more cleverness. Just splitting the types until the source is
legal often leads to an illegal result time, which is then widened and a
scalarization step is introduced which leads to truly horrible code
generation. With the loop vectorizer, these sorts of operations are much
more common, and so it's worth extra effort to do them well.
Add a legalization hook for the operands of a TRUNCATE node, which will
be encountered after the result type has been legalized, but if the
operand type is still illegal. If simple splitting of both types
ends up with the result type of each half still being legal, just
do that (v16i16 -> v16i8 on ARM, for example). If, however, that would
result in an illegal result type (v8i32 -> v8i8 on ARM, for example),
we can get more clever with power-two vectors. Specifically,
split the input type, but also widen the result element size, then
concatenate the halves and truncate again. For example on ARM,
To perform a "%res = v8i8 trunc v8i32 %in" we transform to:
%inlo = v4i32 extract_subvector %in, 0
%inhi = v4i32 extract_subvector %in, 4
%lo16 = v4i16 trunc v4i32 %inlo
%hi16 = v4i16 trunc v4i32 %inhi
%in16 = v8i16 concat_vectors v4i16 %lo16, v4i16 %hi16
%res = v8i8 trunc v8i16 %in16
This allows instruction selection to generate three VMOVN instructions
instead of a sequences of moves, stores and loads.
Update the ARMTargetTransformInfo to take this improved legalization
into account.
Consider the simplified IR:
define <16 x i8> @test1(<16 x i32>* %ap) {
%a = load <16 x i32>* %ap
%tmp = trunc <16 x i32> %a to <16 x i8>
ret <16 x i8> %tmp
}
define <8 x i8> @test2(<8 x i32>* %ap) {
%a = load <8 x i32>* %ap
%tmp = trunc <8 x i32> %a to <8 x i8>
ret <8 x i8> %tmp
}
Previously, we would generate the truly hideous:
.syntax unified
.section __TEXT,__text,regular,pure_instructions
.globl _test1
.align 2
_test1: @ @test1
@ BB#0:
push {r7}
mov r7, sp
sub sp, sp, #20
bic sp, sp, #7
add r1, r0, #48
add r2, r0, #32
vld1.64 {d24, d25}, [r0:128]
vld1.64 {d16, d17}, [r1:128]
vld1.64 {d18, d19}, [r2:128]
add r1, r0, #16
vmovn.i32 d22, q8
vld1.64 {d16, d17}, [r1:128]
vmovn.i32 d20, q9
vmovn.i32 d18, q12
vmov.u16 r0, d22[3]
strb r0, [sp, #15]
vmov.u16 r0, d22[2]
strb r0, [sp, #14]
vmov.u16 r0, d22[1]
strb r0, [sp, #13]
vmov.u16 r0, d22[0]
vmovn.i32 d16, q8
strb r0, [sp, #12]
vmov.u16 r0, d20[3]
strb r0, [sp, #11]
vmov.u16 r0, d20[2]
strb r0, [sp, #10]
vmov.u16 r0, d20[1]
strb r0, [sp, #9]
vmov.u16 r0, d20[0]
strb r0, [sp, #8]
vmov.u16 r0, d18[3]
strb r0, [sp, #3]
vmov.u16 r0, d18[2]
strb r0, [sp, #2]
vmov.u16 r0, d18[1]
strb r0, [sp, #1]
vmov.u16 r0, d18[0]
strb r0, [sp]
vmov.u16 r0, d16[3]
strb r0, [sp, #7]
vmov.u16 r0, d16[2]
strb r0, [sp, #6]
vmov.u16 r0, d16[1]
strb r0, [sp, #5]
vmov.u16 r0, d16[0]
strb r0, [sp, #4]
vldmia sp, {d16, d17}
vmov r0, r1, d16
vmov r2, r3, d17
mov sp, r7
pop {r7}
bx lr
.globl _test2
.align 2
_test2: @ @test2
@ BB#0:
push {r7}
mov r7, sp
sub sp, sp, #12
bic sp, sp, #7
vld1.64 {d16, d17}, [r0:128]
add r0, r0, #16
vld1.64 {d20, d21}, [r0:128]
vmovn.i32 d18, q8
vmov.u16 r0, d18[3]
vmovn.i32 d16, q10
strb r0, [sp, #3]
vmov.u16 r0, d18[2]
strb r0, [sp, #2]
vmov.u16 r0, d18[1]
strb r0, [sp, #1]
vmov.u16 r0, d18[0]
strb r0, [sp]
vmov.u16 r0, d16[3]
strb r0, [sp, #7]
vmov.u16 r0, d16[2]
strb r0, [sp, #6]
vmov.u16 r0, d16[1]
strb r0, [sp, #5]
vmov.u16 r0, d16[0]
strb r0, [sp, #4]
ldm sp, {r0, r1}
mov sp, r7
pop {r7}
bx lr
Now, however, we generate the much more straightforward:
.syntax unified
.section __TEXT,__text,regular,pure_instructions
.globl _test1
.align 2
_test1: @ @test1
@ BB#0:
add r1, r0, #48
add r2, r0, #32
vld1.64 {d20, d21}, [r0:128]
vld1.64 {d16, d17}, [r1:128]
add r1, r0, #16
vld1.64 {d18, d19}, [r2:128]
vld1.64 {d22, d23}, [r1:128]
vmovn.i32 d17, q8
vmovn.i32 d16, q9
vmovn.i32 d18, q10
vmovn.i32 d19, q11
vmovn.i16 d17, q8
vmovn.i16 d16, q9
vmov r0, r1, d16
vmov r2, r3, d17
bx lr
.globl _test2
.align 2
_test2: @ @test2
@ BB#0:
vld1.64 {d16, d17}, [r0:128]
add r0, r0, #16
vld1.64 {d18, d19}, [r0:128]
vmovn.i32 d16, q8
vmovn.i32 d17, q9
vmovn.i16 d16, q8
vmov r0, r1, d16
bx lr
llvm-svn: 179989
The costs are overfitted so that I can still use the legalization factor.
For example the following kernel has about half the throughput vectorized than
unvectorized when compiled with SSE2. Before this patch we would vectorize it.
unsigned short A[1024];
double B[1024];
void f() {
int i;
for (i = 0; i < 1024; ++i) {
B[i] = (double) A[i];
}
}
radar://13599001
llvm-svn: 179033
The code in getTypeConversion attempts to promote the element vector type
before it trys to split or widen the vector.
After it failed finding a legal vector type by promoting it would continue using
the promoted vector element type. Thereby missing legal splitted vector types.
For example the type v32i32 that has a legal split of 4 x v3i32 on x86/sse2
would be transformed to: v32i256 and from there on successively split to:
v16i256, v8i256, v1i256 and then finally ends up as an i64 type.
By resetting the vector element type to the original vector element type that
existed before the promotion the code will attempt to split the vector type to
smaller vector widths of the same type.
llvm-svn: 178999
SSE2 has efficient support for shifts by a scalar. My previous change of making
shifts expensive did not take this into account marking all shifts as expensive.
This would prevent vectorization from happening where it is actually beneficial.
With this change we differentiate between shifts of constants and other shifts.
radar://13576547
llvm-svn: 178808
The default logic does not correctly identify costs of casts because they are
marked as custom on x86.
For some cases, where the shift amount is a scalar we would be able to generate
better code. Unfortunately, when this is the case the value (the splat) will get
hoisted out of the loop, thereby making it invisible to ISel.
radar://13130673
radar://13537826
llvm-svn: 178703
- After moving logic recognizing vector shift with scalar amount from
DAG combining into DAG lowering, we declare to customize all vector
shifts even vector shift on AVX is legal. As a result, the cost model
needs special tuning to identify these legal cases.
llvm-svn: 177586
