VBMI introduced VPERMB, so cost-model i8 replication shuffle using it.
Note that we can still model i8 replication shufflle without VBMI,
by promoting to i16/i32. That will be done in follow-ups.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D113479
BWI introduced VPERMW, so cost-model i16 replication shuffle using it.
Note that we can still model i16 replication shufflle without BWI,
by promoting to i32. That will be done in follow-ups.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D113478
This models lowering to `vpermd`/`vpermq`/`vpermps`/`vpermpd`,
that take a single input vector and a single index vector,
and are cross-lane. So far i haven't seen evidence that
replication ever results in demanding more than a single
input vector per output vector.
This results in *shockingly* lesser costs :)
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D113350
This finally creates proper test coverage for replication shuffles,
that are used by LV for conditional loads, and will allow to add
proper costmodel at least for AVX512.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D113324
Even though AVX512's masked mem ops (unlike AVX1/2) have a mask
that is a `VF x i1`, replication of said masks happens after
promotion of it to `VF x i8`, so we should use `i8`, not `i1`,
when calculating the cost of mask replication.
I don't really buy that masked interleaved memory loads/stores are supported on X86.
There is zero costmodel test coverage, no actual cost modelling for the generation
of the mask repetition, and basically only two LV tests.
Additionally, i'm not very interested in AVX512.
I don't know if this really helps "soft" block over at
https://reviews.llvm.org/D111460#inline-1075467,
but i think it can't make things worse at least.
When we are being told that there is a masking, instead of
completely giving up and falling back to
fully scalarizing `BasicTTIImplBase::getInterleavedMemoryOpCost()`,
let's correctly query the cost of masked memory ops,
keep all the pretty shuffle cost modelling,
but scalarize the cost computation for the mask replication.
I think, not scalarizing the shuffles themselves
may adjust the computed costs a bit,
and maybe hopefully just enough to hide the "regressions"
at https://reviews.llvm.org/D111460#inline-1075467
I do mean hide, because the test coverage is non-existent.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D112873
As it can be seen in `InnerLoopVectorizer::vectorizeInterleaveGroup()`,
in some cases (reported by `UseMaskForGaps`), the gaps in the interleaved load/store group
will be masked away by another constant mask, so there is no need to
account for the cost of replication of the mask for these.
Differential Revision: https://reviews.llvm.org/D112877
Copied from llvm/test/Transforms/LoopVectorize/X86/x86-interleaved-accesses-masked-group.ll
As discussed in D111460 / D112877 / D112873 we have basically no test coverage
for this part of cost model.
By definition, interleaving load of stride N means:
load N*VF elements, and shuffle them into N VF-sized vectors,
with 0'th vector containing elements `[0, VF)*stride + 0`,
and 1'th vector containing elements `[0, VF)*stride + 1`.
Example: https://godbolt.org/z/df561Me5E (i64 stride 4 vf 2 => cost 6)
Now, not fully interleaved load, is when not all of these vectors is demanded.
So at worst, we could just pretend that everything is demanded,
and discard the non-demanded vectors. What this means is that the cost
for not-fully-interleaved group should be not greater than the cost
for the same fully-interleaved group, but perhaps somewhat less.
Examples:
https://godbolt.org/z/a78dK5Geq (i64 stride 4 (indices 012u) vf 2 => cost 4)
https://godbolt.org/z/G91ceo8dM (i64 stride 4 (indices 01uu) vf 2 => cost 2)
https://godbolt.org/z/5joYob9rx (i64 stride 4 (indices 0uuu) vf 2 => cost 1)
Right now, for such not-fully-interleaved loads we just use the costs
for fully-interleaved loads. But at least **in general**,
that is obviously overly pessimistic, because **in general**,
not all the shuffles needed to perform the full interleaving
will end up being live.
So what this does, is naively scales the interleaving cost
by the fraction of the live members. I believe this should still result
in the right ballpark cost estimate, although it may be over/under -estimate.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D112307
This patch introduces a new function:
AArch64Subtarget::getVScaleForTuning
that returns a value for vscale that can be used for tuning the cost
model when using scalable vectors. The VScaleForTuning option in
AArch64Subtarget is initialised according to the following rules:
1. If the user has specified the CPU to tune for we use that, else
2. If the target CPU was specified we use that, else
3. The tuning is set to "generic".
For CPUs of type "generic" I have assumed that vscale=2.
New tests added here:
Analysis/CostModel/AArch64/sve-gather.ll
Analysis/CostModel/AArch64/sve-scatter.ll
Transforms/LoopVectorize/AArch64/sve-strict-fadd-cost.ll
Differential Revision: https://reviews.llvm.org/D110259
A few more tuples are being queried after D111546. Might be good to model them,
They all require a lot of manual assembly surgery.
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/9bnKrefcG - for intels `Block RThroughput: =40.0`; for ryzens, `Block RThroughput: =16.0`
So could pick cost of `40`
For store we have:
https://godbolt.org/z/5s3s14dEY - for intels `Block RThroughput: =40.0`; for ryzens, `Block RThroughput: =16.0`
So we could pick cost of `40`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111945
A few more tuples are being queried after D111546. Might be good to model them,
They all require a lot of manual assembly surgery.
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/MTaKboejM - for intels `Block RThroughput: =32.0`; for ryzens, `Block RThroughput: <=16.0`
So could pick cost of `32`
For store we have:
https://godbolt.org/z/v7xPj3Wd4 - for intels `Block RThroughput: =32.0`; for ryzens, `Block RThroughput: <=32.0`
So we could pick cost of `32`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111944
A few more tuples are being queried after D111546. Might be good to model them,
They all require a lot of manual assembly surgery.
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/11rcvdreP - for intels `Block RThroughput: <=68.0`; for ryzens, `Block RThroughput: <=48.0`
So could pick cost of `68`
For store we have:
https://godbolt.org/z/6aM11fWcP - for intels `Block RThroughput: <=64.0`; for ryzens, `Block RThroughput: <=32.0`
So we could pick cost of `64`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111943
A few more tuples are being queried after D111546. Might be good to model them,
They all require a lot of manual assembly surgery.
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/s5b6E6jsP - for intels `Block RThroughput: <=32.0`; for ryzens, `Block RThroughput: <=24.0`
So could pick cost of `32`
For store we have:
https://godbolt.org/z/efh99d93b - for intels `Block RThroughput: <=48.0`; for ryzens, `Block RThroughput: <=32.0`
So we could pick cost of `48`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111942
A few more tuples are being queried after D111546. Might be good to model them,
They all require a lot of manual assembly surgery.
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/YTeT9M7fW - for intels `Block RThroughput: <=212.0`; for ryzens, `Block RThroughput: <=64.0`
So could pick cost of `212`
For store we have:
https://godbolt.org/z/vc954KEGP - for intels `Block RThroughput: <=90.0`; for ryzens, `Block RThroughput: <=24.0`
So we could pick cost of `90`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111940
These cases use the same codegen as AVX2 (pshuflw/pshufd) for the sub-128bit vector deinterleaving, and unpcklqdq for v2i64.
It's going to take a while to add full interleaved cost coverage, but since these are the same for SSE2 -> AVX2 it should be an easy win.
Fixes PR47437
Differential Revision: https://reviews.llvm.org/D111938
And another attempt to start untangling this ball of threads around gather.
There's `TTI::prefersVectorizedAddressing()`hoop, which confusingly defaults to `true`,
which tells LV to try to vectorize the addresses that lead to loads,
but X86 generally can not deal with vectors of addresses,
the only instructions that support that are GATHER/SCATTER,
but even those aren't available until AVX2, and aren't really usable until AVX512.
This specializes the hook for X86, to return true only if we have AVX512 or AVX2 w/ fast gather.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111546
While i've modelled most of the relevant tuples for AVX2,
that only covered fully-interleaved groups.
By definition, interleaving load of stride N means:
load N*VF elements, and shuffle them into N VF-sized vectors,
with 0'th vector containing elements `[0, VF)*stride + 0`,
and 1'th vector containing elements `[0, VF)*stride + 1`.
Example: https://godbolt.org/z/df561Me5E (i64 stride 4 vf 2 => cost 6)
Now, not fully interleaved load, is when not all of these vectors is demanded.
So at worst, we could just pretend that everything is demanded,
and discard the non-demanded vectors. What this means is that the cost
for not-fully-interleaved group should be not greater than the cost
for the same fully-interleaved group, but perhaps somewhat less.
Examples:
https://godbolt.org/z/a78dK5Geq (i64 stride 4 (indices 012u) vf 2 => cost 4)
https://godbolt.org/z/G91ceo8dM (i64 stride 4 (indices 01uu) vf 2 => cost 2)
https://godbolt.org/z/5joYob9rx (i64 stride 4 (indices 0uuu) vf 2 => cost 1)
As we have established over the course of last ~70 patches, (wow)
`BaseT::getInterleavedMemoryOpCos()` is absolutely bogus,
it is usually almost an order of magnitude overestimation,
so i would claim that we should at least use the hardcoded costs
of fully interleaved load groups.
We could go further and adjust them e.g. by the number of demanded indices,
but then i'm somewhat fearful of underestimating the cost.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111174
Without SSE41 sext/zext instructions the extensions will be split, meaning that the MUL->PMADDWD fold will split the sext_i32(x) into zext_i32(sext_i16(x))
`X86TTIImpl::getGSScalarCost()` has (at least) two issues:
* it naively computes the cost of sequence of `insertelement`/`extractelement`.
If we are operating not on the XMM (but YMM/ZMM),
this widely overestimates the cost of subvector insertions/extractions.
* Gather/scatter takes a vector of pointers, and scalarization results in us performing
scalar memory operation for each of these pointers, but we never account for the cost
of extracting these pointers out of the vector of pointers.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111222
We would like to start pushing -mcpu=generic towards enabling the set of
features that improves performance for some CPUs, without hurting any
others. A blend of the performance options hopefully beneficial to all
CPUs. The largest part of that is enabling in-order scheduling using the
Cortex-A55 schedule model. This is similar to the Arm backend change
from eecb353d0e which made -mcpu=generic perform in-order scheduling
using the cortex-a8 schedule model.
The idea is that in-order cpu's require the most help in instruction
scheduling, whereas out-of-order cpus can for the most part out-of-order
schedule around different codegen. Our benchmarking suggests that
hypothesis holds. When running on an in-order core this improved
performance by 3.8% geomean on a set of DSP workloads, 2% geomean on
some other embedded benchmark and between 1% and 1.8% on a set of
singlecore and multicore workloads, all running on a Cortex-A55 cluster.
On an out-of-order cpu the results are a lot more noisy but show flat
performance or an improvement. On the set of DSP and embedded
benchmarks, run on a Cortex-A78 there was a very noisy 1% speed
improvement. Using the most detailed results I could find, SPEC2006 runs
on a Neoverse N1 show a small increase in instruction count (+0.127%),
but a decrease in cycle counts (-0.155%, on average). The instruction
count is very low noise, the cycle count is more noisy with a 0.15%
decrease not being significant. SPEC2k17 shows a small decrease (-0.2%)
in instruction count leading to a -0.296% decrease in cycle count. These
results are within noise margins but tend to show a small improvement in
general.
When specifying an Apple target, clang will set "-target-cpu apple-a7"
on the command line, so should not be affected by this change when
running from clang. This also doesn't enable more runtime unrolling like
-mcpu=cortex-a55 does, only changing the schedule used.
A lot of existing tests have updated. This is a summary of the important
differences:
- Most changes are the same instructions in a different order.
- Sometimes this leads to very minor inefficiencies, such as requiring
an extra mov to move variables into r0/v0 for the return value of a test
function.
- misched-fusion.ll was no longer fusing the pairs of instructions it
should, as per D110561. I've changed the schedule used in the test
for now.
- neon-mla-mls.ll now uses "mul; sub" as opposed to "neg; mla" due to
the different latencies. This seems fine to me.
- Some SVE tests do not always remove movprfx where they did before due
to different register allocation giving different destructive forms.
- The tests argument-blocks-array-of-struct.ll and arm64-windows-calls.ll
produce two LDR where they previously produced an LDP due to
store-pair-suppress kicking in.
- arm64-ldp.ll and arm64-neon-copy.ll are missing pre/postinc on LPD.
- Some tests such as arm64-neon-mul-div.ll and
ragreedy-local-interval-cost.ll have more, less or just different
spilling.
- In aarch64_generated_funcs.ll.generated.expected one part of the
function is no longer outlined. Interestingly if I switch this to use
any other scheduled even less is outlined.
Some of these are expected to happen, such as differences in outlining
or register spilling. There will be places where these result in worse
codegen, places where they are better, with the SPEC instruction counts
suggesting it is not a decrease overall, on average.
Differential Revision: https://reviews.llvm.org/D110830
Split off ABS cost handling from MIN/MAX and use explicit predicates for each
Our generic expansion of ABS doesn't use NEG+CMP+SELECT any more (its now ASHR+ADD+XOR) so this needs to be updated.
Fix copy+pasta that was checking for smul_fix instead of smul_with_overflow to detected signed values.
The LShr is performed on the extended type as we use it to truncate+extract the upper/hi bits of the extended multiply.
More closely matches the default expansion from TargetLowering::expandMULO
As suggested on D111024, we should treat getCmpSelInstrCost calls without a specific predicate as matching the worst case predicate cost.
These regressions will be addressed with a mixture of D111024 and fixing other specific getCmpSelInstrCost calls to have realistic predicates.
The coverage could have cumulative explosion here,
so i'm adding only the most basic cases,
and hoping it's enough, though more can be added if needed.
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/1jfGddcre - for intels `Block RThroughput: =36.0`; for ryzens, `Block RThroughput: =12.0`
So could pick cost of `36`
For store we have:
https://godbolt.org/z/ao9srMT8r - for intels `Block RThroughput: =30.0`; for ryzens, `Block RThroughput: =12.0`
So we could pick cost of `30`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111094
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/rc8jYxW6M - for intels `Block RThroughput: =18.0`; for ryzens, `Block RThroughput: =6.0`
So could pick cost of `18`.
For store we have:
https://godbolt.org/z/9PhPEr65G - for intels `Block RThroughput: =15.0`; for ryzens, `Block RThroughput: =6.0`
So we could pick cost of `15`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111093
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/onese7rec - for intels `Block RThroughput: =6.0`; for ryzens, `Block RThroughput: =3.0`
So could pick cost of `6`.
For store we have:
https://godbolt.org/z/bMd7dddnT - for intels `Block RThroughput: =8.0`; for ryzens, `Block RThroughput: <=6.0`
So we could pick cost of `8`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111092
Florian Hahn