This patch adds the beginnings of more thorough support in the
legalizers for vector-predicated (VP) operations.
The first step is the ability to widen illegal vectors. The more
complicated scenario in which the result/operands need widening but the
mask doesn't has not been handled here. That would require a lot of code
without an in-tree target on which to test it.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D107904
We normally select these when the root node is a sext_inreg, but
SimplifyDemandedBits can sometimes bypass the sext_inreg for some
users. This can create situation where sext_inreg+add/sub/mul/shl
is selected to a W instruction, and then the add/sub/mul/shl is
separately selected to a non-W instruction with the same inputs.
This patch tries to detect when it would still be ok to use a W
instruction without the sext_inreg by checking the direct users.
This can allow the W instruction to CSE with one created for a
sext_inreg+add/sub/mul/shl. To minimize complexity and cost of
checking, we make no attempt to determine if the CSE will happen
and just always use a W instruction when we can.
Differential Revision: https://reviews.llvm.org/D107658
Currently isReallyTriviallyReMaterializableGeneric() implementation
prevents rematerialization on any virtual register use on the grounds
that is not a trivial rematerialization and that we do not want to
extend liveranges.
It appears that LRE logic does not attempt to extend a liverange of
a source register for rematerialization so that is not an issue.
That is checked in the LiveRangeEdit::allUsesAvailableAt().
The only non-trivial aspect of it is accounting for tied-defs which
normally represent a read-modify-write operation and not rematerializable.
The test for a tied-def situation already exists in the
/CodeGen/AMDGPU/remat-vop.mir,
test_no_remat_v_cvt_f32_i32_sdwa_dst_unused_preserve.
The change has affected ARM/Thumb, Mips, RISCV, and x86. For the targets
where I more or less understand the asm it seems to reduce spilling
(as expected) or be neutral. However, it needs a review by all targets'
specialists.
Differential Revision: https://reviews.llvm.org/D106408
Follow-up to D107068, attempt to fold nested concat_vectors/undefs, as long as both the vector and inner subvector types are legal.
This exposed the same issue in ARM's MVE LowerCONCAT_VECTORS_i1 (raised as PR51365) and AArch64's performConcatVectorsCombine which both assumed concat_vectors only took 2 subvector operands.
Differential Revision: https://reviews.llvm.org/D107597
For unit-stride and strided load/stores we set the SEW operand of
the pseudo instruction equal the EEW in the opcode. The LMUL
of the pseudo instruction is the LMUL we want.
These instructions calculate EMUL=(EEW/SEW) * LMUL. We can use
this to avoid changing vtype if the SEW/LMUL of the previous
vtype matches the EEW/EMUL ratio we need for the instruction.
Due to how the global analysis works, we can only do this
optimization when the previous vsetvli was produced in the block
containing the store. We need to know in the first phase if the
vsetvli will be inserted so we can propagate information to
the successors in the second phase correctly. This means we can't
depend on predecessors.
Reviewed By: rogfer01
Differential Revision: https://reviews.llvm.org/D106601
Shuffles which are broken into separate halves reveal splats in which
a half is accessed via one index; such operations can be optimized to
use "vrgather.vi".
This optimization could be achieved by adding extra patterns to match
`vrgather_vv_vl` which uses a splat as an index operand, but this patch
instead identifies splat earlier. This way, future optimizations can
build on top of the data gathered here, e.g., to splat-gather dominant
indices and insert any leftovers.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D107449
These shuffles all take the form of a "splat" of the LHS and/or RHS to
some degree, with one or two elements needing patched up afterwards. We
currently lower all of these to full LHS/RHS vector-index shuffles with
vrgather.vv.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D107447
We don't have real demanded bits support for MULHU, but we can
still use the known bits based constant folding support at the end
of SimplifyDemandedBits to simplify a MULHU. This helps with cases
where we know the LHS and RHS have enough leading zeros so that
the high multiply result is always 0.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D106471
This patch extends the optimization of VID-sequence BUILD_VECTORs
introduced in D104921 to include simple fractional steps composed of a
separated integer numerator and denominator.
A notable limitation in this sequence detection is that only sequences
with steps N/1 or 1/D are found, meaning that the step between elements
and the frequency with which it changes is consistent across the whole
sequence. Fractional steps such as 2/3 won't be matched as those would
involve more complex tracking of state or some level of backtracking.
As is stands, however, this patch is sufficient to match common
interleave-type shuffle indices, for example matching `<0,0,1,1>` (or
commonly `<0,u,1,u>` or `<u,0,u,1>`) to an index sequence divided by 2.
While the optimization is relatively `undef`-tolerant, due to greedy
pattern-matching there even are some simple patterns which confuse the
sequence detection into identifying either a suboptimal sequence or no
sequence at all.
Currently only fractional-step sequences identified as having a
power-of-two denominator are actually lowered to RVV instructions. This
is to avoid introducing divisions into the generated code.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106533
If a vsetvli instruction is not compatible with the next vector instruction,
and there is no other things that may update or use VL/VTYPE, we could merge
it with the next vsetvli instruction that should be insert for the vector
instruction.
This commit only merge VTYPE with the former vsetvli instruction which has
the same VL.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106857
Currently, the default alignment is much larger than the actual size of
the vector in memory. Fix this to use a sane default.
For SVE, temporarily remove lowering of load/store operations for
predicates with less than 16 elements. The layout the backend was
assuming for SVE predicates with less than 16 elements doesn't agree
with the frontend. More work probably needs to be done here.
This change is, strictly speaking, not backwards-compatible at the
bitcode level. But probably nobody is actually depending on that; i1
vectors in memory are rare, and the code that does use them probably
ends up forcing the alignment to something sane anyway. If we think
this is a concern, I can restrict this to scalable vectors for now
(where it's actually causing issues for me at the moment).
Differential Revision: https://reviews.llvm.org/D88994
This patch aims to improve the performance of BUILD_VECTORs which are
identified as containing a dominant element. Given that most
floating-point constants themselves require a load from the constant
pool, it was possible for the optimization to actually increase the
number of individual loads on small vectors. The exception is the zero
constant -- +0.0 -- which can be materialized efficiently.
While this optimization could do with a proper cost model to weigh the
benfits of a single vector load vs. the manipulation of individual
elements -- even for integer vectors which often require several
instructions to materialize -- without a concrete RVV implementation to
work with any heuristic is likely to be both more obtuse and inaccurate.
Until then, this patch fixes at least one known obvious deficiency.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106963
The second test case added here was pointed out to me by @craig.topper
and shows how we "optimize" a two-element BUILD_VECTOR from being one
load from the constant pool to two loads from the constant pool.
The first test case shows that since materialization for the
floating-point +0.0 value is cheap and doesn't involve a load, the
optimization is more clearly beneficial here.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106962
This patch builds on top of D106575 in which scalable-vector splats were
supported in `ISD::matchBinaryPredicate`. It teaches the DAGCombiner how
to perform a variety of the pre-existing saturating add/sub combines on
scalable-vector types.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106652
This patch adds support for lowering the saturating vector add/sub
intrinsics to RVV instructions, for both fixed-length and
scalable-vector forms alike.
Note that some of the DAG combines are still not triggering for the
scalable-vector tests. These require a bit more work in the DAGCombiner
itself.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106651
These will be optimized by upcoming patches. The tests are primarily not
being optimized due to the lack of support for saturating vector
arithmetic in the RISC-V backend.
On top of that, however, a large percentage of the scalable-vector tests
are also lacking support in the DAGCombiner: either in
`ISD::matchBinaryPredicate` or due to checks specifically for
`BUILD_VECTOR` and not `SPLAT_VECTOR`.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106649
This patch extends support for (scalable-vector) splats in the
DAGCombiner via the `ISD::matchBinaryPredicate` function, which enable a
variety of simple combines of constants.
Users of this function may now have to distinguish between
`BUILD_VECTOR` and `SPLAT_VECTOR` vector operands. The way of dealing
with this in-tree follows the approach added for
`ISD::matchUnaryPredicate` implemented in D94501.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106575
This patch adds a reduced test case which identifies an illegal vsetvli
inserted by the compiler. The compiler emits a vsetvli which is intended
to preserve VL with the SEW/LMUL ratio e32/m1 when in fact the VL could
have been set by e64/m2 in a predecessor block.
Differential Revision: https://reviews.llvm.org/D106286
Since we're changing VTYPE, we may change VLMAX which could
invalidate the previous VL. If we can't tell if it is safe we
should use an AVL of 1 instead of keeping the old VL.
This is a quick fix. We may want to thread VL to the pseudo
instruction instead of making up a value. That will require ISD
opcode changes and changes to the C intrinsic interface.
This fixes the issue raised in D106286.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D106403
These tests show missed opportunities in the SelectionDAG layer when
dealing with scalable-vector splats. All of these are handled for the
equivalent `ISD::BUILD_VECTOR` code, and the tests have largely been
translated from the equivalent X86 tests.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D106574
This relands a6ca88e908 which was originally
reverted due to overflow bugs in e3fa2b1eab.
This patch teaches the compiler to identify a wider variety of
`BUILD_VECTOR`s which form integer arithmetic sequences, and to lower
them to `vid.v` with modifications for non-unit steps and non-zero
addends.
The sequences handled by this optimization must either be monotonically
increasing or decreasing. Consecutive elements holding the same value
indicate a fractional step which, while simple mathematically,
becomes more complex to handle both in the realm of lossy integer
division and in the presence of `undef`s.
For example, a common "interleaving" shuffle index will be lowered by
LLVM to both `<0,u,1,u,2,...>` and `<u,0,u,1,u,...>` `BUILD_VECTOR`
nodes. Either of these would ideally be lowered to `vid.v` shifted right
by 1. Detection of this sequence in presence of general `undef` values
is more complicated, however: `<0,u,u,1,>` could match either
`<0,0,0,1,>` or `<0,0,1,1,>` depending on later values in the sequence.
Both are possible, so backtracking or multiple passes is inevitable.
Sticking to monotonic sequences keeps the logic simpler as it can be
done in one pass. Fractional steps will likely be a separate
optimization in a future patch.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D104921
Prior to this patch, it skipped the instruction defining VNI when checking if the tainted lanes are used.
In the given example, VRGATHER is an illegal instruction because its DstReg overlaps with SrcReg.
Therefore we need to check the defining instruction as well when there is an earlyclobber constraint.
Reviewed By: qcolombet
Differential Revision: https://reviews.llvm.org/D105684
The existing rule about the operand type is strange. Instead, just say
the operand is a TargetConstant with the right width. (Legalization
ignores TargetConstants, so it doesn't matter if that width is legal.)
Highlights:
1. I had to substantially rewrite the AArch64 isel patterns to expect a
TargetConstant. Nothing too exotic, but maybe a little hairy. Maybe
worth considering a target-specific node with some dagcombines instead
of this complicated nest of isel patterns.
2. Our behavior on RV32 for vectors of i64 has changed slightly. In
particular, we correctly preserve the width of the arithmetic through
legalization. This changes the DAG a bit. Maybe room for
improvement here.
3. I explicitly defined the behavior around overflow. This is necessary
to make the DAGCombine transforms legal, and I don't think it causes any
practical issues.
Differential Revision: https://reviews.llvm.org/D105673
If we need to shift left anyway we might be able to take advantage
of LUI implicitly shifting its immediate left by 12 to cover part
of the shift. This allows us to use more bits of the LUI immediate
to avoid an ADDI.
isDesirableToCommuteWithShift now considers compressed instruction
opportunities when deciding if commuting should be allowed.
I believe this is the same or similar to one of the optimizations
from D79492.
Reviewed By: luismarques, arcbbb
Differential Revision: https://reviews.llvm.org/D105417
I don't think the semantics of the llvm masked gather intrinsic care
about the order the elements are loaded. For example, type legalization
by splitting will chain them in parallel. This is different than
scatter which we do chain in order.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D106025
We assume VLENB is a multiple of 8 and previously relied on shift
pairs being optimized to an AND+SHL/SHR and computeKnownBits
removing the AND. This doesn't happen if (vlenb >> 3) gets CSEd
to have multiple uses. This patch manually emits the best shift
to workaround this.
This adds new pseudoinstructions with ForceTailAgnostic set. This
matches what we did for non-widening VMACC. We should move to a
tail policy operand on the pseudos when we expand the intrinsic
interface to include the tail policy.
This patch teaches the compiler to identify a wider variety of
`BUILD_VECTOR`s which form integer arithmetic sequences, and to lower
them to `vid.v` with modifications for non-unit steps and non-zero
addends.
The sequences handled by this optimization must either be monotonically
increasing or decreasing. Consecutive elements holding the same value
indicate a fractional step which, while simple mathematically,
becomes more complex to handle both in the realm of lossy integer
division and in the presence of `undef`s.
For example, a common "interleaving" shuffle index will be lowered by
LLVM to both `<0,u,1,u,2,...>` and `<u,0,u,1,u,...>` `BUILD_VECTOR`
nodes. Either of these would ideally be lowered to `vid.v` shifted right
by 1. Detection of this sequence in presence of general `undef` values
is more complicated, however: `<0,u,u,1,>` could match either
`<0,0,0,1,>` or `<0,0,1,1,>` depending on later values in the sequence.
Both are possible, so backtracking or multiple passes is inevitable.
Sticking to monotonic sequences keeps the logic simpler as it can be
done in one pass. Fractional steps will likely be a separate
optimization in a future patch.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D104921
Using positive zero as the neutral element in 'fadd' reductions, while
it generates better code, is incorrect. The correct neutral element is
negative zero: 0.0 + -0.0 = 0.0, whereas -0.0 + -0.0 = -0.0.
There are perhaps more optimal lowerings of negative zero avoiding
constant-pool loads which could be left as future work.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D105902
Often when lowering vector shuffles, we split the shuffle into two
LHS/RHS shuffles which are then blended together. To do so we split the
original indices into two, indexed into each respective vector. These
two index vectors are then separately lowered as BUILD_VECTORs.
This patch forwards on any undef indices to the BUILD_VECTOR, rather
than having the VECTOR_SHUFFLE lowering decide on an optimal concrete
index. The motiviation for ths change is so that we don't duplicate
optimization logic between the two lowering methods and let BUILD_VECTOR
do what it does best.
Propagating undef in this way allows us, for example, to generate
`vid.v` to produce the LHS indices of commonly-used interleave-type
shuffles. I have designs on further optimizing interleave-type and other
common shuffle patterns in the near future.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D104789
This adds a DAG combine to detect sext/zext inputs and emit a
new ISD opcode. The extends will either be removed or replaced
with narrower extends.
Isel patterns are used to match add and widening mul to vwmacc
similar to the recently added vmacc patterns.
There's still some work to be to match vmulsu.
We should also rewrite splats that were extended as scalars and
then splatted.
Reviewed By: arcbbb
Differential Revision: https://reviews.llvm.org/D104802
This will currently accept the old number of bytes syntax, and convert
it to a scalar. This should be removed in the near future (I think I
converted all of the tests already, but likely missed a few).
Not sure what the exact syntax and policy should be. We can continue
printing the number of bytes for non-generic instructions to avoid
test churn and only allow non-scalar types for generic instructions.
This will currently print the LLT in parentheses, but accept parsing
the existing integers and implicitly converting to scalar. The
parentheses are a bit ugly, but the parser logic seems unable to deal
without either parentheses or some keyword to indicate the start of a
type.
I thought this might help with another optimization I was
thinking about, but I don't think it will. So it just wastes
compile time calling computeKnownBits for no benefit.
This reverts commit 81b2f95971.
This patch teaches the compiler to generate code to handle larger RVV
stack sizes and stack offsets which resolve an amount larger than 2047
vector registers in size.
The previous behaviour was asserting on such large values as it was only
able to materialize the constant by feeding it to the 12-bit immediate
of an `ADDI` instruction. The compiler can now materialize this amount
into a temporary register before continuing with the computation.
A test case for this scenario is included which also checks that the
temporary register used to materialize the amount doesn't require an
additional spill slot over what we're already reserving for RVV code.
Reviewed By: rogfer01
Differential Revision: https://reviews.llvm.org/D104727
This patch optimizes the code generation of vector-type SELECTs (LLVM
select instructions with scalar conditions) by custom-lowering to
VSELECTs (LLVM select instructions with vector conditions) by splatting
the condition to a vector. This avoids the default expansion path which
would either introduce control flow or fully scalarize.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D104772
With regards to overrunning, the langref (llvm/docs/LangRef.rst)
specifies:
(llvm.experimental.vector.insert)
Elements ``idx`` through (``idx`` + num_elements(``subvec``) - 1)
must be valid ``vec`` indices. If this condition cannot be determined
statically but is false at runtime, then the result vector is
undefined.
(llvm.experimental.vector.extract)
Elements ``idx`` through (``idx`` + num_elements(result_type) - 1)
must be valid vector indices. If this condition cannot be determined
statically but is false at runtime, then the result vector is
undefined.
For the non-mixed cases (e.g. inserting/extracting a scalable into/from
another scalable, or inserting/extracting a fixed into/from another
fixed), it is possible to statically check whether or not the above
conditions are met. This was previously missing from the verifier, and
if the conditions were found to be false, the result of the
insertion/extraction would be replaced with an undef.
With regards to invalid indices, the langref (llvm/docs/LangRef.rst)
specifies:
(llvm.experimental.vector.insert)
``idx`` represents the starting element number at which ``subvec``
will be inserted. ``idx`` must be a constant multiple of
``subvec``'s known minimum vector length.
(llvm.experimental.vector.extract)
The ``idx`` specifies the starting element number within ``vec``
from which a subvector is extracted. ``idx`` must be a constant
multiple of the known-minimum vector length of the result type.
Similarly, these conditions were not previously enforced in the
verifier. In some circumstances, invalid indices were permitted
silently, and in other circumstances, an undef was spawned where a
verifier error would have been preferred.
This commit adds verifier checks to enforce the constraints above.
Differential Revision: https://reviews.llvm.org/D104468
Previously we went directly to unknown state on VTYPE mismatch.
If we instead remember the partial match, we can use this to
still use X0, X0 vsetvli in successors if AVL and needed SEW/LMUL
ratio match.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D104069
Fraser Cormack