A previous patch moved the index versions. This moves the rest.
I also removed the custom lowering for VLEFF since we can now
do everything directly in the isel handling.
I had to update getLMUL to handle mask registers to index the
pseudo table correctly for VLE1/VSE1.
This is good for another 15K reduction in llc size.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D97097
We had more combinations of data and index lmuls than we needed.
Also add some asserts to verify that the IndexVT and data VT have
the same element count when we isel these pseudo instructions.
There are many legal combinations of index and data VTs supported
for these intrinsics. This results in a lot of isel patterns in
RISCVGenDAGISel.inc.
By adding a separate table similar to what we use for segment
load/stores, we can more efficiently manually select these
intrinsics. We should also be able to reuse this table scalable
vector gather/scatter.
This reduces the llc binary size by ~56K.
Reviewed By: khchen
Differential Revision: https://reviews.llvm.org/D97033
We were creating more combinations of value and index lmul than
we needed.
I've copied the loop structure used here from VPseudoAMOEI with
all data sew values instead of just 32/64.
Similar can be done for segment loads/store.
Reviewed By: khchen
Differential Revision: https://reviews.llvm.org/D97008
Intrinsic ID is a 32-bit value which made each row of the table 4
byte aligned. The remaining fields used 5 bytes. This meant 3 bytes
of padding per row.
This patch breaks the table into 4 separate tables and indexes them
by properties we know about the intrinsic. NF, masked,
strided, ordered, etc. The indexed load/store tables have no
padding in their rows now.
All together this reduces the size of llc binary by ~28K.
I'm considering adding similar tables for isel of non-segment
load/store as well to cut down the size of the isel table and
probably improve our isel performance. Those tables would need to
indexed from intrinsics, IR loads/stores, gathers/scatters, and
RISCVISD opcodes. So having a table that can be indexed without using
intrinsic ID is more flexible.
Reviewed By: HsiangKai
Differential Revision: https://reviews.llvm.org/D96894
This table is queried in RISCVMCInstLower without knowing
whether the instruction is a vector pseudo. Due to the way the
binary search works, we have to do log2(tablesize) checks just
to determine a non-vector instruction isn't in the table.
Conveniently, all the vector pseudos are pretty tightly
packed within the internal instruction enum. By enabling the
PrimaryKeyEarlyOut, tablegen will emit a check against the
beginning and end of the table before doing the binary search.
This gives a quick early out on the search for the majority
of non-vector instructions.
Differential Revision: https://reviews.llvm.org/D97016
For masked segment load, the destination register should not overlap
with mask register. It could not be V0.
In the original implementation, there is no segment load/store register
class without V0. In this patch, I added these register classes and
modify `GetVRegNoV0` to get the correct one.
Differential Revision: https://reviews.llvm.org/D96937
Unlike scalable vectors, I'm only using a ComplexPattern for
the immediate itself. The vmv_v_x is matched explicitly. We igore
the VL argument when matching a binary operator, but we do check
it when matching splat directly.
I left out tests for vXi64 as they fail on rv32 right now.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D96365
This refines how we determine which masks types are legal and adds
support for loads, stores, and all ones/zeros splats.
I left a fixme in store handling where I think we need to zero
extra bits if the type isn't a multiple of a byte. If I remember
right from X86 there was some case we could have a store of a
1, 2, or 4 bit mask and have a scalar zextload that then expected the
bits to be 0. Its tricky to zero the bits with RVV. We need to do
something like round VL up, zero a register, lower the VL back down,
then do a tail undisturbed move into the zero register. Another
option might be to generate a mask of 1/2/4 bits set with a VL of 8
and use that to mask off the bits.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D96468
In vector v0.10, there are whole vector register load/store
instructions. I suggest to use the whole register load/store
instructions for generic load/store for scalable vector types. It could
save up vset{i}vl{i} for these load/store.
For fractional LMUL, I keep to use vle{eew}.v/vse{eew}.v instructions to
load/store partial vector registers.
Differential Revision: https://reviews.llvm.org/D95853
Building on the fixed vector support from D95705
I've added ISD nodes for vmv.v.x and vfmv.v.f and switched to
lowering the intrinsics to it. This allows us to share the same
isel patterns for both.
This doesn't handle splats of i64 on RV32 yet. The build_vector
gets converted to a vXi32 build_vector+bitcast during type
legalization. Not sure the best way to handle this at the moment.
Differential Revision: https://reviews.llvm.org/D96108
This is an alternative to D95563.
This is modeled after a similar feature for AArch64's SVE that uses
predicated scalable vector instructions.a
Rather than use predication, this patch uses an explicit VL operand.
I've limited it to always use LMUL=1 for now, but we can improve this
in the future.
This requires a bunch of new ISD opcodes to carry the VL operand.
I think we can probably lower intrinsics to these ISD opcodes to
cut down on the size of the isel table. Which is why I've added
patterns for all integer/float types and not just LMUL=1.
I'm only testing one vector width right now, but the width is
programmable via the command line.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D95705
This adds support for commuting operands and converting between
vfmadd and vfmacc to avoid register copies.
To avoid messing up intrinsic behavior, I've added new pseudo
instructions that have the isCommutable flag set. These pseudos also
force a tail agnostic policy. The intrinsic version still use
the tail undisturbed policy.
For best results it looks like we need to start with fmadd and only
pick fmacc if its beneficial. MachineCSE commutes without contraining
the operands and then commutes back if it didn't help with CSE. So
I've made sure that when the operand choice isn't constrained, we
will keep fmadd for MachineCSE and when it does the second commute,
we get back the original instruction.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D95800
The vrgather.vv instruction uses a vector of indices with the same
SEW as operand 0. The vrgather.vx instructions use a scalar index
operand of XLen bits.
By splitting this into 2 intrinsics we are able to use LLVMatchType
in the definition to avoid specifying the type for the index operand
when creating the IR for the intrinsic. For .vv it will match the
operand 0 type. And for .vx it will match the type of the vl operand
we already needed to specify a type for.
I'm considering splitting more intrinsics. This was a somewhat
odd one because the .vx doesn't use the element type, it always
use XLen.
Reviewed By: HsiangKai
Differential Revision: https://reviews.llvm.org/D95979
Not all combinations of SEW and LMUL we need to support. For example, we
only need to support [M1, M2, M4, M8] for SEW = 64. There is no need to
define pseudos for PseudoVLSE64MF8, PseudoVLSE64MF4, and PseudoVLSE64MF2.
Differential Revision: https://reviews.llvm.org/D95667
-Remove the ISD opcode for READ_VL. Just emit the MachineSDNode directly.
-Move segmented fault first only load intrinsic handling completely to
RISCVISelDAGToDAG.cpp and emit the ReadVL MachineSDNode there
instead of lowering to ISD opcodes first.
Move the Suffix string into the VTypeInfo class so we don't need a helper class to get to it.
Adjust pseudo naming scheme for FPRs to put F16/F32/F64 in
place of F in the pseudo instruction name rather than as a suffix.
This avoids special cases like VFMERGE from the original patch.
Differential Revision: https://reviews.llvm.org/D95404
When spilling, the spill size will depend on the size of register class.
For .vf vector instructions, it may spill the floating point scalar
argument. In order to use the correct load/store instructions for
spilling, we need to provide the correct floating point register class
for the .vf vector pseudo instructions.
In this commit, we define the .vf pseudo instructions as three
different kinds of pseudo instructions for half/float/double. For
example, PseudoVFADD_M1 will become as PseudoVFADD_F16_M1,
PseudoVFADD_F32_M1, and PseudoVFADD_F64_M1.
Differential Revision: https://reviews.llvm.org/D95234
These instructions produce 2*SEW result so the input can't have
an LMUL=8 or the result would need a non-existant LMUL=16. So
only create pseudos for LMUL up to 4.
Differential Revision: https://reviews.llvm.org/D95189
The fault-only-first-load instructions can reduce VL if an element
other than element 0 triggers a memory fault. This can be used to
vectorize loops with data dependent exit conditions like strcmp or
strlen.
This patch adds a VL output to these intrinsics so that the new
VL value can be captured by software. This will be expanded to
'csrr gpr, vl' after the vleff instruction during SelectionDAG.
By doing this with one intrinsic we are able to guarantee that the
csrr reads the VL value produced by the vleff instruction. Having
it as a separate intrinsic would make it impossible to guarantee
ordering without making every other vector intrinsic have side
effects.
The intrinsics are expanded during lowering into two ISD nodes
that are glued together. These ISD nodes will go
through isel separately, but should maintain the glue so that they
get emitted adjacently by InstrEmitter.
I've only ran the chain through the vleff instruction, allowing
the READ_VL to be deleted if it is unused.
Reviewed By: HsiangKai
Differential Revision: https://reviews.llvm.org/D94286
Upgrade RISC-V V extension to v1.0-08a0b46.
Indexed load/store have ordered and unordered form.
New whole vector load/store.
Differential Revision: https://reviews.llvm.org/D93614
For Zvlsseg, we need continuous vector registers for the values. We need
to define new register classes for the different combinations of (number
of fields and LMUL). For example,
when the number of fields(NF) = 3, LMUL = 2, the values will be assigned
to (V0M2, V2M2, V4M2), (V2M2, V4M2, V6M2), (V4M2, V6M2, V8M2), ...
We define the vlseg intrinsics with multiple outputs. There is no way to
describe the codegen patterns with multiple outputs in the tablegen
files. We do the codegen in RISCVISelDAGToDAG and use EXTRACT_SUBREG to
extract the values of output.
The multiple scalable vector values will be put into a struct. This
patch is depended on the support for scalable vector struct.
Differential Revision: https://reviews.llvm.org/D94229
Make it easier to reuse for intrinsic vrgatherei16
which needs to encode both LMUL & EMUL in the instruction name,
like PseudoVRGATHEREI16_VV_M1_M1 and PseudoVRGATHEREI16_VV_M1_M2.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D94951
This patch factors out the "VLMax" operand passed to most
scalable-vector ISel patterns into a property of each VType.
This is seen as a preparatory change to allow RVV in the future to
more easily support fixed-length vector types with constrained vector
lengths, with the AVL operand set to the length of the fixed-length
vector. It has no effect on the scalable code generation path.
Reviewed By: HsiangKai
Differential Revision: https://reviews.llvm.org/D94594
This patch custom lowers ISD::VSCALE into a csrr vlenb followed
by a shift right by 3 followed by a multiply by the scale amount.
I've added computeKnownBits support to indicate that the csrr vlenb
always produces 3 trailng bits of 0s so the shift right is "exact".
This allows the shift and multiply sequence to be nicely optimized
into a single shift or removed completely when the scale amount is
a power of 2.
The non power of 2 case multiplying by 24 is still producing
suboptimal code. We could remove the right shift and use a
multiply by 3. Hopefully we can improve DAG combine to fix that
since it's not unique to this sequence.
This replaces D94144.
Reviewed By: HsiangKai
Differential Revision: https://reviews.llvm.org/D94249
We can use a 0 immediate to avoid needing to materialize 0 into
an FPR first.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D94459
Define the `vfclass` IR intrinsics for the respective V instructions.
Authored-by: Roger Ferrer Ibanez <rofirrim@gmail.com>
Co-Authored-by: Evandro Menezes <evandro.menezes@sifive.com>
Differential Revision: https://reviews.llvm.org/D94356