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
The custom expansion of select operations in the RISC-V backend
interferes with the matching of cmov instructions. Legalizing
select when the Zbt extension is available solves that problem.
Reviewed By: lenary, craig.topper
Differential Revision: https://reviews.llvm.org/D93767
Original patch by @rogfer01.
All ordered comparisons except ONE are supported natively, and all
unordered comparisons except UNE are expanded into sequences involving
explicit NaN checks and mask arithmetic.
Additionally, we expand GT,OGT,GE,OGE to their swapped-operand versions, and
pattern-match those back to the "original", swapping operands once more. This
way we catch both operations and both "vf" and "fv" forms with fewer patterns.
Also add support for floating-point splat_vector, with an optimization for
splatting fpimm0.
Authored-by: Roger Ferrer Ibanez <rofirrim@gmail.com>
Co-Authored-by: Fraser Cormack <fraser@codeplay.com>
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D94242
This patch moves all but the BaseInstr to bits in TSFlags.
For the index fields, we can just use a bit to indicate their presence.
The locations of the operands are well defined.
This reduces the llc binary by about 32K on my build. It also
removes the binary search of the table from the custom inserter.
Instead we just check that the SEW op is present.
Reviewed By: rogfer01
Differential Revision: https://reviews.llvm.org/D94375
1. Break MUL with specific constant to a SLLI and an ADD/SUB on riscv32
with the M extension.
2. Break MUL with specific constant to two SLLI and an ADD/SUB, if the
constant needs a pair of LUI/ADDI to construct.
Reviewed by: craig.topper
Differential Revision: https://reviews.llvm.org/D93619
nvxXi1 types are legal with V extension and that's the result
vmseq/vmsne/vmslt/etc instructions return.
No test cases yet because the setcc isel patterns aren't in
and we'll need more than basic tests to observe this. I locally
tested that this plus D947078, D94168, D94142, and D94149
was enough to be able to handle the overflow result from
llvm.sadd.overflow.
If the destination is tied, then user has some control of the
register used for input. They would have the ability to control
the value of any tail elements. By using tail agnostic we take
this option away from them.
Its not clear that the intrinsics are defined such that this isn't
supposed to work. And undisturbed is a valid implementation for agnostic
so code wouldn't even fail to work on all systems if we always used
agnostic.
The vcompress intrinsic is defined to require tail undisturbed so
at minimum we need this for that instruction or need to redefine
the intrinsic.
I've made an exception here for vmv.s.x/fmv.s.f and reduction
instructions which only write to element 0 regardless of the tail
policy. This allows us to keep the agnostic policy on those which
should allow better redundant vsetvli removal.
An enhancement would be to check for undef input and keep the
agnostic policy, but we don't have good test coverage for that yet.
Reviewed By: khchen
Differential Revision: https://reviews.llvm.org/D93878
This patch extends the SDNode ISel support for RVV from only the
vector/vector instructions to include the vector/scalar and
vector/immediate forms.
It uses splat_vector to carry the scalar in each case, except when
XLEN<SEW (RV32 SEW=64) when a custom node `SPLAT_VECTOR_I64` is used for
type-legalization and to encode the fact that the value is sign-extended
to SEW. When the scalar is a full 64-bit value we use a sequence to
materialize the constant into the vector register.
The non-intrinsic ISel patterns have also been split into their own
file.
Authored-by: Roger Ferrer Ibanez <rofirrim@gmail.com>
Co-Authored-by: Fraser Cormack <fraser@codeplay.com>
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D93312
This patch enables jump table lowering in the RISC-V backend.
In addition to the test case included, the new lowering was
tested by compiling the OCaml runtime and running it under qemu.
Differential Revision: https://reviews.llvm.org/D92097
The default behavior for any_extend of a constant is to zero extend.
This occurs inside of getNode rather than allowing type legalization
to promote the constant which would sign extend. By using sign extend
with getNode the constant will be sign extended. This gives a better
chance for isel to find a simm5 immediate since all xlen bits are
examined there.
For instructions that use a uimm5 immediate, this change only affects
constants >= 128 for i8 or >= 32768 for i16. Constants that large
already wouldn't have been eligible for uimm5 and would need to use a
scalar register.
If the instruction isn't able to use simm5 or the immediate is
too large, we'll need to materialize the immediate in a register.
As far as I know constants with all 1s in the upper bits should
materialize as well or better than all 0s.
Longer term we should probably have a SEW aware PatFrag to ignore
the bits above SEW before checking simm5.
I updated about half the test cases in some tests to use a negative
constant to get coverage for this.
Reviewed By: evandro
Differential Revision: https://reviews.llvm.org/D93487
This adds intrinsics for vmv.x.s and vmv.s.x.
I've used stricter type constraints on these intrinsics than what we've been doing on the arithmetic intrinsics so far. This will allow us to not need to pass the scalar type to the Intrinsic::getDeclaration call when creating these intrinsics.
A custom ISD is used for vmv.x.s in order to implement the change in computeNumSignBitsForTargetNode which can remove sign extends on the result.
I also modified the MC layer description of these instructions to show the tied source/dest operand. This is different than what we do for masked instructions where we drop the tied source operand when converting to MC. But it is a more accurate description of the instruction. We can't do this for masked instructions since we use the same MC instruction for masked and unmasked. Tools like llvm-mca operate in the MC layer and rely on ins/outs and Uses/Defs for analysis so I don't know if we'll be able to maintain the current behavior for masked instructions. So I went with the accurate description here since it was easy.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D93365
We work with @rogfer01 from BSC to come out this patch.
Authored-by: Roger Ferrer Ibanez <rofirrim@gmail.com>
Co-Authored-by: ShihPo Hung <shihpo.hung@sifive.com>
Co-Authored-by: Monk Chiang <monk.chiang@sifive.com>
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D93366
If users want to use vector floating point instructions, they need to
specify 'F' extension additionally.
Differential Revision: https://reviews.llvm.org/D93282
The compiler is making no effort to preserve upper elements. To do so would require another source operand tied with the destination and a different intrinsic interface to give control of this source to the programmer.
This patch changes the tail policy to agnostic so that the CPU doesn't need to make an effort to preserve them.
This is consistent with the RVV intrinsic spec here https://github.com/riscv/rvv-intrinsic-doc/blob/master/rvv-intrinsic-rfc.md#configuration-setting
Differential Revision: https://reviews.llvm.org/D93080
If SETUNE isn't legal, UO can use the NOT of the SETO expansion.
Removes some complex isel patterns. Most of the test changes are
from using XORI instead of SEQZ.
Differential Revision: https://reviews.llvm.org/D92008
The register operand was not being marked as a def when it should be. No tests
for this in the main branch as there are not yet any pseudos without a
non-negative VLIndex.
Also change the type of a virtual register operand from unsigned to Register
and adjust formatting.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D92823
This merges the SEW and LMUL enums that each used into singles enums in RISCVBaseInfo.h. The patch also adds a new encoding helper to take SEW, LMUL, tail agnostic, mask agnostic and turn it into a vtype immediate.
I also stopped storing the Encoding in the VTYPE operand in the assembler. It is easy to calculate when adding the operand which should only happen once per instruction.
Differential Revision: https://reviews.llvm.org/D92813
A rotate by half the bitwidth swaps the bottom and top half which is the same as one of the MSB GREVI stage.
We have to do this as a special combine because we prefer to keep (rotl/rotr X, BitWidth/2) as a rotate rather than a single stage GREVI.
Differential Revision: https://reviews.llvm.org/D92286
The companion RFC (http://lists.llvm.org/pipermail/llvm-dev/2020-October/145850.html) gives lots of details on the overall strategy, but we summarize it here:
LLVM IR involving vector types is going to be selected using pseudo instructions (only MachineInstr). These pseudo instructions contain dummy operands to represent the vector type being operated and the vector length for the operation.
These two dummy operands, as set by instruction selection, will be used by the custom inserter to prepend every operation with an appropriate vsetvli instruction that ensures the vector architecture is properly configured for the operation. Not in this patch: later passes will remove the redundant vsetvli instructions.
Register classes of tuples of vector registers are used to represent vector register groups (LMUL > 1).
Those pseudos are eventually lowered into the actual instructions when emitting the MCInsts.
About the patch:
Because there is a bit of initial infrastructure required, this is the minimal patch that allows us to select instructions for 3 LLVM IR instructions: load, add and store vectors of integers. LLVM IR operations have "whole-vector" semantics (as in they generate values for all the elements).
Later patches will extend the information represented in TableGen.
Authored-by: Roger Ferrer Ibanez <rofirrim@gmail.com>
Co-Authored-by: Evandro Menezes <evandro.menezes@sifive.com>
Co-Authored-by: Craig Topper <craig.topper@sifive.com>
Differential Revision: https://reviews.llvm.org/D89449
Rather than having a different opcode for RV32 and RV64. Let's just say the integer type is XLenVT and use a single opcode for both modes.
Differential Revision: https://reviews.llvm.org/D92538
This enables bswap/bitreverse to combine with other GREVI patterns or each other without needing to add more special cases to the DAG combine or new DAG combines.
I've also enabled the existing GREVI combine for GREVIW so that it can pick up the i32 bswap/bitreverse on RV64 after they've been type legalized to GREVIW.
Differential Revision: https://reviews.llvm.org/D92253
GORCI performs an OR between each stage. So we need to ensure only
one stage is active before doing this combine.
Initial attempts at finding a test case for this failed due to
the order things get combined. It's most likely that we'll form
one stage of GREVI then combine to GORCI before the two stages of
GREVI are able to be formed and combined with each other to form
a multi stage GREVI.
Differential Revision: https://reviews.llvm.org/D92289
Start with an assumption that FMA is faster than Fmul+FAdd. If thats not true
on some particular implementation we can add a tuning parameter in the future.
I've update the fmuladd test cases and added new test cases for fast math flag
based contraction.
Differential Revision: https://reviews.llvm.org/D91987
This is the logically correct thing to do. But it generates worse
code for i32 umin/umax on the rv64 due to type legalize requesting
zext even though the arguments are sext. Maybe we can teach type
legalizer to use sext for umin/umax for RISCV.
It's also producing possibly worse code on i64 on RV32 since we
still end up with selects that become branches. But this seems
like something we could improve in type legalization or DAG combine.
Hopefully this makes D92095 work for RISCV with Zbb.
This adds custom opcodes for FSLW/FSRW so we can type legalize
fshl/fshr without needing to match a sign_extend_inreg.
I've used the operand order from fshl/fshr to make the isel
pattern similar to the non-W form. It was also hard to decide
another order since the register instruction has the shift amount
as the second operand, but the immediate instruction has it as
the third operand.
Differential Revision: https://reviews.llvm.org/D91479
This is a special calling convention to be used by the GHC compiler.
Patch by Andreas Schwab (schwab)
Differential Revision: https://reviews.llvm.org/D89788
We generate two 4 byte loads or two stores as part of the expansion.
Previously the MemOperand was set the same for both to cover the
full 8 bytes. Now we set a separate 4 byte mem operand for each
with a 4 byte offset for the high part.
Previously we required a sra to pattern match these properly in isel. If the consumer didn't need the result sign extended we'll have an srl instead of sra and fail to match.
This patch switches to custom legalizing to GREVIW using portions of D91259.
Differential Revision: https://reviews.llvm.org/D91457
This should result in better utilization of RORIW since we
don't need to look for a SIGN_EXTEND_INREG that may not exist.
Also remove rotl/rotr isel matching to GREVI and just prefer RORI.
This is to keep consistency so we don't have to match ROLW/RORW
to GREVIW as well. I imagine RORI/RORIW performance will be the
same or better than GREVI.
Differential Revision: https://reviews.llvm.org/D91449
This moves the recognition of GREVI and GORCI from TableGen patterns
into a DAGCombine. This is done primarily to match "deeper" patterns in
the future, like (grevi (grevi x, 1) 2) -> (grevi x, 3).
TableGen is not best suited to matching patterns such as these as the compile
time of the DAG matchers quickly gets out of hand due to the expansion of
commutative permutations.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D91259
Similar to the X86 and AMDGPU targets, this uses a macro to cut down on
repetitive and error-prone code when converting RISCVISD node names to
strings in getTargetNodeName.
Reviewed By: asb
Differential Revision: https://reviews.llvm.org/D91414
We were creating RISCVISD::SELECT_CC nodes with Glue output that was never being used, and the tablegen SDNode had the SDNPInGlue flag instead of the SDNPOutGlue flag.
Since we don't seem to need the Glue just get rid of it from both places.
Differential Revision: https://reviews.llvm.org/D91199
D80526 added custom lowering to pick the si lib call on RV64, but this custom handling is only enabled when the F and D extension are both disabled. This prevents the si library call from being used for double when F is enabled but D is not.
This patch changes the behavior so we always enable the Custom hook on RV64 and decide in ReplaceNodeResults if we should emit a libcall based on whether the FP type should be softened or not.
Differential Revision: https://reviews.llvm.org/D90817
