This extends the recent MVE lane interleaving passto handle other
non-instruction leaves, for which a new shuffle is added. This helps
especially for constants and potentially for arguments.
Differential Revision: https://reviews.llvm.org/D97289
MVE does not have a single sext/zext or trunc instruction that takes the
bottom half of a vector and extends to a full width, like NEON has with
MOVL. Instead it is expected that this happens through top/bottom
instructions. So the MVE equivalent VMOVLT/B instructions take either
the even or odd elements of the input and extend them to the larger
type, producing a vector with half the number of elements each of double
the bitwidth. As there is no simple instruction for a normal extend, we
often have to expand sext/zext/trunc into a series of lane moves (or
stack loads/stores, which we do not do yet).
This pass takes vector code that starts at truncs, looks for
interconnected blobs of operations that end with sext/zext and
transforms them by adding shuffles so that the lanes are interleaved and
the MVE VMOVL/VMOVN instructions can be used. This is done pre-ISel so
that it can work across basic blocks.
This initial version of the pass just handles a limited set of
instructions, not handling constants or splats or FP, which can all come
as extensions to this base.
Differential Revision: https://reviews.llvm.org/D95804
If a WhileLoopStartLR is reverted due to calls in the preheader, we may
still be able to instead create a DoLoopStart, preserving the low
overhead loop. This adds code for that, only reverting the
WhileLoopStartR to a Br/Cmp, leaving the rest of the low overhead loop
in place.
Differential Revision: https://reviews.llvm.org/D98413
In function ConvertVPTBlocks(), it is assumed that every instruction
within a vector-predicated block is predicated. This is false for debug
instructions, used by LLVM.
Because of this, an assertion failure is reached when an input contains
debug instructions inside VPT blocks. In non-assert builds, an out of
bounds memory access took place.
The present patch properly covers the case of debug instructions.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D99075
This adds an extra pattern for inserting an f16 into a odd vector lane
via an VINS. If the dual-insert-lane pattern does not happen to apply,
this can help with some simple cases.
Differential Revision: https://reviews.llvm.org/D95471
Reuse the existing KnownBits multiplication code to handle the 'extend + multiply + extract high bits' pattern for multiply-high ops.
Noticed while looking at the codegen for D88785 / D98587 - the patch helps division-by-constant expansion code in particular, which suggests that we might have some further KnownBits div/rem cases we could handle - but this was far easier to implement.
Differential Revision: https://reviews.llvm.org/D98857
This also briefly tests a larger set of architectures than the more
exhaustive functionality tests for AArch64 and x86.
As requested in D88785
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D98339
This adjusts the place that the t2DoLoopStart reg allocation hint is
inserted, adding it in the ARMTPAndVPTOptimizaionPass in a similar place
as other tail predicated loop optimizations. This removes the need for
doing so in a custom inserter, and should make the hint more accurate,
only adding it where we expect to create a DLS (not DLSTP or WLS).
Recently we improved the lowering of low overhead loops and tail
predicated loops, but concentrated first on the DLS do style loops. This
extends those improvements over to the WLS while loops, improving the
chance of lowering them successfully. To do this the lowering has to
change a little as the instructions are terminators that produce a value
- something that needs to be treated carefully.
Lowering starts at the Hardware Loop pass, inserting a new
llvm.test.start.loop.iterations that produces both an i1 to control the
loop entry and an i32 similar to the llvm.start.loop.iterations
intrinsic added for do loops. This feeds into the loop phi, properly
gluing the values together:
%wls = call { i32, i1 } @llvm.test.start.loop.iterations.i32(i32 %div)
%wls0 = extractvalue { i32, i1 } %wls, 0
%wls1 = extractvalue { i32, i1 } %wls, 1
br i1 %wls1, label %loop.ph, label %loop.exit
...
loop:
%lsr.iv = phi i32 [ %wls0, %loop.ph ], [ %iv.next, %loop ]
..
%iv.next = call i32 @llvm.loop.decrement.reg.i32(i32 %lsr.iv, i32 1)
%cmp = icmp ne i32 %iv.next, 0
br i1 %cmp, label %loop, label %loop.exit
The llvm.test.start.loop.iterations need to be lowered through ISel
lowering as a pair of WLS and WLSSETUP nodes, which each get converted
to t2WhileLoopSetup and t2WhileLoopStart Pseudos. This helps prevent
t2WhileLoopStart from being a terminator that produces a value,
something difficult to control at that stage in the pipeline. Instead
the t2WhileLoopSetup produces the value of LR (essentially acting as a
lr = subs rn, 0), t2WhileLoopStart consumes that lr value (the Bcc).
These are then converted into a single t2WhileLoopStartLR at the same
point as t2DoLoopStartTP and t2LoopEndDec. Otherwise we revert the loop
to prevent them from progressing further in the pipeline. The
t2WhileLoopStartLR is a single instruction that takes a GPR and produces
LR, similar to the WLS instruction.
%1:gprlr = t2WhileLoopStartLR %0:rgpr, %bb.3
t2B %bb.1
...
bb.2.loop:
%2:gprlr = PHI %1:gprlr, %bb.1, %3:gprlr, %bb.2
...
%3:gprlr = t2LoopEndDec %2:gprlr, %bb.2
t2B %bb.3
The t2WhileLoopStartLR can then be treated similar to the other low
overhead loop pseudos, eventually being lowered to a WLS providing the
branches are within range.
Differential Revision: https://reviews.llvm.org/D97729
These intrinsics, not the icmp+select are the canonical form nowadays,
so we might as well directly emit them.
This should not cause any regressions, but if it does,
then then they would needed to be fixed regardless.
Note that this doesn't deal with `SCEVExpander::isHighCostExpansion()`,
but that is a pessimization, not a correctness issue.
Additionally, the non-intrinsic form has issues with undef,
see https://reviews.llvm.org/D88287#2587863
This adds some simple known bits handling for the three CSINC/NEG/INV
instructions. From the operands known bits we can compute the common
bits of the first operand and incremented/negated/inverted second
operand. The first, especially CSINC ZR, ZR, comes up fair amount in the
tests. The others are more rare so a unit test for them is added.
Differential Revision: https://reviews.llvm.org/D97788
In some rare circumstances we can be using an undef register for a
compare. When folded into a CBZ/CBNZ the undef flags are lost, leading
to machine verifier problems. This propagates the existing flags to the
new instruction.
If we insert undef using a VMOVN, we can just use the original value in
three out of the four possible combinations. Using VMOVT into a undef
vector will still require the lanes to be moved, but otherwise the
non-undef value can be used.
Rather than converting 3 signbits to bools and comparing them,
we can do bitwise logic on the whole vector and convert the
resulting sign bit to a bool at the end.
This is still a different algorithm than what we do in LegalizeDAG
through expandSADDOSSUBO. That algorithm needs to know that the
RHS of SSUBO is > 0, but that's costly when the type is split.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D97325
Currently the load/store optimizer will only fold in increments of the
same size as the load/store. This patch expands that to any legal
immediate for the post-inc instruction.
This is a recommit of 3b34b06fc5 with correctness fixes and extra
tests.
Differential Revision: https://reviews.llvm.org/D95885
This code creates 3 setccs that need to be expanded. It was
creating a sign bit test as setge X, 0 which is non-canonical.
Canonical would be setgt X, -1. This misses the special case in
IntegerExpandSetCCOperands for sign bit tests that assumes
canonical form. If we don't hit this special case we end up
with a multipart setcc instead of just checking the sign of
the high part.
To fix this I've reversed the polarity of all of the setccs to
setlt X, 0 which is canonical. The rest of the logic should
still work. This seems to produce better code on RISCV which
lacks a setgt instruction.
This probably still isn't the best code sequence we could use here.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D97181
This removes the existing patterns for inserting two lanes into an
f16/i16 vector register using VINS, instead using a DAG combine to
pattern match the same code sequences. The tablegen patterns were
already on the large side (foreach LANE = [0, 2, 4, 6]) and were not
handling all the cases they could. Moving that to a DAG combine, whilst
not less code, allows us to better control and expand the selection of
VINSs. Additionally this allows us to remove the AddedComplexity on
VCVTT.
The extra trick that this has learned in the process is to move two
adjacent lanes using a single f32 vmov, allowing some extra
inefficiencies to be removed.
Differenial Revision: https://reviews.llvm.org/D96876
Currently the load/store optimizer will only fold in increments of the
same size as the load/store. This patch expands that to any legal
immediate for the post-inc instruction.
Differential Revision: https://reviews.llvm.org/D95885
From what I can tell, a writeback is unpredictable with LR for both
loads and stores. This changes the operand from a gprnopc to a rGPR in
both cases (which I believe is essentially a NFC due to the tied-def
already being a rGPR.)
Differential Revision: https://reviews.llvm.org/D96723
Currently the findIncDecAfter will only look at the next instruction for
post-inc candidates in the load/store optimizer. This extends that to a
search through the current BB, until an instruction that modifies or
uses the increment reg is found. This allows more post-inc load/stores
and ldm/stm's to be created, especially in cases where a schedule might
move instructions further apart.
We make sure not to look any further for an SP, as that might invalidate
stack slots that are still in use.
Differential Revision: https://reviews.llvm.org/D95881
Given a floating point store from an extracted vector, with an integer
VGETLANE that already exists, storing the existing VGETLANEu directly
can be better for performance. As the value is known to already be in an
integer registers, this can help reduce fp register pressure, removed
the need for the fp extract and allows use of more integer post-inc
stores not available with vstr.
This can be a bit narrow in scope, but helps with certain biquad kernels
that store shuffled vector elements.
Differential Revision: https://reviews.llvm.org/D96159
Our current lowering of VMOVNT goes via a shuffle vector of the form
<0, N, 2, N+2, 4, N+4, ..>. That can of course also be a single input
shuffle of the form <0, 0, 2, 2, 4, 4, ..>, where we use a VMOVNT to
insert a vector into the top lanes of itself. This adds lowering of that
case, re-using the existing isVMOVNMask.
Differential Revision: https://reviews.llvm.org/D96065
With t2DoLoopDec we can be left with some extra MOV's in the preheaders
of tail predicated loops. This removes them, in the same way we remove
other dead variables.
Differential Revision: https://reviews.llvm.org/D91857
We were storing predicate registers, such as a <8 x i1>, in the opposite
order to how the rest of llvm expects. This actually turns out to be
correct for the one place that usually uses it - the
ScalarizeMaskedMemIntrin pass, but only because the pass was incorrect
itself. This fixes the order so that bits are stored in the opposite
order and bitcasts work as expected. This allows the Scalarization pass
to be fixed, as in https://reviews.llvm.org/D94765.
Differential Revision: https://reviews.llvm.org/D94867
A One-Off Identity mask is a shuffle that is mostly an identity mask
from as single source but contains a single element out-of-place, either
from a different vector or from another position in the same vector. As
opposed to lowering this via a ARMISD::BUILD_VECTOR we can generate an
extract/insert pair directly. Under ARM with individually accessible
lane elements this often becomes a simple lane move.
This also alters the LowerVECTOR_SHUFFLEUsingMovs code to use v4f32 (not
v4i32), a more natural type for lane moves.
Differential Revision: https://reviews.llvm.org/D95551
Because we mark all operations as expand for v2f64, scalar_to_vector
would end up lowering through a stack store/reload. But it is pretty
simple to implement, only inserting a D reg into an undef vector. This
helps clear up some inefficient codegen from soft calling conventions.
Differential Revision: https://reviews.llvm.org/D96153
This adds another tablegen fold that converts an i16 odd-lane-insert of
an even-lane-extract into a VINS. We extract the existing f32 value from
the destination register and VINS the new value into it. The rest of the
backend then is able to optimize the INSERT_SUBREG / COPY_TO_REGCLASS /
EXTRACT_SUBREG.
Differential Revision: https://reviews.llvm.org/D95456
This allows the peephole optimizer to know that a MVE_VMOV_to_lane_32 is
the same as an insert subreg, allowing it to optimize some redundant
lane moves.
Differential Revision: https://reviews.llvm.org/D95433
A v4i32 insert of an extract can become a simple lane move, as opposed
to round-tripping via a GPR. This adds a patterns that turns an v4i32
insert-extract pair into a EXTRACT_SUBREG/INSERT_SUBREG, with the
required COPY_TO_REGCLASS. These get better optimized into a simple lane
move by the rest of the backend.
Differential Revision: https://reviews.llvm.org/D95428
This patch adds tablegen patterns for pairs of i16/f16 insert/extracts.
If we are inserting into two adjacent vector lanes (0 and 1 for
example), we can use either a vmov;vins or vmovx;vins to insert the pair
together, avoiding a round-trip from GRP registers. This is quite a
large patterns with a number of EXTRACT_SUBREG/INSERT_SUBREG/
COPY_TO_REGCLASS nodes, but hopefully as most of those become copies all
that will be cleaned up by further optimizations.
The VINS pattern was also adjusted to allow it to represent that it is
inserting into the top half of an existing register.
Differential Revision: https://reviews.llvm.org/D95381
A DLS lr, lr instruction only moves lr to itself. It need not be emitted
on it's own to save a instruction in the loop preheader.
Differential Revision: https://reviews.llvm.org/D78916
Given a shuffle(vqdmulh(shuffle, shuffle), we can flatter the shuffles
out if they become an identity mask. This can come up during lane
interleaving, when we do that better.
Differential Revision: https://reviews.llvm.org/D94034
Under the softfp calling convention, we are often left with
VMOVRRD(extract(bitcast(build_vector(a, b, c, d)))) for the return value
of the function. These can be simplified to a,b or c,d directly,
depending on the value of the extract.
Big endian is a little different because the bitcast switches the lanes
around, meaning we end up with b,a or d,c.
Differential Revision: https://reviews.llvm.org/D94989
This adds a DAG combine for converting sext_inreg of VGetLaneu into
VGetLanes, providing the types match correctly.
Differential Revision: https://reviews.llvm.org/D95073
Under SoftFP calling conventions, we can be left with
extract(bitcast(BUILD_VECTOR(VMOVDRR(a, b), ..))) patterns that can
simplify to a or b, depending on the extract lane.
Differential Revision: https://reviews.llvm.org/D94990
The MVE VLD2/4 and VST2/4 instructions require the pointer to be aligned
to at least the size of the element type. This adds a check for that
into the ARM lowerInterleavedStore and lowerInterleavedLoad functions,
not creating the intrinsics if they are invalid for the alignment of
the load/store.
Unfortunately this is one of those bug fixes that does effect some
useful codegen, as we were able to sometimes do some nice lowering of
q15 types. But they can cause problem with low aligned pointers.
Differential Revision: https://reviews.llvm.org/D95319
David Green